[llvm] [VPlan] Use pointer index type for stride in convertToStridedAccesses (PR #199647)

Mel Chen via llvm-commits llvm-commits at lists.llvm.org
Mon Jun 1 02:09:13 PDT 2026


================
@@ -7030,13 +7031,20 @@ void VPlanTransforms::convertToStridedAccesses(VPlan &Plan,
 
       VPBuilder Builder(LoadR);
       // Create the base pointer of strided access.
+      // TODO: reuse VPDerivedIVRecipe for base pointer computation when it
+      // supports a general VPValue as the start value.
       VPValue *StartVPV = vputils::getOrCreateVPValueForSCEVExpr(Plan, Start);
-      VPValue *StrideInBytes =
-          Plan.getConstantInt(VectorLoop->getCanonicalIVType(),
-                              Step->getSExtValue(), /*IsSigned=*/true);
+      VPValue *StrideInBytes = Plan.getOrAddLiveIn(Step->getValue());
+      Type *StrideTy = TypeInfo.inferScalarType(StrideInBytes);
+      assert(StrideTy == PSE.getSE()->getDataLayout().getIndexType(
+                             TypeInfo.inferScalarType(Ptr)) &&
+             "Stride type from SCEV must match the index type");
+      VPValue *CanIV = Builder.createScalarZExtOrTrunc(
----------------
Mel-Chen wrote:

I've also considered this point, especially since `VPDerivedIVRecipe` uses `sext` rather than `zext`. However, a canonical IV is a non-negative integer I think. If the canonical IV gets large enough in some iteration to have its most significant bit set to 1, using sext would treat it as a negative number, yielding a completely different value than what was intended. Or is there an implicit guarantee that a canonical IV's most significant bit will never be 1, making this a non-issue?

https://github.com/llvm/llvm-project/pull/199647


More information about the llvm-commits mailing list