[llvm] fix `llvm.fma.f16` double rounding issue when there is no native support (PR #171904)

[Trevor Gross via llvm-commits]

tgross35 wrote:

> I could not find examples of a libcall being used for fma, but that's something that could be looked in separately to work around code size issues.

I think this is pretty triple-dependent, e.g. glib and LLVM’s libraries have them but not much else does. I think what would be nice one day is for a frontend to have a way to indicate that a full C23 math library is available (with fmaf16, fmaf32, fmaf64, fmaf128 and similar available rather than just fma/fmaf/fmal), which would help with this plus a few other things (I’ve mentioned something similar to this to @arsenm before).

But for now, unconditionally using f64 ops is definitely easiest.

https://github.com/llvm/llvm-project/pull/171904