Ordering functor fails for optimized build with GCC-4.0.3 on athlon64 dual core

[Erik]

Hello,
I have created a 32-bit type called Coords, consisting of 2 16-bit 
values x and y. Then I created an ordering functor for use with standard 
containers. It just compares the 32-bit values. This should be the 
fastest way to do it. See this code:

#include <SDL_types.h>

struct Coords {
    struct ordering_functor
    {bool operator()(const Coords a, const Coords b) const;};

    int x : 16, y : 16;
};

bool Coords::ordering_functor::operator()(const Coords a, const Coords 
b) const {
    return
        *reinterpret_cast<const Uint32 * const>(&a)
        <
        *reinterpret_cast<const Uint32 * const>(&b);
}


This works great for me on Pentium Mobile and gcc-3.4.5, but someone 
else with athlon64 dual core and gcc-4.0.3 has problems. It works for 
him without optimization, but fails with optimization. I asked him for 
the assembly output of a slow build and an optimized build. This is what 
he sent me:
SLOW:

______________________________________________________________________________________

.LFB2:
	pushl	%ebp
.LCFI0:
	movl	%esp, %ebp
.LCFI1:
	leal	12(%ebp), %eax
	movl	(%eax), %edx
	leal	16(%ebp), %eax
	movl	(%eax), %eax
	cmpl	%eax, %edx
	setb	%al
	movzbl	%al, %eax
	popl	%ebp
	ret
______________________________________________________________________________________


OPTIMIZED:
______________________________________________________________________________________
.LFB2:
	pushl	%ebp
.LCFI0:
	movl	%esp, %ebp
.LCFI1:
	movl	16(%ebp), %eax
	cmpl	%eax, 12(%ebp)
	popl	%ebp
	setb	%al
	movzbl	%al, %eax
	ret
______________________________________________________________________________________


What should be done about this? (I can not use a union, because I need to have a constructor for Coords.)

thanks in advance,
Erik