Polyhedron performance regression

[Richard Guenther]

On 11/13/06, Richard Guenther <richard.guenther@gmail.com> wrote:
> On 11/13/06, Richard Guenther <richard.guenther@gmail.com> wrote:
> > On 11/13/06, Paul Richard Thomas <paul.richard.thomas@gmail.com> wrote:
> > > Richard,
> > >
> > > > I bet we are missing an inline x86_64 variant.  But it should be not too
> > > > hard to implement.
> > >
> > > 'tis what I thought.
> > >
> > > My view is that Steve is right about right and slow; however it is a
> > > very visible right and slow.  Which is the right button to press to
> > > have the inline version crafted for us?
> >
> > I can do expanders for fmod and remainder.
>
> I sort of take that back ;)  But at least for the fortran MODULO you should
> be able to emit A - FLOOR (A / P) * P as in the definition.  Currently
> it seems at least FMOD is called in the expansion.  With the above
> definition used for expansion you'd get
>
>         cvtss2sd        %xmm0, %xmm3
>         movsd   (%rsi), %xmm2
>         movapd  %xmm3, %xmm0
>         divsd   %xmm2, %xmm0
>         cvttsd2si       %xmm0, %eax
>         cvtsi2sd        %eax, %xmm1
>         ucomisd %xmm1, %xmm0
>         jae     .L2
>         subl    $1, %eax
>         cvtsi2sd        %eax, %xmm1
> .L2:
>         mulsd   %xmm2, %xmm1
>         movapd  %xmm3, %xmm0
>         subsd   %xmm1, %xmm0
>
> exact IEC 60559 semantics for remainder will probably not be suitable
> for inlining.  But I can try...

A quick shot for remainder with not honoring signed zeros would be

  remainder (x, y) = x - lrint (x / y) * y

precision is of course an issue, but with cvtsd2siq available we can cover
more than the 52 bits of double precision in the integer part so we only
have intermediate rounding issues left.  So a -fflag-unsafe-math-optimizations
expansion which would yield

        movapd  %xmm0, %xmm2
        divsd   %xmm1, %xmm2
        cvtsd2siq       %xmm2, %rax
        cvtsi2sdq       %rax, %xmm2
        mulsd   %xmm1, %xmm2
        subsd   %xmm2, %xmm0

modulo overflow handling (which is a comparison against 0x8000000,
a jump and loading zero as result).

Anyone spots a serious flaw here?

Richard.