Lines Matching refs:__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 return (__m128d)__builtin_ia32_cmpsd(__b, __a, 1);
252 _mm_cmpge_sd(__m128d __a, __m128d __b)
254 return (__m128d)__builtin_ia32_cmpsd(__b, __a, 2);
258 _mm_cmpord_sd(__m128d __a, __m128d __b)
260 return (__m128d)__builtin_ia32_cmpsd(__a, __b, 7);
264 _mm_cmpunord_sd(__m128d __a, __m128d __b)
266 return (__m128d)__builtin_ia32_cmpsd(__a, __b, 3);
270 _mm_cmpneq_sd(__m128d __a, __m128d __b)
272 return (__m128d)__builtin_ia32_cmpsd(__a, __b, 4);
276 _mm_cmpnlt_sd(__m128d __a, __m128d __b)
278 return (__m128d)__builtin_ia32_cmpsd(__a, __b, 5);
282 _mm_cmpnle_sd(__m128d __a, __m128d __b)
284 return (__m128d)__builtin_ia32_cmpsd(__a, __b, 6);
288 _mm_cmpngt_sd(__m128d __a, __m128d __b)
290 return (__m128d)__builtin_ia32_cmpsd(__b, __a, 5);
294 _mm_cmpnge_sd(__m128d __a, __m128d __b)
296 return (__m128d)__builtin_ia32_cmpsd(__b, __a, 6);
300 _mm_comieq_sd(__m128d __a, __m128d __b)
302 return __builtin_ia32_comisdeq(__a, __b);
306 _mm_comilt_sd(__m128d __a, __m128d __b)
308 return __builtin_ia32_comisdlt(__a, __b);
312 _mm_comile_sd(__m128d __a, __m128d __b)
314 return __builtin_ia32_comisdle(__a, __b);
318 _mm_comigt_sd(__m128d __a, __m128d __b)
320 return __builtin_ia32_comisdgt(__a, __b);
324 _mm_comige_sd(__m128d __a, __m128d __b)
326 return __builtin_ia32_comisdge(__a, __b);
330 _mm_comineq_sd(__m128d __a, __m128d __b)
332 return __builtin_ia32_comisdneq(__a, __b);
336 _mm_ucomieq_sd(__m128d __a, __m128d __b)
338 return __builtin_ia32_ucomisdeq(__a, __b);
342 _mm_ucomilt_sd(__m128d __a, __m128d __b)
344 return __builtin_ia32_ucomisdlt(__a, __b);
348 _mm_ucomile_sd(__m128d __a, __m128d __b)
350 return __builtin_ia32_ucomisdle(__a, __b);
354 _mm_ucomigt_sd(__m128d __a, __m128d __b)
356 return __builtin_ia32_ucomisdgt(__a, __b);
360 _mm_ucomige_sd(__m128d __a, __m128d __b)
362 return __builtin_ia32_ucomisdge(__a, __b);
366 _mm_ucomineq_sd(__m128d __a, __m128d __b)
368 return __builtin_ia32_ucomisdneq(__a, __b);
372 _mm_cvtpd_ps(__m128d __a)
374 return __builtin_ia32_cvtpd2ps(__a);
378 _mm_cvtps_pd(__m128 __a)
380 return __builtin_ia32_cvtps2pd(__a);
384 _mm_cvtepi32_pd(__m128i __a)
386 return __builtin_ia32_cvtdq2pd((__v4si)__a);
390 _mm_cvtpd_epi32(__m128d __a)
392 return __builtin_ia32_cvtpd2dq(__a);
396 _mm_cvtsd_si32(__m128d __a)
398 return __builtin_ia32_cvtsd2si(__a);
402 _mm_cvtsd_ss(__m128 __a, __m128d __b)
404 __a[0] = __b[0];
405 return __a;
409 _mm_cvtsi32_sd(__m128d __a, int __b)
411 __a[0] = __b;
412 return __a;
416 _mm_cvtss_sd(__m128d __a, __m128 __b)
418 __a[0] = __b[0];
419 return __a;
423 _mm_cvttpd_epi32(__m128d __a)
425 return (__m128i)__builtin_ia32_cvttpd2dq(__a);
429 _mm_cvttsd_si32(__m128d __a)
431 return __a[0];
435 _mm_cvtpd_pi32(__m128d __a)
437 return (__m64)__builtin_ia32_cvtpd2pi(__a);
441 _mm_cvttpd_pi32(__m128d __a)
443 return (__m64)__builtin_ia32_cvttpd2pi(__a);
447 _mm_cvtpi32_pd(__m64 __a)
449 return __builtin_ia32_cvtpi2pd((__v2si)__a);
453 _mm_cvtsd_f64(__m128d __a)
455 return __a[0];
503 _mm_loadh_pd(__m128d __a, double const *__dp)
509 return (__m128d){ __a[0], __u };
513 _mm_loadl_pd(__m128d __a, double const *__dp)
519 return (__m128d){ __u, __a[1] };
553 _mm_move_sd(__m128d __a, __m128d __b)
555 return (__m128d){ __b[0], __a[1] };
559 _mm_store_sd(double *__dp, __m128d __a)
564 ((struct __mm_store_sd_struct*)__dp)->__u = __a[0];
568 _mm_store1_pd(double *__dp, __m128d __a)
573 ((struct __mm_store1_pd_struct*)__dp)->__u[0] = __a[0];
574 ((struct __mm_store1_pd_struct*)__dp)->__u[1] = __a[0];
578 _mm_store_pd(double *__dp, __m128d __a)
580 *(__m128d *)__dp = __a;
584 _mm_storeu_pd(double *__dp, __m128d __a)
586 __builtin_ia32_storeupd(__dp, __a);
590 _mm_storer_pd(double *__dp, __m128d __a)
592 __a = __builtin_shufflevector(__a, __a, 1, 0);
593 __a;
597 _mm_storeh_pd(double *__dp, __m128d __a)
602 ((struct __mm_storeh_pd_struct*)__dp)->__u = __a[1];
606 _mm_storel_pd(double *__dp, __m128d __a)
611 ((struct __mm_storeh_pd_struct*)__dp)->__u = __a[0];
615 _mm_add_epi8(__m128i __a, __m128i __b)
617 return (__m128i)((__v16qi)__a + (__v16qi)__b);
621 _mm_add_epi16(__m128i __a, __m128i __b)
623 return (__m128i)((__v8hi)__a + (__v8hi)__b);
627 _mm_add_epi32(__m128i __a, __m128i __b)
629 return (__m128i)((__v4si)__a + (__v4si)__b);
633 _mm_add_si64(__m64 __a, __m64 __b)
635 return __a + __b;
639 _mm_add_epi64(__m128i __a, __m128i __b)
641 return __a + __b;
645 _mm_adds_epi8(__m128i __a, __m128i __b)
647 return (__m128i)__builtin_ia32_paddsb128((__v16qi)__a, (__v16qi)__b);
651 _mm_adds_epi16(__m128i __a, __m128i __b)
653 return (__m128i)__builtin_ia32_paddsw128((__v8hi)__a, (__v8hi)__b);
657 _mm_adds_epu8(__m128i __a, __m128i __b)
659 return (__m128i)__builtin_ia32_paddusb128((__v16qi)__a, (__v16qi)__b);
663 _mm_adds_epu16(__m128i __a, __m128i __b)
665 return (__m128i)__builtin_ia32_paddusw128((__v8hi)__a, (__v8hi)__b);
669 _mm_avg_epu8(__m128i __a, __m128i __b)
671 return (__m128i)__builtin_ia32_pavgb128((__v16qi)__a, (__v16qi)__b);
675 _mm_avg_epu16(__m128i __a, __m128i __b)
677 return (__m128i)__builtin_ia32_pavgw128((__v8hi)__a, (__v8hi)__b);
681 _mm_madd_epi16(__m128i __a, __m128i __b)
683 return (__m128i)__builtin_ia32_pmaddwd128((__v8hi)__a, (__v8hi)__b);
687 _mm_max_epi16(__m128i __a, __m128i __b)
689 return (__m128i)__builtin_ia32_pmaxsw128((__v8hi)__a, (__v8hi)__b);
693 _mm_max_epu8(__m128i __a, __m128i __b)
695 return (__m128i)__builtin_ia32_pmaxub128((__v16qi)__a, (__v16qi)__b);
699 _mm_min_epi16(__m128i __a, __m128i __b)
701 return (__m128i)__builtin_ia32_pminsw128((__v8hi)__a, (__v8hi)__b);
705 _mm_min_epu8(__m128i __a, __m128i __b)
707 return (__m128i)__builtin_ia32_pminub128((__v16qi)__a, (__v16qi)__b);
711 _mm_mulhi_epi16(__m128i __a, __m128i __b)
713 return (__m128i)__builtin_ia32_pmulhw128((__v8hi)__a, (__v8hi)__b);
717 _mm_mulhi_epu16(__m128i __a, __m128i __b)
719 return (__m128i)__builtin_ia32_pmulhuw128((__v8hi)__a, (__v8hi)__b);
723 _mm_mullo_epi16(__m128i __a, __m128i __b)
725 return (__m128i)((__v8hi)__a * (__v8hi)__b);
729 _mm_mul_su32(__m64 __a, __m64 __b)
731 return __builtin_ia32_pmuludq((__v2si)__a, (__v2si)__b);
735 _mm_mul_epu32(__m128i __a, __m128i __b)
737 return __builtin_ia32_pmuludq128((__v4si)__a, (__v4si)__b);
741 _mm_sad_epu8(__m128i __a, __m128i __b)
743 return __builtin_ia32_psadbw128((__v16qi)__a, (__v16qi)__b);
747 _mm_sub_epi8(__m128i __a, __m128i __b)
749 return (__m128i)((__v16qi)__a - (__v16qi)__b);
753 _mm_sub_epi16(__m128i __a, __m128i __b)
755 return (__m128i)((__v8hi)__a - (__v8hi)__b);
759 _mm_sub_epi32(__m128i __a, __m128i __b)
761 return (__m128i)((__v4si)__a - (__v4si)__b);
765 _mm_sub_si64(__m64 __a, __m64 __b)
767 return __a - __b;
771 _mm_sub_epi64(__m128i __a, __m128i __b)
773 return __a - __b;
777 _mm_subs_epi8(__m128i __a, __m128i __b)
779 return (__m128i)__builtin_ia32_psubsb128((__v16qi)__a, (__v16qi)__b);
783 _mm_subs_epi16(__m128i __a, __m128i __b)
785 return (__m128i)__builtin_ia32_psubsw128((__v8hi)__a, (__v8hi)__b);
789 _mm_subs_epu8(__m128i __a, __m128i __b)
791 return (__m128i)__builtin_ia32_psubusb128((__v16qi)__a, (__v16qi)__b);
795 _mm_subs_epu16(__m128i __a, __m128i __b)
797 return (__m128i)__builtin_ia32_psubusw128((__v8hi)__a, (__v8hi)__b);
801 _mm_and_si128(__m128i __a, __m128i __b)
803 return __a & __b;
807 _mm_andnot_si128(__m128i __a, __m128i __b)
809 return ~__a & __b;
813 _mm_or_si128(__m128i __a, __m128i __b)
815 return __a | __b;
819 _mm_xor_si128(__m128i __a, __m128i __b)
821 return __a ^ __b;
825 __m128i __a = (a); \
826 (__m128i)__builtin_ia32_pslldqi128(__a, (count)*8); })
829 _mm_slli_epi16(__m128i __a, int __count)
831 return (__m128i)__builtin_ia32_psllwi128((__v8hi)__a, __count);
835 _mm_sll_epi16(__m128i __a, __m128i __count)
837 return (__m128i)__builtin_ia32_psllw128((__v8hi)__a, (__v8hi)__count);
841 _mm_slli_epi32(__m128i __a, int __count)
843 return (__m128i)__builtin_ia32_pslldi128((__v4si)__a, __count);
847 _mm_sll_epi32(__m128i __a, __m128i __count)
849 return (__m128i)__builtin_ia32_pslld128((__v4si)__a, (__v4si)__count);
853 _mm_slli_epi64(__m128i __a, int __count)
855 return __builtin_ia32_psllqi128(__a, __count);
859 _mm_sll_epi64(__m128i __a, __m128i __count)
861 return __builtin_ia32_psllq128(__a, __count);
865 _mm_srai_epi16(__m128i __a, int __count)
867 return (__m128i)__builtin_ia32_psrawi128((__v8hi)__a, __count);
871 _mm_sra_epi16(__m128i __a, __m128i __count)
873 return (__m128i)__builtin_ia32_psraw128((__v8hi)__a, (__v8hi)__count);
877 _mm_srai_epi32(__m128i __a, int __count)
879 return (__m128i)__builtin_ia32_psradi128((__v4si)__a, __count);
883 _mm_sra_epi32(__m128i __a, __m128i __count)
885 return (__m128i)__builtin_ia32_psrad128((__v4si)__a, (__v4si)__count);
890 __m128i __a = (a); \
891 (__m128i)__builtin_ia32_psrldqi128(__a, (count)*8); })
894 _mm_srli_epi16(__m128i __a, int __count)
896 return (__m128i)__builtin_ia32_psrlwi128((__v8hi)__a, __count);
900 _mm_srl_epi16(__m128i __a, __m128i __count)
902 return (__m128i)__builtin_ia32_psrlw128((__v8hi)__a, (__v8hi)__count);
906 _mm_srli_epi32(__m128i __a, int __count)
908 return (__m128i)__builtin_ia32_psrldi128((__v4si)__a, __count);
912 _mm_srl_epi32(__m128i __a, __m128i __count)
914 return (__m128i)__builtin_ia32_psrld128((__v4si)__a, (__v4si)__count);
918 _mm_srli_epi64(__m128i __a, int __count)
920 return __builtin_ia32_psrlqi128(__a, __count);
924 _mm_srl_epi64(__m128i __a, __m128i __count)
926 return __builtin_ia32_psrlq128(__a, __count);
930 _mm_cmpeq_epi8(__m128i __a, __m128i __b)
932 return (__m128i)((__v16qi)__a == (__v16qi)__b);
936 _mm_cmpeq_epi16(__m128i __a, __m128i __b)
938 return (__m128i)((__v8hi)__a == (__v8hi)__b);
942 _mm_cmpeq_epi32(__m128i __a, __m128i __b)
944 return (__m128i)((__v4si)__a == (__v4si)__b);
948 _mm_cmpgt_epi8(__m128i __a, __m128i __b)
953 return (__m128i)((__v16qs)__a > (__v16qs)__b);
957 _mm_cmpgt_epi16(__m128i __a, __m128i __b)
959 return (__m128i)((__v8hi)__a > (__v8hi)__b);
963 _mm_cmpgt_epi32(__m128i __a, __m128i __b)
965 return (__m128i)((__v4si)__a > (__v4si)__b);
969 _mm_cmplt_epi8(__m128i __a, __m128i __b)
971 return _mm_cmpgt_epi8(__b, __a);
975 _mm_cmplt_epi16(__m128i __a, __m128i __b)
977 return _mm_cmpgt_epi16(__b, __a);
981 _mm_cmplt_epi32(__m128i __a, __m128i __b)
983 return _mm_cmpgt_epi32(__b, __a);
988 _mm_cvtsi64_sd(__m128d __a, long long __b)
990 __a[0] = __b;
991 return __a;
995 _mm_cvtsd_si64(__m128d __a)
997 return __builtin_ia32_cvtsd2si64(__a);
1001 _mm_cvttsd_si64(__m128d __a)
1003 return __a[0];
1008 _mm_cvtepi32_ps(__m128i __a)
1010 return __builtin_ia32_cvtdq2ps((__v4si)__a);
1014 _mm_cvtps_epi32(__m128 __a)
1016 return (__m128i)__builtin_ia32_cvtps2dq(__a);
1020 _mm_cvttps_epi32(__m128 __a)
1022 return (__m128i)__builtin_ia32_cvttps2dq(__a);
1026 _mm_cvtsi32_si128(int __a)
1028 return (__m128i)(__v4si){ __a, 0, 0, 0 };
1033 _mm_cvtsi64_si128(long long __a)
1035 return (__m128i){ __a, 0 };
1040 _mm_cvtsi128_si32(__m128i __a)
1042 __v4si __b = (__v4si)__a;
1048 _mm_cvtsi128_si64(__m128i __a)
1050 return __a[0];
1187 _mm_storel_epi64(__m128i *__p, __m128i __a)
1192 ((struct __mm_storel_epi64_struct*)__p)->__u = __a[0];
1196 _mm_stream_pd(double *__p, __m128d __a)
1198 __builtin_ia32_movntpd(__p, __a);
1202 _mm_stream_si128(__m128i *__p, __m128i __a)
1204 __builtin_ia32_movntdq(__p, __a);
1208 _mm_stream_si32(int *__p, int __a)
1210 __builtin_ia32_movnti(__p, __a);
1232 _mm_packs_epi16(__m128i __a, __m128i __b)
1234 return (__m128i)__builtin_ia32_packsswb128((__v8hi)__a, (__v8hi)__b);
1238 _mm_packs_epi32(__m128i __a, __m128i __b)
1240 return (__m128i)__builtin_ia32_packssdw128((__v4si)__a, (__v4si)__b);
1244 _mm_packus_epi16(__m128i __a, __m128i __b)
1246 return (__m128i)__builtin_ia32_packuswb128((__v8hi)__a, (__v8hi)__b);
1250 _mm_extract_epi16(__m128i __a, int __imm)
1252 __v8hi __b = (__v8hi)__a;
1257 _mm_insert_epi16(__m128i __a, int __b, int __imm)
1259 __v8hi __c = (__v8hi)__a;
1265 _mm_movemask_epi8(__m128i __a)
1267 return __builtin_ia32_pmovmskb128((__v16qi)__a);
1271 __m128i __a = (a); \
1272 (__m128i)__builtin_shufflevector((__v4si)__a, (__v4si) _mm_set1_epi32(0), \
1277 __m128i __a = (a); \
1278 (__m128i)__builtin_shufflevector((__v8hi)__a, (__v8hi) _mm_set1_epi16(0), \
1284 __m128i __a = (a); \
1285 (__m128i)__builtin_shufflevector((__v8hi)__a, (__v8hi) _mm_set1_epi16(0), \
1293 _mm_unpackhi_epi8(__m128i __a, __m128i __b)
1295 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);
1299 _mm_unpackhi_epi16(__m128i __a, __m128i __b)
1301 return (__m128i)__builtin_shufflevector((__v8hi)__a, (__v8hi)__b, 4, 8+4, 5, 8+5, 6, 8+6, 7, 8+7);
1305 _mm_unpackhi_epi32(__m128i __a, __m128i __b)
1307 return (__m128i)__builtin_shufflevector((__v4si)__a, (__v4si)__b, 2, 4+2, 3, 4+3);
1311 _mm_unpackhi_epi64(__m128i __a, __m128i __b)
1313 return (__m128i)__builtin_shufflevector(__a, __b, 1, 2+1);
1317 _mm_unpacklo_epi8(__m128i __a, __m128i __b)
1319 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);
1323 _mm_unpacklo_epi16(__m128i __a, __m128i __b)
1325 return (__m128i)__builtin_shufflevector((__v8hi)__a, (__v8hi)__b, 0, 8+0, 1, 8+1, 2, 8+2, 3, 8+3);
1329 _mm_unpacklo_epi32(__m128i __a, __m128i __b)
1331 return (__m128i)__builtin_shufflevector((__v4si)__a, (__v4si)__b, 0, 4+0, 1, 4+1);
1335 _mm_unpacklo_epi64(__m128i __a, __m128i __b)
1337 return (__m128i)__builtin_shufflevector(__a, __b, 0, 2+0);
1341 _mm_movepi64_pi64(__m128i __a)
1343 return (__m64)__a[0];
1347 _mm_movpi64_pi64(__m64 __a)
1349 return (__m128i){ (long long)__a, 0 };
1353 _mm_move_epi64(__m128i __a)
1355 return __builtin_shufflevector(__a, (__m128i){ 0 }, 0, 2);
1359 _mm_unpackhi_pd(__m128d __a, __m128d __b)
1361 return __builtin_shufflevector(__a, __b, 1, 2+1);
1365 _mm_unpacklo_pd(__m128d __a, __m128d __b)
1367 return __builtin_shufflevector(__a, __b, 0, 2+0);
1371 _mm_movemask_pd(__m128d __a)
1373 return __builtin_ia32_movmskpd(__a);
1377 __m128d __a = (a); \
1379 __builtin_shufflevector(__a, __b, (i) & 1, (((i) & 2) >> 1) + 2); })
1382 _mm_castpd_ps(__m128d __a)
1384 return (__m128)__a;
1388 _mm_castpd_si128(__m128d __a)
1390 return (__m128i)__a;
1394 _mm_castps_pd(__m128 __a)
1396 return (__m128d)__a;
1400 _mm_castps_si128(__m128 __a)
1402 return (__m128i)__a;
1406 _mm_castsi128_ps(__m128i __a)
1408 return (__m128)__a;
1412 _mm_castsi128_pd(__m128i __a)
1414 return (__m128d)__a;
