MLIR for arm SME vectorizing matmul-like ops as part of a broader program

[banach-space]

@nujaa

To sort the issue with maskOp, Should we prevent vectorization from iterating inside vector.maskOp, considering it should already be vectorized since it is a vectorOp? I am working on such fix at the moment.

Correct. This was also pointed out by @dcaballe who architected the masking support in the vectorizer (On Improving Arm SME Lowering Resilience in MLIR - #15 by dcaballe):

@dcaballe

This looks like an easy-to-fix bug and you pointed to the solution: the vectorizer shouldn’t try to vectorize an op that is already vectorized. Both the mask op and the masked op should have vector types so they shouldn’t pass the vectorizer’s preconditions.

You probably need to update vectorizeOpPrecondition.

@nujaa

I am working on such fix at the moment.

Great, please add as as reviewers and/or reach out if you experience any issues with this.

-Andrzej