[Mlir-commits] [mlir] [mlir][Vector] Add fold transpose(shape_cast) -> shape_cast (PR #73951)

[Benjamin Maxwell]

MacDue wrote:

> > > Kicked of a PR on IREE ([openxla/iree#15748](https://github.com/openxla/iree/pull/15748)) that cherry-picks this change and undoes the revert we were carrying locally if it fixes the original issue.
> > 
> > 
> > This does not depend on the previous transpose lowering (which will likely still lead to problems for you)
> > This patch alone is all we need for lowering SME matmuls :)
> 
> Brain fade on my part.... So really we will also need to upstream revert the original patch.... Thats a different discussion. Fixed up the PR for now that just tests this patch.

One thing to note is that with the original patch reverted, the transpose lowering generates invalid code for scalable vectors. It must at least return `failure()` for scalable vectors.

https://github.com/llvm/llvm-project/pull/73951