Math operations for f16 incorrectly implemented

[bjorn3]

libm has no such intrinsics as logf16 and roundf16. Discovered in rust-lang/rust#148206 (comment).

[tgross35]

Those are real functions in C23 and are available with some libms, e.g. https://github.com/llvm/llvm-project/blob/ddb706bbb04a460fb2741268039c69787dc8b2fc/libc/src/math/roundf16.h#L17. I think LLVM will eventually start using them in some lowerings, there are legit usecases (e.g. fmaf16 currently lowers to extend f64 -> fmaf64 -> trunc f16 to get the right precision, but fmaf16 is way cheaper if you don't have hardware f64). So IMO cranelift doesn't need to change anything here, we can just add wrappers in libm that forward to the f32 versions.

[bjorn3]

Compiler-builtins doesn't have logf16 at least.

[tgross35]

Right, currently - that's why I said we we can just add wrappers there.

[tgross35]

Actually, even easier; we can just add fallbacks to the intrinsics that go via the f32 implementations.

[tgross35]

rust-lang/rust#150946