[Mlir-commits] [mlir] [mlir][vector] Add mask elimination transform (PR #99314)

[Diego Caballero]

dcaballe wrote:

Thanks! A couple of high-level comments. I'll come back :)

1. Could we move the trivial pattern-matching cases to the folder/canonicalizer? We really don't want to have any trivial all-one mask around both for the fixed and scalable cases. It makes sense to have a more sophisticated and expensive pass to get rid of the tricky cases that require value bounds or range analysis.

2. Could we use `constant_mask` instead of `arith.constant`? This is not enforced in any way but we've been using the former to represent vector masks (they are kind of redundant, actually).


https://github.com/llvm/llvm-project/pull/99314