[Mlir-commits] [mlir] [mlir][xegpu] Add dynamic memref support in transpose optimization. (PR #170218)

[Charitha Saumya]

charithaintc wrote:

> I feel like extracting a pointer from memref needs to resemble what xegpu-to-xevm has in its [memref materialization](https://github.com/llvm/llvm-project/blob/3c6864ab8879f7274e6c24c3b7c8e8139cd135dd/mlir/lib/Conversion/XeGPUToXeVM/XeGPUToXeVM.cpp#L1003). Yes, right now they are the same, but AFAIK, @Jianhui-Li plans to expand it by accounting for metadata (`ExtractStridedMetadataOp`) so that memrefs coming from reinterpret casts also work correctly.
> 
> Perhaps the upcoming materialization logic can be moved to utils for reusability across passes. Otherwise looks good.

good point! in this PR, I transfer the existing strides and shape info to the new create_nd. for offset  I guess extract_aligned_ptr should handle it? If not we need to revisit this logic. 

https://github.com/llvm/llvm-project/pull/170218