[llvm-branch-commits] [llvm] AMDGPU: Start considering new atomicrmw metadata on integer operations (PR #122138)

[Matt Arsenault via llvm-branch-commits]

arsenm wrote:

> Clang adds !amdgpu.no.fine.grained.memory and !amdgpu.no.remote.memory to any atomic instructions by default. I think this behavior is expected to keep ISA unchanged compared to the ISA before these metatadat were introduced. Did I miss anything?

All of the tests that fail are atomicMin_system and atomicMax_system. I would expect that the explicit system scoped functions would not be using these annotations. With this PR, these tests are switching from CAS expansion to the direct instruction. 

It just happens that the integer min and max are the cases we handled conservatively before, so it's possible the test is just wrong in some way 

https://github.com/llvm/llvm-project/pull/122138