[Bug libstdc++/13450] std::pow(std::complex<double>(-1.,0.),0.5) yields (NaN,0)

["pkienzle at nist dot gov" <gcc-bugzilla at gcc dot gnu dot org>]

------- Additional Comments From pkienzle at nist dot gov  2004-01-12 16:06 -------
(In reply to comment #9)
>   OK, I see what is going wrong.  But the submitter's patch is not right
> either. To properly implement the branch cut, we need to test whether
> the imaginary part is a positive or negative zero (on IEEE-754
> plateform).  Yes that is a crazy notion -- but we got it.

For example, the following is calculated by Matlab on IRIX:

  >> (-1+realmin*1i)^0.5
     0.0000 + 1.0000i

  >> (-1-realmin*1i)^0.5
     0.0000 - 1.0000i

> Therefore we need a primitive that tells the sign of a real number.

C99 has signbit and copysign.  I was able to access them as follows:

#include <iostream>
#include <cmath>
int main(int argc, char *argv[])
{
   double x=1.0, y=-1.0, z;

   std::cout << "1.0/HUGE_VAL=" << x/HUGE_VAL << std::endl;
   std::cout << "signbit(1.0/HUGE_VAL)="
           << __gnu_cxx::signbit(x/HUGE_VAL) << std::endl;

   std::cout << "-1.0/HUGE_VAL=" << y/HUGE_VAL << std::endl;
   std::cout << "signbit(-1.0/HUGE_VAL)="
           << __gnu_cxx::signbit(y/HUGE_VAL) << std::endl;

   return 0;
}

[prog]$ ./a.out 
1.0/HUGE_VAL=0
signbit(1.0/HUGE_VAL)=0
-1.0/HUGE_VAL=-0
signbit(-1.0/HUGE_VAL)=-2147483648

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13450