[Bug regression/93009] AVX512 FMA - wrong code generation

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93009

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #2)
> It is called fma but it does not have to do the fused multiple add.
> For scalar types, you should use std::fma .  BUT NOTE on targets where FMA
> does not exist, the function std::fma can be slow as the multiple is done in
> infinite precision.

That really depends on the libm implementation, e.g. glibc uses
algorithm from the http://www.lri.fr/~melquion/doc/08-tc.pdf paper.