[llvm] [VPlan] Use SCEV to prove non-aliasing for stores at different offsets. (PR #170347)

Ramkumar Ramachandra via llvm-commits llvm-commits at lists.llvm.org
Tue Dec 2 11:55:35 PST 2025


================
@@ -139,14 +139,54 @@ bool VPlanTransforms::tryToConvertVPInstructionsToVPRecipes(
   return true;
 }
 
+// Return true if \p A and \p B are known to not alias for all VFs in the plan,
+// checked via the distance between the accesses
+static bool isNoAliasViaDistance(VPReplicateRecipe *A, VPReplicateRecipe *B,
+                                 ScalarEvolution &SE, const Loop &L,
+                                 VPTypeAnalysis &TypeInfo) {
+  if (A->getOpcode() != Instruction::Store ||
+      B->getOpcode() != Instruction::Store)
+    return false;
+
+  VPValue *AddrA = A->getOperand(1);
+  const SCEV *SCEVA = vputils::getSCEVExprForVPValue(AddrA, SE, &L);
+  VPValue *AddrB = B->getOperand(1);
+  const SCEV *SCEVB = vputils::getSCEVExprForVPValue(AddrB, SE, &L);
+  if (isa<SCEVCouldNotCompute>(SCEVA) || isa<SCEVCouldNotCompute>(SCEVB))
+    return false;
+
+  const SCEV *Distance = SE.getMinusSCEV(SCEVA, SCEVB);
+  auto *ConstDist = dyn_cast<SCEVConstant>(Distance);
+  if (!ConstDist)
+    return false;
----------------
artagnon wrote:

```suggestion
  const APInt *ConstDist;
  if (!match(Distance, m_scev_APInt(ConstDist)))
    return false;
```

... and strip the subsequent DistValue?

https://github.com/llvm/llvm-project/pull/170347


More information about the llvm-commits mailing list