[llvm] [AMDGPU] Insert before and after instructions that always use GDS (PR #131338)

[Jay Foad via llvm-commits]

jayfoad wrote:

> Should this be treated as a type of hazard, rather than something for waitcnt insertion to deal with?

No. The waitcnt _before_ a GWS instruction definitely belongs in this pass, since we need to track whether there is an outstanding GDS operation. For the waitcnt _after_ a GWS instruction it is less clear, but I think it fits fairly naturally into this pass.

https://github.com/llvm/llvm-project/pull/131338