Real.maxFinite bug

[Stephen Weeks]

I have tracked a bug in MLton's implementation of Real.maxFinite down to the
following C example.  (If you try using Real.maxFinite in *any* program, the
resulting executable will fail to terminate.)

The code below runs as I would expect when compiled gcc -O0, but fails to
terminate when compiled gcc -O2.  Furthermore, if you uncomment the fprintf and
compile gcc -O2, the program terminates as expected.  I'm guessing the problem
has something to do with extended precision, but I can't see how, since my
understanding of the C calling convention is that the float argument is passed
in 64 bits on the stack.  Thus I can't understand how uncommenting the fprintf
changes behavior.  

Any ideas?

--------------------------------------------------------------------------------


#include <stdio.h>

typedef double Double;
typedef int Int;
typedef unsigned int Word;

#define EXPONENT_MASK 0x7FF00000

Int Real_isFinite(Double d) {
	Word word1;

/*	fprintf(stderr, "Real_isFinite(%.20e)\n", d); */
	word1 = ((Word *)&d)[1];

	return !((word1 & EXPONENT_MASK) == EXPONENT_MASK);
}

int main ()
{
	Double x, y;
	
	x = 1.0;
loop:
	y = x * 2.0;
	if (Real_isFinite(y)) {
		x = y;
		goto loop;
	}
	fprintf(stderr, "max is %.20e\n", x);
	return 0;
}