We are porting our MSVC/Win32 applications to Clang/GCC/Linux and have just spent the morning tracing why our unit tests fail.
void BadExampleCode()
{
double a = NAN;
ASSERT(!isnan(a));
}
Under MSVC and Clang all good, GCC asserts. We added printf’s and looked at the assembly and the code was hard coded to 0.
Some googling found 2006 posts stating GCC -ffast-math did odd things with isnan, and it’s still a presently reported issue
This came about because we are using Premake, and had the FloatFast flag set, because that’s how our MSVC projects were set, and we don’t want to change those builds, so for now we have tweaked the Premake code for this flag under GCC, as it doesn’t make sense that you would ever want isnan to be hard coded to zero, that’s really not fast maths at all.
We also put a #ifdef for ffast-math to break if this flag is present in the future.
And by we I mean Dave, and I just sat and talking it though with him.
