[llvm] [AMDGPU] Set glc/slc on volatile/nontemporal SMEM loads (PR #77443)

[Carl Ritson via llvm-commits]

https://github.com/perlfu commented:

The code changes make sense to me.
It does seem the memory does not explicitly cover SMEM at the moment.
Adding cache bypass bits certainly shouldn't make the code gen any less correct.

Not really a problem with this patch, but I cannot convince myself that `RT_NT` is the right temporal hint on GFX12 for volatile -- I would have expected perhaps `LU`.  Seems some kind of scope might also be needed.

https://github.com/llvm/llvm-project/pull/77443