[llvm-branch-commits] [llvm] [AMDGPU] Canonicalize NAN values for float reduction intrinsics (PR #175131)

[Matt Arsenault via llvm-branch-commits]

https://github.com/arsenm commented:

We probably should just ignore the issue and document the order of signaling nans is nondeterministic. I think you can avoid the canonicalize inside the loop (plus we should be able to detect most of the time that the source can't be a signaling nan, but that's difficult here). It's not really worth the high effort to handle this accurately. Signaling nans do not matter in the slightest in the real world, and is just a defect IEEE has inflicted upon us 

https://github.com/llvm/llvm-project/pull/175131