[llvm] [AMDGPU] always emit a soft wait even if it is trivially ~0 (PR #147257)

[Jay Foad via llvm-commits]

jayfoad wrote:

> Something still doesn't feel right with this PR for me

Me too.

Currently `S_WAITCNT_soft` has exactly the same execution semantics as `S_WAITCNT`. (The only difference is whether we want `SIInsertWaitcnts` to optimize it away if it can prove that it is a no-op.) What I don't like about your patches is that you're assigning some extra meaning to `S_WAITCNT_soft vmcnt(63)`, or to the mere existence of an `S_WAITCNT_soft` even if the corresponding `S_WAITCNT` would be a no-op. I think it would be much cleaner to use a different pseudo for that.

https://github.com/llvm/llvm-project/pull/147257