[Mlir-commits] [mlir] [vector][mlir] Canonicalize to shape_cast where possible (PR #140583)

[James Newling]

newling wrote:

> Thanks!
> 
> I run the SME e2e tests and all pass. I wasn't able to cherry-pick this in IREE though, getting weird compilation errors. Though upstream tests should be sufficient to surface all potential issues.
> 
> @newling , why not name all "folding" patterns as `fold...`? Wouldn't that be more consistent?

I'll give this a spin with IREE 

> @newling , why not name all "folding" patterns as fold...? Wouldn't that be more consistent?

Yes, I think so. Actually fe3933da15b5b made me wonder if we should split canonicalize.mlir into 2 files (the new one with name fold.mlir containing everything in canonicalize.mlir that only depends on 1-time folds). 

@banach-space and @dcaballe thanks for your feedback! Unfortunately I'm going to put this on hold again temporarily, as I've uncovered some other things which should  be done before this. Moving back into draft mode, will ping when I think it's ready again. 

https://github.com/llvm/llvm-project/pull/140583