On Tue, 20 Nov 2018 at 05:24, Nick Desaulniers <ndesaulni...@google.com> wrote:
> >
> > The only functional change I noticed was this in udiv_qrnnd.
> >
> > __r1 = (n1) % __d1;
> > __q1 = (n1) / __d1;
> >
> > Becomes this:
> >
> > __q1 = (n1) / __d1;
> > __r1 = (n1) - __q1 * __d1;
> >
> > This is equivalent as it instead of calculating the remainder
> > using modulo, it uses the result of integer division to subtract the
> > count of 'whole' d1 from r1.
>
> I don't understand this; why was this functional change made?
I couldn't see why. It pre-dates GMP's mecurial history that .
\
> > - (q) = (UWtype) __q1 * __ll_B | __q0; \
> > + (q) = __q1 * __ll_B | __q0;
> > \
> > (r) = __r0;
> > \
> > } while (0)
>
> This appears to now differ from the upstream source:
> https://github.com/gcc-mirror/gcc/blob/f7289f563a5e447ac21a5901ed75aad0dbd37732/include/longlong.h#L1679
AIUI the upstream source is GMP:
https://gmplib.org/repo/gmp/file/tip/longlong.h
> If we're going to borrow implementations from GCC, let's borrow the
> same implementation. Otherwise it's hard to have confidence in this
> part of the patch.
I agree we should use the upstream source.
Segher, which tree contains the One True Upstream for longlong.h?
Cheers,
Joel
