[llvm] [AMDGPU][ISel] Add more trunc store actions regarding bf16 (PR #90493)
Shilei Tian via llvm-commits
llvm-commits at lists.llvm.org
Mon Apr 29 09:31:27 PDT 2024
https://github.com/shiltian created https://github.com/llvm/llvm-project/pull/90493
None
>From ccc29ebeb16df8d4554bb3355bfb3294545015df Mon Sep 17 00:00:00 2001
From: Shilei Tian <i at tianshilei.me>
Date: Mon, 29 Apr 2024 12:31:09 -0400
Subject: [PATCH] [AMDGPU][ISel] Add more trunc store actions regarding bf16
---
llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp | 6 +
...bf16.ll => fp_trunc_store_fp32_to_bf16.ll} | 34 +
.../AMDGPU/fp_trunc_store_fp64_to_bf16.ll | 632 ++++++++++++++++++
3 files changed, 672 insertions(+)
rename llvm/test/CodeGen/AMDGPU/{fp_trunc_store_bf16.ll => fp_trunc_store_fp32_to_bf16.ll} (93%)
create mode 100644 llvm/test/CodeGen/AMDGPU/fp_trunc_store_fp64_to_bf16.ll
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index 7f4a2437f62eb4..3124fb23fb0be7 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -315,6 +315,7 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM,
setTruncStoreAction(MVT::f32, MVT::f16, Expand);
setTruncStoreAction(MVT::v2f32, MVT::v2bf16, Expand);
setTruncStoreAction(MVT::v2f32, MVT::v2f16, Expand);
+ setTruncStoreAction(MVT::v3f32, MVT::v3bf16, Expand);
setTruncStoreAction(MVT::v3f32, MVT::v3f16, Expand);
setTruncStoreAction(MVT::v4f32, MVT::v4bf16, Expand);
setTruncStoreAction(MVT::v4f32, MVT::v4f16, Expand);
@@ -330,6 +331,7 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM,
setTruncStoreAction(MVT::f64, MVT::f32, Expand);
setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
+ setTruncStoreAction(MVT::v2f64, MVT::v2bf16, Expand);
setTruncStoreAction(MVT::v2f64, MVT::v2f16, Expand);
setTruncStoreAction(MVT::v3i32, MVT::v3i8, Expand);
@@ -339,17 +341,21 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM,
setTruncStoreAction(MVT::v3i64, MVT::v3i8, Expand);
setTruncStoreAction(MVT::v3i64, MVT::v3i1, Expand);
setTruncStoreAction(MVT::v3f64, MVT::v3f32, Expand);
+ setTruncStoreAction(MVT::v3f64, MVT::v3bf16, Expand);
setTruncStoreAction(MVT::v3f64, MVT::v3f16, Expand);
setTruncStoreAction(MVT::v4i64, MVT::v4i32, Expand);
setTruncStoreAction(MVT::v4i64, MVT::v4i16, Expand);
setTruncStoreAction(MVT::v4f64, MVT::v4f32, Expand);
+ setTruncStoreAction(MVT::v4f64, MVT::v4bf16, Expand);
setTruncStoreAction(MVT::v4f64, MVT::v4f16, Expand);
setTruncStoreAction(MVT::v8f64, MVT::v8f32, Expand);
+ setTruncStoreAction(MVT::v8f64, MVT::v8bf16, Expand);
setTruncStoreAction(MVT::v8f64, MVT::v8f16, Expand);
setTruncStoreAction(MVT::v16f64, MVT::v16f32, Expand);
+ setTruncStoreAction(MVT::v16f64, MVT::v16bf16, Expand);
setTruncStoreAction(MVT::v16f64, MVT::v16f16, Expand);
setTruncStoreAction(MVT::v16i64, MVT::v16i16, Expand);
setTruncStoreAction(MVT::v16i64, MVT::v16i16, Expand);
diff --git a/llvm/test/CodeGen/AMDGPU/fp_trunc_store_bf16.ll b/llvm/test/CodeGen/AMDGPU/fp_trunc_store_fp32_to_bf16.ll
similarity index 93%
rename from llvm/test/CodeGen/AMDGPU/fp_trunc_store_bf16.ll
rename to llvm/test/CodeGen/AMDGPU/fp_trunc_store_fp32_to_bf16.ll
index 5aaff773689f96..87a5d34b90fd76 100644
--- a/llvm/test/CodeGen/AMDGPU/fp_trunc_store_bf16.ll
+++ b/llvm/test/CodeGen/AMDGPU/fp_trunc_store_fp32_to_bf16.ll
@@ -27,6 +27,40 @@ entry:
ret void
}
+define void @v3(<3 x float> %num, ptr addrspace(1) %p) {
+; CHECK-LABEL: v3:
+; CHECK: ; %bb.0: ; %entry
+; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT: v_mov_b32_e32 v5, v4
+; CHECK-NEXT: v_mov_b32_e32 v4, v3
+; CHECK-NEXT: v_bfe_u32 v3, v0, 16, 1
+; CHECK-NEXT: s_movk_i32 s4, 0x7fff
+; CHECK-NEXT: v_add3_u32 v3, v3, v0, s4
+; CHECK-NEXT: v_or_b32_e32 v6, 0x400000, v0
+; CHECK-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; CHECK-NEXT: v_cndmask_b32_e32 v0, v3, v6, vcc
+; CHECK-NEXT: v_bfe_u32 v3, v1, 16, 1
+; CHECK-NEXT: v_add3_u32 v3, v3, v1, s4
+; CHECK-NEXT: v_or_b32_e32 v6, 0x400000, v1
+; CHECK-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; CHECK-NEXT: v_cndmask_b32_e32 v1, v3, v6, vcc
+; CHECK-NEXT: s_mov_b32 s5, 0x7060302
+; CHECK-NEXT: v_perm_b32 v0, v1, v0, s5
+; CHECK-NEXT: v_bfe_u32 v1, v2, 16, 1
+; CHECK-NEXT: v_add3_u32 v1, v1, v2, s4
+; CHECK-NEXT: v_or_b32_e32 v3, 0x400000, v2
+; CHECK-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; CHECK-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc
+; CHECK-NEXT: global_store_short_d16_hi v[4:5], v1, off offset:4
+; CHECK-NEXT: global_store_dword v[4:5], v0, off
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: s_setpc_b64 s[30:31]
+entry:
+ %conv = fptrunc <3 x float> %num to <3 x bfloat>
+ store <3 x bfloat> %conv, ptr addrspace(1) %p, align 8
+ ret void
+}
+
define void @v4(<4 x float> %num, ptr addrspace(1) %p) {
; CHECK-LABEL: v4:
; CHECK: ; %bb.0: ; %entry
diff --git a/llvm/test/CodeGen/AMDGPU/fp_trunc_store_fp64_to_bf16.ll b/llvm/test/CodeGen/AMDGPU/fp_trunc_store_fp64_to_bf16.ll
new file mode 100644
index 00000000000000..01e1dfc0ba0c31
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/fp_trunc_store_fp64_to_bf16.ll
@@ -0,0 +1,632 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 4
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a %s -o - | FileCheck %s
+
+define void @v2(<2 x double> %num, ptr addrspace(1) %p) {
+; CHECK-LABEL: v2:
+; CHECK: ; %bb.0: ; %entry
+; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT: v_cvt_f32_f64_e64 v8, |v[0:1]|
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[6:7], v8
+; CHECK-NEXT: v_and_b32_e32 v9, 1, v8
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[6:7], |v[0:1]|, v[6:7]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[4:5], |v[0:1]|, v[6:7]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v9
+; CHECK-NEXT: v_cndmask_b32_e64 v6, -1, 1, s[6:7]
+; CHECK-NEXT: v_add_u32_e32 v6, v8, v6
+; CHECK-NEXT: s_or_b64 vcc, s[4:5], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v6, v6, v8, vcc
+; CHECK-NEXT: s_brev_b32 s8, 1
+; CHECK-NEXT: v_and_or_b32 v7, v1, s8, v6
+; CHECK-NEXT: v_bfe_u32 v6, v6, 16, 1
+; CHECK-NEXT: s_movk_i32 s9, 0x7fff
+; CHECK-NEXT: v_add3_u32 v6, v6, v7, s9
+; CHECK-NEXT: v_or_b32_e32 v7, 0x400000, v7
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[0:1], v[0:1]
+; CHECK-NEXT: v_cndmask_b32_e32 v6, v6, v7, vcc
+; CHECK-NEXT: v_cvt_f32_f64_e64 v7, |v[2:3]|
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[0:1], v7
+; CHECK-NEXT: v_and_b32_e32 v8, 1, v7
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[6:7], |v[2:3]|, v[0:1]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[4:5], |v[2:3]|, v[0:1]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v8
+; CHECK-NEXT: v_cndmask_b32_e64 v0, -1, 1, s[6:7]
+; CHECK-NEXT: v_add_u32_e32 v0, v7, v0
+; CHECK-NEXT: s_or_b64 vcc, s[4:5], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v0, v0, v7, vcc
+; CHECK-NEXT: v_and_or_b32 v1, v3, s8, v0
+; CHECK-NEXT: v_bfe_u32 v0, v0, 16, 1
+; CHECK-NEXT: v_add3_u32 v0, v0, v1, s9
+; CHECK-NEXT: v_or_b32_e32 v1, 0x400000, v1
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[2:3], v[2:3]
+; CHECK-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc
+; CHECK-NEXT: s_mov_b32 s4, 0x7060302
+; CHECK-NEXT: v_perm_b32 v0, v0, v6, s4
+; CHECK-NEXT: global_store_dword v[4:5], v0, off
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: s_setpc_b64 s[30:31]
+entry:
+ %conv = fptrunc <2 x double> %num to <2 x bfloat>
+ store <2 x bfloat> %conv, ptr addrspace(1) %p, align 8
+ ret void
+}
+
+define void @v3(<3 x double> %num, ptr addrspace(1) %p) {
+; CHECK-LABEL: v3:
+; CHECK: ; %bb.0: ; %entry
+; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT: v_cvt_f32_f64_e64 v10, |v[0:1]|
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[8:9], v10
+; CHECK-NEXT: v_and_b32_e32 v11, 1, v10
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[6:7], |v[0:1]|, v[8:9]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[4:5], |v[0:1]|, v[8:9]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v11
+; CHECK-NEXT: v_cndmask_b32_e64 v8, -1, 1, s[6:7]
+; CHECK-NEXT: v_add_u32_e32 v8, v10, v8
+; CHECK-NEXT: s_or_b64 vcc, s[4:5], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v8, v8, v10, vcc
+; CHECK-NEXT: s_brev_b32 s8, 1
+; CHECK-NEXT: v_and_or_b32 v9, v1, s8, v8
+; CHECK-NEXT: v_bfe_u32 v8, v8, 16, 1
+; CHECK-NEXT: s_movk_i32 s9, 0x7fff
+; CHECK-NEXT: v_add3_u32 v8, v8, v9, s9
+; CHECK-NEXT: v_or_b32_e32 v9, 0x400000, v9
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[0:1], v[0:1]
+; CHECK-NEXT: v_cndmask_b32_e32 v8, v8, v9, vcc
+; CHECK-NEXT: v_cvt_f32_f64_e64 v9, |v[2:3]|
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[0:1], v9
+; CHECK-NEXT: v_and_b32_e32 v10, 1, v9
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[6:7], |v[2:3]|, v[0:1]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[4:5], |v[2:3]|, v[0:1]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v10
+; CHECK-NEXT: v_cndmask_b32_e64 v0, -1, 1, s[6:7]
+; CHECK-NEXT: v_add_u32_e32 v0, v9, v0
+; CHECK-NEXT: s_or_b64 vcc, s[4:5], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v0, v0, v9, vcc
+; CHECK-NEXT: v_and_or_b32 v1, v3, s8, v0
+; CHECK-NEXT: v_bfe_u32 v0, v0, 16, 1
+; CHECK-NEXT: v_add3_u32 v0, v0, v1, s9
+; CHECK-NEXT: v_or_b32_e32 v1, 0x400000, v1
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[2:3], v[2:3]
+; CHECK-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc
+; CHECK-NEXT: s_mov_b32 s4, 0x7060302
+; CHECK-NEXT: v_cvt_f32_f64_e64 v3, |v[4:5]|
+; CHECK-NEXT: v_perm_b32 v2, v0, v8, s4
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[0:1], v3
+; CHECK-NEXT: v_and_b32_e32 v8, 1, v3
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[6:7], |v[4:5]|, v[0:1]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[4:5], |v[4:5]|, v[0:1]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v8
+; CHECK-NEXT: v_cndmask_b32_e64 v0, -1, 1, s[6:7]
+; CHECK-NEXT: v_add_u32_e32 v0, v3, v0
+; CHECK-NEXT: s_or_b64 vcc, s[4:5], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
+; CHECK-NEXT: v_and_or_b32 v1, v5, s8, v0
+; CHECK-NEXT: v_bfe_u32 v0, v0, 16, 1
+; CHECK-NEXT: v_add3_u32 v0, v0, v1, s9
+; CHECK-NEXT: v_or_b32_e32 v1, 0x400000, v1
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[4:5], v[4:5]
+; CHECK-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc
+; CHECK-NEXT: global_store_short_d16_hi v[6:7], v0, off offset:4
+; CHECK-NEXT: global_store_dword v[6:7], v2, off
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: s_setpc_b64 s[30:31]
+entry:
+ %conv = fptrunc <3 x double> %num to <3 x bfloat>
+ store <3 x bfloat> %conv, ptr addrspace(1) %p, align 8
+ ret void
+}
+
+define void @v4(<4 x double> %num, ptr addrspace(1) %p) {
+; CHECK-LABEL: v4:
+; CHECK: ; %bb.0: ; %entry
+; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT: v_cvt_f32_f64_e64 v12, |v[4:5]|
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[10:11], v12
+; CHECK-NEXT: v_and_b32_e32 v13, 1, v12
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[6:7], |v[4:5]|, v[10:11]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[4:5], |v[4:5]|, v[10:11]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v13
+; CHECK-NEXT: v_cndmask_b32_e64 v10, -1, 1, s[6:7]
+; CHECK-NEXT: v_add_u32_e32 v10, v12, v10
+; CHECK-NEXT: s_or_b64 vcc, s[4:5], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v10, v10, v12, vcc
+; CHECK-NEXT: s_brev_b32 s8, 1
+; CHECK-NEXT: v_and_or_b32 v11, v5, s8, v10
+; CHECK-NEXT: v_bfe_u32 v10, v10, 16, 1
+; CHECK-NEXT: s_movk_i32 s9, 0x7fff
+; CHECK-NEXT: v_add3_u32 v10, v10, v11, s9
+; CHECK-NEXT: v_or_b32_e32 v11, 0x400000, v11
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[4:5], v[4:5]
+; CHECK-NEXT: v_cndmask_b32_e32 v10, v10, v11, vcc
+; CHECK-NEXT: v_cvt_f32_f64_e64 v11, |v[6:7]|
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[4:5], v11
+; CHECK-NEXT: v_and_b32_e32 v12, 1, v11
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[6:7], |v[6:7]|, v[4:5]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[4:5], |v[6:7]|, v[4:5]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v12
+; CHECK-NEXT: v_cndmask_b32_e64 v4, -1, 1, s[6:7]
+; CHECK-NEXT: v_add_u32_e32 v4, v11, v4
+; CHECK-NEXT: s_or_b64 vcc, s[4:5], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v4, v4, v11, vcc
+; CHECK-NEXT: v_and_or_b32 v5, v7, s8, v4
+; CHECK-NEXT: v_bfe_u32 v4, v4, 16, 1
+; CHECK-NEXT: v_add3_u32 v4, v4, v5, s9
+; CHECK-NEXT: v_or_b32_e32 v5, 0x400000, v5
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[6:7], v[6:7]
+; CHECK-NEXT: v_cndmask_b32_e32 v4, v4, v5, vcc
+; CHECK-NEXT: s_mov_b32 s10, 0x7060302
+; CHECK-NEXT: v_perm_b32 v5, v4, v10, s10
+; CHECK-NEXT: v_cvt_f32_f64_e64 v4, |v[0:1]|
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[6:7], v4
+; CHECK-NEXT: v_and_b32_e32 v10, 1, v4
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[6:7], |v[0:1]|, v[6:7]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[4:5], |v[0:1]|, v[6:7]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v10
+; CHECK-NEXT: v_cndmask_b32_e64 v6, -1, 1, s[6:7]
+; CHECK-NEXT: v_add_u32_e32 v6, v4, v6
+; CHECK-NEXT: s_or_b64 vcc, s[4:5], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v4, v6, v4, vcc
+; CHECK-NEXT: v_and_or_b32 v6, v1, s8, v4
+; CHECK-NEXT: v_bfe_u32 v4, v4, 16, 1
+; CHECK-NEXT: v_add3_u32 v4, v4, v6, s9
+; CHECK-NEXT: v_or_b32_e32 v6, 0x400000, v6
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[0:1], v[0:1]
+; CHECK-NEXT: v_cndmask_b32_e32 v4, v4, v6, vcc
+; CHECK-NEXT: v_cvt_f32_f64_e64 v6, |v[2:3]|
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[0:1], v6
+; CHECK-NEXT: v_and_b32_e32 v7, 1, v6
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[6:7], |v[2:3]|, v[0:1]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[4:5], |v[2:3]|, v[0:1]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v7
+; CHECK-NEXT: v_cndmask_b32_e64 v0, -1, 1, s[6:7]
+; CHECK-NEXT: v_add_u32_e32 v0, v6, v0
+; CHECK-NEXT: s_or_b64 vcc, s[4:5], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v0, v0, v6, vcc
+; CHECK-NEXT: v_and_or_b32 v1, v3, s8, v0
+; CHECK-NEXT: v_bfe_u32 v0, v0, 16, 1
+; CHECK-NEXT: v_add3_u32 v0, v0, v1, s9
+; CHECK-NEXT: v_or_b32_e32 v1, 0x400000, v1
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[2:3], v[2:3]
+; CHECK-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc
+; CHECK-NEXT: v_perm_b32 v4, v0, v4, s10
+; CHECK-NEXT: global_store_dwordx2 v[8:9], v[4:5], off
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: s_setpc_b64 s[30:31]
+entry:
+ %conv = fptrunc <4 x double> %num to <4 x bfloat>
+ store <4 x bfloat> %conv, ptr addrspace(1) %p, align 8
+ ret void
+}
+
+define void @v8(<8 x double> %num, ptr addrspace(1) %p) {
+; CHECK-LABEL: v8:
+; CHECK: ; %bb.0: ; %entry
+; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT: v_cvt_f32_f64_e64 v20, |v[12:13]|
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[18:19], v20
+; CHECK-NEXT: v_and_b32_e32 v21, 1, v20
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[6:7], |v[12:13]|, v[18:19]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[4:5], |v[12:13]|, v[18:19]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v21
+; CHECK-NEXT: v_cndmask_b32_e64 v18, -1, 1, s[6:7]
+; CHECK-NEXT: v_add_u32_e32 v18, v20, v18
+; CHECK-NEXT: s_or_b64 vcc, s[4:5], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v18, v18, v20, vcc
+; CHECK-NEXT: s_brev_b32 s8, 1
+; CHECK-NEXT: v_and_or_b32 v19, v13, s8, v18
+; CHECK-NEXT: v_bfe_u32 v18, v18, 16, 1
+; CHECK-NEXT: s_movk_i32 s9, 0x7fff
+; CHECK-NEXT: v_add3_u32 v18, v18, v19, s9
+; CHECK-NEXT: v_or_b32_e32 v19, 0x400000, v19
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[12:13], v[12:13]
+; CHECK-NEXT: v_cndmask_b32_e32 v18, v18, v19, vcc
+; CHECK-NEXT: v_cvt_f32_f64_e64 v19, |v[14:15]|
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[12:13], v19
+; CHECK-NEXT: v_and_b32_e32 v20, 1, v19
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[6:7], |v[14:15]|, v[12:13]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[4:5], |v[14:15]|, v[12:13]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v20
+; CHECK-NEXT: v_cndmask_b32_e64 v12, -1, 1, s[6:7]
+; CHECK-NEXT: v_add_u32_e32 v12, v19, v12
+; CHECK-NEXT: s_or_b64 vcc, s[4:5], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v12, v12, v19, vcc
+; CHECK-NEXT: v_and_or_b32 v13, v15, s8, v12
+; CHECK-NEXT: v_bfe_u32 v12, v12, 16, 1
+; CHECK-NEXT: v_add3_u32 v12, v12, v13, s9
+; CHECK-NEXT: v_or_b32_e32 v13, 0x400000, v13
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[14:15], v[14:15]
+; CHECK-NEXT: v_cndmask_b32_e32 v12, v12, v13, vcc
+; CHECK-NEXT: s_mov_b32 s10, 0x7060302
+; CHECK-NEXT: v_perm_b32 v13, v12, v18, s10
+; CHECK-NEXT: v_cvt_f32_f64_e64 v12, |v[8:9]|
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[14:15], v12
+; CHECK-NEXT: v_and_b32_e32 v18, 1, v12
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[6:7], |v[8:9]|, v[14:15]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[4:5], |v[8:9]|, v[14:15]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v18
+; CHECK-NEXT: v_cndmask_b32_e64 v14, -1, 1, s[6:7]
+; CHECK-NEXT: v_add_u32_e32 v14, v12, v14
+; CHECK-NEXT: s_or_b64 vcc, s[4:5], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v12, v14, v12, vcc
+; CHECK-NEXT: v_and_or_b32 v14, v9, s8, v12
+; CHECK-NEXT: v_bfe_u32 v12, v12, 16, 1
+; CHECK-NEXT: v_add3_u32 v12, v12, v14, s9
+; CHECK-NEXT: v_or_b32_e32 v14, 0x400000, v14
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[8:9], v[8:9]
+; CHECK-NEXT: v_cndmask_b32_e32 v12, v12, v14, vcc
+; CHECK-NEXT: v_cvt_f32_f64_e64 v14, |v[10:11]|
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[8:9], v14
+; CHECK-NEXT: v_and_b32_e32 v15, 1, v14
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[6:7], |v[10:11]|, v[8:9]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[4:5], |v[10:11]|, v[8:9]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v15
+; CHECK-NEXT: v_cndmask_b32_e64 v8, -1, 1, s[6:7]
+; CHECK-NEXT: v_add_u32_e32 v8, v14, v8
+; CHECK-NEXT: s_or_b64 vcc, s[4:5], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v8, v8, v14, vcc
+; CHECK-NEXT: v_and_or_b32 v9, v11, s8, v8
+; CHECK-NEXT: v_bfe_u32 v8, v8, 16, 1
+; CHECK-NEXT: v_add3_u32 v8, v8, v9, s9
+; CHECK-NEXT: v_or_b32_e32 v9, 0x400000, v9
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[10:11], v[10:11]
+; CHECK-NEXT: v_cndmask_b32_e32 v8, v8, v9, vcc
+; CHECK-NEXT: v_cvt_f32_f64_e64 v10, |v[4:5]|
+; CHECK-NEXT: v_perm_b32 v12, v8, v12, s10
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[8:9], v10
+; CHECK-NEXT: v_and_b32_e32 v11, 1, v10
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[6:7], |v[4:5]|, v[8:9]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[4:5], |v[4:5]|, v[8:9]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v11
+; CHECK-NEXT: v_cndmask_b32_e64 v8, -1, 1, s[6:7]
+; CHECK-NEXT: v_add_u32_e32 v8, v10, v8
+; CHECK-NEXT: s_or_b64 vcc, s[4:5], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v8, v8, v10, vcc
+; CHECK-NEXT: v_and_or_b32 v9, v5, s8, v8
+; CHECK-NEXT: v_bfe_u32 v8, v8, 16, 1
+; CHECK-NEXT: v_add3_u32 v8, v8, v9, s9
+; CHECK-NEXT: v_or_b32_e32 v9, 0x400000, v9
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[4:5], v[4:5]
+; CHECK-NEXT: v_cndmask_b32_e32 v8, v8, v9, vcc
+; CHECK-NEXT: v_cvt_f32_f64_e64 v9, |v[6:7]|
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[4:5], v9
+; CHECK-NEXT: v_and_b32_e32 v10, 1, v9
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[6:7], |v[6:7]|, v[4:5]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[4:5], |v[6:7]|, v[4:5]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v10
+; CHECK-NEXT: v_cndmask_b32_e64 v4, -1, 1, s[6:7]
+; CHECK-NEXT: v_add_u32_e32 v4, v9, v4
+; CHECK-NEXT: s_or_b64 vcc, s[4:5], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v4, v4, v9, vcc
+; CHECK-NEXT: v_and_or_b32 v5, v7, s8, v4
+; CHECK-NEXT: v_bfe_u32 v4, v4, 16, 1
+; CHECK-NEXT: v_add3_u32 v4, v4, v5, s9
+; CHECK-NEXT: v_or_b32_e32 v5, 0x400000, v5
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[6:7], v[6:7]
+; CHECK-NEXT: v_cndmask_b32_e32 v4, v4, v5, vcc
+; CHECK-NEXT: v_cvt_f32_f64_e64 v6, |v[0:1]|
+; CHECK-NEXT: v_perm_b32 v11, v4, v8, s10
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[4:5], v6
+; CHECK-NEXT: v_and_b32_e32 v7, 1, v6
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[6:7], |v[0:1]|, v[4:5]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[4:5], |v[0:1]|, v[4:5]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v7
+; CHECK-NEXT: v_cndmask_b32_e64 v4, -1, 1, s[6:7]
+; CHECK-NEXT: v_add_u32_e32 v4, v6, v4
+; CHECK-NEXT: s_or_b64 vcc, s[4:5], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v4, v4, v6, vcc
+; CHECK-NEXT: v_and_or_b32 v5, v1, s8, v4
+; CHECK-NEXT: v_bfe_u32 v4, v4, 16, 1
+; CHECK-NEXT: v_add3_u32 v4, v4, v5, s9
+; CHECK-NEXT: v_or_b32_e32 v5, 0x400000, v5
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[0:1], v[0:1]
+; CHECK-NEXT: v_cndmask_b32_e32 v4, v4, v5, vcc
+; CHECK-NEXT: v_cvt_f32_f64_e64 v5, |v[2:3]|
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[0:1], v5
+; CHECK-NEXT: v_and_b32_e32 v6, 1, v5
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[6:7], |v[2:3]|, v[0:1]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[4:5], |v[2:3]|, v[0:1]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v6
+; CHECK-NEXT: v_cndmask_b32_e64 v0, -1, 1, s[6:7]
+; CHECK-NEXT: v_add_u32_e32 v0, v5, v0
+; CHECK-NEXT: s_or_b64 vcc, s[4:5], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v0, v0, v5, vcc
+; CHECK-NEXT: v_and_or_b32 v1, v3, s8, v0
+; CHECK-NEXT: v_bfe_u32 v0, v0, 16, 1
+; CHECK-NEXT: v_add3_u32 v0, v0, v1, s9
+; CHECK-NEXT: v_or_b32_e32 v1, 0x400000, v1
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[2:3], v[2:3]
+; CHECK-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc
+; CHECK-NEXT: v_perm_b32 v10, v0, v4, s10
+; CHECK-NEXT: global_store_dwordx4 v[16:17], v[10:13], off
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: s_setpc_b64 s[30:31]
+entry:
+ %conv = fptrunc <8 x double> %num to <8 x bfloat>
+ store <8 x bfloat> %conv, ptr addrspace(1) %p, align 8
+ ret void
+}
+
+define void @v16(<16 x double> %num, ptr addrspace(1) %p) {
+; CHECK-LABEL: v16:
+; CHECK: ; %bb.0: ; %entry
+; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
+; CHECK-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
+; CHECK-NEXT: buffer_load_dword v31, off, s[0:3], s32
+; CHECK-NEXT: v_cvt_f32_f64_e64 v36, |v[12:13]|
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[34:35], v36
+; CHECK-NEXT: v_and_b32_e32 v37, 1, v36
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[6:7], |v[12:13]|, v[34:35]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[4:5], |v[12:13]|, v[34:35]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v37
+; CHECK-NEXT: v_cndmask_b32_e64 v34, -1, 1, s[6:7]
+; CHECK-NEXT: v_add_u32_e32 v34, v36, v34
+; CHECK-NEXT: s_or_b64 vcc, s[4:5], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v34, v34, v36, vcc
+; CHECK-NEXT: s_brev_b32 s4, 1
+; CHECK-NEXT: v_and_or_b32 v35, v13, s4, v34
+; CHECK-NEXT: v_bfe_u32 v34, v34, 16, 1
+; CHECK-NEXT: s_movk_i32 s5, 0x7fff
+; CHECK-NEXT: v_add3_u32 v34, v34, v35, s5
+; CHECK-NEXT: v_or_b32_e32 v35, 0x400000, v35
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[12:13], v[12:13]
+; CHECK-NEXT: v_cndmask_b32_e32 v34, v34, v35, vcc
+; CHECK-NEXT: v_cvt_f32_f64_e64 v35, |v[14:15]|
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[12:13], v35
+; CHECK-NEXT: v_and_b32_e32 v36, 1, v35
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[8:9], |v[14:15]|, v[12:13]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[6:7], |v[14:15]|, v[12:13]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v36
+; CHECK-NEXT: v_cndmask_b32_e64 v12, -1, 1, s[8:9]
+; CHECK-NEXT: v_add_u32_e32 v12, v35, v12
+; CHECK-NEXT: s_or_b64 vcc, s[6:7], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v12, v12, v35, vcc
+; CHECK-NEXT: v_and_or_b32 v13, v15, s4, v12
+; CHECK-NEXT: v_bfe_u32 v12, v12, 16, 1
+; CHECK-NEXT: v_add3_u32 v12, v12, v13, s5
+; CHECK-NEXT: v_or_b32_e32 v13, 0x400000, v13
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[14:15], v[14:15]
+; CHECK-NEXT: v_cndmask_b32_e32 v12, v12, v13, vcc
+; CHECK-NEXT: s_mov_b32 s6, 0x7060302
+; CHECK-NEXT: v_perm_b32 v13, v12, v34, s6
+; CHECK-NEXT: v_cvt_f32_f64_e64 v12, |v[8:9]|
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[14:15], v12
+; CHECK-NEXT: v_and_b32_e32 v34, 1, v12
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[10:11], |v[8:9]|, v[14:15]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[8:9], |v[8:9]|, v[14:15]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v34
+; CHECK-NEXT: v_cndmask_b32_e64 v14, -1, 1, s[10:11]
+; CHECK-NEXT: v_add_u32_e32 v14, v12, v14
+; CHECK-NEXT: s_or_b64 vcc, s[8:9], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v12, v14, v12, vcc
+; CHECK-NEXT: v_and_or_b32 v14, v9, s4, v12
+; CHECK-NEXT: v_bfe_u32 v12, v12, 16, 1
+; CHECK-NEXT: v_add3_u32 v12, v12, v14, s5
+; CHECK-NEXT: v_or_b32_e32 v14, 0x400000, v14
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[8:9], v[8:9]
+; CHECK-NEXT: v_cndmask_b32_e32 v12, v12, v14, vcc
+; CHECK-NEXT: v_cvt_f32_f64_e64 v14, |v[10:11]|
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[8:9], v14
+; CHECK-NEXT: v_and_b32_e32 v15, 1, v14
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[10:11], |v[10:11]|, v[8:9]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[8:9], |v[10:11]|, v[8:9]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v15
+; CHECK-NEXT: v_cndmask_b32_e64 v8, -1, 1, s[10:11]
+; CHECK-NEXT: v_add_u32_e32 v8, v14, v8
+; CHECK-NEXT: s_or_b64 vcc, s[8:9], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v8, v8, v14, vcc
+; CHECK-NEXT: v_and_or_b32 v9, v11, s4, v8
+; CHECK-NEXT: v_bfe_u32 v8, v8, 16, 1
+; CHECK-NEXT: v_add3_u32 v8, v8, v9, s5
+; CHECK-NEXT: v_or_b32_e32 v9, 0x400000, v9
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[10:11], v[10:11]
+; CHECK-NEXT: v_cndmask_b32_e32 v8, v8, v9, vcc
+; CHECK-NEXT: v_cvt_f32_f64_e64 v10, |v[4:5]|
+; CHECK-NEXT: v_perm_b32 v12, v8, v12, s6
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[8:9], v10
+; CHECK-NEXT: v_and_b32_e32 v11, 1, v10
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[10:11], |v[4:5]|, v[8:9]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[8:9], |v[4:5]|, v[8:9]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v11
+; CHECK-NEXT: v_cndmask_b32_e64 v8, -1, 1, s[10:11]
+; CHECK-NEXT: v_add_u32_e32 v8, v10, v8
+; CHECK-NEXT: s_or_b64 vcc, s[8:9], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v8, v8, v10, vcc
+; CHECK-NEXT: v_and_or_b32 v9, v5, s4, v8
+; CHECK-NEXT: v_bfe_u32 v8, v8, 16, 1
+; CHECK-NEXT: v_add3_u32 v8, v8, v9, s5
+; CHECK-NEXT: v_or_b32_e32 v9, 0x400000, v9
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[4:5], v[4:5]
+; CHECK-NEXT: v_cndmask_b32_e32 v8, v8, v9, vcc
+; CHECK-NEXT: v_cvt_f32_f64_e64 v9, |v[6:7]|
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[4:5], v9
+; CHECK-NEXT: v_and_b32_e32 v10, 1, v9
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[10:11], |v[6:7]|, v[4:5]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[8:9], |v[6:7]|, v[4:5]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v10
+; CHECK-NEXT: v_cndmask_b32_e64 v4, -1, 1, s[10:11]
+; CHECK-NEXT: v_add_u32_e32 v4, v9, v4
+; CHECK-NEXT: s_or_b64 vcc, s[8:9], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v4, v4, v9, vcc
+; CHECK-NEXT: v_and_or_b32 v5, v7, s4, v4
+; CHECK-NEXT: v_bfe_u32 v4, v4, 16, 1
+; CHECK-NEXT: v_add3_u32 v4, v4, v5, s5
+; CHECK-NEXT: v_or_b32_e32 v5, 0x400000, v5
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[6:7], v[6:7]
+; CHECK-NEXT: v_cndmask_b32_e32 v4, v4, v5, vcc
+; CHECK-NEXT: v_cvt_f32_f64_e64 v6, |v[0:1]|
+; CHECK-NEXT: v_perm_b32 v11, v4, v8, s6
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[4:5], v6
+; CHECK-NEXT: v_and_b32_e32 v7, 1, v6
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[10:11], |v[0:1]|, v[4:5]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[8:9], |v[0:1]|, v[4:5]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v7
+; CHECK-NEXT: v_cndmask_b32_e64 v4, -1, 1, s[10:11]
+; CHECK-NEXT: v_add_u32_e32 v4, v6, v4
+; CHECK-NEXT: s_or_b64 vcc, s[8:9], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v4, v4, v6, vcc
+; CHECK-NEXT: v_and_or_b32 v5, v1, s4, v4
+; CHECK-NEXT: v_bfe_u32 v4, v4, 16, 1
+; CHECK-NEXT: v_add3_u32 v4, v4, v5, s5
+; CHECK-NEXT: v_or_b32_e32 v5, 0x400000, v5
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[0:1], v[0:1]
+; CHECK-NEXT: v_cndmask_b32_e32 v4, v4, v5, vcc
+; CHECK-NEXT: v_cvt_f32_f64_e64 v5, |v[2:3]|
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[0:1], v5
+; CHECK-NEXT: v_and_b32_e32 v6, 1, v5
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[10:11], |v[2:3]|, v[0:1]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[8:9], |v[2:3]|, v[0:1]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v6
+; CHECK-NEXT: v_cndmask_b32_e64 v0, -1, 1, s[10:11]
+; CHECK-NEXT: v_add_u32_e32 v0, v5, v0
+; CHECK-NEXT: s_or_b64 vcc, s[8:9], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v0, v0, v5, vcc
+; CHECK-NEXT: v_and_or_b32 v1, v3, s4, v0
+; CHECK-NEXT: v_bfe_u32 v0, v0, 16, 1
+; CHECK-NEXT: v_add3_u32 v0, v0, v1, s5
+; CHECK-NEXT: v_or_b32_e32 v1, 0x400000, v1
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[2:3], v[2:3]
+; CHECK-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc
+; CHECK-NEXT: v_cvt_f32_f64_e64 v2, |v[28:29]|
+; CHECK-NEXT: v_perm_b32 v10, v0, v4, s6
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[0:1], v2
+; CHECK-NEXT: v_and_b32_e32 v3, 1, v2
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[10:11], |v[28:29]|, v[0:1]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[8:9], |v[28:29]|, v[0:1]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v3
+; CHECK-NEXT: v_cndmask_b32_e64 v0, -1, 1, s[10:11]
+; CHECK-NEXT: v_add_u32_e32 v0, v2, v0
+; CHECK-NEXT: s_or_b64 vcc, s[8:9], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc
+; CHECK-NEXT: v_and_or_b32 v1, v29, s4, v0
+; CHECK-NEXT: v_bfe_u32 v0, v0, 16, 1
+; CHECK-NEXT: v_add3_u32 v0, v0, v1, s5
+; CHECK-NEXT: v_or_b32_e32 v1, 0x400000, v1
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[28:29], v[28:29]
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: v_cvt_f32_f64_e64 v3, |v[30:31]|
+; CHECK-NEXT: v_cndmask_b32_e32 v2, v0, v1, vcc
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[0:1], v3
+; CHECK-NEXT: v_and_b32_e32 v4, 1, v3
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[10:11], |v[30:31]|, v[0:1]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[8:9], |v[30:31]|, v[0:1]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v4
+; CHECK-NEXT: v_cndmask_b32_e64 v0, -1, 1, s[10:11]
+; CHECK-NEXT: v_add_u32_e32 v0, v3, v0
+; CHECK-NEXT: s_or_b64 vcc, s[8:9], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
+; CHECK-NEXT: v_and_or_b32 v1, v31, s4, v0
+; CHECK-NEXT: v_bfe_u32 v0, v0, 16, 1
+; CHECK-NEXT: v_add3_u32 v0, v0, v1, s5
+; CHECK-NEXT: v_or_b32_e32 v1, 0x400000, v1
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[30:31], v[30:31]
+; CHECK-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc
+; CHECK-NEXT: v_perm_b32 v3, v0, v2, s6
+; CHECK-NEXT: v_cvt_f32_f64_e64 v2, |v[24:25]|
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[0:1], v2
+; CHECK-NEXT: v_and_b32_e32 v4, 1, v2
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[10:11], |v[24:25]|, v[0:1]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[8:9], |v[24:25]|, v[0:1]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v4
+; CHECK-NEXT: v_cndmask_b32_e64 v0, -1, 1, s[10:11]
+; CHECK-NEXT: v_add_u32_e32 v0, v2, v0
+; CHECK-NEXT: s_or_b64 vcc, s[8:9], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc
+; CHECK-NEXT: v_and_or_b32 v1, v25, s4, v0
+; CHECK-NEXT: v_bfe_u32 v0, v0, 16, 1
+; CHECK-NEXT: v_add3_u32 v0, v0, v1, s5
+; CHECK-NEXT: v_or_b32_e32 v1, 0x400000, v1
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[24:25], v[24:25]
+; CHECK-NEXT: v_cvt_f32_f64_e64 v4, |v[26:27]|
+; CHECK-NEXT: v_cndmask_b32_e32 v2, v0, v1, vcc
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[0:1], v4
+; CHECK-NEXT: v_and_b32_e32 v5, 1, v4
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[10:11], |v[26:27]|, v[0:1]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[8:9], |v[26:27]|, v[0:1]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v5
+; CHECK-NEXT: v_cndmask_b32_e64 v0, -1, 1, s[10:11]
+; CHECK-NEXT: v_add_u32_e32 v0, v4, v0
+; CHECK-NEXT: s_or_b64 vcc, s[8:9], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v0, v0, v4, vcc
+; CHECK-NEXT: v_and_or_b32 v1, v27, s4, v0
+; CHECK-NEXT: v_bfe_u32 v0, v0, 16, 1
+; CHECK-NEXT: v_add3_u32 v0, v0, v1, s5
+; CHECK-NEXT: v_or_b32_e32 v1, 0x400000, v1
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[26:27], v[26:27]
+; CHECK-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc
+; CHECK-NEXT: v_cvt_f32_f64_e64 v4, |v[20:21]|
+; CHECK-NEXT: v_perm_b32 v2, v0, v2, s6
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[0:1], v4
+; CHECK-NEXT: v_and_b32_e32 v5, 1, v4
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[10:11], |v[20:21]|, v[0:1]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[8:9], |v[20:21]|, v[0:1]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v5
+; CHECK-NEXT: v_cndmask_b32_e64 v0, -1, 1, s[10:11]
+; CHECK-NEXT: v_add_u32_e32 v0, v4, v0
+; CHECK-NEXT: s_or_b64 vcc, s[8:9], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v0, v0, v4, vcc
+; CHECK-NEXT: v_and_or_b32 v1, v21, s4, v0
+; CHECK-NEXT: v_bfe_u32 v0, v0, 16, 1
+; CHECK-NEXT: v_add3_u32 v0, v0, v1, s5
+; CHECK-NEXT: v_or_b32_e32 v1, 0x400000, v1
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[20:21], v[20:21]
+; CHECK-NEXT: v_cvt_f32_f64_e64 v5, |v[22:23]|
+; CHECK-NEXT: v_cndmask_b32_e32 v4, v0, v1, vcc
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[0:1], v5
+; CHECK-NEXT: v_and_b32_e32 v6, 1, v5
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[10:11], |v[22:23]|, v[0:1]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[8:9], |v[22:23]|, v[0:1]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v6
+; CHECK-NEXT: v_cndmask_b32_e64 v0, -1, 1, s[10:11]
+; CHECK-NEXT: v_add_u32_e32 v0, v5, v0
+; CHECK-NEXT: s_or_b64 vcc, s[8:9], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v0, v0, v5, vcc
+; CHECK-NEXT: v_and_or_b32 v1, v23, s4, v0
+; CHECK-NEXT: v_bfe_u32 v0, v0, 16, 1
+; CHECK-NEXT: v_add3_u32 v0, v0, v1, s5
+; CHECK-NEXT: v_or_b32_e32 v1, 0x400000, v1
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[22:23], v[22:23]
+; CHECK-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc
+; CHECK-NEXT: v_perm_b32 v1, v0, v4, s6
+; CHECK-NEXT: v_cvt_f32_f64_e64 v0, |v[16:17]|
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[4:5], v0
+; CHECK-NEXT: v_and_b32_e32 v6, 1, v0
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[10:11], |v[16:17]|, v[4:5]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[8:9], |v[16:17]|, v[4:5]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v6
+; CHECK-NEXT: v_cndmask_b32_e64 v4, -1, 1, s[10:11]
+; CHECK-NEXT: v_add_u32_e32 v4, v0, v4
+; CHECK-NEXT: s_or_b64 vcc, s[8:9], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v0, v4, v0, vcc
+; CHECK-NEXT: v_and_or_b32 v4, v17, s4, v0
+; CHECK-NEXT: v_bfe_u32 v0, v0, 16, 1
+; CHECK-NEXT: v_add3_u32 v0, v0, v4, s5
+; CHECK-NEXT: v_or_b32_e32 v4, 0x400000, v4
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[16:17], v[16:17]
+; CHECK-NEXT: v_cvt_f32_f64_e64 v6, |v[18:19]|
+; CHECK-NEXT: v_cndmask_b32_e32 v0, v0, v4, vcc
+; CHECK-NEXT: v_cvt_f64_f32_e32 v[4:5], v6
+; CHECK-NEXT: v_and_b32_e32 v7, 1, v6
+; CHECK-NEXT: v_cmp_gt_f64_e64 s[10:11], |v[18:19]|, v[4:5]
+; CHECK-NEXT: v_cmp_nlg_f64_e64 s[8:9], |v[18:19]|, v[4:5]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, v7
+; CHECK-NEXT: v_cndmask_b32_e64 v4, -1, 1, s[10:11]
+; CHECK-NEXT: v_add_u32_e32 v4, v6, v4
+; CHECK-NEXT: s_or_b64 vcc, s[8:9], vcc
+; CHECK-NEXT: v_cndmask_b32_e32 v4, v4, v6, vcc
+; CHECK-NEXT: v_and_or_b32 v5, v19, s4, v4
+; CHECK-NEXT: v_bfe_u32 v4, v4, 16, 1
+; CHECK-NEXT: v_add3_u32 v4, v4, v5, s5
+; CHECK-NEXT: v_or_b32_e32 v5, 0x400000, v5
+; CHECK-NEXT: v_cmp_u_f64_e32 vcc, v[18:19], v[18:19]
+; CHECK-NEXT: v_cndmask_b32_e32 v4, v4, v5, vcc
+; CHECK-NEXT: v_perm_b32 v0, v4, v0, s6
+; CHECK-NEXT: global_store_dwordx4 v[32:33], v[0:3], off offset:16
+; CHECK-NEXT: global_store_dwordx4 v[32:33], v[10:13], off
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: s_setpc_b64 s[30:31]
+entry:
+ %conv = fptrunc <16 x double> %num to <16 x bfloat>
+ store <16 x bfloat> %conv, ptr addrspace(1) %p, align 8
+ ret void
+}
More information about the llvm-commits
mailing list