[llvm] [ConstantFolding] Fold `vector.fmax` and `vector.fmin` (PR #143904)

[Yingwei Zheng via llvm-commits]

https://github.com/dtcxzyw commented:

My main concern is that the order of fp reduction is unspecified in LLVM. On x86 target, it is lowered into a "tree-like" order, and the result may differ (nan handling/signed zeros). Can you please provide Alive2 proof to show minnum/maxnum are safe to reassociate?
See also https://github.com/llvm/llvm-project/issues/82813


https://github.com/llvm/llvm-project/pull/143904