Re: Polyhedron performance regression

[Geert Bosch <bosch at adacore dot com>]

On Nov 13, 2006, at 17:12, Richard Guenther wrote:

> On 11/13/06, Richard Guenther <richard.guenther@gmail.com> wrote:
> > 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

I'm probably missing some context here (are x and y single
precision?), but if I understand correctly, we're talking
about floating point here. Then, doing lrint would
be very wrong because the result of the division may be
large, even though the final remainder is small enough.

So, it's not just an overflow case, but one that needs
to be computed correctly.
> 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?
Yes, it won't work for large arguments. Reducing large arguments
is exactly why one would use remainder, so just returning 0 seems
quite dubious. As example, in single precision,
remainder(773094178816, 360) should be 16 exactly.

  -Geert