Skip to content

[InstCombine] Propagate invariant.load metadata across unpacked loads #152186

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.

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 7 commits into from
Aug 14, 2025
Merged
Show file tree
Hide file tree
Changes from 2 commits
Commits
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
10 changes: 10 additions & 0 deletions llvm/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -718,6 +718,14 @@ static Instruction *combineLoadToOperationType(InstCombinerImpl &IC,
return nullptr;
}

// Check if the aggregate load has a invariant.load metadata
// If aggregate load has invariant.load metadata, add it to the
// unpacked loads as well.
static void copyInvariantLoadMetadata(LoadInst &LI, LoadInst *NewLoad) {
if (MDNode *MD = LI.getMetadata("invariant.load"))
Copy link
Contributor

Choose a reason for hiding this comment

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

Should access by MD_ enum instead of by name. Also isn't this what copyMetadataForLoad is for?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

I don't think that function will correctly propagate metadata for unpacking loads. For example, it copies MD_tbaa as is. But the tbaa nodes would change for scalar loads vs aggregate loads.

Copy link
Contributor Author

Choose a reason for hiding this comment

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

@arsenm now using enums instead of by name.

Copy link
Contributor

Choose a reason for hiding this comment

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

The point is still that most of these metadata should propagate, this isn't unique to invariant.load

Copy link
Member

Choose a reason for hiding this comment

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

We can use copyMetadataForLoad.

Copy link

Choose a reason for hiding this comment

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

copyMetadataForLoad sounds like only designed for load type casting. And some code that checks NewType->isPointerTy() implicitly assumes the OldType is also a pointer, which sounds incorrect for spliting loads.

    // Note, essentially every kind of metadata should be preserved here! This
    // routine is supposed to clone a load instruction changing *only its type*.
    // The only metadata it makes sense to drop is metadata which is invalidated
    // when the pointer type changes. This should essentially never be the case
    // in LLVM, but we explicitly switch over only known metadata to be
    // conservatively correct. If you are adding metadata to LLVM which pertains
    // to loads, you almost certainly want to add it here.

Shall we add a copyMetadataForSplitLoad?

Copy link
Member

Choose a reason for hiding this comment

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

copyMetadataForLoad sounds like only designed for load type casting.

You are right. If we have to repeat it multiple times in the future, it is definitely better to have a helper function.

NewLoad->setMetadata("invariant.load", MD);
}

static Instruction *unpackLoadToAggregate(InstCombinerImpl &IC, LoadInst &LI) {
// FIXME: We could probably with some care handle both volatile and atomic
// stores here but it isn't clear that this is important.
Expand All @@ -737,6 +745,7 @@ static Instruction *unpackLoadToAggregate(InstCombinerImpl &IC, LoadInst &LI) {
LoadInst *NewLoad = IC.combineLoadToNewType(LI, ST->getTypeAtIndex(0U),
".unpack");
NewLoad->setAAMetadata(LI.getAAMetadata());
copyInvariantLoadMetadata(LI, NewLoad);
Copy link
Member

Choose a reason for hiding this comment

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

Suggested change
copyInvariantLoadMetadata(LI, NewLoad);
NewLI->copyMetadata(LI, LLVMContext::MD_invariant_load);

return IC.replaceInstUsesWith(LI, IC.Builder.CreateInsertValue(
PoisonValue::get(T), NewLoad, 0, Name));
}
Expand Down Expand Up @@ -764,6 +773,7 @@ static Instruction *unpackLoadToAggregate(InstCombinerImpl &IC, LoadInst &LI) {
Name + ".unpack");
// Propagate AA metadata. It'll still be valid on the narrowed load.
L->setAAMetadata(LI.getAAMetadata());
copyInvariantLoadMetadata(LI, L);
Copy link
Member

Choose a reason for hiding this comment

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

Suggested change
copyInvariantLoadMetadata(LI, L);
NewLI->copyMetadata(LI, LLVMContext::MD_invariant_load);

Copy link
Contributor Author

@VedantParanjape VedantParanjape Aug 13, 2025

Choose a reason for hiding this comment

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

@dtcxzyw I have fixed these issues, could you please give a lgtm? I will create another PR for adjusting AA metadata.

V = IC.Builder.CreateInsertValue(V, L, i);
}

Expand Down
46 changes: 46 additions & 0 deletions llvm/test/Transforms/InstCombine/invariant-metadata-propagation.ll
Original file line number Diff line number Diff line change
@@ -0,0 +1,46 @@
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
; RUN: opt -S < %s -passes=instcombine | FileCheck %s

%struct.double2 = type { double, double }
%struct.double1 = type { double }

define %struct.double2 @func1(ptr addrspace(1) %a) {
; CHECK-LABEL: define %struct.double2 @func1(
; CHECK-SAME: ptr addrspace(1) [[A:%.*]]) {
; CHECK-NEXT: [[DOTUNPACK:%.*]] = load double, ptr addrspace(1) [[A]], align 16, !invariant.load [[META0:![0-9]+]]
; CHECK-NEXT: [[TMP1:%.*]] = insertvalue [[STRUCT_DOUBLE2:%.*]] poison, double [[DOTUNPACK]], 0
; CHECK-NEXT: [[DOTELT1:%.*]] = getelementptr inbounds nuw i8, ptr addrspace(1) [[A]], i64 8
; CHECK-NEXT: [[DOTUNPACK2:%.*]] = load double, ptr addrspace(1) [[DOTELT1]], align 8, !invariant.load [[META0]]
; CHECK-NEXT: [[TMP2:%.*]] = insertvalue [[STRUCT_DOUBLE2]] [[TMP1]], double [[DOTUNPACK2]], 1
; CHECK-NEXT: ret [[STRUCT_DOUBLE2]] [[TMP2]]
;
%1 = load %struct.double2, ptr addrspace(1) %a, align 16, !invariant.load !1
ret %struct.double2 %1
}

define %struct.double2 @func2(ptr %a) {
; CHECK-LABEL: define %struct.double2 @func2(
; CHECK-SAME: ptr [[A:%.*]]) {
; CHECK-NEXT: [[DOTUNPACK:%.*]] = load double, ptr [[A]], align 16, !invariant.load [[META0]]
; CHECK-NEXT: [[TMP1:%.*]] = insertvalue [[STRUCT_DOUBLE2:%.*]] poison, double [[DOTUNPACK]], 0
; CHECK-NEXT: [[DOTELT1:%.*]] = getelementptr inbounds nuw i8, ptr [[A]], i64 8
; CHECK-NEXT: [[DOTUNPACK2:%.*]] = load double, ptr [[DOTELT1]], align 8, !invariant.load [[META0]]
; CHECK-NEXT: [[TMP2:%.*]] = insertvalue [[STRUCT_DOUBLE2]] [[TMP1]], double [[DOTUNPACK2]], 1
; CHECK-NEXT: ret [[STRUCT_DOUBLE2]] [[TMP2]]
;
%1 = load %struct.double2, ptr %a, align 16, !invariant.load !1
ret %struct.double2 %1
}

define %struct.double1 @func3(ptr %a) {
; CHECK-LABEL: define %struct.double1 @func3(
; CHECK-SAME: ptr [[A:%.*]]) {
; CHECK-NEXT: [[DOTUNPACK:%.*]] = load double, ptr [[A]], align 16, !invariant.load [[META0]]
; CHECK-NEXT: [[TMP1:%.*]] = insertvalue [[STRUCT_DOUBLE1:%.*]] poison, double [[DOTUNPACK]], 0
; CHECK-NEXT: ret [[STRUCT_DOUBLE1]] [[TMP1]]
;
%1 = load %struct.double1, ptr %a, align 16, !invariant.load !1
ret %struct.double1 %1
}

!1 = !{}