[Mlir-commits] [mlir] [mlir][vector] Allow signless integer element types in `vector.step` (PR #205142)

[Andrzej Warzyński]

banach-space wrote:

> @banach-space the changes have actual effect on some pipelines (updated linalg tests). I need to leave for the day, but unless you disagree, tomorrow I'll update the pattern taking `indexBitwidth` as an argument for backwards compatibility.

It makes sense to me as a measure to reduce noise in this PR and as temporary solution to allow folks to transition. 

That said, I would immediately document it as "deprecated" and send a follow-up to remove it some time soon. Basically, we make no promises re C++ API stability and this workaround will weaken the great that you've been doing to clarify the semantics of `vector.step`. Also, with `indexBitwidth`, you'd probably want to update `VectorizeChildrenAndApplyPatternsOp`: https://github.com/llvm/llvm-project/blob/0b5c006208cf7b0fee3709f468152e7de81607ef/mlir/lib/Dialect/Linalg/TransformOps/LinalgTransformOps.cpp?plain=1#L4141

How will you decide the correct `bit-wdith`?

One alternative that i see is to update how `vector.step` is created in e.g. the Linalg vectorizer and replace e.g. `vector.step : vector<8xindex>` with `vector.step : vector<8xi8>`. Hopefully that would be enough to reduce noise in tests.

Either path will be an improvement and I am happy for to choose your preference. We've already diverged a bit and I am keen to unblock you. I will definitely appreciate if you continue improving the situation with `vector.step`, but that's not a pre-condition for landing this. 

https://github.com/llvm/llvm-project/pull/205142