[Mlir-commits] [mlir] [MLIR][Linalg] pack, unpack to take memref inputs (PR #129036)

[Han-Chung Wang]

================
@@ -4951,7 +4993,7 @@ LogicalResult PackOp::canonicalize(PackOp packOp, PatternRewriter &rewriter) {
     rewriter.modifyOpInPlace(packOp, [&] {
       packOp.getSourceMutable().assign(source);
       packOp.getDestMutable().assign(dest);
-      packOp.getResult().setType(cast<RankedTensorType>(dest.getType()));
+      packOp.getResult().setType(cast<ShapedType>(dest.getType()));
----------------
hanhanW wrote:

I believe that this snippet is wrong after we add the memref semantics to pack ops. Because

1. There are no results in memref. Updating the type is incorrect.
2. The above comment states that it inserts `tensor.cast` op, while we need `memref.cast` op for memref version.

Thus, I suggest to kick in the canonicalization pattern only when the pack op has tensor semantics. Otherwise, the code is wrong to me.

https://github.com/llvm/llvm-project/pull/129036