[llvm] f572a59 - [VectorCombine] Ensure canScalarizeAccess handles cases where the index type can't represent all inbounds values

[Simon Pilgrim via llvm-commits]

Author: Simon Pilgrim
Date: 2025-04-24T14:17:55+01:00
New Revision: f572a5951a664d57e909928d5595285212ad6884

URL: https://github.com/llvm/llvm-project/commit/f572a5951a664d57e909928d5595285212ad6884
DIFF: https://github.com/llvm/llvm-project/commit/f572a5951a664d57e909928d5595285212ad6884.diff

LOG: [VectorCombine] Ensure canScalarizeAccess handles cases where the index type can't represent all inbounds values

Fixes #132563

Added: 
    llvm/test/Transforms/VectorCombine/pr132563.ll

Modified: 
    llvm/lib/Transforms/Vectorize/VectorCombine.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Transforms/Vectorize/VectorCombine.cpp b/llvm/lib/Transforms/Vectorize/VectorCombine.cpp
index bd225bcc0635e..04c084ffdda97 100644
--- a/llvm/lib/Transforms/Vectorize/VectorCombine.cpp
+++ b/llvm/lib/Transforms/Vectorize/VectorCombine.cpp
@@ -1438,6 +1438,7 @@ static ScalarizationResult canScalarizeAccess(VectorType *VecTy, Value *Idx,
   // This is the number of elements of fixed vector types,
   // or the minimum number of elements of scalable vector types.
   uint64_t NumElements = VecTy->getElementCount().getKnownMinValue();
+  unsigned IntWidth = Idx->getType()->getScalarSizeInBits();
 
   if (auto *C = dyn_cast<ConstantInt>(Idx)) {
     if (C->getValue().ult(NumElements))
@@ -1445,7 +1446,10 @@ static ScalarizationResult canScalarizeAccess(VectorType *VecTy, Value *Idx,
     return ScalarizationResult::unsafe();
   }
 
-  unsigned IntWidth = Idx->getType()->getScalarSizeInBits();
+  // Always unsafe if the index type can't handle all inbound values.
+  if (!llvm::isUIntN(IntWidth, NumElements))
+    return ScalarizationResult::unsafe();
+
   APInt Zero(IntWidth, 0);
   APInt MaxElts(IntWidth, NumElements);
   ConstantRange ValidIndices(Zero, MaxElts);

diff  --git a/llvm/test/Transforms/VectorCombine/pr132563.ll b/llvm/test/Transforms/VectorCombine/pr132563.ll
new file mode 100644
index 0000000000000..d0f1ebc46780e
--- /dev/null
+++ b/llvm/test/Transforms/VectorCombine/pr132563.ll
@@ -0,0 +1,30 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -passes=vector-combine -S %s | FileCheck %s
+
+; Ensure canScalarizeAccess handles cases where the index type can't represent all inbounds values
+
+define void @src_1_idx(ptr %q, i8 zeroext %s, i1 %idx) {
+; CHECK-LABEL: @src_1_idx(
+; CHECK-NEXT:    [[LD:%.*]] = load <16 x i8>, ptr [[Q:%.*]], align 16
+; CHECK-NEXT:    [[V1:%.*]] = insertelement <16 x i8> [[LD]], i8 [[S:%.*]], i1 [[IDX:%.*]]
+; CHECK-NEXT:    store <16 x i8> [[V1]], ptr [[Q]], align 16
+; CHECK-NEXT:    ret void
+;
+  %ld = load <16 x i8>, ptr %q
+  %v1 = insertelement <16 x i8> %ld, i8 %s, i1 %idx
+  store <16 x i8> %v1, ptr %q
+  ret void
+}
+
+define void @src_2_idx(ptr %q, i8 zeroext %s, i8 %idx) {
+; CHECK-LABEL: @src_2_idx(
+; CHECK-NEXT:    [[LD:%.*]] = load <256 x i8>, ptr [[Q:%.*]], align 256
+; CHECK-NEXT:    [[V1:%.*]] = insertelement <256 x i8> [[LD]], i8 [[S:%.*]], i8 [[IDX:%.*]]
+; CHECK-NEXT:    store <256 x i8> [[V1]], ptr [[Q]], align 256
+; CHECK-NEXT:    ret void
+;
+  %ld = load <256 x i8>, ptr %q
+  %v1 = insertelement <256 x i8> %ld, i8 %s, i8 %idx
+  store <256 x i8> %v1, ptr %q
+  ret void
+}