Re: Slaying the HYPOTamus

Tom Lane wrote:

> Greg Stark <gsstark@mit.edu> writes:
>   
>> If there's a performance advantage then we could add a configure test
>> and define the macro to call hypot(). You said it existed before C99
>> though, how widespread was it? If it's in all the platforms we support
>> it might be reasonable to just go with it.
>>     
>
> For one data point, I see hypot() in HPUX 10.20, released circa 1996.
> I suspect we would want a configure test and a substitute function
> anyway.  Personally I wouldn't have a problem with the substitute being
> the naive sqrt(x*x+y*y), particularly if it's replacing existing code
> that overflows in the same places.
>
>             regards, tom lane
>
>   

A hypot() substitute might look something like this psudo-code, this is
how Python does it if the real hypot() is missing.

double hypot( double dx, double dy )
{   double yx;
   if( isinf(dx) || ifinf(dy) ) {     return INFINITY;   }    dx = fabs(dx);   dy = fabs(dy);   if (dx < dy) {
doubletemp = dx;       dx = dy;       dy = temp;   }   if (x == 0.)       return 0.;   else {       yx = dy/dx;
returndx*sqrt(1.0+yx*yx);   }
 
}

As the following link shows, a lot of care could be put into getting a
substitute hypot() correct.
http://gforge.inria.fr/plugins/scmsvn/viewcvs.php/trunk/hypot.c?rev=5677&root=mpfr&view=markup