_mm_movehl_ps
(r1) = (__m128i)_mm_movehl_ps((__m128)_mm_tmp2, (__m128)_mm_tmp0); \
(r3) = (__m128i)_mm_movehl_ps((__m128)_mm_tmp3, (__m128)_mm_tmp1); \
return (__m128i)_mm_movehl_ps((__m128)rk, _mm_setzero_ps());
return (__m128i)_mm_movehl_ps(hi, lo);
return (__m128i)_mm_movehl_ps(hi, lo);