[Mem2Reg] Don't use single store optimization for potentially poison value #97711
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…value If there is a single store, then loads must either load the stored value or uninitialized memory (undef). If the stored value may be poison, then replacing an uninitialized memory load with it would be incorrect. Fall back to the generic code in that case. This PR only fixes the case where there is a literal poison store -- the case where the value is non-trivially poison will be miscompiled by phi simplification later, see llvm#96631. Fixes llvm#97702.
|
@llvm/pr-subscribers-llvm-transforms Author: Nikita Popov (nikic) ChangesIf there is a single store, then loads must either load the stored value or uninitialized memory (undef). If the stored value may be poison, then replacing an uninitialized memory load with it would be incorrect. Fall back to the generic code in that case. This PR only fixes the case where there is a literal poison store -- the case where the value is non-trivially poison will still get miscompiled by phi simplification later, see #96631. Fixes #97702. Full diff: https://github.com/llvm/llvm-project/pull/97711.diff 2 Files Affected:
diff --git a/llvm/lib/Transforms/Utils/PromoteMemoryToRegister.cpp b/llvm/lib/Transforms/Utils/PromoteMemoryToRegister.cpp
index 6e021a5e2d05a..cd5ab55c2122f 100644
--- a/llvm/lib/Transforms/Utils/PromoteMemoryToRegister.cpp
+++ b/llvm/lib/Transforms/Utils/PromoteMemoryToRegister.cpp
@@ -525,7 +525,14 @@ rewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info, LargeBlockInfo &LBI,
SmallSet<DbgAssignIntrinsic *, 8> *DbgAssignsToDelete,
SmallSet<DbgVariableRecord *, 8> *DVRAssignsToDelete) {
StoreInst *OnlyStore = Info.OnlyStore;
- bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
+ Value *ReplVal = OnlyStore->getOperand(0);
+ // Loads may either load the stored value or uninitialized memory (undef).
+ // If the stored value may be poison, then replacing an uninitialized memory
+ // load with it would be incorrect.
+ if (!isGuaranteedNotToBePoison(ReplVal))
+ return false;
+
+ bool StoringGlobalVal = !isa<Instruction>(ReplVal);
BasicBlock *StoreBB = OnlyStore->getParent();
int StoreIndex = -1;
@@ -565,7 +572,6 @@ rewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info, LargeBlockInfo &LBI,
}
// Otherwise, we *can* safely rewrite this load.
- Value *ReplVal = OnlyStore->getOperand(0);
// If the replacement value is the load, this must occur in unreachable
// code.
if (ReplVal == LI)
diff --git a/llvm/test/Transforms/Mem2Reg/single-store.ll b/llvm/test/Transforms/Mem2Reg/single-store.ll
index b82e26158a361..f864227c49145 100644
--- a/llvm/test/Transforms/Mem2Reg/single-store.ll
+++ b/llvm/test/Transforms/Mem2Reg/single-store.ll
@@ -1,7 +1,6 @@
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
; RUN: opt -S -passes=mem2reg < %s | FileCheck %s
-; FIXME: This is a miscompile.
define i8 @single_store_literal_poison(i1 %cond) {
; CHECK-LABEL: define i8 @single_store_literal_poison(
; CHECK-SAME: i1 [[COND:%.*]]) {
@@ -9,7 +8,7 @@ define i8 @single_store_literal_poison(i1 %cond) {
; CHECK: [[IF]]:
; CHECK-NEXT: br label %[[EXIT]]
; CHECK: [[EXIT]]:
-; CHECK-NEXT: ret i8 poison
+; CHECK-NEXT: ret i8 undef
;
%a = alloca i8, align 1
br i1 %cond, label %if, label %exit
|
dianqk
pushed a commit
to dianqk/llvm-project
that referenced
this pull request
Jul 5, 2024
…value (llvm#97711) If there is a single store, then loads must either load the stored value or uninitialized memory (undef). If the stored value may be poison, then replacing an uninitialized memory load with it would be incorrect. Fall back to the generic code in that case. This PR only fixes the case where there is a literal poison store -- the case where the value is non-trivially poison will still get miscompiled by phi simplification later, see llvm#96631. Fixes llvm#97702. (cherry picked from commit f58930f)
nikic
added a commit
that referenced
this pull request
Jul 5, 2024
In #97711 the single-store optimization was disabled for the case where the value is potentially poison, as this may produce incorrect results for loads of uninitialized memory. However, this resulted in compile-time regressions. Address these by still allowing the single-store optimization to occur in cases where the store dominates the load, as we know that such a load will always read initialized memory.
dianqk
pushed a commit
to dianqk/llvm-project
that referenced
this pull request
Jul 5, 2024
In llvm#97711 the single-store optimization was disabled for the case where the value is potentially poison, as this may produce incorrect results for loads of uninitialized memory. However, this resulted in compile-time regressions. Address these by still allowing the single-store optimization to occur in cases where the store dominates the load, as we know that such a load will always read initialized memory. (cherry picked from commit daaea12)
nikic
added a commit
to rust-lang/llvm-project
that referenced
this pull request
Jul 5, 2024
…value (llvm#97711) If there is a single store, then loads must either load the stored value or uninitialized memory (undef). If the stored value may be poison, then replacing an uninitialized memory load with it would be incorrect. Fall back to the generic code in that case. This PR only fixes the case where there is a literal poison store -- the case where the value is non-trivially poison will still get miscompiled by phi simplification later, see llvm#96631. Fixes llvm#97702. (cherry picked from commit f58930f)
nikic
added a commit
to rust-lang/llvm-project
that referenced
this pull request
Jul 5, 2024
In llvm#97711 the single-store optimization was disabled for the case where the value is potentially poison, as this may produce incorrect results for loads of uninitialized memory. However, this resulted in compile-time regressions. Address these by still allowing the single-store optimization to occur in cases where the store dominates the load, as we know that such a load will always read initialized memory. (cherry picked from commit daaea12)
kbluck
pushed a commit
to kbluck/llvm-project
that referenced
this pull request
Jul 6, 2024
…value (llvm#97711) If there is a single store, then loads must either load the stored value or uninitialized memory (undef). If the stored value may be poison, then replacing an uninitialized memory load with it would be incorrect. Fall back to the generic code in that case. This PR only fixes the case where there is a literal poison store -- the case where the value is non-trivially poison will still get miscompiled by phi simplification later, see llvm#96631. Fixes llvm#97702.
kbluck
pushed a commit
to kbluck/llvm-project
that referenced
this pull request
Jul 6, 2024
In llvm#97711 the single-store optimization was disabled for the case where the value is potentially poison, as this may produce incorrect results for loads of uninitialized memory. However, this resulted in compile-time regressions. Address these by still allowing the single-store optimization to occur in cases where the store dominates the load, as we know that such a load will always read initialized memory.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
If there is a single store, then loads must either load the stored value or uninitialized memory (undef). If the stored value may be poison, then replacing an uninitialized memory load with it would be incorrect. Fall back to the generic code in that case.
This PR only fixes the case where there is a literal poison store -- the case where the value is non-trivially poison will still get miscompiled by phi simplification later, see #96631.
Fixes #97702.