[llvm] [AMDGPU] Remove verification failing cases due to bundle (PR #139868)
via llvm-commits
llvm-commits at lists.llvm.org
Wed May 14 03:01:35 PDT 2025
https://github.com/Shoreshen updated https://github.com/llvm/llvm-project/pull/139868
>From a80467ca0fa1495d72dac4afca3d9c97917de497 Mon Sep 17 00:00:00 2001
From: shore <372660931 at qq.com>
Date: Wed, 14 May 2025 17:36:38 +0800
Subject: [PATCH 1/2] remove failing cases
---
.../CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll | 23764 +---------------
1 file changed, 225 insertions(+), 23539 deletions(-)
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll
index 8d491f6bafbbc..a67679c1293bb 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll
@@ -1,10 +1,10 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
-; RUN: llc -mtriple=amdgcn -mcpu=tahiti < %s | FileCheck -check-prefix=SI %s
-; RUN: llc -mtriple=amdgcn -mcpu=tonga < %s | FileCheck -check-prefix=VI %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefix=GFX9 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-TRUE16 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-FAKE16 %s
+; RUN: llc -mtriple=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefix=SI %s
+; RUN: llc -mtriple=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=VI %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX9 %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11,GFX11-TRUE16 %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11,GFX11-FAKE16 %s
define <32 x float> @bitcast_v32i32_to_v32f32(<32 x i32> %a, i32 %b) {
; SI-LABEL: bitcast_v32i32_to_v32f32:
@@ -203181,23320 +203181,6 @@ end:
ret <128 x i8> %phi
}
-define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
-; SI-LABEL: bitcast_v128i8_to_v64i16:
-; SI: ; %bb.0:
-; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:940 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:392
-; SI-NEXT: s_waitcnt expcnt(1)
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:4
-; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:28
-; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:36
-; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:20
-; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:60
-; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:68
-; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:52
-; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:92
-; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:100
-; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:84
-; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:124
-; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:132
-; SI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:116
-; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:156
-; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:164
-; SI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:148
-; SI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:188
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:128
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 8, v14
-; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v16
-; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v12
-; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 8, v22
-; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v20
-; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 8, v30
-; SI-NEXT: v_lshlrev_b32_e32 v6, 8, v6
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 24, v8
-; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v28
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; SI-NEXT: v_lshlrev_b32_e32 v24, 24, v24
-; SI-NEXT: ; implicit-def: $vgpr43
-; SI-NEXT: ; implicit-def: $vgpr40
-; SI-NEXT: ; implicit-def: $vgpr22
-; SI-NEXT: ; implicit-def: $vgpr20
-; SI-NEXT: ; implicit-def: $vgpr56
-; SI-NEXT: ; implicit-def: $vgpr41
-; SI-NEXT: ; implicit-def: $vgpr39
-; SI-NEXT: ; implicit-def: $vgpr53
-; SI-NEXT: ; implicit-def: $vgpr50
-; SI-NEXT: ; implicit-def: $vgpr49
-; SI-NEXT: ; implicit-def: $vgpr48
-; SI-NEXT: ; implicit-def: $vgpr35
-; SI-NEXT: ; implicit-def: $vgpr55
-; SI-NEXT: ; implicit-def: $vgpr63
-; SI-NEXT: ; implicit-def: $vgpr58
-; SI-NEXT: ; implicit-def: $vgpr61
-; SI-NEXT: ; implicit-def: $vgpr52
-; SI-NEXT: ; implicit-def: $vgpr57
-; SI-NEXT: ; implicit-def: $vgpr28
-; SI-NEXT: ; implicit-def: $vgpr38
-; SI-NEXT: ; implicit-def: $vgpr46
-; SI-NEXT: ; implicit-def: $vgpr36
-; SI-NEXT: ; implicit-def: $vgpr44
-; SI-NEXT: ; kill: killed $vgpr43
-; SI-NEXT: ; implicit-def: $vgpr43
-; SI-NEXT: ; kill: killed $vgpr40
-; SI-NEXT: ; implicit-def: $vgpr40
-; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:360
-; SI-NEXT: s_waitcnt vmcnt(14)
-; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v25
-; SI-NEXT: v_lshlrev_b32_e32 v30, 24, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v7
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 24, v9
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 24, v5
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v13
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 24, v15
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 24, v11
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v19
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 24, v21
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt vmcnt(14)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:152
-; SI-NEXT: v_lshlrev_b32_e32 v3, 24, v17
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v23
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v27
-; SI-NEXT: v_lshlrev_b32_e32 v5, 24, v32
-; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:332
-; SI-NEXT: ; implicit-def: $vgpr23
-; SI-NEXT: ; implicit-def: $vgpr25
-; SI-NEXT: ; implicit-def: $vgpr32
-; SI-NEXT: ; implicit-def: $vgpr27
-; SI-NEXT: ; implicit-def: $vgpr21
-; SI-NEXT: ; kill: killed $vgpr23
-; SI-NEXT: ; implicit-def: $vgpr23
-; SI-NEXT: s_waitcnt vmcnt(3)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:160
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:96
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:88
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:64
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:56
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:112
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:144
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:140
-; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v29
-; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 8, v31
-; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:184
-; SI-NEXT: v_lshlrev_b32_e32 v29, 8, v2
-; SI-NEXT: v_lshlrev_b32_e32 v2, 8, v18
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:108
-; SI-NEXT: ; implicit-def: $vgpr18
-; SI-NEXT: ; implicit-def: $vgpr31
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:180
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:176
-; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v5, 24, v33
-; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v5, 8, v34
-; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; SI-NEXT: v_lshlrev_b32_e32 v33, 8, v10
-; SI-NEXT: ; implicit-def: $vgpr10
-; SI-NEXT: s_waitcnt vmcnt(4)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 24, v6
-; SI-NEXT: s_waitcnt vmcnt(3)
-; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:196
-; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:220
-; SI-NEXT: s_waitcnt expcnt(1)
-; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:192
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:172
-; SI-NEXT: s_waitcnt vmcnt(3)
-; SI-NEXT: v_lshlrev_b32_e32 v7, 24, v7
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:216
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:212
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:208
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 8, v8
-; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt vmcnt(3)
-; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(1)
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:228
-; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:252
-; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:224
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:204
-; SI-NEXT: s_waitcnt vmcnt(3)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 24, v6
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:248
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:244
-; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:240
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 24, v9
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 8, v7
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt vmcnt(3)
-; SI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:236
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:260
-; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:284
-; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:256
-; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 24, v6
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:280
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:276
-; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:272
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 24, v8
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 8, v7
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:928 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt vmcnt(3)
-; SI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:268
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:292
-; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:316
-; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:288
-; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 24, v6
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v34, 8, v7
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:312
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:844 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:308
-; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:304
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 24, v9
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:948 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt vmcnt(3)
-; SI-NEXT: v_lshlrev_b32_e32 v51, 24, v8
-; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(1)
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:324
-; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:348
-; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:320
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:300
-; SI-NEXT: s_waitcnt vmcnt(3)
-; SI-NEXT: v_lshlrev_b32_e32 v37, 24, v6
-; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: v_lshlrev_b32_e32 v7, 8, v7
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:924 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:344
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:956 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:340
-; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:336
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v59, 24, v9
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:960 ; 4-byte Folded Spill
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:356
-; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:380
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:352
-; SI-NEXT: ; implicit-def: $vgpr9
-; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: v_lshlrev_b32_e32 v47, 24, v6
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v8, 8, v8
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:964 ; 4-byte Folded Spill
-; SI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:376
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:372
-; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:368
-; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:364
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:388
-; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:384
-; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v2, 8, v26
-; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; SI-NEXT: ; implicit-def: $vgpr26
-; SI-NEXT: s_waitcnt vmcnt(6)
-; SI-NEXT: v_lshlrev_b32_e32 v14, 24, v14
-; SI-NEXT: s_waitcnt vmcnt(3)
-; SI-NEXT: v_lshlrev_b32_e32 v19, 24, v6
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:12
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:8
-; SI-NEXT: v_lshlrev_b32_e32 v62, 8, v16
-; SI-NEXT: ; implicit-def: $vgpr16
-; SI-NEXT: ; kill: killed $vgpr16
-; SI-NEXT: ; implicit-def: $vgpr16
-; SI-NEXT: ; kill: killed $vgpr16
-; SI-NEXT: ; implicit-def: $vgpr16
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v2, 8, v2
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:120
-; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:44
-; SI-NEXT: v_lshlrev_b32_e32 v1, 8, v1
-; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v2, 8, v2
-; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:76
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:72
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v1, 8, v3
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; SI-NEXT: ; implicit-def: $vgpr3
-; SI-NEXT: s_waitcnt vmcnt(3)
-; SI-NEXT: v_lshlrev_b32_e32 v2, 8, v2
-; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:48
-; SI-NEXT: v_lshlrev_b32_e32 v1, 8, v4
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v1, 8, v5
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:920 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v1, 8, v11
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:952 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v1, 8, v12
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:972 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:32
-; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr5
-; SI-NEXT: ; implicit-def: $vgpr4
-; SI-NEXT: ; implicit-def: $vgpr11
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:24
-; SI-NEXT: v_lshlrev_b32_e32 v15, 8, v15
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; SI-NEXT: ; implicit-def: $vgpr6
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:80
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:40
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:16
-; SI-NEXT: v_lshlrev_b32_e32 v12, 8, v13
-; SI-NEXT: ; implicit-def: $vgpr13
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:104
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:136
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:168
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:200
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:232
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:264
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:296
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:944 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:328
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:968 ; 4-byte Folded Spill
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB96_2
-; SI-NEXT: ; %bb.1: ; %cmp.false
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:972 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(4)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: s_waitcnt vmcnt(3)
-; SI-NEXT: v_or_b32_e32 v20, v1, v2
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v18, v2, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v3, v2, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v1, v1, v2
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v5, v2, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v9, v2, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v41, v1, v2
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: v_or_b32_e32 v55, v24, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; SI-NEXT: v_mov_b32_e32 v24, v33
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: v_or_b32_e32 v13, v2, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v1, v1, v2
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: v_or_b32_e32 v26, v30, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: v_or_b32_e32 v21, v2, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v32, v1, v2
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v28, v2, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v27, v2, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v1, v1, v2
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v35, v2, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v31, v2, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v63, v1, v2
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v48, v2, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v50, v2, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v11, v1, v2
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v2, v2, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: v_or_b32_e32 v53, v4, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v6, v1, v4
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v1, v4, v1
-; SI-NEXT: v_and_b32_e32 v4, 0xff, v17
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v43, v10, v4
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v44, v4, v10
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v16, v10, v4
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v46, v10, v4
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v56, v4, v10
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v17, v10, v4
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v57, v10, v4
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v61, v4, v10
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v30, v10, v4
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v49, v10, v4
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v4, v4, v10
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v22, 0xff, v10
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload
-; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v23, v10, v22
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v22, 0xff, v10
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:948 ; 4-byte Folded Reload
-; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v33, v10, v22
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v22, 0xff, v10
-; SI-NEXT: v_mov_b32_e32 v10, v29
-; SI-NEXT: v_or_b32_e32 v29, v22, v34
-; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v22, 0xff, v22
-; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; SI-NEXT: v_or_b32_e32 v34, v37, v22
-; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v22, 0xff, v22
-; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; SI-NEXT: v_or_b32_e32 v37, v51, v22
-; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v22, 0xff, v22
-; SI-NEXT: v_or_b32_e32 v51, v22, v7
-; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v22, 0xff, v22
-; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; SI-NEXT: v_or_b32_e32 v7, v47, v7
-; SI-NEXT: v_or_b32_e32 v47, v59, v22
-; SI-NEXT: v_and_b32_e32 v22, 0xff, v54
-; SI-NEXT: v_or_b32_e32 v54, v22, v8
-; SI-NEXT: v_and_b32_e32 v8, 0xff, v60
-; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT: v_or_b32_e32 v8, v19, v8
-; SI-NEXT: v_and_b32_e32 v19, 0xff, v42
-; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; SI-NEXT: v_or_b32_e32 v14, v14, v19
-; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v19, 0xff, v19
-; SI-NEXT: v_or_b32_e32 v19, v19, v10
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v22, 0xff, v10
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; SI-NEXT: v_or_b32_e32 v25, v22, v24
-; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v22, 0xff, v10
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v39, v22, v10
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v22, 0xff, v10
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v42, v22, v10
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v22, 0xff, v10
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v59, v22, v10
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v22, 0xff, v10
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v60, v22, v10
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v22, 0xff, v10
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v36, v22, v10
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v22, 0xff, v10
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v10, v22, v10
-; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v22, 0xff, v22
-; SI-NEXT: v_or_b32_e32 v38, v22, v24
-; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v22, 0xff, v22
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v58, v22, v24
-; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v22, 0xff, v22
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v52, v22, v24
-; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v22, 0xff, v22
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v24, v22, v24
-; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v22, 0xff, v22
-; SI-NEXT: v_or_b32_e32 v40, v22, v40
-; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:944 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v22, 0xff, v22
-; SI-NEXT: v_or_b32_e32 v12, v22, v12
-; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:968 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v22, 0xff, v22
-; SI-NEXT: v_or_b32_e32 v15, v22, v15
-; SI-NEXT: v_and_b32_e32 v22, 0xff, v45
-; SI-NEXT: v_or_b32_e32 v45, v22, v62
-; SI-NEXT: v_mov_b32_e32 v62, v18
-; SI-NEXT: v_or_b32_e32 v20, v20, v62
-; SI-NEXT: v_or_b32_e32 v22, v19, v3
-; SI-NEXT: v_alignbit_b32 v3, v20, v3, 16
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; SI-NEXT: v_mov_b32_e32 v19, v5
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT: v_or_b32_e32 v5, v3, v19
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v25
-; SI-NEXT: v_or_b32_e32 v25, v3, v9
-; SI-NEXT: v_alignbit_b32 v3, v5, v9, 16
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v41
-; SI-NEXT: v_or_b32_e32 v41, v3, v55
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v39
-; SI-NEXT: v_or_b32_e32 v18, v3, v13
-; SI-NEXT: v_alignbit_b32 v3, v41, v13, 16
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT: v_or_b32_e32 v13, v3, v26
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v42
-; SI-NEXT: v_or_b32_e32 v39, v3, v21
-; SI-NEXT: v_alignbit_b32 v3, v13, v21, 16
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v32
-; SI-NEXT: v_or_b32_e32 v32, v3, v28
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v59
-; SI-NEXT: v_or_b32_e32 v42, v3, v27
-; SI-NEXT: v_alignbit_b32 v3, v32, v27, 16
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; SI-NEXT: ; implicit-def: $vgpr59
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT: v_or_b32_e32 v27, v3, v35
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v60
-; SI-NEXT: v_or_b32_e32 v3, v3, v31
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v3, v27, v31, 16
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v63
-; SI-NEXT: v_or_b32_e32 v21, v3, v48
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v36
-; SI-NEXT: v_or_b32_e32 v31, v3, v50
-; SI-NEXT: v_alignbit_b32 v3, v21, v50, 16
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v11
-; SI-NEXT: v_or_b32_e32 v11, v3, v2
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v10
-; SI-NEXT: v_or_b32_e32 v3, v3, v53
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v3, v11, v53, 16
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v6
-; SI-NEXT: v_and_b32_e32 v6, 0xffff, v38
-; SI-NEXT: v_or_b32_e32 v3, v3, v1
-; SI-NEXT: v_or_b32_e32 v6, v6, v43
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v6, v3, v43, 16
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v6, 0xffff, v44
-; SI-NEXT: v_or_b32_e32 v9, v6, v16
-; SI-NEXT: v_and_b32_e32 v6, 0xffff, v58
-; SI-NEXT: v_or_b32_e32 v6, v6, v46
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v6, v9, v46, 16
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v6, 0xffff, v56
-; SI-NEXT: v_or_b32_e32 v63, v6, v17
-; SI-NEXT: v_and_b32_e32 v6, 0xffff, v52
-; SI-NEXT: v_or_b32_e32 v6, v6, v57
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v6, v63, v57, 16
-; SI-NEXT: v_or_b32_e32 v57, v4, v23
-; SI-NEXT: v_and_b32_e32 v4, 0xffff, v40
-; SI-NEXT: v_or_b32_e32 v52, v4, v33
-; SI-NEXT: v_alignbit_b32 v4, v57, v33, 16
-; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v4, 0xffff, v29
-; SI-NEXT: v_or_b32_e32 v46, v4, v34
-; SI-NEXT: v_and_b32_e32 v4, 0xffff, v12
-; SI-NEXT: v_or_b32_e32 v38, v4, v37
-; SI-NEXT: v_alignbit_b32 v4, v46, v37, 16
-; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v4, 0xffff, v51
-; SI-NEXT: v_or_b32_e32 v44, v4, v7
-; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v7
-; SI-NEXT: v_lshrrev_b32_e32 v7, 16, v8
-; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: v_and_b32_e32 v4, 0xffff, v15
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: v_or_b32_e32 v36, v4, v47
-; SI-NEXT: v_alignbit_b32 v4, v44, v47, 16
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v4, 0xffff, v54
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: v_or_b32_e32 v43, v4, v8
-; SI-NEXT: v_and_b32_e32 v4, 0xffff, v45
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: v_or_b32_e32 v12, v4, v14
-; SI-NEXT: v_alignbit_b32 v4, v43, v14, 16
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v4, 16, v28
-; SI-NEXT: v_lshrrev_b32_e32 v28, 16, v23
-; SI-NEXT: v_mov_b32_e32 v23, v12
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr12
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr12
-; SI-NEXT: ; implicit-def: $vgpr12
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr12
-; SI-NEXT: ; implicit-def: $vgpr12
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr12
-; SI-NEXT: ; implicit-def: $vgpr12
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr12
-; SI-NEXT: ; implicit-def: $vgpr12
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr12
-; SI-NEXT: ; implicit-def: $vgpr12
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr12
-; SI-NEXT: ; implicit-def: $vgpr12
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr12
-; SI-NEXT: ; implicit-def: $vgpr12
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v6, 0xffff, v61
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr12
-; SI-NEXT: ; implicit-def: $vgpr12
-; SI-NEXT: v_or_b32_e32 v61, v6, v30
-; SI-NEXT: v_and_b32_e32 v6, 0xffff, v24
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr12
-; SI-NEXT: ; implicit-def: $vgpr12
-; SI-NEXT: v_or_b32_e32 v58, v6, v49
-; SI-NEXT: v_alignbit_b32 v6, v61, v49, 16
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; kill: killed $vgpr12
-; SI-NEXT: ; implicit-def: $vgpr12
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v6, 16, v62
-; SI-NEXT: v_lshrrev_b32_e32 v56, 16, v19
-; SI-NEXT: v_lshrrev_b32_e32 v10, 16, v55
-; SI-NEXT: v_lshrrev_b32_e32 v53, 16, v26
-; SI-NEXT: v_mov_b32_e32 v26, v42
-; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v35
-; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v48
-; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v2
-; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v1
-; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v16
-; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v17
-; SI-NEXT: v_lshrrev_b32_e32 v2, 16, v30
-; SI-NEXT: v_lshrrev_b32_e32 v16, 16, v34
-; SI-NEXT: ; implicit-def: $vgpr17
-; SI-NEXT: ; implicit-def: $vgpr45
-; SI-NEXT: ; implicit-def: $vgpr42
-; SI-NEXT: ; implicit-def: $vgpr54
-; SI-NEXT: ; implicit-def: $vgpr60
-; SI-NEXT: ; implicit-def: $vgpr24
-; SI-NEXT: ; implicit-def: $vgpr30
-; SI-NEXT: ; kill: killed $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr34
-; SI-NEXT: ; implicit-def: $vgpr37
-; SI-NEXT: ; implicit-def: $vgpr51
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr47
-; SI-NEXT: ; implicit-def: $vgpr8
-; SI-NEXT: ; implicit-def: $vgpr19
-; SI-NEXT: ; implicit-def: $vgpr14
-; SI-NEXT: ; implicit-def: $vgpr29
-; SI-NEXT: ; implicit-def: $vgpr33
-; SI-NEXT: ; kill: killed $vgpr12
-; SI-NEXT: ; implicit-def: $vgpr12
-; SI-NEXT: ; implicit-def: $vgpr15
-; SI-NEXT: ; implicit-def: $vgpr62
-; SI-NEXT: .LBB96_2: ; %Flow
-; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB96_4
-; SI-NEXT: ; %bb.3: ; %cmp.true
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v45
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v42
-; SI-NEXT: v_or_b32_e32 v1, v62, v1
-; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT: v_add_i32_e32 v1, vcc, 0x300, v1
-; SI-NEXT: v_or_b32_e32 v2, v14, v2
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v1, v2, v1
-; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v54
-; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v60
-; SI-NEXT: v_or_b32_e32 v2, v8, v2
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: s_movk_i32 s6, 0x300
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_add_i32_e32 v2, vcc, s6, v2
-; SI-NEXT: v_or_b32_e32 v3, v19, v3
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v2, v3, v2
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:968 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
-; SI-NEXT: s_mov_b32 s7, 0x3000000
-; SI-NEXT: v_add_i32_e32 v43, vcc, s7, v2
-; SI-NEXT: v_mov_b32_e32 v28, v24
-; SI-NEXT: v_add_i32_e32 v1, vcc, s7, v1
-; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:972 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:948 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(14)
-; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; SI-NEXT: v_or_b32_e32 v3, v15, v3
-; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_add_i32_e32 v3, vcc, s6, v3
-; SI-NEXT: v_or_b32_e32 v4, v59, v4
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; SI-NEXT: v_and_b32_e32 v5, 0xff, v5
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: v_or_b32_e32 v5, v47, v5
-; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; SI-NEXT: v_and_b32_e32 v6, 0xff, v6
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; SI-NEXT: v_or_b32_e32 v6, v51, v6
-; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT: v_or_b32_e32 v4, v7, v4
-; SI-NEXT: v_add_i32_e32 v4, vcc, s6, v4
-; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:944 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v44, vcc, s7, v4
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v44
-; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; SI-NEXT: v_and_b32_e32 v5, 0xff, v5
-; SI-NEXT: v_or_b32_e32 v5, v12, v5
-; SI-NEXT: v_add_i32_e32 v5, vcc, s6, v5
-; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; SI-NEXT: v_or_b32_e32 v5, v6, v5
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; SI-NEXT: v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; SI-NEXT: v_or_b32_e32 v7, v37, v7
-; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v35, vcc, s7, v5
-; SI-NEXT: v_mov_b32_e32 v38, v35
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6
-; SI-NEXT: v_and_b32_e32 v6, 0xff, v6
-; SI-NEXT: v_or_b32_e32 v6, v34, v6
-; SI-NEXT: v_add_i32_e32 v6, vcc, s6, v6
-; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; SI-NEXT: v_or_b32_e32 v6, v7, v6
-; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v46, vcc, s7, v6
-; SI-NEXT: v_add_i32_e32 v34, vcc, s7, v3
-; SI-NEXT: v_mov_b32_e32 v36, v34
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; SI-NEXT: v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT: v_or_b32_e32 v7, v8, v7
-; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v7, vcc, s6, v7
-; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; SI-NEXT: v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT: v_or_b32_e32 v8, v9, v8
-; SI-NEXT: v_or_b32_e32 v7, v8, v7
-; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v6, vcc, s7, v7
-; SI-NEXT: v_mov_b32_e32 v52, v6
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8
-; SI-NEXT: v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v8, v9, v8
-; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v8, vcc, s6, v8
-; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; SI-NEXT: v_and_b32_e32 v9, 0xff, v9
-; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; SI-NEXT: v_or_b32_e32 v9, v10, v9
-; SI-NEXT: v_or_b32_e32 v8, v9, v8
-; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v57, vcc, s7, v8
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v9
-; SI-NEXT: v_and_b32_e32 v9, 0xff, v9
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v9, v10, v9
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v9, vcc, s6, v9
-; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; SI-NEXT: v_and_b32_e32 v10, 0xff, v10
-; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; SI-NEXT: v_or_b32_e32 v10, v11, v10
-; SI-NEXT: v_or_b32_e32 v9, v10, v9
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v7, vcc, s7, v9
-; SI-NEXT: v_mov_b32_e32 v58, v7
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10
-; SI-NEXT: v_and_b32_e32 v10, 0xff, v10
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v10, v11, v10
-; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v10, vcc, s6, v10
-; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; SI-NEXT: v_and_b32_e32 v11, 0xff, v11
-; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; SI-NEXT: v_or_b32_e32 v11, v12, v11
-; SI-NEXT: v_or_b32_e32 v10, v11, v10
-; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v61, vcc, s7, v10
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; SI-NEXT: v_and_b32_e32 v11, 0xff, v11
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v11, v12, v11
-; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v11, vcc, s6, v11
-; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; SI-NEXT: v_and_b32_e32 v12, 0xff, v12
-; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; SI-NEXT: v_or_b32_e32 v12, v13, v12
-; SI-NEXT: v_or_b32_e32 v11, v12, v11
-; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v8, vcc, s7, v11
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12
-; SI-NEXT: v_and_b32_e32 v12, 0xff, v12
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v12, v13, v12
-; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v12, vcc, s6, v12
-; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; SI-NEXT: v_and_b32_e32 v13, 0xff, v13
-; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; SI-NEXT: v_or_b32_e32 v13, v14, v13
-; SI-NEXT: v_or_b32_e32 v12, v13, v12
-; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v63, vcc, s7, v12
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v13
-; SI-NEXT: v_and_b32_e32 v13, 0xff, v13
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v13, v14, v13
-; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v13, vcc, s6, v13
-; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT: v_and_b32_e32 v14, 0xff, v14
-; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; SI-NEXT: v_or_b32_e32 v14, v15, v14
-; SI-NEXT: v_or_b32_e32 v13, v14, v13
-; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v10, vcc, s7, v13
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT: v_and_b32_e32 v14, 0xff, v14
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v14, v15, v14
-; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v14, vcc, s6, v14
-; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; SI-NEXT: v_and_b32_e32 v15, 0xff, v15
-; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; SI-NEXT: v_or_b32_e32 v15, v16, v15
-; SI-NEXT: v_or_b32_e32 v14, v15, v14
-; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v9, vcc, s7, v14
-; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v9
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v15
-; SI-NEXT: v_and_b32_e32 v15, 0xff, v15
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v15, v16, v15
-; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v17
-; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v16, 0xff, v16
-; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; SI-NEXT: v_add_i32_e32 v15, vcc, s6, v15
-; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v16, v17, v16
-; SI-NEXT: v_or_b32_e32 v15, v16, v15
-; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v12, vcc, s7, v15
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16
-; SI-NEXT: v_and_b32_e32 v16, 0xff, v16
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v16, v17, v16
-; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v16, vcc, s6, v16
-; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v17
-; SI-NEXT: v_and_b32_e32 v17, 0xff, v17
-; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; SI-NEXT: v_or_b32_e32 v17, v18, v17
-; SI-NEXT: v_or_b32_e32 v16, v17, v16
-; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v3, vcc, s7, v16
-; SI-NEXT: v_lshrrev_b32_e32 v16, 16, v46
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v17
-; SI-NEXT: v_and_b32_e32 v17, 0xff, v17
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v17, v18, v17
-; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v17, vcc, s6, v17
-; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18
-; SI-NEXT: v_and_b32_e32 v18, 0xff, v18
-; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; SI-NEXT: v_or_b32_e32 v18, v19, v18
-; SI-NEXT: v_or_b32_e32 v17, v18, v17
-; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v14, vcc, s7, v17
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18
-; SI-NEXT: v_and_b32_e32 v18, 0xff, v18
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v18, v19, v18
-; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v18, vcc, s6, v18
-; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v19
-; SI-NEXT: v_and_b32_e32 v19, 0xff, v19
-; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; SI-NEXT: v_or_b32_e32 v19, v20, v19
-; SI-NEXT: v_or_b32_e32 v18, v19, v18
-; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v11, vcc, s7, v18
-; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v11
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v19
-; SI-NEXT: v_and_b32_e32 v19, 0xff, v19
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v19, v20, v19
-; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v19, vcc, s6, v19
-; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
-; SI-NEXT: v_and_b32_e32 v20, 0xff, v20
-; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; SI-NEXT: v_or_b32_e32 v20, v21, v20
-; SI-NEXT: v_or_b32_e32 v19, v20, v19
-; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v31, vcc, s7, v19
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
-; SI-NEXT: v_and_b32_e32 v20, 0xff, v20
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v20, v21, v20
-; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v20, vcc, s6, v20
-; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
-; SI-NEXT: v_and_b32_e32 v21, 0xff, v21
-; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; SI-NEXT: v_or_b32_e32 v21, v22, v21
-; SI-NEXT: v_or_b32_e32 v20, v21, v20
-; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
-; SI-NEXT: v_and_b32_e32 v21, 0xff, v21
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v21, v22, v21
-; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v21, vcc, s6, v21
-; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v22, vcc, 3, v22
-; SI-NEXT: v_and_b32_e32 v22, 0xff, v22
-; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; SI-NEXT: v_or_b32_e32 v22, v23, v22
-; SI-NEXT: v_or_b32_e32 v22, v22, v21
-; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v15, vcc, s7, v22
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
-; SI-NEXT: v_and_b32_e32 v21, 0xff, v21
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v21, v23, v21
-; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v21, vcc, s6, v21
-; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v23, vcc, 3, v23
-; SI-NEXT: v_and_b32_e32 v23, 0xff, v23
-; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; SI-NEXT: v_or_b32_e32 v23, v25, v23
-; SI-NEXT: v_or_b32_e32 v23, v23, v21
-; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v27, vcc, s7, v23
-; SI-NEXT: v_mov_b32_e32 v23, v1
-; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v27
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
-; SI-NEXT: v_and_b32_e32 v21, 0xff, v21
-; SI-NEXT: v_or_b32_e32 v24, v24, v21
-; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v24
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
-; SI-NEXT: v_and_b32_e32 v21, 0xff, v21
-; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; SI-NEXT: v_or_b32_e32 v25, v25, v21
-; SI-NEXT: v_or_b32_e32 v1, v25, v1
-; SI-NEXT: v_add_i32_e32 v26, vcc, s7, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v21, vcc, s7, v20
-; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v21
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: v_or_b32_e32 v1, v2, v1
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT: v_or_b32_e32 v2, v4, v2
-; SI-NEXT: v_or_b32_e32 v1, v2, v1
-; SI-NEXT: v_add_i32_e32 v32, vcc, s7, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_or_b32_e32 v1, v2, v1
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT: v_or_b32_e32 v2, v4, v2
-; SI-NEXT: v_or_b32_e32 v1, v2, v1
-; SI-NEXT: v_add_i32_e32 v39, vcc, s7, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_or_b32_e32 v1, v2, v1
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT: v_or_b32_e32 v2, v30, v2
-; SI-NEXT: v_or_b32_e32 v1, v2, v1
-; SI-NEXT: v_add_i32_e32 v13, vcc, s7, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; SI-NEXT: v_lshrrev_b32_e32 v53, 16, v13
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v1, v2, v1
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT: v_or_b32_e32 v2, v4, v2
-; SI-NEXT: v_or_b32_e32 v1, v2, v1
-; SI-NEXT: v_add_i32_e32 v18, vcc, s7, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_or_b32_e32 v1, v2, v1
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT: v_or_b32_e32 v2, v28, v2
-; SI-NEXT: v_or_b32_e32 v1, v2, v1
-; SI-NEXT: v_add_i32_e32 v41, vcc, s7, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; SI-NEXT: v_lshrrev_b32_e32 v28, 16, v57
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; SI-NEXT: v_or_b32_e32 v1, v33, v1
-; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v2, v4, v2
-; SI-NEXT: v_or_b32_e32 v1, v2, v1
-; SI-NEXT: v_add_i32_e32 v25, vcc, s7, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_or_b32_e32 v1, v2, v1
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT: v_or_b32_e32 v2, v4, v2
-; SI-NEXT: v_or_b32_e32 v1, v2, v1
-; SI-NEXT: v_add_i32_e32 v5, vcc, s7, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; SI-NEXT: v_lshrrev_b32_e32 v56, 16, v5
-; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; SI-NEXT: v_or_b32_e32 v1, v29, v1
-; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v2, v4, v2
-; SI-NEXT: v_or_b32_e32 v1, v2, v1
-; SI-NEXT: v_add_i32_e32 v22, vcc, s7, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v1, v2, v1
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT: v_or_b32_e32 v2, v4, v2
-; SI-NEXT: v_or_b32_e32 v1, v2, v1
-; SI-NEXT: v_add_i32_e32 v20, vcc, s7, v1
-; SI-NEXT: v_alignbit_b32 v1, v20, v22, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v5, v25, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v41, v18, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v13, v39, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v32, v26, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(1)
-; SI-NEXT: v_alignbit_b32 v1, v27, v15, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v21, v31, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(1)
-; SI-NEXT: v_alignbit_b32 v1, v11, v14, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(1)
-; SI-NEXT: v_alignbit_b32 v1, v3, v12, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(1)
-; SI-NEXT: v_alignbit_b32 v1, v9, v10, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(1)
-; SI-NEXT: v_alignbit_b32 v1, v63, v8, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v61, v7, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v57, v6, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v46, v35, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v44, v34, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v43, v23, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
-; SI-NEXT: v_lshrrev_b32_e32 v6, 16, v20
-; SI-NEXT: v_lshrrev_b32_e32 v10, 16, v41
-; SI-NEXT: v_lshrrev_b32_e32 v4, 16, v32
-; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v3
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v63
-; SI-NEXT: v_lshrrev_b32_e32 v2, 16, v61
-; SI-NEXT: v_lshrrev_b32_e32 v7, 16, v43
-; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill
-; SI-NEXT: .LBB96_4: ; %end
-; SI-NEXT: s_or_b64 exec, exec, s[4:5]
-; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:848 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v7, 0xffff, v22
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT: v_or_b32_e32 v7, v7, v8
-; SI-NEXT: buffer_store_dword v7, v0, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v7, 0xffff, v20
-; SI-NEXT: v_or_b32_e32 v6, v7, v6
-; SI-NEXT: v_add_i32_e32 v7, vcc, 4, v0
-; SI-NEXT: buffer_store_dword v6, v7, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v6, 0xffff, v25
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; SI-NEXT: v_or_b32_e32 v6, v6, v7
-; SI-NEXT: v_add_i32_e32 v7, vcc, 8, v0
-; SI-NEXT: buffer_store_dword v6, v7, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v56
-; SI-NEXT: v_or_b32_e32 v5, v5, v6
-; SI-NEXT: v_add_i32_e32 v6, vcc, 12, v0
-; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v5, 0xffff, v18
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; SI-NEXT: v_or_b32_e32 v5, v5, v6
-; SI-NEXT: v_add_i32_e32 v6, vcc, 16, v0
-; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v5, 0xffff, v41
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v10
-; SI-NEXT: v_or_b32_e32 v5, v5, v6
-; SI-NEXT: v_add_i32_e32 v6, vcc, 20, v0
-; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v5, 0xffff, v39
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; SI-NEXT: v_or_b32_e32 v5, v5, v6
-; SI-NEXT: v_add_i32_e32 v6, vcc, 24, v0
-; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v5, 0xffff, v13
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v53
-; SI-NEXT: v_or_b32_e32 v5, v5, v6
-; SI-NEXT: v_add_i32_e32 v6, vcc, 28, v0
-; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:864 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v5, 0xffff, v26
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; SI-NEXT: v_or_b32_e32 v5, v5, v6
-; SI-NEXT: v_add_i32_e32 v6, vcc, 32, v0
-; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v5, 0xffff, v32
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: v_add_i32_e32 v5, vcc, 36, v0
-; SI-NEXT: buffer_store_dword v4, v5, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:868 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: v_or_b32_e32 v4, v4, v5
-; SI-NEXT: v_add_i32_e32 v5, vcc, 40, v0
-; SI-NEXT: buffer_store_dword v4, v5, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v4, 0xffff, v27
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v50
-; SI-NEXT: v_or_b32_e32 v4, v4, v5
-; SI-NEXT: v_add_i32_e32 v5, vcc, 44, v0
-; SI-NEXT: buffer_store_dword v4, v5, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v4, 0xffff, v31
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: v_or_b32_e32 v4, v4, v5
-; SI-NEXT: v_add_i32_e32 v5, vcc, 48, v0
-; SI-NEXT: buffer_store_dword v4, v5, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v4, 0xffff, v21
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v49
-; SI-NEXT: v_or_b32_e32 v4, v4, v5
-; SI-NEXT: v_add_i32_e32 v5, vcc, 52, v0
-; SI-NEXT: buffer_store_dword v4, v5, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: v_or_b32_e32 v4, v4, v5
-; SI-NEXT: v_add_i32_e32 v5, vcc, 56, v0
-; SI-NEXT: buffer_store_dword v4, v5, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v4, 0xffff, v11
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v48
-; SI-NEXT: v_or_b32_e32 v4, v4, v5
-; SI-NEXT: v_add_i32_e32 v5, vcc, 60, v0
-; SI-NEXT: buffer_store_dword v4, v5, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: v_or_b32_e32 v4, v4, v5
-; SI-NEXT: v_add_i32_e32 v5, vcc, 64, v0
-; SI-NEXT: buffer_store_dword v4, v5, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v35
-; SI-NEXT: v_or_b32_e32 v3, v3, v4
-; SI-NEXT: v_add_i32_e32 v4, vcc, 0x44, v0
-; SI-NEXT: buffer_store_dword v3, v4, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:884 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v3, v3, v4
-; SI-NEXT: v_add_i32_e32 v4, vcc, 0x48, v0
-; SI-NEXT: buffer_store_dword v3, v4, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v9
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v55
-; SI-NEXT: v_or_b32_e32 v3, v3, v4
-; SI-NEXT: v_add_i32_e32 v4, vcc, 0x4c, v0
-; SI-NEXT: buffer_store_dword v3, v4, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v3, v3, v4
-; SI-NEXT: v_add_i32_e32 v4, vcc, 0x50, v0
-; SI-NEXT: buffer_store_dword v3, v4, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v63
-; SI-NEXT: v_or_b32_e32 v1, v3, v1
-; SI-NEXT: v_add_i32_e32 v3, vcc, 0x54, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v58
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 0x58, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v61
-; SI-NEXT: v_or_b32_e32 v1, v1, v2
-; SI-NEXT: v_add_i32_e32 v2, vcc, 0x5c, v0
-; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v52
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT: v_or_b32_e32 v1, v1, v2
-; SI-NEXT: v_add_i32_e32 v2, vcc, 0x60, v0
-; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v57
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v28
-; SI-NEXT: v_or_b32_e32 v1, v1, v2
-; SI-NEXT: v_add_i32_e32 v2, vcc, 0x64, v0
-; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:900 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v38
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT: v_or_b32_e32 v1, v1, v2
-; SI-NEXT: v_add_i32_e32 v2, vcc, 0x68, v0
-; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v46
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v16
-; SI-NEXT: v_or_b32_e32 v1, v1, v2
-; SI-NEXT: v_add_i32_e32 v2, vcc, 0x6c, v0
-; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v36
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT: v_or_b32_e32 v1, v1, v2
-; SI-NEXT: v_add_i32_e32 v2, vcc, 0x70, v0
-; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v44
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v40
-; SI-NEXT: v_or_b32_e32 v1, v1, v2
-; SI-NEXT: v_add_i32_e32 v2, vcc, 0x74, v0
-; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v23
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT: v_or_b32_e32 v1, v1, v2
-; SI-NEXT: v_add_i32_e32 v2, vcc, 0x78, v0
-; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:936 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v43
-; SI-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT: v_or_b32_e32 v1, v1, v2
-; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; SI-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 .LBB96_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: .LBB96_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB96_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: .LBB96_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 .LBB96_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: .LBB96_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB96_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: .LBB96_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-TRUE16-LABEL: bitcast_v128i8_to_v64i16:
-; GFX11-TRUE16: ; %bb.0:
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v31, off, s32 offset:384
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v39, off, s32 offset:380
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:376
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:372
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v150, off, s32 offset:368
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v39, off, s32 offset:364
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v150, off, s32 offset:360
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v50, off, s32 offset:356
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v147, off, s32 offset:352
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:348
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v149, off, s32 offset:344
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v49, off, s32 offset:340
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v149, off, s32 offset:336
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:332
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v148, off, s32 offset:328
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v48, off, s32 offset:324
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v145, off, s32 offset:320
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:316
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v148, off, s32 offset:312
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v48, off, s32 offset:308
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v147, off, s32 offset:304
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:300
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v146, off, s32 offset:296
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:292
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v146, off, s32 offset:288
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:284
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v145, off, s32 offset:280
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:276
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v134, off, s32 offset:272
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:268
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v144, off, s32 offset:264
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:260
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v144, off, s32 offset:256
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:252
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v135, off, s32 offset:248
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:244
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v132, off, s32 offset:240
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:236
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v135, off, s32 offset:232
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:228
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v134, off, s32 offset:224
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:220
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v133, off, s32 offset:216
-; GFX11-TRUE16-NEXT: scratch_load_b32 v160, off, s32 offset:388
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v101, off, s32
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v103, off, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v113, off, s32 offset:16
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v113, off, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v114, off, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v114, off, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v115, off, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v116, off, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v116, off, s32 offset:64
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v117, off, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v117, off, s32 offset:80
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v118, off, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v118, off, s32 offset:96
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v119, off, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v119, off, s32 offset:112
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v128, off, s32 offset:120
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v128, off, s32 offset:128
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v129, off, s32 offset:136
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v129, off, s32 offset:144
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v130, off, s32 offset:152
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v130, off, s32 offset:160
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v131, off, s32 offset:168
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v131, off, s32 offset:176
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v132, off, s32 offset:184
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v133, off, s32 offset:192
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v151, off, s32 offset:200
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v151, off, s32 offset:208
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:212
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:204
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:196
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:188
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:180
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v86, off, s32 offset:172
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v96, off, s32 offset:164
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:156
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:148
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32 offset:140
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v84, off, s32 offset:132
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v86, off, s32 offset:124
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:116
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:108
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:100
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v82, off, s32 offset:92
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v83, off, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v68, off, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v70, off, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:20
-; GFX11-TRUE16-NEXT: s_clause 0x1
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:4
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v30.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v28.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.h, v26.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.h, v24.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v22.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v20.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v18.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.h, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.h, v14.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.h, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.h, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v81.l, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v81.h, 8, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.h, 8, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.l, 8, v7.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.h, 8, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.l, 8, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v98.h, 8, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.l, 8, v15.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.h, 8, v17.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.l, 8, v19.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.h, 8, v21.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.l, 8, v23.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.h, 8, v25.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.l, 8, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v112.l, 8, v29.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(62)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v150.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.l, 8, v150.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.h, 8, v147.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v149.h, 8, v149.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v149.l, 8, v149.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.h, 8, v148.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.l, 8, v145.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.l, 8, v148.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v147.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v146.h, 8, v146.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v146.l, 8, v146.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.h, 8, v145.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v134.h, 8, v134.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v144.h, 8, v144.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v144.l, 8, v144.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.h, 8, v135.h
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(61)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.l, 8, v132.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(59)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.l, 8, v135.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(57)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v134.l, 8, v134.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(55)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.h, 8, v133.h
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(54)
-; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v160
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(53)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v101.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(52)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v112.h, 8, v103.h
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(51)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v113.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(50)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.h, 8, v113.h
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(49)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.h, 8, v114.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(48)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v115.l, 8, v114.h
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(47)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v115.h, 8, v115.h
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(46)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.l, 8, v116.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(45)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.h, 8, v116.h
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(44)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.l, 8, v117.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(43)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.l, 8, v117.h
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(42)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v118.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(41)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.h, 8, v118.h
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(40)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v119.l, 8, v119.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(39)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.h, 8, v119.h
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(38)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v119.h, 8, v128.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(37)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.l, 8, v128.h
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(36)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.h, 8, v129.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(35)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.l, 8, v129.h
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(34)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.l, 8, v130.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(33)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.h, 8, v130.h
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(32)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.h, 8, v131.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(31)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v131.l, 8, v131.h
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(30)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v131.h, 8, v132.h
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(29)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.h, 8, v133.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(28)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.h, 8, v151.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(27)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.l, 8, v151.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.l, 8, v31.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v31.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB96_3
-; GFX11-TRUE16-NEXT: ; %bb.1: ; %Flow
-; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB96_4
-; GFX11-TRUE16-NEXT: .LBB96_2: ; %end
-; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT: .LBB96_3: ; %cmp.false
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v52.h
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v51.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v53.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v51.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v54.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v49.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v55.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v53.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v55.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v52.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v65.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v64.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v65.h
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v64.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v67.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v54.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v67.h
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v66.h
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v68.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v66.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v70.h
-; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v68.h
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v71.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v69.h
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v83.h
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v69.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v84.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v82.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v85.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v80.l
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v96.h
-; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v86.h
-; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v84.l
-; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v80.h
-; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v87.h
-; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v71.h
-; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v96.l
-; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v86.l
-; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v97.l
-; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v85.h
-; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v100.l
-; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v98.l
-; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v100.h
-; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v32.h
-; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v34.h
-; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v32.l
-; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v35.l
-; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v33.h
-; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v35.h
-; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v33.l
-; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v37.h
-; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v36.l
-; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v38.l
-; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v36.h
-; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v48.l
-; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v34.l
-; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v48.h
-; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v38.h
-; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v49.l
-; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v37.l
-; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v50.l
-; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v39.l
-; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v50.h
-; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v39.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v81.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v81.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v82.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v83.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v97.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v70.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v98.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v99.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v99.h
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v87.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v101.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v102.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v102.h
-; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v103.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v112.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v101.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v112.h
-; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v113.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v113.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v103.h
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v115.l
-; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v115.h
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v116.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v116.h
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v117.l
-; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v114.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v118.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v118.h
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v119.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v114.h
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v119.h
-; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v128.l
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v128.h
-; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v129.l
-; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v130.l
-; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v117.h
-; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v130.h
-; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v131.l
-; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v131.h
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v129.h
-; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v132.h
-; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v133.l
-; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v21.l, v133.h
-; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v134.l
-; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v135.l
-; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v132.l
-; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v135.h
-; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v144.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v144.h
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v134.h
-; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v25.l, v145.h
-; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v146.l
-; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v26.l, v146.h
-; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v147.l
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v27.l, v148.l
-; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v145.l
-; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v148.h
-; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v149.l
-; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v149.h
-; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v147.h
-; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v30.l, v150.l
-; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v150.h
-; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v31.l, v151.l
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v151.h
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_hi16
-; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB96_2
-; GFX11-TRUE16-NEXT: .LBB96_4: ; %cmp.true
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v50.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v39.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v50.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v39.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v48.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v151.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v151.h, v0.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v150.l, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v150.h, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v148.l, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, 0x300, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v49.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v37.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v48.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v38.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v36.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v149.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v147.h, v0.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v148.h, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v149.l, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, 0x300, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v34.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v38.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v36.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v37.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v146.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v145.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v146.h, v0.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v147.l, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v145.h, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v34.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v35.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v33.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v35.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v33.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v135.l, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v144.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v134.h, v0.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v135.h, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v144.l, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, 0x300, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v32.l, 3
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(26)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v100.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v32.h, 3
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(24)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v100.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v98.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v132.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v133.h, v0.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v134.l, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v132.h, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v133.l, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, 0x300, v1.h
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(22)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v97.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v85.h, 3
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(20)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v96.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v86.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(18)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v87.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v131.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v129.h, v0.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v130.h, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v131.l, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, 0x300, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v71.h, 3
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(16)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v84.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v80.h, 3
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(14)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v96.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v130.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v117.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v128.h, v0.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v129.l, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v119.h, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v86.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v1.h
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(12)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v85.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v80.l, 3
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(10)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v84.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v82.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v128.l, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v119.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v114.h, v0.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v118.l, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v118.h, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(8)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v83.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v69.l, 3
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(6)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v71.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v69.h, 3
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(4)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v70.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v117.l, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v114.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v116.l, v0.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v116.h, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v115.l, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v68.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v1.h
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v68.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v66.l, 3
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v67.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v66.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v115.h, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v113.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v103.h, v0.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v112.h, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v113.l, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v67.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v54.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v65.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v64.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v65.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v112.l, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v101.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v102.h, v0.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v103.l, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v101.h, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v64.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v55.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v52.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v55.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v53.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v102.l, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v99.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v87.l, v0.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v98.h, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v99.l, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v54.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v49.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v53.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v51.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v52.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v51.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v97.h, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v70.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v82.h, v0.h
-; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v83.l, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v81.l, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v33.l, v81.h, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v32.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v32.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v33.l
-; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v128i8_to_v64i16:
-; GFX11-FAKE16: ; %bb.0:
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT: s_clause 0x1f
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v40, s32 offset:580
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v41, s32 offset:576
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v42, s32 offset:572
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v43, s32 offset:568
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v44, s32 offset:564
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v45, s32 offset:560
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v46, s32 offset:556
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v47, s32 offset:552
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v56, s32 offset:548
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v57, s32 offset:544
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v58, s32 offset:540
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v59, s32 offset:536
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v60, s32 offset:532
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v61, s32 offset:528
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v62, s32 offset:524
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v63, s32 offset:520
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v72, s32 offset:516
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v73, s32 offset:512
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v74, s32 offset:508
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v75, s32 offset:504
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v76, s32 offset:500
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v77, s32 offset:496
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v78, s32 offset:492
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v79, s32 offset:488
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v88, s32 offset:484
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v89, s32 offset:480
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v90, s32 offset:476
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v91, s32 offset:472
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v92, s32 offset:468
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v93, s32 offset:464
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v94, s32 offset:460
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v95, s32 offset:456
-; GFX11-FAKE16-NEXT: s_clause 0xf
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v104, s32 offset:452
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v105, s32 offset:448
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v106, s32 offset:444
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v107, s32 offset:440
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v108, s32 offset:436
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v109, s32 offset:432
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v110, s32 offset:428
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v111, s32 offset:424
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v120, s32 offset:420
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v121, s32 offset:416
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v122, s32 offset:412
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v123, s32 offset:408
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v124, s32 offset:404
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v125, s32 offset:400
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v126, s32 offset:396
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v127, s32 offset:392
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v37, v30 :: v_dual_mov_b32 v54, v24
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v55, v28 :: v_dual_mov_b32 v52, v26
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v50, v22 :: v_dual_mov_b32 v53, v20
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v34, v18 :: v_dual_mov_b32 v51, v16
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v36, v14 :: v_dual_mov_b32 v49, v12
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v32, v10 :: v_dual_mov_b32 v39, v4
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v48, v8 :: v_dual_mov_b32 v33, v6
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v35, v2 :: v_dual_mov_b32 v38, v0
-; GFX11-FAKE16-NEXT: s_clause 0x1f
-; GFX11-FAKE16-NEXT: scratch_load_u16 v0, off, s32 offset:384
-; GFX11-FAKE16-NEXT: scratch_load_u16 v118, off, s32 offset:380
-; GFX11-FAKE16-NEXT: scratch_load_u16 v2, off, s32 offset:376
-; GFX11-FAKE16-NEXT: scratch_load_u16 v134, off, s32 offset:372
-; GFX11-FAKE16-NEXT: scratch_load_u16 v4, off, s32 offset:368
-; GFX11-FAKE16-NEXT: scratch_load_u16 v116, off, s32 offset:364
-; GFX11-FAKE16-NEXT: scratch_load_u16 v6, off, s32 offset:360
-; GFX11-FAKE16-NEXT: scratch_load_u16 v131, off, s32 offset:356
-; GFX11-FAKE16-NEXT: scratch_load_u16 v8, off, s32 offset:352
-; GFX11-FAKE16-NEXT: scratch_load_u16 v98, off, s32 offset:348
-; GFX11-FAKE16-NEXT: scratch_load_u16 v10, off, s32 offset:344
-; GFX11-FAKE16-NEXT: scratch_load_u16 v129, off, s32 offset:340
-; GFX11-FAKE16-NEXT: scratch_load_u16 v12, off, s32 offset:336
-; GFX11-FAKE16-NEXT: scratch_load_u16 v99, off, s32 offset:332
-; GFX11-FAKE16-NEXT: scratch_load_u16 v14, off, s32 offset:328
-; GFX11-FAKE16-NEXT: scratch_load_u16 v112, off, s32 offset:324
-; GFX11-FAKE16-NEXT: scratch_load_u16 v16, off, s32 offset:320
-; GFX11-FAKE16-NEXT: scratch_load_u16 v81, off, s32 offset:316
-; GFX11-FAKE16-NEXT: scratch_load_u16 v18, off, s32 offset:312
-; GFX11-FAKE16-NEXT: scratch_load_u16 v103, off, s32 offset:308
-; GFX11-FAKE16-NEXT: scratch_load_u16 v20, off, s32 offset:304
-; GFX11-FAKE16-NEXT: scratch_load_u16 v86, off, s32 offset:300
-; GFX11-FAKE16-NEXT: scratch_load_u16 v22, off, s32 offset:296
-; GFX11-FAKE16-NEXT: scratch_load_u16 v101, off, s32 offset:292
-; GFX11-FAKE16-NEXT: scratch_load_u16 v24, off, s32 offset:288
-; GFX11-FAKE16-NEXT: scratch_load_u16 v83, off, s32 offset:284
-; GFX11-FAKE16-NEXT: scratch_load_u16 v26, off, s32 offset:280
-; GFX11-FAKE16-NEXT: scratch_load_u16 v97, off, s32 offset:276
-; GFX11-FAKE16-NEXT: scratch_load_u16 v28, off, s32 offset:272
-; GFX11-FAKE16-NEXT: scratch_load_u16 v67, off, s32 offset:268
-; GFX11-FAKE16-NEXT: scratch_load_u16 v30, off, s32 offset:264
-; GFX11-FAKE16-NEXT: scratch_load_u16 v85, off, s32 offset:260
-; GFX11-FAKE16-NEXT: s_clause 0x1f
-; GFX11-FAKE16-NEXT: scratch_load_u16 v31, off, s32 offset:256
-; GFX11-FAKE16-NEXT: scratch_load_u16 v68, off, s32 offset:252
-; GFX11-FAKE16-NEXT: scratch_load_u16 v94, off, s32 offset:248
-; GFX11-FAKE16-NEXT: scratch_load_u16 v80, off, s32 offset:244
-; GFX11-FAKE16-NEXT: scratch_load_u16 v88, off, s32 offset:240
-; GFX11-FAKE16-NEXT: scratch_load_u16 v64, off, s32 offset:236
-; GFX11-FAKE16-NEXT: scratch_load_u16 v93, off, s32 offset:232
-; GFX11-FAKE16-NEXT: scratch_load_u16 v69, off, s32 offset:228
-; GFX11-FAKE16-NEXT: scratch_load_u16 v91, off, s32 offset:224
-; GFX11-FAKE16-NEXT: scratch_load_u16 v65, off, s32 offset:220
-; GFX11-FAKE16-NEXT: scratch_load_u16 v92, off, s32 offset:216
-; GFX11-FAKE16-NEXT: scratch_load_b32 v150, off, s32 offset:388
-; GFX11-FAKE16-NEXT: scratch_load_u16 v182, off, s32
-; GFX11-FAKE16-NEXT: scratch_load_u16 v40, off, s32 offset:8
-; GFX11-FAKE16-NEXT: scratch_load_u16 v43, off, s32 offset:16
-; GFX11-FAKE16-NEXT: scratch_load_u16 v44, off, s32 offset:24
-; GFX11-FAKE16-NEXT: scratch_load_u16 v45, off, s32 offset:32
-; GFX11-FAKE16-NEXT: scratch_load_u16 v46, off, s32 offset:40
-; GFX11-FAKE16-NEXT: scratch_load_u16 v47, off, s32 offset:48
-; GFX11-FAKE16-NEXT: scratch_load_u16 v56, off, s32 offset:56
-; GFX11-FAKE16-NEXT: scratch_load_u16 v58, off, s32 offset:64
-; GFX11-FAKE16-NEXT: scratch_load_u16 v59, off, s32 offset:72
-; GFX11-FAKE16-NEXT: scratch_load_u16 v60, off, s32 offset:80
-; GFX11-FAKE16-NEXT: scratch_load_u16 v61, off, s32 offset:88
-; GFX11-FAKE16-NEXT: scratch_load_u16 v62, off, s32 offset:96
-; GFX11-FAKE16-NEXT: scratch_load_u16 v63, off, s32 offset:104
-; GFX11-FAKE16-NEXT: scratch_load_u16 v72, off, s32 offset:112
-; GFX11-FAKE16-NEXT: scratch_load_u16 v73, off, s32 offset:120
-; GFX11-FAKE16-NEXT: scratch_load_u16 v74, off, s32 offset:128
-; GFX11-FAKE16-NEXT: scratch_load_u16 v75, off, s32 offset:136
-; GFX11-FAKE16-NEXT: scratch_load_u16 v76, off, s32 offset:144
-; GFX11-FAKE16-NEXT: scratch_load_u16 v77, off, s32 offset:152
-; GFX11-FAKE16-NEXT: s_clause 0x1f
-; GFX11-FAKE16-NEXT: scratch_load_u16 v78, off, s32 offset:160
-; GFX11-FAKE16-NEXT: scratch_load_u16 v79, off, s32 offset:168
-; GFX11-FAKE16-NEXT: scratch_load_u16 v89, off, s32 offset:176
-; GFX11-FAKE16-NEXT: scratch_load_u16 v90, off, s32 offset:184
-; GFX11-FAKE16-NEXT: scratch_load_u16 v95, off, s32 offset:192
-; GFX11-FAKE16-NEXT: scratch_load_u16 v104, off, s32 offset:200
-; GFX11-FAKE16-NEXT: scratch_load_u16 v105, off, s32 offset:208
-; GFX11-FAKE16-NEXT: scratch_load_u16 v42, off, s32 offset:212
-; GFX11-FAKE16-NEXT: scratch_load_u16 v180, off, s32 offset:204
-; GFX11-FAKE16-NEXT: scratch_load_u16 v183, off, s32 offset:196
-; GFX11-FAKE16-NEXT: scratch_load_u16 v165, off, s32 offset:188
-; GFX11-FAKE16-NEXT: scratch_load_u16 v179, off, s32 offset:180
-; GFX11-FAKE16-NEXT: scratch_load_u16 v163, off, s32 offset:172
-; GFX11-FAKE16-NEXT: scratch_load_u16 v177, off, s32 offset:164
-; GFX11-FAKE16-NEXT: scratch_load_u16 v145, off, s32 offset:156
-; GFX11-FAKE16-NEXT: scratch_load_u16 v166, off, s32 offset:148
-; GFX11-FAKE16-NEXT: scratch_load_u16 v148, off, s32 offset:140
-; GFX11-FAKE16-NEXT: scratch_load_u16 v151, off, s32 offset:132
-; GFX11-FAKE16-NEXT: scratch_load_u16 v164, off, s32 offset:124
-; GFX11-FAKE16-NEXT: scratch_load_u16 v178, off, s32 offset:116
-; GFX11-FAKE16-NEXT: scratch_load_u16 v146, off, s32 offset:108
-; GFX11-FAKE16-NEXT: scratch_load_u16 v162, off, s32 offset:100
-; GFX11-FAKE16-NEXT: scratch_load_u16 v144, off, s32 offset:92
-; GFX11-FAKE16-NEXT: scratch_load_u16 v149, off, s32 offset:84
-; GFX11-FAKE16-NEXT: scratch_load_u16 v119, off, s32 offset:76
-; GFX11-FAKE16-NEXT: scratch_load_u16 v147, off, s32 offset:68
-; GFX11-FAKE16-NEXT: scratch_load_u16 v130, off, s32 offset:60
-; GFX11-FAKE16-NEXT: scratch_load_u16 v135, off, s32 offset:52
-; GFX11-FAKE16-NEXT: scratch_load_u16 v117, off, s32 offset:44
-; GFX11-FAKE16-NEXT: scratch_load_u16 v133, off, s32 offset:36
-; GFX11-FAKE16-NEXT: scratch_load_u16 v96, off, s32 offset:28
-; GFX11-FAKE16-NEXT: scratch_load_u16 v114, off, s32 offset:20
-; GFX11-FAKE16-NEXT: s_clause 0x1
-; GFX11-FAKE16-NEXT: scratch_load_u16 v87, off, s32 offset:12
-; GFX11-FAKE16-NEXT: scratch_load_u16 v102, off, s32 offset:4
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v70, 8, v1
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v71, 8, v3
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v84, 8, v5
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v82, 8, v7
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v115, 8, v9
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v66, 8, v11
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v128, 8, v13
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v113, 8, v15
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v132, 8, v17
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v100, 8, v19
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v161, 8, v21
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v160, 8, v23
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v176, 8, v25
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v167, 8, v27
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v181, 8, v29
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(62)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v127, 8, v0
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v126, 8, v2
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v124, 8, v4
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v125, 8, v6
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v120, 8, v8
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v123, 8, v10
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v121, 8, v12
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v122, 8, v14
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v106, 8, v16
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v111, 8, v18
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v109, 8, v20
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v110, 8, v22
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v107, 8, v24
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v108, 8, v26
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(61)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v88, 8, v88
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(59)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v93, 8, v93
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(57)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v91, 8, v91
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(55)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v92, 8, v92
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(54)
-; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v150
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(53)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v150, 8, v182
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(52)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v41, 8, v40
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(51)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v40, 8, v43
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(50)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v43, 8, v44
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(49)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v182, 8, v45
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(48)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v46, 8, v46
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(47)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v45, 8, v47
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(46)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v57, 8, v56
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(45)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v56, 8, v58
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(44)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v58, 8, v59
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(43)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v44, 8, v60
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(42)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v60, 8, v61
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(41)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v59, 8, v62
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(40)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v62, 8, v63
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(39)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v47, 8, v72
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(38)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v72, 8, v73
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(37)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v63, 8, v74
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(36)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v74, 8, v75
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(35)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v73, 8, v76
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(34)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v75, 8, v77
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(33)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v61, 8, v78
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(32)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v78, 8, v79
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(31)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v77, 8, v89
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(30)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v79, 8, v90
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(29)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v76, 8, v95
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(28)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v90, 8, v104
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(27)
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v89, 8, v105
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v104, 8, v94
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v95, 8, v31
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v105, 8, v30
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v94, 8, v28
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB96_2
-; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v38
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v35
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v39
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v33
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v48
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v70
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v71
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v2, v84
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v3, v82
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xff, v36
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xff, v51
-; GFX11-FAKE16-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xff, v34
-; GFX11-FAKE16-NEXT: v_perm_b32 v1, v3, v2, 0x5040100
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v32
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v49
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xff, v53
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v9, 0xff, v50
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xff, v54
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xff, v52
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v4, v115
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v2, v66
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v3, v128
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v5, v113
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v6, v6, v132
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v7, v100
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v8, v8, v161
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v9, v9, v160
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v10, v10, v176
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v11, v11, v167
-; GFX11-FAKE16-NEXT: v_perm_b32 v2, v2, v4, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v3, v5, v3, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v4, v7, v6, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v5, v9, v8, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v6, v11, v10, 0x5040100
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xff, v55
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xff, v37
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v9, 0xff, v102
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xff, v87
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xff, v114
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xff, v96
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xff, v133
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xff, v117
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xff, v135
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v16, 0xff, v130
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v7, v181
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v8, v8, v150
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v9, v9, v41
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v10, v10, v40
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v11, v11, v43
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v12, v12, v182
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v13, v13, v46
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v14, v14, v45
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v15, v15, v57
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v16, v16, v56
-; GFX11-FAKE16-NEXT: v_perm_b32 v7, v8, v7, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v8, v10, v9, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v9, v12, v11, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v10, v14, v13, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v11, v16, v15, 0x5040100
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xff, v147
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xff, v119
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xff, v149
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xff, v144
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v16, 0xff, v162
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v17, 0xff, v146
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xff, v178
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xff, v164
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xff, v151
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xff, v148
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v12, v12, v58
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v13, v13, v44
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v14, v14, v60
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v15, v15, v59
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v16, v16, v62
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v17, v17, v47
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v18, v18, v72
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v19, v19, v63
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v20, v20, v74
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v21, v21, v73
-; GFX11-FAKE16-NEXT: v_perm_b32 v12, v13, v12, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v13, v15, v14, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v14, v17, v16, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v15, v19, v18, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v16, v21, v20, 0x5040100
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v17, 0xff, v166
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xff, v145
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xff, v177
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xff, v163
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xff, v179
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xff, v165
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v23, 0xff, v183
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v24, 0xff, v180
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v25, 0xff, v42
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v26, 0xff, v65
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v17, v17, v75
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v18, v18, v61
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v19, v19, v78
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v20, v20, v77
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v21, v21, v79
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v22, v22, v76
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v23, v23, v90
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v24, v24, v89
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v25, v25, v92
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v26, v26, v91
-; GFX11-FAKE16-NEXT: v_perm_b32 v17, v18, v17, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v18, v20, v19, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v19, v22, v21, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v20, v24, v23, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v21, v26, v25, 0x5040100
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xff, v69
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v23, 0xff, v64
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v24, 0xff, v80
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v25, 0xff, v68
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v26, 0xff, v85
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v27, 0xff, v67
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v28, 0xff, v97
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v29, 0xff, v83
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v30, 0xff, v101
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v31, 0xff, v86
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v22, v22, v93
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v23, v23, v88
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v24, v24, v104
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v25, v25, v95
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v26, v26, v105
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v27, v27, v94
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v28, v28, v108
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v29, v29, v107
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v30, v30, v110
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v31, v31, v109
-; GFX11-FAKE16-NEXT: v_perm_b32 v22, v23, v22, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v23, v25, v24, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v24, v27, v26, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v25, v29, v28, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v26, v31, v30, 0x5040100
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v27, 0xff, v103
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v28, 0xff, v81
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v29, 0xff, v112
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v30, 0xff, v99
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v31, 0xff, v129
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v32, 0xff, v98
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xff, v131
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v34, 0xff, v116
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v35, 0xff, v134
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v36, 0xff, v118
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v27, v27, v111
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v28, v28, v106
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v29, v29, v122
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v30, v30, v121
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v31, v31, v123
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v32, v32, v120
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v33, v33, v125
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v34, v34, v124
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v35, v35, v126
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v36, v36, v127
-; GFX11-FAKE16-NEXT: v_perm_b32 v27, v28, v27, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v28, v30, v29, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v29, v32, v31, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v30, v34, v33, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v31, v36, v35, 0x5040100
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr48
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr49
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr51
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr53
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr50
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr54
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr52
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr55
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr102
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr87
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr114
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr96
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr133
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr117
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr135
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr130
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr147
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr119
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr149
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr144
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr162
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr146
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr178
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr164
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr151
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr148
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr166
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr145
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr177
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr163
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr179
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr165
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr183
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr180
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr65
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr69
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr64
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr80
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr68
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr85
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr67
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr97
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr83
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr101
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr86
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr103
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr81
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr112
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr99
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr129
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr98
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr131
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr116
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr134
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr118
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr70
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr71
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr84
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr82
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr115
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr66
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr128
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr113
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr132
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr100
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr161
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr160
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr176
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr167
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr181
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr150
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr41
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr40
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr43
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr182
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr46
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr45
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr57
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr56
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr58
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr44
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr60
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr59
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr62
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr47
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr72
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr63
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr74
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr73
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr75
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr61
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr78
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr77
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr79
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr76
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr90
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr89
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr92
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr91
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr93
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr88
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr104
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr95
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr105
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr94
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr108
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr107
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr110
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr109
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr111
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr106
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr122
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr121
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr123
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr120
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr125
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr124
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr126
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr127
-; GFX11-FAKE16-NEXT: .LBB96_2: ; %Flow
-; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB96_4
-; GFX11-FAKE16-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v134, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v1, v118, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v2, v131, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, v116, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, v129, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v35, v35, 3
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v126, v0
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v127, v1
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v125, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v124, v3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v33, v33, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v31, 0x300, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v98, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v116, 0x300, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v30, 0x300, v2
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v98, 0x300, v3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v1, v112, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, v99, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, v103, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v123, v2
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v120, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v122, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v29, 0x300, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v121, v3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v99, 0x300, v0
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v111, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v28, 0x300, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v1, v81, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v81, 0x300, v2
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v2, v101, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v27, 0x300, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v86, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, v97, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, v83, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v106, v1
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v110, v2
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v109, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v83, 0x300, v1
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v108, v3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v26, 0x300, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v107, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v86, 0x300, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v85, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v25, 0x300, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v1, v67, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v67, 0x300, v2
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v2, v80, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, v68, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, v69, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v105, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v94, v1
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v104, v2
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v24, 0x300, v0
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v95, v3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v68, 0x300, v1
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v93, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v23, 0x300, v2
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v2, v64, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v64, 0x300, v0
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(26)
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v42, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v22, 0x300, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v1, v65, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(24)
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, v183, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, v180, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v88, v2
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v92, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v91, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v65, 0x300, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v90, v3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v21, 0x300, v0
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v89, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v69, 0x300, v1
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(22)
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v1, v179, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v20, 0x300, v2
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v2, v165, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v80, 0x300, v0
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(20)
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v177, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, v163, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(18)
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, v166, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v79, v1
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v76, v2
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v78, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v19, 0x300, v1
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v77, v3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v85, 0x300, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v75, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v18, 0x300, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v145, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v97, 0x300, v1
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(16)
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v1, v151, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v17, 0x300, v2
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v2, v148, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(14)
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, v178, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, v164, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v61, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v74, v1
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v73, v2
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v101, 0x300, v0
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v72, v3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v16, 0x300, v1
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v63, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v103, 0x300, v2
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(12)
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v2, v162, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v15, 0x300, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v146, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v112, 0x300, v1
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(10)
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v1, v149, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, v144, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(8)
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, v147, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v62, v2
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v47, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v60, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v14, 0x300, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v59, v3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v118, 0x300, v0
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v58, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v13, 0x300, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v1, v119, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v119, 0x300, v2
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(6)
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v2, v135, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v12, 0x300, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v130, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(4)
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, v133, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, v117, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v44, v1
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v57, v2
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v56, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v117, 0x300, v1
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v46, v3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v11, 0x300, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v45, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v129, 0x300, v0
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(2)
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v114, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v10, 0x300, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v1, v96, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v96, 0x300, v2
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v2, v102, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, v87, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, v55, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v43, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v182, v1
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v41, v2
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v9, 0x300, v0
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v40, v3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v55, 0x300, v1
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v181, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v8, 0x300, v2
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v2, v37, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v37, 0x300, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v54, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v7, 0x300, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v1, v52, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, v53, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, v50, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v150, v2
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v176, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v167, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v50, 0x300, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v161, v3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v6, 0x300, v0
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v160, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v52, 0x300, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v1, v51, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v5, 0x300, v2
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v2, v49, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v49, 0x300, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v48, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, v34, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, v36, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v132, v1
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v128, v2
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v115, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v34, 0x300, v1
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v100, v3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, 0x300, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v113, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, 0x300, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v39, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v36, v38, 3
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v32, v32, 3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v35, 0xff, v35
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v36, 0xff, v36
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v35, v71, v35
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v33, v82, v33
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v84, v0
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v36, v70, v36
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v32, v66, v32
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v35, 0x300, v35
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v33, 0x300, v33
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v38, 0x300, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, 0x300, v36
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v32, 0x300, v32
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v36, 0x300, v2
-; GFX11-FAKE16-NEXT: v_add_nc_u16 v39, 0x300, v1
-; GFX11-FAKE16-NEXT: v_perm_b32 v1, v33, v38, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v0, v35, v0, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v2, v32, v4, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v3, v36, v3, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v4, v39, v34, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v5, v49, v5, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v6, v52, v6, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v7, v50, v7, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v8, v37, v8, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v9, v55, v9, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v10, v96, v10, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v11, v129, v11, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v12, v117, v12, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v13, v119, v13, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v14, v118, v14, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v15, v112, v15, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v16, v103, v16, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v17, v101, v17, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v18, v97, v18, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v19, v85, v19, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v20, v80, v20, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v21, v69, v21, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v22, v65, v22, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v23, v64, v23, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v24, v68, v24, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v25, v67, v25, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v26, v86, v26, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v27, v83, v27, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v28, v81, v28, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v29, v99, v29, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v30, v98, v30, 0x5040100
-; GFX11-FAKE16-NEXT: v_perm_b32 v31, v116, v31, 0x5040100
-; GFX11-FAKE16-NEXT: .LBB96_4: ; %end
-; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT: s_clause 0x1f
-; GFX11-FAKE16-NEXT: scratch_load_b32 v127, off, s32 offset:392
-; GFX11-FAKE16-NEXT: scratch_load_b32 v126, off, s32 offset:396
-; GFX11-FAKE16-NEXT: scratch_load_b32 v125, off, s32 offset:400
-; GFX11-FAKE16-NEXT: scratch_load_b32 v124, off, s32 offset:404
-; GFX11-FAKE16-NEXT: scratch_load_b32 v123, off, s32 offset:408
-; GFX11-FAKE16-NEXT: scratch_load_b32 v122, off, s32 offset:412
-; GFX11-FAKE16-NEXT: scratch_load_b32 v121, off, s32 offset:416
-; GFX11-FAKE16-NEXT: scratch_load_b32 v120, off, s32 offset:420
-; GFX11-FAKE16-NEXT: scratch_load_b32 v111, off, s32 offset:424
-; GFX11-FAKE16-NEXT: scratch_load_b32 v110, off, s32 offset:428
-; GFX11-FAKE16-NEXT: scratch_load_b32 v109, off, s32 offset:432
-; GFX11-FAKE16-NEXT: scratch_load_b32 v108, off, s32 offset:436
-; GFX11-FAKE16-NEXT: scratch_load_b32 v107, off, s32 offset:440
-; GFX11-FAKE16-NEXT: scratch_load_b32 v106, off, s32 offset:444
-; GFX11-FAKE16-NEXT: scratch_load_b32 v105, off, s32 offset:448
-; GFX11-FAKE16-NEXT: scratch_load_b32 v104, off, s32 offset:452
-; GFX11-FAKE16-NEXT: scratch_load_b32 v95, off, s32 offset:456
-; GFX11-FAKE16-NEXT: scratch_load_b32 v94, off, s32 offset:460
-; GFX11-FAKE16-NEXT: scratch_load_b32 v93, off, s32 offset:464
-; GFX11-FAKE16-NEXT: scratch_load_b32 v92, off, s32 offset:468
-; GFX11-FAKE16-NEXT: scratch_load_b32 v91, off, s32 offset:472
-; GFX11-FAKE16-NEXT: scratch_load_b32 v90, off, s32 offset:476
-; GFX11-FAKE16-NEXT: scratch_load_b32 v89, off, s32 offset:480
-; GFX11-FAKE16-NEXT: scratch_load_b32 v88, off, s32 offset:484
-; GFX11-FAKE16-NEXT: scratch_load_b32 v79, off, s32 offset:488
-; GFX11-FAKE16-NEXT: scratch_load_b32 v78, off, s32 offset:492
-; GFX11-FAKE16-NEXT: scratch_load_b32 v77, off, s32 offset:496
-; GFX11-FAKE16-NEXT: scratch_load_b32 v76, off, s32 offset:500
-; GFX11-FAKE16-NEXT: scratch_load_b32 v75, off, s32 offset:504
-; GFX11-FAKE16-NEXT: scratch_load_b32 v74, off, s32 offset:508
-; GFX11-FAKE16-NEXT: scratch_load_b32 v73, off, s32 offset:512
-; GFX11-FAKE16-NEXT: scratch_load_b32 v72, off, s32 offset:516
-; GFX11-FAKE16-NEXT: s_clause 0xf
-; GFX11-FAKE16-NEXT: scratch_load_b32 v63, off, s32 offset:520
-; GFX11-FAKE16-NEXT: scratch_load_b32 v62, off, s32 offset:524
-; GFX11-FAKE16-NEXT: scratch_load_b32 v61, off, s32 offset:528
-; GFX11-FAKE16-NEXT: scratch_load_b32 v60, off, s32 offset:532
-; GFX11-FAKE16-NEXT: scratch_load_b32 v59, off, s32 offset:536
-; GFX11-FAKE16-NEXT: scratch_load_b32 v58, off, s32 offset:540
-; GFX11-FAKE16-NEXT: scratch_load_b32 v57, off, s32 offset:544
-; GFX11-FAKE16-NEXT: scratch_load_b32 v56, off, s32 offset:548
-; GFX11-FAKE16-NEXT: scratch_load_b32 v47, off, s32 offset:552
-; GFX11-FAKE16-NEXT: scratch_load_b32 v46, off, s32 offset:556
-; GFX11-FAKE16-NEXT: scratch_load_b32 v45, off, s32 offset:560
-; GFX11-FAKE16-NEXT: scratch_load_b32 v44, off, s32 offset:564
-; GFX11-FAKE16-NEXT: scratch_load_b32 v43, off, s32 offset:568
-; GFX11-FAKE16-NEXT: scratch_load_b32 v42, off, s32 offset:572
-; GFX11-FAKE16-NEXT: scratch_load_b32 v41, off, s32 offset:576
-; GFX11-FAKE16-NEXT: scratch_load_b32 v40, off, s32 offset:580
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT: 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 inreg <64 x i16> @bitcast_v128i8_to_v64i16_scalar(<128 x i8> inreg %a, i32 inreg %b) {
-; SI-LABEL: bitcast_v128i8_to_v64i16_scalar:
-; SI: ; %bb.0:
-; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; SI-NEXT: s_or_saveexec_b64 s[4:5], -1
-; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; SI-NEXT: s_mov_b64 exec, s[4:5]
-; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; SI-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:332
-; SI-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:328
-; SI-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:324
-; SI-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:320
-; SI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:308
-; SI-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:304
-; SI-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:300
-; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:296
-; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:292
-; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:288
-; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:276
-; SI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:272
-; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:268
-; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:264
-; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:260
-; SI-NEXT: s_waitcnt expcnt(6)
-; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:256
-; SI-NEXT: s_waitcnt expcnt(5)
-; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:244
-; SI-NEXT: s_waitcnt expcnt(4)
-; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:240
-; SI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:236
-; SI-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:232
-; SI-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:228
-; SI-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:224
-; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:212
-; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:208
-; SI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:204
-; SI-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:200
-; SI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:196
-; SI-NEXT: v_writelane_b32 v63, s30, 0
-; SI-NEXT: v_writelane_b32 v63, s31, 1
-; SI-NEXT: v_writelane_b32 v63, s34, 2
-; SI-NEXT: v_writelane_b32 v63, s35, 3
-; SI-NEXT: v_writelane_b32 v63, s36, 4
-; SI-NEXT: v_writelane_b32 v63, s37, 5
-; SI-NEXT: v_writelane_b32 v63, s38, 6
-; SI-NEXT: v_writelane_b32 v63, s39, 7
-; SI-NEXT: v_writelane_b32 v63, s48, 8
-; SI-NEXT: v_writelane_b32 v63, s49, 9
-; SI-NEXT: v_writelane_b32 v63, s50, 10
-; SI-NEXT: v_writelane_b32 v63, s51, 11
-; SI-NEXT: v_writelane_b32 v63, s52, 12
-; SI-NEXT: v_writelane_b32 v63, s53, 13
-; SI-NEXT: v_writelane_b32 v63, s54, 14
-; SI-NEXT: v_writelane_b32 v63, s55, 15
-; SI-NEXT: v_writelane_b32 v63, s64, 16
-; SI-NEXT: v_writelane_b32 v63, s65, 17
-; SI-NEXT: ; implicit-def: $vgpr62 : SGPR spill to VGPR lane
-; SI-NEXT: v_writelane_b32 v63, s66, 18
-; SI-NEXT: v_writelane_b32 v62, s28, 0
-; SI-NEXT: v_writelane_b32 v63, s67, 19
-; SI-NEXT: v_writelane_b32 v62, s27, 1
-; SI-NEXT: v_writelane_b32 v63, s68, 20
-; SI-NEXT: v_writelane_b32 v62, s25, 2
-; SI-NEXT: v_writelane_b32 v63, s69, 21
-; SI-NEXT: v_writelane_b32 v62, s24, 3
-; SI-NEXT: v_writelane_b32 v63, s70, 22
-; SI-NEXT: v_writelane_b32 v62, s23, 4
-; SI-NEXT: v_writelane_b32 v63, s71, 23
-; SI-NEXT: v_writelane_b32 v62, s22, 5
-; SI-NEXT: v_writelane_b32 v63, s80, 24
-; SI-NEXT: v_writelane_b32 v62, s21, 6
-; SI-NEXT: v_writelane_b32 v63, s81, 25
-; SI-NEXT: v_writelane_b32 v62, s20, 7
-; SI-NEXT: v_writelane_b32 v63, s82, 26
-; SI-NEXT: v_writelane_b32 v62, s19, 8
-; SI-NEXT: v_writelane_b32 v63, s83, 27
-; SI-NEXT: v_writelane_b32 v62, s18, 9
-; SI-NEXT: v_writelane_b32 v63, s84, 28
-; SI-NEXT: v_writelane_b32 v62, s16, 10
-; SI-NEXT: v_writelane_b32 v63, s85, 29
-; SI-NEXT: v_writelane_b32 v63, s86, 30
-; SI-NEXT: v_writelane_b32 v63, s87, 31
-; SI-NEXT: v_writelane_b32 v63, s96, 32
-; SI-NEXT: v_writelane_b32 v63, s97, 33
-; SI-NEXT: v_writelane_b32 v63, s98, 34
-; SI-NEXT: s_waitcnt expcnt(2)
-; SI-NEXT: v_mov_b32_e32 v29, v5
-; SI-NEXT: v_readfirstlane_b32 s76, v18
-; SI-NEXT: v_readfirstlane_b32 s40, v25
-; SI-NEXT: v_readfirstlane_b32 s16, v24
-; SI-NEXT: v_readfirstlane_b32 s42, v23
-; SI-NEXT: v_readfirstlane_b32 s52, v20
-; SI-NEXT: v_readfirstlane_b32 s8, v19
-; SI-NEXT: s_waitcnt vmcnt(14)
-; SI-NEXT: v_readfirstlane_b32 s50, v35
-; SI-NEXT: v_readfirstlane_b32 s31, v36
-; SI-NEXT: v_readfirstlane_b32 s53, v37
-; SI-NEXT: v_readfirstlane_b32 s82, v48
-; SI-NEXT: v_readfirstlane_b32 s7, v49
-; SI-NEXT: v_readfirstlane_b32 s79, v52
-; SI-NEXT: v_readfirstlane_b32 s78, v55
-; SI-NEXT: v_readfirstlane_b32 s4, v41
-; SI-NEXT: v_writelane_b32 v62, s4, 11
-; SI-NEXT: v_readfirstlane_b32 s4, v43
-; SI-NEXT: v_writelane_b32 v62, s4, 12
-; SI-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:192
-; SI-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:180
-; SI-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:176
-; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:172
-; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:168
-; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:164
-; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:160
-; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:148
-; SI-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:144
-; SI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:140
-; SI-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:136
-; SI-NEXT: v_readfirstlane_b32 s4, v45
-; SI-NEXT: v_writelane_b32 v62, s4, 13
-; SI-NEXT: s_waitcnt vmcnt(14)
-; SI-NEXT: v_readfirstlane_b32 s4, v56
-; SI-NEXT: v_writelane_b32 v62, s4, 14
-; SI-NEXT: v_readfirstlane_b32 s4, v57
-; SI-NEXT: v_writelane_b32 v62, s4, 15
-; SI-NEXT: v_readfirstlane_b32 s4, v60
-; SI-NEXT: v_readfirstlane_b32 s86, v31
-; SI-NEXT: v_readfirstlane_b32 s36, v32
-; SI-NEXT: v_readfirstlane_b32 s71, v33
-; SI-NEXT: v_readfirstlane_b32 s77, v59
-; SI-NEXT: v_writelane_b32 v62, s4, 16
-; SI-NEXT: v_readfirstlane_b32 s4, v61
-; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:132
-; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:128
-; SI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:116
-; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:112
-; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:108
-; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:104
-; SI-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:100
-; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:96
-; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:84
-; SI-NEXT: v_readfirstlane_b32 s98, v50
-; SI-NEXT: s_waitcnt vmcnt(14)
-; SI-NEXT: v_readfirstlane_b32 s67, v53
-; SI-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:68
-; SI-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:64
-; SI-NEXT: v_writelane_b32 v62, s4, 17
-; SI-NEXT: v_readfirstlane_b32 s4, v38
-; SI-NEXT: v_writelane_b32 v62, s4, 18
-; SI-NEXT: v_readfirstlane_b32 s81, v34
-; SI-NEXT: v_readfirstlane_b32 s75, v39
-; SI-NEXT: v_readfirstlane_b32 s68, v42
-; SI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:80
-; SI-NEXT: v_readfirstlane_b32 s49, v40
-; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:52
-; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:48
-; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:44
-; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:40
-; SI-NEXT: v_readfirstlane_b32 s51, v54
-; SI-NEXT: v_readfirstlane_b32 s97, v51
-; SI-NEXT: v_readfirstlane_b32 s35, v27
-; SI-NEXT: v_writelane_b32 v63, s99, 35
-; SI-NEXT: v_readfirstlane_b32 s28, v28
-; SI-NEXT: v_readfirstlane_b32 s87, v26
-; SI-NEXT: v_lshlrev_b32_e32 v6, 24, v6
-; SI-NEXT: v_lshlrev_b32_e32 v14, 24, v14
-; SI-NEXT: v_readfirstlane_b32 s96, v17
-; SI-NEXT: v_readfirstlane_b32 s99, v16
-; SI-NEXT: v_readfirstlane_b32 s89, v15
-; SI-NEXT: v_readfirstlane_b32 s88, v12
-; SI-NEXT: v_readfirstlane_b32 s30, v11
-; SI-NEXT: v_readfirstlane_b32 s64, v10
-; SI-NEXT: v_readfirstlane_b32 s55, v9
-; SI-NEXT: v_readfirstlane_b32 s65, v8
-; SI-NEXT: v_readfirstlane_b32 s80, v7
-; SI-NEXT: v_readfirstlane_b32 s21, v2
-; SI-NEXT: v_readfirstlane_b32 s74, v1
-; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:156
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:220
-; SI-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:72
-; SI-NEXT: v_readfirstlane_b32 s93, v36
-; SI-NEXT: v_readfirstlane_b32 s24, v37
-; SI-NEXT: v_readfirstlane_b32 s27, v48
-; SI-NEXT: v_readfirstlane_b32 s84, v43
-; SI-NEXT: v_readfirstlane_b32 s83, v44
-; SI-NEXT: v_readfirstlane_b32 s85, v46
-; SI-NEXT: s_waitcnt vmcnt(14)
-; SI-NEXT: v_readfirstlane_b32 s20, v47
-; SI-NEXT: v_readfirstlane_b32 s4, v58
-; SI-NEXT: v_writelane_b32 v62, s4, 19
-; SI-NEXT: v_readfirstlane_b32 s23, v49
-; SI-NEXT: v_readfirstlane_b32 s92, v52
-; SI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:36
-; SI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:32
-; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:20
-; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:16
-; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12
-; SI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:336
-; SI-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:8
-; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:4
-; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32
-; SI-NEXT: v_readfirstlane_b32 s90, v35
-; SI-NEXT: v_readfirstlane_b32 s4, v31
-; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:92
-; SI-NEXT: v_readfirstlane_b32 s38, v32
-; SI-NEXT: v_readfirstlane_b32 s70, v33
-; SI-NEXT: v_readfirstlane_b32 s54, v59
-; SI-NEXT: v_readfirstlane_b32 s57, v60
-; SI-NEXT: s_waitcnt vmcnt(14)
-; SI-NEXT: v_readfirstlane_b32 s56, v61
-; SI-NEXT: v_readfirstlane_b32 s59, v55
-; SI-NEXT: v_readfirstlane_b32 s61, v41
-; SI-NEXT: v_readfirstlane_b32 s19, v45
-; SI-NEXT: v_readfirstlane_b32 s34, v50
-; SI-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:316
-; SI-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:76
-; SI-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:284
-; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:252
-; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:188
-; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:124
-; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:60
-; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:28
-; SI-NEXT: v_writelane_b32 v62, s4, 20
-; SI-NEXT: v_readfirstlane_b32 s25, v53
-; SI-NEXT: v_readfirstlane_b32 s91, v40
-; SI-NEXT: v_readfirstlane_b32 s37, v34
-; SI-NEXT: v_readfirstlane_b32 s47, v56
-; SI-NEXT: s_waitcnt vmcnt(14)
-; SI-NEXT: v_readfirstlane_b32 s46, v57
-; SI-NEXT: v_readfirstlane_b32 s22, v42
-; SI-NEXT: v_lshlrev_b32_e32 v40, 24, v22
-; SI-NEXT: v_lshlrev_b32_e32 v25, 24, v5
-; SI-NEXT: v_readfirstlane_b32 s72, v39
-; SI-NEXT: v_readfirstlane_b32 s94, v51
-; SI-NEXT: v_readfirstlane_b32 s48, v54
-; SI-NEXT: v_readfirstlane_b32 s66, v43
-; SI-NEXT: v_readfirstlane_b32 s69, v44
-; SI-NEXT: s_waitcnt vmcnt(13)
-; SI-NEXT: v_readfirstlane_b32 s45, v46
-; SI-NEXT: s_waitcnt vmcnt(12)
-; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v49
-; SI-NEXT: s_and_b64 s[4:5], vcc, exec
-; SI-NEXT: v_readfirstlane_b32 s4, v4
-; SI-NEXT: v_writelane_b32 v62, s4, 21
-; SI-NEXT: v_readfirstlane_b32 s4, v3
-; SI-NEXT: v_writelane_b32 v62, s4, 22
-; SI-NEXT: v_writelane_b32 v62, s17, 23
-; SI-NEXT: v_writelane_b32 v62, s40, 24
-; SI-NEXT: v_writelane_b32 v62, s16, 25
-; SI-NEXT: v_writelane_b32 v62, s42, 26
-; SI-NEXT: v_writelane_b32 v62, s46, 27
-; SI-NEXT: v_writelane_b32 v62, s47, 28
-; SI-NEXT: v_writelane_b32 v62, s56, 29
-; SI-NEXT: v_writelane_b32 v62, s57, 30
-; SI-NEXT: v_writelane_b32 v62, s45, 31
-; SI-NEXT: s_waitcnt vmcnt(11)
-; SI-NEXT: v_readfirstlane_b32 s58, v52
-; SI-NEXT: v_writelane_b32 v62, s49, 32
-; SI-NEXT: v_writelane_b32 v62, s58, 33
-; SI-NEXT: v_writelane_b32 v62, s59, 34
-; SI-NEXT: v_writelane_b32 v62, s52, 35
-; SI-NEXT: s_waitcnt vmcnt(6)
-; SI-NEXT: v_readfirstlane_b32 s60, v38
-; SI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:312
-; SI-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:280
-; SI-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:248
-; SI-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:216
-; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:184
-; SI-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:152
-; SI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:120
-; SI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:88
-; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:56
-; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:24
-; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; SI-NEXT: v_writelane_b32 v62, s60, 36
-; SI-NEXT: v_writelane_b32 v62, s61, 37
-; SI-NEXT: v_writelane_b32 v62, s93, 38
-; SI-NEXT: v_writelane_b32 v62, s8, 39
-; SI-NEXT: v_readfirstlane_b32 s62, v47
-; SI-NEXT: v_writelane_b32 v62, s72, 40
-; SI-NEXT: v_readfirstlane_b32 s73, v58
-; SI-NEXT: v_writelane_b32 v62, s62, 41
-; SI-NEXT: v_writelane_b32 v62, s73, 42
-; SI-NEXT: v_writelane_b32 v62, s35, 43
-; SI-NEXT: v_writelane_b32 v62, s94, 44
-; SI-NEXT: v_writelane_b32 v62, s48, 45
-; SI-NEXT: v_writelane_b32 v62, s91, 46
-; SI-NEXT: s_waitcnt vmcnt(11)
-; SI-NEXT: v_lshlrev_b32_e32 v61, 24, v61
-; SI-NEXT: v_lshlrev_b32_e32 v60, 24, v60
-; SI-NEXT: v_lshlrev_b32_e32 v31, 24, v31
-; SI-NEXT: v_lshlrev_b32_e32 v24, 24, v59
-; SI-NEXT: v_lshlrev_b32_e32 v27, 24, v45
-; SI-NEXT: v_lshlrev_b32_e32 v43, 24, v13
-; SI-NEXT: v_lshlrev_b32_e32 v44, 24, v41
-; SI-NEXT: v_lshlrev_b32_e32 v45, 24, v55
-; SI-NEXT: v_lshlrev_b32_e32 v46, 24, v50
-; SI-NEXT: v_writelane_b32 v62, s66, 47
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v42, 24, v18
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; SI-NEXT: s_cbranch_scc0 .LBB97_2
-; SI-NEXT: ; %bb.1: ; %cmp.false
-; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; SI-NEXT: v_mov_b32_e32 v21, v29
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v29
-; SI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; SI-NEXT: v_readlane_b32 s4, v62, 7
-; SI-NEXT: v_readlane_b32 s5, v62, 6
-; SI-NEXT: s_and_b32 s4, s4, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 8
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: v_writelane_b32 v62, s4, 48
-; SI-NEXT: v_readlane_b32 s4, v62, 5
-; SI-NEXT: s_and_b32 s4, s4, 0xff
-; SI-NEXT: v_readlane_b32 s5, v62, 4
-; SI-NEXT: s_lshl_b32 s4, s4, 16
-; SI-NEXT: s_lshl_b32 s5, s5, 24
-; SI-NEXT: s_or_b32 s63, s5, s4
-; SI-NEXT: v_readlane_b32 s4, v62, 9
-; SI-NEXT: s_and_b32 s5, s4, 0xff
-; SI-NEXT: v_readlane_b32 s4, v62, 8
-; SI-NEXT: s_lshl_b32 s5, s5, 16
-; SI-NEXT: s_lshl_b32 s9, s4, 24
-; SI-NEXT: v_readlane_b32 s4, v62, 0
-; SI-NEXT: s_or_b32 s9, s9, s5
-; SI-NEXT: s_and_b32 s5, s4, 0xff
-; SI-NEXT: s_lshl_b32 s10, s29, 8
-; SI-NEXT: s_or_b32 s4, s5, s10
-; SI-NEXT: v_writelane_b32 v62, s4, 49
-; SI-NEXT: s_and_b32 s5, s74, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 16
-; SI-NEXT: v_writelane_b32 v62, s54, 50
-; SI-NEXT: s_lshl_b32 s11, s21, 24
-; SI-NEXT: s_mov_b32 s18, s22
-; SI-NEXT: s_mov_b32 s22, s21
-; SI-NEXT: s_or_b32 s21, s11, s5
-; SI-NEXT: s_and_b32 s11, s26, 0xff
-; SI-NEXT: v_readlane_b32 s4, v62, 1
-; SI-NEXT: s_lshl_b32 s11, s11, 16
-; SI-NEXT: s_lshl_b32 s12, s4, 24
-; SI-NEXT: s_or_b32 s14, s12, s11
-; SI-NEXT: s_and_b32 s11, s80, 0xff
-; SI-NEXT: s_lshl_b32 s12, s65, 8
-; SI-NEXT: s_or_b32 s12, s11, s12
-; SI-NEXT: s_and_b32 s11, s55, 0xff
-; SI-NEXT: s_lshl_b32 s11, s11, 16
-; SI-NEXT: s_lshl_b32 s13, s64, 24
-; SI-NEXT: s_or_b32 s41, s13, s11
-; SI-NEXT: s_and_b32 s11, s89, 0xff
-; SI-NEXT: s_lshl_b32 s13, s99, 8
-; SI-NEXT: s_or_b32 s13, s11, s13
-; SI-NEXT: s_and_b32 s11, s96, 0xff
-; SI-NEXT: s_lshl_b32 s11, s11, 16
-; SI-NEXT: s_lshl_b32 s15, s76, 24
-; SI-NEXT: s_or_b32 s43, s15, s11
-; SI-NEXT: s_and_b32 s11, s42, 0xff
-; SI-NEXT: s_lshl_b32 s15, s16, 8
-; SI-NEXT: s_or_b32 s16, s11, s15
-; SI-NEXT: s_and_b32 s11, s40, 0xff
-; SI-NEXT: s_lshl_b32 s11, s11, 16
-; SI-NEXT: s_lshl_b32 s15, s87, 24
-; SI-NEXT: s_or_b32 s44, s15, s11
-; SI-NEXT: s_and_b32 s11, s73, 0xff
-; SI-NEXT: s_lshl_b32 s15, s62, 8
-; SI-NEXT: s_or_b32 s62, s11, s15
-; SI-NEXT: s_and_b32 s11, s58, 0xff
-; SI-NEXT: s_lshl_b32 s11, s11, 16
-; SI-NEXT: s_lshl_b32 s15, s45, 24
-; SI-NEXT: s_or_b32 s45, s15, s11
-; SI-NEXT: s_and_b32 s11, s48, 0xff
-; SI-NEXT: s_lshl_b32 s15, s94, 8
-; SI-NEXT: s_or_b32 s10, s11, s15
-; SI-NEXT: s_and_b32 s11, s46, 0xff
-; SI-NEXT: s_lshl_b32 s11, s11, 16
-; SI-NEXT: s_lshl_b32 s15, s47, 24
-; SI-NEXT: s_or_b32 s46, s15, s11
-; SI-NEXT: s_and_b32 s11, s25, 0xff
-; SI-NEXT: s_lshl_b32 s15, s34, 8
-; SI-NEXT: s_or_b32 s94, s11, s15
-; SI-NEXT: s_and_b32 s11, s72, 0xff
-; SI-NEXT: s_lshl_b32 s11, s11, 16
-; SI-NEXT: s_lshl_b32 s15, s60, 24
-; SI-NEXT: s_or_b32 s47, s15, s11
-; SI-NEXT: s_and_b32 s11, s61, 0xff
-; SI-NEXT: s_lshl_b32 s15, s59, 8
-; SI-NEXT: s_or_b32 s73, s11, s15
-; SI-NEXT: s_and_b32 s11, s56, 0xff
-; SI-NEXT: s_lshl_b32 s11, s11, 16
-; SI-NEXT: s_lshl_b32 s15, s57, 24
-; SI-NEXT: v_readlane_b32 s4, v62, 20
-; SI-NEXT: s_or_b32 s56, s15, s11
-; SI-NEXT: s_and_b32 s11, s38, 0xff
-; SI-NEXT: s_lshl_b32 s15, s4, 8
-; SI-NEXT: s_or_b32 s48, s11, s15
-; SI-NEXT: s_and_b32 s11, s92, 0xff
-; SI-NEXT: s_lshl_b32 s11, s11, 16
-; SI-NEXT: s_lshl_b32 s15, s23, 24
-; SI-NEXT: s_or_b32 vcc_lo, s15, s11
-; SI-NEXT: s_and_b32 s11, s20, 0xff
-; SI-NEXT: s_lshl_b32 s15, s85, 8
-; SI-NEXT: s_or_b32 s72, s11, s15
-; SI-NEXT: s_and_b32 s11, s83, 0xff
-; SI-NEXT: s_lshl_b32 s11, s11, 16
-; SI-NEXT: s_lshl_b32 s15, s84, 24
-; SI-NEXT: s_or_b32 vcc_hi, s15, s11
-; SI-NEXT: s_and_b32 s11, s93, 0xff
-; SI-NEXT: s_lshl_b32 s15, s97, 8
-; SI-NEXT: s_or_b32 s57, s11, s15
-; SI-NEXT: s_and_b32 s11, s67, 0xff
-; SI-NEXT: s_lshl_b32 s11, s11, 16
-; SI-NEXT: s_lshl_b32 s15, s51, 24
-; SI-NEXT: v_writelane_b32 v62, s67, 51
-; SI-NEXT: s_mov_b32 s67, s51
-; SI-NEXT: s_mov_b32 s51, s74
-; SI-NEXT: s_or_b32 s74, s15, s11
-; SI-NEXT: s_and_b32 s11, s98, 0xff
-; SI-NEXT: s_lshl_b32 s15, s75, 8
-; SI-NEXT: v_readlane_b32 s4, v62, 18
-; SI-NEXT: v_writelane_b32 v62, s87, 52
-; SI-NEXT: s_or_b32 s58, s11, s15
-; SI-NEXT: s_and_b32 s11, s4, 0xff
-; SI-NEXT: v_writelane_b32 v62, s25, 53
-; SI-NEXT: s_lshl_b32 s11, s11, 16
-; SI-NEXT: s_lshl_b32 s15, s81, 24
-; SI-NEXT: v_readlane_b32 s4, v62, 15
-; SI-NEXT: s_mov_b32 s54, s75
-; SI-NEXT: s_or_b32 s75, s15, s11
-; SI-NEXT: s_and_b32 s11, s77, 0xff
-; SI-NEXT: s_lshl_b32 s15, s4, 8
-; SI-NEXT: v_readlane_b32 s4, v62, 14
-; SI-NEXT: s_or_b32 s59, s11, s15
-; SI-NEXT: s_and_b32 s11, s4, 0xff
-; SI-NEXT: v_readlane_b32 s4, v62, 13
-; SI-NEXT: s_lshl_b32 s11, s11, 16
-; SI-NEXT: s_lshl_b32 s15, s4, 24
-; SI-NEXT: v_readlane_b32 s4, v62, 12
-; SI-NEXT: s_mov_b32 s95, s69
-; SI-NEXT: s_mov_b32 s69, s76
-; SI-NEXT: s_or_b32 s76, s15, s11
-; SI-NEXT: s_and_b32 s11, s86, 0xff
-; SI-NEXT: s_lshl_b32 s15, s4, 8
-; SI-NEXT: v_readlane_b32 s4, v62, 11
-; SI-NEXT: s_or_b32 s60, s11, s15
-; SI-NEXT: s_and_b32 s11, s4, 0xff
-; SI-NEXT: s_lshl_b32 s11, s11, 16
-; SI-NEXT: s_lshl_b32 s15, s78, 24
-; SI-NEXT: s_mov_b32 s93, s99
-; SI-NEXT: s_mov_b32 s99, s84
-; SI-NEXT: s_mov_b32 s84, s77
-; SI-NEXT: s_or_b32 s77, s15, s11
-; SI-NEXT: s_and_b32 s11, s82, 0xff
-; SI-NEXT: s_lshl_b32 s15, s53, 8
-; SI-NEXT: s_or_b32 s61, s11, s15
-; SI-NEXT: s_and_b32 s11, s31, 0xff
-; SI-NEXT: s_lshl_b32 s11, s11, 16
-; SI-NEXT: s_lshl_b32 s15, s50, 24
-; SI-NEXT: s_mov_b32 s4, s85
-; SI-NEXT: s_mov_b32 s85, s83
-; SI-NEXT: s_mov_b32 s83, s82
-; SI-NEXT: s_mov_b32 s82, s53
-; SI-NEXT: s_mov_b32 s53, s50
-; SI-NEXT: s_mov_b32 s50, s31
-; SI-NEXT: s_mov_b32 s31, s78
-; SI-NEXT: s_or_b32 s78, s15, s11
-; SI-NEXT: v_readlane_b32 s11, v62, 10
-; SI-NEXT: s_and_b32 s11, s11, 0xff
-; SI-NEXT: s_lshl_b32 s15, s17, 8
-; SI-NEXT: s_or_b32 s11, s11, s15
-; SI-NEXT: s_and_b32 s11, s11, 0xffff
-; SI-NEXT: v_mov_b32_e32 v51, s9
-; SI-NEXT: s_or_b32 s6, s11, s9
-; SI-NEXT: v_readlane_b32 s9, v62, 3
-; SI-NEXT: v_readlane_b32 s11, v62, 2
-; SI-NEXT: s_and_b32 s9, s9, 0xff
-; SI-NEXT: s_lshl_b32 s15, s11, 8
-; SI-NEXT: s_or_b32 s9, s9, s15
-; SI-NEXT: s_and_b32 s9, s9, 0xffff
-; SI-NEXT: s_or_b32 s17, s9, s14
-; SI-NEXT: v_readlane_b32 s9, v62, 22
-; SI-NEXT: v_mov_b32_e32 v52, s14
-; SI-NEXT: s_and_b32 s14, s9, 0xff
-; SI-NEXT: v_readlane_b32 s9, v62, 21
-; SI-NEXT: s_lshl_b32 s15, s9, 8
-; SI-NEXT: s_or_b32 s14, s14, s15
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: v_or_b32_e32 v53, v6, v1
-; SI-NEXT: s_and_b32 s14, s14, 0xffff
-; SI-NEXT: v_or_b32_e32 v50, s14, v53
-; SI-NEXT: s_and_b32 s14, s30, 0xff
-; SI-NEXT: s_lshl_b32 s15, s88, 8
-; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v20
-; SI-NEXT: s_or_b32 s14, s14, s15
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: v_or_b32_e32 v54, v14, v1
-; SI-NEXT: s_and_b32 s14, s14, 0xffff
-; SI-NEXT: v_or_b32_e32 v17, s14, v54
-; SI-NEXT: s_and_b32 s14, s8, 0xff
-; SI-NEXT: s_lshl_b32 s15, s52, 8
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v19
-; SI-NEXT: s_or_b32 s14, s14, s15
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: v_or_b32_e32 v55, v40, v1
-; SI-NEXT: s_and_b32 s14, s14, 0xffff
-; SI-NEXT: v_or_b32_e32 v16, s14, v55
-; SI-NEXT: s_and_b32 s14, s35, 0xff
-; SI-NEXT: s_lshl_b32 s15, s28, 8
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v29
-; SI-NEXT: s_or_b32 s14, s14, s15
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: s_waitcnt expcnt(4)
-; SI-NEXT: v_or_b32_e32 v40, v42, v1
-; SI-NEXT: s_and_b32 s14, s14, 0xffff
-; SI-NEXT: v_or_b32_e32 v15, s14, v40
-; SI-NEXT: s_and_b32 s14, s95, 0xff
-; SI-NEXT: s_lshl_b32 s15, s66, 8
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v56
-; SI-NEXT: s_or_b32 s14, s14, s15
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: v_or_b32_e32 v41, v61, v1
-; SI-NEXT: s_and_b32 s14, s14, 0xffff
-; SI-NEXT: v_or_b32_e32 v12, s14, v41
-; SI-NEXT: s_and_b32 s14, s18, 0xff
-; SI-NEXT: s_lshl_b32 s15, s91, 8
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v57
-; SI-NEXT: s_or_b32 s14, s14, s15
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: s_waitcnt expcnt(3)
-; SI-NEXT: v_or_b32_e32 v42, v60, v1
-; SI-NEXT: s_and_b32 s14, s14, 0xffff
-; SI-NEXT: v_or_b32_e32 v11, s14, v42
-; SI-NEXT: s_and_b32 s14, s37, 0xff
-; SI-NEXT: s_lshl_b32 s15, s19, 8
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v33
-; SI-NEXT: s_or_b32 s14, s14, s15
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: s_mov_b32 s91, s6
-; SI-NEXT: v_or_b32_e32 v59, v31, v1
-; SI-NEXT: s_and_b32 s14, s14, 0xffff
-; SI-NEXT: v_readlane_b32 s6, v62, 50
-; SI-NEXT: v_or_b32_e32 v10, s14, v59
-; SI-NEXT: s_and_b32 s14, s6, 0xff
-; SI-NEXT: s_lshl_b32 s15, s70, 8
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v30
-; SI-NEXT: s_or_b32 s14, s14, s15
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: v_or_b32_e32 v5, v24, v1
-; SI-NEXT: s_and_b32 s14, s14, 0xffff
-; SI-NEXT: v_readlane_b32 s6, v62, 19
-; SI-NEXT: v_or_b32_e32 v9, s14, v5
-; SI-NEXT: s_and_b32 s14, s90, 0xff
-; SI-NEXT: s_lshl_b32 s15, s6, 8
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v35
-; SI-NEXT: s_or_b32 s14, s14, s15
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: v_or_b32_e32 v13, v25, v1
-; SI-NEXT: s_and_b32 s14, s14, 0xffff
-; SI-NEXT: v_or_b32_e32 v8, s14, v13
-; SI-NEXT: s_and_b32 s14, s27, 0xff
-; SI-NEXT: s_lshl_b32 s15, s24, 8
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v32
-; SI-NEXT: s_or_b32 s14, s14, s15
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: v_mov_b32_e32 v26, v31
-; SI-NEXT: v_or_b32_e32 v31, v27, v1
-; SI-NEXT: s_and_b32 s14, s14, 0xffff
-; SI-NEXT: v_or_b32_e32 v7, s14, v31
-; SI-NEXT: s_and_b32 s14, s49, 0xff
-; SI-NEXT: s_lshl_b32 s15, s68, 8
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v38
-; SI-NEXT: s_or_b32 s14, s14, s15
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: v_mov_b32_e32 v28, v25
-; SI-NEXT: v_mov_b32_e32 v25, v60
-; SI-NEXT: v_or_b32_e32 v60, v43, v1
-; SI-NEXT: s_and_b32 s14, s14, 0xffff
-; SI-NEXT: v_readlane_b32 s6, v62, 17
-; SI-NEXT: v_or_b32_e32 v4, s14, v60
-; SI-NEXT: s_and_b32 s14, s6, 0xff
-; SI-NEXT: v_readlane_b32 s6, v62, 16
-; SI-NEXT: s_lshl_b32 s15, s6, 8
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v39
-; SI-NEXT: s_or_b32 s14, s14, s15
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: v_mov_b32_e32 v23, v27
-; SI-NEXT: v_mov_b32_e32 v27, v24
-; SI-NEXT: v_mov_b32_e32 v24, v61
-; SI-NEXT: v_or_b32_e32 v61, v44, v1
-; SI-NEXT: s_and_b32 s14, s14, 0xffff
-; SI-NEXT: v_or_b32_e32 v2, s14, v61
-; SI-NEXT: s_and_b32 s14, s71, 0xff
-; SI-NEXT: s_lshl_b32 s15, s36, 8
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v48
-; SI-NEXT: s_or_b32 s14, s14, s15
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: v_readlane_b32 s6, v62, 48
-; SI-NEXT: v_or_b32_e32 v6, v45, v1
-; SI-NEXT: s_and_b32 s14, s14, 0xffff
-; SI-NEXT: s_and_b32 s8, s6, 0xffff
-; SI-NEXT: v_readlane_b32 s6, v62, 49
-; SI-NEXT: v_or_b32_e32 v1, s14, v6
-; SI-NEXT: s_and_b32 s14, s79, 0xff
-; SI-NEXT: s_lshl_b32 s15, s7, 8
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v34
-; SI-NEXT: s_or_b32 s42, s8, s63
-; SI-NEXT: s_and_b32 s8, s6, 0xffff
-; SI-NEXT: s_or_b32 s14, s14, s15
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: s_or_b32 s40, s8, s21
-; SI-NEXT: s_and_b32 s8, s12, 0xffff
-; SI-NEXT: v_or_b32_e32 v14, v46, v3
-; SI-NEXT: s_and_b32 s14, s14, 0xffff
-; SI-NEXT: s_or_b32 s15, s8, s41
-; SI-NEXT: s_and_b32 s8, s13, 0xffff
-; SI-NEXT: v_or_b32_e32 v3, s14, v14
-; SI-NEXT: s_or_b32 s14, s8, s43
-; SI-NEXT: s_and_b32 s8, s16, 0xffff
-; SI-NEXT: s_and_b32 s16, s73, 0xffff
-; SI-NEXT: s_or_b32 s13, s8, s44
-; SI-NEXT: s_and_b32 s8, s62, 0xffff
-; SI-NEXT: s_or_b32 s35, s16, s56
-; SI-NEXT: s_and_b32 s16, s48, 0xffff
-; SI-NEXT: s_or_b32 s12, s8, s45
-; SI-NEXT: s_and_b32 s8, s10, 0xffff
-; SI-NEXT: s_or_b32 s52, s16, vcc_lo
-; SI-NEXT: s_and_b32 s16, s72, 0xffff
-; SI-NEXT: s_or_b32 s10, s8, s46
-; SI-NEXT: s_and_b32 s8, s94, 0xffff
-; SI-NEXT: s_or_b32 s94, s16, vcc_hi
-; SI-NEXT: s_and_b32 s16, s57, 0xffff
-; SI-NEXT: s_or_b32 s49, s16, s74
-; SI-NEXT: s_and_b32 s16, s58, 0xffff
-; SI-NEXT: s_or_b32 s48, s16, s75
-; SI-NEXT: s_and_b32 s16, s59, 0xffff
-; SI-NEXT: s_mov_b32 s25, s23
-; SI-NEXT: s_or_b32 s11, s16, s76
-; SI-NEXT: s_and_b32 s16, s60, 0xffff
-; SI-NEXT: s_and_b32 s23, s61, 0xffff
-; SI-NEXT: s_mov_b32 s87, s34
-; SI-NEXT: s_mov_b32 s34, s55
-; SI-NEXT: s_mov_b32 s55, s22
-; SI-NEXT: s_or_b32 s8, s8, s47
-; SI-NEXT: s_or_b32 s9, s16, s77
-; SI-NEXT: s_or_b32 s16, s23, s78
-; SI-NEXT: s_mov_b32 s22, s18
-; SI-NEXT: v_mov_b32_e32 v49, v30
-; SI-NEXT: v_mov_b32_e32 v36, v35
-; SI-NEXT: v_mov_b32_e32 v37, v32
-; SI-NEXT: v_mov_b32_e32 v18, v43
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_mov_b32_e32 v33, v44
-; SI-NEXT: v_mov_b32_e32 v35, v45
-; SI-NEXT: v_mov_b32_e32 v22, v46
-; SI-NEXT: v_alignbit_b32 v57, s42, v51, 16
-; SI-NEXT: v_alignbit_b32 v58, s40, v52, 16
-; SI-NEXT: v_alignbit_b32 v56, s15, v53, 16
-; SI-NEXT: v_alignbit_b32 v47, s14, v54, 16
-; SI-NEXT: v_alignbit_b32 v46, s13, v55, 16
-; SI-NEXT: v_alignbit_b32 v45, s12, v40, 16
-; SI-NEXT: v_alignbit_b32 v44, s10, v41, 16
-; SI-NEXT: v_alignbit_b32 v43, s8, v42, 16
-; SI-NEXT: v_alignbit_b32 v42, s35, v59, 16
-; SI-NEXT: v_alignbit_b32 v41, s52, v5, 16
-; SI-NEXT: v_alignbit_b32 v40, s94, v13, 16
-; SI-NEXT: v_alignbit_b32 v55, s49, v31, 16
-; SI-NEXT: v_mov_b32_e32 v31, v26
-; SI-NEXT: v_alignbit_b32 v54, s48, v60, 16
-; SI-NEXT: v_mov_b32_e32 v60, v25
-; SI-NEXT: v_mov_b32_e32 v25, v28
-; SI-NEXT: v_alignbit_b32 v53, s11, v61, 16
-; SI-NEXT: v_mov_b32_e32 v61, v24
-; SI-NEXT: v_mov_b32_e32 v24, v27
-; SI-NEXT: v_alignbit_b32 v52, s9, v6, 16
-; SI-NEXT: v_alignbit_b32 v51, s16, v14, 16
-; SI-NEXT: s_lshr_b32 s73, s63, 16
-; SI-NEXT: s_lshr_b32 s72, s21, 16
-; SI-NEXT: s_lshr_b32 s63, s41, 16
-; SI-NEXT: s_lshr_b32 s62, s43, 16
-; SI-NEXT: s_lshr_b32 s61, s44, 16
-; SI-NEXT: s_lshr_b32 s60, s45, 16
-; SI-NEXT: s_lshr_b32 s59, s46, 16
-; SI-NEXT: s_lshr_b32 s58, s47, 16
-; SI-NEXT: s_lshr_b32 s57, s56, 16
-; SI-NEXT: s_lshr_b32 s56, vcc_lo, 16
-; SI-NEXT: s_lshr_b32 s47, vcc_hi, 16
-; SI-NEXT: s_lshr_b32 s46, s74, 16
-; SI-NEXT: s_mov_b32 s74, s51
-; SI-NEXT: s_mov_b32 s51, s67
-; SI-NEXT: v_readlane_b32 s67, v62, 51
-; SI-NEXT: s_lshr_b32 s45, s75, 16
-; SI-NEXT: s_mov_b32 s23, s25
-; SI-NEXT: s_mov_b32 s21, s55
-; SI-NEXT: s_mov_b32 s55, s34
-; SI-NEXT: s_mov_b32 s75, s54
-; SI-NEXT: s_mov_b32 s34, s87
-; SI-NEXT: v_readlane_b32 s25, v62, 53
-; SI-NEXT: v_readlane_b32 s87, v62, 52
-; SI-NEXT: s_lshr_b32 s44, s76, 16
-; SI-NEXT: v_readlane_b32 s54, v62, 50
-; SI-NEXT: s_lshr_b32 s43, s77, 16
-; SI-NEXT: s_mov_b32 s76, s69
-; SI-NEXT: s_mov_b32 s69, s95
-; SI-NEXT: s_mov_b32 s77, s84
-; SI-NEXT: s_mov_b32 s84, s99
-; SI-NEXT: s_mov_b32 s99, s93
-; SI-NEXT: s_lshr_b32 s41, s78, 16
-; SI-NEXT: s_mov_b32 s78, s31
-; SI-NEXT: s_mov_b32 s31, s50
-; SI-NEXT: s_mov_b32 s50, s53
-; SI-NEXT: s_mov_b32 s53, s82
-; SI-NEXT: s_mov_b32 s82, s83
-; SI-NEXT: s_mov_b32 s83, s85
-; SI-NEXT: s_mov_b32 s85, s4
-; SI-NEXT: s_mov_b64 s[4:5], 0
-; SI-NEXT: v_mov_b32_e32 v6, v20
-; SI-NEXT: v_mov_b32_e32 v13, v19
-; SI-NEXT: v_mov_b32_e32 v14, v29
-; SI-NEXT: s_branch .LBB97_3
-; SI-NEXT: .LBB97_2:
-; SI-NEXT: s_waitcnt expcnt(6)
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(5)
-; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; SI-NEXT: v_mov_b32_e32 v22, v46
-; SI-NEXT: v_mov_b32_e32 v37, v32
-; SI-NEXT: v_mov_b32_e32 v36, v35
-; SI-NEXT: v_mov_b32_e32 v35, v45
-; SI-NEXT: v_mov_b32_e32 v49, v30
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_mov_b32_e32 v33, v44
-; SI-NEXT: v_mov_b32_e32 v18, v43
-; SI-NEXT: v_mov_b32_e32 v23, v27
-; SI-NEXT: v_mov_b32_e32 v21, v29
-; SI-NEXT: s_mov_b64 s[4:5], -1
-; SI-NEXT: ; implicit-def: $sgpr91
-; SI-NEXT: ; implicit-def: $vgpr57
-; SI-NEXT: ; implicit-def: $sgpr42
-; SI-NEXT: ; implicit-def: $sgpr73
-; SI-NEXT: ; implicit-def: $sgpr17
-; SI-NEXT: ; implicit-def: $vgpr58
-; SI-NEXT: ; implicit-def: $sgpr40
-; SI-NEXT: ; implicit-def: $sgpr72
-; SI-NEXT: ; implicit-def: $vgpr50
-; SI-NEXT: ; implicit-def: $vgpr56
-; SI-NEXT: ; implicit-def: $sgpr15
-; SI-NEXT: ; implicit-def: $sgpr63
-; SI-NEXT: ; implicit-def: $vgpr17
-; SI-NEXT: ; implicit-def: $vgpr47
-; SI-NEXT: ; implicit-def: $sgpr14
-; SI-NEXT: ; implicit-def: $sgpr62
-; SI-NEXT: ; implicit-def: $vgpr16
-; SI-NEXT: ; implicit-def: $vgpr46
-; SI-NEXT: ; implicit-def: $sgpr13
-; SI-NEXT: ; implicit-def: $sgpr61
-; SI-NEXT: ; implicit-def: $vgpr15
-; SI-NEXT: ; implicit-def: $vgpr45
-; SI-NEXT: ; implicit-def: $sgpr12
-; SI-NEXT: ; implicit-def: $sgpr60
-; SI-NEXT: ; implicit-def: $vgpr12
-; SI-NEXT: ; implicit-def: $vgpr44
-; SI-NEXT: ; implicit-def: $sgpr10
-; SI-NEXT: ; implicit-def: $sgpr59
-; SI-NEXT: ; implicit-def: $vgpr11
-; SI-NEXT: ; implicit-def: $vgpr43
-; SI-NEXT: ; implicit-def: $sgpr8
-; SI-NEXT: ; implicit-def: $sgpr58
-; SI-NEXT: ; implicit-def: $vgpr10
-; SI-NEXT: ; implicit-def: $vgpr42
-; SI-NEXT: ; implicit-def: $sgpr35
-; SI-NEXT: ; implicit-def: $sgpr57
-; SI-NEXT: ; implicit-def: $vgpr9
-; SI-NEXT: ; implicit-def: $vgpr41
-; SI-NEXT: ; implicit-def: $sgpr52
-; SI-NEXT: ; implicit-def: $sgpr56
-; SI-NEXT: ; implicit-def: $vgpr8
-; SI-NEXT: ; implicit-def: $vgpr40
-; SI-NEXT: ; implicit-def: $sgpr94
-; SI-NEXT: ; implicit-def: $sgpr47
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr55
-; SI-NEXT: ; implicit-def: $sgpr49
-; SI-NEXT: ; implicit-def: $sgpr46
-; SI-NEXT: ; implicit-def: $vgpr4
-; SI-NEXT: ; implicit-def: $vgpr54
-; SI-NEXT: ; implicit-def: $sgpr48
-; SI-NEXT: ; implicit-def: $sgpr45
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr53
-; SI-NEXT: ; implicit-def: $sgpr11
-; SI-NEXT: ; implicit-def: $sgpr44
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr52
-; SI-NEXT: ; implicit-def: $sgpr9
-; SI-NEXT: ; implicit-def: $sgpr43
-; SI-NEXT: ; implicit-def: $vgpr3
-; SI-NEXT: ; implicit-def: $vgpr51
-; SI-NEXT: ; implicit-def: $sgpr16
-; SI-NEXT: ; implicit-def: $sgpr41
-; SI-NEXT: .LBB97_3: ; %Flow
-; SI-NEXT: v_mov_b32_e32 v5, v39
-; SI-NEXT: s_andn2_b64 vcc, exec, s[4:5]
-; SI-NEXT: s_cbranch_vccnz .LBB97_5
-; SI-NEXT: ; %bb.4: ; %cmp.true
-; SI-NEXT: s_add_i32 s4, s79, 3
-; SI-NEXT: s_and_b32 s4, s4, 0xff
-; SI-NEXT: s_lshl_b32 s5, s7, 8
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v34
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: v_or_b32_e32 v1, v22, v1
-; SI-NEXT: v_or_b32_e32 v1, s4, v1
-; SI-NEXT: s_add_i32 s4, s82, 3
-; SI-NEXT: s_and_b32 s4, s4, 0xff
-; SI-NEXT: s_lshl_b32 s5, s53, 8
-; SI-NEXT: s_add_i32 s8, s31, 3
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_and_b32 s8, s8, 0xff
-; SI-NEXT: s_lshl_b32 s5, s50, 24
-; SI-NEXT: s_lshl_b32 s8, s8, 16
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: s_or_b32 s5, s5, s8
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_add_i32 s71, s71, 3
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_and_b32 s5, s71, 0xff
-; SI-NEXT: s_lshl_b32 s8, s36, 8
-; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v48
-; SI-NEXT: s_or_b32 s5, s8, s5
-; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT: s_addk_i32 s5, 0x300
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT: v_readlane_b32 s7, v62, 12
-; SI-NEXT: s_and_b32 s5, s5, 0xffff
-; SI-NEXT: v_or_b32_e32 v2, v35, v2
-; SI-NEXT: s_add_i32 s36, s86, 3
-; SI-NEXT: s_lshl_b32 s8, s7, 8
-; SI-NEXT: v_readlane_b32 s7, v62, 11
-; SI-NEXT: v_or_b32_e32 v2, s5, v2
-; SI-NEXT: s_and_b32 s5, s36, 0xff
-; SI-NEXT: s_add_i32 s9, s7, 3
-; SI-NEXT: s_or_b32 s5, s8, s5
-; SI-NEXT: s_and_b32 s9, s9, 0xff
-; SI-NEXT: s_lshl_b32 s8, s78, 24
-; SI-NEXT: s_lshl_b32 s9, s9, 16
-; SI-NEXT: s_addk_i32 s5, 0x300
-; SI-NEXT: s_or_b32 s8, s8, s9
-; SI-NEXT: s_and_b32 s5, s5, 0xffff
-; SI-NEXT: s_or_b32 s5, s8, s5
-; SI-NEXT: s_add_i32 s16, s4, 0x3000000
-; SI-NEXT: v_readlane_b32 s4, v62, 17
-; SI-NEXT: s_add_i32 s9, s5, 0x3000000
-; SI-NEXT: s_add_i32 s79, s4, 3
-; SI-NEXT: v_readlane_b32 s5, v62, 16
-; SI-NEXT: v_add_i32_e32 v3, vcc, 0x3000000, v1
-; SI-NEXT: v_add_i32_e32 v1, vcc, 0x3000000, v2
-; SI-NEXT: s_and_b32 s4, s79, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 8
-; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v5
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: v_or_b32_e32 v2, v33, v2
-; SI-NEXT: v_or_b32_e32 v2, s4, v2
-; SI-NEXT: s_add_i32 s4, s77, 3
-; SI-NEXT: v_readlane_b32 s5, v62, 15
-; SI-NEXT: v_readlane_b32 s6, v62, 14
-; SI-NEXT: s_and_b32 s4, s4, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 8
-; SI-NEXT: s_add_i32 s8, s6, 3
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: v_readlane_b32 s5, v62, 13
-; SI-NEXT: s_and_b32 s8, s8, 0xff
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: s_lshl_b32 s5, s5, 24
-; SI-NEXT: s_lshl_b32 s8, s8, 16
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s5, s8
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s11, s4, 0x3000000
-; SI-NEXT: v_readlane_b32 s4, v62, 32
-; SI-NEXT: s_add_i32 s53, s4, 3
-; SI-NEXT: s_and_b32 s4, s53, 0xff
-; SI-NEXT: s_lshl_b32 s5, s68, 8
-; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v38
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: v_or_b32_e32 v4, v18, v4
-; SI-NEXT: s_add_i32 s93, s98, 3
-; SI-NEXT: v_readlane_b32 s6, v62, 18
-; SI-NEXT: v_or_b32_e32 v4, s4, v4
-; SI-NEXT: s_and_b32 s4, s93, 0xff
-; SI-NEXT: s_lshl_b32 s5, s75, 8
-; SI-NEXT: s_add_i32 s8, s6, 3
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_and_b32 s8, s8, 0xff
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: s_lshl_b32 s5, s81, 24
-; SI-NEXT: s_lshl_b32 s8, s8, 16
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s5, s8
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s86, s27, 3
-; SI-NEXT: s_add_i32 s48, s4, 0x3000000
-; SI-NEXT: s_and_b32 s4, s86, 0xff
-; SI-NEXT: s_lshl_b32 s5, s24, 8
-; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v37
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: v_and_b32_e32 v5, 0xff, v5
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: v_or_b32_e32 v5, v23, v5
-; SI-NEXT: v_or_b32_e32 v5, s4, v5
-; SI-NEXT: v_readlane_b32 s4, v62, 38
-; SI-NEXT: s_add_i32 s68, s4, 3
-; SI-NEXT: s_and_b32 s4, s68, 0xff
-; SI-NEXT: s_lshl_b32 s5, s97, 8
-; SI-NEXT: s_add_i32 s8, s67, 3
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_and_b32 s8, s8, 0xff
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: s_lshl_b32 s5, s51, 24
-; SI-NEXT: s_lshl_b32 s8, s8, 16
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s5, s8
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s50, s90, 3
-; SI-NEXT: v_readlane_b32 s5, v62, 19
-; SI-NEXT: v_add_i32_e32 v7, vcc, 0x3000000, v5
-; SI-NEXT: s_add_i32 s49, s4, 0x3000000
-; SI-NEXT: s_and_b32 s4, s50, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 8
-; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v36
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: v_and_b32_e32 v5, 0xff, v5
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: v_or_b32_e32 v5, v25, v5
-; SI-NEXT: s_add_i32 s94, s20, 3
-; SI-NEXT: v_or_b32_e32 v5, s4, v5
-; SI-NEXT: s_and_b32 s4, s94, 0xff
-; SI-NEXT: s_lshl_b32 s5, s85, 8
-; SI-NEXT: s_add_i32 s8, s83, 3
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_and_b32 s8, s8, 0xff
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: s_lshl_b32 s5, s84, 24
-; SI-NEXT: s_lshl_b32 s8, s8, 16
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s5, s8
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s20, s54, 3
-; SI-NEXT: v_add_i32_e32 v8, vcc, 0x3000000, v5
-; SI-NEXT: s_add_i32 s94, s4, 0x3000000
-; SI-NEXT: s_and_b32 s4, s20, 0xff
-; SI-NEXT: s_lshl_b32 s5, s70, 8
-; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v49
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: v_and_b32_e32 v5, 0xff, v5
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: v_or_b32_e32 v5, v24, v5
-; SI-NEXT: v_or_b32_e32 v5, s4, v5
-; SI-NEXT: v_add_i32_e32 v9, vcc, 0x3000000, v5
-; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; SI-NEXT: s_add_i32 s98, s38, 3
-; SI-NEXT: v_readlane_b32 s5, v62, 20
-; SI-NEXT: s_and_b32 s4, s98, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 8
-; SI-NEXT: s_add_i32 s8, s92, 3
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_and_b32 s8, s8, 0xff
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: s_lshl_b32 s5, s23, 24
-; SI-NEXT: s_lshl_b32 s8, s8, 16
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s5, s8
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s17, s37, 3
-; SI-NEXT: s_add_i32 s52, s4, 0x3000000
-; SI-NEXT: s_and_b32 s4, s17, 0xff
-; SI-NEXT: s_lshl_b32 s5, s19, 8
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: v_readlane_b32 s5, v62, 34
-; SI-NEXT: v_readlane_b32 s6, v62, 29
-; SI-NEXT: s_lshl_b32 s5, s5, 8
-; SI-NEXT: s_add_i32 s38, s6, 3
-; SI-NEXT: s_and_b32 s8, s38, 0xff
-; SI-NEXT: s_lshl_b32 s8, s8, 16
-; SI-NEXT: s_add_i32 s22, s22, 3
-; SI-NEXT: s_add_i32 s85, s25, 3
-; SI-NEXT: v_readlane_b32 s6, v62, 40
-; SI-NEXT: s_add_i32 s70, s6, 3
-; SI-NEXT: s_and_b32 s7, s70, 0xff
-; SI-NEXT: s_lshl_b32 s7, s7, 16
-; SI-NEXT: s_add_i32 s19, s69, 3
-; SI-NEXT: s_add_i32 s51, s30, 3
-; SI-NEXT: s_add_i32 s95, s89, 3
-; SI-NEXT: v_mov_b32_e32 v30, s16
-; SI-NEXT: v_mov_b32_e32 v39, s9
-; SI-NEXT: v_add_i32_e32 v2, vcc, 0x3000000, v2
-; SI-NEXT: v_mov_b32_e32 v28, s11
-; SI-NEXT: v_add_i32_e32 v4, vcc, 0x3000000, v4
-; SI-NEXT: v_mov_b32_e32 v27, s48
-; SI-NEXT: v_mov_b32_e32 v26, s49
-; SI-NEXT: v_mov_b32_e32 v25, s94
-; SI-NEXT: v_mov_b32_e32 v24, s52
-; SI-NEXT: v_alignbit_b32 v41, v24, v9, 16
-; SI-NEXT: v_alignbit_b32 v40, v25, v8, 16
-; SI-NEXT: v_alignbit_b32 v55, v26, v7, 16
-; SI-NEXT: v_alignbit_b32 v54, v27, v4, 16
-; SI-NEXT: v_alignbit_b32 v53, v28, v2, 16
-; SI-NEXT: v_alignbit_b32 v52, v39, v1, 16
-; SI-NEXT: v_alignbit_b32 v51, v30, v3, 16
-; SI-NEXT: s_lshr_b32 s56, s52, 16
-; SI-NEXT: s_lshr_b32 s47, s94, 16
-; SI-NEXT: s_lshr_b32 s46, s49, 16
-; SI-NEXT: s_lshr_b32 s45, s48, 16
-; SI-NEXT: s_lshr_b32 s44, s11, 16
-; SI-NEXT: s_lshr_b32 s43, s9, 16
-; SI-NEXT: s_lshr_b32 s41, s16, 16
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; SI-NEXT: v_and_b32_e32 v5, 0xff, v5
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: v_or_b32_e32 v5, v31, v5
-; SI-NEXT: v_or_b32_e32 v5, s4, v5
-; SI-NEXT: v_add_i32_e32 v10, vcc, 0x3000000, v5
-; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; SI-NEXT: v_readlane_b32 s4, v62, 37
-; SI-NEXT: s_add_i32 s67, s4, 3
-; SI-NEXT: s_and_b32 s4, s67, 0xff
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: v_readlane_b32 s5, v62, 30
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: s_lshl_b32 s5, s5, 24
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s5, s8
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: v_readlane_b32 s5, v62, 46
-; SI-NEXT: s_add_i32 s35, s4, 0x3000000
-; SI-NEXT: s_and_b32 s4, s22, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 8
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s34, 8
-; SI-NEXT: v_mov_b32_e32 v23, s35
-; SI-NEXT: v_alignbit_b32 v42, v23, v10, 16
-; SI-NEXT: s_lshr_b32 s57, s35, 16
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; SI-NEXT: v_and_b32_e32 v5, 0xff, v5
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: v_or_b32_e32 v5, v60, v5
-; SI-NEXT: v_or_b32_e32 v5, s4, v5
-; SI-NEXT: v_add_i32_e32 v11, vcc, 0x3000000, v5
-; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; SI-NEXT: s_and_b32 s4, s85, 0xff
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: v_readlane_b32 s5, v62, 36
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: s_lshl_b32 s5, s5, 24
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s5, s7
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: v_readlane_b32 s5, v62, 47
-; SI-NEXT: s_add_i32 s8, s4, 0x3000000
-; SI-NEXT: s_and_b32 s4, s19, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 8
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: v_readlane_b32 s5, v62, 44
-; SI-NEXT: s_lshl_b32 s5, s5, 8
-; SI-NEXT: v_mov_b32_e32 v22, s8
-; SI-NEXT: v_alignbit_b32 v43, v22, v11, 16
-; SI-NEXT: s_lshr_b32 s58, s8, 16
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5
-; SI-NEXT: v_and_b32_e32 v5, 0xff, v5
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: v_or_b32_e32 v5, v61, v5
-; SI-NEXT: v_or_b32_e32 v5, s4, v5
-; SI-NEXT: v_add_i32_e32 v12, vcc, 0x3000000, v5
-; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v14
-; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; SI-NEXT: v_readlane_b32 s4, v62, 45
-; SI-NEXT: s_add_i32 s6, s4, 3
-; SI-NEXT: s_and_b32 s4, s6, 0xff
-; SI-NEXT: v_readlane_b32 s6, v62, 27
-; SI-NEXT: s_add_i32 s34, s6, 3
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: v_readlane_b32 s5, v62, 28
-; SI-NEXT: s_and_b32 s6, s34, 0xff
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: s_lshl_b32 s5, s5, 24
-; SI-NEXT: s_lshl_b32 s6, s6, 16
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s5, s6
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s10, s4, 0x3000000
-; SI-NEXT: v_readlane_b32 s4, v62, 43
-; SI-NEXT: s_add_i32 s97, s4, 3
-; SI-NEXT: s_and_b32 s4, s97, 0xff
-; SI-NEXT: s_lshl_b32 s5, s28, 8
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: v_and_b32_e32 v5, 0xff, v5
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: v_readlane_b32 s5, v62, 41
-; SI-NEXT: v_readlane_b32 s6, v62, 33
-; SI-NEXT: s_lshl_b32 s5, s5, 8
-; SI-NEXT: s_add_i32 s54, s6, 3
-; SI-NEXT: s_and_b32 s6, s54, 0xff
-; SI-NEXT: s_lshl_b32 s6, s6, 16
-; SI-NEXT: v_mov_b32_e32 v20, s10
-; SI-NEXT: v_alignbit_b32 v44, v20, v12, 16
-; SI-NEXT: s_lshr_b32 s59, s10, 16
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v5, v14, v5
-; SI-NEXT: v_or_b32_e32 v5, s4, v5
-; SI-NEXT: v_add_i32_e32 v15, vcc, 0x3000000, v5
-; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v13
-; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; SI-NEXT: v_readlane_b32 s4, v62, 42
-; SI-NEXT: s_add_i32 s81, s4, 3
-; SI-NEXT: s_and_b32 s4, s81, 0xff
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: v_readlane_b32 s5, v62, 31
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: s_lshl_b32 s5, s5, 24
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s5, s6
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s12, s4, 0x3000000
-; SI-NEXT: v_readlane_b32 s4, v62, 39
-; SI-NEXT: s_add_i32 s69, s4, 3
-; SI-NEXT: v_readlane_b32 s5, v62, 35
-; SI-NEXT: s_and_b32 s4, s69, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 8
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: v_and_b32_e32 v5, 0xff, v5
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: v_readlane_b32 s5, v62, 25
-; SI-NEXT: v_readlane_b32 s6, v62, 24
-; SI-NEXT: s_lshl_b32 s5, s5, 8
-; SI-NEXT: s_add_i32 s92, s6, 3
-; SI-NEXT: s_and_b32 s6, s92, 0xff
-; SI-NEXT: s_lshl_b32 s6, s6, 16
-; SI-NEXT: v_mov_b32_e32 v19, s12
-; SI-NEXT: v_alignbit_b32 v45, v19, v15, 16
-; SI-NEXT: s_lshr_b32 s60, s12, 16
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v5, v13, v5
-; SI-NEXT: v_or_b32_e32 v5, s4, v5
-; SI-NEXT: v_add_i32_e32 v16, vcc, 0x3000000, v5
-; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v6
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v5, 0xff, v5
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: v_readlane_b32 s4, v62, 26
-; SI-NEXT: s_add_i32 s31, s4, 3
-; SI-NEXT: s_and_b32 s4, s31, 0xff
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: s_lshl_b32 s5, s87, 24
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s5, s6
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s13, s4, 0x3000000
-; SI-NEXT: s_and_b32 s4, s51, 0xff
-; SI-NEXT: s_lshl_b32 s5, s88, 8
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s99, 8
-; SI-NEXT: s_add_i32 s6, s96, 3
-; SI-NEXT: s_and_b32 s6, s6, 0xff
-; SI-NEXT: s_lshl_b32 s6, s6, 16
-; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v21
-; SI-NEXT: v_and_b32_e32 v13, 0xff, v13
-; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; SI-NEXT: v_mov_b32_e32 v18, s13
-; SI-NEXT: v_alignbit_b32 v46, v18, v16, 16
-; SI-NEXT: s_lshr_b32 s61, s13, 16
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v5, v6, v5
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; SI-NEXT: v_or_b32_e32 v5, s4, v5
-; SI-NEXT: s_and_b32 s4, s95, 0xff
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: s_lshl_b32 s5, s76, 24
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s5, s6
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s14, s4, 0x3000000
-; SI-NEXT: v_readlane_b32 s4, v62, 22
-; SI-NEXT: s_add_i32 s4, s4, 3
-; SI-NEXT: v_readlane_b32 s5, v62, 21
-; SI-NEXT: s_and_b32 s4, s4, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 8
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s65, 8
-; SI-NEXT: s_add_i32 s6, s55, 3
-; SI-NEXT: s_and_b32 s6, s6, 0xff
-; SI-NEXT: s_lshl_b32 s6, s6, 16
-; SI-NEXT: v_add_i32_e32 v17, vcc, 0x3000000, v5
-; SI-NEXT: v_mov_b32_e32 v5, s14
-; SI-NEXT: v_alignbit_b32 v47, v5, v17, 16
-; SI-NEXT: s_lshr_b32 s62, s14, 16
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v6, v6, v13
-; SI-NEXT: v_or_b32_e32 v6, s4, v6
-; SI-NEXT: s_add_i32 s4, s80, 3
-; SI-NEXT: s_and_b32 s4, s4, 0xff
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: s_lshl_b32 s5, s64, 24
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s5, s6
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s15, s4, 0x3000000
-; SI-NEXT: v_readlane_b32 s4, v62, 3
-; SI-NEXT: s_add_i32 s4, s4, 3
-; SI-NEXT: v_readlane_b32 s5, v62, 2
-; SI-NEXT: s_and_b32 s4, s4, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 8
-; SI-NEXT: s_add_i32 s6, s26, 3
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: v_readlane_b32 s5, v62, 1
-; SI-NEXT: s_and_b32 s6, s6, 0xff
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: s_lshl_b32 s5, s5, 24
-; SI-NEXT: s_lshl_b32 s6, s6, 16
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s5, s6
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s17, s4, 0x3000000
-; SI-NEXT: v_readlane_b32 s4, v62, 0
-; SI-NEXT: s_add_i32 s4, s4, 3
-; SI-NEXT: s_and_b32 s4, s4, 0xff
-; SI-NEXT: s_lshl_b32 s5, s29, 8
-; SI-NEXT: s_add_i32 s6, s74, 3
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_and_b32 s6, s6, 0xff
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: s_lshl_b32 s5, s21, 24
-; SI-NEXT: s_lshl_b32 s6, s6, 16
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s5, s6
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s40, s4, 0x3000000
-; SI-NEXT: v_readlane_b32 s4, v62, 10
-; SI-NEXT: s_add_i32 s4, s4, 3
-; SI-NEXT: v_readlane_b32 s5, v62, 23
-; SI-NEXT: v_readlane_b32 s6, v62, 9
-; SI-NEXT: s_and_b32 s4, s4, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 8
-; SI-NEXT: s_add_i32 s6, s6, 3
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: v_readlane_b32 s5, v62, 8
-; SI-NEXT: s_and_b32 s6, s6, 0xff
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: s_lshl_b32 s5, s5, 24
-; SI-NEXT: s_lshl_b32 s6, s6, 16
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s5, s6
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s91, s4, 0x3000000
-; SI-NEXT: v_readlane_b32 s4, v62, 7
-; SI-NEXT: s_add_i32 s4, s4, 3
-; SI-NEXT: v_readlane_b32 s5, v62, 6
-; SI-NEXT: v_readlane_b32 s6, v62, 5
-; SI-NEXT: s_and_b32 s4, s4, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 8
-; SI-NEXT: s_add_i32 s6, s6, 3
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: v_readlane_b32 s5, v62, 4
-; SI-NEXT: s_and_b32 s6, s6, 0xff
-; SI-NEXT: s_addk_i32 s4, 0x300
-; SI-NEXT: s_lshl_b32 s5, s5, 24
-; SI-NEXT: s_lshl_b32 s6, s6, 16
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s5, s6
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s42, s4, 0x3000000
-; SI-NEXT: v_mov_b32_e32 v13, s91
-; SI-NEXT: v_add_i32_e32 v50, vcc, 0x3000000, v6
-; SI-NEXT: v_mov_b32_e32 v6, s15
-; SI-NEXT: v_alignbit_b32 v57, s42, v13, 16
-; SI-NEXT: v_mov_b32_e32 v13, s17
-; SI-NEXT: v_alignbit_b32 v58, s40, v13, 16
-; SI-NEXT: v_alignbit_b32 v56, v6, v50, 16
-; SI-NEXT: s_lshr_b32 s73, s42, 16
-; SI-NEXT: s_lshr_b32 s72, s40, 16
-; SI-NEXT: s_lshr_b32 s63, s15, 16
-; SI-NEXT: .LBB97_5: ; %end
-; SI-NEXT: s_and_b32 s4, s91, 0xffff
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v57
-; SI-NEXT: v_or_b32_e32 v5, s4, v5
-; SI-NEXT: s_and_b32 s4, s42, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s73, 16
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: buffer_store_dword v5, v0, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v5, vcc, 4, v0
-; SI-NEXT: s_waitcnt vmcnt(3)
-; SI-NEXT: v_mov_b32_e32 v6, s4
-; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
-; SI-NEXT: s_and_b32 s4, s17, 0xffff
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v58
-; SI-NEXT: v_or_b32_e32 v5, s4, v5
-; SI-NEXT: s_and_b32 s4, s40, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s72, 16
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v6, vcc, 8, v0
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v5, vcc, 12, v0
-; SI-NEXT: v_mov_b32_e32 v6, s4
-; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
-; SI-NEXT: v_and_b32_e32 v5, 0xffff, v50
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v56
-; SI-NEXT: s_and_b32 s4, s15, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s63, 16
-; SI-NEXT: v_or_b32_e32 v5, v5, v6
-; SI-NEXT: v_add_i32_e32 v6, vcc, 16, v0
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v5, vcc, 20, v0
-; SI-NEXT: v_mov_b32_e32 v6, s4
-; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
-; SI-NEXT: v_and_b32_e32 v5, 0xffff, v17
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v47
-; SI-NEXT: s_and_b32 s4, s14, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s62, 16
-; SI-NEXT: v_or_b32_e32 v5, v5, v6
-; SI-NEXT: v_add_i32_e32 v6, vcc, 24, v0
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v5, vcc, 28, v0
-; SI-NEXT: v_mov_b32_e32 v6, s4
-; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
-; SI-NEXT: v_and_b32_e32 v5, 0xffff, v16
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v46
-; SI-NEXT: s_and_b32 s4, s13, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s61, 16
-; SI-NEXT: v_or_b32_e32 v5, v5, v6
-; SI-NEXT: v_add_i32_e32 v6, vcc, 32, v0
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v5, vcc, 36, v0
-; SI-NEXT: v_mov_b32_e32 v6, s4
-; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
-; SI-NEXT: v_and_b32_e32 v5, 0xffff, v15
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v45
-; SI-NEXT: s_and_b32 s4, s12, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s60, 16
-; SI-NEXT: v_or_b32_e32 v5, v5, v6
-; SI-NEXT: v_add_i32_e32 v6, vcc, 40, v0
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v5, vcc, 44, v0
-; SI-NEXT: v_mov_b32_e32 v6, s4
-; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
-; SI-NEXT: v_and_b32_e32 v5, 0xffff, v12
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v44
-; SI-NEXT: s_and_b32 s4, s10, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s59, 16
-; SI-NEXT: v_or_b32_e32 v5, v5, v6
-; SI-NEXT: v_add_i32_e32 v6, vcc, 48, v0
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v5, vcc, 52, v0
-; SI-NEXT: v_mov_b32_e32 v6, s4
-; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
-; SI-NEXT: v_and_b32_e32 v5, 0xffff, v11
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v43
-; SI-NEXT: s_and_b32 s4, s8, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s58, 16
-; SI-NEXT: v_or_b32_e32 v5, v5, v6
-; SI-NEXT: v_add_i32_e32 v6, vcc, 56, v0
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v5, vcc, 60, v0
-; SI-NEXT: v_mov_b32_e32 v6, s4
-; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
-; SI-NEXT: v_and_b32_e32 v5, 0xffff, v10
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v42
-; SI-NEXT: s_and_b32 s4, s35, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s57, 16
-; SI-NEXT: v_or_b32_e32 v5, v5, v6
-; SI-NEXT: v_add_i32_e32 v6, vcc, 64, v0
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v5, vcc, 0x44, v0
-; SI-NEXT: v_mov_b32_e32 v6, s4
-; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
-; SI-NEXT: v_and_b32_e32 v5, 0xffff, v9
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v41
-; SI-NEXT: s_and_b32 s4, s52, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s56, 16
-; SI-NEXT: v_or_b32_e32 v5, v5, v6
-; SI-NEXT: v_add_i32_e32 v6, vcc, 0x48, v0
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v5, vcc, 0x4c, v0
-; SI-NEXT: v_mov_b32_e32 v6, s4
-; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
-; SI-NEXT: v_and_b32_e32 v5, 0xffff, v8
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v40
-; SI-NEXT: s_and_b32 s4, s94, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s47, 16
-; SI-NEXT: v_or_b32_e32 v5, v5, v6
-; SI-NEXT: v_add_i32_e32 v6, vcc, 0x50, v0
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v5, vcc, 0x54, v0
-; SI-NEXT: v_mov_b32_e32 v6, s4
-; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
-; SI-NEXT: v_and_b32_e32 v5, 0xffff, v7
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v55
-; SI-NEXT: s_and_b32 s4, s49, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s46, 16
-; SI-NEXT: v_or_b32_e32 v5, v5, v6
-; SI-NEXT: v_add_i32_e32 v6, vcc, 0x58, v0
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v5, vcc, 0x5c, v0
-; SI-NEXT: v_mov_b32_e32 v6, s4
-; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
-; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v54
-; SI-NEXT: s_and_b32 s4, s48, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s45, 16
-; SI-NEXT: v_or_b32_e32 v4, v4, v5
-; SI-NEXT: v_add_i32_e32 v5, vcc, 0x60, v0
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: buffer_store_dword v4, v5, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v4, vcc, 0x64, v0
-; SI-NEXT: v_mov_b32_e32 v5, s4
-; SI-NEXT: buffer_store_dword v5, v4, s[0:3], 0 offen
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v53
-; SI-NEXT: s_and_b32 s4, s11, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s44, 16
-; SI-NEXT: v_or_b32_e32 v2, v2, v4
-; SI-NEXT: v_add_i32_e32 v4, vcc, 0x68, v0
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: buffer_store_dword v2, v4, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v2, vcc, 0x6c, v0
-; SI-NEXT: v_mov_b32_e32 v4, s4
-; SI-NEXT: buffer_store_dword v4, v2, s[0:3], 0 offen
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v52
-; SI-NEXT: s_and_b32 s4, s9, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s43, 16
-; SI-NEXT: v_or_b32_e32 v1, v1, v2
-; SI-NEXT: v_add_i32_e32 v2, vcc, 0x70, v0
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 0x74, v0
-; SI-NEXT: v_mov_b32_e32 v2, s4
-; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v3
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v51
-; SI-NEXT: s_and_b32 s4, s16, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s41, 16
-; SI-NEXT: v_or_b32_e32 v1, v1, v2
-; SI-NEXT: v_add_i32_e32 v2, vcc, 0x78, v0
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; SI-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_mov_b32_e32 v1, s4
-; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; SI-NEXT: v_readlane_b32 s99, v63, 35
-; SI-NEXT: v_readlane_b32 s98, v63, 34
-; SI-NEXT: v_readlane_b32 s97, v63, 33
-; SI-NEXT: v_readlane_b32 s96, v63, 32
-; SI-NEXT: v_readlane_b32 s87, v63, 31
-; SI-NEXT: v_readlane_b32 s86, v63, 30
-; SI-NEXT: v_readlane_b32 s85, v63, 29
-; SI-NEXT: v_readlane_b32 s84, v63, 28
-; SI-NEXT: v_readlane_b32 s83, v63, 27
-; SI-NEXT: v_readlane_b32 s82, v63, 26
-; SI-NEXT: v_readlane_b32 s81, v63, 25
-; SI-NEXT: v_readlane_b32 s80, v63, 24
-; SI-NEXT: v_readlane_b32 s71, v63, 23
-; SI-NEXT: v_readlane_b32 s70, v63, 22
-; SI-NEXT: v_readlane_b32 s69, v63, 21
-; SI-NEXT: v_readlane_b32 s68, v63, 20
-; SI-NEXT: v_readlane_b32 s67, v63, 19
-; SI-NEXT: v_readlane_b32 s66, v63, 18
-; SI-NEXT: v_readlane_b32 s65, v63, 17
-; SI-NEXT: v_readlane_b32 s64, v63, 16
-; SI-NEXT: v_readlane_b32 s55, v63, 15
-; SI-NEXT: v_readlane_b32 s54, v63, 14
-; SI-NEXT: v_readlane_b32 s53, v63, 13
-; SI-NEXT: v_readlane_b32 s52, v63, 12
-; SI-NEXT: v_readlane_b32 s51, v63, 11
-; SI-NEXT: v_readlane_b32 s50, v63, 10
-; SI-NEXT: v_readlane_b32 s49, v63, 9
-; SI-NEXT: v_readlane_b32 s48, v63, 8
-; SI-NEXT: v_readlane_b32 s39, v63, 7
-; SI-NEXT: v_readlane_b32 s38, v63, 6
-; SI-NEXT: v_readlane_b32 s37, v63, 5
-; SI-NEXT: v_readlane_b32 s36, v63, 4
-; SI-NEXT: v_readlane_b32 s35, v63, 3
-; SI-NEXT: v_readlane_b32 s34, v63, 2
-; SI-NEXT: v_readlane_b32 s31, v63, 1
-; SI-NEXT: v_readlane_b32 s30, v63, 0
-; SI-NEXT: s_or_saveexec_b64 s[4:5], -1
-; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; SI-NEXT: s_mov_b64 exec, s[4:5]
-; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; SI-NEXT: s_setpc_b64 s[30:31]
-;
-; VI-LABEL: bitcast_v128i8_to_v64i16_scalar:
-; 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:396 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; VI-NEXT: v_mov_b32_e32 v46, v16
-; VI-NEXT: v_mov_b32_e32 v60, v5
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:332
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:16
-; VI-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:32
-; VI-NEXT: buffer_load_ushort v5, off, s[0:3], s32 offset:48
-; VI-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:64
-; VI-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:80
-; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:96
-; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:112
-; VI-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:128
-; VI-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:144
-; VI-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:160
-; VI-NEXT: buffer_load_ushort v37, off, s[0:3], s32 offset:12
-; VI-NEXT: buffer_load_ushort v36, off, s[0:3], s32 offset:44
-; VI-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:60
-; VI-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:56
-; VI-NEXT: v_lshlrev_b32_e32 v11, 8, v11
-; VI-NEXT: v_mov_b32_e32 v62, v21
-; VI-NEXT: v_mov_b32_e32 v47, v17
-; VI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; VI-NEXT: v_lshlrev_b32_e32 v7, 8, v7
-; VI-NEXT: buffer_load_ushort v54, off, s[0:3], s32 offset:76
-; VI-NEXT: buffer_load_ushort v25, off, s[0:3], s32 offset:124
-; VI-NEXT: buffer_load_ushort v41, off, s[0:3], s32 offset:324
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v1
-; VI-NEXT: s_and_b64 s[4:5], vcc, exec
-; VI-NEXT: s_waitcnt vmcnt(12)
-; VI-NEXT: v_lshlrev_b32_e32 v59, 8, v26
-; VI-NEXT: s_waitcnt vmcnt(10)
-; VI-NEXT: v_lshlrev_b32_e32 v58, 8, v10
-; VI-NEXT: s_waitcnt vmcnt(9)
-; VI-NEXT: v_lshlrev_b32_e32 v21, 8, v18
-; VI-NEXT: s_waitcnt vmcnt(7)
-; VI-NEXT: v_lshlrev_b32_e32 v28, 8, v14
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:52
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:92
-; VI-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:88
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:84
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:108
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:140
-; VI-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:28
-; VI-NEXT: buffer_load_ushort v38, off, s[0:3], s32 offset:24
-; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:20
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v11, 8, v15
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v31
-; VI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v11, 8, v19
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v5
-; VI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v11, 8, v23
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v6
-; VI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v11, 8, v27
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v30
-; VI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v23, off, s[0:3], s32 offset:156
-; VI-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:152
-; VI-NEXT: buffer_load_ushort v5, off, s[0:3], s32 offset:148
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:176
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:192
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:208
-; VI-NEXT: buffer_load_ushort v35, off, s[0:3], s32 offset:172
-; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:188
-; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:184
-; VI-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:180
-; VI-NEXT: buffer_load_ushort v55, off, s[0:3], s32 offset:204
-; VI-NEXT: v_lshlrev_b32_e32 v31, 8, v22
-; VI-NEXT: buffer_load_ushort v27, off, s[0:3], s32 offset:276
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(8)
-; VI-NEXT: v_lshlrev_b32_e32 v16, 8, v2
-; 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:240
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:256
-; VI-NEXT: buffer_load_ushort v11, off, s[0:3], s32 offset:220
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:216
-; VI-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:212
-; VI-NEXT: buffer_load_ushort v11, off, s[0:3], s32 offset:236
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v2
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v11, off, s[0:3], s32 offset:252
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:272
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:288
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:268
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v0
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:284
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v1
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:280
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:304
-; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:320
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:300
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b32_e32 v1, 8, v1
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:316
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:312
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v43, off, s[0:3], s32 offset:308
-; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:328
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; VI-NEXT: v_lshlrev_b32_e32 v2, 8, v0
-; VI-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:120
-; VI-NEXT: buffer_load_ushort v19, off, s[0:3], s32 offset:116
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:248
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:244
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:296
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:292
-; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:264
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; VI-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:260
-; VI-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:232
-; VI-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:228
-; VI-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:200
-; VI-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:196
-; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:168
-; VI-NEXT: buffer_load_ushort v61, off, s[0:3], s32 offset:164
-; VI-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:136
-; VI-NEXT: buffer_load_ushort v40, off, s[0:3], s32 offset:132
-; VI-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:104
-; VI-NEXT: buffer_load_ushort v34, off, s[0:3], s32 offset:100
-; VI-NEXT: buffer_load_ushort v15, off, s[0:3], s32 offset:72
-; VI-NEXT: buffer_load_ushort v29, off, s[0:3], s32 offset:68
-; VI-NEXT: buffer_load_ushort v17, off, s[0:3], s32 offset:40
-; VI-NEXT: buffer_load_ushort v33, off, s[0:3], s32 offset:36
-; VI-NEXT: buffer_load_ushort v11, off, s[0:3], s32 offset:8
-; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:4
-; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: buffer_store_dword v50, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v37, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v52, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v48, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; VI-NEXT: s_cbranch_scc0 .LBB97_2
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b32_e32 v1, 8, v9
-; VI-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v60
-; VI-NEXT: v_mov_b32_e32 v28, v26
-; VI-NEXT: v_mov_b32_e32 v26, v23
-; VI-NEXT: v_mov_b32_e32 v23, v5
-; VI-NEXT: v_or_b32_sdwa v0, v4, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_mov_b32_e32 v31, v22
-; VI-NEXT: v_mov_b32_e32 v22, v6
-; VI-NEXT: v_mov_b32_e32 v59, v10
-; VI-NEXT: v_mov_b32_e32 v58, v43
-; VI-NEXT: v_mov_b32_e32 v43, v27
-; VI-NEXT: v_mov_b32_e32 v27, v14
-; VI-NEXT: s_and_b32 s4, s28, 0xff
-; VI-NEXT: s_lshl_b32 s5, s29, 8
-; VI-NEXT: s_or_b32 s4, s4, s5
-; VI-NEXT: s_and_b32 s4, s4, 0xffff
-; VI-NEXT: s_lshl_b32 s5, s17, 8
-; VI-NEXT: s_lshl_b32 s6, s19, 8
-; VI-NEXT: s_lshl_b32 s7, s23, 8
-; VI-NEXT: s_lshl_b32 s8, s27, 8
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; VI-NEXT: v_or_b32_sdwa v4, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v13
-; VI-NEXT: v_or_b32_sdwa v0, v12, v0 dst_sel:WORD_1 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, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v5, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; VI-NEXT: v_mov_b32_e32 v3, v38
-; VI-NEXT: s_waitcnt vmcnt(0)
-; 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 v6, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; 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:416 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v47
-; VI-NEXT: v_or_b32_sdwa v0, v46, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; 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_or_b32_sdwa v7, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v62
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; 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:404 ; 4-byte Folded Reload
-; 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_or_b32_sdwa v8, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v0
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; 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:400 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(2)
-; 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_or_b32_sdwa v9, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; VI-NEXT: v_mov_b32_e32 v2, v63
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v0
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v46, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v10, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v11
-; VI-NEXT: v_or_b32_sdwa v0, v44, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; 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 v11, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v3
-; VI-NEXT: v_or_b32_sdwa v0, v2, v0 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:484 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v1, v37, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v12, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v17
-; VI-NEXT: v_or_b32_sdwa v0, v33, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v2, v54, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_mov_b32_e32 v33, v36
-; VI-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_or_b32_sdwa v3, v24, v3 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 v1, v52, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; VI-NEXT: v_or_b32_sdwa v13, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v52
-; VI-NEXT: v_or_b32_sdwa v0, v37, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v36, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v14, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v15
-; VI-NEXT: v_or_b32_sdwa v0, v29, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_mov_b32_e32 v36, v49
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v48, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v15, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v63
-; VI-NEXT: v_lshlrev_b32_e32 v1, 8, v49
-; VI-NEXT: v_or_b32_sdwa v0, v62, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v34, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v16, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v17, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; 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:492 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v18
-; VI-NEXT: v_or_b32_sdwa v0, v19, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_mov_b32_e32 v48, v32
-; VI-NEXT: v_mov_b32_e32 v34, v40
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v18, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v57
-; VI-NEXT: v_or_b32_sdwa v0, v40, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v1, v25, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v19, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v22
-; VI-NEXT: v_or_b32_sdwa v0, v23, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_mov_b32_e32 v40, v60
-; VI-NEXT: v_mov_b32_e32 v57, v61
-; VI-NEXT: s_waitcnt vmcnt(0)
-; 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 v20, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v32
-; VI-NEXT: v_or_b32_sdwa v0, v61, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_mov_b32_e32 v61, v38
-; VI-NEXT: s_waitcnt vmcnt(0)
-; 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 v21, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v59
-; VI-NEXT: v_or_b32_sdwa v0, v27, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v35, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v22, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v56
-; VI-NEXT: v_or_b32_sdwa v0, v51, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_mov_b32_e32 v56, v45
-; VI-NEXT: v_mov_b32_e32 v51, v42
-; VI-NEXT: s_waitcnt vmcnt(0)
-; 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 v23, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v28
-; VI-NEXT: v_or_b32_sdwa v0, v31, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v55, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v24, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v45
-; VI-NEXT: v_or_b32_sdwa v0, v50, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; VI-NEXT: v_mov_b32_e32 v50, v53
-; VI-NEXT: v_mov_b32_e32 v55, v63
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v25, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v59
-; VI-NEXT: v_or_b32_sdwa v0, v53, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v26, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v53
-; VI-NEXT: v_or_b32_sdwa v0, v42, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_mov_b32_e32 v42, v44
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v27, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v0
-; VI-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v28, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v49
-; VI-NEXT: v_or_b32_sdwa v0, v39, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_mov_b32_e32 v43, v46
-; VI-NEXT: v_mov_b32_e32 v39, v41
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v29, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(3)
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v0
-; VI-NEXT: v_or_b32_sdwa v0, v58, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v30, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v54
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; VI-NEXT: v_or_b32_sdwa v0, v41, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; VI-NEXT: v_mov_b32_e32 v41, v52
-; VI-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; VI-NEXT: v_mov_b32_e32 v63, v54
-; VI-NEXT: v_mov_b32_e32 v54, v49
-; VI-NEXT: v_mov_b32_e32 v49, v53
-; VI-NEXT: s_waitcnt vmcnt(7)
-; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_sdwa v31, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v38
-; VI-NEXT: v_or_b32_sdwa v0, v47, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT: v_or_b32_e32 v3, s4, v0
-; VI-NEXT: s_and_b32 s4, s16, 0xff
-; VI-NEXT: s_or_b32 s4, s4, s5
-; VI-NEXT: s_and_b32 s5, s18, 0xff
-; VI-NEXT: s_or_b32 s5, s5, s6
-; VI-NEXT: s_and_b32 s4, s4, 0xffff
-; VI-NEXT: s_lshl_b32 s5, s5, 16
-; VI-NEXT: s_or_b32 s4, s4, s5
-; VI-NEXT: s_and_b32 s5, s20, 0xff
-; VI-NEXT: s_lshl_b32 s6, s21, 8
-; VI-NEXT: s_or_b32 s5, s5, s6
-; VI-NEXT: s_and_b32 s6, s22, 0xff
-; VI-NEXT: s_or_b32 s6, s6, s7
-; VI-NEXT: s_and_b32 s5, s5, 0xffff
-; VI-NEXT: s_lshl_b32 s6, s6, 16
-; VI-NEXT: s_or_b32 s5, s5, s6
-; VI-NEXT: s_and_b32 s6, s24, 0xff
-; VI-NEXT: s_lshl_b32 s7, s25, 8
-; VI-NEXT: s_or_b32 s6, s6, s7
-; VI-NEXT: s_and_b32 s7, s26, 0xff
-; VI-NEXT: s_or_b32 s7, s7, s8
-; VI-NEXT: s_and_b32 s6, s6, 0xffff
-; VI-NEXT: s_lshl_b32 s7, s7, 16
-; VI-NEXT: s_or_b32 s6, s6, s7
-; VI-NEXT: v_mov_b32_e32 v0, s4
-; VI-NEXT: v_mov_b32_e32 v1, s5
-; VI-NEXT: v_mov_b32_e32 v2, s6
-; VI-NEXT: s_mov_b64 s[4:5], 0
-; VI-NEXT: s_branch .LBB97_3
-; VI-NEXT: .LBB97_2:
-; VI-NEXT: v_mov_b32_e32 v34, v40
-; VI-NEXT: v_mov_b32_e32 v57, v61
-; VI-NEXT: v_mov_b32_e32 v48, v32
-; VI-NEXT: v_mov_b32_e32 v56, v45
-; VI-NEXT: v_mov_b32_e32 v51, v42
-; VI-NEXT: v_mov_b32_e32 v39, v41
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; VI-NEXT: v_mov_b32_e32 v33, v36
-; VI-NEXT: v_mov_b32_e32 v36, v49
-; VI-NEXT: v_mov_b32_e32 v35, v63
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; VI-NEXT: v_mov_b32_e32 v50, v53
-; VI-NEXT: s_mov_b64 s[4:5], -1
-; VI-NEXT: v_mov_b32_e32 v52, v38
-; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; VI-NEXT: ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; VI-NEXT: .LBB97_3: ; %Flow
-; VI-NEXT: v_mov_b32_e32 v38, v48
-; VI-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT: s_andn2_b64 vcc, exec, s[4:5]
-; VI-NEXT: s_cbranch_vccnz .LBB97_5
-; VI-NEXT: ; %bb.4: ; %cmp.true
-; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(14)
-; VI-NEXT: v_add_u32_e32 v29, vcc, 3, v47
-; VI-NEXT: v_and_b32_e32 v29, 0xff, v29
-; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; VI-NEXT: s_add_i32 s28, s28, 3
-; VI-NEXT: s_lshl_b32 s4, s29, 8
-; VI-NEXT: s_and_b32 s5, s28, 0xff
-; VI-NEXT: s_or_b32 s4, s4, s5
-; VI-NEXT: s_addk_i32 s4, 0x300
-; VI-NEXT: s_and_b32 s4, s4, 0xffff
-; VI-NEXT: v_mov_b32_e32 v53, v34
-; VI-NEXT: s_add_i32 s24, s24, 3
-; VI-NEXT: s_add_i32 s20, s20, 3
-; VI-NEXT: s_lshl_b32 s9, s25, 8
-; VI-NEXT: s_and_b32 s10, s24, 0xff
-; VI-NEXT: s_add_i32 s16, s16, 3
-; VI-NEXT: s_lshl_b32 s8, s21, 8
-; VI-NEXT: s_or_b32 s9, s9, s10
-; VI-NEXT: s_and_b32 s10, s20, 0xff
-; VI-NEXT: s_add_i32 s18, s18, 3
-; VI-NEXT: s_lshl_b32 s7, s17, 8
-; VI-NEXT: s_or_b32 s8, s8, s10
-; VI-NEXT: s_and_b32 s10, s16, 0xff
-; VI-NEXT: s_or_b32 s7, s7, s10
-; VI-NEXT: s_and_b32 s12, s18, 0xff
-; VI-NEXT: s_lshl_b32 s6, s19, 24
-; VI-NEXT: s_add_i32 s22, s22, 3
-; VI-NEXT: s_addk_i32 s7, 0x300
-; VI-NEXT: s_lshl_b32 s12, s12, 16
-; VI-NEXT: s_add_i32 s26, s26, 3
-; VI-NEXT: s_addk_i32 s8, 0x300
-; VI-NEXT: s_and_b32 s11, s22, 0xff
-; VI-NEXT: s_and_b32 s7, s7, 0xffff
-; VI-NEXT: s_or_b32 s6, s6, s12
-; VI-NEXT: s_lshl_b32 s5, s23, 24
-; VI-NEXT: v_add_u32_e32 v39, vcc, 3, v39
-; VI-NEXT: s_and_b32 s10, s26, 0xff
-; VI-NEXT: s_or_b32 s6, s6, s7
-; VI-NEXT: s_and_b32 s7, s8, 0xffff
-; VI-NEXT: s_lshl_b32 s8, s11, 16
-; VI-NEXT: s_addk_i32 s9, 0x300
-; VI-NEXT: v_and_b32_e32 v39, 0xff, v39
-; VI-NEXT: s_or_b32 s5, s5, s8
-; VI-NEXT: s_lshl_b32 s8, s10, 16
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_lshlrev_b32_e32 v31, 24, v63
-; VI-NEXT: s_or_b32 s5, s5, s7
-; VI-NEXT: s_and_b32 s7, s9, 0xffff
-; VI-NEXT: s_add_i32 s6, s6, 0x3000000
-; VI-NEXT: s_add_i32 s5, s5, 0x3000000
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(6)
-; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT: s_waitcnt vmcnt(5)
-; VI-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(2)
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
-; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
-; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: v_or_b32_sdwa v22, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v23, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v24, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v25, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v25, vcc, 0x300, v25
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v26, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v27, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v28, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v28, vcc, 0x300, v28
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v21, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v21, vcc, 0x300, v21
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v20, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v20, vcc, 0x300, v20
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v19, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v19, vcc, 0x300, v19
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v18, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v18, vcc, 0x300, v18
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v16, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v16, vcc, 0x300, v16
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v17, v4, v3 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:480 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v48
-; VI-NEXT: v_add_u32_e32 v17, vcc, 0x300, v17
-; VI-NEXT: v_add_u32_e32 v48, vcc, 0x300, v23
-; VI-NEXT: v_add_u32_e32 v23, vcc, 0x300, v27
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v15, v4, v3 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:476 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v33
-; VI-NEXT: v_add_u32_e32 v15, vcc, 0x300, v15
-; VI-NEXT: v_lshlrev_b32_e32 v33, 24, v59
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v14, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v14, vcc, 0x300, v14
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v13, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v13, vcc, 0x300, v13
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v12, v4, v3 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:460 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v42
-; VI-NEXT: v_add_u32_e32 v12, vcc, 0x300, v12
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v11, v4, v3 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:456 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v43
-; VI-NEXT: v_add_u32_e32 v11, vcc, 0x300, v11
-; VI-NEXT: v_add_u32_e32 v43, vcc, 3, v50
-; VI-NEXT: v_add_u32_e32 v50, vcc, 0x300, v24
-; VI-NEXT: v_add_u32_e32 v24, vcc, 0x300, v26
-; VI-NEXT: v_and_b32_e32 v26, 0xff, v43
-; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v26
-; VI-NEXT: v_or_b32_e32 v26, v33, v26
-; VI-NEXT: v_or_b32_sdwa v26, v26, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: v_add_u32_e32 v26, vcc, 0x3000000, v26
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v10, v4, v3 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:452 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v44
-; VI-NEXT: v_add_u32_e32 v10, vcc, 0x300, v10
-; VI-NEXT: v_add_u32_e32 v44, vcc, 3, v51
-; VI-NEXT: v_and_b32_e32 v27, 0xff, v44
-; VI-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v9, v4, v3 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:448 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v45
-; VI-NEXT: v_add_u32_e32 v9, vcc, 0x300, v9
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v8, v4, v3 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:444 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v46
-; VI-NEXT: v_add_u32_e32 v8, vcc, 0x300, v8
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v7, v4, v3 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:440 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v40
-; VI-NEXT: v_add_u32_e32 v7, vcc, 0x300, v7
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v6, v4, v3 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:740 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v58
-; VI-NEXT: v_add_u32_e32 v6, vcc, 0x300, v6
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v5, v4, v3 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:732 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v32
-; VI-NEXT: v_add_u32_e32 v5, vcc, 0x300, v5
-; VI-NEXT: v_lshlrev_b32_e32 v32, 24, v49
-; VI-NEXT: v_or_b32_e32 v27, v32, v27
-; VI-NEXT: v_or_b32_sdwa v27, v27, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: v_add_u32_e32 v27, vcc, 0x3000000, v27
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT: v_lshlrev_b32_e32 v4, 24, v61
-; VI-NEXT: v_or_b32_e32 v4, v4, v29
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v3, vcc, 0x300, v3
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
-; VI-NEXT: v_or_b32_e32 v29, v29, v30
-; VI-NEXT: v_or_b32_sdwa v29, v29, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: v_or_b32_e32 v3, s4, v4
-; VI-NEXT: v_add_u32_e32 v4, vcc, 0x3000000, v29
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; VI-NEXT: s_lshl_b32 s4, s27, 24
-; VI-NEXT: s_or_b32 s4, s4, s8
-; VI-NEXT: s_or_b32 s4, s4, s7
-; VI-NEXT: s_add_i32 s4, s4, 0x3000000
-; VI-NEXT: v_add_u32_e32 v3, vcc, 0x3000000, v3
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
-; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; VI-NEXT: v_or_b32_e32 v29, v29, v30
-; VI-NEXT: v_or_b32_sdwa v5, v29, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v5, vcc, 0x3000000, v5
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
-; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; VI-NEXT: v_or_b32_e32 v29, v29, v30
-; VI-NEXT: v_or_b32_sdwa v6, v29, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v6, vcc, 0x3000000, v6
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
-; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; VI-NEXT: v_or_b32_e32 v29, v29, v30
-; VI-NEXT: v_or_b32_sdwa v7, v29, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v7, vcc, 0x3000000, v7
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
-; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; VI-NEXT: v_or_b32_e32 v29, v29, v30
-; VI-NEXT: v_or_b32_sdwa v8, v29, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v8, vcc, 0x3000000, v8
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
-; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; VI-NEXT: v_or_b32_e32 v29, v29, v30
-; VI-NEXT: v_or_b32_sdwa v9, v29, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v9, vcc, 0x3000000, v9
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
-; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; VI-NEXT: v_or_b32_e32 v29, v29, v30
-; VI-NEXT: v_or_b32_sdwa v10, v29, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v10, vcc, 0x3000000, v10
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
-; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; VI-NEXT: v_or_b32_e32 v29, v29, v30
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v35
-; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
-; VI-NEXT: v_or_b32_sdwa v11, v29, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v52
-; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; VI-NEXT: v_or_b32_e32 v29, v29, v30
-; VI-NEXT: v_or_b32_sdwa v12, v29, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; VI-NEXT: v_lshlrev_b32_e32 v35, 24, v56
-; VI-NEXT: v_add_u32_e32 v11, vcc, 0x3000000, v11
-; VI-NEXT: v_add_u32_e32 v12, vcc, 0x3000000, v12
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
-; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; VI-NEXT: v_or_b32_e32 v29, v29, v30
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v37
-; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
-; VI-NEXT: v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v41
-; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; VI-NEXT: v_or_b32_e32 v29, v29, v30
-; VI-NEXT: v_or_b32_sdwa v14, v29, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v37, vcc, 0x300, v2
-; VI-NEXT: v_add_u32_e32 v13, vcc, 0x3000000, v13
-; VI-NEXT: v_add_u32_e32 v14, vcc, 0x3000000, v14
-; VI-NEXT: v_mov_b32_e32 v2, s4
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
-; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; VI-NEXT: v_or_b32_e32 v29, v29, v30
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v62
-; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
-; VI-NEXT: v_or_b32_sdwa v15, v29, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v55
-; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; VI-NEXT: v_or_b32_e32 v29, v29, v30
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; VI-NEXT: v_or_b32_sdwa v17, v29, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v36
-; VI-NEXT: v_add_u32_e32 v36, vcc, 0x300, v1
-; VI-NEXT: v_add_u32_e32 v15, vcc, 0x3000000, v15
-; VI-NEXT: v_mov_b32_e32 v1, s5
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
-; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; VI-NEXT: v_or_b32_e32 v29, v29, v30
-; VI-NEXT: v_or_b32_sdwa v29, v29, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: v_add_u32_e32 v16, vcc, 0x3000000, v17
-; VI-NEXT: v_add_u32_e32 v17, vcc, 0x3000000, v29
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
-; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; VI-NEXT: v_or_b32_e32 v29, v29, v30
-; VI-NEXT: v_or_b32_sdwa v18, v29, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v53
-; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
-; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; VI-NEXT: v_add_u32_e32 v18, vcc, 0x3000000, v18
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
-; VI-NEXT: v_or_b32_e32 v29, v29, v30
-; VI-NEXT: v_or_b32_sdwa v19, v29, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v19, vcc, 0x3000000, v19
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
-; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; VI-NEXT: v_or_b32_e32 v29, v29, v30
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v57
-; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
-; VI-NEXT: v_or_b32_sdwa v20, v29, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v38
-; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; VI-NEXT: v_or_b32_e32 v29, v29, v30
-; VI-NEXT: v_or_b32_sdwa v21, v29, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v38, vcc, 0x300, v22
-; VI-NEXT: v_add_u32_e32 v20, vcc, 0x3000000, v20
-; VI-NEXT: v_add_u32_e32 v21, vcc, 0x3000000, v21
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
-; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
-; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
-; VI-NEXT: v_or_b32_e32 v29, v29, v30
-; VI-NEXT: v_or_b32_sdwa v55, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v54
-; VI-NEXT: v_lshlrev_b32_e32 v54, 24, v34
-; VI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; VI-NEXT: v_add_u32_e32 v22, vcc, 0x3000000, v55
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b32_e32 v40, 24, v34
-; VI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b32_e32 v30, 24, v28
-; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_add_u32_e32 v41, vcc, 3, v34
-; VI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(1)
-; VI-NEXT: v_lshlrev_b32_e32 v28, 24, v28
-; VI-NEXT: v_and_b32_e32 v41, 0xff, v41
-; VI-NEXT: v_lshlrev_b32_e32 v55, 16, v41
-; VI-NEXT: v_or_b32_e32 v55, v40, v55
-; VI-NEXT: v_or_b32_sdwa v23, v55, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: v_add_u32_e32 v23, vcc, 0x3000000, v23
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v42, vcc, 3, v34
-; VI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; VI-NEXT: v_and_b32_e32 v42, 0xff, v42
-; VI-NEXT: v_lshlrev_b32_e32 v55, 16, v42
-; VI-NEXT: v_or_b32_e32 v54, v54, v55
-; VI-NEXT: v_or_b32_sdwa v24, v54, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: v_add_u32_e32 v24, vcc, 0x3000000, v24
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v53, vcc, 3, v34
-; VI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; VI-NEXT: v_and_b32_e32 v53, 0xff, v53
-; VI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
-; VI-NEXT: v_or_b32_e32 v35, v35, v53
-; VI-NEXT: v_or_b32_sdwa v25, v35, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: v_add_u32_e32 v25, vcc, 0x3000000, v25
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v52, vcc, 3, v34
-; VI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; VI-NEXT: v_and_b32_e32 v52, 0xff, v52
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v52
-; VI-NEXT: v_or_b32_e32 v28, v28, v32
-; VI-NEXT: v_or_b32_sdwa v28, v28, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: v_add_u32_e32 v28, vcc, 0x3000000, v28
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v51, vcc, 3, v34
-; VI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; VI-NEXT: v_and_b32_e32 v51, 0xff, v51
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v51
-; VI-NEXT: v_or_b32_e32 v29, v29, v32
-; VI-NEXT: v_or_b32_sdwa v29, v29, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: v_add_u32_e32 v29, vcc, 0x3000000, v29
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: v_add_u32_e32 v49, vcc, 3, v34
-; VI-NEXT: v_and_b32_e32 v49, 0xff, v49
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v49
-; VI-NEXT: v_or_b32_e32 v30, v30, v32
-; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v39
-; VI-NEXT: v_add_u32_e32 v34, vcc, 0x300, v0
-; VI-NEXT: v_or_b32_e32 v31, v31, v32
-; VI-NEXT: v_or_b32_sdwa v30, v30, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: v_or_b32_sdwa v31, v31, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; VI-NEXT: v_mov_b32_e32 v0, s6
-; VI-NEXT: v_add_u32_e32 v30, vcc, 0x3000000, v30
-; VI-NEXT: v_add_u32_e32 v31, vcc, 0x3000000, v31
-; VI-NEXT: .LBB97_5: ; %end
-; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: bitcast_v128i8_to_v64i16_scalar:
-; 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:396 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:332
-; GFX9-NEXT: buffer_load_ushort v45, off, s[0:3], s32
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:8
-; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:16
-; GFX9-NEXT: buffer_load_ushort v41, off, s[0:3], s32 offset:24
-; GFX9-NEXT: buffer_load_ushort v40, off, s[0:3], s32 offset:32
-; GFX9-NEXT: buffer_load_ushort v55, off, s[0:3], s32 offset:40
-; GFX9-NEXT: buffer_load_ushort v54, off, s[0:3], s32 offset:48
-; GFX9-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:56
-; GFX9-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:64
-; GFX9-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:72
-; GFX9-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:80
-; GFX9-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:88
-; GFX9-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:96
-; GFX9-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:104
-; GFX9-NEXT: buffer_load_ushort v30, 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 v32, off, s[0:3], s32 offset:128
-; GFX9-NEXT: buffer_load_ushort v33, off, s[0:3], s32 offset:136
-; GFX9-NEXT: buffer_load_ushort v34, off, s[0:3], s32 offset:144
-; GFX9-NEXT: buffer_load_ushort v35, off, s[0:3], s32 offset:152
-; GFX9-NEXT: buffer_load_ushort v36, off, s[0:3], s32 offset:160
-; GFX9-NEXT: buffer_load_ushort v37, off, s[0:3], s32 offset:168
-; GFX9-NEXT: buffer_load_ushort v38, off, s[0:3], s32 offset:176
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 8, v1
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v29
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshlrev_b32_e32 v14, 8, v3
-; GFX9-NEXT: v_lshlrev_b32_e32 v46, 8, v5
-; GFX9-NEXT: v_lshlrev_b32_e32 v22, 8, v7
-; GFX9-NEXT: v_lshlrev_b32_e32 v24, 8, v9
-; GFX9-NEXT: v_lshlrev_b32_e32 v26, 8, v11
-; GFX9-NEXT: v_lshlrev_b32_e32 v20, 8, v13
-; GFX9-NEXT: v_lshlrev_b32_e32 v28, 8, v15
-; GFX9-NEXT: v_lshlrev_b32_e32 v18, 8, v17
-; GFX9-NEXT: v_lshlrev_b32_e32 v17, 8, v25
-; GFX9-NEXT: v_lshlrev_b32_e32 v12, 8, v27
-; GFX9-NEXT: v_lshlrev_b32_e32 v19, 8, v19
-; GFX9-NEXT: v_lshlrev_b32_e32 v21, 8, v21
-; GFX9-NEXT: v_lshlrev_b32_e32 v23, 8, v23
-; GFX9-NEXT: s_waitcnt vmcnt(24)
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v43
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v45
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v44
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v42
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v41
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v40
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v55
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v54
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v53
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v52
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v51
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v50
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v49
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v48
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v39
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v30
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v31
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v32
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v33
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v34
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v35
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v36
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v37
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(23)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v38
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v11, off, s[0:3], s32 offset:184
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:192
-; GFX9-NEXT: buffer_load_ushort v15, off, s[0:3], s32 offset:200
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:208
-; GFX9-NEXT: buffer_load_ushort v13, off, s[0:3], s32 offset:216
-; GFX9-NEXT: buffer_load_ushort v5, off, s[0:3], s32 offset:224
-; GFX9-NEXT: buffer_load_ushort v9, off, s[0:3], s32 offset:232
-; GFX9-NEXT: buffer_load_ushort v7, off, s[0:3], s32 offset:240
-; GFX9-NEXT: s_and_b64 s[4:5], vcc, exec
-; GFX9-NEXT: s_waitcnt vmcnt(7)
-; GFX9-NEXT: v_lshlrev_b32_e32 v11, 8, v11
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v1
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v15
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v3
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v13
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v5
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v9
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v7
-; GFX9-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v11, off, s[0:3], s32 offset:248
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:256
-; GFX9-NEXT: buffer_load_ushort v15, off, s[0:3], s32 offset:264
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:272
-; GFX9-NEXT: buffer_load_ushort v13, off, s[0:3], s32 offset:280
-; GFX9-NEXT: buffer_load_ushort v5, off, s[0:3], s32 offset:288
-; GFX9-NEXT: buffer_load_ushort v9, off, s[0:3], s32 offset:296
-; GFX9-NEXT: buffer_load_ushort v7, off, s[0:3], s32 offset:304
-; GFX9-NEXT: s_waitcnt vmcnt(7)
-; GFX9-NEXT: v_lshlrev_b32_e32 v11, 8, v11
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v1
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v15
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v3
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v13
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v5
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v9
-; GFX9-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:312
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:320
-; GFX9-NEXT: buffer_load_ushort v5, off, s[0:3], s32 offset:328
-; GFX9-NEXT: buffer_load_ushort v11, off, s[0:3], s32 offset:4
-; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:12
-; GFX9-NEXT: buffer_load_ushort v55, off, s[0:3], s32 offset:20
-; GFX9-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:28
-; GFX9-NEXT: buffer_load_ushort v13, off, s[0:3], s32 offset:36
-; GFX9-NEXT: s_waitcnt vmcnt(15)
-; GFX9-NEXT: v_lshlrev_b32_e32 v7, 8, v7
-; GFX9-NEXT: s_waitcnt vmcnt(7)
-; GFX9-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GFX9-NEXT: s_waitcnt vmcnt(6)
-; GFX9-NEXT: v_lshlrev_b32_e32 v9, 8, v1
-; GFX9-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:44
-; GFX9-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:52
-; GFX9-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:60
-; GFX9-NEXT: buffer_load_ushort v15, off, s[0:3], s32 offset:68
-; GFX9-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:76
-; GFX9-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:84
-; GFX9-NEXT: buffer_load_ushort v25, off, s[0:3], s32 offset:92
-; GFX9-NEXT: buffer_load_ushort v34, off, s[0:3], s32 offset:100
-; GFX9-NEXT: buffer_load_ushort v33, off, s[0:3], s32 offset:108
-; GFX9-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:116
-; GFX9-NEXT: buffer_load_ushort v43, off, s[0:3], s32 offset:124
-; GFX9-NEXT: buffer_load_ushort v38, off, s[0:3], s32 offset:132
-; GFX9-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:140
-; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:148
-; GFX9-NEXT: buffer_load_ushort v61, off, s[0:3], s32 offset:156
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:164
-; GFX9-NEXT: s_waitcnt vmcnt(21)
-; GFX9-NEXT: v_lshlrev_b32_e32 v5, 8, v5
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:172
-; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:180
-; GFX9-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:188
-; GFX9-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:196
-; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:204
-; GFX9-NEXT: buffer_load_ushort v35, off, s[0:3], s32 offset:212
-; GFX9-NEXT: buffer_load_ushort v36, off, s[0:3], s32 offset:220
-; GFX9-NEXT: buffer_load_ushort v37, off, s[0:3], s32 offset:228
-; GFX9-NEXT: buffer_load_ushort v41, off, s[0:3], s32 offset:236
-; GFX9-NEXT: buffer_load_ushort v54, off, s[0:3], s32 offset:244
-; GFX9-NEXT: buffer_load_ushort v29, off, s[0:3], s32 offset:252
-; GFX9-NEXT: buffer_load_ushort v27, off, s[0:3], s32 offset:260
-; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:268
-; GFX9-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:276
-; GFX9-NEXT: buffer_load_ushort v59, off, s[0:3], s32 offset:284
-; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:292
-; GFX9-NEXT: buffer_load_ushort v62, off, s[0:3], s32 offset:300
-; GFX9-NEXT: buffer_load_ushort v40, off, s[0:3], s32 offset:308
-; GFX9-NEXT: buffer_load_ushort v58, off, s[0:3], s32 offset:316
-; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:324
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v38, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(28)
-; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(30)
-; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(33)
-; GFX9-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(36)
-; GFX9-NEXT: buffer_store_dword v37, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v36, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(39)
-; GFX9-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(41)
-; GFX9-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(41)
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(41)
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(41)
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; GFX9-NEXT: s_waitcnt vmcnt(41)
-; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v48, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v50, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v52, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; GFX9-NEXT: s_cbranch_scc0 .LBB97_2
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: s_and_b32 s4, s28, 0xff
-; GFX9-NEXT: s_lshl_b32 s5, s29, 8
-; GFX9-NEXT: s_or_b32 s4, s4, s5
-; GFX9-NEXT: v_mov_b32_e32 v1, 0xffff
-; GFX9-NEXT: v_and_b32_e32 v3, s4, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX9-NEXT: v_or_b32_sdwa v2, v0, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v4, v4, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v0, v6, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshl_or_b32 v4, v4, 16, v1
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: v_or_b32_sdwa v1, v8, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshl_or_b32 v5, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GFX9-NEXT: v_or_b32_sdwa v0, v10, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: v_lshl_or_b32 v3, v2, 16, v3
-; GFX9-NEXT: s_and_b32 s4, s16, 0xff
-; GFX9-NEXT: s_lshl_b32 s5, s17, 8
-; GFX9-NEXT: s_or_b32 s4, s4, s5
-; GFX9-NEXT: s_and_b32 s5, s18, 0xff
-; GFX9-NEXT: s_lshl_b32 s6, s19, 8
-; GFX9-NEXT: s_or_b32 s5, s5, s6
-; GFX9-NEXT: s_pack_ll_b32_b16 s4, s4, s5
-; GFX9-NEXT: s_and_b32 s5, s20, 0xff
-; GFX9-NEXT: s_lshl_b32 s6, s21, 8
-; GFX9-NEXT: s_or_b32 s5, s5, s6
-; GFX9-NEXT: s_and_b32 s6, s22, 0xff
-; GFX9-NEXT: s_lshl_b32 s7, s23, 8
-; GFX9-NEXT: s_or_b32 s6, s6, s7
-; GFX9-NEXT: s_pack_ll_b32_b16 s5, s5, s6
-; GFX9-NEXT: s_and_b32 s6, s24, 0xff
-; GFX9-NEXT: s_lshl_b32 s7, s25, 8
-; GFX9-NEXT: s_or_b32 s6, s6, s7
-; GFX9-NEXT: s_and_b32 s7, s26, 0xff
-; GFX9-NEXT: s_lshl_b32 s8, s27, 8
-; GFX9-NEXT: s_or_b32 s7, s7, s8
-; GFX9-NEXT: s_pack_ll_b32_b16 s6, s6, s7
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshl_or_b32 v6, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshl_or_b32 v7, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshl_or_b32 v8, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshl_or_b32 v9, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshl_or_b32 v10, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:416 ; 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 v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v2, v39, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_mov_b32_e32 v39, v16
-; GFX9-NEXT: v_or_b32_sdwa v17, v34, v16 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, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshl_or_b32 v11, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v0, v42, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v55, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshl_or_b32 v12, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GFX9-NEXT: v_mov_b32_e32 v42, v61
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v0, v53, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v55, v1
-; 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_lshl_or_b32 v13, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_mov_b32_e32 v53, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v52, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v50, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshl_or_b32 v14, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_mov_b32_e32 v50, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v49, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v49, v1
-; 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_lshl_or_b32 v15, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v0, v48, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: v_lshl_or_b32 v16, v2, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_mov_b32_e32 v48, v1
-; 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: v_and_b32_e32 v1, 0xffff, v1
-; GFX9-NEXT: v_lshl_or_b32 v17, v17, 16, v1
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v0, v33, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: v_mov_b32_e32 v33, v45
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v45, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshl_or_b32 v18, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v38, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshl_or_b32 v19, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v0, v56, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v47, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshl_or_b32 v20, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v0, v42, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v61, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshl_or_b32 v21, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; 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: v_lshl_or_b32 v22, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v0, v51, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v34, 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: v_lshl_or_b32 v23, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(4)
-; GFX9-NEXT: v_mov_b32_e32 v46, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v44, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_or_b32_sdwa v1, v35, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshl_or_b32 v24, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GFX9-NEXT: v_mov_b32_e32 v35, v45
-; GFX9-NEXT: v_mov_b32_e32 v45, v61
-; GFX9-NEXT: v_mov_b32_e32 v61, v42
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v38, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v36, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v37, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshl_or_b32 v25, v1, 16, v0
-; GFX9-NEXT: v_or_b32_sdwa v1, v54, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_mov_b32_e32 v54, v2
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v0, v41, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: v_lshl_or_b32 v26, v1, 16, v0
-; GFX9-NEXT: v_or_b32_sdwa v1, v27, v2 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 v0, v29, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: v_lshl_or_b32 v27, v1, 16, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v60, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GFX9-NEXT: v_or_b32_sdwa v1, v57, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: v_lshl_or_b32 v28, v1, 16, v0
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v0, v59, v60 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v63, v57 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshl_or_b32 v29, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v0, v62, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v40, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshl_or_b32 v30, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v0, v58, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshl_or_b32 v31, v1, 16, v0
-; GFX9-NEXT: v_mov_b32_e32 v0, s4
-; GFX9-NEXT: v_mov_b32_e32 v1, s5
-; GFX9-NEXT: v_mov_b32_e32 v2, s6
-; GFX9-NEXT: s_mov_b64 s[4:5], 0
-; GFX9-NEXT: s_branch .LBB97_3
-; GFX9-NEXT: .LBB97_2:
-; GFX9-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GFX9-NEXT: v_mov_b32_e32 v33, v45
-; GFX9-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; GFX9-NEXT: s_mov_b64 s[4:5], -1
-; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX9-NEXT: ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX9-NEXT: .LBB97_3: ; %Flow
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; GFX9-NEXT: s_andn2_b64 vcc, exec, s[4:5]
-; GFX9-NEXT: s_cbranch_vccnz .LBB97_5
-; GFX9-NEXT: ; %bb.4: ; %cmp.true
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GFX9-NEXT: s_add_i32 s28, s28, 3
-; GFX9-NEXT: s_and_b32 s4, s28, 0xff
-; GFX9-NEXT: s_lshl_b32 s5, s29, 8
-; GFX9-NEXT: s_or_b32 s4, s5, s4
-; GFX9-NEXT: s_addk_i32 s4, 0x300
-; GFX9-NEXT: s_add_i32 s24, s24, 3
-; GFX9-NEXT: s_lshl_b32 s5, s25, 8
-; GFX9-NEXT: s_add_i32 s26, s26, 3
-; GFX9-NEXT: s_lshl_b32 s6, s27, 8
-; GFX9-NEXT: s_add_i32 s20, s20, 3
-; GFX9-NEXT: s_lshl_b32 s7, s21, 8
-; GFX9-NEXT: s_add_i32 s22, s22, 3
-; GFX9-NEXT: s_lshl_b32 s8, s23, 8
-; GFX9-NEXT: s_add_i32 s16, s16, 3
-; GFX9-NEXT: s_lshl_b32 s9, s17, 8
-; GFX9-NEXT: s_add_i32 s18, s18, 3
-; GFX9-NEXT: s_lshl_b32 s10, s19, 8
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(15)
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: s_waitcnt vmcnt(13)
-; GFX9-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u32_e32 v0, 0x300, v0
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(12)
-; GFX9-NEXT: v_add_u32_e32 v25, 3, v25
-; GFX9-NEXT: s_waitcnt vmcnt(11)
-; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v25, v37, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v37, v51, v26 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:708 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u32_e32 v37, 0x300, v37
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
-; GFX9-NEXT: v_lshl_or_b32 v4, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v38, v38, v26 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:704 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_mov_b32_e32 v3, 0xffff
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: v_add_u32_e32 v2, 0x300, v2
-; GFX9-NEXT: v_and_b32_e32 v3, s4, v3
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshl_or_b32 v3, v2, 16, v3
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u32_e32 v0, 0x300, v0
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_and_b32 s4, s24, 0xff
-; GFX9-NEXT: s_or_b32 s4, s5, s4
-; GFX9-NEXT: s_and_b32 s5, s26, 0xff
-; GFX9-NEXT: s_or_b32 s5, s6, s5
-; GFX9-NEXT: s_and_b32 s6, s20, 0xff
-; GFX9-NEXT: s_or_b32 s6, s7, s6
-; GFX9-NEXT: s_and_b32 s7, s22, 0xff
-; GFX9-NEXT: s_or_b32 s7, s8, s7
-; GFX9-NEXT: s_and_b32 s8, s16, 0xff
-; GFX9-NEXT: s_or_b32 s8, s9, s8
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
-; GFX9-NEXT: s_and_b32 s9, s18, 0xff
-; GFX9-NEXT: s_or_b32 s9, s10, s9
-; GFX9-NEXT: s_addk_i32 s4, 0x300
-; GFX9-NEXT: s_addk_i32 s5, 0x300
-; GFX9-NEXT: s_addk_i32 s6, 0x300
-; GFX9-NEXT: s_addk_i32 s7, 0x300
-; GFX9-NEXT: s_addk_i32 s8, 0x300
-; GFX9-NEXT: s_addk_i32 s9, 0x300
-; GFX9-NEXT: s_pack_ll_b32_b16 s8, s8, s9
-; GFX9-NEXT: s_pack_ll_b32_b16 s6, s6, s7
-; GFX9-NEXT: s_pack_ll_b32_b16 s4, s4, s5
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
-; GFX9-NEXT: v_lshl_or_b32 v5, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u32_e32 v0, 0x300, v0
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
-; GFX9-NEXT: v_lshl_or_b32 v6, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u32_e32 v0, 0x300, v0
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
-; GFX9-NEXT: v_lshl_or_b32 v7, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u32_e32 v0, 0x300, v0
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
-; GFX9-NEXT: v_lshl_or_b32 v8, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u32_e32 v0, 0x300, v0
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
-; GFX9-NEXT: v_lshl_or_b32 v9, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u32_e32 v0, 0x300, v0
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
-; GFX9-NEXT: v_lshl_or_b32 v10, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u32_e32 v0, 0x300, v0
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v58, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
-; GFX9-NEXT: v_lshl_or_b32 v11, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u32_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: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u32_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: v_lshl_or_b32 v12, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: v_or_b32_sdwa v0, v42, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v55, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u32_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: v_lshl_or_b32 v13, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u32_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: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v52, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u32_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: v_lshl_or_b32 v14, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u32_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: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v49, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u32_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: v_lshl_or_b32 v15, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: v_or_b32_sdwa v1, v39, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_or_b32_sdwa v39, v36, v26 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:692 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; 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_add_u32_e32 v0, 0x300, v0
-; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
-; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v48, v46, v26 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:688 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v49, v35, v26 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:676 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v50, v27, v26 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:672 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; 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: buffer_load_dword v16, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u32_e32 v2, 0x300, v2
-; GFX9-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v51, v34, v26 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:660 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u32_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_lshl_or_b32 v17, v1, 16, v0
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u32_e32 v16, 0x300, v16
-; GFX9-NEXT: v_lshl_or_b32 v16, v16, 16, v2
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
-; GFX9-NEXT: v_or_b32_sdwa v52, v27, v26 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:656 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v53, 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:480 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u32_e32 v24, 3, v24
-; GFX9-NEXT: v_add_u32_e32 v26, 3, v61
-; GFX9-NEXT: v_or_b32_sdwa v24, v54, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u32_e32 v36, 0x300, v24
-; GFX9-NEXT: v_add_u32_e32 v24, 0x300, v48
-; GFX9-NEXT: v_add_u32_e32 v48, 0x300, v51
-; GFX9-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v54, 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:476 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u32_e32 v26, 3, v45
-; GFX9-NEXT: v_add_u32_e32 v20, 3, v20
-; GFX9-NEXT: v_or_b32_sdwa v20, v57, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u32_e32 v34, 0x300, v20
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v55, 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:472 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u32_e32 v26, 3, v56
-; GFX9-NEXT: v_add_u32_e32 v21, 3, v21
-; GFX9-NEXT: v_or_b32_sdwa v21, v32, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u32_e32 v28, 0x300, v21
-; GFX9-NEXT: v_add_u32_e32 v21, 0x300, v54
-; GFX9-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GFX9-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v40, 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:468 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u32_e32 v23, 3, v23
-; GFX9-NEXT: v_add_u32_e32 v26, 3, v47
-; GFX9-NEXT: v_or_b32_sdwa v23, v41, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u32_e32 v20, 0x300, v40
-; GFX9-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v41, 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:464 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u32_e32 v26, 3, v43
-; GFX9-NEXT: v_add_u32_e32 v22, 3, v22
-; GFX9-NEXT: v_or_b32_sdwa v22, v44, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u32_e32 v35, 0x300, v22
-; GFX9-NEXT: v_add_u32_e32 v22, 0x300, v52
-; GFX9-NEXT: v_add_u32_e32 v51, 0x300, v41
-; GFX9-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GFX9-NEXT: v_lshl_or_b32 v20, v51, 16, v20
-; GFX9-NEXT: v_lshl_or_b32 v28, v35, 16, v28
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v42, v27, v26 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:636 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v43, v27, v26 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:624 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; GFX9-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u32_e32 v31, 0x300, v0
-; GFX9-NEXT: v_add_u32_e32 v52, 0x300, v43
-; GFX9-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GFX9-NEXT: v_mov_b32_e32 v0, s8
-; GFX9-NEXT: s_waitcnt vmcnt(3)
-; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_or_b32_sdwa v44, 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:444 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u32_e32 v26, 3, v33
-; GFX9-NEXT: v_add_u32_e32 v32, 0x300, v1
-; GFX9-NEXT: v_mov_b32_e32 v1, s6
-; GFX9-NEXT: v_lshl_or_b32 v31, v32, 16, v31
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_or_b32_sdwa v45, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u32_e32 v27, 0x300, v23
-; GFX9-NEXT: v_add_u32_e32 v26, 0x300, v25
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
-; GFX9-NEXT: v_or_b32_sdwa v2, v18, v2 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:760 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u32_e32 v30, 0x300, v2
-; GFX9-NEXT: v_add_u32_e32 v25, 0x300, v38
-; GFX9-NEXT: v_add_u32_e32 v23, 0x300, v50
-; GFX9-NEXT: v_add_u32_e32 v38, 0x300, v39
-; GFX9-NEXT: v_add_u32_e32 v39, 0x300, v49
-; GFX9-NEXT: v_add_u32_e32 v49, 0x300, v53
-; GFX9-NEXT: v_add_u32_e32 v50, 0x300, v55
-; GFX9-NEXT: v_add_u32_e32 v53, 0x300, v45
-; GFX9-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GFX9-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GFX9-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX9-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GFX9-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GFX9-NEXT: v_mov_b32_e32 v2, s4
-; GFX9-NEXT: v_lshl_or_b32 v21, v50, 16, v21
-; GFX9-NEXT: v_lshl_or_b32 v22, v49, 16, v22
-; GFX9-NEXT: v_lshl_or_b32 v23, v48, 16, v23
-; GFX9-NEXT: v_lshl_or_b32 v24, v39, 16, v24
-; GFX9-NEXT: v_lshl_or_b32 v25, v38, 16, v25
-; GFX9-NEXT: v_lshl_or_b32 v26, v37, 16, v26
-; GFX9-NEXT: v_lshl_or_b32 v27, v36, 16, v27
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u32_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:756 ; 4-byte Folded Reload
-; GFX9-NEXT: v_add_u32_e32 v33, 0x300, v18
-; GFX9-NEXT: v_add_u32_e32 v18, 0x300, v44
-; GFX9-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GFX9-NEXT: v_lshl_or_b32 v18, v53, 16, v18
-; GFX9-NEXT: v_lshl_or_b32 v30, v33, 16, v30
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_u32_e32 v19, 3, v19
-; GFX9-NEXT: v_or_b32_sdwa v19, v60, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT: v_add_u32_e32 v29, 0x300, v19
-; GFX9-NEXT: v_add_u32_e32 v19, 0x300, v42
-; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX9-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GFX9-NEXT: v_lshl_or_b32 v19, v52, 16, v19
-; GFX9-NEXT: v_lshl_or_b32 v29, v34, 16, v29
-; GFX9-NEXT: .LBB97_5: ; %end
-; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-TRUE16-LABEL: bitcast_v128i8_to_v64i16_scalar:
-; GFX11-TRUE16: ; %bb.0:
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT: s_clause 0x1d
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v40, s32 offset:436
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v41, s32 offset:432
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v42, s32 offset:428
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v43, s32 offset:424
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v44, s32 offset:420
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v45, s32 offset:416
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v46, s32 offset:412
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v47, s32 offset:408
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v56, s32 offset:404
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v57, s32 offset:400
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v58, s32 offset:396
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v59, s32 offset:392
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v60, s32 offset:388
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v61, s32 offset:384
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v62, s32 offset:380
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v63, s32 offset:376
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v72, s32 offset:372
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v73, s32 offset:368
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v74, s32 offset:364
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v75, s32 offset:360
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v76, s32 offset:356
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v77, s32 offset:352
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v78, s32 offset:348
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v79, s32 offset:344
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v88, s32 offset:340
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v89, s32 offset:336
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v90, s32 offset:332
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v91, s32 offset:328
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v92, s32 offset:324
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v93, s32 offset:320
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_b32 v31, off, s32 offset:316
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v89, off, s32 offset:312
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v92, off, s32 offset:308
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v91, off, s32 offset:304
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v93, off, s32 offset:300
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v78, off, s32 offset:296
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v79, off, s32 offset:292
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v88, off, s32 offset:288
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v90, off, s32 offset:284
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v72, off, s32 offset:280
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v75, off, s32 offset:276
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v76, off, s32 offset:272
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v77, off, s32 offset:268
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v60, off, s32 offset:264
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v63, off, s32 offset:260
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v73, off, s32 offset:256
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v74, off, s32 offset:252
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v57, off, s32 offset:248
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v59, off, s32 offset:244
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v61, off, s32 offset:240
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v62, off, s32 offset:236
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v45, off, s32 offset:232
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v56, off, s32 offset:228
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v47, off, s32 offset:224
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v58, off, s32 offset:220
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v41, off, s32 offset:216
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v44, off, s32 offset:212
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v43, off, s32 offset:208
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v46, off, s32 offset:204
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v180, off, s32 offset:200
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v183, off, s32 offset:196
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v40, off, s32 offset:192
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v42, off, s32 offset:188
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v176, off, s32 offset:184
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v179, off, s32 offset:180
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v181, off, s32 offset:176
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v182, off, s32 offset:172
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v165, off, s32 offset:168
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v167, off, s32 offset:164
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v177, off, s32 offset:160
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v178, off, s32 offset:156
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v161, off, s32 offset:152
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v164, off, s32 offset:148
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v163, off, s32 offset:144
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v166, off, s32 offset:140
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v149, off, s32 offset:136
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v160, off, s32 offset:132
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v151, off, s32 offset:128
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v162, off, s32 offset:124
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v134, off, s32 offset:120
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v147, off, s32 offset:116
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v148, off, s32 offset:112
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v150, off, s32 offset:108
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v119, off, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v132, off, s32 offset:100
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v144, off, s32 offset:96
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v145, off, s32 offset:92
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v114, off, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v117, off, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v129, off, s32 offset:80
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v130, off, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v102, off, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v135, off, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v133, off, s32 offset:64
-; GFX11-TRUE16-NEXT: s_clause 0xf
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v146, off, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v115, off, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v128, off, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v118, off, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v131, off, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v112, off, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v113, off, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v116, off, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v96, off, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v99, off, s32 offset:20
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v101, off, s32 offset:16
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v103, off, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v87, off, s32 offset:4
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v98.l, v30.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v81.l, v29.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v84.l, v28.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v83.l, v27.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v86.l, v26.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v69.l, v25.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v80.l, v24.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v71.l, v23.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v82.l, v22.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v21.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v20.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v68.l, v19.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v70.l, v18.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v17.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v15.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v14.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v13.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v11.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v9.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v7.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v5.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v0.l
-; GFX11-TRUE16-NEXT: s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(62)
-; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v31
-; GFX11-TRUE16-NEXT: s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB97_4
-; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v39
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v37
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v33
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v32
-; GFX11-TRUE16-NEXT: s_and_b32 s5, s28, 0xff
-; GFX11-TRUE16-NEXT: s_lshl_b32 s6, s29, 8
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xff, v35
-; GFX11-TRUE16-NEXT: s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 8, v34
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT: v_and_b32_e64 v2, 0xffff, s5
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xff, v50
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v4, v5
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 8, v38
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v4, v1, 16, v2
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v54
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v53
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v5, v3, 16, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v6, v7
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v48
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 8, v36
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v51
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 8, v49
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xff, v66
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 8, v65
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v3, v6
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v2, v7
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v8, v9
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xff, v55
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 8, v52
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xff, v70
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 8, v68
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v6, v3, 16, v0
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v7, v2, 16, v1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v9, v10
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v8
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v11, v12
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v82
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 8, v71
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xff, v67
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 8, v64
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xff, v80
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 8, v69
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v3, v8
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v8, v0, 16, v1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v9, v10
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v11, v12
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xff, v86
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 8, v83
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v9, v0, 16, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v98
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v97
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v10, v2, 16, v3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v11, v12
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v84
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 8, v81
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v87
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 8, v85
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xff, v103
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 8, v101
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v3, v11
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v12
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v13, v14
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xff, v99
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 8, v96
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xff, v116
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 8, v113
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v11, v3, 16, v2
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v12, v1, 16, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v14, v15
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v13
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v16, v17
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v112
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 8, v100
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xff, v131
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 8, v118
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xff, v128
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 8, v115
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xff, v146
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 8, v133
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v3, v13
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v14, v15
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v16, v17
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xff, v130
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 8, v129
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v18, v19
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xff, v135
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 8, v102
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v16, v17
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v13
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v13, v0, 16, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v145
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v144
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v18, v19
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v14, v3, 16, v2
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v119
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v132
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xff, v117
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 8, v114
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v15, v15, 16, v17
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v134
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v16
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v16, v18, 16, v19
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v18, v1, 16, v0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v150
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v148
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v17, v20, 16, v21
-; GFX11-TRUE16-NEXT: s_and_b32 s5, s0, 0xff
-; GFX11-TRUE16-NEXT: s_lshl_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT: s_and_b32 s7, s2, 0xff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v147
-; GFX11-TRUE16-NEXT: s_lshl_b32 s8, s3, 8
-; GFX11-TRUE16-NEXT: s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT: s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v149
-; GFX11-TRUE16-NEXT: s_and_b32 s7, s16, 0xff
-; GFX11-TRUE16-NEXT: s_lshl_b32 s8, s17, 8
-; GFX11-TRUE16-NEXT: s_and_b32 s9, s18, 0xff
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v19, v1, 16, v0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v162
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v151
-; GFX11-TRUE16-NEXT: s_lshl_b32 s10, s19, 8
-; GFX11-TRUE16-NEXT: s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT: s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s5, s6
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v160
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s7, s8
-; GFX11-TRUE16-NEXT: s_and_b32 s7, s20, 0xff
-; GFX11-TRUE16-NEXT: s_lshl_b32 s8, s21, 8
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v161
-; GFX11-TRUE16-NEXT: s_and_b32 s9, s22, 0xff
-; GFX11-TRUE16-NEXT: s_lshl_b32 s10, s23, 8
-; GFX11-TRUE16-NEXT: s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v20, v1, 16, v0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v166
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v163
-; GFX11-TRUE16-NEXT: s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT: s_and_b32 s9, s24, 0xff
-; GFX11-TRUE16-NEXT: s_lshl_b32 s10, s25, 8
-; GFX11-TRUE16-NEXT: s_and_b32 s11, s26, 0xff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v164
-; GFX11-TRUE16-NEXT: s_lshl_b32 s12, s27, 8
-; GFX11-TRUE16-NEXT: s_or_b32 s9, s9, s10
-; GFX11-TRUE16-NEXT: s_or_b32 s10, s11, s12
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v165
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s7, s8
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s9, s10
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v21, v1, 16, v0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v178
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v177
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, s8
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v167
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v176
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v22, v1, 16, v0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v182
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v181
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v179
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v180
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v23, v1, 16, v0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v42
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v40
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v183
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v41
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v24, v1, 16, v0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v46
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v43
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v44
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v45
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v25, v1, 16, v0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v58
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v47
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v56
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v57
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v26, v1, 16, v0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v62
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v61
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v59
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v60
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v27, v1, 16, v0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v74
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v73
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v63
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v72
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v28, v1, 16, v0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v77
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v76
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v75
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v78
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v29, v1, 16, v0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v90
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v88
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v79
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v89
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v30, v1, 16, v0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v93
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v91
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v92
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, s7
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v31, v1, 16, v0
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6
-; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB97_3
-; GFX11-TRUE16-NEXT: .LBB97_2: ; %cmp.true
-; GFX11-TRUE16-NEXT: s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT: s_lshl_b32 s5, s29, 8
-; GFX11-TRUE16-NEXT: s_and_b32 s4, s28, 0xff
-; GFX11-TRUE16-NEXT: s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT: s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT: s_and_b32 s5, s24, 0xff
-; GFX11-TRUE16-NEXT: s_lshl_b32 s6, s25, 8
-; GFX11-TRUE16-NEXT: s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT: s_or_b32 s5, s6, s5
-; GFX11-TRUE16-NEXT: s_and_b32 s6, s26, 0xff
-; GFX11-TRUE16-NEXT: s_lshl_b32 s7, s27, 8
-; GFX11-TRUE16-NEXT: s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT: s_or_b32 s6, s7, s6
-; GFX11-TRUE16-NEXT: s_and_b32 s7, s20, 0xff
-; GFX11-TRUE16-NEXT: s_lshl_b32 s8, s21, 8
-; GFX11-TRUE16-NEXT: s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT: s_or_b32 s7, s8, s7
-; GFX11-TRUE16-NEXT: s_and_b32 s8, s22, 0xff
-; GFX11-TRUE16-NEXT: s_lshl_b32 s9, s23, 8
-; GFX11-TRUE16-NEXT: s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT: s_or_b32 s8, s9, s8
-; GFX11-TRUE16-NEXT: s_and_b32 s9, s16, 0xff
-; GFX11-TRUE16-NEXT: s_lshl_b32 s10, s17, 8
-; GFX11-TRUE16-NEXT: s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT: s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT: s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT: s_or_b32 s9, s10, s9
-; GFX11-TRUE16-NEXT: s_and_b32 s10, s18, 0xff
-; GFX11-TRUE16-NEXT: s_lshl_b32 s11, s19, 8
-; GFX11-TRUE16-NEXT: s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT: s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT: s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT: s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT: s_or_b32 s10, s11, s10
-; GFX11-TRUE16-NEXT: s_or_b32 s0, s1, s0
-; GFX11-TRUE16-NEXT: s_or_b32 s1, s3, s2
-; GFX11-TRUE16-NEXT: s_addk_i32 s9, 0x300
-; GFX11-TRUE16-NEXT: s_addk_i32 s0, 0x300
-; GFX11-TRUE16-NEXT: s_addk_i32 s1, 0x300
-; GFX11-TRUE16-NEXT: s_addk_i32 s10, 0x300
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s1
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 3, v92
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s9, s10
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v0, 3, v93
-; GFX11-TRUE16-NEXT: s_addk_i32 s5, 0x300
-; GFX11-TRUE16-NEXT: s_addk_i32 s6, 0x300
-; GFX11-TRUE16-NEXT: s_addk_i32 s7, 0x300
-; GFX11-TRUE16-NEXT: s_addk_i32 s8, 0x300
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v90
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v79
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s7, s8
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v89
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s5, s6
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v91
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 8, v88
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 8, v78
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v6
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v2, v0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v77
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v31, 0x300, v1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v3
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v27, 0x300, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v2
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v76
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v75
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v30, 0x300, v1
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v28, 0x300, v0
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v0, 3, v74
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v2
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v72
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v73
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v63
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v29, 0x300, v1
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(59)
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 3, v62
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v60
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 8, v61
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v59
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v59, 0x300, v2
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v60, 0x300, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v6
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v57
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(55)
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v58
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v57, 0x300, v0
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v0, 3, v56
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v56, 0x300, v1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v2
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v47
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v45
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(51)
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v46
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v45, 0x300, v1
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 3, v44
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v43
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 8, v41
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(47)
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v42
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v22, 0x300, v2
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v26, 0x300, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v6
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v40
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v183
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v23, 0x300, v0
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(43)
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v0, 3, v182
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v25, 0x300, v1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v2
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v180
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v181
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v179
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v24, 0x300, v1
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(39)
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 3, v178
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v176
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 8, v177
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v167
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v167, 0x300, v2
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v176, 0x300, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v6
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v165
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(35)
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v166
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v165, 0x300, v0
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v0, 3, v164
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v164, 0x300, v1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v2
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v163
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v161
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(31)
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v162
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v161, 0x300, v1
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 3, v160
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v151
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 8, v149
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(27)
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v150
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v17, 0x300, v2
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v21, 0x300, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v6
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v148
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v147
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v18, 0x300, v0
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(23)
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v0, 3, v145
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v20, 0x300, v1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v2
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v134
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v144
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v132
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v19, 0x300, v1
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(19)
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 3, v130
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v119
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 8, v129
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v117
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v117, 0x300, v2
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v119, 0x300, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v6
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v114
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(15)
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v146
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v114, 0x300, v0
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v0, 3, v135
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v129, 0x300, v1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v2
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v133
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v102
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(11)
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v131
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v102, 0x300, v1
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 3, v128
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v118
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 8, v115
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(7)
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v116
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v16, 0x300, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v113
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v112
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(3)
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v103
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 8, v101
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v8, 3, v99
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v4, v3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xff, v5
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 8, v100
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v9, 3, v98
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v10, 3, v87
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v55, 3, v55
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v5, v4
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v7, v6
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 8, v96
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xff, v9
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 8, v97
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v14, 0x300, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v11, 0x300, v5
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v6
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xff, v10
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v9, v8
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 8, v85
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v8, 3, v86
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v13, 0x300, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v84
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v12, 0x300, v5
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v7, v6
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 8, v83
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 8, v81
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v9, 3, v82
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v15, 0x300, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v80
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v7, v6
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v8, v4
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xff, v9
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 8, v71
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 8, v69
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v10, 3, v70
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v70, 0x300, v4
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v8, v7
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v69, 0x300, v6
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v9, v5
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xff, v10
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 8, v68
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v8, 3, v67
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v9, 0x300, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v66
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v10, 0x300, v5
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v7, v6
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 8, v64
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 8, v65
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v64, 0x300, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v54
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v7, v6
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xff, v55
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v8, v4
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 8, v52
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 8, v53
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v51, 3, v51
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v53, 0x300, v6
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v8, v7
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v7, 0x300, v4
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v52, v5
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xff, v51
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 8, v49
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v49, 0x300, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v50
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v50, 0x300, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v48
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v8, 3, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 8, v38
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v36, 8, v36
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v37, 8, v37
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v35, 3, v35
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v39, 0x300, v5
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v38, v6
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v36, v4
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v37, v8
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xff, v35
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v34, 8, v34
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v35, 0x300, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v33
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v33, 0x300, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 0x300, v6
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v34, v8
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 8, v32
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT: s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v5
-; GFX11-TRUE16-NEXT: v_and_b32_e64 v8, 0xffff, s4
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v64
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v4, v5, 16, v8
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v5, v6, 16, v32
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v50
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v35
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v9, v53, 16, v34
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v7, v39, 16, v8
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v8, v49, 16, v32
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff, v69
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v6, v33, 16, v6
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v10, v10, 16, v35
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v18
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v11, v70, 16, v32
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff, v2
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v13, v13, 16, v33
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v119
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v17
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v20, v20, 16, v34
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v16, v16, 16, v32
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff, v129
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v18, v114, 16, v33
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v21, v21, 16, v35
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v176
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v23
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v17, v102, 16, v32
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff, v164
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v22
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v23, v165, 16, v33
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v22, v161, 16, v32
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v25, v25, 16, v34
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v26, v26, 16, v35
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff, v56
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v60
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v28
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v27
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v12, v15, 16, v12
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v14, v14, 16, v3
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, s3
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v15, v1, 16, v0
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, s2
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v19, v117, 16, v19
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v24, v167, 16, v24
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v27, v45, 16, v32
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v28, v57, 16, v33
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v29, v59, 16, v29
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v30, v30, 16, v34
-; GFX11-TRUE16-NEXT: v_lshl_or_b32 v31, v31, 16, v35
-; GFX11-TRUE16-NEXT: .LBB97_3: ; %end
-; GFX11-TRUE16-NEXT: s_clause 0x1d
-; GFX11-TRUE16-NEXT: scratch_load_b32 v93, off, s32 offset:320
-; GFX11-TRUE16-NEXT: scratch_load_b32 v92, off, s32 offset:324
-; GFX11-TRUE16-NEXT: scratch_load_b32 v91, off, s32 offset:328
-; GFX11-TRUE16-NEXT: scratch_load_b32 v90, off, s32 offset:332
-; GFX11-TRUE16-NEXT: scratch_load_b32 v89, off, s32 offset:336
-; GFX11-TRUE16-NEXT: scratch_load_b32 v88, off, s32 offset:340
-; GFX11-TRUE16-NEXT: scratch_load_b32 v79, off, s32 offset:344
-; GFX11-TRUE16-NEXT: scratch_load_b32 v78, off, s32 offset:348
-; GFX11-TRUE16-NEXT: scratch_load_b32 v77, off, s32 offset:352
-; GFX11-TRUE16-NEXT: scratch_load_b32 v76, off, s32 offset:356
-; GFX11-TRUE16-NEXT: scratch_load_b32 v75, off, s32 offset:360
-; GFX11-TRUE16-NEXT: scratch_load_b32 v74, off, s32 offset:364
-; GFX11-TRUE16-NEXT: scratch_load_b32 v73, off, s32 offset:368
-; GFX11-TRUE16-NEXT: scratch_load_b32 v72, off, s32 offset:372
-; GFX11-TRUE16-NEXT: scratch_load_b32 v63, off, s32 offset:376
-; GFX11-TRUE16-NEXT: scratch_load_b32 v62, off, s32 offset:380
-; GFX11-TRUE16-NEXT: scratch_load_b32 v61, off, s32 offset:384
-; GFX11-TRUE16-NEXT: scratch_load_b32 v60, off, s32 offset:388
-; GFX11-TRUE16-NEXT: scratch_load_b32 v59, off, s32 offset:392
-; GFX11-TRUE16-NEXT: scratch_load_b32 v58, off, s32 offset:396
-; GFX11-TRUE16-NEXT: scratch_load_b32 v57, off, s32 offset:400
-; GFX11-TRUE16-NEXT: scratch_load_b32 v56, off, s32 offset:404
-; GFX11-TRUE16-NEXT: scratch_load_b32 v47, off, s32 offset:408
-; GFX11-TRUE16-NEXT: scratch_load_b32 v46, off, s32 offset:412
-; GFX11-TRUE16-NEXT: scratch_load_b32 v45, off, s32 offset:416
-; GFX11-TRUE16-NEXT: scratch_load_b32 v44, off, s32 offset:420
-; GFX11-TRUE16-NEXT: scratch_load_b32 v43, off, s32 offset:424
-; GFX11-TRUE16-NEXT: scratch_load_b32 v42, off, s32 offset:428
-; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s32 offset:432
-; GFX11-TRUE16-NEXT: scratch_load_b32 v40, off, s32 offset:436
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT: .LBB97_4:
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-TRUE16-NEXT: s_branch .LBB97_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v128i8_to_v64i16_scalar:
-; GFX11-FAKE16: ; %bb.0:
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT: s_clause 0x1e
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v40, s32 offset:440
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v41, s32 offset:436
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v42, s32 offset:432
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v43, s32 offset:428
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v44, s32 offset:424
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v45, s32 offset:420
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v46, s32 offset:416
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v47, s32 offset:412
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v56, s32 offset:408
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v57, s32 offset:404
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v58, s32 offset:400
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v59, s32 offset:396
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v60, s32 offset:392
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v61, s32 offset:388
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v62, s32 offset:384
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v63, s32 offset:380
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v72, s32 offset:376
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v73, s32 offset:372
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v74, s32 offset:368
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v75, s32 offset:364
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v76, s32 offset:360
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v77, s32 offset:356
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v78, s32 offset:352
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v79, s32 offset:348
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v88, s32 offset:344
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v89, s32 offset:340
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v90, s32 offset:336
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v91, s32 offset:332
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v92, s32 offset:328
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v93, s32 offset:324
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v94, s32 offset:320
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v54, v30 :: v_dual_mov_b32 v51, v24
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v52, v28 :: v_dual_mov_b32 v55, v26
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v53, v22 :: v_dual_mov_b32 v48, v20
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v50, v18 :: v_dual_mov_b32 v39, v16
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v49, v14 :: v_dual_mov_b32 v34, v8
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v37, v12 :: v_dual_mov_b32 v38, v10
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v36, v6 :: v_dual_mov_b32 v33, v4
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v35, v2 :: v_dual_mov_b32 v32, v0
-; GFX11-FAKE16-NEXT: s_clause 0x1f
-; GFX11-FAKE16-NEXT: scratch_load_b32 v0, off, s32 offset:316
-; GFX11-FAKE16-NEXT: scratch_load_u16 v2, off, s32
-; GFX11-FAKE16-NEXT: scratch_load_u16 v4, off, s32 offset:8
-; GFX11-FAKE16-NEXT: scratch_load_u16 v6, off, s32 offset:16
-; GFX11-FAKE16-NEXT: scratch_load_u16 v8, off, s32 offset:24
-; GFX11-FAKE16-NEXT: scratch_load_u16 v10, off, s32 offset:32
-; GFX11-FAKE16-NEXT: scratch_load_u16 v12, off, s32 offset:40
-; GFX11-FAKE16-NEXT: scratch_load_u16 v14, off, s32 offset:48
-; GFX11-FAKE16-NEXT: scratch_load_u16 v16, off, s32 offset:56
-; GFX11-FAKE16-NEXT: scratch_load_u16 v18, off, s32 offset:64
-; GFX11-FAKE16-NEXT: scratch_load_u16 v20, off, s32 offset:72
-; GFX11-FAKE16-NEXT: scratch_load_u16 v22, off, s32 offset:80
-; GFX11-FAKE16-NEXT: scratch_load_u16 v24, off, s32 offset:88
-; GFX11-FAKE16-NEXT: scratch_load_u16 v26, off, s32 offset:96
-; GFX11-FAKE16-NEXT: scratch_load_u16 v28, off, s32 offset:104
-; GFX11-FAKE16-NEXT: scratch_load_u16 v30, off, s32 offset:112
-; GFX11-FAKE16-NEXT: scratch_load_u16 v31, off, s32 offset:120
-; GFX11-FAKE16-NEXT: scratch_load_u16 v41, off, s32 offset:128
-; GFX11-FAKE16-NEXT: scratch_load_u16 v44, off, s32 offset:136
-; GFX11-FAKE16-NEXT: scratch_load_u16 v45, off, s32 offset:144
-; GFX11-FAKE16-NEXT: scratch_load_u16 v56, off, s32 offset:152
-; GFX11-FAKE16-NEXT: scratch_load_u16 v59, off, s32 offset:160
-; GFX11-FAKE16-NEXT: scratch_load_u16 v60, off, s32 offset:168
-; GFX11-FAKE16-NEXT: scratch_load_u16 v61, off, s32 offset:176
-; GFX11-FAKE16-NEXT: scratch_load_u16 v62, off, s32 offset:184
-; GFX11-FAKE16-NEXT: scratch_load_u16 v63, off, s32 offset:192
-; GFX11-FAKE16-NEXT: scratch_load_u16 v72, off, s32 offset:200
-; GFX11-FAKE16-NEXT: scratch_load_u16 v73, off, s32 offset:208
-; GFX11-FAKE16-NEXT: scratch_load_u16 v74, off, s32 offset:216
-; GFX11-FAKE16-NEXT: scratch_load_u16 v75, off, s32 offset:224
-; GFX11-FAKE16-NEXT: scratch_load_u16 v76, off, s32 offset:232
-; GFX11-FAKE16-NEXT: scratch_load_u16 v77, off, s32 offset:240
-; GFX11-FAKE16-NEXT: s_clause 0x1f
-; GFX11-FAKE16-NEXT: scratch_load_u16 v78, off, s32 offset:248
-; GFX11-FAKE16-NEXT: scratch_load_u16 v79, off, s32 offset:256
-; GFX11-FAKE16-NEXT: scratch_load_u16 v88, off, s32 offset:264
-; GFX11-FAKE16-NEXT: scratch_load_u16 v89, off, s32 offset:272
-; GFX11-FAKE16-NEXT: scratch_load_u16 v90, off, s32 offset:280
-; GFX11-FAKE16-NEXT: scratch_load_u16 v91, off, s32 offset:288
-; GFX11-FAKE16-NEXT: scratch_load_u16 v92, off, s32 offset:296
-; GFX11-FAKE16-NEXT: scratch_load_u16 v93, off, s32 offset:304
-; GFX11-FAKE16-NEXT: scratch_load_u16 v94, off, s32 offset:312
-; GFX11-FAKE16-NEXT: scratch_load_u16 v57, off, s32 offset:308
-; GFX11-FAKE16-NEXT: scratch_load_u16 v58, off, s32 offset:300
-; GFX11-FAKE16-NEXT: scratch_load_u16 v46, off, s32 offset:292
-; GFX11-FAKE16-NEXT: scratch_load_u16 v47, off, s32 offset:284
-; GFX11-FAKE16-NEXT: scratch_load_u16 v40, off, s32 offset:276
-; GFX11-FAKE16-NEXT: scratch_load_u16 v43, off, s32 offset:268
-; GFX11-FAKE16-NEXT: scratch_load_u16 v182, off, s32 offset:260
-; GFX11-FAKE16-NEXT: scratch_load_u16 v183, off, s32 offset:252
-; GFX11-FAKE16-NEXT: scratch_load_u16 v178, off, s32 offset:244
-; GFX11-FAKE16-NEXT: scratch_load_u16 v181, off, s32 offset:236
-; GFX11-FAKE16-NEXT: scratch_load_u16 v164, off, s32 offset:228
-; GFX11-FAKE16-NEXT: scratch_load_u16 v176, off, s32 offset:220
-; GFX11-FAKE16-NEXT: scratch_load_u16 v160, off, s32 offset:212
-; GFX11-FAKE16-NEXT: scratch_load_u16 v163, off, s32 offset:204
-; GFX11-FAKE16-NEXT: scratch_load_u16 v146, off, s32 offset:196
-; GFX11-FAKE16-NEXT: scratch_load_u16 v150, off, s32 offset:188
-; GFX11-FAKE16-NEXT: scratch_load_u16 v135, off, s32 offset:180
-; GFX11-FAKE16-NEXT: scratch_load_u16 v145, off, s32 offset:172
-; GFX11-FAKE16-NEXT: scratch_load_u16 v118, off, s32 offset:164
-; GFX11-FAKE16-NEXT: scratch_load_u16 v131, off, s32 offset:156
-; GFX11-FAKE16-NEXT: scratch_load_u16 v115, off, s32 offset:148
-; GFX11-FAKE16-NEXT: scratch_load_u16 v179, off, s32 offset:140
-; GFX11-FAKE16-NEXT: scratch_load_u16 v162, off, s32 offset:132
-; GFX11-FAKE16-NEXT: s_clause 0xf
-; GFX11-FAKE16-NEXT: scratch_load_u16 v165, off, s32 offset:124
-; GFX11-FAKE16-NEXT: scratch_load_u16 v149, off, s32 offset:116
-; GFX11-FAKE16-NEXT: scratch_load_u16 v151, off, s32 offset:108
-; GFX11-FAKE16-NEXT: scratch_load_u16 v144, off, s32 offset:100
-; GFX11-FAKE16-NEXT: scratch_load_u16 v148, off, s32 offset:92
-; GFX11-FAKE16-NEXT: scratch_load_u16 v129, off, s32 offset:84
-; GFX11-FAKE16-NEXT: scratch_load_u16 v133, off, s32 offset:76
-; GFX11-FAKE16-NEXT: scratch_load_u16 v117, off, s32 offset:68
-; GFX11-FAKE16-NEXT: scratch_load_u16 v119, off, s32 offset:60
-; GFX11-FAKE16-NEXT: scratch_load_u16 v114, off, s32 offset:52
-; GFX11-FAKE16-NEXT: scratch_load_u16 v116, off, s32 offset:44
-; GFX11-FAKE16-NEXT: scratch_load_u16 v100, off, s32 offset:36
-; GFX11-FAKE16-NEXT: scratch_load_u16 v103, off, s32 offset:28
-; GFX11-FAKE16-NEXT: scratch_load_u16 v98, off, s32 offset:20
-; GFX11-FAKE16-NEXT: scratch_load_u16 v99, off, s32 offset:12
-; GFX11-FAKE16-NEXT: scratch_load_u16 v96, off, s32 offset:4
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v64, 8, v1
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v66, 8, v3
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v65, 8, v5
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v68, 8, v7
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v67, 8, v9
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v70, 8, v11
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v69, 8, v13
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v71, 8, v15
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v80, 8, v17
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v82, 8, v19
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v81, 8, v21
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v84, 8, v23
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v83, 8, v25
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v86, 8, v27
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v85, 8, v29
-; GFX11-FAKE16-NEXT: s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(62)
-; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v97, 8, v2
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v87, 8, v4
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v101, 8, v6
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v102, 8, v8
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v113, 8, v10
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v112, 8, v12
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v130, 8, v14
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v128, 8, v16
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v134, 8, v18
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v132, 8, v20
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v161, 8, v22
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v147, 8, v24
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v166, 8, v26
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v167, 8, v28
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v180, 8, v30
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v177, 8, v31
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v42, 8, v41
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(61)
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v41, 8, v44
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(60)
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v45, 8, v45
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(59)
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v44, 8, v56
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(58)
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v59, 8, v59
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(57)
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v56, 8, v60
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(56)
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v60, 8, v61
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(55)
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v61, 8, v62
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(54)
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v63, 8, v63
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(53)
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v62, 8, v72
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(52)
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v73, 8, v73
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(51)
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v72, 8, v74
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(50)
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v75, 8, v75
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(49)
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v74, 8, v76
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(48)
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v77, 8, v77
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(47)
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v76, 8, v78
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(46)
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v78, 8, v79
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(45)
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v79, 8, v88
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(44)
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v89, 8, v89
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(43)
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v88, 8, v90
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(42)
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v91, 8, v91
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(41)
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v90, 8, v92
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(40)
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v92, 8, v93
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(39)
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v93, 8, v94
-; GFX11-FAKE16-NEXT: s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB97_4
-; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT: s_and_b32 s5, s28, 0xff
-; GFX11-FAKE16-NEXT: s_lshl_b32 s6, s29, 8
-; GFX11-FAKE16-NEXT: s_and_b32 s7, s2, 0xff
-; GFX11-FAKE16-NEXT: s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT: s_lshl_b32 s6, s1, 8
-; GFX11-FAKE16-NEXT: v_and_b32_e64 v5, 0xffff, s5
-; GFX11-FAKE16-NEXT: s_and_b32 s5, s0, 0xff
-; GFX11-FAKE16-NEXT: s_lshl_b32 s8, s3, 8
-; GFX11-FAKE16-NEXT: s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT: s_or_b32 s6, s7, s8
-; GFX11-FAKE16-NEXT: s_and_b32 s7, s16, 0xff
-; GFX11-FAKE16-NEXT: s_lshl_b32 s8, s17, 8
-; GFX11-FAKE16-NEXT: s_and_b32 s9, s18, 0xff
-; GFX11-FAKE16-NEXT: s_lshl_b32 s10, s19, 8
-; GFX11-FAKE16-NEXT: s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT: s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s5, s5, s6
-; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s6, s7, s8
-; GFX11-FAKE16-NEXT: s_and_b32 s7, s20, 0xff
-; GFX11-FAKE16-NEXT: s_lshl_b32 s8, s21, 8
-; GFX11-FAKE16-NEXT: s_and_b32 s9, s22, 0xff
-; GFX11-FAKE16-NEXT: s_lshl_b32 s10, s23, 8
-; GFX11-FAKE16-NEXT: s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT: s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT: s_and_b32 s9, s24, 0xff
-; GFX11-FAKE16-NEXT: s_lshl_b32 s10, s25, 8
-; GFX11-FAKE16-NEXT: s_and_b32 s11, s26, 0xff
-; GFX11-FAKE16-NEXT: s_lshl_b32 s12, s27, 8
-; GFX11-FAKE16-NEXT: s_or_b32 s9, s9, s10
-; GFX11-FAKE16-NEXT: s_or_b32 s10, s11, s12
-; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s7, s7, s8
-; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s8, s9, s10
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v36
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v32
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v35
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v34
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v33
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v2, v68
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v64
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v66
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v6, v4, v67
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v3, v65
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v4, v0, 16, v5
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v38
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xff, v39
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v6, 16, v2
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v49
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v5, v3, 16, v1
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v37
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v70
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v50
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v2, v71
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xff, v48
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v69
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v3, v82
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v9, v7, v80
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v10, v8, v81
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v1, 16, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v53
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v9, 16, v2
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v55
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v51
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v10, 16, v3
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v84
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v52
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xff, v54
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v2, v86
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v83
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xff, v96
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v3, v85
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v12, v10, v97
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v10, v1, 16, v0
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v11, v87
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v99
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v11, v3, 16, v2
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v103
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xff, v114
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v98
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v12, v0, 16, v12
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v100
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v3, v113
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v101
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xff, v116
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v17, v14, v128
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v112
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xff, v117
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v2, v102
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v13, v13, v130
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v16, 0xff, v133
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v20, v14, v132
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v14, v0, 16, v3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v148
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xff, v119
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xff, v129
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v16, v16, v161
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xffff, v13
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v13, v2, 16, v1
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v2, s7
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v166
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v144
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v15, v15, v134
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v18, v18, v147
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xffff, v16
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v167
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v15
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v15, v17, 16, v19
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v17, v18, 16, v22
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v3, s8
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v1, 16, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v151
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v149
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v16, v20, 16, v21
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v180
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v177
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v1, 16, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v165
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v162
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v42
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v41
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v1, 16, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v179
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v115
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v45
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v44
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v1, 16, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v131
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v118
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v59
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v56
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v1, 16, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v145
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v135
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v60
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v61
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v1, 16, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v150
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v146
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v63
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v62
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v24, v1, 16, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v163
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v160
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v73
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v72
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v25, v1, 16, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v176
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v164
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v75
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v74
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v26, v1, 16, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v181
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v178
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v77
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v76
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v27, v1, 16, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v183
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v182
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v78
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v79
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v28, v1, 16, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v43
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v40
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v89
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v88
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v29, v1, 16, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v47
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v46
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v91
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v90
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v30, v1, 16, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v58
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v57
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v92
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v93
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v31, v1, 16, v0
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6
-; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB97_3
-; GFX11-FAKE16-NEXT: .LBB97_2: ; %cmp.true
-; GFX11-FAKE16-NEXT: s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT: s_lshl_b32 s5, s29, 8
-; GFX11-FAKE16-NEXT: s_and_b32 s4, s28, 0xff
-; GFX11-FAKE16-NEXT: s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT: s_or_b32 s4, s5, s4
-; GFX11-FAKE16-NEXT: s_and_b32 s5, s24, 0xff
-; GFX11-FAKE16-NEXT: s_lshl_b32 s6, s25, 8
-; GFX11-FAKE16-NEXT: s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT: s_or_b32 s5, s6, s5
-; GFX11-FAKE16-NEXT: s_and_b32 s6, s26, 0xff
-; GFX11-FAKE16-NEXT: s_lshl_b32 s7, s27, 8
-; GFX11-FAKE16-NEXT: s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT: s_or_b32 s6, s7, s6
-; GFX11-FAKE16-NEXT: s_and_b32 s7, s20, 0xff
-; GFX11-FAKE16-NEXT: s_lshl_b32 s8, s21, 8
-; GFX11-FAKE16-NEXT: s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT: s_or_b32 s7, s8, s7
-; GFX11-FAKE16-NEXT: s_and_b32 s8, s22, 0xff
-; GFX11-FAKE16-NEXT: s_lshl_b32 s9, s23, 8
-; GFX11-FAKE16-NEXT: s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT: s_or_b32 s8, s9, s8
-; GFX11-FAKE16-NEXT: s_and_b32 s9, s16, 0xff
-; GFX11-FAKE16-NEXT: s_lshl_b32 s10, s17, 8
-; GFX11-FAKE16-NEXT: s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT: s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT: s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT: s_or_b32 s9, s10, s9
-; GFX11-FAKE16-NEXT: s_and_b32 s10, s18, 0xff
-; GFX11-FAKE16-NEXT: s_lshl_b32 s11, s19, 8
-; GFX11-FAKE16-NEXT: s_and_b32 s0, s0, 0xff
-; GFX11-FAKE16-NEXT: s_lshl_b32 s1, s1, 8
-; GFX11-FAKE16-NEXT: s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT: s_lshl_b32 s3, s3, 8
-; GFX11-FAKE16-NEXT: s_or_b32 s10, s11, s10
-; GFX11-FAKE16-NEXT: s_or_b32 s0, s1, s0
-; GFX11-FAKE16-NEXT: s_or_b32 s1, s3, s2
-; GFX11-FAKE16-NEXT: s_addk_i32 s9, 0x300
-; GFX11-FAKE16-NEXT: s_addk_i32 s0, 0x300
-; GFX11-FAKE16-NEXT: s_addk_i32 s1, 0x300
-; GFX11-FAKE16-NEXT: s_addk_i32 s10, 0x300
-; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s0, s0, s1
-; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s1, s9, s10
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(37)
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v0, 3, v58
-; GFX11-FAKE16-NEXT: s_addk_i32 s5, 0x300
-; GFX11-FAKE16-NEXT: s_addk_i32 s6, 0x300
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v1, 3, v57
-; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s3, s5, s6
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(35)
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v2, 3, v47
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT: s_addk_i32 s7, 0x300
-; GFX11-FAKE16-NEXT: s_addk_i32 s8, 0x300
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v92, v0
-; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s2, s7, s8
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v3, 3, v46
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v93, v1
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v91, v2
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(33)
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v43
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v27, 0x300, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v0, 3, v40
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v31, 0x300, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v28, 0x300, v2
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v4
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v90, v3
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(31)
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v2, 3, v183
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v182
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v89, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v30, 0x300, v3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v4
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(29)
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v181
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v88, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v181, 0x300, v1
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v78, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v79, v3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v29, 0x300, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v0, 3, v178
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v182, 0x300, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v178, 0x300, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v77, v3
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(27)
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v2, 3, v176
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v3, 3, v164
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(25)
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v163
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v163, 0x300, v1
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v76, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v160
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v75, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v160, 0x300, v0
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v74, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v73, v3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v22, 0x300, v1
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(23)
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v1, 3, v150
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v26, 0x300, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v23, 0x300, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v72, v3
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v2, 3, v146
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(21)
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v3, 3, v145
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v135
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v25, 0x300, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v63, v1
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v4
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(19)
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v131
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v62, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v131, 0x300, v1
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v60, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v61, v3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v24, 0x300, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v0, 3, v118
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v135, 0x300, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v118, 0x300, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v59, v3
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(17)
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v2, 3, v179
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v3, 3, v115
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(15)
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v165
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v115, 0x300, v1
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v56, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v162
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v45, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v145, 0x300, v0
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v44, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v42, v3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v17, 0x300, v1
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(13)
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v1, 3, v151
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v21, 0x300, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v18, 0x300, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v41, v3
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v2, 3, v149
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(11)
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v3, 3, v148
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v144
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v20, 0x300, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v180, v1
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v4
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(9)
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v133
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v177, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v19, 0x300, v1
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v166, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v167, v3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v133, 0x300, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v0, 3, v129
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v129, 0x300, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v144, 0x300, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v161, v3
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(7)
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v2, 3, v119
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v3, 3, v117
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(5)
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v116
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v116, 0x300, v1
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v147, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v114
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(1)
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v5, 3, v99
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v114, 0x300, v0
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v132, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v130, v3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v103
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v6, 3, v98
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v7, 3, v54
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v8, 3, v53
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v9, 3, v39
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v33, 3, v33
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v16, 0x300, v0
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v113, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v0, 0x300, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v128, v3
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v3, 3, v100
-; GFX11-FAKE16-NEXT: s_addk_i32 s4, 0x300
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v11, 0x300, v4
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v101, v5
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v102, v6
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xff, v7
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v7, 3, v96
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v134, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v12, 0x300, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v13, 0x300, v5
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v97, v6
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v5, 3, v55
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xff, v7
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v7, 3, v52
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v14, 0x300, v4
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v5
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v87, v6
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xff, v7
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xff, v8
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v8, 3, v51
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v86, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v15, 0x300, v5
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v85, v6
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v6, v84, v7
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xff, v8
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v51, 0x300, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v50
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v50, 0x300, v5
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v8, 3, v49
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v83, v7
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v7, 3, v48
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v10, 0x300, v5
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xff, v7
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v82, v4
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xff, v8
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xff, v9
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v9, 3, v38
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v81, v5
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v38, 0x300, v4
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v71, v7
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v80, v8
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xff, v9
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v9, 0x300, v5
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v5, 3, v37
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v37, 0x300, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v39, 0x300, v7
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v70, v8
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v7, 3, v36
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v8, 3, v34
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v34, 3, v35
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v35, 0x300, v4
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v7
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v69, v5
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xff, v8
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xff, v34
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v112, v3
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v68, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v34, 0x300, v5
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v67, v7
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v66, v8
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xff, v33
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v33, 0x300, v4
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v32
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v32, 0x300, v5
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v5, 0x300, v7
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v65, v8
-; GFX11-FAKE16-NEXT: v_and_b32_e64 v8, 0xffff, s4
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v36, 0xffff, v6
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v7, 0x300, v7
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v64, v4
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v10, v10, 16, v36
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v5, v7, 16, v5
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xffff, v33
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 0x300, v4
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xffff, v37
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v36, 0xffff, v22
-; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v32, 16, v7
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v4, v4, 16, v8
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xffff, v35
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v32, 0xffff, v51
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v35, 0xffff, v38
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v34, 16, v8
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v39, 16, v33
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xffff, v12
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v34, 0xffff, v11
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v11, v50, 16, v32
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v32, 0xffff, v1
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v9, 16, v35
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v12, v15, 16, v14
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v13, v13, 16, v33
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v14, v3, 16, v34
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v16, v16, 16, v32
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v32, 0xffff, v116
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xffff, v129
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v34, 0xffff, v18
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v35, 0xffff, v17
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v26, v26, 16, v36
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v17, v114, 16, v32
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v144, 16, v33
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v20, 16, v34
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v21, 16, v35
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v32, 0xffff, v115
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xffff, v135
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v34, 0xffff, v131
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v35, 0xffff, v23
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v36, 0xffff, v27
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v145, 16, v32
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v118, 16, v33
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v24, v24, 16, v34
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v25, v25, 16, v35
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v32, 0xffff, v163
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xffff, v182
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v34, 0xffff, v181
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v35, 0xffff, v28
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v15, v2, 16, v0
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v2, s2
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v3, s3
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v133, 16, v19
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v27, v160, 16, v32
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v28, v178, 16, v33
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v29, v29, 16, v34
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v30, v30, 16, v35
-; GFX11-FAKE16-NEXT: v_lshl_or_b32 v31, v31, 16, v36
-; GFX11-FAKE16-NEXT: .LBB97_3: ; %end
-; GFX11-FAKE16-NEXT: s_clause 0x1e
-; GFX11-FAKE16-NEXT: scratch_load_b32 v94, off, s32 offset:320
-; GFX11-FAKE16-NEXT: scratch_load_b32 v93, off, s32 offset:324
-; GFX11-FAKE16-NEXT: scratch_load_b32 v92, off, s32 offset:328
-; GFX11-FAKE16-NEXT: scratch_load_b32 v91, off, s32 offset:332
-; GFX11-FAKE16-NEXT: scratch_load_b32 v90, off, s32 offset:336
-; GFX11-FAKE16-NEXT: scratch_load_b32 v89, off, s32 offset:340
-; GFX11-FAKE16-NEXT: scratch_load_b32 v88, off, s32 offset:344
-; GFX11-FAKE16-NEXT: scratch_load_b32 v79, off, s32 offset:348
-; GFX11-FAKE16-NEXT: scratch_load_b32 v78, off, s32 offset:352
-; GFX11-FAKE16-NEXT: scratch_load_b32 v77, off, s32 offset:356
-; GFX11-FAKE16-NEXT: scratch_load_b32 v76, off, s32 offset:360
-; GFX11-FAKE16-NEXT: scratch_load_b32 v75, off, s32 offset:364
-; GFX11-FAKE16-NEXT: scratch_load_b32 v74, off, s32 offset:368
-; GFX11-FAKE16-NEXT: scratch_load_b32 v73, off, s32 offset:372
-; GFX11-FAKE16-NEXT: scratch_load_b32 v72, off, s32 offset:376
-; GFX11-FAKE16-NEXT: scratch_load_b32 v63, off, s32 offset:380
-; GFX11-FAKE16-NEXT: scratch_load_b32 v62, off, s32 offset:384
-; GFX11-FAKE16-NEXT: scratch_load_b32 v61, off, s32 offset:388
-; GFX11-FAKE16-NEXT: scratch_load_b32 v60, off, s32 offset:392
-; GFX11-FAKE16-NEXT: scratch_load_b32 v59, off, s32 offset:396
-; GFX11-FAKE16-NEXT: scratch_load_b32 v58, off, s32 offset:400
-; GFX11-FAKE16-NEXT: scratch_load_b32 v57, off, s32 offset:404
-; GFX11-FAKE16-NEXT: scratch_load_b32 v56, off, s32 offset:408
-; GFX11-FAKE16-NEXT: scratch_load_b32 v47, off, s32 offset:412
-; GFX11-FAKE16-NEXT: scratch_load_b32 v46, off, s32 offset:416
-; GFX11-FAKE16-NEXT: scratch_load_b32 v45, off, s32 offset:420
-; GFX11-FAKE16-NEXT: scratch_load_b32 v44, off, s32 offset:424
-; GFX11-FAKE16-NEXT: scratch_load_b32 v43, off, s32 offset:428
-; GFX11-FAKE16-NEXT: scratch_load_b32 v42, off, s32 offset:432
-; GFX11-FAKE16-NEXT: scratch_load_b32 v41, off, s32 offset:436
-; GFX11-FAKE16-NEXT: scratch_load_b32 v40, off, s32 offset:440
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT: .LBB97_4:
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-FAKE16-NEXT: s_branch .LBB97_2
- %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) {
-; SI-LABEL: bitcast_v64i16_to_v128i8:
-; SI: ; %bb.0:
-; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:136
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:12
-; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:28
-; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:44
-; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:60
-; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:76
-; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:92
-; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:108
-; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:124
-; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:132
-; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:128
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT: ; implicit-def: $vgpr23
-; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v28
-; SI-NEXT: ; kill: killed $vgpr23
-; SI-NEXT: ; implicit-def: $vgpr23
-; SI-NEXT: ; implicit-def: $vgpr46
-; SI-NEXT: ; implicit-def: $vgpr45
-; SI-NEXT: ; implicit-def: $vgpr44
-; SI-NEXT: ; implicit-def: $vgpr43
-; SI-NEXT: ; implicit-def: $vgpr42
-; SI-NEXT: ; implicit-def: $vgpr41
-; SI-NEXT: ; implicit-def: $vgpr40
-; SI-NEXT: ; implicit-def: $vgpr55
-; SI-NEXT: ; implicit-def: $vgpr54
-; SI-NEXT: ; implicit-def: $vgpr53
-; SI-NEXT: ; implicit-def: $vgpr52
-; SI-NEXT: ; implicit-def: $vgpr51
-; SI-NEXT: ; implicit-def: $vgpr50
-; SI-NEXT: ; implicit-def: $vgpr49
-; SI-NEXT: ; implicit-def: $vgpr48
-; SI-NEXT: ; implicit-def: $vgpr39
-; SI-NEXT: ; implicit-def: $vgpr38
-; SI-NEXT: ; implicit-def: $vgpr37
-; SI-NEXT: ; implicit-def: $vgpr36
-; SI-NEXT: ; implicit-def: $vgpr35
-; SI-NEXT: ; implicit-def: $vgpr34
-; SI-NEXT: ; implicit-def: $vgpr33
-; SI-NEXT: ; implicit-def: $vgpr31
-; SI-NEXT: ; implicit-def: $vgpr32
-; SI-NEXT: ; kill: killed $vgpr23
-; SI-NEXT: ; implicit-def: $vgpr23
-; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:4
-; SI-NEXT: s_waitcnt vmcnt(11)
-; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
-; SI-NEXT: s_waitcnt vmcnt(10)
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v3
-; SI-NEXT: s_waitcnt vmcnt(6)
-; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v11
-; SI-NEXT: s_waitcnt vmcnt(5)
-; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v13
-; SI-NEXT: s_waitcnt vmcnt(3)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v17
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:120
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:116
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:112
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:88
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:84
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:80
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v56
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:56
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:52
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:48
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v57
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:24
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v5
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v7
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v9
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v6
-; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v10
-; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v14
-; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v18
-; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v22
-; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v26
-; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v30
-; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:72
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: s_waitcnt vmcnt(13)
-; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:104
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v15
-; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v58
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v62
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr26
-; SI-NEXT: ; implicit-def: $vgpr30
-; SI-NEXT: ; implicit-def: $vgpr18
-; SI-NEXT: ; implicit-def: $vgpr22
-; SI-NEXT: ; implicit-def: $vgpr14
-; SI-NEXT: ; implicit-def: $vgpr10
-; SI-NEXT: ; implicit-def: $vgpr6
-; SI-NEXT: ; kill: killed $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:68
-; SI-NEXT: s_waitcnt expcnt(1)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:40
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v63
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:36
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v59
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:100
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:96
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v61
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:64
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:32
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32
-; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v60
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v4
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v12
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v20
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v47
-; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB98_2
-; SI-NEXT: ; %bb.1: ; %cmp.false
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v23, 0xffff, v47
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v46, v2, v6
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v45, v2, v6
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v44, v2, v6
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v8
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v43, v2, v6
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v42, v2, v6
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v12
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v41, v2, v6
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v40, v2, v6
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v16
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v55, v2, v6
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v54, v2, v6
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v20
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v53, v2, v6
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v52, v2, v6
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v24
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v51, v2, v6
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v50, v2, v6
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v49, v2, v27
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; SI-NEXT: ; implicit-def: $vgpr27
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v48, v2, v6
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v39, v2, v19
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; SI-NEXT: ; implicit-def: $vgpr19
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v38, v2, v6
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v37, v2, v15
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; SI-NEXT: ; implicit-def: $vgpr15
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v36, v2, v6
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v35, v2, v29
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; SI-NEXT: ; implicit-def: $vgpr29
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v34, v2, v6
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v33, v2, v21
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; SI-NEXT: ; implicit-def: $vgpr21
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v31, v2, v6
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v32, v2, v17
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; SI-NEXT: ; implicit-def: $vgpr17
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v26, v2, v25
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; SI-NEXT: ; implicit-def: $vgpr25
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v30, v2, v13
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; SI-NEXT: ; implicit-def: $vgpr13
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v18, v2, v11
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; SI-NEXT: ; implicit-def: $vgpr11
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v22, v2, v9
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; SI-NEXT: ; implicit-def: $vgpr9
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v14, v2, v7
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v10, v2, v5
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; SI-NEXT: ; implicit-def: $vgpr5
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v6, v2, v3
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; SI-NEXT: ; implicit-def: $vgpr3
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v2, v2, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v4
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v12
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v20
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v28
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v60
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v58
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v59
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v57
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v63
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_bfe_u32 v1, v4, 8, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_bfe_u32 v1, v8, 8, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_bfe_u32 v1, v12, 8, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_bfe_u32 v1, v16, 8, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_bfe_u32 v1, v20, 8, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_bfe_u32 v1, v24, 8, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_bfe_u32 v1, v28, 8, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_bfe_u32 v1, v60, 8, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_bfe_u32 v1, v58, 8, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_bfe_u32 v1, v59, 8, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_bfe_u32 v1, v57, 8, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_bfe_u32 v1, v63, 8, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v56
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_bfe_u32 v1, v56, 8, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v61
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_bfe_u32 v1, v61, 8, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v62
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_bfe_u32 v1, v62, 8, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_bfe_u32 v1, v47, 8, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v45, v46, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v45, v46, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v45, v46, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v43, v44, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v43, v44, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v43, v44, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v41, v42, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v41, v42, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v41, v42, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v55, v40, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v55, v40, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v55, v40, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v53, v54, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v53, v54, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v53, v54, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v51, v52, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v51, v52, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v51, v52, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v49, v50, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v49, v50, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v49, v50, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v39, v48, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v39, v48, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v39, v48, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v37, v38, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v37, v38, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v37, v38, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v35, v36, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v35, v36, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v35, v36, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v33, v34, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v33, v34, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v33, v34, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v32, v31, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v32, v31, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v32, v31, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v30, v26, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v30, v26, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v30, v26, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v22, v18, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v22, v18, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v22, v18, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v10, v14, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v10, v14, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v10, v14, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v2, v6, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v2, v6, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v2, v6, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v45
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v43
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v41
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v55
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v53
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v51
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v49
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v39
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v37
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v35
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v33
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v32
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v30
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v22
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v10
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v2
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: .LBB98_2: ; %Flow
-; SI-NEXT: s_or_saveexec_b64 s[4:5], s[4:5]
-; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; SI-NEXT: s_xor_b64 exec, exec, s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB98_4
-; SI-NEXT: ; %bb.3: ; %cmp.true
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; SI-NEXT: s_mov_b32 s6, 0x30000
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v2, v3, v2
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v6, vcc, 0x30000, v2
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v2, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; SI-NEXT: v_alignbit_b32 v12, v2, v6, 24
-; SI-NEXT: v_alignbit_b32 v20, v2, v6, 16
-; SI-NEXT: v_alignbit_b32 v47, v2, v6, 8
-; SI-NEXT: v_lshrrev_b32_e32 v8, 24, v2
-; SI-NEXT: v_lshrrev_b32_e32 v23, 16, v2
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v1, v11, v1
-; SI-NEXT: v_add_i32_e32 v18, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v1, v9, v1
-; SI-NEXT: v_add_i32_e32 v22, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; SI-NEXT: v_alignbit_b32 v62, v22, v18, 24
-; SI-NEXT: v_alignbit_b32 v63, v22, v18, 16
-; SI-NEXT: v_lshrrev_b32_e32 v59, 24, v22
-; SI-NEXT: v_lshrrev_b32_e32 v60, 16, v22
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v1, v25, v1
-; SI-NEXT: v_add_i32_e32 v26, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT: v_or_b32_e32 v3, v7, v3
-; SI-NEXT: v_add_i32_e32 v14, vcc, s6, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v1, v13, v1
-; SI-NEXT: v_add_i32_e32 v30, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v1, v3, v1
-; SI-NEXT: v_add_i32_e32 v31, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v1, v17, v1
-; SI-NEXT: v_add_i32_e32 v32, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; SI-NEXT: v_lshrrev_b32_e32 v16, 16, v32
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v1, v3, v1
-; SI-NEXT: v_add_i32_e32 v34, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v1, v21, v1
-; SI-NEXT: v_add_i32_e32 v33, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; SI-NEXT: v_lshrrev_b32_e32 v58, 16, v33
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v1, v3, v1
-; SI-NEXT: v_add_i32_e32 v36, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v1, v29, v1
-; SI-NEXT: v_add_i32_e32 v35, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v1, v3, v1
-; SI-NEXT: v_add_i32_e32 v38, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v1, v15, v1
-; SI-NEXT: v_add_i32_e32 v37, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v1, v3, v1
-; SI-NEXT: v_add_i32_e32 v48, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v1, v19, v1
-; SI-NEXT: v_add_i32_e32 v39, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v1, v3, v1
-; SI-NEXT: v_add_i32_e32 v50, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4
-; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: v_add_i32_e32 v10, vcc, s6, v4
-; SI-NEXT: v_alignbit_b32 v56, v10, v14, 24
-; SI-NEXT: v_alignbit_b32 v57, v10, v14, 16
-; SI-NEXT: v_alignbit_b32 v61, v10, v14, 8
-; SI-NEXT: v_lshrrev_b32_e32 v24, 24, v10
-; SI-NEXT: v_lshrrev_b32_e32 v28, 16, v10
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v1, v27, v1
-; SI-NEXT: v_add_i32_e32 v49, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v1, v3, v1
-; SI-NEXT: v_add_i32_e32 v52, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v1, v3, v1
-; SI-NEXT: v_add_i32_e32 v51, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v1, v3, v1
-; SI-NEXT: v_add_i32_e32 v54, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v1, v3, v1
-; SI-NEXT: v_add_i32_e32 v53, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v1, v3, v1
-; SI-NEXT: v_add_i32_e32 v40, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v1, v3, v1
-; SI-NEXT: v_add_i32_e32 v55, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v1, v3, v1
-; SI-NEXT: v_add_i32_e32 v42, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v1, v3, v1
-; SI-NEXT: v_add_i32_e32 v41, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v1, v3, v1
-; SI-NEXT: v_add_i32_e32 v44, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v1, v3, v1
-; SI-NEXT: v_add_i32_e32 v43, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v1, v3, v1
-; SI-NEXT: v_add_i32_e32 v46, vcc, s6, v1
-; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_or_b32_e32 v1, v3, v1
-; SI-NEXT: v_add_i32_e32 v45, vcc, s6, v1
-; SI-NEXT: v_alignbit_b32 v1, v45, v46, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v45, v46, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v43, v44, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v43, v44, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v43, v44, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v41, v42, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v41, v42, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v41, v42, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v55, v40, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v55, v40, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v55, v40, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v53, v54, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v53, v54, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v53, v54, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v51, v52, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v51, v52, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v51, v52, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v49, v50, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v49, v50, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v49, v50, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v39, v48, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v39, v48, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v39, v48, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v37, v38, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v37, v38, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v37, v38, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v35, v36, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v35, v36, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v35, v36, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v33, v34, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v33, v34, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v33, v34, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v32, v31, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v32, v31, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v32, v31, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v30, v26, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v30, v26, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v30, v26, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, v22, v18, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v45
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v45
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v45
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v43
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v43
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v43
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v41
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v41
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v41
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v55
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v55
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v55
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v53
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v53
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v53
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v51
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v51
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v51
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v49
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v49
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v49
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v39
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v39
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v39
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v37
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v37
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v37
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v35
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v35
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v35
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v33
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v33
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v32
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v32
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v30
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v30
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v30
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v22
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v10
-; SI-NEXT: v_alignbit_b32 v4, v45, v46, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v2
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; SI-NEXT: .LBB98_4: ; %end
-; SI-NEXT: s_or_b64 exec, exec, s[4:5]
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v46
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v45
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 4, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v44
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 8, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v43
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 12, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v42
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 16, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v41
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 20, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v40
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 24, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v55
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 28, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v54
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 32, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v53
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 36, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v52
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 40, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v51
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 44, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v50
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 48, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v49
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 52, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v48
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 56, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v39
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 60, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v38
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 64, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v37
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 0x44, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v36
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 0x48, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v35
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 0x4c, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v34
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 0x50, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v33
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v58
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 0x54, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v31
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 0x58, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v32
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v16
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 0x5c, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v26
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 0x60, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v30
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 0x64, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v18
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v62
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v63
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 0x68, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v22
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v59
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v60
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 0x6c, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v14
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v61
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v57
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v56
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 0x70, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v10
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v24
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v28
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 0x74, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v6
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v47
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v20
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v12
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_or_b32_e32 v1, v1, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 0x78, v0
-; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v2
-; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; SI-NEXT: v_lshlrev_b32_e32 v3, 24, v8
-; SI-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v2, 8, v2
-; SI-NEXT: v_or_b32_e32 v1, v1, v2
-; SI-NEXT: v_and_b32_e32 v2, 0xff, v23
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v2, v3, v2
-; SI-NEXT: v_or_b32_e32 v1, v1, v2
-; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; SI-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 .LBB98_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: .LBB98_2: ; %Flow
-; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB98_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: .LBB98_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(2)
-; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; 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: 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 .LBB98_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: .LBB98_2: ; %Flow
-; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; GFX9-NEXT: s_cbranch_execz .LBB98_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: .LBB98_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(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; 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(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; 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(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; 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(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; 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(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; 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(2)
-; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
-; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
-; 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-TRUE16-LABEL: bitcast_v64i16_to_v128i8:
-; GFX11-TRUE16: ; %bb.0:
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT: s_clause 0x2
-; GFX11-TRUE16-NEXT: scratch_load_b32 v33, off, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-TRUE16-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr163_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr162_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2)
-; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v33
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB98_2
-; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[9:10]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[7:8]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[15:16]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[5:6]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[49:50], 24, v[13:14]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[52:53], 24, v[11:12]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[3:4]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v15
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 24, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v5
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v2
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v1
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v31
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 24, v30
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v30
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v29
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 24, v28
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v28
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v27
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v26
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v26
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v25
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v24
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v24
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v23
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v22
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v22
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v21
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v20
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v20
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v19
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[1:2]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[19:20]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[17:18]
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v17
-; GFX11-TRUE16-NEXT: .LBB98_2: ; %Flow
-; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB98_4
-; GFX11-TRUE16-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v32, v32, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[9:10]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[7:8]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[15:16]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[5:6]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[49:50], 24, v[13:14]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[52:53], 24, v[11:12]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[3:4]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[1:2]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[19:20]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[17:18]
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v15
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 24, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v5
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v2
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v1
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v31
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 24, v30
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v30
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v29
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 24, v28
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v28
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v27
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v26
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v26
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v25
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v24
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v24
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v23
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v22
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v22
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v21
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v20
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v20
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v19
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v17
-; GFX11-TRUE16-NEXT: .LBB98_4: ; %end
-; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v163.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v68.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v161.l
-; GFX11-TRUE16-NEXT: v_or_b16 v51.l, v1.h, v34.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v162.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.h, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v54
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v160.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v150.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v68.l, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v69.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v67.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v54, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v68
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v67, 16, v69
-; GFX11-TRUE16-NEXT: v_or_b16 v51.l, v2.l, v2.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v3.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v151.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v54, v67
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v51
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.h, v4.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.h, v33.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v149.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v54
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v68.l, v4.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v66.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v3.h, v4.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v51, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v67
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v68
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v148.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v147.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v4.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v51, v54
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.h, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.h, v33.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v65.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v66
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v146.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v54
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.h, v6.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v51, v54
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v66
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v135.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v134.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v65
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v51, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v65
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v64.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v130.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v7.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v8.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v131.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v64, 0xffff, v64
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v65, 16, v65
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v66
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.h, v8.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v11.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v64, v65
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v65, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v117.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v52.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v7.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v51, v54
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v119.l
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h
-; GFX11-TRUE16-NEXT: s_clause 0x1
-; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off
-; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v103.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v115.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v114.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v13.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v49.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v33.h
-; GFX11-TRUE16-NEXT: v_or_b16 v51.l, v12.h, v34.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v11
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v101.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v48.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v65, v9
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v14.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v99.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v15.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v98.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v6
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v12, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v16.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v87.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v16.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v85.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v5
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v11
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v18.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v145.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v8
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v17.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v39.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v18.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v144.l
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v17.h
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v12, v13
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v14, v15
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v19.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v133.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v19.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v50.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v20.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v129.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v8
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v16, v17
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v20.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v132.l
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v11
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v21.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v128.l
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v13
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v21.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v38.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v10
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v15, v16
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v23.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v113.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v22.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v118.l
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v17, v18
-; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v23.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v37.l
-; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v24.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v112.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v22.h
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v12
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v19, v20
-; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v24.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v102.l
-; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v25.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v100.l
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h
-; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v17.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v15
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v25.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v36.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v13
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v21, v22
-; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v18.h
-; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v19.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v16
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v17
-; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v26.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v97.l
-; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v26.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v96.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v18
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h
-; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v15
-; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v86.l
-; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v27.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v16
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v17
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v20, v21
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v22, v23
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v24, v25
-; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v28.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v84.l
-; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v28.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v83.l
-; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v29.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v82.l
-; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v29.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v34.l
-; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v30.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v81.l
-; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v18.h
-; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h
-; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v21.l, v21.h
-; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v22.h
-; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v23.h
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v24.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v53.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v116.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v35.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v18
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v26, v27
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v20
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v21
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v22
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v23
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v24
-; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v30.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v80.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v31.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v71.l
-; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v31.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v33.l
-; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v32.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v70.l
-; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v32.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v55.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h
-; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v19.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v66
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v64, 16, v67
-; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h
-; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v21.l, v21.h
-; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v22.h
-; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v23.h
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v24.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v54, v64
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v30, 16, v20
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v21
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v22
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v23
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v39, v9
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v38, v14
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v35, v19
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v25, v26
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v27, v28
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v29, v30
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v31, v32
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v33, v24
-; GFX11-TRUE16-NEXT: s_clause 0x5
-; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off offset:32
-; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:48
-; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:64
-; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:80
-; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:96
-; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:112
-; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v64i16_to_v128i8:
-; GFX11-FAKE16: ; %bb.0:
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT: s_clause 0x13
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v40, s32 offset:88
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v41, s32 offset:84
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v42, s32 offset:80
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v43, s32 offset:76
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v44, s32 offset:72
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v45, s32 offset:68
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v46, s32 offset:64
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v47, s32 offset:60
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v56, s32 offset:56
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v57, s32 offset:52
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v58, s32 offset:48
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v59, s32 offset:44
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v60, s32 offset:40
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v61, s32 offset:36
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v62, s32 offset:32
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v63, s32 offset:28
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v72, s32 offset:24
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v73, s32 offset:20
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v74, s32 offset:16
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v75, s32 offset:12
-; GFX11-FAKE16-NEXT: s_clause 0x2
-; GFX11-FAKE16-NEXT: scratch_load_b32 v33, off, s32 offset:8
-; GFX11-FAKE16-NEXT: scratch_load_b32 v32, off, s32 offset:4
-; GFX11-FAKE16-NEXT: scratch_load_b32 v31, off, s32
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr74
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr72
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr64
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr63
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr62
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr60
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr57
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr47
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr54
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr45
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr43
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr183
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr181
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr53
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr179
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr177
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr167
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr165
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr164
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr52
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr161
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr151
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr150
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr147
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr145
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr51
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr144
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr133
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr131
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr129
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr119
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr50
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr117
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr116
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr115
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr102
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr100
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr49
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr98
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr97
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr96
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr86
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr82
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr48
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr80
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr70
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr68
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr75
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr73
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr66
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr61
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr59
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr58
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr56
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr46
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr65
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr44
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr41
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr40
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr182
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr180
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr178
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr176
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr166
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr163
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr162
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr160
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr149
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr148
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr146
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr135
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr134
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr132
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr130
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr128
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr118
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr114
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr113
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr112
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr103
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr101
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr99
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr87
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr85
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr84
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr83
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr81
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr71
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr69
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(2)
-; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v33
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB98_2
-; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[15:16]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[49:50], 24, v[13:14]
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[11:12]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[51:52], 24, v[9:10]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[52:53], 24, v[7:8]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[1:2]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[5:6]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[19:20]
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 24, v16
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v16
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v80, 8, v16
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v82, 16, v15
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v86, 8, v15
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v96, 24, v14
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v97, 16, v14
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v98, 8, v14
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v100, 16, v13
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v102, 8, v13
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v115, 24, v12
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v116, 16, v12
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v117, 8, v12
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v119, 16, v11
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v129, 8, v11
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v131, 24, v10
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v133, 16, v10
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v144, 8, v10
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v145, 16, v9
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v147, 8, v9
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v150, 24, v8
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v151, 16, v8
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v161, 8, v8
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v164, 16, v7
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v165, 8, v7
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v167, 24, v6
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v177, 16, v6
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v179, 8, v6
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v181, 16, v5
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v183, 8, v5
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v42, 24, v4
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v43, 16, v4
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v45, 8, v4
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v47, 16, v3
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v57, 8, v3
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v60, 24, v2
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v62, 16, v2
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v63, 8, v2
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v72, 16, v1
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v74, 8, v1
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 24, v32
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v32
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v81, 8, v32
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v83, 16, v31
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v84, 8, v31
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v85, 24, v30
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v87, 16, v30
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v99, 8, v30
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v101, 16, v29
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v103, 8, v29
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v112, 24, v28
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v113, 16, v28
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v114, 8, v28
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v118, 16, v27
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v128, 8, v27
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v130, 24, v26
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v132, 16, v26
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v134, 8, v26
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v135, 16, v25
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v146, 8, v25
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v148, 24, v24
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v149, 16, v24
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v160, 8, v24
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v162, 16, v23
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v163, 8, v23
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v166, 24, v22
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v176, 16, v22
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v178, 8, v22
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v180, 16, v21
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v182, 8, v21
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v40, 24, v20
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v41, 16, v20
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v44, 8, v20
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v46, 16, v19
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v56, 8, v19
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v58, 24, v18
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v59, 16, v18
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v61, 8, v18
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v73, 16, v17
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v75, 8, v17
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[3:4]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[17:18]
-; GFX11-FAKE16-NEXT: .LBB98_2: ; %Flow
-; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB98_4
-; GFX11-FAKE16-NEXT: ; %bb.3: ; %cmp.true
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(1)
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v32, v32, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[15:16]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[49:50], 24, v[13:14]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[11:12]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[51:52], 24, v[9:10]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[52:53], 24, v[7:8]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[1:2]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[5:6]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[19:20]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[3:4]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[17:18]
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 24, v16
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v16
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v80, 8, v16
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v82, 16, v15
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v86, 8, v15
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v96, 24, v14
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v97, 16, v14
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v98, 8, v14
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v100, 16, v13
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v102, 8, v13
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v115, 24, v12
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v116, 16, v12
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v117, 8, v12
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v119, 16, v11
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v129, 8, v11
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v131, 24, v10
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v133, 16, v10
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v144, 8, v10
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v145, 16, v9
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v147, 8, v9
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v150, 24, v8
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v151, 16, v8
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v161, 8, v8
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v164, 16, v7
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v165, 8, v7
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v167, 24, v6
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v177, 16, v6
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v179, 8, v6
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v181, 16, v5
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v183, 8, v5
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v42, 24, v4
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v43, 16, v4
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v45, 8, v4
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v47, 16, v3
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v57, 8, v3
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v60, 24, v2
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v62, 16, v2
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v63, 8, v2
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v72, 16, v1
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v74, 8, v1
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 24, v32
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v32
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v81, 8, v32
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v83, 16, v31
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v84, 8, v31
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v85, 24, v30
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v87, 16, v30
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v99, 8, v30
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v101, 16, v29
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v103, 8, v29
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v112, 24, v28
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v113, 16, v28
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v114, 8, v28
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v118, 16, v27
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v128, 8, v27
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v130, 24, v26
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v132, 16, v26
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v134, 8, v26
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v135, 16, v25
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v146, 8, v25
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v148, 24, v24
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v149, 16, v24
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v160, 8, v24
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v162, 16, v23
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v163, 8, v23
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v166, 24, v22
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v176, 16, v22
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v178, 8, v22
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v180, 16, v21
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v182, 8, v21
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v40, 24, v20
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v41, 16, v20
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v44, 8, v20
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v46, 16, v19
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v56, 8, v19
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v58, 24, v18
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v59, 16, v18
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v61, 8, v18
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v73, 16, v17
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v75, 8, v17
-; GFX11-FAKE16-NEXT: .LBB98_4: ; %end
-; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v55, 8, v74
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v39, 8, v64
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v64, 8, v60
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v54, 8, v54
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v55
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v55, 0xff, v72
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v67, 0xff, v47
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v39, v55, v39
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v55, 8, v63
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v54, v67, v54
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v67, 8, v42
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v39, 16, v39
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v2, v55
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v55, 0xff, v62
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v53, 8, v53
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v39
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v55, v55, v64
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v64, 8, v57
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v52, 8, v52
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v39, 16, v55
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v3, v64
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v55, 8, v45
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v64, 0xff, v43
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v51, 8, v51
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v2, v39
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v39, 16, v54
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v4, v55
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v54, v64, v67
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v55, 8, v183
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v64, 0xff, v181
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v3, v39
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v39, 16, v54
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v5, v55
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v53, v64, v53
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v54, 8, v179
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v55, 0xff, v177
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v64, 8, v167
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v4, v39
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v39, 16, v53
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v6, v6, v54
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v53, v55, v64
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v54, 8, v165
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v55, 0xff, v164
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v5, v39
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v39, 16, v53
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v7, v54
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v52, v55, v52
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v53, 8, v161
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v54, 0xff, v151
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v55, 8, v150
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v6, v6, v39
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v39, 16, v52
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v8, v8, v53
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v52, v54, v55
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v54, 0xff, v145
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v53, 8, v147
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v55, 8, v144
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v52, 16, v52
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v51, v54, v51
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v9, v9, v53
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v10, v10, v55
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v7, v39
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v8, v8, v52
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v39, 16, v51
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v51, 0xff, v133
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v52, 8, v131
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v53, 8, v129
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v54, 0xff, v119
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v50, 8, v50
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v55, 8, v117
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v64, 0xff, v116
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v67, 8, v115
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v51, v51, v52
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v11, v11, v53
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v50, v54, v50
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v12, v12, v55
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v52, v64, v67
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v50, 16, v50
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v52, 16, v52
-; GFX11-FAKE16-NEXT: s_clause 0x1
-; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[1:4], off
-; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v9, v39
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v10, v51
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v11, v50
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v12, v52
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xff, v13
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v6, 8, v102
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xff, v100
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v8, 8, v49
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v9, 0xff, v14
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v10, 8, v98
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xff, v97
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v12, 8, v96
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xff, v15
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v14, 8, v86
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v6, v7, v8
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v9, v10
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v8, v11, v12
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v9, v13, v14
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xff, v82
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v11, 8, v48
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xff, v16
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v13, 8, v80
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xff, v70
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v15, 8, v68
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v16, 0xff, v17
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v17, 8, v75
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v39, 0xff, v73
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v48, 8, v66
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v14, v39, v48
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v6, v7, v8
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v9, v10
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v8, v11, v12
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v9, v13, v14
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xff, v18
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v11, 8, v61
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xff, v59
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v13, 8, v58
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xff, v19
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v15, 8, v56
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v16, 0xff, v46
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v17, 8, v65
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xff, v20
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v19, 8, v44
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v14, v18, v19
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xff, v41
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v16, 8, v40
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v17, 0xff, v21
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v18, 8, v182
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xff, v180
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v20, 8, v38
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xff, v22
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v22, 8, v178
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v38, 0xff, v176
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v39, 8, v166
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v16, v17, v18
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v17, v19, v20
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v18, v21, v22
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v19, v38, v39
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v10, v10, v11
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v11, v12, v13
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v12, v14, v15
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v13, v16, v17
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v14, v18, v19
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xff, v23
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v16, 8, v163
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v17, 0xff, v162
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v18, 8, v37
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xff, v24
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v20, 8, v160
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xff, v149
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v22, 8, v148
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v23, 0xff, v25
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v24, 8, v146
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v16, v17, v18
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v17, v19, v20
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v18, v21, v22
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v19, v23, v24
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xff, v135
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v21, 8, v36
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xff, v26
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v23, 8, v134
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v24, 0xff, v132
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v25, 8, v130
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v26, 0xff, v27
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v27, 8, v128
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v36, 0xff, v118
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v35, 8, v35
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v21, v22, v23
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v22, v24, v25
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v23, v26, v27
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v24, v36, v35
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v16, v17, v18
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v17, v19, v20
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v18, v21, v22
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v19, v23, v24
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xff, v28
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v21, 8, v114
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xff, v113
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v23, 8, v112
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v24, 0xff, v29
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v25, 8, v103
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v26, 0xff, v101
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v27, 8, v34
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v28, 0xff, v30
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v29, 8, v99
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v21, v22, v23
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v22, v24, v25
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v23, v26, v27
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v24, v28, v29
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v25, 0xff, v87
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v26, 8, v85
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v27, 0xff, v31
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v28, 8, v84
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v29, 0xff, v83
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v30, 8, v33
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v31, 0xff, v32
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v32, 8, v81
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xff, v71
-; GFX11-FAKE16-NEXT: v_lshlrev_b16 v34, 8, v69
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v25, v25, v26
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v26, v27, v28
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v27, v29, v30
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v28, v31, v32
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v29, v33, v34
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v20, v20, v21
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v21, v22, v23
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v22, v24, v25
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v23, v26, v27
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v24, v28, v29
-; GFX11-FAKE16-NEXT: s_clause 0x5
-; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[1:4], off offset:32
-; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:48
-; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:64
-; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:80
-; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:96
-; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:112
-; GFX11-FAKE16-NEXT: s_clause 0x13
-; GFX11-FAKE16-NEXT: scratch_load_b32 v75, off, s32 offset:12
-; GFX11-FAKE16-NEXT: scratch_load_b32 v74, off, s32 offset:16
-; GFX11-FAKE16-NEXT: scratch_load_b32 v73, off, s32 offset:20
-; GFX11-FAKE16-NEXT: scratch_load_b32 v72, off, s32 offset:24
-; GFX11-FAKE16-NEXT: scratch_load_b32 v63, off, s32 offset:28
-; GFX11-FAKE16-NEXT: scratch_load_b32 v62, off, s32 offset:32
-; GFX11-FAKE16-NEXT: scratch_load_b32 v61, off, s32 offset:36
-; GFX11-FAKE16-NEXT: scratch_load_b32 v60, off, s32 offset:40
-; GFX11-FAKE16-NEXT: scratch_load_b32 v59, off, s32 offset:44
-; GFX11-FAKE16-NEXT: scratch_load_b32 v58, off, s32 offset:48
-; GFX11-FAKE16-NEXT: scratch_load_b32 v57, off, s32 offset:52
-; GFX11-FAKE16-NEXT: scratch_load_b32 v56, off, s32 offset:56
-; GFX11-FAKE16-NEXT: scratch_load_b32 v47, off, s32 offset:60
-; GFX11-FAKE16-NEXT: scratch_load_b32 v46, off, s32 offset:64
-; GFX11-FAKE16-NEXT: scratch_load_b32 v45, off, s32 offset:68
-; GFX11-FAKE16-NEXT: scratch_load_b32 v44, off, s32 offset:72
-; GFX11-FAKE16-NEXT: scratch_load_b32 v43, off, s32 offset:76
-; GFX11-FAKE16-NEXT: scratch_load_b32 v42, off, s32 offset:80
-; GFX11-FAKE16-NEXT: scratch_load_b32 v41, off, s32 offset:84
-; GFX11-FAKE16-NEXT: scratch_load_b32 v40, off, s32 offset:88
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT: 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 inreg <128 x i8> @bitcast_v64i16_to_v128i8_scalar(<64 x i16> inreg %a, i32 inreg %b) {
-; SI-LABEL: bitcast_v64i16_to_v128i8_scalar:
-; SI: ; %bb.0:
-; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; SI-NEXT: s_or_saveexec_b64 s[4:5], -1
-; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; SI-NEXT: s_mov_b64 exec, s[4:5]
-; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; SI-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:76
-; SI-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:72
-; SI-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:60
-; SI-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:56
-; SI-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:44
-; SI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:40
-; SI-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:28
-; SI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:24
-; SI-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:12
-; SI-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:8
-; SI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:80
-; SI-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:36
-; SI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:32
-; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:20
-; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:4
-; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:68
-; SI-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:64
-; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:52
-; SI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:48
-; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:16
-; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32
-; SI-NEXT: v_writelane_b32 v63, s30, 0
-; SI-NEXT: v_writelane_b32 v63, s31, 1
-; SI-NEXT: v_writelane_b32 v63, s34, 2
-; SI-NEXT: v_writelane_b32 v63, s35, 3
-; SI-NEXT: v_writelane_b32 v63, s36, 4
-; SI-NEXT: v_writelane_b32 v63, s37, 5
-; SI-NEXT: v_writelane_b32 v63, s38, 6
-; SI-NEXT: v_writelane_b32 v63, s39, 7
-; SI-NEXT: v_writelane_b32 v63, s48, 8
-; SI-NEXT: v_writelane_b32 v63, s49, 9
-; SI-NEXT: v_writelane_b32 v63, s50, 10
-; SI-NEXT: v_writelane_b32 v63, s51, 11
-; SI-NEXT: v_writelane_b32 v63, s52, 12
-; SI-NEXT: v_writelane_b32 v63, s53, 13
-; SI-NEXT: v_writelane_b32 v63, s54, 14
-; SI-NEXT: v_writelane_b32 v63, s55, 15
-; SI-NEXT: v_writelane_b32 v63, s64, 16
-; SI-NEXT: v_writelane_b32 v63, s65, 17
-; SI-NEXT: v_writelane_b32 v63, s66, 18
-; SI-NEXT: v_writelane_b32 v63, s67, 19
-; SI-NEXT: v_writelane_b32 v63, s68, 20
-; SI-NEXT: v_writelane_b32 v63, s69, 21
-; SI-NEXT: v_writelane_b32 v63, s70, 22
-; SI-NEXT: v_writelane_b32 v63, s71, 23
-; SI-NEXT: v_writelane_b32 v63, s80, 24
-; SI-NEXT: v_writelane_b32 v63, s81, 25
-; SI-NEXT: v_writelane_b32 v63, s82, 26
-; SI-NEXT: v_writelane_b32 v63, s83, 27
-; SI-NEXT: v_writelane_b32 v63, s84, 28
-; SI-NEXT: v_writelane_b32 v63, s85, 29
-; SI-NEXT: v_writelane_b32 v63, s86, 30
-; SI-NEXT: v_writelane_b32 v63, s87, 31
-; SI-NEXT: v_writelane_b32 v63, s96, 32
-; SI-NEXT: v_writelane_b32 v63, s97, 33
-; SI-NEXT: v_writelane_b32 v63, s98, 34
-; SI-NEXT: s_mov_b32 s6, s18
-; SI-NEXT: ; implicit-def: $vgpr62 : SGPR spill to VGPR lane
-; SI-NEXT: v_writelane_b32 v63, s99, 35
-; SI-NEXT: v_readfirstlane_b32 s62, v30
-; SI-NEXT: v_readfirstlane_b32 s63, v29
-; SI-NEXT: v_readfirstlane_b32 s59, v26
-; SI-NEXT: v_readfirstlane_b32 s60, v25
-; SI-NEXT: v_readfirstlane_b32 s98, v22
-; SI-NEXT: v_readfirstlane_b32 s61, v21
-; SI-NEXT: v_readfirstlane_b32 s99, v18
-; SI-NEXT: v_readfirstlane_b32 s58, v17
-; SI-NEXT: v_readfirstlane_b32 s96, v14
-; SI-NEXT: v_readfirstlane_b32 s97, v13
-; SI-NEXT: v_readfirstlane_b32 s86, v10
-; SI-NEXT: v_readfirstlane_b32 s87, v9
-; SI-NEXT: v_readfirstlane_b32 s84, v6
-; SI-NEXT: v_readfirstlane_b32 s85, v5
-; SI-NEXT: v_readfirstlane_b32 s81, v2
-; SI-NEXT: v_readfirstlane_b32 s82, v1
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v4
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v8
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v12
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v16
-; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v20
-; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v24
-; SI-NEXT: s_waitcnt vmcnt(14)
-; SI-NEXT: v_readfirstlane_b32 s88, v36
-; SI-NEXT: v_readfirstlane_b32 s18, v37
-; SI-NEXT: v_readfirstlane_b32 s78, v38
-; SI-NEXT: v_readfirstlane_b32 s79, v39
-; SI-NEXT: v_readfirstlane_b32 s76, v48
-; SI-NEXT: v_readfirstlane_b32 s77, v49
-; SI-NEXT: v_readfirstlane_b32 s74, v50
-; SI-NEXT: s_waitcnt vmcnt(13)
-; SI-NEXT: v_readfirstlane_b32 s75, v51
-; SI-NEXT: s_waitcnt vmcnt(12)
-; SI-NEXT: v_readfirstlane_b32 s72, v52
-; SI-NEXT: s_waitcnt vmcnt(11)
-; SI-NEXT: v_readfirstlane_b32 s73, v53
-; SI-NEXT: s_waitcnt vmcnt(10)
-; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v54
-; SI-NEXT: s_and_b64 s[4:5], vcc, exec
-; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v28
-; SI-NEXT: s_waitcnt vmcnt(7)
-; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v40
-; SI-NEXT: s_waitcnt vmcnt(6)
-; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v41
-; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v55
-; SI-NEXT: s_waitcnt vmcnt(5)
-; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v42
-; SI-NEXT: s_waitcnt vmcnt(3)
-; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v43
-; SI-NEXT: v_writelane_b32 v62, s6, 0
-; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; SI-NEXT: s_cbranch_scc0 .LBB99_4
-; SI-NEXT: ; %bb.1: ; %cmp.false
-; SI-NEXT: s_and_b32 s4, s16, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s17, 16
-; SI-NEXT: s_or_b32 s56, s4, s5
-; SI-NEXT: s_and_b32 s4, s6, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s19, 16
-; SI-NEXT: s_or_b32 s57, s4, s5
-; SI-NEXT: v_mov_b32_e32 v1, s56
-; SI-NEXT: s_and_b32 s4, s20, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s21, 16
-; SI-NEXT: v_alignbit_b32 v8, s57, v1, 24
-; SI-NEXT: v_alignbit_b32 v50, s57, v1, 16
-; SI-NEXT: v_alignbit_b32 v1, s57, v1, 8
-; SI-NEXT: s_or_b32 s46, s4, s5
-; SI-NEXT: s_and_b32 s4, s22, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s23, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; SI-NEXT: s_or_b32 s47, s4, s5
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_mov_b32_e32 v1, s46
-; SI-NEXT: s_and_b32 s4, s24, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s25, 16
-; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v8, s47, v1, 24
-; SI-NEXT: s_or_b32 s44, s4, s5
-; SI-NEXT: s_and_b32 s4, s26, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s27, 16
-; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v8, s47, v1, 16
-; SI-NEXT: v_alignbit_b32 v51, s47, v1, 8
-; SI-NEXT: s_or_b32 s45, s4, s5
-; SI-NEXT: v_mov_b32_e32 v1, s44
-; SI-NEXT: s_and_b32 s4, s28, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s29, 16
-; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v8, s45, v1, 24
-; SI-NEXT: s_or_b32 s42, s4, s5
-; SI-NEXT: s_and_b32 s4, s82, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s81, 16
-; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v8, s45, v1, 16
-; SI-NEXT: v_alignbit_b32 v49, s45, v1, 8
-; SI-NEXT: s_or_b32 s43, s4, s5
-; SI-NEXT: v_mov_b32_e32 v1, s42
-; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v8, s43, v1, 24
-; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v8, s43, v1, 16
-; SI-NEXT: v_alignbit_b32 v48, s43, v1, 8
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v3
-; SI-NEXT: s_and_b32 s4, s85, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s84, 16
-; SI-NEXT: v_or_b32_e32 v16, v1, v2
-; SI-NEXT: s_or_b32 s41, s4, s5
-; SI-NEXT: v_alignbit_b32 v1, s41, v16, 24
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, s41, v16, 16
-; SI-NEXT: s_and_b32 s4, s87, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s86, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, s41, v16, 8
-; SI-NEXT: s_or_b32 s40, s4, s5
-; SI-NEXT: s_and_b32 s4, s97, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s96, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v7
-; SI-NEXT: s_or_b32 s15, s4, s5
-; SI-NEXT: s_and_b32 s4, s58, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s99, 16
-; SI-NEXT: v_or_b32_e32 v14, v1, v4
-; SI-NEXT: s_or_b32 s14, s4, s5
-; SI-NEXT: s_and_b32 s4, s61, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s98, 16
-; SI-NEXT: v_alignbit_b32 v1, s40, v14, 24
-; SI-NEXT: s_or_b32 s13, s4, s5
-; SI-NEXT: s_and_b32 s4, s60, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s59, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, s40, v14, 16
-; SI-NEXT: s_or_b32 s12, s4, s5
-; SI-NEXT: s_and_b32 s4, s63, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s62, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, s40, v14, 8
-; SI-NEXT: s_or_b32 s11, s4, s5
-; SI-NEXT: s_and_b32 s4, s73, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s72, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v11
-; SI-NEXT: s_or_b32 s10, s4, s5
-; SI-NEXT: s_and_b32 s4, s75, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s74, 16
-; SI-NEXT: v_or_b32_e32 v12, v1, v5
-; SI-NEXT: s_or_b32 s9, s4, s5
-; SI-NEXT: s_and_b32 s4, s77, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s76, 16
-; SI-NEXT: v_alignbit_b32 v1, s15, v12, 24
-; SI-NEXT: s_or_b32 s8, s4, s5
-; SI-NEXT: s_and_b32 s4, s79, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s78, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, s15, v12, 16
-; SI-NEXT: s_or_b32 s7, s4, s5
-; SI-NEXT: s_and_b32 s4, s18, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s88, 16
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, s15, v12, 8
-; SI-NEXT: s_or_b32 s6, s4, s5
-; SI-NEXT: s_lshr_b32 s4, s11, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v15
-; SI-NEXT: v_writelane_b32 v62, s4, 1
-; SI-NEXT: s_lshr_b32 s4, s10, 8
-; SI-NEXT: v_or_b32_e32 v10, v1, v6
-; SI-NEXT: v_writelane_b32 v62, s4, 3
-; SI-NEXT: s_lshr_b32 s4, s9, 8
-; SI-NEXT: v_alignbit_b32 v1, s14, v10, 24
-; SI-NEXT: v_writelane_b32 v62, s4, 6
-; SI-NEXT: s_lshr_b32 s4, s8, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, s14, v10, 16
-; SI-NEXT: v_writelane_b32 v62, s4, 9
-; SI-NEXT: s_lshr_b32 s4, s7, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v1, s14, v10, 8
-; SI-NEXT: v_writelane_b32 v62, s4, 12
-; SI-NEXT: s_lshr_b32 s4, s6, 8
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v19
-; SI-NEXT: v_writelane_b32 v62, s4, 15
-; SI-NEXT: s_and_b32 s4, s72, 0xffff
-; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_or_b32_e32 v8, v1, v9
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v23
-; SI-NEXT: v_writelane_b32 v62, s4, 2
-; SI-NEXT: s_and_b32 s4, s74, 0xffff
-; SI-NEXT: v_or_b32_e32 v5, v1, v13
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v27
-; SI-NEXT: v_writelane_b32 v62, s4, 5
-; SI-NEXT: s_and_b32 s4, s76, 0xffff
-; SI-NEXT: v_mov_b32_e32 v28, v13
-; SI-NEXT: v_or_b32_e32 v13, v1, v17
-; SI-NEXT: s_waitcnt vmcnt(14)
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v31
-; SI-NEXT: v_writelane_b32 v62, s4, 8
-; SI-NEXT: s_and_b32 s4, s78, 0xffff
-; SI-NEXT: v_mov_b32_e32 v26, v9
-; SI-NEXT: v_or_b32_e32 v9, v1, v18
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v32
-; SI-NEXT: v_writelane_b32 v62, s4, 11
-; SI-NEXT: s_and_b32 s4, s88, 0xffff
-; SI-NEXT: v_mov_b32_e32 v25, v6
-; SI-NEXT: v_or_b32_e32 v6, v1, v20
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v33
-; SI-NEXT: v_writelane_b32 v62, s4, 14
-; SI-NEXT: s_bfe_u32 s4, s74, 0x80008
-; SI-NEXT: v_or_b32_e32 v4, v1, v21
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v34
-; SI-NEXT: v_writelane_b32 v62, s4, 4
-; SI-NEXT: s_bfe_u32 s4, s76, 0x80008
-; SI-NEXT: v_or_b32_e32 v2, v1, v22
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v35
-; SI-NEXT: v_writelane_b32 v62, s4, 7
-; SI-NEXT: s_bfe_u32 s4, s78, 0x80008
-; SI-NEXT: v_or_b32_e32 v1, v1, v24
-; SI-NEXT: v_writelane_b32 v62, s4, 10
-; SI-NEXT: s_bfe_u32 s4, s88, 0x80008
-; SI-NEXT: v_mov_b32_e32 v29, v17
-; SI-NEXT: v_mov_b32_e32 v30, v18
-; SI-NEXT: v_mov_b32_e32 v36, v20
-; SI-NEXT: v_mov_b32_e32 v37, v21
-; SI-NEXT: v_mov_b32_e32 v38, v22
-; SI-NEXT: v_mov_b32_e32 v39, v24
-; SI-NEXT: s_lshr_b32 s68, s57, 8
-; SI-NEXT: s_lshr_b32 s65, s47, 8
-; SI-NEXT: s_lshr_b32 s54, s45, 8
-; SI-NEXT: s_lshr_b32 s51, s43, 8
-; SI-NEXT: s_lshr_b32 s48, s41, 8
-; SI-NEXT: s_lshr_b32 s37, s40, 8
-; SI-NEXT: s_lshr_b32 s34, s15, 8
-; SI-NEXT: s_lshr_b32 s95, s14, 8
-; SI-NEXT: s_lshr_b32 s92, s13, 8
-; SI-NEXT: s_lshr_b32 s89, s12, 8
-; SI-NEXT: s_and_b32 s71, s19, 0xffff
-; SI-NEXT: s_and_b32 s69, s23, 0xffff
-; SI-NEXT: s_and_b32 s66, s27, 0xffff
-; SI-NEXT: s_and_b32 s55, s81, 0xffff
-; SI-NEXT: s_and_b32 s52, s84, 0xffff
-; SI-NEXT: s_and_b32 s49, s86, 0xffff
-; SI-NEXT: s_and_b32 s38, s96, 0xffff
-; SI-NEXT: s_and_b32 s35, s99, 0xffff
-; SI-NEXT: s_and_b32 s30, s98, 0xffff
-; SI-NEXT: s_and_b32 s93, s59, 0xffff
-; SI-NEXT: s_and_b32 s90, s62, 0xffff
-; SI-NEXT: s_bfe_u32 s83, s19, 0x80008
-; SI-NEXT: s_bfe_u32 s80, s23, 0x80008
-; SI-NEXT: s_bfe_u32 s70, s27, 0x80008
-; SI-NEXT: s_bfe_u32 s67, s81, 0x80008
-; SI-NEXT: s_bfe_u32 s64, s84, 0x80008
-; SI-NEXT: s_bfe_u32 s53, s86, 0x80008
-; SI-NEXT: s_bfe_u32 s50, s96, 0x80008
-; SI-NEXT: s_bfe_u32 s39, s99, 0x80008
-; SI-NEXT: s_bfe_u32 s36, s98, 0x80008
-; SI-NEXT: s_bfe_u32 s31, s59, 0x80008
-; SI-NEXT: s_bfe_u32 s94, s62, 0x80008
-; SI-NEXT: s_bfe_u32 s91, s72, 0x80008
-; SI-NEXT: v_writelane_b32 v62, s4, 13
-; SI-NEXT: v_alignbit_b32 v45, s13, v8, 24
-; SI-NEXT: v_alignbit_b32 v47, s13, v8, 16
-; SI-NEXT: v_alignbit_b32 v57, s13, v8, 8
-; SI-NEXT: v_alignbit_b32 v41, s12, v5, 24
-; SI-NEXT: v_alignbit_b32 v43, s12, v5, 16
-; SI-NEXT: v_alignbit_b32 v44, s12, v5, 8
-; SI-NEXT: v_alignbit_b32 v21, s11, v13, 24
-; SI-NEXT: v_alignbit_b32 v22, s11, v13, 16
-; SI-NEXT: v_alignbit_b32 v24, s11, v13, 8
-; SI-NEXT: v_alignbit_b32 v17, s10, v9, 24
-; SI-NEXT: v_alignbit_b32 v18, s10, v9, 16
-; SI-NEXT: v_alignbit_b32 v20, s10, v9, 8
-; SI-NEXT: v_alignbit_b32 v59, s9, v6, 24
-; SI-NEXT: v_alignbit_b32 v60, s9, v6, 16
-; SI-NEXT: v_alignbit_b32 v61, s9, v6, 8
-; SI-NEXT: v_alignbit_b32 v46, s8, v4, 24
-; SI-NEXT: v_alignbit_b32 v56, s8, v4, 16
-; SI-NEXT: v_alignbit_b32 v58, s8, v4, 8
-; SI-NEXT: v_alignbit_b32 v55, s7, v2, 24
-; SI-NEXT: v_alignbit_b32 v40, s7, v2, 16
-; SI-NEXT: v_alignbit_b32 v42, s7, v2, 8
-; SI-NEXT: v_alignbit_b32 v52, s6, v1, 24
-; SI-NEXT: v_alignbit_b32 v53, s6, v1, 16
-; SI-NEXT: v_alignbit_b32 v54, s6, v1, 8
-; SI-NEXT: s_cbranch_execnz .LBB99_3
-; SI-NEXT: .LBB99_2: ; %cmp.true
-; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT: s_add_i32 s18, s18, 3
-; SI-NEXT: s_and_b32 s4, s18, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s88, 16
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s79, s79, 3
-; SI-NEXT: s_add_i32 s6, s4, 0x30000
-; SI-NEXT: s_and_b32 s4, s79, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s78, 16
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s77, s77, 3
-; SI-NEXT: s_add_i32 s7, s4, 0x30000
-; SI-NEXT: s_and_b32 s4, s77, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s76, 16
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s75, s75, 3
-; SI-NEXT: s_add_i32 s8, s4, 0x30000
-; SI-NEXT: s_and_b32 s4, s75, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s74, 16
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s73, s73, 3
-; SI-NEXT: s_add_i32 s9, s4, 0x30000
-; SI-NEXT: s_and_b32 s4, s73, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s72, 16
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s63, s63, 3
-; SI-NEXT: s_add_i32 s10, s4, 0x30000
-; SI-NEXT: s_and_b32 s4, s63, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s62, 16
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s60, s60, 3
-; SI-NEXT: s_add_i32 s11, s4, 0x30000
-; SI-NEXT: s_and_b32 s4, s60, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s59, 16
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s61, s61, 3
-; SI-NEXT: s_add_i32 s12, s4, 0x30000
-; SI-NEXT: s_and_b32 s4, s61, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s98, 16
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s58, s58, 3
-; SI-NEXT: s_add_i32 s13, s4, 0x30000
-; SI-NEXT: s_and_b32 s4, s58, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s99, 16
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s97, s97, 3
-; SI-NEXT: s_add_i32 s14, s4, 0x30000
-; SI-NEXT: s_and_b32 s4, s97, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s96, 16
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s87, s87, 3
-; SI-NEXT: s_add_i32 s15, s4, 0x30000
-; SI-NEXT: s_and_b32 s4, s87, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s86, 16
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s85, s85, 3
-; SI-NEXT: s_add_i32 s40, s4, 0x30000
-; SI-NEXT: s_and_b32 s4, s85, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s84, 16
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s28, s28, 3
-; SI-NEXT: s_add_i32 s41, s4, 0x30000
-; SI-NEXT: s_and_b32 s4, s28, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s29, 16
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s82, s82, 3
-; SI-NEXT: s_add_i32 s42, s4, 0x30000
-; SI-NEXT: s_and_b32 s4, s82, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s81, 16
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s24, s24, 3
-; SI-NEXT: s_add_i32 s43, s4, 0x30000
-; SI-NEXT: s_and_b32 s4, s24, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s25, 16
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s26, s26, 3
-; SI-NEXT: s_add_i32 s44, s4, 0x30000
-; SI-NEXT: s_and_b32 s4, s26, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s27, 16
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s20, s20, 3
-; SI-NEXT: s_add_i32 s45, s4, 0x30000
-; SI-NEXT: s_and_b32 s4, s20, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s21, 16
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s22, s22, 3
-; SI-NEXT: s_add_i32 s46, s4, 0x30000
-; SI-NEXT: s_and_b32 s4, s22, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s23, 16
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s16, s16, 3
-; SI-NEXT: s_add_i32 s47, s4, 0x30000
-; SI-NEXT: s_and_b32 s4, s16, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s17, 16
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s56, s4, 0x30000
-; SI-NEXT: v_readlane_b32 s4, v62, 0
-; SI-NEXT: s_add_i32 s4, s4, 3
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_lshl_b32 s5, s19, 16
-; SI-NEXT: s_or_b32 s4, s5, s4
-; SI-NEXT: s_add_i32 s57, s4, 0x30000
-; SI-NEXT: v_mov_b32_e32 v21, s56
-; SI-NEXT: v_alignbit_b32 v22, s57, v21, 24
-; SI-NEXT: v_alignbit_b32 v50, s57, v21, 16
-; SI-NEXT: v_alignbit_b32 v21, s57, v21, 8
-; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_mov_b32_e32 v21, s46
-; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v22, s47, v21, 24
-; SI-NEXT: s_lshr_b32 s4, s11, 8
-; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v3
-; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v22, s47, v21, 16
-; SI-NEXT: v_alignbit_b32 v51, s47, v21, 8
-; SI-NEXT: v_mov_b32_e32 v21, s44
-; SI-NEXT: v_writelane_b32 v62, s4, 1
-; SI-NEXT: s_lshr_b32 s4, s10, 16
-; SI-NEXT: s_waitcnt vmcnt(10)
-; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v32
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v22, s45, v21, 24
-; SI-NEXT: v_writelane_b32 v62, s4, 2
-; SI-NEXT: s_lshr_b32 s4, s10, 8
-; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v7
-; SI-NEXT: s_waitcnt vmcnt(4)
-; SI-NEXT: v_or_b32_e32 v3, v16, v3
-; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v22, s45, v21, 16
-; SI-NEXT: v_alignbit_b32 v49, s45, v21, 8
-; SI-NEXT: v_mov_b32_e32 v21, s42
-; SI-NEXT: v_writelane_b32 v62, s4, 3
-; SI-NEXT: s_lshr_b32 s4, s9, 24
-; SI-NEXT: v_or_b32_e32 v5, v36, v5
-; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; SI-NEXT: v_add_i32_e32 v16, vcc, 0x30000, v3
-; SI-NEXT: v_mov_b32_e32 v3, s41
-; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v22, s43, v21, 24
-; SI-NEXT: v_writelane_b32 v62, s4, 4
-; SI-NEXT: s_lshr_b32 s4, s9, 16
-; SI-NEXT: v_add_i32_e32 v6, vcc, 0x30000, v5
-; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v31
-; SI-NEXT: v_or_b32_e32 v7, v14, v7
-; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v22, s43, v21, 16
-; SI-NEXT: v_alignbit_b32 v48, s43, v21, 8
-; SI-NEXT: v_alignbit_b32 v21, v3, v16, 24
-; SI-NEXT: v_writelane_b32 v62, s4, 5
-; SI-NEXT: s_lshr_b32 s4, s9, 8
-; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11
-; SI-NEXT: v_add_i32_e32 v14, vcc, 0x30000, v7
-; SI-NEXT: v_mov_b32_e32 v7, s40
-; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v21, v3, v16, 16
-; SI-NEXT: v_alignbit_b32 v3, v3, v16, 8
-; SI-NEXT: v_writelane_b32 v62, s4, 6
-; SI-NEXT: s_lshr_b32 s4, s8, 24
-; SI-NEXT: v_or_b32_e32 v5, v30, v5
-; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v3, v7, v14, 24
-; SI-NEXT: v_writelane_b32 v62, s4, 7
-; SI-NEXT: s_lshr_b32 s4, s8, 16
-; SI-NEXT: v_add_i32_e32 v9, vcc, 0x30000, v5
-; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v27
-; SI-NEXT: v_or_b32_e32 v11, v12, v11
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v3, v7, v14, 16
-; SI-NEXT: v_writelane_b32 v62, s4, 8
-; SI-NEXT: s_lshr_b32 s4, s8, 8
-; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v15
-; SI-NEXT: v_add_i32_e32 v12, vcc, 0x30000, v11
-; SI-NEXT: v_mov_b32_e32 v11, s15
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v3, v7, v14, 8
-; SI-NEXT: v_writelane_b32 v62, s4, 9
-; SI-NEXT: s_lshr_b32 s4, s7, 24
-; SI-NEXT: v_or_b32_e32 v5, v29, v5
-; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v3, v11, v12, 24
-; SI-NEXT: v_writelane_b32 v62, s4, 10
-; SI-NEXT: s_lshr_b32 s4, s7, 16
-; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v35
-; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v34
-; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v33
-; SI-NEXT: v_add_i32_e32 v13, vcc, 0x30000, v5
-; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v23
-; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v19
-; SI-NEXT: v_or_b32_e32 v10, v25, v10
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v3, v11, v12, 16
-; SI-NEXT: v_writelane_b32 v62, s4, 11
-; SI-NEXT: s_lshr_b32 s4, s7, 8
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; SI-NEXT: v_add_i32_e32 v10, vcc, 0x30000, v10
-; SI-NEXT: v_mov_b32_e32 v15, s14
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v3, v11, v12, 8
-; SI-NEXT: v_writelane_b32 v62, s4, 12
-; SI-NEXT: s_lshr_b32 s4, s6, 24
-; SI-NEXT: v_or_b32_e32 v1, v39, v1
-; SI-NEXT: v_or_b32_e32 v2, v38, v2
-; SI-NEXT: v_or_b32_e32 v4, v37, v4
-; SI-NEXT: v_or_b32_e32 v5, v28, v5
-; SI-NEXT: v_or_b32_e32 v8, v26, v8
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v3, v15, v10, 24
-; SI-NEXT: v_writelane_b32 v62, s4, 13
-; SI-NEXT: s_lshr_b32 s4, s6, 16
-; SI-NEXT: v_add_i32_e32 v1, vcc, 0x30000, v1
-; SI-NEXT: v_mov_b32_e32 v35, s6
-; SI-NEXT: v_add_i32_e32 v2, vcc, 0x30000, v2
-; SI-NEXT: v_mov_b32_e32 v34, s7
-; SI-NEXT: v_add_i32_e32 v4, vcc, 0x30000, v4
-; SI-NEXT: v_mov_b32_e32 v33, s8
-; SI-NEXT: v_mov_b32_e32 v32, s9
-; SI-NEXT: v_mov_b32_e32 v20, s10
-; SI-NEXT: v_mov_b32_e32 v17, s11
-; SI-NEXT: v_add_i32_e32 v5, vcc, 0x30000, v5
-; SI-NEXT: v_mov_b32_e32 v18, s12
-; SI-NEXT: v_add_i32_e32 v8, vcc, 0x30000, v8
-; SI-NEXT: v_mov_b32_e32 v19, s13
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v3, v15, v10, 16
-; SI-NEXT: v_writelane_b32 v62, s4, 14
-; SI-NEXT: s_lshr_b32 s4, s6, 8
-; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_alignbit_b32 v3, v15, v10, 8
-; SI-NEXT: v_alignbit_b32 v45, v19, v8, 24
-; SI-NEXT: v_alignbit_b32 v47, v19, v8, 16
-; SI-NEXT: v_alignbit_b32 v57, v19, v8, 8
-; SI-NEXT: v_alignbit_b32 v41, v18, v5, 24
-; SI-NEXT: v_alignbit_b32 v43, v18, v5, 16
-; SI-NEXT: v_alignbit_b32 v44, v18, v5, 8
-; SI-NEXT: v_alignbit_b32 v21, v17, v13, 24
-; SI-NEXT: v_alignbit_b32 v22, v17, v13, 16
-; SI-NEXT: v_alignbit_b32 v24, v17, v13, 8
-; SI-NEXT: v_alignbit_b32 v17, v20, v9, 24
-; SI-NEXT: v_alignbit_b32 v18, v20, v9, 16
-; SI-NEXT: v_alignbit_b32 v20, v20, v9, 8
-; SI-NEXT: v_alignbit_b32 v59, v32, v6, 24
-; SI-NEXT: v_alignbit_b32 v60, v32, v6, 16
-; SI-NEXT: v_alignbit_b32 v61, v32, v6, 8
-; SI-NEXT: v_alignbit_b32 v46, v33, v4, 24
-; SI-NEXT: v_alignbit_b32 v56, v33, v4, 16
-; SI-NEXT: v_alignbit_b32 v58, v33, v4, 8
-; SI-NEXT: v_alignbit_b32 v55, v34, v2, 24
-; SI-NEXT: v_alignbit_b32 v40, v34, v2, 16
-; SI-NEXT: v_alignbit_b32 v42, v34, v2, 8
-; SI-NEXT: v_alignbit_b32 v52, v35, v1, 24
-; SI-NEXT: v_alignbit_b32 v53, v35, v1, 16
-; SI-NEXT: v_alignbit_b32 v54, v35, v1, 8
-; SI-NEXT: s_lshr_b32 s83, s57, 24
-; SI-NEXT: s_lshr_b32 s71, s57, 16
-; SI-NEXT: s_lshr_b32 s68, s57, 8
-; SI-NEXT: s_lshr_b32 s80, s47, 24
-; SI-NEXT: s_lshr_b32 s69, s47, 16
-; SI-NEXT: s_lshr_b32 s65, s47, 8
-; SI-NEXT: s_lshr_b32 s70, s45, 24
-; SI-NEXT: s_lshr_b32 s66, s45, 16
-; SI-NEXT: s_lshr_b32 s54, s45, 8
-; SI-NEXT: s_lshr_b32 s67, s43, 24
-; SI-NEXT: s_lshr_b32 s55, s43, 16
-; SI-NEXT: s_lshr_b32 s51, s43, 8
-; SI-NEXT: s_lshr_b32 s64, s41, 24
-; SI-NEXT: s_lshr_b32 s52, s41, 16
-; SI-NEXT: s_lshr_b32 s48, s41, 8
-; SI-NEXT: s_lshr_b32 s53, s40, 24
-; SI-NEXT: s_lshr_b32 s49, s40, 16
-; SI-NEXT: s_lshr_b32 s37, s40, 8
-; SI-NEXT: s_lshr_b32 s50, s15, 24
-; SI-NEXT: s_lshr_b32 s38, s15, 16
-; SI-NEXT: s_lshr_b32 s34, s15, 8
-; SI-NEXT: s_lshr_b32 s39, s14, 24
-; SI-NEXT: s_lshr_b32 s35, s14, 16
-; SI-NEXT: s_lshr_b32 s95, s14, 8
-; SI-NEXT: s_lshr_b32 s36, s13, 24
-; SI-NEXT: s_lshr_b32 s30, s13, 16
-; SI-NEXT: s_lshr_b32 s92, s13, 8
-; SI-NEXT: s_lshr_b32 s31, s12, 24
-; SI-NEXT: s_lshr_b32 s93, s12, 16
-; SI-NEXT: s_lshr_b32 s89, s12, 8
-; SI-NEXT: s_lshr_b32 s94, s11, 24
-; SI-NEXT: s_lshr_b32 s90, s11, 16
-; SI-NEXT: s_lshr_b32 s91, s10, 24
-; SI-NEXT: v_writelane_b32 v62, s4, 15
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; SI-NEXT: .LBB99_3: ; %end
-; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v50
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: s_and_b32 s4, s56, 0xff
-; SI-NEXT: s_lshl_b32 s5, s68, 8
-; SI-NEXT: s_lshl_b32 s16, s83, 24
-; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT: v_readlane_b32 s99, v63, 35
-; SI-NEXT: v_readlane_b32 s98, v63, 34
-; SI-NEXT: v_readlane_b32 s97, v63, 33
-; SI-NEXT: v_readlane_b32 s96, v63, 32
-; SI-NEXT: v_readlane_b32 s87, v63, 31
-; SI-NEXT: v_readlane_b32 s86, v63, 30
-; SI-NEXT: v_readlane_b32 s85, v63, 29
-; SI-NEXT: v_readlane_b32 s84, v63, 28
-; SI-NEXT: v_readlane_b32 s83, v63, 27
-; SI-NEXT: v_readlane_b32 s82, v63, 26
-; SI-NEXT: v_readlane_b32 s81, v63, 25
-; SI-NEXT: v_readlane_b32 s68, v63, 20
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v7, 24, v7
-; SI-NEXT: v_or_b32_e32 v3, v7, v3
-; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v7, 8, v7
-; SI-NEXT: v_or_b32_e32 v7, s4, v7
-; SI-NEXT: s_and_b32 s4, s57, 0xff
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: s_and_b32 s5, s71, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 16
-; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s16, s5
-; SI-NEXT: v_or_b32_e32 v3, v7, v3
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: buffer_store_dword v3, v0, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v3, vcc, 4, v0
-; SI-NEXT: v_mov_b32_e32 v7, s4
-; SI-NEXT: buffer_store_dword v7, v3, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; SI-NEXT: s_and_b32 s4, s46, 0xff
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v51
-; SI-NEXT: v_or_b32_e32 v3, s4, v3
-; SI-NEXT: s_and_b32 s4, s47, 0xff
-; SI-NEXT: s_lshl_b32 s5, s65, 8
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: s_and_b32 s5, s69, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 16
-; SI-NEXT: s_lshl_b32 s16, s80, 24
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s16, s5
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: s_lshl_b32 s5, s54, 8
-; SI-NEXT: s_lshl_b32 s16, s70, 24
-; SI-NEXT: v_readlane_b32 s80, v63, 24
-; SI-NEXT: v_readlane_b32 s71, v63, 23
-; SI-NEXT: v_readlane_b32 s70, v63, 22
-; SI-NEXT: v_readlane_b32 s69, v63, 21
-; SI-NEXT: v_readlane_b32 s65, v63, 17
-; SI-NEXT: v_readlane_b32 s54, v63, 14
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v11, 24, v11
-; SI-NEXT: v_or_b32_e32 v7, v11, v7
-; SI-NEXT: v_or_b32_e32 v3, v3, v7
-; SI-NEXT: v_add_i32_e32 v7, vcc, 8, v0
-; SI-NEXT: buffer_store_dword v3, v7, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v3, vcc, 12, v0
-; SI-NEXT: v_mov_b32_e32 v7, s4
-; SI-NEXT: buffer_store_dword v7, v3, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT: s_and_b32 s4, s44, 0xff
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v49
-; SI-NEXT: v_or_b32_e32 v3, s4, v3
-; SI-NEXT: s_and_b32 s4, s45, 0xff
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: s_and_b32 s5, s66, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 16
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s16, s5
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: s_lshl_b32 s5, s51, 8
-; SI-NEXT: s_lshl_b32 s16, s67, 24
-; SI-NEXT: v_readlane_b32 s67, v63, 19
-; SI-NEXT: v_readlane_b32 s66, v63, 18
-; SI-NEXT: v_readlane_b32 s51, v63, 11
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v11, 24, v11
-; SI-NEXT: v_or_b32_e32 v7, v11, v7
-; SI-NEXT: v_or_b32_e32 v3, v3, v7
-; SI-NEXT: v_add_i32_e32 v7, vcc, 16, v0
-; SI-NEXT: buffer_store_dword v3, v7, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v3, vcc, 20, v0
-; SI-NEXT: v_mov_b32_e32 v7, s4
-; SI-NEXT: buffer_store_dword v7, v3, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT: s_and_b32 s4, s42, 0xff
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v48
-; SI-NEXT: v_or_b32_e32 v3, s4, v3
-; SI-NEXT: s_and_b32 s4, s43, 0xff
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: s_and_b32 s5, s55, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 16
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s16, s5
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: s_lshl_b32 s5, s48, 8
-; SI-NEXT: s_lshl_b32 s16, s64, 24
-; SI-NEXT: v_readlane_b32 s64, v63, 16
-; SI-NEXT: v_readlane_b32 s55, v63, 15
-; SI-NEXT: v_readlane_b32 s48, v63, 8
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v11, 24, v11
-; SI-NEXT: v_or_b32_e32 v7, v11, v7
-; SI-NEXT: v_or_b32_e32 v3, v3, v7
-; SI-NEXT: v_add_i32_e32 v7, vcc, 24, v0
-; SI-NEXT: buffer_store_dword v3, v7, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v3, vcc, 28, v0
-; SI-NEXT: v_mov_b32_e32 v7, s4
-; SI-NEXT: buffer_store_dword v7, v3, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v16
-; SI-NEXT: s_and_b32 s4, s41, 0xff
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: s_and_b32 s5, s52, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 16
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s16, s5
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: s_lshl_b32 s5, s37, 8
-; SI-NEXT: s_lshl_b32 s16, s53, 24
-; SI-NEXT: v_readlane_b32 s53, v63, 13
-; SI-NEXT: v_readlane_b32 s52, v63, 12
-; SI-NEXT: v_readlane_b32 s37, v63, 5
-; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v7, 8, v7
-; SI-NEXT: v_or_b32_e32 v3, v3, v7
-; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v11, 24, v11
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; SI-NEXT: v_or_b32_e32 v7, v11, v7
-; SI-NEXT: v_or_b32_e32 v3, v3, v7
-; SI-NEXT: v_add_i32_e32 v7, vcc, 32, v0
-; SI-NEXT: buffer_store_dword v3, v7, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v3, vcc, 36, v0
-; SI-NEXT: v_mov_b32_e32 v7, s4
-; SI-NEXT: buffer_store_dword v7, v3, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v14
-; SI-NEXT: s_and_b32 s4, s40, 0xff
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: s_and_b32 s5, s49, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 16
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s16, s5
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: s_lshl_b32 s5, s34, 8
-; SI-NEXT: v_readlane_b32 s49, v63, 9
-; SI-NEXT: v_readlane_b32 s34, v63, 2
-; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v7, 8, v7
-; SI-NEXT: v_or_b32_e32 v3, v3, v7
-; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v11, 24, v11
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; SI-NEXT: v_or_b32_e32 v7, v11, v7
-; SI-NEXT: v_or_b32_e32 v3, v3, v7
-; SI-NEXT: v_add_i32_e32 v7, vcc, 40, v0
-; SI-NEXT: buffer_store_dword v3, v7, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v3, vcc, 44, v0
-; SI-NEXT: v_mov_b32_e32 v7, s4
-; SI-NEXT: buffer_store_dword v7, v3, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v12
-; SI-NEXT: s_and_b32 s4, s15, 0xff
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: s_and_b32 s5, s38, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 16
-; SI-NEXT: s_lshl_b32 s15, s50, 24
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s15, s5
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: s_lshl_b32 s5, s95, 8
-; SI-NEXT: v_readlane_b32 s50, v63, 10
-; SI-NEXT: v_readlane_b32 s38, v63, 6
-; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v7, 8, v7
-; SI-NEXT: v_or_b32_e32 v3, v3, v7
-; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v11, 24, v11
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; SI-NEXT: v_or_b32_e32 v7, v11, v7
-; SI-NEXT: v_or_b32_e32 v3, v3, v7
-; SI-NEXT: v_add_i32_e32 v7, vcc, 48, v0
-; SI-NEXT: buffer_store_dword v3, v7, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v3, vcc, 52, v0
-; SI-NEXT: v_mov_b32_e32 v7, s4
-; SI-NEXT: buffer_store_dword v7, v3, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v10
-; SI-NEXT: s_and_b32 s4, s14, 0xff
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: s_and_b32 s5, s35, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 16
-; SI-NEXT: s_lshl_b32 s14, s39, 24
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s14, s5
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: s_lshl_b32 s5, s92, 8
-; SI-NEXT: v_readlane_b32 s39, v63, 7
-; SI-NEXT: v_readlane_b32 s35, v63, 3
-; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v7, 8, v7
-; SI-NEXT: v_or_b32_e32 v3, v3, v7
-; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v10, 24, v10
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; SI-NEXT: v_or_b32_e32 v7, v10, v7
-; SI-NEXT: v_or_b32_e32 v3, v3, v7
-; SI-NEXT: v_add_i32_e32 v7, vcc, 56, v0
-; SI-NEXT: buffer_store_dword v3, v7, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v3, vcc, 60, v0
-; SI-NEXT: v_mov_b32_e32 v7, s4
-; SI-NEXT: buffer_store_dword v7, v3, s[0:3], 0 offen
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v8
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v7, 8, v57
-; SI-NEXT: s_and_b32 s4, s13, 0xff
-; SI-NEXT: v_or_b32_e32 v3, v3, v7
-; SI-NEXT: v_and_b32_e32 v7, 0xff, v47
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: s_and_b32 s5, s30, 0xff
-; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; SI-NEXT: v_lshlrev_b32_e32 v8, 24, v45
-; SI-NEXT: s_lshl_b32 s5, s5, 16
-; SI-NEXT: s_lshl_b32 s13, s36, 24
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT: v_or_b32_e32 v7, v8, v7
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s13, s5
-; SI-NEXT: v_or_b32_e32 v3, v3, v7
-; SI-NEXT: v_add_i32_e32 v7, vcc, 64, v0
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: buffer_store_dword v3, v7, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v3, vcc, 0x44, v0
-; SI-NEXT: v_mov_b32_e32 v7, s4
-; SI-NEXT: buffer_store_dword v7, v3, s[0:3], 0 offen
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v5
-; SI-NEXT: v_lshlrev_b32_e32 v5, 8, v44
-; SI-NEXT: s_and_b32 s4, s12, 0xff
-; SI-NEXT: s_lshl_b32 s5, s89, 8
-; SI-NEXT: v_or_b32_e32 v3, v3, v5
-; SI-NEXT: v_and_b32_e32 v5, 0xff, v43
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: s_and_b32 s5, s93, 0xff
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v7, 24, v41
-; SI-NEXT: s_lshl_b32 s5, s5, 16
-; SI-NEXT: s_lshl_b32 s12, s31, 24
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT: v_or_b32_e32 v5, v7, v5
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s12, s5
-; SI-NEXT: v_or_b32_e32 v3, v3, v5
-; SI-NEXT: v_add_i32_e32 v5, vcc, 0x48, v0
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: buffer_store_dword v3, v5, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v3, vcc, 0x4c, v0
-; SI-NEXT: v_mov_b32_e32 v5, s4
-; SI-NEXT: v_readlane_b32 s5, v62, 1
-; SI-NEXT: buffer_store_dword v5, v3, s[0:3], 0 offen
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v13
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v5, 8, v24
-; SI-NEXT: s_and_b32 s4, s11, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 8
-; SI-NEXT: v_or_b32_e32 v3, v3, v5
-; SI-NEXT: v_and_b32_e32 v5, 0xff, v22
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: s_and_b32 s5, s90, 0xff
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: v_lshlrev_b32_e32 v7, 24, v21
-; SI-NEXT: s_lshl_b32 s5, s5, 16
-; SI-NEXT: s_lshl_b32 s11, s94, 24
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT: v_or_b32_e32 v5, v7, v5
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s11, s5
-; SI-NEXT: v_or_b32_e32 v3, v3, v5
-; SI-NEXT: v_add_i32_e32 v5, vcc, 0x50, v0
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: v_readlane_b32 s5, v62, 3
-; SI-NEXT: buffer_store_dword v3, v5, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v3, vcc, 0x54, v0
-; SI-NEXT: v_mov_b32_e32 v5, s4
-; SI-NEXT: s_and_b32 s4, s10, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 8
-; SI-NEXT: buffer_store_dword v5, v3, s[0:3], 0 offen
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v9
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v5, 8, v20
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: v_readlane_b32 s5, v62, 2
-; SI-NEXT: v_or_b32_e32 v3, v3, v5
-; SI-NEXT: v_and_b32_e32 v5, 0xff, v18
-; SI-NEXT: s_and_b32 s5, s5, 0xff
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: v_lshlrev_b32_e32 v7, 24, v17
-; SI-NEXT: s_lshl_b32 s5, s5, 16
-; SI-NEXT: s_lshl_b32 s10, s91, 24
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT: v_or_b32_e32 v5, v7, v5
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s10, s5
-; SI-NEXT: v_or_b32_e32 v3, v3, v5
-; SI-NEXT: v_add_i32_e32 v5, vcc, 0x58, v0
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: v_readlane_b32 s5, v62, 6
-; SI-NEXT: buffer_store_dword v3, v5, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v3, vcc, 0x5c, v0
-; SI-NEXT: v_mov_b32_e32 v5, s4
-; SI-NEXT: s_and_b32 s4, s9, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 8
-; SI-NEXT: buffer_store_dword v5, v3, s[0:3], 0 offen
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v6
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v5, 8, v61
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: v_readlane_b32 s5, v62, 5
-; SI-NEXT: v_or_b32_e32 v3, v3, v5
-; SI-NEXT: v_and_b32_e32 v5, 0xff, v60
-; SI-NEXT: s_and_b32 s5, s5, 0xff
-; SI-NEXT: v_readlane_b32 s9, v62, 4
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: v_lshlrev_b32_e32 v6, 24, v59
-; SI-NEXT: s_lshl_b32 s5, s5, 16
-; SI-NEXT: s_lshl_b32 s9, s9, 24
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT: v_or_b32_e32 v5, v6, v5
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s9, s5
-; SI-NEXT: v_or_b32_e32 v3, v3, v5
-; SI-NEXT: v_add_i32_e32 v5, vcc, 0x60, v0
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: v_readlane_b32 s5, v62, 9
-; SI-NEXT: buffer_store_dword v3, v5, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v3, vcc, 0x64, v0
-; SI-NEXT: v_mov_b32_e32 v5, s4
-; SI-NEXT: s_and_b32 s4, s8, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 8
-; SI-NEXT: buffer_store_dword v5, v3, s[0:3], 0 offen
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v4
-; SI-NEXT: v_lshlrev_b32_e32 v4, 8, v58
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: v_readlane_b32 s5, v62, 8
-; SI-NEXT: v_or_b32_e32 v3, v3, v4
-; SI-NEXT: v_and_b32_e32 v4, 0xff, v56
-; SI-NEXT: s_and_b32 s5, s5, 0xff
-; SI-NEXT: v_readlane_b32 s8, v62, 7
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v5, 24, v46
-; SI-NEXT: s_lshl_b32 s5, s5, 16
-; SI-NEXT: s_lshl_b32 s8, s8, 24
-; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s8, s5
-; SI-NEXT: v_or_b32_e32 v3, v3, v4
-; SI-NEXT: v_add_i32_e32 v4, vcc, 0x68, v0
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: v_readlane_b32 s5, v62, 12
-; SI-NEXT: buffer_store_dword v3, v4, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v3, vcc, 0x6c, v0
-; SI-NEXT: v_mov_b32_e32 v4, s4
-; SI-NEXT: s_and_b32 s4, s7, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 8
-; SI-NEXT: buffer_store_dword v4, v3, s[0:3], 0 offen
-; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v42
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: v_readlane_b32 s5, v62, 11
-; SI-NEXT: v_or_b32_e32 v2, v2, v3
-; SI-NEXT: v_and_b32_e32 v3, 0xff, v40
-; SI-NEXT: s_and_b32 s5, s5, 0xff
-; SI-NEXT: v_readlane_b32 s7, v62, 10
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v55
-; SI-NEXT: s_lshl_b32 s5, s5, 16
-; SI-NEXT: s_lshl_b32 s7, s7, 24
-; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s7, s5
-; SI-NEXT: v_or_b32_e32 v2, v2, v3
-; SI-NEXT: v_add_i32_e32 v3, vcc, 0x70, v0
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: v_readlane_b32 s5, v62, 15
-; SI-NEXT: buffer_store_dword v2, v3, s[0:3], 0 offen
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v2, vcc, 0x74, v0
-; SI-NEXT: v_mov_b32_e32 v3, s4
-; SI-NEXT: s_and_b32 s4, s6, 0xff
-; SI-NEXT: s_lshl_b32 s5, s5, 8
-; SI-NEXT: buffer_store_dword v3, v2, s[0:3], 0 offen
-; SI-NEXT: v_lshlrev_b32_e32 v2, 8, v54
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: v_readlane_b32 s5, v62, 14
-; SI-NEXT: v_or_b32_e32 v1, v1, v2
-; SI-NEXT: v_and_b32_e32 v2, 0xff, v53
-; SI-NEXT: s_and_b32 s5, s5, 0xff
-; SI-NEXT: v_readlane_b32 s6, v62, 13
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshlrev_b32_e32 v3, 24, v52
-; SI-NEXT: s_lshl_b32 s5, s5, 16
-; SI-NEXT: s_lshl_b32 s6, s6, 24
-; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT: v_or_b32_e32 v2, v3, v2
-; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_or_b32 s5, s6, s5
-; SI-NEXT: v_or_b32_e32 v1, v1, v2
-; SI-NEXT: v_add_i32_e32 v2, vcc, 0x78, v0
-; SI-NEXT: s_or_b32 s4, s4, s5
-; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
-; SI-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_mov_b32_e32 v1, s4
-; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT: v_readlane_b32 s36, v63, 4
-; SI-NEXT: v_readlane_b32 s31, v63, 1
-; SI-NEXT: v_readlane_b32 s30, v63, 0
-; SI-NEXT: s_or_saveexec_b64 s[4:5], -1
-; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT: s_mov_b64 exec, s[4:5]
-; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; SI-NEXT: s_setpc_b64 s[30:31]
-; SI-NEXT: .LBB99_4:
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $sgpr6
-; SI-NEXT: ; kill: killed $sgpr6
-; SI-NEXT: ; implicit-def: $vcc_lo
-; SI-NEXT: v_mov_b32_e32 v39, v24
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $sgpr6
-; SI-NEXT: ; kill: killed $sgpr6
-; SI-NEXT: v_mov_b32_e32 v38, v22
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $sgpr6
-; SI-NEXT: ; kill: killed $sgpr6
-; SI-NEXT: v_mov_b32_e32 v37, v21
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $sgpr6
-; SI-NEXT: ; kill: killed $sgpr6
-; SI-NEXT: v_mov_b32_e32 v36, v20
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $sgpr6
-; SI-NEXT: ; kill: killed $sgpr6
-; SI-NEXT: v_mov_b32_e32 v30, v18
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $sgpr6
-; SI-NEXT: ; kill: killed $sgpr6
-; SI-NEXT: v_mov_b32_e32 v29, v17
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $sgpr6
-; SI-NEXT: ; kill: killed $sgpr6
-; SI-NEXT: v_mov_b32_e32 v28, v13
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $sgpr6
-; SI-NEXT: ; kill: killed $sgpr6
-; SI-NEXT: v_mov_b32_e32 v26, v9
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $sgpr6
-; SI-NEXT: ; kill: killed $sgpr6
-; SI-NEXT: v_mov_b32_e32 v25, v6
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $sgpr6
-; SI-NEXT: ; kill: killed $sgpr6
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $sgpr6
-; SI-NEXT: ; kill: killed $sgpr6
-; SI-NEXT: ; kill: killed $vcc_lo
-; SI-NEXT: ; implicit-def: $vcc_lo
-; SI-NEXT: ; implicit-def: $sgpr56
-; SI-NEXT: ; implicit-def: $vgpr50
-; SI-NEXT: ; implicit-def: $sgpr57
-; SI-NEXT: ; implicit-def: $sgpr68
-; SI-NEXT: ; implicit-def: $sgpr71
-; SI-NEXT: ; implicit-def: $sgpr83
-; SI-NEXT: ; implicit-def: $sgpr46
-; SI-NEXT: ; implicit-def: $vgpr51
-; SI-NEXT: ; implicit-def: $sgpr47
-; SI-NEXT: ; implicit-def: $sgpr65
-; SI-NEXT: ; implicit-def: $sgpr69
-; SI-NEXT: ; implicit-def: $sgpr80
-; SI-NEXT: ; implicit-def: $sgpr44
-; SI-NEXT: ; implicit-def: $vgpr49
-; SI-NEXT: ; implicit-def: $sgpr45
-; SI-NEXT: ; implicit-def: $sgpr54
-; SI-NEXT: ; implicit-def: $sgpr66
-; SI-NEXT: ; implicit-def: $sgpr70
-; SI-NEXT: ; implicit-def: $sgpr42
-; SI-NEXT: ; implicit-def: $vgpr48
-; SI-NEXT: ; implicit-def: $sgpr43
-; SI-NEXT: ; implicit-def: $sgpr51
-; SI-NEXT: ; implicit-def: $sgpr55
-; SI-NEXT: ; implicit-def: $sgpr67
-; SI-NEXT: ; implicit-def: $vgpr16
-; SI-NEXT: ; implicit-def: $sgpr41
-; SI-NEXT: ; implicit-def: $sgpr48
-; SI-NEXT: ; implicit-def: $sgpr52
-; SI-NEXT: ; implicit-def: $sgpr64
-; SI-NEXT: ; implicit-def: $vgpr14
-; SI-NEXT: ; implicit-def: $sgpr40
-; SI-NEXT: ; implicit-def: $sgpr37
-; SI-NEXT: ; implicit-def: $sgpr49
-; SI-NEXT: ; implicit-def: $sgpr53
-; SI-NEXT: ; implicit-def: $vgpr12
-; SI-NEXT: ; implicit-def: $sgpr15
-; SI-NEXT: ; implicit-def: $sgpr34
-; SI-NEXT: ; implicit-def: $sgpr38
-; SI-NEXT: ; implicit-def: $sgpr50
-; SI-NEXT: ; implicit-def: $vgpr10
-; SI-NEXT: ; implicit-def: $sgpr14
-; SI-NEXT: ; implicit-def: $sgpr95
-; SI-NEXT: ; implicit-def: $sgpr35
-; SI-NEXT: ; implicit-def: $sgpr39
-; SI-NEXT: ; implicit-def: $vgpr8
-; SI-NEXT: ; implicit-def: $vgpr57
-; SI-NEXT: ; implicit-def: $vgpr47
-; SI-NEXT: ; implicit-def: $vgpr45
-; SI-NEXT: ; implicit-def: $sgpr13
-; SI-NEXT: ; implicit-def: $sgpr92
-; SI-NEXT: ; implicit-def: $sgpr30
-; SI-NEXT: ; implicit-def: $sgpr36
-; SI-NEXT: ; implicit-def: $vgpr5
-; SI-NEXT: ; implicit-def: $vgpr44
-; SI-NEXT: ; implicit-def: $vgpr43
-; SI-NEXT: ; implicit-def: $vgpr41
-; SI-NEXT: ; implicit-def: $sgpr12
-; SI-NEXT: ; implicit-def: $sgpr89
-; SI-NEXT: ; implicit-def: $sgpr93
-; SI-NEXT: ; implicit-def: $sgpr31
-; SI-NEXT: ; implicit-def: $sgpr11
-; SI-NEXT: ; implicit-def: $sgpr90
-; SI-NEXT: ; implicit-def: $sgpr94
-; SI-NEXT: ; implicit-def: $sgpr10
-; SI-NEXT: ; implicit-def: $sgpr91
-; SI-NEXT: ; implicit-def: $sgpr9
-; SI-NEXT: ; implicit-def: $sgpr8
-; SI-NEXT: ; implicit-def: $sgpr7
-; SI-NEXT: ; kill: killed $vcc_lo
-; SI-NEXT: ; implicit-def: $vcc_lo
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $sgpr6
-; SI-NEXT: ; kill: killed $sgpr6
-; SI-NEXT: ; kill: killed $vcc_lo
-; SI-NEXT: ; implicit-def: $vgpr13
-; SI-NEXT: ; implicit-def: $vgpr24
-; SI-NEXT: ; implicit-def: $vgpr22
-; SI-NEXT: ; implicit-def: $vgpr21
-; SI-NEXT: ; implicit-def: $vgpr9
-; SI-NEXT: ; implicit-def: $vgpr20
-; SI-NEXT: ; implicit-def: $vgpr18
-; SI-NEXT: ; implicit-def: $vgpr17
-; SI-NEXT: ; implicit-def: $vgpr6
-; SI-NEXT: ; implicit-def: $vgpr61
-; SI-NEXT: ; implicit-def: $vgpr60
-; SI-NEXT: ; implicit-def: $vgpr59
-; SI-NEXT: ; implicit-def: $vgpr4
-; SI-NEXT: ; implicit-def: $vgpr58
-; SI-NEXT: ; implicit-def: $vgpr56
-; SI-NEXT: ; implicit-def: $vgpr46
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr42
-; SI-NEXT: ; implicit-def: $vgpr40
-; SI-NEXT: ; implicit-def: $vgpr55
-; SI-NEXT: ; implicit-def: $vgpr54
-; SI-NEXT: ; implicit-def: $vgpr53
-; SI-NEXT: ; implicit-def: $vgpr52
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $sgpr6
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: s_branch .LBB99_2
-;
-; VI-LABEL: bitcast_v64i16_to_v128i8_scalar:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 ; 4-byte Folded Spill
-; VI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; VI-NEXT: s_mov_b64 exec, s[4:5]
-; VI-NEXT: v_writelane_b32 v20, s30, 0
-; VI-NEXT: v_writelane_b32 v20, s31, 1
-; VI-NEXT: v_writelane_b32 v20, s34, 2
-; VI-NEXT: v_writelane_b32 v20, s35, 3
-; VI-NEXT: v_writelane_b32 v20, s36, 4
-; VI-NEXT: v_writelane_b32 v20, s37, 5
-; VI-NEXT: v_writelane_b32 v20, s38, 6
-; VI-NEXT: v_writelane_b32 v20, s39, 7
-; VI-NEXT: v_writelane_b32 v20, s48, 8
-; VI-NEXT: v_writelane_b32 v20, s49, 9
-; VI-NEXT: v_writelane_b32 v20, s50, 10
-; VI-NEXT: v_writelane_b32 v20, s51, 11
-; VI-NEXT: v_writelane_b32 v20, s52, 12
-; VI-NEXT: v_writelane_b32 v20, s53, 13
-; VI-NEXT: v_writelane_b32 v20, s54, 14
-; VI-NEXT: v_writelane_b32 v20, s55, 15
-; VI-NEXT: v_writelane_b32 v20, s64, 16
-; VI-NEXT: v_writelane_b32 v20, s65, 17
-; VI-NEXT: v_writelane_b32 v20, s66, 18
-; VI-NEXT: v_writelane_b32 v20, s67, 19
-; VI-NEXT: v_writelane_b32 v20, s68, 20
-; VI-NEXT: v_writelane_b32 v20, s69, 21
-; VI-NEXT: v_writelane_b32 v20, s70, 22
-; VI-NEXT: v_writelane_b32 v20, s71, 23
-; VI-NEXT: v_writelane_b32 v20, s80, 24
-; VI-NEXT: v_writelane_b32 v20, s81, 25
-; VI-NEXT: v_writelane_b32 v20, s82, 26
-; VI-NEXT: v_writelane_b32 v20, s83, 27
-; VI-NEXT: v_writelane_b32 v20, s84, 28
-; VI-NEXT: v_writelane_b32 v20, s85, 29
-; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v19
-; VI-NEXT: v_writelane_b32 v20, s86, 30
-; VI-NEXT: v_readfirstlane_b32 s42, v3
-; VI-NEXT: v_readfirstlane_b32 s43, v4
-; VI-NEXT: v_readfirstlane_b32 s40, v5
-; VI-NEXT: v_readfirstlane_b32 s41, v6
-; VI-NEXT: v_readfirstlane_b32 s14, v7
-; VI-NEXT: v_readfirstlane_b32 s15, v8
-; VI-NEXT: v_readfirstlane_b32 s12, v9
-; VI-NEXT: v_readfirstlane_b32 s13, v10
-; VI-NEXT: v_readfirstlane_b32 s10, v11
-; VI-NEXT: v_readfirstlane_b32 s11, v12
-; VI-NEXT: v_readfirstlane_b32 s8, v13
-; VI-NEXT: v_readfirstlane_b32 s9, v14
-; VI-NEXT: v_readfirstlane_b32 s6, v15
-; VI-NEXT: v_readfirstlane_b32 s7, v16
-; VI-NEXT: v_readfirstlane_b32 s4, v17
-; VI-NEXT: v_readfirstlane_b32 s5, v18
-; VI-NEXT: v_readfirstlane_b32 s44, v1
-; VI-NEXT: s_and_b64 s[46:47], vcc, exec
-; VI-NEXT: v_readfirstlane_b32 s45, v2
-; VI-NEXT: v_writelane_b32 v20, s87, 31
-; VI-NEXT: ; implicit-def: $vgpr21 : SGPR spill to VGPR lane
-; VI-NEXT: s_cbranch_scc0 .LBB99_4
-; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: s_lshr_b32 s46, s45, 24
-; VI-NEXT: v_writelane_b32 v21, s46, 0
-; VI-NEXT: s_lshr_b32 s46, s45, 16
-; VI-NEXT: v_writelane_b32 v21, s46, 1
-; VI-NEXT: s_lshr_b32 s46, s45, 8
-; VI-NEXT: v_writelane_b32 v21, s46, 2
-; VI-NEXT: s_lshr_b32 s46, s44, 16
-; VI-NEXT: v_writelane_b32 v21, s46, 3
-; VI-NEXT: s_lshr_b32 s46, s44, 8
-; VI-NEXT: v_writelane_b32 v21, s46, 4
-; VI-NEXT: s_lshr_b32 s46, s29, 24
-; VI-NEXT: v_writelane_b32 v21, s46, 5
-; VI-NEXT: s_lshr_b32 s46, s29, 16
-; VI-NEXT: v_writelane_b32 v21, s46, 6
-; VI-NEXT: s_lshr_b32 s46, s29, 8
-; VI-NEXT: v_writelane_b32 v21, s46, 7
-; VI-NEXT: s_lshr_b32 s46, s28, 16
-; VI-NEXT: v_writelane_b32 v21, s46, 8
-; VI-NEXT: s_lshr_b32 s46, s28, 8
-; VI-NEXT: v_writelane_b32 v21, s46, 9
-; VI-NEXT: s_lshr_b32 s46, s27, 24
-; VI-NEXT: v_writelane_b32 v21, s46, 10
-; VI-NEXT: s_lshr_b32 s46, s27, 16
-; VI-NEXT: v_writelane_b32 v21, s46, 11
-; VI-NEXT: s_lshr_b32 s46, s27, 8
-; VI-NEXT: v_writelane_b32 v21, s46, 12
-; VI-NEXT: s_lshr_b32 s46, s26, 16
-; VI-NEXT: v_writelane_b32 v21, s46, 13
-; VI-NEXT: s_lshr_b32 s46, s26, 8
-; VI-NEXT: v_writelane_b32 v21, s46, 14
-; VI-NEXT: s_lshr_b32 s46, s25, 24
-; VI-NEXT: v_writelane_b32 v21, s46, 15
-; VI-NEXT: s_lshr_b32 s46, s25, 16
-; VI-NEXT: v_writelane_b32 v21, s46, 16
-; VI-NEXT: s_lshr_b32 s46, s25, 8
-; VI-NEXT: v_writelane_b32 v21, s46, 17
-; VI-NEXT: s_lshr_b32 s46, s24, 16
-; VI-NEXT: v_writelane_b32 v21, s46, 18
-; VI-NEXT: s_lshr_b32 s46, s24, 8
-; VI-NEXT: v_writelane_b32 v21, s46, 19
-; VI-NEXT: s_lshr_b32 s46, s23, 24
-; VI-NEXT: v_writelane_b32 v21, s46, 20
-; VI-NEXT: s_lshr_b32 s46, s23, 16
-; VI-NEXT: v_writelane_b32 v21, s46, 21
-; VI-NEXT: s_lshr_b32 s46, s23, 8
-; VI-NEXT: v_writelane_b32 v21, s46, 22
-; VI-NEXT: s_lshr_b32 s46, s22, 16
-; VI-NEXT: v_writelane_b32 v21, s46, 23
-; VI-NEXT: s_lshr_b32 s46, s22, 8
-; VI-NEXT: v_writelane_b32 v21, s46, 24
-; VI-NEXT: s_lshr_b32 s46, s21, 24
-; VI-NEXT: v_writelane_b32 v21, s46, 25
-; VI-NEXT: s_lshr_b32 s46, s5, 24
-; VI-NEXT: v_writelane_b32 v21, s46, 26
-; VI-NEXT: s_lshr_b32 s46, s5, 16
-; VI-NEXT: v_writelane_b32 v21, s46, 27
-; VI-NEXT: s_lshr_b32 s46, s5, 8
-; VI-NEXT: v_writelane_b32 v21, s46, 28
-; VI-NEXT: s_lshr_b32 s46, s4, 16
-; VI-NEXT: v_writelane_b32 v21, s46, 29
-; VI-NEXT: s_lshr_b32 s46, s4, 8
-; VI-NEXT: v_writelane_b32 v21, s46, 30
-; VI-NEXT: s_lshr_b32 s46, s7, 24
-; VI-NEXT: v_writelane_b32 v21, s46, 31
-; VI-NEXT: s_lshr_b32 s46, s7, 16
-; VI-NEXT: v_writelane_b32 v21, s46, 32
-; VI-NEXT: s_lshr_b32 s46, s7, 8
-; VI-NEXT: v_writelane_b32 v21, s46, 33
-; VI-NEXT: s_lshr_b32 s46, s6, 16
-; VI-NEXT: v_writelane_b32 v21, s46, 34
-; VI-NEXT: s_lshr_b32 s46, s6, 8
-; VI-NEXT: v_writelane_b32 v21, s46, 35
-; VI-NEXT: s_lshr_b32 s46, s9, 24
-; VI-NEXT: v_writelane_b32 v21, s46, 36
-; VI-NEXT: s_lshr_b32 s46, s9, 16
-; VI-NEXT: v_writelane_b32 v21, s46, 37
-; VI-NEXT: s_lshr_b32 s46, s9, 8
-; VI-NEXT: v_writelane_b32 v21, s46, 38
-; VI-NEXT: s_lshr_b32 s46, s8, 16
-; VI-NEXT: v_writelane_b32 v21, s46, 39
-; VI-NEXT: s_lshr_b32 s46, s8, 8
-; VI-NEXT: v_writelane_b32 v21, s46, 40
-; VI-NEXT: s_lshr_b32 s46, s11, 24
-; VI-NEXT: v_writelane_b32 v21, s46, 41
-; VI-NEXT: s_lshr_b32 s46, s11, 16
-; VI-NEXT: v_writelane_b32 v21, s46, 42
-; VI-NEXT: s_lshr_b32 s46, s11, 8
-; VI-NEXT: v_writelane_b32 v21, s46, 43
-; VI-NEXT: s_lshr_b32 s46, s10, 16
-; VI-NEXT: v_writelane_b32 v21, s46, 44
-; VI-NEXT: s_lshr_b32 s46, s10, 8
-; VI-NEXT: v_writelane_b32 v21, s46, 45
-; VI-NEXT: s_lshr_b32 s46, s13, 24
-; VI-NEXT: v_writelane_b32 v21, s46, 46
-; VI-NEXT: s_lshr_b32 s46, s13, 16
-; VI-NEXT: v_writelane_b32 v21, s46, 47
-; VI-NEXT: s_lshr_b32 s46, s13, 8
-; VI-NEXT: v_writelane_b32 v21, s46, 48
-; VI-NEXT: s_lshr_b32 s46, s12, 16
-; VI-NEXT: v_writelane_b32 v21, s46, 49
-; VI-NEXT: s_lshr_b32 s46, s12, 8
-; VI-NEXT: v_writelane_b32 v21, s46, 50
-; VI-NEXT: s_lshr_b32 s46, s15, 24
-; VI-NEXT: v_writelane_b32 v21, s46, 51
-; VI-NEXT: s_lshr_b32 s46, s15, 16
-; VI-NEXT: v_writelane_b32 v21, s46, 52
-; VI-NEXT: s_lshr_b32 s46, s15, 8
-; VI-NEXT: v_writelane_b32 v21, s46, 53
-; VI-NEXT: s_lshr_b32 s46, s14, 16
-; VI-NEXT: v_writelane_b32 v21, s46, 54
-; VI-NEXT: s_lshr_b32 s46, s14, 8
-; VI-NEXT: v_writelane_b32 v21, s46, 55
-; VI-NEXT: s_lshr_b32 s46, s41, 24
-; VI-NEXT: v_writelane_b32 v21, s46, 56
-; VI-NEXT: s_lshr_b32 s46, s41, 16
-; VI-NEXT: s_lshr_b32 s80, s21, 16
-; VI-NEXT: s_lshr_b32 s82, s21, 8
-; VI-NEXT: s_lshr_b32 s84, s20, 16
-; VI-NEXT: s_lshr_b32 s86, s20, 8
-; VI-NEXT: s_lshr_b32 s51, s19, 24
-; VI-NEXT: s_lshr_b32 s53, s19, 16
-; VI-NEXT: s_lshr_b32 s54, s19, 8
-; VI-NEXT: s_lshr_b32 s65, s18, 16
-; VI-NEXT: s_lshr_b32 s66, s18, 8
-; VI-NEXT: s_lshr_b32 s67, s17, 24
-; VI-NEXT: s_lshr_b32 s68, s17, 16
-; VI-NEXT: s_lshr_b32 s69, s17, 8
-; VI-NEXT: s_lshr_b32 s70, s16, 16
-; VI-NEXT: s_lshr_b32 s71, s16, 8
-; VI-NEXT: v_writelane_b32 v21, s46, 57
-; VI-NEXT: s_lshr_b32 s81, s41, 8
-; VI-NEXT: s_lshr_b32 s83, s40, 16
-; VI-NEXT: s_lshr_b32 s85, s40, 8
-; VI-NEXT: s_lshr_b32 s87, s43, 24
-; VI-NEXT: s_lshr_b32 s50, s43, 16
-; VI-NEXT: s_lshr_b32 s52, s43, 8
-; VI-NEXT: s_lshr_b32 s55, s42, 16
-; VI-NEXT: s_lshr_b32 s64, s42, 8
-; VI-NEXT: s_lshr_b64 s[76:77], s[44:45], 24
-; VI-NEXT: s_lshr_b64 s[88:89], s[28:29], 24
-; VI-NEXT: s_lshr_b64 s[90:91], s[26:27], 24
-; VI-NEXT: s_lshr_b64 s[30:31], s[24:25], 24
-; VI-NEXT: s_lshr_b64 s[34:35], s[22:23], 24
-; VI-NEXT: s_lshr_b64 s[36:37], s[20:21], 24
-; VI-NEXT: s_lshr_b64 s[38:39], s[18:19], 24
-; VI-NEXT: s_lshr_b64 s[48:49], s[16:17], 24
-; VI-NEXT: s_lshr_b64 s[46:47], s[4:5], 24
-; VI-NEXT: s_lshr_b64 s[56:57], s[6:7], 24
-; VI-NEXT: s_lshr_b64 s[58:59], s[8:9], 24
-; VI-NEXT: s_lshr_b64 s[60:61], s[10:11], 24
-; VI-NEXT: s_lshr_b64 s[62:63], s[12:13], 24
-; VI-NEXT: s_lshr_b64 s[72:73], s[14:15], 24
-; VI-NEXT: s_lshr_b64 s[74:75], s[40:41], 24
-; VI-NEXT: s_lshr_b64 s[78:79], s[42:43], 24
-; VI-NEXT: s_cbranch_execnz .LBB99_3
-; VI-NEXT: .LBB99_2: ; %cmp.true
-; VI-NEXT: s_and_b32 s47, s45, 0xffff0000
-; VI-NEXT: s_add_i32 s45, s45, 3
-; VI-NEXT: s_and_b32 s45, s45, 0xffff
-; VI-NEXT: s_and_b32 s46, s44, 0xffff0000
-; VI-NEXT: s_add_i32 s44, s44, 3
-; VI-NEXT: s_or_b32 s45, s47, s45
-; VI-NEXT: s_add_i32 s45, s45, 0x30000
-; VI-NEXT: s_and_b32 s44, s44, 0xffff
-; VI-NEXT: s_or_b32 s44, s46, s44
-; VI-NEXT: s_lshr_b32 s46, s45, 24
-; VI-NEXT: s_and_b32 s57, s29, 0xffff0000
-; VI-NEXT: s_add_i32 s29, s29, 3
-; VI-NEXT: v_writelane_b32 v21, s46, 0
-; VI-NEXT: s_lshr_b32 s46, s45, 16
-; VI-NEXT: s_and_b32 s29, s29, 0xffff
-; VI-NEXT: s_add_i32 s44, s44, 0x30000
-; VI-NEXT: v_writelane_b32 v21, s46, 1
-; VI-NEXT: s_lshr_b32 s46, s45, 8
-; VI-NEXT: s_or_b32 s29, s57, s29
-; VI-NEXT: v_writelane_b32 v21, s46, 2
-; VI-NEXT: s_lshr_b32 s46, s44, 16
-; VI-NEXT: s_and_b32 s56, s28, 0xffff0000
-; VI-NEXT: s_add_i32 s28, s28, 3
-; VI-NEXT: s_add_i32 s29, s29, 0x30000
-; VI-NEXT: v_writelane_b32 v21, s46, 3
-; VI-NEXT: s_lshr_b32 s46, s44, 8
-; VI-NEXT: s_and_b32 s28, s28, 0xffff
-; VI-NEXT: v_writelane_b32 v21, s46, 4
-; VI-NEXT: s_lshr_b32 s46, s29, 24
-; VI-NEXT: s_and_b32 s59, s27, 0xffff0000
-; VI-NEXT: s_add_i32 s27, s27, 3
-; VI-NEXT: s_or_b32 s28, s56, s28
-; VI-NEXT: v_writelane_b32 v21, s46, 5
-; VI-NEXT: s_lshr_b32 s46, s29, 16
-; VI-NEXT: s_and_b32 s27, s27, 0xffff
-; VI-NEXT: s_add_i32 s28, s28, 0x30000
-; VI-NEXT: v_writelane_b32 v21, s46, 6
-; VI-NEXT: s_lshr_b32 s46, s29, 8
-; VI-NEXT: s_or_b32 s27, s59, s27
-; VI-NEXT: v_writelane_b32 v21, s46, 7
-; VI-NEXT: s_lshr_b32 s46, s28, 16
-; VI-NEXT: s_and_b32 s58, s26, 0xffff0000
-; VI-NEXT: s_add_i32 s26, s26, 3
-; VI-NEXT: s_add_i32 s27, s27, 0x30000
-; VI-NEXT: v_writelane_b32 v21, s46, 8
-; VI-NEXT: s_lshr_b32 s46, s28, 8
-; VI-NEXT: s_and_b32 s26, s26, 0xffff
-; VI-NEXT: v_writelane_b32 v21, s46, 9
-; VI-NEXT: s_lshr_b32 s46, s27, 24
-; VI-NEXT: s_and_b32 s61, s25, 0xffff0000
-; VI-NEXT: s_add_i32 s25, s25, 3
-; VI-NEXT: s_or_b32 s26, s58, s26
-; VI-NEXT: v_writelane_b32 v21, s46, 10
-; VI-NEXT: s_lshr_b32 s46, s27, 16
-; VI-NEXT: s_and_b32 s25, s25, 0xffff
-; VI-NEXT: s_add_i32 s26, s26, 0x30000
-; VI-NEXT: v_writelane_b32 v21, s46, 11
-; VI-NEXT: s_lshr_b32 s46, s27, 8
-; VI-NEXT: s_or_b32 s25, s61, s25
-; VI-NEXT: v_writelane_b32 v21, s46, 12
-; VI-NEXT: s_lshr_b32 s46, s26, 16
-; VI-NEXT: s_and_b32 s60, s24, 0xffff0000
-; VI-NEXT: s_add_i32 s24, s24, 3
-; VI-NEXT: s_add_i32 s25, s25, 0x30000
-; VI-NEXT: v_writelane_b32 v21, s46, 13
-; VI-NEXT: s_lshr_b32 s46, s26, 8
-; VI-NEXT: s_and_b32 s24, s24, 0xffff
-; VI-NEXT: v_writelane_b32 v21, s46, 14
-; VI-NEXT: s_lshr_b32 s46, s25, 24
-; VI-NEXT: s_and_b32 s63, s23, 0xffff0000
-; VI-NEXT: s_add_i32 s23, s23, 3
-; VI-NEXT: s_or_b32 s24, s60, s24
-; VI-NEXT: v_writelane_b32 v21, s46, 15
-; VI-NEXT: s_lshr_b32 s46, s25, 16
-; VI-NEXT: s_and_b32 s23, s23, 0xffff
-; VI-NEXT: s_add_i32 s24, s24, 0x30000
-; VI-NEXT: v_writelane_b32 v21, s46, 16
-; VI-NEXT: s_lshr_b32 s46, s25, 8
-; VI-NEXT: s_or_b32 s23, s63, s23
-; VI-NEXT: v_writelane_b32 v21, s46, 17
-; VI-NEXT: s_lshr_b32 s46, s24, 16
-; VI-NEXT: s_and_b32 s62, s22, 0xffff0000
-; VI-NEXT: s_add_i32 s22, s22, 3
-; VI-NEXT: s_add_i32 s23, s23, 0x30000
-; VI-NEXT: v_writelane_b32 v21, s46, 18
-; VI-NEXT: s_lshr_b32 s46, s24, 8
-; VI-NEXT: s_and_b32 s22, s22, 0xffff
-; VI-NEXT: v_writelane_b32 v21, s46, 19
-; VI-NEXT: s_lshr_b32 s46, s23, 24
-; VI-NEXT: s_and_b32 s73, s21, 0xffff0000
-; VI-NEXT: s_add_i32 s21, s21, 3
-; VI-NEXT: s_or_b32 s22, s62, s22
-; VI-NEXT: v_writelane_b32 v21, s46, 20
-; VI-NEXT: s_lshr_b32 s46, s23, 16
-; VI-NEXT: s_and_b32 s79, s5, 0xffff0000
-; VI-NEXT: s_add_i32 s5, s5, 3
-; VI-NEXT: s_and_b32 s21, s21, 0xffff
-; VI-NEXT: s_add_i32 s22, s22, 0x30000
-; VI-NEXT: v_writelane_b32 v21, s46, 21
-; VI-NEXT: s_lshr_b32 s46, s23, 8
-; VI-NEXT: s_and_b32 s5, s5, 0xffff
-; VI-NEXT: s_or_b32 s21, s73, s21
-; VI-NEXT: v_writelane_b32 v21, s46, 22
-; VI-NEXT: s_lshr_b32 s46, s22, 16
-; VI-NEXT: s_or_b32 s5, s79, s5
-; VI-NEXT: s_add_i32 s21, s21, 0x30000
-; VI-NEXT: v_writelane_b32 v21, s46, 23
-; VI-NEXT: s_lshr_b32 s46, s22, 8
-; VI-NEXT: s_and_b32 s78, s4, 0xffff0000
-; VI-NEXT: s_add_i32 s4, s4, 3
-; VI-NEXT: s_add_i32 s5, s5, 0x30000
-; VI-NEXT: v_writelane_b32 v21, s46, 24
-; VI-NEXT: s_lshr_b32 s46, s21, 24
-; VI-NEXT: s_and_b32 s4, s4, 0xffff
-; VI-NEXT: v_writelane_b32 v21, s46, 25
-; VI-NEXT: s_lshr_b32 s46, s5, 24
-; VI-NEXT: s_and_b32 s89, s7, 0xffff0000
-; VI-NEXT: s_add_i32 s7, s7, 3
-; VI-NEXT: s_or_b32 s4, s78, s4
-; VI-NEXT: v_writelane_b32 v21, s46, 26
-; VI-NEXT: s_lshr_b32 s46, s5, 16
-; VI-NEXT: s_and_b32 s7, s7, 0xffff
-; VI-NEXT: s_add_i32 s4, s4, 0x30000
-; VI-NEXT: v_writelane_b32 v21, s46, 27
-; VI-NEXT: s_lshr_b32 s46, s5, 8
-; VI-NEXT: s_or_b32 s7, s89, s7
-; VI-NEXT: v_writelane_b32 v21, s46, 28
-; VI-NEXT: s_lshr_b32 s46, s4, 16
-; VI-NEXT: s_and_b32 s88, s6, 0xffff0000
-; VI-NEXT: s_add_i32 s6, s6, 3
-; VI-NEXT: s_add_i32 s7, s7, 0x30000
-; VI-NEXT: v_writelane_b32 v21, s46, 29
-; VI-NEXT: s_lshr_b32 s46, s4, 8
-; VI-NEXT: s_and_b32 s6, s6, 0xffff
-; VI-NEXT: v_writelane_b32 v21, s46, 30
-; VI-NEXT: s_lshr_b32 s46, s7, 24
-; VI-NEXT: s_and_b32 s91, s9, 0xffff0000
-; VI-NEXT: s_add_i32 s9, s9, 3
-; VI-NEXT: s_or_b32 s6, s88, s6
-; VI-NEXT: v_writelane_b32 v21, s46, 31
-; VI-NEXT: s_lshr_b32 s46, s7, 16
-; VI-NEXT: s_and_b32 s9, s9, 0xffff
-; VI-NEXT: s_add_i32 s6, s6, 0x30000
-; VI-NEXT: v_writelane_b32 v21, s46, 32
-; VI-NEXT: s_lshr_b32 s46, s7, 8
-; VI-NEXT: s_or_b32 s9, s91, s9
-; VI-NEXT: v_writelane_b32 v21, s46, 33
-; VI-NEXT: s_lshr_b32 s46, s6, 16
-; VI-NEXT: s_and_b32 s90, s8, 0xffff0000
-; VI-NEXT: s_add_i32 s8, s8, 3
-; VI-NEXT: s_add_i32 s9, s9, 0x30000
-; VI-NEXT: v_writelane_b32 v21, s46, 34
-; VI-NEXT: s_lshr_b32 s46, s6, 8
-; VI-NEXT: s_and_b32 s8, s8, 0xffff
-; VI-NEXT: v_writelane_b32 v21, s46, 35
-; VI-NEXT: s_lshr_b32 s46, s9, 24
-; VI-NEXT: s_and_b32 vcc_hi, s11, 0xffff0000
-; VI-NEXT: s_add_i32 s11, s11, 3
-; VI-NEXT: s_or_b32 s8, s90, s8
-; VI-NEXT: v_writelane_b32 v21, s46, 36
-; VI-NEXT: s_lshr_b32 s46, s9, 16
-; VI-NEXT: s_and_b32 s11, s11, 0xffff
-; VI-NEXT: s_add_i32 s8, s8, 0x30000
-; VI-NEXT: v_writelane_b32 v21, s46, 37
-; VI-NEXT: s_lshr_b32 s46, s9, 8
-; VI-NEXT: s_or_b32 s11, vcc_hi, s11
-; VI-NEXT: v_writelane_b32 v21, s46, 38
-; VI-NEXT: s_lshr_b32 s46, s8, 16
-; VI-NEXT: s_and_b32 vcc_lo, s10, 0xffff0000
-; VI-NEXT: s_add_i32 s10, s10, 3
-; VI-NEXT: s_add_i32 s11, s11, 0x30000
-; VI-NEXT: v_writelane_b32 v21, s46, 39
-; VI-NEXT: s_lshr_b32 s46, s8, 8
-; VI-NEXT: s_and_b32 s10, s10, 0xffff
-; VI-NEXT: v_writelane_b32 v21, s46, 40
-; VI-NEXT: s_lshr_b32 s46, s11, 24
-; VI-NEXT: s_and_b32 s31, s13, 0xffff0000
-; VI-NEXT: s_add_i32 s13, s13, 3
-; VI-NEXT: s_or_b32 s10, vcc_lo, s10
-; VI-NEXT: v_writelane_b32 v21, s46, 41
-; VI-NEXT: s_lshr_b32 s46, s11, 16
-; VI-NEXT: s_and_b32 s13, s13, 0xffff
-; VI-NEXT: s_add_i32 s10, s10, 0x30000
-; VI-NEXT: v_writelane_b32 v21, s46, 42
-; VI-NEXT: s_lshr_b32 s46, s11, 8
-; VI-NEXT: s_or_b32 s13, s31, s13
-; VI-NEXT: v_writelane_b32 v21, s46, 43
-; VI-NEXT: s_lshr_b32 s46, s10, 16
-; VI-NEXT: s_and_b32 s30, s12, 0xffff0000
-; VI-NEXT: s_add_i32 s12, s12, 3
-; VI-NEXT: s_add_i32 s13, s13, 0x30000
-; VI-NEXT: v_writelane_b32 v21, s46, 44
-; VI-NEXT: s_lshr_b32 s46, s10, 8
-; VI-NEXT: s_and_b32 s12, s12, 0xffff
-; VI-NEXT: v_writelane_b32 v21, s46, 45
-; VI-NEXT: s_lshr_b32 s46, s13, 24
-; VI-NEXT: s_and_b32 s35, s15, 0xffff0000
-; VI-NEXT: s_add_i32 s15, s15, 3
-; VI-NEXT: s_or_b32 s12, s30, s12
-; VI-NEXT: v_writelane_b32 v21, s46, 46
-; VI-NEXT: s_lshr_b32 s46, s13, 16
-; VI-NEXT: s_and_b32 s15, s15, 0xffff
-; VI-NEXT: s_add_i32 s12, s12, 0x30000
-; VI-NEXT: v_writelane_b32 v21, s46, 47
-; VI-NEXT: s_lshr_b32 s46, s13, 8
-; VI-NEXT: s_or_b32 s15, s35, s15
-; VI-NEXT: v_writelane_b32 v21, s46, 48
-; VI-NEXT: s_lshr_b32 s46, s12, 16
-; VI-NEXT: s_and_b32 s34, s14, 0xffff0000
-; VI-NEXT: s_add_i32 s14, s14, 3
-; VI-NEXT: s_add_i32 s15, s15, 0x30000
-; VI-NEXT: v_writelane_b32 v21, s46, 49
-; VI-NEXT: s_lshr_b32 s46, s12, 8
-; VI-NEXT: s_and_b32 s14, s14, 0xffff
-; VI-NEXT: v_writelane_b32 v21, s46, 50
-; VI-NEXT: s_lshr_b32 s46, s15, 24
-; VI-NEXT: s_and_b32 s37, s41, 0xffff0000
-; VI-NEXT: s_add_i32 s41, s41, 3
-; VI-NEXT: s_or_b32 s14, s34, s14
-; VI-NEXT: v_writelane_b32 v21, s46, 51
-; VI-NEXT: s_lshr_b32 s46, s15, 16
-; VI-NEXT: s_and_b32 s41, s41, 0xffff
-; VI-NEXT: s_add_i32 s14, s14, 0x30000
-; VI-NEXT: v_writelane_b32 v21, s46, 52
-; VI-NEXT: s_lshr_b32 s46, s15, 8
-; VI-NEXT: s_and_b32 s72, s20, 0xffff0000
-; VI-NEXT: s_add_i32 s20, s20, 3
-; VI-NEXT: s_and_b32 s74, s18, 0xffff0000
-; VI-NEXT: s_add_i32 s18, s18, 3
-; VI-NEXT: s_and_b32 s75, s19, 0xffff0000
-; VI-NEXT: s_add_i32 s19, s19, 3
-; VI-NEXT: s_and_b32 s76, s16, 0xffff0000
-; VI-NEXT: s_add_i32 s16, s16, 3
-; VI-NEXT: s_and_b32 s77, s17, 0xffff0000
-; VI-NEXT: s_add_i32 s17, s17, 3
-; VI-NEXT: s_and_b32 s36, s40, 0xffff0000
-; VI-NEXT: s_add_i32 s40, s40, 3
-; VI-NEXT: s_and_b32 s38, s42, 0xffff0000
-; VI-NEXT: s_add_i32 s42, s42, 3
-; VI-NEXT: s_and_b32 s39, s43, 0xffff0000
-; VI-NEXT: s_add_i32 s43, s43, 3
-; VI-NEXT: s_or_b32 s41, s37, s41
-; VI-NEXT: v_writelane_b32 v21, s46, 53
-; VI-NEXT: s_lshr_b32 s46, s14, 16
-; VI-NEXT: s_and_b32 s43, s43, 0xffff
-; VI-NEXT: s_and_b32 s42, s42, 0xffff
-; VI-NEXT: s_add_i32 s41, s41, 0x30000
-; VI-NEXT: s_and_b32 s40, s40, 0xffff
-; VI-NEXT: s_and_b32 s17, s17, 0xffff
-; VI-NEXT: s_and_b32 s16, s16, 0xffff
-; VI-NEXT: s_and_b32 s19, s19, 0xffff
-; VI-NEXT: s_and_b32 s18, s18, 0xffff
-; VI-NEXT: s_and_b32 s20, s20, 0xffff
-; VI-NEXT: v_writelane_b32 v21, s46, 54
-; VI-NEXT: s_lshr_b32 s46, s14, 8
-; VI-NEXT: s_or_b32 s43, s39, s43
-; VI-NEXT: s_or_b32 s42, s38, s42
-; VI-NEXT: s_or_b32 s40, s36, s40
-; VI-NEXT: s_or_b32 s17, s77, s17
-; VI-NEXT: s_or_b32 s16, s76, s16
-; VI-NEXT: s_or_b32 s19, s75, s19
-; VI-NEXT: s_or_b32 s18, s74, s18
-; VI-NEXT: s_or_b32 s20, s72, s20
-; VI-NEXT: v_writelane_b32 v21, s46, 55
-; VI-NEXT: s_lshr_b32 s46, s41, 24
-; VI-NEXT: s_add_i32 s43, s43, 0x30000
-; VI-NEXT: s_add_i32 s42, s42, 0x30000
-; VI-NEXT: s_add_i32 s40, s40, 0x30000
-; VI-NEXT: s_add_i32 s17, s17, 0x30000
-; VI-NEXT: s_add_i32 s16, s16, 0x30000
-; VI-NEXT: s_add_i32 s19, s19, 0x30000
-; VI-NEXT: s_add_i32 s18, s18, 0x30000
-; VI-NEXT: s_add_i32 s20, s20, 0x30000
-; VI-NEXT: v_writelane_b32 v21, s46, 56
-; VI-NEXT: s_lshr_b32 s46, s41, 16
-; VI-NEXT: s_lshr_b32 s80, s21, 16
-; VI-NEXT: s_lshr_b32 s82, s21, 8
-; VI-NEXT: s_lshr_b32 s84, s20, 16
-; VI-NEXT: s_lshr_b32 s86, s20, 8
-; VI-NEXT: s_lshr_b32 s51, s19, 24
-; VI-NEXT: s_lshr_b32 s53, s19, 16
-; VI-NEXT: s_lshr_b32 s54, s19, 8
-; VI-NEXT: s_lshr_b32 s65, s18, 16
-; VI-NEXT: s_lshr_b32 s66, s18, 8
-; VI-NEXT: s_lshr_b32 s67, s17, 24
-; VI-NEXT: s_lshr_b32 s68, s17, 16
-; VI-NEXT: s_lshr_b32 s69, s17, 8
-; VI-NEXT: s_lshr_b32 s70, s16, 16
-; VI-NEXT: s_lshr_b32 s71, s16, 8
-; VI-NEXT: v_writelane_b32 v21, s46, 57
-; VI-NEXT: s_lshr_b32 s81, s41, 8
-; VI-NEXT: s_lshr_b32 s83, s40, 16
-; VI-NEXT: s_lshr_b32 s85, s40, 8
-; VI-NEXT: s_lshr_b32 s87, s43, 24
-; VI-NEXT: s_lshr_b32 s50, s43, 16
-; VI-NEXT: s_lshr_b32 s52, s43, 8
-; VI-NEXT: s_lshr_b32 s55, s42, 16
-; VI-NEXT: s_lshr_b32 s64, s42, 8
-; VI-NEXT: s_lshr_b64 s[76:77], s[44:45], 24
-; VI-NEXT: s_lshr_b64 s[88:89], s[28:29], 24
-; VI-NEXT: s_lshr_b64 s[90:91], s[26:27], 24
-; VI-NEXT: s_lshr_b64 s[30:31], s[24:25], 24
-; VI-NEXT: s_lshr_b64 s[34:35], s[22:23], 24
-; VI-NEXT: s_lshr_b64 s[36:37], s[20:21], 24
-; VI-NEXT: s_lshr_b64 s[38:39], s[18:19], 24
-; VI-NEXT: s_lshr_b64 s[48:49], s[16:17], 24
-; VI-NEXT: s_lshr_b64 s[46:47], s[4:5], 24
-; VI-NEXT: s_lshr_b64 s[56:57], s[6:7], 24
-; VI-NEXT: s_lshr_b64 s[58:59], s[8:9], 24
-; VI-NEXT: s_lshr_b64 s[60:61], s[10:11], 24
-; VI-NEXT: s_lshr_b64 s[62:63], s[12:13], 24
-; VI-NEXT: s_lshr_b64 s[72:73], s[14:15], 24
-; VI-NEXT: s_lshr_b64 s[74:75], s[40:41], 24
-; VI-NEXT: s_lshr_b64 s[78:79], s[42:43], 24
-; VI-NEXT: .LBB99_3: ; %end
-; VI-NEXT: s_lshl_b32 s47, s71, 8
-; VI-NEXT: s_and_b32 s16, s16, 0xff
-; VI-NEXT: s_or_b32 s16, s16, s47
-; VI-NEXT: s_lshl_b32 s47, s48, 8
-; VI-NEXT: s_and_b32 s57, s70, 0xff
-; VI-NEXT: s_or_b32 s47, s57, s47
-; VI-NEXT: s_and_b32 s16, s16, 0xffff
-; VI-NEXT: s_lshl_b32 s47, s47, 16
-; VI-NEXT: s_or_b32 s16, s16, s47
-; VI-NEXT: v_mov_b32_e32 v1, s16
-; VI-NEXT: s_and_b32 s16, s17, 0xff
-; VI-NEXT: s_lshl_b32 s17, s69, 8
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: s_and_b32 s17, s68, 0xff
-; VI-NEXT: s_lshl_b32 s47, s67, 8
-; VI-NEXT: s_or_b32 s17, s17, s47
-; VI-NEXT: s_and_b32 s16, s16, 0xffff
-; VI-NEXT: s_lshl_b32 s17, s17, 16
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: v_mov_b32_e32 v2, s16
-; VI-NEXT: s_lshl_b32 s16, s66, 8
-; VI-NEXT: s_and_b32 s17, s18, 0xff
-; VI-NEXT: s_or_b32 s16, s17, s16
-; VI-NEXT: s_lshl_b32 s17, s38, 8
-; VI-NEXT: s_and_b32 s18, s65, 0xff
-; VI-NEXT: s_or_b32 s17, s18, s17
-; VI-NEXT: s_and_b32 s16, s16, 0xffff
-; VI-NEXT: s_lshl_b32 s17, s17, 16
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: v_mov_b32_e32 v3, s16
-; VI-NEXT: s_and_b32 s16, s19, 0xff
-; VI-NEXT: s_lshl_b32 s17, s54, 8
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: s_and_b32 s17, s53, 0xff
-; VI-NEXT: s_lshl_b32 s18, s51, 8
-; VI-NEXT: s_or_b32 s17, s17, s18
-; VI-NEXT: s_and_b32 s16, s16, 0xffff
-; VI-NEXT: s_lshl_b32 s17, s17, 16
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: v_mov_b32_e32 v4, s16
-; VI-NEXT: s_lshl_b32 s16, s86, 8
-; VI-NEXT: s_and_b32 s17, s20, 0xff
-; VI-NEXT: s_or_b32 s16, s17, s16
-; VI-NEXT: s_lshl_b32 s17, s36, 8
-; VI-NEXT: s_and_b32 s18, s84, 0xff
-; VI-NEXT: s_or_b32 s17, s18, s17
-; VI-NEXT: s_and_b32 s16, s16, 0xffff
-; VI-NEXT: s_lshl_b32 s17, s17, 16
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: v_mov_b32_e32 v5, s16
-; VI-NEXT: s_and_b32 s16, s21, 0xff
-; VI-NEXT: s_lshl_b32 s17, s82, 8
-; VI-NEXT: v_readlane_b32 s18, v21, 25
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: s_and_b32 s17, s80, 0xff
-; VI-NEXT: s_lshl_b32 s18, s18, 8
-; VI-NEXT: s_or_b32 s17, s17, s18
-; VI-NEXT: s_and_b32 s16, s16, 0xffff
-; VI-NEXT: s_lshl_b32 s17, s17, 16
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: v_mov_b32_e32 v6, s16
-; VI-NEXT: v_readlane_b32 s16, v21, 24
-; VI-NEXT: s_lshl_b32 s16, s16, 8
-; VI-NEXT: s_and_b32 s17, s22, 0xff
-; VI-NEXT: v_readlane_b32 s18, v21, 23
-; VI-NEXT: s_or_b32 s16, s17, s16
-; VI-NEXT: s_lshl_b32 s17, s34, 8
-; VI-NEXT: s_and_b32 s18, s18, 0xff
-; VI-NEXT: s_or_b32 s17, s18, s17
-; VI-NEXT: s_and_b32 s16, s16, 0xffff
-; VI-NEXT: s_lshl_b32 s17, s17, 16
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: v_readlane_b32 s17, v21, 22
-; VI-NEXT: v_mov_b32_e32 v7, s16
-; VI-NEXT: s_and_b32 s16, s23, 0xff
-; VI-NEXT: s_lshl_b32 s17, s17, 8
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: v_readlane_b32 s17, v21, 21
-; VI-NEXT: v_readlane_b32 s18, v21, 20
-; VI-NEXT: s_and_b32 s17, s17, 0xff
-; VI-NEXT: s_lshl_b32 s18, s18, 8
-; VI-NEXT: s_or_b32 s17, s17, s18
-; VI-NEXT: s_and_b32 s16, s16, 0xffff
-; VI-NEXT: s_lshl_b32 s17, s17, 16
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: v_mov_b32_e32 v8, s16
-; VI-NEXT: v_readlane_b32 s16, v21, 19
-; VI-NEXT: s_lshl_b32 s16, s16, 8
-; VI-NEXT: s_and_b32 s17, s24, 0xff
-; VI-NEXT: v_readlane_b32 s18, v21, 18
-; VI-NEXT: s_or_b32 s16, s17, s16
-; VI-NEXT: s_lshl_b32 s17, s30, 8
-; VI-NEXT: s_and_b32 s18, s18, 0xff
-; VI-NEXT: s_or_b32 s17, s18, s17
-; VI-NEXT: s_and_b32 s16, s16, 0xffff
-; VI-NEXT: s_lshl_b32 s17, s17, 16
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: v_readlane_b32 s17, v21, 17
-; VI-NEXT: v_mov_b32_e32 v9, s16
-; VI-NEXT: s_and_b32 s16, s25, 0xff
-; VI-NEXT: s_lshl_b32 s17, s17, 8
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: v_readlane_b32 s17, v21, 16
-; VI-NEXT: v_readlane_b32 s18, v21, 15
-; VI-NEXT: s_and_b32 s17, s17, 0xff
-; VI-NEXT: s_lshl_b32 s18, s18, 8
-; VI-NEXT: s_or_b32 s17, s17, s18
-; VI-NEXT: s_and_b32 s16, s16, 0xffff
-; VI-NEXT: s_lshl_b32 s17, s17, 16
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: v_mov_b32_e32 v10, s16
-; VI-NEXT: v_readlane_b32 s16, v21, 14
-; VI-NEXT: s_lshl_b32 s16, s16, 8
-; VI-NEXT: s_and_b32 s17, s26, 0xff
-; VI-NEXT: v_readlane_b32 s18, v21, 13
-; VI-NEXT: s_or_b32 s16, s17, s16
-; VI-NEXT: s_lshl_b32 s17, s90, 8
-; VI-NEXT: s_and_b32 s18, s18, 0xff
-; VI-NEXT: s_or_b32 s17, s18, s17
-; VI-NEXT: s_and_b32 s16, s16, 0xffff
-; VI-NEXT: s_lshl_b32 s17, s17, 16
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: v_readlane_b32 s17, v21, 12
-; VI-NEXT: v_mov_b32_e32 v11, s16
-; VI-NEXT: s_and_b32 s16, s27, 0xff
-; VI-NEXT: s_lshl_b32 s17, s17, 8
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: v_readlane_b32 s17, v21, 11
-; VI-NEXT: v_readlane_b32 s18, v21, 10
-; VI-NEXT: s_and_b32 s17, s17, 0xff
-; VI-NEXT: s_lshl_b32 s18, s18, 8
-; VI-NEXT: s_or_b32 s17, s17, s18
-; VI-NEXT: s_and_b32 s16, s16, 0xffff
-; VI-NEXT: s_lshl_b32 s17, s17, 16
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: v_mov_b32_e32 v12, s16
-; VI-NEXT: v_readlane_b32 s16, v21, 9
-; VI-NEXT: s_lshl_b32 s16, s16, 8
-; VI-NEXT: s_and_b32 s17, s28, 0xff
-; VI-NEXT: v_readlane_b32 s18, v21, 8
-; VI-NEXT: s_or_b32 s16, s17, s16
-; VI-NEXT: s_lshl_b32 s17, s88, 8
-; VI-NEXT: s_and_b32 s18, s18, 0xff
-; VI-NEXT: s_or_b32 s17, s18, s17
-; VI-NEXT: s_and_b32 s16, s16, 0xffff
-; VI-NEXT: s_lshl_b32 s17, s17, 16
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: v_readlane_b32 s17, v21, 7
-; VI-NEXT: v_mov_b32_e32 v13, s16
-; VI-NEXT: s_and_b32 s16, s29, 0xff
-; VI-NEXT: s_lshl_b32 s17, s17, 8
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: v_readlane_b32 s17, v21, 6
-; VI-NEXT: v_readlane_b32 s18, v21, 5
-; VI-NEXT: s_and_b32 s17, s17, 0xff
-; VI-NEXT: s_lshl_b32 s18, s18, 8
-; VI-NEXT: s_or_b32 s17, s17, s18
-; VI-NEXT: s_and_b32 s16, s16, 0xffff
-; VI-NEXT: s_lshl_b32 s17, s17, 16
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: v_mov_b32_e32 v14, s16
-; VI-NEXT: v_readlane_b32 s16, v21, 4
-; VI-NEXT: s_lshl_b32 s16, s16, 8
-; VI-NEXT: s_and_b32 s17, s44, 0xff
-; VI-NEXT: v_readlane_b32 s18, v21, 3
-; VI-NEXT: s_or_b32 s16, s17, s16
-; VI-NEXT: s_lshl_b32 s17, s76, 8
-; VI-NEXT: s_and_b32 s18, s18, 0xff
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: v_add_u32_e32 v1, vcc, 4, v0
-; VI-NEXT: s_or_b32 s17, s18, s17
-; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; VI-NEXT: v_add_u32_e32 v1, vcc, 8, v0
-; VI-NEXT: s_and_b32 s16, s16, 0xffff
-; VI-NEXT: s_lshl_b32 s17, s17, 16
-; VI-NEXT: buffer_store_dword v3, v1, s[0:3], 0 offen
-; VI-NEXT: v_add_u32_e32 v1, vcc, 12, v0
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: v_readlane_b32 s17, v21, 2
-; VI-NEXT: buffer_store_dword v4, v1, s[0:3], 0 offen
-; VI-NEXT: v_add_u32_e32 v1, vcc, 16, v0
-; VI-NEXT: v_mov_b32_e32 v15, s16
-; VI-NEXT: s_and_b32 s16, s45, 0xff
-; VI-NEXT: s_lshl_b32 s17, s17, 8
-; VI-NEXT: buffer_store_dword v5, v1, s[0:3], 0 offen
-; VI-NEXT: v_add_u32_e32 v1, vcc, 20, v0
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: buffer_store_dword v6, v1, s[0:3], 0 offen
-; VI-NEXT: v_add_u32_e32 v1, vcc, 24, v0
-; VI-NEXT: v_readlane_b32 s17, v21, 1
-; VI-NEXT: v_readlane_b32 s18, v21, 0
-; VI-NEXT: buffer_store_dword v7, v1, s[0:3], 0 offen
-; VI-NEXT: v_add_u32_e32 v1, vcc, 28, v0
-; VI-NEXT: s_and_b32 s17, s17, 0xff
-; VI-NEXT: s_lshl_b32 s18, s18, 8
-; VI-NEXT: buffer_store_dword v8, v1, s[0:3], 0 offen
-; VI-NEXT: v_add_u32_e32 v1, vcc, 32, v0
-; VI-NEXT: s_or_b32 s17, s17, s18
-; VI-NEXT: buffer_store_dword v9, v1, s[0:3], 0 offen
-; VI-NEXT: v_add_u32_e32 v1, vcc, 36, v0
-; VI-NEXT: s_and_b32 s16, s16, 0xffff
-; VI-NEXT: s_lshl_b32 s17, s17, 16
-; VI-NEXT: buffer_store_dword v10, v1, s[0:3], 0 offen
-; VI-NEXT: v_add_u32_e32 v1, vcc, 40, v0
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: buffer_store_dword v11, v1, s[0:3], 0 offen
-; VI-NEXT: v_add_u32_e32 v1, vcc, 44, v0
-; VI-NEXT: v_mov_b32_e32 v2, s16
-; VI-NEXT: s_and_b32 s16, s42, 0xff
-; VI-NEXT: s_lshl_b32 s17, s64, 8
-; VI-NEXT: buffer_store_dword v12, v1, s[0:3], 0 offen
-; VI-NEXT: v_add_u32_e32 v1, vcc, 48, v0
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: s_and_b32 s17, s55, 0xff
-; VI-NEXT: s_lshl_b32 s18, s78, 8
-; VI-NEXT: buffer_store_dword v13, v1, s[0:3], 0 offen
-; VI-NEXT: v_add_u32_e32 v1, vcc, 52, v0
-; VI-NEXT: s_or_b32 s17, s17, s18
-; VI-NEXT: buffer_store_dword v14, v1, s[0:3], 0 offen
-; VI-NEXT: v_add_u32_e32 v1, vcc, 56, v0
-; VI-NEXT: s_and_b32 s16, s16, 0xffff
-; VI-NEXT: s_lshl_b32 s17, s17, 16
-; VI-NEXT: buffer_store_dword v15, v1, s[0:3], 0 offen
-; VI-NEXT: v_add_u32_e32 v1, vcc, 60, v0
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; VI-NEXT: v_mov_b32_e32 v2, s16
-; VI-NEXT: s_and_b32 s16, s43, 0xff
-; VI-NEXT: s_lshl_b32 s17, s52, 8
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: s_and_b32 s17, s50, 0xff
-; VI-NEXT: s_lshl_b32 s18, s87, 8
-; VI-NEXT: s_or_b32 s17, s17, s18
-; VI-NEXT: s_and_b32 s16, s16, 0xffff
-; VI-NEXT: s_lshl_b32 s17, s17, 16
-; VI-NEXT: v_add_u32_e32 v1, vcc, 64, v0
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; VI-NEXT: v_mov_b32_e32 v2, s16
-; VI-NEXT: s_and_b32 s16, s40, 0xff
-; VI-NEXT: s_lshl_b32 s17, s85, 8
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: s_and_b32 s17, s83, 0xff
-; VI-NEXT: s_lshl_b32 s18, s74, 8
-; VI-NEXT: s_or_b32 s17, s17, s18
-; VI-NEXT: s_and_b32 s16, s16, 0xffff
-; VI-NEXT: s_lshl_b32 s17, s17, 16
-; VI-NEXT: v_add_u32_e32 v1, vcc, 0x44, v0
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; VI-NEXT: v_mov_b32_e32 v2, s16
-; VI-NEXT: s_and_b32 s16, s41, 0xff
-; VI-NEXT: s_lshl_b32 s17, s81, 8
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: v_readlane_b32 s17, v21, 57
-; VI-NEXT: v_readlane_b32 s18, v21, 56
-; VI-NEXT: s_and_b32 s17, s17, 0xff
-; VI-NEXT: s_lshl_b32 s18, s18, 8
-; VI-NEXT: s_or_b32 s17, s17, s18
-; VI-NEXT: s_and_b32 s16, s16, 0xffff
-; VI-NEXT: s_lshl_b32 s17, s17, 16
-; VI-NEXT: v_add_u32_e32 v1, vcc, 0x48, v0
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; VI-NEXT: v_mov_b32_e32 v2, s16
-; VI-NEXT: v_readlane_b32 s16, v21, 55
-; VI-NEXT: s_and_b32 s14, s14, 0xff
-; VI-NEXT: s_lshl_b32 s16, s16, 8
-; VI-NEXT: s_or_b32 s14, s14, s16
-; VI-NEXT: v_readlane_b32 s16, v21, 54
-; VI-NEXT: s_and_b32 s16, s16, 0xff
-; VI-NEXT: s_lshl_b32 s17, s72, 8
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: s_and_b32 s14, s14, 0xffff
-; VI-NEXT: s_lshl_b32 s16, s16, 16
-; VI-NEXT: v_add_u32_e32 v1, vcc, 0x4c, v0
-; VI-NEXT: s_or_b32 s14, s14, s16
-; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; VI-NEXT: v_mov_b32_e32 v2, s14
-; VI-NEXT: s_and_b32 s14, s15, 0xff
-; VI-NEXT: v_readlane_b32 s15, v21, 53
-; VI-NEXT: s_lshl_b32 s15, s15, 8
-; VI-NEXT: s_or_b32 s14, s14, s15
-; VI-NEXT: v_readlane_b32 s15, v21, 52
-; VI-NEXT: v_readlane_b32 s16, v21, 51
-; VI-NEXT: s_and_b32 s15, s15, 0xff
-; VI-NEXT: s_lshl_b32 s16, s16, 8
-; VI-NEXT: s_or_b32 s15, s15, s16
-; VI-NEXT: s_and_b32 s14, s14, 0xffff
-; VI-NEXT: s_lshl_b32 s15, s15, 16
-; VI-NEXT: v_add_u32_e32 v1, vcc, 0x50, v0
-; VI-NEXT: s_or_b32 s14, s14, s15
-; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; VI-NEXT: v_mov_b32_e32 v2, s14
-; VI-NEXT: v_readlane_b32 s14, v21, 50
-; VI-NEXT: s_and_b32 s12, s12, 0xff
-; VI-NEXT: s_lshl_b32 s14, s14, 8
-; VI-NEXT: s_or_b32 s12, s12, s14
-; VI-NEXT: v_readlane_b32 s14, v21, 49
-; VI-NEXT: s_and_b32 s14, s14, 0xff
-; VI-NEXT: s_lshl_b32 s15, s62, 8
-; VI-NEXT: s_or_b32 s14, s14, s15
-; VI-NEXT: s_and_b32 s12, s12, 0xffff
-; VI-NEXT: s_lshl_b32 s14, s14, 16
-; VI-NEXT: v_add_u32_e32 v1, vcc, 0x54, v0
-; VI-NEXT: s_or_b32 s12, s12, s14
-; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; VI-NEXT: v_mov_b32_e32 v2, s12
-; VI-NEXT: s_and_b32 s12, s13, 0xff
-; VI-NEXT: v_readlane_b32 s13, v21, 48
-; VI-NEXT: s_lshl_b32 s13, s13, 8
-; VI-NEXT: s_or_b32 s12, s12, s13
-; VI-NEXT: v_readlane_b32 s13, v21, 47
-; VI-NEXT: v_readlane_b32 s14, v21, 46
-; VI-NEXT: s_and_b32 s13, s13, 0xff
-; VI-NEXT: s_lshl_b32 s14, s14, 8
-; VI-NEXT: s_or_b32 s13, s13, s14
-; VI-NEXT: s_and_b32 s12, s12, 0xffff
-; VI-NEXT: s_lshl_b32 s13, s13, 16
-; VI-NEXT: v_add_u32_e32 v1, vcc, 0x58, v0
-; VI-NEXT: s_or_b32 s12, s12, s13
-; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; VI-NEXT: v_mov_b32_e32 v2, s12
-; VI-NEXT: v_readlane_b32 s12, v21, 45
-; VI-NEXT: s_and_b32 s10, s10, 0xff
-; VI-NEXT: s_lshl_b32 s12, s12, 8
-; VI-NEXT: s_or_b32 s10, s10, s12
-; VI-NEXT: v_readlane_b32 s12, v21, 44
-; VI-NEXT: s_and_b32 s12, s12, 0xff
-; VI-NEXT: s_lshl_b32 s13, s60, 8
-; VI-NEXT: s_or_b32 s12, s12, s13
-; VI-NEXT: s_and_b32 s10, s10, 0xffff
-; VI-NEXT: s_lshl_b32 s12, s12, 16
-; VI-NEXT: v_add_u32_e32 v1, vcc, 0x5c, v0
-; VI-NEXT: s_or_b32 s10, s10, s12
-; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; VI-NEXT: v_mov_b32_e32 v2, s10
-; VI-NEXT: s_and_b32 s10, s11, 0xff
-; VI-NEXT: v_readlane_b32 s11, v21, 43
-; VI-NEXT: s_lshl_b32 s11, s11, 8
-; VI-NEXT: s_or_b32 s10, s10, s11
-; VI-NEXT: v_readlane_b32 s11, v21, 42
-; VI-NEXT: v_readlane_b32 s12, v21, 41
-; VI-NEXT: s_and_b32 s11, s11, 0xff
-; VI-NEXT: s_lshl_b32 s12, s12, 8
-; VI-NEXT: s_or_b32 s11, s11, s12
-; VI-NEXT: s_and_b32 s10, s10, 0xffff
-; VI-NEXT: s_lshl_b32 s11, s11, 16
-; VI-NEXT: v_add_u32_e32 v1, vcc, 0x60, v0
-; VI-NEXT: s_or_b32 s10, s10, s11
-; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; VI-NEXT: v_mov_b32_e32 v2, s10
-; VI-NEXT: v_readlane_b32 s10, v21, 40
-; VI-NEXT: s_and_b32 s8, s8, 0xff
-; VI-NEXT: s_lshl_b32 s10, s10, 8
-; VI-NEXT: s_or_b32 s8, s8, s10
-; VI-NEXT: v_readlane_b32 s10, v21, 39
-; VI-NEXT: s_and_b32 s10, s10, 0xff
-; VI-NEXT: s_lshl_b32 s11, s58, 8
-; VI-NEXT: s_or_b32 s10, s10, s11
-; VI-NEXT: s_and_b32 s8, s8, 0xffff
-; VI-NEXT: s_lshl_b32 s10, s10, 16
-; VI-NEXT: v_add_u32_e32 v1, vcc, 0x64, v0
-; VI-NEXT: s_or_b32 s8, s8, s10
-; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; VI-NEXT: v_mov_b32_e32 v2, s8
-; VI-NEXT: s_and_b32 s8, s9, 0xff
-; VI-NEXT: v_readlane_b32 s9, v21, 38
-; VI-NEXT: s_lshl_b32 s9, s9, 8
-; VI-NEXT: s_or_b32 s8, s8, s9
-; VI-NEXT: v_readlane_b32 s9, v21, 37
-; VI-NEXT: v_readlane_b32 s10, v21, 36
-; VI-NEXT: s_and_b32 s9, s9, 0xff
-; VI-NEXT: s_lshl_b32 s10, s10, 8
-; VI-NEXT: s_or_b32 s9, s9, s10
-; VI-NEXT: s_and_b32 s8, s8, 0xffff
-; VI-NEXT: s_lshl_b32 s9, s9, 16
-; VI-NEXT: v_add_u32_e32 v1, vcc, 0x68, v0
-; VI-NEXT: s_or_b32 s8, s8, s9
-; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; VI-NEXT: v_mov_b32_e32 v2, s8
-; VI-NEXT: v_readlane_b32 s8, v21, 35
-; VI-NEXT: s_and_b32 s6, s6, 0xff
-; VI-NEXT: s_lshl_b32 s8, s8, 8
-; VI-NEXT: s_or_b32 s6, s6, s8
-; VI-NEXT: v_readlane_b32 s8, v21, 34
-; VI-NEXT: s_and_b32 s8, s8, 0xff
-; VI-NEXT: s_lshl_b32 s9, s56, 8
-; VI-NEXT: s_or_b32 s8, s8, s9
-; VI-NEXT: s_and_b32 s6, s6, 0xffff
-; VI-NEXT: s_lshl_b32 s8, s8, 16
-; VI-NEXT: v_add_u32_e32 v1, vcc, 0x6c, v0
-; VI-NEXT: s_or_b32 s6, s6, s8
-; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; VI-NEXT: v_mov_b32_e32 v2, s6
-; VI-NEXT: s_and_b32 s6, s7, 0xff
-; VI-NEXT: v_readlane_b32 s7, v21, 33
-; VI-NEXT: s_lshl_b32 s7, s7, 8
-; VI-NEXT: s_or_b32 s6, s6, s7
-; VI-NEXT: v_readlane_b32 s7, v21, 32
-; VI-NEXT: v_readlane_b32 s8, v21, 31
-; VI-NEXT: s_and_b32 s7, s7, 0xff
-; VI-NEXT: s_lshl_b32 s8, s8, 8
-; VI-NEXT: s_or_b32 s7, s7, s8
-; VI-NEXT: s_and_b32 s6, s6, 0xffff
-; VI-NEXT: s_lshl_b32 s7, s7, 16
-; VI-NEXT: v_add_u32_e32 v1, vcc, 0x70, v0
-; VI-NEXT: s_or_b32 s6, s6, s7
-; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; VI-NEXT: v_mov_b32_e32 v2, s6
-; VI-NEXT: v_readlane_b32 s6, v21, 30
-; VI-NEXT: s_and_b32 s4, s4, 0xff
-; VI-NEXT: s_lshl_b32 s6, s6, 8
-; VI-NEXT: s_or_b32 s4, s4, s6
-; VI-NEXT: v_readlane_b32 s6, v21, 29
-; VI-NEXT: s_and_b32 s6, s6, 0xff
-; VI-NEXT: s_lshl_b32 s7, s46, 8
-; VI-NEXT: s_or_b32 s6, s6, s7
-; VI-NEXT: s_and_b32 s4, s4, 0xffff
-; VI-NEXT: s_lshl_b32 s6, s6, 16
-; VI-NEXT: v_add_u32_e32 v1, vcc, 0x74, v0
-; VI-NEXT: s_or_b32 s4, s4, s6
-; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; VI-NEXT: v_mov_b32_e32 v2, s4
-; VI-NEXT: s_and_b32 s4, s5, 0xff
-; VI-NEXT: v_readlane_b32 s5, v21, 28
-; VI-NEXT: s_lshl_b32 s5, s5, 8
-; VI-NEXT: s_or_b32 s4, s4, s5
-; VI-NEXT: v_readlane_b32 s5, v21, 27
-; VI-NEXT: v_readlane_b32 s6, v21, 26
-; VI-NEXT: s_and_b32 s5, s5, 0xff
-; VI-NEXT: s_lshl_b32 s6, s6, 8
-; VI-NEXT: s_or_b32 s5, s5, s6
-; VI-NEXT: s_and_b32 s4, s4, 0xffff
-; VI-NEXT: s_lshl_b32 s5, s5, 16
-; VI-NEXT: v_add_u32_e32 v1, vcc, 0x78, v0
-; VI-NEXT: s_or_b32 s4, s4, s5
-; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
-; VI-NEXT: v_add_u32_e32 v0, vcc, 0x7c, v0
-; VI-NEXT: v_mov_b32_e32 v1, s4
-; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; VI-NEXT: v_readlane_b32 s87, v20, 31
-; VI-NEXT: v_readlane_b32 s86, v20, 30
-; VI-NEXT: v_readlane_b32 s85, v20, 29
-; VI-NEXT: v_readlane_b32 s84, v20, 28
-; VI-NEXT: v_readlane_b32 s83, v20, 27
-; VI-NEXT: v_readlane_b32 s82, v20, 26
-; VI-NEXT: v_readlane_b32 s81, v20, 25
-; VI-NEXT: v_readlane_b32 s80, v20, 24
-; VI-NEXT: v_readlane_b32 s71, v20, 23
-; VI-NEXT: v_readlane_b32 s70, v20, 22
-; VI-NEXT: v_readlane_b32 s69, v20, 21
-; VI-NEXT: v_readlane_b32 s68, v20, 20
-; VI-NEXT: v_readlane_b32 s67, v20, 19
-; VI-NEXT: v_readlane_b32 s66, v20, 18
-; VI-NEXT: v_readlane_b32 s65, v20, 17
-; VI-NEXT: v_readlane_b32 s64, v20, 16
-; VI-NEXT: v_readlane_b32 s55, v20, 15
-; VI-NEXT: v_readlane_b32 s54, v20, 14
-; VI-NEXT: v_readlane_b32 s53, v20, 13
-; VI-NEXT: v_readlane_b32 s52, v20, 12
-; VI-NEXT: v_readlane_b32 s51, v20, 11
-; VI-NEXT: v_readlane_b32 s50, v20, 10
-; VI-NEXT: v_readlane_b32 s49, v20, 9
-; VI-NEXT: v_readlane_b32 s48, v20, 8
-; VI-NEXT: v_readlane_b32 s39, v20, 7
-; VI-NEXT: v_readlane_b32 s38, v20, 6
-; VI-NEXT: v_readlane_b32 s37, v20, 5
-; VI-NEXT: v_readlane_b32 s36, v20, 4
-; VI-NEXT: v_readlane_b32 s35, v20, 3
-; VI-NEXT: v_readlane_b32 s34, v20, 2
-; VI-NEXT: v_readlane_b32 s31, v20, 1
-; VI-NEXT: v_readlane_b32 s30, v20, 0
-; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 ; 4-byte Folded Reload
-; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; VI-NEXT: s_mov_b64 exec, s[4:5]
-; VI-NEXT: s_waitcnt vmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB99_4:
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr71
-; VI-NEXT: ; implicit-def: $sgpr70
-; VI-NEXT: ; implicit-def: $sgpr69
-; VI-NEXT: ; implicit-def: $sgpr68
-; VI-NEXT: ; implicit-def: $sgpr67
-; VI-NEXT: ; implicit-def: $sgpr66
-; VI-NEXT: ; implicit-def: $sgpr65
-; VI-NEXT: ; implicit-def: $sgpr54
-; VI-NEXT: ; implicit-def: $sgpr53
-; VI-NEXT: ; implicit-def: $sgpr51
-; VI-NEXT: ; implicit-def: $sgpr86
-; VI-NEXT: ; implicit-def: $sgpr84
-; VI-NEXT: ; implicit-def: $sgpr82
-; VI-NEXT: ; implicit-def: $sgpr80
-; VI-NEXT: ; implicit-def: $sgpr64
-; VI-NEXT: ; implicit-def: $sgpr55
-; VI-NEXT: ; implicit-def: $sgpr52
-; VI-NEXT: ; implicit-def: $sgpr50
-; VI-NEXT: ; implicit-def: $sgpr87
-; VI-NEXT: ; implicit-def: $sgpr85
-; VI-NEXT: ; implicit-def: $sgpr83
-; VI-NEXT: ; implicit-def: $sgpr81
-; VI-NEXT: ; implicit-def: $sgpr48
-; VI-NEXT: ; implicit-def: $sgpr38
-; VI-NEXT: ; implicit-def: $sgpr36
-; VI-NEXT: ; implicit-def: $sgpr34
-; VI-NEXT: ; implicit-def: $sgpr30
-; VI-NEXT: ; implicit-def: $sgpr90
-; VI-NEXT: ; implicit-def: $sgpr88
-; VI-NEXT: ; implicit-def: $sgpr76
-; VI-NEXT: ; implicit-def: $sgpr78
-; VI-NEXT: ; implicit-def: $sgpr74
-; VI-NEXT: ; implicit-def: $sgpr72
-; VI-NEXT: ; implicit-def: $sgpr62
-; VI-NEXT: ; implicit-def: $sgpr60
-; VI-NEXT: ; implicit-def: $sgpr58
-; VI-NEXT: ; implicit-def: $sgpr56
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: ; kill: killed $sgpr46
-; VI-NEXT: ; implicit-def: $sgpr46
-; VI-NEXT: s_branch .LBB99_2
-;
-; GFX9-LABEL: bitcast_v64i16_to_v128i8_scalar:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: s_or_saveexec_b64 s[4:5], -1
-; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; GFX9-NEXT: s_mov_b64 exec, s[4:5]
-; GFX9-NEXT: v_writelane_b32 v63, s30, 0
-; GFX9-NEXT: v_writelane_b32 v63, s31, 1
-; GFX9-NEXT: v_writelane_b32 v63, s34, 2
-; GFX9-NEXT: v_writelane_b32 v63, s35, 3
-; GFX9-NEXT: v_writelane_b32 v63, s36, 4
-; GFX9-NEXT: v_writelane_b32 v63, s37, 5
-; GFX9-NEXT: v_writelane_b32 v63, s38, 6
-; GFX9-NEXT: v_writelane_b32 v63, s39, 7
-; GFX9-NEXT: v_writelane_b32 v63, s48, 8
-; GFX9-NEXT: v_writelane_b32 v63, s49, 9
-; GFX9-NEXT: v_writelane_b32 v63, s50, 10
-; GFX9-NEXT: v_writelane_b32 v63, s51, 11
-; GFX9-NEXT: v_writelane_b32 v63, s52, 12
-; GFX9-NEXT: v_writelane_b32 v63, s53, 13
-; GFX9-NEXT: v_writelane_b32 v63, s54, 14
-; GFX9-NEXT: v_writelane_b32 v63, s55, 15
-; GFX9-NEXT: v_writelane_b32 v63, s64, 16
-; GFX9-NEXT: v_writelane_b32 v63, s65, 17
-; GFX9-NEXT: v_writelane_b32 v63, s66, 18
-; GFX9-NEXT: v_writelane_b32 v63, s67, 19
-; GFX9-NEXT: v_writelane_b32 v63, s68, 20
-; GFX9-NEXT: v_writelane_b32 v63, s69, 21
-; GFX9-NEXT: v_writelane_b32 v63, s70, 22
-; GFX9-NEXT: v_writelane_b32 v63, s71, 23
-; GFX9-NEXT: v_writelane_b32 v63, s80, 24
-; GFX9-NEXT: v_writelane_b32 v63, s81, 25
-; GFX9-NEXT: v_writelane_b32 v63, s82, 26
-; GFX9-NEXT: v_writelane_b32 v63, s83, 27
-; GFX9-NEXT: v_writelane_b32 v63, s84, 28
-; GFX9-NEXT: v_writelane_b32 v63, s85, 29
-; GFX9-NEXT: v_writelane_b32 v63, s86, 30
-; GFX9-NEXT: v_writelane_b32 v63, s87, 31
-; GFX9-NEXT: v_writelane_b32 v63, s96, 32
-; GFX9-NEXT: v_writelane_b32 v63, s97, 33
-; GFX9-NEXT: v_writelane_b32 v63, s98, 34
-; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v19
-; GFX9-NEXT: v_writelane_b32 v63, s99, 35
-; GFX9-NEXT: v_readfirstlane_b32 s44, v3
-; GFX9-NEXT: v_readfirstlane_b32 s45, v4
-; GFX9-NEXT: v_readfirstlane_b32 s42, v5
-; GFX9-NEXT: v_readfirstlane_b32 s43, v6
-; GFX9-NEXT: v_readfirstlane_b32 s40, v7
-; GFX9-NEXT: v_readfirstlane_b32 s41, v8
-; GFX9-NEXT: v_readfirstlane_b32 s14, v9
-; GFX9-NEXT: v_readfirstlane_b32 s15, v10
-; GFX9-NEXT: v_readfirstlane_b32 s12, v11
-; GFX9-NEXT: v_readfirstlane_b32 s13, v12
-; GFX9-NEXT: v_readfirstlane_b32 s10, v13
-; GFX9-NEXT: v_readfirstlane_b32 s11, v14
-; GFX9-NEXT: v_readfirstlane_b32 s8, v15
-; GFX9-NEXT: v_readfirstlane_b32 s9, v16
-; GFX9-NEXT: v_readfirstlane_b32 s6, v17
-; GFX9-NEXT: v_readfirstlane_b32 s7, v18
-; GFX9-NEXT: v_readfirstlane_b32 s4, v1
-; GFX9-NEXT: s_and_b64 s[46:47], vcc, exec
-; GFX9-NEXT: v_readfirstlane_b32 s5, v2
-; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 ; 4-byte Folded Spill
-; GFX9-NEXT: ; implicit-def: $vgpr62 : SGPR spill to VGPR lane
-; GFX9-NEXT: s_cbranch_scc0 .LBB99_3
-; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: s_lshr_b32 s46, s5, 24
-; GFX9-NEXT: v_writelane_b32 v62, s46, 49
-; GFX9-NEXT: s_lshr_b32 s46, s5, 16
-; GFX9-NEXT: v_writelane_b32 v62, s46, 48
-; GFX9-NEXT: s_lshr_b32 s46, s5, 8
-; GFX9-NEXT: v_writelane_b32 v62, s46, 47
-; GFX9-NEXT: s_lshr_b32 s46, s4, 16
-; GFX9-NEXT: v_writelane_b32 v62, s46, 46
-; GFX9-NEXT: s_lshr_b32 s46, s4, 8
-; GFX9-NEXT: v_writelane_b32 v62, s46, 45
-; GFX9-NEXT: s_lshr_b32 s46, s29, 24
-; GFX9-NEXT: v_writelane_b32 v62, s46, 44
-; GFX9-NEXT: s_lshr_b32 s46, s29, 16
-; GFX9-NEXT: v_writelane_b32 v62, s46, 43
-; GFX9-NEXT: s_lshr_b32 s46, s29, 8
-; GFX9-NEXT: v_writelane_b32 v62, s46, 42
-; GFX9-NEXT: s_lshr_b32 s46, s28, 16
-; GFX9-NEXT: v_writelane_b32 v62, s46, 41
-; GFX9-NEXT: s_lshr_b32 s46, s28, 8
-; GFX9-NEXT: v_writelane_b32 v62, s46, 40
-; GFX9-NEXT: s_lshr_b32 s46, s27, 24
-; GFX9-NEXT: v_writelane_b32 v62, s46, 39
-; GFX9-NEXT: s_lshr_b32 s46, s27, 16
-; GFX9-NEXT: v_writelane_b32 v62, s46, 38
-; GFX9-NEXT: s_lshr_b32 s46, s27, 8
-; GFX9-NEXT: v_writelane_b32 v62, s46, 37
-; GFX9-NEXT: s_lshr_b32 s46, s26, 16
-; GFX9-NEXT: v_writelane_b32 v62, s46, 36
-; GFX9-NEXT: s_lshr_b32 s46, s26, 8
-; GFX9-NEXT: v_writelane_b32 v62, s46, 35
-; GFX9-NEXT: s_lshr_b32 s46, s25, 24
-; GFX9-NEXT: v_writelane_b32 v62, s46, 34
-; GFX9-NEXT: s_lshr_b32 s46, s25, 16
-; GFX9-NEXT: v_writelane_b32 v62, s46, 33
-; GFX9-NEXT: s_lshr_b32 s46, s25, 8
-; GFX9-NEXT: v_writelane_b32 v62, s46, 32
-; GFX9-NEXT: s_lshr_b32 s46, s24, 16
-; GFX9-NEXT: v_writelane_b32 v62, s46, 31
-; GFX9-NEXT: s_lshr_b32 s46, s24, 8
-; GFX9-NEXT: v_writelane_b32 v62, s46, 30
-; GFX9-NEXT: s_lshr_b32 s46, s23, 24
-; GFX9-NEXT: v_writelane_b32 v62, s46, 29
-; GFX9-NEXT: s_lshr_b32 s46, s23, 16
-; GFX9-NEXT: v_writelane_b32 v62, s46, 28
-; GFX9-NEXT: s_lshr_b32 s46, s23, 8
-; GFX9-NEXT: v_writelane_b32 v62, s46, 27
-; GFX9-NEXT: s_lshr_b32 s46, s22, 16
-; GFX9-NEXT: v_writelane_b32 v62, s46, 26
-; GFX9-NEXT: s_lshr_b32 s46, s22, 8
-; GFX9-NEXT: v_writelane_b32 v62, s46, 25
-; GFX9-NEXT: s_lshr_b32 s46, s21, 24
-; GFX9-NEXT: v_writelane_b32 v62, s46, 24
-; GFX9-NEXT: s_lshr_b32 s46, s21, 16
-; GFX9-NEXT: v_writelane_b32 v62, s46, 23
-; GFX9-NEXT: s_lshr_b32 s46, s21, 8
-; GFX9-NEXT: v_writelane_b32 v62, s46, 22
-; GFX9-NEXT: s_lshr_b32 s46, s20, 16
-; GFX9-NEXT: v_writelane_b32 v62, s46, 21
-; GFX9-NEXT: s_lshr_b32 s46, s20, 8
-; GFX9-NEXT: v_writelane_b32 v62, s46, 20
-; GFX9-NEXT: s_lshr_b32 s46, s19, 24
-; GFX9-NEXT: v_writelane_b32 v62, s46, 19
-; GFX9-NEXT: s_lshr_b32 s46, s19, 16
-; GFX9-NEXT: v_writelane_b32 v62, s46, 18
-; GFX9-NEXT: s_lshr_b32 s46, s19, 8
-; GFX9-NEXT: v_writelane_b32 v62, s46, 17
-; GFX9-NEXT: s_lshr_b32 s46, s18, 16
-; GFX9-NEXT: v_writelane_b32 v62, s46, 16
-; GFX9-NEXT: s_lshr_b32 s46, s18, 8
-; GFX9-NEXT: v_writelane_b32 v62, s46, 15
-; GFX9-NEXT: s_lshr_b32 s46, s17, 24
-; GFX9-NEXT: v_writelane_b32 v62, s46, 14
-; GFX9-NEXT: s_lshr_b32 s46, s17, 16
-; GFX9-NEXT: v_writelane_b32 v62, s46, 13
-; GFX9-NEXT: s_lshr_b32 s46, s17, 8
-; GFX9-NEXT: v_writelane_b32 v62, s46, 12
-; GFX9-NEXT: s_lshr_b32 s46, s16, 16
-; GFX9-NEXT: v_writelane_b32 v62, s46, 11
-; GFX9-NEXT: s_lshr_b32 s46, s16, 8
-; GFX9-NEXT: v_writelane_b32 v62, s46, 10
-; GFX9-NEXT: s_lshr_b32 s46, s7, 24
-; GFX9-NEXT: v_writelane_b32 v62, s46, 9
-; GFX9-NEXT: s_lshr_b32 s46, s7, 16
-; GFX9-NEXT: v_writelane_b32 v62, s46, 8
-; GFX9-NEXT: s_lshr_b32 s46, s7, 8
-; GFX9-NEXT: v_writelane_b32 v62, s46, 7
-; GFX9-NEXT: s_lshr_b32 s46, s6, 16
-; GFX9-NEXT: v_writelane_b32 v62, s46, 6
-; GFX9-NEXT: s_lshr_b32 s46, s6, 8
-; GFX9-NEXT: v_writelane_b32 v62, s46, 5
-; GFX9-NEXT: s_lshr_b32 s46, s9, 24
-; GFX9-NEXT: v_writelane_b32 v62, s46, 4
-; GFX9-NEXT: s_lshr_b32 s46, s9, 16
-; GFX9-NEXT: v_writelane_b32 v62, s46, 3
-; GFX9-NEXT: s_lshr_b32 s46, s9, 8
-; GFX9-NEXT: v_writelane_b32 v62, s46, 2
-; GFX9-NEXT: s_lshr_b32 s46, s8, 16
-; GFX9-NEXT: v_writelane_b32 v62, s46, 1
-; GFX9-NEXT: s_lshr_b32 s46, s8, 8
-; GFX9-NEXT: v_writelane_b32 v62, s46, 0
-; GFX9-NEXT: s_lshr_b32 s82, s11, 24
-; GFX9-NEXT: s_lshr_b32 s83, s11, 16
-; GFX9-NEXT: s_lshr_b32 s85, s11, 8
-; GFX9-NEXT: s_lshr_b32 s84, s10, 16
-; GFX9-NEXT: s_lshr_b32 s86, s10, 8
-; GFX9-NEXT: s_lshr_b32 s87, s13, 24
-; GFX9-NEXT: s_lshr_b32 s96, s13, 16
-; GFX9-NEXT: s_lshr_b32 s98, s13, 8
-; GFX9-NEXT: s_lshr_b32 s97, s12, 16
-; GFX9-NEXT: s_lshr_b32 s99, s12, 8
-; GFX9-NEXT: s_lshr_b32 s38, s15, 24
-; GFX9-NEXT: s_lshr_b32 s39, s15, 16
-; GFX9-NEXT: s_lshr_b32 s49, s15, 8
-; GFX9-NEXT: s_lshr_b32 s48, s14, 16
-; GFX9-NEXT: s_lshr_b32 s50, s14, 8
-; GFX9-NEXT: s_lshr_b32 s51, s41, 24
-; GFX9-NEXT: s_lshr_b32 s52, s41, 16
-; GFX9-NEXT: s_lshr_b32 s54, s41, 8
-; GFX9-NEXT: s_lshr_b32 s53, s40, 16
-; GFX9-NEXT: s_lshr_b32 s55, s40, 8
-; GFX9-NEXT: s_lshr_b32 s64, s43, 24
-; GFX9-NEXT: s_lshr_b32 s65, s43, 16
-; GFX9-NEXT: s_lshr_b32 s67, s43, 8
-; GFX9-NEXT: s_lshr_b32 s66, s42, 16
-; GFX9-NEXT: s_lshr_b32 s68, s42, 8
-; GFX9-NEXT: s_lshr_b32 s69, s45, 24
-; GFX9-NEXT: s_lshr_b32 s70, s45, 16
-; GFX9-NEXT: s_lshr_b32 s80, s45, 8
-; GFX9-NEXT: s_lshr_b32 s71, s44, 16
-; GFX9-NEXT: s_lshr_b32 s81, s44, 8
-; GFX9-NEXT: s_lshr_b64 s[46:47], s[4:5], 24
-; GFX9-NEXT: s_lshr_b64 s[56:57], s[28:29], 24
-; GFX9-NEXT: s_lshr_b64 s[58:59], s[26:27], 24
-; GFX9-NEXT: s_lshr_b64 s[60:61], s[24:25], 24
-; GFX9-NEXT: s_lshr_b64 s[62:63], s[22:23], 24
-; GFX9-NEXT: s_lshr_b64 s[72:73], s[20:21], 24
-; GFX9-NEXT: s_lshr_b64 s[74:75], s[18:19], 24
-; GFX9-NEXT: s_lshr_b64 s[76:77], s[16:17], 24
-; GFX9-NEXT: s_lshr_b64 s[78:79], s[6:7], 24
-; GFX9-NEXT: s_lshr_b64 s[88:89], s[8:9], 24
-; GFX9-NEXT: s_lshr_b64 s[90:91], s[10:11], 24
-; GFX9-NEXT: s_lshr_b64 s[92:93], s[12:13], 24
-; GFX9-NEXT: s_lshr_b64 s[94:95], s[14:15], 24
-; GFX9-NEXT: s_lshr_b64 s[30:31], s[40:41], 24
-; GFX9-NEXT: s_lshr_b64 s[34:35], s[42:43], 24
-; GFX9-NEXT: s_lshr_b64 s[36:37], s[44:45], 24
-; GFX9-NEXT: s_cbranch_execnz .LBB99_4
-; GFX9-NEXT: .LBB99_2: ; %cmp.true
-; GFX9-NEXT: v_pk_add_u16 v26, s5, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v25, s4, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[25:26]
-; GFX9-NEXT: v_pk_add_u16 v28, s29, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v27, s28, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[27:28]
-; GFX9-NEXT: v_pk_add_u16 v30, s27, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v29, s26, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[29:30]
-; GFX9-NEXT: v_pk_add_u16 v32, s25, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v31, s24, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[31:32]
-; GFX9-NEXT: v_pk_add_u16 v34, s23, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v33, s22, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[33:34]
-; GFX9-NEXT: v_pk_add_u16 v36, s21, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v35, s20, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[35:36]
-; GFX9-NEXT: v_pk_add_u16 v38, s19, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v37, s18, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[37:38]
-; GFX9-NEXT: v_pk_add_u16 v49, s17, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v48, s16, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[48:49]
-; GFX9-NEXT: v_pk_add_u16 v2, s7, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v1, s6, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
-; GFX9-NEXT: v_pk_add_u16 v4, s9, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v3, s8, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[3:4]
-; GFX9-NEXT: v_pk_add_u16 v6, s11, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v5, s10, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[5:6]
-; GFX9-NEXT: v_pk_add_u16 v8, s13, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v7, s12, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[7:8]
-; GFX9-NEXT: v_pk_add_u16 v10, s15, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v9, s14, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[9:10]
-; GFX9-NEXT: v_pk_add_u16 v12, s41, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v11, s40, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[11:12]
-; GFX9-NEXT: v_pk_add_u16 v14, s43, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v13, s42, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[13:14]
-; GFX9-NEXT: v_pk_add_u16 v22, s45, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: v_pk_add_u16 v21, s44, 3 op_sel_hi:[1,0]
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[21:22]
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 24, v4
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v3
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v3
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 24, v6
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v6
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v6
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v5
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v5
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 24, v8
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v8
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v8
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v7
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v26
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v7
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v26
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 24, v10
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v25
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v10
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v28
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v10
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v28
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v9
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v27
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v9
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v30
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 24, v12
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v30
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v12
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v29
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v12
-; 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, v32
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v11
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v32
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v11
-; 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, v31
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 24, v14
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v34
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v14
-; 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, v34
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v14
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v33
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v13
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v36
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v13
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v36
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 24, v22
-; 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, v35
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v22
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v38
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v22
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 8, v38
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v21
-; GFX9-NEXT: v_lshrrev_b32_e32 v44, 8, v26
-; GFX9-NEXT: v_lshrrev_b32_e32 v45, 8, v25
-; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v28
-; GFX9-NEXT: v_lshrrev_b32_e32 v47, 8, v27
-; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v30
-; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v29
-; GFX9-NEXT: v_lshrrev_b32_e32 v57, 8, v32
-; GFX9-NEXT: v_lshrrev_b32_e32 v58, 8, v31
-; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v34
-; GFX9-NEXT: v_lshrrev_b32_e32 v23, 8, v33
-; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v36
-; GFX9-NEXT: v_lshrrev_b32_e32 v61, 8, v35
-; GFX9-NEXT: v_lshrrev_b32_e32 v54, 24, v38
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v37
-; GFX9-NEXT: v_lshrrev_b32_e32 v15, 8, v37
-; GFX9-NEXT: v_lshrrev_b32_e32 v43, 24, v49
-; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v49
-; GFX9-NEXT: v_lshrrev_b32_e32 v18, 8, v49
-; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v48
-; GFX9-NEXT: v_lshrrev_b32_e32 v20, 8, v48
-; GFX9-NEXT: v_lshrrev_b32_e32 v24, 24, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v50, 8, v2
-; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v1
-; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v4
-; GFX9-NEXT: v_lshrrev_b32_e32 v55, 8, v4
-; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v21
-; GFX9-NEXT: s_branch .LBB99_5
-; GFX9-NEXT: .LBB99_3:
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr81
-; GFX9-NEXT: ; implicit-def: $sgpr71
-; GFX9-NEXT: ; implicit-def: $sgpr80
-; GFX9-NEXT: ; implicit-def: $sgpr70
-; GFX9-NEXT: ; implicit-def: $sgpr69
-; GFX9-NEXT: ; implicit-def: $sgpr68
-; GFX9-NEXT: ; implicit-def: $sgpr66
-; GFX9-NEXT: ; implicit-def: $sgpr67
-; GFX9-NEXT: ; implicit-def: $sgpr65
-; GFX9-NEXT: ; implicit-def: $sgpr64
-; GFX9-NEXT: ; implicit-def: $sgpr55
-; GFX9-NEXT: ; implicit-def: $sgpr53
-; GFX9-NEXT: ; implicit-def: $sgpr54
-; GFX9-NEXT: ; implicit-def: $sgpr52
-; GFX9-NEXT: ; implicit-def: $sgpr51
-; GFX9-NEXT: ; implicit-def: $sgpr50
-; GFX9-NEXT: ; implicit-def: $sgpr48
-; GFX9-NEXT: ; implicit-def: $sgpr49
-; GFX9-NEXT: ; implicit-def: $sgpr39
-; GFX9-NEXT: ; implicit-def: $sgpr38
-; GFX9-NEXT: ; implicit-def: $sgpr99
-; GFX9-NEXT: ; implicit-def: $sgpr97
-; GFX9-NEXT: ; implicit-def: $sgpr98
-; GFX9-NEXT: ; implicit-def: $sgpr96
-; GFX9-NEXT: ; implicit-def: $sgpr87
-; GFX9-NEXT: ; implicit-def: $sgpr86
-; GFX9-NEXT: ; implicit-def: $sgpr84
-; GFX9-NEXT: ; implicit-def: $sgpr85
-; GFX9-NEXT: ; implicit-def: $sgpr83
-; GFX9-NEXT: ; implicit-def: $sgpr82
-; GFX9-NEXT: ; implicit-def: $sgpr76
-; GFX9-NEXT: ; implicit-def: $sgpr74
-; GFX9-NEXT: ; implicit-def: $sgpr72
-; GFX9-NEXT: ; implicit-def: $sgpr62
-; GFX9-NEXT: ; implicit-def: $sgpr60
-; GFX9-NEXT: ; implicit-def: $sgpr58
-; GFX9-NEXT: ; implicit-def: $sgpr56
-; GFX9-NEXT: ; implicit-def: $sgpr36
-; GFX9-NEXT: ; implicit-def: $sgpr34
-; GFX9-NEXT: ; implicit-def: $sgpr30
-; GFX9-NEXT: ; implicit-def: $sgpr94
-; GFX9-NEXT: ; implicit-def: $sgpr92
-; GFX9-NEXT: ; implicit-def: $sgpr90
-; GFX9-NEXT: ; implicit-def: $sgpr88
-; GFX9-NEXT: ; implicit-def: $sgpr78
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: ; kill: killed $sgpr46
-; GFX9-NEXT: ; implicit-def: $sgpr46
-; GFX9-NEXT: s_branch .LBB99_2
-; GFX9-NEXT: .LBB99_4:
-; GFX9-NEXT: v_mov_b32_e32 v15, s71
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s80
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s70
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s69
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s68
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s66
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s67
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s65
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s64
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s55
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s53
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s54
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s52
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s51
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s50
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s48
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s49
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s39
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s38
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s99
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s97
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s98
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s96
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s87
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s86
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s84
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s85
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s83
-; GFX9-NEXT: v_mov_b32_e32 v25, s4
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s82
-; GFX9-NEXT: v_readlane_b32 s4, v62, 0
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 1
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 2
-; GFX9-NEXT: v_mov_b32_e32 v19, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 3
-; GFX9-NEXT: v_mov_b32_e32 v55, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 4
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 5
-; GFX9-NEXT: v_mov_b32_e32 v53, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 6
-; GFX9-NEXT: v_mov_b32_e32 v52, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 7
-; GFX9-NEXT: v_mov_b32_e32 v51, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 8
-; GFX9-NEXT: v_mov_b32_e32 v50, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 9
-; GFX9-NEXT: v_mov_b32_e32 v24, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 10
-; GFX9-NEXT: v_mov_b32_e32 v20, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 11
-; GFX9-NEXT: v_mov_b32_e32 v42, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 12
-; GFX9-NEXT: v_mov_b32_e32 v18, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 13
-; GFX9-NEXT: v_mov_b32_e32 v39, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 14
-; GFX9-NEXT: v_mov_b32_e32 v43, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 15
-; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v15, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 16
-; GFX9-NEXT: v_mov_b32_e32 v40, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 17
-; GFX9-NEXT: v_mov_b32_e32 v17, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 18
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v17, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 19
-; GFX9-NEXT: v_mov_b32_e32 v54, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 20
-; GFX9-NEXT: v_mov_b32_e32 v61, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 21
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v17, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 22
-; GFX9-NEXT: v_mov_b32_e32 v60, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 23
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v17, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 24
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v17, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 25
-; GFX9-NEXT: v_mov_b32_e32 v23, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 26
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v17, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 27
-; GFX9-NEXT: v_mov_b32_e32 v59, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 28
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v17, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 29
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v17, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 30
-; GFX9-NEXT: v_mov_b32_e32 v58, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 31
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v17, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 32
-; GFX9-NEXT: v_mov_b32_e32 v57, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 33
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v17, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 34
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v17, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 35
-; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v17, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 36
-; GFX9-NEXT: v_mov_b32_e32 v41, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 37
-; GFX9-NEXT: v_mov_b32_e32 v56, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 38
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v41, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 39
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v41, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 40
-; GFX9-NEXT: v_mov_b32_e32 v47, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 41
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v41, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 42
-; GFX9-NEXT: v_mov_b32_e32 v46, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 43
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v41, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 44
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v41, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 45
-; GFX9-NEXT: v_mov_b32_e32 v45, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 46
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v41, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 47
-; GFX9-NEXT: v_mov_b32_e32 v44, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 48
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v41, s4
-; GFX9-NEXT: v_readlane_b32 s4, v62, 49
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v41, s4
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v41, s46
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v41, s56
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v41, s58
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v41, s60
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v41, s62
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v41, s72
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v41, s74
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v41, s76
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v41, s78
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v41, s88
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v41, s90
-; 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_mov_b32_e32 v41, s92
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v41, s94
-; 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_mov_b32_e32 v41, s30
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v41, s34
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v41, s36
-; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GFX9-NEXT: v_mov_b32_e32 v21, s44
-; GFX9-NEXT: v_mov_b32_e32 v22, s45
-; GFX9-NEXT: v_mov_b32_e32 v13, s42
-; GFX9-NEXT: v_mov_b32_e32 v14, s43
-; GFX9-NEXT: v_mov_b32_e32 v11, s40
-; GFX9-NEXT: v_mov_b32_e32 v12, s41
-; GFX9-NEXT: v_mov_b32_e32 v9, s14
-; GFX9-NEXT: v_mov_b32_e32 v10, s15
-; GFX9-NEXT: v_mov_b32_e32 v7, s12
-; GFX9-NEXT: v_mov_b32_e32 v8, s13
-; GFX9-NEXT: v_mov_b32_e32 v5, s10
-; GFX9-NEXT: v_mov_b32_e32 v6, s11
-; GFX9-NEXT: v_mov_b32_e32 v3, s8
-; GFX9-NEXT: v_mov_b32_e32 v4, s9
-; GFX9-NEXT: v_mov_b32_e32 v1, s6
-; GFX9-NEXT: v_mov_b32_e32 v2, s7
-; GFX9-NEXT: v_mov_b32_e32 v48, s16
-; GFX9-NEXT: v_mov_b32_e32 v49, s17
-; GFX9-NEXT: v_mov_b32_e32 v37, s18
-; GFX9-NEXT: v_mov_b32_e32 v38, s19
-; GFX9-NEXT: v_mov_b32_e32 v35, s20
-; GFX9-NEXT: v_mov_b32_e32 v36, s21
-; GFX9-NEXT: v_mov_b32_e32 v33, s22
-; GFX9-NEXT: v_mov_b32_e32 v34, s23
-; GFX9-NEXT: v_mov_b32_e32 v31, s24
-; GFX9-NEXT: v_mov_b32_e32 v32, s25
-; GFX9-NEXT: v_mov_b32_e32 v29, s26
-; GFX9-NEXT: v_mov_b32_e32 v30, s27
-; GFX9-NEXT: v_mov_b32_e32 v27, s28
-; GFX9-NEXT: v_mov_b32_e32 v28, s29
-; GFX9-NEXT: v_mov_b32_e32 v26, s5
-; GFX9-NEXT: v_mov_b32_e32 v41, v50
-; GFX9-NEXT: v_mov_b32_e32 v50, v51
-; GFX9-NEXT: v_mov_b32_e32 v51, v52
-; GFX9-NEXT: v_mov_b32_e32 v52, v53
-; GFX9-NEXT: v_mov_b32_e32 v53, v55
-; GFX9-NEXT: v_mov_b32_e32 v55, v19
-; GFX9-NEXT: v_mov_b32_e32 v19, s81
-; GFX9-NEXT: .LBB99_5: ; %end
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 8, v15
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b32_e32 v19, 8, v19
-; GFX9-NEXT: v_or_b32_sdwa v16, v37, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b32_e32 v37, 8, v61
-; GFX9-NEXT: v_or_b32_sdwa v21, v21, 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:368 ; 4-byte Folded Reload
-; GFX9-NEXT: v_or_b32_sdwa v35, v35, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b32_e32 v37, 8, v60
-; GFX9-NEXT: v_or_b32_sdwa v36, v36, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b32_e32 v20, 8, v20
-; GFX9-NEXT: v_or_b32_sdwa v20, v48, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b32_e32 v18, 8, v18
-; GFX9-NEXT: v_or_b32_sdwa v18, v49, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b32_e32 v23, 8, v23
-; GFX9-NEXT: v_or_b32_sdwa v23, v33, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b32_e32 v33, 8, v59
-; GFX9-NEXT: v_or_b32_sdwa v33, v34, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b32_e32 v34, 8, v58
-; GFX9-NEXT: v_or_b32_sdwa v31, v31, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b32_e32 v34, 8, v57
-; GFX9-NEXT: v_or_b32_sdwa v32, v32, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b32_e32 v17, 8, v17
-; GFX9-NEXT: v_or_b32_sdwa v17, v29, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b32_e32 v29, 8, v56
-; GFX9-NEXT: v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b32_e32 v30, 8, v47
-; GFX9-NEXT: v_or_b32_sdwa v27, v27, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b32_e32 v30, 8, v46
-; GFX9-NEXT: v_or_b32_sdwa v28, v28, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b32_e32 v30, 8, v45
-; GFX9-NEXT: v_or_b32_sdwa v25, v25, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_lshlrev_b32_e32 v30, 8, v44
-; GFX9-NEXT: v_or_b32_sdwa v26, v26, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_readlane_b32 s99, v63, 35
-; GFX9-NEXT: v_readlane_b32 s98, v63, 34
-; GFX9-NEXT: v_readlane_b32 s97, v63, 33
-; GFX9-NEXT: v_readlane_b32 s96, v63, 32
-; GFX9-NEXT: v_readlane_b32 s87, v63, 31
-; GFX9-NEXT: v_readlane_b32 s86, v63, 30
-; GFX9-NEXT: v_readlane_b32 s85, v63, 29
-; GFX9-NEXT: v_readlane_b32 s84, v63, 28
-; GFX9-NEXT: v_readlane_b32 s83, v63, 27
-; GFX9-NEXT: v_readlane_b32 s82, v63, 26
-; GFX9-NEXT: v_readlane_b32 s81, v63, 25
-; GFX9-NEXT: v_readlane_b32 s80, v63, 24
-; GFX9-NEXT: v_readlane_b32 s71, v63, 23
-; GFX9-NEXT: v_readlane_b32 s70, v63, 22
-; GFX9-NEXT: v_readlane_b32 s69, v63, 21
-; GFX9-NEXT: v_readlane_b32 s68, v63, 20
-; GFX9-NEXT: v_readlane_b32 s67, v63, 19
-; GFX9-NEXT: v_readlane_b32 s66, v63, 18
-; GFX9-NEXT: v_readlane_b32 s65, v63, 17
-; GFX9-NEXT: v_readlane_b32 s64, v63, 16
-; GFX9-NEXT: v_readlane_b32 s55, v63, 15
-; GFX9-NEXT: v_readlane_b32 s54, v63, 14
-; GFX9-NEXT: v_readlane_b32 s53, v63, 13
-; GFX9-NEXT: v_readlane_b32 s52, v63, 12
-; GFX9-NEXT: v_readlane_b32 s51, v63, 11
-; GFX9-NEXT: v_readlane_b32 s50, v63, 10
-; GFX9-NEXT: v_readlane_b32 s49, v63, 9
-; GFX9-NEXT: v_readlane_b32 s48, v63, 8
-; GFX9-NEXT: v_readlane_b32 s39, v63, 7
-; GFX9-NEXT: v_readlane_b32 s38, v63, 6
-; GFX9-NEXT: v_readlane_b32 s37, v63, 5
-; GFX9-NEXT: v_readlane_b32 s36, v63, 4
-; GFX9-NEXT: v_readlane_b32 s35, v63, 3
-; GFX9-NEXT: v_readlane_b32 s34, v63, 2
-; GFX9-NEXT: v_readlane_b32 s31, v63, 1
-; GFX9-NEXT: v_readlane_b32 s30, v63, 0
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
-; GFX9-NEXT: v_or_b32_sdwa v15, v38, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b32_e32 v19, 8, v19
-; GFX9-NEXT: v_or_b32_sdwa v22, v22, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b32_e32 v19, 8, v37
-; GFX9-NEXT: v_or_b32_sdwa v19, v42, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v19, v0, s[0:3], 0 offen
-; GFX9-NEXT: v_lshlrev_b32_e32 v19, 8, v43
-; GFX9-NEXT: v_or_b32_sdwa v19, v39, v19 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_store_dword v18, v0, s[0:3], 0 offen offset:4
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b32_e32 v18, 8, v18
-; GFX9-NEXT: v_or_b32_sdwa v18, v40, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v16, v16, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v16, v0, s[0:3], 0 offen offset:8
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 8, v54
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v16, v18, v16 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_store_dword v15, v0, s[0:3], 0 offen offset:12
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v35, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:16
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v36, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:20
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v23, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:24
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v33, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:28
-; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v18
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v31, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:32
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v32, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:36
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v17, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:40
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v29, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:44
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v27, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:48
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v28, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:52
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v25, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:56
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v26, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:60
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v21, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:64
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v15, v22, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:68
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GFX9-NEXT: s_nop 0
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b32_e32 v16, 8, v16
-; GFX9-NEXT: v_or_b32_sdwa v13, v13, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v13, v13, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v13, v0, s[0:3], 0 offen offset:72
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b32_e32 v13, 8, v13
-; GFX9-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b32_e32 v14, 8, v14
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v14, v15, v14 dst_sel:WORD_1 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: buffer_store_dword v13, v0, s[0:3], 0 offen offset:76
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b32_e32 v13, 8, v13
-; GFX9-NEXT: v_or_b32_sdwa v11, v11, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b32_e32 v13, 8, v13
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v11, v11, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v11, v0, s[0:3], 0 offen offset:80
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b32_e32 v11, 8, v11
-; GFX9-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b32_e32 v12, 8, v12
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:WORD_1 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_store_dword v11, v0, s[0:3], 0 offen offset:84
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b32_e32 v11, 8, v11
-; GFX9-NEXT: v_or_b32_sdwa v9, v9, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b32_e32 v11, 8, v11
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v9, v9, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v9, v0, s[0:3], 0 offen offset:88
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b32_e32 v9, 8, v9
-; GFX9-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b32_e32 v10, 8, v10
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v10, v11, v10 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: buffer_store_dword v9, v0, s[0:3], 0 offen offset:92
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b32_e32 v9, 8, v9
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b32_e32 v9, 8, v9
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v7, v7, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v7, v0, s[0:3], 0 offen offset:96
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b32_e32 v7, 8, v7
-; GFX9-NEXT: v_or_b32_sdwa v7, v8, v7 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:140 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b32_e32 v8, 8, v8
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v8, v9, v8 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: buffer_store_dword v7, v0, s[0:3], 0 offen offset:100
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b32_e32 v7, 8, v7
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b32_e32 v7, 8, v7
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v7, v8, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v5, v5, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v5, v0, s[0:3], 0 offen offset:104
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b32_e32 v5, 8, v5
-; GFX9-NEXT: v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b32_e32 v6, 8, v6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v6, v7, v6 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: buffer_store_dword v5, v0, s[0:3], 0 offen offset:108
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b32_e32 v5, 8, v5
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(2)
-; GFX9-NEXT: v_lshlrev_b32_e32 v5, 8, v5
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_or_b32_sdwa v5, v6, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v3, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_store_dword v3, v0, s[0:3], 0 offen offset:112
-; GFX9-NEXT: v_lshlrev_b32_e32 v3, 8, v55
-; GFX9-NEXT: v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_lshlrev_b32_e32 v4, 8, v4
-; GFX9-NEXT: v_or_b32_sdwa v4, v53, v4 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: buffer_store_dword v3, v0, s[0:3], 0 offen offset:116
-; GFX9-NEXT: v_lshlrev_b32_e32 v3, 8, v52
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GFX9-NEXT: s_waitcnt vmcnt(1)
-; GFX9-NEXT: v_lshlrev_b32_e32 v3, 8, v3
-; GFX9-NEXT: v_or_b32_sdwa v3, v51, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT: v_or_b32_sdwa v1, v1, v3 dst_sel:DWORD 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: v_lshlrev_b32_e32 v1, 8, v50
-; 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_b32_e32 v2, 8, v24
-; GFX9-NEXT: v_or_b32_sdwa v2, v41, 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 v61, off, s[0:3], s32 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GFX9-NEXT: s_or_saveexec_b64 s[4:5], -1
-; 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: s_mov_b64 exec, s[4:5]
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-TRUE16-LABEL: bitcast_v64i16_to_v128i8_scalar:
-; GFX11-TRUE16: ; %bb.0:
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT: s_or_saveexec_b32 s4, -1
-; GFX11-TRUE16-NEXT: s_clause 0x3
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v75, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v76, s32 offset:80
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v77, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v78, s32 offset:88
-; GFX11-TRUE16-NEXT: s_mov_b32 exec_lo, s4
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s30, 0
-; GFX11-TRUE16-NEXT: v_writelane_b32 v76, s96, 0
-; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v15
-; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s40, v1
-; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s41, v2
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s31, 1
-; GFX11-TRUE16-NEXT: v_writelane_b32 v76, s97, 1
-; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s14, v3
-; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s15, v4
-; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s12, v5
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s34, 2
-; GFX11-TRUE16-NEXT: v_writelane_b32 v76, s98, 2
-; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s13, v6
-; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s10, v7
-; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s11, v8
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s35, 3
-; GFX11-TRUE16-NEXT: v_writelane_b32 v76, s99, 3
-; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s8, v9
-; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s9, v10
-; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s6, v11
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s36, 4
-; GFX11-TRUE16-NEXT: v_writelane_b32 v76, s100, 4
-; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s7, v12
-; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s4, v13
-; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s5, v14
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s37, 5
-; GFX11-TRUE16-NEXT: v_writelane_b32 v76, s101, 5
-; GFX11-TRUE16-NEXT: s_mov_b32 vcc_hi, 0
-; GFX11-TRUE16-NEXT: s_and_b32 s42, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT: s_clause 0x12
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v40, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v41, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v42, s32 offset:64
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v43, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v44, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v45, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v46, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v47, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v56, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v57, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v58, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v59, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v60, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v61, s32 offset:20
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v62, s32 offset:16
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v63, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v72, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v73, s32 offset:4
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v74, s32
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s38, 6
-; GFX11-TRUE16-NEXT: v_writelane_b32 v76, s102, 6
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr78 : SGPR spill to VGPR lane
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr77 : SGPR spill to VGPR lane
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s39, 7
-; GFX11-TRUE16-NEXT: v_writelane_b32 v76, s103, 7
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s48, 8
-; GFX11-TRUE16-NEXT: v_writelane_b32 v76, s104, 8
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s49, 9
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s50, 10
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s51, 11
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s52, 12
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s53, 13
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s54, 14
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s55, 15
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s64, 16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s65, 17
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s66, 18
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s67, 19
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s68, 20
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s69, 21
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s70, 22
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s71, 23
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s80, 24
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s81, 25
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s82, 26
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s83, 27
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s84, 28
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s85, 29
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s86, 30
-; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s87, 31
-; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB99_3
-; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s27, 16
-; GFX11-TRUE16-NEXT: s_lshr_b64 s[74:75], s[0:1], 24
-; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 8
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s27, 8
-; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s27, 24
-; GFX11-TRUE16-NEXT: s_lshr_b32 s35, s5, 24
-; GFX11-TRUE16-NEXT: s_lshr_b32 s36, s5, 16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 7
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s26, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s37, s5, 8
-; GFX11-TRUE16-NEXT: s_lshr_b32 s38, s4, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s39, s4, 8
-; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 6
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s26, 8
-; GFX11-TRUE16-NEXT: s_lshr_b32 s48, s7, 24
-; GFX11-TRUE16-NEXT: s_lshr_b32 s49, s7, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s50, s7, 8
-; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 5
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s25, 24
-; GFX11-TRUE16-NEXT: s_lshr_b32 s51, s6, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s52, s6, 8
-; GFX11-TRUE16-NEXT: s_lshr_b32 s53, s9, 24
-; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 4
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s25, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s54, s9, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s55, s9, 8
-; GFX11-TRUE16-NEXT: s_lshr_b32 s64, s8, 16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 3
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s25, 8
-; GFX11-TRUE16-NEXT: s_lshr_b32 s65, s8, 8
-; GFX11-TRUE16-NEXT: s_lshr_b32 s66, s11, 24
-; GFX11-TRUE16-NEXT: s_lshr_b32 s67, s11, 16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 2
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s24, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s68, s11, 8
-; GFX11-TRUE16-NEXT: s_lshr_b32 s69, s10, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s70, s10, 8
-; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 1
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s24, 8
-; GFX11-TRUE16-NEXT: s_lshr_b32 s71, s13, 24
-; GFX11-TRUE16-NEXT: s_lshr_b32 s80, s13, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s81, s13, 8
-; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 0
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s23, 24
-; GFX11-TRUE16-NEXT: s_lshr_b32 s82, s12, 16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 31
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s23, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s83, s12, 8
-; GFX11-TRUE16-NEXT: s_lshr_b32 s84, s15, 24
-; GFX11-TRUE16-NEXT: s_lshr_b32 s85, s15, 16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 30
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s23, 8
-; GFX11-TRUE16-NEXT: s_lshr_b32 s86, s15, 8
-; GFX11-TRUE16-NEXT: s_lshr_b32 s87, s14, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s96, s14, 8
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 29
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s22, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s97, s41, 24
-; GFX11-TRUE16-NEXT: s_lshr_b32 s98, s41, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s99, s41, 8
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 28
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s22, 8
-; GFX11-TRUE16-NEXT: s_lshr_b32 s100, s40, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s101, s40, 8
-; GFX11-TRUE16-NEXT: s_lshr_b32 s102, s29, 24
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 27
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s21, 24
-; GFX11-TRUE16-NEXT: s_lshr_b32 s103, s29, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s104, s29, 8
-; GFX11-TRUE16-NEXT: s_lshr_b32 s34, s28, 16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 26
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s21, 16
-; GFX11-TRUE16-NEXT: s_lshr_b64 s[62:63], s[26:27], 24
-; GFX11-TRUE16-NEXT: s_lshr_b64 s[72:73], s[24:25], 24
-; GFX11-TRUE16-NEXT: s_lshr_b64 s[60:61], s[22:23], 24
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 25
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s21, 8
-; GFX11-TRUE16-NEXT: s_lshr_b64 s[58:59], s[20:21], 24
-; GFX11-TRUE16-NEXT: s_lshr_b64 s[56:57], s[18:19], 24
-; GFX11-TRUE16-NEXT: s_lshr_b64 s[46:47], s[16:17], 24
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 24
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s20, 16
-; GFX11-TRUE16-NEXT: s_lshr_b64 s[44:45], s[2:3], 24
-; GFX11-TRUE16-NEXT: s_lshr_b64 s[76:77], s[6:7], 24
-; GFX11-TRUE16-NEXT: s_lshr_b64 s[78:79], s[8:9], 24
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 23
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s20, 8
-; GFX11-TRUE16-NEXT: s_lshr_b64 s[88:89], s[10:11], 24
-; GFX11-TRUE16-NEXT: s_lshr_b64 s[90:91], s[12:13], 24
-; GFX11-TRUE16-NEXT: s_lshr_b64 s[92:93], s[14:15], 24
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 22
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s19, 24
-; GFX11-TRUE16-NEXT: s_lshr_b64 s[94:95], s[40:41], 24
-; GFX11-TRUE16-NEXT: s_lshr_b64 s[30:31], s[28:29], 24
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 21
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s19, 16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 20
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s19, 8
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 19
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s18, 16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 18
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s18, 8
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 17
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s17, 24
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s17, 16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 15
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s17, 8
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 14
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s16, 16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 13
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s16, 8
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 12
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s3, 24
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 11
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 10
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s3, 8
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 9
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s2, 16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 8
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s2, 8
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 7
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s1, 24
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 6
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s1, 16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 5
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s1, 8
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 4
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s0, 16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 1
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s0, 8
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 0
-; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s28, 8
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s74, 2
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s75, 3
-; GFX11-TRUE16-NEXT: s_lshr_b64 s[74:75], s[4:5], 24
-; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, vcc_hi
-; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB99_4
-; GFX11-TRUE16-NEXT: .LBB99_2: ; %cmp.true
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, s7, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, s6, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, s9, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, s8, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v51, s3, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v50, s2, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v36, s19, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v4, s21, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v28, s27, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v27, s26, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v3, s20, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v35, s18, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v22, s29, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v21, s28, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v18, s41, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, s40, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, s15, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, s14, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, s13, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, s12, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, s11, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, s10, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, s5, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, s4, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v55, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v54, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v49, s17, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v48, s16, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, s23, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v32, s25, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v31, s24, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v1, s22, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[23:24], 24, v[7:8]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[37:38], 24, v[27:28]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[3:4]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[35:36]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[80:81], 24, v[50:51]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[9:10]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[31:32]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[52:53], 24, v[1:2]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[70:71], 24, v[48:49]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[81:82], 24, v[54:55]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[19:20], 24, v[5:6]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[25:26], 24, v[11:12]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[29:30], 24, v[13:14]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[33:34], 24, v[15:16]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[17:18]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[21:22]
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 24, v28
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 16, v28
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v28
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 16, v27
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v27
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 16, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 16, v31
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v164, 8, v31
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v165, 24, v2
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v166, 16, v2
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v167, 8, v2
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v176, 16, v1
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v177, 8, v1
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v178, 24, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v179, 16, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v180, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v181, 16, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v182, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v183, 24, v36
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v40, 16, v36
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v41, 8, v36
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v42, 16, v35
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v43, 8, v35
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v44, 24, v49
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v45, 16, v49
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v46, 8, v49
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v47, 16, v48
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v56, 8, v48
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v57, 24, v51
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v58, 16, v51
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v59, 8, v51
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v60, 16, v50
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v61, 8, v50
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v62, 24, v55
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v63, 16, v55
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v72, 8, v55
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v73, 16, v54
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v74, 8, v54
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 24, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 16, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v5
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 24, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 16, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 16, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 16, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 16, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 24, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 16, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 16, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 8, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 24, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 16, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 16, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 16, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 16, v15
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v15
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 16, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 16, v17
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v17
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v22
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 16, v22
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 8, v22
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 16, v21
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v21
-; GFX11-TRUE16-NEXT: s_branch .LBB99_5
-; GFX11-TRUE16-NEXT: .LBB99_3:
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 0
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr34_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr104_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr103_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr102_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr101_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr100_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr99_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr98_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr97_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr96_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr87_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr86_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr85_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr84_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr83_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr82_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr81_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr80_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr71_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr70_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr69_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr68_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr67_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr66_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr65_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr64_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr55_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr54_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr53_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr52_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr51_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr50_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr49_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr48_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr39_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr38_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr37_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr36_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr35_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr30_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr94_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr92_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr90_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr88_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr78_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 1
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 2
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s43, 3
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 4
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 5
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 6
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 7
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 8
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 9
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 10
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 11
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 12
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 13
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 14
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 15
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 16
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 17
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 18
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 19
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 20
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 21
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 22
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 23
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 24
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 25
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 26
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 27
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 28
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 29
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 30
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 31
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 0
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 1
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 2
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 3
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 4
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 5
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 6
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 7
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 8
-; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT: s_branch .LBB99_2
-; GFX11-TRUE16-NEXT: .LBB99_4:
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 0
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, s1
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, s28
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, s29
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, s40
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v74.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 1
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, s41
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, s14
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, s15
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, s12
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v73.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 4
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, s13
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, s10
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, s11
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, s8
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v72.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 5
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, s9
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, s6
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, s7
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, s4
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v63.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 6
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, s5
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, s2
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, s3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, s16
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v62.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 7
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, s17
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, s18
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, s19
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, s20
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v61.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 8
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, s21
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, s22
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, s23
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, s24
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v60.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 9
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, s25
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, s26
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, s27
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v150.l, s42
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v59.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 10
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.l, s34
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.l, s104
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v135.l, s103
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.l, s102
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v58.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 11
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.l, s101
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v132.l, s100
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, s99
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.l, s98
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v57.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 12
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.l, s97
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.l, s96
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.l, s87
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v118.l, s86
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v56.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 13
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v117.l, s85
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v116.l, s84
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v115.l, s83
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v114.l, s82
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v47.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 14
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v113.l, s81
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v112.l, s80
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v103.l, s71
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v102.l, s70
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v46.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 15
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v101.l, s69
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v100.l, s68
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v99.l, s67
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v98.l, s66
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v45.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 16
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v97.l, s65
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v96.l, s64
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v87.l, s55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v86.l, s54
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v44.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 17
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v85.l, s53
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v84.l, s52
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v83.l, s51
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v82.l, s50
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v43.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 18
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v71.l, s49
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, s48
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, s39
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, s38
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v42.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 19
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, s37
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, s36
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, s35
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v145.l, s43
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v41.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 20
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, s62
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, s72
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, s60
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, s58
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v40.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 21
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, s56
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v70.l, s46
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v80.l, s44
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, s74
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v183.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 22
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, s76
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, s78
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, s88
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, s90
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v182.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 23
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, s92
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, s94
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v68.l, s30
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v181.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 24
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v180.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 25
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v179.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 26
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v178.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 27
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v177.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 28
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v176.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 29
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v167.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 30
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v166.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 31
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v165.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v77, 0
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v164.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v77, 1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v163.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v77, 2
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v162.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v77, 3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v161.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v77, 4
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v160.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v77, 5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v151.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v77, 6
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v148.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v77, 7
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v147.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v77, 8
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.l, s0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 2
-; GFX11-TRUE16-NEXT: v_readlane_b32 s1, v78, 3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v81.l, s0
-; GFX11-TRUE16-NEXT: .LBB99_5: ; %end
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v69, 8, v74
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xff, v54
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v66, 8, v81
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xff, v55
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 8, v62
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v80, 8, v80
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v54, v54, v69
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v69, 0xff, v73
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xff, v50
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v60, 0xff, v60
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xff, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v54
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v66, v69, v66
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v69, 8, v72
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v80, v60, v80
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v57, 8, v57
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xff, v48
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v66, 16, v66
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v55, v55, v69
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v69, 0xff, v63
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v70, 8, v70
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xff, v49
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v54, v54, v66
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v55
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v69, v69, v81
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 8, v61
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xff, v35
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v67, 8, v67
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xff, v36
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v66, 16, v69
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v50, v50, v81
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v69, 8, v59
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v81, 0xff, v58
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v64, 8, v64
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v55, v55, v66
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff, v50
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v66, 16, v80
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v51, v51, v69
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v69, v81, v57
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v80, 8, v56
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v81, 0xff, v47
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v56, v50, v66
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff, v51
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v69
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v48, v48, v80
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v66, v81, v70
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v69, 8, v46
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v70, 0xff, v45
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v80, 8, v44
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v57, v50, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff, v48
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v66
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v49, v49, v69
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v51, v70, v80
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v66, 8, v43
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v69, 0xff, v42
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v48, v48, v50
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v49
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v51
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v35, v35, v66
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v51, v69, v67
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v66, 8, v41
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v67, 0xff, v40
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v69, 8, v183
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v49, v49, v50
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v51
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v36, v36, v66
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v35
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v51, v67, v69
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v67, 0xff, v181
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v36
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v66, 8, v182
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v64, v67, v64
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v69, 8, v180
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v50, v35, v50
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v51, v36, v51
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v35, 16, v64
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xff, v179
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v64, 8, v178
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v3, v66
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v4, v69
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v66, 8, v177
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v67, 0xff, v176
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 8, v52
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v69, 8, v167
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v36, v36, v64
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v66
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v52, v67, v52
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v2, v69
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GFX11-TRUE16-NEXT: s_clause 0x1
-; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[54:57], off
-; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[48:51], off offset:16
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v66, 0xffff, v1
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v52
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v67, 0xffff, v2
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v35
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v36
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xff, v31
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v35, 8, v164
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xff, v163
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 8, v38
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v48, 8, v162
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xff, v161
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 8, v160
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xff, v27
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 8, v151
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v66, v52
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v31, v35
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v35, v36, v38
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v32, v48
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v36, v49, v50
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v27, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xff, v148
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v37, 8, v37
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v48, 8, v147
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xff, v146
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 8, v145
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 8, v150
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xff, v149
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 8, v68
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, v38, v37
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v28, v48
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, v49, v50
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v21, v51
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v48, v52, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v31
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v35, 16, v35
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v37, 16, v37
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 16, v38
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v48, 16, v48
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v36, 16, v36
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v35, v31, v35
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, v27, v37
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, v28, v38
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v48, v21, v48
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xff, v22
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 8, v144
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xff, v135
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 8, v134
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xff, v17
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 8, v133
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v36, v32, v36
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xff, v132
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v49, 8, v65
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xff, v18
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 8, v131
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v21, v22
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v27, v28
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v17, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xff, v130
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 8, v129
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v32, v49
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v18, v50
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 8, v128
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xff, v119
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v33, 8, v33
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v28, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xff, v16
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 8, v118
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xff, v117
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 8, v116
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v15, v32
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v49, v33
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v16, v50
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v51, v52
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v31
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v49, v21, v22
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v50, v17, v27
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v51, v18, v28
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 8, v115
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xff, v114
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 8, v29
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 8, v113
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v29, 8, v102
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v15, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xff, v112
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 8, v103
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v13, v17
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v18, v21
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v14, v22
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v11, v29
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xff, v101
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 8, v25
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v29, 8, v97
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xff, v96
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 8, v24
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v27, v28
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 8, v100
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xff, v99
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 8, v98
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v21, v22
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v9, v29
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v31, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v12, v25
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v27, v28
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v13, v17
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v14, v18
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v11, v21
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v9, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xff, v10
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 8, v87
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xff, v86
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 8, v85
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xff, v83
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 8, v23
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v70, 0xff, v166
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v80, 8, v165
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v12, v22
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 8, v84
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 8, v82
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v9, v10
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v14, v21
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v24, v23
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v64, v70, v80
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v7, v22
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v8, v25
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 8, v53
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v14
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xff, v71
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 8, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xff, v34
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 8, v19
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 8, v30
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xff, v26
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 8, v20
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v64, 16, v64
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v14, v22
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v5, v23
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v24, v19
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v6, v25
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v26, v20
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v67, v64
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v14
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v16, v32
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v9, v10
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v7, v21
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v8, v22
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v23, v19
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v24, v20
-; GFX11-TRUE16-NEXT: s_clause 0x5
-; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off offset:32
-; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[35:38], off offset:48
-; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[48:51], off offset:64
-; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[15:18], off offset:80
-; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[11:14], off offset:96
-; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:112
-; GFX11-TRUE16-NEXT: s_clause 0x12
-; GFX11-TRUE16-NEXT: scratch_load_b32 v74, off, s32
-; GFX11-TRUE16-NEXT: scratch_load_b32 v73, off, s32 offset:4
-; GFX11-TRUE16-NEXT: scratch_load_b32 v72, off, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_load_b32 v63, off, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_load_b32 v62, off, s32 offset:16
-; GFX11-TRUE16-NEXT: scratch_load_b32 v61, off, s32 offset:20
-; GFX11-TRUE16-NEXT: scratch_load_b32 v60, off, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_load_b32 v59, off, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_load_b32 v58, off, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_load_b32 v57, off, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_load_b32 v56, off, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_load_b32 v47, off, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_load_b32 v46, off, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_load_b32 v45, off, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_load_b32 v44, off, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_load_b32 v43, off, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_load_b32 v42, off, s32 offset:64
-; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_load_b32 v40, off, s32 offset:72
-; GFX11-TRUE16-NEXT: v_readlane_b32 s104, v76, 8
-; GFX11-TRUE16-NEXT: v_readlane_b32 s103, v76, 7
-; GFX11-TRUE16-NEXT: v_readlane_b32 s102, v76, 6
-; GFX11-TRUE16-NEXT: v_readlane_b32 s101, v76, 5
-; GFX11-TRUE16-NEXT: v_readlane_b32 s100, v76, 4
-; GFX11-TRUE16-NEXT: v_readlane_b32 s99, v76, 3
-; GFX11-TRUE16-NEXT: v_readlane_b32 s98, v76, 2
-; GFX11-TRUE16-NEXT: v_readlane_b32 s97, v76, 1
-; GFX11-TRUE16-NEXT: v_readlane_b32 s96, v76, 0
-; GFX11-TRUE16-NEXT: v_readlane_b32 s87, v75, 31
-; GFX11-TRUE16-NEXT: v_readlane_b32 s86, v75, 30
-; GFX11-TRUE16-NEXT: v_readlane_b32 s85, v75, 29
-; GFX11-TRUE16-NEXT: v_readlane_b32 s84, v75, 28
-; GFX11-TRUE16-NEXT: v_readlane_b32 s83, v75, 27
-; GFX11-TRUE16-NEXT: v_readlane_b32 s82, v75, 26
-; GFX11-TRUE16-NEXT: v_readlane_b32 s81, v75, 25
-; GFX11-TRUE16-NEXT: v_readlane_b32 s80, v75, 24
-; GFX11-TRUE16-NEXT: v_readlane_b32 s71, v75, 23
-; GFX11-TRUE16-NEXT: v_readlane_b32 s70, v75, 22
-; GFX11-TRUE16-NEXT: v_readlane_b32 s69, v75, 21
-; GFX11-TRUE16-NEXT: v_readlane_b32 s68, v75, 20
-; GFX11-TRUE16-NEXT: v_readlane_b32 s67, v75, 19
-; GFX11-TRUE16-NEXT: v_readlane_b32 s66, v75, 18
-; GFX11-TRUE16-NEXT: v_readlane_b32 s65, v75, 17
-; GFX11-TRUE16-NEXT: v_readlane_b32 s64, v75, 16
-; GFX11-TRUE16-NEXT: v_readlane_b32 s55, v75, 15
-; GFX11-TRUE16-NEXT: v_readlane_b32 s54, v75, 14
-; GFX11-TRUE16-NEXT: v_readlane_b32 s53, v75, 13
-; GFX11-TRUE16-NEXT: v_readlane_b32 s52, v75, 12
-; GFX11-TRUE16-NEXT: v_readlane_b32 s51, v75, 11
-; GFX11-TRUE16-NEXT: v_readlane_b32 s50, v75, 10
-; GFX11-TRUE16-NEXT: v_readlane_b32 s49, v75, 9
-; GFX11-TRUE16-NEXT: v_readlane_b32 s48, v75, 8
-; GFX11-TRUE16-NEXT: v_readlane_b32 s39, v75, 7
-; GFX11-TRUE16-NEXT: v_readlane_b32 s38, v75, 6
-; GFX11-TRUE16-NEXT: v_readlane_b32 s37, v75, 5
-; GFX11-TRUE16-NEXT: v_readlane_b32 s36, v75, 4
-; GFX11-TRUE16-NEXT: v_readlane_b32 s35, v75, 3
-; GFX11-TRUE16-NEXT: v_readlane_b32 s34, v75, 2
-; GFX11-TRUE16-NEXT: v_readlane_b32 s31, v75, 1
-; GFX11-TRUE16-NEXT: v_readlane_b32 s30, v75, 0
-; GFX11-TRUE16-NEXT: s_or_saveexec_b32 s0, -1
-; GFX11-TRUE16-NEXT: s_clause 0x3
-; GFX11-TRUE16-NEXT: scratch_load_b32 v75, off, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_load_b32 v76, off, s32 offset:80
-; GFX11-TRUE16-NEXT: scratch_load_b32 v77, off, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_load_b32 v78, off, s32 offset:88
-; GFX11-TRUE16-NEXT: s_mov_b32 exec_lo, s0
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v64i16_to_v128i8_scalar:
-; GFX11-FAKE16: ; %bb.0:
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT: s_or_saveexec_b32 s4, -1
-; GFX11-FAKE16-NEXT: s_clause 0x3
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v75, s32 offset:76
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v76, s32 offset:80
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v77, s32 offset:84
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v78, s32 offset:88
-; GFX11-FAKE16-NEXT: s_mov_b32 exec_lo, s4
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s30, 0
-; GFX11-FAKE16-NEXT: v_writelane_b32 v76, s96, 0
-; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v15
-; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s40, v1
-; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s41, v2
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s31, 1
-; GFX11-FAKE16-NEXT: v_writelane_b32 v76, s97, 1
-; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s14, v3
-; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s15, v4
-; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s12, v5
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s34, 2
-; GFX11-FAKE16-NEXT: v_writelane_b32 v76, s98, 2
-; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s13, v6
-; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s10, v7
-; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s11, v8
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s35, 3
-; GFX11-FAKE16-NEXT: v_writelane_b32 v76, s99, 3
-; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s8, v9
-; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s9, v10
-; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s6, v11
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s36, 4
-; GFX11-FAKE16-NEXT: v_writelane_b32 v76, s100, 4
-; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s7, v12
-; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s4, v13
-; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s5, v14
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s37, 5
-; GFX11-FAKE16-NEXT: v_writelane_b32 v76, s101, 5
-; GFX11-FAKE16-NEXT: s_mov_b32 s99, 0
-; GFX11-FAKE16-NEXT: s_and_b32 s42, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT: s_clause 0x12
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v40, s32 offset:72
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v41, s32 offset:68
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v42, s32 offset:64
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v43, s32 offset:60
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v44, s32 offset:56
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v45, s32 offset:52
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v46, s32 offset:48
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v47, s32 offset:44
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v56, s32 offset:40
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v57, s32 offset:36
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v58, s32 offset:32
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v59, s32 offset:28
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v60, s32 offset:24
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v61, s32 offset:20
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v62, s32 offset:16
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v63, s32 offset:12
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v72, s32 offset:8
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v73, s32 offset:4
-; GFX11-FAKE16-NEXT: scratch_store_b32 off, v74, s32
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s38, 6
-; GFX11-FAKE16-NEXT: v_writelane_b32 v76, s102, 6
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr78 : SGPR spill to VGPR lane
-; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr77 : SGPR spill to VGPR lane
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s39, 7
-; GFX11-FAKE16-NEXT: v_writelane_b32 v76, s103, 7
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s48, 8
-; GFX11-FAKE16-NEXT: v_writelane_b32 v76, s104, 8
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s49, 9
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s50, 10
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s51, 11
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s52, 12
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s53, 13
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s54, 14
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s55, 15
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s64, 16
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s65, 17
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s66, 18
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s67, 19
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s68, 20
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s69, 21
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s70, 22
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s71, 23
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s80, 24
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s81, 25
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s82, 26
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s83, 27
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s84, 28
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s85, 29
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s86, 30
-; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s87, 31
-; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB99_3
-; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s27, 16
-; GFX11-FAKE16-NEXT: s_lshr_b64 s[74:75], s[0:1], 24
-; GFX11-FAKE16-NEXT: v_writelane_b32 v77, s42, 8
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s27, 8
-; GFX11-FAKE16-NEXT: s_lshr_b32 s43, s27, 24
-; GFX11-FAKE16-NEXT: s_lshr_b32 s34, s5, 24
-; GFX11-FAKE16-NEXT: s_lshr_b32 s35, s5, 16
-; GFX11-FAKE16-NEXT: v_writelane_b32 v77, s42, 7
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s26, 16
-; GFX11-FAKE16-NEXT: s_lshr_b32 s37, s5, 8
-; GFX11-FAKE16-NEXT: s_lshr_b32 s36, s4, 16
-; GFX11-FAKE16-NEXT: s_lshr_b32 s38, s4, 8
-; GFX11-FAKE16-NEXT: v_writelane_b32 v77, s42, 6
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s26, 8
-; GFX11-FAKE16-NEXT: s_lshr_b32 s39, s7, 24
-; GFX11-FAKE16-NEXT: s_lshr_b32 s48, s7, 16
-; GFX11-FAKE16-NEXT: s_lshr_b32 s50, s7, 8
-; GFX11-FAKE16-NEXT: v_writelane_b32 v77, s42, 5
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s25, 24
-; GFX11-FAKE16-NEXT: s_lshr_b32 s49, s6, 16
-; GFX11-FAKE16-NEXT: s_lshr_b32 s51, s6, 8
-; GFX11-FAKE16-NEXT: s_lshr_b32 s52, s9, 24
-; GFX11-FAKE16-NEXT: v_writelane_b32 v77, s42, 4
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s25, 16
-; GFX11-FAKE16-NEXT: s_lshr_b32 s53, s9, 16
-; GFX11-FAKE16-NEXT: s_lshr_b32 s55, s9, 8
-; GFX11-FAKE16-NEXT: s_lshr_b32 s54, s8, 16
-; GFX11-FAKE16-NEXT: v_writelane_b32 v77, s42, 3
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s25, 8
-; GFX11-FAKE16-NEXT: s_lshr_b32 s64, s8, 8
-; GFX11-FAKE16-NEXT: s_lshr_b32 s65, s11, 24
-; GFX11-FAKE16-NEXT: s_lshr_b32 s66, s11, 16
-; GFX11-FAKE16-NEXT: v_writelane_b32 v77, s42, 2
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s24, 16
-; GFX11-FAKE16-NEXT: s_lshr_b32 s68, s11, 8
-; GFX11-FAKE16-NEXT: s_lshr_b32 s67, s10, 16
-; GFX11-FAKE16-NEXT: s_lshr_b32 s69, s10, 8
-; GFX11-FAKE16-NEXT: v_writelane_b32 v77, s42, 1
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s24, 8
-; GFX11-FAKE16-NEXT: s_lshr_b32 s70, s13, 24
-; GFX11-FAKE16-NEXT: s_lshr_b32 s71, s13, 16
-; GFX11-FAKE16-NEXT: s_lshr_b32 s81, s13, 8
-; GFX11-FAKE16-NEXT: v_writelane_b32 v77, s42, 0
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s23, 24
-; GFX11-FAKE16-NEXT: s_lshr_b32 s80, s12, 16
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 31
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s23, 16
-; GFX11-FAKE16-NEXT: s_lshr_b32 s82, s12, 8
-; GFX11-FAKE16-NEXT: s_lshr_b32 s83, s15, 24
-; GFX11-FAKE16-NEXT: s_lshr_b32 s84, s15, 16
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 30
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s23, 8
-; GFX11-FAKE16-NEXT: s_lshr_b32 s86, s15, 8
-; GFX11-FAKE16-NEXT: s_lshr_b32 s85, s14, 16
-; GFX11-FAKE16-NEXT: s_lshr_b32 s87, s14, 8
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 29
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s22, 16
-; GFX11-FAKE16-NEXT: s_lshr_b32 s96, s41, 24
-; GFX11-FAKE16-NEXT: s_lshr_b32 s97, s41, 16
-; GFX11-FAKE16-NEXT: s_lshr_b32 s100, s41, 8
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 28
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s22, 8
-; GFX11-FAKE16-NEXT: s_lshr_b32 s98, s40, 16
-; GFX11-FAKE16-NEXT: s_lshr_b32 s101, s40, 8
-; GFX11-FAKE16-NEXT: s_lshr_b32 s102, s29, 24
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 27
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s21, 24
-; GFX11-FAKE16-NEXT: s_lshr_b32 s103, s29, 16
-; GFX11-FAKE16-NEXT: s_lshr_b32 vcc_hi, s29, 8
-; GFX11-FAKE16-NEXT: s_lshr_b32 s104, s28, 16
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 26
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s21, 16
-; GFX11-FAKE16-NEXT: s_lshr_b64 s[62:63], s[26:27], 24
-; GFX11-FAKE16-NEXT: s_lshr_b64 s[72:73], s[24:25], 24
-; GFX11-FAKE16-NEXT: s_lshr_b64 s[60:61], s[22:23], 24
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 25
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s21, 8
-; GFX11-FAKE16-NEXT: s_lshr_b64 s[58:59], s[20:21], 24
-; GFX11-FAKE16-NEXT: s_lshr_b64 s[56:57], s[18:19], 24
-; GFX11-FAKE16-NEXT: s_lshr_b64 s[46:47], s[16:17], 24
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 24
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s20, 16
-; GFX11-FAKE16-NEXT: s_lshr_b64 s[44:45], s[2:3], 24
-; GFX11-FAKE16-NEXT: s_lshr_b64 s[76:77], s[6:7], 24
-; GFX11-FAKE16-NEXT: s_lshr_b64 s[78:79], s[8:9], 24
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 23
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s20, 8
-; GFX11-FAKE16-NEXT: s_lshr_b64 s[88:89], s[10:11], 24
-; GFX11-FAKE16-NEXT: s_lshr_b64 s[90:91], s[12:13], 24
-; GFX11-FAKE16-NEXT: s_lshr_b64 s[92:93], s[14:15], 24
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 22
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s19, 24
-; GFX11-FAKE16-NEXT: s_lshr_b64 s[94:95], s[40:41], 24
-; GFX11-FAKE16-NEXT: s_lshr_b64 s[30:31], s[28:29], 24
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 21
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s19, 16
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 20
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s19, 8
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 19
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s18, 16
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 18
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s18, 8
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 17
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s17, 24
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 16
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s17, 16
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 15
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s17, 8
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 14
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s16, 16
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 13
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s16, 8
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 12
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s3, 24
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 11
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s3, 16
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 10
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s3, 8
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 9
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s2, 16
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 8
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s2, 8
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 7
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s1, 24
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 6
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s1, 16
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 5
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s1, 8
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 4
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s0, 16
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 3
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s0, 8
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 2
-; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s28, 8
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s74, 0
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s75, 1
-; GFX11-FAKE16-NEXT: s_lshr_b64 s[74:75], s[4:5], 24
-; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s99
-; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB99_4
-; GFX11-FAKE16-NEXT: .LBB99_2: ; %cmp.true
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v39, s17, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v38, s16, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v2, s5, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v1, s4, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v51, s3, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v50, s2, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v33, s21, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v29, s23, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v28, s22, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v32, s20, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v16, s29, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v15, s28, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v14, s41, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v13, s40, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v12, s15, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v11, s14, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v10, s13, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v9, s12, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v8, s11, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v7, s10, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v6, s9, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v5, s8, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v4, s7, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v3, s6, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v53, s1, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v52, s0, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v37, s19, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v25, s25, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v21, s27, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v20, s26, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v24, s24, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_pk_add_u16 v36, s18, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[80:81], 24, v[38:39]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[28:29]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[32:33]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[81:82], 24, v[50:51]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[17:18], 24, v[1:2]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[20:21]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[24:25]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[70:71], 24, v[36:37]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[82:83], 24, v[52:53]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[18:19], 24, v[3:4]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[26:27], 24, v[7:8]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[30:31], 24, v[9:10]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[34:35], 24, v[11:12]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[13:14]
-; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[15:16]
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v147, 24, v21
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v148, 16, v21
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v149, 8, v21
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v150, 16, v20
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v151, 8, v20
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v161, 24, v25
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v160, 16, v25
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v162, 8, v25
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v163, 16, v24
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v164, 8, v24
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v166, 24, v29
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v165, 16, v29
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v167, 8, v29
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v176, 16, v28
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v177, 8, v28
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v179, 24, v33
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v178, 16, v33
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v180, 8, v33
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v181, 16, v32
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v182, 8, v32
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v40, 24, v37
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v183, 16, v37
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v41, 8, v37
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v42, 16, v36
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v43, 8, v36
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v45, 24, v39
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v44, 16, v39
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v46, 8, v39
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v47, 16, v38
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v56, 8, v38
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v58, 24, v51
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v57, 16, v51
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v59, 8, v51
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v61, 16, v50
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v60, 8, v50
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v63, 24, v53
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v62, 16, v53
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v72, 8, v53
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v73, 16, v52
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v74, 8, v52
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v19, 24, v2
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v2
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 8, v2
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v1
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 8, v1
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 24, v4
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v4
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 8, v4
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v83, 16, v3
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v84, 8, v3
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v85, 24, v6
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v86, 16, v6
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v87, 8, v6
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v96, 16, v5
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v97, 8, v5
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v98, 24, v8
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v99, 16, v8
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v100, 8, v8
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v101, 16, v7
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v102, 8, v7
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v103, 24, v10
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v112, 16, v10
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v113, 8, v10
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v114, 16, v9
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v115, 8, v9
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v116, 24, v12
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v117, 16, v12
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v118, 8, v12
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v119, 16, v11
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v128, 8, v11
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v129, 24, v14
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v130, 16, v14
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v131, 8, v14
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v132, 16, v13
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v133, 8, v13
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v134, 24, v16
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v135, 16, v16
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v144, 8, v16
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v145, 16, v15
-; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v146, 8, v15
-; GFX11-FAKE16-NEXT: s_branch .LBB99_5
-; GFX11-FAKE16-NEXT: .LBB99_3:
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: s_mov_b32 s99, -1
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr104
-; GFX11-FAKE16-NEXT: ; implicit-def: $vcc_hi
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr103
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr102
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr101
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr98
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr100
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr97
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr96
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr87
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr85
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr86
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr84
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr83
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr82
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr80
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr81
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr71
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr70
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr69
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr67
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr68
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr66
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr65
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr64
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr54
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr55
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr53
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr52
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr51
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr49
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr50
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr48
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr39
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr38
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr36
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr37
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr35
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr34
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr30
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr94
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr92
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr90
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr88
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr78
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 0
-; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s43, 1
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT: s_branch .LBB99_2
-; GFX11-FAKE16-NEXT: .LBB99_4:
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v52, s0 :: v_dual_mov_b32 v53, s1
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 2
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v71, s50
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v15, s28 :: v_dual_mov_b32 v16, s29
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v13, s40 :: v_dual_mov_b32 v14, s41
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v74, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 3
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v11, s14 :: v_dual_mov_b32 v12, s15
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v9, s12 :: v_dual_mov_b32 v10, s13
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v73, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 4
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v55, s48
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v7, s10 :: v_dual_mov_b32 v8, s11
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v5, s8 :: v_dual_mov_b32 v6, s9
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v72, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 5
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v49, s39
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v3, s6 :: v_dual_mov_b32 v4, s7
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v1, s4 :: v_dual_mov_b32 v2, s5
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v62, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 6
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v50, s2 :: v_dual_mov_b32 v51, s3
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v38, s16 :: v_dual_mov_b32 v39, s17
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v63, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 7
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v35, s38 :: v_dual_mov_b32 v36, s18
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v37, s19 :: v_dual_mov_b32 v32, s20
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v33, s21 :: v_dual_mov_b32 v60, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 8
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v28, s22 :: v_dual_mov_b32 v29, s23
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v61, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 9
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v20, s26 :: v_dual_mov_b32 v21, s27
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v146, s42 :: v_dual_mov_b32 v145, s104
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v59, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 10
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v144, vcc_hi :: v_dual_mov_b32 v135, s103
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v134, s102 :: v_dual_mov_b32 v133, s101
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v57, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 11
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v31, s36 :: v_dual_mov_b32 v132, s98
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v131, s100 :: v_dual_mov_b32 v130, s97
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v129, s96 :: v_dual_mov_b32 v58, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 12
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v27, s37 :: v_dual_mov_b32 v128, s87
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v119, s85 :: v_dual_mov_b32 v118, s86
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v117, s84 :: v_dual_mov_b32 v56, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 13
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v116, s83 :: v_dual_mov_b32 v115, s82
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v114, s80 :: v_dual_mov_b32 v113, s81
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v47, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 14
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v23, s35 :: v_dual_mov_b32 v112, s71
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v103, s70 :: v_dual_mov_b32 v102, s69
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v101, s67 :: v_dual_mov_b32 v46, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 15
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v19, s34 :: v_dual_mov_b32 v100, s68
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v99, s66 :: v_dual_mov_b32 v98, s65
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v97, s64 :: v_dual_mov_b32 v44, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 16
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v96, s54 :: v_dual_mov_b32 v87, s55
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v86, s53 :: v_dual_mov_b32 v85, s52
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v45, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 17
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v84, s51 :: v_dual_mov_b32 v83, s49
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v147, s43 :: v_dual_mov_b32 v22, s78
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v43, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 18
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v67, s58 :: v_dual_mov_b32 v26, s88
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v81, s44 :: v_dual_mov_b32 v30, s90
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v42, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 19
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v17, s74 :: v_dual_mov_b32 v34, s92
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v65, s94 :: v_dual_mov_b32 v68, s30
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v41, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 20
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v48, s62
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v54, s72
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v64, s60
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT: v_dual_mov_b32 v70, s56 :: v_dual_mov_b32 v183, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 21
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v80, s46
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v18, s76
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v40, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 22
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v182, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 23
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v181, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 24
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v180, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 25
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v178, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 26
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v179, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 27
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v177, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 28
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v176, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 29
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v167, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 30
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v165, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 31
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v166, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v77, 0
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v164, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v77, 1
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v163, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v77, 2
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v162, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v77, 3
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v160, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v77, 4
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v161, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v77, 5
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v151, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v77, 6
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v150, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v77, 7
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v149, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v77, 8
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v148, s0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s1, v78, 1
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v82, s0
-; GFX11-FAKE16-NEXT: .LBB99_5: ; %end
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v69, 8, v74
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v52, 0xff, v52
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v66, 8, v82
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v53, 0xff, v53
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v82, 8, v63
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v81, 8, v81
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v52, v52, v69
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v69, 0xff, v73
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v50, 0xff, v50
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v57, 0xff, v57
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v58, 8, v58
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v52, 0xffff, v52
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v66, v69, v66
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v69, 8, v72
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v38, 0xff, v38
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v80, 8, v80
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v39, 0xff, v39
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v66, 16, v66
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v53, v53, v69
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v69, 0xff, v62
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v70, 8, v70
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v67, 8, v67
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v32, 0xff, v32
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xff, v33
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v69, v69, v82
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v82, 8, v60
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v60, 0xff, v61
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v28, 0xff, v28
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v64, 8, v64
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v29, 0xff, v29
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v82, v50, v82
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v81, v60, v81
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v50, v52, v66
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v52, 0xffff, v53
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v53, 16, v69
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v66, 0xff, v51
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v69, 8, v59
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v24, 0xff, v24
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v25, 0xff, v25
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v51, v52, v53
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v52, 0xffff, v82
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v53, 16, v81
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v66, v66, v69
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v69, v57, v58
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v81, 8, v56
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v82, 0xff, v47
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v52, v52, v53
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v53, 0xffff, v66
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v66, 16, v69
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v38, v38, v81
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v69, v82, v80
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v80, 8, v46
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v81, 0xff, v44
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v82, 8, v45
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v53, v53, v66
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v38, 0xffff, v38
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v66, 16, v69
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v39, v39, v80
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v69, v81, v82
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v80, 0xff, v36
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v81, 8, v43
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v82, 0xff, v42
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v36, v38, v66
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v38, 0xffff, v39
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v39, 16, v69
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v66, v80, v81
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v69, v82, v70
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v70, 0xff, v37
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v80, 8, v41
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v81, 0xff, v183
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v82, 8, v40
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v37, v38, v39
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v38, 0xffff, v66
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v39, 16, v69
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v66, v70, v80
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v69, v81, v82
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v80, 0xff, v181
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v70, 8, v182
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v81, 8, v180
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v66, 0xffff, v66
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v69, 16, v69
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v67, v80, v67
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v32, v32, v70
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v33, v33, v81
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v38, v38, v39
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v39, v66, v69
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v66, 16, v67
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v67, 0xff, v178
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v69, 8, v179
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v70, 8, v177
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v80, 0xff, v176
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v81, 8, v167
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v82, 0xff, v165
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v165, 8, v166
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v67, v67, v69
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v28, v28, v70
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v64, v80, v64
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v29, v29, v81
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v69, v82, v165
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v32, 0xffff, v32
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xffff, v33
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v67, 16, v67
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v64, 16, v64
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v69, 16, v69
-; GFX11-FAKE16-NEXT: s_clause 0x1
-; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[50:53], off
-; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[36:39], off offset:16
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v36, v32, v66
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v37, v33, v67
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v38, v28, v64
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v39, v29, v69
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v28, 8, v164
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v29, 0xff, v163
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v32, 8, v54
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v33, 8, v162
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v50, 0xff, v160
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v51, 8, v161
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xff, v20
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v52, 8, v151
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v24, v24, v28
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v28, v29, v32
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v25, v25, v33
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v29, v50, v51
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v20, v20, v52
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v32, 0xff, v150
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v33, 8, v48
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xff, v21
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v48, 8, v149
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v50, 0xff, v148
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v51, 8, v147
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xff, v15
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v52, 8, v146
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v53, 0xff, v145
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v54, 8, v68
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v32, v32, v33
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v21, v21, v48
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v33, v50, v51
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v15, v15, v52
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v48, v53, v54
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v33, 16, v33
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v48, 16, v48
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v29, 16, v29
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v50, v24, v28
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v52, v20, v32
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v53, v21, v33
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v64, v15, v48
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xff, v16
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v16, 8, v144
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xff, v135
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v21, 8, v134
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xff, v13
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v24, 8, v133
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v51, v25, v29
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v25, 0xff, v132
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v28, 8, v65
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xff, v14
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v29, 8, v131
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v16, v20, v21
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v13, v13, v24
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xff, v130
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v24, 8, v129
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v20, v25, v28
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v14, v14, v29
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xff, v11
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v25, 8, v128
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v28, 0xff, v119
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v29, 8, v34
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xff, v12
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v32, 8, v118
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xff, v117
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v34, 8, v116
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v21, v21, v24
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v11, v11, v25
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v24, v28, v29
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v12, v12, v32
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v25, v33, v34
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v65, v15, v16
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v66, v13, v20
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v67, v14, v21
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v9, 0xff, v9
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v13, 8, v115
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xff, v114
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v15, 8, v30
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xff, v10
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v16, 8, v113
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xff, v112
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v21, 8, v103
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v11, v11, v24
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v12, v12, v25
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xff, v7
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v24, 8, v102
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v9, v9, v13
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v13, v14, v15
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v10, v10, v16
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v14, v20, v21
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xff, v101
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v16, 8, v26
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xff, v8
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v20, 8, v100
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v25, 8, v97
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v26, 0xff, v96
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v22, 8, v22
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v7, v24
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xff, v99
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v24, 8, v98
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v15, v15, v16
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v8, v8, v20
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v5, v25
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v20, v26, v22
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v16, v21, v24
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v13, v9, v13
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v14, v10, v14
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v7, v15
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v9, v5, v20
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xff, v6
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v6, 8, v87
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xff, v86
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v15, 8, v85
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xff, v83
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v18, 8, v18
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v8, v8, v16
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v16, 8, v84
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v21, 8, v71
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v5, v6
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v6, v10, v15
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v10, v20, v18
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v3, v16
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v4, v21
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v16, 8, v49
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v15, 16, v10
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xff, v55
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v18, 8, v35
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xff, v31
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v17, 8, v17
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v21, 8, v27
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xff, v23
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v19, 8, v19
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v10, v10, v16
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v18
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v16, v20, v17
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v2, v21
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v17, v22, v19
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v18, 16, v10
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xffff, v1
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xffff, v2
-; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v10, v5, v6
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v3, v15
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v4, v18
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v19, v16
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v20, v17
-; GFX11-FAKE16-NEXT: s_clause 0x5
-; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[36:39], off offset:32
-; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[50:53], off offset:48
-; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[64:67], off offset:64
-; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[11:14], off offset:80
-; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[7:10], off offset:96
-; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[1:4], off offset:112
-; GFX11-FAKE16-NEXT: s_clause 0x12
-; GFX11-FAKE16-NEXT: scratch_load_b32 v74, off, s32
-; GFX11-FAKE16-NEXT: scratch_load_b32 v73, off, s32 offset:4
-; GFX11-FAKE16-NEXT: scratch_load_b32 v72, off, s32 offset:8
-; GFX11-FAKE16-NEXT: scratch_load_b32 v63, off, s32 offset:12
-; GFX11-FAKE16-NEXT: scratch_load_b32 v62, off, s32 offset:16
-; GFX11-FAKE16-NEXT: scratch_load_b32 v61, off, s32 offset:20
-; GFX11-FAKE16-NEXT: scratch_load_b32 v60, off, s32 offset:24
-; GFX11-FAKE16-NEXT: scratch_load_b32 v59, off, s32 offset:28
-; GFX11-FAKE16-NEXT: scratch_load_b32 v58, off, s32 offset:32
-; GFX11-FAKE16-NEXT: scratch_load_b32 v57, off, s32 offset:36
-; GFX11-FAKE16-NEXT: scratch_load_b32 v56, off, s32 offset:40
-; GFX11-FAKE16-NEXT: scratch_load_b32 v47, off, s32 offset:44
-; GFX11-FAKE16-NEXT: scratch_load_b32 v46, off, s32 offset:48
-; GFX11-FAKE16-NEXT: scratch_load_b32 v45, off, s32 offset:52
-; GFX11-FAKE16-NEXT: scratch_load_b32 v44, off, s32 offset:56
-; GFX11-FAKE16-NEXT: scratch_load_b32 v43, off, s32 offset:60
-; GFX11-FAKE16-NEXT: scratch_load_b32 v42, off, s32 offset:64
-; GFX11-FAKE16-NEXT: scratch_load_b32 v41, off, s32 offset:68
-; GFX11-FAKE16-NEXT: scratch_load_b32 v40, off, s32 offset:72
-; GFX11-FAKE16-NEXT: v_readlane_b32 s104, v76, 8
-; GFX11-FAKE16-NEXT: v_readlane_b32 s103, v76, 7
-; GFX11-FAKE16-NEXT: v_readlane_b32 s102, v76, 6
-; GFX11-FAKE16-NEXT: v_readlane_b32 s101, v76, 5
-; GFX11-FAKE16-NEXT: v_readlane_b32 s100, v76, 4
-; GFX11-FAKE16-NEXT: v_readlane_b32 s99, v76, 3
-; GFX11-FAKE16-NEXT: v_readlane_b32 s98, v76, 2
-; GFX11-FAKE16-NEXT: v_readlane_b32 s97, v76, 1
-; GFX11-FAKE16-NEXT: v_readlane_b32 s96, v76, 0
-; GFX11-FAKE16-NEXT: v_readlane_b32 s87, v75, 31
-; GFX11-FAKE16-NEXT: v_readlane_b32 s86, v75, 30
-; GFX11-FAKE16-NEXT: v_readlane_b32 s85, v75, 29
-; GFX11-FAKE16-NEXT: v_readlane_b32 s84, v75, 28
-; GFX11-FAKE16-NEXT: v_readlane_b32 s83, v75, 27
-; GFX11-FAKE16-NEXT: v_readlane_b32 s82, v75, 26
-; GFX11-FAKE16-NEXT: v_readlane_b32 s81, v75, 25
-; GFX11-FAKE16-NEXT: v_readlane_b32 s80, v75, 24
-; GFX11-FAKE16-NEXT: v_readlane_b32 s71, v75, 23
-; GFX11-FAKE16-NEXT: v_readlane_b32 s70, v75, 22
-; GFX11-FAKE16-NEXT: v_readlane_b32 s69, v75, 21
-; GFX11-FAKE16-NEXT: v_readlane_b32 s68, v75, 20
-; GFX11-FAKE16-NEXT: v_readlane_b32 s67, v75, 19
-; GFX11-FAKE16-NEXT: v_readlane_b32 s66, v75, 18
-; GFX11-FAKE16-NEXT: v_readlane_b32 s65, v75, 17
-; GFX11-FAKE16-NEXT: v_readlane_b32 s64, v75, 16
-; GFX11-FAKE16-NEXT: v_readlane_b32 s55, v75, 15
-; GFX11-FAKE16-NEXT: v_readlane_b32 s54, v75, 14
-; GFX11-FAKE16-NEXT: v_readlane_b32 s53, v75, 13
-; GFX11-FAKE16-NEXT: v_readlane_b32 s52, v75, 12
-; GFX11-FAKE16-NEXT: v_readlane_b32 s51, v75, 11
-; GFX11-FAKE16-NEXT: v_readlane_b32 s50, v75, 10
-; GFX11-FAKE16-NEXT: v_readlane_b32 s49, v75, 9
-; GFX11-FAKE16-NEXT: v_readlane_b32 s48, v75, 8
-; GFX11-FAKE16-NEXT: v_readlane_b32 s39, v75, 7
-; GFX11-FAKE16-NEXT: v_readlane_b32 s38, v75, 6
-; GFX11-FAKE16-NEXT: v_readlane_b32 s37, v75, 5
-; GFX11-FAKE16-NEXT: v_readlane_b32 s36, v75, 4
-; GFX11-FAKE16-NEXT: v_readlane_b32 s35, v75, 3
-; GFX11-FAKE16-NEXT: v_readlane_b32 s34, v75, 2
-; GFX11-FAKE16-NEXT: v_readlane_b32 s31, v75, 1
-; GFX11-FAKE16-NEXT: v_readlane_b32 s30, v75, 0
-; GFX11-FAKE16-NEXT: s_or_saveexec_b32 s0, -1
-; GFX11-FAKE16-NEXT: s_clause 0x3
-; GFX11-FAKE16-NEXT: scratch_load_b32 v75, off, s32 offset:76
-; GFX11-FAKE16-NEXT: scratch_load_b32 v76, off, s32 offset:80
-; GFX11-FAKE16-NEXT: scratch_load_b32 v77, off, s32 offset:84
-; GFX11-FAKE16-NEXT: scratch_load_b32 v78, off, s32 offset:88
-; GFX11-FAKE16-NEXT: s_mov_b32 exec_lo, s0
-; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT: 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) {
; SI-LABEL: bitcast_v64bf16_to_v64f16:
; SI: ; %bb.0:
@@ -226763,7 +203449,7 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
; SI-NEXT: ; kill: killed $vgpr2
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB100_2
+; SI-NEXT: s_cbranch_execz .LBB96_2
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
@@ -227089,9 +203775,9 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
; SI-NEXT: ; implicit-def: $vgpr61
; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; SI-NEXT: .LBB100_2: ; %Flow
+; SI-NEXT: .LBB96_2: ; %Flow
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB100_4
+; SI-NEXT: s_cbranch_execz .LBB96_4
; SI-NEXT: ; %bb.3: ; %cmp.true
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_and_b32_e32 v2, 0xffff0000, v31
@@ -227481,7 +204167,7 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; SI-NEXT: .LBB100_4: ; %end
+; SI-NEXT: .LBB96_4: ; %end
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
@@ -227879,7 +204565,7 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT: 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 .LBB100_2
+; VI-NEXT: s_cbranch_execz .LBB96_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
@@ -228465,7 +205151,7 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
; 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: .LBB100_2: ; %end
+; VI-NEXT: .LBB96_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
@@ -228512,7 +205198,7 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT: 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 .LBB100_2
+; GFX9-NEXT: s_cbranch_execz .LBB96_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
@@ -229002,7 +205688,7 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
; 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: .LBB100_2: ; %end
+; GFX9-NEXT: .LBB96_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
@@ -229035,7 +205721,7 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB100_2
+; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB96_2
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff0000, v16
; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff0000, v20
@@ -229557,7 +206243,7 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v85, v85, v97, vcc_lo
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_bfi_b32 v13, 0xffff, v13, v85
-; GFX11-TRUE16-NEXT: .LBB100_2: ; %end
+; GFX11-TRUE16-NEXT: .LBB96_2: ; %end
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
@@ -229574,7 +206260,7 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB100_2
+; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB96_2
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.true
; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v33, 16, v17
; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v32, 16, v16
@@ -230084,7 +206770,7 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
; GFX11-FAKE16-NEXT: v_cndmask_b32_e32 v11, v85, v96, vcc_lo
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v11, v83, 0x7060302
-; GFX11-FAKE16-NEXT: .LBB100_2: ; %end
+; GFX11-FAKE16-NEXT: .LBB96_2: ; %end
; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
@@ -230274,7 +206960,7 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; SI-NEXT: buffer_store_dword v38, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; SI-NEXT: s_cbranch_scc0 .LBB101_4
+; SI-NEXT: s_cbranch_scc0 .LBB97_4
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
@@ -230475,8 +207161,8 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; SI-NEXT: v_lshrrev_b32_e32 v13, 16, v52
; SI-NEXT: v_cvt_f32_f16_e32 v13, v13
; SI-NEXT: v_cvt_f32_f16_e32 v43, v43
-; SI-NEXT: s_cbranch_execnz .LBB101_3
-; SI-NEXT: .LBB101_2: ; %cmp.true
+; SI-NEXT: s_cbranch_execnz .LBB97_3
+; SI-NEXT: .LBB97_2: ; %cmp.true
; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: v_and_b32_e32 v1, 0xffff0000, v54
; SI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
@@ -230894,7 +207580,7 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; SI-NEXT: v_cvt_f32_f16_e32 v10, v10
; SI-NEXT: s_waitcnt vmcnt(1) expcnt(0)
; SI-NEXT: v_cvt_f32_f16_e32 v43, v14
-; SI-NEXT: .LBB101_3: ; %end
+; SI-NEXT: .LBB97_3: ; %end
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
@@ -231177,7 +207863,7 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT: s_setpc_b64 s[30:31]
-; SI-NEXT: .LBB101_4:
+; SI-NEXT: .LBB97_4:
; SI-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_mov_b32_e32 v53, v54
@@ -231282,7 +207968,7 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; SI-NEXT: ; implicit-def: $vgpr13
; SI-NEXT: ; kill: killed $vgpr43
; SI-NEXT: ; implicit-def: $vgpr43
-; SI-NEXT: s_branch .LBB101_2
+; SI-NEXT: s_branch .LBB97_2
;
; VI-LABEL: bitcast_v64bf16_to_v64f16_scalar:
; VI: ; %bb.0:
@@ -231314,10 +208000,10 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; VI-NEXT: v_readfirstlane_b32 s31, v1
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 ; 4-byte Folded Spill
-; VI-NEXT: s_cbranch_scc0 .LBB101_3
+; VI-NEXT: s_cbranch_scc0 .LBB97_3
; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: s_cbranch_execnz .LBB101_4
-; VI-NEXT: .LBB101_2: ; %cmp.true
+; VI-NEXT: s_cbranch_execnz .LBB97_4
+; VI-NEXT: .LBB97_2: ; %cmp.true
; VI-NEXT: v_mov_b32_e32 v0, 0x40c00000
; VI-NEXT: s_lshl_b32 s4, s30, 16
; VI-NEXT: v_add_f32_e32 v1, s4, v0
@@ -231895,10 +208581,10 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; VI-NEXT: v_alignbit_b32 v32, v32, v34, 16
; VI-NEXT: v_alignbit_b32 v17, v17, v33, 16
; VI-NEXT: v_alignbit_b32 v16, v16, v18, 16
-; VI-NEXT: s_branch .LBB101_5
-; VI-NEXT: .LBB101_3:
-; VI-NEXT: s_branch .LBB101_2
-; VI-NEXT: .LBB101_4:
+; VI-NEXT: s_branch .LBB97_5
+; VI-NEXT: .LBB97_3:
+; VI-NEXT: s_branch .LBB97_2
+; VI-NEXT: .LBB97_4:
; VI-NEXT: v_mov_b32_e32 v0, s16
; VI-NEXT: v_mov_b32_e32 v1, s17
; VI-NEXT: v_mov_b32_e32 v2, s18
@@ -231915,7 +208601,7 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; VI-NEXT: v_mov_b32_e32 v13, s29
; VI-NEXT: v_mov_b32_e32 v14, s30
; VI-NEXT: v_mov_b32_e32 v15, s31
-; VI-NEXT: .LBB101_5: ; %end
+; VI-NEXT: .LBB97_5: ; %end
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; VI-NEXT: v_mov_b32_e32 v18, v32
@@ -231958,10 +208644,10 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 ; 4-byte Folded Spill
-; GFX9-NEXT: s_cbranch_scc0 .LBB101_3
+; GFX9-NEXT: s_cbranch_scc0 .LBB97_3
; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: s_cbranch_execnz .LBB101_4
-; GFX9-NEXT: .LBB101_2: ; %cmp.true
+; GFX9-NEXT: s_cbranch_execnz .LBB97_4
+; GFX9-NEXT: .LBB97_2: ; %cmp.true
; GFX9-NEXT: v_mov_b32_e32 v0, 0x40c00000
; GFX9-NEXT: s_and_b32 s4, s30, 0xffff0000
; GFX9-NEXT: v_add_f32_e32 v1, s4, v0
@@ -232572,10 +209258,10 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; GFX9-NEXT: v_lshl_or_b32 v32, v35, 16, v32
; GFX9-NEXT: v_lshl_or_b32 v17, v34, 16, v17
; GFX9-NEXT: v_lshl_or_b32 v16, v18, 16, v16
-; GFX9-NEXT: s_branch .LBB101_5
-; GFX9-NEXT: .LBB101_3:
-; GFX9-NEXT: s_branch .LBB101_2
-; GFX9-NEXT: .LBB101_4:
+; GFX9-NEXT: s_branch .LBB97_5
+; GFX9-NEXT: .LBB97_3:
+; GFX9-NEXT: s_branch .LBB97_2
+; GFX9-NEXT: .LBB97_4:
; GFX9-NEXT: v_mov_b32_e32 v0, s16
; GFX9-NEXT: v_mov_b32_e32 v1, s17
; GFX9-NEXT: v_mov_b32_e32 v2, s18
@@ -232592,7 +209278,7 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; GFX9-NEXT: v_mov_b32_e32 v13, s29
; GFX9-NEXT: v_mov_b32_e32 v14, s30
; GFX9-NEXT: v_mov_b32_e32 v15, s31
-; GFX9-NEXT: .LBB101_5: ; %end
+; GFX9-NEXT: .LBB97_5: ; %end
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
@@ -232623,11 +209309,11 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; GFX11-TRUE16-NEXT: s_mov_b32 s12, s0
; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0
; GFX11-TRUE16-NEXT: s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB101_3
+; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB97_3
; GFX11-TRUE16-NEXT: ; %bb.1: ; %Flow
; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB101_4
-; GFX11-TRUE16-NEXT: .LBB101_2: ; %cmp.true
+; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB97_4
+; GFX11-TRUE16-NEXT: .LBB97_2: ; %cmp.true
; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v17
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v16
; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v18
@@ -233287,9 +209973,9 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; GFX11-TRUE16-NEXT: v_lshl_or_b32 v18, v33, 16, v37
; GFX11-TRUE16-NEXT: v_lshl_or_b32 v17, v17, 16, v38
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT: .LBB101_3:
-; GFX11-TRUE16-NEXT: s_branch .LBB101_2
-; GFX11-TRUE16-NEXT: .LBB101_4:
+; GFX11-TRUE16-NEXT: .LBB97_3:
+; GFX11-TRUE16-NEXT: s_branch .LBB97_2
+; GFX11-TRUE16-NEXT: .LBB97_4:
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
@@ -233318,11 +210004,11 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; GFX11-FAKE16-NEXT: s_mov_b32 s12, s0
; GFX11-FAKE16-NEXT: s_mov_b32 s0, 0
; GFX11-FAKE16-NEXT: s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB101_3
+; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB97_3
; GFX11-FAKE16-NEXT: ; %bb.1: ; %Flow
; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB101_4
-; GFX11-FAKE16-NEXT: .LBB101_2: ; %cmp.true
+; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB97_4
+; GFX11-FAKE16-NEXT: .LBB97_2: ; %cmp.true
; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v17
; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v1, 16, v16
; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v18
@@ -233982,9 +210668,9 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v33, 16, v37
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v17, v17, 16, v38
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT: .LBB101_3:
-; GFX11-FAKE16-NEXT: s_branch .LBB101_2
-; GFX11-FAKE16-NEXT: .LBB101_4:
+; GFX11-FAKE16-NEXT: .LBB97_3:
+; GFX11-FAKE16-NEXT: s_branch .LBB97_2
+; GFX11-FAKE16-NEXT: .LBB97_4:
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
@@ -234281,7 +210967,7 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
; SI-NEXT: ; kill: killed $vgpr2
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB102_2
+; SI-NEXT: s_cbranch_execz .LBB98_2
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
@@ -234542,9 +211228,9 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
; SI-NEXT: ; implicit-def: $vgpr60
; SI-NEXT: ; implicit-def: $vgpr61
; SI-NEXT: ; implicit-def: $vgpr62
-; SI-NEXT: .LBB102_2: ; %Flow
+; SI-NEXT: .LBB98_2: ; %Flow
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB102_4
+; SI-NEXT: s_cbranch_execz .LBB98_4
; SI-NEXT: ; %bb.3: ; %cmp.true
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt expcnt(0)
@@ -234938,7 +211624,7 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v4
; SI-NEXT: v_mov_b32_e32 v4, v7
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; SI-NEXT: .LBB102_4: ; %end
+; SI-NEXT: .LBB98_4: ; %end
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
@@ -235316,7 +212002,7 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT: 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 .LBB102_2
+; VI-NEXT: s_cbranch_execz .LBB98_2
; VI-NEXT: ; %bb.1: ; %cmp.true
; VI-NEXT: v_mov_b32_e32 v32, 0x200
; VI-NEXT: v_add_f16_e32 v33, 0x200, v15
@@ -235416,7 +212102,7 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
; 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: .LBB102_2: ; %end
+; VI-NEXT: .LBB98_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]
@@ -235431,7 +212117,7 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT: 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 .LBB102_2
+; GFX9-NEXT: s_cbranch_execz .LBB98_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]
@@ -235467,7 +212153,7 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
; 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: .LBB102_2: ; %end
+; GFX9-NEXT: .LBB98_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]
@@ -235484,7 +212170,7 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
; GFX11-NEXT: s_xor_b32 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 .LBB102_2
+; GFX11-NEXT: s_cbranch_execz .LBB98_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]
@@ -235519,7 +212205,7 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
; 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: .LBB102_2: ; %end
+; GFX11-NEXT: .LBB98_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]
@@ -235701,7 +212387,7 @@ define inreg <64 x bfloat> @bitcast_v64f16_to_v64bf16_scalar(<64 x half> inreg %
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v36, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; SI-NEXT: s_cbranch_scc0 .LBB103_2
+; SI-NEXT: s_cbranch_scc0 .LBB99_2
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
@@ -235851,8 +212537,8 @@ define inreg <64 x bfloat> @bitcast_v64f16_to_v64bf16_scalar(<64 x half> inreg %
; SI-NEXT: v_mov_b32_e32 v44, v18
; SI-NEXT: v_mov_b32_e32 v5, v43
; SI-NEXT: v_mov_b32_e32 v18, v6
-; SI-NEXT: s_branch .LBB103_3
-; SI-NEXT: .LBB103_2:
+; SI-NEXT: s_branch .LBB99_3
+; SI-NEXT: .LBB99_2:
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; SI-NEXT: ; implicit-def: $vgpr5
; SI-NEXT: ; implicit-def: $vgpr3
@@ -235960,7 +212646,7 @@ define inreg <64 x bfloat> @bitcast_v64f16_to_v64bf16_scalar(<64 x half> inreg %
; SI-NEXT: ; implicit-def: $vgpr15
; SI-NEXT: ; implicit-def: $vgpr7
; SI-NEXT: ; implicit-def: $vgpr14
-; SI-NEXT: .LBB103_3: ; %Flow
+; SI-NEXT: .LBB99_3: ; %Flow
; SI-NEXT: s_waitcnt expcnt(1)
; SI-NEXT: v_mov_b32_e32 v36, v2
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
@@ -235969,7 +212655,7 @@ define inreg <64 x bfloat> @bitcast_v64f16_to_v64bf16_scalar(<64 x half> inreg %
; SI-NEXT: s_andn2_b64 vcc, exec, s[4:5]
; SI-NEXT: v_mov_b32_e32 v31, v11
; SI-NEXT: v_mov_b32_e32 v9, v17
-; SI-NEXT: s_cbranch_vccnz .LBB103_5
+; SI-NEXT: s_cbranch_vccnz .LBB99_5
; SI-NEXT: ; %bb.4: ; %cmp.true
; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
@@ -236392,7 +213078,7 @@ define inreg <64 x bfloat> @bitcast_v64f16_to_v64bf16_scalar(<64 x half> inreg %
; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v2
-; SI-NEXT: .LBB103_5: ; %end
+; SI-NEXT: .LBB99_5: ; %end
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt vmcnt(1)
@@ -236737,10 +213423,10 @@ define inreg <64 x bfloat> @bitcast_v64f16_to_v64bf16_scalar(<64 x half> inreg %
; VI-NEXT: v_mov_b32_e32 v11, s27
; VI-NEXT: v_mov_b32_e32 v12, s28
; VI-NEXT: v_mov_b32_e32 v13, s29
-; VI-NEXT: s_cbranch_scc0 .LBB103_4
+; VI-NEXT: s_cbranch_scc0 .LBB99_4
; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: s_cbranch_execnz .LBB103_3
-; VI-NEXT: .LBB103_2: ; %cmp.true
+; VI-NEXT: s_cbranch_execnz .LBB99_3
+; VI-NEXT: .LBB99_2: ; %cmp.true
; VI-NEXT: v_mov_b32_e32 v18, 0x200
; VI-NEXT: v_add_f16_e32 v33, 0x200, v15
; VI-NEXT: v_add_f16_sdwa v15, v15, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
@@ -236838,11 +213524,11 @@ define inreg <64 x bfloat> @bitcast_v64f16_to_v64bf16_scalar(<64 x half> inreg %
; 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, v18
-; VI-NEXT: .LBB103_3: ; %end
+; VI-NEXT: .LBB99_3: ; %end
; VI-NEXT: v_mov_b32_e32 v18, v32
; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB103_4:
-; VI-NEXT: s_branch .LBB103_2
+; VI-NEXT: .LBB99_4:
+; VI-NEXT: s_branch .LBB99_2
;
; GFX9-LABEL: bitcast_v64f16_to_v64bf16_scalar:
; GFX9: ; %bb.0:
@@ -236881,10 +213567,10 @@ define inreg <64 x bfloat> @bitcast_v64f16_to_v64bf16_scalar(<64 x half> inreg %
; GFX9-NEXT: v_mov_b32_e32 v11, s27
; GFX9-NEXT: v_mov_b32_e32 v12, s28
; GFX9-NEXT: v_mov_b32_e32 v13, s29
-; GFX9-NEXT: s_cbranch_scc0 .LBB103_4
+; GFX9-NEXT: s_cbranch_scc0 .LBB99_4
; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: s_cbranch_execnz .LBB103_3
-; GFX9-NEXT: .LBB103_2: ; %cmp.true
+; GFX9-NEXT: s_cbranch_execnz .LBB99_3
+; GFX9-NEXT: .LBB99_2: ; %cmp.true
; GFX9-NEXT: s_movk_i32 s4, 0x200
; GFX9-NEXT: v_pk_add_f16 v15, v15, s4 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v14, v14, s4 op_sel_hi:[1,0]
@@ -236918,11 +213604,11 @@ define inreg <64 x bfloat> @bitcast_v64f16_to_v64bf16_scalar(<64 x half> inreg %
; GFX9-NEXT: v_pk_add_f16 v32, v32, s4 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v17, v17, s4 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v16, v16, s4 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB103_3: ; %end
+; GFX9-NEXT: .LBB99_3: ; %end
; GFX9-NEXT: v_mov_b32_e32 v18, v32
; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB103_4:
-; GFX9-NEXT: s_branch .LBB103_2
+; GFX9-NEXT: .LBB99_4:
+; GFX9-NEXT: s_branch .LBB99_2
;
; GFX11-LABEL: bitcast_v64f16_to_v64bf16_scalar:
; GFX11: ; %bb.0:
@@ -236942,11 +213628,11 @@ define inreg <64 x bfloat> @bitcast_v64f16_to_v64bf16_scalar(<64 x half> inreg %
; GFX11-NEXT: s_mov_b32 s12, s0
; GFX11-NEXT: s_mov_b32 s0, 0
; GFX11-NEXT: s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-NEXT: s_cbranch_scc0 .LBB103_3
+; GFX11-NEXT: s_cbranch_scc0 .LBB99_3
; GFX11-NEXT: ; %bb.1: ; %Flow
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_vccnz .LBB103_4
-; GFX11-NEXT: .LBB103_2: ; %cmp.true
+; GFX11-NEXT: s_cbranch_vccnz .LBB99_4
+; GFX11-NEXT: .LBB99_2: ; %cmp.true
; GFX11-NEXT: v_pk_add_f16 v15, 0x200, s27 op_sel_hi:[0,1]
; GFX11-NEXT: v_pk_add_f16 v14, 0x200, s26 op_sel_hi:[0,1]
; GFX11-NEXT: v_pk_add_f16 v13, 0x200, s25 op_sel_hi:[0,1]
@@ -236980,9 +213666,9 @@ define inreg <64 x bfloat> @bitcast_v64f16_to_v64bf16_scalar(<64 x half> inreg %
; 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: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB103_3:
-; GFX11-NEXT: s_branch .LBB103_2
-; GFX11-NEXT: .LBB103_4:
+; GFX11-NEXT: .LBB99_3:
+; GFX11-NEXT: s_branch .LBB99_2
+; GFX11-NEXT: .LBB99_4:
; GFX11-NEXT: v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-NEXT: v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
@@ -237295,7 +213981,7 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
; SI-NEXT: ; implicit-def: $vgpr26
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB104_2
+; SI-NEXT: s_cbranch_execz .LBB100_2
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: v_lshrrev_b32_e32 v26, 16, v33
; SI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
@@ -237556,9 +214242,9 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v11
; SI-NEXT: ; implicit-def: $vgpr11
-; SI-NEXT: .LBB104_2: ; %Flow
+; SI-NEXT: .LBB100_2: ; %Flow
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB104_4
+; SI-NEXT: s_cbranch_execz .LBB100_4
; SI-NEXT: ; %bb.3: ; %cmp.true
; SI-NEXT: v_and_b32_e32 v27, 0xffff0000, v43
; SI-NEXT: v_and_b32_e32 v26, 0xffff0000, v33
@@ -237892,7 +214578,7 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_alignbit_b32 v1, v27, v3, 16
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; SI-NEXT: .LBB104_4: ; %end
+; SI-NEXT: .LBB100_4: ; %end
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
@@ -238222,7 +214908,7 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT: 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 .LBB104_2
+; VI-NEXT: s_cbranch_execz .LBB100_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
@@ -238808,7 +215494,7 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
; 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: .LBB104_2: ; %end
+; VI-NEXT: .LBB100_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
@@ -238855,7 +215541,7 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT: 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 .LBB104_2
+; GFX9-NEXT: s_cbranch_execz .LBB100_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
@@ -239345,7 +216031,7 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
; 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: .LBB104_2: ; %end
+; GFX9-NEXT: .LBB100_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
@@ -239378,7 +216064,7 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB104_2
+; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB100_2
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v33, 16, v17
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v16
@@ -239944,7 +216630,7 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshl_or_b32 v17, v37, 16, v38
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT: v_lshl_or_b32 v16, v39, 16, v16
-; GFX11-TRUE16-NEXT: .LBB104_2: ; %end
+; GFX11-TRUE16-NEXT: .LBB100_2: ; %end
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
@@ -239961,7 +216647,7 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB104_2
+; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB100_2
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.true
; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v33, 16, v17
; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v32, 16, v16
@@ -240471,7 +217157,7 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
; GFX11-FAKE16-NEXT: v_cndmask_b32_e32 v11, v85, v96, vcc_lo
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v11, v83, 0x7060302
-; GFX11-FAKE16-NEXT: .LBB104_2: ; %end
+; GFX11-FAKE16-NEXT: .LBB100_2: ; %end
; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
@@ -240657,7 +217343,7 @@ define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a
; SI-NEXT: buffer_store_dword v52, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; SI-NEXT: s_cbranch_scc0 .LBB105_2
+; SI-NEXT: s_cbranch_scc0 .LBB101_2
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
@@ -240801,8 +217487,8 @@ define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a
; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v8
; SI-NEXT: s_waitcnt vmcnt(3)
; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v34
-; SI-NEXT: s_branch .LBB105_3
-; SI-NEXT: .LBB105_2:
+; SI-NEXT: s_branch .LBB101_3
+; SI-NEXT: .LBB101_2:
; SI-NEXT: s_waitcnt expcnt(4)
; SI-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
@@ -240914,10 +217600,10 @@ define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a
; SI-NEXT: ; implicit-def: $vgpr5
; SI-NEXT: ; implicit-def: $vgpr36
; SI-NEXT: ; implicit-def: $vgpr35
-; SI-NEXT: .LBB105_3: ; %Flow
+; SI-NEXT: .LBB101_3: ; %Flow
; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT: s_andn2_b64 vcc, exec, s[4:5]
-; SI-NEXT: s_cbranch_vccnz .LBB105_5
+; SI-NEXT: s_cbranch_vccnz .LBB101_5
; SI-NEXT: ; %bb.4: ; %cmp.true
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
@@ -241275,7 +217961,7 @@ define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a
; SI-NEXT: v_alignbit_b32 v4, v13, v14, 16
; SI-NEXT: v_mov_b32_e32 v14, v51
; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; SI-NEXT: .LBB105_5: ; %end
+; SI-NEXT: .LBB101_5: ; %end
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
@@ -241562,10 +218248,10 @@ define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a
; VI-NEXT: v_readfirstlane_b32 s31, v1
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 ; 4-byte Folded Spill
-; VI-NEXT: s_cbranch_scc0 .LBB105_3
+; VI-NEXT: s_cbranch_scc0 .LBB101_3
; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: s_cbranch_execnz .LBB105_4
-; VI-NEXT: .LBB105_2: ; %cmp.true
+; VI-NEXT: s_cbranch_execnz .LBB101_4
+; VI-NEXT: .LBB101_2: ; %cmp.true
; VI-NEXT: v_mov_b32_e32 v0, 0x40c00000
; VI-NEXT: s_lshl_b32 s4, s30, 16
; VI-NEXT: v_add_f32_e32 v1, s4, v0
@@ -242143,10 +218829,10 @@ define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a
; VI-NEXT: v_alignbit_b32 v32, v32, v34, 16
; VI-NEXT: v_alignbit_b32 v17, v17, v33, 16
; VI-NEXT: v_alignbit_b32 v16, v16, v18, 16
-; VI-NEXT: s_branch .LBB105_5
-; VI-NEXT: .LBB105_3:
-; VI-NEXT: s_branch .LBB105_2
-; VI-NEXT: .LBB105_4:
+; VI-NEXT: s_branch .LBB101_5
+; VI-NEXT: .LBB101_3:
+; VI-NEXT: s_branch .LBB101_2
+; VI-NEXT: .LBB101_4:
; VI-NEXT: v_mov_b32_e32 v0, s16
; VI-NEXT: v_mov_b32_e32 v1, s17
; VI-NEXT: v_mov_b32_e32 v2, s18
@@ -242163,7 +218849,7 @@ define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a
; VI-NEXT: v_mov_b32_e32 v13, s29
; VI-NEXT: v_mov_b32_e32 v14, s30
; VI-NEXT: v_mov_b32_e32 v15, s31
-; VI-NEXT: .LBB105_5: ; %end
+; VI-NEXT: .LBB101_5: ; %end
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; VI-NEXT: v_mov_b32_e32 v18, v32
@@ -242206,10 +218892,10 @@ define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 ; 4-byte Folded Spill
-; GFX9-NEXT: s_cbranch_scc0 .LBB105_3
+; GFX9-NEXT: s_cbranch_scc0 .LBB101_3
; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: s_cbranch_execnz .LBB105_4
-; GFX9-NEXT: .LBB105_2: ; %cmp.true
+; GFX9-NEXT: s_cbranch_execnz .LBB101_4
+; GFX9-NEXT: .LBB101_2: ; %cmp.true
; GFX9-NEXT: v_mov_b32_e32 v0, 0x40c00000
; GFX9-NEXT: s_and_b32 s4, s30, 0xffff0000
; GFX9-NEXT: v_add_f32_e32 v1, s4, v0
@@ -242788,10 +219474,10 @@ define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a
; GFX9-NEXT: v_and_or_b32 v32, v35, v18, v32
; GFX9-NEXT: v_and_or_b32 v17, v34, v18, v17
; GFX9-NEXT: v_and_or_b32 v16, v33, v18, v16
-; GFX9-NEXT: s_branch .LBB105_5
-; GFX9-NEXT: .LBB105_3:
-; GFX9-NEXT: s_branch .LBB105_2
-; GFX9-NEXT: .LBB105_4:
+; GFX9-NEXT: s_branch .LBB101_5
+; GFX9-NEXT: .LBB101_3:
+; GFX9-NEXT: s_branch .LBB101_2
+; GFX9-NEXT: .LBB101_4:
; GFX9-NEXT: v_mov_b32_e32 v0, s16
; GFX9-NEXT: v_mov_b32_e32 v1, s17
; GFX9-NEXT: v_mov_b32_e32 v2, s18
@@ -242808,7 +219494,7 @@ define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a
; GFX9-NEXT: v_mov_b32_e32 v13, s29
; GFX9-NEXT: v_mov_b32_e32 v14, s30
; GFX9-NEXT: v_mov_b32_e32 v15, s31
-; GFX9-NEXT: .LBB105_5: ; %end
+; GFX9-NEXT: .LBB101_5: ; %end
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
@@ -242839,11 +219525,11 @@ define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a
; GFX11-NEXT: s_mov_b32 s12, s0
; GFX11-NEXT: s_mov_b32 s0, 0
; GFX11-NEXT: s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-NEXT: s_cbranch_scc0 .LBB105_3
+; GFX11-NEXT: s_cbranch_scc0 .LBB101_3
; GFX11-NEXT: ; %bb.1: ; %Flow
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_vccnz .LBB105_4
-; GFX11-NEXT: .LBB105_2: ; %cmp.true
+; GFX11-NEXT: s_cbranch_vccnz .LBB101_4
+; GFX11-NEXT: .LBB101_2: ; %cmp.true
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v17
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v16
; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v18
@@ -243444,9 +220130,9 @@ define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a
; GFX11-NEXT: v_and_or_b32 v17, 0xffff0000, v32, v39
; GFX11-NEXT: v_and_or_b32 v16, 0xffff0000, v16, v48
; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB105_3:
-; GFX11-NEXT: s_branch .LBB105_2
-; GFX11-NEXT: .LBB105_4:
+; GFX11-NEXT: .LBB101_3:
+; GFX11-NEXT: s_branch .LBB101_2
+; GFX11-NEXT: .LBB101_4:
; GFX11-NEXT: v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-NEXT: v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
@@ -243631,7 +220317,7 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
; SI-NEXT: ; implicit-def: $vgpr3
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB106_2
+; SI-NEXT: s_cbranch_execz .LBB102_2
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt vmcnt(8)
@@ -243778,9 +220464,9 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
; SI-NEXT: ; implicit-def: $vgpr63
; SI-NEXT: ; implicit-def: $vgpr40
; SI-NEXT: ; implicit-def: $vgpr62
-; SI-NEXT: .LBB106_2: ; %Flow
+; SI-NEXT: .LBB102_2: ; %Flow
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB106_4
+; SI-NEXT: s_cbranch_execz .LBB102_4
; SI-NEXT: ; %bb.3: ; %cmp.true
; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt vmcnt(9)
@@ -244044,7 +220730,7 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
; SI-NEXT: v_and_b32_e32 v14, 0xffff0000, v16
; SI-NEXT: v_and_b32_e32 v16, 0xffff0000, v11
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; SI-NEXT: .LBB106_4: ; %end
+; SI-NEXT: .LBB102_4: ; %end
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
; SI-NEXT: s_waitcnt vmcnt(7) expcnt(0)
; SI-NEXT: v_mul_f32_e32 v1, 1.0, v2
@@ -244334,7 +221020,7 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT: 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 .LBB106_2
+; VI-NEXT: s_cbranch_execz .LBB102_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
@@ -244434,7 +221120,7 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
; 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: .LBB106_2: ; %end
+; VI-NEXT: .LBB102_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]
@@ -244449,7 +221135,7 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT: 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 .LBB106_2
+; GFX9-NEXT: s_cbranch_execz .LBB102_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]
@@ -244484,7 +221170,7 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
; 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: .LBB106_2: ; %end
+; GFX9-NEXT: .LBB102_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]
@@ -244501,7 +221187,7 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
; GFX11-NEXT: s_xor_b32 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 .LBB106_2
+; GFX11-NEXT: s_cbranch_execz .LBB102_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]
@@ -244536,7 +221222,7 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
; 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: .LBB106_2: ; %end
+; GFX11-NEXT: .LBB102_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]
@@ -244729,7 +221415,7 @@ define inreg <64 x bfloat> @bitcast_v64i16_to_v64bf16_scalar(<64 x i16> inreg %a
; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: v_readfirstlane_b32 s41, v37
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
-; SI-NEXT: s_cbranch_scc0 .LBB107_2
+; SI-NEXT: s_cbranch_scc0 .LBB103_2
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: s_lshl_b32 s4, s60, 16
; SI-NEXT: v_writelane_b32 v41, s4, 25
@@ -244812,8 +221498,8 @@ define inreg <64 x bfloat> @bitcast_v64i16_to_v64bf16_scalar(<64 x i16> inreg %a
; SI-NEXT: s_mov_b32 s6, s62
; SI-NEXT: s_lshl_b32 s26, s62, 16
; SI-NEXT: s_mov_b64 s[4:5], 0
-; SI-NEXT: s_branch .LBB107_3
-; SI-NEXT: .LBB107_2:
+; SI-NEXT: s_branch .LBB103_3
+; SI-NEXT: .LBB103_2:
; SI-NEXT: ; implicit-def: $sgpr17
; SI-NEXT: ; kill: killed $sgpr17
; SI-NEXT: s_mov_b32 s16, s68
@@ -244895,12 +221581,12 @@ define inreg <64 x bfloat> @bitcast_v64i16_to_v64bf16_scalar(<64 x i16> inreg %a
; SI-NEXT: ; implicit-def: $sgpr17
; SI-NEXT: ; kill: killed $sgpr17
; SI-NEXT: ; implicit-def: $sgpr17
-; SI-NEXT: .LBB107_3: ; %Flow
+; SI-NEXT: .LBB103_3: ; %Flow
; SI-NEXT: s_andn2_b64 vcc, exec, s[4:5]
; SI-NEXT: s_mov_b32 s5, s17
; SI-NEXT: s_mov_b32 s17, s86
; SI-NEXT: s_mov_b32 s86, s7
-; SI-NEXT: s_cbranch_vccnz .LBB107_5
+; SI-NEXT: s_cbranch_vccnz .LBB103_5
; SI-NEXT: ; %bb.4: ; %cmp.true
; SI-NEXT: s_lshl_b32 s5, s6, 16
; SI-NEXT: v_readlane_b32 s6, v41, 24
@@ -245163,7 +221849,7 @@ define inreg <64 x bfloat> @bitcast_v64i16_to_v64bf16_scalar(<64 x i16> inreg %a
; SI-NEXT: s_and_b32 s26, s4, 0xffff0000
; SI-NEXT: s_lshl_b32 s5, s4, 16
; SI-NEXT: v_writelane_b32 v41, s6, 36
-; SI-NEXT: .LBB107_5: ; %end
+; SI-NEXT: .LBB103_5: ; %end
; SI-NEXT: v_mul_f32_e64 v1, 1.0, s86
; SI-NEXT: v_readlane_b32 s4, v41, 25
; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
@@ -245475,10 +222161,10 @@ define inreg <64 x bfloat> @bitcast_v64i16_to_v64bf16_scalar(<64 x i16> inreg %a
; VI-NEXT: s_and_b64 s[4:5], vcc, exec
; VI-NEXT: v_readfirstlane_b32 s7, v1
; VI-NEXT: v_writelane_b32 v32, s39, 7
-; VI-NEXT: s_cbranch_scc0 .LBB107_4
+; VI-NEXT: s_cbranch_scc0 .LBB103_4
; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: s_cbranch_execnz .LBB107_3
-; VI-NEXT: .LBB107_2: ; %cmp.true
+; VI-NEXT: s_cbranch_execnz .LBB103_3
+; VI-NEXT: .LBB103_2: ; %cmp.true
; VI-NEXT: s_and_b32 s4, s47, 0xffff0000
; VI-NEXT: s_add_i32 s5, s47, 3
; VI-NEXT: s_and_b32 s47, s46, 0xffff0000
@@ -245639,7 +222325,7 @@ define inreg <64 x bfloat> @bitcast_v64i16_to_v64bf16_scalar(<64 x i16> inreg %a
; VI-NEXT: s_add_i32 s45, s45, 0x30000
; VI-NEXT: s_add_i32 s46, s46, 0x30000
; VI-NEXT: s_add_i32 s47, s4, 0x30000
-; VI-NEXT: .LBB107_3: ; %end
+; VI-NEXT: .LBB103_3: ; %end
; VI-NEXT: v_mov_b32_e32 v0, s16
; VI-NEXT: v_mov_b32_e32 v1, s17
; VI-NEXT: v_mov_b32_e32 v2, s18
@@ -245685,8 +222371,8 @@ define inreg <64 x bfloat> @bitcast_v64i16_to_v64bf16_scalar(<64 x i16> inreg %a
; VI-NEXT: s_mov_b64 exec, s[4:5]
; VI-NEXT: s_waitcnt vmcnt(0)
; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB107_4:
-; VI-NEXT: s_branch .LBB107_2
+; VI-NEXT: .LBB103_4:
+; VI-NEXT: s_branch .LBB103_2
;
; GFX9-LABEL: bitcast_v64i16_to_v64bf16_scalar:
; GFX9: ; %bb.0:
@@ -245725,10 +222411,10 @@ define inreg <64 x bfloat> @bitcast_v64i16_to_v64bf16_scalar(<64 x i16> inreg %a
; GFX9-NEXT: v_mov_b32_e32 v11, s27
; GFX9-NEXT: v_mov_b32_e32 v12, s28
; GFX9-NEXT: v_mov_b32_e32 v13, s29
-; GFX9-NEXT: s_cbranch_scc0 .LBB107_4
+; GFX9-NEXT: s_cbranch_scc0 .LBB103_4
; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: s_cbranch_execnz .LBB107_3
-; GFX9-NEXT: .LBB107_2: ; %cmp.true
+; GFX9-NEXT: s_cbranch_execnz .LBB103_3
+; GFX9-NEXT: .LBB103_2: ; %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]
@@ -245761,11 +222447,11 @@ define inreg <64 x bfloat> @bitcast_v64i16_to_v64bf16_scalar(<64 x i16> inreg %a
; GFX9-NEXT: v_pk_add_u16 v32, v32, 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: .LBB107_3: ; %end
+; GFX9-NEXT: .LBB103_3: ; %end
; GFX9-NEXT: v_mov_b32_e32 v18, v32
; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB107_4:
-; GFX9-NEXT: s_branch .LBB107_2
+; GFX9-NEXT: .LBB103_4:
+; GFX9-NEXT: s_branch .LBB103_2
;
; GFX11-LABEL: bitcast_v64i16_to_v64bf16_scalar:
; GFX11: ; %bb.0:
@@ -245785,11 +222471,11 @@ define inreg <64 x bfloat> @bitcast_v64i16_to_v64bf16_scalar(<64 x i16> inreg %a
; GFX11-NEXT: s_mov_b32 s12, s0
; GFX11-NEXT: s_mov_b32 s0, 0
; GFX11-NEXT: s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-NEXT: s_cbranch_scc0 .LBB107_3
+; GFX11-NEXT: s_cbranch_scc0 .LBB103_3
; GFX11-NEXT: ; %bb.1: ; %Flow
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_vccnz .LBB107_4
-; GFX11-NEXT: .LBB107_2: ; %cmp.true
+; GFX11-NEXT: s_cbranch_vccnz .LBB103_4
+; GFX11-NEXT: .LBB103_2: ; %cmp.true
; GFX11-NEXT: v_pk_add_u16 v15, s27, 3 op_sel_hi:[1,0]
; GFX11-NEXT: v_pk_add_u16 v14, s26, 3 op_sel_hi:[1,0]
; GFX11-NEXT: v_pk_add_u16 v13, s25, 3 op_sel_hi:[1,0]
@@ -245823,9 +222509,9 @@ define inreg <64 x bfloat> @bitcast_v64i16_to_v64bf16_scalar(<64 x i16> inreg %a
; 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: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB107_3:
-; GFX11-NEXT: s_branch .LBB107_2
-; GFX11-NEXT: .LBB107_4:
+; GFX11-NEXT: .LBB103_3:
+; GFX11-NEXT: s_branch .LBB103_2
+; GFX11-NEXT: .LBB103_4:
; GFX11-NEXT: v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-NEXT: v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
@@ -246017,7 +222703,7 @@ define <64 x i16> @bitcast_v64f16_to_v64i16(<64 x half> %a, i32 %b) {
; SI-NEXT: v_mov_b32_e32 v62, v3
; SI-NEXT: v_mov_b32_e32 v63, v4
; SI-NEXT: s_xor_b64 exec, exec, s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB108_2
+; SI-NEXT: s_cbranch_execz .LBB104_2
; SI-NEXT: ; %bb.1: ; %cmp.true
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT: v_cvt_f32_f16_e32 v10, v10
@@ -246321,7 +223007,7 @@ define <64 x i16> @bitcast_v64f16_to_v64i16(<64 x half> %a, i32 %b) {
; SI-NEXT: v_alignbit_b32 v2, v21, v27, 16
; SI-NEXT: v_alignbit_b32 v35, v33, v35, 16
; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; SI-NEXT: .LBB108_2: ; %end
+; SI-NEXT: .LBB104_2: ; %end
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
@@ -246558,7 +223244,7 @@ define <64 x i16> @bitcast_v64f16_to_v64i16(<64 x half> %a, i32 %b) {
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT: 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 .LBB108_2
+; VI-NEXT: s_cbranch_execz .LBB104_2
; VI-NEXT: ; %bb.1: ; %cmp.true
; VI-NEXT: v_mov_b32_e32 v32, 0x200
; VI-NEXT: v_add_f16_e32 v33, 0x200, v15
@@ -246658,7 +223344,7 @@ define <64 x i16> @bitcast_v64f16_to_v64i16(<64 x half> %a, i32 %b) {
; 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: .LBB108_2: ; %end
+; VI-NEXT: .LBB104_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]
@@ -246673,7 +223359,7 @@ define <64 x i16> @bitcast_v64f16_to_v64i16(<64 x half> %a, i32 %b) {
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT: 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 .LBB108_2
+; GFX9-NEXT: s_cbranch_execz .LBB104_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]
@@ -246709,7 +223395,7 @@ define <64 x i16> @bitcast_v64f16_to_v64i16(<64 x half> %a, i32 %b) {
; 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: .LBB108_2: ; %end
+; GFX9-NEXT: .LBB104_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]
@@ -246726,7 +223412,7 @@ define <64 x i16> @bitcast_v64f16_to_v64i16(<64 x half> %a, i32 %b) {
; GFX11-NEXT: s_xor_b32 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 .LBB108_2
+; GFX11-NEXT: s_cbranch_execz .LBB104_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]
@@ -246761,7 +223447,7 @@ define <64 x i16> @bitcast_v64f16_to_v64i16(<64 x half> %a, i32 %b) {
; 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: .LBB108_2: ; %end
+; GFX11-NEXT: .LBB104_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]
@@ -246918,10 +223604,10 @@ define inreg <64 x i16> @bitcast_v64f16_to_v64i16_scalar(<64 x half> inreg %a, i
; SI-NEXT: v_cvt_f16_f32_e32 v45, s28
; SI-NEXT: v_cvt_f16_f32_e32 v61, s29
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
-; SI-NEXT: s_cbranch_scc0 .LBB109_4
+; SI-NEXT: s_cbranch_scc0 .LBB105_4
; SI-NEXT: ; %bb.1: ; %cmp.false
-; SI-NEXT: s_cbranch_execnz .LBB109_3
-; SI-NEXT: .LBB109_2: ; %cmp.true
+; SI-NEXT: s_cbranch_execnz .LBB105_3
+; SI-NEXT: .LBB105_2: ; %cmp.true
; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
@@ -247228,7 +223914,7 @@ define inreg <64 x i16> @bitcast_v64f16_to_v64i16_scalar(<64 x half> inreg %a, i
; SI-NEXT: v_alignbit_b32 v24, v5, v24, 16
; SI-NEXT: s_waitcnt vmcnt(1)
; SI-NEXT: v_alignbit_b32 v37, v1, v37, 16
-; SI-NEXT: .LBB109_3: ; %end
+; SI-NEXT: .LBB105_3: ; %end
; SI-NEXT: v_and_b32_e32 v48, 0xffff, v60
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT: v_or_b32_e32 v4, v48, v4
@@ -247450,8 +224136,8 @@ define inreg <64 x i16> @bitcast_v64f16_to_v64i16_scalar(<64 x half> inreg %a, i
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT: s_setpc_b64 s[30:31]
-; SI-NEXT: .LBB109_4:
-; SI-NEXT: s_branch .LBB109_2
+; SI-NEXT: .LBB105_4:
+; SI-NEXT: s_branch .LBB105_2
;
; VI-LABEL: bitcast_v64f16_to_v64i16_scalar:
; VI: ; %bb.0:
@@ -247490,10 +224176,10 @@ define inreg <64 x i16> @bitcast_v64f16_to_v64i16_scalar(<64 x half> inreg %a, i
; VI-NEXT: v_mov_b32_e32 v11, s27
; VI-NEXT: v_mov_b32_e32 v12, s28
; VI-NEXT: v_mov_b32_e32 v13, s29
-; VI-NEXT: s_cbranch_scc0 .LBB109_4
+; VI-NEXT: s_cbranch_scc0 .LBB105_4
; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: s_cbranch_execnz .LBB109_3
-; VI-NEXT: .LBB109_2: ; %cmp.true
+; VI-NEXT: s_cbranch_execnz .LBB105_3
+; VI-NEXT: .LBB105_2: ; %cmp.true
; VI-NEXT: v_mov_b32_e32 v18, 0x200
; VI-NEXT: v_add_f16_e32 v33, 0x200, v15
; VI-NEXT: v_add_f16_sdwa v15, v15, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
@@ -247591,11 +224277,11 @@ define inreg <64 x i16> @bitcast_v64f16_to_v64i16_scalar(<64 x half> inreg %a, i
; 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, v18
-; VI-NEXT: .LBB109_3: ; %end
+; VI-NEXT: .LBB105_3: ; %end
; VI-NEXT: v_mov_b32_e32 v18, v32
; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB109_4:
-; VI-NEXT: s_branch .LBB109_2
+; VI-NEXT: .LBB105_4:
+; VI-NEXT: s_branch .LBB105_2
;
; GFX9-LABEL: bitcast_v64f16_to_v64i16_scalar:
; GFX9: ; %bb.0:
@@ -247634,10 +224320,10 @@ define inreg <64 x i16> @bitcast_v64f16_to_v64i16_scalar(<64 x half> inreg %a, i
; GFX9-NEXT: v_mov_b32_e32 v11, s27
; GFX9-NEXT: v_mov_b32_e32 v12, s28
; GFX9-NEXT: v_mov_b32_e32 v13, s29
-; GFX9-NEXT: s_cbranch_scc0 .LBB109_4
+; GFX9-NEXT: s_cbranch_scc0 .LBB105_4
; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: s_cbranch_execnz .LBB109_3
-; GFX9-NEXT: .LBB109_2: ; %cmp.true
+; GFX9-NEXT: s_cbranch_execnz .LBB105_3
+; GFX9-NEXT: .LBB105_2: ; %cmp.true
; GFX9-NEXT: s_movk_i32 s4, 0x200
; GFX9-NEXT: v_pk_add_f16 v15, v15, s4 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v14, v14, s4 op_sel_hi:[1,0]
@@ -247671,11 +224357,11 @@ define inreg <64 x i16> @bitcast_v64f16_to_v64i16_scalar(<64 x half> inreg %a, i
; GFX9-NEXT: v_pk_add_f16 v32, v32, s4 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v17, v17, s4 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v16, v16, s4 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB109_3: ; %end
+; GFX9-NEXT: .LBB105_3: ; %end
; GFX9-NEXT: v_mov_b32_e32 v18, v32
; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB109_4:
-; GFX9-NEXT: s_branch .LBB109_2
+; GFX9-NEXT: .LBB105_4:
+; GFX9-NEXT: s_branch .LBB105_2
;
; GFX11-LABEL: bitcast_v64f16_to_v64i16_scalar:
; GFX11: ; %bb.0:
@@ -247695,11 +224381,11 @@ define inreg <64 x i16> @bitcast_v64f16_to_v64i16_scalar(<64 x half> inreg %a, i
; GFX11-NEXT: s_mov_b32 s12, s0
; GFX11-NEXT: s_mov_b32 s0, 0
; GFX11-NEXT: s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-NEXT: s_cbranch_scc0 .LBB109_3
+; GFX11-NEXT: s_cbranch_scc0 .LBB105_3
; GFX11-NEXT: ; %bb.1: ; %Flow
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_vccnz .LBB109_4
-; GFX11-NEXT: .LBB109_2: ; %cmp.true
+; GFX11-NEXT: s_cbranch_vccnz .LBB105_4
+; GFX11-NEXT: .LBB105_2: ; %cmp.true
; GFX11-NEXT: v_pk_add_f16 v15, 0x200, s27 op_sel_hi:[0,1]
; GFX11-NEXT: v_pk_add_f16 v14, 0x200, s26 op_sel_hi:[0,1]
; GFX11-NEXT: v_pk_add_f16 v13, 0x200, s25 op_sel_hi:[0,1]
@@ -247733,9 +224419,9 @@ define inreg <64 x i16> @bitcast_v64f16_to_v64i16_scalar(<64 x half> inreg %a, i
; 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: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB109_3:
-; GFX11-NEXT: s_branch .LBB109_2
-; GFX11-NEXT: .LBB109_4:
+; GFX11-NEXT: .LBB105_3:
+; GFX11-NEXT: s_branch .LBB105_2
+; GFX11-NEXT: .LBB105_4:
; GFX11-NEXT: v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-NEXT: v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
@@ -247953,7 +224639,7 @@ define <64 x half> @bitcast_v64i16_to_v64f16(<64 x i16> %a, i32 %b) {
; SI-NEXT: ; kill: killed $vgpr1
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB110_2
+; SI-NEXT: s_cbranch_execz .LBB106_2
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
@@ -248217,9 +224903,9 @@ define <64 x half> @bitcast_v64i16_to_v64f16(<64 x i16> %a, i32 %b) {
; SI-NEXT: ; kill: killed $vgpr1
; SI-NEXT: ; implicit-def: $vgpr1
; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: .LBB110_2: ; %Flow
+; SI-NEXT: .LBB106_2: ; %Flow
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB110_4
+; SI-NEXT: s_cbranch_execz .LBB106_4
; SI-NEXT: ; %bb.3: ; %cmp.true
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v43
@@ -248481,7 +225167,7 @@ define <64 x half> @bitcast_v64i16_to_v64f16(<64 x i16> %a, i32 %b) {
; SI-NEXT: s_waitcnt vmcnt(14)
; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; SI-NEXT: .LBB110_4: ; %end
+; SI-NEXT: .LBB106_4: ; %end
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
@@ -248863,7 +225549,7 @@ define <64 x half> @bitcast_v64i16_to_v64f16(<64 x i16> %a, i32 %b) {
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT: 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 .LBB110_2
+; VI-NEXT: s_cbranch_execz .LBB106_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
@@ -248963,7 +225649,7 @@ define <64 x half> @bitcast_v64i16_to_v64f16(<64 x i16> %a, i32 %b) {
; 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: .LBB110_2: ; %end
+; VI-NEXT: .LBB106_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]
@@ -248978,7 +225664,7 @@ define <64 x half> @bitcast_v64i16_to_v64f16(<64 x i16> %a, i32 %b) {
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT: 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 .LBB110_2
+; GFX9-NEXT: s_cbranch_execz .LBB106_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]
@@ -249013,7 +225699,7 @@ define <64 x half> @bitcast_v64i16_to_v64f16(<64 x i16> %a, i32 %b) {
; 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: .LBB110_2: ; %end
+; GFX9-NEXT: .LBB106_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]
@@ -249030,7 +225716,7 @@ define <64 x half> @bitcast_v64i16_to_v64f16(<64 x i16> %a, i32 %b) {
; GFX11-NEXT: s_xor_b32 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 .LBB110_2
+; GFX11-NEXT: s_cbranch_execz .LBB106_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]
@@ -249065,7 +225751,7 @@ define <64 x half> @bitcast_v64i16_to_v64f16(<64 x i16> %a, i32 %b) {
; 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: .LBB110_2: ; %end
+; GFX11-NEXT: .LBB106_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]
@@ -249159,7 +225845,7 @@ define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i
; SI-NEXT: s_waitcnt vmcnt(14)
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v41
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
-; SI-NEXT: s_cbranch_scc0 .LBB111_2
+; SI-NEXT: s_cbranch_scc0 .LBB107_2
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: v_cvt_f32_f16_e32 v4, v1
; SI-NEXT: v_cvt_f32_f16_e32 v41, s16
@@ -249294,8 +225980,8 @@ define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i
; SI-NEXT: v_cvt_f32_f16_e32 v4, v49
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT: v_cvt_f32_f16_e32 v43, v33
-; SI-NEXT: s_branch .LBB111_3
-; SI-NEXT: .LBB111_2:
+; SI-NEXT: s_branch .LBB107_3
+; SI-NEXT: .LBB107_2:
; SI-NEXT: ; implicit-def: $vgpr4
; SI-NEXT: ; kill: killed $vgpr4
; SI-NEXT: ; implicit-def: $vgpr41
@@ -249396,7 +226082,7 @@ define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i
; SI-NEXT: ; implicit-def: $vgpr4
; SI-NEXT: ; kill: killed $vgpr4
; SI-NEXT: ; implicit-def: $vgpr4
-; SI-NEXT: .LBB111_3: ; %Flow
+; SI-NEXT: .LBB107_3: ; %Flow
; SI-NEXT: s_andn2_b64 vcc, exec, s[4:5]
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_mov_b32_e32 v41, v42
@@ -249413,7 +226099,7 @@ define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i
; SI-NEXT: v_mov_b32_e32 v15, v17
; SI-NEXT: v_mov_b32_e32 v17, v19
; SI-NEXT: v_mov_b32_e32 v19, v1
-; SI-NEXT: s_cbranch_vccnz .LBB111_5
+; SI-NEXT: s_cbranch_vccnz .LBB107_5
; SI-NEXT: ; %bb.4: ; %cmp.true
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
@@ -249655,7 +226341,7 @@ define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f32_f16_e32 v1, v34
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; SI-NEXT: .LBB111_5: ; %end
+; SI-NEXT: .LBB107_5: ; %end
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
@@ -249994,10 +226680,10 @@ define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i
; VI-NEXT: s_and_b64 s[4:5], vcc, exec
; VI-NEXT: v_readfirstlane_b32 s7, v1
; VI-NEXT: v_writelane_b32 v32, s39, 7
-; VI-NEXT: s_cbranch_scc0 .LBB111_4
+; VI-NEXT: s_cbranch_scc0 .LBB107_4
; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: s_cbranch_execnz .LBB111_3
-; VI-NEXT: .LBB111_2: ; %cmp.true
+; VI-NEXT: s_cbranch_execnz .LBB107_3
+; VI-NEXT: .LBB107_2: ; %cmp.true
; VI-NEXT: s_and_b32 s4, s47, 0xffff0000
; VI-NEXT: s_add_i32 s5, s47, 3
; VI-NEXT: s_and_b32 s47, s46, 0xffff0000
@@ -250158,7 +226844,7 @@ define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i
; VI-NEXT: s_add_i32 s45, s45, 0x30000
; VI-NEXT: s_add_i32 s46, s46, 0x30000
; VI-NEXT: s_add_i32 s47, s4, 0x30000
-; VI-NEXT: .LBB111_3: ; %end
+; VI-NEXT: .LBB107_3: ; %end
; VI-NEXT: v_mov_b32_e32 v0, s16
; VI-NEXT: v_mov_b32_e32 v1, s17
; VI-NEXT: v_mov_b32_e32 v2, s18
@@ -250204,8 +226890,8 @@ define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i
; VI-NEXT: s_mov_b64 exec, s[4:5]
; VI-NEXT: s_waitcnt vmcnt(0)
; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB111_4:
-; VI-NEXT: s_branch .LBB111_2
+; VI-NEXT: .LBB107_4:
+; VI-NEXT: s_branch .LBB107_2
;
; GFX9-LABEL: bitcast_v64i16_to_v64f16_scalar:
; GFX9: ; %bb.0:
@@ -250244,10 +226930,10 @@ define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i
; GFX9-NEXT: v_mov_b32_e32 v11, s27
; GFX9-NEXT: v_mov_b32_e32 v12, s28
; GFX9-NEXT: v_mov_b32_e32 v13, s29
-; GFX9-NEXT: s_cbranch_scc0 .LBB111_4
+; GFX9-NEXT: s_cbranch_scc0 .LBB107_4
; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: s_cbranch_execnz .LBB111_3
-; GFX9-NEXT: .LBB111_2: ; %cmp.true
+; GFX9-NEXT: s_cbranch_execnz .LBB107_3
+; GFX9-NEXT: .LBB107_2: ; %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]
@@ -250280,11 +226966,11 @@ define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i
; GFX9-NEXT: v_pk_add_u16 v32, v32, 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: .LBB111_3: ; %end
+; GFX9-NEXT: .LBB107_3: ; %end
; GFX9-NEXT: v_mov_b32_e32 v18, v32
; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB111_4:
-; GFX9-NEXT: s_branch .LBB111_2
+; GFX9-NEXT: .LBB107_4:
+; GFX9-NEXT: s_branch .LBB107_2
;
; GFX11-LABEL: bitcast_v64i16_to_v64f16_scalar:
; GFX11: ; %bb.0:
@@ -250304,11 +226990,11 @@ define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i
; GFX11-NEXT: s_mov_b32 s12, s0
; GFX11-NEXT: s_mov_b32 s0, 0
; GFX11-NEXT: s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-NEXT: s_cbranch_scc0 .LBB111_3
+; GFX11-NEXT: s_cbranch_scc0 .LBB107_3
; GFX11-NEXT: ; %bb.1: ; %Flow
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_vccnz .LBB111_4
-; GFX11-NEXT: .LBB111_2: ; %cmp.true
+; GFX11-NEXT: s_cbranch_vccnz .LBB107_4
+; GFX11-NEXT: .LBB107_2: ; %cmp.true
; GFX11-NEXT: v_pk_add_u16 v15, s27, 3 op_sel_hi:[1,0]
; GFX11-NEXT: v_pk_add_u16 v14, s26, 3 op_sel_hi:[1,0]
; GFX11-NEXT: v_pk_add_u16 v13, s25, 3 op_sel_hi:[1,0]
@@ -250342,9 +227028,9 @@ define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i
; 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: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB111_3:
-; GFX11-NEXT: s_branch .LBB111_2
-; GFX11-NEXT: .LBB111_4:
+; GFX11-NEXT: .LBB107_3:
+; GFX11-NEXT: s_branch .LBB107_2
+; GFX11-NEXT: .LBB107_4:
; GFX11-NEXT: v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-NEXT: v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
>From cc8ee65438c1d359b9c0a69e425f3bff20f1cb61 Mon Sep 17 00:00:00 2001
From: shore <372660931 at qq.com>
Date: Wed, 14 May 2025 18:01:08 +0800
Subject: [PATCH 2/2] fix comments by Matthew, using -verify-machineinstrs=0
---
.../CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll | 23764 +++++++++++++++-
1 file changed, 23539 insertions(+), 225 deletions(-)
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll
index a67679c1293bb..eb725815db4b1 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll
@@ -1,10 +1,10 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
-; RUN: llc -mtriple=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefix=SI %s
-; RUN: llc -mtriple=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=VI %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX9 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11,GFX11-TRUE16 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11,GFX11-FAKE16 %s
+; RUN: llc -mtriple=amdgcn -mcpu=tahiti -verify-machineinstrs=0 < %s | FileCheck -check-prefix=SI %s
+; RUN: llc -mtriple=amdgcn -mcpu=tonga -verify-machineinstrs=0 < %s | FileCheck -check-prefix=VI %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx900 -verify-machineinstrs=0 < %s | FileCheck -check-prefix=GFX9 %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 -verify-machineinstrs=0 < %s | FileCheck -check-prefixes=GFX11,GFX11-TRUE16 %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -verify-machineinstrs=0 < %s | FileCheck -check-prefixes=GFX11,GFX11-FAKE16 %s
define <32 x float> @bitcast_v32i32_to_v32f32(<32 x i32> %a, i32 %b) {
; SI-LABEL: bitcast_v32i32_to_v32f32:
@@ -203181,6 +203181,23320 @@ end:
ret <128 x i8> %phi
}
+define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
+; SI-LABEL: bitcast_v128i8_to_v64i16:
+; SI: ; %bb.0:
+; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:940 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
+; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:392
+; SI-NEXT: s_waitcnt expcnt(1)
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:4
+; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:28
+; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:36
+; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:20
+; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:60
+; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:68
+; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:52
+; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:92
+; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:100
+; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:84
+; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:124
+; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:132
+; SI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:116
+; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:156
+; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:164
+; SI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:148
+; SI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:188
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:128
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 8, v14
+; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v16
+; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v12
+; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 8, v22
+; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v20
+; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 8, v30
+; SI-NEXT: v_lshlrev_b32_e32 v6, 8, v6
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 24, v8
+; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v28
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
+; SI-NEXT: v_lshlrev_b32_e32 v24, 24, v24
+; SI-NEXT: ; implicit-def: $vgpr43
+; SI-NEXT: ; implicit-def: $vgpr40
+; SI-NEXT: ; implicit-def: $vgpr22
+; SI-NEXT: ; implicit-def: $vgpr20
+; SI-NEXT: ; implicit-def: $vgpr56
+; SI-NEXT: ; implicit-def: $vgpr41
+; SI-NEXT: ; implicit-def: $vgpr39
+; SI-NEXT: ; implicit-def: $vgpr53
+; SI-NEXT: ; implicit-def: $vgpr50
+; SI-NEXT: ; implicit-def: $vgpr49
+; SI-NEXT: ; implicit-def: $vgpr48
+; SI-NEXT: ; implicit-def: $vgpr35
+; SI-NEXT: ; implicit-def: $vgpr55
+; SI-NEXT: ; implicit-def: $vgpr63
+; SI-NEXT: ; implicit-def: $vgpr58
+; SI-NEXT: ; implicit-def: $vgpr61
+; SI-NEXT: ; implicit-def: $vgpr52
+; SI-NEXT: ; implicit-def: $vgpr57
+; SI-NEXT: ; implicit-def: $vgpr28
+; SI-NEXT: ; implicit-def: $vgpr38
+; SI-NEXT: ; implicit-def: $vgpr46
+; SI-NEXT: ; implicit-def: $vgpr36
+; SI-NEXT: ; implicit-def: $vgpr44
+; SI-NEXT: ; kill: killed $vgpr43
+; SI-NEXT: ; implicit-def: $vgpr43
+; SI-NEXT: ; kill: killed $vgpr40
+; SI-NEXT: ; implicit-def: $vgpr40
+; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:360
+; SI-NEXT: s_waitcnt vmcnt(14)
+; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v25
+; SI-NEXT: v_lshlrev_b32_e32 v30, 24, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v7
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 24, v9
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 24, v5
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v13
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 24, v15
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 24, v11
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v19
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 24, v21
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt vmcnt(14)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:152
+; SI-NEXT: v_lshlrev_b32_e32 v3, 24, v17
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v23
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v27
+; SI-NEXT: v_lshlrev_b32_e32 v5, 24, v32
+; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:332
+; SI-NEXT: ; implicit-def: $vgpr23
+; SI-NEXT: ; implicit-def: $vgpr25
+; SI-NEXT: ; implicit-def: $vgpr32
+; SI-NEXT: ; implicit-def: $vgpr27
+; SI-NEXT: ; implicit-def: $vgpr21
+; SI-NEXT: ; kill: killed $vgpr23
+; SI-NEXT: ; implicit-def: $vgpr23
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:160
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:96
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:88
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:64
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:56
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:112
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:144
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:140
+; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v29
+; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 8, v31
+; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:184
+; SI-NEXT: v_lshlrev_b32_e32 v29, 8, v2
+; SI-NEXT: v_lshlrev_b32_e32 v2, 8, v18
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:108
+; SI-NEXT: ; implicit-def: $vgpr18
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:180
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:176
+; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v5, 24, v33
+; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v5, 8, v34
+; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
+; SI-NEXT: v_lshlrev_b32_e32 v33, 8, v10
+; SI-NEXT: ; implicit-def: $vgpr10
+; SI-NEXT: s_waitcnt vmcnt(4)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 24, v6
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
+; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:196
+; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:220
+; SI-NEXT: s_waitcnt expcnt(1)
+; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:192
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:172
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: v_lshlrev_b32_e32 v7, 24, v7
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:216
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
+; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:212
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:208
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 8, v8
+; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(1)
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:228
+; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:252
+; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:224
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:204
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 24, v6
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:248
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:244
+; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:240
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 24, v9
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 8, v7
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:236
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:260
+; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:284
+; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:256
+; SI-NEXT: s_waitcnt vmcnt(2)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 24, v6
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:280
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:276
+; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:272
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 24, v8
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 8, v7
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:928 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:268
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:292
+; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:316
+; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:288
+; SI-NEXT: s_waitcnt vmcnt(2)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 24, v6
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v34, 8, v7
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:312
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:844 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:308
+; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:304
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 24, v9
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:948 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: v_lshlrev_b32_e32 v51, 24, v8
+; SI-NEXT: s_waitcnt vmcnt(2)
+; SI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(1)
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:324
+; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:348
+; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:320
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:300
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: v_lshlrev_b32_e32 v37, 24, v6
+; SI-NEXT: s_waitcnt vmcnt(2)
+; SI-NEXT: v_lshlrev_b32_e32 v7, 8, v7
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:924 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:344
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:956 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:340
+; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:336
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v59, 24, v9
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:960 ; 4-byte Folded Spill
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:356
+; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:380
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:352
+; SI-NEXT: ; implicit-def: $vgpr9
+; SI-NEXT: s_waitcnt vmcnt(2)
+; SI-NEXT: v_lshlrev_b32_e32 v47, 24, v6
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v8, 8, v8
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:964 ; 4-byte Folded Spill
+; SI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:376
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:372
+; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:368
+; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:364
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:388
+; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:384
+; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v2, 8, v26
+; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; SI-NEXT: ; implicit-def: $vgpr26
+; SI-NEXT: s_waitcnt vmcnt(6)
+; SI-NEXT: v_lshlrev_b32_e32 v14, 24, v14
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: v_lshlrev_b32_e32 v19, 24, v6
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:12
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:8
+; SI-NEXT: v_lshlrev_b32_e32 v62, 8, v16
+; SI-NEXT: ; implicit-def: $vgpr16
+; SI-NEXT: ; kill: killed $vgpr16
+; SI-NEXT: ; implicit-def: $vgpr16
+; SI-NEXT: ; kill: killed $vgpr16
+; SI-NEXT: ; implicit-def: $vgpr16
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v2, 8, v2
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:120
+; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:44
+; SI-NEXT: v_lshlrev_b32_e32 v1, 8, v1
+; SI-NEXT: s_waitcnt vmcnt(2)
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v2, 8, v2
+; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:76
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:72
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v1, 8, v3
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
+; SI-NEXT: ; implicit-def: $vgpr3
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: v_lshlrev_b32_e32 v2, 8, v2
+; SI-NEXT: s_waitcnt vmcnt(2)
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:48
+; SI-NEXT: v_lshlrev_b32_e32 v1, 8, v4
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v1, 8, v5
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:920 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v1, 8, v11
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:952 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v1, 8, v12
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:972 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:32
+; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr5
+; SI-NEXT: ; implicit-def: $vgpr4
+; SI-NEXT: ; implicit-def: $vgpr11
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:24
+; SI-NEXT: v_lshlrev_b32_e32 v15, 8, v15
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
+; SI-NEXT: ; implicit-def: $vgpr6
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:80
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:40
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:16
+; SI-NEXT: v_lshlrev_b32_e32 v12, 8, v13
+; SI-NEXT: ; implicit-def: $vgpr13
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:104
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:136
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:168
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:200
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:232
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:264
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:296
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:944 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:328
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:968 ; 4-byte Folded Spill
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; SI-NEXT: s_cbranch_execz .LBB96_2
+; SI-NEXT: ; %bb.1: ; %cmp.false
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:972 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(4)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: v_or_b32_e32 v20, v1, v2
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v18, v2, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v3, v2, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v1, v1, v2
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v5, v2, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v9, v2, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v41, v1, v2
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v55, v24, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
+; SI-NEXT: v_mov_b32_e32 v24, v33
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v13, v2, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v1, v1, v2
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v26, v30, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v21, v2, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v32, v1, v2
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v28, v2, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v27, v2, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v1, v1, v2
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v35, v2, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v31, v2, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v63, v1, v2
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v48, v2, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v50, v2, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v11, v1, v2
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v2, v2, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v53, v4, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v6, v1, v4
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v1, v4, v1
+; SI-NEXT: v_and_b32_e32 v4, 0xff, v17
+; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT: v_or_b32_e32 v43, v10, v4
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v44, v4, v10
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
+; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v16, v10, v4
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
+; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v46, v10, v4
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v56, v4, v10
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
+; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v17, v10, v4
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
+; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v57, v10, v4
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v61, v4, v10
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
+; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v30, v10, v4
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
+; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v49, v10, v4
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v4, v4, v10
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v22, 0xff, v10
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload
+; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v23, v10, v22
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v22, 0xff, v10
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:948 ; 4-byte Folded Reload
+; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v33, v10, v22
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v22, 0xff, v10
+; SI-NEXT: v_mov_b32_e32 v10, v29
+; SI-NEXT: v_or_b32_e32 v29, v22, v34
+; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v22, 0xff, v22
+; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
+; SI-NEXT: v_or_b32_e32 v34, v37, v22
+; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v22, 0xff, v22
+; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
+; SI-NEXT: v_or_b32_e32 v37, v51, v22
+; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v22, 0xff, v22
+; SI-NEXT: v_or_b32_e32 v51, v22, v7
+; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v7, 0xff, v7
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v22, 0xff, v22
+; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
+; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
+; SI-NEXT: v_or_b32_e32 v7, v47, v7
+; SI-NEXT: v_or_b32_e32 v47, v59, v22
+; SI-NEXT: v_and_b32_e32 v22, 0xff, v54
+; SI-NEXT: v_or_b32_e32 v54, v22, v8
+; SI-NEXT: v_and_b32_e32 v8, 0xff, v60
+; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
+; SI-NEXT: v_or_b32_e32 v8, v19, v8
+; SI-NEXT: v_and_b32_e32 v19, 0xff, v42
+; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
+; SI-NEXT: v_or_b32_e32 v14, v14, v19
+; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v19, 0xff, v19
+; SI-NEXT: v_or_b32_e32 v19, v19, v10
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v22, 0xff, v10
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
+; SI-NEXT: v_or_b32_e32 v25, v22, v24
+; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v22, 0xff, v10
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v39, v22, v10
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v22, 0xff, v10
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v42, v22, v10
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v22, 0xff, v10
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v59, v22, v10
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v22, 0xff, v10
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v60, v22, v10
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v22, 0xff, v10
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v36, v22, v10
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v22, 0xff, v10
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v10, v22, v10
+; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v22, 0xff, v22
+; SI-NEXT: v_or_b32_e32 v38, v22, v24
+; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v22, 0xff, v22
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v58, v22, v24
+; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v22, 0xff, v22
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v52, v22, v24
+; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v22, 0xff, v22
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v24, v22, v24
+; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v22, 0xff, v22
+; SI-NEXT: v_or_b32_e32 v40, v22, v40
+; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:944 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v22, 0xff, v22
+; SI-NEXT: v_or_b32_e32 v12, v22, v12
+; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:968 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v22, 0xff, v22
+; SI-NEXT: v_or_b32_e32 v15, v22, v15
+; SI-NEXT: v_and_b32_e32 v22, 0xff, v45
+; SI-NEXT: v_or_b32_e32 v45, v22, v62
+; SI-NEXT: v_mov_b32_e32 v62, v18
+; SI-NEXT: v_or_b32_e32 v20, v20, v62
+; SI-NEXT: v_or_b32_e32 v22, v19, v3
+; SI-NEXT: v_alignbit_b32 v3, v20, v3, 16
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
+; SI-NEXT: v_mov_b32_e32 v19, v5
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT: v_or_b32_e32 v5, v3, v19
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v25
+; SI-NEXT: v_or_b32_e32 v25, v3, v9
+; SI-NEXT: v_alignbit_b32 v3, v5, v9, 16
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v41
+; SI-NEXT: v_or_b32_e32 v41, v3, v55
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v39
+; SI-NEXT: v_or_b32_e32 v18, v3, v13
+; SI-NEXT: v_alignbit_b32 v3, v41, v13, 16
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT: v_or_b32_e32 v13, v3, v26
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v42
+; SI-NEXT: v_or_b32_e32 v39, v3, v21
+; SI-NEXT: v_alignbit_b32 v3, v13, v21, 16
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v32
+; SI-NEXT: v_or_b32_e32 v32, v3, v28
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v59
+; SI-NEXT: v_or_b32_e32 v42, v3, v27
+; SI-NEXT: v_alignbit_b32 v3, v32, v27, 16
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
+; SI-NEXT: ; implicit-def: $vgpr59
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT: v_or_b32_e32 v27, v3, v35
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v60
+; SI-NEXT: v_or_b32_e32 v3, v3, v31
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v3, v27, v31, 16
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v63
+; SI-NEXT: v_or_b32_e32 v21, v3, v48
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v36
+; SI-NEXT: v_or_b32_e32 v31, v3, v50
+; SI-NEXT: v_alignbit_b32 v3, v21, v50, 16
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v11
+; SI-NEXT: v_or_b32_e32 v11, v3, v2
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v10
+; SI-NEXT: v_or_b32_e32 v3, v3, v53
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v3, v11, v53, 16
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v6
+; SI-NEXT: v_and_b32_e32 v6, 0xffff, v38
+; SI-NEXT: v_or_b32_e32 v3, v3, v1
+; SI-NEXT: v_or_b32_e32 v6, v6, v43
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v6, v3, v43, 16
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v6, 0xffff, v44
+; SI-NEXT: v_or_b32_e32 v9, v6, v16
+; SI-NEXT: v_and_b32_e32 v6, 0xffff, v58
+; SI-NEXT: v_or_b32_e32 v6, v6, v46
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v6, v9, v46, 16
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v6, 0xffff, v56
+; SI-NEXT: v_or_b32_e32 v63, v6, v17
+; SI-NEXT: v_and_b32_e32 v6, 0xffff, v52
+; SI-NEXT: v_or_b32_e32 v6, v6, v57
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v6, v63, v57, 16
+; SI-NEXT: v_or_b32_e32 v57, v4, v23
+; SI-NEXT: v_and_b32_e32 v4, 0xffff, v40
+; SI-NEXT: v_or_b32_e32 v52, v4, v33
+; SI-NEXT: v_alignbit_b32 v4, v57, v33, 16
+; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v4, 0xffff, v29
+; SI-NEXT: v_or_b32_e32 v46, v4, v34
+; SI-NEXT: v_and_b32_e32 v4, 0xffff, v12
+; SI-NEXT: v_or_b32_e32 v38, v4, v37
+; SI-NEXT: v_alignbit_b32 v4, v46, v37, 16
+; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v4, 0xffff, v51
+; SI-NEXT: v_or_b32_e32 v44, v4, v7
+; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v7
+; SI-NEXT: v_lshrrev_b32_e32 v7, 16, v8
+; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: v_and_b32_e32 v4, 0xffff, v15
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: v_or_b32_e32 v36, v4, v47
+; SI-NEXT: v_alignbit_b32 v4, v44, v47, 16
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v4, 0xffff, v54
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: v_or_b32_e32 v43, v4, v8
+; SI-NEXT: v_and_b32_e32 v4, 0xffff, v45
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: v_or_b32_e32 v12, v4, v14
+; SI-NEXT: v_alignbit_b32 v4, v43, v14, 16
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v4, 16, v28
+; SI-NEXT: v_lshrrev_b32_e32 v28, 16, v23
+; SI-NEXT: v_mov_b32_e32 v23, v12
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr12
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr12
+; SI-NEXT: ; implicit-def: $vgpr12
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr12
+; SI-NEXT: ; implicit-def: $vgpr12
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr12
+; SI-NEXT: ; implicit-def: $vgpr12
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr12
+; SI-NEXT: ; implicit-def: $vgpr12
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr12
+; SI-NEXT: ; implicit-def: $vgpr12
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr12
+; SI-NEXT: ; implicit-def: $vgpr12
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr12
+; SI-NEXT: ; implicit-def: $vgpr12
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v6, 0xffff, v61
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr12
+; SI-NEXT: ; implicit-def: $vgpr12
+; SI-NEXT: v_or_b32_e32 v61, v6, v30
+; SI-NEXT: v_and_b32_e32 v6, 0xffff, v24
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr12
+; SI-NEXT: ; implicit-def: $vgpr12
+; SI-NEXT: v_or_b32_e32 v58, v6, v49
+; SI-NEXT: v_alignbit_b32 v6, v61, v49, 16
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; kill: killed $vgpr12
+; SI-NEXT: ; implicit-def: $vgpr12
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v6, 16, v62
+; SI-NEXT: v_lshrrev_b32_e32 v56, 16, v19
+; SI-NEXT: v_lshrrev_b32_e32 v10, 16, v55
+; SI-NEXT: v_lshrrev_b32_e32 v53, 16, v26
+; SI-NEXT: v_mov_b32_e32 v26, v42
+; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v35
+; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v48
+; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v2
+; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v1
+; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v16
+; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v17
+; SI-NEXT: v_lshrrev_b32_e32 v2, 16, v30
+; SI-NEXT: v_lshrrev_b32_e32 v16, 16, v34
+; SI-NEXT: ; implicit-def: $vgpr17
+; SI-NEXT: ; implicit-def: $vgpr45
+; SI-NEXT: ; implicit-def: $vgpr42
+; SI-NEXT: ; implicit-def: $vgpr54
+; SI-NEXT: ; implicit-def: $vgpr60
+; SI-NEXT: ; implicit-def: $vgpr24
+; SI-NEXT: ; implicit-def: $vgpr30
+; SI-NEXT: ; kill: killed $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr34
+; SI-NEXT: ; implicit-def: $vgpr37
+; SI-NEXT: ; implicit-def: $vgpr51
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr47
+; SI-NEXT: ; implicit-def: $vgpr8
+; SI-NEXT: ; implicit-def: $vgpr19
+; SI-NEXT: ; implicit-def: $vgpr14
+; SI-NEXT: ; implicit-def: $vgpr29
+; SI-NEXT: ; implicit-def: $vgpr33
+; SI-NEXT: ; kill: killed $vgpr12
+; SI-NEXT: ; implicit-def: $vgpr12
+; SI-NEXT: ; implicit-def: $vgpr15
+; SI-NEXT: ; implicit-def: $vgpr62
+; SI-NEXT: .LBB96_2: ; %Flow
+; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; SI-NEXT: s_cbranch_execz .LBB96_4
+; SI-NEXT: ; %bb.3: ; %cmp.true
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v45
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v42
+; SI-NEXT: v_or_b32_e32 v1, v62, v1
+; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT: v_add_i32_e32 v1, vcc, 0x300, v1
+; SI-NEXT: v_or_b32_e32 v2, v14, v2
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v54
+; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
+; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v60
+; SI-NEXT: v_or_b32_e32 v2, v8, v2
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: s_movk_i32 s6, 0x300
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_add_i32_e32 v2, vcc, s6, v2
+; SI-NEXT: v_or_b32_e32 v3, v19, v3
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v2, v3, v2
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:968 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
+; SI-NEXT: s_mov_b32 s7, 0x3000000
+; SI-NEXT: v_add_i32_e32 v43, vcc, s7, v2
+; SI-NEXT: v_mov_b32_e32 v28, v24
+; SI-NEXT: v_add_i32_e32 v1, vcc, s7, v1
+; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:972 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:948 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(14)
+; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v3
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4
+; SI-NEXT: v_or_b32_e32 v3, v15, v3
+; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
+; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT: v_add_i32_e32 v3, vcc, s6, v3
+; SI-NEXT: v_or_b32_e32 v4, v59, v4
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5
+; SI-NEXT: v_and_b32_e32 v5, 0xff, v5
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT: v_or_b32_e32 v5, v47, v5
+; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6
+; SI-NEXT: v_and_b32_e32 v6, 0xff, v6
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6
+; SI-NEXT: v_or_b32_e32 v6, v51, v6
+; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4
+; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
+; SI-NEXT: v_or_b32_e32 v4, v7, v4
+; SI-NEXT: v_add_i32_e32 v4, vcc, s6, v4
+; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; SI-NEXT: v_or_b32_e32 v4, v5, v4
+; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:944 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v44, vcc, s7, v4
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v44
+; SI-NEXT: s_waitcnt vmcnt(2)
+; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5
+; SI-NEXT: v_and_b32_e32 v5, 0xff, v5
+; SI-NEXT: v_or_b32_e32 v5, v12, v5
+; SI-NEXT: v_add_i32_e32 v5, vcc, s6, v5
+; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; SI-NEXT: v_or_b32_e32 v5, v6, v5
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(2)
+; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v7
+; SI-NEXT: v_and_b32_e32 v7, 0xff, v7
+; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
+; SI-NEXT: v_or_b32_e32 v7, v37, v7
+; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v35, vcc, s7, v5
+; SI-NEXT: v_mov_b32_e32 v38, v35
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6
+; SI-NEXT: v_and_b32_e32 v6, 0xff, v6
+; SI-NEXT: v_or_b32_e32 v6, v34, v6
+; SI-NEXT: v_add_i32_e32 v6, vcc, s6, v6
+; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6
+; SI-NEXT: v_or_b32_e32 v6, v7, v6
+; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v46, vcc, s7, v6
+; SI-NEXT: v_add_i32_e32 v34, vcc, s7, v3
+; SI-NEXT: v_mov_b32_e32 v36, v34
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v7
+; SI-NEXT: v_and_b32_e32 v7, 0xff, v7
+; SI-NEXT: v_or_b32_e32 v7, v8, v7
+; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v7, vcc, s6, v7
+; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8
+; SI-NEXT: v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
+; SI-NEXT: v_or_b32_e32 v8, v9, v8
+; SI-NEXT: v_or_b32_e32 v7, v8, v7
+; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v6, vcc, s7, v7
+; SI-NEXT: v_mov_b32_e32 v52, v6
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8
+; SI-NEXT: v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v8, v9, v8
+; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v8, vcc, s6, v8
+; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v9
+; SI-NEXT: v_and_b32_e32 v9, 0xff, v9
+; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9
+; SI-NEXT: v_or_b32_e32 v9, v10, v9
+; SI-NEXT: v_or_b32_e32 v8, v9, v8
+; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v57, vcc, s7, v8
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v9
+; SI-NEXT: v_and_b32_e32 v9, 0xff, v9
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v9, v10, v9
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v9, vcc, s6, v9
+; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10
+; SI-NEXT: v_and_b32_e32 v10, 0xff, v10
+; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
+; SI-NEXT: v_or_b32_e32 v10, v11, v10
+; SI-NEXT: v_or_b32_e32 v9, v10, v9
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v7, vcc, s7, v9
+; SI-NEXT: v_mov_b32_e32 v58, v7
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10
+; SI-NEXT: v_and_b32_e32 v10, 0xff, v10
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v10, v11, v10
+; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v10, vcc, s6, v10
+; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11
+; SI-NEXT: v_and_b32_e32 v11, 0xff, v11
+; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11
+; SI-NEXT: v_or_b32_e32 v11, v12, v11
+; SI-NEXT: v_or_b32_e32 v10, v11, v10
+; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v61, vcc, s7, v10
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11
+; SI-NEXT: v_and_b32_e32 v11, 0xff, v11
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v11, v12, v11
+; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v11, vcc, s6, v11
+; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12
+; SI-NEXT: v_and_b32_e32 v12, 0xff, v12
+; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12
+; SI-NEXT: v_or_b32_e32 v12, v13, v12
+; SI-NEXT: v_or_b32_e32 v11, v12, v11
+; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v8, vcc, s7, v11
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12
+; SI-NEXT: v_and_b32_e32 v12, 0xff, v12
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v12, v13, v12
+; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v12, vcc, s6, v12
+; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v13
+; SI-NEXT: v_and_b32_e32 v13, 0xff, v13
+; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
+; SI-NEXT: v_or_b32_e32 v13, v14, v13
+; SI-NEXT: v_or_b32_e32 v12, v13, v12
+; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v63, vcc, s7, v12
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v13
+; SI-NEXT: v_and_b32_e32 v13, 0xff, v13
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v13, v14, v13
+; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v13, vcc, s6, v13
+; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14
+; SI-NEXT: v_and_b32_e32 v14, 0xff, v14
+; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
+; SI-NEXT: v_or_b32_e32 v14, v15, v14
+; SI-NEXT: v_or_b32_e32 v13, v14, v13
+; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v10, vcc, s7, v13
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14
+; SI-NEXT: v_and_b32_e32 v14, 0xff, v14
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v14, v15, v14
+; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v14, vcc, s6, v14
+; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v15
+; SI-NEXT: v_and_b32_e32 v15, 0xff, v15
+; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
+; SI-NEXT: v_or_b32_e32 v15, v16, v15
+; SI-NEXT: v_or_b32_e32 v14, v15, v14
+; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v9, vcc, s7, v14
+; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v9
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v15
+; SI-NEXT: v_and_b32_e32 v15, 0xff, v15
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v15, v16, v15
+; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v17
+; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v16, 0xff, v16
+; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
+; SI-NEXT: v_add_i32_e32 v15, vcc, s6, v15
+; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v16, v17, v16
+; SI-NEXT: v_or_b32_e32 v15, v16, v15
+; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v12, vcc, s7, v15
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16
+; SI-NEXT: v_and_b32_e32 v16, 0xff, v16
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v16, v17, v16
+; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v16, vcc, s6, v16
+; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v17
+; SI-NEXT: v_and_b32_e32 v17, 0xff, v17
+; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17
+; SI-NEXT: v_or_b32_e32 v17, v18, v17
+; SI-NEXT: v_or_b32_e32 v16, v17, v16
+; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v3, vcc, s7, v16
+; SI-NEXT: v_lshrrev_b32_e32 v16, 16, v46
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v17
+; SI-NEXT: v_and_b32_e32 v17, 0xff, v17
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v17, v18, v17
+; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v17, vcc, s6, v17
+; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18
+; SI-NEXT: v_and_b32_e32 v18, 0xff, v18
+; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
+; SI-NEXT: v_or_b32_e32 v18, v19, v18
+; SI-NEXT: v_or_b32_e32 v17, v18, v17
+; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v14, vcc, s7, v17
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18
+; SI-NEXT: v_and_b32_e32 v18, 0xff, v18
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v18, v19, v18
+; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v18, vcc, s6, v18
+; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v19
+; SI-NEXT: v_and_b32_e32 v19, 0xff, v19
+; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
+; SI-NEXT: v_or_b32_e32 v19, v20, v19
+; SI-NEXT: v_or_b32_e32 v18, v19, v18
+; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v11, vcc, s7, v18
+; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v11
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v19
+; SI-NEXT: v_and_b32_e32 v19, 0xff, v19
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v19, v20, v19
+; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v19, vcc, s6, v19
+; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
+; SI-NEXT: v_and_b32_e32 v20, 0xff, v20
+; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20
+; SI-NEXT: v_or_b32_e32 v20, v21, v20
+; SI-NEXT: v_or_b32_e32 v19, v20, v19
+; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v31, vcc, s7, v19
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
+; SI-NEXT: v_and_b32_e32 v20, 0xff, v20
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v20, v21, v20
+; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v20, vcc, s6, v20
+; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
+; SI-NEXT: v_and_b32_e32 v21, 0xff, v21
+; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21
+; SI-NEXT: v_or_b32_e32 v21, v22, v21
+; SI-NEXT: v_or_b32_e32 v20, v21, v20
+; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
+; SI-NEXT: v_and_b32_e32 v21, 0xff, v21
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v21, v22, v21
+; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v21, vcc, s6, v21
+; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v22, vcc, 3, v22
+; SI-NEXT: v_and_b32_e32 v22, 0xff, v22
+; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
+; SI-NEXT: v_or_b32_e32 v22, v23, v22
+; SI-NEXT: v_or_b32_e32 v22, v22, v21
+; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v15, vcc, s7, v22
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
+; SI-NEXT: v_and_b32_e32 v21, 0xff, v21
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v21, v23, v21
+; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v21, vcc, s6, v21
+; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v23, vcc, 3, v23
+; SI-NEXT: v_and_b32_e32 v23, 0xff, v23
+; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23
+; SI-NEXT: v_or_b32_e32 v23, v25, v23
+; SI-NEXT: v_or_b32_e32 v23, v23, v21
+; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v27, vcc, s7, v23
+; SI-NEXT: v_mov_b32_e32 v23, v1
+; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v27
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
+; SI-NEXT: v_and_b32_e32 v21, 0xff, v21
+; SI-NEXT: v_or_b32_e32 v24, v24, v21
+; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v24
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
+; SI-NEXT: v_and_b32_e32 v21, 0xff, v21
+; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21
+; SI-NEXT: v_or_b32_e32 v25, v25, v21
+; SI-NEXT: v_or_b32_e32 v1, v25, v1
+; SI-NEXT: v_add_i32_e32 v26, vcc, s7, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v21, vcc, s7, v20
+; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v21
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
+; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT: v_or_b32_e32 v2, v4, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: v_add_i32_e32 v32, vcc, s7, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(2)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
+; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT: v_or_b32_e32 v2, v4, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: v_add_i32_e32 v39, vcc, s7, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(2)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
+; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT: v_or_b32_e32 v2, v30, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: v_add_i32_e32 v13, vcc, s7, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; SI-NEXT: v_lshrrev_b32_e32 v53, 16, v13
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
+; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT: v_or_b32_e32 v2, v4, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: v_add_i32_e32 v18, vcc, s7, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(2)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
+; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT: v_or_b32_e32 v2, v28, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: v_add_i32_e32 v41, vcc, s7, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
+; SI-NEXT: v_lshrrev_b32_e32 v28, 16, v57
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
+; SI-NEXT: v_or_b32_e32 v1, v33, v1
+; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
+; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v2, v4, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: v_add_i32_e32 v25, vcc, s7, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(2)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
+; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT: v_or_b32_e32 v2, v4, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: v_add_i32_e32 v5, vcc, s7, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
+; SI-NEXT: v_lshrrev_b32_e32 v56, 16, v5
+; SI-NEXT: s_waitcnt vmcnt(2)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
+; SI-NEXT: v_or_b32_e32 v1, v29, v1
+; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
+; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v2, v4, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: v_add_i32_e32 v22, vcc, s7, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(2)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
+; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT: v_or_b32_e32 v2, v4, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: v_add_i32_e32 v20, vcc, s7, v1
+; SI-NEXT: v_alignbit_b32 v1, v20, v22, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v5, v25, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v41, v18, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v13, v39, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v32, v26, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(1)
+; SI-NEXT: v_alignbit_b32 v1, v27, v15, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v21, v31, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(1)
+; SI-NEXT: v_alignbit_b32 v1, v11, v14, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(1)
+; SI-NEXT: v_alignbit_b32 v1, v3, v12, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(1)
+; SI-NEXT: v_alignbit_b32 v1, v9, v10, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(1)
+; SI-NEXT: v_alignbit_b32 v1, v63, v8, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v61, v7, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v57, v6, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v46, v35, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v44, v34, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v43, v23, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
+; SI-NEXT: v_lshrrev_b32_e32 v6, 16, v20
+; SI-NEXT: v_lshrrev_b32_e32 v10, 16, v41
+; SI-NEXT: v_lshrrev_b32_e32 v4, 16, v32
+; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v3
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v63
+; SI-NEXT: v_lshrrev_b32_e32 v2, 16, v61
+; SI-NEXT: v_lshrrev_b32_e32 v7, 16, v43
+; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill
+; SI-NEXT: .LBB96_4: ; %end
+; SI-NEXT: s_or_b64 exec, exec, s[4:5]
+; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:848 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v7, 0xffff, v22
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6
+; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
+; SI-NEXT: v_or_b32_e32 v7, v7, v8
+; SI-NEXT: buffer_store_dword v7, v0, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v7, 0xffff, v20
+; SI-NEXT: v_or_b32_e32 v6, v7, v6
+; SI-NEXT: v_add_i32_e32 v7, vcc, 4, v0
+; SI-NEXT: buffer_store_dword v6, v7, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v6, 0xffff, v25
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
+; SI-NEXT: v_or_b32_e32 v6, v6, v7
+; SI-NEXT: v_add_i32_e32 v7, vcc, 8, v0
+; SI-NEXT: buffer_store_dword v6, v7, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v56
+; SI-NEXT: v_or_b32_e32 v5, v5, v6
+; SI-NEXT: v_add_i32_e32 v6, vcc, 12, v0
+; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v5, 0xffff, v18
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6
+; SI-NEXT: v_or_b32_e32 v5, v5, v6
+; SI-NEXT: v_add_i32_e32 v6, vcc, 16, v0
+; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v5, 0xffff, v41
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v10
+; SI-NEXT: v_or_b32_e32 v5, v5, v6
+; SI-NEXT: v_add_i32_e32 v6, vcc, 20, v0
+; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v5, 0xffff, v39
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6
+; SI-NEXT: v_or_b32_e32 v5, v5, v6
+; SI-NEXT: v_add_i32_e32 v6, vcc, 24, v0
+; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v5, 0xffff, v13
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v53
+; SI-NEXT: v_or_b32_e32 v5, v5, v6
+; SI-NEXT: v_add_i32_e32 v6, vcc, 28, v0
+; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:864 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v5, 0xffff, v26
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6
+; SI-NEXT: v_or_b32_e32 v5, v5, v6
+; SI-NEXT: v_add_i32_e32 v6, vcc, 32, v0
+; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v5, 0xffff, v32
+; SI-NEXT: v_or_b32_e32 v4, v5, v4
+; SI-NEXT: v_add_i32_e32 v5, vcc, 36, v0
+; SI-NEXT: buffer_store_dword v4, v5, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:868 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT: v_or_b32_e32 v4, v4, v5
+; SI-NEXT: v_add_i32_e32 v5, vcc, 40, v0
+; SI-NEXT: buffer_store_dword v4, v5, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v4, 0xffff, v27
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v50
+; SI-NEXT: v_or_b32_e32 v4, v4, v5
+; SI-NEXT: v_add_i32_e32 v5, vcc, 44, v0
+; SI-NEXT: buffer_store_dword v4, v5, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v4, 0xffff, v31
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT: v_or_b32_e32 v4, v4, v5
+; SI-NEXT: v_add_i32_e32 v5, vcc, 48, v0
+; SI-NEXT: buffer_store_dword v4, v5, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v4, 0xffff, v21
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v49
+; SI-NEXT: v_or_b32_e32 v4, v4, v5
+; SI-NEXT: v_add_i32_e32 v5, vcc, 52, v0
+; SI-NEXT: buffer_store_dword v4, v5, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT: v_or_b32_e32 v4, v4, v5
+; SI-NEXT: v_add_i32_e32 v5, vcc, 56, v0
+; SI-NEXT: buffer_store_dword v4, v5, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v4, 0xffff, v11
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v48
+; SI-NEXT: v_or_b32_e32 v4, v4, v5
+; SI-NEXT: v_add_i32_e32 v5, vcc, 60, v0
+; SI-NEXT: buffer_store_dword v4, v5, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT: v_or_b32_e32 v4, v4, v5
+; SI-NEXT: v_add_i32_e32 v5, vcc, 64, v0
+; SI-NEXT: buffer_store_dword v4, v5, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v35
+; SI-NEXT: v_or_b32_e32 v3, v3, v4
+; SI-NEXT: v_add_i32_e32 v4, vcc, 0x44, v0
+; SI-NEXT: buffer_store_dword v3, v4, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:884 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT: v_or_b32_e32 v3, v3, v4
+; SI-NEXT: v_add_i32_e32 v4, vcc, 0x48, v0
+; SI-NEXT: buffer_store_dword v3, v4, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v9
+; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v55
+; SI-NEXT: v_or_b32_e32 v3, v3, v4
+; SI-NEXT: v_add_i32_e32 v4, vcc, 0x4c, v0
+; SI-NEXT: buffer_store_dword v3, v4, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT: v_or_b32_e32 v3, v3, v4
+; SI-NEXT: v_add_i32_e32 v4, vcc, 0x50, v0
+; SI-NEXT: buffer_store_dword v3, v4, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v63
+; SI-NEXT: v_or_b32_e32 v1, v3, v1
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x54, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v58
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x58, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v61
+; SI-NEXT: v_or_b32_e32 v1, v1, v2
+; SI-NEXT: v_add_i32_e32 v2, vcc, 0x5c, v0
+; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v52
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT: v_or_b32_e32 v1, v1, v2
+; SI-NEXT: v_add_i32_e32 v2, vcc, 0x60, v0
+; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v57
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v28
+; SI-NEXT: v_or_b32_e32 v1, v1, v2
+; SI-NEXT: v_add_i32_e32 v2, vcc, 0x64, v0
+; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:900 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v38
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT: v_or_b32_e32 v1, v1, v2
+; SI-NEXT: v_add_i32_e32 v2, vcc, 0x68, v0
+; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v46
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v16
+; SI-NEXT: v_or_b32_e32 v1, v1, v2
+; SI-NEXT: v_add_i32_e32 v2, vcc, 0x6c, v0
+; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v36
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT: v_or_b32_e32 v1, v1, v2
+; SI-NEXT: v_add_i32_e32 v2, vcc, 0x70, v0
+; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v44
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v40
+; SI-NEXT: v_or_b32_e32 v1, v1, v2
+; SI-NEXT: v_add_i32_e32 v2, vcc, 0x74, v0
+; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v23
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT: v_or_b32_e32 v1, v1, v2
+; SI-NEXT: v_add_i32_e32 v2, vcc, 0x78, v0
+; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:936 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v43
+; SI-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT: v_or_b32_e32 v1, v1, v2
+; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; SI-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 .LBB96_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: .LBB96_2: ; %Flow
+; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; VI-NEXT: s_cbranch_execz .LBB96_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: .LBB96_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 .LBB96_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: .LBB96_2: ; %Flow
+; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX9-NEXT: s_cbranch_execz .LBB96_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: .LBB96_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-TRUE16-LABEL: bitcast_v128i8_to_v64i16:
+; GFX11-TRUE16: ; %bb.0:
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT: s_clause 0x1f
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v31, off, s32 offset:384
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v39, off, s32 offset:380
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:376
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:372
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v150, off, s32 offset:368
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v39, off, s32 offset:364
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v150, off, s32 offset:360
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v50, off, s32 offset:356
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v147, off, s32 offset:352
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:348
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v149, off, s32 offset:344
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v49, off, s32 offset:340
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v149, off, s32 offset:336
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:332
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v148, off, s32 offset:328
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v48, off, s32 offset:324
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v145, off, s32 offset:320
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:316
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v148, off, s32 offset:312
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v48, off, s32 offset:308
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v147, off, s32 offset:304
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:300
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v146, off, s32 offset:296
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:292
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v146, off, s32 offset:288
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:284
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v145, off, s32 offset:280
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:276
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v134, off, s32 offset:272
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:268
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v144, off, s32 offset:264
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:260
+; GFX11-TRUE16-NEXT: s_clause 0x1f
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v144, off, s32 offset:256
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:252
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v135, off, s32 offset:248
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:244
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v132, off, s32 offset:240
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:236
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v135, off, s32 offset:232
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:228
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v134, off, s32 offset:224
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:220
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v133, off, s32 offset:216
+; GFX11-TRUE16-NEXT: scratch_load_b32 v160, off, s32 offset:388
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v101, off, s32
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v103, off, s32 offset:8
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v113, off, s32 offset:16
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v113, off, s32 offset:24
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v114, off, s32 offset:32
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v114, off, s32 offset:40
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v115, off, s32 offset:48
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v116, off, s32 offset:56
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v116, off, s32 offset:64
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v117, off, s32 offset:72
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v117, off, s32 offset:80
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v118, off, s32 offset:88
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v118, off, s32 offset:96
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v119, off, s32 offset:104
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v119, off, s32 offset:112
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v128, off, s32 offset:120
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v128, off, s32 offset:128
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v129, off, s32 offset:136
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v129, off, s32 offset:144
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v130, off, s32 offset:152
+; GFX11-TRUE16-NEXT: s_clause 0x1f
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v130, off, s32 offset:160
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v131, off, s32 offset:168
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v131, off, s32 offset:176
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v132, off, s32 offset:184
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v133, off, s32 offset:192
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v151, off, s32 offset:200
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v151, off, s32 offset:208
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:212
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:204
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:196
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:188
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:180
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v86, off, s32 offset:172
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v96, off, s32 offset:164
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:156
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:148
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32 offset:140
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v84, off, s32 offset:132
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v86, off, s32 offset:124
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:116
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:108
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:100
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v82, off, s32 offset:92
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:84
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:76
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v83, off, s32 offset:68
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:60
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:52
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v68, off, s32 offset:44
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v70, off, s32 offset:36
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:28
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:20
+; GFX11-TRUE16-NEXT: s_clause 0x1
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:12
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:4
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v30.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v28.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.h, v26.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.h, v24.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v22.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v20.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.h, v16.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.h, v14.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.h, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.h, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.h, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, v0.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v81.l, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v81.h, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.h, 8, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.l, 8, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.h, 8, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.l, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v98.h, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.l, 8, v15.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.h, 8, v17.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.l, 8, v19.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.l, 8, v23.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.h, 8, v25.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.l, 8, v27.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v112.l, 8, v29.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(62)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v150.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.l, 8, v150.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.h, 8, v147.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v149.h, 8, v149.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v149.l, 8, v149.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.h, 8, v148.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.l, 8, v145.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.l, 8, v148.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v147.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v146.h, 8, v146.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v146.l, 8, v146.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.h, 8, v145.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v134.h, 8, v134.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v144.h, 8, v144.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v144.l, 8, v144.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.h, 8, v135.h
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(61)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.l, 8, v132.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(59)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.l, 8, v135.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(57)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v134.l, 8, v134.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(55)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.h, 8, v133.h
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(54)
+; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v160
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(53)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v101.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(52)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v112.h, 8, v103.h
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(51)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v113.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(50)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.h, 8, v113.h
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(49)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.h, 8, v114.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(48)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v115.l, 8, v114.h
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(47)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v115.h, 8, v115.h
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(46)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.l, 8, v116.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(45)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.h, 8, v116.h
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(44)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.l, 8, v117.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(43)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.l, 8, v117.h
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(42)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v118.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(41)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.h, 8, v118.h
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(40)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v119.l, 8, v119.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(39)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.h, 8, v119.h
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(38)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v119.h, 8, v128.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(37)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.l, 8, v128.h
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(36)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.h, 8, v129.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(35)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.l, 8, v129.h
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(34)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.l, 8, v130.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(33)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.h, 8, v130.h
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(32)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.h, 8, v131.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(31)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v131.l, 8, v131.h
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(30)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v131.h, 8, v132.h
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(29)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.h, 8, v133.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(28)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.h, 8, v151.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(27)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.l, 8, v151.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.l, 8, v31.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v31.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
+; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
+; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB96_3
+; GFX11-TRUE16-NEXT: ; %bb.1: ; %Flow
+; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
+; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB96_4
+; GFX11-TRUE16-NEXT: .LBB96_2: ; %end
+; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
+; GFX11-TRUE16-NEXT: .LBB96_3: ; %cmp.false
+; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v52.h
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v51.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v53.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v51.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v54.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v49.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v55.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v53.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v55.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v52.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v65.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v64.l
+; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v65.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v64.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v67.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v54.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v67.h
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v66.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v68.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v66.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v70.h
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v68.h
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v71.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v69.h
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v83.h
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v69.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v84.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v82.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v85.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v80.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v96.h
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v86.h
+; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v84.l
+; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v80.h
+; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v87.h
+; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v71.h
+; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v96.l
+; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v86.l
+; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v97.l
+; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v85.h
+; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v100.l
+; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v98.l
+; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v100.h
+; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v32.h
+; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v34.h
+; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v32.l
+; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v35.l
+; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v33.h
+; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v35.h
+; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v33.l
+; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v37.h
+; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v36.l
+; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v38.l
+; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v36.h
+; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v48.l
+; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v34.l
+; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v48.h
+; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v38.h
+; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v49.l
+; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v37.l
+; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v50.l
+; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v39.l
+; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v50.h
+; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v39.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v81.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v81.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v82.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v83.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v97.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v70.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v98.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v99.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v99.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v87.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v101.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v102.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v102.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v103.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v112.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v101.l
+; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v112.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v113.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v113.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v103.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v115.l
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v115.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v116.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v116.h
+; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v117.l
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v114.l
+; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v118.l
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v118.h
+; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v119.l
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v114.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v119.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v128.l
+; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v128.h
+; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v129.l
+; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v130.l
+; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v117.h
+; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v130.h
+; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v131.l
+; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v131.h
+; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v129.h
+; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v132.h
+; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v133.l
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v21.l, v133.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v134.l
+; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v135.l
+; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v132.l
+; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v135.h
+; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v144.l
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v144.h
+; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v134.h
+; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v25.l, v145.h
+; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v146.l
+; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v26.l, v146.h
+; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v147.l
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v27.l, v148.l
+; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v145.l
+; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v148.h
+; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v149.l
+; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v147.h
+; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v30.l, v150.l
+; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v150.h
+; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v31.l, v151.l
+; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v151.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_hi16
+; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
+; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB96_2
+; GFX11-TRUE16-NEXT: .LBB96_4: ; %cmp.true
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v50.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v39.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v50.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v39.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v48.l, 3
+; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v151.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v151.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v150.l, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v150.h, v1.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v148.l, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v49.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v37.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v48.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v38.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, 0x300, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v36.l, 3
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v149.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v147.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v148.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v149.l, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v34.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v38.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v36.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v37.h, 3
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v146.l, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, 0x300, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v145.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v146.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v147.l, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v145.h, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v34.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v35.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v33.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v35.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v33.h, 3
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v135.l, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v144.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v134.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v135.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v144.l, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, 0x300, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v32.l, 3
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(26)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v100.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v32.h, 3
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(24)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v100.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v98.l, 3
+; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v132.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v133.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v134.l, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v132.h, v1.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v133.l, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(22)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v97.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v85.h, 3
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(20)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v96.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v86.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, 0x300, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(18)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v87.h, 3
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v131.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v129.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v130.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v131.l, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v71.h, 3
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(16)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v84.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v80.h, 3
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(14)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v96.h, 3
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v130.l, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v117.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v128.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v129.l, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v119.h, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v86.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(12)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v85.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v80.l, 3
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(10)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v84.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v82.l, 3
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v128.l, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v119.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v114.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v118.l, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v118.h, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v2.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(8)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v83.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v69.l, 3
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(6)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v71.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v69.h, 3
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(4)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v70.h, 3
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v117.l, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v114.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v116.l, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v116.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v115.l, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v68.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v68.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v66.l, 3
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v67.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v66.h, 3
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v115.h, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v113.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v103.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v112.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v113.l, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v67.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v54.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v65.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v64.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v65.l, 3
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v112.l, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v101.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v102.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v103.l, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v101.h, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v64.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v55.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v52.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v55.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v53.h, 3
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v102.l, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v99.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v87.l, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v98.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v99.l, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v54.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v49.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v53.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v51.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v52.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v51.h, 3
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v97.h, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v70.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v82.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v83.l, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v81.l, v1.h
+; GFX11-TRUE16-NEXT: v_or_b16 v33.l, v81.h, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v32.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v32.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v33.l
+; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-FAKE16-LABEL: bitcast_v128i8_to_v64i16:
+; GFX11-FAKE16: ; %bb.0:
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-FAKE16-NEXT: s_clause 0x1f
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v40, s32 offset:580
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v41, s32 offset:576
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v42, s32 offset:572
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v43, s32 offset:568
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v44, s32 offset:564
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v45, s32 offset:560
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v46, s32 offset:556
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v47, s32 offset:552
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v56, s32 offset:548
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v57, s32 offset:544
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v58, s32 offset:540
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v59, s32 offset:536
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v60, s32 offset:532
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v61, s32 offset:528
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v62, s32 offset:524
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v63, s32 offset:520
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v72, s32 offset:516
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v73, s32 offset:512
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v74, s32 offset:508
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v75, s32 offset:504
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v76, s32 offset:500
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v77, s32 offset:496
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v78, s32 offset:492
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v79, s32 offset:488
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v88, s32 offset:484
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v89, s32 offset:480
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v90, s32 offset:476
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v91, s32 offset:472
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v92, s32 offset:468
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v93, s32 offset:464
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v94, s32 offset:460
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v95, s32 offset:456
+; GFX11-FAKE16-NEXT: s_clause 0xf
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v104, s32 offset:452
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v105, s32 offset:448
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v106, s32 offset:444
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v107, s32 offset:440
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v108, s32 offset:436
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v109, s32 offset:432
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v110, s32 offset:428
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v111, s32 offset:424
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v120, s32 offset:420
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v121, s32 offset:416
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v122, s32 offset:412
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v123, s32 offset:408
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v124, s32 offset:404
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v125, s32 offset:400
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v126, s32 offset:396
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v127, s32 offset:392
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v37, v30 :: v_dual_mov_b32 v54, v24
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v55, v28 :: v_dual_mov_b32 v52, v26
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v50, v22 :: v_dual_mov_b32 v53, v20
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v34, v18 :: v_dual_mov_b32 v51, v16
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v36, v14 :: v_dual_mov_b32 v49, v12
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v32, v10 :: v_dual_mov_b32 v39, v4
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v48, v8 :: v_dual_mov_b32 v33, v6
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v35, v2 :: v_dual_mov_b32 v38, v0
+; GFX11-FAKE16-NEXT: s_clause 0x1f
+; GFX11-FAKE16-NEXT: scratch_load_u16 v0, off, s32 offset:384
+; GFX11-FAKE16-NEXT: scratch_load_u16 v118, off, s32 offset:380
+; GFX11-FAKE16-NEXT: scratch_load_u16 v2, off, s32 offset:376
+; GFX11-FAKE16-NEXT: scratch_load_u16 v134, off, s32 offset:372
+; GFX11-FAKE16-NEXT: scratch_load_u16 v4, off, s32 offset:368
+; GFX11-FAKE16-NEXT: scratch_load_u16 v116, off, s32 offset:364
+; GFX11-FAKE16-NEXT: scratch_load_u16 v6, off, s32 offset:360
+; GFX11-FAKE16-NEXT: scratch_load_u16 v131, off, s32 offset:356
+; GFX11-FAKE16-NEXT: scratch_load_u16 v8, off, s32 offset:352
+; GFX11-FAKE16-NEXT: scratch_load_u16 v98, off, s32 offset:348
+; GFX11-FAKE16-NEXT: scratch_load_u16 v10, off, s32 offset:344
+; GFX11-FAKE16-NEXT: scratch_load_u16 v129, off, s32 offset:340
+; GFX11-FAKE16-NEXT: scratch_load_u16 v12, off, s32 offset:336
+; GFX11-FAKE16-NEXT: scratch_load_u16 v99, off, s32 offset:332
+; GFX11-FAKE16-NEXT: scratch_load_u16 v14, off, s32 offset:328
+; GFX11-FAKE16-NEXT: scratch_load_u16 v112, off, s32 offset:324
+; GFX11-FAKE16-NEXT: scratch_load_u16 v16, off, s32 offset:320
+; GFX11-FAKE16-NEXT: scratch_load_u16 v81, off, s32 offset:316
+; GFX11-FAKE16-NEXT: scratch_load_u16 v18, off, s32 offset:312
+; GFX11-FAKE16-NEXT: scratch_load_u16 v103, off, s32 offset:308
+; GFX11-FAKE16-NEXT: scratch_load_u16 v20, off, s32 offset:304
+; GFX11-FAKE16-NEXT: scratch_load_u16 v86, off, s32 offset:300
+; GFX11-FAKE16-NEXT: scratch_load_u16 v22, off, s32 offset:296
+; GFX11-FAKE16-NEXT: scratch_load_u16 v101, off, s32 offset:292
+; GFX11-FAKE16-NEXT: scratch_load_u16 v24, off, s32 offset:288
+; GFX11-FAKE16-NEXT: scratch_load_u16 v83, off, s32 offset:284
+; GFX11-FAKE16-NEXT: scratch_load_u16 v26, off, s32 offset:280
+; GFX11-FAKE16-NEXT: scratch_load_u16 v97, off, s32 offset:276
+; GFX11-FAKE16-NEXT: scratch_load_u16 v28, off, s32 offset:272
+; GFX11-FAKE16-NEXT: scratch_load_u16 v67, off, s32 offset:268
+; GFX11-FAKE16-NEXT: scratch_load_u16 v30, off, s32 offset:264
+; GFX11-FAKE16-NEXT: scratch_load_u16 v85, off, s32 offset:260
+; GFX11-FAKE16-NEXT: s_clause 0x1f
+; GFX11-FAKE16-NEXT: scratch_load_u16 v31, off, s32 offset:256
+; GFX11-FAKE16-NEXT: scratch_load_u16 v68, off, s32 offset:252
+; GFX11-FAKE16-NEXT: scratch_load_u16 v94, off, s32 offset:248
+; GFX11-FAKE16-NEXT: scratch_load_u16 v80, off, s32 offset:244
+; GFX11-FAKE16-NEXT: scratch_load_u16 v88, off, s32 offset:240
+; GFX11-FAKE16-NEXT: scratch_load_u16 v64, off, s32 offset:236
+; GFX11-FAKE16-NEXT: scratch_load_u16 v93, off, s32 offset:232
+; GFX11-FAKE16-NEXT: scratch_load_u16 v69, off, s32 offset:228
+; GFX11-FAKE16-NEXT: scratch_load_u16 v91, off, s32 offset:224
+; GFX11-FAKE16-NEXT: scratch_load_u16 v65, off, s32 offset:220
+; GFX11-FAKE16-NEXT: scratch_load_u16 v92, off, s32 offset:216
+; GFX11-FAKE16-NEXT: scratch_load_b32 v150, off, s32 offset:388
+; GFX11-FAKE16-NEXT: scratch_load_u16 v182, off, s32
+; GFX11-FAKE16-NEXT: scratch_load_u16 v40, off, s32 offset:8
+; GFX11-FAKE16-NEXT: scratch_load_u16 v43, off, s32 offset:16
+; GFX11-FAKE16-NEXT: scratch_load_u16 v44, off, s32 offset:24
+; GFX11-FAKE16-NEXT: scratch_load_u16 v45, off, s32 offset:32
+; GFX11-FAKE16-NEXT: scratch_load_u16 v46, off, s32 offset:40
+; GFX11-FAKE16-NEXT: scratch_load_u16 v47, off, s32 offset:48
+; GFX11-FAKE16-NEXT: scratch_load_u16 v56, off, s32 offset:56
+; GFX11-FAKE16-NEXT: scratch_load_u16 v58, off, s32 offset:64
+; GFX11-FAKE16-NEXT: scratch_load_u16 v59, off, s32 offset:72
+; GFX11-FAKE16-NEXT: scratch_load_u16 v60, off, s32 offset:80
+; GFX11-FAKE16-NEXT: scratch_load_u16 v61, off, s32 offset:88
+; GFX11-FAKE16-NEXT: scratch_load_u16 v62, off, s32 offset:96
+; GFX11-FAKE16-NEXT: scratch_load_u16 v63, off, s32 offset:104
+; GFX11-FAKE16-NEXT: scratch_load_u16 v72, off, s32 offset:112
+; GFX11-FAKE16-NEXT: scratch_load_u16 v73, off, s32 offset:120
+; GFX11-FAKE16-NEXT: scratch_load_u16 v74, off, s32 offset:128
+; GFX11-FAKE16-NEXT: scratch_load_u16 v75, off, s32 offset:136
+; GFX11-FAKE16-NEXT: scratch_load_u16 v76, off, s32 offset:144
+; GFX11-FAKE16-NEXT: scratch_load_u16 v77, off, s32 offset:152
+; GFX11-FAKE16-NEXT: s_clause 0x1f
+; GFX11-FAKE16-NEXT: scratch_load_u16 v78, off, s32 offset:160
+; GFX11-FAKE16-NEXT: scratch_load_u16 v79, off, s32 offset:168
+; GFX11-FAKE16-NEXT: scratch_load_u16 v89, off, s32 offset:176
+; GFX11-FAKE16-NEXT: scratch_load_u16 v90, off, s32 offset:184
+; GFX11-FAKE16-NEXT: scratch_load_u16 v95, off, s32 offset:192
+; GFX11-FAKE16-NEXT: scratch_load_u16 v104, off, s32 offset:200
+; GFX11-FAKE16-NEXT: scratch_load_u16 v105, off, s32 offset:208
+; GFX11-FAKE16-NEXT: scratch_load_u16 v42, off, s32 offset:212
+; GFX11-FAKE16-NEXT: scratch_load_u16 v180, off, s32 offset:204
+; GFX11-FAKE16-NEXT: scratch_load_u16 v183, off, s32 offset:196
+; GFX11-FAKE16-NEXT: scratch_load_u16 v165, off, s32 offset:188
+; GFX11-FAKE16-NEXT: scratch_load_u16 v179, off, s32 offset:180
+; GFX11-FAKE16-NEXT: scratch_load_u16 v163, off, s32 offset:172
+; GFX11-FAKE16-NEXT: scratch_load_u16 v177, off, s32 offset:164
+; GFX11-FAKE16-NEXT: scratch_load_u16 v145, off, s32 offset:156
+; GFX11-FAKE16-NEXT: scratch_load_u16 v166, off, s32 offset:148
+; GFX11-FAKE16-NEXT: scratch_load_u16 v148, off, s32 offset:140
+; GFX11-FAKE16-NEXT: scratch_load_u16 v151, off, s32 offset:132
+; GFX11-FAKE16-NEXT: scratch_load_u16 v164, off, s32 offset:124
+; GFX11-FAKE16-NEXT: scratch_load_u16 v178, off, s32 offset:116
+; GFX11-FAKE16-NEXT: scratch_load_u16 v146, off, s32 offset:108
+; GFX11-FAKE16-NEXT: scratch_load_u16 v162, off, s32 offset:100
+; GFX11-FAKE16-NEXT: scratch_load_u16 v144, off, s32 offset:92
+; GFX11-FAKE16-NEXT: scratch_load_u16 v149, off, s32 offset:84
+; GFX11-FAKE16-NEXT: scratch_load_u16 v119, off, s32 offset:76
+; GFX11-FAKE16-NEXT: scratch_load_u16 v147, off, s32 offset:68
+; GFX11-FAKE16-NEXT: scratch_load_u16 v130, off, s32 offset:60
+; GFX11-FAKE16-NEXT: scratch_load_u16 v135, off, s32 offset:52
+; GFX11-FAKE16-NEXT: scratch_load_u16 v117, off, s32 offset:44
+; GFX11-FAKE16-NEXT: scratch_load_u16 v133, off, s32 offset:36
+; GFX11-FAKE16-NEXT: scratch_load_u16 v96, off, s32 offset:28
+; GFX11-FAKE16-NEXT: scratch_load_u16 v114, off, s32 offset:20
+; GFX11-FAKE16-NEXT: s_clause 0x1
+; GFX11-FAKE16-NEXT: scratch_load_u16 v87, off, s32 offset:12
+; GFX11-FAKE16-NEXT: scratch_load_u16 v102, off, s32 offset:4
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v70, 8, v1
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v71, 8, v3
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v84, 8, v5
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v82, 8, v7
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v115, 8, v9
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v66, 8, v11
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v128, 8, v13
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v113, 8, v15
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v132, 8, v17
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v100, 8, v19
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v161, 8, v21
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v160, 8, v23
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v176, 8, v25
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v167, 8, v27
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v181, 8, v29
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(62)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v127, 8, v0
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v126, 8, v2
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v124, 8, v4
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v125, 8, v6
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v120, 8, v8
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v123, 8, v10
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v121, 8, v12
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v122, 8, v14
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v106, 8, v16
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v111, 8, v18
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v109, 8, v20
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v110, 8, v22
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v107, 8, v24
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v108, 8, v26
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(61)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v88, 8, v88
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(59)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v93, 8, v93
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(57)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v91, 8, v91
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(55)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v92, 8, v92
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(54)
+; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v150
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(53)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v150, 8, v182
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(52)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v41, 8, v40
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(51)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v40, 8, v43
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(50)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v43, 8, v44
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(49)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v182, 8, v45
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(48)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v46, 8, v46
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(47)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v45, 8, v47
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(46)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v57, 8, v56
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(45)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v56, 8, v58
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(44)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v58, 8, v59
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(43)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v44, 8, v60
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(42)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v60, 8, v61
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(41)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v59, 8, v62
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(40)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v62, 8, v63
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(39)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v47, 8, v72
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(38)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v72, 8, v73
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(37)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v63, 8, v74
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(36)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v74, 8, v75
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(35)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v73, 8, v76
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(34)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v75, 8, v77
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(33)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v61, 8, v78
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(32)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v78, 8, v79
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(31)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v77, 8, v89
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(30)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v79, 8, v90
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(29)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v76, 8, v95
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(28)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v90, 8, v104
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(27)
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v89, 8, v105
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v104, 8, v94
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v95, 8, v31
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v105, 8, v30
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v94, 8, v28
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
+; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
+; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB96_2
+; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v38
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v35
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v39
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v33
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v48
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v70
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v71
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v2, v84
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v3, v82
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xff, v36
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xff, v51
+; GFX11-FAKE16-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xff, v34
+; GFX11-FAKE16-NEXT: v_perm_b32 v1, v3, v2, 0x5040100
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v32
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v49
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xff, v53
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v9, 0xff, v50
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xff, v54
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xff, v52
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v4, v115
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v2, v66
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v3, v128
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v5, v113
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v6, v6, v132
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v7, v100
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v8, v8, v161
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v9, v9, v160
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v10, v10, v176
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v11, v11, v167
+; GFX11-FAKE16-NEXT: v_perm_b32 v2, v2, v4, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v3, v5, v3, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v4, v7, v6, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v5, v9, v8, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v6, v11, v10, 0x5040100
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xff, v55
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xff, v37
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v9, 0xff, v102
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xff, v87
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xff, v114
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xff, v96
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xff, v133
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xff, v117
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xff, v135
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v16, 0xff, v130
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v7, v181
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v8, v8, v150
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v9, v9, v41
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v10, v10, v40
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v11, v11, v43
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v12, v12, v182
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v13, v13, v46
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v14, v14, v45
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v15, v15, v57
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v16, v16, v56
+; GFX11-FAKE16-NEXT: v_perm_b32 v7, v8, v7, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v8, v10, v9, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v9, v12, v11, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v10, v14, v13, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v11, v16, v15, 0x5040100
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xff, v147
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xff, v119
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xff, v149
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xff, v144
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v16, 0xff, v162
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v17, 0xff, v146
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xff, v178
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xff, v164
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xff, v151
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xff, v148
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v12, v12, v58
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v13, v13, v44
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v14, v14, v60
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v15, v15, v59
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v16, v16, v62
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v17, v17, v47
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v18, v18, v72
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v19, v19, v63
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v20, v20, v74
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v21, v21, v73
+; GFX11-FAKE16-NEXT: v_perm_b32 v12, v13, v12, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v13, v15, v14, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v14, v17, v16, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v15, v19, v18, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v16, v21, v20, 0x5040100
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v17, 0xff, v166
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xff, v145
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xff, v177
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xff, v163
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xff, v179
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xff, v165
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v23, 0xff, v183
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v24, 0xff, v180
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v25, 0xff, v42
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v26, 0xff, v65
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v17, v17, v75
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v18, v18, v61
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v19, v19, v78
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v20, v20, v77
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v21, v21, v79
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v22, v22, v76
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v23, v23, v90
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v24, v24, v89
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v25, v25, v92
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v26, v26, v91
+; GFX11-FAKE16-NEXT: v_perm_b32 v17, v18, v17, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v18, v20, v19, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v19, v22, v21, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v20, v24, v23, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v21, v26, v25, 0x5040100
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xff, v69
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v23, 0xff, v64
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v24, 0xff, v80
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v25, 0xff, v68
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v26, 0xff, v85
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v27, 0xff, v67
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v28, 0xff, v97
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v29, 0xff, v83
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v30, 0xff, v101
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v31, 0xff, v86
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v22, v22, v93
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v23, v23, v88
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v24, v24, v104
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v25, v25, v95
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v26, v26, v105
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v27, v27, v94
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v28, v28, v108
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v29, v29, v107
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v30, v30, v110
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v31, v31, v109
+; GFX11-FAKE16-NEXT: v_perm_b32 v22, v23, v22, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v23, v25, v24, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v24, v27, v26, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v25, v29, v28, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v26, v31, v30, 0x5040100
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v27, 0xff, v103
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v28, 0xff, v81
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v29, 0xff, v112
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v30, 0xff, v99
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v31, 0xff, v129
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v32, 0xff, v98
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xff, v131
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v34, 0xff, v116
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v35, 0xff, v134
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v36, 0xff, v118
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v27, v27, v111
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v28, v28, v106
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v29, v29, v122
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v30, v30, v121
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v31, v31, v123
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v32, v32, v120
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v33, v33, v125
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v34, v34, v124
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v35, v35, v126
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v36, v36, v127
+; GFX11-FAKE16-NEXT: v_perm_b32 v27, v28, v27, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v28, v30, v29, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v29, v32, v31, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v30, v34, v33, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v31, v36, v35, 0x5040100
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr38
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr35
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr39
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr33
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr48
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr32
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr49
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr36
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr51
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr34
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr53
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr50
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr54
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr52
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr55
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr37
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr102
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr87
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr114
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr96
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr133
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr117
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr135
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr130
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr147
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr119
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr149
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr144
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr162
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr146
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr178
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr164
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr151
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr148
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr166
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr145
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr177
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr163
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr179
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr165
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr183
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr180
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr65
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr69
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr64
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr80
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr68
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr85
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr67
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr97
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr83
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr101
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr86
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr103
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr81
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr112
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr99
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr129
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr98
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr131
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr116
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr134
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr118
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr70
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr71
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr84
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr82
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr115
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr66
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr128
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr113
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr132
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr100
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr161
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr160
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr176
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr167
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr181
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr150
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr41
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr40
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr43
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr182
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr46
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr45
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr57
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr56
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr58
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr44
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr60
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr59
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr62
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr47
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr72
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr63
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr74
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr73
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr75
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr61
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr78
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr77
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr79
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr76
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr90
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr89
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr92
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr91
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr93
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr88
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr104
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr95
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr105
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr94
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr108
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr107
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr110
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr109
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr111
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr106
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr122
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr121
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr123
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr120
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr125
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr124
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr126
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr127
+; GFX11-FAKE16-NEXT: .LBB96_2: ; %Flow
+; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
+; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB96_4
+; GFX11-FAKE16-NEXT: ; %bb.3: ; %cmp.true
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v134, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v1, v118, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v2, v131, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, v116, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, v129, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v35, v35, 3
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v126, v0
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v127, v1
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v125, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v124, v3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v33, v33, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v31, 0x300, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v98, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v116, 0x300, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v30, 0x300, v2
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v98, 0x300, v3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v1, v112, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, v99, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, v103, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v123, v2
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v120, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v122, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v29, 0x300, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v121, v3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v99, 0x300, v0
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v111, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v28, 0x300, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v1, v81, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v81, 0x300, v2
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v2, v101, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v27, 0x300, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v86, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, v97, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, v83, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v106, v1
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v110, v2
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v109, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v83, 0x300, v1
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v108, v3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v26, 0x300, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v107, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v86, 0x300, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v85, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v25, 0x300, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v1, v67, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v67, 0x300, v2
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v2, v80, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, v68, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, v69, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v105, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v94, v1
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v104, v2
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v24, 0x300, v0
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v95, v3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v68, 0x300, v1
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v93, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v23, 0x300, v2
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v2, v64, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v64, 0x300, v0
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(26)
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v42, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v22, 0x300, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v1, v65, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(24)
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, v183, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, v180, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v88, v2
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v92, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v91, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v65, 0x300, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v90, v3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v21, 0x300, v0
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v89, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v69, 0x300, v1
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(22)
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v1, v179, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v20, 0x300, v2
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v2, v165, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v80, 0x300, v0
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(20)
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v177, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, v163, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(18)
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, v166, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v79, v1
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v76, v2
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v78, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v19, 0x300, v1
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v77, v3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v85, 0x300, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v75, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v18, 0x300, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v145, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v97, 0x300, v1
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(16)
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v1, v151, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v17, 0x300, v2
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v2, v148, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(14)
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, v178, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, v164, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v61, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v74, v1
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v73, v2
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v101, 0x300, v0
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v72, v3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v16, 0x300, v1
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v63, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v103, 0x300, v2
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(12)
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v2, v162, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v15, 0x300, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v146, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v112, 0x300, v1
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(10)
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v1, v149, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, v144, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(8)
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, v147, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v62, v2
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v47, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v60, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v14, 0x300, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v59, v3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v118, 0x300, v0
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v58, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v13, 0x300, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v1, v119, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v119, 0x300, v2
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(6)
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v2, v135, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v12, 0x300, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v130, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(4)
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, v133, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, v117, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v44, v1
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v57, v2
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v56, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v117, 0x300, v1
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v46, v3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v11, 0x300, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v45, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v129, 0x300, v0
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(2)
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v114, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v10, 0x300, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v1, v96, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v96, 0x300, v2
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v2, v102, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, v87, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, v55, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v43, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v182, v1
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v41, v2
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v9, 0x300, v0
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v40, v3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v55, 0x300, v1
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v181, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v8, 0x300, v2
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v2, v37, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v37, 0x300, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v54, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v7, 0x300, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v1, v52, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, v53, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, v50, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v150, v2
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v176, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v167, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v50, 0x300, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v161, v3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v6, 0x300, v0
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v160, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v52, 0x300, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v1, v51, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v5, 0x300, v2
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v2, v49, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v49, 0x300, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v48, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, v34, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, v36, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v132, v1
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v128, v2
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v115, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v34, 0x300, v1
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v100, v3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v3, 0x300, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v113, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v4, 0x300, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, v39, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v36, v38, 3
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v32, v32, 3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v35, 0xff, v35
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xff, v33
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v36, 0xff, v36
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v32, 0xff, v32
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v35, v71, v35
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v33, v82, v33
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v84, v0
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v36, v70, v36
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v32, v66, v32
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v35, 0x300, v35
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v33, 0x300, v33
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v38, 0x300, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v0, 0x300, v36
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v32, 0x300, v32
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v36, 0x300, v2
+; GFX11-FAKE16-NEXT: v_add_nc_u16 v39, 0x300, v1
+; GFX11-FAKE16-NEXT: v_perm_b32 v1, v33, v38, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v0, v35, v0, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v2, v32, v4, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v3, v36, v3, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v4, v39, v34, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v5, v49, v5, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v6, v52, v6, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v7, v50, v7, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v8, v37, v8, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v9, v55, v9, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v10, v96, v10, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v11, v129, v11, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v12, v117, v12, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v13, v119, v13, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v14, v118, v14, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v15, v112, v15, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v16, v103, v16, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v17, v101, v17, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v18, v97, v18, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v19, v85, v19, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v20, v80, v20, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v21, v69, v21, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v22, v65, v22, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v23, v64, v23, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v24, v68, v24, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v25, v67, v25, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v26, v86, v26, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v27, v83, v27, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v28, v81, v28, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v29, v99, v29, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v30, v98, v30, 0x5040100
+; GFX11-FAKE16-NEXT: v_perm_b32 v31, v116, v31, 0x5040100
+; GFX11-FAKE16-NEXT: .LBB96_4: ; %end
+; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-FAKE16-NEXT: s_clause 0x1f
+; GFX11-FAKE16-NEXT: scratch_load_b32 v127, off, s32 offset:392
+; GFX11-FAKE16-NEXT: scratch_load_b32 v126, off, s32 offset:396
+; GFX11-FAKE16-NEXT: scratch_load_b32 v125, off, s32 offset:400
+; GFX11-FAKE16-NEXT: scratch_load_b32 v124, off, s32 offset:404
+; GFX11-FAKE16-NEXT: scratch_load_b32 v123, off, s32 offset:408
+; GFX11-FAKE16-NEXT: scratch_load_b32 v122, off, s32 offset:412
+; GFX11-FAKE16-NEXT: scratch_load_b32 v121, off, s32 offset:416
+; GFX11-FAKE16-NEXT: scratch_load_b32 v120, off, s32 offset:420
+; GFX11-FAKE16-NEXT: scratch_load_b32 v111, off, s32 offset:424
+; GFX11-FAKE16-NEXT: scratch_load_b32 v110, off, s32 offset:428
+; GFX11-FAKE16-NEXT: scratch_load_b32 v109, off, s32 offset:432
+; GFX11-FAKE16-NEXT: scratch_load_b32 v108, off, s32 offset:436
+; GFX11-FAKE16-NEXT: scratch_load_b32 v107, off, s32 offset:440
+; GFX11-FAKE16-NEXT: scratch_load_b32 v106, off, s32 offset:444
+; GFX11-FAKE16-NEXT: scratch_load_b32 v105, off, s32 offset:448
+; GFX11-FAKE16-NEXT: scratch_load_b32 v104, off, s32 offset:452
+; GFX11-FAKE16-NEXT: scratch_load_b32 v95, off, s32 offset:456
+; GFX11-FAKE16-NEXT: scratch_load_b32 v94, off, s32 offset:460
+; GFX11-FAKE16-NEXT: scratch_load_b32 v93, off, s32 offset:464
+; GFX11-FAKE16-NEXT: scratch_load_b32 v92, off, s32 offset:468
+; GFX11-FAKE16-NEXT: scratch_load_b32 v91, off, s32 offset:472
+; GFX11-FAKE16-NEXT: scratch_load_b32 v90, off, s32 offset:476
+; GFX11-FAKE16-NEXT: scratch_load_b32 v89, off, s32 offset:480
+; GFX11-FAKE16-NEXT: scratch_load_b32 v88, off, s32 offset:484
+; GFX11-FAKE16-NEXT: scratch_load_b32 v79, off, s32 offset:488
+; GFX11-FAKE16-NEXT: scratch_load_b32 v78, off, s32 offset:492
+; GFX11-FAKE16-NEXT: scratch_load_b32 v77, off, s32 offset:496
+; GFX11-FAKE16-NEXT: scratch_load_b32 v76, off, s32 offset:500
+; GFX11-FAKE16-NEXT: scratch_load_b32 v75, off, s32 offset:504
+; GFX11-FAKE16-NEXT: scratch_load_b32 v74, off, s32 offset:508
+; GFX11-FAKE16-NEXT: scratch_load_b32 v73, off, s32 offset:512
+; GFX11-FAKE16-NEXT: scratch_load_b32 v72, off, s32 offset:516
+; GFX11-FAKE16-NEXT: s_clause 0xf
+; GFX11-FAKE16-NEXT: scratch_load_b32 v63, off, s32 offset:520
+; GFX11-FAKE16-NEXT: scratch_load_b32 v62, off, s32 offset:524
+; GFX11-FAKE16-NEXT: scratch_load_b32 v61, off, s32 offset:528
+; GFX11-FAKE16-NEXT: scratch_load_b32 v60, off, s32 offset:532
+; GFX11-FAKE16-NEXT: scratch_load_b32 v59, off, s32 offset:536
+; GFX11-FAKE16-NEXT: scratch_load_b32 v58, off, s32 offset:540
+; GFX11-FAKE16-NEXT: scratch_load_b32 v57, off, s32 offset:544
+; GFX11-FAKE16-NEXT: scratch_load_b32 v56, off, s32 offset:548
+; GFX11-FAKE16-NEXT: scratch_load_b32 v47, off, s32 offset:552
+; GFX11-FAKE16-NEXT: scratch_load_b32 v46, off, s32 offset:556
+; GFX11-FAKE16-NEXT: scratch_load_b32 v45, off, s32 offset:560
+; GFX11-FAKE16-NEXT: scratch_load_b32 v44, off, s32 offset:564
+; GFX11-FAKE16-NEXT: scratch_load_b32 v43, off, s32 offset:568
+; GFX11-FAKE16-NEXT: scratch_load_b32 v42, off, s32 offset:572
+; GFX11-FAKE16-NEXT: scratch_load_b32 v41, off, s32 offset:576
+; GFX11-FAKE16-NEXT: scratch_load_b32 v40, off, s32 offset:580
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-FAKE16-NEXT: 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 inreg <64 x i16> @bitcast_v128i8_to_v64i16_scalar(<128 x i8> inreg %a, i32 inreg %b) {
+; SI-LABEL: bitcast_v128i8_to_v64i16_scalar:
+; SI: ; %bb.0:
+; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; SI-NEXT: s_or_saveexec_b64 s[4:5], -1
+; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
+; SI-NEXT: s_mov_b64 exec, s[4:5]
+; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
+; SI-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:332
+; SI-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:328
+; SI-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:324
+; SI-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:320
+; SI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:308
+; SI-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:304
+; SI-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:300
+; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:296
+; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:292
+; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:288
+; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:276
+; SI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:272
+; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:268
+; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:264
+; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:260
+; SI-NEXT: s_waitcnt expcnt(6)
+; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:256
+; SI-NEXT: s_waitcnt expcnt(5)
+; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:244
+; SI-NEXT: s_waitcnt expcnt(4)
+; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:240
+; SI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:236
+; SI-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:232
+; SI-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:228
+; SI-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:224
+; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:212
+; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:208
+; SI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:204
+; SI-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:200
+; SI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:196
+; SI-NEXT: v_writelane_b32 v63, s30, 0
+; SI-NEXT: v_writelane_b32 v63, s31, 1
+; SI-NEXT: v_writelane_b32 v63, s34, 2
+; SI-NEXT: v_writelane_b32 v63, s35, 3
+; SI-NEXT: v_writelane_b32 v63, s36, 4
+; SI-NEXT: v_writelane_b32 v63, s37, 5
+; SI-NEXT: v_writelane_b32 v63, s38, 6
+; SI-NEXT: v_writelane_b32 v63, s39, 7
+; SI-NEXT: v_writelane_b32 v63, s48, 8
+; SI-NEXT: v_writelane_b32 v63, s49, 9
+; SI-NEXT: v_writelane_b32 v63, s50, 10
+; SI-NEXT: v_writelane_b32 v63, s51, 11
+; SI-NEXT: v_writelane_b32 v63, s52, 12
+; SI-NEXT: v_writelane_b32 v63, s53, 13
+; SI-NEXT: v_writelane_b32 v63, s54, 14
+; SI-NEXT: v_writelane_b32 v63, s55, 15
+; SI-NEXT: v_writelane_b32 v63, s64, 16
+; SI-NEXT: v_writelane_b32 v63, s65, 17
+; SI-NEXT: ; implicit-def: $vgpr62 : SGPR spill to VGPR lane
+; SI-NEXT: v_writelane_b32 v63, s66, 18
+; SI-NEXT: v_writelane_b32 v62, s28, 0
+; SI-NEXT: v_writelane_b32 v63, s67, 19
+; SI-NEXT: v_writelane_b32 v62, s27, 1
+; SI-NEXT: v_writelane_b32 v63, s68, 20
+; SI-NEXT: v_writelane_b32 v62, s25, 2
+; SI-NEXT: v_writelane_b32 v63, s69, 21
+; SI-NEXT: v_writelane_b32 v62, s24, 3
+; SI-NEXT: v_writelane_b32 v63, s70, 22
+; SI-NEXT: v_writelane_b32 v62, s23, 4
+; SI-NEXT: v_writelane_b32 v63, s71, 23
+; SI-NEXT: v_writelane_b32 v62, s22, 5
+; SI-NEXT: v_writelane_b32 v63, s80, 24
+; SI-NEXT: v_writelane_b32 v62, s21, 6
+; SI-NEXT: v_writelane_b32 v63, s81, 25
+; SI-NEXT: v_writelane_b32 v62, s20, 7
+; SI-NEXT: v_writelane_b32 v63, s82, 26
+; SI-NEXT: v_writelane_b32 v62, s19, 8
+; SI-NEXT: v_writelane_b32 v63, s83, 27
+; SI-NEXT: v_writelane_b32 v62, s18, 9
+; SI-NEXT: v_writelane_b32 v63, s84, 28
+; SI-NEXT: v_writelane_b32 v62, s16, 10
+; SI-NEXT: v_writelane_b32 v63, s85, 29
+; SI-NEXT: v_writelane_b32 v63, s86, 30
+; SI-NEXT: v_writelane_b32 v63, s87, 31
+; SI-NEXT: v_writelane_b32 v63, s96, 32
+; SI-NEXT: v_writelane_b32 v63, s97, 33
+; SI-NEXT: v_writelane_b32 v63, s98, 34
+; SI-NEXT: s_waitcnt expcnt(2)
+; SI-NEXT: v_mov_b32_e32 v29, v5
+; SI-NEXT: v_readfirstlane_b32 s76, v18
+; SI-NEXT: v_readfirstlane_b32 s40, v25
+; SI-NEXT: v_readfirstlane_b32 s16, v24
+; SI-NEXT: v_readfirstlane_b32 s42, v23
+; SI-NEXT: v_readfirstlane_b32 s52, v20
+; SI-NEXT: v_readfirstlane_b32 s8, v19
+; SI-NEXT: s_waitcnt vmcnt(14)
+; SI-NEXT: v_readfirstlane_b32 s50, v35
+; SI-NEXT: v_readfirstlane_b32 s31, v36
+; SI-NEXT: v_readfirstlane_b32 s53, v37
+; SI-NEXT: v_readfirstlane_b32 s82, v48
+; SI-NEXT: v_readfirstlane_b32 s7, v49
+; SI-NEXT: v_readfirstlane_b32 s79, v52
+; SI-NEXT: v_readfirstlane_b32 s78, v55
+; SI-NEXT: v_readfirstlane_b32 s4, v41
+; SI-NEXT: v_writelane_b32 v62, s4, 11
+; SI-NEXT: v_readfirstlane_b32 s4, v43
+; SI-NEXT: v_writelane_b32 v62, s4, 12
+; SI-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:192
+; SI-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:180
+; SI-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:176
+; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:172
+; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:168
+; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:164
+; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:160
+; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:148
+; SI-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:144
+; SI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:140
+; SI-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:136
+; SI-NEXT: v_readfirstlane_b32 s4, v45
+; SI-NEXT: v_writelane_b32 v62, s4, 13
+; SI-NEXT: s_waitcnt vmcnt(14)
+; SI-NEXT: v_readfirstlane_b32 s4, v56
+; SI-NEXT: v_writelane_b32 v62, s4, 14
+; SI-NEXT: v_readfirstlane_b32 s4, v57
+; SI-NEXT: v_writelane_b32 v62, s4, 15
+; SI-NEXT: v_readfirstlane_b32 s4, v60
+; SI-NEXT: v_readfirstlane_b32 s86, v31
+; SI-NEXT: v_readfirstlane_b32 s36, v32
+; SI-NEXT: v_readfirstlane_b32 s71, v33
+; SI-NEXT: v_readfirstlane_b32 s77, v59
+; SI-NEXT: v_writelane_b32 v62, s4, 16
+; SI-NEXT: v_readfirstlane_b32 s4, v61
+; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:132
+; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:128
+; SI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:116
+; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:112
+; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:108
+; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:104
+; SI-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:100
+; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:96
+; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:84
+; SI-NEXT: v_readfirstlane_b32 s98, v50
+; SI-NEXT: s_waitcnt vmcnt(14)
+; SI-NEXT: v_readfirstlane_b32 s67, v53
+; SI-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:68
+; SI-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:64
+; SI-NEXT: v_writelane_b32 v62, s4, 17
+; SI-NEXT: v_readfirstlane_b32 s4, v38
+; SI-NEXT: v_writelane_b32 v62, s4, 18
+; SI-NEXT: v_readfirstlane_b32 s81, v34
+; SI-NEXT: v_readfirstlane_b32 s75, v39
+; SI-NEXT: v_readfirstlane_b32 s68, v42
+; SI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:80
+; SI-NEXT: v_readfirstlane_b32 s49, v40
+; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:52
+; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:48
+; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:44
+; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:40
+; SI-NEXT: v_readfirstlane_b32 s51, v54
+; SI-NEXT: v_readfirstlane_b32 s97, v51
+; SI-NEXT: v_readfirstlane_b32 s35, v27
+; SI-NEXT: v_writelane_b32 v63, s99, 35
+; SI-NEXT: v_readfirstlane_b32 s28, v28
+; SI-NEXT: v_readfirstlane_b32 s87, v26
+; SI-NEXT: v_lshlrev_b32_e32 v6, 24, v6
+; SI-NEXT: v_lshlrev_b32_e32 v14, 24, v14
+; SI-NEXT: v_readfirstlane_b32 s96, v17
+; SI-NEXT: v_readfirstlane_b32 s99, v16
+; SI-NEXT: v_readfirstlane_b32 s89, v15
+; SI-NEXT: v_readfirstlane_b32 s88, v12
+; SI-NEXT: v_readfirstlane_b32 s30, v11
+; SI-NEXT: v_readfirstlane_b32 s64, v10
+; SI-NEXT: v_readfirstlane_b32 s55, v9
+; SI-NEXT: v_readfirstlane_b32 s65, v8
+; SI-NEXT: v_readfirstlane_b32 s80, v7
+; SI-NEXT: v_readfirstlane_b32 s21, v2
+; SI-NEXT: v_readfirstlane_b32 s74, v1
+; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:156
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:220
+; SI-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:72
+; SI-NEXT: v_readfirstlane_b32 s93, v36
+; SI-NEXT: v_readfirstlane_b32 s24, v37
+; SI-NEXT: v_readfirstlane_b32 s27, v48
+; SI-NEXT: v_readfirstlane_b32 s84, v43
+; SI-NEXT: v_readfirstlane_b32 s83, v44
+; SI-NEXT: v_readfirstlane_b32 s85, v46
+; SI-NEXT: s_waitcnt vmcnt(14)
+; SI-NEXT: v_readfirstlane_b32 s20, v47
+; SI-NEXT: v_readfirstlane_b32 s4, v58
+; SI-NEXT: v_writelane_b32 v62, s4, 19
+; SI-NEXT: v_readfirstlane_b32 s23, v49
+; SI-NEXT: v_readfirstlane_b32 s92, v52
+; SI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:36
+; SI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:32
+; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:20
+; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:16
+; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12
+; SI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:336
+; SI-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:8
+; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:4
+; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32
+; SI-NEXT: v_readfirstlane_b32 s90, v35
+; SI-NEXT: v_readfirstlane_b32 s4, v31
+; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:92
+; SI-NEXT: v_readfirstlane_b32 s38, v32
+; SI-NEXT: v_readfirstlane_b32 s70, v33
+; SI-NEXT: v_readfirstlane_b32 s54, v59
+; SI-NEXT: v_readfirstlane_b32 s57, v60
+; SI-NEXT: s_waitcnt vmcnt(14)
+; SI-NEXT: v_readfirstlane_b32 s56, v61
+; SI-NEXT: v_readfirstlane_b32 s59, v55
+; SI-NEXT: v_readfirstlane_b32 s61, v41
+; SI-NEXT: v_readfirstlane_b32 s19, v45
+; SI-NEXT: v_readfirstlane_b32 s34, v50
+; SI-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:316
+; SI-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:76
+; SI-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:284
+; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:252
+; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:188
+; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:124
+; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:60
+; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:28
+; SI-NEXT: v_writelane_b32 v62, s4, 20
+; SI-NEXT: v_readfirstlane_b32 s25, v53
+; SI-NEXT: v_readfirstlane_b32 s91, v40
+; SI-NEXT: v_readfirstlane_b32 s37, v34
+; SI-NEXT: v_readfirstlane_b32 s47, v56
+; SI-NEXT: s_waitcnt vmcnt(14)
+; SI-NEXT: v_readfirstlane_b32 s46, v57
+; SI-NEXT: v_readfirstlane_b32 s22, v42
+; SI-NEXT: v_lshlrev_b32_e32 v40, 24, v22
+; SI-NEXT: v_lshlrev_b32_e32 v25, 24, v5
+; SI-NEXT: v_readfirstlane_b32 s72, v39
+; SI-NEXT: v_readfirstlane_b32 s94, v51
+; SI-NEXT: v_readfirstlane_b32 s48, v54
+; SI-NEXT: v_readfirstlane_b32 s66, v43
+; SI-NEXT: v_readfirstlane_b32 s69, v44
+; SI-NEXT: s_waitcnt vmcnt(13)
+; SI-NEXT: v_readfirstlane_b32 s45, v46
+; SI-NEXT: s_waitcnt vmcnt(12)
+; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v49
+; SI-NEXT: s_and_b64 s[4:5], vcc, exec
+; SI-NEXT: v_readfirstlane_b32 s4, v4
+; SI-NEXT: v_writelane_b32 v62, s4, 21
+; SI-NEXT: v_readfirstlane_b32 s4, v3
+; SI-NEXT: v_writelane_b32 v62, s4, 22
+; SI-NEXT: v_writelane_b32 v62, s17, 23
+; SI-NEXT: v_writelane_b32 v62, s40, 24
+; SI-NEXT: v_writelane_b32 v62, s16, 25
+; SI-NEXT: v_writelane_b32 v62, s42, 26
+; SI-NEXT: v_writelane_b32 v62, s46, 27
+; SI-NEXT: v_writelane_b32 v62, s47, 28
+; SI-NEXT: v_writelane_b32 v62, s56, 29
+; SI-NEXT: v_writelane_b32 v62, s57, 30
+; SI-NEXT: v_writelane_b32 v62, s45, 31
+; SI-NEXT: s_waitcnt vmcnt(11)
+; SI-NEXT: v_readfirstlane_b32 s58, v52
+; SI-NEXT: v_writelane_b32 v62, s49, 32
+; SI-NEXT: v_writelane_b32 v62, s58, 33
+; SI-NEXT: v_writelane_b32 v62, s59, 34
+; SI-NEXT: v_writelane_b32 v62, s52, 35
+; SI-NEXT: s_waitcnt vmcnt(6)
+; SI-NEXT: v_readfirstlane_b32 s60, v38
+; SI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:312
+; SI-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:280
+; SI-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:248
+; SI-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:216
+; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:184
+; SI-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:152
+; SI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:120
+; SI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:88
+; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:56
+; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:24
+; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
+; SI-NEXT: v_writelane_b32 v62, s60, 36
+; SI-NEXT: v_writelane_b32 v62, s61, 37
+; SI-NEXT: v_writelane_b32 v62, s93, 38
+; SI-NEXT: v_writelane_b32 v62, s8, 39
+; SI-NEXT: v_readfirstlane_b32 s62, v47
+; SI-NEXT: v_writelane_b32 v62, s72, 40
+; SI-NEXT: v_readfirstlane_b32 s73, v58
+; SI-NEXT: v_writelane_b32 v62, s62, 41
+; SI-NEXT: v_writelane_b32 v62, s73, 42
+; SI-NEXT: v_writelane_b32 v62, s35, 43
+; SI-NEXT: v_writelane_b32 v62, s94, 44
+; SI-NEXT: v_writelane_b32 v62, s48, 45
+; SI-NEXT: v_writelane_b32 v62, s91, 46
+; SI-NEXT: s_waitcnt vmcnt(11)
+; SI-NEXT: v_lshlrev_b32_e32 v61, 24, v61
+; SI-NEXT: v_lshlrev_b32_e32 v60, 24, v60
+; SI-NEXT: v_lshlrev_b32_e32 v31, 24, v31
+; SI-NEXT: v_lshlrev_b32_e32 v24, 24, v59
+; SI-NEXT: v_lshlrev_b32_e32 v27, 24, v45
+; SI-NEXT: v_lshlrev_b32_e32 v43, 24, v13
+; SI-NEXT: v_lshlrev_b32_e32 v44, 24, v41
+; SI-NEXT: v_lshlrev_b32_e32 v45, 24, v55
+; SI-NEXT: v_lshlrev_b32_e32 v46, 24, v50
+; SI-NEXT: v_writelane_b32 v62, s66, 47
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v42, 24, v18
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
+; SI-NEXT: s_cbranch_scc0 .LBB97_2
+; SI-NEXT: ; %bb.1: ; %cmp.false
+; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
+; SI-NEXT: v_mov_b32_e32 v21, v29
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v29
+; SI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
+; SI-NEXT: v_readlane_b32 s4, v62, 7
+; SI-NEXT: v_readlane_b32 s5, v62, 6
+; SI-NEXT: s_and_b32 s4, s4, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 8
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: v_writelane_b32 v62, s4, 48
+; SI-NEXT: v_readlane_b32 s4, v62, 5
+; SI-NEXT: s_and_b32 s4, s4, 0xff
+; SI-NEXT: v_readlane_b32 s5, v62, 4
+; SI-NEXT: s_lshl_b32 s4, s4, 16
+; SI-NEXT: s_lshl_b32 s5, s5, 24
+; SI-NEXT: s_or_b32 s63, s5, s4
+; SI-NEXT: v_readlane_b32 s4, v62, 9
+; SI-NEXT: s_and_b32 s5, s4, 0xff
+; SI-NEXT: v_readlane_b32 s4, v62, 8
+; SI-NEXT: s_lshl_b32 s5, s5, 16
+; SI-NEXT: s_lshl_b32 s9, s4, 24
+; SI-NEXT: v_readlane_b32 s4, v62, 0
+; SI-NEXT: s_or_b32 s9, s9, s5
+; SI-NEXT: s_and_b32 s5, s4, 0xff
+; SI-NEXT: s_lshl_b32 s10, s29, 8
+; SI-NEXT: s_or_b32 s4, s5, s10
+; SI-NEXT: v_writelane_b32 v62, s4, 49
+; SI-NEXT: s_and_b32 s5, s74, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 16
+; SI-NEXT: v_writelane_b32 v62, s54, 50
+; SI-NEXT: s_lshl_b32 s11, s21, 24
+; SI-NEXT: s_mov_b32 s18, s22
+; SI-NEXT: s_mov_b32 s22, s21
+; SI-NEXT: s_or_b32 s21, s11, s5
+; SI-NEXT: s_and_b32 s11, s26, 0xff
+; SI-NEXT: v_readlane_b32 s4, v62, 1
+; SI-NEXT: s_lshl_b32 s11, s11, 16
+; SI-NEXT: s_lshl_b32 s12, s4, 24
+; SI-NEXT: s_or_b32 s14, s12, s11
+; SI-NEXT: s_and_b32 s11, s80, 0xff
+; SI-NEXT: s_lshl_b32 s12, s65, 8
+; SI-NEXT: s_or_b32 s12, s11, s12
+; SI-NEXT: s_and_b32 s11, s55, 0xff
+; SI-NEXT: s_lshl_b32 s11, s11, 16
+; SI-NEXT: s_lshl_b32 s13, s64, 24
+; SI-NEXT: s_or_b32 s41, s13, s11
+; SI-NEXT: s_and_b32 s11, s89, 0xff
+; SI-NEXT: s_lshl_b32 s13, s99, 8
+; SI-NEXT: s_or_b32 s13, s11, s13
+; SI-NEXT: s_and_b32 s11, s96, 0xff
+; SI-NEXT: s_lshl_b32 s11, s11, 16
+; SI-NEXT: s_lshl_b32 s15, s76, 24
+; SI-NEXT: s_or_b32 s43, s15, s11
+; SI-NEXT: s_and_b32 s11, s42, 0xff
+; SI-NEXT: s_lshl_b32 s15, s16, 8
+; SI-NEXT: s_or_b32 s16, s11, s15
+; SI-NEXT: s_and_b32 s11, s40, 0xff
+; SI-NEXT: s_lshl_b32 s11, s11, 16
+; SI-NEXT: s_lshl_b32 s15, s87, 24
+; SI-NEXT: s_or_b32 s44, s15, s11
+; SI-NEXT: s_and_b32 s11, s73, 0xff
+; SI-NEXT: s_lshl_b32 s15, s62, 8
+; SI-NEXT: s_or_b32 s62, s11, s15
+; SI-NEXT: s_and_b32 s11, s58, 0xff
+; SI-NEXT: s_lshl_b32 s11, s11, 16
+; SI-NEXT: s_lshl_b32 s15, s45, 24
+; SI-NEXT: s_or_b32 s45, s15, s11
+; SI-NEXT: s_and_b32 s11, s48, 0xff
+; SI-NEXT: s_lshl_b32 s15, s94, 8
+; SI-NEXT: s_or_b32 s10, s11, s15
+; SI-NEXT: s_and_b32 s11, s46, 0xff
+; SI-NEXT: s_lshl_b32 s11, s11, 16
+; SI-NEXT: s_lshl_b32 s15, s47, 24
+; SI-NEXT: s_or_b32 s46, s15, s11
+; SI-NEXT: s_and_b32 s11, s25, 0xff
+; SI-NEXT: s_lshl_b32 s15, s34, 8
+; SI-NEXT: s_or_b32 s94, s11, s15
+; SI-NEXT: s_and_b32 s11, s72, 0xff
+; SI-NEXT: s_lshl_b32 s11, s11, 16
+; SI-NEXT: s_lshl_b32 s15, s60, 24
+; SI-NEXT: s_or_b32 s47, s15, s11
+; SI-NEXT: s_and_b32 s11, s61, 0xff
+; SI-NEXT: s_lshl_b32 s15, s59, 8
+; SI-NEXT: s_or_b32 s73, s11, s15
+; SI-NEXT: s_and_b32 s11, s56, 0xff
+; SI-NEXT: s_lshl_b32 s11, s11, 16
+; SI-NEXT: s_lshl_b32 s15, s57, 24
+; SI-NEXT: v_readlane_b32 s4, v62, 20
+; SI-NEXT: s_or_b32 s56, s15, s11
+; SI-NEXT: s_and_b32 s11, s38, 0xff
+; SI-NEXT: s_lshl_b32 s15, s4, 8
+; SI-NEXT: s_or_b32 s48, s11, s15
+; SI-NEXT: s_and_b32 s11, s92, 0xff
+; SI-NEXT: s_lshl_b32 s11, s11, 16
+; SI-NEXT: s_lshl_b32 s15, s23, 24
+; SI-NEXT: s_or_b32 vcc_lo, s15, s11
+; SI-NEXT: s_and_b32 s11, s20, 0xff
+; SI-NEXT: s_lshl_b32 s15, s85, 8
+; SI-NEXT: s_or_b32 s72, s11, s15
+; SI-NEXT: s_and_b32 s11, s83, 0xff
+; SI-NEXT: s_lshl_b32 s11, s11, 16
+; SI-NEXT: s_lshl_b32 s15, s84, 24
+; SI-NEXT: s_or_b32 vcc_hi, s15, s11
+; SI-NEXT: s_and_b32 s11, s93, 0xff
+; SI-NEXT: s_lshl_b32 s15, s97, 8
+; SI-NEXT: s_or_b32 s57, s11, s15
+; SI-NEXT: s_and_b32 s11, s67, 0xff
+; SI-NEXT: s_lshl_b32 s11, s11, 16
+; SI-NEXT: s_lshl_b32 s15, s51, 24
+; SI-NEXT: v_writelane_b32 v62, s67, 51
+; SI-NEXT: s_mov_b32 s67, s51
+; SI-NEXT: s_mov_b32 s51, s74
+; SI-NEXT: s_or_b32 s74, s15, s11
+; SI-NEXT: s_and_b32 s11, s98, 0xff
+; SI-NEXT: s_lshl_b32 s15, s75, 8
+; SI-NEXT: v_readlane_b32 s4, v62, 18
+; SI-NEXT: v_writelane_b32 v62, s87, 52
+; SI-NEXT: s_or_b32 s58, s11, s15
+; SI-NEXT: s_and_b32 s11, s4, 0xff
+; SI-NEXT: v_writelane_b32 v62, s25, 53
+; SI-NEXT: s_lshl_b32 s11, s11, 16
+; SI-NEXT: s_lshl_b32 s15, s81, 24
+; SI-NEXT: v_readlane_b32 s4, v62, 15
+; SI-NEXT: s_mov_b32 s54, s75
+; SI-NEXT: s_or_b32 s75, s15, s11
+; SI-NEXT: s_and_b32 s11, s77, 0xff
+; SI-NEXT: s_lshl_b32 s15, s4, 8
+; SI-NEXT: v_readlane_b32 s4, v62, 14
+; SI-NEXT: s_or_b32 s59, s11, s15
+; SI-NEXT: s_and_b32 s11, s4, 0xff
+; SI-NEXT: v_readlane_b32 s4, v62, 13
+; SI-NEXT: s_lshl_b32 s11, s11, 16
+; SI-NEXT: s_lshl_b32 s15, s4, 24
+; SI-NEXT: v_readlane_b32 s4, v62, 12
+; SI-NEXT: s_mov_b32 s95, s69
+; SI-NEXT: s_mov_b32 s69, s76
+; SI-NEXT: s_or_b32 s76, s15, s11
+; SI-NEXT: s_and_b32 s11, s86, 0xff
+; SI-NEXT: s_lshl_b32 s15, s4, 8
+; SI-NEXT: v_readlane_b32 s4, v62, 11
+; SI-NEXT: s_or_b32 s60, s11, s15
+; SI-NEXT: s_and_b32 s11, s4, 0xff
+; SI-NEXT: s_lshl_b32 s11, s11, 16
+; SI-NEXT: s_lshl_b32 s15, s78, 24
+; SI-NEXT: s_mov_b32 s93, s99
+; SI-NEXT: s_mov_b32 s99, s84
+; SI-NEXT: s_mov_b32 s84, s77
+; SI-NEXT: s_or_b32 s77, s15, s11
+; SI-NEXT: s_and_b32 s11, s82, 0xff
+; SI-NEXT: s_lshl_b32 s15, s53, 8
+; SI-NEXT: s_or_b32 s61, s11, s15
+; SI-NEXT: s_and_b32 s11, s31, 0xff
+; SI-NEXT: s_lshl_b32 s11, s11, 16
+; SI-NEXT: s_lshl_b32 s15, s50, 24
+; SI-NEXT: s_mov_b32 s4, s85
+; SI-NEXT: s_mov_b32 s85, s83
+; SI-NEXT: s_mov_b32 s83, s82
+; SI-NEXT: s_mov_b32 s82, s53
+; SI-NEXT: s_mov_b32 s53, s50
+; SI-NEXT: s_mov_b32 s50, s31
+; SI-NEXT: s_mov_b32 s31, s78
+; SI-NEXT: s_or_b32 s78, s15, s11
+; SI-NEXT: v_readlane_b32 s11, v62, 10
+; SI-NEXT: s_and_b32 s11, s11, 0xff
+; SI-NEXT: s_lshl_b32 s15, s17, 8
+; SI-NEXT: s_or_b32 s11, s11, s15
+; SI-NEXT: s_and_b32 s11, s11, 0xffff
+; SI-NEXT: v_mov_b32_e32 v51, s9
+; SI-NEXT: s_or_b32 s6, s11, s9
+; SI-NEXT: v_readlane_b32 s9, v62, 3
+; SI-NEXT: v_readlane_b32 s11, v62, 2
+; SI-NEXT: s_and_b32 s9, s9, 0xff
+; SI-NEXT: s_lshl_b32 s15, s11, 8
+; SI-NEXT: s_or_b32 s9, s9, s15
+; SI-NEXT: s_and_b32 s9, s9, 0xffff
+; SI-NEXT: s_or_b32 s17, s9, s14
+; SI-NEXT: v_readlane_b32 s9, v62, 22
+; SI-NEXT: v_mov_b32_e32 v52, s14
+; SI-NEXT: s_and_b32 s14, s9, 0xff
+; SI-NEXT: v_readlane_b32 s9, v62, 21
+; SI-NEXT: s_lshl_b32 s15, s9, 8
+; SI-NEXT: s_or_b32 s14, s14, s15
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v53, v6, v1
+; SI-NEXT: s_and_b32 s14, s14, 0xffff
+; SI-NEXT: v_or_b32_e32 v50, s14, v53
+; SI-NEXT: s_and_b32 s14, s30, 0xff
+; SI-NEXT: s_lshl_b32 s15, s88, 8
+; SI-NEXT: s_waitcnt vmcnt(2)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v20
+; SI-NEXT: s_or_b32 s14, s14, s15
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v54, v14, v1
+; SI-NEXT: s_and_b32 s14, s14, 0xffff
+; SI-NEXT: v_or_b32_e32 v17, s14, v54
+; SI-NEXT: s_and_b32 s14, s8, 0xff
+; SI-NEXT: s_lshl_b32 s15, s52, 8
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v19
+; SI-NEXT: s_or_b32 s14, s14, s15
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v55, v40, v1
+; SI-NEXT: s_and_b32 s14, s14, 0xffff
+; SI-NEXT: v_or_b32_e32 v16, s14, v55
+; SI-NEXT: s_and_b32 s14, s35, 0xff
+; SI-NEXT: s_lshl_b32 s15, s28, 8
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v29
+; SI-NEXT: s_or_b32 s14, s14, s15
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt expcnt(4)
+; SI-NEXT: v_or_b32_e32 v40, v42, v1
+; SI-NEXT: s_and_b32 s14, s14, 0xffff
+; SI-NEXT: v_or_b32_e32 v15, s14, v40
+; SI-NEXT: s_and_b32 s14, s95, 0xff
+; SI-NEXT: s_lshl_b32 s15, s66, 8
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v56
+; SI-NEXT: s_or_b32 s14, s14, s15
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v41, v61, v1
+; SI-NEXT: s_and_b32 s14, s14, 0xffff
+; SI-NEXT: v_or_b32_e32 v12, s14, v41
+; SI-NEXT: s_and_b32 s14, s18, 0xff
+; SI-NEXT: s_lshl_b32 s15, s91, 8
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v57
+; SI-NEXT: s_or_b32 s14, s14, s15
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt expcnt(3)
+; SI-NEXT: v_or_b32_e32 v42, v60, v1
+; SI-NEXT: s_and_b32 s14, s14, 0xffff
+; SI-NEXT: v_or_b32_e32 v11, s14, v42
+; SI-NEXT: s_and_b32 s14, s37, 0xff
+; SI-NEXT: s_lshl_b32 s15, s19, 8
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v33
+; SI-NEXT: s_or_b32 s14, s14, s15
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_mov_b32 s91, s6
+; SI-NEXT: v_or_b32_e32 v59, v31, v1
+; SI-NEXT: s_and_b32 s14, s14, 0xffff
+; SI-NEXT: v_readlane_b32 s6, v62, 50
+; SI-NEXT: v_or_b32_e32 v10, s14, v59
+; SI-NEXT: s_and_b32 s14, s6, 0xff
+; SI-NEXT: s_lshl_b32 s15, s70, 8
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v30
+; SI-NEXT: s_or_b32 s14, s14, s15
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v5, v24, v1
+; SI-NEXT: s_and_b32 s14, s14, 0xffff
+; SI-NEXT: v_readlane_b32 s6, v62, 19
+; SI-NEXT: v_or_b32_e32 v9, s14, v5
+; SI-NEXT: s_and_b32 s14, s90, 0xff
+; SI-NEXT: s_lshl_b32 s15, s6, 8
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v35
+; SI-NEXT: s_or_b32 s14, s14, s15
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v13, v25, v1
+; SI-NEXT: s_and_b32 s14, s14, 0xffff
+; SI-NEXT: v_or_b32_e32 v8, s14, v13
+; SI-NEXT: s_and_b32 s14, s27, 0xff
+; SI-NEXT: s_lshl_b32 s15, s24, 8
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v32
+; SI-NEXT: s_or_b32 s14, s14, s15
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_mov_b32_e32 v26, v31
+; SI-NEXT: v_or_b32_e32 v31, v27, v1
+; SI-NEXT: s_and_b32 s14, s14, 0xffff
+; SI-NEXT: v_or_b32_e32 v7, s14, v31
+; SI-NEXT: s_and_b32 s14, s49, 0xff
+; SI-NEXT: s_lshl_b32 s15, s68, 8
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v38
+; SI-NEXT: s_or_b32 s14, s14, s15
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_mov_b32_e32 v28, v25
+; SI-NEXT: v_mov_b32_e32 v25, v60
+; SI-NEXT: v_or_b32_e32 v60, v43, v1
+; SI-NEXT: s_and_b32 s14, s14, 0xffff
+; SI-NEXT: v_readlane_b32 s6, v62, 17
+; SI-NEXT: v_or_b32_e32 v4, s14, v60
+; SI-NEXT: s_and_b32 s14, s6, 0xff
+; SI-NEXT: v_readlane_b32 s6, v62, 16
+; SI-NEXT: s_lshl_b32 s15, s6, 8
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v39
+; SI-NEXT: s_or_b32 s14, s14, s15
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_mov_b32_e32 v23, v27
+; SI-NEXT: v_mov_b32_e32 v27, v24
+; SI-NEXT: v_mov_b32_e32 v24, v61
+; SI-NEXT: v_or_b32_e32 v61, v44, v1
+; SI-NEXT: s_and_b32 s14, s14, 0xffff
+; SI-NEXT: v_or_b32_e32 v2, s14, v61
+; SI-NEXT: s_and_b32 s14, s71, 0xff
+; SI-NEXT: s_lshl_b32 s15, s36, 8
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v48
+; SI-NEXT: s_or_b32 s14, s14, s15
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_readlane_b32 s6, v62, 48
+; SI-NEXT: v_or_b32_e32 v6, v45, v1
+; SI-NEXT: s_and_b32 s14, s14, 0xffff
+; SI-NEXT: s_and_b32 s8, s6, 0xffff
+; SI-NEXT: v_readlane_b32 s6, v62, 49
+; SI-NEXT: v_or_b32_e32 v1, s14, v6
+; SI-NEXT: s_and_b32 s14, s79, 0xff
+; SI-NEXT: s_lshl_b32 s15, s7, 8
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v34
+; SI-NEXT: s_or_b32 s42, s8, s63
+; SI-NEXT: s_and_b32 s8, s6, 0xffff
+; SI-NEXT: s_or_b32 s14, s14, s15
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: s_or_b32 s40, s8, s21
+; SI-NEXT: s_and_b32 s8, s12, 0xffff
+; SI-NEXT: v_or_b32_e32 v14, v46, v3
+; SI-NEXT: s_and_b32 s14, s14, 0xffff
+; SI-NEXT: s_or_b32 s15, s8, s41
+; SI-NEXT: s_and_b32 s8, s13, 0xffff
+; SI-NEXT: v_or_b32_e32 v3, s14, v14
+; SI-NEXT: s_or_b32 s14, s8, s43
+; SI-NEXT: s_and_b32 s8, s16, 0xffff
+; SI-NEXT: s_and_b32 s16, s73, 0xffff
+; SI-NEXT: s_or_b32 s13, s8, s44
+; SI-NEXT: s_and_b32 s8, s62, 0xffff
+; SI-NEXT: s_or_b32 s35, s16, s56
+; SI-NEXT: s_and_b32 s16, s48, 0xffff
+; SI-NEXT: s_or_b32 s12, s8, s45
+; SI-NEXT: s_and_b32 s8, s10, 0xffff
+; SI-NEXT: s_or_b32 s52, s16, vcc_lo
+; SI-NEXT: s_and_b32 s16, s72, 0xffff
+; SI-NEXT: s_or_b32 s10, s8, s46
+; SI-NEXT: s_and_b32 s8, s94, 0xffff
+; SI-NEXT: s_or_b32 s94, s16, vcc_hi
+; SI-NEXT: s_and_b32 s16, s57, 0xffff
+; SI-NEXT: s_or_b32 s49, s16, s74
+; SI-NEXT: s_and_b32 s16, s58, 0xffff
+; SI-NEXT: s_or_b32 s48, s16, s75
+; SI-NEXT: s_and_b32 s16, s59, 0xffff
+; SI-NEXT: s_mov_b32 s25, s23
+; SI-NEXT: s_or_b32 s11, s16, s76
+; SI-NEXT: s_and_b32 s16, s60, 0xffff
+; SI-NEXT: s_and_b32 s23, s61, 0xffff
+; SI-NEXT: s_mov_b32 s87, s34
+; SI-NEXT: s_mov_b32 s34, s55
+; SI-NEXT: s_mov_b32 s55, s22
+; SI-NEXT: s_or_b32 s8, s8, s47
+; SI-NEXT: s_or_b32 s9, s16, s77
+; SI-NEXT: s_or_b32 s16, s23, s78
+; SI-NEXT: s_mov_b32 s22, s18
+; SI-NEXT: v_mov_b32_e32 v49, v30
+; SI-NEXT: v_mov_b32_e32 v36, v35
+; SI-NEXT: v_mov_b32_e32 v37, v32
+; SI-NEXT: v_mov_b32_e32 v18, v43
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_mov_b32_e32 v33, v44
+; SI-NEXT: v_mov_b32_e32 v35, v45
+; SI-NEXT: v_mov_b32_e32 v22, v46
+; SI-NEXT: v_alignbit_b32 v57, s42, v51, 16
+; SI-NEXT: v_alignbit_b32 v58, s40, v52, 16
+; SI-NEXT: v_alignbit_b32 v56, s15, v53, 16
+; SI-NEXT: v_alignbit_b32 v47, s14, v54, 16
+; SI-NEXT: v_alignbit_b32 v46, s13, v55, 16
+; SI-NEXT: v_alignbit_b32 v45, s12, v40, 16
+; SI-NEXT: v_alignbit_b32 v44, s10, v41, 16
+; SI-NEXT: v_alignbit_b32 v43, s8, v42, 16
+; SI-NEXT: v_alignbit_b32 v42, s35, v59, 16
+; SI-NEXT: v_alignbit_b32 v41, s52, v5, 16
+; SI-NEXT: v_alignbit_b32 v40, s94, v13, 16
+; SI-NEXT: v_alignbit_b32 v55, s49, v31, 16
+; SI-NEXT: v_mov_b32_e32 v31, v26
+; SI-NEXT: v_alignbit_b32 v54, s48, v60, 16
+; SI-NEXT: v_mov_b32_e32 v60, v25
+; SI-NEXT: v_mov_b32_e32 v25, v28
+; SI-NEXT: v_alignbit_b32 v53, s11, v61, 16
+; SI-NEXT: v_mov_b32_e32 v61, v24
+; SI-NEXT: v_mov_b32_e32 v24, v27
+; SI-NEXT: v_alignbit_b32 v52, s9, v6, 16
+; SI-NEXT: v_alignbit_b32 v51, s16, v14, 16
+; SI-NEXT: s_lshr_b32 s73, s63, 16
+; SI-NEXT: s_lshr_b32 s72, s21, 16
+; SI-NEXT: s_lshr_b32 s63, s41, 16
+; SI-NEXT: s_lshr_b32 s62, s43, 16
+; SI-NEXT: s_lshr_b32 s61, s44, 16
+; SI-NEXT: s_lshr_b32 s60, s45, 16
+; SI-NEXT: s_lshr_b32 s59, s46, 16
+; SI-NEXT: s_lshr_b32 s58, s47, 16
+; SI-NEXT: s_lshr_b32 s57, s56, 16
+; SI-NEXT: s_lshr_b32 s56, vcc_lo, 16
+; SI-NEXT: s_lshr_b32 s47, vcc_hi, 16
+; SI-NEXT: s_lshr_b32 s46, s74, 16
+; SI-NEXT: s_mov_b32 s74, s51
+; SI-NEXT: s_mov_b32 s51, s67
+; SI-NEXT: v_readlane_b32 s67, v62, 51
+; SI-NEXT: s_lshr_b32 s45, s75, 16
+; SI-NEXT: s_mov_b32 s23, s25
+; SI-NEXT: s_mov_b32 s21, s55
+; SI-NEXT: s_mov_b32 s55, s34
+; SI-NEXT: s_mov_b32 s75, s54
+; SI-NEXT: s_mov_b32 s34, s87
+; SI-NEXT: v_readlane_b32 s25, v62, 53
+; SI-NEXT: v_readlane_b32 s87, v62, 52
+; SI-NEXT: s_lshr_b32 s44, s76, 16
+; SI-NEXT: v_readlane_b32 s54, v62, 50
+; SI-NEXT: s_lshr_b32 s43, s77, 16
+; SI-NEXT: s_mov_b32 s76, s69
+; SI-NEXT: s_mov_b32 s69, s95
+; SI-NEXT: s_mov_b32 s77, s84
+; SI-NEXT: s_mov_b32 s84, s99
+; SI-NEXT: s_mov_b32 s99, s93
+; SI-NEXT: s_lshr_b32 s41, s78, 16
+; SI-NEXT: s_mov_b32 s78, s31
+; SI-NEXT: s_mov_b32 s31, s50
+; SI-NEXT: s_mov_b32 s50, s53
+; SI-NEXT: s_mov_b32 s53, s82
+; SI-NEXT: s_mov_b32 s82, s83
+; SI-NEXT: s_mov_b32 s83, s85
+; SI-NEXT: s_mov_b32 s85, s4
+; SI-NEXT: s_mov_b64 s[4:5], 0
+; SI-NEXT: v_mov_b32_e32 v6, v20
+; SI-NEXT: v_mov_b32_e32 v13, v19
+; SI-NEXT: v_mov_b32_e32 v14, v29
+; SI-NEXT: s_branch .LBB97_3
+; SI-NEXT: .LBB97_2:
+; SI-NEXT: s_waitcnt expcnt(6)
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(5)
+; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
+; SI-NEXT: v_mov_b32_e32 v22, v46
+; SI-NEXT: v_mov_b32_e32 v37, v32
+; SI-NEXT: v_mov_b32_e32 v36, v35
+; SI-NEXT: v_mov_b32_e32 v35, v45
+; SI-NEXT: v_mov_b32_e32 v49, v30
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_mov_b32_e32 v33, v44
+; SI-NEXT: v_mov_b32_e32 v18, v43
+; SI-NEXT: v_mov_b32_e32 v23, v27
+; SI-NEXT: v_mov_b32_e32 v21, v29
+; SI-NEXT: s_mov_b64 s[4:5], -1
+; SI-NEXT: ; implicit-def: $sgpr91
+; SI-NEXT: ; implicit-def: $vgpr57
+; SI-NEXT: ; implicit-def: $sgpr42
+; SI-NEXT: ; implicit-def: $sgpr73
+; SI-NEXT: ; implicit-def: $sgpr17
+; SI-NEXT: ; implicit-def: $vgpr58
+; SI-NEXT: ; implicit-def: $sgpr40
+; SI-NEXT: ; implicit-def: $sgpr72
+; SI-NEXT: ; implicit-def: $vgpr50
+; SI-NEXT: ; implicit-def: $vgpr56
+; SI-NEXT: ; implicit-def: $sgpr15
+; SI-NEXT: ; implicit-def: $sgpr63
+; SI-NEXT: ; implicit-def: $vgpr17
+; SI-NEXT: ; implicit-def: $vgpr47
+; SI-NEXT: ; implicit-def: $sgpr14
+; SI-NEXT: ; implicit-def: $sgpr62
+; SI-NEXT: ; implicit-def: $vgpr16
+; SI-NEXT: ; implicit-def: $vgpr46
+; SI-NEXT: ; implicit-def: $sgpr13
+; SI-NEXT: ; implicit-def: $sgpr61
+; SI-NEXT: ; implicit-def: $vgpr15
+; SI-NEXT: ; implicit-def: $vgpr45
+; SI-NEXT: ; implicit-def: $sgpr12
+; SI-NEXT: ; implicit-def: $sgpr60
+; SI-NEXT: ; implicit-def: $vgpr12
+; SI-NEXT: ; implicit-def: $vgpr44
+; SI-NEXT: ; implicit-def: $sgpr10
+; SI-NEXT: ; implicit-def: $sgpr59
+; SI-NEXT: ; implicit-def: $vgpr11
+; SI-NEXT: ; implicit-def: $vgpr43
+; SI-NEXT: ; implicit-def: $sgpr8
+; SI-NEXT: ; implicit-def: $sgpr58
+; SI-NEXT: ; implicit-def: $vgpr10
+; SI-NEXT: ; implicit-def: $vgpr42
+; SI-NEXT: ; implicit-def: $sgpr35
+; SI-NEXT: ; implicit-def: $sgpr57
+; SI-NEXT: ; implicit-def: $vgpr9
+; SI-NEXT: ; implicit-def: $vgpr41
+; SI-NEXT: ; implicit-def: $sgpr52
+; SI-NEXT: ; implicit-def: $sgpr56
+; SI-NEXT: ; implicit-def: $vgpr8
+; SI-NEXT: ; implicit-def: $vgpr40
+; SI-NEXT: ; implicit-def: $sgpr94
+; SI-NEXT: ; implicit-def: $sgpr47
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr55
+; SI-NEXT: ; implicit-def: $sgpr49
+; SI-NEXT: ; implicit-def: $sgpr46
+; SI-NEXT: ; implicit-def: $vgpr4
+; SI-NEXT: ; implicit-def: $vgpr54
+; SI-NEXT: ; implicit-def: $sgpr48
+; SI-NEXT: ; implicit-def: $sgpr45
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr53
+; SI-NEXT: ; implicit-def: $sgpr11
+; SI-NEXT: ; implicit-def: $sgpr44
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr52
+; SI-NEXT: ; implicit-def: $sgpr9
+; SI-NEXT: ; implicit-def: $sgpr43
+; SI-NEXT: ; implicit-def: $vgpr3
+; SI-NEXT: ; implicit-def: $vgpr51
+; SI-NEXT: ; implicit-def: $sgpr16
+; SI-NEXT: ; implicit-def: $sgpr41
+; SI-NEXT: .LBB97_3: ; %Flow
+; SI-NEXT: v_mov_b32_e32 v5, v39
+; SI-NEXT: s_andn2_b64 vcc, exec, s[4:5]
+; SI-NEXT: s_cbranch_vccnz .LBB97_5
+; SI-NEXT: ; %bb.4: ; %cmp.true
+; SI-NEXT: s_add_i32 s4, s79, 3
+; SI-NEXT: s_and_b32 s4, s4, 0xff
+; SI-NEXT: s_lshl_b32 s5, s7, 8
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v34
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: v_or_b32_e32 v1, v22, v1
+; SI-NEXT: v_or_b32_e32 v1, s4, v1
+; SI-NEXT: s_add_i32 s4, s82, 3
+; SI-NEXT: s_and_b32 s4, s4, 0xff
+; SI-NEXT: s_lshl_b32 s5, s53, 8
+; SI-NEXT: s_add_i32 s8, s31, 3
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_and_b32 s8, s8, 0xff
+; SI-NEXT: s_lshl_b32 s5, s50, 24
+; SI-NEXT: s_lshl_b32 s8, s8, 16
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: s_or_b32 s5, s5, s8
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_add_i32 s71, s71, 3
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_and_b32 s5, s71, 0xff
+; SI-NEXT: s_lshl_b32 s8, s36, 8
+; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v48
+; SI-NEXT: s_or_b32 s5, s8, s5
+; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
+; SI-NEXT: s_addk_i32 s5, 0x300
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT: v_readlane_b32 s7, v62, 12
+; SI-NEXT: s_and_b32 s5, s5, 0xffff
+; SI-NEXT: v_or_b32_e32 v2, v35, v2
+; SI-NEXT: s_add_i32 s36, s86, 3
+; SI-NEXT: s_lshl_b32 s8, s7, 8
+; SI-NEXT: v_readlane_b32 s7, v62, 11
+; SI-NEXT: v_or_b32_e32 v2, s5, v2
+; SI-NEXT: s_and_b32 s5, s36, 0xff
+; SI-NEXT: s_add_i32 s9, s7, 3
+; SI-NEXT: s_or_b32 s5, s8, s5
+; SI-NEXT: s_and_b32 s9, s9, 0xff
+; SI-NEXT: s_lshl_b32 s8, s78, 24
+; SI-NEXT: s_lshl_b32 s9, s9, 16
+; SI-NEXT: s_addk_i32 s5, 0x300
+; SI-NEXT: s_or_b32 s8, s8, s9
+; SI-NEXT: s_and_b32 s5, s5, 0xffff
+; SI-NEXT: s_or_b32 s5, s8, s5
+; SI-NEXT: s_add_i32 s16, s4, 0x3000000
+; SI-NEXT: v_readlane_b32 s4, v62, 17
+; SI-NEXT: s_add_i32 s9, s5, 0x3000000
+; SI-NEXT: s_add_i32 s79, s4, 3
+; SI-NEXT: v_readlane_b32 s5, v62, 16
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x3000000, v1
+; SI-NEXT: v_add_i32_e32 v1, vcc, 0x3000000, v2
+; SI-NEXT: s_and_b32 s4, s79, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 8
+; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v5
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: v_or_b32_e32 v2, v33, v2
+; SI-NEXT: v_or_b32_e32 v2, s4, v2
+; SI-NEXT: s_add_i32 s4, s77, 3
+; SI-NEXT: v_readlane_b32 s5, v62, 15
+; SI-NEXT: v_readlane_b32 s6, v62, 14
+; SI-NEXT: s_and_b32 s4, s4, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 8
+; SI-NEXT: s_add_i32 s8, s6, 3
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: v_readlane_b32 s5, v62, 13
+; SI-NEXT: s_and_b32 s8, s8, 0xff
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: s_lshl_b32 s5, s5, 24
+; SI-NEXT: s_lshl_b32 s8, s8, 16
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s5, s8
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s11, s4, 0x3000000
+; SI-NEXT: v_readlane_b32 s4, v62, 32
+; SI-NEXT: s_add_i32 s53, s4, 3
+; SI-NEXT: s_and_b32 s4, s53, 0xff
+; SI-NEXT: s_lshl_b32 s5, s68, 8
+; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v38
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: v_and_b32_e32 v4, 0xff, v4
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: v_or_b32_e32 v4, v18, v4
+; SI-NEXT: s_add_i32 s93, s98, 3
+; SI-NEXT: v_readlane_b32 s6, v62, 18
+; SI-NEXT: v_or_b32_e32 v4, s4, v4
+; SI-NEXT: s_and_b32 s4, s93, 0xff
+; SI-NEXT: s_lshl_b32 s5, s75, 8
+; SI-NEXT: s_add_i32 s8, s6, 3
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_and_b32 s8, s8, 0xff
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: s_lshl_b32 s5, s81, 24
+; SI-NEXT: s_lshl_b32 s8, s8, 16
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s5, s8
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s86, s27, 3
+; SI-NEXT: s_add_i32 s48, s4, 0x3000000
+; SI-NEXT: s_and_b32 s4, s86, 0xff
+; SI-NEXT: s_lshl_b32 s5, s24, 8
+; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v37
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: v_and_b32_e32 v5, 0xff, v5
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: v_or_b32_e32 v5, v23, v5
+; SI-NEXT: v_or_b32_e32 v5, s4, v5
+; SI-NEXT: v_readlane_b32 s4, v62, 38
+; SI-NEXT: s_add_i32 s68, s4, 3
+; SI-NEXT: s_and_b32 s4, s68, 0xff
+; SI-NEXT: s_lshl_b32 s5, s97, 8
+; SI-NEXT: s_add_i32 s8, s67, 3
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_and_b32 s8, s8, 0xff
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: s_lshl_b32 s5, s51, 24
+; SI-NEXT: s_lshl_b32 s8, s8, 16
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s5, s8
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s50, s90, 3
+; SI-NEXT: v_readlane_b32 s5, v62, 19
+; SI-NEXT: v_add_i32_e32 v7, vcc, 0x3000000, v5
+; SI-NEXT: s_add_i32 s49, s4, 0x3000000
+; SI-NEXT: s_and_b32 s4, s50, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 8
+; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v36
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: v_and_b32_e32 v5, 0xff, v5
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: v_or_b32_e32 v5, v25, v5
+; SI-NEXT: s_add_i32 s94, s20, 3
+; SI-NEXT: v_or_b32_e32 v5, s4, v5
+; SI-NEXT: s_and_b32 s4, s94, 0xff
+; SI-NEXT: s_lshl_b32 s5, s85, 8
+; SI-NEXT: s_add_i32 s8, s83, 3
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_and_b32 s8, s8, 0xff
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: s_lshl_b32 s5, s84, 24
+; SI-NEXT: s_lshl_b32 s8, s8, 16
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s5, s8
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s20, s54, 3
+; SI-NEXT: v_add_i32_e32 v8, vcc, 0x3000000, v5
+; SI-NEXT: s_add_i32 s94, s4, 0x3000000
+; SI-NEXT: s_and_b32 s4, s20, 0xff
+; SI-NEXT: s_lshl_b32 s5, s70, 8
+; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v49
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: v_and_b32_e32 v5, 0xff, v5
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: v_or_b32_e32 v5, v24, v5
+; SI-NEXT: v_or_b32_e32 v5, s4, v5
+; SI-NEXT: v_add_i32_e32 v9, vcc, 0x3000000, v5
+; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
+; SI-NEXT: s_add_i32 s98, s38, 3
+; SI-NEXT: v_readlane_b32 s5, v62, 20
+; SI-NEXT: s_and_b32 s4, s98, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 8
+; SI-NEXT: s_add_i32 s8, s92, 3
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_and_b32 s8, s8, 0xff
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: s_lshl_b32 s5, s23, 24
+; SI-NEXT: s_lshl_b32 s8, s8, 16
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s5, s8
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s17, s37, 3
+; SI-NEXT: s_add_i32 s52, s4, 0x3000000
+; SI-NEXT: s_and_b32 s4, s17, 0xff
+; SI-NEXT: s_lshl_b32 s5, s19, 8
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: v_readlane_b32 s5, v62, 34
+; SI-NEXT: v_readlane_b32 s6, v62, 29
+; SI-NEXT: s_lshl_b32 s5, s5, 8
+; SI-NEXT: s_add_i32 s38, s6, 3
+; SI-NEXT: s_and_b32 s8, s38, 0xff
+; SI-NEXT: s_lshl_b32 s8, s8, 16
+; SI-NEXT: s_add_i32 s22, s22, 3
+; SI-NEXT: s_add_i32 s85, s25, 3
+; SI-NEXT: v_readlane_b32 s6, v62, 40
+; SI-NEXT: s_add_i32 s70, s6, 3
+; SI-NEXT: s_and_b32 s7, s70, 0xff
+; SI-NEXT: s_lshl_b32 s7, s7, 16
+; SI-NEXT: s_add_i32 s19, s69, 3
+; SI-NEXT: s_add_i32 s51, s30, 3
+; SI-NEXT: s_add_i32 s95, s89, 3
+; SI-NEXT: v_mov_b32_e32 v30, s16
+; SI-NEXT: v_mov_b32_e32 v39, s9
+; SI-NEXT: v_add_i32_e32 v2, vcc, 0x3000000, v2
+; SI-NEXT: v_mov_b32_e32 v28, s11
+; SI-NEXT: v_add_i32_e32 v4, vcc, 0x3000000, v4
+; SI-NEXT: v_mov_b32_e32 v27, s48
+; SI-NEXT: v_mov_b32_e32 v26, s49
+; SI-NEXT: v_mov_b32_e32 v25, s94
+; SI-NEXT: v_mov_b32_e32 v24, s52
+; SI-NEXT: v_alignbit_b32 v41, v24, v9, 16
+; SI-NEXT: v_alignbit_b32 v40, v25, v8, 16
+; SI-NEXT: v_alignbit_b32 v55, v26, v7, 16
+; SI-NEXT: v_alignbit_b32 v54, v27, v4, 16
+; SI-NEXT: v_alignbit_b32 v53, v28, v2, 16
+; SI-NEXT: v_alignbit_b32 v52, v39, v1, 16
+; SI-NEXT: v_alignbit_b32 v51, v30, v3, 16
+; SI-NEXT: s_lshr_b32 s56, s52, 16
+; SI-NEXT: s_lshr_b32 s47, s94, 16
+; SI-NEXT: s_lshr_b32 s46, s49, 16
+; SI-NEXT: s_lshr_b32 s45, s48, 16
+; SI-NEXT: s_lshr_b32 s44, s11, 16
+; SI-NEXT: s_lshr_b32 s43, s9, 16
+; SI-NEXT: s_lshr_b32 s41, s16, 16
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5
+; SI-NEXT: v_and_b32_e32 v5, 0xff, v5
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT: v_or_b32_e32 v5, v31, v5
+; SI-NEXT: v_or_b32_e32 v5, s4, v5
+; SI-NEXT: v_add_i32_e32 v10, vcc, 0x3000000, v5
+; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
+; SI-NEXT: v_readlane_b32 s4, v62, 37
+; SI-NEXT: s_add_i32 s67, s4, 3
+; SI-NEXT: s_and_b32 s4, s67, 0xff
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: v_readlane_b32 s5, v62, 30
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: s_lshl_b32 s5, s5, 24
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s5, s8
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: v_readlane_b32 s5, v62, 46
+; SI-NEXT: s_add_i32 s35, s4, 0x3000000
+; SI-NEXT: s_and_b32 s4, s22, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 8
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s34, 8
+; SI-NEXT: v_mov_b32_e32 v23, s35
+; SI-NEXT: v_alignbit_b32 v42, v23, v10, 16
+; SI-NEXT: s_lshr_b32 s57, s35, 16
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5
+; SI-NEXT: v_and_b32_e32 v5, 0xff, v5
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT: v_or_b32_e32 v5, v60, v5
+; SI-NEXT: v_or_b32_e32 v5, s4, v5
+; SI-NEXT: v_add_i32_e32 v11, vcc, 0x3000000, v5
+; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
+; SI-NEXT: s_and_b32 s4, s85, 0xff
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: v_readlane_b32 s5, v62, 36
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: s_lshl_b32 s5, s5, 24
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s5, s7
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: v_readlane_b32 s5, v62, 47
+; SI-NEXT: s_add_i32 s8, s4, 0x3000000
+; SI-NEXT: s_and_b32 s4, s19, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 8
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: v_readlane_b32 s5, v62, 44
+; SI-NEXT: s_lshl_b32 s5, s5, 8
+; SI-NEXT: v_mov_b32_e32 v22, s8
+; SI-NEXT: v_alignbit_b32 v43, v22, v11, 16
+; SI-NEXT: s_lshr_b32 s58, s8, 16
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5
+; SI-NEXT: v_and_b32_e32 v5, 0xff, v5
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT: v_or_b32_e32 v5, v61, v5
+; SI-NEXT: v_or_b32_e32 v5, s4, v5
+; SI-NEXT: v_add_i32_e32 v12, vcc, 0x3000000, v5
+; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v14
+; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
+; SI-NEXT: v_readlane_b32 s4, v62, 45
+; SI-NEXT: s_add_i32 s6, s4, 3
+; SI-NEXT: s_and_b32 s4, s6, 0xff
+; SI-NEXT: v_readlane_b32 s6, v62, 27
+; SI-NEXT: s_add_i32 s34, s6, 3
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: v_readlane_b32 s5, v62, 28
+; SI-NEXT: s_and_b32 s6, s34, 0xff
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: s_lshl_b32 s5, s5, 24
+; SI-NEXT: s_lshl_b32 s6, s6, 16
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s5, s6
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s10, s4, 0x3000000
+; SI-NEXT: v_readlane_b32 s4, v62, 43
+; SI-NEXT: s_add_i32 s97, s4, 3
+; SI-NEXT: s_and_b32 s4, s97, 0xff
+; SI-NEXT: s_lshl_b32 s5, s28, 8
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: v_and_b32_e32 v5, 0xff, v5
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: v_readlane_b32 s5, v62, 41
+; SI-NEXT: v_readlane_b32 s6, v62, 33
+; SI-NEXT: s_lshl_b32 s5, s5, 8
+; SI-NEXT: s_add_i32 s54, s6, 3
+; SI-NEXT: s_and_b32 s6, s54, 0xff
+; SI-NEXT: s_lshl_b32 s6, s6, 16
+; SI-NEXT: v_mov_b32_e32 v20, s10
+; SI-NEXT: v_alignbit_b32 v44, v20, v12, 16
+; SI-NEXT: s_lshr_b32 s59, s10, 16
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v5, v14, v5
+; SI-NEXT: v_or_b32_e32 v5, s4, v5
+; SI-NEXT: v_add_i32_e32 v15, vcc, 0x3000000, v5
+; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v13
+; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
+; SI-NEXT: v_readlane_b32 s4, v62, 42
+; SI-NEXT: s_add_i32 s81, s4, 3
+; SI-NEXT: s_and_b32 s4, s81, 0xff
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: v_readlane_b32 s5, v62, 31
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: s_lshl_b32 s5, s5, 24
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s5, s6
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s12, s4, 0x3000000
+; SI-NEXT: v_readlane_b32 s4, v62, 39
+; SI-NEXT: s_add_i32 s69, s4, 3
+; SI-NEXT: v_readlane_b32 s5, v62, 35
+; SI-NEXT: s_and_b32 s4, s69, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 8
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: v_and_b32_e32 v5, 0xff, v5
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: v_readlane_b32 s5, v62, 25
+; SI-NEXT: v_readlane_b32 s6, v62, 24
+; SI-NEXT: s_lshl_b32 s5, s5, 8
+; SI-NEXT: s_add_i32 s92, s6, 3
+; SI-NEXT: s_and_b32 s6, s92, 0xff
+; SI-NEXT: s_lshl_b32 s6, s6, 16
+; SI-NEXT: v_mov_b32_e32 v19, s12
+; SI-NEXT: v_alignbit_b32 v45, v19, v15, 16
+; SI-NEXT: s_lshr_b32 s60, s12, 16
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v5, v13, v5
+; SI-NEXT: v_or_b32_e32 v5, s4, v5
+; SI-NEXT: v_add_i32_e32 v16, vcc, 0x3000000, v5
+; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v6
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v5, 0xff, v5
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT: v_readlane_b32 s4, v62, 26
+; SI-NEXT: s_add_i32 s31, s4, 3
+; SI-NEXT: s_and_b32 s4, s31, 0xff
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: s_lshl_b32 s5, s87, 24
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s5, s6
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s13, s4, 0x3000000
+; SI-NEXT: s_and_b32 s4, s51, 0xff
+; SI-NEXT: s_lshl_b32 s5, s88, 8
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s99, 8
+; SI-NEXT: s_add_i32 s6, s96, 3
+; SI-NEXT: s_and_b32 s6, s6, 0xff
+; SI-NEXT: s_lshl_b32 s6, s6, 16
+; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v21
+; SI-NEXT: v_and_b32_e32 v13, 0xff, v13
+; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
+; SI-NEXT: v_mov_b32_e32 v18, s13
+; SI-NEXT: v_alignbit_b32 v46, v18, v16, 16
+; SI-NEXT: s_lshr_b32 s61, s13, 16
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v5, v6, v5
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
+; SI-NEXT: v_or_b32_e32 v5, s4, v5
+; SI-NEXT: s_and_b32 s4, s95, 0xff
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: s_lshl_b32 s5, s76, 24
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s5, s6
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s14, s4, 0x3000000
+; SI-NEXT: v_readlane_b32 s4, v62, 22
+; SI-NEXT: s_add_i32 s4, s4, 3
+; SI-NEXT: v_readlane_b32 s5, v62, 21
+; SI-NEXT: s_and_b32 s4, s4, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 8
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s65, 8
+; SI-NEXT: s_add_i32 s6, s55, 3
+; SI-NEXT: s_and_b32 s6, s6, 0xff
+; SI-NEXT: s_lshl_b32 s6, s6, 16
+; SI-NEXT: v_add_i32_e32 v17, vcc, 0x3000000, v5
+; SI-NEXT: v_mov_b32_e32 v5, s14
+; SI-NEXT: v_alignbit_b32 v47, v5, v17, 16
+; SI-NEXT: s_lshr_b32 s62, s14, 16
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v6, v6, v13
+; SI-NEXT: v_or_b32_e32 v6, s4, v6
+; SI-NEXT: s_add_i32 s4, s80, 3
+; SI-NEXT: s_and_b32 s4, s4, 0xff
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: s_lshl_b32 s5, s64, 24
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s5, s6
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s15, s4, 0x3000000
+; SI-NEXT: v_readlane_b32 s4, v62, 3
+; SI-NEXT: s_add_i32 s4, s4, 3
+; SI-NEXT: v_readlane_b32 s5, v62, 2
+; SI-NEXT: s_and_b32 s4, s4, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 8
+; SI-NEXT: s_add_i32 s6, s26, 3
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: v_readlane_b32 s5, v62, 1
+; SI-NEXT: s_and_b32 s6, s6, 0xff
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: s_lshl_b32 s5, s5, 24
+; SI-NEXT: s_lshl_b32 s6, s6, 16
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s5, s6
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s17, s4, 0x3000000
+; SI-NEXT: v_readlane_b32 s4, v62, 0
+; SI-NEXT: s_add_i32 s4, s4, 3
+; SI-NEXT: s_and_b32 s4, s4, 0xff
+; SI-NEXT: s_lshl_b32 s5, s29, 8
+; SI-NEXT: s_add_i32 s6, s74, 3
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_and_b32 s6, s6, 0xff
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: s_lshl_b32 s5, s21, 24
+; SI-NEXT: s_lshl_b32 s6, s6, 16
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s5, s6
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s40, s4, 0x3000000
+; SI-NEXT: v_readlane_b32 s4, v62, 10
+; SI-NEXT: s_add_i32 s4, s4, 3
+; SI-NEXT: v_readlane_b32 s5, v62, 23
+; SI-NEXT: v_readlane_b32 s6, v62, 9
+; SI-NEXT: s_and_b32 s4, s4, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 8
+; SI-NEXT: s_add_i32 s6, s6, 3
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: v_readlane_b32 s5, v62, 8
+; SI-NEXT: s_and_b32 s6, s6, 0xff
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: s_lshl_b32 s5, s5, 24
+; SI-NEXT: s_lshl_b32 s6, s6, 16
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s5, s6
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s91, s4, 0x3000000
+; SI-NEXT: v_readlane_b32 s4, v62, 7
+; SI-NEXT: s_add_i32 s4, s4, 3
+; SI-NEXT: v_readlane_b32 s5, v62, 6
+; SI-NEXT: v_readlane_b32 s6, v62, 5
+; SI-NEXT: s_and_b32 s4, s4, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 8
+; SI-NEXT: s_add_i32 s6, s6, 3
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: v_readlane_b32 s5, v62, 4
+; SI-NEXT: s_and_b32 s6, s6, 0xff
+; SI-NEXT: s_addk_i32 s4, 0x300
+; SI-NEXT: s_lshl_b32 s5, s5, 24
+; SI-NEXT: s_lshl_b32 s6, s6, 16
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s5, s6
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s42, s4, 0x3000000
+; SI-NEXT: v_mov_b32_e32 v13, s91
+; SI-NEXT: v_add_i32_e32 v50, vcc, 0x3000000, v6
+; SI-NEXT: v_mov_b32_e32 v6, s15
+; SI-NEXT: v_alignbit_b32 v57, s42, v13, 16
+; SI-NEXT: v_mov_b32_e32 v13, s17
+; SI-NEXT: v_alignbit_b32 v58, s40, v13, 16
+; SI-NEXT: v_alignbit_b32 v56, v6, v50, 16
+; SI-NEXT: s_lshr_b32 s73, s42, 16
+; SI-NEXT: s_lshr_b32 s72, s40, 16
+; SI-NEXT: s_lshr_b32 s63, s15, 16
+; SI-NEXT: .LBB97_5: ; %end
+; SI-NEXT: s_and_b32 s4, s91, 0xffff
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v57
+; SI-NEXT: v_or_b32_e32 v5, s4, v5
+; SI-NEXT: s_and_b32 s4, s42, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s73, 16
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: buffer_store_dword v5, v0, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v5, vcc, 4, v0
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: v_mov_b32_e32 v6, s4
+; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
+; SI-NEXT: s_and_b32 s4, s17, 0xffff
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v58
+; SI-NEXT: v_or_b32_e32 v5, s4, v5
+; SI-NEXT: s_and_b32 s4, s40, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s72, 16
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v6, vcc, 8, v0
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v5, vcc, 12, v0
+; SI-NEXT: v_mov_b32_e32 v6, s4
+; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
+; SI-NEXT: v_and_b32_e32 v5, 0xffff, v50
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v56
+; SI-NEXT: s_and_b32 s4, s15, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s63, 16
+; SI-NEXT: v_or_b32_e32 v5, v5, v6
+; SI-NEXT: v_add_i32_e32 v6, vcc, 16, v0
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v5, vcc, 20, v0
+; SI-NEXT: v_mov_b32_e32 v6, s4
+; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
+; SI-NEXT: v_and_b32_e32 v5, 0xffff, v17
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v47
+; SI-NEXT: s_and_b32 s4, s14, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s62, 16
+; SI-NEXT: v_or_b32_e32 v5, v5, v6
+; SI-NEXT: v_add_i32_e32 v6, vcc, 24, v0
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v5, vcc, 28, v0
+; SI-NEXT: v_mov_b32_e32 v6, s4
+; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
+; SI-NEXT: v_and_b32_e32 v5, 0xffff, v16
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v46
+; SI-NEXT: s_and_b32 s4, s13, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s61, 16
+; SI-NEXT: v_or_b32_e32 v5, v5, v6
+; SI-NEXT: v_add_i32_e32 v6, vcc, 32, v0
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v5, vcc, 36, v0
+; SI-NEXT: v_mov_b32_e32 v6, s4
+; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
+; SI-NEXT: v_and_b32_e32 v5, 0xffff, v15
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v45
+; SI-NEXT: s_and_b32 s4, s12, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s60, 16
+; SI-NEXT: v_or_b32_e32 v5, v5, v6
+; SI-NEXT: v_add_i32_e32 v6, vcc, 40, v0
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v5, vcc, 44, v0
+; SI-NEXT: v_mov_b32_e32 v6, s4
+; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
+; SI-NEXT: v_and_b32_e32 v5, 0xffff, v12
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v44
+; SI-NEXT: s_and_b32 s4, s10, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s59, 16
+; SI-NEXT: v_or_b32_e32 v5, v5, v6
+; SI-NEXT: v_add_i32_e32 v6, vcc, 48, v0
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v5, vcc, 52, v0
+; SI-NEXT: v_mov_b32_e32 v6, s4
+; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
+; SI-NEXT: v_and_b32_e32 v5, 0xffff, v11
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v43
+; SI-NEXT: s_and_b32 s4, s8, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s58, 16
+; SI-NEXT: v_or_b32_e32 v5, v5, v6
+; SI-NEXT: v_add_i32_e32 v6, vcc, 56, v0
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v5, vcc, 60, v0
+; SI-NEXT: v_mov_b32_e32 v6, s4
+; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
+; SI-NEXT: v_and_b32_e32 v5, 0xffff, v10
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v42
+; SI-NEXT: s_and_b32 s4, s35, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s57, 16
+; SI-NEXT: v_or_b32_e32 v5, v5, v6
+; SI-NEXT: v_add_i32_e32 v6, vcc, 64, v0
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v5, vcc, 0x44, v0
+; SI-NEXT: v_mov_b32_e32 v6, s4
+; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
+; SI-NEXT: v_and_b32_e32 v5, 0xffff, v9
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v41
+; SI-NEXT: s_and_b32 s4, s52, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s56, 16
+; SI-NEXT: v_or_b32_e32 v5, v5, v6
+; SI-NEXT: v_add_i32_e32 v6, vcc, 0x48, v0
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v5, vcc, 0x4c, v0
+; SI-NEXT: v_mov_b32_e32 v6, s4
+; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
+; SI-NEXT: v_and_b32_e32 v5, 0xffff, v8
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v40
+; SI-NEXT: s_and_b32 s4, s94, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s47, 16
+; SI-NEXT: v_or_b32_e32 v5, v5, v6
+; SI-NEXT: v_add_i32_e32 v6, vcc, 0x50, v0
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v5, vcc, 0x54, v0
+; SI-NEXT: v_mov_b32_e32 v6, s4
+; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
+; SI-NEXT: v_and_b32_e32 v5, 0xffff, v7
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v55
+; SI-NEXT: s_and_b32 s4, s49, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s46, 16
+; SI-NEXT: v_or_b32_e32 v5, v5, v6
+; SI-NEXT: v_add_i32_e32 v6, vcc, 0x58, v0
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v5, vcc, 0x5c, v0
+; SI-NEXT: v_mov_b32_e32 v6, s4
+; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
+; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v54
+; SI-NEXT: s_and_b32 s4, s48, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s45, 16
+; SI-NEXT: v_or_b32_e32 v4, v4, v5
+; SI-NEXT: v_add_i32_e32 v5, vcc, 0x60, v0
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: buffer_store_dword v4, v5, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v4, vcc, 0x64, v0
+; SI-NEXT: v_mov_b32_e32 v5, s4
+; SI-NEXT: buffer_store_dword v5, v4, s[0:3], 0 offen
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v53
+; SI-NEXT: s_and_b32 s4, s11, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s44, 16
+; SI-NEXT: v_or_b32_e32 v2, v2, v4
+; SI-NEXT: v_add_i32_e32 v4, vcc, 0x68, v0
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: buffer_store_dword v2, v4, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v2, vcc, 0x6c, v0
+; SI-NEXT: v_mov_b32_e32 v4, s4
+; SI-NEXT: buffer_store_dword v4, v2, s[0:3], 0 offen
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v52
+; SI-NEXT: s_and_b32 s4, s9, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s43, 16
+; SI-NEXT: v_or_b32_e32 v1, v1, v2
+; SI-NEXT: v_add_i32_e32 v2, vcc, 0x70, v0
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 0x74, v0
+; SI-NEXT: v_mov_b32_e32 v2, s4
+; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v3
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v51
+; SI-NEXT: s_and_b32 s4, s16, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s41, 16
+; SI-NEXT: v_or_b32_e32 v1, v1, v2
+; SI-NEXT: v_add_i32_e32 v2, vcc, 0x78, v0
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
+; SI-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_mov_b32_e32 v1, s4
+; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
+; SI-NEXT: v_readlane_b32 s99, v63, 35
+; SI-NEXT: v_readlane_b32 s98, v63, 34
+; SI-NEXT: v_readlane_b32 s97, v63, 33
+; SI-NEXT: v_readlane_b32 s96, v63, 32
+; SI-NEXT: v_readlane_b32 s87, v63, 31
+; SI-NEXT: v_readlane_b32 s86, v63, 30
+; SI-NEXT: v_readlane_b32 s85, v63, 29
+; SI-NEXT: v_readlane_b32 s84, v63, 28
+; SI-NEXT: v_readlane_b32 s83, v63, 27
+; SI-NEXT: v_readlane_b32 s82, v63, 26
+; SI-NEXT: v_readlane_b32 s81, v63, 25
+; SI-NEXT: v_readlane_b32 s80, v63, 24
+; SI-NEXT: v_readlane_b32 s71, v63, 23
+; SI-NEXT: v_readlane_b32 s70, v63, 22
+; SI-NEXT: v_readlane_b32 s69, v63, 21
+; SI-NEXT: v_readlane_b32 s68, v63, 20
+; SI-NEXT: v_readlane_b32 s67, v63, 19
+; SI-NEXT: v_readlane_b32 s66, v63, 18
+; SI-NEXT: v_readlane_b32 s65, v63, 17
+; SI-NEXT: v_readlane_b32 s64, v63, 16
+; SI-NEXT: v_readlane_b32 s55, v63, 15
+; SI-NEXT: v_readlane_b32 s54, v63, 14
+; SI-NEXT: v_readlane_b32 s53, v63, 13
+; SI-NEXT: v_readlane_b32 s52, v63, 12
+; SI-NEXT: v_readlane_b32 s51, v63, 11
+; SI-NEXT: v_readlane_b32 s50, v63, 10
+; SI-NEXT: v_readlane_b32 s49, v63, 9
+; SI-NEXT: v_readlane_b32 s48, v63, 8
+; SI-NEXT: v_readlane_b32 s39, v63, 7
+; SI-NEXT: v_readlane_b32 s38, v63, 6
+; SI-NEXT: v_readlane_b32 s37, v63, 5
+; SI-NEXT: v_readlane_b32 s36, v63, 4
+; SI-NEXT: v_readlane_b32 s35, v63, 3
+; SI-NEXT: v_readlane_b32 s34, v63, 2
+; SI-NEXT: v_readlane_b32 s31, v63, 1
+; SI-NEXT: v_readlane_b32 s30, v63, 0
+; SI-NEXT: s_or_saveexec_b64 s[4:5], -1
+; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
+; SI-NEXT: s_mov_b64 exec, s[4:5]
+; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; SI-NEXT: s_setpc_b64 s[30:31]
+;
+; VI-LABEL: bitcast_v128i8_to_v64i16_scalar:
+; 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:396 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; VI-NEXT: v_mov_b32_e32 v46, v16
+; VI-NEXT: v_mov_b32_e32 v60, v5
+; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:332
+; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32
+; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:16
+; VI-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:32
+; VI-NEXT: buffer_load_ushort v5, off, s[0:3], s32 offset:48
+; VI-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:64
+; VI-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:80
+; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:96
+; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:112
+; VI-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:128
+; VI-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:144
+; VI-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:160
+; VI-NEXT: buffer_load_ushort v37, off, s[0:3], s32 offset:12
+; VI-NEXT: buffer_load_ushort v36, off, s[0:3], s32 offset:44
+; VI-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:60
+; VI-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:56
+; VI-NEXT: v_lshlrev_b32_e32 v11, 8, v11
+; VI-NEXT: v_mov_b32_e32 v62, v21
+; VI-NEXT: v_mov_b32_e32 v47, v17
+; VI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; VI-NEXT: v_lshlrev_b32_e32 v7, 8, v7
+; VI-NEXT: buffer_load_ushort v54, off, s[0:3], s32 offset:76
+; VI-NEXT: buffer_load_ushort v25, off, s[0:3], s32 offset:124
+; VI-NEXT: buffer_load_ushort v41, off, s[0:3], s32 offset:324
+; VI-NEXT: s_waitcnt vmcnt(14)
+; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v1
+; VI-NEXT: s_and_b64 s[4:5], vcc, exec
+; VI-NEXT: s_waitcnt vmcnt(12)
+; VI-NEXT: v_lshlrev_b32_e32 v59, 8, v26
+; VI-NEXT: s_waitcnt vmcnt(10)
+; VI-NEXT: v_lshlrev_b32_e32 v58, 8, v10
+; VI-NEXT: s_waitcnt vmcnt(9)
+; VI-NEXT: v_lshlrev_b32_e32 v21, 8, v18
+; VI-NEXT: s_waitcnt vmcnt(7)
+; VI-NEXT: v_lshlrev_b32_e32 v28, 8, v14
+; VI-NEXT: s_waitcnt vmcnt(3)
+; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
+; VI-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:52
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
+; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:92
+; VI-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:88
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
+; VI-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:84
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
+; VI-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:108
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
+; VI-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:140
+; VI-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:28
+; VI-NEXT: buffer_load_ushort v38, off, s[0:3], s32 offset:24
+; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:20
+; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v2
+; VI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
+; VI-NEXT: v_lshlrev_b32_e32 v11, 8, v15
+; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v31
+; VI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
+; VI-NEXT: v_lshlrev_b32_e32 v11, 8, v19
+; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v5
+; VI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
+; VI-NEXT: v_lshlrev_b32_e32 v11, 8, v23
+; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v6
+; VI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
+; VI-NEXT: v_lshlrev_b32_e32 v11, 8, v27
+; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v30
+; VI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
+; VI-NEXT: buffer_load_ushort v23, off, s[0:3], s32 offset:156
+; VI-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:152
+; VI-NEXT: buffer_load_ushort v5, off, s[0:3], s32 offset:148
+; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:176
+; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:192
+; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:208
+; VI-NEXT: buffer_load_ushort v35, off, s[0:3], s32 offset:172
+; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:188
+; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:184
+; VI-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:180
+; VI-NEXT: buffer_load_ushort v55, off, s[0:3], s32 offset:204
+; VI-NEXT: v_lshlrev_b32_e32 v31, 8, v22
+; VI-NEXT: buffer_load_ushort v27, off, s[0:3], s32 offset:276
+; VI-NEXT: s_waitcnt vmcnt(8)
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v0
+; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
+; VI-NEXT: s_waitcnt vmcnt(8)
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v1
+; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
+; VI-NEXT: s_waitcnt vmcnt(8)
+; VI-NEXT: v_lshlrev_b32_e32 v16, 8, v2
+; 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:240
+; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:256
+; VI-NEXT: buffer_load_ushort v11, off, s[0:3], s32 offset:220
+; VI-NEXT: s_waitcnt vmcnt(3)
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v0
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
+; VI-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:216
+; VI-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:212
+; VI-NEXT: buffer_load_ushort v11, off, s[0:3], s32 offset:236
+; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v1
+; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v2
+; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
+; VI-NEXT: s_waitcnt vmcnt(3)
+; VI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
+; VI-NEXT: buffer_load_ushort v11, off, s[0:3], s32 offset:252
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
+; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:272
+; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:288
+; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:268
+; VI-NEXT: s_waitcnt vmcnt(2)
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v0
+; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
+; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:284
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v1
+; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
+; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:280
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:304
+; VI-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:320
+; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:300
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_lshlrev_b32_e32 v1, 8, v1
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
+; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:316
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
+; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:312
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
+; VI-NEXT: buffer_load_ushort v43, off, s[0:3], s32 offset:308
+; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:328
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
+; VI-NEXT: v_lshlrev_b32_e32 v2, 8, v0
+; VI-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:120
+; VI-NEXT: buffer_load_ushort v19, off, s[0:3], s32 offset:116
+; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:248
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; VI-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:244
+; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:296
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
+; VI-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:292
+; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 offset:264
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
+; VI-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:260
+; VI-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:232
+; VI-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:228
+; VI-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:200
+; VI-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:196
+; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:168
+; VI-NEXT: buffer_load_ushort v61, off, s[0:3], s32 offset:164
+; VI-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:136
+; VI-NEXT: buffer_load_ushort v40, off, s[0:3], s32 offset:132
+; VI-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:104
+; VI-NEXT: buffer_load_ushort v34, off, s[0:3], s32 offset:100
+; VI-NEXT: buffer_load_ushort v15, off, s[0:3], s32 offset:72
+; VI-NEXT: buffer_load_ushort v29, off, s[0:3], s32 offset:68
+; VI-NEXT: buffer_load_ushort v17, off, s[0:3], s32 offset:40
+; VI-NEXT: buffer_load_ushort v33, off, s[0:3], s32 offset:36
+; VI-NEXT: buffer_load_ushort v11, off, s[0:3], s32 offset:8
+; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:4
+; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
+; VI-NEXT: s_waitcnt vmcnt(14)
+; VI-NEXT: buffer_store_dword v50, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v37, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v52, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v48, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
+; VI-NEXT: s_cbranch_scc0 .LBB97_2
+; VI-NEXT: ; %bb.1: ; %cmp.false
+; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
+; VI-NEXT: v_lshlrev_b32_e32 v1, 8, v9
+; VI-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v60
+; VI-NEXT: v_mov_b32_e32 v28, v26
+; VI-NEXT: v_mov_b32_e32 v26, v23
+; VI-NEXT: v_mov_b32_e32 v23, v5
+; VI-NEXT: v_or_b32_sdwa v0, v4, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_mov_b32_e32 v31, v22
+; VI-NEXT: v_mov_b32_e32 v22, v6
+; VI-NEXT: v_mov_b32_e32 v59, v10
+; VI-NEXT: v_mov_b32_e32 v58, v43
+; VI-NEXT: v_mov_b32_e32 v43, v27
+; VI-NEXT: v_mov_b32_e32 v27, v14
+; VI-NEXT: s_and_b32 s4, s28, 0xff
+; VI-NEXT: s_lshl_b32 s5, s29, 8
+; VI-NEXT: s_or_b32 s4, s4, s5
+; VI-NEXT: s_and_b32 s4, s4, 0xffff
+; VI-NEXT: s_lshl_b32 s5, s17, 8
+; VI-NEXT: s_lshl_b32 s6, s19, 8
+; VI-NEXT: s_lshl_b32 s7, s23, 8
+; VI-NEXT: s_lshl_b32 s8, s27, 8
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
+; VI-NEXT: v_or_b32_sdwa v4, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v13
+; VI-NEXT: v_or_b32_sdwa v0, v12, v0 dst_sel:WORD_1 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, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_or_b32_sdwa v5, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
+; VI-NEXT: v_mov_b32_e32 v3, v38
+; VI-NEXT: s_waitcnt vmcnt(0)
+; 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 v6, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; 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:416 ; 4-byte Folded Reload
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v47
+; VI-NEXT: v_or_b32_sdwa v0, v46, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; 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_or_b32_sdwa v7, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v62
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; 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:404 ; 4-byte Folded Reload
+; 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_or_b32_sdwa v8, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v0
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; 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:400 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
+; VI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
+; VI-NEXT: s_waitcnt vmcnt(2)
+; 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_or_b32_sdwa v9, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
+; VI-NEXT: v_mov_b32_e32 v2, v63
+; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
+; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
+; VI-NEXT: s_waitcnt vmcnt(3)
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v0
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v1, v46, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_or_b32_sdwa v10, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v11
+; VI-NEXT: v_or_b32_sdwa v0, v44, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(0)
+; 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 v11, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v3
+; VI-NEXT: v_or_b32_sdwa v0, v2, v0 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:484 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(2)
+; VI-NEXT: v_or_b32_sdwa v1, v37, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_or_b32_sdwa v12, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v17
+; VI-NEXT: v_or_b32_sdwa v0, v33, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: s_waitcnt vmcnt(2)
+; VI-NEXT: v_or_b32_sdwa v2, v54, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_mov_b32_e32 v33, v36
+; VI-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(2)
+; VI-NEXT: v_or_b32_sdwa v3, v24, v3 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 v1, v52, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; VI-NEXT: v_or_b32_sdwa v13, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v52
+; VI-NEXT: v_or_b32_sdwa v0, v37, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v1, v36, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_or_b32_sdwa v14, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v15
+; VI-NEXT: v_or_b32_sdwa v0, v29, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_mov_b32_e32 v36, v49
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v1, v48, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_or_b32_sdwa v15, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v63
+; VI-NEXT: v_lshlrev_b32_e32 v1, 8, v49
+; VI-NEXT: v_or_b32_sdwa v0, v62, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_or_b32_sdwa v1, v34, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_or_b32_sdwa v16, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: v_or_b32_sdwa v17, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; 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:492 ; 4-byte Folded Reload
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v18
+; VI-NEXT: v_or_b32_sdwa v0, v19, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_mov_b32_e32 v48, v32
+; VI-NEXT: v_mov_b32_e32 v34, v40
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_or_b32_sdwa v18, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v57
+; VI-NEXT: v_or_b32_sdwa v0, v40, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_or_b32_sdwa v1, v25, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_or_b32_sdwa v19, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v22
+; VI-NEXT: v_or_b32_sdwa v0, v23, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_mov_b32_e32 v40, v60
+; VI-NEXT: v_mov_b32_e32 v57, v61
+; VI-NEXT: s_waitcnt vmcnt(0)
+; 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 v20, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v32
+; VI-NEXT: v_or_b32_sdwa v0, v61, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_mov_b32_e32 v61, v38
+; VI-NEXT: s_waitcnt vmcnt(0)
+; 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 v21, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v59
+; VI-NEXT: v_or_b32_sdwa v0, v27, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v1, v35, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_or_b32_sdwa v22, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v56
+; VI-NEXT: v_or_b32_sdwa v0, v51, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_mov_b32_e32 v56, v45
+; VI-NEXT: v_mov_b32_e32 v51, v42
+; VI-NEXT: s_waitcnt vmcnt(0)
+; 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 v23, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v28
+; VI-NEXT: v_or_b32_sdwa v0, v31, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v1, v55, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_or_b32_sdwa v24, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v45
+; VI-NEXT: v_or_b32_sdwa v0, v50, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; VI-NEXT: v_mov_b32_e32 v50, v53
+; VI-NEXT: v_mov_b32_e32 v55, v63
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_or_b32_sdwa v25, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(2)
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v59
+; VI-NEXT: v_or_b32_sdwa v0, v53, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_or_b32_sdwa v26, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(2)
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v53
+; VI-NEXT: v_or_b32_sdwa v0, v42, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_mov_b32_e32 v42, v44
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_or_b32_sdwa v27, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(3)
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v0
+; VI-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_or_b32_sdwa v28, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(2)
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v49
+; VI-NEXT: v_or_b32_sdwa v0, v39, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_mov_b32_e32 v43, v46
+; VI-NEXT: v_mov_b32_e32 v39, v41
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_or_b32_sdwa v29, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(3)
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v0
+; VI-NEXT: v_or_b32_sdwa v0, v58, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_or_b32_sdwa v30, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(2)
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v54
+; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
+; VI-NEXT: v_or_b32_sdwa v0, v41, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
+; VI-NEXT: v_mov_b32_e32 v41, v52
+; VI-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; VI-NEXT: v_mov_b32_e32 v63, v54
+; VI-NEXT: v_mov_b32_e32 v54, v49
+; VI-NEXT: v_mov_b32_e32 v49, v53
+; VI-NEXT: s_waitcnt vmcnt(7)
+; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_or_b32_sdwa v31, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: v_lshlrev_b32_e32 v0, 8, v38
+; VI-NEXT: v_or_b32_sdwa v0, v47, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT: v_or_b32_e32 v3, s4, v0
+; VI-NEXT: s_and_b32 s4, s16, 0xff
+; VI-NEXT: s_or_b32 s4, s4, s5
+; VI-NEXT: s_and_b32 s5, s18, 0xff
+; VI-NEXT: s_or_b32 s5, s5, s6
+; VI-NEXT: s_and_b32 s4, s4, 0xffff
+; VI-NEXT: s_lshl_b32 s5, s5, 16
+; VI-NEXT: s_or_b32 s4, s4, s5
+; VI-NEXT: s_and_b32 s5, s20, 0xff
+; VI-NEXT: s_lshl_b32 s6, s21, 8
+; VI-NEXT: s_or_b32 s5, s5, s6
+; VI-NEXT: s_and_b32 s6, s22, 0xff
+; VI-NEXT: s_or_b32 s6, s6, s7
+; VI-NEXT: s_and_b32 s5, s5, 0xffff
+; VI-NEXT: s_lshl_b32 s6, s6, 16
+; VI-NEXT: s_or_b32 s5, s5, s6
+; VI-NEXT: s_and_b32 s6, s24, 0xff
+; VI-NEXT: s_lshl_b32 s7, s25, 8
+; VI-NEXT: s_or_b32 s6, s6, s7
+; VI-NEXT: s_and_b32 s7, s26, 0xff
+; VI-NEXT: s_or_b32 s7, s7, s8
+; VI-NEXT: s_and_b32 s6, s6, 0xffff
+; VI-NEXT: s_lshl_b32 s7, s7, 16
+; VI-NEXT: s_or_b32 s6, s6, s7
+; VI-NEXT: v_mov_b32_e32 v0, s4
+; VI-NEXT: v_mov_b32_e32 v1, s5
+; VI-NEXT: v_mov_b32_e32 v2, s6
+; VI-NEXT: s_mov_b64 s[4:5], 0
+; VI-NEXT: s_branch .LBB97_3
+; VI-NEXT: .LBB97_2:
+; VI-NEXT: v_mov_b32_e32 v34, v40
+; VI-NEXT: v_mov_b32_e32 v57, v61
+; VI-NEXT: v_mov_b32_e32 v48, v32
+; VI-NEXT: v_mov_b32_e32 v56, v45
+; VI-NEXT: v_mov_b32_e32 v51, v42
+; VI-NEXT: v_mov_b32_e32 v39, v41
+; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; VI-NEXT: v_mov_b32_e32 v33, v36
+; VI-NEXT: v_mov_b32_e32 v36, v49
+; VI-NEXT: v_mov_b32_e32 v35, v63
+; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; VI-NEXT: v_mov_b32_e32 v50, v53
+; VI-NEXT: s_mov_b64 s[4:5], -1
+; VI-NEXT: v_mov_b32_e32 v52, v38
+; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
+; VI-NEXT: ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
+; VI-NEXT: .LBB97_3: ; %Flow
+; VI-NEXT: v_mov_b32_e32 v38, v48
+; VI-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
+; VI-NEXT: s_andn2_b64 vcc, exec, s[4:5]
+; VI-NEXT: s_cbranch_vccnz .LBB97_5
+; VI-NEXT: ; %bb.4: ; %cmp.true
+; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(14)
+; VI-NEXT: v_add_u32_e32 v29, vcc, 3, v47
+; VI-NEXT: v_and_b32_e32 v29, 0xff, v29
+; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
+; VI-NEXT: s_add_i32 s28, s28, 3
+; VI-NEXT: s_lshl_b32 s4, s29, 8
+; VI-NEXT: s_and_b32 s5, s28, 0xff
+; VI-NEXT: s_or_b32 s4, s4, s5
+; VI-NEXT: s_addk_i32 s4, 0x300
+; VI-NEXT: s_and_b32 s4, s4, 0xffff
+; VI-NEXT: v_mov_b32_e32 v53, v34
+; VI-NEXT: s_add_i32 s24, s24, 3
+; VI-NEXT: s_add_i32 s20, s20, 3
+; VI-NEXT: s_lshl_b32 s9, s25, 8
+; VI-NEXT: s_and_b32 s10, s24, 0xff
+; VI-NEXT: s_add_i32 s16, s16, 3
+; VI-NEXT: s_lshl_b32 s8, s21, 8
+; VI-NEXT: s_or_b32 s9, s9, s10
+; VI-NEXT: s_and_b32 s10, s20, 0xff
+; VI-NEXT: s_add_i32 s18, s18, 3
+; VI-NEXT: s_lshl_b32 s7, s17, 8
+; VI-NEXT: s_or_b32 s8, s8, s10
+; VI-NEXT: s_and_b32 s10, s16, 0xff
+; VI-NEXT: s_or_b32 s7, s7, s10
+; VI-NEXT: s_and_b32 s12, s18, 0xff
+; VI-NEXT: s_lshl_b32 s6, s19, 24
+; VI-NEXT: s_add_i32 s22, s22, 3
+; VI-NEXT: s_addk_i32 s7, 0x300
+; VI-NEXT: s_lshl_b32 s12, s12, 16
+; VI-NEXT: s_add_i32 s26, s26, 3
+; VI-NEXT: s_addk_i32 s8, 0x300
+; VI-NEXT: s_and_b32 s11, s22, 0xff
+; VI-NEXT: s_and_b32 s7, s7, 0xffff
+; VI-NEXT: s_or_b32 s6, s6, s12
+; VI-NEXT: s_lshl_b32 s5, s23, 24
+; VI-NEXT: v_add_u32_e32 v39, vcc, 3, v39
+; VI-NEXT: s_and_b32 s10, s26, 0xff
+; VI-NEXT: s_or_b32 s6, s6, s7
+; VI-NEXT: s_and_b32 s7, s8, 0xffff
+; VI-NEXT: s_lshl_b32 s8, s11, 16
+; VI-NEXT: s_addk_i32 s9, 0x300
+; VI-NEXT: v_and_b32_e32 v39, 0xff, v39
+; VI-NEXT: s_or_b32 s5, s5, s8
+; VI-NEXT: s_lshl_b32 s8, s10, 16
+; VI-NEXT: s_waitcnt vmcnt(6)
+; VI-NEXT: v_lshlrev_b32_e32 v31, 24, v63
+; VI-NEXT: s_or_b32 s5, s5, s7
+; VI-NEXT: s_and_b32 s7, s9, 0xffff
+; VI-NEXT: s_add_i32 s6, s6, 0x3000000
+; VI-NEXT: s_add_i32 s5, s5, 0x3000000
+; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(6)
+; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
+; VI-NEXT: s_waitcnt vmcnt(5)
+; VI-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(2)
+; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
+; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
+; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
+; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
+; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
+; VI-NEXT: v_or_b32_sdwa v22, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v23, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v24, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v25, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v25, vcc, 0x300, v25
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v26, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v27, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v28, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v28, vcc, 0x300, v28
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v21, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v21, vcc, 0x300, v21
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v20, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v20, vcc, 0x300, v20
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v19, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v19, vcc, 0x300, v19
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v18, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v18, vcc, 0x300, v18
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v16, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v16, vcc, 0x300, v16
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v17, v4, v3 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:480 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v48
+; VI-NEXT: v_add_u32_e32 v17, vcc, 0x300, v17
+; VI-NEXT: v_add_u32_e32 v48, vcc, 0x300, v23
+; VI-NEXT: v_add_u32_e32 v23, vcc, 0x300, v27
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v15, v4, v3 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:476 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v33
+; VI-NEXT: v_add_u32_e32 v15, vcc, 0x300, v15
+; VI-NEXT: v_lshlrev_b32_e32 v33, 24, v59
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v14, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v14, vcc, 0x300, v14
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v13, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v13, vcc, 0x300, v13
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v12, v4, v3 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:460 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v42
+; VI-NEXT: v_add_u32_e32 v12, vcc, 0x300, v12
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v11, v4, v3 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:456 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v43
+; VI-NEXT: v_add_u32_e32 v11, vcc, 0x300, v11
+; VI-NEXT: v_add_u32_e32 v43, vcc, 3, v50
+; VI-NEXT: v_add_u32_e32 v50, vcc, 0x300, v24
+; VI-NEXT: v_add_u32_e32 v24, vcc, 0x300, v26
+; VI-NEXT: v_and_b32_e32 v26, 0xff, v43
+; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v26
+; VI-NEXT: v_or_b32_e32 v26, v33, v26
+; VI-NEXT: v_or_b32_sdwa v26, v26, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: v_add_u32_e32 v26, vcc, 0x3000000, v26
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v10, v4, v3 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:452 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v44
+; VI-NEXT: v_add_u32_e32 v10, vcc, 0x300, v10
+; VI-NEXT: v_add_u32_e32 v44, vcc, 3, v51
+; VI-NEXT: v_and_b32_e32 v27, 0xff, v44
+; VI-NEXT: v_lshlrev_b32_e32 v27, 16, v27
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v9, v4, v3 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:448 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v45
+; VI-NEXT: v_add_u32_e32 v9, vcc, 0x300, v9
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v8, v4, v3 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:444 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v46
+; VI-NEXT: v_add_u32_e32 v8, vcc, 0x300, v8
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v7, v4, v3 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:440 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v40
+; VI-NEXT: v_add_u32_e32 v7, vcc, 0x300, v7
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v6, v4, v3 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:740 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v58
+; VI-NEXT: v_add_u32_e32 v6, vcc, 0x300, v6
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v5, v4, v3 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:732 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v32
+; VI-NEXT: v_add_u32_e32 v5, vcc, 0x300, v5
+; VI-NEXT: v_lshlrev_b32_e32 v32, 24, v49
+; VI-NEXT: v_or_b32_e32 v27, v32, v27
+; VI-NEXT: v_or_b32_sdwa v27, v27, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: v_add_u32_e32 v27, vcc, 0x3000000, v27
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT: v_lshlrev_b32_e32 v4, 24, v61
+; VI-NEXT: v_or_b32_e32 v4, v4, v29
+; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v3, vcc, 0x300, v3
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
+; VI-NEXT: v_or_b32_e32 v29, v29, v30
+; VI-NEXT: v_or_b32_sdwa v29, v29, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: v_or_b32_e32 v3, s4, v4
+; VI-NEXT: v_add_u32_e32 v4, vcc, 0x3000000, v29
+; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
+; VI-NEXT: s_lshl_b32 s4, s27, 24
+; VI-NEXT: s_or_b32 s4, s4, s8
+; VI-NEXT: s_or_b32 s4, s4, s7
+; VI-NEXT: s_add_i32 s4, s4, 0x3000000
+; VI-NEXT: v_add_u32_e32 v3, vcc, 0x3000000, v3
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
+; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
+; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
+; VI-NEXT: v_or_b32_e32 v29, v29, v30
+; VI-NEXT: v_or_b32_sdwa v5, v29, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v5, vcc, 0x3000000, v5
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
+; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
+; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
+; VI-NEXT: v_or_b32_e32 v29, v29, v30
+; VI-NEXT: v_or_b32_sdwa v6, v29, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v6, vcc, 0x3000000, v6
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
+; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
+; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
+; VI-NEXT: v_or_b32_e32 v29, v29, v30
+; VI-NEXT: v_or_b32_sdwa v7, v29, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v7, vcc, 0x3000000, v7
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
+; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
+; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
+; VI-NEXT: v_or_b32_e32 v29, v29, v30
+; VI-NEXT: v_or_b32_sdwa v8, v29, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v8, vcc, 0x3000000, v8
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
+; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
+; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
+; VI-NEXT: v_or_b32_e32 v29, v29, v30
+; VI-NEXT: v_or_b32_sdwa v9, v29, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v9, vcc, 0x3000000, v9
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
+; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
+; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
+; VI-NEXT: v_or_b32_e32 v29, v29, v30
+; VI-NEXT: v_or_b32_sdwa v10, v29, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v10, vcc, 0x3000000, v10
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
+; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
+; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
+; VI-NEXT: v_or_b32_e32 v29, v29, v30
+; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v35
+; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
+; VI-NEXT: v_or_b32_sdwa v11, v29, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v52
+; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
+; VI-NEXT: v_or_b32_e32 v29, v29, v30
+; VI-NEXT: v_or_b32_sdwa v12, v29, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; VI-NEXT: v_lshlrev_b32_e32 v35, 24, v56
+; VI-NEXT: v_add_u32_e32 v11, vcc, 0x3000000, v11
+; VI-NEXT: v_add_u32_e32 v12, vcc, 0x3000000, v12
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
+; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
+; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
+; VI-NEXT: v_or_b32_e32 v29, v29, v30
+; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v37
+; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
+; VI-NEXT: v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v41
+; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
+; VI-NEXT: v_or_b32_e32 v29, v29, v30
+; VI-NEXT: v_or_b32_sdwa v14, v29, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v37, vcc, 0x300, v2
+; VI-NEXT: v_add_u32_e32 v13, vcc, 0x3000000, v13
+; VI-NEXT: v_add_u32_e32 v14, vcc, 0x3000000, v14
+; VI-NEXT: v_mov_b32_e32 v2, s4
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
+; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
+; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
+; VI-NEXT: v_or_b32_e32 v29, v29, v30
+; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v62
+; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
+; VI-NEXT: v_or_b32_sdwa v15, v29, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v55
+; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
+; VI-NEXT: v_or_b32_e32 v29, v29, v30
+; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; VI-NEXT: v_or_b32_sdwa v17, v29, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v36
+; VI-NEXT: v_add_u32_e32 v36, vcc, 0x300, v1
+; VI-NEXT: v_add_u32_e32 v15, vcc, 0x3000000, v15
+; VI-NEXT: v_mov_b32_e32 v1, s5
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
+; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
+; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
+; VI-NEXT: v_or_b32_e32 v29, v29, v30
+; VI-NEXT: v_or_b32_sdwa v29, v29, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: v_add_u32_e32 v16, vcc, 0x3000000, v17
+; VI-NEXT: v_add_u32_e32 v17, vcc, 0x3000000, v29
+; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
+; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
+; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
+; VI-NEXT: v_or_b32_e32 v29, v29, v30
+; VI-NEXT: v_or_b32_sdwa v18, v29, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v53
+; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
+; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
+; VI-NEXT: v_add_u32_e32 v18, vcc, 0x3000000, v18
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
+; VI-NEXT: v_or_b32_e32 v29, v29, v30
+; VI-NEXT: v_or_b32_sdwa v19, v29, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v19, vcc, 0x3000000, v19
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
+; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
+; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
+; VI-NEXT: v_or_b32_e32 v29, v29, v30
+; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v57
+; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
+; VI-NEXT: v_or_b32_sdwa v20, v29, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v38
+; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
+; VI-NEXT: v_or_b32_e32 v29, v29, v30
+; VI-NEXT: v_or_b32_sdwa v21, v29, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v38, vcc, 0x300, v22
+; VI-NEXT: v_add_u32_e32 v20, vcc, 0x3000000, v20
+; VI-NEXT: v_add_u32_e32 v21, vcc, 0x3000000, v21
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v29
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_add_u32_e32 v30, vcc, 3, v30
+; VI-NEXT: v_and_b32_e32 v30, 0xff, v30
+; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
+; VI-NEXT: v_or_b32_e32 v29, v29, v30
+; VI-NEXT: v_or_b32_sdwa v55, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: v_lshlrev_b32_e32 v29, 24, v54
+; VI-NEXT: v_lshlrev_b32_e32 v54, 24, v34
+; VI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; VI-NEXT: v_add_u32_e32 v22, vcc, 0x3000000, v55
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_lshlrev_b32_e32 v40, 24, v34
+; VI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_lshlrev_b32_e32 v30, 24, v28
+; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_add_u32_e32 v41, vcc, 3, v34
+; VI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(1)
+; VI-NEXT: v_lshlrev_b32_e32 v28, 24, v28
+; VI-NEXT: v_and_b32_e32 v41, 0xff, v41
+; VI-NEXT: v_lshlrev_b32_e32 v55, 16, v41
+; VI-NEXT: v_or_b32_e32 v55, v40, v55
+; VI-NEXT: v_or_b32_sdwa v23, v55, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: v_add_u32_e32 v23, vcc, 0x3000000, v23
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_add_u32_e32 v42, vcc, 3, v34
+; VI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
+; VI-NEXT: v_and_b32_e32 v42, 0xff, v42
+; VI-NEXT: v_lshlrev_b32_e32 v55, 16, v42
+; VI-NEXT: v_or_b32_e32 v54, v54, v55
+; VI-NEXT: v_or_b32_sdwa v24, v54, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: v_add_u32_e32 v24, vcc, 0x3000000, v24
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_add_u32_e32 v53, vcc, 3, v34
+; VI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
+; VI-NEXT: v_and_b32_e32 v53, 0xff, v53
+; VI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
+; VI-NEXT: v_or_b32_e32 v35, v35, v53
+; VI-NEXT: v_or_b32_sdwa v25, v35, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: v_add_u32_e32 v25, vcc, 0x3000000, v25
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_add_u32_e32 v52, vcc, 3, v34
+; VI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
+; VI-NEXT: v_and_b32_e32 v52, 0xff, v52
+; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v52
+; VI-NEXT: v_or_b32_e32 v28, v28, v32
+; VI-NEXT: v_or_b32_sdwa v28, v28, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: v_add_u32_e32 v28, vcc, 0x3000000, v28
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_add_u32_e32 v51, vcc, 3, v34
+; VI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
+; VI-NEXT: v_and_b32_e32 v51, 0xff, v51
+; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v51
+; VI-NEXT: v_or_b32_e32 v29, v29, v32
+; VI-NEXT: v_or_b32_sdwa v29, v29, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: v_add_u32_e32 v29, vcc, 0x3000000, v29
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: v_add_u32_e32 v49, vcc, 3, v34
+; VI-NEXT: v_and_b32_e32 v49, 0xff, v49
+; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v49
+; VI-NEXT: v_or_b32_e32 v30, v30, v32
+; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v39
+; VI-NEXT: v_add_u32_e32 v34, vcc, 0x300, v0
+; VI-NEXT: v_or_b32_e32 v31, v31, v32
+; VI-NEXT: v_or_b32_sdwa v30, v30, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: v_or_b32_sdwa v31, v31, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; VI-NEXT: v_mov_b32_e32 v0, s6
+; VI-NEXT: v_add_u32_e32 v30, vcc, 0x3000000, v30
+; VI-NEXT: v_add_u32_e32 v31, vcc, 0x3000000, v31
+; VI-NEXT: .LBB97_5: ; %end
+; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: bitcast_v128i8_to_v64i16_scalar:
+; 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:396 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:332
+; GFX9-NEXT: buffer_load_ushort v45, off, s[0:3], s32
+; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:8
+; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:16
+; GFX9-NEXT: buffer_load_ushort v41, off, s[0:3], s32 offset:24
+; GFX9-NEXT: buffer_load_ushort v40, off, s[0:3], s32 offset:32
+; GFX9-NEXT: buffer_load_ushort v55, off, s[0:3], s32 offset:40
+; GFX9-NEXT: buffer_load_ushort v54, off, s[0:3], s32 offset:48
+; GFX9-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:56
+; GFX9-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:64
+; GFX9-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:72
+; GFX9-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:80
+; GFX9-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:88
+; GFX9-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:96
+; GFX9-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:104
+; GFX9-NEXT: buffer_load_ushort v30, 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 v32, off, s[0:3], s32 offset:128
+; GFX9-NEXT: buffer_load_ushort v33, off, s[0:3], s32 offset:136
+; GFX9-NEXT: buffer_load_ushort v34, off, s[0:3], s32 offset:144
+; GFX9-NEXT: buffer_load_ushort v35, off, s[0:3], s32 offset:152
+; GFX9-NEXT: buffer_load_ushort v36, off, s[0:3], s32 offset:160
+; GFX9-NEXT: buffer_load_ushort v37, off, s[0:3], s32 offset:168
+; GFX9-NEXT: buffer_load_ushort v38, off, s[0:3], s32 offset:176
+; GFX9-NEXT: v_lshlrev_b32_e32 v16, 8, v1
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v29
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshlrev_b32_e32 v14, 8, v3
+; GFX9-NEXT: v_lshlrev_b32_e32 v46, 8, v5
+; GFX9-NEXT: v_lshlrev_b32_e32 v22, 8, v7
+; GFX9-NEXT: v_lshlrev_b32_e32 v24, 8, v9
+; GFX9-NEXT: v_lshlrev_b32_e32 v26, 8, v11
+; GFX9-NEXT: v_lshlrev_b32_e32 v20, 8, v13
+; GFX9-NEXT: v_lshlrev_b32_e32 v28, 8, v15
+; GFX9-NEXT: v_lshlrev_b32_e32 v18, 8, v17
+; GFX9-NEXT: v_lshlrev_b32_e32 v17, 8, v25
+; GFX9-NEXT: v_lshlrev_b32_e32 v12, 8, v27
+; GFX9-NEXT: v_lshlrev_b32_e32 v19, 8, v19
+; GFX9-NEXT: v_lshlrev_b32_e32 v21, 8, v21
+; GFX9-NEXT: v_lshlrev_b32_e32 v23, 8, v23
+; GFX9-NEXT: s_waitcnt vmcnt(24)
+; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v43
+; GFX9-NEXT: s_waitcnt vmcnt(23)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v45
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(23)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v44
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(23)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v42
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(23)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v41
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(23)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v40
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(23)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v55
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(23)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v54
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(23)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v53
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(23)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v52
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(23)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v51
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(23)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v50
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(23)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v49
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(23)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v48
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(23)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v39
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(23)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v30
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(23)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v31
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(23)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v32
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(23)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v33
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(23)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v34
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(23)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v35
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(23)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v36
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(23)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v37
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(23)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v38
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_load_ushort v11, off, s[0:3], s32 offset:184
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:192
+; GFX9-NEXT: buffer_load_ushort v15, off, s[0:3], s32 offset:200
+; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:208
+; GFX9-NEXT: buffer_load_ushort v13, off, s[0:3], s32 offset:216
+; GFX9-NEXT: buffer_load_ushort v5, off, s[0:3], s32 offset:224
+; GFX9-NEXT: buffer_load_ushort v9, off, s[0:3], s32 offset:232
+; GFX9-NEXT: buffer_load_ushort v7, off, s[0:3], s32 offset:240
+; GFX9-NEXT: s_and_b64 s[4:5], vcc, exec
+; GFX9-NEXT: s_waitcnt vmcnt(7)
+; GFX9-NEXT: v_lshlrev_b32_e32 v11, 8, v11
+; GFX9-NEXT: s_waitcnt vmcnt(6)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v1
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(6)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v15
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(6)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v3
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(6)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v13
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(6)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v5
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(6)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v9
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(6)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v7
+; GFX9-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_load_ushort v11, off, s[0:3], s32 offset:248
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:256
+; GFX9-NEXT: buffer_load_ushort v15, off, s[0:3], s32 offset:264
+; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:272
+; GFX9-NEXT: buffer_load_ushort v13, off, s[0:3], s32 offset:280
+; GFX9-NEXT: buffer_load_ushort v5, off, s[0:3], s32 offset:288
+; GFX9-NEXT: buffer_load_ushort v9, off, s[0:3], s32 offset:296
+; GFX9-NEXT: buffer_load_ushort v7, off, s[0:3], s32 offset:304
+; GFX9-NEXT: s_waitcnt vmcnt(7)
+; GFX9-NEXT: v_lshlrev_b32_e32 v11, 8, v11
+; GFX9-NEXT: s_waitcnt vmcnt(6)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v1
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(6)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v15
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(6)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v3
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(6)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v13
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(6)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v5
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(6)
+; GFX9-NEXT: v_lshlrev_b32_e32 v1, 8, v9
+; GFX9-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_load_ushort v3, off, s[0:3], s32 offset:312
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:320
+; GFX9-NEXT: buffer_load_ushort v5, off, s[0:3], s32 offset:328
+; GFX9-NEXT: buffer_load_ushort v11, off, s[0:3], s32 offset:4
+; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:12
+; GFX9-NEXT: buffer_load_ushort v55, off, s[0:3], s32 offset:20
+; GFX9-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:28
+; GFX9-NEXT: buffer_load_ushort v13, off, s[0:3], s32 offset:36
+; GFX9-NEXT: s_waitcnt vmcnt(15)
+; GFX9-NEXT: v_lshlrev_b32_e32 v7, 8, v7
+; GFX9-NEXT: s_waitcnt vmcnt(7)
+; GFX9-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; GFX9-NEXT: s_waitcnt vmcnt(6)
+; GFX9-NEXT: v_lshlrev_b32_e32 v9, 8, v1
+; GFX9-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:44
+; GFX9-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:52
+; GFX9-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:60
+; GFX9-NEXT: buffer_load_ushort v15, off, s[0:3], s32 offset:68
+; GFX9-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:76
+; GFX9-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:84
+; GFX9-NEXT: buffer_load_ushort v25, off, s[0:3], s32 offset:92
+; GFX9-NEXT: buffer_load_ushort v34, off, s[0:3], s32 offset:100
+; GFX9-NEXT: buffer_load_ushort v33, off, s[0:3], s32 offset:108
+; GFX9-NEXT: buffer_load_ushort v45, off, s[0:3], s32 offset:116
+; GFX9-NEXT: buffer_load_ushort v43, off, s[0:3], s32 offset:124
+; GFX9-NEXT: buffer_load_ushort v38, off, s[0:3], s32 offset:132
+; GFX9-NEXT: buffer_load_ushort v56, off, s[0:3], s32 offset:140
+; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:148
+; GFX9-NEXT: buffer_load_ushort v61, off, s[0:3], s32 offset:156
+; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:164
+; GFX9-NEXT: s_waitcnt vmcnt(21)
+; GFX9-NEXT: v_lshlrev_b32_e32 v5, 8, v5
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:172
+; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:180
+; GFX9-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:188
+; GFX9-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:196
+; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:204
+; GFX9-NEXT: buffer_load_ushort v35, off, s[0:3], s32 offset:212
+; GFX9-NEXT: buffer_load_ushort v36, off, s[0:3], s32 offset:220
+; GFX9-NEXT: buffer_load_ushort v37, off, s[0:3], s32 offset:228
+; GFX9-NEXT: buffer_load_ushort v41, off, s[0:3], s32 offset:236
+; GFX9-NEXT: buffer_load_ushort v54, off, s[0:3], s32 offset:244
+; GFX9-NEXT: buffer_load_ushort v29, off, s[0:3], s32 offset:252
+; GFX9-NEXT: buffer_load_ushort v27, off, s[0:3], s32 offset:260
+; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:268
+; GFX9-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:276
+; GFX9-NEXT: buffer_load_ushort v59, off, s[0:3], s32 offset:284
+; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:292
+; GFX9-NEXT: buffer_load_ushort v62, off, s[0:3], s32 offset:300
+; GFX9-NEXT: buffer_load_ushort v40, off, s[0:3], s32 offset:308
+; GFX9-NEXT: buffer_load_ushort v58, off, s[0:3], s32 offset:316
+; GFX9-NEXT: buffer_load_ushort v1, off, s[0:3], s32 offset:324
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v38, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(28)
+; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(30)
+; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(33)
+; GFX9-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(36)
+; GFX9-NEXT: buffer_store_dword v37, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v36, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(39)
+; GFX9-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(41)
+; GFX9-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(41)
+; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(41)
+; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(41)
+; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; GFX9-NEXT: s_waitcnt vmcnt(41)
+; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v48, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v50, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v52, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
+; GFX9-NEXT: s_cbranch_scc0 .LBB97_2
+; GFX9-NEXT: ; %bb.1: ; %cmp.false
+; GFX9-NEXT: s_and_b32 s4, s28, 0xff
+; GFX9-NEXT: s_lshl_b32 s5, s29, 8
+; GFX9-NEXT: s_or_b32 s4, s4, s5
+; GFX9-NEXT: v_mov_b32_e32 v1, 0xffff
+; GFX9-NEXT: v_and_b32_e32 v3, s4, v1
+; GFX9-NEXT: v_or_b32_sdwa v1, v2, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX9-NEXT: v_or_b32_sdwa v2, v0, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v4, v4, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v0, v6, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshl_or_b32 v4, v4, 16, v1
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: v_or_b32_sdwa v1, v8, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshl_or_b32 v5, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
+; GFX9-NEXT: v_or_b32_sdwa v0, v10, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: v_lshl_or_b32 v3, v2, 16, v3
+; GFX9-NEXT: s_and_b32 s4, s16, 0xff
+; GFX9-NEXT: s_lshl_b32 s5, s17, 8
+; GFX9-NEXT: s_or_b32 s4, s4, s5
+; GFX9-NEXT: s_and_b32 s5, s18, 0xff
+; GFX9-NEXT: s_lshl_b32 s6, s19, 8
+; GFX9-NEXT: s_or_b32 s5, s5, s6
+; GFX9-NEXT: s_pack_ll_b32_b16 s4, s4, s5
+; GFX9-NEXT: s_and_b32 s5, s20, 0xff
+; GFX9-NEXT: s_lshl_b32 s6, s21, 8
+; GFX9-NEXT: s_or_b32 s5, s5, s6
+; GFX9-NEXT: s_and_b32 s6, s22, 0xff
+; GFX9-NEXT: s_lshl_b32 s7, s23, 8
+; GFX9-NEXT: s_or_b32 s6, s6, s7
+; GFX9-NEXT: s_pack_ll_b32_b16 s5, s5, s6
+; GFX9-NEXT: s_and_b32 s6, s24, 0xff
+; GFX9-NEXT: s_lshl_b32 s7, s25, 8
+; GFX9-NEXT: s_or_b32 s6, s6, s7
+; GFX9-NEXT: s_and_b32 s7, s26, 0xff
+; GFX9-NEXT: s_lshl_b32 s8, s27, 8
+; GFX9-NEXT: s_or_b32 s7, s7, s8
+; GFX9-NEXT: s_pack_ll_b32_b16 s6, s6, s7
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v1, v1, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshl_or_b32 v6, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_or_b32_sdwa v0, v0, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v1, v1, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshl_or_b32 v7, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_or_b32_sdwa v0, v0, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v1, v1, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshl_or_b32 v8, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_or_b32_sdwa v0, v0, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v1, v1, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshl_or_b32 v9, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_or_b32_sdwa v0, v0, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshl_or_b32 v10, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:416 ; 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 v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_or_b32_sdwa v2, v39, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_mov_b32_e32 v39, v16
+; GFX9-NEXT: v_or_b32_sdwa v17, v34, v16 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, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshl_or_b32 v11, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_or_b32_sdwa v0, v42, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v1, v55, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshl_or_b32 v12, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; GFX9-NEXT: v_mov_b32_e32 v42, v61
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_or_b32_sdwa v0, v53, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_mov_b32_e32 v55, v1
+; 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_lshl_or_b32 v13, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_mov_b32_e32 v53, v0
+; GFX9-NEXT: v_or_b32_sdwa v0, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_mov_b32_e32 v52, v1
+; GFX9-NEXT: v_or_b32_sdwa v1, v50, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshl_or_b32 v14, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_mov_b32_e32 v50, v0
+; GFX9-NEXT: v_or_b32_sdwa v0, v49, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_mov_b32_e32 v49, v1
+; 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_lshl_or_b32 v15, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_or_b32_sdwa v0, v48, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: v_lshl_or_b32 v16, v2, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_mov_b32_e32 v48, v1
+; 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: v_and_b32_e32 v1, 0xffff, v1
+; GFX9-NEXT: v_lshl_or_b32 v17, v17, 16, v1
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_or_b32_sdwa v0, v33, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: v_mov_b32_e32 v33, v45
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v1, v45, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshl_or_b32 v18, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v1, v38, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshl_or_b32 v19, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_or_b32_sdwa v0, v56, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v1, v47, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshl_or_b32 v20, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_or_b32_sdwa v0, v42, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v1, v61, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshl_or_b32 v21, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; 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: v_lshl_or_b32 v22, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_or_b32_sdwa v0, v51, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_mov_b32_e32 v34, 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: v_lshl_or_b32 v23, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(4)
+; GFX9-NEXT: v_mov_b32_e32 v46, v0
+; GFX9-NEXT: v_or_b32_sdwa v0, v44, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(3)
+; GFX9-NEXT: v_or_b32_sdwa v1, v35, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshl_or_b32 v24, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; GFX9-NEXT: v_mov_b32_e32 v35, v45
+; GFX9-NEXT: v_mov_b32_e32 v45, v61
+; GFX9-NEXT: v_mov_b32_e32 v61, v42
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_mov_b32_e32 v38, v0
+; GFX9-NEXT: v_or_b32_sdwa v0, v36, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v1, v37, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; GFX9-NEXT: v_lshl_or_b32 v25, v1, 16, v0
+; GFX9-NEXT: v_or_b32_sdwa v1, v54, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_mov_b32_e32 v54, v2
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v0, v41, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: v_lshl_or_b32 v26, v1, 16, v0
+; GFX9-NEXT: v_or_b32_sdwa v1, v27, v2 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 v0, v29, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: v_lshl_or_b32 v27, v1, 16, v0
+; GFX9-NEXT: v_or_b32_sdwa v0, v60, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; GFX9-NEXT: v_or_b32_sdwa v1, v57, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: v_lshl_or_b32 v28, v1, 16, v0
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_or_b32_sdwa v0, v59, v60 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v1, v63, v57 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshl_or_b32 v29, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_or_b32_sdwa v0, v62, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v1, v40, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshl_or_b32 v30, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_or_b32_sdwa v0, v58, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshl_or_b32 v31, v1, 16, v0
+; GFX9-NEXT: v_mov_b32_e32 v0, s4
+; GFX9-NEXT: v_mov_b32_e32 v1, s5
+; GFX9-NEXT: v_mov_b32_e32 v2, s6
+; GFX9-NEXT: s_mov_b64 s[4:5], 0
+; GFX9-NEXT: s_branch .LBB97_3
+; GFX9-NEXT: .LBB97_2:
+; GFX9-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; GFX9-NEXT: v_mov_b32_e32 v33, v45
+; GFX9-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; GFX9-NEXT: s_mov_b64 s[4:5], -1
+; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
+; GFX9-NEXT: ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
+; GFX9-NEXT: .LBB97_3: ; %Flow
+; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
+; GFX9-NEXT: s_andn2_b64 vcc, exec, s[4:5]
+; GFX9-NEXT: s_cbranch_vccnz .LBB97_5
+; GFX9-NEXT: ; %bb.4: ; %cmp.true
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; GFX9-NEXT: s_add_i32 s28, s28, 3
+; GFX9-NEXT: s_and_b32 s4, s28, 0xff
+; GFX9-NEXT: s_lshl_b32 s5, s29, 8
+; GFX9-NEXT: s_or_b32 s4, s5, s4
+; GFX9-NEXT: s_addk_i32 s4, 0x300
+; GFX9-NEXT: s_add_i32 s24, s24, 3
+; GFX9-NEXT: s_lshl_b32 s5, s25, 8
+; GFX9-NEXT: s_add_i32 s26, s26, 3
+; GFX9-NEXT: s_lshl_b32 s6, s27, 8
+; GFX9-NEXT: s_add_i32 s20, s20, 3
+; GFX9-NEXT: s_lshl_b32 s7, s21, 8
+; GFX9-NEXT: s_add_i32 s22, s22, 3
+; GFX9-NEXT: s_lshl_b32 s8, s23, 8
+; GFX9-NEXT: s_add_i32 s16, s16, 3
+; GFX9-NEXT: s_lshl_b32 s9, s17, 8
+; GFX9-NEXT: s_add_i32 s18, s18, 3
+; GFX9-NEXT: s_lshl_b32 s10, s19, 8
+; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(15)
+; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
+; GFX9-NEXT: s_waitcnt vmcnt(13)
+; GFX9-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
+; GFX9-NEXT: v_add_u32_e32 v0, 0x300, v0
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(12)
+; GFX9-NEXT: v_add_u32_e32 v25, 3, v25
+; GFX9-NEXT: s_waitcnt vmcnt(11)
+; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
+; GFX9-NEXT: v_or_b32_sdwa v25, v37, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: v_or_b32_sdwa v37, v51, v26 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:708 ; 4-byte Folded Reload
+; GFX9-NEXT: v_add_u32_e32 v37, 0x300, v37
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
+; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
+; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
+; GFX9-NEXT: v_lshl_or_b32 v4, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(3)
+; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
+; GFX9-NEXT: v_or_b32_sdwa v38, v38, v26 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:704 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(3)
+; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
+; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: v_mov_b32_e32 v3, 0xffff
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
+; GFX9-NEXT: v_add_u32_e32 v2, 0x300, v2
+; GFX9-NEXT: v_and_b32_e32 v3, s4, v3
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
+; GFX9-NEXT: v_lshl_or_b32 v3, v2, 16, v3
+; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; GFX9-NEXT: v_add_u32_e32 v0, 0x300, v0
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_and_b32 s4, s24, 0xff
+; GFX9-NEXT: s_or_b32 s4, s5, s4
+; GFX9-NEXT: s_and_b32 s5, s26, 0xff
+; GFX9-NEXT: s_or_b32 s5, s6, s5
+; GFX9-NEXT: s_and_b32 s6, s20, 0xff
+; GFX9-NEXT: s_or_b32 s6, s7, s6
+; GFX9-NEXT: s_and_b32 s7, s22, 0xff
+; GFX9-NEXT: s_or_b32 s7, s8, s7
+; GFX9-NEXT: s_and_b32 s8, s16, 0xff
+; GFX9-NEXT: s_or_b32 s8, s9, s8
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
+; GFX9-NEXT: s_and_b32 s9, s18, 0xff
+; GFX9-NEXT: s_or_b32 s9, s10, s9
+; GFX9-NEXT: s_addk_i32 s4, 0x300
+; GFX9-NEXT: s_addk_i32 s5, 0x300
+; GFX9-NEXT: s_addk_i32 s6, 0x300
+; GFX9-NEXT: s_addk_i32 s7, 0x300
+; GFX9-NEXT: s_addk_i32 s8, 0x300
+; GFX9-NEXT: s_addk_i32 s9, 0x300
+; GFX9-NEXT: s_pack_ll_b32_b16 s8, s8, s9
+; GFX9-NEXT: s_pack_ll_b32_b16 s6, s6, s7
+; GFX9-NEXT: s_pack_ll_b32_b16 s4, s4, s5
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
+; GFX9-NEXT: v_lshl_or_b32 v5, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
+; GFX9-NEXT: v_add_u32_e32 v0, 0x300, v0
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
+; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
+; GFX9-NEXT: v_lshl_or_b32 v6, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
+; GFX9-NEXT: v_add_u32_e32 v0, 0x300, v0
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
+; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
+; GFX9-NEXT: v_lshl_or_b32 v7, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
+; GFX9-NEXT: v_add_u32_e32 v0, 0x300, v0
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
+; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
+; GFX9-NEXT: v_lshl_or_b32 v8, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
+; GFX9-NEXT: v_add_u32_e32 v0, 0x300, v0
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
+; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
+; GFX9-NEXT: v_lshl_or_b32 v9, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
+; GFX9-NEXT: v_add_u32_e32 v0, 0x300, v0
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
+; GFX9-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
+; GFX9-NEXT: v_lshl_or_b32 v10, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; GFX9-NEXT: v_add_u32_e32 v0, 0x300, v0
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
+; GFX9-NEXT: v_or_b32_sdwa v1, v58, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
+; GFX9-NEXT: v_lshl_or_b32 v11, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_add_u32_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: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
+; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: v_add_u32_e32 v0, 0x300, v0
+; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: v_lshl_or_b32 v12, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
+; GFX9-NEXT: v_or_b32_sdwa v0, v42, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
+; GFX9-NEXT: v_or_b32_sdwa v1, v55, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: v_add_u32_e32 v0, 0x300, v0
+; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: v_lshl_or_b32 v13, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_add_u32_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: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
+; GFX9-NEXT: v_or_b32_sdwa v1, v52, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: v_add_u32_e32 v0, 0x300, v0
+; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: v_lshl_or_b32 v14, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_add_u32_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: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
+; GFX9-NEXT: v_or_b32_sdwa v1, v49, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: v_add_u32_e32 v0, 0x300, v0
+; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: v_lshl_or_b32 v15, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
+; GFX9-NEXT: v_or_b32_sdwa v1, v39, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: v_or_b32_sdwa v39, v36, v26 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:692 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
+; 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_add_u32_e32 v0, 0x300, v0
+; GFX9-NEXT: v_add_u32_e32 v1, 0x300, v1
+; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
+; GFX9-NEXT: v_or_b32_sdwa v48, v46, v26 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:688 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
+; GFX9-NEXT: v_or_b32_sdwa v49, v35, v26 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:676 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
+; GFX9-NEXT: v_or_b32_sdwa v50, v27, v26 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:672 ; 4-byte Folded Reload
+; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
+; 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: buffer_load_dword v16, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; GFX9-NEXT: v_add_u32_e32 v2, 0x300, v2
+; GFX9-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
+; GFX9-NEXT: v_or_b32_sdwa v51, v34, v26 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:660 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_add_u32_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_lshl_or_b32 v17, v1, 16, v0
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; GFX9-NEXT: v_add_u32_e32 v16, 0x300, v16
+; GFX9-NEXT: v_lshl_or_b32 v16, v16, 16, v2
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
+; GFX9-NEXT: v_or_b32_sdwa v52, v27, v26 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:656 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v53, 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:480 ; 4-byte Folded Reload
+; GFX9-NEXT: v_add_u32_e32 v24, 3, v24
+; GFX9-NEXT: v_add_u32_e32 v26, 3, v61
+; GFX9-NEXT: v_or_b32_sdwa v24, v54, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: v_add_u32_e32 v36, 0x300, v24
+; GFX9-NEXT: v_add_u32_e32 v24, 0x300, v48
+; GFX9-NEXT: v_add_u32_e32 v48, 0x300, v51
+; GFX9-NEXT: v_and_b32_e32 v24, 0xffff, v24
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v54, 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:476 ; 4-byte Folded Reload
+; GFX9-NEXT: v_add_u32_e32 v26, 3, v45
+; GFX9-NEXT: v_add_u32_e32 v20, 3, v20
+; GFX9-NEXT: v_or_b32_sdwa v20, v57, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: v_add_u32_e32 v34, 0x300, v20
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v55, 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:472 ; 4-byte Folded Reload
+; GFX9-NEXT: v_add_u32_e32 v26, 3, v56
+; GFX9-NEXT: v_add_u32_e32 v21, 3, v21
+; GFX9-NEXT: v_or_b32_sdwa v21, v32, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: v_add_u32_e32 v28, 0x300, v21
+; GFX9-NEXT: v_add_u32_e32 v21, 0x300, v54
+; GFX9-NEXT: v_and_b32_e32 v21, 0xffff, v21
+; GFX9-NEXT: v_and_b32_e32 v28, 0xffff, v28
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v40, 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:468 ; 4-byte Folded Reload
+; GFX9-NEXT: v_add_u32_e32 v23, 3, v23
+; GFX9-NEXT: v_add_u32_e32 v26, 3, v47
+; GFX9-NEXT: v_or_b32_sdwa v23, v41, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: v_add_u32_e32 v20, 0x300, v40
+; GFX9-NEXT: v_and_b32_e32 v20, 0xffff, v20
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v41, 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:464 ; 4-byte Folded Reload
+; GFX9-NEXT: v_add_u32_e32 v26, 3, v43
+; GFX9-NEXT: v_add_u32_e32 v22, 3, v22
+; GFX9-NEXT: v_or_b32_sdwa v22, v44, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: v_add_u32_e32 v35, 0x300, v22
+; GFX9-NEXT: v_add_u32_e32 v22, 0x300, v52
+; GFX9-NEXT: v_add_u32_e32 v51, 0x300, v41
+; GFX9-NEXT: v_and_b32_e32 v22, 0xffff, v22
+; GFX9-NEXT: v_lshl_or_b32 v20, v51, 16, v20
+; GFX9-NEXT: v_lshl_or_b32 v28, v35, 16, v28
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v42, v27, v26 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:636 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
+; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v43, v27, v26 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:624 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
+; GFX9-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; GFX9-NEXT: v_add_u32_e32 v31, 0x300, v0
+; GFX9-NEXT: v_add_u32_e32 v52, 0x300, v43
+; GFX9-NEXT: v_and_b32_e32 v31, 0xffff, v31
+; GFX9-NEXT: v_mov_b32_e32 v0, s8
+; GFX9-NEXT: s_waitcnt vmcnt(3)
+; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_or_b32_sdwa v44, 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:444 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; GFX9-NEXT: v_add_u32_e32 v26, 3, v33
+; GFX9-NEXT: v_add_u32_e32 v32, 0x300, v1
+; GFX9-NEXT: v_mov_b32_e32 v1, s6
+; GFX9-NEXT: v_lshl_or_b32 v31, v32, 16, v31
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_or_b32_sdwa v45, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: v_add_u32_e32 v27, 0x300, v23
+; GFX9-NEXT: v_add_u32_e32 v26, 0x300, v25
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
+; GFX9-NEXT: v_or_b32_sdwa v2, v18, v2 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:760 ; 4-byte Folded Reload
+; GFX9-NEXT: v_add_u32_e32 v30, 0x300, v2
+; GFX9-NEXT: v_add_u32_e32 v25, 0x300, v38
+; GFX9-NEXT: v_add_u32_e32 v23, 0x300, v50
+; GFX9-NEXT: v_add_u32_e32 v38, 0x300, v39
+; GFX9-NEXT: v_add_u32_e32 v39, 0x300, v49
+; GFX9-NEXT: v_add_u32_e32 v49, 0x300, v53
+; GFX9-NEXT: v_add_u32_e32 v50, 0x300, v55
+; GFX9-NEXT: v_add_u32_e32 v53, 0x300, v45
+; GFX9-NEXT: v_and_b32_e32 v23, 0xffff, v23
+; GFX9-NEXT: v_and_b32_e32 v25, 0xffff, v25
+; GFX9-NEXT: v_and_b32_e32 v26, 0xffff, v26
+; GFX9-NEXT: v_and_b32_e32 v27, 0xffff, v27
+; GFX9-NEXT: v_and_b32_e32 v30, 0xffff, v30
+; GFX9-NEXT: v_mov_b32_e32 v2, s4
+; GFX9-NEXT: v_lshl_or_b32 v21, v50, 16, v21
+; GFX9-NEXT: v_lshl_or_b32 v22, v49, 16, v22
+; GFX9-NEXT: v_lshl_or_b32 v23, v48, 16, v23
+; GFX9-NEXT: v_lshl_or_b32 v24, v39, 16, v24
+; GFX9-NEXT: v_lshl_or_b32 v25, v38, 16, v25
+; GFX9-NEXT: v_lshl_or_b32 v26, v37, 16, v26
+; GFX9-NEXT: v_lshl_or_b32 v27, v36, 16, v27
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_add_u32_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:756 ; 4-byte Folded Reload
+; GFX9-NEXT: v_add_u32_e32 v33, 0x300, v18
+; GFX9-NEXT: v_add_u32_e32 v18, 0x300, v44
+; GFX9-NEXT: v_and_b32_e32 v18, 0xffff, v18
+; GFX9-NEXT: v_lshl_or_b32 v18, v53, 16, v18
+; GFX9-NEXT: v_lshl_or_b32 v30, v33, 16, v30
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_add_u32_e32 v19, 3, v19
+; GFX9-NEXT: v_or_b32_sdwa v19, v60, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT: v_add_u32_e32 v29, 0x300, v19
+; GFX9-NEXT: v_add_u32_e32 v19, 0x300, v42
+; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v19
+; GFX9-NEXT: v_and_b32_e32 v29, 0xffff, v29
+; GFX9-NEXT: v_lshl_or_b32 v19, v52, 16, v19
+; GFX9-NEXT: v_lshl_or_b32 v29, v34, 16, v29
+; GFX9-NEXT: .LBB97_5: ; %end
+; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-TRUE16-LABEL: bitcast_v128i8_to_v64i16_scalar:
+; GFX11-TRUE16: ; %bb.0:
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT: s_clause 0x1d
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v40, s32 offset:436
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v41, s32 offset:432
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v42, s32 offset:428
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v43, s32 offset:424
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v44, s32 offset:420
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v45, s32 offset:416
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v46, s32 offset:412
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v47, s32 offset:408
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v56, s32 offset:404
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v57, s32 offset:400
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v58, s32 offset:396
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v59, s32 offset:392
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v60, s32 offset:388
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v61, s32 offset:384
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v62, s32 offset:380
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v63, s32 offset:376
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v72, s32 offset:372
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v73, s32 offset:368
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v74, s32 offset:364
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v75, s32 offset:360
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v76, s32 offset:356
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v77, s32 offset:352
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v78, s32 offset:348
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v79, s32 offset:344
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v88, s32 offset:340
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v89, s32 offset:336
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v90, s32 offset:332
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v91, s32 offset:328
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v92, s32 offset:324
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v93, s32 offset:320
+; GFX11-TRUE16-NEXT: s_clause 0x1f
+; GFX11-TRUE16-NEXT: scratch_load_b32 v31, off, s32 offset:316
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v89, off, s32 offset:312
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v92, off, s32 offset:308
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v91, off, s32 offset:304
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v93, off, s32 offset:300
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v78, off, s32 offset:296
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v79, off, s32 offset:292
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v88, off, s32 offset:288
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v90, off, s32 offset:284
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v72, off, s32 offset:280
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v75, off, s32 offset:276
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v76, off, s32 offset:272
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v77, off, s32 offset:268
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v60, off, s32 offset:264
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v63, off, s32 offset:260
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v73, off, s32 offset:256
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v74, off, s32 offset:252
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v57, off, s32 offset:248
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v59, off, s32 offset:244
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v61, off, s32 offset:240
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v62, off, s32 offset:236
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v45, off, s32 offset:232
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v56, off, s32 offset:228
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v47, off, s32 offset:224
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v58, off, s32 offset:220
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v41, off, s32 offset:216
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v44, off, s32 offset:212
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v43, off, s32 offset:208
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v46, off, s32 offset:204
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v180, off, s32 offset:200
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v183, off, s32 offset:196
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v40, off, s32 offset:192
+; GFX11-TRUE16-NEXT: s_clause 0x1f
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v42, off, s32 offset:188
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v176, off, s32 offset:184
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v179, off, s32 offset:180
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v181, off, s32 offset:176
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v182, off, s32 offset:172
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v165, off, s32 offset:168
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v167, off, s32 offset:164
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v177, off, s32 offset:160
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v178, off, s32 offset:156
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v161, off, s32 offset:152
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v164, off, s32 offset:148
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v163, off, s32 offset:144
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v166, off, s32 offset:140
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v149, off, s32 offset:136
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v160, off, s32 offset:132
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v151, off, s32 offset:128
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v162, off, s32 offset:124
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v134, off, s32 offset:120
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v147, off, s32 offset:116
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v148, off, s32 offset:112
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v150, off, s32 offset:108
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v119, off, s32 offset:104
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v132, off, s32 offset:100
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v144, off, s32 offset:96
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v145, off, s32 offset:92
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v114, off, s32 offset:88
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v117, off, s32 offset:84
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v129, off, s32 offset:80
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v130, off, s32 offset:76
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v102, off, s32 offset:72
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v135, off, s32 offset:68
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v133, off, s32 offset:64
+; GFX11-TRUE16-NEXT: s_clause 0xf
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v146, off, s32 offset:60
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v115, off, s32 offset:56
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v128, off, s32 offset:52
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v118, off, s32 offset:48
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v131, off, s32 offset:44
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:40
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v112, off, s32 offset:36
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v113, off, s32 offset:32
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v116, off, s32 offset:28
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v96, off, s32 offset:24
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v99, off, s32 offset:20
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v101, off, s32 offset:16
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v103, off, s32 offset:12
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:8
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v87, off, s32 offset:4
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v98.l, v30.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v81.l, v29.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v84.l, v28.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v83.l, v27.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v86.l, v26.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v69.l, v25.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v80.l, v24.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v71.l, v23.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v82.l, v22.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v21.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v20.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v68.l, v19.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v70.l, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v17.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v16.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v15.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v14.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v13.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v11.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v9.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v7.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v5.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v1.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v0.l
+; GFX11-TRUE16-NEXT: s_mov_b32 s4, 0
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(62)
+; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v31
+; GFX11-TRUE16-NEXT: s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB97_4
+; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v39
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v37
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v33
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v32
+; GFX11-TRUE16-NEXT: s_and_b32 s5, s28, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s6, s29, 8
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xff, v35
+; GFX11-TRUE16-NEXT: s_or_b32 s5, s5, s6
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 8, v34
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v2, v3
+; GFX11-TRUE16-NEXT: v_and_b32_e64 v2, 0xffff, s5
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xff, v50
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v4, v5
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 8, v38
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v4, v1, 16, v2
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v54
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v53
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v5, v3, 16, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v6, v7
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v48
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 8, v36
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v51
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 8, v49
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xff, v66
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 8, v65
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v3, v6
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v2, v7
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v8, v9
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xff, v55
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 8, v52
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xff, v70
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 8, v68
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v6, v3, 16, v0
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v7, v2, 16, v1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v9, v10
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v8
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v11, v12
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v82
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 8, v71
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xff, v67
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 8, v64
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xff, v80
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 8, v69
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v3, v8
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v8, v0, 16, v1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v9, v10
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v2
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v11, v12
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xff, v86
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 8, v83
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v9, v0, 16, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v98
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v97
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v10, v2, 16, v3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v11, v12
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v84
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 8, v81
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v87
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 8, v85
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xff, v103
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 8, v101
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v3, v11
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v12
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v13, v14
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xff, v99
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 8, v96
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xff, v116
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 8, v113
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v11, v3, 16, v2
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v12, v1, 16, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v14, v15
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v13
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v16, v17
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v112
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 8, v100
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xff, v131
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 8, v118
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xff, v128
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 8, v115
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xff, v146
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 8, v133
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v3, v13
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v14, v15
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v16, v17
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xff, v130
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 8, v129
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v18, v19
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xff, v135
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 8, v102
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v16, v17
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v13
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v13, v0, 16, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v145
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v144
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v18, v19
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v14
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v14, v3, 16, v2
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v119
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v132
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xff, v117
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 8, v114
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v15, v15, 16, v17
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v134
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v20, v21
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v16
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v16, v18, 16, v19
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v18, v1, 16, v0
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v150
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v148
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v17, v20, 16, v21
+; GFX11-TRUE16-NEXT: s_and_b32 s5, s0, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s6, s1, 8
+; GFX11-TRUE16-NEXT: s_and_b32 s7, s2, 0xff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v147
+; GFX11-TRUE16-NEXT: s_lshl_b32 s8, s3, 8
+; GFX11-TRUE16-NEXT: s_or_b32 s5, s5, s6
+; GFX11-TRUE16-NEXT: s_or_b32 s6, s7, s8
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v149
+; GFX11-TRUE16-NEXT: s_and_b32 s7, s16, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s8, s17, 8
+; GFX11-TRUE16-NEXT: s_and_b32 s9, s18, 0xff
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v19, v1, 16, v0
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v162
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v151
+; GFX11-TRUE16-NEXT: s_lshl_b32 s10, s19, 8
+; GFX11-TRUE16-NEXT: s_or_b32 s7, s7, s8
+; GFX11-TRUE16-NEXT: s_or_b32 s8, s9, s10
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s5, s6
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v160
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s7, s8
+; GFX11-TRUE16-NEXT: s_and_b32 s7, s20, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s8, s21, 8
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v161
+; GFX11-TRUE16-NEXT: s_and_b32 s9, s22, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s10, s23, 8
+; GFX11-TRUE16-NEXT: s_or_b32 s7, s7, s8
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v20, v1, 16, v0
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v166
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v163
+; GFX11-TRUE16-NEXT: s_or_b32 s8, s9, s10
+; GFX11-TRUE16-NEXT: s_and_b32 s9, s24, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s10, s25, 8
+; GFX11-TRUE16-NEXT: s_and_b32 s11, s26, 0xff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v164
+; GFX11-TRUE16-NEXT: s_lshl_b32 s12, s27, 8
+; GFX11-TRUE16-NEXT: s_or_b32 s9, s9, s10
+; GFX11-TRUE16-NEXT: s_or_b32 s10, s11, s12
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v165
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s7, s8
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s9, s10
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v21, v1, 16, v0
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v178
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v177
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, s8
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v167
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v176
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v22, v1, 16, v0
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v182
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v181
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v179
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v180
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v23, v1, 16, v0
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v42
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v40
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v183
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v41
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v24, v1, 16, v0
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v46
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v43
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v44
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v45
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v25, v1, 16, v0
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v58
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v47
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v56
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v57
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v26, v1, 16, v0
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v62
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v61
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v59
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v60
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v27, v1, 16, v0
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v74
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v73
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v63
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v72
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v28, v1, 16, v0
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v77
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v76
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v75
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v78
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v29, v1, 16, v0
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v90
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v88
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v79
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v89
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v30, v1, 16, v0
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v93
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 8, v91
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v92
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v2
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, s7
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v31, v1, 16, v0
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6
+; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB97_3
+; GFX11-TRUE16-NEXT: .LBB97_2: ; %cmp.true
+; GFX11-TRUE16-NEXT: s_add_i32 s28, s28, 3
+; GFX11-TRUE16-NEXT: s_lshl_b32 s5, s29, 8
+; GFX11-TRUE16-NEXT: s_and_b32 s4, s28, 0xff
+; GFX11-TRUE16-NEXT: s_add_i32 s24, s24, 3
+; GFX11-TRUE16-NEXT: s_or_b32 s4, s5, s4
+; GFX11-TRUE16-NEXT: s_and_b32 s5, s24, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s6, s25, 8
+; GFX11-TRUE16-NEXT: s_add_i32 s26, s26, 3
+; GFX11-TRUE16-NEXT: s_or_b32 s5, s6, s5
+; GFX11-TRUE16-NEXT: s_and_b32 s6, s26, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s7, s27, 8
+; GFX11-TRUE16-NEXT: s_add_i32 s20, s20, 3
+; GFX11-TRUE16-NEXT: s_or_b32 s6, s7, s6
+; GFX11-TRUE16-NEXT: s_and_b32 s7, s20, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s8, s21, 8
+; GFX11-TRUE16-NEXT: s_add_i32 s22, s22, 3
+; GFX11-TRUE16-NEXT: s_or_b32 s7, s8, s7
+; GFX11-TRUE16-NEXT: s_and_b32 s8, s22, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s9, s23, 8
+; GFX11-TRUE16-NEXT: s_add_i32 s16, s16, 3
+; GFX11-TRUE16-NEXT: s_or_b32 s8, s9, s8
+; GFX11-TRUE16-NEXT: s_and_b32 s9, s16, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s10, s17, 8
+; GFX11-TRUE16-NEXT: s_add_i32 s18, s18, 3
+; GFX11-TRUE16-NEXT: s_add_i32 s0, s0, 3
+; GFX11-TRUE16-NEXT: s_add_i32 s2, s2, 3
+; GFX11-TRUE16-NEXT: s_or_b32 s9, s10, s9
+; GFX11-TRUE16-NEXT: s_and_b32 s10, s18, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s11, s19, 8
+; GFX11-TRUE16-NEXT: s_and_b32 s0, s0, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s1, s1, 8
+; GFX11-TRUE16-NEXT: s_and_b32 s2, s2, 0xff
+; GFX11-TRUE16-NEXT: s_lshl_b32 s3, s3, 8
+; GFX11-TRUE16-NEXT: s_or_b32 s10, s11, s10
+; GFX11-TRUE16-NEXT: s_or_b32 s0, s1, s0
+; GFX11-TRUE16-NEXT: s_or_b32 s1, s3, s2
+; GFX11-TRUE16-NEXT: s_addk_i32 s9, 0x300
+; GFX11-TRUE16-NEXT: s_addk_i32 s0, 0x300
+; GFX11-TRUE16-NEXT: s_addk_i32 s1, 0x300
+; GFX11-TRUE16-NEXT: s_addk_i32 s10, 0x300
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s1
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 3, v92
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s9, s10
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v0, 3, v93
+; GFX11-TRUE16-NEXT: s_addk_i32 s5, 0x300
+; GFX11-TRUE16-NEXT: s_addk_i32 s6, 0x300
+; GFX11-TRUE16-NEXT: s_addk_i32 s7, 0x300
+; GFX11-TRUE16-NEXT: s_addk_i32 s8, 0x300
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v90
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v79
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s7, s8
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v89
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s5, s6
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 8, v91
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 8, v88
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 8, v78
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v6
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v2, v0
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v77
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v31, 0x300, v1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v27, 0x300, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v2
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v4
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v76
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v75
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v30, 0x300, v1
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v28, 0x300, v0
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v0, 3, v74
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v2
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v4
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v72
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v73
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v63
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v29, 0x300, v1
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(59)
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 3, v62
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v0
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v5
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v60
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 8, v61
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v59
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v59, 0x300, v2
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v60, 0x300, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v6
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v57
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(55)
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v58
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v57, 0x300, v0
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v0, 3, v56
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v56, 0x300, v1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v2
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v4
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v47
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v45
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(51)
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v46
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v45, 0x300, v1
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 3, v44
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v0
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v5
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v43
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 8, v41
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(47)
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v42
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v22, 0x300, v2
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v26, 0x300, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v6
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v40
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v183
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v23, 0x300, v0
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(43)
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v0, 3, v182
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v25, 0x300, v1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v2
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v4
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v180
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v181
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v179
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v24, 0x300, v1
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(39)
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 3, v178
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v0
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v5
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v176
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 8, v177
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v167
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v167, 0x300, v2
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v176, 0x300, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v6
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v165
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(35)
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v166
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v165, 0x300, v0
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v0, 3, v164
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v164, 0x300, v1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v2
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v4
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v163
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v161
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(31)
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v162
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v161, 0x300, v1
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 3, v160
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v0
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v5
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v151
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 8, v149
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(27)
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v150
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v17, 0x300, v2
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v21, 0x300, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v6
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v148
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v147
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v18, 0x300, v0
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(23)
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v0, 3, v145
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v20, 0x300, v1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v2
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v4
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v134
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v144
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v132
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v19, 0x300, v1
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(19)
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 3, v130
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v0
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v5
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v119
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 8, v129
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v117
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v117, 0x300, v2
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v119, 0x300, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v6
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v114
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(15)
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v146
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v114, 0x300, v0
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v0, 3, v135
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v129, 0x300, v1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v2
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v4
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 8, v133
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v102
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(11)
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v131
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v102, 0x300, v1
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 3, v128
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v0
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v5
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v118
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 8, v115
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(7)
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v116
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v16, 0x300, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 8, v113
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v112
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(3)
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v103
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 8, v101
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v8, 3, v99
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v4, v3
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xff, v5
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 8, v100
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xff, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v9, 3, v98
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(1)
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v10, 3, v87
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v55, 3, v55
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v7, v6
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xff, v8
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 8, v96
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xff, v9
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 8, v97
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v14, 0x300, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v11, 0x300, v5
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v6
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xff, v10
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v9, v8
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 8, v85
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v8, 3, v86
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v13, 0x300, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v84
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v12, 0x300, v5
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v7, v6
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xff, v8
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 8, v83
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xff, v4
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 8, v81
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v9, 3, v82
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v15, 0x300, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v80
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v7, v6
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v8, v4
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xff, v9
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 8, v71
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xff, v5
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 8, v69
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v10, 3, v70
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v70, 0x300, v4
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v8, v7
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v69, 0x300, v6
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v9, v5
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xff, v10
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 8, v68
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v8, 3, v67
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v9, 0x300, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v66
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v10, 0x300, v5
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v7, v6
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xff, v8
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 8, v64
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xff, v4
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 8, v65
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v64, 0x300, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v54
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v7, v6
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xff, v55
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v8, v4
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 8, v52
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xff, v5
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 8, v53
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v51, 3, v51
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v53, 0x300, v6
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v8, v7
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v7, 0x300, v4
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v52, v5
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xff, v51
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 8, v49
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v49, 0x300, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v50
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v50, 0x300, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v48
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v8, 3, v39
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xff, v6
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 8, v38
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xff, v4
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v36, 8, v36
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xff, v8
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v37, 8, v37
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v35, 3, v35
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v39, 0x300, v5
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v38, v6
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v36, v4
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v37, v8
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xff, v35
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v34, 8, v34
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v35, 0x300, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v33
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v33, 0x300, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 0x300, v6
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v34, v8
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 8, v32
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xff, v5
+; GFX11-TRUE16-NEXT: s_addk_i32 s4, 0x300
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v5
+; GFX11-TRUE16-NEXT: v_and_b32_e64 v8, 0xffff, s4
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v64
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v0, 0x300, v0
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 0x300, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v4, v5, 16, v8
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v5, v6, 16, v32
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v50
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff, v7
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v35
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v9
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v9, v53, 16, v34
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v7, v39, 16, v8
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v8, v49, 16, v32
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff, v69
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v6, v33, 16, v6
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v11
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v10, v10, 16, v35
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v18
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v11, v70, 16, v32
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff, v2
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v13, v13, 16, v33
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v119
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v17
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v20, v20, 16, v34
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v16, v16, 16, v32
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff, v129
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v18, v114, 16, v33
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v21, v21, 16, v35
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v176
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v23
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v17, v102, 16, v32
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff, v164
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v22
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v19
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v24
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v23, v165, 16, v33
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v22, v161, 16, v32
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v25, v25, 16, v34
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v26, v26, 16, v35
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff, v56
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v60
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v29
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v28
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v27
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v12, v15, 16, v12
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v14, v14, 16, v3
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, s3
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v15, v1, 16, v0
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, s2
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v19, v117, 16, v19
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v24, v167, 16, v24
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v27, v45, 16, v32
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v28, v57, 16, v33
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v29, v59, 16, v29
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v30, v30, 16, v34
+; GFX11-TRUE16-NEXT: v_lshl_or_b32 v31, v31, 16, v35
+; GFX11-TRUE16-NEXT: .LBB97_3: ; %end
+; GFX11-TRUE16-NEXT: s_clause 0x1d
+; GFX11-TRUE16-NEXT: scratch_load_b32 v93, off, s32 offset:320
+; GFX11-TRUE16-NEXT: scratch_load_b32 v92, off, s32 offset:324
+; GFX11-TRUE16-NEXT: scratch_load_b32 v91, off, s32 offset:328
+; GFX11-TRUE16-NEXT: scratch_load_b32 v90, off, s32 offset:332
+; GFX11-TRUE16-NEXT: scratch_load_b32 v89, off, s32 offset:336
+; GFX11-TRUE16-NEXT: scratch_load_b32 v88, off, s32 offset:340
+; GFX11-TRUE16-NEXT: scratch_load_b32 v79, off, s32 offset:344
+; GFX11-TRUE16-NEXT: scratch_load_b32 v78, off, s32 offset:348
+; GFX11-TRUE16-NEXT: scratch_load_b32 v77, off, s32 offset:352
+; GFX11-TRUE16-NEXT: scratch_load_b32 v76, off, s32 offset:356
+; GFX11-TRUE16-NEXT: scratch_load_b32 v75, off, s32 offset:360
+; GFX11-TRUE16-NEXT: scratch_load_b32 v74, off, s32 offset:364
+; GFX11-TRUE16-NEXT: scratch_load_b32 v73, off, s32 offset:368
+; GFX11-TRUE16-NEXT: scratch_load_b32 v72, off, s32 offset:372
+; GFX11-TRUE16-NEXT: scratch_load_b32 v63, off, s32 offset:376
+; GFX11-TRUE16-NEXT: scratch_load_b32 v62, off, s32 offset:380
+; GFX11-TRUE16-NEXT: scratch_load_b32 v61, off, s32 offset:384
+; GFX11-TRUE16-NEXT: scratch_load_b32 v60, off, s32 offset:388
+; GFX11-TRUE16-NEXT: scratch_load_b32 v59, off, s32 offset:392
+; GFX11-TRUE16-NEXT: scratch_load_b32 v58, off, s32 offset:396
+; GFX11-TRUE16-NEXT: scratch_load_b32 v57, off, s32 offset:400
+; GFX11-TRUE16-NEXT: scratch_load_b32 v56, off, s32 offset:404
+; GFX11-TRUE16-NEXT: scratch_load_b32 v47, off, s32 offset:408
+; GFX11-TRUE16-NEXT: scratch_load_b32 v46, off, s32 offset:412
+; GFX11-TRUE16-NEXT: scratch_load_b32 v45, off, s32 offset:416
+; GFX11-TRUE16-NEXT: scratch_load_b32 v44, off, s32 offset:420
+; GFX11-TRUE16-NEXT: scratch_load_b32 v43, off, s32 offset:424
+; GFX11-TRUE16-NEXT: scratch_load_b32 v42, off, s32 offset:428
+; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s32 offset:432
+; GFX11-TRUE16-NEXT: scratch_load_b32 v40, off, s32 offset:436
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
+; GFX11-TRUE16-NEXT: .LBB97_4:
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
+; GFX11-TRUE16-NEXT: s_branch .LBB97_2
+;
+; GFX11-FAKE16-LABEL: bitcast_v128i8_to_v64i16_scalar:
+; GFX11-FAKE16: ; %bb.0:
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-FAKE16-NEXT: s_clause 0x1e
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v40, s32 offset:440
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v41, s32 offset:436
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v42, s32 offset:432
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v43, s32 offset:428
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v44, s32 offset:424
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v45, s32 offset:420
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v46, s32 offset:416
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v47, s32 offset:412
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v56, s32 offset:408
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v57, s32 offset:404
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v58, s32 offset:400
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v59, s32 offset:396
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v60, s32 offset:392
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v61, s32 offset:388
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v62, s32 offset:384
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v63, s32 offset:380
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v72, s32 offset:376
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v73, s32 offset:372
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v74, s32 offset:368
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v75, s32 offset:364
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v76, s32 offset:360
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v77, s32 offset:356
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v78, s32 offset:352
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v79, s32 offset:348
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v88, s32 offset:344
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v89, s32 offset:340
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v90, s32 offset:336
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v91, s32 offset:332
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v92, s32 offset:328
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v93, s32 offset:324
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v94, s32 offset:320
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v54, v30 :: v_dual_mov_b32 v51, v24
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v52, v28 :: v_dual_mov_b32 v55, v26
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v53, v22 :: v_dual_mov_b32 v48, v20
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v50, v18 :: v_dual_mov_b32 v39, v16
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v49, v14 :: v_dual_mov_b32 v34, v8
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v37, v12 :: v_dual_mov_b32 v38, v10
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v36, v6 :: v_dual_mov_b32 v33, v4
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v35, v2 :: v_dual_mov_b32 v32, v0
+; GFX11-FAKE16-NEXT: s_clause 0x1f
+; GFX11-FAKE16-NEXT: scratch_load_b32 v0, off, s32 offset:316
+; GFX11-FAKE16-NEXT: scratch_load_u16 v2, off, s32
+; GFX11-FAKE16-NEXT: scratch_load_u16 v4, off, s32 offset:8
+; GFX11-FAKE16-NEXT: scratch_load_u16 v6, off, s32 offset:16
+; GFX11-FAKE16-NEXT: scratch_load_u16 v8, off, s32 offset:24
+; GFX11-FAKE16-NEXT: scratch_load_u16 v10, off, s32 offset:32
+; GFX11-FAKE16-NEXT: scratch_load_u16 v12, off, s32 offset:40
+; GFX11-FAKE16-NEXT: scratch_load_u16 v14, off, s32 offset:48
+; GFX11-FAKE16-NEXT: scratch_load_u16 v16, off, s32 offset:56
+; GFX11-FAKE16-NEXT: scratch_load_u16 v18, off, s32 offset:64
+; GFX11-FAKE16-NEXT: scratch_load_u16 v20, off, s32 offset:72
+; GFX11-FAKE16-NEXT: scratch_load_u16 v22, off, s32 offset:80
+; GFX11-FAKE16-NEXT: scratch_load_u16 v24, off, s32 offset:88
+; GFX11-FAKE16-NEXT: scratch_load_u16 v26, off, s32 offset:96
+; GFX11-FAKE16-NEXT: scratch_load_u16 v28, off, s32 offset:104
+; GFX11-FAKE16-NEXT: scratch_load_u16 v30, off, s32 offset:112
+; GFX11-FAKE16-NEXT: scratch_load_u16 v31, off, s32 offset:120
+; GFX11-FAKE16-NEXT: scratch_load_u16 v41, off, s32 offset:128
+; GFX11-FAKE16-NEXT: scratch_load_u16 v44, off, s32 offset:136
+; GFX11-FAKE16-NEXT: scratch_load_u16 v45, off, s32 offset:144
+; GFX11-FAKE16-NEXT: scratch_load_u16 v56, off, s32 offset:152
+; GFX11-FAKE16-NEXT: scratch_load_u16 v59, off, s32 offset:160
+; GFX11-FAKE16-NEXT: scratch_load_u16 v60, off, s32 offset:168
+; GFX11-FAKE16-NEXT: scratch_load_u16 v61, off, s32 offset:176
+; GFX11-FAKE16-NEXT: scratch_load_u16 v62, off, s32 offset:184
+; GFX11-FAKE16-NEXT: scratch_load_u16 v63, off, s32 offset:192
+; GFX11-FAKE16-NEXT: scratch_load_u16 v72, off, s32 offset:200
+; GFX11-FAKE16-NEXT: scratch_load_u16 v73, off, s32 offset:208
+; GFX11-FAKE16-NEXT: scratch_load_u16 v74, off, s32 offset:216
+; GFX11-FAKE16-NEXT: scratch_load_u16 v75, off, s32 offset:224
+; GFX11-FAKE16-NEXT: scratch_load_u16 v76, off, s32 offset:232
+; GFX11-FAKE16-NEXT: scratch_load_u16 v77, off, s32 offset:240
+; GFX11-FAKE16-NEXT: s_clause 0x1f
+; GFX11-FAKE16-NEXT: scratch_load_u16 v78, off, s32 offset:248
+; GFX11-FAKE16-NEXT: scratch_load_u16 v79, off, s32 offset:256
+; GFX11-FAKE16-NEXT: scratch_load_u16 v88, off, s32 offset:264
+; GFX11-FAKE16-NEXT: scratch_load_u16 v89, off, s32 offset:272
+; GFX11-FAKE16-NEXT: scratch_load_u16 v90, off, s32 offset:280
+; GFX11-FAKE16-NEXT: scratch_load_u16 v91, off, s32 offset:288
+; GFX11-FAKE16-NEXT: scratch_load_u16 v92, off, s32 offset:296
+; GFX11-FAKE16-NEXT: scratch_load_u16 v93, off, s32 offset:304
+; GFX11-FAKE16-NEXT: scratch_load_u16 v94, off, s32 offset:312
+; GFX11-FAKE16-NEXT: scratch_load_u16 v57, off, s32 offset:308
+; GFX11-FAKE16-NEXT: scratch_load_u16 v58, off, s32 offset:300
+; GFX11-FAKE16-NEXT: scratch_load_u16 v46, off, s32 offset:292
+; GFX11-FAKE16-NEXT: scratch_load_u16 v47, off, s32 offset:284
+; GFX11-FAKE16-NEXT: scratch_load_u16 v40, off, s32 offset:276
+; GFX11-FAKE16-NEXT: scratch_load_u16 v43, off, s32 offset:268
+; GFX11-FAKE16-NEXT: scratch_load_u16 v182, off, s32 offset:260
+; GFX11-FAKE16-NEXT: scratch_load_u16 v183, off, s32 offset:252
+; GFX11-FAKE16-NEXT: scratch_load_u16 v178, off, s32 offset:244
+; GFX11-FAKE16-NEXT: scratch_load_u16 v181, off, s32 offset:236
+; GFX11-FAKE16-NEXT: scratch_load_u16 v164, off, s32 offset:228
+; GFX11-FAKE16-NEXT: scratch_load_u16 v176, off, s32 offset:220
+; GFX11-FAKE16-NEXT: scratch_load_u16 v160, off, s32 offset:212
+; GFX11-FAKE16-NEXT: scratch_load_u16 v163, off, s32 offset:204
+; GFX11-FAKE16-NEXT: scratch_load_u16 v146, off, s32 offset:196
+; GFX11-FAKE16-NEXT: scratch_load_u16 v150, off, s32 offset:188
+; GFX11-FAKE16-NEXT: scratch_load_u16 v135, off, s32 offset:180
+; GFX11-FAKE16-NEXT: scratch_load_u16 v145, off, s32 offset:172
+; GFX11-FAKE16-NEXT: scratch_load_u16 v118, off, s32 offset:164
+; GFX11-FAKE16-NEXT: scratch_load_u16 v131, off, s32 offset:156
+; GFX11-FAKE16-NEXT: scratch_load_u16 v115, off, s32 offset:148
+; GFX11-FAKE16-NEXT: scratch_load_u16 v179, off, s32 offset:140
+; GFX11-FAKE16-NEXT: scratch_load_u16 v162, off, s32 offset:132
+; GFX11-FAKE16-NEXT: s_clause 0xf
+; GFX11-FAKE16-NEXT: scratch_load_u16 v165, off, s32 offset:124
+; GFX11-FAKE16-NEXT: scratch_load_u16 v149, off, s32 offset:116
+; GFX11-FAKE16-NEXT: scratch_load_u16 v151, off, s32 offset:108
+; GFX11-FAKE16-NEXT: scratch_load_u16 v144, off, s32 offset:100
+; GFX11-FAKE16-NEXT: scratch_load_u16 v148, off, s32 offset:92
+; GFX11-FAKE16-NEXT: scratch_load_u16 v129, off, s32 offset:84
+; GFX11-FAKE16-NEXT: scratch_load_u16 v133, off, s32 offset:76
+; GFX11-FAKE16-NEXT: scratch_load_u16 v117, off, s32 offset:68
+; GFX11-FAKE16-NEXT: scratch_load_u16 v119, off, s32 offset:60
+; GFX11-FAKE16-NEXT: scratch_load_u16 v114, off, s32 offset:52
+; GFX11-FAKE16-NEXT: scratch_load_u16 v116, off, s32 offset:44
+; GFX11-FAKE16-NEXT: scratch_load_u16 v100, off, s32 offset:36
+; GFX11-FAKE16-NEXT: scratch_load_u16 v103, off, s32 offset:28
+; GFX11-FAKE16-NEXT: scratch_load_u16 v98, off, s32 offset:20
+; GFX11-FAKE16-NEXT: scratch_load_u16 v99, off, s32 offset:12
+; GFX11-FAKE16-NEXT: scratch_load_u16 v96, off, s32 offset:4
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v64, 8, v1
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v66, 8, v3
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v65, 8, v5
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v68, 8, v7
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v67, 8, v9
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v70, 8, v11
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v69, 8, v13
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v71, 8, v15
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v80, 8, v17
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v82, 8, v19
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v81, 8, v21
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v84, 8, v23
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v83, 8, v25
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v86, 8, v27
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v85, 8, v29
+; GFX11-FAKE16-NEXT: s_mov_b32 s4, 0
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(62)
+; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v97, 8, v2
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v87, 8, v4
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v101, 8, v6
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v102, 8, v8
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v113, 8, v10
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v112, 8, v12
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v130, 8, v14
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v128, 8, v16
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v134, 8, v18
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v132, 8, v20
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v161, 8, v22
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v147, 8, v24
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v166, 8, v26
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v167, 8, v28
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v180, 8, v30
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v177, 8, v31
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v42, 8, v41
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(61)
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v41, 8, v44
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(60)
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v45, 8, v45
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(59)
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v44, 8, v56
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(58)
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v59, 8, v59
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(57)
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v56, 8, v60
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(56)
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v60, 8, v61
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(55)
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v61, 8, v62
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(54)
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v63, 8, v63
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(53)
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v62, 8, v72
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(52)
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v73, 8, v73
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(51)
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v72, 8, v74
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(50)
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v75, 8, v75
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(49)
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v74, 8, v76
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(48)
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v77, 8, v77
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(47)
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v76, 8, v78
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(46)
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v78, 8, v79
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(45)
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v79, 8, v88
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(44)
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v89, 8, v89
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(43)
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v88, 8, v90
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(42)
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v91, 8, v91
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(41)
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v90, 8, v92
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(40)
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v92, 8, v93
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(39)
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v93, 8, v94
+; GFX11-FAKE16-NEXT: s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB97_4
+; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
+; GFX11-FAKE16-NEXT: s_and_b32 s5, s28, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s6, s29, 8
+; GFX11-FAKE16-NEXT: s_and_b32 s7, s2, 0xff
+; GFX11-FAKE16-NEXT: s_or_b32 s5, s5, s6
+; GFX11-FAKE16-NEXT: s_lshl_b32 s6, s1, 8
+; GFX11-FAKE16-NEXT: v_and_b32_e64 v5, 0xffff, s5
+; GFX11-FAKE16-NEXT: s_and_b32 s5, s0, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s8, s3, 8
+; GFX11-FAKE16-NEXT: s_or_b32 s5, s5, s6
+; GFX11-FAKE16-NEXT: s_or_b32 s6, s7, s8
+; GFX11-FAKE16-NEXT: s_and_b32 s7, s16, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s8, s17, 8
+; GFX11-FAKE16-NEXT: s_and_b32 s9, s18, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s10, s19, 8
+; GFX11-FAKE16-NEXT: s_or_b32 s7, s7, s8
+; GFX11-FAKE16-NEXT: s_or_b32 s8, s9, s10
+; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s5, s5, s6
+; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s6, s7, s8
+; GFX11-FAKE16-NEXT: s_and_b32 s7, s20, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s8, s21, 8
+; GFX11-FAKE16-NEXT: s_and_b32 s9, s22, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s10, s23, 8
+; GFX11-FAKE16-NEXT: s_or_b32 s7, s7, s8
+; GFX11-FAKE16-NEXT: s_or_b32 s8, s9, s10
+; GFX11-FAKE16-NEXT: s_and_b32 s9, s24, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s10, s25, 8
+; GFX11-FAKE16-NEXT: s_and_b32 s11, s26, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s12, s27, 8
+; GFX11-FAKE16-NEXT: s_or_b32 s9, s9, s10
+; GFX11-FAKE16-NEXT: s_or_b32 s10, s11, s12
+; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s7, s7, s8
+; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s8, s9, s10
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v36
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v32
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v35
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v34
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v33
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v2, v68
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v64
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v66
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v6, v4, v67
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v3, v65
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v4, v0, 16, v5
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v38
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xff, v39
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v6, 16, v2
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v49
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v5, v3, 16, v1
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v37
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v70
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v50
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v2, v71
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xff, v48
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v69
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v3, v82
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v9, v7, v80
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v10, v8, v81
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v1, 16, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v53
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v9, 16, v2
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v55
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v51
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v10, 16, v3
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v84
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v52
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xff, v54
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v2, v86
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v83
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xff, v96
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v3, v85
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v12, v10, v97
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v10, v1, 16, v0
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v11, v87
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v99
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v11, v3, 16, v2
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v103
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xff, v114
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v98
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v12, v0, 16, v12
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v100
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v3, v113
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v101
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xff, v116
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v17, v14, v128
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v112
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xff, v117
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v2, v102
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v13, v13, v130
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v16, 0xff, v133
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v20, v14, v132
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v14, v0, 16, v3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v148
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xff, v119
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xff, v129
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v16, v16, v161
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xffff, v13
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v13, v2, 16, v1
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v2, s7
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v166
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v144
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v15, v15, v134
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v18, v18, v147
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xffff, v16
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v167
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v15
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v15, v17, 16, v19
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v17, v18, 16, v22
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v3, s8
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v1, 16, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v151
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v149
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v16, v20, 16, v21
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v180
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v177
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v1, 16, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v165
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v162
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v42
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v41
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v1, 16, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v179
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v115
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v45
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v44
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v1, 16, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v131
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v118
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v59
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v56
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v1, 16, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v145
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v135
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v60
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v61
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v1, 16, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v150
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v146
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v63
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v62
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v24, v1, 16, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v163
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v160
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v73
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v72
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v25, v1, 16, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v176
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v164
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v75
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v74
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v26, v1, 16, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v181
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v178
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v77
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v76
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v27, v1, 16, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v183
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v182
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v78
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v79
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v28, v1, 16, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v43
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v40
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v89
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v88
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v29, v1, 16, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v47
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v46
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v91
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v90
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v30, v1, 16, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v58
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v57
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v0, v92
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v93
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v31, v1, 16, v0
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6
+; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB97_3
+; GFX11-FAKE16-NEXT: .LBB97_2: ; %cmp.true
+; GFX11-FAKE16-NEXT: s_add_i32 s28, s28, 3
+; GFX11-FAKE16-NEXT: s_lshl_b32 s5, s29, 8
+; GFX11-FAKE16-NEXT: s_and_b32 s4, s28, 0xff
+; GFX11-FAKE16-NEXT: s_add_i32 s24, s24, 3
+; GFX11-FAKE16-NEXT: s_or_b32 s4, s5, s4
+; GFX11-FAKE16-NEXT: s_and_b32 s5, s24, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s6, s25, 8
+; GFX11-FAKE16-NEXT: s_add_i32 s26, s26, 3
+; GFX11-FAKE16-NEXT: s_or_b32 s5, s6, s5
+; GFX11-FAKE16-NEXT: s_and_b32 s6, s26, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s7, s27, 8
+; GFX11-FAKE16-NEXT: s_add_i32 s20, s20, 3
+; GFX11-FAKE16-NEXT: s_or_b32 s6, s7, s6
+; GFX11-FAKE16-NEXT: s_and_b32 s7, s20, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s8, s21, 8
+; GFX11-FAKE16-NEXT: s_add_i32 s22, s22, 3
+; GFX11-FAKE16-NEXT: s_or_b32 s7, s8, s7
+; GFX11-FAKE16-NEXT: s_and_b32 s8, s22, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s9, s23, 8
+; GFX11-FAKE16-NEXT: s_add_i32 s16, s16, 3
+; GFX11-FAKE16-NEXT: s_or_b32 s8, s9, s8
+; GFX11-FAKE16-NEXT: s_and_b32 s9, s16, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s10, s17, 8
+; GFX11-FAKE16-NEXT: s_add_i32 s18, s18, 3
+; GFX11-FAKE16-NEXT: s_add_i32 s0, s0, 3
+; GFX11-FAKE16-NEXT: s_add_i32 s2, s2, 3
+; GFX11-FAKE16-NEXT: s_or_b32 s9, s10, s9
+; GFX11-FAKE16-NEXT: s_and_b32 s10, s18, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s11, s19, 8
+; GFX11-FAKE16-NEXT: s_and_b32 s0, s0, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s1, s1, 8
+; GFX11-FAKE16-NEXT: s_and_b32 s2, s2, 0xff
+; GFX11-FAKE16-NEXT: s_lshl_b32 s3, s3, 8
+; GFX11-FAKE16-NEXT: s_or_b32 s10, s11, s10
+; GFX11-FAKE16-NEXT: s_or_b32 s0, s1, s0
+; GFX11-FAKE16-NEXT: s_or_b32 s1, s3, s2
+; GFX11-FAKE16-NEXT: s_addk_i32 s9, 0x300
+; GFX11-FAKE16-NEXT: s_addk_i32 s0, 0x300
+; GFX11-FAKE16-NEXT: s_addk_i32 s1, 0x300
+; GFX11-FAKE16-NEXT: s_addk_i32 s10, 0x300
+; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s0, s0, s1
+; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s1, s9, s10
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(37)
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v0, 3, v58
+; GFX11-FAKE16-NEXT: s_addk_i32 s5, 0x300
+; GFX11-FAKE16-NEXT: s_addk_i32 s6, 0x300
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v1, 3, v57
+; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s3, s5, s6
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(35)
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v2, 3, v47
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-FAKE16-NEXT: s_addk_i32 s7, 0x300
+; GFX11-FAKE16-NEXT: s_addk_i32 s8, 0x300
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v92, v0
+; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s2, s7, s8
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v3, 3, v46
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v93, v1
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v91, v2
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(33)
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v43
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v27, 0x300, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v0, 3, v40
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v31, 0x300, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v28, 0x300, v2
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v4
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v90, v3
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(31)
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v2, 3, v183
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v182
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v89, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v30, 0x300, v3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v4
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(29)
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v181
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v88, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v181, 0x300, v1
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v78, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v79, v3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v29, 0x300, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v0, 3, v178
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v182, 0x300, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v178, 0x300, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v77, v3
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(27)
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v2, 3, v176
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v3, 3, v164
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(25)
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v163
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v163, 0x300, v1
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v76, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v160
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v75, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v160, 0x300, v0
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v74, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v73, v3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v22, 0x300, v1
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(23)
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v1, 3, v150
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v26, 0x300, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v23, 0x300, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v72, v3
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v2, 3, v146
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(21)
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v3, 3, v145
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v135
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v25, 0x300, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v63, v1
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v4
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(19)
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v131
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v62, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v131, 0x300, v1
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v60, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v61, v3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v24, 0x300, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v0, 3, v118
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v135, 0x300, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v118, 0x300, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v59, v3
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(17)
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v2, 3, v179
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v3, 3, v115
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(15)
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v165
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v115, 0x300, v1
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v56, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v162
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v45, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v145, 0x300, v0
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v44, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v42, v3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v17, 0x300, v1
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(13)
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v1, 3, v151
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v21, 0x300, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v18, 0x300, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v41, v3
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v2, 3, v149
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(11)
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v3, 3, v148
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v144
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v20, 0x300, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v180, v1
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v4
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(9)
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v133
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v177, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v19, 0x300, v1
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v166, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v167, v3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v133, 0x300, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v0, 3, v129
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v129, 0x300, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v144, 0x300, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v161, v3
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(7)
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v2, 3, v119
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xff, v0
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v3, 3, v117
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(5)
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v116
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v116, 0x300, v1
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v147, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v114
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(1)
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v5, 3, v99
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v114, 0x300, v0
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v132, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v130, v3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v103
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v6, 3, v98
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xff, v5
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v7, 3, v54
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v8, 3, v53
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xff, v6
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v9, 3, v39
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v33, 3, v33
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v16, 0x300, v0
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v113, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v0, 0x300, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v128, v3
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v3, 3, v100
+; GFX11-FAKE16-NEXT: s_addk_i32 s4, 0x300
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v11, 0x300, v4
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v101, v5
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v102, v6
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xff, v7
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v7, 3, v96
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v134, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v12, 0x300, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v13, 0x300, v5
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v97, v6
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v5, 3, v55
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xff, v7
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v7, 3, v52
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v14, 0x300, v4
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v5
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v87, v6
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xff, v7
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xff, v8
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v8, 3, v51
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v86, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v15, 0x300, v5
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v85, v6
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v6, v84, v7
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xff, v8
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v51, 0x300, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v50
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v50, 0x300, v5
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v8, 3, v49
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v83, v7
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v7, 3, v48
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v10, 0x300, v5
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xff, v7
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v82, v4
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xff, v8
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xff, v9
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v9, 3, v38
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v81, v5
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v38, 0x300, v4
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v71, v7
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v80, v8
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xff, v9
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v9, 0x300, v5
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v5, 3, v37
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v37, 0x300, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v39, 0x300, v7
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v70, v8
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v7, 3, v36
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xff, v5
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v8, 3, v34
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v34, 3, v35
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v35, 0x300, v4
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v7
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v69, v5
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xff, v8
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xff, v34
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v112, v3
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v68, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v34, 0x300, v5
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v67, v7
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v66, v8
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xff, v33
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v33, 0x300, v4
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v32
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v32, 0x300, v5
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v5, 0x300, v7
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v65, v8
+; GFX11-FAKE16-NEXT: v_and_b32_e64 v8, 0xffff, s4
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v36, 0xffff, v6
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v7, 0x300, v7
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v64, v4
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xffff, v14
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v10, v10, 16, v36
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v5, v7, 16, v5
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xffff, v33
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 0x300, v4
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xffff, v37
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v36, 0xffff, v22
+; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v32, 16, v7
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v4, v4, 16, v8
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xffff, v35
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v32, 0xffff, v51
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v35, 0xffff, v38
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xffff, v19
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v34, 16, v8
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v39, 16, v33
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xffff, v12
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v34, 0xffff, v11
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v11, v50, 16, v32
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v32, 0xffff, v1
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v9, 16, v35
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v12, v15, 16, v14
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v13, v13, 16, v33
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v14, v3, 16, v34
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v16, v16, 16, v32
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v32, 0xffff, v116
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xffff, v129
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v34, 0xffff, v18
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v35, 0xffff, v17
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v26, v26, 16, v36
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v17, v114, 16, v32
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v144, 16, v33
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v20, 16, v34
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v21, 16, v35
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v32, 0xffff, v115
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xffff, v135
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v34, 0xffff, v131
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v35, 0xffff, v23
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v36, 0xffff, v27
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v145, 16, v32
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v118, 16, v33
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v24, v24, 16, v34
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v25, v25, 16, v35
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v32, 0xffff, v163
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xffff, v182
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v34, 0xffff, v181
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v35, 0xffff, v28
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v15, v2, 16, v0
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v2, s2
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v3, s3
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v133, 16, v19
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v27, v160, 16, v32
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v28, v178, 16, v33
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v29, v29, 16, v34
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v30, v30, 16, v35
+; GFX11-FAKE16-NEXT: v_lshl_or_b32 v31, v31, 16, v36
+; GFX11-FAKE16-NEXT: .LBB97_3: ; %end
+; GFX11-FAKE16-NEXT: s_clause 0x1e
+; GFX11-FAKE16-NEXT: scratch_load_b32 v94, off, s32 offset:320
+; GFX11-FAKE16-NEXT: scratch_load_b32 v93, off, s32 offset:324
+; GFX11-FAKE16-NEXT: scratch_load_b32 v92, off, s32 offset:328
+; GFX11-FAKE16-NEXT: scratch_load_b32 v91, off, s32 offset:332
+; GFX11-FAKE16-NEXT: scratch_load_b32 v90, off, s32 offset:336
+; GFX11-FAKE16-NEXT: scratch_load_b32 v89, off, s32 offset:340
+; GFX11-FAKE16-NEXT: scratch_load_b32 v88, off, s32 offset:344
+; GFX11-FAKE16-NEXT: scratch_load_b32 v79, off, s32 offset:348
+; GFX11-FAKE16-NEXT: scratch_load_b32 v78, off, s32 offset:352
+; GFX11-FAKE16-NEXT: scratch_load_b32 v77, off, s32 offset:356
+; GFX11-FAKE16-NEXT: scratch_load_b32 v76, off, s32 offset:360
+; GFX11-FAKE16-NEXT: scratch_load_b32 v75, off, s32 offset:364
+; GFX11-FAKE16-NEXT: scratch_load_b32 v74, off, s32 offset:368
+; GFX11-FAKE16-NEXT: scratch_load_b32 v73, off, s32 offset:372
+; GFX11-FAKE16-NEXT: scratch_load_b32 v72, off, s32 offset:376
+; GFX11-FAKE16-NEXT: scratch_load_b32 v63, off, s32 offset:380
+; GFX11-FAKE16-NEXT: scratch_load_b32 v62, off, s32 offset:384
+; GFX11-FAKE16-NEXT: scratch_load_b32 v61, off, s32 offset:388
+; GFX11-FAKE16-NEXT: scratch_load_b32 v60, off, s32 offset:392
+; GFX11-FAKE16-NEXT: scratch_load_b32 v59, off, s32 offset:396
+; GFX11-FAKE16-NEXT: scratch_load_b32 v58, off, s32 offset:400
+; GFX11-FAKE16-NEXT: scratch_load_b32 v57, off, s32 offset:404
+; GFX11-FAKE16-NEXT: scratch_load_b32 v56, off, s32 offset:408
+; GFX11-FAKE16-NEXT: scratch_load_b32 v47, off, s32 offset:412
+; GFX11-FAKE16-NEXT: scratch_load_b32 v46, off, s32 offset:416
+; GFX11-FAKE16-NEXT: scratch_load_b32 v45, off, s32 offset:420
+; GFX11-FAKE16-NEXT: scratch_load_b32 v44, off, s32 offset:424
+; GFX11-FAKE16-NEXT: scratch_load_b32 v43, off, s32 offset:428
+; GFX11-FAKE16-NEXT: scratch_load_b32 v42, off, s32 offset:432
+; GFX11-FAKE16-NEXT: scratch_load_b32 v41, off, s32 offset:436
+; GFX11-FAKE16-NEXT: scratch_load_b32 v40, off, s32 offset:440
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
+; GFX11-FAKE16-NEXT: .LBB97_4:
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
+; GFX11-FAKE16-NEXT: s_branch .LBB97_2
+ %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) {
+; SI-LABEL: bitcast_v64i16_to_v128i8:
+; SI: ; %bb.0:
+; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:136
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:12
+; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:28
+; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:44
+; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:60
+; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:76
+; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:92
+; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:108
+; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:124
+; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:132
+; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:128
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT: ; implicit-def: $vgpr23
+; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v28
+; SI-NEXT: ; kill: killed $vgpr23
+; SI-NEXT: ; implicit-def: $vgpr23
+; SI-NEXT: ; implicit-def: $vgpr46
+; SI-NEXT: ; implicit-def: $vgpr45
+; SI-NEXT: ; implicit-def: $vgpr44
+; SI-NEXT: ; implicit-def: $vgpr43
+; SI-NEXT: ; implicit-def: $vgpr42
+; SI-NEXT: ; implicit-def: $vgpr41
+; SI-NEXT: ; implicit-def: $vgpr40
+; SI-NEXT: ; implicit-def: $vgpr55
+; SI-NEXT: ; implicit-def: $vgpr54
+; SI-NEXT: ; implicit-def: $vgpr53
+; SI-NEXT: ; implicit-def: $vgpr52
+; SI-NEXT: ; implicit-def: $vgpr51
+; SI-NEXT: ; implicit-def: $vgpr50
+; SI-NEXT: ; implicit-def: $vgpr49
+; SI-NEXT: ; implicit-def: $vgpr48
+; SI-NEXT: ; implicit-def: $vgpr39
+; SI-NEXT: ; implicit-def: $vgpr38
+; SI-NEXT: ; implicit-def: $vgpr37
+; SI-NEXT: ; implicit-def: $vgpr36
+; SI-NEXT: ; implicit-def: $vgpr35
+; SI-NEXT: ; implicit-def: $vgpr34
+; SI-NEXT: ; implicit-def: $vgpr33
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr32
+; SI-NEXT: ; kill: killed $vgpr23
+; SI-NEXT: ; implicit-def: $vgpr23
+; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:4
+; SI-NEXT: s_waitcnt vmcnt(11)
+; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
+; SI-NEXT: s_waitcnt vmcnt(10)
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v3
+; SI-NEXT: s_waitcnt vmcnt(6)
+; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v11
+; SI-NEXT: s_waitcnt vmcnt(5)
+; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v13
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v17
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:120
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
+; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:116
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:112
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:88
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:84
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:80
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v56
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:56
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
+; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:52
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:48
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v57
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:24
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
+; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v5
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v7
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v9
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v10
+; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v14
+; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v18
+; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v22
+; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v26
+; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v30
+; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:72
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: s_waitcnt vmcnt(13)
+; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:104
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v15
+; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v58
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v62
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr26
+; SI-NEXT: ; implicit-def: $vgpr30
+; SI-NEXT: ; implicit-def: $vgpr18
+; SI-NEXT: ; implicit-def: $vgpr22
+; SI-NEXT: ; implicit-def: $vgpr14
+; SI-NEXT: ; implicit-def: $vgpr10
+; SI-NEXT: ; implicit-def: $vgpr6
+; SI-NEXT: ; kill: killed $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: s_waitcnt vmcnt(2)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
+; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:68
+; SI-NEXT: s_waitcnt expcnt(1)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:40
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v63
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
+; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:36
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v59
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
+; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:100
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:96
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v61
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:64
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:32
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32
+; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v60
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v4
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v12
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v20
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v47
+; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; SI-NEXT: s_cbranch_execz .LBB98_2
+; SI-NEXT: ; %bb.1: ; %cmp.false
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v23, 0xffff, v47
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v46, v2, v6
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v45, v2, v6
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v44, v2, v6
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v8
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v43, v2, v6
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v42, v2, v6
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v12
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v41, v2, v6
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v40, v2, v6
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v16
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v55, v2, v6
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v54, v2, v6
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v20
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v53, v2, v6
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v52, v2, v6
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v24
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v51, v2, v6
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v50, v2, v6
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v49, v2, v27
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; SI-NEXT: ; implicit-def: $vgpr27
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v48, v2, v6
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v39, v2, v19
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
+; SI-NEXT: ; implicit-def: $vgpr19
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v38, v2, v6
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v37, v2, v15
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; SI-NEXT: ; implicit-def: $vgpr15
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v36, v2, v6
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v35, v2, v29
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; SI-NEXT: ; implicit-def: $vgpr29
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v34, v2, v6
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v33, v2, v21
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; SI-NEXT: ; implicit-def: $vgpr21
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v31, v2, v6
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v32, v2, v17
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
+; SI-NEXT: ; implicit-def: $vgpr17
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v26, v2, v25
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; SI-NEXT: ; implicit-def: $vgpr25
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v30, v2, v13
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; SI-NEXT: ; implicit-def: $vgpr13
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v18, v2, v11
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT: ; implicit-def: $vgpr11
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v22, v2, v9
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
+; SI-NEXT: ; implicit-def: $vgpr9
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v14, v2, v7
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v10, v2, v5
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; SI-NEXT: ; implicit-def: $vgpr5
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v6, v2, v3
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; SI-NEXT: ; implicit-def: $vgpr3
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v2, v2, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v4
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v12
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v20
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v28
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v60
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v58
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v59
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v57
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v63
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_bfe_u32 v1, v4, 8, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_bfe_u32 v1, v8, 8, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_bfe_u32 v1, v12, 8, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_bfe_u32 v1, v16, 8, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_bfe_u32 v1, v20, 8, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_bfe_u32 v1, v24, 8, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_bfe_u32 v1, v28, 8, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_bfe_u32 v1, v60, 8, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_bfe_u32 v1, v58, 8, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_bfe_u32 v1, v59, 8, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_bfe_u32 v1, v57, 8, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_bfe_u32 v1, v63, 8, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v56
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_bfe_u32 v1, v56, 8, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v61
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_bfe_u32 v1, v61, 8, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v62
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_bfe_u32 v1, v62, 8, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_bfe_u32 v1, v47, 8, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v45, v46, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v45, v46, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v45, v46, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v43, v44, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v43, v44, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v43, v44, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v41, v42, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v41, v42, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v41, v42, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v55, v40, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v55, v40, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v55, v40, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v53, v54, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v53, v54, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v53, v54, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v51, v52, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v51, v52, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v51, v52, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v49, v50, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v49, v50, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v49, v50, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v39, v48, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v39, v48, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v39, v48, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v37, v38, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v37, v38, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v37, v38, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v35, v36, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v35, v36, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v35, v36, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v33, v34, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v33, v34, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v33, v34, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v32, v31, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v32, v31, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v32, v31, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v30, v26, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v30, v26, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v30, v26, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v22, v18, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v22, v18, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v22, v18, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v10, v14, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v10, v14, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v10, v14, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v2, v6, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v2, v6, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v2, v6, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v45
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v43
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v41
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v55
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v53
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v51
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v49
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v39
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v37
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v35
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v33
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v32
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v30
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v22
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v10
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v2
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: .LBB98_2: ; %Flow
+; SI-NEXT: s_or_saveexec_b64 s[4:5], s[4:5]
+; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; SI-NEXT: s_xor_b64 exec, exec, s[4:5]
+; SI-NEXT: s_cbranch_execz .LBB98_4
+; SI-NEXT: ; %bb.3: ; %cmp.true
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; SI-NEXT: s_mov_b32 s6, 0x30000
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v2, v3, v2
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v6, vcc, 0x30000, v2
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v3
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v2, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
+; SI-NEXT: v_alignbit_b32 v12, v2, v6, 24
+; SI-NEXT: v_alignbit_b32 v20, v2, v6, 16
+; SI-NEXT: v_alignbit_b32 v47, v2, v6, 8
+; SI-NEXT: v_lshrrev_b32_e32 v8, 24, v2
+; SI-NEXT: v_lshrrev_b32_e32 v23, 16, v2
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v1, v11, v1
+; SI-NEXT: v_add_i32_e32 v18, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v1, v9, v1
+; SI-NEXT: v_add_i32_e32 v22, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
+; SI-NEXT: v_alignbit_b32 v62, v22, v18, 24
+; SI-NEXT: v_alignbit_b32 v63, v22, v18, 16
+; SI-NEXT: v_lshrrev_b32_e32 v59, 24, v22
+; SI-NEXT: v_lshrrev_b32_e32 v60, 16, v22
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v1, v25, v1
+; SI-NEXT: v_add_i32_e32 v26, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v3
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT: v_or_b32_e32 v3, v7, v3
+; SI-NEXT: v_add_i32_e32 v14, vcc, s6, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v1, v13, v1
+; SI-NEXT: v_add_i32_e32 v30, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v1, v3, v1
+; SI-NEXT: v_add_i32_e32 v31, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v1, v17, v1
+; SI-NEXT: v_add_i32_e32 v32, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; SI-NEXT: v_lshrrev_b32_e32 v16, 16, v32
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v1, v3, v1
+; SI-NEXT: v_add_i32_e32 v34, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v1, v21, v1
+; SI-NEXT: v_add_i32_e32 v33, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; SI-NEXT: v_lshrrev_b32_e32 v58, 16, v33
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v1, v3, v1
+; SI-NEXT: v_add_i32_e32 v36, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v1, v29, v1
+; SI-NEXT: v_add_i32_e32 v35, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v1, v3, v1
+; SI-NEXT: v_add_i32_e32 v38, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v1, v15, v1
+; SI-NEXT: v_add_i32_e32 v37, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v1, v3, v1
+; SI-NEXT: v_add_i32_e32 v48, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v1, v19, v1
+; SI-NEXT: v_add_i32_e32 v39, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v1, v3, v1
+; SI-NEXT: v_add_i32_e32 v50, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4
+; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; SI-NEXT: v_or_b32_e32 v4, v5, v4
+; SI-NEXT: v_add_i32_e32 v10, vcc, s6, v4
+; SI-NEXT: v_alignbit_b32 v56, v10, v14, 24
+; SI-NEXT: v_alignbit_b32 v57, v10, v14, 16
+; SI-NEXT: v_alignbit_b32 v61, v10, v14, 8
+; SI-NEXT: v_lshrrev_b32_e32 v24, 24, v10
+; SI-NEXT: v_lshrrev_b32_e32 v28, 16, v10
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v1, v27, v1
+; SI-NEXT: v_add_i32_e32 v49, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v1, v3, v1
+; SI-NEXT: v_add_i32_e32 v52, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v1, v3, v1
+; SI-NEXT: v_add_i32_e32 v51, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v1, v3, v1
+; SI-NEXT: v_add_i32_e32 v54, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v1, v3, v1
+; SI-NEXT: v_add_i32_e32 v53, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v1, v3, v1
+; SI-NEXT: v_add_i32_e32 v40, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v1, v3, v1
+; SI-NEXT: v_add_i32_e32 v55, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v1, v3, v1
+; SI-NEXT: v_add_i32_e32 v42, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v1, v3, v1
+; SI-NEXT: v_add_i32_e32 v41, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v1, v3, v1
+; SI-NEXT: v_add_i32_e32 v44, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v1, v3, v1
+; SI-NEXT: v_add_i32_e32 v43, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v1, v3, v1
+; SI-NEXT: v_add_i32_e32 v46, vcc, s6, v1
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_or_b32_e32 v1, v3, v1
+; SI-NEXT: v_add_i32_e32 v45, vcc, s6, v1
+; SI-NEXT: v_alignbit_b32 v1, v45, v46, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v45, v46, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v43, v44, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v43, v44, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v43, v44, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v41, v42, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v41, v42, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v41, v42, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v55, v40, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v55, v40, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v55, v40, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v53, v54, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v53, v54, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v53, v54, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v51, v52, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v51, v52, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v51, v52, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v49, v50, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v49, v50, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v49, v50, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v39, v48, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v39, v48, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v39, v48, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v37, v38, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v37, v38, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v37, v38, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v35, v36, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v35, v36, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v35, v36, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v33, v34, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v33, v34, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v33, v34, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v32, v31, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v32, v31, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v32, v31, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v30, v26, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v30, v26, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v30, v26, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, v22, v18, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v45
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v45
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v45
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v43
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v43
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v43
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v41
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v41
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v41
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v55
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v55
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v55
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v53
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v53
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v53
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v51
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v51
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v51
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v49
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v49
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v49
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v39
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v39
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v39
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v37
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v37
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v37
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v35
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v35
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v35
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v33
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v33
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v32
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v32
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 24, v30
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v30
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v30
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v22
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v10
+; SI-NEXT: v_alignbit_b32 v4, v45, v46, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 8, v2
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
+; SI-NEXT: .LBB98_4: ; %end
+; SI-NEXT: s_or_b64 exec, exec, s[4:5]
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v46
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v45
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 4, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v44
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 8, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v43
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 12, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v42
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 16, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v41
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 20, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v40
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 24, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v55
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 28, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v54
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 32, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v53
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 36, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v52
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 40, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v51
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 44, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v50
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 48, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v49
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 52, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v48
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 56, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v39
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 60, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v38
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 64, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v37
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x44, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v36
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x48, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v35
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x4c, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v34
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x50, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v33
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v58
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x54, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v31
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x58, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v32
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v16
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x5c, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v26
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x60, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v30
+; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x64, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v18
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v62
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v63
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x68, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v22
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v59
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v60
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x6c, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v14
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v61
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v57
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v56
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x70, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v10
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v24
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v28
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x74, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v6
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v47
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v20
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v12
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: v_or_b32_e32 v1, v1, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x78, v0
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v2
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT: v_lshlrev_b32_e32 v3, 24, v8
+; SI-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v2, 8, v2
+; SI-NEXT: v_or_b32_e32 v1, v1, v2
+; SI-NEXT: v_and_b32_e32 v2, 0xff, v23
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v2, v3, v2
+; SI-NEXT: v_or_b32_e32 v1, v1, v2
+; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; SI-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 .LBB98_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: .LBB98_2: ; %Flow
+; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; VI-NEXT: s_cbranch_execz .LBB98_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: .LBB98_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(2)
+; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2
+; 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: 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 .LBB98_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: .LBB98_2: ; %Flow
+; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX9-NEXT: s_cbranch_execz .LBB98_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: .LBB98_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(2)
+; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
+; 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(2)
+; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
+; 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(2)
+; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
+; 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(2)
+; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
+; 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(2)
+; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
+; 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(2)
+; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1
+; 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-TRUE16-LABEL: bitcast_v64i16_to_v128i8:
+; GFX11-TRUE16: ; %bb.0:
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT: s_clause 0x2
+; GFX11-TRUE16-NEXT: scratch_load_b32 v33, off, s32 offset:8
+; GFX11-TRUE16-NEXT: scratch_load_b32 v32, off, s32 offset:4
+; GFX11-TRUE16-NEXT: scratch_load_b32 v31, off, s32
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr163_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr162_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2)
+; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v33
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
+; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
+; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB98_2
+; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[9:10]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[7:8]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[15:16]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[5:6]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[49:50], 24, v[13:14]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[52:53], 24, v[11:12]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[3:4]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v15
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v11
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v9
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v5
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v1
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v32
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v32
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v31
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 24, v30
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v30
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v29
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 24, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v27
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v26
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v26
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v25
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v24
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v24
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v23
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v21
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v20
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v20
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v19
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[1:2]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[19:20]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[17:18]
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v17
+; GFX11-TRUE16-NEXT: .LBB98_2: ; %Flow
+; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
+; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB98_4
+; GFX11-TRUE16-NEXT: ; %bb.3: ; %cmp.true
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(1)
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v32, v32, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[9:10]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[7:8]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[15:16]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[5:6]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[49:50], 24, v[13:14]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[52:53], 24, v[11:12]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[3:4]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[1:2]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[19:20]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[17:18]
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v15
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v11
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v9
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v5
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v1
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v32
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v32
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v31
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 24, v30
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v30
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v29
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 24, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v27
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v26
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v26
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v25
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v24
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v24
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v23
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v21
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v20
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v20
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v19
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v17
+; GFX11-TRUE16-NEXT: .LBB98_4: ; %end
+; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v163.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v68.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v161.l
+; GFX11-TRUE16-NEXT: v_or_b16 v51.l, v1.h, v34.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v162.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.h, v33.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v54
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v160.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v150.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v68.l, v1.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v69.l, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v67.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v54, v51
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v68
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v67, 16, v69
+; GFX11-TRUE16-NEXT: v_or_b16 v51.l, v2.l, v2.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v3.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v151.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v54, v67
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v51
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.h, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.h, v33.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v149.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v54
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v68.l, v4.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v66.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v3.h, v4.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v51, v54
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v67
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v68
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v148.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v147.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v4.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v51, v54
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.h, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.h, v33.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v65.l
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v66
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v146.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v54
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.h, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v51, v54
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v66
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v135.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v134.l
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v65
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v9.l
+; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v51, v54
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v65
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v64.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v130.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v7.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v8.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v131.l
+; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v9.l, v9.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.l, v33.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v64, 0xffff, v64
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v65, 16, v65
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v66
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.h, v8.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v11.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v64, v65
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v65, 0xffff, v9
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v117.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v52.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v7.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v51, v54
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v119.l
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h
+; GFX11-TRUE16-NEXT: s_clause 0x1
+; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off
+; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v103.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v115.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v114.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11
+; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v13.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v49.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v51.l, v12.h, v34.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v11
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v14.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v101.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v5
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v48.l
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v65, v9
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v99.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v15.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v98.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v12, v51
+; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v7
+; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v16.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v87.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v85.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v5
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v11
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v18.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v145.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v8
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v9
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v17.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v39.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v18.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v144.l
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v6
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v12, v13
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v14, v15
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v19.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v133.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v19.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v50.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v10
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v20.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v129.l
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v8
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v16, v17
+; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v20.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v132.l
+; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v11
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v21.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v128.l
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v12
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v13
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v21.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v38.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v10
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v15, v16
+; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v23.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v113.l
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v14
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v22.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v118.l
+; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v11
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v17, v18
+; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v23.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v24.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v112.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h
+; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v22.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v12
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v19, v20
+; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v102.l
+; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v25.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v100.l
+; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h
+; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v17.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v15
+; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v25.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v36.l
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v13
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v21, v22
+; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v18.h
+; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v19.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v16
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v17
+; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v26.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v97.l
+; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v26.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v96.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v18
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v19
+; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h
+; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v17.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v15
+; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v27.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v86.l
+; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v27.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v16
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v17
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v20, v21
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v22, v23
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v24, v25
+; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v28.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v84.l
+; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v28.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v83.l
+; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v29.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v82.l
+; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v29.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v34.l
+; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v30.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v81.l
+; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v18.h
+; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v21.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v22.h
+; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v23.h
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v53.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v116.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v35.l
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v18
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v26, v27
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v20
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v21
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v22
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v23
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v24
+; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v30.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v80.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v31.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v71.l
+; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v31.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v33.l
+; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v32.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v70.l
+; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v32.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v55.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h
+; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h
+; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v19.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v66
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v64, 16, v67
+; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v21.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v22.h
+; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v23.h
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v19
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v54, v64
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v30, 16, v20
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v21
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v22
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v23
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v39, v9
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v38, v14
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v35, v19
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v25, v26
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v27, v28
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v29, v30
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v31, v32
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v33, v24
+; GFX11-TRUE16-NEXT: s_clause 0x5
+; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off offset:32
+; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:48
+; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:64
+; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:80
+; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:96
+; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:112
+; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-FAKE16-LABEL: bitcast_v64i16_to_v128i8:
+; GFX11-FAKE16: ; %bb.0:
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-FAKE16-NEXT: s_clause 0x13
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v40, s32 offset:88
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v41, s32 offset:84
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v42, s32 offset:80
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v43, s32 offset:76
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v44, s32 offset:72
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v45, s32 offset:68
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v46, s32 offset:64
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v47, s32 offset:60
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v56, s32 offset:56
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v57, s32 offset:52
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v58, s32 offset:48
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v59, s32 offset:44
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v60, s32 offset:40
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v61, s32 offset:36
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v62, s32 offset:32
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v63, s32 offset:28
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v72, s32 offset:24
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v73, s32 offset:20
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v74, s32 offset:16
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v75, s32 offset:12
+; GFX11-FAKE16-NEXT: s_clause 0x2
+; GFX11-FAKE16-NEXT: scratch_load_b32 v33, off, s32 offset:8
+; GFX11-FAKE16-NEXT: scratch_load_b32 v32, off, s32 offset:4
+; GFX11-FAKE16-NEXT: scratch_load_b32 v31, off, s32
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr74
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr72
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr64
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr63
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr62
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr60
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr57
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr47
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr54
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr45
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr43
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr183
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr181
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr53
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr179
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr177
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr167
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr165
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr164
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr52
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr161
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr151
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr150
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr147
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr145
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr51
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr144
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr133
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr131
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr129
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr119
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr50
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr117
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr116
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr115
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr102
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr100
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr49
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr98
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr97
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr96
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr86
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr82
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr48
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr80
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr70
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr68
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr75
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr73
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr66
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr61
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr59
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr58
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr56
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr46
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr65
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr44
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr41
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr40
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr182
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr180
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr38
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr178
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr176
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr166
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr163
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr162
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr37
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr160
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr149
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr148
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr146
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr135
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr36
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr134
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr132
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr130
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr128
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr118
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr35
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr114
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr113
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr112
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr103
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr101
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr99
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr87
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr85
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr84
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr83
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr81
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr71
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr69
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr34
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(2)
+; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v33
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr33
+; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
+; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB98_2
+; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[15:16]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[49:50], 24, v[13:14]
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[11:12]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[51:52], 24, v[9:10]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[52:53], 24, v[7:8]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[1:2]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[5:6]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[19:20]
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 24, v16
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v16
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v80, 8, v16
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v82, 16, v15
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v86, 8, v15
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v96, 24, v14
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v97, 16, v14
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v98, 8, v14
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v100, 16, v13
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v102, 8, v13
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v115, 24, v12
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v116, 16, v12
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v117, 8, v12
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v119, 16, v11
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v129, 8, v11
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v131, 24, v10
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v133, 16, v10
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v144, 8, v10
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v145, 16, v9
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v147, 8, v9
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v150, 24, v8
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v151, 16, v8
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v161, 8, v8
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v164, 16, v7
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v165, 8, v7
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v167, 24, v6
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v177, 16, v6
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v179, 8, v6
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v181, 16, v5
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v183, 8, v5
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v42, 24, v4
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v43, 16, v4
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v45, 8, v4
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v47, 16, v3
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v57, 8, v3
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v60, 24, v2
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v62, 16, v2
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v63, 8, v2
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v72, 16, v1
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v74, 8, v1
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 24, v32
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v32
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v81, 8, v32
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v83, 16, v31
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v84, 8, v31
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v85, 24, v30
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v87, 16, v30
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v99, 8, v30
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v101, 16, v29
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v103, 8, v29
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v112, 24, v28
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v113, 16, v28
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v114, 8, v28
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v118, 16, v27
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v128, 8, v27
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v130, 24, v26
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v132, 16, v26
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v134, 8, v26
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v135, 16, v25
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v146, 8, v25
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v148, 24, v24
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v149, 16, v24
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v160, 8, v24
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v162, 16, v23
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v163, 8, v23
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v166, 24, v22
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v176, 16, v22
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v178, 8, v22
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v180, 16, v21
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v182, 8, v21
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v40, 24, v20
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v41, 16, v20
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v44, 8, v20
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v46, 16, v19
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v56, 8, v19
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v58, 24, v18
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v59, 16, v18
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v61, 8, v18
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v73, 16, v17
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v75, 8, v17
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[3:4]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[17:18]
+; GFX11-FAKE16-NEXT: .LBB98_2: ; %Flow
+; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
+; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB98_4
+; GFX11-FAKE16-NEXT: ; %bb.3: ; %cmp.true
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(1)
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v32, v32, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[15:16]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[49:50], 24, v[13:14]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[11:12]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[51:52], 24, v[9:10]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[52:53], 24, v[7:8]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[1:2]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[5:6]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[19:20]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[3:4]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[17:18]
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 24, v16
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v16
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v80, 8, v16
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v82, 16, v15
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v86, 8, v15
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v96, 24, v14
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v97, 16, v14
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v98, 8, v14
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v100, 16, v13
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v102, 8, v13
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v115, 24, v12
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v116, 16, v12
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v117, 8, v12
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v119, 16, v11
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v129, 8, v11
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v131, 24, v10
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v133, 16, v10
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v144, 8, v10
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v145, 16, v9
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v147, 8, v9
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v150, 24, v8
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v151, 16, v8
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v161, 8, v8
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v164, 16, v7
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v165, 8, v7
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v167, 24, v6
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v177, 16, v6
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v179, 8, v6
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v181, 16, v5
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v183, 8, v5
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v42, 24, v4
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v43, 16, v4
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v45, 8, v4
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v47, 16, v3
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v57, 8, v3
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v60, 24, v2
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v62, 16, v2
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v63, 8, v2
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v72, 16, v1
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v74, 8, v1
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 24, v32
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v32
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v81, 8, v32
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v83, 16, v31
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v84, 8, v31
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v85, 24, v30
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v87, 16, v30
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v99, 8, v30
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v101, 16, v29
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v103, 8, v29
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v112, 24, v28
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v113, 16, v28
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v114, 8, v28
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v118, 16, v27
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v128, 8, v27
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v130, 24, v26
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v132, 16, v26
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v134, 8, v26
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v135, 16, v25
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v146, 8, v25
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v148, 24, v24
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v149, 16, v24
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v160, 8, v24
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v162, 16, v23
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v163, 8, v23
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v166, 24, v22
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v176, 16, v22
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v178, 8, v22
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v180, 16, v21
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v182, 8, v21
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v40, 24, v20
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v41, 16, v20
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v44, 8, v20
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v46, 16, v19
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v56, 8, v19
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v58, 24, v18
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v59, 16, v18
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v61, 8, v18
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v73, 16, v17
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v75, 8, v17
+; GFX11-FAKE16-NEXT: .LBB98_4: ; %end
+; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v55, 8, v74
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v39, 8, v64
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v64, 8, v60
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v54, 8, v54
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v55
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v55, 0xff, v72
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v67, 0xff, v47
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v39, v55, v39
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v55, 8, v63
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v54, v67, v54
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v67, 8, v42
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xff, v5
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v39, 16, v39
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v2, v55
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v55, 0xff, v62
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v53, 8, v53
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xff, v6
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v39
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v55, v55, v64
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v64, 8, v57
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xff, v7
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v52, 8, v52
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xff, v8
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v39, 16, v55
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v3, v64
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v55, 8, v45
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v64, 0xff, v43
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v51, 8, v51
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v2, v39
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v39, 16, v54
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v4, v55
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v54, v64, v67
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v55, 8, v183
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v64, 0xff, v181
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v3, v39
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v39, 16, v54
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v5, v55
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v53, v64, v53
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v54, 8, v179
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v55, 0xff, v177
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v64, 8, v167
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v4, v39
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v39, 16, v53
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v6, v6, v54
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v53, v55, v64
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v54, 8, v165
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v55, 0xff, v164
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v5, v39
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v39, 16, v53
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v7, v54
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v52, v55, v52
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v53, 8, v161
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v54, 0xff, v151
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v55, 8, v150
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v6, v6, v39
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v39, 16, v52
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v8, v8, v53
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v52, v54, v55
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v54, 0xff, v145
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v9, 0xff, v9
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v53, 8, v147
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xff, v10
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v55, 8, v144
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xffff, v8
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v52, 16, v52
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v51, v54, v51
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v9, v9, v53
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v10, v10, v55
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v7, v39
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v8, v8, v52
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v39, 16, v51
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v51, 0xff, v133
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v52, 8, v131
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xff, v11
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v53, 8, v129
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v54, 0xff, v119
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v50, 8, v50
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xff, v12
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v55, 8, v117
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v64, 0xff, v116
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v67, 8, v115
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v51, v51, v52
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v11, v11, v53
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v50, v54, v50
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v12, v12, v55
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v52, v64, v67
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v9, 0xffff, v9
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v50, 16, v50
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v52, 16, v52
+; GFX11-FAKE16-NEXT: s_clause 0x1
+; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[1:4], off
+; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v9, v39
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v10, v51
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v11, v50
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v12, v52
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xff, v13
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v6, 8, v102
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xff, v100
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v8, 8, v49
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v9, 0xff, v14
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v10, 8, v98
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xff, v97
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v12, 8, v96
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xff, v15
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v14, 8, v86
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v5, v6
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v6, v7, v8
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v9, v10
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v8, v11, v12
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v9, v13, v14
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xff, v82
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v11, 8, v48
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xff, v16
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v13, 8, v80
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xff, v70
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v15, 8, v68
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v16, 0xff, v17
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v17, 8, v75
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v39, 0xff, v73
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v48, 8, v66
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v10, v10, v11
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v11, v12, v13
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v12, v14, v15
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v13, v16, v17
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v14, v39, v48
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v9, 0xffff, v9
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xffff, v13
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v5, v6
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v6, v7, v8
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v9, v10
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v8, v11, v12
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v9, v13, v14
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xff, v18
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v11, 8, v61
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xff, v59
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v13, 8, v58
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xff, v19
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v15, 8, v56
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v16, 0xff, v46
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v17, 8, v65
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xff, v20
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v19, 8, v44
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v10, v10, v11
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v11, v12, v13
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v12, v14, v15
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v13, v16, v17
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v14, v18, v19
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xff, v41
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v16, 8, v40
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v17, 0xff, v21
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v18, 8, v182
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xff, v180
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v20, 8, v38
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xff, v22
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v22, 8, v178
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v38, 0xff, v176
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v39, 8, v166
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v15, v15, v16
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v16, v17, v18
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v17, v19, v20
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v18, v21, v22
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v19, v38, v39
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xffff, v14
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v15, 16, v15
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v16, 0xffff, v16
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xffff, v18
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v19, 16, v19
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v10, v10, v11
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v11, v12, v13
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v12, v14, v15
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v13, v16, v17
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v14, v18, v19
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xff, v23
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v16, 8, v163
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v17, 0xff, v162
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v18, 8, v37
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xff, v24
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v20, 8, v160
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xff, v149
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v22, 8, v148
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v23, 0xff, v25
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v24, 8, v146
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v15, v15, v16
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v16, v17, v18
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v17, v19, v20
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v18, v21, v22
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v19, v23, v24
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xff, v135
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v21, 8, v36
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xff, v26
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v23, 8, v134
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v24, 0xff, v132
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v25, 8, v130
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v26, 0xff, v27
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v27, 8, v128
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v36, 0xff, v118
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v35, 8, v35
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v20, v20, v21
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v21, v22, v23
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v22, v24, v25
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v23, v26, v27
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v24, v36, v35
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xffff, v15
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v17, 0xffff, v17
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xffff, v19
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v20, 16, v20
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v21
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v23, 0xffff, v23
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v15, v15, v16
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v16, v17, v18
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v17, v19, v20
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v18, v21, v22
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v19, v23, v24
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xff, v28
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v21, 8, v114
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xff, v113
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v23, 8, v112
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v24, 0xff, v29
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v25, 8, v103
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v26, 0xff, v101
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v27, 8, v34
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v28, 0xff, v30
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v29, 8, v99
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v20, v20, v21
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v21, v22, v23
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v22, v24, v25
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v23, v26, v27
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v24, v28, v29
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v25, 0xff, v87
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v26, 8, v85
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v27, 0xff, v31
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v28, 8, v84
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v29, 0xff, v83
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v30, 8, v33
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v31, 0xff, v32
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v32, 8, v81
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xff, v71
+; GFX11-FAKE16-NEXT: v_lshlrev_b16 v34, 8, v69
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v25, v25, v26
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v26, v27, v28
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v27, v29, v30
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v28, v31, v32
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v29, v33, v34
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xffff, v20
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xffff, v22
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v24, 0xffff, v24
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v25, 16, v25
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v26, 0xffff, v26
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v27, 16, v27
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v28, 0xffff, v28
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v29, 16, v29
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v20, v20, v21
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v21, v22, v23
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v22, v24, v25
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v23, v26, v27
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v24, v28, v29
+; GFX11-FAKE16-NEXT: s_clause 0x5
+; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[1:4], off offset:32
+; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:48
+; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:64
+; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:80
+; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:96
+; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:112
+; GFX11-FAKE16-NEXT: s_clause 0x13
+; GFX11-FAKE16-NEXT: scratch_load_b32 v75, off, s32 offset:12
+; GFX11-FAKE16-NEXT: scratch_load_b32 v74, off, s32 offset:16
+; GFX11-FAKE16-NEXT: scratch_load_b32 v73, off, s32 offset:20
+; GFX11-FAKE16-NEXT: scratch_load_b32 v72, off, s32 offset:24
+; GFX11-FAKE16-NEXT: scratch_load_b32 v63, off, s32 offset:28
+; GFX11-FAKE16-NEXT: scratch_load_b32 v62, off, s32 offset:32
+; GFX11-FAKE16-NEXT: scratch_load_b32 v61, off, s32 offset:36
+; GFX11-FAKE16-NEXT: scratch_load_b32 v60, off, s32 offset:40
+; GFX11-FAKE16-NEXT: scratch_load_b32 v59, off, s32 offset:44
+; GFX11-FAKE16-NEXT: scratch_load_b32 v58, off, s32 offset:48
+; GFX11-FAKE16-NEXT: scratch_load_b32 v57, off, s32 offset:52
+; GFX11-FAKE16-NEXT: scratch_load_b32 v56, off, s32 offset:56
+; GFX11-FAKE16-NEXT: scratch_load_b32 v47, off, s32 offset:60
+; GFX11-FAKE16-NEXT: scratch_load_b32 v46, off, s32 offset:64
+; GFX11-FAKE16-NEXT: scratch_load_b32 v45, off, s32 offset:68
+; GFX11-FAKE16-NEXT: scratch_load_b32 v44, off, s32 offset:72
+; GFX11-FAKE16-NEXT: scratch_load_b32 v43, off, s32 offset:76
+; GFX11-FAKE16-NEXT: scratch_load_b32 v42, off, s32 offset:80
+; GFX11-FAKE16-NEXT: scratch_load_b32 v41, off, s32 offset:84
+; GFX11-FAKE16-NEXT: scratch_load_b32 v40, off, s32 offset:88
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-FAKE16-NEXT: 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 inreg <128 x i8> @bitcast_v64i16_to_v128i8_scalar(<64 x i16> inreg %a, i32 inreg %b) {
+; SI-LABEL: bitcast_v64i16_to_v128i8_scalar:
+; SI: ; %bb.0:
+; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; SI-NEXT: s_or_saveexec_b64 s[4:5], -1
+; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT: s_mov_b64 exec, s[4:5]
+; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
+; SI-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:76
+; SI-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:72
+; SI-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:60
+; SI-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:56
+; SI-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:44
+; SI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:40
+; SI-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:28
+; SI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:24
+; SI-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:12
+; SI-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:8
+; SI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:80
+; SI-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:36
+; SI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:32
+; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:20
+; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:4
+; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:68
+; SI-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:64
+; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:52
+; SI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:48
+; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:16
+; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32
+; SI-NEXT: v_writelane_b32 v63, s30, 0
+; SI-NEXT: v_writelane_b32 v63, s31, 1
+; SI-NEXT: v_writelane_b32 v63, s34, 2
+; SI-NEXT: v_writelane_b32 v63, s35, 3
+; SI-NEXT: v_writelane_b32 v63, s36, 4
+; SI-NEXT: v_writelane_b32 v63, s37, 5
+; SI-NEXT: v_writelane_b32 v63, s38, 6
+; SI-NEXT: v_writelane_b32 v63, s39, 7
+; SI-NEXT: v_writelane_b32 v63, s48, 8
+; SI-NEXT: v_writelane_b32 v63, s49, 9
+; SI-NEXT: v_writelane_b32 v63, s50, 10
+; SI-NEXT: v_writelane_b32 v63, s51, 11
+; SI-NEXT: v_writelane_b32 v63, s52, 12
+; SI-NEXT: v_writelane_b32 v63, s53, 13
+; SI-NEXT: v_writelane_b32 v63, s54, 14
+; SI-NEXT: v_writelane_b32 v63, s55, 15
+; SI-NEXT: v_writelane_b32 v63, s64, 16
+; SI-NEXT: v_writelane_b32 v63, s65, 17
+; SI-NEXT: v_writelane_b32 v63, s66, 18
+; SI-NEXT: v_writelane_b32 v63, s67, 19
+; SI-NEXT: v_writelane_b32 v63, s68, 20
+; SI-NEXT: v_writelane_b32 v63, s69, 21
+; SI-NEXT: v_writelane_b32 v63, s70, 22
+; SI-NEXT: v_writelane_b32 v63, s71, 23
+; SI-NEXT: v_writelane_b32 v63, s80, 24
+; SI-NEXT: v_writelane_b32 v63, s81, 25
+; SI-NEXT: v_writelane_b32 v63, s82, 26
+; SI-NEXT: v_writelane_b32 v63, s83, 27
+; SI-NEXT: v_writelane_b32 v63, s84, 28
+; SI-NEXT: v_writelane_b32 v63, s85, 29
+; SI-NEXT: v_writelane_b32 v63, s86, 30
+; SI-NEXT: v_writelane_b32 v63, s87, 31
+; SI-NEXT: v_writelane_b32 v63, s96, 32
+; SI-NEXT: v_writelane_b32 v63, s97, 33
+; SI-NEXT: v_writelane_b32 v63, s98, 34
+; SI-NEXT: s_mov_b32 s6, s18
+; SI-NEXT: ; implicit-def: $vgpr62 : SGPR spill to VGPR lane
+; SI-NEXT: v_writelane_b32 v63, s99, 35
+; SI-NEXT: v_readfirstlane_b32 s62, v30
+; SI-NEXT: v_readfirstlane_b32 s63, v29
+; SI-NEXT: v_readfirstlane_b32 s59, v26
+; SI-NEXT: v_readfirstlane_b32 s60, v25
+; SI-NEXT: v_readfirstlane_b32 s98, v22
+; SI-NEXT: v_readfirstlane_b32 s61, v21
+; SI-NEXT: v_readfirstlane_b32 s99, v18
+; SI-NEXT: v_readfirstlane_b32 s58, v17
+; SI-NEXT: v_readfirstlane_b32 s96, v14
+; SI-NEXT: v_readfirstlane_b32 s97, v13
+; SI-NEXT: v_readfirstlane_b32 s86, v10
+; SI-NEXT: v_readfirstlane_b32 s87, v9
+; SI-NEXT: v_readfirstlane_b32 s84, v6
+; SI-NEXT: v_readfirstlane_b32 s85, v5
+; SI-NEXT: v_readfirstlane_b32 s81, v2
+; SI-NEXT: v_readfirstlane_b32 s82, v1
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v12
+; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v16
+; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v20
+; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v24
+; SI-NEXT: s_waitcnt vmcnt(14)
+; SI-NEXT: v_readfirstlane_b32 s88, v36
+; SI-NEXT: v_readfirstlane_b32 s18, v37
+; SI-NEXT: v_readfirstlane_b32 s78, v38
+; SI-NEXT: v_readfirstlane_b32 s79, v39
+; SI-NEXT: v_readfirstlane_b32 s76, v48
+; SI-NEXT: v_readfirstlane_b32 s77, v49
+; SI-NEXT: v_readfirstlane_b32 s74, v50
+; SI-NEXT: s_waitcnt vmcnt(13)
+; SI-NEXT: v_readfirstlane_b32 s75, v51
+; SI-NEXT: s_waitcnt vmcnt(12)
+; SI-NEXT: v_readfirstlane_b32 s72, v52
+; SI-NEXT: s_waitcnt vmcnt(11)
+; SI-NEXT: v_readfirstlane_b32 s73, v53
+; SI-NEXT: s_waitcnt vmcnt(10)
+; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v54
+; SI-NEXT: s_and_b64 s[4:5], vcc, exec
+; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v28
+; SI-NEXT: s_waitcnt vmcnt(7)
+; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v40
+; SI-NEXT: s_waitcnt vmcnt(6)
+; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v41
+; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v55
+; SI-NEXT: s_waitcnt vmcnt(5)
+; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v42
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v43
+; SI-NEXT: v_writelane_b32 v62, s6, 0
+; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT: s_cbranch_scc0 .LBB99_4
+; SI-NEXT: ; %bb.1: ; %cmp.false
+; SI-NEXT: s_and_b32 s4, s16, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s17, 16
+; SI-NEXT: s_or_b32 s56, s4, s5
+; SI-NEXT: s_and_b32 s4, s6, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s19, 16
+; SI-NEXT: s_or_b32 s57, s4, s5
+; SI-NEXT: v_mov_b32_e32 v1, s56
+; SI-NEXT: s_and_b32 s4, s20, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s21, 16
+; SI-NEXT: v_alignbit_b32 v8, s57, v1, 24
+; SI-NEXT: v_alignbit_b32 v50, s57, v1, 16
+; SI-NEXT: v_alignbit_b32 v1, s57, v1, 8
+; SI-NEXT: s_or_b32 s46, s4, s5
+; SI-NEXT: s_and_b32 s4, s22, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s23, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT: s_or_b32 s47, s4, s5
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_mov_b32_e32 v1, s46
+; SI-NEXT: s_and_b32 s4, s24, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s25, 16
+; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v8, s47, v1, 24
+; SI-NEXT: s_or_b32 s44, s4, s5
+; SI-NEXT: s_and_b32 s4, s26, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s27, 16
+; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v8, s47, v1, 16
+; SI-NEXT: v_alignbit_b32 v51, s47, v1, 8
+; SI-NEXT: s_or_b32 s45, s4, s5
+; SI-NEXT: v_mov_b32_e32 v1, s44
+; SI-NEXT: s_and_b32 s4, s28, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s29, 16
+; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v8, s45, v1, 24
+; SI-NEXT: s_or_b32 s42, s4, s5
+; SI-NEXT: s_and_b32 s4, s82, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s81, 16
+; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v8, s45, v1, 16
+; SI-NEXT: v_alignbit_b32 v49, s45, v1, 8
+; SI-NEXT: s_or_b32 s43, s4, s5
+; SI-NEXT: v_mov_b32_e32 v1, s42
+; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v8, s43, v1, 24
+; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v8, s43, v1, 16
+; SI-NEXT: v_alignbit_b32 v48, s43, v1, 8
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v3
+; SI-NEXT: s_and_b32 s4, s85, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s84, 16
+; SI-NEXT: v_or_b32_e32 v16, v1, v2
+; SI-NEXT: s_or_b32 s41, s4, s5
+; SI-NEXT: v_alignbit_b32 v1, s41, v16, 24
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, s41, v16, 16
+; SI-NEXT: s_and_b32 s4, s87, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s86, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, s41, v16, 8
+; SI-NEXT: s_or_b32 s40, s4, s5
+; SI-NEXT: s_and_b32 s4, s97, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s96, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v7
+; SI-NEXT: s_or_b32 s15, s4, s5
+; SI-NEXT: s_and_b32 s4, s58, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s99, 16
+; SI-NEXT: v_or_b32_e32 v14, v1, v4
+; SI-NEXT: s_or_b32 s14, s4, s5
+; SI-NEXT: s_and_b32 s4, s61, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s98, 16
+; SI-NEXT: v_alignbit_b32 v1, s40, v14, 24
+; SI-NEXT: s_or_b32 s13, s4, s5
+; SI-NEXT: s_and_b32 s4, s60, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s59, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, s40, v14, 16
+; SI-NEXT: s_or_b32 s12, s4, s5
+; SI-NEXT: s_and_b32 s4, s63, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s62, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, s40, v14, 8
+; SI-NEXT: s_or_b32 s11, s4, s5
+; SI-NEXT: s_and_b32 s4, s73, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s72, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v11
+; SI-NEXT: s_or_b32 s10, s4, s5
+; SI-NEXT: s_and_b32 s4, s75, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s74, 16
+; SI-NEXT: v_or_b32_e32 v12, v1, v5
+; SI-NEXT: s_or_b32 s9, s4, s5
+; SI-NEXT: s_and_b32 s4, s77, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s76, 16
+; SI-NEXT: v_alignbit_b32 v1, s15, v12, 24
+; SI-NEXT: s_or_b32 s8, s4, s5
+; SI-NEXT: s_and_b32 s4, s79, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s78, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, s15, v12, 16
+; SI-NEXT: s_or_b32 s7, s4, s5
+; SI-NEXT: s_and_b32 s4, s18, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s88, 16
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, s15, v12, 8
+; SI-NEXT: s_or_b32 s6, s4, s5
+; SI-NEXT: s_lshr_b32 s4, s11, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v15
+; SI-NEXT: v_writelane_b32 v62, s4, 1
+; SI-NEXT: s_lshr_b32 s4, s10, 8
+; SI-NEXT: v_or_b32_e32 v10, v1, v6
+; SI-NEXT: v_writelane_b32 v62, s4, 3
+; SI-NEXT: s_lshr_b32 s4, s9, 8
+; SI-NEXT: v_alignbit_b32 v1, s14, v10, 24
+; SI-NEXT: v_writelane_b32 v62, s4, 6
+; SI-NEXT: s_lshr_b32 s4, s8, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, s14, v10, 16
+; SI-NEXT: v_writelane_b32 v62, s4, 9
+; SI-NEXT: s_lshr_b32 s4, s7, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v1, s14, v10, 8
+; SI-NEXT: v_writelane_b32 v62, s4, 12
+; SI-NEXT: s_lshr_b32 s4, s6, 8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v19
+; SI-NEXT: v_writelane_b32 v62, s4, 15
+; SI-NEXT: s_and_b32 s4, s72, 0xffff
+; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_or_b32_e32 v8, v1, v9
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v23
+; SI-NEXT: v_writelane_b32 v62, s4, 2
+; SI-NEXT: s_and_b32 s4, s74, 0xffff
+; SI-NEXT: v_or_b32_e32 v5, v1, v13
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v27
+; SI-NEXT: v_writelane_b32 v62, s4, 5
+; SI-NEXT: s_and_b32 s4, s76, 0xffff
+; SI-NEXT: v_mov_b32_e32 v28, v13
+; SI-NEXT: v_or_b32_e32 v13, v1, v17
+; SI-NEXT: s_waitcnt vmcnt(14)
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v31
+; SI-NEXT: v_writelane_b32 v62, s4, 8
+; SI-NEXT: s_and_b32 s4, s78, 0xffff
+; SI-NEXT: v_mov_b32_e32 v26, v9
+; SI-NEXT: v_or_b32_e32 v9, v1, v18
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v32
+; SI-NEXT: v_writelane_b32 v62, s4, 11
+; SI-NEXT: s_and_b32 s4, s88, 0xffff
+; SI-NEXT: v_mov_b32_e32 v25, v6
+; SI-NEXT: v_or_b32_e32 v6, v1, v20
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v33
+; SI-NEXT: v_writelane_b32 v62, s4, 14
+; SI-NEXT: s_bfe_u32 s4, s74, 0x80008
+; SI-NEXT: v_or_b32_e32 v4, v1, v21
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v34
+; SI-NEXT: v_writelane_b32 v62, s4, 4
+; SI-NEXT: s_bfe_u32 s4, s76, 0x80008
+; SI-NEXT: v_or_b32_e32 v2, v1, v22
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v35
+; SI-NEXT: v_writelane_b32 v62, s4, 7
+; SI-NEXT: s_bfe_u32 s4, s78, 0x80008
+; SI-NEXT: v_or_b32_e32 v1, v1, v24
+; SI-NEXT: v_writelane_b32 v62, s4, 10
+; SI-NEXT: s_bfe_u32 s4, s88, 0x80008
+; SI-NEXT: v_mov_b32_e32 v29, v17
+; SI-NEXT: v_mov_b32_e32 v30, v18
+; SI-NEXT: v_mov_b32_e32 v36, v20
+; SI-NEXT: v_mov_b32_e32 v37, v21
+; SI-NEXT: v_mov_b32_e32 v38, v22
+; SI-NEXT: v_mov_b32_e32 v39, v24
+; SI-NEXT: s_lshr_b32 s68, s57, 8
+; SI-NEXT: s_lshr_b32 s65, s47, 8
+; SI-NEXT: s_lshr_b32 s54, s45, 8
+; SI-NEXT: s_lshr_b32 s51, s43, 8
+; SI-NEXT: s_lshr_b32 s48, s41, 8
+; SI-NEXT: s_lshr_b32 s37, s40, 8
+; SI-NEXT: s_lshr_b32 s34, s15, 8
+; SI-NEXT: s_lshr_b32 s95, s14, 8
+; SI-NEXT: s_lshr_b32 s92, s13, 8
+; SI-NEXT: s_lshr_b32 s89, s12, 8
+; SI-NEXT: s_and_b32 s71, s19, 0xffff
+; SI-NEXT: s_and_b32 s69, s23, 0xffff
+; SI-NEXT: s_and_b32 s66, s27, 0xffff
+; SI-NEXT: s_and_b32 s55, s81, 0xffff
+; SI-NEXT: s_and_b32 s52, s84, 0xffff
+; SI-NEXT: s_and_b32 s49, s86, 0xffff
+; SI-NEXT: s_and_b32 s38, s96, 0xffff
+; SI-NEXT: s_and_b32 s35, s99, 0xffff
+; SI-NEXT: s_and_b32 s30, s98, 0xffff
+; SI-NEXT: s_and_b32 s93, s59, 0xffff
+; SI-NEXT: s_and_b32 s90, s62, 0xffff
+; SI-NEXT: s_bfe_u32 s83, s19, 0x80008
+; SI-NEXT: s_bfe_u32 s80, s23, 0x80008
+; SI-NEXT: s_bfe_u32 s70, s27, 0x80008
+; SI-NEXT: s_bfe_u32 s67, s81, 0x80008
+; SI-NEXT: s_bfe_u32 s64, s84, 0x80008
+; SI-NEXT: s_bfe_u32 s53, s86, 0x80008
+; SI-NEXT: s_bfe_u32 s50, s96, 0x80008
+; SI-NEXT: s_bfe_u32 s39, s99, 0x80008
+; SI-NEXT: s_bfe_u32 s36, s98, 0x80008
+; SI-NEXT: s_bfe_u32 s31, s59, 0x80008
+; SI-NEXT: s_bfe_u32 s94, s62, 0x80008
+; SI-NEXT: s_bfe_u32 s91, s72, 0x80008
+; SI-NEXT: v_writelane_b32 v62, s4, 13
+; SI-NEXT: v_alignbit_b32 v45, s13, v8, 24
+; SI-NEXT: v_alignbit_b32 v47, s13, v8, 16
+; SI-NEXT: v_alignbit_b32 v57, s13, v8, 8
+; SI-NEXT: v_alignbit_b32 v41, s12, v5, 24
+; SI-NEXT: v_alignbit_b32 v43, s12, v5, 16
+; SI-NEXT: v_alignbit_b32 v44, s12, v5, 8
+; SI-NEXT: v_alignbit_b32 v21, s11, v13, 24
+; SI-NEXT: v_alignbit_b32 v22, s11, v13, 16
+; SI-NEXT: v_alignbit_b32 v24, s11, v13, 8
+; SI-NEXT: v_alignbit_b32 v17, s10, v9, 24
+; SI-NEXT: v_alignbit_b32 v18, s10, v9, 16
+; SI-NEXT: v_alignbit_b32 v20, s10, v9, 8
+; SI-NEXT: v_alignbit_b32 v59, s9, v6, 24
+; SI-NEXT: v_alignbit_b32 v60, s9, v6, 16
+; SI-NEXT: v_alignbit_b32 v61, s9, v6, 8
+; SI-NEXT: v_alignbit_b32 v46, s8, v4, 24
+; SI-NEXT: v_alignbit_b32 v56, s8, v4, 16
+; SI-NEXT: v_alignbit_b32 v58, s8, v4, 8
+; SI-NEXT: v_alignbit_b32 v55, s7, v2, 24
+; SI-NEXT: v_alignbit_b32 v40, s7, v2, 16
+; SI-NEXT: v_alignbit_b32 v42, s7, v2, 8
+; SI-NEXT: v_alignbit_b32 v52, s6, v1, 24
+; SI-NEXT: v_alignbit_b32 v53, s6, v1, 16
+; SI-NEXT: v_alignbit_b32 v54, s6, v1, 8
+; SI-NEXT: s_cbranch_execnz .LBB99_3
+; SI-NEXT: .LBB99_2: ; %cmp.true
+; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; SI-NEXT: s_add_i32 s18, s18, 3
+; SI-NEXT: s_and_b32 s4, s18, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s88, 16
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s79, s79, 3
+; SI-NEXT: s_add_i32 s6, s4, 0x30000
+; SI-NEXT: s_and_b32 s4, s79, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s78, 16
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s77, s77, 3
+; SI-NEXT: s_add_i32 s7, s4, 0x30000
+; SI-NEXT: s_and_b32 s4, s77, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s76, 16
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s75, s75, 3
+; SI-NEXT: s_add_i32 s8, s4, 0x30000
+; SI-NEXT: s_and_b32 s4, s75, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s74, 16
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s73, s73, 3
+; SI-NEXT: s_add_i32 s9, s4, 0x30000
+; SI-NEXT: s_and_b32 s4, s73, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s72, 16
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s63, s63, 3
+; SI-NEXT: s_add_i32 s10, s4, 0x30000
+; SI-NEXT: s_and_b32 s4, s63, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s62, 16
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s60, s60, 3
+; SI-NEXT: s_add_i32 s11, s4, 0x30000
+; SI-NEXT: s_and_b32 s4, s60, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s59, 16
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s61, s61, 3
+; SI-NEXT: s_add_i32 s12, s4, 0x30000
+; SI-NEXT: s_and_b32 s4, s61, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s98, 16
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s58, s58, 3
+; SI-NEXT: s_add_i32 s13, s4, 0x30000
+; SI-NEXT: s_and_b32 s4, s58, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s99, 16
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s97, s97, 3
+; SI-NEXT: s_add_i32 s14, s4, 0x30000
+; SI-NEXT: s_and_b32 s4, s97, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s96, 16
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s87, s87, 3
+; SI-NEXT: s_add_i32 s15, s4, 0x30000
+; SI-NEXT: s_and_b32 s4, s87, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s86, 16
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s85, s85, 3
+; SI-NEXT: s_add_i32 s40, s4, 0x30000
+; SI-NEXT: s_and_b32 s4, s85, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s84, 16
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s28, s28, 3
+; SI-NEXT: s_add_i32 s41, s4, 0x30000
+; SI-NEXT: s_and_b32 s4, s28, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s29, 16
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s82, s82, 3
+; SI-NEXT: s_add_i32 s42, s4, 0x30000
+; SI-NEXT: s_and_b32 s4, s82, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s81, 16
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s24, s24, 3
+; SI-NEXT: s_add_i32 s43, s4, 0x30000
+; SI-NEXT: s_and_b32 s4, s24, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s25, 16
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s26, s26, 3
+; SI-NEXT: s_add_i32 s44, s4, 0x30000
+; SI-NEXT: s_and_b32 s4, s26, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s27, 16
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s20, s20, 3
+; SI-NEXT: s_add_i32 s45, s4, 0x30000
+; SI-NEXT: s_and_b32 s4, s20, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s21, 16
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s22, s22, 3
+; SI-NEXT: s_add_i32 s46, s4, 0x30000
+; SI-NEXT: s_and_b32 s4, s22, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s23, 16
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s16, s16, 3
+; SI-NEXT: s_add_i32 s47, s4, 0x30000
+; SI-NEXT: s_and_b32 s4, s16, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s17, 16
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s56, s4, 0x30000
+; SI-NEXT: v_readlane_b32 s4, v62, 0
+; SI-NEXT: s_add_i32 s4, s4, 3
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_lshl_b32 s5, s19, 16
+; SI-NEXT: s_or_b32 s4, s5, s4
+; SI-NEXT: s_add_i32 s57, s4, 0x30000
+; SI-NEXT: v_mov_b32_e32 v21, s56
+; SI-NEXT: v_alignbit_b32 v22, s57, v21, 24
+; SI-NEXT: v_alignbit_b32 v50, s57, v21, 16
+; SI-NEXT: v_alignbit_b32 v21, s57, v21, 8
+; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_mov_b32_e32 v21, s46
+; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v22, s47, v21, 24
+; SI-NEXT: s_lshr_b32 s4, s11, 8
+; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v3
+; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v22, s47, v21, 16
+; SI-NEXT: v_alignbit_b32 v51, s47, v21, 8
+; SI-NEXT: v_mov_b32_e32 v21, s44
+; SI-NEXT: v_writelane_b32 v62, s4, 1
+; SI-NEXT: s_lshr_b32 s4, s10, 16
+; SI-NEXT: s_waitcnt vmcnt(10)
+; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v32
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v22, s45, v21, 24
+; SI-NEXT: v_writelane_b32 v62, s4, 2
+; SI-NEXT: s_lshr_b32 s4, s10, 8
+; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v7
+; SI-NEXT: s_waitcnt vmcnt(4)
+; SI-NEXT: v_or_b32_e32 v3, v16, v3
+; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v22, s45, v21, 16
+; SI-NEXT: v_alignbit_b32 v49, s45, v21, 8
+; SI-NEXT: v_mov_b32_e32 v21, s42
+; SI-NEXT: v_writelane_b32 v62, s4, 3
+; SI-NEXT: s_lshr_b32 s4, s9, 24
+; SI-NEXT: v_or_b32_e32 v5, v36, v5
+; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
+; SI-NEXT: v_add_i32_e32 v16, vcc, 0x30000, v3
+; SI-NEXT: v_mov_b32_e32 v3, s41
+; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v22, s43, v21, 24
+; SI-NEXT: v_writelane_b32 v62, s4, 4
+; SI-NEXT: s_lshr_b32 s4, s9, 16
+; SI-NEXT: v_add_i32_e32 v6, vcc, 0x30000, v5
+; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v31
+; SI-NEXT: v_or_b32_e32 v7, v14, v7
+; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v22, s43, v21, 16
+; SI-NEXT: v_alignbit_b32 v48, s43, v21, 8
+; SI-NEXT: v_alignbit_b32 v21, v3, v16, 24
+; SI-NEXT: v_writelane_b32 v62, s4, 5
+; SI-NEXT: s_lshr_b32 s4, s9, 8
+; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11
+; SI-NEXT: v_add_i32_e32 v14, vcc, 0x30000, v7
+; SI-NEXT: v_mov_b32_e32 v7, s40
+; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v21, v3, v16, 16
+; SI-NEXT: v_alignbit_b32 v3, v3, v16, 8
+; SI-NEXT: v_writelane_b32 v62, s4, 6
+; SI-NEXT: s_lshr_b32 s4, s8, 24
+; SI-NEXT: v_or_b32_e32 v5, v30, v5
+; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v3, v7, v14, 24
+; SI-NEXT: v_writelane_b32 v62, s4, 7
+; SI-NEXT: s_lshr_b32 s4, s8, 16
+; SI-NEXT: v_add_i32_e32 v9, vcc, 0x30000, v5
+; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v27
+; SI-NEXT: v_or_b32_e32 v11, v12, v11
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v3, v7, v14, 16
+; SI-NEXT: v_writelane_b32 v62, s4, 8
+; SI-NEXT: s_lshr_b32 s4, s8, 8
+; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v15
+; SI-NEXT: v_add_i32_e32 v12, vcc, 0x30000, v11
+; SI-NEXT: v_mov_b32_e32 v11, s15
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v3, v7, v14, 8
+; SI-NEXT: v_writelane_b32 v62, s4, 9
+; SI-NEXT: s_lshr_b32 s4, s7, 24
+; SI-NEXT: v_or_b32_e32 v5, v29, v5
+; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v3, v11, v12, 24
+; SI-NEXT: v_writelane_b32 v62, s4, 10
+; SI-NEXT: s_lshr_b32 s4, s7, 16
+; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v35
+; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v34
+; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v33
+; SI-NEXT: v_add_i32_e32 v13, vcc, 0x30000, v5
+; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v23
+; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v19
+; SI-NEXT: v_or_b32_e32 v10, v25, v10
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v3, v11, v12, 16
+; SI-NEXT: v_writelane_b32 v62, s4, 11
+; SI-NEXT: s_lshr_b32 s4, s7, 8
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
+; SI-NEXT: v_add_i32_e32 v10, vcc, 0x30000, v10
+; SI-NEXT: v_mov_b32_e32 v15, s14
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v3, v11, v12, 8
+; SI-NEXT: v_writelane_b32 v62, s4, 12
+; SI-NEXT: s_lshr_b32 s4, s6, 24
+; SI-NEXT: v_or_b32_e32 v1, v39, v1
+; SI-NEXT: v_or_b32_e32 v2, v38, v2
+; SI-NEXT: v_or_b32_e32 v4, v37, v4
+; SI-NEXT: v_or_b32_e32 v5, v28, v5
+; SI-NEXT: v_or_b32_e32 v8, v26, v8
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v3, v15, v10, 24
+; SI-NEXT: v_writelane_b32 v62, s4, 13
+; SI-NEXT: s_lshr_b32 s4, s6, 16
+; SI-NEXT: v_add_i32_e32 v1, vcc, 0x30000, v1
+; SI-NEXT: v_mov_b32_e32 v35, s6
+; SI-NEXT: v_add_i32_e32 v2, vcc, 0x30000, v2
+; SI-NEXT: v_mov_b32_e32 v34, s7
+; SI-NEXT: v_add_i32_e32 v4, vcc, 0x30000, v4
+; SI-NEXT: v_mov_b32_e32 v33, s8
+; SI-NEXT: v_mov_b32_e32 v32, s9
+; SI-NEXT: v_mov_b32_e32 v20, s10
+; SI-NEXT: v_mov_b32_e32 v17, s11
+; SI-NEXT: v_add_i32_e32 v5, vcc, 0x30000, v5
+; SI-NEXT: v_mov_b32_e32 v18, s12
+; SI-NEXT: v_add_i32_e32 v8, vcc, 0x30000, v8
+; SI-NEXT: v_mov_b32_e32 v19, s13
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v3, v15, v10, 16
+; SI-NEXT: v_writelane_b32 v62, s4, 14
+; SI-NEXT: s_lshr_b32 s4, s6, 8
+; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_alignbit_b32 v3, v15, v10, 8
+; SI-NEXT: v_alignbit_b32 v45, v19, v8, 24
+; SI-NEXT: v_alignbit_b32 v47, v19, v8, 16
+; SI-NEXT: v_alignbit_b32 v57, v19, v8, 8
+; SI-NEXT: v_alignbit_b32 v41, v18, v5, 24
+; SI-NEXT: v_alignbit_b32 v43, v18, v5, 16
+; SI-NEXT: v_alignbit_b32 v44, v18, v5, 8
+; SI-NEXT: v_alignbit_b32 v21, v17, v13, 24
+; SI-NEXT: v_alignbit_b32 v22, v17, v13, 16
+; SI-NEXT: v_alignbit_b32 v24, v17, v13, 8
+; SI-NEXT: v_alignbit_b32 v17, v20, v9, 24
+; SI-NEXT: v_alignbit_b32 v18, v20, v9, 16
+; SI-NEXT: v_alignbit_b32 v20, v20, v9, 8
+; SI-NEXT: v_alignbit_b32 v59, v32, v6, 24
+; SI-NEXT: v_alignbit_b32 v60, v32, v6, 16
+; SI-NEXT: v_alignbit_b32 v61, v32, v6, 8
+; SI-NEXT: v_alignbit_b32 v46, v33, v4, 24
+; SI-NEXT: v_alignbit_b32 v56, v33, v4, 16
+; SI-NEXT: v_alignbit_b32 v58, v33, v4, 8
+; SI-NEXT: v_alignbit_b32 v55, v34, v2, 24
+; SI-NEXT: v_alignbit_b32 v40, v34, v2, 16
+; SI-NEXT: v_alignbit_b32 v42, v34, v2, 8
+; SI-NEXT: v_alignbit_b32 v52, v35, v1, 24
+; SI-NEXT: v_alignbit_b32 v53, v35, v1, 16
+; SI-NEXT: v_alignbit_b32 v54, v35, v1, 8
+; SI-NEXT: s_lshr_b32 s83, s57, 24
+; SI-NEXT: s_lshr_b32 s71, s57, 16
+; SI-NEXT: s_lshr_b32 s68, s57, 8
+; SI-NEXT: s_lshr_b32 s80, s47, 24
+; SI-NEXT: s_lshr_b32 s69, s47, 16
+; SI-NEXT: s_lshr_b32 s65, s47, 8
+; SI-NEXT: s_lshr_b32 s70, s45, 24
+; SI-NEXT: s_lshr_b32 s66, s45, 16
+; SI-NEXT: s_lshr_b32 s54, s45, 8
+; SI-NEXT: s_lshr_b32 s67, s43, 24
+; SI-NEXT: s_lshr_b32 s55, s43, 16
+; SI-NEXT: s_lshr_b32 s51, s43, 8
+; SI-NEXT: s_lshr_b32 s64, s41, 24
+; SI-NEXT: s_lshr_b32 s52, s41, 16
+; SI-NEXT: s_lshr_b32 s48, s41, 8
+; SI-NEXT: s_lshr_b32 s53, s40, 24
+; SI-NEXT: s_lshr_b32 s49, s40, 16
+; SI-NEXT: s_lshr_b32 s37, s40, 8
+; SI-NEXT: s_lshr_b32 s50, s15, 24
+; SI-NEXT: s_lshr_b32 s38, s15, 16
+; SI-NEXT: s_lshr_b32 s34, s15, 8
+; SI-NEXT: s_lshr_b32 s39, s14, 24
+; SI-NEXT: s_lshr_b32 s35, s14, 16
+; SI-NEXT: s_lshr_b32 s95, s14, 8
+; SI-NEXT: s_lshr_b32 s36, s13, 24
+; SI-NEXT: s_lshr_b32 s30, s13, 16
+; SI-NEXT: s_lshr_b32 s92, s13, 8
+; SI-NEXT: s_lshr_b32 s31, s12, 24
+; SI-NEXT: s_lshr_b32 s93, s12, 16
+; SI-NEXT: s_lshr_b32 s89, s12, 8
+; SI-NEXT: s_lshr_b32 s94, s11, 24
+; SI-NEXT: s_lshr_b32 s90, s11, 16
+; SI-NEXT: s_lshr_b32 s91, s10, 24
+; SI-NEXT: v_writelane_b32 v62, s4, 15
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
+; SI-NEXT: .LBB99_3: ; %end
+; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v50
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: s_and_b32 s4, s56, 0xff
+; SI-NEXT: s_lshl_b32 s5, s68, 8
+; SI-NEXT: s_lshl_b32 s16, s83, 24
+; SI-NEXT: v_and_b32_e32 v2, 0xff, v2
+; SI-NEXT: v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT: v_readlane_b32 s99, v63, 35
+; SI-NEXT: v_readlane_b32 s98, v63, 34
+; SI-NEXT: v_readlane_b32 s97, v63, 33
+; SI-NEXT: v_readlane_b32 s96, v63, 32
+; SI-NEXT: v_readlane_b32 s87, v63, 31
+; SI-NEXT: v_readlane_b32 s86, v63, 30
+; SI-NEXT: v_readlane_b32 s85, v63, 29
+; SI-NEXT: v_readlane_b32 s84, v63, 28
+; SI-NEXT: v_readlane_b32 s83, v63, 27
+; SI-NEXT: v_readlane_b32 s82, v63, 26
+; SI-NEXT: v_readlane_b32 s81, v63, 25
+; SI-NEXT: v_readlane_b32 s68, v63, 20
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v7, 24, v7
+; SI-NEXT: v_or_b32_e32 v3, v7, v3
+; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v7, 8, v7
+; SI-NEXT: v_or_b32_e32 v7, s4, v7
+; SI-NEXT: s_and_b32 s4, s57, 0xff
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: s_and_b32 s5, s71, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 16
+; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s16, s5
+; SI-NEXT: v_or_b32_e32 v3, v7, v3
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: buffer_store_dword v3, v0, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v3, vcc, 4, v0
+; SI-NEXT: v_mov_b32_e32 v7, s4
+; SI-NEXT: buffer_store_dword v7, v3, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT: s_and_b32 s4, s46, 0xff
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v51
+; SI-NEXT: v_or_b32_e32 v3, s4, v3
+; SI-NEXT: s_and_b32 s4, s47, 0xff
+; SI-NEXT: s_lshl_b32 s5, s65, 8
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: s_and_b32 s5, s69, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 16
+; SI-NEXT: s_lshl_b32 s16, s80, 24
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s16, s5
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: s_lshl_b32 s5, s54, 8
+; SI-NEXT: s_lshl_b32 s16, s70, 24
+; SI-NEXT: v_readlane_b32 s80, v63, 24
+; SI-NEXT: v_readlane_b32 s71, v63, 23
+; SI-NEXT: v_readlane_b32 s70, v63, 22
+; SI-NEXT: v_readlane_b32 s69, v63, 21
+; SI-NEXT: v_readlane_b32 s65, v63, 17
+; SI-NEXT: v_readlane_b32 s54, v63, 14
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v7, 0xff, v7
+; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v11, 24, v11
+; SI-NEXT: v_or_b32_e32 v7, v11, v7
+; SI-NEXT: v_or_b32_e32 v3, v3, v7
+; SI-NEXT: v_add_i32_e32 v7, vcc, 8, v0
+; SI-NEXT: buffer_store_dword v3, v7, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v3, vcc, 12, v0
+; SI-NEXT: v_mov_b32_e32 v7, s4
+; SI-NEXT: buffer_store_dword v7, v3, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT: s_and_b32 s4, s44, 0xff
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v49
+; SI-NEXT: v_or_b32_e32 v3, s4, v3
+; SI-NEXT: s_and_b32 s4, s45, 0xff
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: s_and_b32 s5, s66, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 16
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s16, s5
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: s_lshl_b32 s5, s51, 8
+; SI-NEXT: s_lshl_b32 s16, s67, 24
+; SI-NEXT: v_readlane_b32 s67, v63, 19
+; SI-NEXT: v_readlane_b32 s66, v63, 18
+; SI-NEXT: v_readlane_b32 s51, v63, 11
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v7, 0xff, v7
+; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v11, 24, v11
+; SI-NEXT: v_or_b32_e32 v7, v11, v7
+; SI-NEXT: v_or_b32_e32 v3, v3, v7
+; SI-NEXT: v_add_i32_e32 v7, vcc, 16, v0
+; SI-NEXT: buffer_store_dword v3, v7, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v3, vcc, 20, v0
+; SI-NEXT: v_mov_b32_e32 v7, s4
+; SI-NEXT: buffer_store_dword v7, v3, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT: s_and_b32 s4, s42, 0xff
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v48
+; SI-NEXT: v_or_b32_e32 v3, s4, v3
+; SI-NEXT: s_and_b32 s4, s43, 0xff
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: s_and_b32 s5, s55, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 16
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s16, s5
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: s_lshl_b32 s5, s48, 8
+; SI-NEXT: s_lshl_b32 s16, s64, 24
+; SI-NEXT: v_readlane_b32 s64, v63, 16
+; SI-NEXT: v_readlane_b32 s55, v63, 15
+; SI-NEXT: v_readlane_b32 s48, v63, 8
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_and_b32_e32 v7, 0xff, v7
+; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v11, 24, v11
+; SI-NEXT: v_or_b32_e32 v7, v11, v7
+; SI-NEXT: v_or_b32_e32 v3, v3, v7
+; SI-NEXT: v_add_i32_e32 v7, vcc, 24, v0
+; SI-NEXT: buffer_store_dword v3, v7, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v3, vcc, 28, v0
+; SI-NEXT: v_mov_b32_e32 v7, s4
+; SI-NEXT: buffer_store_dword v7, v3, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v16
+; SI-NEXT: s_and_b32 s4, s41, 0xff
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: s_and_b32 s5, s52, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 16
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s16, s5
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: s_lshl_b32 s5, s37, 8
+; SI-NEXT: s_lshl_b32 s16, s53, 24
+; SI-NEXT: v_readlane_b32 s53, v63, 13
+; SI-NEXT: v_readlane_b32 s52, v63, 12
+; SI-NEXT: v_readlane_b32 s37, v63, 5
+; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v7, 8, v7
+; SI-NEXT: v_or_b32_e32 v3, v3, v7
+; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v11, 24, v11
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v7, 0xff, v7
+; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
+; SI-NEXT: v_or_b32_e32 v7, v11, v7
+; SI-NEXT: v_or_b32_e32 v3, v3, v7
+; SI-NEXT: v_add_i32_e32 v7, vcc, 32, v0
+; SI-NEXT: buffer_store_dword v3, v7, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v3, vcc, 36, v0
+; SI-NEXT: v_mov_b32_e32 v7, s4
+; SI-NEXT: buffer_store_dword v7, v3, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v14
+; SI-NEXT: s_and_b32 s4, s40, 0xff
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: s_and_b32 s5, s49, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 16
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s16, s5
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: s_lshl_b32 s5, s34, 8
+; SI-NEXT: v_readlane_b32 s49, v63, 9
+; SI-NEXT: v_readlane_b32 s34, v63, 2
+; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v7, 8, v7
+; SI-NEXT: v_or_b32_e32 v3, v3, v7
+; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v11, 24, v11
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v7, 0xff, v7
+; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
+; SI-NEXT: v_or_b32_e32 v7, v11, v7
+; SI-NEXT: v_or_b32_e32 v3, v3, v7
+; SI-NEXT: v_add_i32_e32 v7, vcc, 40, v0
+; SI-NEXT: buffer_store_dword v3, v7, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v3, vcc, 44, v0
+; SI-NEXT: v_mov_b32_e32 v7, s4
+; SI-NEXT: buffer_store_dword v7, v3, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v12
+; SI-NEXT: s_and_b32 s4, s15, 0xff
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: s_and_b32 s5, s38, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 16
+; SI-NEXT: s_lshl_b32 s15, s50, 24
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s15, s5
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: s_lshl_b32 s5, s95, 8
+; SI-NEXT: v_readlane_b32 s50, v63, 10
+; SI-NEXT: v_readlane_b32 s38, v63, 6
+; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v7, 8, v7
+; SI-NEXT: v_or_b32_e32 v3, v3, v7
+; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v11, 24, v11
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v7, 0xff, v7
+; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
+; SI-NEXT: v_or_b32_e32 v7, v11, v7
+; SI-NEXT: v_or_b32_e32 v3, v3, v7
+; SI-NEXT: v_add_i32_e32 v7, vcc, 48, v0
+; SI-NEXT: buffer_store_dword v3, v7, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v3, vcc, 52, v0
+; SI-NEXT: v_mov_b32_e32 v7, s4
+; SI-NEXT: buffer_store_dword v7, v3, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v10
+; SI-NEXT: s_and_b32 s4, s14, 0xff
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: s_and_b32 s5, s35, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 16
+; SI-NEXT: s_lshl_b32 s14, s39, 24
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s14, s5
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: s_lshl_b32 s5, s92, 8
+; SI-NEXT: v_readlane_b32 s39, v63, 7
+; SI-NEXT: v_readlane_b32 s35, v63, 3
+; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v7, 8, v7
+; SI-NEXT: v_or_b32_e32 v3, v3, v7
+; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v10, 24, v10
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_and_b32_e32 v7, 0xff, v7
+; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
+; SI-NEXT: v_or_b32_e32 v7, v10, v7
+; SI-NEXT: v_or_b32_e32 v3, v3, v7
+; SI-NEXT: v_add_i32_e32 v7, vcc, 56, v0
+; SI-NEXT: buffer_store_dword v3, v7, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v3, vcc, 60, v0
+; SI-NEXT: v_mov_b32_e32 v7, s4
+; SI-NEXT: buffer_store_dword v7, v3, s[0:3], 0 offen
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v8
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v7, 8, v57
+; SI-NEXT: s_and_b32 s4, s13, 0xff
+; SI-NEXT: v_or_b32_e32 v3, v3, v7
+; SI-NEXT: v_and_b32_e32 v7, 0xff, v47
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: s_and_b32 s5, s30, 0xff
+; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
+; SI-NEXT: v_lshlrev_b32_e32 v8, 24, v45
+; SI-NEXT: s_lshl_b32 s5, s5, 16
+; SI-NEXT: s_lshl_b32 s13, s36, 24
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT: v_or_b32_e32 v7, v8, v7
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s13, s5
+; SI-NEXT: v_or_b32_e32 v3, v3, v7
+; SI-NEXT: v_add_i32_e32 v7, vcc, 64, v0
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: buffer_store_dword v3, v7, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x44, v0
+; SI-NEXT: v_mov_b32_e32 v7, s4
+; SI-NEXT: buffer_store_dword v7, v3, s[0:3], 0 offen
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v5
+; SI-NEXT: v_lshlrev_b32_e32 v5, 8, v44
+; SI-NEXT: s_and_b32 s4, s12, 0xff
+; SI-NEXT: s_lshl_b32 s5, s89, 8
+; SI-NEXT: v_or_b32_e32 v3, v3, v5
+; SI-NEXT: v_and_b32_e32 v5, 0xff, v43
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: s_and_b32 s5, s93, 0xff
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v7, 24, v41
+; SI-NEXT: s_lshl_b32 s5, s5, 16
+; SI-NEXT: s_lshl_b32 s12, s31, 24
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT: v_or_b32_e32 v5, v7, v5
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s12, s5
+; SI-NEXT: v_or_b32_e32 v3, v3, v5
+; SI-NEXT: v_add_i32_e32 v5, vcc, 0x48, v0
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: buffer_store_dword v3, v5, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x4c, v0
+; SI-NEXT: v_mov_b32_e32 v5, s4
+; SI-NEXT: v_readlane_b32 s5, v62, 1
+; SI-NEXT: buffer_store_dword v5, v3, s[0:3], 0 offen
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v13
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v5, 8, v24
+; SI-NEXT: s_and_b32 s4, s11, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 8
+; SI-NEXT: v_or_b32_e32 v3, v3, v5
+; SI-NEXT: v_and_b32_e32 v5, 0xff, v22
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: s_and_b32 s5, s90, 0xff
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT: v_lshlrev_b32_e32 v7, 24, v21
+; SI-NEXT: s_lshl_b32 s5, s5, 16
+; SI-NEXT: s_lshl_b32 s11, s94, 24
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT: v_or_b32_e32 v5, v7, v5
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s11, s5
+; SI-NEXT: v_or_b32_e32 v3, v3, v5
+; SI-NEXT: v_add_i32_e32 v5, vcc, 0x50, v0
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: v_readlane_b32 s5, v62, 3
+; SI-NEXT: buffer_store_dword v3, v5, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x54, v0
+; SI-NEXT: v_mov_b32_e32 v5, s4
+; SI-NEXT: s_and_b32 s4, s10, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 8
+; SI-NEXT: buffer_store_dword v5, v3, s[0:3], 0 offen
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v9
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v5, 8, v20
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: v_readlane_b32 s5, v62, 2
+; SI-NEXT: v_or_b32_e32 v3, v3, v5
+; SI-NEXT: v_and_b32_e32 v5, 0xff, v18
+; SI-NEXT: s_and_b32 s5, s5, 0xff
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT: v_lshlrev_b32_e32 v7, 24, v17
+; SI-NEXT: s_lshl_b32 s5, s5, 16
+; SI-NEXT: s_lshl_b32 s10, s91, 24
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT: v_or_b32_e32 v5, v7, v5
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s10, s5
+; SI-NEXT: v_or_b32_e32 v3, v3, v5
+; SI-NEXT: v_add_i32_e32 v5, vcc, 0x58, v0
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: v_readlane_b32 s5, v62, 6
+; SI-NEXT: buffer_store_dword v3, v5, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x5c, v0
+; SI-NEXT: v_mov_b32_e32 v5, s4
+; SI-NEXT: s_and_b32 s4, s9, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 8
+; SI-NEXT: buffer_store_dword v5, v3, s[0:3], 0 offen
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v6
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v5, 8, v61
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: v_readlane_b32 s5, v62, 5
+; SI-NEXT: v_or_b32_e32 v3, v3, v5
+; SI-NEXT: v_and_b32_e32 v5, 0xff, v60
+; SI-NEXT: s_and_b32 s5, s5, 0xff
+; SI-NEXT: v_readlane_b32 s9, v62, 4
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT: v_lshlrev_b32_e32 v6, 24, v59
+; SI-NEXT: s_lshl_b32 s5, s5, 16
+; SI-NEXT: s_lshl_b32 s9, s9, 24
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT: v_or_b32_e32 v5, v6, v5
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s9, s5
+; SI-NEXT: v_or_b32_e32 v3, v3, v5
+; SI-NEXT: v_add_i32_e32 v5, vcc, 0x60, v0
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: v_readlane_b32 s5, v62, 9
+; SI-NEXT: buffer_store_dword v3, v5, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x64, v0
+; SI-NEXT: v_mov_b32_e32 v5, s4
+; SI-NEXT: s_and_b32 s4, s8, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 8
+; SI-NEXT: buffer_store_dword v5, v3, s[0:3], 0 offen
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v4
+; SI-NEXT: v_lshlrev_b32_e32 v4, 8, v58
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: v_readlane_b32 s5, v62, 8
+; SI-NEXT: v_or_b32_e32 v3, v3, v4
+; SI-NEXT: v_and_b32_e32 v4, 0xff, v56
+; SI-NEXT: s_and_b32 s5, s5, 0xff
+; SI-NEXT: v_readlane_b32 s8, v62, 7
+; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v5, 24, v46
+; SI-NEXT: s_lshl_b32 s5, s5, 16
+; SI-NEXT: s_lshl_b32 s8, s8, 24
+; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT: v_or_b32_e32 v4, v5, v4
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s8, s5
+; SI-NEXT: v_or_b32_e32 v3, v3, v4
+; SI-NEXT: v_add_i32_e32 v4, vcc, 0x68, v0
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: v_readlane_b32 s5, v62, 12
+; SI-NEXT: buffer_store_dword v3, v4, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x6c, v0
+; SI-NEXT: v_mov_b32_e32 v4, s4
+; SI-NEXT: s_and_b32 s4, s7, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 8
+; SI-NEXT: buffer_store_dword v4, v3, s[0:3], 0 offen
+; SI-NEXT: v_lshlrev_b32_e32 v3, 8, v42
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: v_readlane_b32 s5, v62, 11
+; SI-NEXT: v_or_b32_e32 v2, v2, v3
+; SI-NEXT: v_and_b32_e32 v3, 0xff, v40
+; SI-NEXT: s_and_b32 s5, s5, 0xff
+; SI-NEXT: v_readlane_b32 s7, v62, 10
+; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v4, 24, v55
+; SI-NEXT: s_lshl_b32 s5, s5, 16
+; SI-NEXT: s_lshl_b32 s7, s7, 24
+; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT: v_or_b32_e32 v3, v4, v3
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s7, s5
+; SI-NEXT: v_or_b32_e32 v2, v2, v3
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x70, v0
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: v_readlane_b32 s5, v62, 15
+; SI-NEXT: buffer_store_dword v2, v3, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_add_i32_e32 v2, vcc, 0x74, v0
+; SI-NEXT: v_mov_b32_e32 v3, s4
+; SI-NEXT: s_and_b32 s4, s6, 0xff
+; SI-NEXT: s_lshl_b32 s5, s5, 8
+; SI-NEXT: buffer_store_dword v3, v2, s[0:3], 0 offen
+; SI-NEXT: v_lshlrev_b32_e32 v2, 8, v54
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: v_readlane_b32 s5, v62, 14
+; SI-NEXT: v_or_b32_e32 v1, v1, v2
+; SI-NEXT: v_and_b32_e32 v2, 0xff, v53
+; SI-NEXT: s_and_b32 s5, s5, 0xff
+; SI-NEXT: v_readlane_b32 s6, v62, 13
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v3, 24, v52
+; SI-NEXT: s_lshl_b32 s5, s5, 16
+; SI-NEXT: s_lshl_b32 s6, s6, 24
+; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT: v_or_b32_e32 v2, v3, v2
+; SI-NEXT: s_and_b32 s4, s4, 0xffff
+; SI-NEXT: s_or_b32 s5, s6, s5
+; SI-NEXT: v_or_b32_e32 v1, v1, v2
+; SI-NEXT: v_add_i32_e32 v2, vcc, 0x78, v0
+; SI-NEXT: s_or_b32 s4, s4, s5
+; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
+; SI-NEXT: v_add_i32_e32 v0, vcc, 0x7c, v0
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_mov_b32_e32 v1, s4
+; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
+; SI-NEXT: v_readlane_b32 s36, v63, 4
+; SI-NEXT: v_readlane_b32 s31, v63, 1
+; SI-NEXT: v_readlane_b32 s30, v63, 0
+; SI-NEXT: s_or_saveexec_b64 s[4:5], -1
+; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
+; SI-NEXT: s_mov_b64 exec, s[4:5]
+; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; SI-NEXT: s_setpc_b64 s[30:31]
+; SI-NEXT: .LBB99_4:
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $sgpr6
+; SI-NEXT: ; kill: killed $sgpr6
+; SI-NEXT: ; implicit-def: $vcc_lo
+; SI-NEXT: v_mov_b32_e32 v39, v24
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $sgpr6
+; SI-NEXT: ; kill: killed $sgpr6
+; SI-NEXT: v_mov_b32_e32 v38, v22
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $sgpr6
+; SI-NEXT: ; kill: killed $sgpr6
+; SI-NEXT: v_mov_b32_e32 v37, v21
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $sgpr6
+; SI-NEXT: ; kill: killed $sgpr6
+; SI-NEXT: v_mov_b32_e32 v36, v20
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $sgpr6
+; SI-NEXT: ; kill: killed $sgpr6
+; SI-NEXT: v_mov_b32_e32 v30, v18
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $sgpr6
+; SI-NEXT: ; kill: killed $sgpr6
+; SI-NEXT: v_mov_b32_e32 v29, v17
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $sgpr6
+; SI-NEXT: ; kill: killed $sgpr6
+; SI-NEXT: v_mov_b32_e32 v28, v13
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $sgpr6
+; SI-NEXT: ; kill: killed $sgpr6
+; SI-NEXT: v_mov_b32_e32 v26, v9
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $sgpr6
+; SI-NEXT: ; kill: killed $sgpr6
+; SI-NEXT: v_mov_b32_e32 v25, v6
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $sgpr6
+; SI-NEXT: ; kill: killed $sgpr6
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $sgpr6
+; SI-NEXT: ; kill: killed $sgpr6
+; SI-NEXT: ; kill: killed $vcc_lo
+; SI-NEXT: ; implicit-def: $vcc_lo
+; SI-NEXT: ; implicit-def: $sgpr56
+; SI-NEXT: ; implicit-def: $vgpr50
+; SI-NEXT: ; implicit-def: $sgpr57
+; SI-NEXT: ; implicit-def: $sgpr68
+; SI-NEXT: ; implicit-def: $sgpr71
+; SI-NEXT: ; implicit-def: $sgpr83
+; SI-NEXT: ; implicit-def: $sgpr46
+; SI-NEXT: ; implicit-def: $vgpr51
+; SI-NEXT: ; implicit-def: $sgpr47
+; SI-NEXT: ; implicit-def: $sgpr65
+; SI-NEXT: ; implicit-def: $sgpr69
+; SI-NEXT: ; implicit-def: $sgpr80
+; SI-NEXT: ; implicit-def: $sgpr44
+; SI-NEXT: ; implicit-def: $vgpr49
+; SI-NEXT: ; implicit-def: $sgpr45
+; SI-NEXT: ; implicit-def: $sgpr54
+; SI-NEXT: ; implicit-def: $sgpr66
+; SI-NEXT: ; implicit-def: $sgpr70
+; SI-NEXT: ; implicit-def: $sgpr42
+; SI-NEXT: ; implicit-def: $vgpr48
+; SI-NEXT: ; implicit-def: $sgpr43
+; SI-NEXT: ; implicit-def: $sgpr51
+; SI-NEXT: ; implicit-def: $sgpr55
+; SI-NEXT: ; implicit-def: $sgpr67
+; SI-NEXT: ; implicit-def: $vgpr16
+; SI-NEXT: ; implicit-def: $sgpr41
+; SI-NEXT: ; implicit-def: $sgpr48
+; SI-NEXT: ; implicit-def: $sgpr52
+; SI-NEXT: ; implicit-def: $sgpr64
+; SI-NEXT: ; implicit-def: $vgpr14
+; SI-NEXT: ; implicit-def: $sgpr40
+; SI-NEXT: ; implicit-def: $sgpr37
+; SI-NEXT: ; implicit-def: $sgpr49
+; SI-NEXT: ; implicit-def: $sgpr53
+; SI-NEXT: ; implicit-def: $vgpr12
+; SI-NEXT: ; implicit-def: $sgpr15
+; SI-NEXT: ; implicit-def: $sgpr34
+; SI-NEXT: ; implicit-def: $sgpr38
+; SI-NEXT: ; implicit-def: $sgpr50
+; SI-NEXT: ; implicit-def: $vgpr10
+; SI-NEXT: ; implicit-def: $sgpr14
+; SI-NEXT: ; implicit-def: $sgpr95
+; SI-NEXT: ; implicit-def: $sgpr35
+; SI-NEXT: ; implicit-def: $sgpr39
+; SI-NEXT: ; implicit-def: $vgpr8
+; SI-NEXT: ; implicit-def: $vgpr57
+; SI-NEXT: ; implicit-def: $vgpr47
+; SI-NEXT: ; implicit-def: $vgpr45
+; SI-NEXT: ; implicit-def: $sgpr13
+; SI-NEXT: ; implicit-def: $sgpr92
+; SI-NEXT: ; implicit-def: $sgpr30
+; SI-NEXT: ; implicit-def: $sgpr36
+; SI-NEXT: ; implicit-def: $vgpr5
+; SI-NEXT: ; implicit-def: $vgpr44
+; SI-NEXT: ; implicit-def: $vgpr43
+; SI-NEXT: ; implicit-def: $vgpr41
+; SI-NEXT: ; implicit-def: $sgpr12
+; SI-NEXT: ; implicit-def: $sgpr89
+; SI-NEXT: ; implicit-def: $sgpr93
+; SI-NEXT: ; implicit-def: $sgpr31
+; SI-NEXT: ; implicit-def: $sgpr11
+; SI-NEXT: ; implicit-def: $sgpr90
+; SI-NEXT: ; implicit-def: $sgpr94
+; SI-NEXT: ; implicit-def: $sgpr10
+; SI-NEXT: ; implicit-def: $sgpr91
+; SI-NEXT: ; implicit-def: $sgpr9
+; SI-NEXT: ; implicit-def: $sgpr8
+; SI-NEXT: ; implicit-def: $sgpr7
+; SI-NEXT: ; kill: killed $vcc_lo
+; SI-NEXT: ; implicit-def: $vcc_lo
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $sgpr6
+; SI-NEXT: ; kill: killed $sgpr6
+; SI-NEXT: ; kill: killed $vcc_lo
+; SI-NEXT: ; implicit-def: $vgpr13
+; SI-NEXT: ; implicit-def: $vgpr24
+; SI-NEXT: ; implicit-def: $vgpr22
+; SI-NEXT: ; implicit-def: $vgpr21
+; SI-NEXT: ; implicit-def: $vgpr9
+; SI-NEXT: ; implicit-def: $vgpr20
+; SI-NEXT: ; implicit-def: $vgpr18
+; SI-NEXT: ; implicit-def: $vgpr17
+; SI-NEXT: ; implicit-def: $vgpr6
+; SI-NEXT: ; implicit-def: $vgpr61
+; SI-NEXT: ; implicit-def: $vgpr60
+; SI-NEXT: ; implicit-def: $vgpr59
+; SI-NEXT: ; implicit-def: $vgpr4
+; SI-NEXT: ; implicit-def: $vgpr58
+; SI-NEXT: ; implicit-def: $vgpr56
+; SI-NEXT: ; implicit-def: $vgpr46
+; SI-NEXT: ; implicit-def: $vgpr2
+; SI-NEXT: ; implicit-def: $vgpr42
+; SI-NEXT: ; implicit-def: $vgpr40
+; SI-NEXT: ; implicit-def: $vgpr55
+; SI-NEXT: ; implicit-def: $vgpr54
+; SI-NEXT: ; implicit-def: $vgpr53
+; SI-NEXT: ; implicit-def: $vgpr52
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $sgpr6
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; kill: killed $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: s_branch .LBB99_2
+;
+; VI-LABEL: bitcast_v64i16_to_v128i8_scalar:
+; VI: ; %bb.0:
+; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; VI-NEXT: s_mov_b64 exec, s[4:5]
+; VI-NEXT: v_writelane_b32 v20, s30, 0
+; VI-NEXT: v_writelane_b32 v20, s31, 1
+; VI-NEXT: v_writelane_b32 v20, s34, 2
+; VI-NEXT: v_writelane_b32 v20, s35, 3
+; VI-NEXT: v_writelane_b32 v20, s36, 4
+; VI-NEXT: v_writelane_b32 v20, s37, 5
+; VI-NEXT: v_writelane_b32 v20, s38, 6
+; VI-NEXT: v_writelane_b32 v20, s39, 7
+; VI-NEXT: v_writelane_b32 v20, s48, 8
+; VI-NEXT: v_writelane_b32 v20, s49, 9
+; VI-NEXT: v_writelane_b32 v20, s50, 10
+; VI-NEXT: v_writelane_b32 v20, s51, 11
+; VI-NEXT: v_writelane_b32 v20, s52, 12
+; VI-NEXT: v_writelane_b32 v20, s53, 13
+; VI-NEXT: v_writelane_b32 v20, s54, 14
+; VI-NEXT: v_writelane_b32 v20, s55, 15
+; VI-NEXT: v_writelane_b32 v20, s64, 16
+; VI-NEXT: v_writelane_b32 v20, s65, 17
+; VI-NEXT: v_writelane_b32 v20, s66, 18
+; VI-NEXT: v_writelane_b32 v20, s67, 19
+; VI-NEXT: v_writelane_b32 v20, s68, 20
+; VI-NEXT: v_writelane_b32 v20, s69, 21
+; VI-NEXT: v_writelane_b32 v20, s70, 22
+; VI-NEXT: v_writelane_b32 v20, s71, 23
+; VI-NEXT: v_writelane_b32 v20, s80, 24
+; VI-NEXT: v_writelane_b32 v20, s81, 25
+; VI-NEXT: v_writelane_b32 v20, s82, 26
+; VI-NEXT: v_writelane_b32 v20, s83, 27
+; VI-NEXT: v_writelane_b32 v20, s84, 28
+; VI-NEXT: v_writelane_b32 v20, s85, 29
+; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v19
+; VI-NEXT: v_writelane_b32 v20, s86, 30
+; VI-NEXT: v_readfirstlane_b32 s42, v3
+; VI-NEXT: v_readfirstlane_b32 s43, v4
+; VI-NEXT: v_readfirstlane_b32 s40, v5
+; VI-NEXT: v_readfirstlane_b32 s41, v6
+; VI-NEXT: v_readfirstlane_b32 s14, v7
+; VI-NEXT: v_readfirstlane_b32 s15, v8
+; VI-NEXT: v_readfirstlane_b32 s12, v9
+; VI-NEXT: v_readfirstlane_b32 s13, v10
+; VI-NEXT: v_readfirstlane_b32 s10, v11
+; VI-NEXT: v_readfirstlane_b32 s11, v12
+; VI-NEXT: v_readfirstlane_b32 s8, v13
+; VI-NEXT: v_readfirstlane_b32 s9, v14
+; VI-NEXT: v_readfirstlane_b32 s6, v15
+; VI-NEXT: v_readfirstlane_b32 s7, v16
+; VI-NEXT: v_readfirstlane_b32 s4, v17
+; VI-NEXT: v_readfirstlane_b32 s5, v18
+; VI-NEXT: v_readfirstlane_b32 s44, v1
+; VI-NEXT: s_and_b64 s[46:47], vcc, exec
+; VI-NEXT: v_readfirstlane_b32 s45, v2
+; VI-NEXT: v_writelane_b32 v20, s87, 31
+; VI-NEXT: ; implicit-def: $vgpr21 : SGPR spill to VGPR lane
+; VI-NEXT: s_cbranch_scc0 .LBB99_4
+; VI-NEXT: ; %bb.1: ; %cmp.false
+; VI-NEXT: s_lshr_b32 s46, s45, 24
+; VI-NEXT: v_writelane_b32 v21, s46, 0
+; VI-NEXT: s_lshr_b32 s46, s45, 16
+; VI-NEXT: v_writelane_b32 v21, s46, 1
+; VI-NEXT: s_lshr_b32 s46, s45, 8
+; VI-NEXT: v_writelane_b32 v21, s46, 2
+; VI-NEXT: s_lshr_b32 s46, s44, 16
+; VI-NEXT: v_writelane_b32 v21, s46, 3
+; VI-NEXT: s_lshr_b32 s46, s44, 8
+; VI-NEXT: v_writelane_b32 v21, s46, 4
+; VI-NEXT: s_lshr_b32 s46, s29, 24
+; VI-NEXT: v_writelane_b32 v21, s46, 5
+; VI-NEXT: s_lshr_b32 s46, s29, 16
+; VI-NEXT: v_writelane_b32 v21, s46, 6
+; VI-NEXT: s_lshr_b32 s46, s29, 8
+; VI-NEXT: v_writelane_b32 v21, s46, 7
+; VI-NEXT: s_lshr_b32 s46, s28, 16
+; VI-NEXT: v_writelane_b32 v21, s46, 8
+; VI-NEXT: s_lshr_b32 s46, s28, 8
+; VI-NEXT: v_writelane_b32 v21, s46, 9
+; VI-NEXT: s_lshr_b32 s46, s27, 24
+; VI-NEXT: v_writelane_b32 v21, s46, 10
+; VI-NEXT: s_lshr_b32 s46, s27, 16
+; VI-NEXT: v_writelane_b32 v21, s46, 11
+; VI-NEXT: s_lshr_b32 s46, s27, 8
+; VI-NEXT: v_writelane_b32 v21, s46, 12
+; VI-NEXT: s_lshr_b32 s46, s26, 16
+; VI-NEXT: v_writelane_b32 v21, s46, 13
+; VI-NEXT: s_lshr_b32 s46, s26, 8
+; VI-NEXT: v_writelane_b32 v21, s46, 14
+; VI-NEXT: s_lshr_b32 s46, s25, 24
+; VI-NEXT: v_writelane_b32 v21, s46, 15
+; VI-NEXT: s_lshr_b32 s46, s25, 16
+; VI-NEXT: v_writelane_b32 v21, s46, 16
+; VI-NEXT: s_lshr_b32 s46, s25, 8
+; VI-NEXT: v_writelane_b32 v21, s46, 17
+; VI-NEXT: s_lshr_b32 s46, s24, 16
+; VI-NEXT: v_writelane_b32 v21, s46, 18
+; VI-NEXT: s_lshr_b32 s46, s24, 8
+; VI-NEXT: v_writelane_b32 v21, s46, 19
+; VI-NEXT: s_lshr_b32 s46, s23, 24
+; VI-NEXT: v_writelane_b32 v21, s46, 20
+; VI-NEXT: s_lshr_b32 s46, s23, 16
+; VI-NEXT: v_writelane_b32 v21, s46, 21
+; VI-NEXT: s_lshr_b32 s46, s23, 8
+; VI-NEXT: v_writelane_b32 v21, s46, 22
+; VI-NEXT: s_lshr_b32 s46, s22, 16
+; VI-NEXT: v_writelane_b32 v21, s46, 23
+; VI-NEXT: s_lshr_b32 s46, s22, 8
+; VI-NEXT: v_writelane_b32 v21, s46, 24
+; VI-NEXT: s_lshr_b32 s46, s21, 24
+; VI-NEXT: v_writelane_b32 v21, s46, 25
+; VI-NEXT: s_lshr_b32 s46, s5, 24
+; VI-NEXT: v_writelane_b32 v21, s46, 26
+; VI-NEXT: s_lshr_b32 s46, s5, 16
+; VI-NEXT: v_writelane_b32 v21, s46, 27
+; VI-NEXT: s_lshr_b32 s46, s5, 8
+; VI-NEXT: v_writelane_b32 v21, s46, 28
+; VI-NEXT: s_lshr_b32 s46, s4, 16
+; VI-NEXT: v_writelane_b32 v21, s46, 29
+; VI-NEXT: s_lshr_b32 s46, s4, 8
+; VI-NEXT: v_writelane_b32 v21, s46, 30
+; VI-NEXT: s_lshr_b32 s46, s7, 24
+; VI-NEXT: v_writelane_b32 v21, s46, 31
+; VI-NEXT: s_lshr_b32 s46, s7, 16
+; VI-NEXT: v_writelane_b32 v21, s46, 32
+; VI-NEXT: s_lshr_b32 s46, s7, 8
+; VI-NEXT: v_writelane_b32 v21, s46, 33
+; VI-NEXT: s_lshr_b32 s46, s6, 16
+; VI-NEXT: v_writelane_b32 v21, s46, 34
+; VI-NEXT: s_lshr_b32 s46, s6, 8
+; VI-NEXT: v_writelane_b32 v21, s46, 35
+; VI-NEXT: s_lshr_b32 s46, s9, 24
+; VI-NEXT: v_writelane_b32 v21, s46, 36
+; VI-NEXT: s_lshr_b32 s46, s9, 16
+; VI-NEXT: v_writelane_b32 v21, s46, 37
+; VI-NEXT: s_lshr_b32 s46, s9, 8
+; VI-NEXT: v_writelane_b32 v21, s46, 38
+; VI-NEXT: s_lshr_b32 s46, s8, 16
+; VI-NEXT: v_writelane_b32 v21, s46, 39
+; VI-NEXT: s_lshr_b32 s46, s8, 8
+; VI-NEXT: v_writelane_b32 v21, s46, 40
+; VI-NEXT: s_lshr_b32 s46, s11, 24
+; VI-NEXT: v_writelane_b32 v21, s46, 41
+; VI-NEXT: s_lshr_b32 s46, s11, 16
+; VI-NEXT: v_writelane_b32 v21, s46, 42
+; VI-NEXT: s_lshr_b32 s46, s11, 8
+; VI-NEXT: v_writelane_b32 v21, s46, 43
+; VI-NEXT: s_lshr_b32 s46, s10, 16
+; VI-NEXT: v_writelane_b32 v21, s46, 44
+; VI-NEXT: s_lshr_b32 s46, s10, 8
+; VI-NEXT: v_writelane_b32 v21, s46, 45
+; VI-NEXT: s_lshr_b32 s46, s13, 24
+; VI-NEXT: v_writelane_b32 v21, s46, 46
+; VI-NEXT: s_lshr_b32 s46, s13, 16
+; VI-NEXT: v_writelane_b32 v21, s46, 47
+; VI-NEXT: s_lshr_b32 s46, s13, 8
+; VI-NEXT: v_writelane_b32 v21, s46, 48
+; VI-NEXT: s_lshr_b32 s46, s12, 16
+; VI-NEXT: v_writelane_b32 v21, s46, 49
+; VI-NEXT: s_lshr_b32 s46, s12, 8
+; VI-NEXT: v_writelane_b32 v21, s46, 50
+; VI-NEXT: s_lshr_b32 s46, s15, 24
+; VI-NEXT: v_writelane_b32 v21, s46, 51
+; VI-NEXT: s_lshr_b32 s46, s15, 16
+; VI-NEXT: v_writelane_b32 v21, s46, 52
+; VI-NEXT: s_lshr_b32 s46, s15, 8
+; VI-NEXT: v_writelane_b32 v21, s46, 53
+; VI-NEXT: s_lshr_b32 s46, s14, 16
+; VI-NEXT: v_writelane_b32 v21, s46, 54
+; VI-NEXT: s_lshr_b32 s46, s14, 8
+; VI-NEXT: v_writelane_b32 v21, s46, 55
+; VI-NEXT: s_lshr_b32 s46, s41, 24
+; VI-NEXT: v_writelane_b32 v21, s46, 56
+; VI-NEXT: s_lshr_b32 s46, s41, 16
+; VI-NEXT: s_lshr_b32 s80, s21, 16
+; VI-NEXT: s_lshr_b32 s82, s21, 8
+; VI-NEXT: s_lshr_b32 s84, s20, 16
+; VI-NEXT: s_lshr_b32 s86, s20, 8
+; VI-NEXT: s_lshr_b32 s51, s19, 24
+; VI-NEXT: s_lshr_b32 s53, s19, 16
+; VI-NEXT: s_lshr_b32 s54, s19, 8
+; VI-NEXT: s_lshr_b32 s65, s18, 16
+; VI-NEXT: s_lshr_b32 s66, s18, 8
+; VI-NEXT: s_lshr_b32 s67, s17, 24
+; VI-NEXT: s_lshr_b32 s68, s17, 16
+; VI-NEXT: s_lshr_b32 s69, s17, 8
+; VI-NEXT: s_lshr_b32 s70, s16, 16
+; VI-NEXT: s_lshr_b32 s71, s16, 8
+; VI-NEXT: v_writelane_b32 v21, s46, 57
+; VI-NEXT: s_lshr_b32 s81, s41, 8
+; VI-NEXT: s_lshr_b32 s83, s40, 16
+; VI-NEXT: s_lshr_b32 s85, s40, 8
+; VI-NEXT: s_lshr_b32 s87, s43, 24
+; VI-NEXT: s_lshr_b32 s50, s43, 16
+; VI-NEXT: s_lshr_b32 s52, s43, 8
+; VI-NEXT: s_lshr_b32 s55, s42, 16
+; VI-NEXT: s_lshr_b32 s64, s42, 8
+; VI-NEXT: s_lshr_b64 s[76:77], s[44:45], 24
+; VI-NEXT: s_lshr_b64 s[88:89], s[28:29], 24
+; VI-NEXT: s_lshr_b64 s[90:91], s[26:27], 24
+; VI-NEXT: s_lshr_b64 s[30:31], s[24:25], 24
+; VI-NEXT: s_lshr_b64 s[34:35], s[22:23], 24
+; VI-NEXT: s_lshr_b64 s[36:37], s[20:21], 24
+; VI-NEXT: s_lshr_b64 s[38:39], s[18:19], 24
+; VI-NEXT: s_lshr_b64 s[48:49], s[16:17], 24
+; VI-NEXT: s_lshr_b64 s[46:47], s[4:5], 24
+; VI-NEXT: s_lshr_b64 s[56:57], s[6:7], 24
+; VI-NEXT: s_lshr_b64 s[58:59], s[8:9], 24
+; VI-NEXT: s_lshr_b64 s[60:61], s[10:11], 24
+; VI-NEXT: s_lshr_b64 s[62:63], s[12:13], 24
+; VI-NEXT: s_lshr_b64 s[72:73], s[14:15], 24
+; VI-NEXT: s_lshr_b64 s[74:75], s[40:41], 24
+; VI-NEXT: s_lshr_b64 s[78:79], s[42:43], 24
+; VI-NEXT: s_cbranch_execnz .LBB99_3
+; VI-NEXT: .LBB99_2: ; %cmp.true
+; VI-NEXT: s_and_b32 s47, s45, 0xffff0000
+; VI-NEXT: s_add_i32 s45, s45, 3
+; VI-NEXT: s_and_b32 s45, s45, 0xffff
+; VI-NEXT: s_and_b32 s46, s44, 0xffff0000
+; VI-NEXT: s_add_i32 s44, s44, 3
+; VI-NEXT: s_or_b32 s45, s47, s45
+; VI-NEXT: s_add_i32 s45, s45, 0x30000
+; VI-NEXT: s_and_b32 s44, s44, 0xffff
+; VI-NEXT: s_or_b32 s44, s46, s44
+; VI-NEXT: s_lshr_b32 s46, s45, 24
+; VI-NEXT: s_and_b32 s57, s29, 0xffff0000
+; VI-NEXT: s_add_i32 s29, s29, 3
+; VI-NEXT: v_writelane_b32 v21, s46, 0
+; VI-NEXT: s_lshr_b32 s46, s45, 16
+; VI-NEXT: s_and_b32 s29, s29, 0xffff
+; VI-NEXT: s_add_i32 s44, s44, 0x30000
+; VI-NEXT: v_writelane_b32 v21, s46, 1
+; VI-NEXT: s_lshr_b32 s46, s45, 8
+; VI-NEXT: s_or_b32 s29, s57, s29
+; VI-NEXT: v_writelane_b32 v21, s46, 2
+; VI-NEXT: s_lshr_b32 s46, s44, 16
+; VI-NEXT: s_and_b32 s56, s28, 0xffff0000
+; VI-NEXT: s_add_i32 s28, s28, 3
+; VI-NEXT: s_add_i32 s29, s29, 0x30000
+; VI-NEXT: v_writelane_b32 v21, s46, 3
+; VI-NEXT: s_lshr_b32 s46, s44, 8
+; VI-NEXT: s_and_b32 s28, s28, 0xffff
+; VI-NEXT: v_writelane_b32 v21, s46, 4
+; VI-NEXT: s_lshr_b32 s46, s29, 24
+; VI-NEXT: s_and_b32 s59, s27, 0xffff0000
+; VI-NEXT: s_add_i32 s27, s27, 3
+; VI-NEXT: s_or_b32 s28, s56, s28
+; VI-NEXT: v_writelane_b32 v21, s46, 5
+; VI-NEXT: s_lshr_b32 s46, s29, 16
+; VI-NEXT: s_and_b32 s27, s27, 0xffff
+; VI-NEXT: s_add_i32 s28, s28, 0x30000
+; VI-NEXT: v_writelane_b32 v21, s46, 6
+; VI-NEXT: s_lshr_b32 s46, s29, 8
+; VI-NEXT: s_or_b32 s27, s59, s27
+; VI-NEXT: v_writelane_b32 v21, s46, 7
+; VI-NEXT: s_lshr_b32 s46, s28, 16
+; VI-NEXT: s_and_b32 s58, s26, 0xffff0000
+; VI-NEXT: s_add_i32 s26, s26, 3
+; VI-NEXT: s_add_i32 s27, s27, 0x30000
+; VI-NEXT: v_writelane_b32 v21, s46, 8
+; VI-NEXT: s_lshr_b32 s46, s28, 8
+; VI-NEXT: s_and_b32 s26, s26, 0xffff
+; VI-NEXT: v_writelane_b32 v21, s46, 9
+; VI-NEXT: s_lshr_b32 s46, s27, 24
+; VI-NEXT: s_and_b32 s61, s25, 0xffff0000
+; VI-NEXT: s_add_i32 s25, s25, 3
+; VI-NEXT: s_or_b32 s26, s58, s26
+; VI-NEXT: v_writelane_b32 v21, s46, 10
+; VI-NEXT: s_lshr_b32 s46, s27, 16
+; VI-NEXT: s_and_b32 s25, s25, 0xffff
+; VI-NEXT: s_add_i32 s26, s26, 0x30000
+; VI-NEXT: v_writelane_b32 v21, s46, 11
+; VI-NEXT: s_lshr_b32 s46, s27, 8
+; VI-NEXT: s_or_b32 s25, s61, s25
+; VI-NEXT: v_writelane_b32 v21, s46, 12
+; VI-NEXT: s_lshr_b32 s46, s26, 16
+; VI-NEXT: s_and_b32 s60, s24, 0xffff0000
+; VI-NEXT: s_add_i32 s24, s24, 3
+; VI-NEXT: s_add_i32 s25, s25, 0x30000
+; VI-NEXT: v_writelane_b32 v21, s46, 13
+; VI-NEXT: s_lshr_b32 s46, s26, 8
+; VI-NEXT: s_and_b32 s24, s24, 0xffff
+; VI-NEXT: v_writelane_b32 v21, s46, 14
+; VI-NEXT: s_lshr_b32 s46, s25, 24
+; VI-NEXT: s_and_b32 s63, s23, 0xffff0000
+; VI-NEXT: s_add_i32 s23, s23, 3
+; VI-NEXT: s_or_b32 s24, s60, s24
+; VI-NEXT: v_writelane_b32 v21, s46, 15
+; VI-NEXT: s_lshr_b32 s46, s25, 16
+; VI-NEXT: s_and_b32 s23, s23, 0xffff
+; VI-NEXT: s_add_i32 s24, s24, 0x30000
+; VI-NEXT: v_writelane_b32 v21, s46, 16
+; VI-NEXT: s_lshr_b32 s46, s25, 8
+; VI-NEXT: s_or_b32 s23, s63, s23
+; VI-NEXT: v_writelane_b32 v21, s46, 17
+; VI-NEXT: s_lshr_b32 s46, s24, 16
+; VI-NEXT: s_and_b32 s62, s22, 0xffff0000
+; VI-NEXT: s_add_i32 s22, s22, 3
+; VI-NEXT: s_add_i32 s23, s23, 0x30000
+; VI-NEXT: v_writelane_b32 v21, s46, 18
+; VI-NEXT: s_lshr_b32 s46, s24, 8
+; VI-NEXT: s_and_b32 s22, s22, 0xffff
+; VI-NEXT: v_writelane_b32 v21, s46, 19
+; VI-NEXT: s_lshr_b32 s46, s23, 24
+; VI-NEXT: s_and_b32 s73, s21, 0xffff0000
+; VI-NEXT: s_add_i32 s21, s21, 3
+; VI-NEXT: s_or_b32 s22, s62, s22
+; VI-NEXT: v_writelane_b32 v21, s46, 20
+; VI-NEXT: s_lshr_b32 s46, s23, 16
+; VI-NEXT: s_and_b32 s79, s5, 0xffff0000
+; VI-NEXT: s_add_i32 s5, s5, 3
+; VI-NEXT: s_and_b32 s21, s21, 0xffff
+; VI-NEXT: s_add_i32 s22, s22, 0x30000
+; VI-NEXT: v_writelane_b32 v21, s46, 21
+; VI-NEXT: s_lshr_b32 s46, s23, 8
+; VI-NEXT: s_and_b32 s5, s5, 0xffff
+; VI-NEXT: s_or_b32 s21, s73, s21
+; VI-NEXT: v_writelane_b32 v21, s46, 22
+; VI-NEXT: s_lshr_b32 s46, s22, 16
+; VI-NEXT: s_or_b32 s5, s79, s5
+; VI-NEXT: s_add_i32 s21, s21, 0x30000
+; VI-NEXT: v_writelane_b32 v21, s46, 23
+; VI-NEXT: s_lshr_b32 s46, s22, 8
+; VI-NEXT: s_and_b32 s78, s4, 0xffff0000
+; VI-NEXT: s_add_i32 s4, s4, 3
+; VI-NEXT: s_add_i32 s5, s5, 0x30000
+; VI-NEXT: v_writelane_b32 v21, s46, 24
+; VI-NEXT: s_lshr_b32 s46, s21, 24
+; VI-NEXT: s_and_b32 s4, s4, 0xffff
+; VI-NEXT: v_writelane_b32 v21, s46, 25
+; VI-NEXT: s_lshr_b32 s46, s5, 24
+; VI-NEXT: s_and_b32 s89, s7, 0xffff0000
+; VI-NEXT: s_add_i32 s7, s7, 3
+; VI-NEXT: s_or_b32 s4, s78, s4
+; VI-NEXT: v_writelane_b32 v21, s46, 26
+; VI-NEXT: s_lshr_b32 s46, s5, 16
+; VI-NEXT: s_and_b32 s7, s7, 0xffff
+; VI-NEXT: s_add_i32 s4, s4, 0x30000
+; VI-NEXT: v_writelane_b32 v21, s46, 27
+; VI-NEXT: s_lshr_b32 s46, s5, 8
+; VI-NEXT: s_or_b32 s7, s89, s7
+; VI-NEXT: v_writelane_b32 v21, s46, 28
+; VI-NEXT: s_lshr_b32 s46, s4, 16
+; VI-NEXT: s_and_b32 s88, s6, 0xffff0000
+; VI-NEXT: s_add_i32 s6, s6, 3
+; VI-NEXT: s_add_i32 s7, s7, 0x30000
+; VI-NEXT: v_writelane_b32 v21, s46, 29
+; VI-NEXT: s_lshr_b32 s46, s4, 8
+; VI-NEXT: s_and_b32 s6, s6, 0xffff
+; VI-NEXT: v_writelane_b32 v21, s46, 30
+; VI-NEXT: s_lshr_b32 s46, s7, 24
+; VI-NEXT: s_and_b32 s91, s9, 0xffff0000
+; VI-NEXT: s_add_i32 s9, s9, 3
+; VI-NEXT: s_or_b32 s6, s88, s6
+; VI-NEXT: v_writelane_b32 v21, s46, 31
+; VI-NEXT: s_lshr_b32 s46, s7, 16
+; VI-NEXT: s_and_b32 s9, s9, 0xffff
+; VI-NEXT: s_add_i32 s6, s6, 0x30000
+; VI-NEXT: v_writelane_b32 v21, s46, 32
+; VI-NEXT: s_lshr_b32 s46, s7, 8
+; VI-NEXT: s_or_b32 s9, s91, s9
+; VI-NEXT: v_writelane_b32 v21, s46, 33
+; VI-NEXT: s_lshr_b32 s46, s6, 16
+; VI-NEXT: s_and_b32 s90, s8, 0xffff0000
+; VI-NEXT: s_add_i32 s8, s8, 3
+; VI-NEXT: s_add_i32 s9, s9, 0x30000
+; VI-NEXT: v_writelane_b32 v21, s46, 34
+; VI-NEXT: s_lshr_b32 s46, s6, 8
+; VI-NEXT: s_and_b32 s8, s8, 0xffff
+; VI-NEXT: v_writelane_b32 v21, s46, 35
+; VI-NEXT: s_lshr_b32 s46, s9, 24
+; VI-NEXT: s_and_b32 vcc_hi, s11, 0xffff0000
+; VI-NEXT: s_add_i32 s11, s11, 3
+; VI-NEXT: s_or_b32 s8, s90, s8
+; VI-NEXT: v_writelane_b32 v21, s46, 36
+; VI-NEXT: s_lshr_b32 s46, s9, 16
+; VI-NEXT: s_and_b32 s11, s11, 0xffff
+; VI-NEXT: s_add_i32 s8, s8, 0x30000
+; VI-NEXT: v_writelane_b32 v21, s46, 37
+; VI-NEXT: s_lshr_b32 s46, s9, 8
+; VI-NEXT: s_or_b32 s11, vcc_hi, s11
+; VI-NEXT: v_writelane_b32 v21, s46, 38
+; VI-NEXT: s_lshr_b32 s46, s8, 16
+; VI-NEXT: s_and_b32 vcc_lo, s10, 0xffff0000
+; VI-NEXT: s_add_i32 s10, s10, 3
+; VI-NEXT: s_add_i32 s11, s11, 0x30000
+; VI-NEXT: v_writelane_b32 v21, s46, 39
+; VI-NEXT: s_lshr_b32 s46, s8, 8
+; VI-NEXT: s_and_b32 s10, s10, 0xffff
+; VI-NEXT: v_writelane_b32 v21, s46, 40
+; VI-NEXT: s_lshr_b32 s46, s11, 24
+; VI-NEXT: s_and_b32 s31, s13, 0xffff0000
+; VI-NEXT: s_add_i32 s13, s13, 3
+; VI-NEXT: s_or_b32 s10, vcc_lo, s10
+; VI-NEXT: v_writelane_b32 v21, s46, 41
+; VI-NEXT: s_lshr_b32 s46, s11, 16
+; VI-NEXT: s_and_b32 s13, s13, 0xffff
+; VI-NEXT: s_add_i32 s10, s10, 0x30000
+; VI-NEXT: v_writelane_b32 v21, s46, 42
+; VI-NEXT: s_lshr_b32 s46, s11, 8
+; VI-NEXT: s_or_b32 s13, s31, s13
+; VI-NEXT: v_writelane_b32 v21, s46, 43
+; VI-NEXT: s_lshr_b32 s46, s10, 16
+; VI-NEXT: s_and_b32 s30, s12, 0xffff0000
+; VI-NEXT: s_add_i32 s12, s12, 3
+; VI-NEXT: s_add_i32 s13, s13, 0x30000
+; VI-NEXT: v_writelane_b32 v21, s46, 44
+; VI-NEXT: s_lshr_b32 s46, s10, 8
+; VI-NEXT: s_and_b32 s12, s12, 0xffff
+; VI-NEXT: v_writelane_b32 v21, s46, 45
+; VI-NEXT: s_lshr_b32 s46, s13, 24
+; VI-NEXT: s_and_b32 s35, s15, 0xffff0000
+; VI-NEXT: s_add_i32 s15, s15, 3
+; VI-NEXT: s_or_b32 s12, s30, s12
+; VI-NEXT: v_writelane_b32 v21, s46, 46
+; VI-NEXT: s_lshr_b32 s46, s13, 16
+; VI-NEXT: s_and_b32 s15, s15, 0xffff
+; VI-NEXT: s_add_i32 s12, s12, 0x30000
+; VI-NEXT: v_writelane_b32 v21, s46, 47
+; VI-NEXT: s_lshr_b32 s46, s13, 8
+; VI-NEXT: s_or_b32 s15, s35, s15
+; VI-NEXT: v_writelane_b32 v21, s46, 48
+; VI-NEXT: s_lshr_b32 s46, s12, 16
+; VI-NEXT: s_and_b32 s34, s14, 0xffff0000
+; VI-NEXT: s_add_i32 s14, s14, 3
+; VI-NEXT: s_add_i32 s15, s15, 0x30000
+; VI-NEXT: v_writelane_b32 v21, s46, 49
+; VI-NEXT: s_lshr_b32 s46, s12, 8
+; VI-NEXT: s_and_b32 s14, s14, 0xffff
+; VI-NEXT: v_writelane_b32 v21, s46, 50
+; VI-NEXT: s_lshr_b32 s46, s15, 24
+; VI-NEXT: s_and_b32 s37, s41, 0xffff0000
+; VI-NEXT: s_add_i32 s41, s41, 3
+; VI-NEXT: s_or_b32 s14, s34, s14
+; VI-NEXT: v_writelane_b32 v21, s46, 51
+; VI-NEXT: s_lshr_b32 s46, s15, 16
+; VI-NEXT: s_and_b32 s41, s41, 0xffff
+; VI-NEXT: s_add_i32 s14, s14, 0x30000
+; VI-NEXT: v_writelane_b32 v21, s46, 52
+; VI-NEXT: s_lshr_b32 s46, s15, 8
+; VI-NEXT: s_and_b32 s72, s20, 0xffff0000
+; VI-NEXT: s_add_i32 s20, s20, 3
+; VI-NEXT: s_and_b32 s74, s18, 0xffff0000
+; VI-NEXT: s_add_i32 s18, s18, 3
+; VI-NEXT: s_and_b32 s75, s19, 0xffff0000
+; VI-NEXT: s_add_i32 s19, s19, 3
+; VI-NEXT: s_and_b32 s76, s16, 0xffff0000
+; VI-NEXT: s_add_i32 s16, s16, 3
+; VI-NEXT: s_and_b32 s77, s17, 0xffff0000
+; VI-NEXT: s_add_i32 s17, s17, 3
+; VI-NEXT: s_and_b32 s36, s40, 0xffff0000
+; VI-NEXT: s_add_i32 s40, s40, 3
+; VI-NEXT: s_and_b32 s38, s42, 0xffff0000
+; VI-NEXT: s_add_i32 s42, s42, 3
+; VI-NEXT: s_and_b32 s39, s43, 0xffff0000
+; VI-NEXT: s_add_i32 s43, s43, 3
+; VI-NEXT: s_or_b32 s41, s37, s41
+; VI-NEXT: v_writelane_b32 v21, s46, 53
+; VI-NEXT: s_lshr_b32 s46, s14, 16
+; VI-NEXT: s_and_b32 s43, s43, 0xffff
+; VI-NEXT: s_and_b32 s42, s42, 0xffff
+; VI-NEXT: s_add_i32 s41, s41, 0x30000
+; VI-NEXT: s_and_b32 s40, s40, 0xffff
+; VI-NEXT: s_and_b32 s17, s17, 0xffff
+; VI-NEXT: s_and_b32 s16, s16, 0xffff
+; VI-NEXT: s_and_b32 s19, s19, 0xffff
+; VI-NEXT: s_and_b32 s18, s18, 0xffff
+; VI-NEXT: s_and_b32 s20, s20, 0xffff
+; VI-NEXT: v_writelane_b32 v21, s46, 54
+; VI-NEXT: s_lshr_b32 s46, s14, 8
+; VI-NEXT: s_or_b32 s43, s39, s43
+; VI-NEXT: s_or_b32 s42, s38, s42
+; VI-NEXT: s_or_b32 s40, s36, s40
+; VI-NEXT: s_or_b32 s17, s77, s17
+; VI-NEXT: s_or_b32 s16, s76, s16
+; VI-NEXT: s_or_b32 s19, s75, s19
+; VI-NEXT: s_or_b32 s18, s74, s18
+; VI-NEXT: s_or_b32 s20, s72, s20
+; VI-NEXT: v_writelane_b32 v21, s46, 55
+; VI-NEXT: s_lshr_b32 s46, s41, 24
+; VI-NEXT: s_add_i32 s43, s43, 0x30000
+; VI-NEXT: s_add_i32 s42, s42, 0x30000
+; VI-NEXT: s_add_i32 s40, s40, 0x30000
+; VI-NEXT: s_add_i32 s17, s17, 0x30000
+; VI-NEXT: s_add_i32 s16, s16, 0x30000
+; VI-NEXT: s_add_i32 s19, s19, 0x30000
+; VI-NEXT: s_add_i32 s18, s18, 0x30000
+; VI-NEXT: s_add_i32 s20, s20, 0x30000
+; VI-NEXT: v_writelane_b32 v21, s46, 56
+; VI-NEXT: s_lshr_b32 s46, s41, 16
+; VI-NEXT: s_lshr_b32 s80, s21, 16
+; VI-NEXT: s_lshr_b32 s82, s21, 8
+; VI-NEXT: s_lshr_b32 s84, s20, 16
+; VI-NEXT: s_lshr_b32 s86, s20, 8
+; VI-NEXT: s_lshr_b32 s51, s19, 24
+; VI-NEXT: s_lshr_b32 s53, s19, 16
+; VI-NEXT: s_lshr_b32 s54, s19, 8
+; VI-NEXT: s_lshr_b32 s65, s18, 16
+; VI-NEXT: s_lshr_b32 s66, s18, 8
+; VI-NEXT: s_lshr_b32 s67, s17, 24
+; VI-NEXT: s_lshr_b32 s68, s17, 16
+; VI-NEXT: s_lshr_b32 s69, s17, 8
+; VI-NEXT: s_lshr_b32 s70, s16, 16
+; VI-NEXT: s_lshr_b32 s71, s16, 8
+; VI-NEXT: v_writelane_b32 v21, s46, 57
+; VI-NEXT: s_lshr_b32 s81, s41, 8
+; VI-NEXT: s_lshr_b32 s83, s40, 16
+; VI-NEXT: s_lshr_b32 s85, s40, 8
+; VI-NEXT: s_lshr_b32 s87, s43, 24
+; VI-NEXT: s_lshr_b32 s50, s43, 16
+; VI-NEXT: s_lshr_b32 s52, s43, 8
+; VI-NEXT: s_lshr_b32 s55, s42, 16
+; VI-NEXT: s_lshr_b32 s64, s42, 8
+; VI-NEXT: s_lshr_b64 s[76:77], s[44:45], 24
+; VI-NEXT: s_lshr_b64 s[88:89], s[28:29], 24
+; VI-NEXT: s_lshr_b64 s[90:91], s[26:27], 24
+; VI-NEXT: s_lshr_b64 s[30:31], s[24:25], 24
+; VI-NEXT: s_lshr_b64 s[34:35], s[22:23], 24
+; VI-NEXT: s_lshr_b64 s[36:37], s[20:21], 24
+; VI-NEXT: s_lshr_b64 s[38:39], s[18:19], 24
+; VI-NEXT: s_lshr_b64 s[48:49], s[16:17], 24
+; VI-NEXT: s_lshr_b64 s[46:47], s[4:5], 24
+; VI-NEXT: s_lshr_b64 s[56:57], s[6:7], 24
+; VI-NEXT: s_lshr_b64 s[58:59], s[8:9], 24
+; VI-NEXT: s_lshr_b64 s[60:61], s[10:11], 24
+; VI-NEXT: s_lshr_b64 s[62:63], s[12:13], 24
+; VI-NEXT: s_lshr_b64 s[72:73], s[14:15], 24
+; VI-NEXT: s_lshr_b64 s[74:75], s[40:41], 24
+; VI-NEXT: s_lshr_b64 s[78:79], s[42:43], 24
+; VI-NEXT: .LBB99_3: ; %end
+; VI-NEXT: s_lshl_b32 s47, s71, 8
+; VI-NEXT: s_and_b32 s16, s16, 0xff
+; VI-NEXT: s_or_b32 s16, s16, s47
+; VI-NEXT: s_lshl_b32 s47, s48, 8
+; VI-NEXT: s_and_b32 s57, s70, 0xff
+; VI-NEXT: s_or_b32 s47, s57, s47
+; VI-NEXT: s_and_b32 s16, s16, 0xffff
+; VI-NEXT: s_lshl_b32 s47, s47, 16
+; VI-NEXT: s_or_b32 s16, s16, s47
+; VI-NEXT: v_mov_b32_e32 v1, s16
+; VI-NEXT: s_and_b32 s16, s17, 0xff
+; VI-NEXT: s_lshl_b32 s17, s69, 8
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: s_and_b32 s17, s68, 0xff
+; VI-NEXT: s_lshl_b32 s47, s67, 8
+; VI-NEXT: s_or_b32 s17, s17, s47
+; VI-NEXT: s_and_b32 s16, s16, 0xffff
+; VI-NEXT: s_lshl_b32 s17, s17, 16
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: v_mov_b32_e32 v2, s16
+; VI-NEXT: s_lshl_b32 s16, s66, 8
+; VI-NEXT: s_and_b32 s17, s18, 0xff
+; VI-NEXT: s_or_b32 s16, s17, s16
+; VI-NEXT: s_lshl_b32 s17, s38, 8
+; VI-NEXT: s_and_b32 s18, s65, 0xff
+; VI-NEXT: s_or_b32 s17, s18, s17
+; VI-NEXT: s_and_b32 s16, s16, 0xffff
+; VI-NEXT: s_lshl_b32 s17, s17, 16
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: v_mov_b32_e32 v3, s16
+; VI-NEXT: s_and_b32 s16, s19, 0xff
+; VI-NEXT: s_lshl_b32 s17, s54, 8
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: s_and_b32 s17, s53, 0xff
+; VI-NEXT: s_lshl_b32 s18, s51, 8
+; VI-NEXT: s_or_b32 s17, s17, s18
+; VI-NEXT: s_and_b32 s16, s16, 0xffff
+; VI-NEXT: s_lshl_b32 s17, s17, 16
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: v_mov_b32_e32 v4, s16
+; VI-NEXT: s_lshl_b32 s16, s86, 8
+; VI-NEXT: s_and_b32 s17, s20, 0xff
+; VI-NEXT: s_or_b32 s16, s17, s16
+; VI-NEXT: s_lshl_b32 s17, s36, 8
+; VI-NEXT: s_and_b32 s18, s84, 0xff
+; VI-NEXT: s_or_b32 s17, s18, s17
+; VI-NEXT: s_and_b32 s16, s16, 0xffff
+; VI-NEXT: s_lshl_b32 s17, s17, 16
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: v_mov_b32_e32 v5, s16
+; VI-NEXT: s_and_b32 s16, s21, 0xff
+; VI-NEXT: s_lshl_b32 s17, s82, 8
+; VI-NEXT: v_readlane_b32 s18, v21, 25
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: s_and_b32 s17, s80, 0xff
+; VI-NEXT: s_lshl_b32 s18, s18, 8
+; VI-NEXT: s_or_b32 s17, s17, s18
+; VI-NEXT: s_and_b32 s16, s16, 0xffff
+; VI-NEXT: s_lshl_b32 s17, s17, 16
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: v_mov_b32_e32 v6, s16
+; VI-NEXT: v_readlane_b32 s16, v21, 24
+; VI-NEXT: s_lshl_b32 s16, s16, 8
+; VI-NEXT: s_and_b32 s17, s22, 0xff
+; VI-NEXT: v_readlane_b32 s18, v21, 23
+; VI-NEXT: s_or_b32 s16, s17, s16
+; VI-NEXT: s_lshl_b32 s17, s34, 8
+; VI-NEXT: s_and_b32 s18, s18, 0xff
+; VI-NEXT: s_or_b32 s17, s18, s17
+; VI-NEXT: s_and_b32 s16, s16, 0xffff
+; VI-NEXT: s_lshl_b32 s17, s17, 16
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: v_readlane_b32 s17, v21, 22
+; VI-NEXT: v_mov_b32_e32 v7, s16
+; VI-NEXT: s_and_b32 s16, s23, 0xff
+; VI-NEXT: s_lshl_b32 s17, s17, 8
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: v_readlane_b32 s17, v21, 21
+; VI-NEXT: v_readlane_b32 s18, v21, 20
+; VI-NEXT: s_and_b32 s17, s17, 0xff
+; VI-NEXT: s_lshl_b32 s18, s18, 8
+; VI-NEXT: s_or_b32 s17, s17, s18
+; VI-NEXT: s_and_b32 s16, s16, 0xffff
+; VI-NEXT: s_lshl_b32 s17, s17, 16
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: v_mov_b32_e32 v8, s16
+; VI-NEXT: v_readlane_b32 s16, v21, 19
+; VI-NEXT: s_lshl_b32 s16, s16, 8
+; VI-NEXT: s_and_b32 s17, s24, 0xff
+; VI-NEXT: v_readlane_b32 s18, v21, 18
+; VI-NEXT: s_or_b32 s16, s17, s16
+; VI-NEXT: s_lshl_b32 s17, s30, 8
+; VI-NEXT: s_and_b32 s18, s18, 0xff
+; VI-NEXT: s_or_b32 s17, s18, s17
+; VI-NEXT: s_and_b32 s16, s16, 0xffff
+; VI-NEXT: s_lshl_b32 s17, s17, 16
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: v_readlane_b32 s17, v21, 17
+; VI-NEXT: v_mov_b32_e32 v9, s16
+; VI-NEXT: s_and_b32 s16, s25, 0xff
+; VI-NEXT: s_lshl_b32 s17, s17, 8
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: v_readlane_b32 s17, v21, 16
+; VI-NEXT: v_readlane_b32 s18, v21, 15
+; VI-NEXT: s_and_b32 s17, s17, 0xff
+; VI-NEXT: s_lshl_b32 s18, s18, 8
+; VI-NEXT: s_or_b32 s17, s17, s18
+; VI-NEXT: s_and_b32 s16, s16, 0xffff
+; VI-NEXT: s_lshl_b32 s17, s17, 16
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: v_mov_b32_e32 v10, s16
+; VI-NEXT: v_readlane_b32 s16, v21, 14
+; VI-NEXT: s_lshl_b32 s16, s16, 8
+; VI-NEXT: s_and_b32 s17, s26, 0xff
+; VI-NEXT: v_readlane_b32 s18, v21, 13
+; VI-NEXT: s_or_b32 s16, s17, s16
+; VI-NEXT: s_lshl_b32 s17, s90, 8
+; VI-NEXT: s_and_b32 s18, s18, 0xff
+; VI-NEXT: s_or_b32 s17, s18, s17
+; VI-NEXT: s_and_b32 s16, s16, 0xffff
+; VI-NEXT: s_lshl_b32 s17, s17, 16
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: v_readlane_b32 s17, v21, 12
+; VI-NEXT: v_mov_b32_e32 v11, s16
+; VI-NEXT: s_and_b32 s16, s27, 0xff
+; VI-NEXT: s_lshl_b32 s17, s17, 8
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: v_readlane_b32 s17, v21, 11
+; VI-NEXT: v_readlane_b32 s18, v21, 10
+; VI-NEXT: s_and_b32 s17, s17, 0xff
+; VI-NEXT: s_lshl_b32 s18, s18, 8
+; VI-NEXT: s_or_b32 s17, s17, s18
+; VI-NEXT: s_and_b32 s16, s16, 0xffff
+; VI-NEXT: s_lshl_b32 s17, s17, 16
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: v_mov_b32_e32 v12, s16
+; VI-NEXT: v_readlane_b32 s16, v21, 9
+; VI-NEXT: s_lshl_b32 s16, s16, 8
+; VI-NEXT: s_and_b32 s17, s28, 0xff
+; VI-NEXT: v_readlane_b32 s18, v21, 8
+; VI-NEXT: s_or_b32 s16, s17, s16
+; VI-NEXT: s_lshl_b32 s17, s88, 8
+; VI-NEXT: s_and_b32 s18, s18, 0xff
+; VI-NEXT: s_or_b32 s17, s18, s17
+; VI-NEXT: s_and_b32 s16, s16, 0xffff
+; VI-NEXT: s_lshl_b32 s17, s17, 16
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: v_readlane_b32 s17, v21, 7
+; VI-NEXT: v_mov_b32_e32 v13, s16
+; VI-NEXT: s_and_b32 s16, s29, 0xff
+; VI-NEXT: s_lshl_b32 s17, s17, 8
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: v_readlane_b32 s17, v21, 6
+; VI-NEXT: v_readlane_b32 s18, v21, 5
+; VI-NEXT: s_and_b32 s17, s17, 0xff
+; VI-NEXT: s_lshl_b32 s18, s18, 8
+; VI-NEXT: s_or_b32 s17, s17, s18
+; VI-NEXT: s_and_b32 s16, s16, 0xffff
+; VI-NEXT: s_lshl_b32 s17, s17, 16
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: v_mov_b32_e32 v14, s16
+; VI-NEXT: v_readlane_b32 s16, v21, 4
+; VI-NEXT: s_lshl_b32 s16, s16, 8
+; VI-NEXT: s_and_b32 s17, s44, 0xff
+; VI-NEXT: v_readlane_b32 s18, v21, 3
+; VI-NEXT: s_or_b32 s16, s17, s16
+; VI-NEXT: s_lshl_b32 s17, s76, 8
+; VI-NEXT: s_and_b32 s18, s18, 0xff
+; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; VI-NEXT: v_add_u32_e32 v1, vcc, 4, v0
+; VI-NEXT: s_or_b32 s17, s18, s17
+; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
+; VI-NEXT: v_add_u32_e32 v1, vcc, 8, v0
+; VI-NEXT: s_and_b32 s16, s16, 0xffff
+; VI-NEXT: s_lshl_b32 s17, s17, 16
+; VI-NEXT: buffer_store_dword v3, v1, s[0:3], 0 offen
+; VI-NEXT: v_add_u32_e32 v1, vcc, 12, v0
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: v_readlane_b32 s17, v21, 2
+; VI-NEXT: buffer_store_dword v4, v1, s[0:3], 0 offen
+; VI-NEXT: v_add_u32_e32 v1, vcc, 16, v0
+; VI-NEXT: v_mov_b32_e32 v15, s16
+; VI-NEXT: s_and_b32 s16, s45, 0xff
+; VI-NEXT: s_lshl_b32 s17, s17, 8
+; VI-NEXT: buffer_store_dword v5, v1, s[0:3], 0 offen
+; VI-NEXT: v_add_u32_e32 v1, vcc, 20, v0
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: buffer_store_dword v6, v1, s[0:3], 0 offen
+; VI-NEXT: v_add_u32_e32 v1, vcc, 24, v0
+; VI-NEXT: v_readlane_b32 s17, v21, 1
+; VI-NEXT: v_readlane_b32 s18, v21, 0
+; VI-NEXT: buffer_store_dword v7, v1, s[0:3], 0 offen
+; VI-NEXT: v_add_u32_e32 v1, vcc, 28, v0
+; VI-NEXT: s_and_b32 s17, s17, 0xff
+; VI-NEXT: s_lshl_b32 s18, s18, 8
+; VI-NEXT: buffer_store_dword v8, v1, s[0:3], 0 offen
+; VI-NEXT: v_add_u32_e32 v1, vcc, 32, v0
+; VI-NEXT: s_or_b32 s17, s17, s18
+; VI-NEXT: buffer_store_dword v9, v1, s[0:3], 0 offen
+; VI-NEXT: v_add_u32_e32 v1, vcc, 36, v0
+; VI-NEXT: s_and_b32 s16, s16, 0xffff
+; VI-NEXT: s_lshl_b32 s17, s17, 16
+; VI-NEXT: buffer_store_dword v10, v1, s[0:3], 0 offen
+; VI-NEXT: v_add_u32_e32 v1, vcc, 40, v0
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: buffer_store_dword v11, v1, s[0:3], 0 offen
+; VI-NEXT: v_add_u32_e32 v1, vcc, 44, v0
+; VI-NEXT: v_mov_b32_e32 v2, s16
+; VI-NEXT: s_and_b32 s16, s42, 0xff
+; VI-NEXT: s_lshl_b32 s17, s64, 8
+; VI-NEXT: buffer_store_dword v12, v1, s[0:3], 0 offen
+; VI-NEXT: v_add_u32_e32 v1, vcc, 48, v0
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: s_and_b32 s17, s55, 0xff
+; VI-NEXT: s_lshl_b32 s18, s78, 8
+; VI-NEXT: buffer_store_dword v13, v1, s[0:3], 0 offen
+; VI-NEXT: v_add_u32_e32 v1, vcc, 52, v0
+; VI-NEXT: s_or_b32 s17, s17, s18
+; VI-NEXT: buffer_store_dword v14, v1, s[0:3], 0 offen
+; VI-NEXT: v_add_u32_e32 v1, vcc, 56, v0
+; VI-NEXT: s_and_b32 s16, s16, 0xffff
+; VI-NEXT: s_lshl_b32 s17, s17, 16
+; VI-NEXT: buffer_store_dword v15, v1, s[0:3], 0 offen
+; VI-NEXT: v_add_u32_e32 v1, vcc, 60, v0
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
+; VI-NEXT: v_mov_b32_e32 v2, s16
+; VI-NEXT: s_and_b32 s16, s43, 0xff
+; VI-NEXT: s_lshl_b32 s17, s52, 8
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: s_and_b32 s17, s50, 0xff
+; VI-NEXT: s_lshl_b32 s18, s87, 8
+; VI-NEXT: s_or_b32 s17, s17, s18
+; VI-NEXT: s_and_b32 s16, s16, 0xffff
+; VI-NEXT: s_lshl_b32 s17, s17, 16
+; VI-NEXT: v_add_u32_e32 v1, vcc, 64, v0
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
+; VI-NEXT: v_mov_b32_e32 v2, s16
+; VI-NEXT: s_and_b32 s16, s40, 0xff
+; VI-NEXT: s_lshl_b32 s17, s85, 8
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: s_and_b32 s17, s83, 0xff
+; VI-NEXT: s_lshl_b32 s18, s74, 8
+; VI-NEXT: s_or_b32 s17, s17, s18
+; VI-NEXT: s_and_b32 s16, s16, 0xffff
+; VI-NEXT: s_lshl_b32 s17, s17, 16
+; VI-NEXT: v_add_u32_e32 v1, vcc, 0x44, v0
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
+; VI-NEXT: v_mov_b32_e32 v2, s16
+; VI-NEXT: s_and_b32 s16, s41, 0xff
+; VI-NEXT: s_lshl_b32 s17, s81, 8
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: v_readlane_b32 s17, v21, 57
+; VI-NEXT: v_readlane_b32 s18, v21, 56
+; VI-NEXT: s_and_b32 s17, s17, 0xff
+; VI-NEXT: s_lshl_b32 s18, s18, 8
+; VI-NEXT: s_or_b32 s17, s17, s18
+; VI-NEXT: s_and_b32 s16, s16, 0xffff
+; VI-NEXT: s_lshl_b32 s17, s17, 16
+; VI-NEXT: v_add_u32_e32 v1, vcc, 0x48, v0
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
+; VI-NEXT: v_mov_b32_e32 v2, s16
+; VI-NEXT: v_readlane_b32 s16, v21, 55
+; VI-NEXT: s_and_b32 s14, s14, 0xff
+; VI-NEXT: s_lshl_b32 s16, s16, 8
+; VI-NEXT: s_or_b32 s14, s14, s16
+; VI-NEXT: v_readlane_b32 s16, v21, 54
+; VI-NEXT: s_and_b32 s16, s16, 0xff
+; VI-NEXT: s_lshl_b32 s17, s72, 8
+; VI-NEXT: s_or_b32 s16, s16, s17
+; VI-NEXT: s_and_b32 s14, s14, 0xffff
+; VI-NEXT: s_lshl_b32 s16, s16, 16
+; VI-NEXT: v_add_u32_e32 v1, vcc, 0x4c, v0
+; VI-NEXT: s_or_b32 s14, s14, s16
+; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
+; VI-NEXT: v_mov_b32_e32 v2, s14
+; VI-NEXT: s_and_b32 s14, s15, 0xff
+; VI-NEXT: v_readlane_b32 s15, v21, 53
+; VI-NEXT: s_lshl_b32 s15, s15, 8
+; VI-NEXT: s_or_b32 s14, s14, s15
+; VI-NEXT: v_readlane_b32 s15, v21, 52
+; VI-NEXT: v_readlane_b32 s16, v21, 51
+; VI-NEXT: s_and_b32 s15, s15, 0xff
+; VI-NEXT: s_lshl_b32 s16, s16, 8
+; VI-NEXT: s_or_b32 s15, s15, s16
+; VI-NEXT: s_and_b32 s14, s14, 0xffff
+; VI-NEXT: s_lshl_b32 s15, s15, 16
+; VI-NEXT: v_add_u32_e32 v1, vcc, 0x50, v0
+; VI-NEXT: s_or_b32 s14, s14, s15
+; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
+; VI-NEXT: v_mov_b32_e32 v2, s14
+; VI-NEXT: v_readlane_b32 s14, v21, 50
+; VI-NEXT: s_and_b32 s12, s12, 0xff
+; VI-NEXT: s_lshl_b32 s14, s14, 8
+; VI-NEXT: s_or_b32 s12, s12, s14
+; VI-NEXT: v_readlane_b32 s14, v21, 49
+; VI-NEXT: s_and_b32 s14, s14, 0xff
+; VI-NEXT: s_lshl_b32 s15, s62, 8
+; VI-NEXT: s_or_b32 s14, s14, s15
+; VI-NEXT: s_and_b32 s12, s12, 0xffff
+; VI-NEXT: s_lshl_b32 s14, s14, 16
+; VI-NEXT: v_add_u32_e32 v1, vcc, 0x54, v0
+; VI-NEXT: s_or_b32 s12, s12, s14
+; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
+; VI-NEXT: v_mov_b32_e32 v2, s12
+; VI-NEXT: s_and_b32 s12, s13, 0xff
+; VI-NEXT: v_readlane_b32 s13, v21, 48
+; VI-NEXT: s_lshl_b32 s13, s13, 8
+; VI-NEXT: s_or_b32 s12, s12, s13
+; VI-NEXT: v_readlane_b32 s13, v21, 47
+; VI-NEXT: v_readlane_b32 s14, v21, 46
+; VI-NEXT: s_and_b32 s13, s13, 0xff
+; VI-NEXT: s_lshl_b32 s14, s14, 8
+; VI-NEXT: s_or_b32 s13, s13, s14
+; VI-NEXT: s_and_b32 s12, s12, 0xffff
+; VI-NEXT: s_lshl_b32 s13, s13, 16
+; VI-NEXT: v_add_u32_e32 v1, vcc, 0x58, v0
+; VI-NEXT: s_or_b32 s12, s12, s13
+; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
+; VI-NEXT: v_mov_b32_e32 v2, s12
+; VI-NEXT: v_readlane_b32 s12, v21, 45
+; VI-NEXT: s_and_b32 s10, s10, 0xff
+; VI-NEXT: s_lshl_b32 s12, s12, 8
+; VI-NEXT: s_or_b32 s10, s10, s12
+; VI-NEXT: v_readlane_b32 s12, v21, 44
+; VI-NEXT: s_and_b32 s12, s12, 0xff
+; VI-NEXT: s_lshl_b32 s13, s60, 8
+; VI-NEXT: s_or_b32 s12, s12, s13
+; VI-NEXT: s_and_b32 s10, s10, 0xffff
+; VI-NEXT: s_lshl_b32 s12, s12, 16
+; VI-NEXT: v_add_u32_e32 v1, vcc, 0x5c, v0
+; VI-NEXT: s_or_b32 s10, s10, s12
+; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
+; VI-NEXT: v_mov_b32_e32 v2, s10
+; VI-NEXT: s_and_b32 s10, s11, 0xff
+; VI-NEXT: v_readlane_b32 s11, v21, 43
+; VI-NEXT: s_lshl_b32 s11, s11, 8
+; VI-NEXT: s_or_b32 s10, s10, s11
+; VI-NEXT: v_readlane_b32 s11, v21, 42
+; VI-NEXT: v_readlane_b32 s12, v21, 41
+; VI-NEXT: s_and_b32 s11, s11, 0xff
+; VI-NEXT: s_lshl_b32 s12, s12, 8
+; VI-NEXT: s_or_b32 s11, s11, s12
+; VI-NEXT: s_and_b32 s10, s10, 0xffff
+; VI-NEXT: s_lshl_b32 s11, s11, 16
+; VI-NEXT: v_add_u32_e32 v1, vcc, 0x60, v0
+; VI-NEXT: s_or_b32 s10, s10, s11
+; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
+; VI-NEXT: v_mov_b32_e32 v2, s10
+; VI-NEXT: v_readlane_b32 s10, v21, 40
+; VI-NEXT: s_and_b32 s8, s8, 0xff
+; VI-NEXT: s_lshl_b32 s10, s10, 8
+; VI-NEXT: s_or_b32 s8, s8, s10
+; VI-NEXT: v_readlane_b32 s10, v21, 39
+; VI-NEXT: s_and_b32 s10, s10, 0xff
+; VI-NEXT: s_lshl_b32 s11, s58, 8
+; VI-NEXT: s_or_b32 s10, s10, s11
+; VI-NEXT: s_and_b32 s8, s8, 0xffff
+; VI-NEXT: s_lshl_b32 s10, s10, 16
+; VI-NEXT: v_add_u32_e32 v1, vcc, 0x64, v0
+; VI-NEXT: s_or_b32 s8, s8, s10
+; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
+; VI-NEXT: v_mov_b32_e32 v2, s8
+; VI-NEXT: s_and_b32 s8, s9, 0xff
+; VI-NEXT: v_readlane_b32 s9, v21, 38
+; VI-NEXT: s_lshl_b32 s9, s9, 8
+; VI-NEXT: s_or_b32 s8, s8, s9
+; VI-NEXT: v_readlane_b32 s9, v21, 37
+; VI-NEXT: v_readlane_b32 s10, v21, 36
+; VI-NEXT: s_and_b32 s9, s9, 0xff
+; VI-NEXT: s_lshl_b32 s10, s10, 8
+; VI-NEXT: s_or_b32 s9, s9, s10
+; VI-NEXT: s_and_b32 s8, s8, 0xffff
+; VI-NEXT: s_lshl_b32 s9, s9, 16
+; VI-NEXT: v_add_u32_e32 v1, vcc, 0x68, v0
+; VI-NEXT: s_or_b32 s8, s8, s9
+; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
+; VI-NEXT: v_mov_b32_e32 v2, s8
+; VI-NEXT: v_readlane_b32 s8, v21, 35
+; VI-NEXT: s_and_b32 s6, s6, 0xff
+; VI-NEXT: s_lshl_b32 s8, s8, 8
+; VI-NEXT: s_or_b32 s6, s6, s8
+; VI-NEXT: v_readlane_b32 s8, v21, 34
+; VI-NEXT: s_and_b32 s8, s8, 0xff
+; VI-NEXT: s_lshl_b32 s9, s56, 8
+; VI-NEXT: s_or_b32 s8, s8, s9
+; VI-NEXT: s_and_b32 s6, s6, 0xffff
+; VI-NEXT: s_lshl_b32 s8, s8, 16
+; VI-NEXT: v_add_u32_e32 v1, vcc, 0x6c, v0
+; VI-NEXT: s_or_b32 s6, s6, s8
+; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
+; VI-NEXT: v_mov_b32_e32 v2, s6
+; VI-NEXT: s_and_b32 s6, s7, 0xff
+; VI-NEXT: v_readlane_b32 s7, v21, 33
+; VI-NEXT: s_lshl_b32 s7, s7, 8
+; VI-NEXT: s_or_b32 s6, s6, s7
+; VI-NEXT: v_readlane_b32 s7, v21, 32
+; VI-NEXT: v_readlane_b32 s8, v21, 31
+; VI-NEXT: s_and_b32 s7, s7, 0xff
+; VI-NEXT: s_lshl_b32 s8, s8, 8
+; VI-NEXT: s_or_b32 s7, s7, s8
+; VI-NEXT: s_and_b32 s6, s6, 0xffff
+; VI-NEXT: s_lshl_b32 s7, s7, 16
+; VI-NEXT: v_add_u32_e32 v1, vcc, 0x70, v0
+; VI-NEXT: s_or_b32 s6, s6, s7
+; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
+; VI-NEXT: v_mov_b32_e32 v2, s6
+; VI-NEXT: v_readlane_b32 s6, v21, 30
+; VI-NEXT: s_and_b32 s4, s4, 0xff
+; VI-NEXT: s_lshl_b32 s6, s6, 8
+; VI-NEXT: s_or_b32 s4, s4, s6
+; VI-NEXT: v_readlane_b32 s6, v21, 29
+; VI-NEXT: s_and_b32 s6, s6, 0xff
+; VI-NEXT: s_lshl_b32 s7, s46, 8
+; VI-NEXT: s_or_b32 s6, s6, s7
+; VI-NEXT: s_and_b32 s4, s4, 0xffff
+; VI-NEXT: s_lshl_b32 s6, s6, 16
+; VI-NEXT: v_add_u32_e32 v1, vcc, 0x74, v0
+; VI-NEXT: s_or_b32 s4, s4, s6
+; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
+; VI-NEXT: v_mov_b32_e32 v2, s4
+; VI-NEXT: s_and_b32 s4, s5, 0xff
+; VI-NEXT: v_readlane_b32 s5, v21, 28
+; VI-NEXT: s_lshl_b32 s5, s5, 8
+; VI-NEXT: s_or_b32 s4, s4, s5
+; VI-NEXT: v_readlane_b32 s5, v21, 27
+; VI-NEXT: v_readlane_b32 s6, v21, 26
+; VI-NEXT: s_and_b32 s5, s5, 0xff
+; VI-NEXT: s_lshl_b32 s6, s6, 8
+; VI-NEXT: s_or_b32 s5, s5, s6
+; VI-NEXT: s_and_b32 s4, s4, 0xffff
+; VI-NEXT: s_lshl_b32 s5, s5, 16
+; VI-NEXT: v_add_u32_e32 v1, vcc, 0x78, v0
+; VI-NEXT: s_or_b32 s4, s4, s5
+; VI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
+; VI-NEXT: v_add_u32_e32 v0, vcc, 0x7c, v0
+; VI-NEXT: v_mov_b32_e32 v1, s4
+; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; VI-NEXT: v_readlane_b32 s87, v20, 31
+; VI-NEXT: v_readlane_b32 s86, v20, 30
+; VI-NEXT: v_readlane_b32 s85, v20, 29
+; VI-NEXT: v_readlane_b32 s84, v20, 28
+; VI-NEXT: v_readlane_b32 s83, v20, 27
+; VI-NEXT: v_readlane_b32 s82, v20, 26
+; VI-NEXT: v_readlane_b32 s81, v20, 25
+; VI-NEXT: v_readlane_b32 s80, v20, 24
+; VI-NEXT: v_readlane_b32 s71, v20, 23
+; VI-NEXT: v_readlane_b32 s70, v20, 22
+; VI-NEXT: v_readlane_b32 s69, v20, 21
+; VI-NEXT: v_readlane_b32 s68, v20, 20
+; VI-NEXT: v_readlane_b32 s67, v20, 19
+; VI-NEXT: v_readlane_b32 s66, v20, 18
+; VI-NEXT: v_readlane_b32 s65, v20, 17
+; VI-NEXT: v_readlane_b32 s64, v20, 16
+; VI-NEXT: v_readlane_b32 s55, v20, 15
+; VI-NEXT: v_readlane_b32 s54, v20, 14
+; VI-NEXT: v_readlane_b32 s53, v20, 13
+; VI-NEXT: v_readlane_b32 s52, v20, 12
+; VI-NEXT: v_readlane_b32 s51, v20, 11
+; VI-NEXT: v_readlane_b32 s50, v20, 10
+; VI-NEXT: v_readlane_b32 s49, v20, 9
+; VI-NEXT: v_readlane_b32 s48, v20, 8
+; VI-NEXT: v_readlane_b32 s39, v20, 7
+; VI-NEXT: v_readlane_b32 s38, v20, 6
+; VI-NEXT: v_readlane_b32 s37, v20, 5
+; VI-NEXT: v_readlane_b32 s36, v20, 4
+; VI-NEXT: v_readlane_b32 s35, v20, 3
+; VI-NEXT: v_readlane_b32 s34, v20, 2
+; VI-NEXT: v_readlane_b32 s31, v20, 1
+; VI-NEXT: v_readlane_b32 s30, v20, 0
+; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; VI-NEXT: s_mov_b64 exec, s[4:5]
+; VI-NEXT: s_waitcnt vmcnt(0)
+; VI-NEXT: s_setpc_b64 s[30:31]
+; VI-NEXT: .LBB99_4:
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr71
+; VI-NEXT: ; implicit-def: $sgpr70
+; VI-NEXT: ; implicit-def: $sgpr69
+; VI-NEXT: ; implicit-def: $sgpr68
+; VI-NEXT: ; implicit-def: $sgpr67
+; VI-NEXT: ; implicit-def: $sgpr66
+; VI-NEXT: ; implicit-def: $sgpr65
+; VI-NEXT: ; implicit-def: $sgpr54
+; VI-NEXT: ; implicit-def: $sgpr53
+; VI-NEXT: ; implicit-def: $sgpr51
+; VI-NEXT: ; implicit-def: $sgpr86
+; VI-NEXT: ; implicit-def: $sgpr84
+; VI-NEXT: ; implicit-def: $sgpr82
+; VI-NEXT: ; implicit-def: $sgpr80
+; VI-NEXT: ; implicit-def: $sgpr64
+; VI-NEXT: ; implicit-def: $sgpr55
+; VI-NEXT: ; implicit-def: $sgpr52
+; VI-NEXT: ; implicit-def: $sgpr50
+; VI-NEXT: ; implicit-def: $sgpr87
+; VI-NEXT: ; implicit-def: $sgpr85
+; VI-NEXT: ; implicit-def: $sgpr83
+; VI-NEXT: ; implicit-def: $sgpr81
+; VI-NEXT: ; implicit-def: $sgpr48
+; VI-NEXT: ; implicit-def: $sgpr38
+; VI-NEXT: ; implicit-def: $sgpr36
+; VI-NEXT: ; implicit-def: $sgpr34
+; VI-NEXT: ; implicit-def: $sgpr30
+; VI-NEXT: ; implicit-def: $sgpr90
+; VI-NEXT: ; implicit-def: $sgpr88
+; VI-NEXT: ; implicit-def: $sgpr76
+; VI-NEXT: ; implicit-def: $sgpr78
+; VI-NEXT: ; implicit-def: $sgpr74
+; VI-NEXT: ; implicit-def: $sgpr72
+; VI-NEXT: ; implicit-def: $sgpr62
+; VI-NEXT: ; implicit-def: $sgpr60
+; VI-NEXT: ; implicit-def: $sgpr58
+; VI-NEXT: ; implicit-def: $sgpr56
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; kill: killed $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: s_branch .LBB99_2
+;
+; GFX9-LABEL: bitcast_v64i16_to_v128i8_scalar:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
+; GFX9-NEXT: s_mov_b64 exec, s[4:5]
+; GFX9-NEXT: v_writelane_b32 v63, s30, 0
+; GFX9-NEXT: v_writelane_b32 v63, s31, 1
+; GFX9-NEXT: v_writelane_b32 v63, s34, 2
+; GFX9-NEXT: v_writelane_b32 v63, s35, 3
+; GFX9-NEXT: v_writelane_b32 v63, s36, 4
+; GFX9-NEXT: v_writelane_b32 v63, s37, 5
+; GFX9-NEXT: v_writelane_b32 v63, s38, 6
+; GFX9-NEXT: v_writelane_b32 v63, s39, 7
+; GFX9-NEXT: v_writelane_b32 v63, s48, 8
+; GFX9-NEXT: v_writelane_b32 v63, s49, 9
+; GFX9-NEXT: v_writelane_b32 v63, s50, 10
+; GFX9-NEXT: v_writelane_b32 v63, s51, 11
+; GFX9-NEXT: v_writelane_b32 v63, s52, 12
+; GFX9-NEXT: v_writelane_b32 v63, s53, 13
+; GFX9-NEXT: v_writelane_b32 v63, s54, 14
+; GFX9-NEXT: v_writelane_b32 v63, s55, 15
+; GFX9-NEXT: v_writelane_b32 v63, s64, 16
+; GFX9-NEXT: v_writelane_b32 v63, s65, 17
+; GFX9-NEXT: v_writelane_b32 v63, s66, 18
+; GFX9-NEXT: v_writelane_b32 v63, s67, 19
+; GFX9-NEXT: v_writelane_b32 v63, s68, 20
+; GFX9-NEXT: v_writelane_b32 v63, s69, 21
+; GFX9-NEXT: v_writelane_b32 v63, s70, 22
+; GFX9-NEXT: v_writelane_b32 v63, s71, 23
+; GFX9-NEXT: v_writelane_b32 v63, s80, 24
+; GFX9-NEXT: v_writelane_b32 v63, s81, 25
+; GFX9-NEXT: v_writelane_b32 v63, s82, 26
+; GFX9-NEXT: v_writelane_b32 v63, s83, 27
+; GFX9-NEXT: v_writelane_b32 v63, s84, 28
+; GFX9-NEXT: v_writelane_b32 v63, s85, 29
+; GFX9-NEXT: v_writelane_b32 v63, s86, 30
+; GFX9-NEXT: v_writelane_b32 v63, s87, 31
+; GFX9-NEXT: v_writelane_b32 v63, s96, 32
+; GFX9-NEXT: v_writelane_b32 v63, s97, 33
+; GFX9-NEXT: v_writelane_b32 v63, s98, 34
+; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v19
+; GFX9-NEXT: v_writelane_b32 v63, s99, 35
+; GFX9-NEXT: v_readfirstlane_b32 s44, v3
+; GFX9-NEXT: v_readfirstlane_b32 s45, v4
+; GFX9-NEXT: v_readfirstlane_b32 s42, v5
+; GFX9-NEXT: v_readfirstlane_b32 s43, v6
+; GFX9-NEXT: v_readfirstlane_b32 s40, v7
+; GFX9-NEXT: v_readfirstlane_b32 s41, v8
+; GFX9-NEXT: v_readfirstlane_b32 s14, v9
+; GFX9-NEXT: v_readfirstlane_b32 s15, v10
+; GFX9-NEXT: v_readfirstlane_b32 s12, v11
+; GFX9-NEXT: v_readfirstlane_b32 s13, v12
+; GFX9-NEXT: v_readfirstlane_b32 s10, v13
+; GFX9-NEXT: v_readfirstlane_b32 s11, v14
+; GFX9-NEXT: v_readfirstlane_b32 s8, v15
+; GFX9-NEXT: v_readfirstlane_b32 s9, v16
+; GFX9-NEXT: v_readfirstlane_b32 s6, v17
+; GFX9-NEXT: v_readfirstlane_b32 s7, v18
+; GFX9-NEXT: v_readfirstlane_b32 s4, v1
+; GFX9-NEXT: s_and_b64 s[46:47], vcc, exec
+; GFX9-NEXT: v_readfirstlane_b32 s5, v2
+; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX9-NEXT: ; implicit-def: $vgpr62 : SGPR spill to VGPR lane
+; GFX9-NEXT: s_cbranch_scc0 .LBB99_3
+; GFX9-NEXT: ; %bb.1: ; %cmp.false
+; GFX9-NEXT: s_lshr_b32 s46, s5, 24
+; GFX9-NEXT: v_writelane_b32 v62, s46, 49
+; GFX9-NEXT: s_lshr_b32 s46, s5, 16
+; GFX9-NEXT: v_writelane_b32 v62, s46, 48
+; GFX9-NEXT: s_lshr_b32 s46, s5, 8
+; GFX9-NEXT: v_writelane_b32 v62, s46, 47
+; GFX9-NEXT: s_lshr_b32 s46, s4, 16
+; GFX9-NEXT: v_writelane_b32 v62, s46, 46
+; GFX9-NEXT: s_lshr_b32 s46, s4, 8
+; GFX9-NEXT: v_writelane_b32 v62, s46, 45
+; GFX9-NEXT: s_lshr_b32 s46, s29, 24
+; GFX9-NEXT: v_writelane_b32 v62, s46, 44
+; GFX9-NEXT: s_lshr_b32 s46, s29, 16
+; GFX9-NEXT: v_writelane_b32 v62, s46, 43
+; GFX9-NEXT: s_lshr_b32 s46, s29, 8
+; GFX9-NEXT: v_writelane_b32 v62, s46, 42
+; GFX9-NEXT: s_lshr_b32 s46, s28, 16
+; GFX9-NEXT: v_writelane_b32 v62, s46, 41
+; GFX9-NEXT: s_lshr_b32 s46, s28, 8
+; GFX9-NEXT: v_writelane_b32 v62, s46, 40
+; GFX9-NEXT: s_lshr_b32 s46, s27, 24
+; GFX9-NEXT: v_writelane_b32 v62, s46, 39
+; GFX9-NEXT: s_lshr_b32 s46, s27, 16
+; GFX9-NEXT: v_writelane_b32 v62, s46, 38
+; GFX9-NEXT: s_lshr_b32 s46, s27, 8
+; GFX9-NEXT: v_writelane_b32 v62, s46, 37
+; GFX9-NEXT: s_lshr_b32 s46, s26, 16
+; GFX9-NEXT: v_writelane_b32 v62, s46, 36
+; GFX9-NEXT: s_lshr_b32 s46, s26, 8
+; GFX9-NEXT: v_writelane_b32 v62, s46, 35
+; GFX9-NEXT: s_lshr_b32 s46, s25, 24
+; GFX9-NEXT: v_writelane_b32 v62, s46, 34
+; GFX9-NEXT: s_lshr_b32 s46, s25, 16
+; GFX9-NEXT: v_writelane_b32 v62, s46, 33
+; GFX9-NEXT: s_lshr_b32 s46, s25, 8
+; GFX9-NEXT: v_writelane_b32 v62, s46, 32
+; GFX9-NEXT: s_lshr_b32 s46, s24, 16
+; GFX9-NEXT: v_writelane_b32 v62, s46, 31
+; GFX9-NEXT: s_lshr_b32 s46, s24, 8
+; GFX9-NEXT: v_writelane_b32 v62, s46, 30
+; GFX9-NEXT: s_lshr_b32 s46, s23, 24
+; GFX9-NEXT: v_writelane_b32 v62, s46, 29
+; GFX9-NEXT: s_lshr_b32 s46, s23, 16
+; GFX9-NEXT: v_writelane_b32 v62, s46, 28
+; GFX9-NEXT: s_lshr_b32 s46, s23, 8
+; GFX9-NEXT: v_writelane_b32 v62, s46, 27
+; GFX9-NEXT: s_lshr_b32 s46, s22, 16
+; GFX9-NEXT: v_writelane_b32 v62, s46, 26
+; GFX9-NEXT: s_lshr_b32 s46, s22, 8
+; GFX9-NEXT: v_writelane_b32 v62, s46, 25
+; GFX9-NEXT: s_lshr_b32 s46, s21, 24
+; GFX9-NEXT: v_writelane_b32 v62, s46, 24
+; GFX9-NEXT: s_lshr_b32 s46, s21, 16
+; GFX9-NEXT: v_writelane_b32 v62, s46, 23
+; GFX9-NEXT: s_lshr_b32 s46, s21, 8
+; GFX9-NEXT: v_writelane_b32 v62, s46, 22
+; GFX9-NEXT: s_lshr_b32 s46, s20, 16
+; GFX9-NEXT: v_writelane_b32 v62, s46, 21
+; GFX9-NEXT: s_lshr_b32 s46, s20, 8
+; GFX9-NEXT: v_writelane_b32 v62, s46, 20
+; GFX9-NEXT: s_lshr_b32 s46, s19, 24
+; GFX9-NEXT: v_writelane_b32 v62, s46, 19
+; GFX9-NEXT: s_lshr_b32 s46, s19, 16
+; GFX9-NEXT: v_writelane_b32 v62, s46, 18
+; GFX9-NEXT: s_lshr_b32 s46, s19, 8
+; GFX9-NEXT: v_writelane_b32 v62, s46, 17
+; GFX9-NEXT: s_lshr_b32 s46, s18, 16
+; GFX9-NEXT: v_writelane_b32 v62, s46, 16
+; GFX9-NEXT: s_lshr_b32 s46, s18, 8
+; GFX9-NEXT: v_writelane_b32 v62, s46, 15
+; GFX9-NEXT: s_lshr_b32 s46, s17, 24
+; GFX9-NEXT: v_writelane_b32 v62, s46, 14
+; GFX9-NEXT: s_lshr_b32 s46, s17, 16
+; GFX9-NEXT: v_writelane_b32 v62, s46, 13
+; GFX9-NEXT: s_lshr_b32 s46, s17, 8
+; GFX9-NEXT: v_writelane_b32 v62, s46, 12
+; GFX9-NEXT: s_lshr_b32 s46, s16, 16
+; GFX9-NEXT: v_writelane_b32 v62, s46, 11
+; GFX9-NEXT: s_lshr_b32 s46, s16, 8
+; GFX9-NEXT: v_writelane_b32 v62, s46, 10
+; GFX9-NEXT: s_lshr_b32 s46, s7, 24
+; GFX9-NEXT: v_writelane_b32 v62, s46, 9
+; GFX9-NEXT: s_lshr_b32 s46, s7, 16
+; GFX9-NEXT: v_writelane_b32 v62, s46, 8
+; GFX9-NEXT: s_lshr_b32 s46, s7, 8
+; GFX9-NEXT: v_writelane_b32 v62, s46, 7
+; GFX9-NEXT: s_lshr_b32 s46, s6, 16
+; GFX9-NEXT: v_writelane_b32 v62, s46, 6
+; GFX9-NEXT: s_lshr_b32 s46, s6, 8
+; GFX9-NEXT: v_writelane_b32 v62, s46, 5
+; GFX9-NEXT: s_lshr_b32 s46, s9, 24
+; GFX9-NEXT: v_writelane_b32 v62, s46, 4
+; GFX9-NEXT: s_lshr_b32 s46, s9, 16
+; GFX9-NEXT: v_writelane_b32 v62, s46, 3
+; GFX9-NEXT: s_lshr_b32 s46, s9, 8
+; GFX9-NEXT: v_writelane_b32 v62, s46, 2
+; GFX9-NEXT: s_lshr_b32 s46, s8, 16
+; GFX9-NEXT: v_writelane_b32 v62, s46, 1
+; GFX9-NEXT: s_lshr_b32 s46, s8, 8
+; GFX9-NEXT: v_writelane_b32 v62, s46, 0
+; GFX9-NEXT: s_lshr_b32 s82, s11, 24
+; GFX9-NEXT: s_lshr_b32 s83, s11, 16
+; GFX9-NEXT: s_lshr_b32 s85, s11, 8
+; GFX9-NEXT: s_lshr_b32 s84, s10, 16
+; GFX9-NEXT: s_lshr_b32 s86, s10, 8
+; GFX9-NEXT: s_lshr_b32 s87, s13, 24
+; GFX9-NEXT: s_lshr_b32 s96, s13, 16
+; GFX9-NEXT: s_lshr_b32 s98, s13, 8
+; GFX9-NEXT: s_lshr_b32 s97, s12, 16
+; GFX9-NEXT: s_lshr_b32 s99, s12, 8
+; GFX9-NEXT: s_lshr_b32 s38, s15, 24
+; GFX9-NEXT: s_lshr_b32 s39, s15, 16
+; GFX9-NEXT: s_lshr_b32 s49, s15, 8
+; GFX9-NEXT: s_lshr_b32 s48, s14, 16
+; GFX9-NEXT: s_lshr_b32 s50, s14, 8
+; GFX9-NEXT: s_lshr_b32 s51, s41, 24
+; GFX9-NEXT: s_lshr_b32 s52, s41, 16
+; GFX9-NEXT: s_lshr_b32 s54, s41, 8
+; GFX9-NEXT: s_lshr_b32 s53, s40, 16
+; GFX9-NEXT: s_lshr_b32 s55, s40, 8
+; GFX9-NEXT: s_lshr_b32 s64, s43, 24
+; GFX9-NEXT: s_lshr_b32 s65, s43, 16
+; GFX9-NEXT: s_lshr_b32 s67, s43, 8
+; GFX9-NEXT: s_lshr_b32 s66, s42, 16
+; GFX9-NEXT: s_lshr_b32 s68, s42, 8
+; GFX9-NEXT: s_lshr_b32 s69, s45, 24
+; GFX9-NEXT: s_lshr_b32 s70, s45, 16
+; GFX9-NEXT: s_lshr_b32 s80, s45, 8
+; GFX9-NEXT: s_lshr_b32 s71, s44, 16
+; GFX9-NEXT: s_lshr_b32 s81, s44, 8
+; GFX9-NEXT: s_lshr_b64 s[46:47], s[4:5], 24
+; GFX9-NEXT: s_lshr_b64 s[56:57], s[28:29], 24
+; GFX9-NEXT: s_lshr_b64 s[58:59], s[26:27], 24
+; GFX9-NEXT: s_lshr_b64 s[60:61], s[24:25], 24
+; GFX9-NEXT: s_lshr_b64 s[62:63], s[22:23], 24
+; GFX9-NEXT: s_lshr_b64 s[72:73], s[20:21], 24
+; GFX9-NEXT: s_lshr_b64 s[74:75], s[18:19], 24
+; GFX9-NEXT: s_lshr_b64 s[76:77], s[16:17], 24
+; GFX9-NEXT: s_lshr_b64 s[78:79], s[6:7], 24
+; GFX9-NEXT: s_lshr_b64 s[88:89], s[8:9], 24
+; GFX9-NEXT: s_lshr_b64 s[90:91], s[10:11], 24
+; GFX9-NEXT: s_lshr_b64 s[92:93], s[12:13], 24
+; GFX9-NEXT: s_lshr_b64 s[94:95], s[14:15], 24
+; GFX9-NEXT: s_lshr_b64 s[30:31], s[40:41], 24
+; GFX9-NEXT: s_lshr_b64 s[34:35], s[42:43], 24
+; GFX9-NEXT: s_lshr_b64 s[36:37], s[44:45], 24
+; GFX9-NEXT: s_cbranch_execnz .LBB99_4
+; GFX9-NEXT: .LBB99_2: ; %cmp.true
+; GFX9-NEXT: v_pk_add_u16 v26, s5, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: v_pk_add_u16 v25, s4, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[25:26]
+; GFX9-NEXT: v_pk_add_u16 v28, s29, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: v_pk_add_u16 v27, s28, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[27:28]
+; GFX9-NEXT: v_pk_add_u16 v30, s27, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: v_pk_add_u16 v29, s26, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[29:30]
+; GFX9-NEXT: v_pk_add_u16 v32, s25, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: v_pk_add_u16 v31, s24, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[31:32]
+; GFX9-NEXT: v_pk_add_u16 v34, s23, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: v_pk_add_u16 v33, s22, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[33:34]
+; GFX9-NEXT: v_pk_add_u16 v36, s21, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: v_pk_add_u16 v35, s20, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[35:36]
+; GFX9-NEXT: v_pk_add_u16 v38, s19, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: v_pk_add_u16 v37, s18, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[37:38]
+; GFX9-NEXT: v_pk_add_u16 v49, s17, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: v_pk_add_u16 v48, s16, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[48:49]
+; GFX9-NEXT: v_pk_add_u16 v2, s7, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: v_pk_add_u16 v1, s6, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2]
+; GFX9-NEXT: v_pk_add_u16 v4, s9, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: v_pk_add_u16 v3, s8, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[3:4]
+; GFX9-NEXT: v_pk_add_u16 v6, s11, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: v_pk_add_u16 v5, s10, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[5:6]
+; GFX9-NEXT: v_pk_add_u16 v8, s13, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: v_pk_add_u16 v7, s12, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[7:8]
+; GFX9-NEXT: v_pk_add_u16 v10, s15, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: v_pk_add_u16 v9, s14, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[9:10]
+; GFX9-NEXT: v_pk_add_u16 v12, s41, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: v_pk_add_u16 v11, s40, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[11:12]
+; GFX9-NEXT: v_pk_add_u16 v14, s43, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: v_pk_add_u16 v13, s42, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[13:14]
+; GFX9-NEXT: v_pk_add_u16 v22, s45, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: v_pk_add_u16 v21, s44, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b64 v[15:16], 24, v[21:22]
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 24, v4
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v3
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v3
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 24, v6
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v6
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v6
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v5
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v5
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 24, v8
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v8
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v8
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v7
+; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v26
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v7
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v26
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 24, v10
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v25
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v10
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v28
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v10
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v28
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v9
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v27
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v9
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v30
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 24, v12
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v30
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v12
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v29
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v12
+; 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, v32
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v11
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v32
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v11
+; 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, v31
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 24, v14
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v34
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v14
+; 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, v34
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v14
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v33
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v13
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v15, 24, v36
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v13
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v36
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 24, v22
+; 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, v35
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v22
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v38
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v22
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v15, 8, v38
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v21
+; GFX9-NEXT: v_lshrrev_b32_e32 v44, 8, v26
+; GFX9-NEXT: v_lshrrev_b32_e32 v45, 8, v25
+; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v28
+; GFX9-NEXT: v_lshrrev_b32_e32 v47, 8, v27
+; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v30
+; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v29
+; GFX9-NEXT: v_lshrrev_b32_e32 v57, 8, v32
+; GFX9-NEXT: v_lshrrev_b32_e32 v58, 8, v31
+; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v34
+; GFX9-NEXT: v_lshrrev_b32_e32 v23, 8, v33
+; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v36
+; GFX9-NEXT: v_lshrrev_b32_e32 v61, 8, v35
+; GFX9-NEXT: v_lshrrev_b32_e32 v54, 24, v38
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v37
+; GFX9-NEXT: v_lshrrev_b32_e32 v15, 8, v37
+; GFX9-NEXT: v_lshrrev_b32_e32 v43, 24, v49
+; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v49
+; GFX9-NEXT: v_lshrrev_b32_e32 v18, 8, v49
+; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v48
+; GFX9-NEXT: v_lshrrev_b32_e32 v20, 8, v48
+; GFX9-NEXT: v_lshrrev_b32_e32 v24, 24, v2
+; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v2
+; GFX9-NEXT: v_lshrrev_b32_e32 v50, 8, v2
+; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v1
+; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v1
+; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v4
+; GFX9-NEXT: v_lshrrev_b32_e32 v55, 8, v4
+; GFX9-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; GFX9-NEXT: v_lshrrev_b32_e32 v19, 8, v21
+; GFX9-NEXT: s_branch .LBB99_5
+; GFX9-NEXT: .LBB99_3:
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr81
+; GFX9-NEXT: ; implicit-def: $sgpr71
+; GFX9-NEXT: ; implicit-def: $sgpr80
+; GFX9-NEXT: ; implicit-def: $sgpr70
+; GFX9-NEXT: ; implicit-def: $sgpr69
+; GFX9-NEXT: ; implicit-def: $sgpr68
+; GFX9-NEXT: ; implicit-def: $sgpr66
+; GFX9-NEXT: ; implicit-def: $sgpr67
+; GFX9-NEXT: ; implicit-def: $sgpr65
+; GFX9-NEXT: ; implicit-def: $sgpr64
+; GFX9-NEXT: ; implicit-def: $sgpr55
+; GFX9-NEXT: ; implicit-def: $sgpr53
+; GFX9-NEXT: ; implicit-def: $sgpr54
+; GFX9-NEXT: ; implicit-def: $sgpr52
+; GFX9-NEXT: ; implicit-def: $sgpr51
+; GFX9-NEXT: ; implicit-def: $sgpr50
+; GFX9-NEXT: ; implicit-def: $sgpr48
+; GFX9-NEXT: ; implicit-def: $sgpr49
+; GFX9-NEXT: ; implicit-def: $sgpr39
+; GFX9-NEXT: ; implicit-def: $sgpr38
+; GFX9-NEXT: ; implicit-def: $sgpr99
+; GFX9-NEXT: ; implicit-def: $sgpr97
+; GFX9-NEXT: ; implicit-def: $sgpr98
+; GFX9-NEXT: ; implicit-def: $sgpr96
+; GFX9-NEXT: ; implicit-def: $sgpr87
+; GFX9-NEXT: ; implicit-def: $sgpr86
+; GFX9-NEXT: ; implicit-def: $sgpr84
+; GFX9-NEXT: ; implicit-def: $sgpr85
+; GFX9-NEXT: ; implicit-def: $sgpr83
+; GFX9-NEXT: ; implicit-def: $sgpr82
+; GFX9-NEXT: ; implicit-def: $sgpr76
+; GFX9-NEXT: ; implicit-def: $sgpr74
+; GFX9-NEXT: ; implicit-def: $sgpr72
+; GFX9-NEXT: ; implicit-def: $sgpr62
+; GFX9-NEXT: ; implicit-def: $sgpr60
+; GFX9-NEXT: ; implicit-def: $sgpr58
+; GFX9-NEXT: ; implicit-def: $sgpr56
+; GFX9-NEXT: ; implicit-def: $sgpr36
+; GFX9-NEXT: ; implicit-def: $sgpr34
+; GFX9-NEXT: ; implicit-def: $sgpr30
+; GFX9-NEXT: ; implicit-def: $sgpr94
+; GFX9-NEXT: ; implicit-def: $sgpr92
+; GFX9-NEXT: ; implicit-def: $sgpr90
+; GFX9-NEXT: ; implicit-def: $sgpr88
+; GFX9-NEXT: ; implicit-def: $sgpr78
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: ; kill: killed $sgpr46
+; GFX9-NEXT: ; implicit-def: $sgpr46
+; GFX9-NEXT: s_branch .LBB99_2
+; GFX9-NEXT: .LBB99_4:
+; GFX9-NEXT: v_mov_b32_e32 v15, s71
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s80
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s70
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s69
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s68
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s66
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s67
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s65
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s64
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s55
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s53
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s54
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s52
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s51
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s50
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s48
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s49
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s39
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s38
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s99
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s97
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s98
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s96
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s87
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s86
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s84
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s85
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s83
+; GFX9-NEXT: v_mov_b32_e32 v25, s4
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s82
+; GFX9-NEXT: v_readlane_b32 s4, v62, 0
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 1
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 2
+; GFX9-NEXT: v_mov_b32_e32 v19, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 3
+; GFX9-NEXT: v_mov_b32_e32 v55, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 4
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 5
+; GFX9-NEXT: v_mov_b32_e32 v53, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 6
+; GFX9-NEXT: v_mov_b32_e32 v52, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 7
+; GFX9-NEXT: v_mov_b32_e32 v51, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 8
+; GFX9-NEXT: v_mov_b32_e32 v50, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 9
+; GFX9-NEXT: v_mov_b32_e32 v24, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 10
+; GFX9-NEXT: v_mov_b32_e32 v20, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 11
+; GFX9-NEXT: v_mov_b32_e32 v42, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 12
+; GFX9-NEXT: v_mov_b32_e32 v18, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 13
+; GFX9-NEXT: v_mov_b32_e32 v39, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 14
+; GFX9-NEXT: v_mov_b32_e32 v43, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 15
+; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v15, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 16
+; GFX9-NEXT: v_mov_b32_e32 v40, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 17
+; GFX9-NEXT: v_mov_b32_e32 v17, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 18
+; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v17, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 19
+; GFX9-NEXT: v_mov_b32_e32 v54, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 20
+; GFX9-NEXT: v_mov_b32_e32 v61, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 21
+; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v17, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 22
+; GFX9-NEXT: v_mov_b32_e32 v60, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 23
+; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v17, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 24
+; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v17, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 25
+; GFX9-NEXT: v_mov_b32_e32 v23, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 26
+; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v17, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 27
+; GFX9-NEXT: v_mov_b32_e32 v59, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 28
+; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v17, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 29
+; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v17, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 30
+; GFX9-NEXT: v_mov_b32_e32 v58, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 31
+; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v17, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 32
+; GFX9-NEXT: v_mov_b32_e32 v57, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 33
+; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v17, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 34
+; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v17, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 35
+; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v17, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 36
+; GFX9-NEXT: v_mov_b32_e32 v41, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 37
+; GFX9-NEXT: v_mov_b32_e32 v56, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 38
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v41, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 39
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v41, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 40
+; GFX9-NEXT: v_mov_b32_e32 v47, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 41
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v41, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 42
+; GFX9-NEXT: v_mov_b32_e32 v46, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 43
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v41, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 44
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v41, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 45
+; GFX9-NEXT: v_mov_b32_e32 v45, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 46
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v41, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 47
+; GFX9-NEXT: v_mov_b32_e32 v44, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 48
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v41, s4
+; GFX9-NEXT: v_readlane_b32 s4, v62, 49
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v41, s4
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v41, s46
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v41, s56
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v41, s58
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v41, s60
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v41, s62
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v41, s72
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v41, s74
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v41, s76
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v41, s78
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v41, s88
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v41, s90
+; 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_mov_b32_e32 v41, s92
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v41, s94
+; 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_mov_b32_e32 v41, s30
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v41, s34
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v41, s36
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
+; GFX9-NEXT: v_mov_b32_e32 v21, s44
+; GFX9-NEXT: v_mov_b32_e32 v22, s45
+; GFX9-NEXT: v_mov_b32_e32 v13, s42
+; GFX9-NEXT: v_mov_b32_e32 v14, s43
+; GFX9-NEXT: v_mov_b32_e32 v11, s40
+; GFX9-NEXT: v_mov_b32_e32 v12, s41
+; GFX9-NEXT: v_mov_b32_e32 v9, s14
+; GFX9-NEXT: v_mov_b32_e32 v10, s15
+; GFX9-NEXT: v_mov_b32_e32 v7, s12
+; GFX9-NEXT: v_mov_b32_e32 v8, s13
+; GFX9-NEXT: v_mov_b32_e32 v5, s10
+; GFX9-NEXT: v_mov_b32_e32 v6, s11
+; GFX9-NEXT: v_mov_b32_e32 v3, s8
+; GFX9-NEXT: v_mov_b32_e32 v4, s9
+; GFX9-NEXT: v_mov_b32_e32 v1, s6
+; GFX9-NEXT: v_mov_b32_e32 v2, s7
+; GFX9-NEXT: v_mov_b32_e32 v48, s16
+; GFX9-NEXT: v_mov_b32_e32 v49, s17
+; GFX9-NEXT: v_mov_b32_e32 v37, s18
+; GFX9-NEXT: v_mov_b32_e32 v38, s19
+; GFX9-NEXT: v_mov_b32_e32 v35, s20
+; GFX9-NEXT: v_mov_b32_e32 v36, s21
+; GFX9-NEXT: v_mov_b32_e32 v33, s22
+; GFX9-NEXT: v_mov_b32_e32 v34, s23
+; GFX9-NEXT: v_mov_b32_e32 v31, s24
+; GFX9-NEXT: v_mov_b32_e32 v32, s25
+; GFX9-NEXT: v_mov_b32_e32 v29, s26
+; GFX9-NEXT: v_mov_b32_e32 v30, s27
+; GFX9-NEXT: v_mov_b32_e32 v27, s28
+; GFX9-NEXT: v_mov_b32_e32 v28, s29
+; GFX9-NEXT: v_mov_b32_e32 v26, s5
+; GFX9-NEXT: v_mov_b32_e32 v41, v50
+; GFX9-NEXT: v_mov_b32_e32 v50, v51
+; GFX9-NEXT: v_mov_b32_e32 v51, v52
+; GFX9-NEXT: v_mov_b32_e32 v52, v53
+; GFX9-NEXT: v_mov_b32_e32 v53, v55
+; GFX9-NEXT: v_mov_b32_e32 v55, v19
+; GFX9-NEXT: v_mov_b32_e32 v19, s81
+; GFX9-NEXT: .LBB99_5: ; %end
+; GFX9-NEXT: v_lshlrev_b32_e32 v16, 8, v15
+; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
+; GFX9-NEXT: v_lshlrev_b32_e32 v19, 8, v19
+; GFX9-NEXT: v_or_b32_sdwa v16, v37, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshlrev_b32_e32 v37, 8, v61
+; GFX9-NEXT: v_or_b32_sdwa v21, v21, 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:368 ; 4-byte Folded Reload
+; GFX9-NEXT: v_or_b32_sdwa v35, v35, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshlrev_b32_e32 v37, 8, v60
+; GFX9-NEXT: v_or_b32_sdwa v36, v36, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshlrev_b32_e32 v20, 8, v20
+; GFX9-NEXT: v_or_b32_sdwa v20, v48, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshlrev_b32_e32 v18, 8, v18
+; GFX9-NEXT: v_or_b32_sdwa v18, v49, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshlrev_b32_e32 v23, 8, v23
+; GFX9-NEXT: v_or_b32_sdwa v23, v33, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshlrev_b32_e32 v33, 8, v59
+; GFX9-NEXT: v_or_b32_sdwa v33, v34, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshlrev_b32_e32 v34, 8, v58
+; GFX9-NEXT: v_or_b32_sdwa v31, v31, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshlrev_b32_e32 v34, 8, v57
+; GFX9-NEXT: v_or_b32_sdwa v32, v32, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshlrev_b32_e32 v17, 8, v17
+; GFX9-NEXT: v_or_b32_sdwa v17, v29, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshlrev_b32_e32 v29, 8, v56
+; GFX9-NEXT: v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshlrev_b32_e32 v30, 8, v47
+; GFX9-NEXT: v_or_b32_sdwa v27, v27, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshlrev_b32_e32 v30, 8, v46
+; GFX9-NEXT: v_or_b32_sdwa v28, v28, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshlrev_b32_e32 v30, 8, v45
+; GFX9-NEXT: v_or_b32_sdwa v25, v25, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_lshlrev_b32_e32 v30, 8, v44
+; GFX9-NEXT: v_or_b32_sdwa v26, v26, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_readlane_b32 s99, v63, 35
+; GFX9-NEXT: v_readlane_b32 s98, v63, 34
+; GFX9-NEXT: v_readlane_b32 s97, v63, 33
+; GFX9-NEXT: v_readlane_b32 s96, v63, 32
+; GFX9-NEXT: v_readlane_b32 s87, v63, 31
+; GFX9-NEXT: v_readlane_b32 s86, v63, 30
+; GFX9-NEXT: v_readlane_b32 s85, v63, 29
+; GFX9-NEXT: v_readlane_b32 s84, v63, 28
+; GFX9-NEXT: v_readlane_b32 s83, v63, 27
+; GFX9-NEXT: v_readlane_b32 s82, v63, 26
+; GFX9-NEXT: v_readlane_b32 s81, v63, 25
+; GFX9-NEXT: v_readlane_b32 s80, v63, 24
+; GFX9-NEXT: v_readlane_b32 s71, v63, 23
+; GFX9-NEXT: v_readlane_b32 s70, v63, 22
+; GFX9-NEXT: v_readlane_b32 s69, v63, 21
+; GFX9-NEXT: v_readlane_b32 s68, v63, 20
+; GFX9-NEXT: v_readlane_b32 s67, v63, 19
+; GFX9-NEXT: v_readlane_b32 s66, v63, 18
+; GFX9-NEXT: v_readlane_b32 s65, v63, 17
+; GFX9-NEXT: v_readlane_b32 s64, v63, 16
+; GFX9-NEXT: v_readlane_b32 s55, v63, 15
+; GFX9-NEXT: v_readlane_b32 s54, v63, 14
+; GFX9-NEXT: v_readlane_b32 s53, v63, 13
+; GFX9-NEXT: v_readlane_b32 s52, v63, 12
+; GFX9-NEXT: v_readlane_b32 s51, v63, 11
+; GFX9-NEXT: v_readlane_b32 s50, v63, 10
+; GFX9-NEXT: v_readlane_b32 s49, v63, 9
+; GFX9-NEXT: v_readlane_b32 s48, v63, 8
+; GFX9-NEXT: v_readlane_b32 s39, v63, 7
+; GFX9-NEXT: v_readlane_b32 s38, v63, 6
+; GFX9-NEXT: v_readlane_b32 s37, v63, 5
+; GFX9-NEXT: v_readlane_b32 s36, v63, 4
+; GFX9-NEXT: v_readlane_b32 s35, v63, 3
+; GFX9-NEXT: v_readlane_b32 s34, v63, 2
+; GFX9-NEXT: v_readlane_b32 s31, v63, 1
+; GFX9-NEXT: v_readlane_b32 s30, v63, 0
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
+; GFX9-NEXT: v_or_b32_sdwa v15, v38, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_lshlrev_b32_e32 v19, 8, v19
+; GFX9-NEXT: v_or_b32_sdwa v22, v22, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_lshlrev_b32_e32 v19, 8, v37
+; GFX9-NEXT: v_or_b32_sdwa v19, v42, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_store_dword v19, v0, s[0:3], 0 offen
+; GFX9-NEXT: v_lshlrev_b32_e32 v19, 8, v43
+; GFX9-NEXT: v_or_b32_sdwa v19, v39, v19 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_store_dword v18, v0, s[0:3], 0 offen offset:4
+; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_lshlrev_b32_e32 v18, 8, v18
+; GFX9-NEXT: v_or_b32_sdwa v18, v40, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v16, v16, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_store_dword v16, v0, s[0:3], 0 offen offset:8
+; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
+; GFX9-NEXT: v_lshlrev_b32_e32 v16, 8, v54
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v16, v18, v16 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_store_dword v15, v0, s[0:3], 0 offen offset:12
+; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v15, v35, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:16
+; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v15, v36, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:20
+; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v15, v23, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:24
+; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v15, v33, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:28
+; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v18
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v15, v31, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:32
+; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v15, v32, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:36
+; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v15, v17, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:40
+; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v15, v29, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:44
+; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v15, v27, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:48
+; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v15, v28, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:52
+; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v15, v25, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:56
+; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v15, v26, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:60
+; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v15, v21, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:64
+; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v15, v22, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:68
+; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_lshlrev_b32_e32 v15, 8, v15
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v15, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_lshlrev_b32_e32 v16, 8, v16
+; GFX9-NEXT: v_or_b32_sdwa v13, v13, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v13, v13, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_store_dword v13, v0, s[0:3], 0 offen offset:72
+; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_lshlrev_b32_e32 v13, 8, v13
+; GFX9-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_lshlrev_b32_e32 v14, 8, v14
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v14, v15, v14 dst_sel:WORD_1 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: buffer_store_dword v13, v0, s[0:3], 0 offen offset:76
+; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_lshlrev_b32_e32 v13, 8, v13
+; GFX9-NEXT: v_or_b32_sdwa v11, v11, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_lshlrev_b32_e32 v13, 8, v13
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v11, v11, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_store_dword v11, v0, s[0:3], 0 offen offset:80
+; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_lshlrev_b32_e32 v11, 8, v11
+; GFX9-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_lshlrev_b32_e32 v12, 8, v12
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v12, v13, v12 dst_sel:WORD_1 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_store_dword v11, v0, s[0:3], 0 offen offset:84
+; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_lshlrev_b32_e32 v11, 8, v11
+; GFX9-NEXT: v_or_b32_sdwa v9, v9, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_lshlrev_b32_e32 v11, 8, v11
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v11, v12, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v9, v9, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_store_dword v9, v0, s[0:3], 0 offen offset:88
+; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_lshlrev_b32_e32 v9, 8, v9
+; GFX9-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_lshlrev_b32_e32 v10, 8, v10
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v10, v11, v10 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: buffer_store_dword v9, v0, s[0:3], 0 offen offset:92
+; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_lshlrev_b32_e32 v9, 8, v9
+; GFX9-NEXT: v_or_b32_sdwa v7, v7, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_lshlrev_b32_e32 v9, 8, v9
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v9, v10, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v7, v7, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_store_dword v7, v0, s[0:3], 0 offen offset:96
+; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_lshlrev_b32_e32 v7, 8, v7
+; GFX9-NEXT: v_or_b32_sdwa v7, v8, v7 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:140 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_lshlrev_b32_e32 v8, 8, v8
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v8, v9, v8 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: buffer_store_dword v7, v0, s[0:3], 0 offen offset:100
+; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_lshlrev_b32_e32 v7, 8, v7
+; GFX9-NEXT: v_or_b32_sdwa v5, v5, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_lshlrev_b32_e32 v7, 8, v7
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v7, v8, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v5, v5, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_store_dword v5, v0, s[0:3], 0 offen offset:104
+; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_lshlrev_b32_e32 v5, 8, v5
+; GFX9-NEXT: v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_lshlrev_b32_e32 v6, 8, v6
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v6, v7, v6 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: buffer_store_dword v5, v0, s[0:3], 0 offen offset:108
+; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_lshlrev_b32_e32 v5, 8, v5
+; GFX9-NEXT: v_or_b32_sdwa v3, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(2)
+; GFX9-NEXT: v_lshlrev_b32_e32 v5, 8, v5
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_or_b32_sdwa v5, v6, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v3, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_store_dword v3, v0, s[0:3], 0 offen offset:112
+; GFX9-NEXT: v_lshlrev_b32_e32 v3, 8, v55
+; GFX9-NEXT: v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_lshlrev_b32_e32 v4, 8, v4
+; GFX9-NEXT: v_or_b32_sdwa v4, v53, v4 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: buffer_store_dword v3, v0, s[0:3], 0 offen offset:116
+; GFX9-NEXT: v_lshlrev_b32_e32 v3, 8, v52
+; GFX9-NEXT: v_or_b32_sdwa v1, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_lshlrev_b32_e32 v3, 8, v3
+; GFX9-NEXT: v_or_b32_sdwa v3, v51, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT: v_or_b32_sdwa v1, v1, v3 dst_sel:DWORD 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: v_lshlrev_b32_e32 v1, 8, v50
+; 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_b32_e32 v2, 8, v24
+; GFX9-NEXT: v_or_b32_sdwa v2, v41, 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 v61, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
+; GFX9-NEXT: s_or_saveexec_b64 s[4:5], -1
+; 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: s_mov_b64 exec, s[4:5]
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-TRUE16-LABEL: bitcast_v64i16_to_v128i8_scalar:
+; GFX11-TRUE16: ; %bb.0:
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT: s_or_saveexec_b32 s4, -1
+; GFX11-TRUE16-NEXT: s_clause 0x3
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v75, s32 offset:76
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v76, s32 offset:80
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v77, s32 offset:84
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v78, s32 offset:88
+; GFX11-TRUE16-NEXT: s_mov_b32 exec_lo, s4
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s30, 0
+; GFX11-TRUE16-NEXT: v_writelane_b32 v76, s96, 0
+; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v15
+; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s40, v1
+; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s41, v2
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s31, 1
+; GFX11-TRUE16-NEXT: v_writelane_b32 v76, s97, 1
+; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s14, v3
+; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s15, v4
+; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s12, v5
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s34, 2
+; GFX11-TRUE16-NEXT: v_writelane_b32 v76, s98, 2
+; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s13, v6
+; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s10, v7
+; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s11, v8
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s35, 3
+; GFX11-TRUE16-NEXT: v_writelane_b32 v76, s99, 3
+; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s8, v9
+; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s9, v10
+; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s6, v11
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s36, 4
+; GFX11-TRUE16-NEXT: v_writelane_b32 v76, s100, 4
+; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s7, v12
+; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s4, v13
+; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s5, v14
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s37, 5
+; GFX11-TRUE16-NEXT: v_writelane_b32 v76, s101, 5
+; GFX11-TRUE16-NEXT: s_mov_b32 vcc_hi, 0
+; GFX11-TRUE16-NEXT: s_and_b32 s42, vcc_lo, exec_lo
+; GFX11-TRUE16-NEXT: s_clause 0x12
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v40, s32 offset:72
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v41, s32 offset:68
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v42, s32 offset:64
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v43, s32 offset:60
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v44, s32 offset:56
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v45, s32 offset:52
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v46, s32 offset:48
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v47, s32 offset:44
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v56, s32 offset:40
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v57, s32 offset:36
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v58, s32 offset:32
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v59, s32 offset:28
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v60, s32 offset:24
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v61, s32 offset:20
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v62, s32 offset:16
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v63, s32 offset:12
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v72, s32 offset:8
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v73, s32 offset:4
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v74, s32
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s38, 6
+; GFX11-TRUE16-NEXT: v_writelane_b32 v76, s102, 6
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr78 : SGPR spill to VGPR lane
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr77 : SGPR spill to VGPR lane
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s39, 7
+; GFX11-TRUE16-NEXT: v_writelane_b32 v76, s103, 7
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s48, 8
+; GFX11-TRUE16-NEXT: v_writelane_b32 v76, s104, 8
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s49, 9
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s50, 10
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s51, 11
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s52, 12
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s53, 13
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s54, 14
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s55, 15
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s64, 16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s65, 17
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s66, 18
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s67, 19
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s68, 20
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s69, 21
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s70, 22
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s71, 23
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s80, 24
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s81, 25
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s82, 26
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s83, 27
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s84, 28
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s85, 29
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s86, 30
+; GFX11-TRUE16-NEXT: v_writelane_b32 v75, s87, 31
+; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB99_3
+; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT: s_lshr_b64 s[74:75], s[0:1], 24
+; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 8
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s27, 8
+; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s27, 24
+; GFX11-TRUE16-NEXT: s_lshr_b32 s35, s5, 24
+; GFX11-TRUE16-NEXT: s_lshr_b32 s36, s5, 16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 7
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s26, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s37, s5, 8
+; GFX11-TRUE16-NEXT: s_lshr_b32 s38, s4, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s39, s4, 8
+; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 6
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s26, 8
+; GFX11-TRUE16-NEXT: s_lshr_b32 s48, s7, 24
+; GFX11-TRUE16-NEXT: s_lshr_b32 s49, s7, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s50, s7, 8
+; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 5
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s25, 24
+; GFX11-TRUE16-NEXT: s_lshr_b32 s51, s6, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s52, s6, 8
+; GFX11-TRUE16-NEXT: s_lshr_b32 s53, s9, 24
+; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 4
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s25, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s54, s9, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s55, s9, 8
+; GFX11-TRUE16-NEXT: s_lshr_b32 s64, s8, 16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 3
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s25, 8
+; GFX11-TRUE16-NEXT: s_lshr_b32 s65, s8, 8
+; GFX11-TRUE16-NEXT: s_lshr_b32 s66, s11, 24
+; GFX11-TRUE16-NEXT: s_lshr_b32 s67, s11, 16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 2
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s24, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s68, s11, 8
+; GFX11-TRUE16-NEXT: s_lshr_b32 s69, s10, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s70, s10, 8
+; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 1
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s24, 8
+; GFX11-TRUE16-NEXT: s_lshr_b32 s71, s13, 24
+; GFX11-TRUE16-NEXT: s_lshr_b32 s80, s13, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s81, s13, 8
+; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 0
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s23, 24
+; GFX11-TRUE16-NEXT: s_lshr_b32 s82, s12, 16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 31
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s23, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s83, s12, 8
+; GFX11-TRUE16-NEXT: s_lshr_b32 s84, s15, 24
+; GFX11-TRUE16-NEXT: s_lshr_b32 s85, s15, 16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 30
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s23, 8
+; GFX11-TRUE16-NEXT: s_lshr_b32 s86, s15, 8
+; GFX11-TRUE16-NEXT: s_lshr_b32 s87, s14, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s96, s14, 8
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 29
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s22, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s97, s41, 24
+; GFX11-TRUE16-NEXT: s_lshr_b32 s98, s41, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s99, s41, 8
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 28
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s22, 8
+; GFX11-TRUE16-NEXT: s_lshr_b32 s100, s40, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s101, s40, 8
+; GFX11-TRUE16-NEXT: s_lshr_b32 s102, s29, 24
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 27
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s21, 24
+; GFX11-TRUE16-NEXT: s_lshr_b32 s103, s29, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s104, s29, 8
+; GFX11-TRUE16-NEXT: s_lshr_b32 s34, s28, 16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 26
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s21, 16
+; GFX11-TRUE16-NEXT: s_lshr_b64 s[62:63], s[26:27], 24
+; GFX11-TRUE16-NEXT: s_lshr_b64 s[72:73], s[24:25], 24
+; GFX11-TRUE16-NEXT: s_lshr_b64 s[60:61], s[22:23], 24
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 25
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s21, 8
+; GFX11-TRUE16-NEXT: s_lshr_b64 s[58:59], s[20:21], 24
+; GFX11-TRUE16-NEXT: s_lshr_b64 s[56:57], s[18:19], 24
+; GFX11-TRUE16-NEXT: s_lshr_b64 s[46:47], s[16:17], 24
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 24
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s20, 16
+; GFX11-TRUE16-NEXT: s_lshr_b64 s[44:45], s[2:3], 24
+; GFX11-TRUE16-NEXT: s_lshr_b64 s[76:77], s[6:7], 24
+; GFX11-TRUE16-NEXT: s_lshr_b64 s[78:79], s[8:9], 24
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 23
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s20, 8
+; GFX11-TRUE16-NEXT: s_lshr_b64 s[88:89], s[10:11], 24
+; GFX11-TRUE16-NEXT: s_lshr_b64 s[90:91], s[12:13], 24
+; GFX11-TRUE16-NEXT: s_lshr_b64 s[92:93], s[14:15], 24
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 22
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s19, 24
+; GFX11-TRUE16-NEXT: s_lshr_b64 s[94:95], s[40:41], 24
+; GFX11-TRUE16-NEXT: s_lshr_b64 s[30:31], s[28:29], 24
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 21
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s19, 16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 20
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s19, 8
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 19
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s18, 16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 18
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s18, 8
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 17
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s17, 24
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s17, 16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 15
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s17, 8
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 14
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s16, 16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 13
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s16, 8
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 12
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s3, 24
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 11
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s3, 16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 10
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s3, 8
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 9
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s2, 16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 8
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s2, 8
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 7
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s1, 24
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 6
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s1, 16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 5
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s1, 8
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 4
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s0, 16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 1
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s0, 8
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 0
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s28, 8
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s74, 2
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s75, 3
+; GFX11-TRUE16-NEXT: s_lshr_b64 s[74:75], s[4:5], 24
+; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, vcc_hi
+; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB99_4
+; GFX11-TRUE16-NEXT: .LBB99_2: ; %cmp.true
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v51, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v50, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v36, s19, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v4, s21, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v28, s27, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v27, s26, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v3, s20, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v35, s18, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v22, s29, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v21, s28, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v18, s41, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, s40, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v55, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v54, s0, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v49, s17, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v48, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, s23, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v32, s25, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v31, s24, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v1, s22, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[23:24], 24, v[7:8]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[37:38], 24, v[27:28]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[3:4]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[35:36]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[80:81], 24, v[50:51]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[9:10]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[31:32]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[52:53], 24, v[1:2]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[70:71], 24, v[48:49]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[81:82], 24, v[54:55]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[19:20], 24, v[5:6]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[25:26], 24, v[11:12]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[29:30], 24, v[13:14]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[33:34], 24, v[15:16]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[17:18]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[21:22]
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 24, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 16, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 16, v27
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v27
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v32
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 16, v32
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v32
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 16, v31
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v164, 8, v31
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v165, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v166, 16, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v167, 8, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v176, 16, v1
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v177, 8, v1
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v178, 24, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v179, 16, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v180, 8, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v181, 16, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v182, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v183, 24, v36
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v40, 16, v36
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v41, 8, v36
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v42, 16, v35
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v43, 8, v35
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v44, 24, v49
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v45, 16, v49
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v46, 8, v49
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v47, 16, v48
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v56, 8, v48
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v57, 24, v51
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v58, 16, v51
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v59, 8, v51
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v60, 16, v50
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v61, 8, v50
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v62, 24, v55
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v63, 16, v55
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v72, 8, v55
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v73, 16, v54
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v74, 8, v54
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 16, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v5
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 24, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 16, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 16, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 16, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 16, v9
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v9
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 16, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 16, v11
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 8, v11
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 24, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 16, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 16, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 16, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 16, v15
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v15
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 16, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 16, v17
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v17
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 16, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 8, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 16, v21
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v21
+; GFX11-TRUE16-NEXT: s_branch .LBB99_5
+; GFX11-TRUE16-NEXT: .LBB99_3:
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 0
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr44_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr46_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr56_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr58_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr60_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr34_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr104_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr103_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr102_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr101_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr100_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr99_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr98_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr97_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr96_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr87_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr86_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr85_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr84_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr83_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr82_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr81_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr80_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr71_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr70_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr69_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr68_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr67_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr66_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr65_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr64_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr55_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr54_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr53_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr52_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr51_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr50_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr49_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr48_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr39_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr38_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr37_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr36_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr35_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr72_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr62_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr30_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr94_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr92_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr90_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr88_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr78_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr76_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr74_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 1
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 2
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s43, 3
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr43_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 4
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 5
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 6
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 7
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 8
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 9
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 10
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 11
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 12
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 13
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 14
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 15
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 16
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 17
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 18
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 19
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 20
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 21
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 22
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 23
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 24
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 25
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 26
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 27
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 28
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 29
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 30
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v78, s42, 31
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 0
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 1
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 2
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 3
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 4
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 5
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 6
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 7
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: v_writelane_b32 v77, s42, 8
+; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42_lo16
+; GFX11-TRUE16-NEXT: s_branch .LBB99_2
+; GFX11-TRUE16-NEXT: .LBB99_4:
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, s1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, s28
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, s29
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, s40
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v74.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, s41
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, s14
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, s15
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, s12
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v73.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 4
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, s13
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, s10
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, s11
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, s8
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v72.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 5
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, s9
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, s6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, s7
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, s4
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v63.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, s5
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, s2
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, s3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, s16
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v62.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 7
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, s17
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, s18
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, s19
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, s20
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v61.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 8
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, s21
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, s22
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, s23
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, s24
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v60.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 9
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, s25
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, s26
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, s27
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v150.l, s42
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v59.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 10
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.l, s34
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.l, s104
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v135.l, s103
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.l, s102
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v58.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 11
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.l, s101
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v132.l, s100
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, s99
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.l, s98
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v57.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 12
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.l, s97
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.l, s96
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.l, s87
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v118.l, s86
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v56.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 13
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v117.l, s85
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v116.l, s84
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v115.l, s83
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v114.l, s82
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v47.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 14
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v113.l, s81
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v112.l, s80
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v103.l, s71
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v102.l, s70
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v46.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 15
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v101.l, s69
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v100.l, s68
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v99.l, s67
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v98.l, s66
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v45.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 16
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v97.l, s65
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v96.l, s64
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v87.l, s55
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v86.l, s54
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v44.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 17
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v85.l, s53
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v84.l, s52
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v83.l, s51
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v82.l, s50
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v43.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 18
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v71.l, s49
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, s48
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, s39
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, s38
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v42.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 19
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, s37
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, s36
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, s35
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v145.l, s43
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v41.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 20
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, s62
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, s72
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, s60
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, s58
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v40.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 21
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, s56
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v70.l, s46
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v80.l, s44
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, s74
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v183.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 22
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, s76
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, s78
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, s88
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, s90
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v182.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 23
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, s92
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, s94
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v68.l, s30
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v181.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 24
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v180.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 25
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v179.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 26
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v178.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 27
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v177.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 28
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v176.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 29
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v167.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 30
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v166.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 31
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v165.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v77, 0
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v164.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v77, 1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v163.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v77, 2
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v162.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v77, 3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v161.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v77, 4
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v160.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v77, 5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v151.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v77, 6
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v148.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v77, 7
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v147.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v77, 8
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.l, s0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v78, 2
+; GFX11-TRUE16-NEXT: v_readlane_b32 s1, v78, 3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v81.l, s0
+; GFX11-TRUE16-NEXT: .LBB99_5: ; %end
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v69, 8, v74
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xff, v54
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v66, 8, v81
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xff, v55
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 8, v62
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v80, 8, v80
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v54, v54, v69
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v69, 0xff, v73
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xff, v50
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v60, 0xff, v60
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xff, v51
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v54
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v66, v69, v66
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v69, 8, v72
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v80, v60, v80
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v57, 8, v57
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xff, v48
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v66, 16, v66
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v55, v55, v69
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v69, 0xff, v63
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v70, 8, v70
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xff, v49
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v54, v54, v66
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v55
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v69, v69, v81
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 8, v61
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xff, v35
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v67, 8, v67
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xff, v36
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v66, 16, v69
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v50, v50, v81
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v69, 8, v59
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v81, 0xff, v58
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v64, 8, v64
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v55, v55, v66
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff, v50
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v66, 16, v80
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v51, v51, v69
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v69, v81, v57
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v80, 8, v56
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v81, 0xff, v47
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v56, v50, v66
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff, v51
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v69
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v48, v48, v80
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v66, v81, v70
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v69, 8, v46
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v70, 0xff, v45
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v80, 8, v44
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v57, v50, v51
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff, v48
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v66
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v49, v49, v69
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v51, v70, v80
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v66, 8, v43
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v69, 0xff, v42
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v48, v48, v50
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v49
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v51
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v35, v35, v66
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v51, v69, v67
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v66, 8, v41
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v67, 0xff, v40
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v69, 8, v183
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v49, v49, v50
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v51
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v36, v36, v66
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v35
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v51, v67, v69
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v67, 0xff, v181
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v36
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v66, 8, v182
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v64, v67, v64
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xff, v4
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v69, 8, v180
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v50, v35, v50
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v51, v36, v51
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v35, 16, v64
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xff, v179
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v64, 8, v178
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v3, v66
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v4, v69
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v66, 8, v177
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v67, 0xff, v176
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 8, v52
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v69, 8, v167
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v36, v36, v64
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v66
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v52, v67, v52
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v2, v69
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v36, 16, v36
+; GFX11-TRUE16-NEXT: s_clause 0x1
+; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[54:57], off
+; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[48:51], off offset:16
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v66, 0xffff, v1
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v52
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v67, 0xffff, v2
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v35
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v36
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xff, v31
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v35, 8, v164
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xff, v163
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 8, v38
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xff, v32
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v48, 8, v162
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xff, v161
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 8, v160
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xff, v27
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 8, v151
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v66, v52
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v31, v35
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v35, v36, v38
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v32, v48
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v36, v49, v50
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v27, v51
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xff, v148
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v37, 8, v37
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xff, v28
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v48, 8, v147
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xff, v146
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 8, v145
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xff, v21
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 8, v150
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xff, v149
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 8, v68
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, v38, v37
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v28, v48
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, v49, v50
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v21, v51
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v48, v52, v54
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v31
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v35, 16, v35
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v27
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v37, 16, v37
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff, v28
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 16, v38
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v48, 16, v48
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff, v32
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v36, 16, v36
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v35, v31, v35
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, v27, v37
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, v28, v38
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v48, v21, v48
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xff, v22
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 8, v144
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xff, v135
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 8, v134
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xff, v17
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 8, v133
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v36, v32, v36
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xff, v132
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v49, 8, v65
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xff, v18
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 8, v131
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v21, v22
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v27, v28
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v17, v31
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xff, v130
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 8, v129
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v32, v49
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v18, v50
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xff, v15
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 8, v128
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xff, v119
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v33, 8, v33
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v28, v31
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v17
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v27
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v18
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xff, v16
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 8, v118
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xff, v117
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 8, v116
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v15, v32
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v49, v33
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v28
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v16, v50
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v51, v52
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v31
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v49, v21, v22
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v50, v17, v27
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v51, v18, v28
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xff, v13
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 8, v115
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xff, v114
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 8, v29
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xff, v14
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 8, v113
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xff, v11
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v29, 8, v102
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v15, v31
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xff, v112
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 8, v103
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v13, v17
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v18, v21
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v14, v22
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v11, v29
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xff, v101
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 8, v25
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xff, v9
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v29, 8, v97
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xff, v96
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 8, v24
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v27, v28
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xff, v12
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 8, v100
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xff, v99
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 8, v98
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v21, v22
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v9, v29
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v31, v24
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v14
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v12, v25
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v27, v28
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v13, v17
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v14, v18
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v11, v21
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v9, v24
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xff, v10
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 8, v87
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xff, v86
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 8, v85
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xff, v83
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 8, v23
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v70, 0xff, v166
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v80, 8, v165
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v12, v22
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xff, v7
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 8, v84
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xff, v8
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 8, v82
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v9, v10
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v14, v21
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v24, v23
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v64, v70, v80
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v7, v22
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v8, v25
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 8, v53
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v14
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xff, v71
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xff, v5
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 8, v39
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xff, v34
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 8, v19
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xff, v6
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 8, v30
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xff, v26
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 8, v20
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v64, 16, v64
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v14, v22
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v5, v23
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v24, v19
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v6, v25
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v26, v20
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v32
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v67, v64
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v14
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v5
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v19
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v6
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v20
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v16, v32
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v9, v10
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v7, v21
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v8, v22
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v23, v19
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v24, v20
+; GFX11-TRUE16-NEXT: s_clause 0x5
+; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off offset:32
+; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[35:38], off offset:48
+; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[48:51], off offset:64
+; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[15:18], off offset:80
+; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[11:14], off offset:96
+; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:112
+; GFX11-TRUE16-NEXT: s_clause 0x12
+; GFX11-TRUE16-NEXT: scratch_load_b32 v74, off, s32
+; GFX11-TRUE16-NEXT: scratch_load_b32 v73, off, s32 offset:4
+; GFX11-TRUE16-NEXT: scratch_load_b32 v72, off, s32 offset:8
+; GFX11-TRUE16-NEXT: scratch_load_b32 v63, off, s32 offset:12
+; GFX11-TRUE16-NEXT: scratch_load_b32 v62, off, s32 offset:16
+; GFX11-TRUE16-NEXT: scratch_load_b32 v61, off, s32 offset:20
+; GFX11-TRUE16-NEXT: scratch_load_b32 v60, off, s32 offset:24
+; GFX11-TRUE16-NEXT: scratch_load_b32 v59, off, s32 offset:28
+; GFX11-TRUE16-NEXT: scratch_load_b32 v58, off, s32 offset:32
+; GFX11-TRUE16-NEXT: scratch_load_b32 v57, off, s32 offset:36
+; GFX11-TRUE16-NEXT: scratch_load_b32 v56, off, s32 offset:40
+; GFX11-TRUE16-NEXT: scratch_load_b32 v47, off, s32 offset:44
+; GFX11-TRUE16-NEXT: scratch_load_b32 v46, off, s32 offset:48
+; GFX11-TRUE16-NEXT: scratch_load_b32 v45, off, s32 offset:52
+; GFX11-TRUE16-NEXT: scratch_load_b32 v44, off, s32 offset:56
+; GFX11-TRUE16-NEXT: scratch_load_b32 v43, off, s32 offset:60
+; GFX11-TRUE16-NEXT: scratch_load_b32 v42, off, s32 offset:64
+; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s32 offset:68
+; GFX11-TRUE16-NEXT: scratch_load_b32 v40, off, s32 offset:72
+; GFX11-TRUE16-NEXT: v_readlane_b32 s104, v76, 8
+; GFX11-TRUE16-NEXT: v_readlane_b32 s103, v76, 7
+; GFX11-TRUE16-NEXT: v_readlane_b32 s102, v76, 6
+; GFX11-TRUE16-NEXT: v_readlane_b32 s101, v76, 5
+; GFX11-TRUE16-NEXT: v_readlane_b32 s100, v76, 4
+; GFX11-TRUE16-NEXT: v_readlane_b32 s99, v76, 3
+; GFX11-TRUE16-NEXT: v_readlane_b32 s98, v76, 2
+; GFX11-TRUE16-NEXT: v_readlane_b32 s97, v76, 1
+; GFX11-TRUE16-NEXT: v_readlane_b32 s96, v76, 0
+; GFX11-TRUE16-NEXT: v_readlane_b32 s87, v75, 31
+; GFX11-TRUE16-NEXT: v_readlane_b32 s86, v75, 30
+; GFX11-TRUE16-NEXT: v_readlane_b32 s85, v75, 29
+; GFX11-TRUE16-NEXT: v_readlane_b32 s84, v75, 28
+; GFX11-TRUE16-NEXT: v_readlane_b32 s83, v75, 27
+; GFX11-TRUE16-NEXT: v_readlane_b32 s82, v75, 26
+; GFX11-TRUE16-NEXT: v_readlane_b32 s81, v75, 25
+; GFX11-TRUE16-NEXT: v_readlane_b32 s80, v75, 24
+; GFX11-TRUE16-NEXT: v_readlane_b32 s71, v75, 23
+; GFX11-TRUE16-NEXT: v_readlane_b32 s70, v75, 22
+; GFX11-TRUE16-NEXT: v_readlane_b32 s69, v75, 21
+; GFX11-TRUE16-NEXT: v_readlane_b32 s68, v75, 20
+; GFX11-TRUE16-NEXT: v_readlane_b32 s67, v75, 19
+; GFX11-TRUE16-NEXT: v_readlane_b32 s66, v75, 18
+; GFX11-TRUE16-NEXT: v_readlane_b32 s65, v75, 17
+; GFX11-TRUE16-NEXT: v_readlane_b32 s64, v75, 16
+; GFX11-TRUE16-NEXT: v_readlane_b32 s55, v75, 15
+; GFX11-TRUE16-NEXT: v_readlane_b32 s54, v75, 14
+; GFX11-TRUE16-NEXT: v_readlane_b32 s53, v75, 13
+; GFX11-TRUE16-NEXT: v_readlane_b32 s52, v75, 12
+; GFX11-TRUE16-NEXT: v_readlane_b32 s51, v75, 11
+; GFX11-TRUE16-NEXT: v_readlane_b32 s50, v75, 10
+; GFX11-TRUE16-NEXT: v_readlane_b32 s49, v75, 9
+; GFX11-TRUE16-NEXT: v_readlane_b32 s48, v75, 8
+; GFX11-TRUE16-NEXT: v_readlane_b32 s39, v75, 7
+; GFX11-TRUE16-NEXT: v_readlane_b32 s38, v75, 6
+; GFX11-TRUE16-NEXT: v_readlane_b32 s37, v75, 5
+; GFX11-TRUE16-NEXT: v_readlane_b32 s36, v75, 4
+; GFX11-TRUE16-NEXT: v_readlane_b32 s35, v75, 3
+; GFX11-TRUE16-NEXT: v_readlane_b32 s34, v75, 2
+; GFX11-TRUE16-NEXT: v_readlane_b32 s31, v75, 1
+; GFX11-TRUE16-NEXT: v_readlane_b32 s30, v75, 0
+; GFX11-TRUE16-NEXT: s_or_saveexec_b32 s0, -1
+; GFX11-TRUE16-NEXT: s_clause 0x3
+; GFX11-TRUE16-NEXT: scratch_load_b32 v75, off, s32 offset:76
+; GFX11-TRUE16-NEXT: scratch_load_b32 v76, off, s32 offset:80
+; GFX11-TRUE16-NEXT: scratch_load_b32 v77, off, s32 offset:84
+; GFX11-TRUE16-NEXT: scratch_load_b32 v78, off, s32 offset:88
+; GFX11-TRUE16-NEXT: s_mov_b32 exec_lo, s0
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-FAKE16-LABEL: bitcast_v64i16_to_v128i8_scalar:
+; GFX11-FAKE16: ; %bb.0:
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-FAKE16-NEXT: s_or_saveexec_b32 s4, -1
+; GFX11-FAKE16-NEXT: s_clause 0x3
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v75, s32 offset:76
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v76, s32 offset:80
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v77, s32 offset:84
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v78, s32 offset:88
+; GFX11-FAKE16-NEXT: s_mov_b32 exec_lo, s4
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s30, 0
+; GFX11-FAKE16-NEXT: v_writelane_b32 v76, s96, 0
+; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v15
+; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s40, v1
+; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s41, v2
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s31, 1
+; GFX11-FAKE16-NEXT: v_writelane_b32 v76, s97, 1
+; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s14, v3
+; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s15, v4
+; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s12, v5
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s34, 2
+; GFX11-FAKE16-NEXT: v_writelane_b32 v76, s98, 2
+; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s13, v6
+; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s10, v7
+; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s11, v8
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s35, 3
+; GFX11-FAKE16-NEXT: v_writelane_b32 v76, s99, 3
+; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s8, v9
+; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s9, v10
+; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s6, v11
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s36, 4
+; GFX11-FAKE16-NEXT: v_writelane_b32 v76, s100, 4
+; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s7, v12
+; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s4, v13
+; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s5, v14
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s37, 5
+; GFX11-FAKE16-NEXT: v_writelane_b32 v76, s101, 5
+; GFX11-FAKE16-NEXT: s_mov_b32 s99, 0
+; GFX11-FAKE16-NEXT: s_and_b32 s42, vcc_lo, exec_lo
+; GFX11-FAKE16-NEXT: s_clause 0x12
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v40, s32 offset:72
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v41, s32 offset:68
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v42, s32 offset:64
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v43, s32 offset:60
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v44, s32 offset:56
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v45, s32 offset:52
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v46, s32 offset:48
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v47, s32 offset:44
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v56, s32 offset:40
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v57, s32 offset:36
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v58, s32 offset:32
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v59, s32 offset:28
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v60, s32 offset:24
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v61, s32 offset:20
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v62, s32 offset:16
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v63, s32 offset:12
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v72, s32 offset:8
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v73, s32 offset:4
+; GFX11-FAKE16-NEXT: scratch_store_b32 off, v74, s32
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s38, 6
+; GFX11-FAKE16-NEXT: v_writelane_b32 v76, s102, 6
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr78 : SGPR spill to VGPR lane
+; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr77 : SGPR spill to VGPR lane
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s39, 7
+; GFX11-FAKE16-NEXT: v_writelane_b32 v76, s103, 7
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s48, 8
+; GFX11-FAKE16-NEXT: v_writelane_b32 v76, s104, 8
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s49, 9
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s50, 10
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s51, 11
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s52, 12
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s53, 13
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s54, 14
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s55, 15
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s64, 16
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s65, 17
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s66, 18
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s67, 19
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s68, 20
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s69, 21
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s70, 22
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s71, 23
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s80, 24
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s81, 25
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s82, 26
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s83, 27
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s84, 28
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s85, 29
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s86, 30
+; GFX11-FAKE16-NEXT: v_writelane_b32 v75, s87, 31
+; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB99_3
+; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s27, 16
+; GFX11-FAKE16-NEXT: s_lshr_b64 s[74:75], s[0:1], 24
+; GFX11-FAKE16-NEXT: v_writelane_b32 v77, s42, 8
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s27, 8
+; GFX11-FAKE16-NEXT: s_lshr_b32 s43, s27, 24
+; GFX11-FAKE16-NEXT: s_lshr_b32 s34, s5, 24
+; GFX11-FAKE16-NEXT: s_lshr_b32 s35, s5, 16
+; GFX11-FAKE16-NEXT: v_writelane_b32 v77, s42, 7
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s26, 16
+; GFX11-FAKE16-NEXT: s_lshr_b32 s37, s5, 8
+; GFX11-FAKE16-NEXT: s_lshr_b32 s36, s4, 16
+; GFX11-FAKE16-NEXT: s_lshr_b32 s38, s4, 8
+; GFX11-FAKE16-NEXT: v_writelane_b32 v77, s42, 6
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s26, 8
+; GFX11-FAKE16-NEXT: s_lshr_b32 s39, s7, 24
+; GFX11-FAKE16-NEXT: s_lshr_b32 s48, s7, 16
+; GFX11-FAKE16-NEXT: s_lshr_b32 s50, s7, 8
+; GFX11-FAKE16-NEXT: v_writelane_b32 v77, s42, 5
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s25, 24
+; GFX11-FAKE16-NEXT: s_lshr_b32 s49, s6, 16
+; GFX11-FAKE16-NEXT: s_lshr_b32 s51, s6, 8
+; GFX11-FAKE16-NEXT: s_lshr_b32 s52, s9, 24
+; GFX11-FAKE16-NEXT: v_writelane_b32 v77, s42, 4
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s25, 16
+; GFX11-FAKE16-NEXT: s_lshr_b32 s53, s9, 16
+; GFX11-FAKE16-NEXT: s_lshr_b32 s55, s9, 8
+; GFX11-FAKE16-NEXT: s_lshr_b32 s54, s8, 16
+; GFX11-FAKE16-NEXT: v_writelane_b32 v77, s42, 3
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s25, 8
+; GFX11-FAKE16-NEXT: s_lshr_b32 s64, s8, 8
+; GFX11-FAKE16-NEXT: s_lshr_b32 s65, s11, 24
+; GFX11-FAKE16-NEXT: s_lshr_b32 s66, s11, 16
+; GFX11-FAKE16-NEXT: v_writelane_b32 v77, s42, 2
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s24, 16
+; GFX11-FAKE16-NEXT: s_lshr_b32 s68, s11, 8
+; GFX11-FAKE16-NEXT: s_lshr_b32 s67, s10, 16
+; GFX11-FAKE16-NEXT: s_lshr_b32 s69, s10, 8
+; GFX11-FAKE16-NEXT: v_writelane_b32 v77, s42, 1
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s24, 8
+; GFX11-FAKE16-NEXT: s_lshr_b32 s70, s13, 24
+; GFX11-FAKE16-NEXT: s_lshr_b32 s71, s13, 16
+; GFX11-FAKE16-NEXT: s_lshr_b32 s81, s13, 8
+; GFX11-FAKE16-NEXT: v_writelane_b32 v77, s42, 0
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s23, 24
+; GFX11-FAKE16-NEXT: s_lshr_b32 s80, s12, 16
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 31
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s23, 16
+; GFX11-FAKE16-NEXT: s_lshr_b32 s82, s12, 8
+; GFX11-FAKE16-NEXT: s_lshr_b32 s83, s15, 24
+; GFX11-FAKE16-NEXT: s_lshr_b32 s84, s15, 16
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 30
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s23, 8
+; GFX11-FAKE16-NEXT: s_lshr_b32 s86, s15, 8
+; GFX11-FAKE16-NEXT: s_lshr_b32 s85, s14, 16
+; GFX11-FAKE16-NEXT: s_lshr_b32 s87, s14, 8
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 29
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s22, 16
+; GFX11-FAKE16-NEXT: s_lshr_b32 s96, s41, 24
+; GFX11-FAKE16-NEXT: s_lshr_b32 s97, s41, 16
+; GFX11-FAKE16-NEXT: s_lshr_b32 s100, s41, 8
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 28
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s22, 8
+; GFX11-FAKE16-NEXT: s_lshr_b32 s98, s40, 16
+; GFX11-FAKE16-NEXT: s_lshr_b32 s101, s40, 8
+; GFX11-FAKE16-NEXT: s_lshr_b32 s102, s29, 24
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 27
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s21, 24
+; GFX11-FAKE16-NEXT: s_lshr_b32 s103, s29, 16
+; GFX11-FAKE16-NEXT: s_lshr_b32 vcc_hi, s29, 8
+; GFX11-FAKE16-NEXT: s_lshr_b32 s104, s28, 16
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 26
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s21, 16
+; GFX11-FAKE16-NEXT: s_lshr_b64 s[62:63], s[26:27], 24
+; GFX11-FAKE16-NEXT: s_lshr_b64 s[72:73], s[24:25], 24
+; GFX11-FAKE16-NEXT: s_lshr_b64 s[60:61], s[22:23], 24
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 25
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s21, 8
+; GFX11-FAKE16-NEXT: s_lshr_b64 s[58:59], s[20:21], 24
+; GFX11-FAKE16-NEXT: s_lshr_b64 s[56:57], s[18:19], 24
+; GFX11-FAKE16-NEXT: s_lshr_b64 s[46:47], s[16:17], 24
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 24
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s20, 16
+; GFX11-FAKE16-NEXT: s_lshr_b64 s[44:45], s[2:3], 24
+; GFX11-FAKE16-NEXT: s_lshr_b64 s[76:77], s[6:7], 24
+; GFX11-FAKE16-NEXT: s_lshr_b64 s[78:79], s[8:9], 24
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 23
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s20, 8
+; GFX11-FAKE16-NEXT: s_lshr_b64 s[88:89], s[10:11], 24
+; GFX11-FAKE16-NEXT: s_lshr_b64 s[90:91], s[12:13], 24
+; GFX11-FAKE16-NEXT: s_lshr_b64 s[92:93], s[14:15], 24
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 22
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s19, 24
+; GFX11-FAKE16-NEXT: s_lshr_b64 s[94:95], s[40:41], 24
+; GFX11-FAKE16-NEXT: s_lshr_b64 s[30:31], s[28:29], 24
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 21
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s19, 16
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 20
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s19, 8
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 19
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s18, 16
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 18
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s18, 8
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 17
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s17, 24
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 16
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s17, 16
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 15
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s17, 8
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 14
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s16, 16
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 13
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s16, 8
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 12
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s3, 24
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 11
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s3, 16
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 10
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s3, 8
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 9
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s2, 16
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 8
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s2, 8
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 7
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s1, 24
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 6
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s1, 16
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 5
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s1, 8
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 4
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s0, 16
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 3
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s0, 8
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 2
+; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s28, 8
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s74, 0
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s75, 1
+; GFX11-FAKE16-NEXT: s_lshr_b64 s[74:75], s[4:5], 24
+; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s99
+; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB99_4
+; GFX11-FAKE16-NEXT: .LBB99_2: ; %cmp.true
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v39, s17, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v38, s16, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v2, s5, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v1, s4, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v51, s3, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v50, s2, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v33, s21, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v29, s23, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v28, s22, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v32, s20, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v16, s29, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v15, s28, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v14, s41, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v13, s40, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v12, s15, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v11, s14, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v10, s13, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v9, s12, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v8, s11, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v7, s10, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v6, s9, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v5, s8, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v4, s7, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v3, s6, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v53, s1, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v52, s0, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v37, s19, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v25, s25, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v21, s27, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v20, s26, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v24, s24, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_pk_add_u16 v36, s18, 3 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[80:81], 24, v[38:39]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[28:29]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[32:33]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[81:82], 24, v[50:51]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[17:18], 24, v[1:2]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[20:21]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[24:25]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[70:71], 24, v[36:37]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[82:83], 24, v[52:53]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[18:19], 24, v[3:4]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[26:27], 24, v[7:8]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[30:31], 24, v[9:10]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[34:35], 24, v[11:12]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[13:14]
+; GFX11-FAKE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[15:16]
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v147, 24, v21
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v148, 16, v21
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v149, 8, v21
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v150, 16, v20
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v151, 8, v20
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v161, 24, v25
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v160, 16, v25
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v162, 8, v25
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v163, 16, v24
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v164, 8, v24
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v166, 24, v29
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v165, 16, v29
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v167, 8, v29
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v176, 16, v28
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v177, 8, v28
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v179, 24, v33
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v178, 16, v33
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v180, 8, v33
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v181, 16, v32
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v182, 8, v32
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v40, 24, v37
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v183, 16, v37
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v41, 8, v37
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v42, 16, v36
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v43, 8, v36
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v45, 24, v39
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v44, 16, v39
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v46, 8, v39
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v47, 16, v38
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v56, 8, v38
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v58, 24, v51
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v57, 16, v51
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v59, 8, v51
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v61, 16, v50
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v60, 8, v50
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v63, 24, v53
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v62, 16, v53
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v72, 8, v53
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v73, 16, v52
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v74, 8, v52
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v19, 24, v2
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v2
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 8, v2
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v1
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 8, v1
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 24, v4
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v4
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 8, v4
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v83, 16, v3
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v84, 8, v3
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v85, 24, v6
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v86, 16, v6
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v87, 8, v6
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v96, 16, v5
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v97, 8, v5
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v98, 24, v8
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v99, 16, v8
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v100, 8, v8
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v101, 16, v7
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v102, 8, v7
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v103, 24, v10
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v112, 16, v10
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v113, 8, v10
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v114, 16, v9
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v115, 8, v9
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v116, 24, v12
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v117, 16, v12
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v118, 8, v12
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v119, 16, v11
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v128, 8, v11
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v129, 24, v14
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v130, 16, v14
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v131, 8, v14
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v132, 16, v13
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v133, 8, v13
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v134, 24, v16
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v135, 16, v16
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v144, 8, v16
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v145, 16, v15
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v146, 8, v15
+; GFX11-FAKE16-NEXT: s_branch .LBB99_5
+; GFX11-FAKE16-NEXT: .LBB99_3:
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: s_mov_b32 s99, -1
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr44
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr46
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr56
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr58
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr60
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr104
+; GFX11-FAKE16-NEXT: ; implicit-def: $vcc_hi
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr103
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr102
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr101
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr98
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr100
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr97
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr96
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr87
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr85
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr86
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr84
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr83
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr82
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr80
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr81
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr71
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr70
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr69
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr67
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr68
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr66
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr65
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr64
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr54
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr55
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr53
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr52
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr51
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr49
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr50
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr48
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr39
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr38
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr36
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr37
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr35
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr34
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr72
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr62
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr30
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr94
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr92
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr90
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr88
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr78
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr76
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr74
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s42, 0
+; GFX11-FAKE16-NEXT: v_writelane_b32 v78, s43, 1
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: ; kill: killed $sgpr42
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr43
+; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
+; GFX11-FAKE16-NEXT: s_branch .LBB99_2
+; GFX11-FAKE16-NEXT: .LBB99_4:
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v52, s0 :: v_dual_mov_b32 v53, s1
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 2
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v71, s50
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v15, s28 :: v_dual_mov_b32 v16, s29
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v13, s40 :: v_dual_mov_b32 v14, s41
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v74, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 3
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v11, s14 :: v_dual_mov_b32 v12, s15
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v9, s12 :: v_dual_mov_b32 v10, s13
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v73, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 4
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v55, s48
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v7, s10 :: v_dual_mov_b32 v8, s11
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v5, s8 :: v_dual_mov_b32 v6, s9
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v72, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 5
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v49, s39
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v3, s6 :: v_dual_mov_b32 v4, s7
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v1, s4 :: v_dual_mov_b32 v2, s5
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v62, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 6
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v50, s2 :: v_dual_mov_b32 v51, s3
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v38, s16 :: v_dual_mov_b32 v39, s17
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v63, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 7
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v35, s38 :: v_dual_mov_b32 v36, s18
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v37, s19 :: v_dual_mov_b32 v32, s20
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v33, s21 :: v_dual_mov_b32 v60, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 8
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v28, s22 :: v_dual_mov_b32 v29, s23
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v61, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 9
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v20, s26 :: v_dual_mov_b32 v21, s27
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v146, s42 :: v_dual_mov_b32 v145, s104
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v59, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 10
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v144, vcc_hi :: v_dual_mov_b32 v135, s103
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v134, s102 :: v_dual_mov_b32 v133, s101
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v57, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 11
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v31, s36 :: v_dual_mov_b32 v132, s98
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v131, s100 :: v_dual_mov_b32 v130, s97
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v129, s96 :: v_dual_mov_b32 v58, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 12
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v27, s37 :: v_dual_mov_b32 v128, s87
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v119, s85 :: v_dual_mov_b32 v118, s86
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v117, s84 :: v_dual_mov_b32 v56, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 13
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v116, s83 :: v_dual_mov_b32 v115, s82
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v114, s80 :: v_dual_mov_b32 v113, s81
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v47, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 14
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v23, s35 :: v_dual_mov_b32 v112, s71
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v103, s70 :: v_dual_mov_b32 v102, s69
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v101, s67 :: v_dual_mov_b32 v46, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 15
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v19, s34 :: v_dual_mov_b32 v100, s68
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v99, s66 :: v_dual_mov_b32 v98, s65
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v97, s64 :: v_dual_mov_b32 v44, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 16
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v96, s54 :: v_dual_mov_b32 v87, s55
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v86, s53 :: v_dual_mov_b32 v85, s52
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v45, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 17
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v84, s51 :: v_dual_mov_b32 v83, s49
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v147, s43 :: v_dual_mov_b32 v22, s78
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v43, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 18
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v67, s58 :: v_dual_mov_b32 v26, s88
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v81, s44 :: v_dual_mov_b32 v30, s90
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v42, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 19
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v17, s74 :: v_dual_mov_b32 v34, s92
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v65, s94 :: v_dual_mov_b32 v68, s30
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v41, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 20
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v48, s62
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v54, s72
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v64, s60
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-FAKE16-NEXT: v_dual_mov_b32 v70, s56 :: v_dual_mov_b32 v183, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 21
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v80, s46
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v18, s76
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v40, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 22
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v182, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 23
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v181, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 24
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v180, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 25
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v178, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 26
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v179, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 27
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v177, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 28
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v176, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 29
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v167, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 30
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v165, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 31
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v166, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v77, 0
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v164, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v77, 1
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v163, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v77, 2
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v162, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v77, 3
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v160, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v77, 4
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v161, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v77, 5
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v151, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v77, 6
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v150, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v77, 7
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v149, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v77, 8
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v148, s0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s0, v78, 0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s1, v78, 1
+; GFX11-FAKE16-NEXT: v_mov_b32_e32 v82, s0
+; GFX11-FAKE16-NEXT: .LBB99_5: ; %end
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v69, 8, v74
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v52, 0xff, v52
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v66, 8, v82
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v53, 0xff, v53
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v82, 8, v63
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v81, 8, v81
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v52, v52, v69
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v69, 0xff, v73
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v50, 0xff, v50
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v57, 0xff, v57
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v58, 8, v58
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v52, 0xffff, v52
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v66, v69, v66
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v69, 8, v72
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v38, 0xff, v38
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v80, 8, v80
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v39, 0xff, v39
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v66, 16, v66
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v53, v53, v69
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v69, 0xff, v62
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v70, 8, v70
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v67, 8, v67
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v32, 0xff, v32
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xff, v33
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v69, v69, v82
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v82, 8, v60
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v60, 0xff, v61
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v28, 0xff, v28
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v64, 8, v64
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v29, 0xff, v29
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v82, v50, v82
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v81, v60, v81
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v50, v52, v66
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v52, 0xffff, v53
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v53, 16, v69
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v66, 0xff, v51
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v69, 8, v59
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v24, 0xff, v24
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v25, 0xff, v25
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v51, v52, v53
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v52, 0xffff, v82
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v53, 16, v81
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v66, v66, v69
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v69, v57, v58
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v81, 8, v56
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v82, 0xff, v47
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v52, v52, v53
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v53, 0xffff, v66
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v66, 16, v69
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v38, v38, v81
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v69, v82, v80
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v80, 8, v46
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v81, 0xff, v44
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v82, 8, v45
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v53, v53, v66
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v38, 0xffff, v38
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v66, 16, v69
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v39, v39, v80
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v69, v81, v82
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v80, 0xff, v36
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v81, 8, v43
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v82, 0xff, v42
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v36, v38, v66
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v38, 0xffff, v39
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v39, 16, v69
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v66, v80, v81
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v69, v82, v70
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v70, 0xff, v37
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v80, 8, v41
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v81, 0xff, v183
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v82, 8, v40
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v37, v38, v39
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v38, 0xffff, v66
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v39, 16, v69
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v66, v70, v80
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v69, v81, v82
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v80, 0xff, v181
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v70, 8, v182
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v81, 8, v180
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v66, 0xffff, v66
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v69, 16, v69
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v67, v80, v67
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v32, v32, v70
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v33, v33, v81
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v38, v38, v39
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v39, v66, v69
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v66, 16, v67
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v67, 0xff, v178
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v69, 8, v179
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v70, 8, v177
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v80, 0xff, v176
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v81, 8, v167
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v82, 0xff, v165
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v165, 8, v166
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v67, v67, v69
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v28, v28, v70
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v64, v80, v64
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v29, v29, v81
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v69, v82, v165
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v32, 0xffff, v32
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xffff, v33
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v67, 16, v67
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v28, 0xffff, v28
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v64, 16, v64
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v29, 0xffff, v29
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v69, 16, v69
+; GFX11-FAKE16-NEXT: s_clause 0x1
+; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[50:53], off
+; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[36:39], off offset:16
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v36, v32, v66
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v37, v33, v67
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v38, v28, v64
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v39, v29, v69
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v28, 8, v164
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v29, 0xff, v163
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v32, 8, v54
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v33, 8, v162
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v50, 0xff, v160
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v51, 8, v161
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xff, v20
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v52, 8, v151
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v24, v24, v28
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v28, v29, v32
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v25, v25, v33
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v29, v50, v51
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v20, v20, v52
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v32, 0xff, v150
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v33, 8, v48
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xff, v21
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v48, 8, v149
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v50, 0xff, v148
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v51, 8, v147
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xff, v15
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v52, 8, v146
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v53, 0xff, v145
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v54, 8, v68
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v32, v32, v33
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v21, v21, v48
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v33, v50, v51
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v15, v15, v52
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v48, v53, v54
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v24, 0xffff, v24
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v28, 16, v28
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xffff, v20
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v32, 16, v32
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v21
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v33, 16, v33
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xffff, v15
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v48, 16, v48
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v25, 0xffff, v25
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v29, 16, v29
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v50, v24, v28
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v52, v20, v32
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v53, v21, v33
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v64, v15, v48
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xff, v16
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v16, 8, v144
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xff, v135
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v21, 8, v134
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xff, v13
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v24, 8, v133
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v51, v25, v29
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v25, 0xff, v132
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v28, 8, v65
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xff, v14
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v29, 8, v131
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v15, v15, v16
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v16, v20, v21
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v13, v13, v24
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xff, v130
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v24, 8, v129
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v20, v25, v28
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v14, v14, v29
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xff, v11
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v25, 8, v128
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v28, 0xff, v119
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v29, 8, v34
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xff, v12
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v32, 8, v118
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xff, v117
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v34, 8, v116
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v21, v21, v24
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xffff, v15
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xffff, v13
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v20, 16, v20
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xffff, v14
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v11, v11, v25
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v24, v28, v29
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v12, v12, v32
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v25, v33, v34
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v25, 16, v25
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v65, v15, v16
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v66, v13, v20
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v67, v14, v21
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v9, 0xff, v9
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v13, 8, v115
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xff, v114
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v15, 8, v30
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xff, v10
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v16, 8, v113
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xff, v112
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v21, 8, v103
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v11, v11, v24
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v12, v12, v25
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xff, v7
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v24, 8, v102
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v9, v9, v13
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v13, v14, v15
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v10, v10, v16
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v14, v20, v21
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xff, v101
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v16, 8, v26
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xff, v8
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v20, 8, v100
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xff, v5
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v25, 8, v97
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v26, 0xff, v96
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v22, 8, v22
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v7, v24
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xff, v99
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v24, 8, v98
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v15, v15, v16
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v8, v8, v20
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v5, v25
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v20, v26, v22
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v9, 0xffff, v9
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v16, v21, v24
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v15, 16, v15
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v20, 16, v20
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xffff, v8
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v13, v9, v13
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v14, v10, v14
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v7, v7, v15
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v9, v5, v20
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xff, v6
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v6, 8, v87
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xff, v86
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v15, 8, v85
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xff, v83
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v18, 8, v18
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v8, v8, v16
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xff, v3
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v16, 8, v84
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xff, v4
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v21, 8, v71
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, v5, v6
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v6, v10, v15
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v10, v20, v18
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v3, v16
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v4, v21
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v16, 8, v49
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xff, v1
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v15, 16, v10
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xff, v55
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v18, 8, v35
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xff, v31
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v17, 8, v17
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xff, v2
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v21, 8, v27
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xff, v23
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v19, 8, v19
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v10, v10, v16
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v1, v18
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v16, v20, v17
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v2, v21
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v17, v22, v19
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v18, 16, v10
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xffff, v1
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xffff, v2
+; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v10, v5, v6
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v1, v3, v15
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v2, v4, v18
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v3, v19, v16
+; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, v20, v17
+; GFX11-FAKE16-NEXT: s_clause 0x5
+; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[36:39], off offset:32
+; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[50:53], off offset:48
+; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[64:67], off offset:64
+; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[11:14], off offset:80
+; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[7:10], off offset:96
+; GFX11-FAKE16-NEXT: scratch_store_b128 v0, v[1:4], off offset:112
+; GFX11-FAKE16-NEXT: s_clause 0x12
+; GFX11-FAKE16-NEXT: scratch_load_b32 v74, off, s32
+; GFX11-FAKE16-NEXT: scratch_load_b32 v73, off, s32 offset:4
+; GFX11-FAKE16-NEXT: scratch_load_b32 v72, off, s32 offset:8
+; GFX11-FAKE16-NEXT: scratch_load_b32 v63, off, s32 offset:12
+; GFX11-FAKE16-NEXT: scratch_load_b32 v62, off, s32 offset:16
+; GFX11-FAKE16-NEXT: scratch_load_b32 v61, off, s32 offset:20
+; GFX11-FAKE16-NEXT: scratch_load_b32 v60, off, s32 offset:24
+; GFX11-FAKE16-NEXT: scratch_load_b32 v59, off, s32 offset:28
+; GFX11-FAKE16-NEXT: scratch_load_b32 v58, off, s32 offset:32
+; GFX11-FAKE16-NEXT: scratch_load_b32 v57, off, s32 offset:36
+; GFX11-FAKE16-NEXT: scratch_load_b32 v56, off, s32 offset:40
+; GFX11-FAKE16-NEXT: scratch_load_b32 v47, off, s32 offset:44
+; GFX11-FAKE16-NEXT: scratch_load_b32 v46, off, s32 offset:48
+; GFX11-FAKE16-NEXT: scratch_load_b32 v45, off, s32 offset:52
+; GFX11-FAKE16-NEXT: scratch_load_b32 v44, off, s32 offset:56
+; GFX11-FAKE16-NEXT: scratch_load_b32 v43, off, s32 offset:60
+; GFX11-FAKE16-NEXT: scratch_load_b32 v42, off, s32 offset:64
+; GFX11-FAKE16-NEXT: scratch_load_b32 v41, off, s32 offset:68
+; GFX11-FAKE16-NEXT: scratch_load_b32 v40, off, s32 offset:72
+; GFX11-FAKE16-NEXT: v_readlane_b32 s104, v76, 8
+; GFX11-FAKE16-NEXT: v_readlane_b32 s103, v76, 7
+; GFX11-FAKE16-NEXT: v_readlane_b32 s102, v76, 6
+; GFX11-FAKE16-NEXT: v_readlane_b32 s101, v76, 5
+; GFX11-FAKE16-NEXT: v_readlane_b32 s100, v76, 4
+; GFX11-FAKE16-NEXT: v_readlane_b32 s99, v76, 3
+; GFX11-FAKE16-NEXT: v_readlane_b32 s98, v76, 2
+; GFX11-FAKE16-NEXT: v_readlane_b32 s97, v76, 1
+; GFX11-FAKE16-NEXT: v_readlane_b32 s96, v76, 0
+; GFX11-FAKE16-NEXT: v_readlane_b32 s87, v75, 31
+; GFX11-FAKE16-NEXT: v_readlane_b32 s86, v75, 30
+; GFX11-FAKE16-NEXT: v_readlane_b32 s85, v75, 29
+; GFX11-FAKE16-NEXT: v_readlane_b32 s84, v75, 28
+; GFX11-FAKE16-NEXT: v_readlane_b32 s83, v75, 27
+; GFX11-FAKE16-NEXT: v_readlane_b32 s82, v75, 26
+; GFX11-FAKE16-NEXT: v_readlane_b32 s81, v75, 25
+; GFX11-FAKE16-NEXT: v_readlane_b32 s80, v75, 24
+; GFX11-FAKE16-NEXT: v_readlane_b32 s71, v75, 23
+; GFX11-FAKE16-NEXT: v_readlane_b32 s70, v75, 22
+; GFX11-FAKE16-NEXT: v_readlane_b32 s69, v75, 21
+; GFX11-FAKE16-NEXT: v_readlane_b32 s68, v75, 20
+; GFX11-FAKE16-NEXT: v_readlane_b32 s67, v75, 19
+; GFX11-FAKE16-NEXT: v_readlane_b32 s66, v75, 18
+; GFX11-FAKE16-NEXT: v_readlane_b32 s65, v75, 17
+; GFX11-FAKE16-NEXT: v_readlane_b32 s64, v75, 16
+; GFX11-FAKE16-NEXT: v_readlane_b32 s55, v75, 15
+; GFX11-FAKE16-NEXT: v_readlane_b32 s54, v75, 14
+; GFX11-FAKE16-NEXT: v_readlane_b32 s53, v75, 13
+; GFX11-FAKE16-NEXT: v_readlane_b32 s52, v75, 12
+; GFX11-FAKE16-NEXT: v_readlane_b32 s51, v75, 11
+; GFX11-FAKE16-NEXT: v_readlane_b32 s50, v75, 10
+; GFX11-FAKE16-NEXT: v_readlane_b32 s49, v75, 9
+; GFX11-FAKE16-NEXT: v_readlane_b32 s48, v75, 8
+; GFX11-FAKE16-NEXT: v_readlane_b32 s39, v75, 7
+; GFX11-FAKE16-NEXT: v_readlane_b32 s38, v75, 6
+; GFX11-FAKE16-NEXT: v_readlane_b32 s37, v75, 5
+; GFX11-FAKE16-NEXT: v_readlane_b32 s36, v75, 4
+; GFX11-FAKE16-NEXT: v_readlane_b32 s35, v75, 3
+; GFX11-FAKE16-NEXT: v_readlane_b32 s34, v75, 2
+; GFX11-FAKE16-NEXT: v_readlane_b32 s31, v75, 1
+; GFX11-FAKE16-NEXT: v_readlane_b32 s30, v75, 0
+; GFX11-FAKE16-NEXT: s_or_saveexec_b32 s0, -1
+; GFX11-FAKE16-NEXT: s_clause 0x3
+; GFX11-FAKE16-NEXT: scratch_load_b32 v75, off, s32 offset:76
+; GFX11-FAKE16-NEXT: scratch_load_b32 v76, off, s32 offset:80
+; GFX11-FAKE16-NEXT: scratch_load_b32 v77, off, s32 offset:84
+; GFX11-FAKE16-NEXT: scratch_load_b32 v78, off, s32 offset:88
+; GFX11-FAKE16-NEXT: s_mov_b32 exec_lo, s0
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-FAKE16-NEXT: 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) {
; SI-LABEL: bitcast_v64bf16_to_v64f16:
; SI: ; %bb.0:
@@ -203449,7 +226763,7 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
; SI-NEXT: ; kill: killed $vgpr2
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB96_2
+; SI-NEXT: s_cbranch_execz .LBB100_2
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
@@ -203775,9 +227089,9 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
; SI-NEXT: ; implicit-def: $vgpr61
; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; SI-NEXT: .LBB96_2: ; %Flow
+; SI-NEXT: .LBB100_2: ; %Flow
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB96_4
+; SI-NEXT: s_cbranch_execz .LBB100_4
; SI-NEXT: ; %bb.3: ; %cmp.true
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_and_b32_e32 v2, 0xffff0000, v31
@@ -204167,7 +227481,7 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; SI-NEXT: .LBB96_4: ; %end
+; SI-NEXT: .LBB100_4: ; %end
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
@@ -204565,7 +227879,7 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT: 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 .LBB96_2
+; VI-NEXT: s_cbranch_execz .LBB100_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
@@ -205151,7 +228465,7 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
; 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: .LBB96_2: ; %end
+; VI-NEXT: .LBB100_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
@@ -205198,7 +228512,7 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT: 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 .LBB96_2
+; GFX9-NEXT: s_cbranch_execz .LBB100_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
@@ -205688,7 +229002,7 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
; 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: .LBB96_2: ; %end
+; GFX9-NEXT: .LBB100_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
@@ -205721,7 +229035,7 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB96_2
+; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB100_2
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff0000, v16
; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff0000, v20
@@ -206243,7 +229557,7 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v85, v85, v97, vcc_lo
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_bfi_b32 v13, 0xffff, v13, v85
-; GFX11-TRUE16-NEXT: .LBB96_2: ; %end
+; GFX11-TRUE16-NEXT: .LBB100_2: ; %end
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
@@ -206260,7 +229574,7 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB96_2
+; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB100_2
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.true
; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v33, 16, v17
; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v32, 16, v16
@@ -206770,7 +230084,7 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
; GFX11-FAKE16-NEXT: v_cndmask_b32_e32 v11, v85, v96, vcc_lo
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v11, v83, 0x7060302
-; GFX11-FAKE16-NEXT: .LBB96_2: ; %end
+; GFX11-FAKE16-NEXT: .LBB100_2: ; %end
; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
@@ -206960,7 +230274,7 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; SI-NEXT: buffer_store_dword v38, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; SI-NEXT: s_cbranch_scc0 .LBB97_4
+; SI-NEXT: s_cbranch_scc0 .LBB101_4
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
@@ -207161,8 +230475,8 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; SI-NEXT: v_lshrrev_b32_e32 v13, 16, v52
; SI-NEXT: v_cvt_f32_f16_e32 v13, v13
; SI-NEXT: v_cvt_f32_f16_e32 v43, v43
-; SI-NEXT: s_cbranch_execnz .LBB97_3
-; SI-NEXT: .LBB97_2: ; %cmp.true
+; SI-NEXT: s_cbranch_execnz .LBB101_3
+; SI-NEXT: .LBB101_2: ; %cmp.true
; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: v_and_b32_e32 v1, 0xffff0000, v54
; SI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
@@ -207580,7 +230894,7 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; SI-NEXT: v_cvt_f32_f16_e32 v10, v10
; SI-NEXT: s_waitcnt vmcnt(1) expcnt(0)
; SI-NEXT: v_cvt_f32_f16_e32 v43, v14
-; SI-NEXT: .LBB97_3: ; %end
+; SI-NEXT: .LBB101_3: ; %end
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
@@ -207863,7 +231177,7 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT: s_setpc_b64 s[30:31]
-; SI-NEXT: .LBB97_4:
+; SI-NEXT: .LBB101_4:
; SI-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_mov_b32_e32 v53, v54
@@ -207968,7 +231282,7 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; SI-NEXT: ; implicit-def: $vgpr13
; SI-NEXT: ; kill: killed $vgpr43
; SI-NEXT: ; implicit-def: $vgpr43
-; SI-NEXT: s_branch .LBB97_2
+; SI-NEXT: s_branch .LBB101_2
;
; VI-LABEL: bitcast_v64bf16_to_v64f16_scalar:
; VI: ; %bb.0:
@@ -208000,10 +231314,10 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; VI-NEXT: v_readfirstlane_b32 s31, v1
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 ; 4-byte Folded Spill
-; VI-NEXT: s_cbranch_scc0 .LBB97_3
+; VI-NEXT: s_cbranch_scc0 .LBB101_3
; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: s_cbranch_execnz .LBB97_4
-; VI-NEXT: .LBB97_2: ; %cmp.true
+; VI-NEXT: s_cbranch_execnz .LBB101_4
+; VI-NEXT: .LBB101_2: ; %cmp.true
; VI-NEXT: v_mov_b32_e32 v0, 0x40c00000
; VI-NEXT: s_lshl_b32 s4, s30, 16
; VI-NEXT: v_add_f32_e32 v1, s4, v0
@@ -208581,10 +231895,10 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; VI-NEXT: v_alignbit_b32 v32, v32, v34, 16
; VI-NEXT: v_alignbit_b32 v17, v17, v33, 16
; VI-NEXT: v_alignbit_b32 v16, v16, v18, 16
-; VI-NEXT: s_branch .LBB97_5
-; VI-NEXT: .LBB97_3:
-; VI-NEXT: s_branch .LBB97_2
-; VI-NEXT: .LBB97_4:
+; VI-NEXT: s_branch .LBB101_5
+; VI-NEXT: .LBB101_3:
+; VI-NEXT: s_branch .LBB101_2
+; VI-NEXT: .LBB101_4:
; VI-NEXT: v_mov_b32_e32 v0, s16
; VI-NEXT: v_mov_b32_e32 v1, s17
; VI-NEXT: v_mov_b32_e32 v2, s18
@@ -208601,7 +231915,7 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; VI-NEXT: v_mov_b32_e32 v13, s29
; VI-NEXT: v_mov_b32_e32 v14, s30
; VI-NEXT: v_mov_b32_e32 v15, s31
-; VI-NEXT: .LBB97_5: ; %end
+; VI-NEXT: .LBB101_5: ; %end
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; VI-NEXT: v_mov_b32_e32 v18, v32
@@ -208644,10 +231958,10 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 ; 4-byte Folded Spill
-; GFX9-NEXT: s_cbranch_scc0 .LBB97_3
+; GFX9-NEXT: s_cbranch_scc0 .LBB101_3
; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: s_cbranch_execnz .LBB97_4
-; GFX9-NEXT: .LBB97_2: ; %cmp.true
+; GFX9-NEXT: s_cbranch_execnz .LBB101_4
+; GFX9-NEXT: .LBB101_2: ; %cmp.true
; GFX9-NEXT: v_mov_b32_e32 v0, 0x40c00000
; GFX9-NEXT: s_and_b32 s4, s30, 0xffff0000
; GFX9-NEXT: v_add_f32_e32 v1, s4, v0
@@ -209258,10 +232572,10 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; GFX9-NEXT: v_lshl_or_b32 v32, v35, 16, v32
; GFX9-NEXT: v_lshl_or_b32 v17, v34, 16, v17
; GFX9-NEXT: v_lshl_or_b32 v16, v18, 16, v16
-; GFX9-NEXT: s_branch .LBB97_5
-; GFX9-NEXT: .LBB97_3:
-; GFX9-NEXT: s_branch .LBB97_2
-; GFX9-NEXT: .LBB97_4:
+; GFX9-NEXT: s_branch .LBB101_5
+; GFX9-NEXT: .LBB101_3:
+; GFX9-NEXT: s_branch .LBB101_2
+; GFX9-NEXT: .LBB101_4:
; GFX9-NEXT: v_mov_b32_e32 v0, s16
; GFX9-NEXT: v_mov_b32_e32 v1, s17
; GFX9-NEXT: v_mov_b32_e32 v2, s18
@@ -209278,7 +232592,7 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; GFX9-NEXT: v_mov_b32_e32 v13, s29
; GFX9-NEXT: v_mov_b32_e32 v14, s30
; GFX9-NEXT: v_mov_b32_e32 v15, s31
-; GFX9-NEXT: .LBB97_5: ; %end
+; GFX9-NEXT: .LBB101_5: ; %end
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
@@ -209309,11 +232623,11 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; GFX11-TRUE16-NEXT: s_mov_b32 s12, s0
; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0
; GFX11-TRUE16-NEXT: s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB97_3
+; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB101_3
; GFX11-TRUE16-NEXT: ; %bb.1: ; %Flow
; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB97_4
-; GFX11-TRUE16-NEXT: .LBB97_2: ; %cmp.true
+; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB101_4
+; GFX11-TRUE16-NEXT: .LBB101_2: ; %cmp.true
; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v17
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v16
; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v18
@@ -209973,9 +233287,9 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; GFX11-TRUE16-NEXT: v_lshl_or_b32 v18, v33, 16, v37
; GFX11-TRUE16-NEXT: v_lshl_or_b32 v17, v17, 16, v38
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT: .LBB97_3:
-; GFX11-TRUE16-NEXT: s_branch .LBB97_2
-; GFX11-TRUE16-NEXT: .LBB97_4:
+; GFX11-TRUE16-NEXT: .LBB101_3:
+; GFX11-TRUE16-NEXT: s_branch .LBB101_2
+; GFX11-TRUE16-NEXT: .LBB101_4:
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
@@ -210004,11 +233318,11 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; GFX11-FAKE16-NEXT: s_mov_b32 s12, s0
; GFX11-FAKE16-NEXT: s_mov_b32 s0, 0
; GFX11-FAKE16-NEXT: s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB97_3
+; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB101_3
; GFX11-FAKE16-NEXT: ; %bb.1: ; %Flow
; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB97_4
-; GFX11-FAKE16-NEXT: .LBB97_2: ; %cmp.true
+; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB101_4
+; GFX11-FAKE16-NEXT: .LBB101_2: ; %cmp.true
; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v17
; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v1, 16, v16
; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v18
@@ -210668,9 +233982,9 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v33, 16, v37
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v17, v17, 16, v38
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT: .LBB97_3:
-; GFX11-FAKE16-NEXT: s_branch .LBB97_2
-; GFX11-FAKE16-NEXT: .LBB97_4:
+; GFX11-FAKE16-NEXT: .LBB101_3:
+; GFX11-FAKE16-NEXT: s_branch .LBB101_2
+; GFX11-FAKE16-NEXT: .LBB101_4:
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
@@ -210967,7 +234281,7 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
; SI-NEXT: ; kill: killed $vgpr2
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB98_2
+; SI-NEXT: s_cbranch_execz .LBB102_2
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
@@ -211228,9 +234542,9 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
; SI-NEXT: ; implicit-def: $vgpr60
; SI-NEXT: ; implicit-def: $vgpr61
; SI-NEXT: ; implicit-def: $vgpr62
-; SI-NEXT: .LBB98_2: ; %Flow
+; SI-NEXT: .LBB102_2: ; %Flow
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB98_4
+; SI-NEXT: s_cbranch_execz .LBB102_4
; SI-NEXT: ; %bb.3: ; %cmp.true
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt expcnt(0)
@@ -211624,7 +234938,7 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v4
; SI-NEXT: v_mov_b32_e32 v4, v7
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; SI-NEXT: .LBB98_4: ; %end
+; SI-NEXT: .LBB102_4: ; %end
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
@@ -212002,7 +235316,7 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT: 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 .LBB98_2
+; VI-NEXT: s_cbranch_execz .LBB102_2
; VI-NEXT: ; %bb.1: ; %cmp.true
; VI-NEXT: v_mov_b32_e32 v32, 0x200
; VI-NEXT: v_add_f16_e32 v33, 0x200, v15
@@ -212102,7 +235416,7 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
; 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: .LBB98_2: ; %end
+; VI-NEXT: .LBB102_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]
@@ -212117,7 +235431,7 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT: 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 .LBB98_2
+; GFX9-NEXT: s_cbranch_execz .LBB102_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]
@@ -212153,7 +235467,7 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
; 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: .LBB98_2: ; %end
+; GFX9-NEXT: .LBB102_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]
@@ -212170,7 +235484,7 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
; GFX11-NEXT: s_xor_b32 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 .LBB98_2
+; GFX11-NEXT: s_cbranch_execz .LBB102_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]
@@ -212205,7 +235519,7 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
; 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: .LBB98_2: ; %end
+; GFX11-NEXT: .LBB102_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]
@@ -212387,7 +235701,7 @@ define inreg <64 x bfloat> @bitcast_v64f16_to_v64bf16_scalar(<64 x half> inreg %
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v36, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; SI-NEXT: s_cbranch_scc0 .LBB99_2
+; SI-NEXT: s_cbranch_scc0 .LBB103_2
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
@@ -212537,8 +235851,8 @@ define inreg <64 x bfloat> @bitcast_v64f16_to_v64bf16_scalar(<64 x half> inreg %
; SI-NEXT: v_mov_b32_e32 v44, v18
; SI-NEXT: v_mov_b32_e32 v5, v43
; SI-NEXT: v_mov_b32_e32 v18, v6
-; SI-NEXT: s_branch .LBB99_3
-; SI-NEXT: .LBB99_2:
+; SI-NEXT: s_branch .LBB103_3
+; SI-NEXT: .LBB103_2:
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; SI-NEXT: ; implicit-def: $vgpr5
; SI-NEXT: ; implicit-def: $vgpr3
@@ -212646,7 +235960,7 @@ define inreg <64 x bfloat> @bitcast_v64f16_to_v64bf16_scalar(<64 x half> inreg %
; SI-NEXT: ; implicit-def: $vgpr15
; SI-NEXT: ; implicit-def: $vgpr7
; SI-NEXT: ; implicit-def: $vgpr14
-; SI-NEXT: .LBB99_3: ; %Flow
+; SI-NEXT: .LBB103_3: ; %Flow
; SI-NEXT: s_waitcnt expcnt(1)
; SI-NEXT: v_mov_b32_e32 v36, v2
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
@@ -212655,7 +235969,7 @@ define inreg <64 x bfloat> @bitcast_v64f16_to_v64bf16_scalar(<64 x half> inreg %
; SI-NEXT: s_andn2_b64 vcc, exec, s[4:5]
; SI-NEXT: v_mov_b32_e32 v31, v11
; SI-NEXT: v_mov_b32_e32 v9, v17
-; SI-NEXT: s_cbranch_vccnz .LBB99_5
+; SI-NEXT: s_cbranch_vccnz .LBB103_5
; SI-NEXT: ; %bb.4: ; %cmp.true
; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
@@ -213078,7 +236392,7 @@ define inreg <64 x bfloat> @bitcast_v64f16_to_v64bf16_scalar(<64 x half> inreg %
; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v2
-; SI-NEXT: .LBB99_5: ; %end
+; SI-NEXT: .LBB103_5: ; %end
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt vmcnt(1)
@@ -213423,10 +236737,10 @@ define inreg <64 x bfloat> @bitcast_v64f16_to_v64bf16_scalar(<64 x half> inreg %
; VI-NEXT: v_mov_b32_e32 v11, s27
; VI-NEXT: v_mov_b32_e32 v12, s28
; VI-NEXT: v_mov_b32_e32 v13, s29
-; VI-NEXT: s_cbranch_scc0 .LBB99_4
+; VI-NEXT: s_cbranch_scc0 .LBB103_4
; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: s_cbranch_execnz .LBB99_3
-; VI-NEXT: .LBB99_2: ; %cmp.true
+; VI-NEXT: s_cbranch_execnz .LBB103_3
+; VI-NEXT: .LBB103_2: ; %cmp.true
; VI-NEXT: v_mov_b32_e32 v18, 0x200
; VI-NEXT: v_add_f16_e32 v33, 0x200, v15
; VI-NEXT: v_add_f16_sdwa v15, v15, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
@@ -213524,11 +236838,11 @@ define inreg <64 x bfloat> @bitcast_v64f16_to_v64bf16_scalar(<64 x half> inreg %
; 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, v18
-; VI-NEXT: .LBB99_3: ; %end
+; VI-NEXT: .LBB103_3: ; %end
; VI-NEXT: v_mov_b32_e32 v18, v32
; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB99_4:
-; VI-NEXT: s_branch .LBB99_2
+; VI-NEXT: .LBB103_4:
+; VI-NEXT: s_branch .LBB103_2
;
; GFX9-LABEL: bitcast_v64f16_to_v64bf16_scalar:
; GFX9: ; %bb.0:
@@ -213567,10 +236881,10 @@ define inreg <64 x bfloat> @bitcast_v64f16_to_v64bf16_scalar(<64 x half> inreg %
; GFX9-NEXT: v_mov_b32_e32 v11, s27
; GFX9-NEXT: v_mov_b32_e32 v12, s28
; GFX9-NEXT: v_mov_b32_e32 v13, s29
-; GFX9-NEXT: s_cbranch_scc0 .LBB99_4
+; GFX9-NEXT: s_cbranch_scc0 .LBB103_4
; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: s_cbranch_execnz .LBB99_3
-; GFX9-NEXT: .LBB99_2: ; %cmp.true
+; GFX9-NEXT: s_cbranch_execnz .LBB103_3
+; GFX9-NEXT: .LBB103_2: ; %cmp.true
; GFX9-NEXT: s_movk_i32 s4, 0x200
; GFX9-NEXT: v_pk_add_f16 v15, v15, s4 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v14, v14, s4 op_sel_hi:[1,0]
@@ -213604,11 +236918,11 @@ define inreg <64 x bfloat> @bitcast_v64f16_to_v64bf16_scalar(<64 x half> inreg %
; GFX9-NEXT: v_pk_add_f16 v32, v32, s4 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v17, v17, s4 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v16, v16, s4 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB99_3: ; %end
+; GFX9-NEXT: .LBB103_3: ; %end
; GFX9-NEXT: v_mov_b32_e32 v18, v32
; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB99_4:
-; GFX9-NEXT: s_branch .LBB99_2
+; GFX9-NEXT: .LBB103_4:
+; GFX9-NEXT: s_branch .LBB103_2
;
; GFX11-LABEL: bitcast_v64f16_to_v64bf16_scalar:
; GFX11: ; %bb.0:
@@ -213628,11 +236942,11 @@ define inreg <64 x bfloat> @bitcast_v64f16_to_v64bf16_scalar(<64 x half> inreg %
; GFX11-NEXT: s_mov_b32 s12, s0
; GFX11-NEXT: s_mov_b32 s0, 0
; GFX11-NEXT: s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-NEXT: s_cbranch_scc0 .LBB99_3
+; GFX11-NEXT: s_cbranch_scc0 .LBB103_3
; GFX11-NEXT: ; %bb.1: ; %Flow
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_vccnz .LBB99_4
-; GFX11-NEXT: .LBB99_2: ; %cmp.true
+; GFX11-NEXT: s_cbranch_vccnz .LBB103_4
+; GFX11-NEXT: .LBB103_2: ; %cmp.true
; GFX11-NEXT: v_pk_add_f16 v15, 0x200, s27 op_sel_hi:[0,1]
; GFX11-NEXT: v_pk_add_f16 v14, 0x200, s26 op_sel_hi:[0,1]
; GFX11-NEXT: v_pk_add_f16 v13, 0x200, s25 op_sel_hi:[0,1]
@@ -213666,9 +236980,9 @@ define inreg <64 x bfloat> @bitcast_v64f16_to_v64bf16_scalar(<64 x half> inreg %
; 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: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB99_3:
-; GFX11-NEXT: s_branch .LBB99_2
-; GFX11-NEXT: .LBB99_4:
+; GFX11-NEXT: .LBB103_3:
+; GFX11-NEXT: s_branch .LBB103_2
+; GFX11-NEXT: .LBB103_4:
; GFX11-NEXT: v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-NEXT: v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
@@ -213981,7 +237295,7 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
; SI-NEXT: ; implicit-def: $vgpr26
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB100_2
+; SI-NEXT: s_cbranch_execz .LBB104_2
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: v_lshrrev_b32_e32 v26, 16, v33
; SI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
@@ -214242,9 +237556,9 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v11
; SI-NEXT: ; implicit-def: $vgpr11
-; SI-NEXT: .LBB100_2: ; %Flow
+; SI-NEXT: .LBB104_2: ; %Flow
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB100_4
+; SI-NEXT: s_cbranch_execz .LBB104_4
; SI-NEXT: ; %bb.3: ; %cmp.true
; SI-NEXT: v_and_b32_e32 v27, 0xffff0000, v43
; SI-NEXT: v_and_b32_e32 v26, 0xffff0000, v33
@@ -214578,7 +237892,7 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_alignbit_b32 v1, v27, v3, 16
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; SI-NEXT: .LBB100_4: ; %end
+; SI-NEXT: .LBB104_4: ; %end
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
@@ -214908,7 +238222,7 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT: 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 .LBB100_2
+; VI-NEXT: s_cbranch_execz .LBB104_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
@@ -215494,7 +238808,7 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
; 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: .LBB100_2: ; %end
+; VI-NEXT: .LBB104_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
@@ -215541,7 +238855,7 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT: 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 .LBB100_2
+; GFX9-NEXT: s_cbranch_execz .LBB104_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
@@ -216031,7 +239345,7 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
; 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: .LBB100_2: ; %end
+; GFX9-NEXT: .LBB104_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
@@ -216064,7 +239378,7 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB100_2
+; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB104_2
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v33, 16, v17
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v16
@@ -216630,7 +239944,7 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshl_or_b32 v17, v37, 16, v38
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT: v_lshl_or_b32 v16, v39, 16, v16
-; GFX11-TRUE16-NEXT: .LBB100_2: ; %end
+; GFX11-TRUE16-NEXT: .LBB104_2: ; %end
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
@@ -216647,7 +239961,7 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
-; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB100_2
+; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB104_2
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.true
; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v33, 16, v17
; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v32, 16, v16
@@ -217157,7 +240471,7 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
; GFX11-FAKE16-NEXT: v_cndmask_b32_e32 v11, v85, v96, vcc_lo
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v11, v83, 0x7060302
-; GFX11-FAKE16-NEXT: .LBB100_2: ; %end
+; GFX11-FAKE16-NEXT: .LBB104_2: ; %end
; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
@@ -217343,7 +240657,7 @@ define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a
; SI-NEXT: buffer_store_dword v52, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; SI-NEXT: s_cbranch_scc0 .LBB101_2
+; SI-NEXT: s_cbranch_scc0 .LBB105_2
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
@@ -217487,8 +240801,8 @@ define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a
; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v8
; SI-NEXT: s_waitcnt vmcnt(3)
; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v34
-; SI-NEXT: s_branch .LBB101_3
-; SI-NEXT: .LBB101_2:
+; SI-NEXT: s_branch .LBB105_3
+; SI-NEXT: .LBB105_2:
; SI-NEXT: s_waitcnt expcnt(4)
; SI-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
@@ -217600,10 +240914,10 @@ define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a
; SI-NEXT: ; implicit-def: $vgpr5
; SI-NEXT: ; implicit-def: $vgpr36
; SI-NEXT: ; implicit-def: $vgpr35
-; SI-NEXT: .LBB101_3: ; %Flow
+; SI-NEXT: .LBB105_3: ; %Flow
; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT: s_andn2_b64 vcc, exec, s[4:5]
-; SI-NEXT: s_cbranch_vccnz .LBB101_5
+; SI-NEXT: s_cbranch_vccnz .LBB105_5
; SI-NEXT: ; %bb.4: ; %cmp.true
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
@@ -217961,7 +241275,7 @@ define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a
; SI-NEXT: v_alignbit_b32 v4, v13, v14, 16
; SI-NEXT: v_mov_b32_e32 v14, v51
; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; SI-NEXT: .LBB101_5: ; %end
+; SI-NEXT: .LBB105_5: ; %end
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
@@ -218248,10 +241562,10 @@ define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a
; VI-NEXT: v_readfirstlane_b32 s31, v1
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 ; 4-byte Folded Spill
-; VI-NEXT: s_cbranch_scc0 .LBB101_3
+; VI-NEXT: s_cbranch_scc0 .LBB105_3
; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: s_cbranch_execnz .LBB101_4
-; VI-NEXT: .LBB101_2: ; %cmp.true
+; VI-NEXT: s_cbranch_execnz .LBB105_4
+; VI-NEXT: .LBB105_2: ; %cmp.true
; VI-NEXT: v_mov_b32_e32 v0, 0x40c00000
; VI-NEXT: s_lshl_b32 s4, s30, 16
; VI-NEXT: v_add_f32_e32 v1, s4, v0
@@ -218829,10 +242143,10 @@ define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a
; VI-NEXT: v_alignbit_b32 v32, v32, v34, 16
; VI-NEXT: v_alignbit_b32 v17, v17, v33, 16
; VI-NEXT: v_alignbit_b32 v16, v16, v18, 16
-; VI-NEXT: s_branch .LBB101_5
-; VI-NEXT: .LBB101_3:
-; VI-NEXT: s_branch .LBB101_2
-; VI-NEXT: .LBB101_4:
+; VI-NEXT: s_branch .LBB105_5
+; VI-NEXT: .LBB105_3:
+; VI-NEXT: s_branch .LBB105_2
+; VI-NEXT: .LBB105_4:
; VI-NEXT: v_mov_b32_e32 v0, s16
; VI-NEXT: v_mov_b32_e32 v1, s17
; VI-NEXT: v_mov_b32_e32 v2, s18
@@ -218849,7 +242163,7 @@ define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a
; VI-NEXT: v_mov_b32_e32 v13, s29
; VI-NEXT: v_mov_b32_e32 v14, s30
; VI-NEXT: v_mov_b32_e32 v15, s31
-; VI-NEXT: .LBB101_5: ; %end
+; VI-NEXT: .LBB105_5: ; %end
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; VI-NEXT: v_mov_b32_e32 v18, v32
@@ -218892,10 +242206,10 @@ define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 ; 4-byte Folded Spill
-; GFX9-NEXT: s_cbranch_scc0 .LBB101_3
+; GFX9-NEXT: s_cbranch_scc0 .LBB105_3
; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: s_cbranch_execnz .LBB101_4
-; GFX9-NEXT: .LBB101_2: ; %cmp.true
+; GFX9-NEXT: s_cbranch_execnz .LBB105_4
+; GFX9-NEXT: .LBB105_2: ; %cmp.true
; GFX9-NEXT: v_mov_b32_e32 v0, 0x40c00000
; GFX9-NEXT: s_and_b32 s4, s30, 0xffff0000
; GFX9-NEXT: v_add_f32_e32 v1, s4, v0
@@ -219474,10 +242788,10 @@ define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a
; GFX9-NEXT: v_and_or_b32 v32, v35, v18, v32
; GFX9-NEXT: v_and_or_b32 v17, v34, v18, v17
; GFX9-NEXT: v_and_or_b32 v16, v33, v18, v16
-; GFX9-NEXT: s_branch .LBB101_5
-; GFX9-NEXT: .LBB101_3:
-; GFX9-NEXT: s_branch .LBB101_2
-; GFX9-NEXT: .LBB101_4:
+; GFX9-NEXT: s_branch .LBB105_5
+; GFX9-NEXT: .LBB105_3:
+; GFX9-NEXT: s_branch .LBB105_2
+; GFX9-NEXT: .LBB105_4:
; GFX9-NEXT: v_mov_b32_e32 v0, s16
; GFX9-NEXT: v_mov_b32_e32 v1, s17
; GFX9-NEXT: v_mov_b32_e32 v2, s18
@@ -219494,7 +242808,7 @@ define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a
; GFX9-NEXT: v_mov_b32_e32 v13, s29
; GFX9-NEXT: v_mov_b32_e32 v14, s30
; GFX9-NEXT: v_mov_b32_e32 v15, s31
-; GFX9-NEXT: .LBB101_5: ; %end
+; GFX9-NEXT: .LBB105_5: ; %end
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
@@ -219525,11 +242839,11 @@ define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a
; GFX11-NEXT: s_mov_b32 s12, s0
; GFX11-NEXT: s_mov_b32 s0, 0
; GFX11-NEXT: s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-NEXT: s_cbranch_scc0 .LBB101_3
+; GFX11-NEXT: s_cbranch_scc0 .LBB105_3
; GFX11-NEXT: ; %bb.1: ; %Flow
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_vccnz .LBB101_4
-; GFX11-NEXT: .LBB101_2: ; %cmp.true
+; GFX11-NEXT: s_cbranch_vccnz .LBB105_4
+; GFX11-NEXT: .LBB105_2: ; %cmp.true
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v17
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v16
; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v18
@@ -220130,9 +243444,9 @@ define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a
; GFX11-NEXT: v_and_or_b32 v17, 0xffff0000, v32, v39
; GFX11-NEXT: v_and_or_b32 v16, 0xffff0000, v16, v48
; GFX11-NEXT: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB101_3:
-; GFX11-NEXT: s_branch .LBB101_2
-; GFX11-NEXT: .LBB101_4:
+; GFX11-NEXT: .LBB105_3:
+; GFX11-NEXT: s_branch .LBB105_2
+; GFX11-NEXT: .LBB105_4:
; GFX11-NEXT: v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-NEXT: v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
@@ -220317,7 +243631,7 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
; SI-NEXT: ; implicit-def: $vgpr3
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB102_2
+; SI-NEXT: s_cbranch_execz .LBB106_2
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt vmcnt(8)
@@ -220464,9 +243778,9 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
; SI-NEXT: ; implicit-def: $vgpr63
; SI-NEXT: ; implicit-def: $vgpr40
; SI-NEXT: ; implicit-def: $vgpr62
-; SI-NEXT: .LBB102_2: ; %Flow
+; SI-NEXT: .LBB106_2: ; %Flow
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB102_4
+; SI-NEXT: s_cbranch_execz .LBB106_4
; SI-NEXT: ; %bb.3: ; %cmp.true
; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt vmcnt(9)
@@ -220730,7 +244044,7 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
; SI-NEXT: v_and_b32_e32 v14, 0xffff0000, v16
; SI-NEXT: v_and_b32_e32 v16, 0xffff0000, v11
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; SI-NEXT: .LBB102_4: ; %end
+; SI-NEXT: .LBB106_4: ; %end
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
; SI-NEXT: s_waitcnt vmcnt(7) expcnt(0)
; SI-NEXT: v_mul_f32_e32 v1, 1.0, v2
@@ -221020,7 +244334,7 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT: 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 .LBB102_2
+; VI-NEXT: s_cbranch_execz .LBB106_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
@@ -221120,7 +244434,7 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
; 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: .LBB102_2: ; %end
+; VI-NEXT: .LBB106_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]
@@ -221135,7 +244449,7 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT: 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 .LBB102_2
+; GFX9-NEXT: s_cbranch_execz .LBB106_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]
@@ -221170,7 +244484,7 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
; 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: .LBB102_2: ; %end
+; GFX9-NEXT: .LBB106_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]
@@ -221187,7 +244501,7 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
; GFX11-NEXT: s_xor_b32 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 .LBB102_2
+; GFX11-NEXT: s_cbranch_execz .LBB106_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]
@@ -221222,7 +244536,7 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
; 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: .LBB102_2: ; %end
+; GFX11-NEXT: .LBB106_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]
@@ -221415,7 +244729,7 @@ define inreg <64 x bfloat> @bitcast_v64i16_to_v64bf16_scalar(<64 x i16> inreg %a
; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: v_readfirstlane_b32 s41, v37
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
-; SI-NEXT: s_cbranch_scc0 .LBB103_2
+; SI-NEXT: s_cbranch_scc0 .LBB107_2
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: s_lshl_b32 s4, s60, 16
; SI-NEXT: v_writelane_b32 v41, s4, 25
@@ -221498,8 +244812,8 @@ define inreg <64 x bfloat> @bitcast_v64i16_to_v64bf16_scalar(<64 x i16> inreg %a
; SI-NEXT: s_mov_b32 s6, s62
; SI-NEXT: s_lshl_b32 s26, s62, 16
; SI-NEXT: s_mov_b64 s[4:5], 0
-; SI-NEXT: s_branch .LBB103_3
-; SI-NEXT: .LBB103_2:
+; SI-NEXT: s_branch .LBB107_3
+; SI-NEXT: .LBB107_2:
; SI-NEXT: ; implicit-def: $sgpr17
; SI-NEXT: ; kill: killed $sgpr17
; SI-NEXT: s_mov_b32 s16, s68
@@ -221581,12 +244895,12 @@ define inreg <64 x bfloat> @bitcast_v64i16_to_v64bf16_scalar(<64 x i16> inreg %a
; SI-NEXT: ; implicit-def: $sgpr17
; SI-NEXT: ; kill: killed $sgpr17
; SI-NEXT: ; implicit-def: $sgpr17
-; SI-NEXT: .LBB103_3: ; %Flow
+; SI-NEXT: .LBB107_3: ; %Flow
; SI-NEXT: s_andn2_b64 vcc, exec, s[4:5]
; SI-NEXT: s_mov_b32 s5, s17
; SI-NEXT: s_mov_b32 s17, s86
; SI-NEXT: s_mov_b32 s86, s7
-; SI-NEXT: s_cbranch_vccnz .LBB103_5
+; SI-NEXT: s_cbranch_vccnz .LBB107_5
; SI-NEXT: ; %bb.4: ; %cmp.true
; SI-NEXT: s_lshl_b32 s5, s6, 16
; SI-NEXT: v_readlane_b32 s6, v41, 24
@@ -221849,7 +245163,7 @@ define inreg <64 x bfloat> @bitcast_v64i16_to_v64bf16_scalar(<64 x i16> inreg %a
; SI-NEXT: s_and_b32 s26, s4, 0xffff0000
; SI-NEXT: s_lshl_b32 s5, s4, 16
; SI-NEXT: v_writelane_b32 v41, s6, 36
-; SI-NEXT: .LBB103_5: ; %end
+; SI-NEXT: .LBB107_5: ; %end
; SI-NEXT: v_mul_f32_e64 v1, 1.0, s86
; SI-NEXT: v_readlane_b32 s4, v41, 25
; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
@@ -222161,10 +245475,10 @@ define inreg <64 x bfloat> @bitcast_v64i16_to_v64bf16_scalar(<64 x i16> inreg %a
; VI-NEXT: s_and_b64 s[4:5], vcc, exec
; VI-NEXT: v_readfirstlane_b32 s7, v1
; VI-NEXT: v_writelane_b32 v32, s39, 7
-; VI-NEXT: s_cbranch_scc0 .LBB103_4
+; VI-NEXT: s_cbranch_scc0 .LBB107_4
; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: s_cbranch_execnz .LBB103_3
-; VI-NEXT: .LBB103_2: ; %cmp.true
+; VI-NEXT: s_cbranch_execnz .LBB107_3
+; VI-NEXT: .LBB107_2: ; %cmp.true
; VI-NEXT: s_and_b32 s4, s47, 0xffff0000
; VI-NEXT: s_add_i32 s5, s47, 3
; VI-NEXT: s_and_b32 s47, s46, 0xffff0000
@@ -222325,7 +245639,7 @@ define inreg <64 x bfloat> @bitcast_v64i16_to_v64bf16_scalar(<64 x i16> inreg %a
; VI-NEXT: s_add_i32 s45, s45, 0x30000
; VI-NEXT: s_add_i32 s46, s46, 0x30000
; VI-NEXT: s_add_i32 s47, s4, 0x30000
-; VI-NEXT: .LBB103_3: ; %end
+; VI-NEXT: .LBB107_3: ; %end
; VI-NEXT: v_mov_b32_e32 v0, s16
; VI-NEXT: v_mov_b32_e32 v1, s17
; VI-NEXT: v_mov_b32_e32 v2, s18
@@ -222371,8 +245685,8 @@ define inreg <64 x bfloat> @bitcast_v64i16_to_v64bf16_scalar(<64 x i16> inreg %a
; VI-NEXT: s_mov_b64 exec, s[4:5]
; VI-NEXT: s_waitcnt vmcnt(0)
; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB103_4:
-; VI-NEXT: s_branch .LBB103_2
+; VI-NEXT: .LBB107_4:
+; VI-NEXT: s_branch .LBB107_2
;
; GFX9-LABEL: bitcast_v64i16_to_v64bf16_scalar:
; GFX9: ; %bb.0:
@@ -222411,10 +245725,10 @@ define inreg <64 x bfloat> @bitcast_v64i16_to_v64bf16_scalar(<64 x i16> inreg %a
; GFX9-NEXT: v_mov_b32_e32 v11, s27
; GFX9-NEXT: v_mov_b32_e32 v12, s28
; GFX9-NEXT: v_mov_b32_e32 v13, s29
-; GFX9-NEXT: s_cbranch_scc0 .LBB103_4
+; GFX9-NEXT: s_cbranch_scc0 .LBB107_4
; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: s_cbranch_execnz .LBB103_3
-; GFX9-NEXT: .LBB103_2: ; %cmp.true
+; GFX9-NEXT: s_cbranch_execnz .LBB107_3
+; GFX9-NEXT: .LBB107_2: ; %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]
@@ -222447,11 +245761,11 @@ define inreg <64 x bfloat> @bitcast_v64i16_to_v64bf16_scalar(<64 x i16> inreg %a
; GFX9-NEXT: v_pk_add_u16 v32, v32, 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: .LBB103_3: ; %end
+; GFX9-NEXT: .LBB107_3: ; %end
; GFX9-NEXT: v_mov_b32_e32 v18, v32
; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB103_4:
-; GFX9-NEXT: s_branch .LBB103_2
+; GFX9-NEXT: .LBB107_4:
+; GFX9-NEXT: s_branch .LBB107_2
;
; GFX11-LABEL: bitcast_v64i16_to_v64bf16_scalar:
; GFX11: ; %bb.0:
@@ -222471,11 +245785,11 @@ define inreg <64 x bfloat> @bitcast_v64i16_to_v64bf16_scalar(<64 x i16> inreg %a
; GFX11-NEXT: s_mov_b32 s12, s0
; GFX11-NEXT: s_mov_b32 s0, 0
; GFX11-NEXT: s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-NEXT: s_cbranch_scc0 .LBB103_3
+; GFX11-NEXT: s_cbranch_scc0 .LBB107_3
; GFX11-NEXT: ; %bb.1: ; %Flow
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_vccnz .LBB103_4
-; GFX11-NEXT: .LBB103_2: ; %cmp.true
+; GFX11-NEXT: s_cbranch_vccnz .LBB107_4
+; GFX11-NEXT: .LBB107_2: ; %cmp.true
; GFX11-NEXT: v_pk_add_u16 v15, s27, 3 op_sel_hi:[1,0]
; GFX11-NEXT: v_pk_add_u16 v14, s26, 3 op_sel_hi:[1,0]
; GFX11-NEXT: v_pk_add_u16 v13, s25, 3 op_sel_hi:[1,0]
@@ -222509,9 +245823,9 @@ define inreg <64 x bfloat> @bitcast_v64i16_to_v64bf16_scalar(<64 x i16> inreg %a
; 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: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB103_3:
-; GFX11-NEXT: s_branch .LBB103_2
-; GFX11-NEXT: .LBB103_4:
+; GFX11-NEXT: .LBB107_3:
+; GFX11-NEXT: s_branch .LBB107_2
+; GFX11-NEXT: .LBB107_4:
; GFX11-NEXT: v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-NEXT: v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
@@ -222703,7 +246017,7 @@ define <64 x i16> @bitcast_v64f16_to_v64i16(<64 x half> %a, i32 %b) {
; SI-NEXT: v_mov_b32_e32 v62, v3
; SI-NEXT: v_mov_b32_e32 v63, v4
; SI-NEXT: s_xor_b64 exec, exec, s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB104_2
+; SI-NEXT: s_cbranch_execz .LBB108_2
; SI-NEXT: ; %bb.1: ; %cmp.true
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT: v_cvt_f32_f16_e32 v10, v10
@@ -223007,7 +246321,7 @@ define <64 x i16> @bitcast_v64f16_to_v64i16(<64 x half> %a, i32 %b) {
; SI-NEXT: v_alignbit_b32 v2, v21, v27, 16
; SI-NEXT: v_alignbit_b32 v35, v33, v35, 16
; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; SI-NEXT: .LBB104_2: ; %end
+; SI-NEXT: .LBB108_2: ; %end
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
@@ -223244,7 +246558,7 @@ define <64 x i16> @bitcast_v64f16_to_v64i16(<64 x half> %a, i32 %b) {
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT: 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 .LBB104_2
+; VI-NEXT: s_cbranch_execz .LBB108_2
; VI-NEXT: ; %bb.1: ; %cmp.true
; VI-NEXT: v_mov_b32_e32 v32, 0x200
; VI-NEXT: v_add_f16_e32 v33, 0x200, v15
@@ -223344,7 +246658,7 @@ define <64 x i16> @bitcast_v64f16_to_v64i16(<64 x half> %a, i32 %b) {
; 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: .LBB104_2: ; %end
+; VI-NEXT: .LBB108_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]
@@ -223359,7 +246673,7 @@ define <64 x i16> @bitcast_v64f16_to_v64i16(<64 x half> %a, i32 %b) {
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT: 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 .LBB104_2
+; GFX9-NEXT: s_cbranch_execz .LBB108_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]
@@ -223395,7 +246709,7 @@ define <64 x i16> @bitcast_v64f16_to_v64i16(<64 x half> %a, i32 %b) {
; 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: .LBB104_2: ; %end
+; GFX9-NEXT: .LBB108_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]
@@ -223412,7 +246726,7 @@ define <64 x i16> @bitcast_v64f16_to_v64i16(<64 x half> %a, i32 %b) {
; GFX11-NEXT: s_xor_b32 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 .LBB104_2
+; GFX11-NEXT: s_cbranch_execz .LBB108_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]
@@ -223447,7 +246761,7 @@ define <64 x i16> @bitcast_v64f16_to_v64i16(<64 x half> %a, i32 %b) {
; 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: .LBB104_2: ; %end
+; GFX11-NEXT: .LBB108_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]
@@ -223604,10 +246918,10 @@ define inreg <64 x i16> @bitcast_v64f16_to_v64i16_scalar(<64 x half> inreg %a, i
; SI-NEXT: v_cvt_f16_f32_e32 v45, s28
; SI-NEXT: v_cvt_f16_f32_e32 v61, s29
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
-; SI-NEXT: s_cbranch_scc0 .LBB105_4
+; SI-NEXT: s_cbranch_scc0 .LBB109_4
; SI-NEXT: ; %bb.1: ; %cmp.false
-; SI-NEXT: s_cbranch_execnz .LBB105_3
-; SI-NEXT: .LBB105_2: ; %cmp.true
+; SI-NEXT: s_cbranch_execnz .LBB109_3
+; SI-NEXT: .LBB109_2: ; %cmp.true
; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
@@ -223914,7 +247228,7 @@ define inreg <64 x i16> @bitcast_v64f16_to_v64i16_scalar(<64 x half> inreg %a, i
; SI-NEXT: v_alignbit_b32 v24, v5, v24, 16
; SI-NEXT: s_waitcnt vmcnt(1)
; SI-NEXT: v_alignbit_b32 v37, v1, v37, 16
-; SI-NEXT: .LBB105_3: ; %end
+; SI-NEXT: .LBB109_3: ; %end
; SI-NEXT: v_and_b32_e32 v48, 0xffff, v60
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT: v_or_b32_e32 v4, v48, v4
@@ -224136,8 +247450,8 @@ define inreg <64 x i16> @bitcast_v64f16_to_v64i16_scalar(<64 x half> inreg %a, i
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT: s_setpc_b64 s[30:31]
-; SI-NEXT: .LBB105_4:
-; SI-NEXT: s_branch .LBB105_2
+; SI-NEXT: .LBB109_4:
+; SI-NEXT: s_branch .LBB109_2
;
; VI-LABEL: bitcast_v64f16_to_v64i16_scalar:
; VI: ; %bb.0:
@@ -224176,10 +247490,10 @@ define inreg <64 x i16> @bitcast_v64f16_to_v64i16_scalar(<64 x half> inreg %a, i
; VI-NEXT: v_mov_b32_e32 v11, s27
; VI-NEXT: v_mov_b32_e32 v12, s28
; VI-NEXT: v_mov_b32_e32 v13, s29
-; VI-NEXT: s_cbranch_scc0 .LBB105_4
+; VI-NEXT: s_cbranch_scc0 .LBB109_4
; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: s_cbranch_execnz .LBB105_3
-; VI-NEXT: .LBB105_2: ; %cmp.true
+; VI-NEXT: s_cbranch_execnz .LBB109_3
+; VI-NEXT: .LBB109_2: ; %cmp.true
; VI-NEXT: v_mov_b32_e32 v18, 0x200
; VI-NEXT: v_add_f16_e32 v33, 0x200, v15
; VI-NEXT: v_add_f16_sdwa v15, v15, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
@@ -224277,11 +247591,11 @@ define inreg <64 x i16> @bitcast_v64f16_to_v64i16_scalar(<64 x half> inreg %a, i
; 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, v18
-; VI-NEXT: .LBB105_3: ; %end
+; VI-NEXT: .LBB109_3: ; %end
; VI-NEXT: v_mov_b32_e32 v18, v32
; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB105_4:
-; VI-NEXT: s_branch .LBB105_2
+; VI-NEXT: .LBB109_4:
+; VI-NEXT: s_branch .LBB109_2
;
; GFX9-LABEL: bitcast_v64f16_to_v64i16_scalar:
; GFX9: ; %bb.0:
@@ -224320,10 +247634,10 @@ define inreg <64 x i16> @bitcast_v64f16_to_v64i16_scalar(<64 x half> inreg %a, i
; GFX9-NEXT: v_mov_b32_e32 v11, s27
; GFX9-NEXT: v_mov_b32_e32 v12, s28
; GFX9-NEXT: v_mov_b32_e32 v13, s29
-; GFX9-NEXT: s_cbranch_scc0 .LBB105_4
+; GFX9-NEXT: s_cbranch_scc0 .LBB109_4
; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: s_cbranch_execnz .LBB105_3
-; GFX9-NEXT: .LBB105_2: ; %cmp.true
+; GFX9-NEXT: s_cbranch_execnz .LBB109_3
+; GFX9-NEXT: .LBB109_2: ; %cmp.true
; GFX9-NEXT: s_movk_i32 s4, 0x200
; GFX9-NEXT: v_pk_add_f16 v15, v15, s4 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v14, v14, s4 op_sel_hi:[1,0]
@@ -224357,11 +247671,11 @@ define inreg <64 x i16> @bitcast_v64f16_to_v64i16_scalar(<64 x half> inreg %a, i
; GFX9-NEXT: v_pk_add_f16 v32, v32, s4 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v17, v17, s4 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v16, v16, s4 op_sel_hi:[1,0]
-; GFX9-NEXT: .LBB105_3: ; %end
+; GFX9-NEXT: .LBB109_3: ; %end
; GFX9-NEXT: v_mov_b32_e32 v18, v32
; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB105_4:
-; GFX9-NEXT: s_branch .LBB105_2
+; GFX9-NEXT: .LBB109_4:
+; GFX9-NEXT: s_branch .LBB109_2
;
; GFX11-LABEL: bitcast_v64f16_to_v64i16_scalar:
; GFX11: ; %bb.0:
@@ -224381,11 +247695,11 @@ define inreg <64 x i16> @bitcast_v64f16_to_v64i16_scalar(<64 x half> inreg %a, i
; GFX11-NEXT: s_mov_b32 s12, s0
; GFX11-NEXT: s_mov_b32 s0, 0
; GFX11-NEXT: s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-NEXT: s_cbranch_scc0 .LBB105_3
+; GFX11-NEXT: s_cbranch_scc0 .LBB109_3
; GFX11-NEXT: ; %bb.1: ; %Flow
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_vccnz .LBB105_4
-; GFX11-NEXT: .LBB105_2: ; %cmp.true
+; GFX11-NEXT: s_cbranch_vccnz .LBB109_4
+; GFX11-NEXT: .LBB109_2: ; %cmp.true
; GFX11-NEXT: v_pk_add_f16 v15, 0x200, s27 op_sel_hi:[0,1]
; GFX11-NEXT: v_pk_add_f16 v14, 0x200, s26 op_sel_hi:[0,1]
; GFX11-NEXT: v_pk_add_f16 v13, 0x200, s25 op_sel_hi:[0,1]
@@ -224419,9 +247733,9 @@ define inreg <64 x i16> @bitcast_v64f16_to_v64i16_scalar(<64 x half> inreg %a, i
; 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: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB105_3:
-; GFX11-NEXT: s_branch .LBB105_2
-; GFX11-NEXT: .LBB105_4:
+; GFX11-NEXT: .LBB109_3:
+; GFX11-NEXT: s_branch .LBB109_2
+; GFX11-NEXT: .LBB109_4:
; GFX11-NEXT: v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-NEXT: v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
@@ -224639,7 +247953,7 @@ define <64 x half> @bitcast_v64i16_to_v64f16(<64 x i16> %a, i32 %b) {
; SI-NEXT: ; kill: killed $vgpr1
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB106_2
+; SI-NEXT: s_cbranch_execz .LBB110_2
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
@@ -224903,9 +248217,9 @@ define <64 x half> @bitcast_v64i16_to_v64f16(<64 x i16> %a, i32 %b) {
; SI-NEXT: ; kill: killed $vgpr1
; SI-NEXT: ; implicit-def: $vgpr1
; SI-NEXT: ; kill: killed $vgpr1
-; SI-NEXT: .LBB106_2: ; %Flow
+; SI-NEXT: .LBB110_2: ; %Flow
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
-; SI-NEXT: s_cbranch_execz .LBB106_4
+; SI-NEXT: s_cbranch_execz .LBB110_4
; SI-NEXT: ; %bb.3: ; %cmp.true
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v43
@@ -225167,7 +248481,7 @@ define <64 x half> @bitcast_v64i16_to_v64f16(<64 x i16> %a, i32 %b) {
; SI-NEXT: s_waitcnt vmcnt(14)
; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; SI-NEXT: .LBB106_4: ; %end
+; SI-NEXT: .LBB110_4: ; %end
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
@@ -225549,7 +248863,7 @@ define <64 x half> @bitcast_v64i16_to_v64f16(<64 x i16> %a, i32 %b) {
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT: 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 .LBB106_2
+; VI-NEXT: s_cbranch_execz .LBB110_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
@@ -225649,7 +248963,7 @@ define <64 x half> @bitcast_v64i16_to_v64f16(<64 x i16> %a, i32 %b) {
; 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: .LBB106_2: ; %end
+; VI-NEXT: .LBB110_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]
@@ -225664,7 +248978,7 @@ define <64 x half> @bitcast_v64i16_to_v64f16(<64 x i16> %a, i32 %b) {
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT: 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 .LBB106_2
+; GFX9-NEXT: s_cbranch_execz .LBB110_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]
@@ -225699,7 +249013,7 @@ define <64 x half> @bitcast_v64i16_to_v64f16(<64 x i16> %a, i32 %b) {
; 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: .LBB106_2: ; %end
+; GFX9-NEXT: .LBB110_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]
@@ -225716,7 +249030,7 @@ define <64 x half> @bitcast_v64i16_to_v64f16(<64 x i16> %a, i32 %b) {
; GFX11-NEXT: s_xor_b32 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 .LBB106_2
+; GFX11-NEXT: s_cbranch_execz .LBB110_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]
@@ -225751,7 +249065,7 @@ define <64 x half> @bitcast_v64i16_to_v64f16(<64 x i16> %a, i32 %b) {
; 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: .LBB106_2: ; %end
+; GFX11-NEXT: .LBB110_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]
@@ -225845,7 +249159,7 @@ define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i
; SI-NEXT: s_waitcnt vmcnt(14)
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v41
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
-; SI-NEXT: s_cbranch_scc0 .LBB107_2
+; SI-NEXT: s_cbranch_scc0 .LBB111_2
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: v_cvt_f32_f16_e32 v4, v1
; SI-NEXT: v_cvt_f32_f16_e32 v41, s16
@@ -225980,8 +249294,8 @@ define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i
; SI-NEXT: v_cvt_f32_f16_e32 v4, v49
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT: v_cvt_f32_f16_e32 v43, v33
-; SI-NEXT: s_branch .LBB107_3
-; SI-NEXT: .LBB107_2:
+; SI-NEXT: s_branch .LBB111_3
+; SI-NEXT: .LBB111_2:
; SI-NEXT: ; implicit-def: $vgpr4
; SI-NEXT: ; kill: killed $vgpr4
; SI-NEXT: ; implicit-def: $vgpr41
@@ -226082,7 +249396,7 @@ define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i
; SI-NEXT: ; implicit-def: $vgpr4
; SI-NEXT: ; kill: killed $vgpr4
; SI-NEXT: ; implicit-def: $vgpr4
-; SI-NEXT: .LBB107_3: ; %Flow
+; SI-NEXT: .LBB111_3: ; %Flow
; SI-NEXT: s_andn2_b64 vcc, exec, s[4:5]
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_mov_b32_e32 v41, v42
@@ -226099,7 +249413,7 @@ define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i
; SI-NEXT: v_mov_b32_e32 v15, v17
; SI-NEXT: v_mov_b32_e32 v17, v19
; SI-NEXT: v_mov_b32_e32 v19, v1
-; SI-NEXT: s_cbranch_vccnz .LBB107_5
+; SI-NEXT: s_cbranch_vccnz .LBB111_5
; SI-NEXT: ; %bb.4: ; %cmp.true
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
@@ -226341,7 +249655,7 @@ define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f32_f16_e32 v1, v34
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; SI-NEXT: .LBB107_5: ; %end
+; SI-NEXT: .LBB111_5: ; %end
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
@@ -226680,10 +249994,10 @@ define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i
; VI-NEXT: s_and_b64 s[4:5], vcc, exec
; VI-NEXT: v_readfirstlane_b32 s7, v1
; VI-NEXT: v_writelane_b32 v32, s39, 7
-; VI-NEXT: s_cbranch_scc0 .LBB107_4
+; VI-NEXT: s_cbranch_scc0 .LBB111_4
; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: s_cbranch_execnz .LBB107_3
-; VI-NEXT: .LBB107_2: ; %cmp.true
+; VI-NEXT: s_cbranch_execnz .LBB111_3
+; VI-NEXT: .LBB111_2: ; %cmp.true
; VI-NEXT: s_and_b32 s4, s47, 0xffff0000
; VI-NEXT: s_add_i32 s5, s47, 3
; VI-NEXT: s_and_b32 s47, s46, 0xffff0000
@@ -226844,7 +250158,7 @@ define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i
; VI-NEXT: s_add_i32 s45, s45, 0x30000
; VI-NEXT: s_add_i32 s46, s46, 0x30000
; VI-NEXT: s_add_i32 s47, s4, 0x30000
-; VI-NEXT: .LBB107_3: ; %end
+; VI-NEXT: .LBB111_3: ; %end
; VI-NEXT: v_mov_b32_e32 v0, s16
; VI-NEXT: v_mov_b32_e32 v1, s17
; VI-NEXT: v_mov_b32_e32 v2, s18
@@ -226890,8 +250204,8 @@ define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i
; VI-NEXT: s_mov_b64 exec, s[4:5]
; VI-NEXT: s_waitcnt vmcnt(0)
; VI-NEXT: s_setpc_b64 s[30:31]
-; VI-NEXT: .LBB107_4:
-; VI-NEXT: s_branch .LBB107_2
+; VI-NEXT: .LBB111_4:
+; VI-NEXT: s_branch .LBB111_2
;
; GFX9-LABEL: bitcast_v64i16_to_v64f16_scalar:
; GFX9: ; %bb.0:
@@ -226930,10 +250244,10 @@ define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i
; GFX9-NEXT: v_mov_b32_e32 v11, s27
; GFX9-NEXT: v_mov_b32_e32 v12, s28
; GFX9-NEXT: v_mov_b32_e32 v13, s29
-; GFX9-NEXT: s_cbranch_scc0 .LBB107_4
+; GFX9-NEXT: s_cbranch_scc0 .LBB111_4
; GFX9-NEXT: ; %bb.1: ; %cmp.false
-; GFX9-NEXT: s_cbranch_execnz .LBB107_3
-; GFX9-NEXT: .LBB107_2: ; %cmp.true
+; GFX9-NEXT: s_cbranch_execnz .LBB111_3
+; GFX9-NEXT: .LBB111_2: ; %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]
@@ -226966,11 +250280,11 @@ define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i
; GFX9-NEXT: v_pk_add_u16 v32, v32, 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: .LBB107_3: ; %end
+; GFX9-NEXT: .LBB111_3: ; %end
; GFX9-NEXT: v_mov_b32_e32 v18, v32
; GFX9-NEXT: s_setpc_b64 s[30:31]
-; GFX9-NEXT: .LBB107_4:
-; GFX9-NEXT: s_branch .LBB107_2
+; GFX9-NEXT: .LBB111_4:
+; GFX9-NEXT: s_branch .LBB111_2
;
; GFX11-LABEL: bitcast_v64i16_to_v64f16_scalar:
; GFX11: ; %bb.0:
@@ -226990,11 +250304,11 @@ define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i
; GFX11-NEXT: s_mov_b32 s12, s0
; GFX11-NEXT: s_mov_b32 s0, 0
; GFX11-NEXT: s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-NEXT: s_cbranch_scc0 .LBB107_3
+; GFX11-NEXT: s_cbranch_scc0 .LBB111_3
; GFX11-NEXT: ; %bb.1: ; %Flow
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_vccnz .LBB107_4
-; GFX11-NEXT: .LBB107_2: ; %cmp.true
+; GFX11-NEXT: s_cbranch_vccnz .LBB111_4
+; GFX11-NEXT: .LBB111_2: ; %cmp.true
; GFX11-NEXT: v_pk_add_u16 v15, s27, 3 op_sel_hi:[1,0]
; GFX11-NEXT: v_pk_add_u16 v14, s26, 3 op_sel_hi:[1,0]
; GFX11-NEXT: v_pk_add_u16 v13, s25, 3 op_sel_hi:[1,0]
@@ -227028,9 +250342,9 @@ define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i
; 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: s_setpc_b64 s[30:31]
-; GFX11-NEXT: .LBB107_3:
-; GFX11-NEXT: s_branch .LBB107_2
-; GFX11-NEXT: .LBB107_4:
+; GFX11-NEXT: .LBB111_3:
+; GFX11-NEXT: s_branch .LBB111_2
+; GFX11-NEXT: .LBB111_4:
; GFX11-NEXT: v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-NEXT: v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
More information about the llvm-commits
mailing list