[llvm] [AMDGPU][gfx1250] Also add a wait on xcnt before volatile accesses (PR #168852)

[Pierre van Houtryve via llvm-commits]

Pierre-vh wrote:

> I think I get it - you don't want these accesses to be repeated just because some _other_ (prior) access failed address translation? Makes sense.

Exactly, if a previous access fails address translation, we could re-try the atomic store/volatile access too, which would be incorrect as they can only be done (observed) once. The xcnt wait ensures that can't happen.

https://github.com/llvm/llvm-project/pull/168852