With float128 x = nan, the test x > 0 evaluates to true

[charles@…]

On Linux with boost 1.57, executing the following code

#include <iostream>
#include <limits>
#include <boost/multiprecision/float128.hpp>

int main() {
  typedef boost::multiprecision::float128 real;
  real x = std::numeric_limits<real>::quiet_NaN();
  std::cout << "x: " << x << "\n"
            << "x == 0: " << (x == 0) << "\n"
            << "x >  0: " << (x >  0) << "\n"
            << "x <  0: " << (x <  0) << "\n"
            << "x >= 0: " << (x >= 0) << "\n"
            << "x <= 0: " << (x <= 0) << "\n"
            << "x != 0: " << (x != 0) << "\n";
  return 0;
}

gives

x: nan
x == 0: 0
x >  0: 1
x <  0: 0
x >= 0: 1
x <= 0: 0
x != 0: 1

I should get

x: nan
x == 0: 0
x >  0: 0
x <  0: 0
x >= 0: 0
x <= 0: 0
x != 0: 1

[charles@…]

I took a look at the code and see, for example in float128.hpp:

   int compare(const float128_backend& o)const
   {
      return m_value == o.m_value ? 0 : m_value < o.m_value ? -1 : 1;
   }

Clearly, this approach to comparisons will fail with NaNs. There will need to be a 4th possible value to indicate "unordered".

[John Maddock]

Oh :(

I need to look into this some more with all the floating-point backends.

[John Maddock]

Fixed in ​https://github.com/boostorg/multiprecision/commit/4f16c26457cd707d238133bc602db2ef05db91a8