[llvm] [LV] isScalarWithPredication to check for interleaved accesses (PR #194579)

Gaƫtan Bossu via llvm-commits llvm-commits at lists.llvm.org
Thu Apr 30 02:24:44 PDT 2026


================
@@ -2465,8 +2465,13 @@ bool LoopVectorizationCostModel::isScalarWithPredication(Instruction *I,
     return getCallWideningDecision(cast<CallInst>(I), VF).Kind == CM_Scalarize;
   case Instruction::Load:
   case Instruction::Store: {
-    return !Config.isLegalMaskedLoadOrStore(I, VF) &&
-           !Config.isLegalGatherOrScatter(I, VF);
+    LoopVectorizationCostModel::InstWidening WidenKind =
+        getWideningDecision(I, VF);
+    assert(WidenKind != CM_Unknown);
+    // According to the legacy cost model, a masked access is never considered
+    // profitable for scalarization when gather/scatter is available, even if
+    // the chosen widening is CM_Scalarize.
+    return WidenKind == CM_Scalarize && !Config.isLegalGatherOrScatter(I, VF);
----------------
gbossu wrote:

Ultimately, I think this should just be `return WidenKind == CM_Scalarize;`, but I'd like this patch to focus on fixing inconsistencies, not changing the legacy cost model.

https://github.com/llvm/llvm-project/pull/194579


More information about the llvm-commits mailing list