[Mlir-commits] [mlir] [mlir][vector] Fix incorrect byte-alignment assumption in ConvertVectorStore (PR #189235)

[Mehdi Amini]

================
@@ -658,10 +659,17 @@ struct ConvertVectorStore final : OpConversionPattern<vector::StoreOp> {
             stridedMetadata.getConstifiedMixedStrides(),
             getAsOpFoldResult(adaptor.getIndices()));
 
+    // Prefer the exact intraDataOffset when it can be folded (e.g. all-constant
+    // indices). Fall back to 0 only when the trailing dimension exactly matches
+    // the vector size (trailingDimsMatch), because in that case a dynamic last
+    // index implies byte-alignment (the caller is responsible for passing a
+    // valid, aligned index). If neither condition holds, bail out.
----------------
joker-eph wrote:

Right! Thanks. I pushed an update, PTAL?

https://github.com/llvm/llvm-project/pull/189235