int64 with MMX

struct int64 {
    int64(): val(_mm_setzero_si64()) {}
    int64(__m64 v): val(v) {}
    int64(int v): val(_mm_cvtsi32_si64(v)) {}

    operator __m64() { return val; }
    operator int() { return _mm_cvtsi64_si32(val); }

    friend int64 operator +(int64 a, int64 b) {
        return _mm_add_si64(a, b);
    }
    int64& operator +=(int64 rhs) {
        return *this = *this + rhs;
    }

    friend int64 operator -(int64 a, int64 b) {
        return _mm_sub_si64(a, b);
    }
    int64& operator -=(int64 rhs) {
        return *this = *this - rhs;
    }

    friend bool operator ==(int64 a, int64 b) {
        return _mm_movemask_pi8(_mm_cmpeq_pi32(a, b)) == ~0;
    }

    friend bool operator !=(int64 a, int64 b) {
        return !(a == b);
    }

    friend bool operator >(int64 a, int64 b) {
        if (a.hi > b.hi) {
            return true;
        } else {
            if (a.hi == b.hi) {
                return a.lo > b.lo;
            }
        }
        return false;
    }

    friend int64 operator *(int64 a, int64 b) {
        int64 res = int64::mul32(a, b);
        int t = a.hi * b.lo + a.lo * b.hi;
        return res + _mm_set_pi32(t, 0);
    }
    int64& operator *=(int64 rhs) {
        return *this = *this * rhs;
    }

    static int64 mul32(int a, int b) {
        return _mm_mul_su32(int64(a), int64(b));
    }

    union {
        struct {
            int hi, lo;
        };
        __m64 val;
    };
};

int main() {
    using namespace std;
    int64 a = 23456234, b = 623462345;
    int64 c = int64::mul32(a, b);
    cout << (_mm_empty(), c.val.m64_i64) << endl;
    cout << (c==c);
    cout << (_mm_empty(), c.val.m64_i64);
}