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

[Nikita Popov via llvm-commits]

================

----------------
nikic wrote:

I see. I guess in that case you could explicitly set the promoted type to f32 in the RISCV backend instead if `+d` is not available...

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