isNan and isInf in Constant.cpp produce wrong results on big-endian systems

[mhillenbrand]

Both attempt to decode a double into high and low 32-bit ints via a union:

glslang/glslang/MachineIndependent/Constant.cpp

Lines 49 to 52 in 8f56d73

	typedef union {
	double d;
	int i[2];
	} DoubleIntUnion;

On little-endian systems that works as intended. On big-endian systems, such as s390x, the 32-bit ints end up in reversed order and thus the two functions fail to detect Infs or NaNs.

Reproduction: glslangValidator -C -i Test/constFold.frag and compare out10 and out11 to expected values -- on s390x, they are all 0 since none of the NaN or Inf conditions are detected correctly.

[dneto0]

I believe that the use of the union exercised undefined behaviour. It's a type-based alias analysis violation.
C++14 3.10 "Lvalues and rvalues" says what types can be loaded from memory locations, depending on what types were used to write those locations.

Thanks for fixing this!

[greg-lunarg]

Had to revert #2803 with #2809; see issue #2808.

[mhillenbrand]

In my patch, I missed moving the #include <cfloat> that is required on Windows. Fixing and testing...
https://docs.microsoft.com/en-us/cpp/c-runtime-library/reference/fpclass-fpclassf?view=msvc-170