Skip to content

Commit 064f02d

Browse files
authored
[VectorCombine] Preserve scoped alias metadata (#153714)
Right now if a load op is scalarized, the `!alias.scope` and `!noalias` metadata are dropped. This PR is to keep them if exist.
1 parent 8d25673 commit 064f02d

File tree

2 files changed

+68
-4
lines changed

2 files changed

+68
-4
lines changed

llvm/lib/Transforms/Vectorize/VectorCombine.cpp

Lines changed: 12 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -1812,6 +1812,8 @@ bool VectorCombine::scalarizeLoadExtract(Instruction &I) {
18121812
// erased in the correct order.
18131813
Worklist.push(LI);
18141814

1815+
Type *ElemType = VecTy->getElementType();
1816+
18151817
// Replace extracts with narrow scalar loads.
18161818
for (User *U : LI->users()) {
18171819
auto *EI = cast<ExtractElementInst>(U);
@@ -1825,13 +1827,19 @@ bool VectorCombine::scalarizeLoadExtract(Instruction &I) {
18251827
Builder.SetInsertPoint(EI);
18261828
Value *GEP =
18271829
Builder.CreateInBoundsGEP(VecTy, Ptr, {Builder.getInt32(0), Idx});
1828-
auto *NewLoad = cast<LoadInst>(Builder.CreateLoad(
1829-
VecTy->getElementType(), GEP, EI->getName() + ".scalar"));
1830+
auto *NewLoad = cast<LoadInst>(
1831+
Builder.CreateLoad(ElemType, GEP, EI->getName() + ".scalar"));
18301832

1831-
Align ScalarOpAlignment = computeAlignmentAfterScalarization(
1832-
LI->getAlign(), VecTy->getElementType(), Idx, *DL);
1833+
Align ScalarOpAlignment =
1834+
computeAlignmentAfterScalarization(LI->getAlign(), ElemType, Idx, *DL);
18331835
NewLoad->setAlignment(ScalarOpAlignment);
18341836

1837+
if (auto *ConstIdx = dyn_cast<ConstantInt>(Idx)) {
1838+
size_t Offset = ConstIdx->getZExtValue() * DL->getTypeStoreSize(ElemType);
1839+
AAMDNodes OldAAMD = LI->getAAMetadata();
1840+
NewLoad->setAAMetadata(OldAAMD.adjustForAccess(Offset, ElemType, *DL));
1841+
}
1842+
18351843
replaceValue(*EI, *NewLoad, false);
18361844
}
18371845

Lines changed: 56 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,56 @@
1+
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
2+
; RUN: opt < %s -passes=vector-combine -S | FileCheck %s --check-prefixes=CHECK
3+
4+
define <4 x i32> @quux(ptr addrspace(3) %arg) {
5+
; CHECK-LABEL: define <4 x i32> @quux(
6+
; CHECK-SAME: ptr addrspace(3) [[ARG:%.*]]) {
7+
; CHECK-NEXT: [[BB:.*:]]
8+
; CHECK-NEXT: [[EXTRACTELEMENT:%.*]] = load i8, ptr addrspace(3) [[ARG]], align 4, !tbaa [[TBAA0:![0-9]+]], !alias.scope [[META0:![0-9]+]], !noalias [[META0]]
9+
; CHECK-NEXT: [[TMP0:%.*]] = getelementptr inbounds <4 x i8>, ptr addrspace(3) [[ARG]], i32 0, i64 1
10+
; CHECK-NEXT: [[EXTRACTELEMENT1:%.*]] = load i8, ptr addrspace(3) [[TMP0]], align 1, !tbaa [[TBAA0]], !alias.scope [[META0]], !noalias [[META0]]
11+
; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds <4 x i8>, ptr addrspace(3) [[ARG]], i32 0, i64 2
12+
; CHECK-NEXT: [[EXTRACTELEMENT2:%.*]] = load i8, ptr addrspace(3) [[TMP1]], align 2, !tbaa [[TBAA0]], !alias.scope [[META0]], !noalias [[META0]]
13+
; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds <4 x i8>, ptr addrspace(3) [[ARG]], i32 0, i64 3
14+
; CHECK-NEXT: [[EXTRACTELEMENT3:%.*]] = load i8, ptr addrspace(3) [[TMP2]], align 1, !tbaa [[TBAA0]], !alias.scope [[META0]], !noalias [[META0]]
15+
; CHECK-NEXT: [[ZEXT:%.*]] = zext i8 [[EXTRACTELEMENT]] to i32
16+
; CHECK-NEXT: [[ZEXT4:%.*]] = zext i8 [[EXTRACTELEMENT1]] to i32
17+
; CHECK-NEXT: [[ZEXT5:%.*]] = zext i8 [[EXTRACTELEMENT2]] to i32
18+
; CHECK-NEXT: [[ZEXT6:%.*]] = zext i8 [[EXTRACTELEMENT3]] to i32
19+
; CHECK-NEXT: [[INSERTELEMENT:%.*]] = insertelement <4 x i32> poison, i32 [[ZEXT]], i64 0
20+
; CHECK-NEXT: [[INSERTELEMENT7:%.*]] = insertelement <4 x i32> [[INSERTELEMENT]], i32 [[ZEXT4]], i64 1
21+
; CHECK-NEXT: [[INSERTELEMENT8:%.*]] = insertelement <4 x i32> [[INSERTELEMENT7]], i32 [[ZEXT5]], i64 2
22+
; CHECK-NEXT: [[INSERTELEMENT9:%.*]] = insertelement <4 x i32> [[INSERTELEMENT8]], i32 [[ZEXT6]], i64 3
23+
; CHECK-NEXT: ret <4 x i32> [[INSERTELEMENT9]]
24+
;
25+
bb:
26+
%load = load <4 x i8>, ptr addrspace(3) %arg, align 4, !alias.scope !0, !noalias !0, !tbaa !5
27+
%extractelement = extractelement <4 x i8> %load, i64 0
28+
%extractelement1 = extractelement <4 x i8> %load, i64 1
29+
%extractelement2 = extractelement <4 x i8> %load, i64 2
30+
%extractelement3 = extractelement <4 x i8> %load, i64 3
31+
%zext = zext i8 %extractelement to i32
32+
%zext4 = zext i8 %extractelement1 to i32
33+
%zext5 = zext i8 %extractelement2 to i32
34+
%zext6 = zext i8 %extractelement3 to i32
35+
%insertelement = insertelement <4 x i32> poison, i32 %zext, i64 0
36+
%insertelement7 = insertelement <4 x i32> %insertelement, i32 %zext4, i64 1
37+
%insertelement8 = insertelement <4 x i32> %insertelement7, i32 %zext5, i64 2
38+
%insertelement9 = insertelement <4 x i32> %insertelement8, i32 %zext6, i64 3
39+
ret <4 x i32> %insertelement9
40+
}
41+
42+
!0 = !{!1}
43+
!1 = distinct !{!1, !2}
44+
!2 = distinct !{!2}
45+
!3 = !{!"Simple C/C++ TBAA"}
46+
!4 = !{!"omnipotent char", !3, i64 0}
47+
!5 = !{!"i8", !4, i64 0}
48+
;.
49+
; CHECK: [[TBAA0]] = !{[[META3:![0-9]+]], [[META3]], i64 0, i64 0}
50+
; CHECK: [[META3]] = !{!"i8", [[META4:![0-9]+]]}
51+
; CHECK: [[META4]] = !{!"omnipotent char", [[META5:![0-9]+]], i64 0}
52+
; CHECK: [[META5]] = !{!"Simple C/C++ TBAA"}
53+
; CHECK: [[META0]] = !{[[META1:![0-9]+]]}
54+
; CHECK: [[META1]] = distinct !{[[META1]], [[META2:![0-9]+]]}
55+
; CHECK: [[META2]] = distinct !{[[META2]]}
56+
;.

0 commit comments

Comments
 (0)