34 typedef long long __m128i __attribute__((__vector_size__(16)));  typedef
372 _mm_cvtepi32_pd(__m128i a)  in _mm_cvtepi32_pd()
377 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
410 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
413   return (__m128i)__builtin_ia32_cvttpd2dq(a);  in _mm_cvttpd_epi32()
573 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
574 _mm_add_epi8(__m128i a, __m128i b)  in _mm_add_epi8()
576   return (__m128i)((__v16qi)a + (__v16qi)b);  in _mm_add_epi8()
579 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
580 _mm_add_epi16(__m128i a, __m128i b)  in _mm_add_epi16()
582   return (__m128i)((__v8hi)a + (__v8hi)b);  in _mm_add_epi16()
585 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
586 _mm_add_epi32(__m128i a, __m128i b)  in _mm_add_epi32()
588   return (__m128i)((__v4si)a + (__v4si)b);  in _mm_add_epi32()
597 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
598 _mm_add_epi64(__m128i a, __m128i b)  in _mm_add_epi64()
603 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
604 _mm_adds_epi8(__m128i a, __m128i b)  in _mm_adds_epi8()
606   return (__m128i)__builtin_ia32_paddsb128((__v16qi)a, (__v16qi)b);  in _mm_adds_epi8()
609 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
610 _mm_adds_epi16(__m128i a, __m128i b)  in _mm_adds_epi16()
612   return (__m128i)__builtin_ia32_paddsw128((__v8hi)a, (__v8hi)b);  in _mm_adds_epi16()
615 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
616 _mm_adds_epu8(__m128i a, __m128i b)  in _mm_adds_epu8()
618   return (__m128i)__builtin_ia32_paddusb128((__v16qi)a, (__v16qi)b);  in _mm_adds_epu8()
621 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
622 _mm_adds_epu16(__m128i a, __m128i b)  in _mm_adds_epu16()
624   return (__m128i)__builtin_ia32_paddusw128((__v8hi)a, (__v8hi)b);  in _mm_adds_epu16()
627 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
628 _mm_avg_epu8(__m128i a, __m128i b)  in _mm_avg_epu8()
630   return (__m128i)__builtin_ia32_pavgb128((__v16qi)a, (__v16qi)b);  in _mm_avg_epu8()
633 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
634 _mm_avg_epu16(__m128i a, __m128i b)  in _mm_avg_epu16()
636   return (__m128i)__builtin_ia32_pavgw128((__v8hi)a, (__v8hi)b);  in _mm_avg_epu16()
639 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
640 _mm_madd_epi16(__m128i a, __m128i b)  in _mm_madd_epi16()
642   return (__m128i)__builtin_ia32_pmaddwd128((__v8hi)a, (__v8hi)b);  in _mm_madd_epi16()
645 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
646 _mm_max_epi16(__m128i a, __m128i b)  in _mm_max_epi16()
648   return (__m128i)__builtin_ia32_pmaxsw128((__v8hi)a, (__v8hi)b);  in _mm_max_epi16()
651 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
652 _mm_max_epu8(__m128i a, __m128i b)  in _mm_max_epu8()
654   return (__m128i)__builtin_ia32_pmaxub128((__v16qi)a, (__v16qi)b);  in _mm_max_epu8()
657 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
658 _mm_min_epi16(__m128i a, __m128i b)  in _mm_min_epi16()
660   return (__m128i)__builtin_ia32_pminsw128((__v8hi)a, (__v8hi)b);  in _mm_min_epi16()
663 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
664 _mm_min_epu8(__m128i a, __m128i b)  in _mm_min_epu8()
666   return (__m128i)__builtin_ia32_pminub128((__v16qi)a, (__v16qi)b);  in _mm_min_epu8()
669 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
670 _mm_mulhi_epi16(__m128i a, __m128i b)  in _mm_mulhi_epi16()
672   return (__m128i)__builtin_ia32_pmulhw128((__v8hi)a, (__v8hi)b);  in _mm_mulhi_epi16()
675 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
676 _mm_mulhi_epu16(__m128i a, __m128i b)  in _mm_mulhi_epu16()
678   return (__m128i)__builtin_ia32_pmulhuw128((__v8hi)a, (__v8hi)b);  in _mm_mulhi_epu16()
681 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
682 _mm_mullo_epi16(__m128i a, __m128i b)  in _mm_mullo_epi16()
684   return (__m128i)((__v8hi)a * (__v8hi)b);  in _mm_mullo_epi16()
693 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
694 _mm_mul_epu32(__m128i a, __m128i b)  in _mm_mul_epu32()
699 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
700 _mm_sad_epu8(__m128i a, __m128i b)  in _mm_sad_epu8()
705 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
706 _mm_sub_epi8(__m128i a, __m128i b)  in _mm_sub_epi8()
708   return (__m128i)((__v16qi)a - (__v16qi)b);  in _mm_sub_epi8()
711 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
712 _mm_sub_epi16(__m128i a, __m128i b)  in _mm_sub_epi16()
714   return (__m128i)((__v8hi)a - (__v8hi)b);  in _mm_sub_epi16()
717 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
718 _mm_sub_epi32(__m128i a, __m128i b)  in _mm_sub_epi32()
720   return (__m128i)((__v4si)a - (__v4si)b);  in _mm_sub_epi32()
729 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
730 _mm_sub_epi64(__m128i a, __m128i b)  in _mm_sub_epi64()
735 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
736 _mm_subs_epi8(__m128i a, __m128i b)  in _mm_subs_epi8()
738   return (__m128i)__builtin_ia32_psubsb128((__v16qi)a, (__v16qi)b);  in _mm_subs_epi8()
741 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
742 _mm_subs_epi16(__m128i a, __m128i b)  in _mm_subs_epi16()
744   return (__m128i)__builtin_ia32_psubsw128((__v8hi)a, (__v8hi)b);  in _mm_subs_epi16()
747 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
748 _mm_subs_epu8(__m128i a, __m128i b)  in _mm_subs_epu8()
750   return (__m128i)__builtin_ia32_psubusb128((__v16qi)a, (__v16qi)b);  in _mm_subs_epu8()
753 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
754 _mm_subs_epu16(__m128i a, __m128i b)  in _mm_subs_epu16()
756   return (__m128i)__builtin_ia32_psubusw128((__v8hi)a, (__v8hi)b);  in _mm_subs_epu16()
759 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
760 _mm_and_si128(__m128i a, __m128i b)  in _mm_and_si128()
765 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
766 _mm_andnot_si128(__m128i a, __m128i b)  in _mm_andnot_si128()
771 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
772 _mm_or_si128(__m128i a, __m128i b)  in _mm_or_si128()
777 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
778 _mm_xor_si128(__m128i a, __m128i b)  in _mm_xor_si128()
784   ((__m128i)__builtin_ia32_pslldqi128((__m128i)(VEC), (IMM)*8))
786 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
787 _mm_slli_epi16(__m128i a, int count)  in _mm_slli_epi16()
789   return (__m128i)__builtin_ia32_psllwi128((__v8hi)a, count);  in _mm_slli_epi16()
792 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
793 _mm_sll_epi16(__m128i a, __m128i count)  in _mm_sll_epi16()
795   return (__m128i)__builtin_ia32_psllw128((__v8hi)a, (__v8hi)count);  in _mm_sll_epi16()
798 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
799 _mm_slli_epi32(__m128i a, int count)  in _mm_slli_epi32()
801   return (__m128i)__builtin_ia32_pslldi128((__v4si)a, count);  in _mm_slli_epi32()
804 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
805 _mm_sll_epi32(__m128i a, __m128i count)  in _mm_sll_epi32()
807   return (__m128i)__builtin_ia32_pslld128((__v4si)a, (__v4si)count);  in _mm_sll_epi32()
810 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
811 _mm_slli_epi64(__m128i a, int count)  in _mm_slli_epi64()
816 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
817 _mm_sll_epi64(__m128i a, __m128i count)  in _mm_sll_epi64()
822 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
823 _mm_srai_epi16(__m128i a, int count)  in _mm_srai_epi16()
825   return (__m128i)__builtin_ia32_psrawi128((__v8hi)a, count);  in _mm_srai_epi16()
828 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
829 _mm_sra_epi16(__m128i a, __m128i count)  in _mm_sra_epi16()
831   return (__m128i)__builtin_ia32_psraw128((__v8hi)a, (__v8hi)count);  in _mm_sra_epi16()
834 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
835 _mm_srai_epi32(__m128i a, int count)  in _mm_srai_epi32()
837   return (__m128i)__builtin_ia32_psradi128((__v4si)a, count);  in _mm_srai_epi32()
840 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
841 _mm_sra_epi32(__m128i a, __m128i count)  in _mm_sra_epi32()
843   return (__m128i)__builtin_ia32_psrad128((__v4si)a, (__v4si)count);  in _mm_sra_epi32()
848   ((__m128i)__builtin_ia32_psrldqi128((__m128i)(VEC), (IMM)*8))
850 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
851 _mm_srli_epi16(__m128i a, int count)  in _mm_srli_epi16()
853   return (__m128i)__builtin_ia32_psrlwi128((__v8hi)a, count);  in _mm_srli_epi16()
856 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
857 _mm_srl_epi16(__m128i a, __m128i count)  in _mm_srl_epi16()
859   return (__m128i)__builtin_ia32_psrlw128((__v8hi)a, (__v8hi)count);  in _mm_srl_epi16()
862 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
863 _mm_srli_epi32(__m128i a, int count)  in _mm_srli_epi32()
865   return (__m128i)__builtin_ia32_psrldi128((__v4si)a, count);  in _mm_srli_epi32()
868 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
869 _mm_srl_epi32(__m128i a, __m128i count)  in _mm_srl_epi32()
871   return (__m128i)__builtin_ia32_psrld128((__v4si)a, (__v4si)count);  in _mm_srl_epi32()
874 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
875 _mm_srli_epi64(__m128i a, int count)  in _mm_srli_epi64()
880 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
881 _mm_srl_epi64(__m128i a, __m128i count)  in _mm_srl_epi64()
886 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
887 _mm_cmpeq_epi8(__m128i a, __m128i b)  in _mm_cmpeq_epi8()
889   return (__m128i)((__v16qi)a == (__v16qi)b);  in _mm_cmpeq_epi8()
892 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
893 _mm_cmpeq_epi16(__m128i a, __m128i b)  in _mm_cmpeq_epi16()
895   return (__m128i)((__v8hi)a == (__v8hi)b);  in _mm_cmpeq_epi16()
898 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
899 _mm_cmpeq_epi32(__m128i a, __m128i b)  in _mm_cmpeq_epi32()
901   return (__m128i)((__v4si)a == (__v4si)b);  in _mm_cmpeq_epi32()
904 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
905 _mm_cmpgt_epi8(__m128i a, __m128i b)  in _mm_cmpgt_epi8()
907   return (__m128i)((__v16qi)a > (__v16qi)b);  in _mm_cmpgt_epi8()
910 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
911 _mm_cmpgt_epi16(__m128i a, __m128i b)  in _mm_cmpgt_epi16()
913   return (__m128i)((__v8hi)a > (__v8hi)b);  in _mm_cmpgt_epi16()
916 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
917 _mm_cmpgt_epi32(__m128i a, __m128i b)  in _mm_cmpgt_epi32()
919   return (__m128i)((__v4si)a > (__v4si)b);  in _mm_cmpgt_epi32()
922 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
923 _mm_cmplt_epi8(__m128i a, __m128i b)  in _mm_cmplt_epi8()
928 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
929 _mm_cmplt_epi16(__m128i a, __m128i b)  in _mm_cmplt_epi16()
934 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
935 _mm_cmplt_epi32(__m128i a, __m128i b)  in _mm_cmplt_epi32()
962 _mm_cvtepi32_ps(__m128i a)  in _mm_cvtepi32_ps()
967 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
970   return (__m128i)__builtin_ia32_cvtps2dq(a);  in _mm_cvtps_epi32()
973 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
976   return (__m128i)__builtin_ia32_cvttps2dq(a);  in _mm_cvttps_epi32()
979 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
982   return (__m128i)(__v4si){ a, 0, 0, 0 };  in _mm_cvtsi32_si128()
986 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
989   return (__m128i){ a, 0 };  in _mm_cvtsi64_si128()
994 _mm_cvtsi128_si32(__m128i a)  in _mm_cvtsi128_si32()
1002 _mm_cvtsi128_si64(__m128i a)  in _mm_cvtsi128_si64()
1008 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1009 _mm_load_si128(__m128i const *p)  in _mm_load_si128()
1014 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1015 _mm_loadu_si128(__m128i const *p)  in _mm_loadu_si128()
1018     __m128i v;  in _mm_loadu_si128()
1023 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1024 _mm_loadl_epi64(__m128i const *p)  in _mm_loadl_epi64()
1026   return (__m128i) { *(long long*)p, 0};  in _mm_loadl_epi64()
1029 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1032   return (__m128i){ q0, q1 };  in _mm_set_epi64x()
1035 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1038   return (__m128i){ (long long)q0, (long long)q1 };  in _mm_set_epi64()
1041 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1044   return (__m128i)(__v4si){ i0, i1, i2, i3};  in _mm_set_epi32()
1047 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1050   return (__m128i)(__v8hi){ w0, w1, w2, w3, w4, w5, w6, w7 };  in _mm_set_epi16()
1053 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1056   return (__m128i)(__v16qi){ b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, b10, b11, b12, b13, b14, b15 };  in _mm_set_epi8()
1059 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1062   return (__m128i){ q, q };  in _mm_set1_epi64x()
1065 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1068   return (__m128i){ (long long)q, (long long)q };  in _mm_set1_epi64()
1071 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1074   return (__m128i)(__v4si){ i, i, i, i };  in _mm_set1_epi32()
1077 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1080   return (__m128i)(__v8hi){ w, w, w, w, w, w, w, w };  in _mm_set1_epi16()
1083 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1086   return (__m128i)(__v16qi){ b, b, b, b, b, b, b, b, b, b, b, b, b, b, b, b };  in _mm_set1_epi8()
1089 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1092   return (__m128i){ (long long)q0, (long long)q1 };  in _mm_setr_epi64()
1095 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1098   return (__m128i)(__v4si){ i0, i1, i2, i3};  in _mm_setr_epi32()
1101 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1104   return (__m128i)(__v8hi){ w0, w1, w2, w3, w4, w5, w6, w7 };  in _mm_setr_epi16()
1107 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1110   return (__m128i)(__v16qi){ b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, b10, b11, b12, b13, b14, b15 };  in _mm_setr_epi8()
1113 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1116   return (__m128i){ 0LL, 0LL };  in _mm_setzero_si128()
1120 _mm_store_si128(__m128i *p, __m128i b)  in _mm_store_si128()
1126 _mm_storeu_si128(__m128i *p, __m128i b)  in _mm_storeu_si128()
1132 _mm_maskmoveu_si128(__m128i d, __m128i n, char *p)  in _mm_maskmoveu_si128()
1138 _mm_storel_epi64(__m128i *p, __m128i a)  in _mm_storel_epi64()
1150 _mm_stream_si128(__m128i *p, __m128i a)  in _mm_stream_si128()
1179 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1180 _mm_packs_epi16(__m128i a, __m128i b)  in _mm_packs_epi16()
1182   return (__m128i)__builtin_ia32_packsswb128((__v8hi)a, (__v8hi)b);  in _mm_packs_epi16()
1185 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1186 _mm_packs_epi32(__m128i a, __m128i b)  in _mm_packs_epi32()
1188   return (__m128i)__builtin_ia32_packssdw128((__v4si)a, (__v4si)b);  in _mm_packs_epi32()
1191 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1192 _mm_packus_epi16(__m128i a, __m128i b)  in _mm_packus_epi16()
1194   return (__m128i)__builtin_ia32_packuswb128((__v8hi)a, (__v8hi)b);  in _mm_packus_epi16()
1198 _mm_extract_epi16(__m128i a, int imm)  in _mm_extract_epi16()
1204 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1205 _mm_insert_epi16(__m128i a, int b, int imm)  in _mm_insert_epi16()
1209   return (__m128i)c;  in _mm_insert_epi16()
1213 _mm_movemask_epi8(__m128i a)  in _mm_movemask_epi8()
1219   ((__m128i)__builtin_shufflevector((__v4si)(a), (__v4si) _mm_set1_epi32(0), \
1225   ((__m128i)__builtin_shufflevector((__v8hi)(a), (__v8hi) _mm_set1_epi16(0), \
1230   ((__m128i)__builtin_shufflevector((__v8hi)(a), (__v8hi) _mm_set1_epi16(0), 0, 1, 2, 3, \
1236 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1237 _mm_unpackhi_epi8(__m128i a, __m128i b)  in _mm_unpackhi_epi8()
1239 …return (__m128i)__builtin_shufflevector((__v16qi)a, (__v16qi)b, 8, 16+8, 9, 16+9, 10, 16+10, 11, 1…  in _mm_unpackhi_epi8()
1242 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1243 _mm_unpackhi_epi16(__m128i a, __m128i b)  in _mm_unpackhi_epi16()
1245   return (__m128i)__builtin_shufflevector((__v8hi)a, (__v8hi)b, 4, 8+4, 5, 8+5, 6, 8+6, 7, 8+7);  in _mm_unpackhi_epi16()
1248 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1249 _mm_unpackhi_epi32(__m128i a, __m128i b)  in _mm_unpackhi_epi32()
1251   return (__m128i)__builtin_shufflevector((__v4si)a, (__v4si)b, 2, 4+2, 3, 4+3);  in _mm_unpackhi_epi32()
1254 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1255 _mm_unpackhi_epi64(__m128i a, __m128i b)  in _mm_unpackhi_epi64()
1257   return (__m128i)__builtin_shufflevector(a, b, 1, 2+1);  in _mm_unpackhi_epi64()
1260 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1261 _mm_unpacklo_epi8(__m128i a, __m128i b)  in _mm_unpacklo_epi8()
1263 …return (__m128i)__builtin_shufflevector((__v16qi)a, (__v16qi)b, 0, 16+0, 1, 16+1, 2, 16+2, 3, 16+3…  in _mm_unpacklo_epi8()
1266 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1267 _mm_unpacklo_epi16(__m128i a, __m128i b)  in _mm_unpacklo_epi16()
1269   return (__m128i)__builtin_shufflevector((__v8hi)a, (__v8hi)b, 0, 8+0, 1, 8+1, 2, 8+2, 3, 8+3);  in _mm_unpacklo_epi16()
1272 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1273 _mm_unpacklo_epi32(__m128i a, __m128i b)  in _mm_unpacklo_epi32()
1275   return (__m128i)__builtin_shufflevector((__v4si)a, (__v4si)b, 0, 4+0, 1, 4+1);  in _mm_unpacklo_epi32()
1278 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1279 _mm_unpacklo_epi64(__m128i a, __m128i b)  in _mm_unpacklo_epi64()
1281   return (__m128i)__builtin_shufflevector(a, b, 0, 2+0);  in _mm_unpacklo_epi64()
1285 _mm_movepi64_pi64(__m128i a)  in _mm_movepi64_pi64()
1290 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1293   return (__m128i){ (long long)a, 0 };  in _mm_movpi64_pi64()
1296 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1297 _mm_move_epi64(__m128i a)  in _mm_move_epi64()
1299   return __builtin_shufflevector(a, (__m128i){ 0 }, 0, 2);  in _mm_move_epi64()
1330 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1333   return (__m128i)in;  in _mm_castpd_si128()
1342 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1345   return (__m128i)in;  in _mm_castps_si128()
1349 _mm_castsi128_ps(__m128i in)  in _mm_castsi128_ps()
1355 _mm_castsi128_pd(__m128i in)  in _mm_castsi128_pd()