[llvm] [LV] Convert gather loads with invariant stride into strided loads (PR #147297)

[Mel Chen via llvm-commits]

================
@@ -4204,3 +4212,184 @@ void VPlanTransforms::addBranchWeightToMiddleTerminator(
       MDB.createBranchWeights({1, VectorStep - 1}, /*IsExpected=*/false);
   MiddleTerm->addMetadata(LLVMContext::MD_prof, BranchWeights);
 }
+
+static std::pair<VPValue *, VPValue *> matchStridedStart(VPValue *CurIndex) {
+  // TODO: Support VPWidenPointerInductionRecipe.
+  if (auto *WidenIV = dyn_cast<VPWidenIntOrFpInductionRecipe>(CurIndex))
+    return {WidenIV, WidenIV->getStepValue()};
+
+  auto *WidenR = dyn_cast<VPWidenRecipe>(CurIndex);
+  if (!WidenR || !CurIndex->getUnderlyingValue())
----------------
Mel-Chen wrote:

I’ve changed it like this for now.
206436986b57f1275315cff5457de6bf210906f5
We need UnderlyingInstr here to narrow all base address recipes into scalar recipes. As discussed earlier, we don’t want to rely on the legacy cost model, so we also can’t depend on its scalarize analysis. The functionality needed here is actually similar to narrowToSingleScalarRecipes; in both cases, we require UnderlyingInstr in order to create a VPReplicateRecipe.

https://github.com/llvm/llvm-project/pull/147297