[llvm] [SDAGBuilder] Fix incorrect fcmp+select to minnum/maxnum transform (PR #184590)

[Nikita Popov via llvm-commits]

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

----------------
nikic wrote:

I think the s390x regressions maybe avoidable. From my reading of the s390x PoO, for M=4, vfmaxdb with an sNaN operand returns the other operand rather than qNaN. @uweigand Is that correct? In that case we should mark FMINIMUMNUM/FMAXIMUMNUM as legal for the same types as FMINNUM/FMAXNUM and select them the same way.

https://github.com/llvm/llvm-project/pull/184590