__m128
(r0) = (__m128i)_mm_movelh_ps((__m128)_mm_tmp0, (__m128)_mm_tmp2); \
(r1) = (__m128i)_mm_movehl_ps((__m128)_mm_tmp2, (__m128)_mm_tmp0); \
(r2) = (__m128i)_mm_movelh_ps((__m128)_mm_tmp1, (__m128)_mm_tmp3); \
(r3) = (__m128i)_mm_movehl_ps((__m128)_mm_tmp3, (__m128)_mm_tmp1); \
return (__m128i)_mm_movehl_ps((__m128)rk, _mm_setzero_ps());
static __inline __m128
return __extension__ (__m128)(__v4sf) { *__p, *__p, *__p, *__p };
static __inline __m128
_mm_movehl_ps(__m128 __v0, __m128 __v1)
return (__m128)__builtin_ia32_movhlps((__v4sf)__v0, (__v4sf)__v1);
static __inline __m128
_mm_movelh_ps(__m128 __v0, __m128 __v1)
return (__m128)__builtin_ia32_movlhps((__v4sf)__v0, (__v4sf)__v1);
static __inline __m128
return __extension__ (__m128){ 0, 0, 0, 0 };
(__m128)__builtin_ia32_shufps((__v4sf)(__m128)(x), \
(__v4sf)(__m128)(y), (int)(m)) \
__m128 lo = (__m128)_mm_unpacklo_epi32(a, b); /* (a[0], b[0], ...) */
__m128 hi = (__m128)_mm_unpacklo_epi32(c, d); /* (c[0], d[0], ...) */
__m128 lo = (__m128)_mm_unpacklo_epi32(a, b); /* (..., a[1], b[1]) */
__m128 hi = (__m128)_mm_unpacklo_epi32(c, d); /* (..., c[1], d[1]) */
__m128 lo = (__m128)_mm_unpackhi_epi32(a, b); /* (a[2], b[2], ...) */
__m128 hi = (__m128)_mm_unpackhi_epi32(c, d); /* (c[2], d[2], ...) */
__m128 lo = (__m128)_mm_unpackhi_epi32(a, b); /* (..., a[3], b[3]) */
__m128 hi = (__m128)_mm_unpackhi_epi32(c, d); /* (..., c[3], d[3]) */