FEMMS required in amd64 environments?

[seeker010]

#include <iostream>

using namespace std;

typedef float v2sf __attribute__ ((vector_size (2*sizeof(float))));

union f2vector{
	v2sf v;
	float f[2];
};

const f2vector operator*(const f2vector &a, const f2vector &b){
	f2vector tmp;
        tmp.v =  __builtin_ia32_pfmul (a.v,b.v);
        return tmp;
}

int main(){
	union f2vector a, b, c;
	a.f[0] = 1; a.f[1] = 2;
	b.f[0] = 5; b.f[1] = 6;
	c = a * b;
	cout << c.f[0] << ' ' << c.f[1] << endl;
	return 0;
}

when compiled with -m32 -m3dnow returns nan 12, but

#include <iostream>

using namespace std;

typedef float v2sf __attribute__ ((vector_size (2*sizeof(float))));

union f2vector{
	v2sf v;
	float f[2];
};

const f2vector operator*(const f2vector &a, const f2vector &b){
	__builtin_ia32_femms();
	f2vector tmp;
        tmp.v =  __builtin_ia32_pfmul (a.v,b.v);
	__builtin_ia32_femms();
        return tmp;
}

int main(){
	union f2vector a, b, c;
	a.f[0] = 1; a.f[1] = 2;
	b.f[0] = 5; b.f[1] = 6;
	c = a * b;
	cout << c.f[0] << ' ' << c.f[1] << endl;
	return 0;
}

returns 5 12

I thought FEMMS was optional on athlon hardware because 3dnow no longer maps on to the fpu registers?