Lines Matching refs:__a
53 _mm_add_sd(__m128d __a, __m128d __b) in _mm_add_sd() argument
55 __a[0] += __b[0]; in _mm_add_sd()
56 return __a; in _mm_add_sd()
60 _mm_add_pd(__m128d __a, __m128d __b) in _mm_add_pd() argument
62 return (__m128d)((__v2df)__a + (__v2df)__b); in _mm_add_pd()
66 _mm_sub_sd(__m128d __a, __m128d __b) in _mm_sub_sd() argument
68 __a[0] -= __b[0]; in _mm_sub_sd()
69 return __a; in _mm_sub_sd()
73 _mm_sub_pd(__m128d __a, __m128d __b) in _mm_sub_pd() argument
75 return (__m128d)((__v2df)__a - (__v2df)__b); in _mm_sub_pd()
79 _mm_mul_sd(__m128d __a, __m128d __b) in _mm_mul_sd() argument
81 __a[0] *= __b[0]; in _mm_mul_sd()
82 return __a; in _mm_mul_sd()
86 _mm_mul_pd(__m128d __a, __m128d __b) in _mm_mul_pd() argument
88 return (__m128d)((__v2df)__a * (__v2df)__b); in _mm_mul_pd()
92 _mm_div_sd(__m128d __a, __m128d __b) in _mm_div_sd() argument
94 __a[0] /= __b[0]; in _mm_div_sd()
95 return __a; in _mm_div_sd()
99 _mm_div_pd(__m128d __a, __m128d __b) in _mm_div_pd() argument
101 return (__m128d)((__v2df)__a / (__v2df)__b); in _mm_div_pd()
105 _mm_sqrt_sd(__m128d __a, __m128d __b) in _mm_sqrt_sd() argument
108 return (__m128d) { __c[0], __a[1] }; in _mm_sqrt_sd()
112 _mm_sqrt_pd(__m128d __a) in _mm_sqrt_pd() argument
114 return __builtin_ia32_sqrtpd((__v2df)__a); in _mm_sqrt_pd()
118 _mm_min_sd(__m128d __a, __m128d __b) in _mm_min_sd() argument
120 return __builtin_ia32_minsd((__v2df)__a, (__v2df)__b); in _mm_min_sd()
124 _mm_min_pd(__m128d __a, __m128d __b) in _mm_min_pd() argument
126 return __builtin_ia32_minpd((__v2df)__a, (__v2df)__b); in _mm_min_pd()
130 _mm_max_sd(__m128d __a, __m128d __b) in _mm_max_sd() argument
132 return __builtin_ia32_maxsd((__v2df)__a, (__v2df)__b); in _mm_max_sd()
136 _mm_max_pd(__m128d __a, __m128d __b) in _mm_max_pd() argument
138 return __builtin_ia32_maxpd((__v2df)__a, (__v2df)__b); in _mm_max_pd()
142 _mm_and_pd(__m128d __a, __m128d __b) in _mm_and_pd() argument
144 return (__m128d)((__v4su)__a & (__v4su)__b); in _mm_and_pd()
148 _mm_andnot_pd(__m128d __a, __m128d __b) in _mm_andnot_pd() argument
150 return (__m128d)(~(__v4su)__a & (__v4su)__b); in _mm_andnot_pd()
154 _mm_or_pd(__m128d __a, __m128d __b) in _mm_or_pd() argument
156 return (__m128d)((__v4su)__a | (__v4su)__b); in _mm_or_pd()
160 _mm_xor_pd(__m128d __a, __m128d __b) in _mm_xor_pd() argument
162 return (__m128d)((__v4su)__a ^ (__v4su)__b); in _mm_xor_pd()
166 _mm_cmpeq_pd(__m128d __a, __m128d __b) in _mm_cmpeq_pd() argument
168 return (__m128d)__builtin_ia32_cmpeqpd((__v2df)__a, (__v2df)__b); in _mm_cmpeq_pd()
172 _mm_cmplt_pd(__m128d __a, __m128d __b) in _mm_cmplt_pd() argument
174 return (__m128d)__builtin_ia32_cmpltpd((__v2df)__a, (__v2df)__b); in _mm_cmplt_pd()
178 _mm_cmple_pd(__m128d __a, __m128d __b) in _mm_cmple_pd() argument
180 return (__m128d)__builtin_ia32_cmplepd((__v2df)__a, (__v2df)__b); in _mm_cmple_pd()
184 _mm_cmpgt_pd(__m128d __a, __m128d __b) in _mm_cmpgt_pd() argument
186 return (__m128d)__builtin_ia32_cmpltpd((__v2df)__b, (__v2df)__a); in _mm_cmpgt_pd()
190 _mm_cmpge_pd(__m128d __a, __m128d __b) in _mm_cmpge_pd() argument
192 return (__m128d)__builtin_ia32_cmplepd((__v2df)__b, (__v2df)__a); in _mm_cmpge_pd()
196 _mm_cmpord_pd(__m128d __a, __m128d __b) in _mm_cmpord_pd() argument
198 return (__m128d)__builtin_ia32_cmpordpd((__v2df)__a, (__v2df)__b); in _mm_cmpord_pd()
202 _mm_cmpunord_pd(__m128d __a, __m128d __b) in _mm_cmpunord_pd() argument
204 return (__m128d)__builtin_ia32_cmpunordpd((__v2df)__a, (__v2df)__b); in _mm_cmpunord_pd()
208 _mm_cmpneq_pd(__m128d __a, __m128d __b) in _mm_cmpneq_pd() argument
210 return (__m128d)__builtin_ia32_cmpneqpd((__v2df)__a, (__v2df)__b); in _mm_cmpneq_pd()
214 _mm_cmpnlt_pd(__m128d __a, __m128d __b) in _mm_cmpnlt_pd() argument
216 return (__m128d)__builtin_ia32_cmpnltpd((__v2df)__a, (__v2df)__b); in _mm_cmpnlt_pd()
220 _mm_cmpnle_pd(__m128d __a, __m128d __b) in _mm_cmpnle_pd() argument
222 return (__m128d)__builtin_ia32_cmpnlepd((__v2df)__a, (__v2df)__b); in _mm_cmpnle_pd()
226 _mm_cmpngt_pd(__m128d __a, __m128d __b) in _mm_cmpngt_pd() argument
228 return (__m128d)__builtin_ia32_cmpnltpd((__v2df)__b, (__v2df)__a); in _mm_cmpngt_pd()
232 _mm_cmpnge_pd(__m128d __a, __m128d __b) in _mm_cmpnge_pd() argument
234 return (__m128d)__builtin_ia32_cmpnlepd((__v2df)__b, (__v2df)__a); in _mm_cmpnge_pd()
238 _mm_cmpeq_sd(__m128d __a, __m128d __b) in _mm_cmpeq_sd() argument
240 return (__m128d)__builtin_ia32_cmpeqsd((__v2df)__a, (__v2df)__b); in _mm_cmpeq_sd()
244 _mm_cmplt_sd(__m128d __a, __m128d __b) in _mm_cmplt_sd() argument
246 return (__m128d)__builtin_ia32_cmpltsd((__v2df)__a, (__v2df)__b); in _mm_cmplt_sd()
250 _mm_cmple_sd(__m128d __a, __m128d __b) in _mm_cmple_sd() argument
252 return (__m128d)__builtin_ia32_cmplesd((__v2df)__a, (__v2df)__b); in _mm_cmple_sd()
256 _mm_cmpgt_sd(__m128d __a, __m128d __b) in _mm_cmpgt_sd() argument
258 __m128d __c = __builtin_ia32_cmpltsd((__v2df)__b, (__v2df)__a); in _mm_cmpgt_sd()
259 return (__m128d) { __c[0], __a[1] }; in _mm_cmpgt_sd()
263 _mm_cmpge_sd(__m128d __a, __m128d __b) in _mm_cmpge_sd() argument
265 __m128d __c = __builtin_ia32_cmplesd((__v2df)__b, (__v2df)__a); in _mm_cmpge_sd()
266 return (__m128d) { __c[0], __a[1] }; in _mm_cmpge_sd()
270 _mm_cmpord_sd(__m128d __a, __m128d __b) in _mm_cmpord_sd() argument
272 return (__m128d)__builtin_ia32_cmpordsd((__v2df)__a, (__v2df)__b); in _mm_cmpord_sd()
276 _mm_cmpunord_sd(__m128d __a, __m128d __b) in _mm_cmpunord_sd() argument
278 return (__m128d)__builtin_ia32_cmpunordsd((__v2df)__a, (__v2df)__b); in _mm_cmpunord_sd()
282 _mm_cmpneq_sd(__m128d __a, __m128d __b) in _mm_cmpneq_sd() argument
284 return (__m128d)__builtin_ia32_cmpneqsd((__v2df)__a, (__v2df)__b); in _mm_cmpneq_sd()
288 _mm_cmpnlt_sd(__m128d __a, __m128d __b) in _mm_cmpnlt_sd() argument
290 return (__m128d)__builtin_ia32_cmpnltsd((__v2df)__a, (__v2df)__b); in _mm_cmpnlt_sd()
294 _mm_cmpnle_sd(__m128d __a, __m128d __b) in _mm_cmpnle_sd() argument
296 return (__m128d)__builtin_ia32_cmpnlesd((__v2df)__a, (__v2df)__b); in _mm_cmpnle_sd()
300 _mm_cmpngt_sd(__m128d __a, __m128d __b) in _mm_cmpngt_sd() argument
302 __m128d __c = __builtin_ia32_cmpnltsd((__v2df)__b, (__v2df)__a); in _mm_cmpngt_sd()
303 return (__m128d) { __c[0], __a[1] }; in _mm_cmpngt_sd()
307 _mm_cmpnge_sd(__m128d __a, __m128d __b) in _mm_cmpnge_sd() argument
309 __m128d __c = __builtin_ia32_cmpnlesd((__v2df)__b, (__v2df)__a); in _mm_cmpnge_sd()
310 return (__m128d) { __c[0], __a[1] }; in _mm_cmpnge_sd()
314 _mm_comieq_sd(__m128d __a, __m128d __b) in _mm_comieq_sd() argument
316 return __builtin_ia32_comisdeq((__v2df)__a, (__v2df)__b); in _mm_comieq_sd()
320 _mm_comilt_sd(__m128d __a, __m128d __b) in _mm_comilt_sd() argument
322 return __builtin_ia32_comisdlt((__v2df)__a, (__v2df)__b); in _mm_comilt_sd()
326 _mm_comile_sd(__m128d __a, __m128d __b) in _mm_comile_sd() argument
328 return __builtin_ia32_comisdle((__v2df)__a, (__v2df)__b); in _mm_comile_sd()
332 _mm_comigt_sd(__m128d __a, __m128d __b) in _mm_comigt_sd() argument
334 return __builtin_ia32_comisdgt((__v2df)__a, (__v2df)__b); in _mm_comigt_sd()
338 _mm_comige_sd(__m128d __a, __m128d __b) in _mm_comige_sd() argument
340 return __builtin_ia32_comisdge((__v2df)__a, (__v2df)__b); in _mm_comige_sd()
344 _mm_comineq_sd(__m128d __a, __m128d __b) in _mm_comineq_sd() argument
346 return __builtin_ia32_comisdneq((__v2df)__a, (__v2df)__b); in _mm_comineq_sd()
350 _mm_ucomieq_sd(__m128d __a, __m128d __b) in _mm_ucomieq_sd() argument
352 return __builtin_ia32_ucomisdeq((__v2df)__a, (__v2df)__b); in _mm_ucomieq_sd()
356 _mm_ucomilt_sd(__m128d __a, __m128d __b) in _mm_ucomilt_sd() argument
358 return __builtin_ia32_ucomisdlt((__v2df)__a, (__v2df)__b); in _mm_ucomilt_sd()
362 _mm_ucomile_sd(__m128d __a, __m128d __b) in _mm_ucomile_sd() argument
364 return __builtin_ia32_ucomisdle((__v2df)__a, (__v2df)__b); in _mm_ucomile_sd()
368 _mm_ucomigt_sd(__m128d __a, __m128d __b) in _mm_ucomigt_sd() argument
370 return __builtin_ia32_ucomisdgt((__v2df)__a, (__v2df)__b); in _mm_ucomigt_sd()
374 _mm_ucomige_sd(__m128d __a, __m128d __b) in _mm_ucomige_sd() argument
376 return __builtin_ia32_ucomisdge((__v2df)__a, (__v2df)__b); in _mm_ucomige_sd()
380 _mm_ucomineq_sd(__m128d __a, __m128d __b) in _mm_ucomineq_sd() argument
382 return __builtin_ia32_ucomisdneq((__v2df)__a, (__v2df)__b); in _mm_ucomineq_sd()
386 _mm_cvtpd_ps(__m128d __a) in _mm_cvtpd_ps() argument
388 return __builtin_ia32_cvtpd2ps((__v2df)__a); in _mm_cvtpd_ps()
392 _mm_cvtps_pd(__m128 __a) in _mm_cvtps_pd() argument
395 __builtin_shufflevector((__v4sf)__a, (__v4sf)__a, 0, 1), __v2df); in _mm_cvtps_pd()
399 _mm_cvtepi32_pd(__m128i __a) in _mm_cvtepi32_pd() argument
402 __builtin_shufflevector((__v4si)__a, (__v4si)__a, 0, 1), __v2df); in _mm_cvtepi32_pd()
406 _mm_cvtpd_epi32(__m128d __a) in _mm_cvtpd_epi32() argument
408 return __builtin_ia32_cvtpd2dq((__v2df)__a); in _mm_cvtpd_epi32()
412 _mm_cvtsd_si32(__m128d __a) in _mm_cvtsd_si32() argument
414 return __builtin_ia32_cvtsd2si((__v2df)__a); in _mm_cvtsd_si32()
418 _mm_cvtsd_ss(__m128 __a, __m128d __b) in _mm_cvtsd_ss() argument
420 __a[0] = __b[0]; in _mm_cvtsd_ss()
421 return __a; in _mm_cvtsd_ss()
425 _mm_cvtsi32_sd(__m128d __a, int __b) in _mm_cvtsi32_sd() argument
427 __a[0] = __b; in _mm_cvtsi32_sd()
428 return __a; in _mm_cvtsi32_sd()
432 _mm_cvtss_sd(__m128d __a, __m128 __b) in _mm_cvtss_sd() argument
434 __a[0] = __b[0]; in _mm_cvtss_sd()
435 return __a; in _mm_cvtss_sd()
439 _mm_cvttpd_epi32(__m128d __a) in _mm_cvttpd_epi32() argument
441 return (__m128i)__builtin_ia32_cvttpd2dq((__v2df)__a); in _mm_cvttpd_epi32()
445 _mm_cvttsd_si32(__m128d __a) in _mm_cvttsd_si32() argument
447 return __a[0]; in _mm_cvttsd_si32()
451 _mm_cvtpd_pi32(__m128d __a) in _mm_cvtpd_pi32() argument
453 return (__m64)__builtin_ia32_cvtpd2pi((__v2df)__a); in _mm_cvtpd_pi32()
457 _mm_cvttpd_pi32(__m128d __a) in _mm_cvttpd_pi32() argument
459 return (__m64)__builtin_ia32_cvttpd2pi((__v2df)__a); in _mm_cvttpd_pi32()
463 _mm_cvtpi32_pd(__m64 __a) in _mm_cvtpi32_pd() argument
465 return __builtin_ia32_cvtpi2pd((__v2si)__a); in _mm_cvtpi32_pd()
469 _mm_cvtsd_f64(__m128d __a) in _mm_cvtsd_f64() argument
471 return __a[0]; in _mm_cvtsd_f64()
509 _mm_loadu_si64(void const *__a) in _mm_loadu_si64() argument
514 long long __u = ((struct __loadu_si64*)__a)->__v; in _mm_loadu_si64()
529 _mm_loadh_pd(__m128d __a, double const *__dp) in _mm_loadh_pd() argument
535 return (__m128d){ __a[0], __u }; in _mm_loadh_pd()
539 _mm_loadl_pd(__m128d __a, double const *__dp) in _mm_loadl_pd() argument
545 return (__m128d){ __u, __a[1] }; in _mm_loadl_pd()
585 _mm_move_sd(__m128d __a, __m128d __b) in _mm_move_sd() argument
587 return (__m128d){ __b[0], __a[1] }; in _mm_move_sd()
591 _mm_store_sd(double *__dp, __m128d __a) in _mm_store_sd() argument
596 ((struct __mm_store_sd_struct*)__dp)->__u = __a[0]; in _mm_store_sd()
600 _mm_store_pd(double *__dp, __m128d __a) in _mm_store_pd() argument
602 *(__m128d*)__dp = __a; in _mm_store_pd()
606 _mm_store1_pd(double *__dp, __m128d __a) in _mm_store1_pd() argument
608 __a = __builtin_shufflevector((__v2df)__a, (__v2df)__a, 0, 0); in _mm_store1_pd()
609 _mm_store_pd(__dp, __a); in _mm_store1_pd()
613 _mm_store_pd1(double *__dp, __m128d __a) in _mm_store_pd1() argument
615 return _mm_store1_pd(__dp, __a); in _mm_store_pd1()
619 _mm_storeu_pd(double *__dp, __m128d __a) in _mm_storeu_pd() argument
624 ((struct __storeu_pd*)__dp)->__v = __a; in _mm_storeu_pd()
628 _mm_storer_pd(double *__dp, __m128d __a) in _mm_storer_pd() argument
630 __a = __builtin_shufflevector((__v2df)__a, (__v2df)__a, 1, 0); in _mm_storer_pd()
631 *(__m128d *)__dp = __a; in _mm_storer_pd()
635 _mm_storeh_pd(double *__dp, __m128d __a) in _mm_storeh_pd() argument
640 ((struct __mm_storeh_pd_struct*)__dp)->__u = __a[1]; in _mm_storeh_pd()
644 _mm_storel_pd(double *__dp, __m128d __a) in _mm_storel_pd() argument
649 ((struct __mm_storeh_pd_struct*)__dp)->__u = __a[0]; in _mm_storel_pd()
653 _mm_add_epi8(__m128i __a, __m128i __b) in _mm_add_epi8() argument
655 return (__m128i)((__v16qu)__a + (__v16qu)__b); in _mm_add_epi8()
659 _mm_add_epi16(__m128i __a, __m128i __b) in _mm_add_epi16() argument
661 return (__m128i)((__v8hu)__a + (__v8hu)__b); in _mm_add_epi16()
665 _mm_add_epi32(__m128i __a, __m128i __b) in _mm_add_epi32() argument
667 return (__m128i)((__v4su)__a + (__v4su)__b); in _mm_add_epi32()
671 _mm_add_si64(__m64 __a, __m64 __b) in _mm_add_si64() argument
673 return (__m64)__builtin_ia32_paddq((__v1di)__a, (__v1di)__b); in _mm_add_si64()
677 _mm_add_epi64(__m128i __a, __m128i __b) in _mm_add_epi64() argument
679 return (__m128i)((__v2du)__a + (__v2du)__b); in _mm_add_epi64()
683 _mm_adds_epi8(__m128i __a, __m128i __b) in _mm_adds_epi8() argument
685 return (__m128i)__builtin_ia32_paddsb128((__v16qi)__a, (__v16qi)__b); in _mm_adds_epi8()
689 _mm_adds_epi16(__m128i __a, __m128i __b) in _mm_adds_epi16() argument
691 return (__m128i)__builtin_ia32_paddsw128((__v8hi)__a, (__v8hi)__b); in _mm_adds_epi16()
695 _mm_adds_epu8(__m128i __a, __m128i __b) in _mm_adds_epu8() argument
697 return (__m128i)__builtin_ia32_paddusb128((__v16qi)__a, (__v16qi)__b); in _mm_adds_epu8()
701 _mm_adds_epu16(__m128i __a, __m128i __b) in _mm_adds_epu16() argument
703 return (__m128i)__builtin_ia32_paddusw128((__v8hi)__a, (__v8hi)__b); in _mm_adds_epu16()
707 _mm_avg_epu8(__m128i __a, __m128i __b) in _mm_avg_epu8() argument
709 return (__m128i)__builtin_ia32_pavgb128((__v16qi)__a, (__v16qi)__b); in _mm_avg_epu8()
713 _mm_avg_epu16(__m128i __a, __m128i __b) in _mm_avg_epu16() argument
715 return (__m128i)__builtin_ia32_pavgw128((__v8hi)__a, (__v8hi)__b); in _mm_avg_epu16()
719 _mm_madd_epi16(__m128i __a, __m128i __b) in _mm_madd_epi16() argument
721 return (__m128i)__builtin_ia32_pmaddwd128((__v8hi)__a, (__v8hi)__b); in _mm_madd_epi16()
725 _mm_max_epi16(__m128i __a, __m128i __b) in _mm_max_epi16() argument
727 return (__m128i)__builtin_ia32_pmaxsw128((__v8hi)__a, (__v8hi)__b); in _mm_max_epi16()
731 _mm_max_epu8(__m128i __a, __m128i __b) in _mm_max_epu8() argument
733 return (__m128i)__builtin_ia32_pmaxub128((__v16qi)__a, (__v16qi)__b); in _mm_max_epu8()
737 _mm_min_epi16(__m128i __a, __m128i __b) in _mm_min_epi16() argument
739 return (__m128i)__builtin_ia32_pminsw128((__v8hi)__a, (__v8hi)__b); in _mm_min_epi16()
743 _mm_min_epu8(__m128i __a, __m128i __b) in _mm_min_epu8() argument
745 return (__m128i)__builtin_ia32_pminub128((__v16qi)__a, (__v16qi)__b); in _mm_min_epu8()
749 _mm_mulhi_epi16(__m128i __a, __m128i __b) in _mm_mulhi_epi16() argument
751 return (__m128i)__builtin_ia32_pmulhw128((__v8hi)__a, (__v8hi)__b); in _mm_mulhi_epi16()
755 _mm_mulhi_epu16(__m128i __a, __m128i __b) in _mm_mulhi_epu16() argument
757 return (__m128i)__builtin_ia32_pmulhuw128((__v8hi)__a, (__v8hi)__b); in _mm_mulhi_epu16()
774 _mm_mullo_epi16(__m128i __a, __m128i __b) in _mm_mullo_epi16() argument
776 return (__m128i)((__v8hu)__a * (__v8hu)__b); in _mm_mullo_epi16()
793 _mm_mul_su32(__m64 __a, __m64 __b) in _mm_mul_su32() argument
795 return __builtin_ia32_pmuludq((__v2si)__a, (__v2si)__b); in _mm_mul_su32()
812 _mm_mul_epu32(__m128i __a, __m128i __b) in _mm_mul_epu32() argument
814 return __builtin_ia32_pmuludq128((__v4si)__a, (__v4si)__b); in _mm_mul_epu32()
834 _mm_sad_epu8(__m128i __a, __m128i __b) in _mm_sad_epu8() argument
836 return __builtin_ia32_psadbw128((__v16qi)__a, (__v16qi)__b); in _mm_sad_epu8()
852 _mm_sub_epi8(__m128i __a, __m128i __b) in _mm_sub_epi8() argument
854 return (__m128i)((__v16qu)__a - (__v16qu)__b); in _mm_sub_epi8()
870 _mm_sub_epi16(__m128i __a, __m128i __b) in _mm_sub_epi16() argument
872 return (__m128i)((__v8hu)__a - (__v8hu)__b); in _mm_sub_epi16()
888 _mm_sub_epi32(__m128i __a, __m128i __b) in _mm_sub_epi32() argument
890 return (__m128i)((__v4su)__a - (__v4su)__b); in _mm_sub_epi32()
907 _mm_sub_si64(__m64 __a, __m64 __b) in _mm_sub_si64() argument
909 return (__m64)__builtin_ia32_psubq((__v1di)__a, (__v1di)__b); in _mm_sub_si64()
925 _mm_sub_epi64(__m128i __a, __m128i __b) in _mm_sub_epi64() argument
927 return (__m128i)((__v2du)__a - (__v2du)__b); in _mm_sub_epi64()
946 _mm_subs_epi8(__m128i __a, __m128i __b) in _mm_subs_epi8() argument
948 return (__m128i)__builtin_ia32_psubsb128((__v16qi)__a, (__v16qi)__b); in _mm_subs_epi8()
967 _mm_subs_epi16(__m128i __a, __m128i __b) in _mm_subs_epi16() argument
969 return (__m128i)__builtin_ia32_psubsw128((__v8hi)__a, (__v8hi)__b); in _mm_subs_epi16()
987 _mm_subs_epu8(__m128i __a, __m128i __b) in _mm_subs_epu8() argument
989 return (__m128i)__builtin_ia32_psubusb128((__v16qi)__a, (__v16qi)__b); in _mm_subs_epu8()
1007 _mm_subs_epu16(__m128i __a, __m128i __b) in _mm_subs_epu16() argument
1009 return (__m128i)__builtin_ia32_psubusw128((__v8hi)__a, (__v8hi)__b); in _mm_subs_epu16()
1025 _mm_and_si128(__m128i __a, __m128i __b) in _mm_and_si128() argument
1027 return (__m128i)((__v2du)__a & (__v2du)__b); in _mm_and_si128()
1045 _mm_andnot_si128(__m128i __a, __m128i __b) in _mm_andnot_si128() argument
1047 return (__m128i)(~(__v2du)__a & (__v2du)__b); in _mm_andnot_si128()
1062 _mm_or_si128(__m128i __a, __m128i __b) in _mm_or_si128() argument
1064 return (__m128i)((__v2du)__a | (__v2du)__b); in _mm_or_si128()
1080 _mm_xor_si128(__m128i __a, __m128i __b) in _mm_xor_si128() argument
1082 return (__m128i)((__v2du)__a ^ (__v2du)__b); in _mm_xor_si128()
1140 _mm_slli_epi16(__m128i __a, int __count) in _mm_slli_epi16() argument
1142 return (__m128i)__builtin_ia32_psllwi128((__v8hi)__a, __count); in _mm_slli_epi16()
1159 _mm_sll_epi16(__m128i __a, __m128i __count) in _mm_sll_epi16() argument
1161 return (__m128i)__builtin_ia32_psllw128((__v8hi)__a, (__v8hi)__count); in _mm_sll_epi16()
1178 _mm_slli_epi32(__m128i __a, int __count) in _mm_slli_epi32() argument
1180 return (__m128i)__builtin_ia32_pslldi128((__v4si)__a, __count); in _mm_slli_epi32()
1197 _mm_sll_epi32(__m128i __a, __m128i __count) in _mm_sll_epi32() argument
1199 return (__m128i)__builtin_ia32_pslld128((__v4si)__a, (__v4si)__count); in _mm_sll_epi32()
1216 _mm_slli_epi64(__m128i __a, int __count) in _mm_slli_epi64() argument
1218 return __builtin_ia32_psllqi128((__v2di)__a, __count); in _mm_slli_epi64()
1235 _mm_sll_epi64(__m128i __a, __m128i __count) in _mm_sll_epi64() argument
1237 return __builtin_ia32_psllq128((__v2di)__a, (__v2di)__count); in _mm_sll_epi64()
1255 _mm_srai_epi16(__m128i __a, int __count) in _mm_srai_epi16() argument
1257 return (__m128i)__builtin_ia32_psrawi128((__v8hi)__a, __count); in _mm_srai_epi16()
1275 _mm_sra_epi16(__m128i __a, __m128i __count) in _mm_sra_epi16() argument
1277 return (__m128i)__builtin_ia32_psraw128((__v8hi)__a, (__v8hi)__count); in _mm_sra_epi16()
1295 _mm_srai_epi32(__m128i __a, int __count) in _mm_srai_epi32() argument
1297 return (__m128i)__builtin_ia32_psradi128((__v4si)__a, __count); in _mm_srai_epi32()
1315 _mm_sra_epi32(__m128i __a, __m128i __count) in _mm_sra_epi32() argument
1317 return (__m128i)__builtin_ia32_psrad128((__v4si)__a, (__v4si)__count); in _mm_sra_epi32()
1375 _mm_srli_epi16(__m128i __a, int __count) in _mm_srli_epi16() argument
1377 return (__m128i)__builtin_ia32_psrlwi128((__v8hi)__a, __count); in _mm_srli_epi16()
1394 _mm_srl_epi16(__m128i __a, __m128i __count) in _mm_srl_epi16() argument
1396 return (__m128i)__builtin_ia32_psrlw128((__v8hi)__a, (__v8hi)__count); in _mm_srl_epi16()
1413 _mm_srli_epi32(__m128i __a, int __count) in _mm_srli_epi32() argument
1415 return (__m128i)__builtin_ia32_psrldi128((__v4si)__a, __count); in _mm_srli_epi32()
1432 _mm_srl_epi32(__m128i __a, __m128i __count) in _mm_srl_epi32() argument
1434 return (__m128i)__builtin_ia32_psrld128((__v4si)__a, (__v4si)__count); in _mm_srl_epi32()
1451 _mm_srli_epi64(__m128i __a, int __count) in _mm_srli_epi64() argument
1453 return __builtin_ia32_psrlqi128((__v2di)__a, __count); in _mm_srli_epi64()
1470 _mm_srl_epi64(__m128i __a, __m128i __count) in _mm_srl_epi64() argument
1472 return __builtin_ia32_psrlq128((__v2di)__a, (__v2di)__count); in _mm_srl_epi64()
1489 _mm_cmpeq_epi8(__m128i __a, __m128i __b) in _mm_cmpeq_epi8() argument
1491 return (__m128i)((__v16qi)__a == (__v16qi)__b); in _mm_cmpeq_epi8()
1508 _mm_cmpeq_epi16(__m128i __a, __m128i __b) in _mm_cmpeq_epi16() argument
1510 return (__m128i)((__v8hi)__a == (__v8hi)__b); in _mm_cmpeq_epi16()
1527 _mm_cmpeq_epi32(__m128i __a, __m128i __b) in _mm_cmpeq_epi32() argument
1529 return (__m128i)((__v4si)__a == (__v4si)__b); in _mm_cmpeq_epi32()
1547 _mm_cmpgt_epi8(__m128i __a, __m128i __b) in _mm_cmpgt_epi8() argument
1551 return (__m128i)((__v16qs)__a > (__v16qs)__b); in _mm_cmpgt_epi8()
1569 _mm_cmpgt_epi16(__m128i __a, __m128i __b) in _mm_cmpgt_epi16() argument
1571 return (__m128i)((__v8hi)__a > (__v8hi)__b); in _mm_cmpgt_epi16()
1589 _mm_cmpgt_epi32(__m128i __a, __m128i __b) in _mm_cmpgt_epi32() argument
1591 return (__m128i)((__v4si)__a > (__v4si)__b); in _mm_cmpgt_epi32()
1609 _mm_cmplt_epi8(__m128i __a, __m128i __b) in _mm_cmplt_epi8() argument
1611 return _mm_cmpgt_epi8(__b, __a); in _mm_cmplt_epi8()
1629 _mm_cmplt_epi16(__m128i __a, __m128i __b) in _mm_cmplt_epi16() argument
1631 return _mm_cmpgt_epi16(__b, __a); in _mm_cmplt_epi16()
1649 _mm_cmplt_epi32(__m128i __a, __m128i __b) in _mm_cmplt_epi32() argument
1651 return _mm_cmpgt_epi32(__b, __a); in _mm_cmplt_epi32()
1673 _mm_cvtsi64_sd(__m128d __a, long long __b) in _mm_cvtsi64_sd() argument
1675 __a[0] = __b; in _mm_cvtsi64_sd()
1676 return __a; in _mm_cvtsi64_sd()
1691 _mm_cvtsd_si64(__m128d __a) in _mm_cvtsd_si64() argument
1693 return __builtin_ia32_cvtsd2si64((__v2df)__a); in _mm_cvtsd_si64()
1708 _mm_cvttsd_si64(__m128d __a) in _mm_cvttsd_si64() argument
1710 return __a[0]; in _mm_cvttsd_si64()
1724 _mm_cvtepi32_ps(__m128i __a) in _mm_cvtepi32_ps() argument
1726 return __builtin_ia32_cvtdq2ps((__v4si)__a); in _mm_cvtepi32_ps()
1740 _mm_cvtps_epi32(__m128 __a) in _mm_cvtps_epi32() argument
1742 return (__m128i)__builtin_ia32_cvtps2dq((__v4sf)__a); in _mm_cvtps_epi32()
1756 _mm_cvttps_epi32(__m128 __a) in _mm_cvttps_epi32() argument
1758 return (__m128i)__builtin_convertvector((__v4sf)__a, __v4si); in _mm_cvttps_epi32()
1772 _mm_cvtsi32_si128(int __a) in _mm_cvtsi32_si128() argument
1774 return (__m128i)(__v4si){ __a, 0, 0, 0 }; in _mm_cvtsi32_si128()
1789 _mm_cvtsi64_si128(long long __a) in _mm_cvtsi64_si128() argument
1791 return (__m128i){ __a, 0 }; in _mm_cvtsi64_si128()
1807 _mm_cvtsi128_si32(__m128i __a) in _mm_cvtsi128_si32() argument
1809 __v4si __b = (__v4si)__a; in _mm_cvtsi128_si32()
1826 _mm_cvtsi128_si64(__m128i __a) in _mm_cvtsi128_si64() argument
1828 return __a[0]; in _mm_cvtsi128_si64()
2210 _mm_storel_epi64(__m128i *__p, __m128i __a) in _mm_storel_epi64() argument
2215 ((struct __mm_storel_epi64_struct*)__p)->__u = __a[0]; in _mm_storel_epi64()
2219 _mm_stream_pd(double *__p, __m128d __a) in _mm_stream_pd() argument
2221 __builtin_nontemporal_store((__v2df)__a, (__v2df*)__p); in _mm_stream_pd()
2225 _mm_stream_si128(__m128i *__p, __m128i __a) in _mm_stream_si128() argument
2227 __builtin_nontemporal_store((__v2di)__a, (__v2di*)__p); in _mm_stream_si128()
2231 _mm_stream_si32(int *__p, int __a) in _mm_stream_si32() argument
2233 __builtin_ia32_movnti(__p, __a); in _mm_stream_si32()
2238 _mm_stream_si64(long long *__p, long long __a) in _mm_stream_si64() argument
2240 __builtin_ia32_movnti64(__p, __a); in _mm_stream_si64()
2263 _mm_packs_epi16(__m128i __a, __m128i __b) in _mm_packs_epi16() argument
2265 return (__m128i)__builtin_ia32_packsswb128((__v8hi)__a, (__v8hi)__b); in _mm_packs_epi16()
2269 _mm_packs_epi32(__m128i __a, __m128i __b) in _mm_packs_epi32() argument
2271 return (__m128i)__builtin_ia32_packssdw128((__v4si)__a, (__v4si)__b); in _mm_packs_epi32()
2275 _mm_packus_epi16(__m128i __a, __m128i __b) in _mm_packus_epi16() argument
2277 return (__m128i)__builtin_ia32_packuswb128((__v8hi)__a, (__v8hi)__b); in _mm_packus_epi16()
2281 _mm_extract_epi16(__m128i __a, int __imm) in _mm_extract_epi16() argument
2283 __v8hi __b = (__v8hi)__a; in _mm_extract_epi16()
2288 _mm_insert_epi16(__m128i __a, int __b, int __imm) in _mm_insert_epi16() argument
2290 __v8hi __c = (__v8hi)__a; in _mm_insert_epi16()
2296 _mm_movemask_epi8(__m128i __a) in _mm_movemask_epi8() argument
2298 return __builtin_ia32_pmovmskb128((__v16qi)__a); in _mm_movemask_epi8()
2324 _mm_unpackhi_epi8(__m128i __a, __m128i __b) in _mm_unpackhi_epi8() argument
2326 …return (__m128i)__builtin_shufflevector((__v16qi)__a, (__v16qi)__b, 8, 16+8, 9, 16+9, 10, 16+10, 1… in _mm_unpackhi_epi8()
2330 _mm_unpackhi_epi16(__m128i __a, __m128i __b) in _mm_unpackhi_epi16() argument
2332 return (__m128i)__builtin_shufflevector((__v8hi)__a, (__v8hi)__b, 4, 8+4, 5, 8+5, 6, 8+6, 7, 8+7); in _mm_unpackhi_epi16()
2336 _mm_unpackhi_epi32(__m128i __a, __m128i __b) in _mm_unpackhi_epi32() argument
2338 return (__m128i)__builtin_shufflevector((__v4si)__a, (__v4si)__b, 2, 4+2, 3, 4+3); in _mm_unpackhi_epi32()
2342 _mm_unpackhi_epi64(__m128i __a, __m128i __b) in _mm_unpackhi_epi64() argument
2344 return (__m128i)__builtin_shufflevector((__v2di)__a, (__v2di)__b, 1, 2+1); in _mm_unpackhi_epi64()
2348 _mm_unpacklo_epi8(__m128i __a, __m128i __b) in _mm_unpacklo_epi8() argument
2350 …return (__m128i)__builtin_shufflevector((__v16qi)__a, (__v16qi)__b, 0, 16+0, 1, 16+1, 2, 16+2, 3, … in _mm_unpacklo_epi8()
2354 _mm_unpacklo_epi16(__m128i __a, __m128i __b) in _mm_unpacklo_epi16() argument
2356 return (__m128i)__builtin_shufflevector((__v8hi)__a, (__v8hi)__b, 0, 8+0, 1, 8+1, 2, 8+2, 3, 8+3); in _mm_unpacklo_epi16()
2360 _mm_unpacklo_epi32(__m128i __a, __m128i __b) in _mm_unpacklo_epi32() argument
2362 return (__m128i)__builtin_shufflevector((__v4si)__a, (__v4si)__b, 0, 4+0, 1, 4+1); in _mm_unpacklo_epi32()
2366 _mm_unpacklo_epi64(__m128i __a, __m128i __b) in _mm_unpacklo_epi64() argument
2368 return (__m128i)__builtin_shufflevector((__v2di)__a, (__v2di)__b, 0, 2+0); in _mm_unpacklo_epi64()
2372 _mm_movepi64_pi64(__m128i __a) in _mm_movepi64_pi64() argument
2374 return (__m64)__a[0]; in _mm_movepi64_pi64()
2378 _mm_movpi64_epi64(__m64 __a) in _mm_movpi64_epi64() argument
2380 return (__m128i){ (long long)__a, 0 }; in _mm_movpi64_epi64()
2384 _mm_move_epi64(__m128i __a) in _mm_move_epi64() argument
2386 return __builtin_shufflevector((__v2di)__a, (__m128i){ 0 }, 0, 2); in _mm_move_epi64()
2390 _mm_unpackhi_pd(__m128d __a, __m128d __b) in _mm_unpackhi_pd() argument
2392 return __builtin_shufflevector((__v2df)__a, (__v2df)__b, 1, 2+1); in _mm_unpackhi_pd()
2396 _mm_unpacklo_pd(__m128d __a, __m128d __b) in _mm_unpacklo_pd() argument
2398 return __builtin_shufflevector((__v2df)__a, (__v2df)__b, 0, 2+0); in _mm_unpacklo_pd()
2402 _mm_movemask_pd(__m128d __a) in _mm_movemask_pd() argument
2404 return __builtin_ia32_movmskpd((__v2df)__a); in _mm_movemask_pd()
2413 _mm_castpd_ps(__m128d __a) in _mm_castpd_ps() argument
2415 return (__m128)__a; in _mm_castpd_ps()
2419 _mm_castpd_si128(__m128d __a) in _mm_castpd_si128() argument
2421 return (__m128i)__a; in _mm_castpd_si128()
2425 _mm_castps_pd(__m128 __a) in _mm_castps_pd() argument
2427 return (__m128d)__a; in _mm_castps_pd()
2431 _mm_castps_si128(__m128 __a) in _mm_castps_si128() argument
2433 return (__m128i)__a; in _mm_castps_si128()
2437 _mm_castsi128_ps(__m128i __a) in _mm_castsi128_ps() argument
2439 return (__m128)__a; in _mm_castsi128_ps()
2443 _mm_castsi128_pd(__m128i __a) in _mm_castsi128_pd() argument
2445 return (__m128d)__a; in _mm_castsi128_pd()