Re: [bug-gawk] Overflow to Infinity

[Andrew J. Schorr]

On Fri, Jul 06, 2018 at 11:46:36AM +0300, Eli Zaretskii wrote:
> > +static char *
> > +format_nan_inf(double val)
> > +{
> > +   static char buf[100];
> > +
> > +   if (isnan(val)) {
> > +           sprintf(buf, "%g", val);
> > +           if (strcmp(buf, "nan") == 0) // "-nan" or "nan"
> > +                   strcpy(buf, "+nan");
> 
> This is non-portable.  On MS-Windows, the string you may get is
> "1.#IND", not "nan", depending on the version of the C runtime.  It
> can also have a leading sign, as in "-1.#IND".
> 
> Why do you need to sprintf the value, if it is known that it's a NaN?
> Why not do what you do with Inf:
> 
> > +   } else if (isinf(val)) {
> > +           strcpy(buf, val < 0 ? "-inf" : "+inf");

I don't think you get a different result when comparing +nan or -nan to 0,
do you? Is there a way to test for NaN signedness?
Perhaps this would work:


+       if (isnan(val)) {
+               sprintf(buf, "%g", val);
+               strcpy(buf, buf[0] == '-' ? "-nan" : "+nan");

But it seems like there ought to be a math library function to do this.
After all, sprintf figures out the sign somehow. Does the 'signbit'
function work with NaN values? I guess one needs to dig into the glibc
sprintf code, but I don't have time now.

Regards,
Andy