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

[Trevor Gross via llvm-commits]

tgross35 wrote:

> the test failures are kind of annoying
> 
> the `Generic` tests sort of require all backends, and I don't really want to build _everything_. The AMDGPU one succeeds locally, but also its name suggests it deals with `bf16`, so maybe it actually should not change at all?

The failing check in `half-op.ll` is an xfail for the issue that this fixes, so I think you should just be able to replace BAD-FMA with CHECK-FMA in all the RUN instructions and almost all targets should pass.

That will still fail if there are targets that do f64->f32->f16 rather than f64->f16. Maybe replace BAD-FMA with SKIP-FMA in these cases?

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