OK.  The problem is that __nanf is not correctly aligned, so that when 
read as a double we get garbage.  If you add the following to the top of 
your program, then this overrides the libc version of __nanf and we get 
the correct results.

const char __nanf[] __attribute__((aligned(2))) = { 0, 0, (char)0xc0, 0x7f 
};

The other way to write it (of course) is

const unsigned __nanf[] = { 0x7fc00000 };

This should be send-pr'ed.  A similar fix is also required for Infinity: 
we have just been lucky there so far.

R.

> 
> #include <math.h>
> #include <stdio.h>
> 
> int main()
> {
>   printf("+/-inifinity, EARLY:\n");
>   printf(" HUGE_VAL: % g\n", HUGE_VAL);
>   printf("-HUGE_VAL: % g\n",-HUGE_VAL);
> 
>   printf("invalid:\n");
> #ifdef NAN
>   printf("NAN: %g, isnan(NAN): %i\n",NAN,isnan(NAN));
> #else
>   printf("2min: %g\n",DBL_MIN+DBL_MIN);
> #endif
> 
>   return 0;
> }
> 
> 
> +/-inifinity, EARLY:
>  HUGE_VAL:  Inf
> -HUGE_VAL: -Inf
> invalid:
> NAN: 3.31756e-39, isnan(NAN): 0
> 
> 
> and from /sys/arch/arm32/include/math.h
> 
> #define NAN             (*(__const float *)(__const void *)__nanf)
> 
> cf i386:
> 
> +/-inifinity, EARLY:
>  HUGE_VAL:  Inf
> -HUGE_VAL: -Inf
> invalid:
> NAN: NaN, isnan(NAN): 1
> 
> 
> Cheers,
> 
> Patrick