https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119092

--- Comment #10 from Matthias Kretz (Vir) <mkretz at gcc dot gnu.org> ---
> lower them in the frontend to scalar operations

But that's not useful IMHO. I can already do that with a simple inline
function. The __builtin_elementwise_math functions I'm looking for need to be
different because they ignore errno.

Right now -ffast-math is necessary to get glibc vector math calls. I need to be
able to call vector math implementations without turning on fast-math.

> The important part is whether stuff like __builtin_elementwise_sqrt reliably 
> produces native sqrt instructions without overhead.

vectorized sqrt and a few other math functions are the low-hanging fruit
because most (all?) targets can translate that to a single instruction. IMHO,
we should have this subset of builtins ASAP :-)

Reply via email to