248 static __inline __m128 __attribute__((__always_inline__, __nodebug__))
249 _mm_permutevar_ps(__m128 a, __m128i c)  in _mm_permutevar_ps()
251   return (__m128)__builtin_ia32_vpermilvarps((__v4sf)a, (__v4si)c);  in _mm_permutevar_ps()
274   __m128 __A = (A); \
275   (__m128)__builtin_shufflevector((__v4sf)__A, (__v4sf) _mm_setzero_ps(), \
392   __m128 __a = (a); \
393   __m128 __b = (b); \
394   (__m128)__builtin_ia32_cmpps((__v4sf)__a, (__v4sf)__b, (c)); })
412   __m128 __a = (a); \
413   __m128 __b = (b); \
414   (__m128)__builtin_ia32_cmpss((__v4sf)__a, (__v4sf)__b, (c)); })
423   (__m128)__builtin_ia32_vextractf128_ps256((__v8sf)__A, (O)); })
467   __m128 __V2 = (V2); \
522 static __inline __m128 __attribute__((__always_inline__, __nodebug__))
525   return (__m128)__builtin_ia32_cvtpd2ps256((__v4df) a);  in _mm256_cvtpd_ps()
535 _mm256_cvtps_pd(__m128 a)  in _mm256_cvtps_pd()
622 _mm_testz_ps(__m128 a, __m128 b)  in _mm_testz_ps()
628 _mm_testc_ps(__m128 a, __m128 b)  in _mm_testc_ps()
634 _mm_testnzc_ps(__m128 a, __m128 b)  in _mm_testnzc_ps()
720 static __inline __m128 __attribute__((__always_inline__, __nodebug__))
723   return (__m128)__builtin_ia32_vbroadcastss(a);  in _mm_broadcast_ss()
745 _mm256_broadcast_ps(__m128 const *a)  in _mm256_broadcast_ps()
852 static __inline __m128 __attribute__((__always_inline__, __nodebug__))
853 _mm_maskload_ps(float const *p, __m128 m)  in _mm_maskload_ps()
855   return (__m128)__builtin_ia32_maskloadps((const __v4sf *)p, (__v4sf)m);  in _mm_maskload_ps()
884 _mm_maskstore_ps(float *p, __m128 m, __m128 a)  in _mm_maskstore_ps()
1117 static __inline __m128 __attribute__((__always_inline__, __nodebug__))
1137 _mm256_castps128_ps256(__m128 in)  in _mm256_castps128_ps256()
1139   __m128 zero = _mm_setzero_ps();  in _mm256_castps128_ps256()
1155     __m128 v;  in _mm256_loadu2_m128()
1187   __m128 v128;  in _mm256_storeu2_m128()