33 typedef double __m128d __attribute__((__vector_size__(16)));  typedef
42 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
43 _mm_add_sd(__m128d a, __m128d b)  in _mm_add_sd()
49 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
50 _mm_add_pd(__m128d a, __m128d b)  in _mm_add_pd()
55 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
56 _mm_sub_sd(__m128d a, __m128d b)  in _mm_sub_sd()
62 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
63 _mm_sub_pd(__m128d a, __m128d b)  in _mm_sub_pd()
68 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
69 _mm_mul_sd(__m128d a, __m128d b)  in _mm_mul_sd()
75 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
76 _mm_mul_pd(__m128d a, __m128d b)  in _mm_mul_pd()
81 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
82 _mm_div_sd(__m128d a, __m128d b)  in _mm_div_sd()
88 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
89 _mm_div_pd(__m128d a, __m128d b)  in _mm_div_pd()
94 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
95 _mm_sqrt_sd(__m128d a, __m128d b)  in _mm_sqrt_sd()
97   __m128d c = __builtin_ia32_sqrtsd(b);  in _mm_sqrt_sd()
98   return (__m128d) { c[0], a[1] };  in _mm_sqrt_sd()
101 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
102 _mm_sqrt_pd(__m128d a)  in _mm_sqrt_pd()
107 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
108 _mm_min_sd(__m128d a, __m128d b)  in _mm_min_sd()
113 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
114 _mm_min_pd(__m128d a, __m128d b)  in _mm_min_pd()
119 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
120 _mm_max_sd(__m128d a, __m128d b)  in _mm_max_sd()
125 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
126 _mm_max_pd(__m128d a, __m128d b)  in _mm_max_pd()
131 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
132 _mm_and_pd(__m128d a, __m128d b)  in _mm_and_pd()
134   return (__m128d)((__v4si)a & (__v4si)b);  in _mm_and_pd()
137 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
138 _mm_andnot_pd(__m128d a, __m128d b)  in _mm_andnot_pd()
140   return (__m128d)(~(__v4si)a & (__v4si)b);  in _mm_andnot_pd()
143 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
144 _mm_or_pd(__m128d a, __m128d b)  in _mm_or_pd()
146   return (__m128d)((__v4si)a | (__v4si)b);  in _mm_or_pd()
149 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
150 _mm_xor_pd(__m128d a, __m128d b)  in _mm_xor_pd()
152   return (__m128d)((__v4si)a ^ (__v4si)b);  in _mm_xor_pd()
155 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
156 _mm_cmpeq_pd(__m128d a, __m128d b)  in _mm_cmpeq_pd()
158   return (__m128d)__builtin_ia32_cmppd(a, b, 0);  in _mm_cmpeq_pd()
161 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
162 _mm_cmplt_pd(__m128d a, __m128d b)  in _mm_cmplt_pd()
164   return (__m128d)__builtin_ia32_cmppd(a, b, 1);  in _mm_cmplt_pd()
167 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
168 _mm_cmple_pd(__m128d a, __m128d b)  in _mm_cmple_pd()
170   return (__m128d)__builtin_ia32_cmppd(a, b, 2);  in _mm_cmple_pd()
173 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
174 _mm_cmpgt_pd(__m128d a, __m128d b)  in _mm_cmpgt_pd()
176   return (__m128d)__builtin_ia32_cmppd(b, a, 1);  in _mm_cmpgt_pd()
179 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
180 _mm_cmpge_pd(__m128d a, __m128d b)  in _mm_cmpge_pd()
182   return (__m128d)__builtin_ia32_cmppd(b, a, 2);  in _mm_cmpge_pd()
185 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
186 _mm_cmpord_pd(__m128d a, __m128d b)  in _mm_cmpord_pd()
188   return (__m128d)__builtin_ia32_cmppd(a, b, 7);  in _mm_cmpord_pd()
191 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
192 _mm_cmpunord_pd(__m128d a, __m128d b)  in _mm_cmpunord_pd()
194   return (__m128d)__builtin_ia32_cmppd(a, b, 3);  in _mm_cmpunord_pd()
197 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
198 _mm_cmpneq_pd(__m128d a, __m128d b)  in _mm_cmpneq_pd()
200   return (__m128d)__builtin_ia32_cmppd(a, b, 4);  in _mm_cmpneq_pd()
203 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
204 _mm_cmpnlt_pd(__m128d a, __m128d b)  in _mm_cmpnlt_pd()
206   return (__m128d)__builtin_ia32_cmppd(a, b, 5);  in _mm_cmpnlt_pd()
209 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
210 _mm_cmpnle_pd(__m128d a, __m128d b)  in _mm_cmpnle_pd()
212   return (__m128d)__builtin_ia32_cmppd(a, b, 6);  in _mm_cmpnle_pd()
215 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
216 _mm_cmpngt_pd(__m128d a, __m128d b)  in _mm_cmpngt_pd()
218   return (__m128d)__builtin_ia32_cmppd(b, a, 5);  in _mm_cmpngt_pd()
221 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
222 _mm_cmpnge_pd(__m128d a, __m128d b)  in _mm_cmpnge_pd()
224   return (__m128d)__builtin_ia32_cmppd(b, a, 6);  in _mm_cmpnge_pd()
227 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
228 _mm_cmpeq_sd(__m128d a, __m128d b)  in _mm_cmpeq_sd()
230   return (__m128d)__builtin_ia32_cmpsd(a, b, 0);  in _mm_cmpeq_sd()
233 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
234 _mm_cmplt_sd(__m128d a, __m128d b)  in _mm_cmplt_sd()
236   return (__m128d)__builtin_ia32_cmpsd(a, b, 1);  in _mm_cmplt_sd()
239 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
240 _mm_cmple_sd(__m128d a, __m128d b)  in _mm_cmple_sd()
242   return (__m128d)__builtin_ia32_cmpsd(a, b, 2);  in _mm_cmple_sd()
245 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
246 _mm_cmpgt_sd(__m128d a, __m128d b)  in _mm_cmpgt_sd()
248   return (__m128d)__builtin_ia32_cmpsd(b, a, 1);  in _mm_cmpgt_sd()
251 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
252 _mm_cmpge_sd(__m128d a, __m128d b)  in _mm_cmpge_sd()
254   return (__m128d)__builtin_ia32_cmpsd(b, a, 2);  in _mm_cmpge_sd()
257 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
258 _mm_cmpord_sd(__m128d a, __m128d b)  in _mm_cmpord_sd()
260   return (__m128d)__builtin_ia32_cmpsd(a, b, 7);  in _mm_cmpord_sd()
263 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
264 _mm_cmpunord_sd(__m128d a, __m128d b)  in _mm_cmpunord_sd()
266   return (__m128d)__builtin_ia32_cmpsd(a, b, 3);  in _mm_cmpunord_sd()
269 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
270 _mm_cmpneq_sd(__m128d a, __m128d b)  in _mm_cmpneq_sd()
272   return (__m128d)__builtin_ia32_cmpsd(a, b, 4);  in _mm_cmpneq_sd()
275 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
276 _mm_cmpnlt_sd(__m128d a, __m128d b)  in _mm_cmpnlt_sd()
278   return (__m128d)__builtin_ia32_cmpsd(a, b, 5);  in _mm_cmpnlt_sd()
281 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
282 _mm_cmpnle_sd(__m128d a, __m128d b)  in _mm_cmpnle_sd()
284   return (__m128d)__builtin_ia32_cmpsd(a, b, 6);  in _mm_cmpnle_sd()
287 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
288 _mm_cmpngt_sd(__m128d a, __m128d b)  in _mm_cmpngt_sd()
290   return (__m128d)__builtin_ia32_cmpsd(b, a, 5);  in _mm_cmpngt_sd()
293 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
294 _mm_cmpnge_sd(__m128d a, __m128d b)  in _mm_cmpnge_sd()
296   return (__m128d)__builtin_ia32_cmpsd(b, a, 6);  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_comige_sd(__m128d a, __m128d b)  in _mm_comige_sd()
330 _mm_comineq_sd(__m128d a, __m128d b)  in _mm_comineq_sd()
336 _mm_ucomieq_sd(__m128d a, __m128d b)  in _mm_ucomieq_sd()
342 _mm_ucomilt_sd(__m128d a, __m128d b)  in _mm_ucomilt_sd()
348 _mm_ucomile_sd(__m128d a, __m128d b)  in _mm_ucomile_sd()
354 _mm_ucomigt_sd(__m128d a, __m128d b)  in _mm_ucomigt_sd()
360 _mm_ucomige_sd(__m128d a, __m128d b)  in _mm_ucomige_sd()
366 _mm_ucomineq_sd(__m128d a, __m128d b)  in _mm_ucomineq_sd()
372 _mm_cvtpd_ps(__m128d a)  in _mm_cvtpd_ps()
377 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
383 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
390 _mm_cvtpd_epi32(__m128d a)  in _mm_cvtpd_epi32()
396 _mm_cvtsd_si32(__m128d a)  in _mm_cvtsd_si32()
402 _mm_cvtsd_ss(__m128 a, __m128d b)  in _mm_cvtsd_ss()
408 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
409 _mm_cvtsi32_sd(__m128d a, int b)  in _mm_cvtsi32_sd()
415 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
416 _mm_cvtss_sd(__m128d a, __m128 b)  in _mm_cvtss_sd()
423 _mm_cvttpd_epi32(__m128d a)  in _mm_cvttpd_epi32()
429 _mm_cvttsd_si32(__m128d a)  in _mm_cvttsd_si32()
435 _mm_cvtpd_pi32(__m128d a)  in _mm_cvtpd_pi32()
441 _mm_cvttpd_pi32(__m128d a)  in _mm_cvttpd_pi32()
446 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
453 _mm_cvtsd_f64(__m128d a)  in _mm_cvtsd_f64()
458 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
461   return *(__m128d*)dp;  in _mm_load_pd()
464 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
471   return (__m128d){ u, u };  in _mm_load1_pd()
476 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
479   __m128d u = *(__m128d*)dp;  in _mm_loadr_pd()
483 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
487     __m128d v;  in _mm_loadu_pd()
492 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
499   return (__m128d){ u, 0 };  in _mm_load_sd()
502 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
503 _mm_loadh_pd(__m128d a, double const *dp)  in _mm_loadh_pd()
509   return (__m128d){ a[0], u };  in _mm_loadh_pd()
512 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
513 _mm_loadl_pd(__m128d a, double const *dp)  in _mm_loadl_pd()
519   return (__m128d){ u, a[1] };   in _mm_loadl_pd()
522 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
525   return (__m128d){ w, 0 };  in _mm_set_sd()
528 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
531   return (__m128d){ w, w };  in _mm_set1_pd()
534 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
537   return (__m128d){ x, w };  in _mm_set_pd()
540 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
543   return (__m128d){ w, x };  in _mm_setr_pd()
546 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
549   return (__m128d){ 0, 0 };  in _mm_setzero_pd()
552 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
553 _mm_move_sd(__m128d a, __m128d b)  in _mm_move_sd()
555   return (__m128d){ b[0], a[1] };  in _mm_move_sd()
559 _mm_store_sd(double *dp, __m128d a)  in _mm_store_sd()
568 _mm_store1_pd(double *dp, __m128d a)  in _mm_store1_pd()
578 _mm_store_pd(double *dp, __m128d a)  in _mm_store_pd()
580   *(__m128d *)dp = a;  in _mm_store_pd()
584 _mm_storeu_pd(double *dp, __m128d a)  in _mm_storeu_pd()
590 _mm_storer_pd(double *dp, __m128d a)  in _mm_storer_pd()
593   *(__m128d *)dp = a;  in _mm_storer_pd()
597 _mm_storeh_pd(double *dp, __m128d a)  in _mm_storeh_pd()
606 _mm_storel_pd(double *dp, __m128d a)  in _mm_storel_pd()
987 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
988 _mm_cvtsi64_sd(__m128d a, long long b)  in _mm_cvtsi64_sd()
995 _mm_cvtsd_si64(__m128d a)  in _mm_cvtsd_si64()
1001 _mm_cvttsd_si64(__m128d a)  in _mm_cvttsd_si64()
1196 _mm_stream_pd(double *p, __m128d a)  in _mm_stream_pd()
1358 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
1359 _mm_unpackhi_pd(__m128d a, __m128d b)  in _mm_unpackhi_pd()
1364 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
1365 _mm_unpacklo_pd(__m128d a, __m128d b)  in _mm_unpacklo_pd()
1371 _mm_movemask_pd(__m128d a)  in _mm_movemask_pd()
1377   __m128d __a = (a); \
1378   __m128d __b = (b); \
1382 _mm_castpd_ps(__m128d in)  in _mm_castpd_ps()
1388 _mm_castpd_si128(__m128d in)  in _mm_castpd_si128()
1393 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
1396   return (__m128d)in;  in _mm_castps_pd()
1411 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
1414   return (__m128d)in;  in _mm_castsi128_pd()