Skip to content

[LLVM][MemCpyOpt] Unify alias tags if we optimize allocas #129537

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 10 commits into from
Apr 10, 2025
Merged

[LLVM][MemCpyOpt] Unify alias tags if we optimize allocas #129537

Show file tree
Hide file tree
Changes from 6 commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
4c49512
[LLVM][MemCpyOpt] Unify alias tags if we optimize allocas
DominikAdamski Feb 24, 2025
5e04c39
Simplified test
DominikAdamski Mar 3, 2025
dddcc8a
Fix style
DominikAdamski Mar 18, 2025
ae4139a
Fix style
DominikAdamski Mar 18, 2025
3ab8a71
Applied remarks
DominikAdamski Mar 18, 2025
0824f68
Applied remarks
DominikAdamski Mar 21, 2025
b9cb648
Applied remarks
DominikAdamski Apr 7, 2025
7fc13ac
Applied remarks
DominikAdamski Apr 8, 2025
2f00aa7
Remove alias.scope
DominikAdamski Apr 8, 2025
92584ea
Merge branch 'main' into fix_memcpy_tbaa_tags
DominikAdamski Apr 8, 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
20 changes: 16 additions & 4 deletions llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1516,6 +1516,7 @@ bool MemCpyOptPass::performStackMoveOptzn(Instruction *Load, Instruction *Store,
SmallVector<Instruction *, 4> LifetimeMarkers;
SmallSet<Instruction *, 4> NoAliasInstrs;
bool SrcNotDom = false;
SmallSet<Instruction *, 4> OptimizedAllocaInstUsers;

// Recursively track the user and check whether modified alias exist.
auto IsDereferenceableOrNull = [](Value *V, const DataLayout &DL) -> bool {
Expand All @@ -1524,8 +1525,8 @@ bool MemCpyOptPass::performStackMoveOptzn(Instruction *Load, Instruction *Store,
};

auto CaptureTrackingWithModRef =
[&](Instruction *AI,
function_ref<bool(Instruction *)> ModRefCallback) -> bool {
[&](Instruction *AI, function_ref<bool(Instruction *)> ModRefCallback,
SmallSet<Instruction *, 4> &AllocaInstUsersWithTBAA) -> bool {
Copy link
Contributor

@nikic nikic Apr 6, 2025

Choose a reason for hiding this comment

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

No need for the argument, the variable is captured.

Copy link
Contributor Author

@DominikAdamski DominikAdamski Apr 7, 2025

Choose a reason for hiding this comment

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

Done

SmallVector<Instruction *, 8> Worklist;
Worklist.push_back(AI);
unsigned MaxUsesToExplore = getDefaultMaxUsesToExploreForCaptureTracking();
Expand Down Expand Up @@ -1569,6 +1570,9 @@ bool MemCpyOptPass::performStackMoveOptzn(Instruction *Load, Instruction *Store,
continue;
}
}
if (UI != Store && (UI->hasMetadata(LLVMContext::MD_tbaa) ||
UI->hasMetadata(LLVMContext::MD_tbaa_struct)))
AllocaInstUsersWithTBAA.insert(UI);
if (UI->hasMetadata(LLVMContext::MD_noalias))
NoAliasInstrs.insert(UI);
if (!ModRefCallback(UI))
Expand Down Expand Up @@ -1621,7 +1625,8 @@ bool MemCpyOptPass::performStackMoveOptzn(Instruction *Load, Instruction *Store,
return true;
};

if (!CaptureTrackingWithModRef(DestAlloca, DestModRefCallback))
if (!CaptureTrackingWithModRef(DestAlloca, DestModRefCallback,
OptimizedAllocaInstUsers))
return false;
// Bailout if Dest may have any ModRef before Store.
if (!ReachabilityWorklist.empty() &&
Expand All @@ -1647,7 +1652,8 @@ bool MemCpyOptPass::performStackMoveOptzn(Instruction *Load, Instruction *Store,
return true;
};

if (!CaptureTrackingWithModRef(SrcAlloca, SrcModRefCallback))
if (!CaptureTrackingWithModRef(SrcAlloca, SrcModRefCallback,
OptimizedAllocaInstUsers))
return false;

// We can do the transformation. First, move the SrcAlloca to the start of the
Expand Down Expand Up @@ -1681,6 +1687,12 @@ bool MemCpyOptPass::performStackMoveOptzn(Instruction *Load, Instruction *Store,
for (Instruction *I : NoAliasInstrs)
I->setMetadata(LLVMContext::MD_noalias, nullptr);

// Remove !tbaa and !tbaa_struct from the metadata, since they are invalid.
for (Instruction *I : OptimizedAllocaInstUsers) {
Copy link
Contributor

@nikic nikic Apr 6, 2025

Choose a reason for hiding this comment

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

Can we combine NoAliasInstrs and OptimizedAllocaInstUsers? The purpose is basically the same, I don't think we need separate sets for different metadata.

Copy link
Contributor Author

@DominikAdamski DominikAdamski Apr 7, 2025

Choose a reason for hiding this comment

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

Yes, I combined them.

I->setMetadata(LLVMContext::MD_tbaa, nullptr);
I->setMetadata(LLVMContext::MD_tbaa_struct, nullptr);
}

LLVM_DEBUG(dbgs() << "Stack Move: Performed staack-move optimization\n");
NumStackMove++;
return true;
Expand Down
77 changes: 77 additions & 0 deletions llvm/test/Transforms/MemCpyOpt/memcpy-tbaa.ll
View file
Open in desktop
Copy link
Contributor

@nikic nikic Mar 18, 2025

Choose a reason for hiding this comment

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

This test still doesn't look minimal. Wouldn't it be sufficient to have a store with one tbaa tag, a memcpy and a load with another tag?

Copy link
Contributor Author

@DominikAdamski DominikAdamski Mar 21, 2025

Choose a reason for hiding this comment

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

I removed unnecessary instructions.

Original file line number Diff line number Diff line change
@@ -0,0 +1,77 @@
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
; RUN: opt < %s -passes=memcpyopt,dse -S -verify-memoryssa | FileCheck %s

define void @test() local_unnamed_addr {
; CHECK-LABEL: define void @test() local_unnamed_addr {
; CHECK-NEXT: [[TEST_ARRAY_B:%.*]] = alloca [31 x float], align 4
; CHECK-NEXT: [[TMP1:%.*]] = getelementptr float, ptr [[TEST_ARRAY_B]], i64 1
; CHECK-NEXT: store float 0x3E6AA51880000000, ptr [[TMP1]], align 4
; CHECK-NEXT: [[TMP2:%.*]] = getelementptr float, ptr [[TEST_ARRAY_B]], i64 1
; CHECK-NEXT: [[TMP3:%.*]] = load float, ptr [[TMP2]], align 4
; CHECK-NEXT: ret void
;
%test_array_a = alloca [31 x float], align 4
%test_array_b = alloca [31 x float], align 4
%1 = getelementptr float, ptr %test_array_b, i64 1
store float 0x3E6AA51880000000, ptr %1, align 4, !tbaa !4
call void @llvm.memcpy.p0.p0.i64(ptr noundef nonnull align 4 dereferenceable(124) %test_array_a, ptr noundef nonnull align 4 dereferenceable(124) %test_array_b, i64 124, i1 false)
%2 = getelementptr float, ptr %test_array_a, i64 1
%3 = load float, ptr %2, align 4, !tbaa !7
ret void
}

%struct.Outer = type { float, double, %struct.Inner }
%struct.Inner = type { i32, float }

; Function Attrs: nounwind uwtable
define dso_local float @f() {
; CHECK-LABEL: define dso_local float @f() {
; CHECK-NEXT: [[ENTRY:.*:]]
; CHECK-NEXT: [[TEST1:%.*]] = alloca [[STRUCT_OUTER:%.*]], align 8
; CHECK-NEXT: [[F:%.*]] = getelementptr inbounds nuw [[STRUCT_OUTER]], ptr [[TEST1]], i32 0, i32 0
; CHECK-NEXT: store float 0.000000e+00, ptr [[F]], align 8
; CHECK-NEXT: [[F1:%.*]] = getelementptr inbounds nuw [[STRUCT_OUTER]], ptr [[TEST1]], i32 0, i32 0
; CHECK-NEXT: [[TMP0:%.*]] = load float, ptr [[F1]], align 8
; CHECK-NEXT: [[ADD:%.*]] = fadd float [[TMP0]], 2.000000e+00
; CHECK-NEXT: store float [[ADD]], ptr [[F1]], align 8
; CHECK-NEXT: [[F2:%.*]] = getelementptr inbounds nuw [[STRUCT_OUTER]], ptr [[TEST1]], i32 0, i32 0
; CHECK-NEXT: [[TMP1:%.*]] = load float, ptr [[F2]], align 8
; CHECK-NEXT: ret float [[TMP1]]
;
entry:
%test = alloca %struct.Outer, align 8
%test1 = alloca %struct.Outer, align 8
%f = getelementptr inbounds nuw %struct.Outer, ptr %test1, i32 0, i32 0
store float 0.000000e+00, ptr %f, align 8, !tbaa !9
%inner_a = getelementptr inbounds nuw %struct.Outer, ptr %test1, i32 0, i32 2
%i = getelementptr inbounds nuw %struct.Inner, ptr %inner_a, i32 0, i32 0
store i32 0, ptr %i, align 8, !tbaa !17
call void @llvm.memcpy.p0.p0.i64(ptr align 8 %test, ptr align 8 %test1, i64 24, i1 false)
%f1 = getelementptr inbounds nuw %struct.Outer, ptr %test, i32 0, i32 0
%0 = load float, ptr %f1, align 8, !tbaa !9
%add = fadd float %0, 2.000000e+00
store float %add, ptr %f1, align 8, !tbaa !9
%f2 = getelementptr inbounds nuw %struct.Outer, ptr %test, i32 0, i32 0
%1 = load float, ptr %f2, align 8, !tbaa !9
ret float %1
}

!1 = !{!"any data access", !2, i64 0}
!2 = !{!"any access", !3, i64 0}
!3 = !{!"Flang function root test"}
!4 = !{!5, !5, i64 0}
!5 = !{!"allocated data/test_array_a", !6, i64 0}
!6 = !{!"allocated data", !1, i64 0}
!7 = !{!8, !8, i64 0}
!8 = !{!"allocated data/test_array_b", !6, i64 0}
!9 = !{!10, !11, i64 0}
!10 = !{!"Outer", !11, i64 0, !14, i64 8, !15, i64 16}
!11 = !{!"float", !12, i64 0}
!12 = !{!"omnipotent char", !13, i64 0}
!13 = !{!"Simple C/C++ TBAA"}
!14 = !{!"double", !12, i64 0}
!15 = !{!"Inner", !16, i64 0, !11, i64 4}
!16 = !{!"int", !12, i64 0}
!17 = !{!10, !16, i64 16}