[llvm] [SDAG] Fix fmaximum legalization errors (PR #142170)

[Nikita Popov via llvm-commits]

nikic wrote:

> > One thing this makes me wonder is why we're not preferring libcall legalization for scalar fmaximum. The generated code is quite complex and probably not better than using a libcall.
> 
> Are there even libcalls? We unfortunately don't know the target runtime versions and these are really new calls

I'd expect that if the ops exist in IR, then the libcalls do as well. Either because it originated from a libcall in new enough C, or because it's a non-C language that is capable of providing its own builtins (a requirement whenever you're exposing more ops or types than C does).

https://github.com/llvm/llvm-project/pull/142170