[clang] [llvm] [AMDGPU] Implement AsyncMark Stages (PR #220442)

[Ryan Mitchell via cfe-commits]

RyanRio wrote:

Thanks, I believe the PR is once again in a reviewable spot, at least I settled on a positive/negative decision myself (awaiting feedback).

To aid your reading - the intrinsics take the negative, as suggested by krzys, i.e. what stages to omit when creating marks, and what stages to ignore when deciding what marks to remove from the sequence list. However, as this would result in generally quite lengthy prints, (as you generally want to ignore everything but 1 or 2 stages), I made the prints show the positive, i.e. what sequences are appended to/removed from when calling asyncmark/asyncmark_wait.

I believe I've also updated the documentation to reflect your feedback Sameer.

I'll update the PR description and fix the merge conflicts.

https://github.com/llvm/llvm-project/pull/220442