[llvm] [AMDGPU] Add waterfall intrinsics (PR #192409)

[Jay Foad via llvm-commits]

jayfoad wrote:

> llvm.amdgcn.waterfall.begin
> llvm.amdgcn.waterfall.readfirstlane
> llvm.amdgcn.waterfall.end
> llvm.amdgcn.waterfall.last.use
> llvm.amdgcn.waterfall.last.use.vgpr
> llvm.amdgcn.waterfall.loop_end

Suggestion: for the initial PR, maybe just include the begin/readfirstlane/end intrinsics? These are all you need to demonstrate basic use cases.

The other intrinsics are for more complex use cases and could be added later. In particular I remember last.use.vgpr being a pretty strange case, which probably needs a review all to itself.

https://github.com/llvm/llvm-project/pull/192409