[llvm] [IVDescriptors] Don't require nsz/nnan for (min|max)num. (PR #137003)

[Florian Hahn via llvm-commits]

fhahn wrote:

Ok, so it sounds like the new intrinsics are the way to go. I'll add support for maximumnum/minimumnum reductions and update Clang to emit use `maximumnum/minimumnum` for `fmax/fmin`.

One thing to note is that GCC uses fmaxnm on AArch64 for fmax, which appears incorrect as per the discussion above: https://godbolt.org/z/W3z1G1z3c

https://github.com/llvm/llvm-project/pull/137003