https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83661
Wilco <wilco at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |wilco at gcc dot gnu.org
--- Comment #8 from Wilco <wilco at gcc dot gnu.org> ---
(In reply to Christophe Monat from comment #6)
> (In reply to Alexander Monakov from comment #5)
> > sincos performs range reduction for the argument just once, which is fairly
> > important. A well-optimized sincos also shares some computations for the
> > sin/cos parts, as done in
> > https://github.com/ARM-software/optimized-routines/blob/master/math/sincosf.c
>
> Thanks for the pointer, indeed this implementation looks great.
>
> @Pratamesh #Linaro: is there synchronization between the ARM optimized
> routines and the usuals libc (glibc, newlib, musl,...) ?
I already committed the rewritten sinf/cosf/sincosf to GLIBC and newlib. Musl
still uses a slow version.
> > (fwiw I'm against adding such transforms to the compiler)
>
> I would favor upgrading the routines if it's not already done, and not do
> any such transformation in the compiler.
One of the key issues with sincos is that it has a slow interface - you can get
a large speedup by avoiding all the unncessary memory accesses.
