Missed optimization when computing fma

[icedrocket]

It optimizes well when the fma is computed explicitly with f32::mul_add(), but not when it is not, like x * y + z.

https://godbolt.org/z/fh74x4W88

[workingjubilee]

That is a result-changing optimization, so the optimization is not permissible. Otherwise, this is a duplicate of e.g. #21690

[workingjubilee]

And there is no LLVM bug here: we don't ask LLVM to do results-changing optimizations, so it doesn't. We would need to define this optimization as acceptable on the Rust language, first.