[llvm] [LV] Vectorize fmin/fmax reductions (PR #198300)

[Florian Hahn via llvm-commits]

================

----------------
fhahn wrote:

I think the last remaining potential mis-compile is `FMaximumNum`/`FMinimumNum` with `NaN` as start value. If the compare in the loop is ordered, I think the `NaN` handling of the intrinsics causes an issue.

Maybe best to bail out here if the start value for `FMaximumNum`/`FMinimumNum` may be `NaN` + test

https://github.com/llvm/llvm-project/pull/198300