[Mlir-commits] [mlir] [MLIR][Linalg] Add boundary fill option to pad fusion pass (PR #223792)

[Renato Golin]

https://github.com/rengolin requested changes to this pull request.

This will produce an uneven tensor, could possibly add undefs and poison, and UB to otherwise legal code. Why can't you just fill the padded tile at the tile level in some downstream transform?

I'm also concerned about the amount of Claude assistance. Weird tests, weird code, weird solution to a corner case problem. Code verified by human is ok, harness to make decisions upon, not so much . Even long term developed harness fail at corner cases, how can we trust anything that the numbers tell us?

Finally, special cases are not generally a great way of doing it at this level unless it spans across a wide variety of usages / hardware / arches. Then we add some special op or something, but this case does not warrant it at all.

I'm not sure this PR achieves much of that, so I'm skeptical of merging it.

https://github.com/llvm/llvm-project/pull/223792