[Mlir-commits] [mlir] [mlir][linalg] Document and diagnose pack/unpack memref limits (PR #225773)

[Adam Siemieniuk]

================

----------------
adam-smnk wrote:

I agree with the general direction to discourage implementing core rewrites on memref just for sake of completness. These are bound to bring more edge cases than expected without clear payoff.

Why I focus on vectorization and tiling specifically is that there's a clear mismatch in overall message. If tiling memrefs is fine in general, why it's suddenly discouraged just for pack.
In these cases, existing TODOs are not unreasonable. AFAIK, they already imply "it could be done but not really needed; patches welcome".

I'd try to emphasize more "adding support requires strong motivation" where relevant. I'm fine with the "memrefs not recommend" message in more specialized specific rewrites like layout propagation.
Now, phrasing it correctly is hard, no clear suggestion in my mind yet 😅

https://github.com/llvm/llvm-project/pull/225773