43 _mm_add_sd(__m128d a, __m128d b)  in _mm_add_sd()
50 _mm_add_pd(__m128d a, __m128d b)  in _mm_add_pd()
56 _mm_sub_sd(__m128d a, __m128d b)  in _mm_sub_sd()
63 _mm_sub_pd(__m128d a, __m128d b)  in _mm_sub_pd()
69 _mm_mul_sd(__m128d a, __m128d b)  in _mm_mul_sd()
76 _mm_mul_pd(__m128d a, __m128d b)  in _mm_mul_pd()
82 _mm_div_sd(__m128d a, __m128d b)  in _mm_div_sd()
89 _mm_div_pd(__m128d a, __m128d b)  in _mm_div_pd()
95 _mm_sqrt_sd(__m128d a, __m128d b)  in _mm_sqrt_sd()
102 _mm_sqrt_pd(__m128d a)  in _mm_sqrt_pd()
108 _mm_min_sd(__m128d a, __m128d b)  in _mm_min_sd()
114 _mm_min_pd(__m128d a, __m128d b)  in _mm_min_pd()
120 _mm_max_sd(__m128d a, __m128d b)  in _mm_max_sd()
126 _mm_max_pd(__m128d a, __m128d b)  in _mm_max_pd()
132 _mm_and_pd(__m128d a, __m128d b)  in _mm_and_pd()
138 _mm_andnot_pd(__m128d a, __m128d b)  in _mm_andnot_pd()
144 _mm_or_pd(__m128d a, __m128d b)  in _mm_or_pd()
150 _mm_xor_pd(__m128d a, __m128d b)  in _mm_xor_pd()
156 _mm_cmpeq_pd(__m128d a, __m128d b)  in _mm_cmpeq_pd()
162 _mm_cmplt_pd(__m128d a, __m128d b)  in _mm_cmplt_pd()
168 _mm_cmple_pd(__m128d a, __m128d b)  in _mm_cmple_pd()
174 _mm_cmpgt_pd(__m128d a, __m128d b)  in _mm_cmpgt_pd()
180 _mm_cmpge_pd(__m128d a, __m128d b)  in _mm_cmpge_pd()
186 _mm_cmpord_pd(__m128d a, __m128d b)  in _mm_cmpord_pd()
192 _mm_cmpunord_pd(__m128d a, __m128d b)  in _mm_cmpunord_pd()
198 _mm_cmpneq_pd(__m128d a, __m128d b)  in _mm_cmpneq_pd()
204 _mm_cmpnlt_pd(__m128d a, __m128d b)  in _mm_cmpnlt_pd()
210 _mm_cmpnle_pd(__m128d a, __m128d b)  in _mm_cmpnle_pd()
216 _mm_cmpngt_pd(__m128d a, __m128d b)  in _mm_cmpngt_pd()
222 _mm_cmpnge_pd(__m128d a, __m128d b)  in _mm_cmpnge_pd()
228 _mm_cmpeq_sd(__m128d a, __m128d b)  in _mm_cmpeq_sd()
234 _mm_cmplt_sd(__m128d a, __m128d b)  in _mm_cmplt_sd()
240 _mm_cmple_sd(__m128d a, __m128d b)  in _mm_cmple_sd()
246 _mm_cmpgt_sd(__m128d a, __m128d b)  in _mm_cmpgt_sd()
252 _mm_cmpge_sd(__m128d a, __m128d b)  in _mm_cmpge_sd()
258 _mm_cmpord_sd(__m128d a, __m128d b)  in _mm_cmpord_sd()
264 _mm_cmpunord_sd(__m128d a, __m128d b)  in _mm_cmpunord_sd()
270 _mm_cmpneq_sd(__m128d a, __m128d b)  in _mm_cmpneq_sd()
276 _mm_cmpnlt_sd(__m128d a, __m128d b)  in _mm_cmpnlt_sd()
282 _mm_cmpnle_sd(__m128d a, __m128d b)  in _mm_cmpnle_sd()
288 _mm_cmpngt_sd(__m128d a, __m128d b)  in _mm_cmpngt_sd()
294 _mm_cmpnge_sd(__m128d a, __m128d b)  in _mm_cmpnge_sd()
300 _mm_comieq_sd(__m128d a, __m128d b)  in _mm_comieq_sd()
306 _mm_comilt_sd(__m128d a, __m128d b)  in _mm_comilt_sd()
312 _mm_comile_sd(__m128d a, __m128d b)  in _mm_comile_sd()
318 _mm_comigt_sd(__m128d a, __m128d b)  in _mm_comigt_sd()
324 _mm_comineq_sd(__m128d a, __m128d b)  in _mm_comineq_sd()
330 _mm_ucomieq_sd(__m128d a, __m128d b)  in _mm_ucomieq_sd()
336 _mm_ucomilt_sd(__m128d a, __m128d b)  in _mm_ucomilt_sd()
342 _mm_ucomile_sd(__m128d a, __m128d b)  in _mm_ucomile_sd()
348 _mm_ucomigt_sd(__m128d a, __m128d b)  in _mm_ucomigt_sd()
354 _mm_ucomineq_sd(__m128d a, __m128d b)  in _mm_ucomineq_sd()
360 _mm_cvtpd_ps(__m128d a)  in _mm_cvtpd_ps()
366 _mm_cvtps_pd(__m128 a)  in _mm_cvtps_pd()
372 _mm_cvtepi32_pd(__m128i a)  in _mm_cvtepi32_pd()
378 _mm_cvtpd_epi32(__m128d a)  in _mm_cvtpd_epi32()
384 _mm_cvtsd_si32(__m128d a)  in _mm_cvtsd_si32()
390 _mm_cvtsd_ss(__m128 a, __m128d b)  in _mm_cvtsd_ss()
397 _mm_cvtsi32_sd(__m128d a, int b)  in _mm_cvtsi32_sd()
404 _mm_cvtss_sd(__m128d a, __m128 b)  in _mm_cvtss_sd()
411 _mm_cvttpd_epi32(__m128d a)  in _mm_cvttpd_epi32()
417 _mm_cvttsd_si32(__m128d a)  in _mm_cvttsd_si32()
423 _mm_cvtpd_pi32(__m128d a)  in _mm_cvtpd_pi32()
429 _mm_cvttpd_pi32(__m128d a)  in _mm_cvttpd_pi32()
435 _mm_cvtpi32_pd(__m64 a)  in _mm_cvtpi32_pd()
441 _mm_cvtsd_f64(__m128d a)  in _mm_cvtsd_f64()
482 _mm_loadh_pd(__m128d a, double const *dp)  in _mm_loadh_pd()
488 _mm_loadl_pd(__m128d a, double const *dp)  in _mm_loadl_pd()
524 _mm_move_sd(__m128d a, __m128d b)  in _mm_move_sd()
530 _mm_store_sd(double *dp, __m128d a)  in _mm_store_sd()
536 _mm_store1_pd(double *dp, __m128d a)  in _mm_store1_pd()
543 _mm_store_pd(double *dp, __m128d a)  in _mm_store_pd()
549 _mm_storeu_pd(double *dp, __m128d a)  in _mm_storeu_pd()
555 _mm_storer_pd(double *dp, __m128d a)  in _mm_storer_pd()
562 _mm_storeh_pd(double *dp, __m128d a)  in _mm_storeh_pd()
568 _mm_storel_pd(double *dp, __m128d a)  in _mm_storel_pd()
574 _mm_add_epi8(__m128i a, __m128i b)  in _mm_add_epi8()
580 _mm_add_epi16(__m128i a, __m128i b)  in _mm_add_epi16()
586 _mm_add_epi32(__m128i a, __m128i b)  in _mm_add_epi32()
592 _mm_add_si64(__m64 a, __m64 b)  in _mm_add_si64()
598 _mm_add_epi64(__m128i a, __m128i b)  in _mm_add_epi64()
604 _mm_adds_epi8(__m128i a, __m128i b)  in _mm_adds_epi8()
610 _mm_adds_epi16(__m128i a, __m128i b)  in _mm_adds_epi16()
616 _mm_adds_epu8(__m128i a, __m128i b)  in _mm_adds_epu8()
622 _mm_adds_epu16(__m128i a, __m128i b)  in _mm_adds_epu16()
628 _mm_avg_epu8(__m128i a, __m128i b)  in _mm_avg_epu8()
634 _mm_avg_epu16(__m128i a, __m128i b)  in _mm_avg_epu16()
640 _mm_madd_epi16(__m128i a, __m128i b)  in _mm_madd_epi16()
646 _mm_max_epi16(__m128i a, __m128i b)  in _mm_max_epi16()
652 _mm_max_epu8(__m128i a, __m128i b)  in _mm_max_epu8()
658 _mm_min_epi16(__m128i a, __m128i b)  in _mm_min_epi16()
664 _mm_min_epu8(__m128i a, __m128i b)  in _mm_min_epu8()
670 _mm_mulhi_epi16(__m128i a, __m128i b)  in _mm_mulhi_epi16()
676 _mm_mulhi_epu16(__m128i a, __m128i b)  in _mm_mulhi_epu16()
682 _mm_mullo_epi16(__m128i a, __m128i b)  in _mm_mullo_epi16()
688 _mm_mul_su32(__m64 a, __m64 b)  in _mm_mul_su32()
694 _mm_mul_epu32(__m128i a, __m128i b)  in _mm_mul_epu32()
700 _mm_sad_epu8(__m128i a, __m128i b)  in _mm_sad_epu8()
706 _mm_sub_epi8(__m128i a, __m128i b)  in _mm_sub_epi8()
712 _mm_sub_epi16(__m128i a, __m128i b)  in _mm_sub_epi16()
718 _mm_sub_epi32(__m128i a, __m128i b)  in _mm_sub_epi32()
724 _mm_sub_si64(__m64 a, __m64 b)  in _mm_sub_si64()
730 _mm_sub_epi64(__m128i a, __m128i b)  in _mm_sub_epi64()
736 _mm_subs_epi8(__m128i a, __m128i b)  in _mm_subs_epi8()
742 _mm_subs_epi16(__m128i a, __m128i b)  in _mm_subs_epi16()
748 _mm_subs_epu8(__m128i a, __m128i b)  in _mm_subs_epu8()
754 _mm_subs_epu16(__m128i a, __m128i b)  in _mm_subs_epu16()
760 _mm_and_si128(__m128i a, __m128i b)  in _mm_and_si128()
766 _mm_andnot_si128(__m128i a, __m128i b)  in _mm_andnot_si128()
772 _mm_or_si128(__m128i a, __m128i b)  in _mm_or_si128()
778 _mm_xor_si128(__m128i a, __m128i b)  in _mm_xor_si128()
787 _mm_slli_epi16(__m128i a, int count)  in _mm_slli_epi16()
793 _mm_sll_epi16(__m128i a, __m128i count)  in _mm_sll_epi16()
799 _mm_slli_epi32(__m128i a, int count)  in _mm_slli_epi32()
805 _mm_sll_epi32(__m128i a, __m128i count)  in _mm_sll_epi32()
811 _mm_slli_epi64(__m128i a, int count)  in _mm_slli_epi64()
817 _mm_sll_epi64(__m128i a, __m128i count)  in _mm_sll_epi64()
823 _mm_srai_epi16(__m128i a, int count)  in _mm_srai_epi16()
829 _mm_sra_epi16(__m128i a, __m128i count)  in _mm_sra_epi16()
835 _mm_srai_epi32(__m128i a, int count)  in _mm_srai_epi32()
841 _mm_sra_epi32(__m128i a, __m128i count)  in _mm_sra_epi32()
851 _mm_srli_epi16(__m128i a, int count)  in _mm_srli_epi16()
857 _mm_srl_epi16(__m128i a, __m128i count)  in _mm_srl_epi16()
863 _mm_srli_epi32(__m128i a, int count)  in _mm_srli_epi32()
869 _mm_srl_epi32(__m128i a, __m128i count)  in _mm_srl_epi32()
875 _mm_srli_epi64(__m128i a, int count)  in _mm_srli_epi64()
881 _mm_srl_epi64(__m128i a, __m128i count)  in _mm_srl_epi64()
887 _mm_cmpeq_epi8(__m128i a, __m128i b)  in _mm_cmpeq_epi8()
893 _mm_cmpeq_epi16(__m128i a, __m128i b)  in _mm_cmpeq_epi16()
899 _mm_cmpeq_epi32(__m128i a, __m128i b)  in _mm_cmpeq_epi32()
905 _mm_cmpgt_epi8(__m128i a, __m128i b)  in _mm_cmpgt_epi8()
911 _mm_cmpgt_epi16(__m128i a, __m128i b)  in _mm_cmpgt_epi16()
917 _mm_cmpgt_epi32(__m128i a, __m128i b)  in _mm_cmpgt_epi32()
923 _mm_cmplt_epi8(__m128i a, __m128i b)  in _mm_cmplt_epi8()
929 _mm_cmplt_epi16(__m128i a, __m128i b)  in _mm_cmplt_epi16()
935 _mm_cmplt_epi32(__m128i a, __m128i b)  in _mm_cmplt_epi32()
942 _mm_cvtsi64_sd(__m128d a, long long b)  in _mm_cvtsi64_sd()
949 _mm_cvtsd_si64(__m128d a)  in _mm_cvtsd_si64()
955 _mm_cvttsd_si64(__m128d a)  in _mm_cvttsd_si64()
962 _mm_cvtepi32_ps(__m128i a)  in _mm_cvtepi32_ps()
968 _mm_cvtps_epi32(__m128 a)  in _mm_cvtps_epi32()
974 _mm_cvttps_epi32(__m128 a)  in _mm_cvttps_epi32()
980 _mm_cvtsi32_si128(int a)  in _mm_cvtsi32_si128()
987 _mm_cvtsi64_si128(long long a)  in _mm_cvtsi64_si128()
994 _mm_cvtsi128_si32(__m128i a)  in _mm_cvtsi128_si32()
996   __v4si b = (__v4si)a;  in _mm_cvtsi128_si32()  local
1002 _mm_cvtsi128_si64(__m128i a)  in _mm_cvtsi128_si64()
1084 _mm_set1_epi8(char b)  in _mm_set1_epi8()
1120 _mm_store_si128(__m128i *p, __m128i b)  in _mm_store_si128()
1126 _mm_storeu_si128(__m128i *p, __m128i b)  in _mm_storeu_si128()
1138 _mm_storel_epi64(__m128i *p, __m128i a)  in _mm_storel_epi64()
1144 _mm_stream_pd(double *p, __m128d a)  in _mm_stream_pd()
1150 _mm_stream_si128(__m128i *p, __m128i a)  in _mm_stream_si128()
1156 _mm_stream_si32(int *p, int a)  in _mm_stream_si32()
1180 _mm_packs_epi16(__m128i a, __m128i b)  in _mm_packs_epi16()
1186 _mm_packs_epi32(__m128i a, __m128i b)  in _mm_packs_epi32()
1192 _mm_packus_epi16(__m128i a, __m128i b)  in _mm_packus_epi16()
1198 _mm_extract_epi16(__m128i a, int imm)  in _mm_extract_epi16()
1200   __v8hi b = (__v8hi)a;  in _mm_extract_epi16()  local
1205 _mm_insert_epi16(__m128i a, int b, int imm)  in _mm_insert_epi16()
1213 _mm_movemask_epi8(__m128i a)  in _mm_movemask_epi8()
1218 #define _mm_shuffle_epi32(a, imm) \  argument
1224 #define _mm_shufflelo_epi16(a, imm) \  argument
1229 #define _mm_shufflehi_epi16(a, imm) \  argument
1237 _mm_unpackhi_epi8(__m128i a, __m128i b)  in _mm_unpackhi_epi8()
1243 _mm_unpackhi_epi16(__m128i a, __m128i b)  in _mm_unpackhi_epi16()
1249 _mm_unpackhi_epi32(__m128i a, __m128i b)  in _mm_unpackhi_epi32()
1255 _mm_unpackhi_epi64(__m128i a, __m128i b)  in _mm_unpackhi_epi64()
1261 _mm_unpacklo_epi8(__m128i a, __m128i b)  in _mm_unpacklo_epi8()
1267 _mm_unpacklo_epi16(__m128i a, __m128i b)  in _mm_unpacklo_epi16()
1273 _mm_unpacklo_epi32(__m128i a, __m128i b)  in _mm_unpacklo_epi32()
1279 _mm_unpacklo_epi64(__m128i a, __m128i b)  in _mm_unpacklo_epi64()
1285 _mm_movepi64_pi64(__m128i a)  in _mm_movepi64_pi64()
1291 _mm_movpi64_pi64(__m64 a)  in _mm_movpi64_pi64()
1297 _mm_move_epi64(__m128i a)  in _mm_move_epi64()
1303 _mm_unpackhi_pd(__m128d a, __m128d b)  in _mm_unpackhi_pd()
1309 _mm_unpacklo_pd(__m128d a, __m128d b)  in _mm_unpacklo_pd()
1315 _mm_movemask_pd(__m128d a)  in _mm_movemask_pd()
1320 #define _mm_shuffle_pd(a, b, i) \  argument