[DA] Fix zero coeff bug in Strong SIV test with runtime assumptions (#149991) #155037
+148
−31
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
llvmbot
added
the
llvm:analysis
Member
|
@llvm/pr-subscribers-llvm-analysis Author: Sebastian Pop (sebpop) ChangesFix GitHub issue #149991 where Strong SIV test incorrectly concludes 'none!' for symbolic coefficients that could be zero, leading to 0/0 undef behavior. The issue occurs in subscripts like {base,+,coeff} where coeff is symbolic:
The Strong SIV test's Delta=0 case assumed 0/X=0 where X is the coefficient, but when X could be zero, we have 0/0 which is undefined. The analysis needs to be conservative when the coefficient might be zero. When coefficient is SCEVUnknown and cannot be proven non-zero at compile time, use SCEV range analysis to attempt proving coefficient > 0. If this fails, add a runtime assumption 'coeff > 0' to the dependence result. This allows precise analysis when possible (none! under assumption coeff > 0) while maintaining correctness by exposing the required assumption. Test cases:
Full diff: https://github.com/llvm/llvm-project/pull/155037.diff 3 Files Affected:
diff --git a/llvm/lib/Analysis/DependenceAnalysis.cpp b/llvm/lib/Analysis/DependenceAnalysis.cpp
index f33e04e804e3d..50e0fb7b3bcd3 100644
--- a/llvm/lib/Analysis/DependenceAnalysis.cpp
+++ b/llvm/lib/Analysis/DependenceAnalysis.cpp
@@ -1282,7 +1282,28 @@ bool DependenceInfo::strongSIVtest(const SCEV *Coeff, const SCEV *SrcConst,
Result.DV[Level].Direction &= Dependence::DVEntry::EQ;
++StrongSIVsuccesses;
} else if (Delta->isZero()) {
- // since 0/X == 0
+ // Check if coefficient could be zero. If so, 0/0 is undefined and we
+ // cannot conclude that only same-iteration dependencies exist.
+ // When coeff=0, all iterations access the same location.
+ if (isa<SCEVUnknown>(Coeff) && !SE->isKnownNonZero(Coeff)) {
+ // Use SCEV range analysis to prove coefficient > 0 in loop context.
+ const SCEV *Zero = SE->getZero(Coeff->getType());
+
+ // Ask SCEV's range analysis if it can prove Coeff > Zero
+ if (SE->isKnownPredicate(ICmpInst::ICMP_SGT, Coeff, Zero)) {
+ LLVM_DEBUG(
+ dbgs()
+ << "\t Coefficient proven positive by SCEV range analysis\n");
+ } else {
+ // Cannot prove at compile time, add runtime assumption
+ const SCEVPredicate *Pred =
+ SE->getComparePredicate(ICmpInst::ICMP_SGT, Coeff, Zero);
+ const_cast<DependenceInfo *>(this)->Assumptions.push_back(Pred);
+ LLVM_DEBUG(dbgs() << "\t Added runtime assumption: " << *Coeff
+ << " > 0\n");
+ }
+ }
+ // since 0/X == 0 (where X is known non-zero)
Result.DV[Level].Distance = Delta;
NewConstraint.setDistance(Delta, CurLoop);
Result.DV[Level].Direction &= Dependence::DVEntry::EQ;
@@ -3574,8 +3595,8 @@ DependenceInfo::depends(Instruction *Src, Instruction *Dst,
if (!isLoadOrStore(Src) || !isLoadOrStore(Dst)) {
// can only analyze simple loads and stores, i.e., no calls, invokes, etc.
LLVM_DEBUG(dbgs() << "can only handle simple loads and stores\n");
- return std::make_unique<Dependence>(Src, Dst,
- SCEVUnionPredicate(Assume, *SE));
+ return std::make_unique<Dependence>(
+ Src, Dst, SCEVUnionPredicate(this->Assumptions, *SE));
}
const MemoryLocation &DstLoc = MemoryLocation::get(Dst);
@@ -3586,8 +3607,8 @@ DependenceInfo::depends(Instruction *Src, Instruction *Dst,
case AliasResult::PartialAlias:
// cannot analyse objects if we don't understand their aliasing.
LLVM_DEBUG(dbgs() << "can't analyze may or partial alias\n");
- return std::make_unique<Dependence>(Src, Dst,
- SCEVUnionPredicate(Assume, *SE));
+ return std::make_unique<Dependence>(
+ Src, Dst, SCEVUnionPredicate(this->Assumptions, *SE));
case AliasResult::NoAlias:
// If the objects noalias, they are distinct, accesses are independent.
LLVM_DEBUG(dbgs() << "no alias\n");
@@ -3601,8 +3622,8 @@ DependenceInfo::depends(Instruction *Src, Instruction *Dst,
// The dependence test gets confused if the size of the memory accesses
// differ.
LLVM_DEBUG(dbgs() << "can't analyze must alias with different sizes\n");
- return std::make_unique<Dependence>(Src, Dst,
- SCEVUnionPredicate(Assume, *SE));
+ return std::make_unique<Dependence>(
+ Src, Dst, SCEVUnionPredicate(this->Assumptions, *SE));
}
Value *SrcPtr = getLoadStorePointerOperand(Src);
@@ -3621,8 +3642,8 @@ DependenceInfo::depends(Instruction *Src, Instruction *Dst,
// We check this upfront so we don't crash in cases where getMinusSCEV()
// returns a SCEVCouldNotCompute.
LLVM_DEBUG(dbgs() << "can't analyze SCEV with different pointer base\n");
- return std::make_unique<Dependence>(Src, Dst,
- SCEVUnionPredicate(Assume, *SE));
+ return std::make_unique<Dependence>(
+ Src, Dst, SCEVUnionPredicate(this->Assumptions, *SE));
}
// Even if the base pointers are the same, they may not be loop-invariant. It
@@ -3634,8 +3655,8 @@ DependenceInfo::depends(Instruction *Src, Instruction *Dst,
if (!isLoopInvariant(SrcBase, SrcLoop) ||
!isLoopInvariant(DstBase, DstLoop)) {
LLVM_DEBUG(dbgs() << "The base pointer is not loop invariant.\n");
- return std::make_unique<Dependence>(Src, Dst,
- SCEVUnionPredicate(Assume, *SE));
+ return std::make_unique<Dependence>(
+ Src, Dst, SCEVUnionPredicate(this->Assumptions, *SE));
}
uint64_t EltSize = SrcLoc.Size.toRaw();
@@ -3646,8 +3667,8 @@ DependenceInfo::depends(Instruction *Src, Instruction *Dst,
if (!SE->isKnownMultipleOf(SrcEv, EltSize, Assume) ||
!SE->isKnownMultipleOf(DstEv, EltSize, Assume)) {
LLVM_DEBUG(dbgs() << "can't analyze SCEV with different offsets\n");
- return std::make_unique<Dependence>(Src, Dst,
- SCEVUnionPredicate(Assume, *SE));
+ return std::make_unique<Dependence>(
+ Src, Dst, SCEVUnionPredicate(this->Assumptions, *SE));
}
if (!Assume.empty()) {
@@ -3670,7 +3691,7 @@ DependenceInfo::depends(Instruction *Src, Instruction *Dst,
LLVM_DEBUG(dbgs() << " common nesting levels = " << CommonLevels << "\n");
LLVM_DEBUG(dbgs() << " maximum nesting levels = " << MaxLevels << "\n");
- FullDependence Result(Src, Dst, SCEVUnionPredicate(Assume, *SE),
+ FullDependence Result(Src, Dst, SCEVUnionPredicate(this->Assumptions, *SE),
PossiblyLoopIndependent, CommonLevels);
++TotalArrayPairs;
diff --git a/llvm/test/Analysis/DependenceAnalysis/DADelin.ll b/llvm/test/Analysis/DependenceAnalysis/DADelin.ll
index 8f94a455d3724..c7b9865329214 100644
--- a/llvm/test/Analysis/DependenceAnalysis/DADelin.ll
+++ b/llvm/test/Analysis/DependenceAnalysis/DADelin.ll
@@ -649,8 +649,13 @@ define void @coeff_may_negative(ptr %a, i32 %k) {
; CHECK-NEXT: da analyze - none!
; CHECK-NEXT: Src: store i8 42, ptr %idx.0, align 1 --> Dst: store i8 42, ptr %idx.1, align 1
; CHECK-NEXT: da analyze - output [*|<]!
+; CHECK-NEXT: Runtime Assumptions:
+; CHECK-NEXT: Compare predicate: %k sgt) 0
; CHECK-NEXT: Src: store i8 42, ptr %idx.1, align 1 --> Dst: store i8 42, ptr %idx.1, align 1
; CHECK-NEXT: da analyze - none!
+; CHECK-NEXT: Runtime Assumptions:
+; CHECK-NEXT: Compare predicate: %k sgt) 0
+; CHECK-NEXT: Compare predicate: %k sgt) 0
;
entry:
br label %loop
@@ -688,8 +693,13 @@ define void @coeff_positive(ptr %a, i32 %k) {
; CHECK-NEXT: da analyze - none!
; CHECK-NEXT: Src: store i8 42, ptr %idx.0, align 1 --> Dst: store i8 42, ptr %idx.1, align 1
; CHECK-NEXT: da analyze - output [*|<]!
+; CHECK-NEXT: Runtime Assumptions:
+; CHECK-NEXT: Compare predicate: %k sgt) 0
; CHECK-NEXT: Src: store i8 42, ptr %idx.1, align 1 --> Dst: store i8 42, ptr %idx.1, align 1
; CHECK-NEXT: da analyze - none!
+; CHECK-NEXT: Runtime Assumptions:
+; CHECK-NEXT: Compare predicate: %k sgt) 0
+; CHECK-NEXT: Compare predicate: %k sgt) 0
;
entry:
br label %loop
diff --git a/llvm/test/Analysis/DependenceAnalysis/zero-coefficient.ll b/llvm/test/Analysis/DependenceAnalysis/zero-coefficient.ll
new file mode 100644
index 0000000000000..d6ad924da722b
--- /dev/null
+++ b/llvm/test/Analysis/DependenceAnalysis/zero-coefficient.ll
@@ -0,0 +1,30 @@
+; NOTE: Assertions have been autogenerated by utils/update_analyze_test_checks.py UTC_ARGS: --version 5
+; RUN: opt < %s -disable-output "-passes=print<da>" 2>&1 | FileCheck %s
+
+; Test case for GitHub issue #149991: Strong SIV test with symbolic coefficient
+; that could be zero. Fixed using runtime assumptions: assume coefficient > 0.
+
+target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
+
+define void @test_zero_coefficient(ptr %a, i64 %k) {
+; CHECK-LABEL: 'test_zero_coefficient'
+; CHECK-NEXT: Src: store i8 42, ptr %idx, align 1 --> Dst: store i8 42, ptr %idx, align 1
+; CHECK-NEXT: da analyze - none!
+; CHECK-NEXT: Runtime Assumptions:
+; CHECK-NEXT: Compare predicate: %k sgt) 0
+;
+entry:
+ br label %loop
+
+loop:
+ %i = phi i64 [ 0, %entry ], [ %i.next, %loop ]
+ %subscript = mul i64 %i, %k ; When %k=0, all iterations access %a[0]
+ %idx = getelementptr i8, ptr %a, i64 %subscript
+ store i8 42, ptr %idx
+ %i.next = add i64 %i, 1
+ %cond.exit = icmp eq i64 %i.next, 100
+ br i1 %cond.exit, label %exit, label %loop
+
+exit:
+ ret void
+}
|
Meinersbur
reviewed
Aug 25, 2025
sebpop
force-pushed
the
pr149991
branch
3 times, most recently
from
22d4899 to
f7365db
Compare
sebpop
changed the title
Meinersbur
reviewed
Nov 4, 2025
sjoerdmeijer
approved these changes
Nov 27, 2025
Collaborator
sjoerdmeijer
left a comment
sjoerdmeijer
left a comment
There was a problem hiding this comment.
LGTM, provided that @Meinersbur previous minor comments have been addressed, which I think is the case.
…tions The Strong SIV test was incorrectly concluding "no dependence" when the coefficient is symbolic and the delta (difference between source and destination) is zero. When delta=0, the Strong SIV test divides delta/coeff to get the distance. The bug occurs when coeff is an unknown symbolic value: if coeff=0 at runtime, then 0/0 is undefined and all iterations access the same memory location, creating a true dependence that was being missed.
kasuga-fj
reviewed
Nov 28, 2025
Comment on lines
+3711
to
+3713
| bool AddedAssumptionsDuringTests = | ||
| Result.Assumptions.getPredicates().size() > AssumptionsBeforeTests; | ||
| if (AllEqual && !AddedAssumptionsDuringTests) |
Contributor
There was a problem hiding this comment.
Why is it allowed to ignore the assumptions added before tests?
Contributor
Author
There was a problem hiding this comment.
Currently we return nullptr when dependence vector is "AllEqual".
This code is here to guard against returning nullptr when the dependence is conditional to runtime assumptions: if we return nullptr the assumptions would be discarded.
We may need to revisit what we return for "no dependence" because we can't attach assumptions to a "nullptr".
Contributor
There was a problem hiding this comment.
I think the condition should be Result.Assumptions.getPredicates().empty(), not !AddedAssumptionsDuringTests. If I understand correctly, current implementation silently discards the assumptions added before tests.
| // Check if coefficient could be zero. If so, 0/0 is undefined and we | ||
| // cannot conclude that only same-iteration dependencies exist. | ||
| // When coeff=0, all iterations access the same location. | ||
| if (isa<SCEVUnknown>(Coeff)) { |
Contributor
There was a problem hiding this comment.
Why only SCEVUnknown? What happens if Coeff is like %m * %n? I think this may also be zero.
Contributor
Author
There was a problem hiding this comment.
I rewrote the code to avoid enumerating all possible expressions that Coeff can take at this location:
if (isa<SCEVConstant>(Coeff) || SE->isKnownNonZero(Coeff)) {
LLVM_DEBUG(
dbgs() << "\t Coefficient proven non-zero by SCEV analysis\n");
} else {
// Cannot prove at compile time, would need runtime assumption.
[...]
}
Contributor
There was a problem hiding this comment.
isa<SCEVConstant>(Coeff) would be redundant.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
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.
Fix GitHub issue #149991 where Strong SIV test incorrectly concludes 'none!' for symbolic coefficients that could be zero, leading to 0/0 undef behavior.
The Strong SIV test was incorrectly concluding "no dependence" when the
coefficient is symbolic and the delta (difference between source and
destination) is zero.
When delta=0, the Strong SIV test divides delta/coeff to get the distance.
The bug occurs when coeff is an unknown symbolic value: if coeff=0 at runtime,
then 0/0 is undefined and all iterations access the same memory location,
creating a true dependence that was being missed.