[llvm] [AMDGPU] Emit a waitcnt instruction after each memory instruction (PR #68932)

[Jun Wang via llvm-commits]

jwanggit86 wrote:

Thanks for the comments. Doing the insertion of "waitcnt 0" in `insertWaitcntInBlock`, however, would mean that we have to check the newly-created flag (`EmitForAllMemOpFlag`) for _each instruction_. The proposed code change, on the other hand, checks this flag only once for each function. Because this new command-line option is for debug purposes only and not used in normal cases, we probably should limit the extra overhead for normal uses. 

https://github.com/llvm/llvm-project/pull/68932