[Mlir-commits] [mlir] [mlir][vector] Drop innermost unit dims on transfer_write. (PR #78554)

[Andrzej Warzyński]

================
@@ -1152,8 +1152,71 @@ struct FoldI1Select : public OpRewritePattern<arith::SelectOp> {
   }
 };
 
-// Drop inner most contiguous unit dimensions from transfer_read operand.
-class DropInnerMostUnitDims : public OpRewritePattern<vector::TransferReadOp> {
+/// Returns the number of dims can be folded away from transfer ops. It returns
+/// a failure if strides and offsets can not be resolved.
+static FailureOr<size_t>
+getTransferFoldableInnerUnitDims(MemRefType srcType, VectorType vectorType) {
+  SmallVector<int64_t> srcStrides;
+  int64_t srcOffset;
+  if (failed(getStridesAndOffset(srcType, srcStrides, srcOffset)))
+    return failure();
+
+  // According to vector.transfer_read/write semantics, the vector can be a
+  // slice. It pads the indices with `1` starting from beginning. Thus, we have
+  // to offset the check index with `rankDiff` in `srcStrides` and source dim
+  // sizes.
----------------
banach-space wrote:

IIUC, both `vector.transfer_read` and `vector.transfer_write` require all indices to be specified. See https://github.com/llvm/llvm-project/pull/78608

https://github.com/llvm/llvm-project/pull/78554