[llvm] [LAA] Allow vectorizing `A[NonZeroNonConstantStride*I] += 1` (PR #186262)

[Andrei Elovikov via llvm-commits]

================
@@ -54,7 +54,12 @@ static std::optional<int64_t> getConstantStride(VPValue *Addr, Type *AccessTy,
   if (!AddRec)
     return {};
 
-  return getStrideFromAddRec(AddRec, L, AccessTy, /*Ptr=*/nullptr, PSE);
+  const auto *Stride = dyn_cast_or_null<SCEVConstant>(
+      getStrideFromAddRec(AddRec, L, AccessTy, /*Ptr=*/nullptr, PSE));
+  if (!Stride)
+    return {};
+
+  return Stride->getAPInt().trySExtValue();
----------------
eas wrote:

I'm not sure how valuable that would be. My idea is to uplift as much as possible (ideally everything) to operate on SCEV strides instead of constant strides...

https://github.com/llvm/llvm-project/pull/186262