caio/clang-p2996
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[GISel] Fix ShuffleVector assert (#139769)

Browse Source 
Fixes issue: https://github.com/llvm/llvm-project/issues/139752

When G_SHUFFLE_VECTOR has only 1 element then it is possible the vector
is decayed into a scalar.
This commit is contained in:
Alan Li
2025-05-20 21:25:31 -04:00
committed by GitHub
parent 744a469bba
commit ada2fbfe36
3 changed files with 84 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp
View File

@@ -420,8 +420,11 @@ void CombinerHelper::applyCombineShuffleToBuildVector(MachineInstr &MI) const {
else
Extracts.push_back(Unmerge2.getReg(Val - Width));
}
Builder.buildBuildVector(MI.getOperand(0).getReg(), Extracts);
assert(Extracts.size() > 0 && "Expected at least one element in the shuffle");
if (Extracts.size() == 1)
Builder.buildCopy(MI.getOperand(0).getReg(), Extracts[0]);
else
Builder.buildBuildVector(MI.getOperand(0).getReg(), Extracts);
MI.eraseFromParent();
}

35
llvm/test/CodeGen/AMDGPU/GlobalISel/bug_shuffle_vector_to_scalar.ll Normal file
View File

@@ -0,0 +1,35 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck %s
; Description: an end-to-end IR test for https://github.com/llvm/llvm-project/issues/139752
; To test combine shuffle_vectors into build_vector
define amdgpu_gs <4 x float> @_amdgpu_gs_main() {
; CHECK-LABEL: _amdgpu_gs_main:
; CHECK: ; %bb.0: ; %bb
; CHECK-NEXT: v_mov_b32_e32 v0, 16
; CHECK-NEXT: ds_read2_b32 v[0:1], v0 offset1:1
; CHECK-NEXT: s_mov_b32 s0, 0
; CHECK-NEXT: s_mov_b32 s1, s0
; CHECK-NEXT: s_mov_b32 s2, s0
; CHECK-NEXT: s_mov_b32 s3, s0
; CHECK-NEXT: v_mov_b32_e32 v2, 0
; CHECK-NEXT: s_waitcnt lgkmcnt(0)
; CHECK-NEXT: buffer_store_dwordx4 v[0:3], v2, s[0:3], 0 idxen
; CHECK-NEXT: s_nop 0
; CHECK-NEXT: v_mov_b32_e32 v0, v1
; CHECK-NEXT: s_waitcnt vmcnt(0)
; CHECK-NEXT: ; return to shader part epilog
bb:
%i = load <1 x float>, ptr addrspace(3) getelementptr inbounds nuw (i8, ptr addrspace(3) null, i32 16), align 4
%i1 = load <1 x float>, ptr addrspace(3) getelementptr inbounds nuw (i8, ptr addrspace(3) null, i32 20), align 4
%i2 = shufflevector <1 x float> %i, <1 x float> zeroinitializer, <4 x i32> <i32 0, i32 poison, i32 poison, i32 poison>
call void @llvm.amdgcn.struct.buffer.store.v4f32(<4 x float> %i2, <4 x i32> zeroinitializer, i32 0, i32 0, i32 0, i32 0)
%i3 = shufflevector <1 x float> %i1, <1 x float> zeroinitializer, <4 x i32> <i32 0, i32 poison, i32 poison, i32 poison>
ret <4 x float> %i3
}
; Function Attrs: nocallback nofree nosync nounwind willreturn memory(write)
declare void @llvm.amdgcn.struct.buffer.store.v4f32(<4 x float>, <4 x i32>, i32, i32, i32, i32 immarg) #0

64
llvm/test/CodeGen/AMDGPU/GlobalISel/prelegalizer-combiner-shuffle.mir
View File

@@ -20,10 +20,10 @@ body: |
; CHECK-NEXT: SI_RETURN
%0:_(p3) = COPY $vgpr0
%1:_(p3) = COPY $vgpr1
%12:_(<8 x s16>) = G_IMPLICIT_DEF
%10:_(<8 x s16>) = G_LOAD %0(p3) :: (load (<8 x s16>), align 8, addrspace 3)
%11:_(<4 x s16>) = G_SHUFFLE_VECTOR %10(<8 x s16>), %12, shufflemask(4, 5, 6, 7)
G_STORE %11(<4 x s16>), %1(p3) :: (store (<4 x s16>), addrspace 3)
%2:_(<8 x s16>) = G_IMPLICIT_DEF
%3:_(<8 x s16>) = G_LOAD %0(p3) :: (load (<8 x s16>), align 8, addrspace 3)
%4:_(<4 x s16>) = G_SHUFFLE_VECTOR %3(<8 x s16>), %2, shufflemask(4, 5, 6, 7)
G_STORE %4(<4 x s16>), %1(p3) :: (store (<4 x s16>), addrspace 3)
SI_RETURN
...
@@ -46,10 +46,10 @@ body: |
; CHECK-NEXT: SI_RETURN
%0:_(p3) = COPY $vgpr0
%1:_(p3) = COPY $vgpr1
%12:_(<8 x s16>) = G_IMPLICIT_DEF
%10:_(<8 x s16>) = G_LOAD %0(p3) :: (load (<8 x s16>), align 8, addrspace 3)
%11:_(<2 x s16>) = G_SHUFFLE_VECTOR %10(<8 x s16>), %12, shufflemask(3, 4)
G_STORE %11(<2 x s16>), %1(p3) :: (store (<2 x s16>), addrspace 3)
%2:_(<8 x s16>) = G_IMPLICIT_DEF
%3:_(<8 x s16>) = G_LOAD %0(p3) :: (load (<8 x s16>), align 8, addrspace 3)
%4:_(<2 x s16>) = G_SHUFFLE_VECTOR %3(<8 x s16>), %2, shufflemask(3, 4)
G_STORE %4(<2 x s16>), %1(p3) :: (store (<2 x s16>), addrspace 3)
SI_RETURN
...
@@ -73,10 +73,10 @@ body: |
; CHECK-NEXT: SI_RETURN
%0:_(p3) = COPY $vgpr0
%1:_(p3) = COPY $vgpr1
%12:_(<8 x s16>) = G_IMPLICIT_DEF
%10:_(<8 x s16>) = G_LOAD %0(p3) :: (load (<8 x s16>), align 8, addrspace 3)
%11:_(<3 x s16>) = G_SHUFFLE_VECTOR %10(<8 x s16>), %12, shufflemask(0, 1, 2)
G_STORE %11(<3 x s16>), %1(p3) :: (store (<3 x s16>), addrspace 3)
%2:_(<8 x s16>) = G_IMPLICIT_DEF
%3:_(<8 x s16>) = G_LOAD %0(p3) :: (load (<8 x s16>), align 8, addrspace 3)
%4:_(<3 x s16>) = G_SHUFFLE_VECTOR %3(<8 x s16>), %2, shufflemask(0, 1, 2)
G_STORE %4(<3 x s16>), %1(p3) :: (store (<3 x s16>), addrspace 3)
SI_RETURN
...
@@ -101,10 +101,10 @@ body: |
; CHECK-NEXT: SI_RETURN
%0:_(p3) = COPY $vgpr0
%1:_(p3) = COPY $vgpr1
%12:_(<8 x s16>) = G_IMPLICIT_DEF
%10:_(<8 x s16>) = G_LOAD %0(p3) :: (load (<8 x s16>), align 8, addrspace 3)
%11:_(<4 x s16>) = G_SHUFFLE_VECTOR %10(<8 x s16>), %12, shufflemask(4, 5, -1, 7)
G_STORE %11(<4 x s16>), %1(p3) :: (store (<4 x s16>), addrspace 3)
%2:_(<8 x s16>) = G_IMPLICIT_DEF
%3:_(<8 x s16>) = G_LOAD %0(p3) :: (load (<8 x s16>), align 8, addrspace 3)
%4:_(<4 x s16>) = G_SHUFFLE_VECTOR %3(<8 x s16>), %2, shufflemask(4, 5, -1, 7)
G_STORE %4(<4 x s16>), %1(p3) :: (store (<4 x s16>), addrspace 3)
SI_RETURN
...
@@ -128,10 +128,34 @@ body: |
; CHECK-NEXT: SI_RETURN
%0:_(p3) = COPY $vgpr0
%1:_(p3) = COPY $vgpr1
%12:_(<8 x s16>) = G_IMPLICIT_DEF
%10:_(<8 x s16>) = G_LOAD %0(p3) :: (load (<8 x s16>), align 8, addrspace 3)
%11:_(<4 x s16>) = G_SHUFFLE_VECTOR %10(<8 x s16>), %12, shufflemask(6, 7, 8, 9)
G_STORE %11(<4 x s16>), %1(p3) :: (store (<4 x s16>), addrspace 3)
%2:_(<8 x s16>) = G_IMPLICIT_DEF
%3:_(<8 x s16>) = G_LOAD %0(p3) :: (load (<8 x s16>), align 8, addrspace 3)
%4:_(<4 x s16>) = G_SHUFFLE_VECTOR %3(<8 x s16>), %2, shufflemask(6, 7, 8, 9)
G_STORE %4(<4 x s16>), %1(p3) :: (store (<4 x s16>), addrspace 3)
SI_RETURN
...
---
name: shuffle_vector_to_copy
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1
; CHECK-LABEL: name: shuffle_vector_to_copy
; CHECK: liveins: $vgpr0, $vgpr1
; CHECK-NEXT: {{ $}}
; CHECK-NEXT: [[COPY:%[0-9]+]]:_(p3) = COPY $vgpr0
; CHECK-NEXT: [[COPY1:%[0-9]+]]:_(p3) = COPY $vgpr1
; CHECK-NEXT: [[LOAD:%[0-9]+]]:_(<8 x s16>) = G_LOAD [[COPY]](p3) :: (load (<8 x s16>), align 8, addrspace 3)
; CHECK-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 [[LOAD]](<8 x s16>)
; CHECK-NEXT: G_STORE [[UV4]](s16), [[COPY1]](p3) :: (store (s16), addrspace 3)
; CHECK-NEXT: SI_RETURN
%0:_(p3) = COPY $vgpr0
%1:_(p3) = COPY $vgpr1
%2:_(<8 x s16>) = G_IMPLICIT_DEF
%3:_(<8 x s16>) = G_LOAD %0(p3) :: (load (<8 x s16>), align 8, addrspace 3)
%4:_(s16) = G_SHUFFLE_VECTOR %3(<8 x s16>), %2, shufflemask(4)
G_STORE %4(s16), %1(p3) :: (store (s16), addrspace 3)
SI_RETURN
...