On Mon, Apr 24, 2023 at 12:05:43PM -0400, Siddhesh Poyarekar wrote:
> On 2023-04-24 12:03, Jakub Jelinek wrote:
> > On Thu, Apr 20, 2023 at 01:57:55PM -0400, Siddhesh Poyarekar wrote:
> > > > Similarly for other functions which have other ranges, perhaps not with
> > > > so
> > > > nice round numbers. Say asin has [-pi/2, pi/2] range, those numbers
> > > > aren't
> > > > exactly representable, but is it any worse to round those values to
> > > > -inf or
> > > > +inf or worse give something 1-5 ulps further from that interval
> > > > comparing
> > > > to other 1-5ulps errors?
> >
> > I've extended my hack^^^test to include sqrt and this time it seems
> > that the boundary in that case holds even for non-standard rounding modes
> > for glibc:
>
> IIRC the standard _requires_ sqrt to be correctly rounded.
Well, we still need to make some effort to verify it is the case.
BTW, what my hacks didn't check yet and I should eventually is libmvec,
because if ranger makes some assumptions on libm calls and then we vectorize
it, it will be libmvec rather than libm.
Jakub
