Re: Weird constant folding of __builtin_remainder() with -frounding-math

[Marc Glisse <marc dot glisse at inria dot fr>]

On Wed, 20 Nov 2013, Marc Glisse wrote:

> On Wed, 20 Nov 2013, FX wrote:
>
> > I reduced my problem to the following code:
> >
> > int main (void)
> > {
> >  double x;
> >
> >  x = 1 / 3.;
> >  __builtin_printf ("%.30lg %.30lg\n", __builtin_remainder(1., 1/3.), 1/3.);
> >  __builtin_printf ("%.30lg %.30lg\n", __builtin_remainder(1., x), x);
> > }
> >
> > This calls __builtin_remainder() twice, once with constant arguments, once 
> > with a variable. Without rounding math, the two results agree (5.55e-17), 
> > but with -frounding-math, the constant case changes (-2.71e-20).
> >
> > On x86_64-apple-darwin, things are even weirder than x86_64-linux (the 
> > above): constant-case with -frounding-math -m64 is unchanged, while 
> > constant case with -frounding-math -m32 follows the linux results.
> >
> > I donât think this is right, since remainder() is supposed to be unaffected 
> > by the rounding mode. Yet, before I open a PR, I wanted to see if someone 
> > understands whatâs going onâ
>
> Did you really test on x86_64-linux? I am only seeing this -2.71e-20 when 
> using -mfpmath=387.

Oups, didn't finish my email.

So 1./3 is computed with 80 bits, the first computation is very nice, then 
the call to printf forces that number to memory where it is rounded to 64 
bits (anything can randomly cause that rounding at any time), and that is 
then used for the second computation.

-frounding-math is a red herring, -O0 could have disabled the constant 
folding.

--
Marc Glisse