_mm_movelh_ps
(r0) = (__m128i)_mm_movelh_ps((__m128)_mm_tmp0, (__m128)_mm_tmp2); \
(r2) = (__m128i)_mm_movelh_ps((__m128)_mm_tmp1, (__m128)_mm_tmp3); \
return (__m128i)_mm_movelh_ps(lo, hi);
return (__m128i)_mm_movelh_ps(lo, hi);