Skip to content

[VectorCombine] Preserve scoped alias metadata #153714

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 9 commits into from
Aug 18, 2025
Merged

[VectorCombine] Preserve scoped alias metadata #153714

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
25cbc66
Make VectorCombine Pass Alias Info
knwng Aug 15, 2025
3c3d2b8
Merge remote-tracking branch 'upstream/main' into fix_vector_combine
knwng Aug 15, 2025
adce747
add test; address comments
knwng Aug 15, 2025
2148978
address comments
knwng Aug 16, 2025
6574cf4
fix lint
knwng Aug 17, 2025
9f4a0a0
fix lint and test
knwng Aug 17, 2025
8dd4467
fix lint
knwng Aug 17, 2025
91f0c8d
Merge remote-tracking branch 'upstream/main' into fix_vector_combine
knwng Aug 18, 2025
c2a88cf
Merge remote-tracking branch 'upstream/main' into fix_vector_combine
knwng Aug 18, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
16 changes: 12 additions & 4 deletions llvm/lib/Transforms/Vectorize/VectorCombine.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1812,6 +1812,8 @@ bool VectorCombine::scalarizeLoadExtract(Instruction &I) {
// erased in the correct order.
Worklist.push(LI);

Type *ElemType = VecTy->getElementType();

// Replace extracts with narrow scalar loads.
for (User *U : LI->users()) {
auto *EI = cast<ExtractElementInst>(U);
Expand All @@ -1825,13 +1827,19 @@ bool VectorCombine::scalarizeLoadExtract(Instruction &I) {
Builder.SetInsertPoint(EI);
Value *GEP =
Builder.CreateInBoundsGEP(VecTy, Ptr, {Builder.getInt32(0), Idx});
auto *NewLoad = cast<LoadInst>(Builder.CreateLoad(
VecTy->getElementType(), GEP, EI->getName() + ".scalar"));
auto *NewLoad = cast<LoadInst>(
Builder.CreateLoad(ElemType, GEP, EI->getName() + ".scalar"));

Align ScalarOpAlignment = computeAlignmentAfterScalarization(
LI->getAlign(), VecTy->getElementType(), Idx, *DL);
Align ScalarOpAlignment =
computeAlignmentAfterScalarization(LI->getAlign(), ElemType, Idx, *DL);
NewLoad->setAlignment(ScalarOpAlignment);

if (auto *ConstIdx = dyn_cast<ConstantInt>(Idx)) {
size_t Offset = ConstIdx->getZExtValue() * DL->getTypeStoreSize(ElemType);
AAMDNodes OldAAMD = LI->getAAMetadata();
NewLoad->setAAMetadata(OldAAMD.adjustForAccess(Offset, ElemType, *DL));
}
Copy link
Contributor

@nikic nikic Aug 15, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If possible, it would be better to use getAAMetadata() and adjustForAccess() here.

Copy link
Contributor Author

@knwng knwng Aug 15, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hi @nikic, thank you for the advice. I know it's a more thoughtful solution. But I'm not familiar with AA in LLVM. Can you guide me on how to use adjustForAccess(), like how to correctly calculate offset and type? Thanks!

Copy link
Contributor

@nikic nikic Aug 16, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The type is VecTy->getElementType(). The offset would be Idx * the type size like here:

llvm-project/llvm/lib/Transforms/Vectorize/VectorCombine.cpp

Lines 1652 to 1654 in adce747

if (auto *C = dyn_cast<ConstantInt>(Idx))
return commonAlignment(VectorAlignment,
C->getZExtValue() * DL.getTypeStoreSize(ScalarType));

The problem is if Idx is a variable. In that case the offset is unknown. So I would make the whole logic conditional on the dyn_cast<ConstantInt>(Idx) case.


replaceValue(*EI, *NewLoad, false);
}

Expand Down
56 changes: 56 additions & 0 deletions llvm/test/Transforms/VectorCombine/alias.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,56 @@
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
; RUN: opt < %s -passes=vector-combine -S | FileCheck %s --check-prefixes=CHECK

define <4 x i32> @quux(ptr addrspace(3) %arg) {
; CHECK-LABEL: define <4 x i32> @quux(
; CHECK-SAME: ptr addrspace(3) [[ARG:%.*]]) {
; CHECK-NEXT: [[BB:.*:]]
; CHECK-NEXT: [[EXTRACTELEMENT:%.*]] = load i8, ptr addrspace(3) [[ARG]], align 4, !tbaa [[TBAA0:![0-9]+]], !alias.scope [[META0:![0-9]+]], !noalias [[META0]]
; CHECK-NEXT: [[TMP0:%.*]] = getelementptr inbounds <4 x i8>, ptr addrspace(3) [[ARG]], i32 0, i64 1
; CHECK-NEXT: [[EXTRACTELEMENT1:%.*]] = load i8, ptr addrspace(3) [[TMP0]], align 1, !tbaa [[TBAA0]], !alias.scope [[META0]], !noalias [[META0]]
; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds <4 x i8>, ptr addrspace(3) [[ARG]], i32 0, i64 2
; CHECK-NEXT: [[EXTRACTELEMENT2:%.*]] = load i8, ptr addrspace(3) [[TMP1]], align 2, !tbaa [[TBAA0]], !alias.scope [[META0]], !noalias [[META0]]
; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds <4 x i8>, ptr addrspace(3) [[ARG]], i32 0, i64 3
; CHECK-NEXT: [[EXTRACTELEMENT3:%.*]] = load i8, ptr addrspace(3) [[TMP2]], align 1, !tbaa [[TBAA0]], !alias.scope [[META0]], !noalias [[META0]]
; CHECK-NEXT: [[ZEXT:%.*]] = zext i8 [[EXTRACTELEMENT]] to i32
; CHECK-NEXT: [[ZEXT4:%.*]] = zext i8 [[EXTRACTELEMENT1]] to i32
; CHECK-NEXT: [[ZEXT5:%.*]] = zext i8 [[EXTRACTELEMENT2]] to i32
; CHECK-NEXT: [[ZEXT6:%.*]] = zext i8 [[EXTRACTELEMENT3]] to i32
; CHECK-NEXT: [[INSERTELEMENT:%.*]] = insertelement <4 x i32> poison, i32 [[ZEXT]], i64 0
; CHECK-NEXT: [[INSERTELEMENT7:%.*]] = insertelement <4 x i32> [[INSERTELEMENT]], i32 [[ZEXT4]], i64 1
; CHECK-NEXT: [[INSERTELEMENT8:%.*]] = insertelement <4 x i32> [[INSERTELEMENT7]], i32 [[ZEXT5]], i64 2
; CHECK-NEXT: [[INSERTELEMENT9:%.*]] = insertelement <4 x i32> [[INSERTELEMENT8]], i32 [[ZEXT6]], i64 3
; CHECK-NEXT: ret <4 x i32> [[INSERTELEMENT9]]
;
bb:
%load = load <4 x i8>, ptr addrspace(3) %arg, align 4, !alias.scope !0, !noalias !0, !tbaa !5
%extractelement = extractelement <4 x i8> %load, i64 0
%extractelement1 = extractelement <4 x i8> %load, i64 1
%extractelement2 = extractelement <4 x i8> %load, i64 2
%extractelement3 = extractelement <4 x i8> %load, i64 3
%zext = zext i8 %extractelement to i32
%zext4 = zext i8 %extractelement1 to i32
%zext5 = zext i8 %extractelement2 to i32
%zext6 = zext i8 %extractelement3 to i32
%insertelement = insertelement <4 x i32> poison, i32 %zext, i64 0
%insertelement7 = insertelement <4 x i32> %insertelement, i32 %zext4, i64 1
%insertelement8 = insertelement <4 x i32> %insertelement7, i32 %zext5, i64 2
%insertelement9 = insertelement <4 x i32> %insertelement8, i32 %zext6, i64 3
ret <4 x i32> %insertelement9
}

!0 = !{!1}
!1 = distinct !{!1, !2}
!2 = distinct !{!2}
!3 = !{!"Simple C/C++ TBAA"}
!4 = !{!"omnipotent char", !3, i64 0}
!5 = !{!"i8", !4, i64 0}
;.
; CHECK: [[TBAA0]] = !{[[META3:![0-9]+]], [[META3]], i64 0, i64 0}
; CHECK: [[META3]] = !{!"i8", [[META4:![0-9]+]]}
; CHECK: [[META4]] = !{!"omnipotent char", [[META5:![0-9]+]], i64 0}
; CHECK: [[META5]] = !{!"Simple C/C++ TBAA"}
; CHECK: [[META0]] = !{[[META1:![0-9]+]]}
; CHECK: [[META1]] = distinct !{[[META1]], [[META2:![0-9]+]]}
; CHECK: [[META2]] = distinct !{[[META2]]}
;.