Can I make gcc tell me when a calculation results in NaN or inf at runtime?

Question

Is there a way to tell gcc to throw a SIGFPE or something similar in response to a calculation that results in NaN or (-)inf at runtime, like it would for a divide-by-zero?

I've tried the -fsignaling-nans flag, which doesn't seem to help.

Geotaxis

-fsignaling-nans is runtime, not compile time. – Quechua

Answer 1

Almost any floating-point operation or math library function that produces a NaN from non-NaN inputs should also signal the 'invalid operation' floating-point exception; similarly, a calculation that produces an infinity from finite inputs will typically signal either the 'divide-by-zero' or 'overflow' floating-point exception. So you want some way of turning these exceptions into a SIGFPE.

I suspect the answer will be highly system-dependent, since control of floating-point traps and flags is likely to be supplied by the platform C library rather than by gcc itself. But here's an example that works for me, on Linux. It uses the feenableexcept function from fenv.h. The _GNU_SOURCE define is necessary for this function to be declared.

#define _GNU_SOURCE
#include <fenv.h>

int main(void) {
    double x, y, z;
    feenableexcept(FE_DIVBYZERO | FE_INVALID | FE_OVERFLOW);

    x = 1e300;
    y = 1e300;
    z = x * y; /* should cause an FPE */

    return 0;
}

A caveat: I think it's possible with some setups that the exception isn't actually generated until the next floating-point operation after the one that (in theory) should have caused it, so you sometimes need a no-op floating-point operation (e.g. multiplying by 1.0) to trigger the exception.

Osuna

Answer 2

On MinGW 4.8.1 (GCC for Win32) I see that the feenableexcept isn't defined. The workaround is to use the Win32 platform's _controlfp thus:

#undef __STRICT_ANSI__ // _controlfp is a non-standard function documented in MSDN
#include <float.h>
#include <stdio.h>

int main()
{
   _clearfp();
   unsigned unused_current_word = 0;
   // clearing the bits unmasks (throws) the exception
   _controlfp_s(&unused_current_word, 0, _EM_OVERFLOW | _EM_ZERODIVIDE);  // _controlfp_s is the secure version of _controlfp

   float num = 1.0f, den = 0.0f;
   float quo = num / den;
   printf("%.8f\n", quo);    // the control should never reach here, due to the exception thrown above
}

Ardennes

This doesn't seem to work on MinGW 4.6.2 ('_clearfp()' was not declared in this scope) – Varietal