[Mlir-commits] [mlir] [mlir][vector] Canonicalize transfer_{read, write} vector<1xT> (PR #196598)

[Erick Ochoa Lopez]

================
@@ -5625,11 +5625,54 @@ struct TransferReadAfterWriteToBroadcast
     return success();
   }
 };
+/// Canonicalize a rank-1, single-element vector transfer so that its
+/// permutation map is the minor identity.
+///
+/// When the vector type is `vector<1xT>`, the permutation map is irrelevant
+/// to which element is accessed: the single vector lane has iteration offset 0,
+/// so the element is always at `indices` regardless of which source dimension
+/// the map points at. Replacing the map with the minor identity unblocks
+/// lowering to vector.load / vector.store.
+template <typename TransferOp>
+struct CanonicalizeSize1TransferPermutationMap final
+    : OpRewritePattern<TransferOp> {
+  using OpRewritePattern<TransferOp>::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(TransferOp op,
+                                PatternRewriter &rewriter) const override {
+    VectorType vecType = op.getVectorType();
+    if (vecType.getRank() != 1 || vecType.getShape()[0] != 1)
----------------
amd-eochoalo wrote:

You are right! Thanks! https://github.com/llvm/llvm-project/pull/196598/commits/0bd5e3310b015acd164f86b72f165883c4b96281

https://github.com/llvm/llvm-project/pull/196598