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

[Diego Caballero]

dcaballe wrote:

> The simple pattern matching is done while collecting the unknown dimensions, and functions as an early-exit, as if we can tell that a mask is not going to be all-true we can bail out before doing any more expensive value-bounds analysis. Patterns like this already exist as canonizations, but just give-up on unknown dims rather than trying to resolve them

Sorry, I'm not sure what this comment implies but, yes, I'm meant that the extra pattern matching should be part of canonicalization. If you also need those patterns here to make the pass reach a fixed point, it should be fine to also populate them. As noted, I would expect this to also simplify fixed-length vectors so if the existing canonicalization ones also include fixed-length pattern it should be ok to also have them here.

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