Lines Matching defs:__a
43 _mm_add_sd(__m128d __a, __m128d __b)
45 __a[0] += __b[0];
46 return __a;
50 _mm_add_pd(__m128d __a, __m128d __b)
52 return __a + __b;
56 _mm_sub_sd(__m128d __a, __m128d __b)
58 __a[0] -= __b[0];
59 return __a;
63 _mm_sub_pd(__m128d __a, __m128d __b)
65 return __a - __b;
69 _mm_mul_sd(__m128d __a, __m128d __b)
71 __a[0] *= __b[0];
72 return __a;
76 _mm_mul_pd(__m128d __a, __m128d __b)
78 return __a * __b;
82 _mm_div_sd(__m128d __a, __m128d __b)
84 __a[0] /= __b[0];
85 return __a;
89 _mm_div_pd(__m128d __a, __m128d __b)
91 return __a / __b;
95 _mm_sqrt_sd(__m128d __a, __m128d __b)
98 return (__m128d) { __c[0], __a[1] };
102 _mm_sqrt_pd(__m128d __a)
104 return __builtin_ia32_sqrtpd(__a);
108 _mm_min_sd(__m128d __a, __m128d __b)
110 return __builtin_ia32_minsd(__a, __b);
114 _mm_min_pd(__m128d __a, __m128d __b)
116 return __builtin_ia32_minpd(__a, __b);
120 _mm_max_sd(__m128d __a, __m128d __b)
122 return __builtin_ia32_maxsd(__a, __b);
126 _mm_max_pd(__m128d __a, __m128d __b)
128 return __builtin_ia32_maxpd(__a, __b);
132 _mm_and_pd(__m128d __a, __m128d __b)
134 return (__m128d)((__v4si)__a & (__v4si)__b);
138 _mm_andnot_pd(__m128d __a, __m128d __b)
140 return (__m128d)(~(__v4si)__a & (__v4si)__b);
144 _mm_or_pd(__m128d __a, __m128d __b)
146 return (__m128d)((__v4si)__a | (__v4si)__b);
150 _mm_xor_pd(__m128d __a, __m128d __b)
152 return (__m128d)((__v4si)__a ^ (__v4si)__b);
156 _mm_cmpeq_pd(__m128d __a, __m128d __b)
158 return (__m128d)__builtin_ia32_cmppd(__a, __b, 0);
162 _mm_cmplt_pd(__m128d __a, __m128d __b)
164 return (__m128d)__builtin_ia32_cmppd(__a, __b, 1);
168 _mm_cmple_pd(__m128d __a, __m128d __b)
170 return (__m128d)__builtin_ia32_cmppd(__a, __b, 2);
174 _mm_cmpgt_pd(__m128d __a, __m128d __b)
176 return (__m128d)__builtin_ia32_cmppd(__b, __a, 1);
180 _mm_cmpge_pd(__m128d __a, __m128d __b)
182 return (__m128d)__builtin_ia32_cmppd(__b, __a, 2);
186 _mm_cmpord_pd(__m128d __a, __m128d __b)
188 return (__m128d)__builtin_ia32_cmppd(__a, __b, 7);
192 _mm_cmpunord_pd(__m128d __a, __m128d __b)
194 return (__m128d)__builtin_ia32_cmppd(__a, __b, 3);
198 _mm_cmpneq_pd(__m128d __a, __m128d __b)
200 return (__m128d)__builtin_ia32_cmppd(__a, __b, 4);
204 _mm_cmpnlt_pd(__m128d __a, __m128d __b)
206 return (__m128d)__builtin_ia32_cmppd(__a, __b, 5);
210 _mm_cmpnle_pd(__m128d __a, __m128d __b)
212 return (__m128d)__builtin_ia32_cmppd(__a, __b, 6);
216 _mm_cmpngt_pd(__m128d __a, __m128d __b)
218 return (__m128d)__builtin_ia32_cmppd(__b, __a, 5);
222 _mm_cmpnge_pd(__m128d __a, __m128d __b)
224 return (__m128d)__builtin_ia32_cmppd(__b, __a, 6);
228 _mm_cmpeq_sd(__m128d __a, __m128d __b)
230 return (__m128d)__builtin_ia32_cmpsd(__a, __b, 0);
234 _mm_cmplt_sd(__m128d __a, __m128d __b)
236 return (__m128d)__builtin_ia32_cmpsd(__a, __b, 1);
240 _mm_cmple_sd(__m128d __a, __m128d __b)
242 return (__m128d)__builtin_ia32_cmpsd(__a, __b, 2);
246 _mm_cmpgt_sd(__m128d __a, __m128d __b)
248 __m128d __c = __builtin_ia32_cmpsd(__b, __a, 1);
249 return (__m128d) { __c[0], __a[1] };
253 _mm_cmpge_sd(__m128d __a, __m128d __b)
255 __m128d __c = __builtin_ia32_cmpsd(__b, __a, 2);
256 return (__m128d) { __c[0], __a[1] };
260 _mm_cmpord_sd(__m128d __a, __m128d __b)
262 return (__m128d)__builtin_ia32_cmpsd(__a, __b, 7);
266 _mm_cmpunord_sd(__m128d __a, __m128d __b)
268 return (__m128d)__builtin_ia32_cmpsd(__a, __b, 3);
272 _mm_cmpneq_sd(__m128d __a, __m128d __b)
274 return (__m128d)__builtin_ia32_cmpsd(__a, __b, 4);
278 _mm_cmpnlt_sd(__m128d __a, __m128d __b)
280 return (__m128d)__builtin_ia32_cmpsd(__a, __b, 5);
284 _mm_cmpnle_sd(__m128d __a, __m128d __b)
286 return (__m128d)__builtin_ia32_cmpsd(__a, __b, 6);
290 _mm_cmpngt_sd(__m128d __a, __m128d __b)
292 __m128d __c = __builtin_ia32_cmpsd(__b, __a, 5);
293 return (__m128d) { __c[0], __a[1] };
297 _mm_cmpnge_sd(__m128d __a, __m128d __b)
299 __m128d __c = __builtin_ia32_cmpsd(__b, __a, 6);
300 return (__m128d) { __c[0], __a[1] };
304 _mm_comieq_sd(__m128d __a, __m128d __b)
306 return __builtin_ia32_comisdeq(__a, __b);
310 _mm_comilt_sd(__m128d __a, __m128d __b)
312 return __builtin_ia32_comisdlt(__a, __b);
316 _mm_comile_sd(__m128d __a, __m128d __b)
318 return __builtin_ia32_comisdle(__a, __b);
322 _mm_comigt_sd(__m128d __a, __m128d __b)
324 return __builtin_ia32_comisdgt(__a, __b);
328 _mm_comige_sd(__m128d __a, __m128d __b)
330 return __builtin_ia32_comisdge(__a, __b);
334 _mm_comineq_sd(__m128d __a, __m128d __b)
336 return __builtin_ia32_comisdneq(__a, __b);
340 _mm_ucomieq_sd(__m128d __a, __m128d __b)
342 return __builtin_ia32_ucomisdeq(__a, __b);
346 _mm_ucomilt_sd(__m128d __a, __m128d __b)
348 return __builtin_ia32_ucomisdlt(__a, __b);
352 _mm_ucomile_sd(__m128d __a, __m128d __b)
354 return __builtin_ia32_ucomisdle(__a, __b);
358 _mm_ucomigt_sd(__m128d __a, __m128d __b)
360 return __builtin_ia32_ucomisdgt(__a, __b);
364 _mm_ucomige_sd(__m128d __a, __m128d __b)
366 return __builtin_ia32_ucomisdge(__a, __b);
370 _mm_ucomineq_sd(__m128d __a, __m128d __b)
372 return __builtin_ia32_ucomisdneq(__a, __b);
376 _mm_cvtpd_ps(__m128d __a)
378 return __builtin_ia32_cvtpd2ps(__a);
382 _mm_cvtps_pd(__m128 __a)
384 return __builtin_ia32_cvtps2pd(__a);
388 _mm_cvtepi32_pd(__m128i __a)
390 return __builtin_ia32_cvtdq2pd((__v4si)__a);
394 _mm_cvtpd_epi32(__m128d __a)
396 return __builtin_ia32_cvtpd2dq(__a);
400 _mm_cvtsd_si32(__m128d __a)
402 return __builtin_ia32_cvtsd2si(__a);
406 _mm_cvtsd_ss(__m128 __a, __m128d __b)
408 __a[0] = __b[0];
409 return __a;
413 _mm_cvtsi32_sd(__m128d __a, int __b)
415 __a[0] = __b;
416 return __a;
420 _mm_cvtss_sd(__m128d __a, __m128 __b)
422 __a[0] = __b[0];
423 return __a;
427 _mm_cvttpd_epi32(__m128d __a)
429 return (__m128i)__builtin_ia32_cvttpd2dq(__a);
433 _mm_cvttsd_si32(__m128d __a)
435 return __a[0];
439 _mm_cvtpd_pi32(__m128d __a)
441 return (__m64)__builtin_ia32_cvtpd2pi(__a);
445 _mm_cvttpd_pi32(__m128d __a)
447 return (__m64)__builtin_ia32_cvttpd2pi(__a);
451 _mm_cvtpi32_pd(__m64 __a)
453 return __builtin_ia32_cvtpi2pd((__v2si)__a);
457 _mm_cvtsd_f64(__m128d __a)
459 return __a[0];
507 _mm_loadh_pd(__m128d __a, double const *__dp)
513 return (__m128d){ __a[0], __u };
517 _mm_loadl_pd(__m128d __a, double const *__dp)
523 return (__m128d){ __u, __a[1] };
557 _mm_move_sd(__m128d __a, __m128d __b)
559 return (__m128d){ __b[0], __a[1] };
563 _mm_store_sd(double *__dp, __m128d __a)
568 ((struct __mm_store_sd_struct*)__dp)->__u = __a[0];
572 _mm_store1_pd(double *__dp, __m128d __a)
577 ((struct __mm_store1_pd_struct*)__dp)->__u[0] = __a[0];
578 ((struct __mm_store1_pd_struct*)__dp)->__u[1] = __a[0];
582 _mm_store_pd(double *__dp, __m128d __a)
584 *(__m128d *)__dp = __a;
588 _mm_storeu_pd(double *__dp, __m128d __a)
590 __builtin_ia32_storeupd(__dp, __a);
594 _mm_storer_pd(double *__dp, __m128d __a)
596 __a = __builtin_shufflevector(__a, __a, 1, 0);
597 *(__m128d *)__dp = __a;
601 _mm_storeh_pd(double *__dp, __m128d __a)
606 ((struct __mm_storeh_pd_struct*)__dp)->__u = __a[1];
610 _mm_storel_pd(double *__dp, __m128d __a)
615 ((struct __mm_storeh_pd_struct*)__dp)->__u = __a[0];
619 _mm_add_epi8(__m128i __a, __m128i __b)
621 return (__m128i)((__v16qi)__a + (__v16qi)__b);
625 _mm_add_epi16(__m128i __a, __m128i __b)
627 return (__m128i)((__v8hi)__a + (__v8hi)__b);
631 _mm_add_epi32(__m128i __a, __m128i __b)
633 return (__m128i)((__v4si)__a + (__v4si)__b);
637 _mm_add_si64(__m64 __a, __m64 __b)
639 return __a + __b;
643 _mm_add_epi64(__m128i __a, __m128i __b)
645 return __a + __b;
649 _mm_adds_epi8(__m128i __a, __m128i __b)
651 return (__m128i)__builtin_ia32_paddsb128((__v16qi)__a, (__v16qi)__b);
655 _mm_adds_epi16(__m128i __a, __m128i __b)
657 return (__m128i)__builtin_ia32_paddsw128((__v8hi)__a, (__v8hi)__b);
661 _mm_adds_epu8(__m128i __a, __m128i __b)
663 return (__m128i)__builtin_ia32_paddusb128((__v16qi)__a, (__v16qi)__b);
667 _mm_adds_epu16(__m128i __a, __m128i __b)
669 return (__m128i)__builtin_ia32_paddusw128((__v8hi)__a, (__v8hi)__b);
673 _mm_avg_epu8(__m128i __a, __m128i __b)
675 return (__m128i)__builtin_ia32_pavgb128((__v16qi)__a, (__v16qi)__b);
679 _mm_avg_epu16(__m128i __a, __m128i __b)
681 return (__m128i)__builtin_ia32_pavgw128((__v8hi)__a, (__v8hi)__b);
685 _mm_madd_epi16(__m128i __a, __m128i __b)
687 return (__m128i)__builtin_ia32_pmaddwd128((__v8hi)__a, (__v8hi)__b);
691 _mm_max_epi16(__m128i __a, __m128i __b)
693 return (__m128i)__builtin_ia32_pmaxsw128((__v8hi)__a, (__v8hi)__b);
697 _mm_max_epu8(__m128i __a, __m128i __b)
699 return (__m128i)__builtin_ia32_pmaxub128((__v16qi)__a, (__v16qi)__b);
703 _mm_min_epi16(__m128i __a, __m128i __b)
705 return (__m128i)__builtin_ia32_pminsw128((__v8hi)__a, (__v8hi)__b);
709 _mm_min_epu8(__m128i __a, __m128i __b)
711 return (__m128i)__builtin_ia32_pminub128((__v16qi)__a, (__v16qi)__b);
715 _mm_mulhi_epi16(__m128i __a, __m128i __b)
717 return (__m128i)__builtin_ia32_pmulhw128((__v8hi)__a, (__v8hi)__b);
721 _mm_mulhi_epu16(__m128i __a, __m128i __b)
723 return (__m128i)__builtin_ia32_pmulhuw128((__v8hi)__a, (__v8hi)__b);
727 _mm_mullo_epi16(__m128i __a, __m128i __b)
729 return (__m128i)((__v8hi)__a * (__v8hi)__b);
733 _mm_mul_su32(__m64 __a, __m64 __b)
735 return __builtin_ia32_pmuludq((__v2si)__a, (__v2si)__b);
739 _mm_mul_epu32(__m128i __a, __m128i __b)
741 return __builtin_ia32_pmuludq128((__v4si)__a, (__v4si)__b);
745 _mm_sad_epu8(__m128i __a, __m128i __b)
747 return __builtin_ia32_psadbw128((__v16qi)__a, (__v16qi)__b);
751 _mm_sub_epi8(__m128i __a, __m128i __b)
753 return (__m128i)((__v16qi)__a - (__v16qi)__b);
757 _mm_sub_epi16(__m128i __a, __m128i __b)
759 return (__m128i)((__v8hi)__a - (__v8hi)__b);
763 _mm_sub_epi32(__m128i __a, __m128i __b)
765 return (__m128i)((__v4si)__a - (__v4si)__b);
769 _mm_sub_si64(__m64 __a, __m64 __b)
771 return __a - __b;
775 _mm_sub_epi64(__m128i __a, __m128i __b)
777 return __a - __b;
781 _mm_subs_epi8(__m128i __a, __m128i __b)
783 return (__m128i)__builtin_ia32_psubsb128((__v16qi)__a, (__v16qi)__b);
787 _mm_subs_epi16(__m128i __a, __m128i __b)
789 return (__m128i)__builtin_ia32_psubsw128((__v8hi)__a, (__v8hi)__b);
793 _mm_subs_epu8(__m128i __a, __m128i __b)
795 return (__m128i)__builtin_ia32_psubusb128((__v16qi)__a, (__v16qi)__b);
799 _mm_subs_epu16(__m128i __a, __m128i __b)
801 return (__m128i)__builtin_ia32_psubusw128((__v8hi)__a, (__v8hi)__b);
805 _mm_and_si128(__m128i __a, __m128i __b)
807 return __a & __b;
811 _mm_andnot_si128(__m128i __a, __m128i __b)
813 return ~__a & __b;
817 _mm_or_si128(__m128i __a, __m128i __b)
819 return __a | __b;
823 _mm_xor_si128(__m128i __a, __m128i __b)
825 return __a ^ __b;
830 __m128i __a = (a); \
832 (__m128i)__builtin_ia32_pslldqi128(__a, (count)*8); })
835 _mm_slli_epi16(__m128i __a, int __count)
837 return (__m128i)__builtin_ia32_psllwi128((__v8hi)__a, __count);
841 _mm_sll_epi16(__m128i __a, __m128i __count)
843 return (__m128i)__builtin_ia32_psllw128((__v8hi)__a, (__v8hi)__count);
847 _mm_slli_epi32(__m128i __a, int __count)
849 return (__m128i)__builtin_ia32_pslldi128((__v4si)__a, __count);
853 _mm_sll_epi32(__m128i __a, __m128i __count)
855 return (__m128i)__builtin_ia32_pslld128((__v4si)__a, (__v4si)__count);
859 _mm_slli_epi64(__m128i __a, int __count)
861 return __builtin_ia32_psllqi128(__a, __count);
865 _mm_sll_epi64(__m128i __a, __m128i __count)
867 return __builtin_ia32_psllq128(__a, __count);
871 _mm_srai_epi16(__m128i __a, int __count)
873 return (__m128i)__builtin_ia32_psrawi128((__v8hi)__a, __count);
877 _mm_sra_epi16(__m128i __a, __m128i __count)
879 return (__m128i)__builtin_ia32_psraw128((__v8hi)__a, (__v8hi)__count);
883 _mm_srai_epi32(__m128i __a, int __count)
885 return (__m128i)__builtin_ia32_psradi128((__v4si)__a, __count);
889 _mm_sra_epi32(__m128i __a, __m128i __count)
891 return (__m128i)__builtin_ia32_psrad128((__v4si)__a, (__v4si)__count);
897 __m128i __a = (a); \
899 (__m128i)__builtin_ia32_psrldqi128(__a, (count)*8); })
902 _mm_srli_epi16(__m128i __a, int __count)
904 return (__m128i)__builtin_ia32_psrlwi128((__v8hi)__a, __count);
908 _mm_srl_epi16(__m128i __a, __m128i __count)
910 return (__m128i)__builtin_ia32_psrlw128((__v8hi)__a, (__v8hi)__count);
914 _mm_srli_epi32(__m128i __a, int __count)
916 return (__m128i)__builtin_ia32_psrldi128((__v4si)__a, __count);
920 _mm_srl_epi32(__m128i __a, __m128i __count)
922 return (__m128i)__builtin_ia32_psrld128((__v4si)__a, (__v4si)__count);
926 _mm_srli_epi64(__m128i __a, int __count)
928 return __builtin_ia32_psrlqi128(__a, __count);
932 _mm_srl_epi64(__m128i __a, __m128i __count)
934 return __builtin_ia32_psrlq128(__a, __count);
938 _mm_cmpeq_epi8(__m128i __a, __m128i __b)
940 return (__m128i)((__v16qi)__a == (__v16qi)__b);
944 _mm_cmpeq_epi16(__m128i __a, __m128i __b)
946 return (__m128i)((__v8hi)__a == (__v8hi)__b);
950 _mm_cmpeq_epi32(__m128i __a, __m128i __b)
952 return (__m128i)((__v4si)__a == (__v4si)__b);
956 _mm_cmpgt_epi8(__m128i __a, __m128i __b)
961 return (__m128i)((__v16qs)__a > (__v16qs)__b);
965 _mm_cmpgt_epi16(__m128i __a, __m128i __b)
967 return (__m128i)((__v8hi)__a > (__v8hi)__b);
971 _mm_cmpgt_epi32(__m128i __a, __m128i __b)
973 return (__m128i)((__v4si)__a > (__v4si)__b);
977 _mm_cmplt_epi8(__m128i __a, __m128i __b)
979 return _mm_cmpgt_epi8(__b, __a);
983 _mm_cmplt_epi16(__m128i __a, __m128i __b)
985 return _mm_cmpgt_epi16(__b, __a);
989 _mm_cmplt_epi32(__m128i __a, __m128i __b)
991 return _mm_cmpgt_epi32(__b, __a);
996 _mm_cvtsi64_sd(__m128d __a, long long __b)
998 __a[0] = __b;
999 return __a;
1003 _mm_cvtsd_si64(__m128d __a)
1005 return __builtin_ia32_cvtsd2si64(__a);
1009 _mm_cvttsd_si64(__m128d __a)
1011 return __a[0];
1016 _mm_cvtepi32_ps(__m128i __a)
1018 return __builtin_ia32_cvtdq2ps((__v4si)__a);
1022 _mm_cvtps_epi32(__m128 __a)
1024 return (__m128i)__builtin_ia32_cvtps2dq(__a);
1028 _mm_cvttps_epi32(__m128 __a)
1030 return (__m128i)__builtin_ia32_cvttps2dq(__a);
1034 _mm_cvtsi32_si128(int __a)
1036 return (__m128i)(__v4si){ __a, 0, 0, 0 };
1041 _mm_cvtsi64_si128(long long __a)
1043 return (__m128i){ __a, 0 };
1048 _mm_cvtsi128_si32(__m128i __a)
1050 __v4si __b = (__v4si)__a;
1056 _mm_cvtsi128_si64(__m128i __a)
1058 return __a[0];
1195 _mm_storel_epi64(__m128i *__p, __m128i __a)
1200 ((struct __mm_storel_epi64_struct*)__p)->__u = __a[0];
1204 _mm_stream_pd(double *__p, __m128d __a)
1206 __builtin_ia32_movntpd(__p, __a);
1210 _mm_stream_si128(__m128i *__p, __m128i __a)
1212 __builtin_ia32_movntdq(__p, __a);
1216 _mm_stream_si32(int *__p, int __a)
1218 __builtin_ia32_movnti(__p, __a);
1223 _mm_stream_si64(long long *__p, long long __a)
1225 __builtin_ia32_movnti64(__p, __a);
1248 _mm_packs_epi16(__m128i __a, __m128i __b)
1250 return (__m128i)__builtin_ia32_packsswb128((__v8hi)__a, (__v8hi)__b);
1254 _mm_packs_epi32(__m128i __a, __m128i __b)
1256 return (__m128i)__builtin_ia32_packssdw128((__v4si)__a, (__v4si)__b);
1260 _mm_packus_epi16(__m128i __a, __m128i __b)
1262 return (__m128i)__builtin_ia32_packuswb128((__v8hi)__a, (__v8hi)__b);
1266 _mm_extract_epi16(__m128i __a, int __imm)
1268 __v8hi __b = (__v8hi)__a;
1273 _mm_insert_epi16(__m128i __a, int __b, int __imm)
1275 __v8hi __c = (__v8hi)__a;
1281 _mm_movemask_epi8(__m128i __a)
1283 return __builtin_ia32_pmovmskb128((__v16qi)__a);
1288 __m128i __a = (a); \
1290 (__m128i)__builtin_shufflevector((__v4si)__a, (__v4si) _mm_set1_epi32(0), \
1296 __m128i __a = (a); \
1298 (__m128i)__builtin_shufflevector((__v8hi)__a, (__v8hi) _mm_set1_epi16(0), \
1305 __m128i __a = (a); \
1307 (__m128i)__builtin_shufflevector((__v8hi)__a, (__v8hi) _mm_set1_epi16(0), \
1315 _mm_unpackhi_epi8(__m128i __a, __m128i __b)
1317 return (__m128i)__builtin_shufflevector((__v16qi)__a, (__v16qi)__b, 8, 16+8, 9, 16+9, 10, 16+10, 11, 16+11, 12, 16+12, 13, 16+13, 14, 16+14, 15, 16+15);
1321 _mm_unpackhi_epi16(__m128i __a, __m128i __b)
1323 return (__m128i)__builtin_shufflevector((__v8hi)__a, (__v8hi)__b, 4, 8+4, 5, 8+5, 6, 8+6, 7, 8+7);
1327 _mm_unpackhi_epi32(__m128i __a, __m128i __b)
1329 return (__m128i)__builtin_shufflevector((__v4si)__a, (__v4si)__b, 2, 4+2, 3, 4+3);
1333 _mm_unpackhi_epi64(__m128i __a, __m128i __b)
1335 return (__m128i)__builtin_shufflevector(__a, __b, 1, 2+1);
1339 _mm_unpacklo_epi8(__m128i __a, __m128i __b)
1341 return (__m128i)__builtin_shufflevector((__v16qi)__a, (__v16qi)__b, 0, 16+0, 1, 16+1, 2, 16+2, 3, 16+3, 4, 16+4, 5, 16+5, 6, 16+6, 7, 16+7);
1345 _mm_unpacklo_epi16(__m128i __a, __m128i __b)
1347 return (__m128i)__builtin_shufflevector((__v8hi)__a, (__v8hi)__b, 0, 8+0, 1, 8+1, 2, 8+2, 3, 8+3);
1351 _mm_unpacklo_epi32(__m128i __a, __m128i __b)
1353 return (__m128i)__builtin_shufflevector((__v4si)__a, (__v4si)__b, 0, 4+0, 1, 4+1);
1357 _mm_unpacklo_epi64(__m128i __a, __m128i __b)
1359 return (__m128i)__builtin_shufflevector(__a, __b, 0, 2+0);
1363 _mm_movepi64_pi64(__m128i __a)
1365 return (__m64)__a[0];
1369 _mm_movpi64_epi64(__m64 __a)
1371 return (__m128i){ (long long)__a, 0 };
1375 _mm_move_epi64(__m128i __a)
1377 return __builtin_shufflevector(__a, (__m128i){ 0 }, 0, 2);
1381 _mm_unpackhi_pd(__m128d __a, __m128d __b)
1383 return __builtin_shufflevector(__a, __b, 1, 2+1);
1387 _mm_unpacklo_pd(__m128d __a, __m128d __b)
1389 return __builtin_shufflevector(__a, __b, 0, 2+0);
1393 _mm_movemask_pd(__m128d __a)
1395 return __builtin_ia32_movmskpd(__a);
1400 __m128d __a = (a); \
1403 __builtin_shufflevector(__a, __b, (i) & 1, (((i) & 2) >> 1) + 2); })
1406 _mm_castpd_ps(__m128d __a)
1408 return (__m128)__a;
1412 _mm_castpd_si128(__m128d __a)
1414 return (__m128i)__a;
1418 _mm_castps_pd(__m128 __a)
1420 return (__m128d)__a;
1424 _mm_castps_si128(__m128 __a)
1426 return (__m128i)__a;
1430 _mm_castsi128_ps(__m128i __a)
1432 return (__m128)__a;
1436 _mm_castsi128_pd(__m128i __a)
1438 return (__m128d)__a;