llvm.amdgcn.global.atomic.csub intrinsic should be removed

[arsenm]

This should be superseded by using atomicrmw usub_sat. Bitcode auto upgrade should replace it.

Additionally we have llvm.amdgcn.atomic.cond.sub.u32. Not sure what the difference is but it probably should also be removed

[llvmbot]

@llvm/issue-subscribers-backend-amdgpu

Author: Matt Arsenault (arsenm)

This should be superseded by using `atomicrmw usub_sat`. Bitcode auto upgrade should replace it.

[jayfoad]

This should be superseded by using atomicrmw usub_sat.

Should be, but it's not yet, because we don't codegen that to global_atomic_sub_clamp_u32 (formerly known as global_atomic_csub{,_u32}). That needs to be implemented first.

Additionally we have llvm.amdgcn.atomic.cond.sub.u32. Not sure what the difference is

It's a different operation and should be superseded by atomicrmw usub_cond, once the codegen is implemented.

[anjenner]

I have a PR for this: #105553 but it needs a little more work.