Google GitGoogle Git
Sign in
llvm/llvm-project/refs/heads/users/sebpop/pr149977^!/.
commit
7ad8fffa551a9d56a69dc18045afe8ede1cbcfb7
[log]
author
Sebastian Pop <spop@nvidia.com>
Fri Aug 22 18:50:18 2025 -0500
committer
Sebastian Pop <spop@nvidia.com>
Tue Sep 09 14:40:22 2025 -0500
tree
a27684022584412df6db7ca68495b86db4492813
parent
ca28604e723af7940ab764a139fa512f5fa59ebc [diff]
[DA] Fix Strong SIV test for symbolic coefficients and deltas (#149977)

Fixes GitHub issue #149977 where Strong SIV test incorrectly rejected
dependencies with symbolic coefficients and deltas due to overly conservative
bound checking.

Root cause: The bound constraint check |Delta| > UpperBound * |Coeff| would
prematurely reject dependencies when SCEV couldn't prove the relationship
definitively for symbolic expressions, preventing the analysis from reaching
the division logic.

Solution:
1. Make bound check less conservative for symbolic expressions by adding
   runtime assumptions when SCEV cannot determine the relationship.
2. Enable symbolic division using SE->getUDivExactExpr for Delta/Coeff.
3. Add runtime assumptions where symbolic division cannot be computed.

This enables precise dependence analysis for cases like:
- Coefficient: -k (symbolic)
- Delta: -(2*k + 1) (symbolic)
- Distance: (2*k + 1)/k (computed symbolically)

Test case validates:
- When k = -1: distance = 1, clear flow dependence detected.
- Runtime assumptions ensure bounds are satisfied.

diff --git a/llvm/lib/Analysis/DependenceAnalysis.cpp b/llvm/lib/Analysis/DependenceAnalysis.cpp
index ad5415d..d1ae9cd 100644
--- a/llvm/lib/Analysis/DependenceAnalysis.cpp
+++ b/llvm/lib/Analysis/DependenceAnalysis.cpp

@@ -1249,10 +1249,33 @@
         SE->isKnownNonNegative(Coeff) ? Coeff : SE->getNegativeSCEV(Coeff);
     const SCEV *Product = SE->getMulExpr(UpperBound, AbsCoeff);
     if (isKnownPredicate(CmpInst::ICMP_SGT, AbsDelta, Product)) {
-      // Distance greater than trip count - no dependence
-      ++StrongSIVindependence;
-      ++StrongSIVsuccesses;
-      return true;
+      // Check if this involves symbolic expressions where we might be too
+      // conservative.
+      if (isa<SCEVUnknown>(Delta) || isa<SCEVUnknown>(Coeff) ||
+          !isa<SCEVConstant>(AbsDelta) || !isa<SCEVConstant>(Product)) {
+        // For symbolic expressions, add runtime assumption rather than
+        // rejecting.
+        const SCEVPredicate *BoundPred =
+            SE->getComparePredicate(ICmpInst::ICMP_SLE, AbsDelta, Product);
+        if (UnderRuntimeAssumptions) {
+          SmallVector<const SCEVPredicate *, 4> NewPreds(
+              Assumptions.getPredicates());
+          NewPreds.push_back(BoundPred);
+          const_cast<DependenceInfo *>(this)->Assumptions =
+              SCEVUnionPredicate(NewPreds, *SE);
+          LLVM_DEBUG(dbgs() << "\t    Added runtime bound assumption\n");
+        } else {
+          // Cannot add runtime assumptions, let more complex tests try.
+          LLVM_DEBUG(dbgs() << "\t    Would need runtime bound assumption but "
+                               "not allowed. Failing this test.\n");
+          return false;
+        }
+      } else {
+        // Distance definitely greater than trip count - no dependence
+        ++StrongSIVindependence;
+        ++StrongSIVsuccesses;
+        return true;
+      }
     }
   }
 
@@ -1293,9 +1316,40 @@
       Result.DV[Level].Distance = Delta; // since X/1 == X
       NewConstraint.setDistance(Delta, CurLoop);
     } else {
-      Result.Consistent = false;
-      NewConstraint.setLine(Coeff, SE->getNegativeSCEV(Coeff),
-                            SE->getNegativeSCEV(Delta), CurLoop);
+      // Try symbolic division: Distance = Delta / Coeff.
+      if (const SCEV *Distance = SE->getUDivExactExpr(Delta, Coeff)) {
+        LLVM_DEBUG(dbgs() << "\t    Symbolic distance = " << *Distance << "\n");
+        Result.DV[Level].Distance = Distance;
+        NewConstraint.setDistance(Distance, CurLoop);
+      } else {
+        // Cannot compute exact division - check if we can add runtime
+        // assumptions.
+        if (isa<SCEVUnknown>(Coeff) && !SE->isKnownNonZero(Coeff)) {
+          // Add runtime assumption that coefficient is non-zero for division.
+          const SCEV *Zero = SE->getZero(Coeff->getType());
+          const SCEVPredicate *NonZeroPred =
+              SE->getComparePredicate(ICmpInst::ICMP_NE, Coeff, Zero);
+          if (UnderRuntimeAssumptions) {
+            SmallVector<const SCEVPredicate *, 4> NewPreds(
+                Assumptions.getPredicates());
+            NewPreds.push_back(NonZeroPred);
+            const_cast<DependenceInfo *>(this)->Assumptions =
+                SCEVUnionPredicate(NewPreds, *SE);
+            LLVM_DEBUG(dbgs() << "\t    Added runtime assumption: " << *Coeff
+                              << " != 0 for symbolic division\n");
+          } else {
+            // Cannot add runtime assumptions, this test fails.
+            LLVM_DEBUG(dbgs()
+                       << "\t    Would need runtime assumption " << *Coeff
+                       << " != 0 but not allowed. Failing this test.\n");
+            return false;
+          }
+        }
+
+        Result.Consistent = false;
+        NewConstraint.setLine(Coeff, SE->getNegativeSCEV(Coeff),
+                              SE->getNegativeSCEV(Delta), CurLoop);
+      }
     }
 
     // maybe we can get a useful direction

diff --git a/llvm/test/Analysis/DependenceAnalysis/BasePtrBug.ll b/llvm/test/Analysis/DependenceAnalysis/BasePtrBug.ll
index 81e461a..8933c5f 100644
--- a/llvm/test/Analysis/DependenceAnalysis/BasePtrBug.ll
+++ b/llvm/test/Analysis/DependenceAnalysis/BasePtrBug.ll

@@ -18,7 +18,7 @@
 ; CHECK-NEXT:  Src: %0 = load i32, ptr %gep.0, align 4 --> Dst: %0 = load i32, ptr %gep.0, align 4
 ; CHECK-NEXT:    da analyze - none!
 ; CHECK-NEXT:  Src: %0 = load i32, ptr %gep.0, align 4 --> Dst: %1 = load i32, ptr %gep.1, align 4
-; CHECK-NEXT:    da analyze - input [*|<]!
+; CHECK-NEXT:    da analyze - consistent input [((-4 * (sext i32 %div to i64))<nsw> /u 4)|<]!
 ; CHECK-NEXT:  Src: %0 = load i32, ptr %gep.0, align 4 --> Dst: store i32 %add, ptr %gep.B, align 4
 ; CHECK-NEXT:    da analyze - confused!
 ; CHECK-NEXT:  Src: %1 = load i32, ptr %gep.1, align 4 --> Dst: %1 = load i32, ptr %gep.1, align 4

diff --git a/llvm/test/Analysis/DependenceAnalysis/DADelin.ll b/llvm/test/Analysis/DependenceAnalysis/DADelin.ll
index 8f94a45..6e50712 100644
--- a/llvm/test/Analysis/DependenceAnalysis/DADelin.ll
+++ b/llvm/test/Analysis/DependenceAnalysis/DADelin.ll

@@ -648,7 +648,7 @@
 ; CHECK-NEXT:  Src: store i8 42, ptr %idx.0, align 1 --> Dst: store i8 42, ptr %idx.0, align 1
 ; 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:    da analyze - consistent output [((-1 * %k) /u %k)|<]!
 ; 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!
 ;
@@ -687,7 +687,7 @@
 ; CHECK-NEXT:  Src: store i8 42, ptr %idx.0, align 1 --> Dst: store i8 42, ptr %idx.0, align 1
 ; 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:    da analyze - consistent output [((-1 * %k) /u %k)|<]!
 ; 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!
 ;

diff --git a/llvm/test/Analysis/DependenceAnalysis/PR149977.ll b/llvm/test/Analysis/DependenceAnalysis/PR149977.ll
new file mode 100644
index 0000000..1576fe9
--- /dev/null
+++ b/llvm/test/Analysis/DependenceAnalysis/PR149977.ll

@@ -0,0 +1,45 @@
+; 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 #149977: Strong SIV test with symbolic coefficients and deltas
+; The issue was that the bound constraint check was overly conservative with symbolic expressions,
+; causing valid dependencies to be rejected before reaching the division logic.
+;
+; Mathematical analysis:
+; - Access patterns: a[-k*i] vs a[-k*i + (2*k + 1)]
+; - Strong SIV equation: -k*(i2-i1) = (2*k + 1)
+; - Distance: (2*k + 1)/k
+; - For k=-1: distance = -1/-1 = 1 (clear dependence)
+
+define void @f(ptr %a, i64 %k) {
+; CHECK-LABEL: 'f'
+; CHECK-NEXT:  Src: store i8 42, ptr %idx.0, align 1 --> Dst: store i8 42, ptr %idx.0, align 1
+; 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 - consistent output [((-1 + (-2 * %k)) /u (-1 * %k))]!
+; CHECK-NEXT:    Runtime Assumptions:
+; CHECK-NEXT:    Compare predicate: (1 + (2 * %k))<nuw><nsw> sle) (2 * %k)
+; 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!
+;
+entry:
+  %mk = sub i64 0, %k          ; mk = -k
+  %kk = mul i64 %k, 2          ; kk = 2*k
+  %kk.inc = add i64 1, %kk     ; kk.inc = 2*k + 1
+  br label %loop
+
+loop:
+  %i = phi i64 [ 0, %entry ], [ %i.next, %loop ]
+  %subscript.0 = mul i64 %mk, %i                    ; -k * i
+  %subscript.1 = add i64 %subscript.0, %kk.inc     ; -k * i + (2*k + 1)
+  %idx.0 = getelementptr i8, ptr %a, i64 %subscript.0
+  %idx.1 = getelementptr i8, ptr %a, i64 %subscript.1
+  store i8 42, ptr %idx.0
+  store i8 42, ptr %idx.1
+  %i.next = add i64 %i, 1
+  %cond.exit = icmp eq i64 %i.next, 3
+  br i1 %cond.exit, label %exit, label %loop
+
+exit:
+  ret void
+}

diff --git a/llvm/test/Analysis/DependenceAnalysis/StrongSIV.ll b/llvm/test/Analysis/DependenceAnalysis/StrongSIV.ll
index 44bd9b7..b7d2587 100644
--- a/llvm/test/Analysis/DependenceAnalysis/StrongSIV.ll
+++ b/llvm/test/Analysis/DependenceAnalysis/StrongSIV.ll

@@ -385,7 +385,7 @@
 ; CHECK-NEXT:  Src: store i32 %conv, ptr %arrayidx, align 4 --> Dst: store i32 %conv, ptr %arrayidx, align 4
 ; CHECK-NEXT:    da analyze - none!
 ; CHECK-NEXT:  Src: store i32 %conv, ptr %arrayidx, align 4 --> Dst: %0 = load i32, ptr %arrayidx1, align 4
-; CHECK-NEXT:    da analyze - flow [*|<]!
+; CHECK-NEXT:    da analyze - consistent flow [((4 * %n) /u 4)|<]!
 ; CHECK-NEXT:  Src: store i32 %conv, ptr %arrayidx, align 4 --> Dst: store i32 %0, ptr %B.addr.01, align 4
 ; CHECK-NEXT:    da analyze - confused!
 ; CHECK-NEXT:  Src: %0 = load i32, ptr %arrayidx1, align 4 --> Dst: %0 = load i32, ptr %arrayidx1, align 4

diff --git a/llvm/test/Analysis/DependenceAnalysis/SymbolicSIV.ll b/llvm/test/Analysis/DependenceAnalysis/SymbolicSIV.ll
index cdfaec7..4b8a06a 100644
--- a/llvm/test/Analysis/DependenceAnalysis/SymbolicSIV.ll
+++ b/llvm/test/Analysis/DependenceAnalysis/SymbolicSIV.ll

@@ -440,7 +440,7 @@
 ; CHECK-NEXT:  Src: store i32 %conv, ptr %arrayidx, align 4 --> Dst: store i32 %conv, ptr %arrayidx, align 4
 ; CHECK-NEXT:    da analyze - none!
 ; CHECK-NEXT:  Src: store i32 %conv, ptr %arrayidx, align 4 --> Dst: %1 = load i32, ptr %arrayidx6, align 4
-; CHECK-NEXT:    da analyze - flow [<>]!
+; CHECK-NEXT:    da analyze - consistent flow [((-8 + (16 * %M)) /u (8 * %N))]!
 ; CHECK-NEXT:  Src: store i32 %conv, ptr %arrayidx, align 4 --> Dst: store i32 %1, ptr %B.addr.02, align 4
 ; CHECK-NEXT:    da analyze - confused!
 ; CHECK-NEXT:  Src: %1 = load i32, ptr %arrayidx6, align 4 --> Dst: %1 = load i32, ptr %arrayidx6, align 4