[llvm] 4d71f20 - [GlobalISel] prevent G_UNMERGE_VALUES for vectors with different elements (#133335)

[via llvm-commits]

Author: Robert Imschweiler
Date: 2025-06-18T09:07:08+02:00
New Revision: 4d71f20b287e398f10bbff55d52bec9683ef89d2

URL: https://github.com/llvm/llvm-project/commit/4d71f20b287e398f10bbff55d52bec9683ef89d2
DIFF: https://github.com/llvm/llvm-project/commit/4d71f20b287e398f10bbff55d52bec9683ef89d2.diff

LOG: [GlobalISel] prevent G_UNMERGE_VALUES for vectors with different elements (#133335)

This commit prevents building a G_UNMERGE_VALUES instruction with
different source and destination vector elements in
`LegalizationArtifactCombiner::ArtifactValueFinder::tryCombineMergeLike()`,
e.g.:
`%1:_(<2 x s8>), %2:_(<2 x s8>) = G_UNMERGE_VALUES %0:_(<2 x s16>)`

This LLVM defect was identified via the AMD Fuzzing project.

Added: 
    

Modified: 
    llvm/include/llvm/CodeGen/GlobalISel/LegalizationArtifactCombiner.h
    llvm/test/CodeGen/AMDGPU/GlobalISel/insertelement.ll

Removed: 
    


################################################################################
diff  --git a/llvm/include/llvm/CodeGen/GlobalISel/LegalizationArtifactCombiner.h b/llvm/include/llvm/CodeGen/GlobalISel/LegalizationArtifactCombiner.h
index 22f6a5fde546a..8f560c42082f9 100644
--- a/llvm/include/llvm/CodeGen/GlobalISel/LegalizationArtifactCombiner.h
+++ b/llvm/include/llvm/CodeGen/GlobalISel/LegalizationArtifactCombiner.h
@@ -997,6 +997,7 @@ class LegalizationArtifactCombiner {
 
       // Recognize UnmergeSrc that can be unmerged to DstTy directly.
       // Types have to be either both vector or both non-vector types.
+      // In case of vector types, the scalar elements need to match.
       // Merge-like opcodes are combined one at the time. First one creates new
       // unmerge, following should use the same unmerge (builder performs CSE).
       //
@@ -1005,7 +1006,9 @@ class LegalizationArtifactCombiner {
       // %AnotherDst:_(DstTy) = G_merge_like_opcode %2:_(EltTy), %3
       //
       // %Dst:_(DstTy), %AnotherDst = G_UNMERGE_VALUES %UnmergeSrc
-      if ((DstTy.isVector() == UnmergeSrcTy.isVector()) &&
+      if (((!DstTy.isVector() && !UnmergeSrcTy.isVector()) ||
+           (DstTy.isVector() && UnmergeSrcTy.isVector() &&
+            DstTy.getScalarType() == UnmergeSrcTy.getScalarType())) &&
           (Elt0UnmergeIdx % NumMIElts == 0) &&
           getCoverTy(UnmergeSrcTy, DstTy) == UnmergeSrcTy) {
         if (!isSequenceFromUnmerge(MI, 0, Unmerge, Elt0UnmergeIdx, NumMIElts,

diff  --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/insertelement.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/insertelement.ll
index 8134eb3ca2afc..132a89478c5fd 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/insertelement.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/insertelement.ll
@@ -6506,3 +6506,58 @@ entry:
   %insert = insertelement <5 x double> %vec, double %val, i32 %idx
   ret <5 x double> %insert
 }
+
+; Found by fuzzer, reduced with llvm-reduce.
+define amdgpu_kernel void @insert_very_small_from_very_large(<32 x i16> %L3, ptr %ptr) {
+; GPRIDX-LABEL: insert_very_small_from_very_large:
+; GPRIDX:       ; %bb.0: ; %bb
+; GPRIDX-NEXT:    s_load_dwordx16 s[12:27], s[8:9], 0x0
+; GPRIDX-NEXT:    s_load_dwordx2 s[0:1], s[8:9], 0x40
+; GPRIDX-NEXT:    s_waitcnt lgkmcnt(0)
+; GPRIDX-NEXT:    s_lshr_b32 s2, s12, 1
+; GPRIDX-NEXT:    s_and_b32 s2, s2, 1
+; GPRIDX-NEXT:    s_lshl_b32 s2, s2, 1
+; GPRIDX-NEXT:    v_mov_b32_e32 v0, s0
+; GPRIDX-NEXT:    v_mov_b32_e32 v2, s2
+; GPRIDX-NEXT:    v_mov_b32_e32 v1, s1
+; GPRIDX-NEXT:    flat_store_byte v[0:1], v2
+; GPRIDX-NEXT:    s_endpgm
+;
+; GFX10-LABEL: insert_very_small_from_very_large:
+; GFX10:       ; %bb.0: ; %bb
+; GFX10-NEXT:    s_clause 0x1
+; GFX10-NEXT:    s_load_dwordx16 s[12:27], s[8:9], 0x0
+; GFX10-NEXT:    s_load_dwordx2 s[0:1], s[8:9], 0x40
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_lshr_b32 s2, s12, 1
+; GFX10-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-NEXT:    s_and_b32 s2, s2, 1
+; GFX10-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-NEXT:    s_lshl_b32 s2, s2, 1
+; GFX10-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-NEXT:    flat_store_byte v[0:1], v2
+; GFX10-NEXT:    s_endpgm
+;
+; GFX11-LABEL: insert_very_small_from_very_large:
+; GFX11:       ; %bb.0: ; %bb
+; GFX11-NEXT:    s_clause 0x1
+; GFX11-NEXT:    s_load_b512 s[8:23], s[4:5], 0x0
+; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x40
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_lshr_b32 s2, s8, 1
+; GFX11-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-NEXT:    s_and_b32 s2, s2, 1
+; GFX11-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 1
+; GFX11-NEXT:    v_mov_b32_e32 v2, s2
+; GFX11-NEXT:    flat_store_b8 v[0:1], v2
+; GFX11-NEXT:    s_endpgm
+bb:
+  %a = bitcast <32 x i16> %L3 to i512
+  %b = trunc i512 %a to i8
+  %c = trunc i8 %b to i2
+  %d = bitcast i2 %c to <2 x i1>
+  %insert = insertelement <2 x i1> %d, i1 false, i32 0
+  store <2 x i1> %insert, ptr %ptr, align 1
+  ret void
+}