Mat::operator=(const Scalar&) incorrectly handles NaNs with floating-point images when GCC's -ffast-math flag is enabled (Bug #2056)

The

if( s[0] == 0 && s[1] == 0 && s[2] == 0 && s[3] == 0 )

test in modules/core/src/copy.cpp's definition of Mat::operator=(const Scalar& s) evaluates to true when s contains NaNs and -ffast-math is enabled, resulting in the input Mat being filled with zeros rather than NaNs. Compiling the following code against a version of OpenCV built with -ffast-math reproduces the problem:

// Define a NaN value
union {
  float f;
  uint32_t i;
} nan;
nan.i=0x7FC00000;

// Assign nan.f to image
cv::Mat nanTest(1,1,CV_32FC1);
nanTest=cv::Scalar(nan.f);

// Print results
std::cout<<"nan.f="<<nan.f<<std::endl;
std::cout<<"nanTest.at<float>(0,0)="<<nanTest.at<float>(0,0)<<std::endl;

To show that -ffast-math is the cause of the problem, compile the following program:

#include <iostream>
#include <cstdint>

int main(int argc, char** argv) {
  union {
    float f;
    uint32_t i;
  } nan;
  nan.i=0x7FC00000;

  std::cout<<"nan.f==0 = "<<(nan.f==0)<<std::endl;

  return 0;
}

with

g++ -std=c++0x -ffast-math -o TestFastMath Test.cpp && g++ -std=c++0x -o Test Test.cpp

./Test produces the output nan.f==0 = 0, whereas ./TestFastMath produces the output nan.f==0 = 1.

The problem occurs with the build of OpenCV 2.3.1 that's available from the Ubuntu 12.04 repository, which I presume was built with -ffast-math enabled. I have just downloaded and built OpenCV 2.4.1, and -ffast-math was disabled by default, so the problem does not occur. If you must allow users to build OpenCV with -ffast-math, please consider using something like the following function to test whether or not the cv::Scalar is zero instead:

inline bool isZero(const double v) {
  union {double d; uint64_t i;} u;
  u.d=v;
  return u.i==0;
}