[llvm] AMDGPU: Drop and upgrade llvm.amdgcn.atomic.csub/cond.sub to atomicrmw (PR #105553)

[via llvm-commits]

anjenner wrote:

> Can you add some IR tests in test/Transforms/AtomicExpand? Need the 32/64 cases tested, and later I'll need to add the no fine grained memory vs. not cases

Could you be more specific about what tests are needed? If you could point me to the corresponding tests for one of the other atomicrmw operations (so that I can copy them and change the operation to usub_cond/usub_sat), that would be ideal.

https://github.com/llvm/llvm-project/pull/105553