[Mlir-commits] [mlir] [mlir][ArmSME] Add rudimentary support for tile spills to the stack (PR #76086)

[Cullen Rhodes]

https://github.com/c-rhodes approved this pull request.

LGTM cheers Ben!

I know you mentioned it's not exactly optimal or elegant, and I agree, but we have to play the hand we're dealt and in this case it's a set of intrinsics that force us to do things we probably shouldn't be doing in MLIR. Hopefully a bit further down the road these intrinsics will either evolve or a new set will come along where the tile is a proper type with spilling/filling support in the backend and dataflow is correctly modeled.

In the meantime, the functionality you've implemented here allows code that uses more tiles than exist at the architecture level to correctly run and there's value in that. It's much better than a crash anyway! 😄 



https://github.com/llvm/llvm-project/pull/76086