[Mlir-commits] [mlir] [mlir][linalg] Extract `GeneralizePadOpPattern` into a standalone transformation (PR #117329)

[Andrzej Warzyński]

banach-space wrote:

Thanks for taking a look!

> Side comment: this is another "layering violation", where `tensor.pad` "lowers" to `linalg.fill`. Tensor dialect should not know about linalg at all. Yet another candidate to move to linalg. (https://discourse.llvm.org/t/rfc-move-tensor-pack-and-tensor-unpack-into-linalg/83096/)

Note that your RFC covers `tensor.pack` and `tensor.unpack`, and this PR is for `tensor.pad`. I've actually alluded to similar patterns for `tensor.pack` and `tensor.unpack` in your RFC: https://discourse.llvm.org/t/rfc-move-tensor-pack-and-tensor-unpack-into-linalg/83096/8?u=banach-space. And yes, I am aware that we _decompose_ `tensor.pack` into `tensor.pad` :)

Now, given that this is merely moving things around within (effectively) Linalg and making sure that we don't mix "decomposition" and "vectorization", I am hoping that this is non-controversial 😅 

https://github.com/llvm/llvm-project/pull/117329