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

[Jay Foad via llvm-commits]

jayfoad wrote:

@b-sumner @pravinjagtap if the performance of the "uniform fadd/fsub with result" case is important then there might be ways to reinstate the uniform case with an extra bug fix, like writing -0.0 into the first active lane of `%y * +0.0`.

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