[llvm] AMDGPU: Directly use v2bf16 as register type for bf16 vectors. (PR #175761)
via llvm-commits
llvm-commits at lists.llvm.org
Tue Jan 13 06:31:30 PST 2026
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-amdgpu
Author: Matt Arsenault (arsenm)
<details>
<summary>Changes</summary>
Previously we were casting v2bf16 to i32, unlike the f16 case. Simplify
this by using the natural vector type. This is probably a leftover from
before v2bf16 was treated as legal. This is preparation for fixing a
miscompile in globalisel.
---
Patch is 184.57 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/175761.diff
9 Files Affected:
- (modified) llvm/lib/Target/AMDGPU/SIISelLowering.cpp (+3-6)
- (modified) llvm/test/CodeGen/AMDGPU/GlobalISel/irtranslate-bf16.ll (+121-282)
- (modified) llvm/test/CodeGen/AMDGPU/GlobalISel/irtranslator-function-args.ll (+2-3)
- (modified) llvm/test/CodeGen/AMDGPU/bf16.ll (+461-458)
- (modified) llvm/test/CodeGen/AMDGPU/constant-address-space-32bit.ll (-2)
- (modified) llvm/test/CodeGen/AMDGPU/llvm.exp2.bf16.ll (+122-133)
- (modified) llvm/test/CodeGen/AMDGPU/llvm.log2.bf16.ll (+14-24)
- (modified) llvm/test/CodeGen/AMDGPU/maximumnum.bf16.ll (+74-74)
- (modified) llvm/test/CodeGen/AMDGPU/minimumnum.bf16.ll (+74-74)
``````````diff
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 139e2d101a077..61c75e1c107ed 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -1111,11 +1111,8 @@ MVT SITargetLowering::getRegisterTypeForCallingConv(LLVMContext &Context,
EVT ScalarVT = VT.getScalarType();
unsigned Size = ScalarVT.getSizeInBits();
if (Size == 16) {
- if (Subtarget->has16BitInsts()) {
- if (VT.isInteger())
- return MVT::v2i16;
- return (ScalarVT == MVT::bf16 ? MVT::i32 : MVT::v2f16);
- }
+ if (Subtarget->has16BitInsts())
+ return MVT::getVectorVT(ScalarVT.getSimpleVT(), 2);
return VT.isInteger() ? MVT::i32 : MVT::f32;
}
@@ -1168,7 +1165,7 @@ unsigned SITargetLowering::getVectorTypeBreakdownForCallingConv(
// inconsistent.
if (Size == 16 && Subtarget->has16BitInsts()) {
if (ScalarVT == MVT::bf16) {
- RegisterVT = MVT::i32;
+ RegisterVT = MVT::v2bf16;;
IntermediateVT = MVT::v2bf16;
} else {
RegisterVT = VT.isInteger() ? MVT::v2i16 : MVT::v2f16;
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/irtranslate-bf16.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/irtranslate-bf16.ll
index 3206f8e55f44e..0213bb35cc3a2 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/irtranslate-bf16.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/irtranslate-bf16.ll
@@ -8,24 +8,20 @@ define <3 x bfloat> @v3bf16(<3 x bfloat> %arg0) {
; GFX9: bb.1 (%ir-block.0):
; GFX9-NEXT: liveins: $vgpr0, $vgpr1
; GFX9-NEXT: {{ $}}
- ; GFX9-NEXT: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
- ; GFX9-NEXT: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
- ; GFX9-NEXT: [[UV:%[0-9]+]]:_(s16), [[UV1:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[COPY]](s32)
- ; GFX9-NEXT: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[UV]](s16)
- ; GFX9-NEXT: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[UV1]](s16)
- ; GFX9-NEXT: [[UV2:%[0-9]+]]:_(s16), [[UV3:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[COPY1]](s32)
- ; GFX9-NEXT: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[UV2]](s16)
- ; GFX9-NEXT: [[ANYEXT3:%[0-9]+]]:_(s32) = G_ANYEXT [[UV3]](s16)
- ; GFX9-NEXT: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[ANYEXT]](s32), [[ANYEXT1]](s32), [[ANYEXT2]](s32)
- ; GFX9-NEXT: [[TRUNC:%[0-9]+]]:_(<3 x s16>) = G_TRUNC [[BUILD_VECTOR]](<3 x s32>)
+ ; GFX9-NEXT: [[COPY:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr0
+ ; GFX9-NEXT: [[COPY1:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr1
+ ; GFX9-NEXT: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[COPY]](<2 x s16>), [[COPY1]](<2 x s16>)
+ ; GFX9-NEXT: [[UV:%[0-9]+]]:_(s16), [[UV1:%[0-9]+]]:_(s16), [[UV2:%[0-9]+]]:_(s16), [[UV3:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<4 x s16>)
+ ; GFX9-NEXT: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s16>) = G_BUILD_VECTOR [[UV]](s16), [[UV1]](s16), [[UV2]](s16)
; GFX9-NEXT: [[C:%[0-9]+]]:_(s16) = G_FCONSTANT bfloat 0xR0000
; GFX9-NEXT: [[BUILD_VECTOR1:%[0-9]+]]:_(<3 x s16>) = G_BUILD_VECTOR [[C]](s16), [[C]](s16), [[C]](s16)
- ; GFX9-NEXT: [[SHUF:%[0-9]+]]:_(<3 x s16>) = G_SHUFFLE_VECTOR [[TRUNC]](<3 x s16>), [[BUILD_VECTOR1]], shufflemask(3, 1, 2)
+ ; GFX9-NEXT: [[SHUF:%[0-9]+]]:_(<3 x s16>) = G_SHUFFLE_VECTOR [[BUILD_VECTOR]](<3 x s16>), [[BUILD_VECTOR1]], shufflemask(3, 1, 2)
; GFX9-NEXT: [[UV4:%[0-9]+]]:_(s16), [[UV5:%[0-9]+]]:_(s16), [[UV6:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[SHUF]](<3 x s16>)
- ; GFX9-NEXT: [[ANYEXT4:%[0-9]+]]:_(s32) = G_ANYEXT [[UV4]](s16)
- ; GFX9-NEXT: [[ANYEXT5:%[0-9]+]]:_(s32) = G_ANYEXT [[UV5]](s16)
- ; GFX9-NEXT: $vgpr0 = COPY [[ANYEXT4]](s32)
- ; GFX9-NEXT: $vgpr1 = COPY [[ANYEXT5]](s32)
+ ; GFX9-NEXT: [[DEF:%[0-9]+]]:_(s16) = G_IMPLICIT_DEF
+ ; GFX9-NEXT: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s16>) = G_BUILD_VECTOR [[UV4]](s16), [[UV5]](s16), [[UV6]](s16), [[DEF]](s16)
+ ; GFX9-NEXT: [[UV7:%[0-9]+]]:_(<2 x s16>), [[UV8:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[BUILD_VECTOR2]](<4 x s16>)
+ ; GFX9-NEXT: $vgpr0 = COPY [[UV7]](<2 x s16>)
+ ; GFX9-NEXT: $vgpr1 = COPY [[UV8]](<2 x s16>)
; GFX9-NEXT: SI_RETURN implicit $vgpr0, implicit $vgpr1
%res = shufflevector <3 x bfloat> %arg0, <3 x bfloat> zeroinitializer, <3 x i32> <i32 3, i32 1, i32 2>
ret <3 x bfloat> %res
@@ -36,24 +32,15 @@ define <4 x bfloat> @v4bf16(<4 x bfloat> %arg0) {
; GFX9: bb.1 (%ir-block.0):
; GFX9-NEXT: liveins: $vgpr0, $vgpr1
; GFX9-NEXT: {{ $}}
- ; GFX9-NEXT: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
- ; GFX9-NEXT: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
- ; GFX9-NEXT: [[UV:%[0-9]+]]:_(s16), [[UV1:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[COPY]](s32)
- ; GFX9-NEXT: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[UV]](s16)
- ; GFX9-NEXT: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[UV1]](s16)
- ; GFX9-NEXT: [[UV2:%[0-9]+]]:_(s16), [[UV3:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[COPY1]](s32)
- ; GFX9-NEXT: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[UV2]](s16)
- ; GFX9-NEXT: [[ANYEXT3:%[0-9]+]]:_(s32) = G_ANYEXT [[UV3]](s16)
- ; GFX9-NEXT: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[ANYEXT]](s32), [[ANYEXT1]](s32), [[ANYEXT2]](s32), [[ANYEXT3]](s32)
- ; GFX9-NEXT: [[TRUNC:%[0-9]+]]:_(<4 x s16>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+ ; GFX9-NEXT: [[COPY:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr0
+ ; GFX9-NEXT: [[COPY1:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr1
+ ; GFX9-NEXT: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[COPY]](<2 x s16>), [[COPY1]](<2 x s16>)
; GFX9-NEXT: [[C:%[0-9]+]]:_(s16) = G_FCONSTANT bfloat 0xR0000
- ; GFX9-NEXT: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s16>) = G_BUILD_VECTOR [[C]](s16), [[C]](s16), [[C]](s16), [[C]](s16)
- ; GFX9-NEXT: [[SHUF:%[0-9]+]]:_(<4 x s16>) = G_SHUFFLE_VECTOR [[TRUNC]](<4 x s16>), [[BUILD_VECTOR1]], shufflemask(3, 1, 2, 0)
- ; GFX9-NEXT: [[UV4:%[0-9]+]]:_(s16), [[UV5:%[0-9]+]]:_(s16), [[UV6:%[0-9]+]]:_(s16), [[UV7:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[SHUF]](<4 x s16>)
- ; GFX9-NEXT: [[ANYEXT4:%[0-9]+]]:_(s32) = G_ANYEXT [[UV4]](s16)
- ; GFX9-NEXT: [[ANYEXT5:%[0-9]+]]:_(s32) = G_ANYEXT [[UV5]](s16)
- ; GFX9-NEXT: $vgpr0 = COPY [[ANYEXT4]](s32)
- ; GFX9-NEXT: $vgpr1 = COPY [[ANYEXT5]](s32)
+ ; GFX9-NEXT: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s16>) = G_BUILD_VECTOR [[C]](s16), [[C]](s16), [[C]](s16), [[C]](s16)
+ ; GFX9-NEXT: [[SHUF:%[0-9]+]]:_(<4 x s16>) = G_SHUFFLE_VECTOR [[CONCAT_VECTORS]](<4 x s16>), [[BUILD_VECTOR]], shufflemask(3, 1, 2, 0)
+ ; GFX9-NEXT: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[SHUF]](<4 x s16>)
+ ; GFX9-NEXT: $vgpr0 = COPY [[UV]](<2 x s16>)
+ ; GFX9-NEXT: $vgpr1 = COPY [[UV1]](<2 x s16>)
; GFX9-NEXT: SI_RETURN implicit $vgpr0, implicit $vgpr1
%res = shufflevector <4 x bfloat> %arg0, <4 x bfloat> zeroinitializer, <4 x i32> <i32 3, i32 1, i32 2, i32 0>
ret <4 x bfloat> %res
@@ -64,30 +51,22 @@ define <5 x bfloat> @v5bf16(<5 x bfloat> %arg0) {
; GFX9: bb.1 (%ir-block.0):
; GFX9-NEXT: liveins: $vgpr0, $vgpr1, $vgpr2
; GFX9-NEXT: {{ $}}
- ; GFX9-NEXT: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
- ; GFX9-NEXT: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
- ; GFX9-NEXT: [[COPY2:%[0-9]+]]:_(s32) = COPY $vgpr2
- ; GFX9-NEXT: [[UV:%[0-9]+]]:_(s16), [[UV1:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[COPY]](s32)
- ; GFX9-NEXT: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[UV]](s16)
- ; GFX9-NEXT: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[UV1]](s16)
- ; GFX9-NEXT: [[UV2:%[0-9]+]]:_(s16), [[UV3:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[COPY1]](s32)
- ; GFX9-NEXT: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[UV2]](s16)
- ; GFX9-NEXT: [[ANYEXT3:%[0-9]+]]:_(s32) = G_ANYEXT [[UV3]](s16)
- ; GFX9-NEXT: [[UV4:%[0-9]+]]:_(s16), [[UV5:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[COPY2]](s32)
- ; GFX9-NEXT: [[ANYEXT4:%[0-9]+]]:_(s32) = G_ANYEXT [[UV4]](s16)
- ; GFX9-NEXT: [[ANYEXT5:%[0-9]+]]:_(s32) = G_ANYEXT [[UV5]](s16)
- ; GFX9-NEXT: [[BUILD_VECTOR:%[0-9]+]]:_(<5 x s32>) = G_BUILD_VECTOR [[ANYEXT]](s32), [[ANYEXT1]](s32), [[ANYEXT2]](s32), [[ANYEXT3]](s32), [[ANYEXT4]](s32)
- ; GFX9-NEXT: [[TRUNC:%[0-9]+]]:_(<5 x s16>) = G_TRUNC [[BUILD_VECTOR]](<5 x s32>)
+ ; GFX9-NEXT: [[COPY:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr0
+ ; GFX9-NEXT: [[COPY1:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr1
+ ; GFX9-NEXT: [[COPY2:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr2
+ ; GFX9-NEXT: [[CONCAT_VECTORS:%[0-9]+]]:_(<6 x s16>) = G_CONCAT_VECTORS [[COPY]](<2 x s16>), [[COPY1]](<2 x s16>), [[COPY2]](<2 x s16>)
+ ; GFX9-NEXT: [[UV:%[0-9]+]]:_(s16), [[UV1:%[0-9]+]]:_(s16), [[UV2:%[0-9]+]]:_(s16), [[UV3:%[0-9]+]]:_(s16), [[UV4:%[0-9]+]]:_(s16), [[UV5:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<6 x s16>)
+ ; GFX9-NEXT: [[BUILD_VECTOR:%[0-9]+]]:_(<5 x s16>) = G_BUILD_VECTOR [[UV]](s16), [[UV1]](s16), [[UV2]](s16), [[UV3]](s16), [[UV4]](s16)
; GFX9-NEXT: [[C:%[0-9]+]]:_(s16) = G_FCONSTANT bfloat 0xR0000
; GFX9-NEXT: [[BUILD_VECTOR1:%[0-9]+]]:_(<5 x s16>) = G_BUILD_VECTOR [[C]](s16), [[C]](s16), [[C]](s16), [[C]](s16), [[C]](s16)
- ; GFX9-NEXT: [[SHUF:%[0-9]+]]:_(<5 x s16>) = G_SHUFFLE_VECTOR [[TRUNC]](<5 x s16>), [[BUILD_VECTOR1]], shufflemask(3, 1, 2, 0, 4)
+ ; GFX9-NEXT: [[SHUF:%[0-9]+]]:_(<5 x s16>) = G_SHUFFLE_VECTOR [[BUILD_VECTOR]](<5 x s16>), [[BUILD_VECTOR1]], shufflemask(3, 1, 2, 0, 4)
; GFX9-NEXT: [[UV6:%[0-9]+]]:_(s16), [[UV7:%[0-9]+]]:_(s16), [[UV8:%[0-9]+]]:_(s16), [[UV9:%[0-9]+]]:_(s16), [[UV10:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[SHUF]](<5 x s16>)
- ; GFX9-NEXT: [[ANYEXT6:%[0-9]+]]:_(s32) = G_ANYEXT [[UV6]](s16)
- ; GFX9-NEXT: [[ANYEXT7:%[0-9]+]]:_(s32) = G_ANYEXT [[UV7]](s16)
- ; GFX9-NEXT: [[ANYEXT8:%[0-9]+]]:_(s32) = G_ANYEXT [[UV8]](s16)
- ; GFX9-NEXT: $vgpr0 = COPY [[ANYEXT6]](s32)
- ; GFX9-NEXT: $vgpr1 = COPY [[ANYEXT7]](s32)
- ; GFX9-NEXT: $vgpr2 = COPY [[ANYEXT8]](s32)
+ ; GFX9-NEXT: [[DEF:%[0-9]+]]:_(s16) = G_IMPLICIT_DEF
+ ; GFX9-NEXT: [[BUILD_VECTOR2:%[0-9]+]]:_(<6 x s16>) = G_BUILD_VECTOR [[UV6]](s16), [[UV7]](s16), [[UV8]](s16), [[UV9]](s16), [[UV10]](s16), [[DEF]](s16)
+ ; GFX9-NEXT: [[UV11:%[0-9]+]]:_(<2 x s16>), [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[BUILD_VECTOR2]](<6 x s16>)
+ ; GFX9-NEXT: $vgpr0 = COPY [[UV11]](<2 x s16>)
+ ; GFX9-NEXT: $vgpr1 = COPY [[UV12]](<2 x s16>)
+ ; GFX9-NEXT: $vgpr2 = COPY [[UV13]](<2 x s16>)
; GFX9-NEXT: SI_RETURN implicit $vgpr0, implicit $vgpr1, implicit $vgpr2
%res = shufflevector <5 x bfloat> %arg0, <5 x bfloat> zeroinitializer, <5 x i32> <i32 3, i32 1, i32 2, i32 0, i32 4>
ret <5 x bfloat> %res
@@ -98,30 +77,17 @@ define <6 x bfloat> @v6bf16(<6 x bfloat> %arg0) {
; GFX9: bb.1 (%ir-block.0):
; GFX9-NEXT: liveins: $vgpr0, $vgpr1, $vgpr2
; GFX9-NEXT: {{ $}}
- ; GFX9-NEXT: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
- ; GFX9-NEXT: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
- ; GFX9-NEXT: [[COPY2:%[0-9]+]]:_(s32) = COPY $vgpr2
- ; GFX9-NEXT: [[UV:%[0-9]+]]:_(s16), [[UV1:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[COPY]](s32)
- ; GFX9-NEXT: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[UV]](s16)
- ; GFX9-NEXT: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[UV1]](s16)
- ; GFX9-NEXT: [[UV2:%[0-9]+]]:_(s16), [[UV3:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[COPY1]](s32)
- ; GFX9-NEXT: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[UV2]](s16)
- ; GFX9-NEXT: [[ANYEXT3:%[0-9]+]]:_(s32) = G_ANYEXT [[UV3]](s16)
- ; GFX9-NEXT: [[UV4:%[0-9]+]]:_(s16), [[UV5:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[COPY2]](s32)
- ; GFX9-NEXT: [[ANYEXT4:%[0-9]+]]:_(s32) = G_ANYEXT [[UV4]](s16)
- ; GFX9-NEXT: [[ANYEXT5:%[0-9]+]]:_(s32) = G_ANYEXT [[UV5]](s16)
- ; GFX9-NEXT: [[BUILD_VECTOR:%[0-9]+]]:_(<6 x s32>) = G_BUILD_VECTOR [[ANYEXT]](s32), [[ANYEXT1]](s32), [[ANYEXT2]](s32), [[ANYEXT3]](s32), [[ANYEXT4]](s32), [[ANYEXT5]](s32)
- ; GFX9-NEXT: [[TRUNC:%[0-9]+]]:_(<6 x s16>) = G_TRUNC [[BUILD_VECTOR]](<6 x s32>)
+ ; GFX9-NEXT: [[COPY:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr0
+ ; GFX9-NEXT: [[COPY1:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr1
+ ; GFX9-NEXT: [[COPY2:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr2
+ ; GFX9-NEXT: [[CONCAT_VECTORS:%[0-9]+]]:_(<6 x s16>) = G_CONCAT_VECTORS [[COPY]](<2 x s16>), [[COPY1]](<2 x s16>), [[COPY2]](<2 x s16>)
; GFX9-NEXT: [[C:%[0-9]+]]:_(s16) = G_FCONSTANT bfloat 0xR0000
- ; GFX9-NEXT: [[BUILD_VECTOR1:%[0-9]+]]:_(<6 x s16>) = G_BUILD_VECTOR [[C]](s16), [[C]](s16), [[C]](s16), [[C]](s16), [[C]](s16), [[C]](s16)
- ; GFX9-NEXT: [[SHUF:%[0-9]+]]:_(<6 x s16>) = G_SHUFFLE_VECTOR [[TRUNC]](<6 x s16>), [[BUILD_VECTOR1]], shufflemask(3, 1, 2, 0, 4, 5)
- ; GFX9-NEXT: [[UV6:%[0-9]+]]:_(s16), [[UV7:%[0-9]+]]:_(s16), [[UV8:%[0-9]+]]:_(s16), [[UV9:%[0-9]+]]:_(s16), [[UV10:%[0-9]+]]:_(s16), [[UV11:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[SHUF]](<6 x s16>)
- ; GFX9-NEXT: [[ANYEXT6:%[0-9]+]]:_(s32) = G_ANYEXT [[UV6]](s16)
- ; GFX9-NEXT: [[ANYEXT7:%[0-9]+]]:_(s32) = G_ANYEXT [[UV7]](s16)
- ; GFX9-NEXT: [[ANYEXT8:%[0-9]+]]:_(s32) = G_ANYEXT [[UV8]](s16)
- ; GFX9-NEXT: $vgpr0 = COPY [[ANYEXT6]](s32)
- ; GFX9-NEXT: $vgpr1 = COPY [[ANYEXT7]](s32)
- ; GFX9-NEXT: $vgpr2 = COPY [[ANYEXT8]](s32)
+ ; GFX9-NEXT: [[BUILD_VECTOR:%[0-9]+]]:_(<6 x s16>) = G_BUILD_VECTOR [[C]](s16), [[C]](s16), [[C]](s16), [[C]](s16), [[C]](s16), [[C]](s16)
+ ; GFX9-NEXT: [[SHUF:%[0-9]+]]:_(<6 x s16>) = G_SHUFFLE_VECTOR [[CONCAT_VECTORS]](<6 x s16>), [[BUILD_VECTOR]], shufflemask(3, 1, 2, 0, 4, 5)
+ ; GFX9-NEXT: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>), [[UV2:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[SHUF]](<6 x s16>)
+ ; GFX9-NEXT: $vgpr0 = COPY [[UV]](<2 x s16>)
+ ; GFX9-NEXT: $vgpr1 = COPY [[UV1]](<2 x s16>)
+ ; GFX9-NEXT: $vgpr2 = COPY [[UV2]](<2 x s16>)
; GFX9-NEXT: SI_RETURN implicit $vgpr0, implicit $vgpr1, implicit $vgpr2
%res = shufflevector <6 x bfloat> %arg0, <6 x bfloat> zeroinitializer, <6 x i32> <i32 3, i32 1, i32 2, i32 0, i32 4, i32 5>
ret <6 x bfloat> %res
@@ -132,36 +98,24 @@ define <7 x bfloat> @v7bf16(<7 x bfloat> %arg0) {
; GFX9: bb.1 (%ir-block.0):
; GFX9-NEXT: liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
; GFX9-NEXT: {{ $}}
- ; GFX9-NEXT: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
- ; GFX9-NEXT: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
- ; GFX9-NEXT: [[COPY2:%[0-9]+]]:_(s32) = COPY $vgpr2
- ; GFX9-NEXT: [[COPY3:%[0-9]+]]:_(s32) = COPY $vgpr3
- ; GFX9-NEXT: [[UV:%[0-9]+]]:_(s16), [[UV1:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[COPY]](s32)
- ; GFX9-NEXT: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[UV]](s16)
- ; GFX9-NEXT: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[UV1]](s16)
- ; GFX9-NEXT: [[UV2:%[0-9]+]]:_(s16), [[UV3:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[COPY1]](s32)
- ; GFX9-NEXT: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[UV2]](s16)
- ; GFX9-NEXT: [[ANYEXT3:%[0-9]+]]:_(s32) = G_ANYEXT [[UV3]](s16)
- ; GFX9-NEXT: [[UV4:%[0-9]+]]:_(s16), [[UV5:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[COPY2]](s32)
- ; GFX9-NEXT: [[ANYEXT4:%[0-9]+]]:_(s32) = G_ANYEXT [[UV4]](s16)
- ; GFX9-NEXT: [[ANYEXT5:%[0-9]+]]:_(s32) = G_ANYEXT [[UV5]](s16)
- ; GFX9-NEXT: [[UV6:%[0-9]+]]:_(s16), [[UV7:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[COPY3]](s32)
- ; GFX9-NEXT: [[ANYEXT6:%[0-9]+]]:_(s32) = G_ANYEXT [[UV6]](s16)
- ; GFX9-NEXT: [[ANYEXT7:%[0-9]+]]:_(s32) = G_ANYEXT [[UV7]](s16)
- ; GFX9-NEXT: [[BUILD_VECTOR:%[0-9]+]]:_(<7 x s32>) = G_BUILD_VECTOR [[ANYEXT]](s32), [[ANYEXT1]](s32), [[ANYEXT2]](s32), [[ANYEXT3]](s32), [[ANYEXT4]](s32), [[ANYEXT5]](s32), [[ANYEXT6]](s32)
- ; GFX9-NEXT: [[TRUNC:%[0-9]+]]:_(<7 x s16>) = G_TRUNC [[BUILD_VECTOR]](<7 x s32>)
+ ; GFX9-NEXT: [[COPY:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr0
+ ; GFX9-NEXT: [[COPY1:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr1
+ ; GFX9-NEXT: [[COPY2:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr2
+ ; GFX9-NEXT: [[COPY3:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr3
+ ; GFX9-NEXT: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s16>) = G_CONCAT_VECTORS [[COPY]](<2 x s16>), [[COPY1]](<2 x s16>), [[COPY2]](<2 x s16>), [[COPY3]](<2 x s16>)
+ ; GFX9-NEXT: [[UV:%[0-9]+]]:_(s16), [[UV1:%[0-9]+]]:_(s16), [[UV2:%[0-9]+]]:_(s16), [[UV3:%[0-9]+]]:_(s16), [[UV4:%[0-9]+]]:_(s16), [[UV5:%[0-9]+]]:_(s16), [[UV6:%[0-9]+]]:_(s16), [[UV7:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<8 x s16>)
+ ; GFX9-NEXT: [[BUILD_VECTOR:%[0-9]+]]:_(<7 x s16>) = G_BUILD_VECTOR [[UV]](s16), [[UV1]](s16), [[UV2]](s16), [[UV3]](s16), [[UV4]](s16), [[UV5]](s16), [[UV6]](s16)
; GFX9-NEXT: [[C:%[0-9]+]]:_(s16) = G_FCONSTANT bfloat 0xR0000
; GFX9-NEXT: [[BUILD_VECTOR1:%[0-9]+]]:_(<7 x s16>) = G_BUILD_VECTOR [[C]](s16), [[C]](s16), [[C]](s16), [[C]](s16), [[C]](s16), [[C]](s16), [[C]](s16)
- ; GFX9-NEXT: [[SHUF:%[0-9]+]]:_(<7 x s16>) = G_SHUFFLE_VECTOR [[TRUNC]](<7 x s16>), [[BUILD_VECTOR1]], shufflemask(3, 1, 2, 0, 4, 5, 6)
+ ; GFX9-NEXT: [[SHUF:%[0-9]+]]:_(<7 x s16>) = G_SHUFFLE_VECTOR [[BUILD_VECTOR]](<7 x s16>), [[BUILD_VECTOR1]], shufflemask(3, 1, 2, 0, 4, 5, 6)
; GFX9-NEXT: [[UV8:%[0-9]+]]:_(s16), [[UV9:%[0-9]+]]:_(s16), [[UV10:%[0-9]+]]:_(s16), [[UV11:%[0-9]+]]:_(s16), [[UV12:%[0-9]+]]:_(s16), [[UV13:%[0-9]+]]:_(s16), [[UV14:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[SHUF]](<7 x s16>)
- ; GFX9-NEXT: [[ANYEXT8:%[0-9]+]]:_(s32) = G_ANYEXT [[UV8]](s16)
- ; GFX9-NEXT: [[ANYEXT9:%[0-9]+]]:_(s32) = G_ANYEXT [[UV9]](s16)
- ; GFX9-NEXT: [[ANYEXT10:%[0-9]+]]:_(s32) = G_ANYEXT [[UV10]](s16)
- ; GFX9-NEXT: [[ANYEXT11:%[0-9]+]]:_(s32) = G_ANYEXT [[UV11]](s16)
- ; GFX9-NEXT: $vgpr0 = COPY [[ANYEXT8]](s32)
- ; GFX9-NEXT: $vgpr1 = COPY [[ANYEXT9]](s32)
- ; GFX9-NEXT: $vgpr2 = COPY [[ANYEXT10]](s32)
- ; GFX9-NEXT: $vgpr3 = COPY [[ANYEXT11]](s32)
+ ; GFX9-NEXT: [[DEF:%[0-9]+]]:_(s16) = G_IMPLICIT_DEF
+ ; GFX9-NEXT: [[BUILD_VECTOR2:%[0-9]+]]:_(<8 x s16>) = G_BUILD_VECTOR [[UV8]](s16), [[UV9]](s16), [[UV10]](s16), [[UV11]](s16), [[UV12]](s16), [[UV13]](s16), [[UV14]](s16), [[DEF]](s16)
+ ; GFX9-NEXT: [[UV15:%[0-9]+]]:_(<2 x s16>), [[UV16:%[0-9]+]]:_(<2 x s16>), [[UV17:%[0-9]+]]:_(<2 x s16>), [[UV18:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[BUILD_VECTOR2]](<8 x s16>)
+ ; GFX9-NEXT: $vgpr0 = COPY [[UV15]](<2 x s16>)
+ ; GFX9-NEXT: $vgpr1 = COPY [[UV16]](<2 x s16>)
+ ; GFX9-NEXT: $vgpr2 = COPY [[UV17]](<2 x s16>)
+ ; GFX9-NEXT: $vgpr3 = COPY [[UV18]](<2 x s16>)
; GFX9-NEXT: SI_RETURN implicit $vgpr0, implicit $vgpr1, implicit $vgpr2, implicit $vgpr3
%res = shufflevector <7 x bfloat> %arg0, <7 x bfloat> zeroinitializer, <7 x i32> <i32 3, i32 1, i32 2, i32 0, i32 4, i32 5, i32 6>
ret <7 x bfloat> %res
@@ -172,36 +126,19 @@ define <8 x bfloat> @v8bf16(<8 x bfloat> %arg0) {
; GFX9: bb.1 (%ir-block.0):
; GFX9-NEXT: liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
; GFX9-NEXT: {{ $}}
- ; GFX9-NEXT: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
- ; GFX9-NEXT: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
- ; GFX9-NEXT: [[COPY2:%[0-9]+]]:_(s32) = COPY $vgpr2
- ; GFX9-NEXT: [[COPY3:%[0-9]+]]:_(s32) = COPY $vgpr3
- ; GFX9-NEXT: [[UV:%[0-9]+]]:_(s16), [[UV1:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[COPY]](s32)
- ; GFX9-NEXT: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[UV]](s16)
- ; GFX9-NEXT: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[UV1]](s16)
- ; GFX9-NEXT: [[UV2:%[0-9]+]]:_(s16), [[UV3:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[COPY1]](s32)
- ; GFX9-NEXT: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[UV2]](s16)
- ; GFX9-NEXT: [[ANYEXT3:%[0-9]+]]:_(s32) = G_ANYEXT [[UV3]](s16)
- ; GFX9-NEXT: [[UV4:%[0-9]+]]:_(s16), [[UV5:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[COPY2]](s32)
- ; GFX9-NEXT: [[ANYEXT4:%[0-9]+]]:_(s32) = G_ANYEXT [[UV4]](s16)
- ; GFX9-NEXT: [[ANYEXT5:%[0-9]+]]:_(s32) = G_ANYEXT [[UV5]](s16)
- ; GFX9-NEXT: [[UV6:%[0-9]+]]:_(s16), [[UV7:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[COPY3]](s32)
- ; GFX9-NEXT: [[ANYEXT6:%[0-9]+]]:_(s32) = G_ANYEXT [[UV6]](s16)
- ; GFX9-NEXT: [[ANYEXT7:%[0-9]+]]:_(s32) = G_ANYEXT [[UV7]](s16)
- ; GFX9-NEXT: [[BUILD_VECTOR:%[0-9]+]]:_(<8 x s32>) = G_BUILD_VECTOR [[ANYEXT]](s32), [[ANYEXT1]](s32), [[ANYEXT2]](s32), [[ANYEXT3]](s32), [[ANYEXT4]](s32), [[ANYEXT5]](s32), [[ANYEXT6]](s32), [[ANYEXT7]](s32)
- ; GFX9-NEXT: [[TRUNC:%[0-9]+]]:_(<8 x s16>) = G_TRUNC [[BUILD_VECTOR]](<8 x s32>)
+ ; GFX9-NEXT: [[COPY:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr0
+ ; GFX9-NEXT: [[COPY1:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr1
+ ; GFX9-NEXT: [[COPY2:%[0-9...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/175761
More information about the llvm-commits
mailing list