Requires floating point excess precision on x86 with gcc

[pjsg]

It turns out that the tests fail on linux on Intel if you add the following code:

void 
set_fpu (unsigned int mode)
{
  asm ("fldcw %0" : : "m" (*&mode));
}

and then

set_fpu(x27f);

at the start of main. This (according to https://www.linuxtopia.org/online_books/an_introduction_to_gcc/gccintro_70.html ) sets excess precision to be disabled. This makes it compatible with other platforms that do not support excess precision.

Indeed, the results that I then get from x86 compare to those that I get from a microcontroller platform.

The failed tests seem to indicate that the system sprintf is doing something different. The worst cases appear to be with very small exponents. For example:

Results don't match, format string: %.16e
         sprintf(3): [1.6784992078596053e-265] (23)
        snprintf(3): [1.6784992078596017e-265] (23)
Results don't match, format string: %.16e
         sprintf(3): [1.6784992078596089e-265] (23)
        snprintf(3): [1.6784992078596053e-265] (23)

I am a bit puzzled by the fact that the x86 code is using long double and the microcontroller platform is not.