[llvm] [AMDGPU] Disable atomic optimization of fadd/fsub with result (PR #96479)

[Jay Foad via llvm-commits]

jayfoad wrote:

> And maybe it's only wrong when %y is uniform -- I have not thought too much about the "scan" path when %y is divergent.

I think the divergent path is OK, so I've changed the patch to just avoid the uniform path for fadd/fsub-with-result.

> Add a todo to fix the 0->-0 case and check no-nan/no-infs?

I've added an explanatory FIXME comment by the code that generates the offending fmul.

https://github.com/llvm/llvm-project/pull/96479