On Wed, Mar 13, 2019 at 02:07:29PM +0100, Jakub Jelinek wrote:
> On Wed, Mar 13, 2019 at 01:00:58PM +0000, Joseph Myers wrote:
> > > Yeah, an alias doesn't really cost much, and has the advantage that if in
> > > the vector version you need at some point to differentiate between the
> > > finite only vs. full implementations, you can just by tweaking libmvec
> > > implementation, the callers will have proper calls depending on if they
> > > were
> > > compiled with -Ofast or -O3 etc.
> >
> > Experience is showing that some or all of the finite-only versions in
> > glibc were mistaken premature optimization - that proper optimized
> > implementations do not gain anything from adding a finite-only
> > restriction. There is no good basis to suppose that if additional
> > variants of the vector functions were useful in future, finite-only would
> > be the right conditional (or that the right set of variants would be the
> > same as the right set of variants for scalar functions).
>
> If the finite only doesn't buy anything, then another option is to drop the
> math-finite.h stuff or portions thereof.
> But adding a new GCC extension that other compilers will need to implement
> too, instead of just adding a couple of aliases seems to be overkill to me.
Not to mention that fixing it on the glibc side will make it work also with
GCC 8, GCC 7 and GCC 6.
Jakub
