1 /* Copyright (C) 2003, 2004, 2005 Free Software Foundation, Inc.
3 This file is part of GCC.
5 GCC is free software; you can redistribute it and/or modify
6 it under the terms of the GNU General Public License as published by
7 the Free Software Foundation; either version 2, or (at your option)
10 GCC is distributed in the hope that it will be useful,
11 but WITHOUT ANY WARRANTY; without even the implied warranty of
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 GNU General Public License for more details.
15 You should have received a copy of the GNU General Public License
16 along with GCC; see the file COPYING. If not, write to
17 the Free Software Foundation, 59 Temple Place - Suite 330,
18 Boston, MA 02111-1307, USA. */
20 /* As a special exception, if you include this header file into source
21 files compiled by GCC, this header file does not by itself cause
22 the resulting executable to be covered by the GNU General Public
23 License. This exception does not however invalidate any other
24 reasons why the executable file might be covered by the GNU General
27 /* Implemented from the specification included in the Intel C++ Compiler
28 User Guide and Reference, version 8.0. */
30 #ifndef _EMMINTRIN_H_INCLUDED
31 #define _EMMINTRIN_H_INCLUDED
34 #include <xmmintrin.h>
37 typedef double __v2df
__attribute__ ((__vector_size__ (16)));
38 typedef long long __v2di
__attribute__ ((__vector_size__ (16)));
39 typedef int __v4si
__attribute__ ((__vector_size__ (16)));
40 typedef short __v8hi
__attribute__ ((__vector_size__ (16)));
41 typedef char __v16qi
__attribute__ ((__vector_size__ (16)));
43 typedef __v2di __m128i
;
44 typedef __v2df __m128d
;
46 /* Create a selector for use with the SHUFPD instruction. */
47 #define _MM_SHUFFLE2(fp1,fp0) \
48 (((fp1) << 1) | (fp0))
50 /* Create a vector with element 0 as F and the rest zero. */
51 static __inline __m128d
52 _mm_set_sd (double __F
)
54 return __extension__ (__m128d
){ __F
, 0 };
57 /* Create a vector with both elements equal to F. */
58 static __inline __m128d
59 _mm_set1_pd (double __F
)
61 return __extension__ (__m128d
){ __F
, __F
};
64 static __inline __m128d
65 _mm_set_pd1 (double __F
)
67 return _mm_set1_pd (__F
);
70 /* Create a vector with the lower value X and upper value W. */
71 static __inline __m128d
72 _mm_set_pd (double __W
, double __X
)
74 return __extension__ (__m128d
){ __X
, __W
};
77 /* Create a vector with the lower value W and upper value X. */
78 static __inline __m128d
79 _mm_setr_pd (double __W
, double __X
)
81 return __extension__ (__m128d
){ __W
, __X
};
84 /* Create a vector of zeros. */
85 static __inline __m128d
88 return __extension__ (__m128d
){ 0.0, 0.0 };
91 /* Sets the low DPFP value of A from the low value of B. */
92 static __inline __m128d
93 _mm_move_sd (__m128d __A
, __m128d __B
)
95 return (__m128d
) __builtin_ia32_movsd ((__v2df
)__A
, (__v2df
)__B
);
98 /* Load two DPFP values from P. The address must be 16-byte aligned. */
99 static __inline __m128d
100 _mm_load_pd (double const *__P
)
102 return *(__m128d
*)__P
;
105 /* Load two DPFP values from P. The address need not be 16-byte aligned. */
106 static __inline __m128d
107 _mm_loadu_pd (double const *__P
)
109 return __builtin_ia32_loadupd (__P
);
112 /* Create a vector with all two elements equal to *P. */
113 static __inline __m128d
114 _mm_load1_pd (double const *__P
)
116 return _mm_set1_pd (*__P
);
119 /* Create a vector with element 0 as *P and the rest zero. */
120 static __inline __m128d
121 _mm_load_sd (double const *__P
)
123 return _mm_set_sd (*__P
);
126 static __inline __m128d
127 _mm_load_pd1 (double const *__P
)
129 return _mm_load1_pd (__P
);
132 /* Load two DPFP values in reverse order. The address must be aligned. */
133 static __inline __m128d
134 _mm_loadr_pd (double const *__P
)
136 __m128d __tmp
= _mm_load_pd (__P
);
137 return __builtin_ia32_shufpd (__tmp
, __tmp
, _MM_SHUFFLE2 (0,1));
140 /* Store two DPFP values. The address must be 16-byte aligned. */
142 _mm_store_pd (double *__P
, __m128d __A
)
144 *(__m128d
*)__P
= __A
;
147 /* Store two DPFP values. The address need not be 16-byte aligned. */
149 _mm_storeu_pd (double *__P
, __m128d __A
)
151 __builtin_ia32_storeupd (__P
, __A
);
154 /* Stores the lower DPFP value. */
156 _mm_store_sd (double *__P
, __m128d __A
)
158 *__P
= __builtin_ia32_vec_ext_v2df (__A
, 0);
162 _mm_storel_pd (double *__P
, __m128d __A
)
164 _mm_store_sd (__P
, __A
);
167 /* Stores the upper DPFP value. */
169 _mm_storeh_pd (double *__P
, __m128d __A
)
171 *__P
= __builtin_ia32_vec_ext_v2df (__A
, 1);
174 /* Store the lower DPFP value across two words.
175 The address must be 16-byte aligned. */
177 _mm_store1_pd (double *__P
, __m128d __A
)
179 _mm_store_pd (__P
, __builtin_ia32_shufpd (__A
, __A
, _MM_SHUFFLE2 (0,0)));
183 _mm_store_pd1 (double *__P
, __m128d __A
)
185 _mm_store1_pd (__P
, __A
);
188 /* Store two DPFP values in reverse order. The address must be aligned. */
190 _mm_storer_pd (double *__P
, __m128d __A
)
192 _mm_store_pd (__P
, __builtin_ia32_shufpd (__A
, __A
, _MM_SHUFFLE2 (0,1)));
196 _mm_cvtsi128_si32 (__m128i __A
)
198 return __builtin_ia32_vec_ext_v4si ((__v4si
)__A
, 0);
202 static __inline
long long
203 _mm_cvtsi128_si64x (__m128i __A
)
205 return __builtin_ia32_vec_ext_v2di ((__v2di
)__A
, 0);
209 static __inline __m128d
210 _mm_add_pd (__m128d __A
, __m128d __B
)
212 return (__m128d
)__builtin_ia32_addpd ((__v2df
)__A
, (__v2df
)__B
);
215 static __inline __m128d
216 _mm_add_sd (__m128d __A
, __m128d __B
)
218 return (__m128d
)__builtin_ia32_addsd ((__v2df
)__A
, (__v2df
)__B
);
221 static __inline __m128d
222 _mm_sub_pd (__m128d __A
, __m128d __B
)
224 return (__m128d
)__builtin_ia32_subpd ((__v2df
)__A
, (__v2df
)__B
);
227 static __inline __m128d
228 _mm_sub_sd (__m128d __A
, __m128d __B
)
230 return (__m128d
)__builtin_ia32_subsd ((__v2df
)__A
, (__v2df
)__B
);
233 static __inline __m128d
234 _mm_mul_pd (__m128d __A
, __m128d __B
)
236 return (__m128d
)__builtin_ia32_mulpd ((__v2df
)__A
, (__v2df
)__B
);
239 static __inline __m128d
240 _mm_mul_sd (__m128d __A
, __m128d __B
)
242 return (__m128d
)__builtin_ia32_mulsd ((__v2df
)__A
, (__v2df
)__B
);
245 static __inline __m128d
246 _mm_div_pd (__m128d __A
, __m128d __B
)
248 return (__m128d
)__builtin_ia32_divpd ((__v2df
)__A
, (__v2df
)__B
);
251 static __inline __m128d
252 _mm_div_sd (__m128d __A
, __m128d __B
)
254 return (__m128d
)__builtin_ia32_divsd ((__v2df
)__A
, (__v2df
)__B
);
257 static __inline __m128d
258 _mm_sqrt_pd (__m128d __A
)
260 return (__m128d
)__builtin_ia32_sqrtpd ((__v2df
)__A
);
263 /* Return pair {sqrt (A[0), B[1]}. */
264 static __inline __m128d
265 _mm_sqrt_sd (__m128d __A
, __m128d __B
)
267 __v2df __tmp
= __builtin_ia32_movsd ((__v2df
)__A
, (__v2df
)__B
);
268 return (__m128d
)__builtin_ia32_sqrtsd ((__v2df
)__tmp
);
271 static __inline __m128d
272 _mm_min_pd (__m128d __A
, __m128d __B
)
274 return (__m128d
)__builtin_ia32_minpd ((__v2df
)__A
, (__v2df
)__B
);
277 static __inline __m128d
278 _mm_min_sd (__m128d __A
, __m128d __B
)
280 return (__m128d
)__builtin_ia32_minsd ((__v2df
)__A
, (__v2df
)__B
);
283 static __inline __m128d
284 _mm_max_pd (__m128d __A
, __m128d __B
)
286 return (__m128d
)__builtin_ia32_maxpd ((__v2df
)__A
, (__v2df
)__B
);
289 static __inline __m128d
290 _mm_max_sd (__m128d __A
, __m128d __B
)
292 return (__m128d
)__builtin_ia32_maxsd ((__v2df
)__A
, (__v2df
)__B
);
295 static __inline __m128d
296 _mm_and_pd (__m128d __A
, __m128d __B
)
298 return (__m128d
)__builtin_ia32_andpd ((__v2df
)__A
, (__v2df
)__B
);
301 static __inline __m128d
302 _mm_andnot_pd (__m128d __A
, __m128d __B
)
304 return (__m128d
)__builtin_ia32_andnpd ((__v2df
)__A
, (__v2df
)__B
);
307 static __inline __m128d
308 _mm_or_pd (__m128d __A
, __m128d __B
)
310 return (__m128d
)__builtin_ia32_orpd ((__v2df
)__A
, (__v2df
)__B
);
313 static __inline __m128d
314 _mm_xor_pd (__m128d __A
, __m128d __B
)
316 return (__m128d
)__builtin_ia32_xorpd ((__v2df
)__A
, (__v2df
)__B
);
319 static __inline __m128d
320 _mm_cmpeq_pd (__m128d __A
, __m128d __B
)
322 return (__m128d
)__builtin_ia32_cmpeqpd ((__v2df
)__A
, (__v2df
)__B
);
325 static __inline __m128d
326 _mm_cmplt_pd (__m128d __A
, __m128d __B
)
328 return (__m128d
)__builtin_ia32_cmpltpd ((__v2df
)__A
, (__v2df
)__B
);
331 static __inline __m128d
332 _mm_cmple_pd (__m128d __A
, __m128d __B
)
334 return (__m128d
)__builtin_ia32_cmplepd ((__v2df
)__A
, (__v2df
)__B
);
337 static __inline __m128d
338 _mm_cmpgt_pd (__m128d __A
, __m128d __B
)
340 return (__m128d
)__builtin_ia32_cmpgtpd ((__v2df
)__A
, (__v2df
)__B
);
343 static __inline __m128d
344 _mm_cmpge_pd (__m128d __A
, __m128d __B
)
346 return (__m128d
)__builtin_ia32_cmpgepd ((__v2df
)__A
, (__v2df
)__B
);
349 static __inline __m128d
350 _mm_cmpneq_pd (__m128d __A
, __m128d __B
)
352 return (__m128d
)__builtin_ia32_cmpneqpd ((__v2df
)__A
, (__v2df
)__B
);
355 static __inline __m128d
356 _mm_cmpnlt_pd (__m128d __A
, __m128d __B
)
358 return (__m128d
)__builtin_ia32_cmpnltpd ((__v2df
)__A
, (__v2df
)__B
);
361 static __inline __m128d
362 _mm_cmpnle_pd (__m128d __A
, __m128d __B
)
364 return (__m128d
)__builtin_ia32_cmpnlepd ((__v2df
)__A
, (__v2df
)__B
);
367 static __inline __m128d
368 _mm_cmpngt_pd (__m128d __A
, __m128d __B
)
370 return (__m128d
)__builtin_ia32_cmpngtpd ((__v2df
)__A
, (__v2df
)__B
);
373 static __inline __m128d
374 _mm_cmpnge_pd (__m128d __A
, __m128d __B
)
376 return (__m128d
)__builtin_ia32_cmpngepd ((__v2df
)__A
, (__v2df
)__B
);
379 static __inline __m128d
380 _mm_cmpord_pd (__m128d __A
, __m128d __B
)
382 return (__m128d
)__builtin_ia32_cmpordpd ((__v2df
)__A
, (__v2df
)__B
);
385 static __inline __m128d
386 _mm_cmpunord_pd (__m128d __A
, __m128d __B
)
388 return (__m128d
)__builtin_ia32_cmpunordpd ((__v2df
)__A
, (__v2df
)__B
);
391 static __inline __m128d
392 _mm_cmpeq_sd (__m128d __A
, __m128d __B
)
394 return (__m128d
)__builtin_ia32_cmpeqsd ((__v2df
)__A
, (__v2df
)__B
);
397 static __inline __m128d
398 _mm_cmplt_sd (__m128d __A
, __m128d __B
)
400 return (__m128d
)__builtin_ia32_cmpltsd ((__v2df
)__A
, (__v2df
)__B
);
403 static __inline __m128d
404 _mm_cmple_sd (__m128d __A
, __m128d __B
)
406 return (__m128d
)__builtin_ia32_cmplesd ((__v2df
)__A
, (__v2df
)__B
);
409 static __inline __m128d
410 _mm_cmpgt_sd (__m128d __A
, __m128d __B
)
412 return (__m128d
) __builtin_ia32_movsd ((__v2df
) __A
,
414 __builtin_ia32_cmpltsd ((__v2df
) __B
,
419 static __inline __m128d
420 _mm_cmpge_sd (__m128d __A
, __m128d __B
)
422 return (__m128d
) __builtin_ia32_movsd ((__v2df
) __A
,
424 __builtin_ia32_cmplesd ((__v2df
) __B
,
429 static __inline __m128d
430 _mm_cmpneq_sd (__m128d __A
, __m128d __B
)
432 return (__m128d
)__builtin_ia32_cmpneqsd ((__v2df
)__A
, (__v2df
)__B
);
435 static __inline __m128d
436 _mm_cmpnlt_sd (__m128d __A
, __m128d __B
)
438 return (__m128d
)__builtin_ia32_cmpnltsd ((__v2df
)__A
, (__v2df
)__B
);
441 static __inline __m128d
442 _mm_cmpnle_sd (__m128d __A
, __m128d __B
)
444 return (__m128d
)__builtin_ia32_cmpnlesd ((__v2df
)__A
, (__v2df
)__B
);
447 static __inline __m128d
448 _mm_cmpngt_sd (__m128d __A
, __m128d __B
)
450 return (__m128d
) __builtin_ia32_movsd ((__v2df
) __A
,
452 __builtin_ia32_cmpnltsd ((__v2df
) __B
,
457 static __inline __m128d
458 _mm_cmpnge_sd (__m128d __A
, __m128d __B
)
460 return (__m128d
) __builtin_ia32_movsd ((__v2df
) __A
,
462 __builtin_ia32_cmpnlesd ((__v2df
) __B
,
467 static __inline __m128d
468 _mm_cmpord_sd (__m128d __A
, __m128d __B
)
470 return (__m128d
)__builtin_ia32_cmpordsd ((__v2df
)__A
, (__v2df
)__B
);
473 static __inline __m128d
474 _mm_cmpunord_sd (__m128d __A
, __m128d __B
)
476 return (__m128d
)__builtin_ia32_cmpunordsd ((__v2df
)__A
, (__v2df
)__B
);
480 _mm_comieq_sd (__m128d __A
, __m128d __B
)
482 return __builtin_ia32_comisdeq ((__v2df
)__A
, (__v2df
)__B
);
486 _mm_comilt_sd (__m128d __A
, __m128d __B
)
488 return __builtin_ia32_comisdlt ((__v2df
)__A
, (__v2df
)__B
);
492 _mm_comile_sd (__m128d __A
, __m128d __B
)
494 return __builtin_ia32_comisdle ((__v2df
)__A
, (__v2df
)__B
);
498 _mm_comigt_sd (__m128d __A
, __m128d __B
)
500 return __builtin_ia32_comisdgt ((__v2df
)__A
, (__v2df
)__B
);
504 _mm_comige_sd (__m128d __A
, __m128d __B
)
506 return __builtin_ia32_comisdge ((__v2df
)__A
, (__v2df
)__B
);
510 _mm_comineq_sd (__m128d __A
, __m128d __B
)
512 return __builtin_ia32_comisdneq ((__v2df
)__A
, (__v2df
)__B
);
516 _mm_ucomieq_sd (__m128d __A
, __m128d __B
)
518 return __builtin_ia32_ucomisdeq ((__v2df
)__A
, (__v2df
)__B
);
522 _mm_ucomilt_sd (__m128d __A
, __m128d __B
)
524 return __builtin_ia32_ucomisdlt ((__v2df
)__A
, (__v2df
)__B
);
528 _mm_ucomile_sd (__m128d __A
, __m128d __B
)
530 return __builtin_ia32_ucomisdle ((__v2df
)__A
, (__v2df
)__B
);
534 _mm_ucomigt_sd (__m128d __A
, __m128d __B
)
536 return __builtin_ia32_ucomisdgt ((__v2df
)__A
, (__v2df
)__B
);
540 _mm_ucomige_sd (__m128d __A
, __m128d __B
)
542 return __builtin_ia32_ucomisdge ((__v2df
)__A
, (__v2df
)__B
);
546 _mm_ucomineq_sd (__m128d __A
, __m128d __B
)
548 return __builtin_ia32_ucomisdneq ((__v2df
)__A
, (__v2df
)__B
);
551 /* Create a vector of Qi, where i is the element number. */
553 static __inline __m128i
554 _mm_set_epi64x (long long __q1
, long long __q0
)
556 return __extension__ (__m128i
)(__v2di
){ __q0
, __q1
};
559 static __inline __m128i
560 _mm_set_epi64 (__m64 __q1
, __m64 __q0
)
562 return _mm_set_epi64x ((long long)__q1
, (long long)__q0
);
565 static __inline __m128i
566 _mm_set_epi32 (int __q3
, int __q2
, int __q1
, int __q0
)
568 return __extension__ (__m128i
)(__v4si
){ __q0
, __q1
, __q2
, __q3
};
571 static __inline __m128i
572 _mm_set_epi16 (short __q7
, short __q6
, short __q5
, short __q4
,
573 short __q3
, short __q2
, short __q1
, short __q0
)
575 return __extension__ (__m128i
)(__v8hi
){
576 __q0
, __q1
, __q2
, __q3
, __q4
, __q5
, __q6
, __q7
};
579 static __inline __m128i
580 _mm_set_epi8 (char __q15
, char __q14
, char __q13
, char __q12
,
581 char __q11
, char __q10
, char __q09
, char __q08
,
582 char __q07
, char __q06
, char __q05
, char __q04
,
583 char __q03
, char __q02
, char __q01
, char __q00
)
585 return __extension__ (__m128i
)(__v16qi
){
586 __q00
, __q01
, __q02
, __q03
, __q04
, __q05
, __q06
, __q07
,
587 __q08
, __q09
, __q10
, __q11
, __q12
, __q13
, __q14
, __q15
591 /* Set all of the elements of the vector to A. */
593 static __inline __m128i
594 _mm_set1_epi64x (long long __A
)
596 return _mm_set_epi64x (__A
, __A
);
599 static __inline __m128i
600 _mm_set1_epi64 (__m64 __A
)
602 return _mm_set_epi64 (__A
, __A
);
605 static __inline __m128i
606 _mm_set1_epi32 (int __A
)
608 return _mm_set_epi32 (__A
, __A
, __A
, __A
);
611 static __inline __m128i
612 _mm_set1_epi16 (short __A
)
614 return _mm_set_epi16 (__A
, __A
, __A
, __A
, __A
, __A
, __A
, __A
);
617 static __inline __m128i
618 _mm_set1_epi8 (char __A
)
620 return _mm_set_epi8 (__A
, __A
, __A
, __A
, __A
, __A
, __A
, __A
,
621 __A
, __A
, __A
, __A
, __A
, __A
, __A
, __A
);
624 /* Create a vector of Qi, where i is the element number.
625 The parameter order is reversed from the _mm_set_epi* functions. */
627 static __inline __m128i
628 _mm_setr_epi64 (__m64 __q0
, __m64 __q1
)
630 return _mm_set_epi64 (__q1
, __q0
);
633 static __inline __m128i
634 _mm_setr_epi32 (int __q0
, int __q1
, int __q2
, int __q3
)
636 return _mm_set_epi32 (__q3
, __q2
, __q1
, __q0
);
639 static __inline __m128i
640 _mm_setr_epi16 (short __q0
, short __q1
, short __q2
, short __q3
,
641 short __q4
, short __q5
, short __q6
, short __q7
)
643 return _mm_set_epi16 (__q7
, __q6
, __q5
, __q4
, __q3
, __q2
, __q1
, __q0
);
646 static __inline __m128i
647 _mm_setr_epi8 (char __q00
, char __q01
, char __q02
, char __q03
,
648 char __q04
, char __q05
, char __q06
, char __q07
,
649 char __q08
, char __q09
, char __q10
, char __q11
,
650 char __q12
, char __q13
, char __q14
, char __q15
)
652 return _mm_set_epi8 (__q15
, __q14
, __q13
, __q12
, __q11
, __q10
, __q09
, __q08
,
653 __q07
, __q06
, __q05
, __q04
, __q03
, __q02
, __q01
, __q00
);
656 /* Create a vector with element 0 as *P and the rest zero. */
658 static __inline __m128i
659 _mm_load_si128 (__m128i
const *__P
)
664 static __inline __m128i
665 _mm_loadu_si128 (__m128i
const *__P
)
667 return (__m128i
) __builtin_ia32_loaddqu ((char const *)__P
);
670 static __inline __m128i
671 _mm_loadl_epi64 (__m128i
const *__P
)
673 return _mm_set_epi64 ((__m64
)0LL, *(__m64
*)__P
);
677 _mm_store_si128 (__m128i
*__P
, __m128i __B
)
683 _mm_storeu_si128 (__m128i
*__P
, __m128i __B
)
685 __builtin_ia32_storedqu ((char *)__P
, (__v16qi
)__B
);
689 _mm_storel_epi64 (__m128i
*__P
, __m128i __B
)
691 *(long long *)__P
= __builtin_ia32_vec_ext_v2di ((__v2di
)__B
, 0);
694 static __inline __m64
695 _mm_movepi64_pi64 (__m128i __B
)
697 return (__m64
) __builtin_ia32_vec_ext_v2di ((__v2di
)__B
, 0);
700 static __inline __m128i
701 _mm_movpi64_epi64 (__m64 __A
)
703 return _mm_set_epi64 ((__m64
)0LL, __A
);
706 static __inline __m128i
707 _mm_move_epi64 (__m128i __A
)
709 return _mm_set_epi64 ((__m64
)0LL, _mm_movepi64_pi64 (__A
));
712 /* Create a vector of zeros. */
713 static __inline __m128i
714 _mm_setzero_si128 (void)
716 return __extension__ (__m128i
)(__v4si
){ 0, 0, 0, 0 };
719 static __inline __m128d
720 _mm_cvtepi32_pd (__m128i __A
)
722 return (__m128d
)__builtin_ia32_cvtdq2pd ((__v4si
) __A
);
725 static __inline __m128
726 _mm_cvtepi32_ps (__m128i __A
)
728 return (__m128
)__builtin_ia32_cvtdq2ps ((__v4si
) __A
);
731 static __inline __m128i
732 _mm_cvtpd_epi32 (__m128d __A
)
734 return (__m128i
)__builtin_ia32_cvtpd2dq ((__v2df
) __A
);
737 static __inline __m64
738 _mm_cvtpd_pi32 (__m128d __A
)
740 return (__m64
)__builtin_ia32_cvtpd2pi ((__v2df
) __A
);
743 static __inline __m128
744 _mm_cvtpd_ps (__m128d __A
)
746 return (__m128
)__builtin_ia32_cvtpd2ps ((__v2df
) __A
);
749 static __inline __m128i
750 _mm_cvttpd_epi32 (__m128d __A
)
752 return (__m128i
)__builtin_ia32_cvttpd2dq ((__v2df
) __A
);
755 static __inline __m64
756 _mm_cvttpd_pi32 (__m128d __A
)
758 return (__m64
)__builtin_ia32_cvttpd2pi ((__v2df
) __A
);
761 static __inline __m128d
762 _mm_cvtpi32_pd (__m64 __A
)
764 return (__m128d
)__builtin_ia32_cvtpi2pd ((__v2si
) __A
);
767 static __inline __m128i
768 _mm_cvtps_epi32 (__m128 __A
)
770 return (__m128i
)__builtin_ia32_cvtps2dq ((__v4sf
) __A
);
773 static __inline __m128i
774 _mm_cvttps_epi32 (__m128 __A
)
776 return (__m128i
)__builtin_ia32_cvttps2dq ((__v4sf
) __A
);
779 static __inline __m128d
780 _mm_cvtps_pd (__m128 __A
)
782 return (__m128d
)__builtin_ia32_cvtps2pd ((__v4sf
) __A
);
786 _mm_cvtsd_si32 (__m128d __A
)
788 return __builtin_ia32_cvtsd2si ((__v2df
) __A
);
792 static __inline
long long
793 _mm_cvtsd_si64x (__m128d __A
)
795 return __builtin_ia32_cvtsd2si64 ((__v2df
) __A
);
800 _mm_cvttsd_si32 (__m128d __A
)
802 return __builtin_ia32_cvttsd2si ((__v2df
) __A
);
806 static __inline
long long
807 _mm_cvttsd_si64x (__m128d __A
)
809 return __builtin_ia32_cvttsd2si64 ((__v2df
) __A
);
813 static __inline __m128
814 _mm_cvtsd_ss (__m128 __A
, __m128d __B
)
816 return (__m128
)__builtin_ia32_cvtsd2ss ((__v4sf
) __A
, (__v2df
) __B
);
819 static __inline __m128d
820 _mm_cvtsi32_sd (__m128d __A
, int __B
)
822 return (__m128d
)__builtin_ia32_cvtsi2sd ((__v2df
) __A
, __B
);
826 static __inline __m128d
827 _mm_cvtsi64x_sd (__m128d __A
, long long __B
)
829 return (__m128d
)__builtin_ia32_cvtsi642sd ((__v2df
) __A
, __B
);
833 static __inline __m128d
834 _mm_cvtss_sd (__m128d __A
, __m128 __B
)
836 return (__m128d
)__builtin_ia32_cvtss2sd ((__v2df
) __A
, (__v4sf
)__B
);
839 #define _mm_shuffle_pd(__A, __B, __C) ((__m128d)__builtin_ia32_shufpd ((__v2df)__A, (__v2df)__B, (__C)))
841 static __inline __m128d
842 _mm_unpackhi_pd (__m128d __A
, __m128d __B
)
844 return (__m128d
)__builtin_ia32_unpckhpd ((__v2df
)__A
, (__v2df
)__B
);
847 static __inline __m128d
848 _mm_unpacklo_pd (__m128d __A
, __m128d __B
)
850 return (__m128d
)__builtin_ia32_unpcklpd ((__v2df
)__A
, (__v2df
)__B
);
853 static __inline __m128d
854 _mm_loadh_pd (__m128d __A
, double const *__B
)
856 return (__m128d
)__builtin_ia32_loadhpd ((__v2df
)__A
, __B
);
859 static __inline __m128d
860 _mm_loadl_pd (__m128d __A
, double const *__B
)
862 return (__m128d
)__builtin_ia32_loadlpd ((__v2df
)__A
, __B
);
866 _mm_movemask_pd (__m128d __A
)
868 return __builtin_ia32_movmskpd ((__v2df
)__A
);
871 static __inline __m128i
872 _mm_packs_epi16 (__m128i __A
, __m128i __B
)
874 return (__m128i
)__builtin_ia32_packsswb128 ((__v8hi
)__A
, (__v8hi
)__B
);
877 static __inline __m128i
878 _mm_packs_epi32 (__m128i __A
, __m128i __B
)
880 return (__m128i
)__builtin_ia32_packssdw128 ((__v4si
)__A
, (__v4si
)__B
);
883 static __inline __m128i
884 _mm_packus_epi16 (__m128i __A
, __m128i __B
)
886 return (__m128i
)__builtin_ia32_packuswb128 ((__v8hi
)__A
, (__v8hi
)__B
);
889 static __inline __m128i
890 _mm_unpackhi_epi8 (__m128i __A
, __m128i __B
)
892 return (__m128i
)__builtin_ia32_punpckhbw128 ((__v16qi
)__A
, (__v16qi
)__B
);
895 static __inline __m128i
896 _mm_unpackhi_epi16 (__m128i __A
, __m128i __B
)
898 return (__m128i
)__builtin_ia32_punpckhwd128 ((__v8hi
)__A
, (__v8hi
)__B
);
901 static __inline __m128i
902 _mm_unpackhi_epi32 (__m128i __A
, __m128i __B
)
904 return (__m128i
)__builtin_ia32_punpckhdq128 ((__v4si
)__A
, (__v4si
)__B
);
907 static __inline __m128i
908 _mm_unpackhi_epi64 (__m128i __A
, __m128i __B
)
910 return (__m128i
)__builtin_ia32_punpckhqdq128 ((__v2di
)__A
, (__v2di
)__B
);
913 static __inline __m128i
914 _mm_unpacklo_epi8 (__m128i __A
, __m128i __B
)
916 return (__m128i
)__builtin_ia32_punpcklbw128 ((__v16qi
)__A
, (__v16qi
)__B
);
919 static __inline __m128i
920 _mm_unpacklo_epi16 (__m128i __A
, __m128i __B
)
922 return (__m128i
)__builtin_ia32_punpcklwd128 ((__v8hi
)__A
, (__v8hi
)__B
);
925 static __inline __m128i
926 _mm_unpacklo_epi32 (__m128i __A
, __m128i __B
)
928 return (__m128i
)__builtin_ia32_punpckldq128 ((__v4si
)__A
, (__v4si
)__B
);
931 static __inline __m128i
932 _mm_unpacklo_epi64 (__m128i __A
, __m128i __B
)
934 return (__m128i
)__builtin_ia32_punpcklqdq128 ((__v2di
)__A
, (__v2di
)__B
);
937 static __inline __m128i
938 _mm_add_epi8 (__m128i __A
, __m128i __B
)
940 return (__m128i
)__builtin_ia32_paddb128 ((__v16qi
)__A
, (__v16qi
)__B
);
943 static __inline __m128i
944 _mm_add_epi16 (__m128i __A
, __m128i __B
)
946 return (__m128i
)__builtin_ia32_paddw128 ((__v8hi
)__A
, (__v8hi
)__B
);
949 static __inline __m128i
950 _mm_add_epi32 (__m128i __A
, __m128i __B
)
952 return (__m128i
)__builtin_ia32_paddd128 ((__v4si
)__A
, (__v4si
)__B
);
955 static __inline __m128i
956 _mm_add_epi64 (__m128i __A
, __m128i __B
)
958 return (__m128i
)__builtin_ia32_paddq128 ((__v2di
)__A
, (__v2di
)__B
);
961 static __inline __m128i
962 _mm_adds_epi8 (__m128i __A
, __m128i __B
)
964 return (__m128i
)__builtin_ia32_paddsb128 ((__v16qi
)__A
, (__v16qi
)__B
);
967 static __inline __m128i
968 _mm_adds_epi16 (__m128i __A
, __m128i __B
)
970 return (__m128i
)__builtin_ia32_paddsw128 ((__v8hi
)__A
, (__v8hi
)__B
);
973 static __inline __m128i
974 _mm_adds_epu8 (__m128i __A
, __m128i __B
)
976 return (__m128i
)__builtin_ia32_paddusb128 ((__v16qi
)__A
, (__v16qi
)__B
);
979 static __inline __m128i
980 _mm_adds_epu16 (__m128i __A
, __m128i __B
)
982 return (__m128i
)__builtin_ia32_paddusw128 ((__v8hi
)__A
, (__v8hi
)__B
);
985 static __inline __m128i
986 _mm_sub_epi8 (__m128i __A
, __m128i __B
)
988 return (__m128i
)__builtin_ia32_psubb128 ((__v16qi
)__A
, (__v16qi
)__B
);
991 static __inline __m128i
992 _mm_sub_epi16 (__m128i __A
, __m128i __B
)
994 return (__m128i
)__builtin_ia32_psubw128 ((__v8hi
)__A
, (__v8hi
)__B
);
997 static __inline __m128i
998 _mm_sub_epi32 (__m128i __A
, __m128i __B
)
1000 return (__m128i
)__builtin_ia32_psubd128 ((__v4si
)__A
, (__v4si
)__B
);
1003 static __inline __m128i
1004 _mm_sub_epi64 (__m128i __A
, __m128i __B
)
1006 return (__m128i
)__builtin_ia32_psubq128 ((__v2di
)__A
, (__v2di
)__B
);
1009 static __inline __m128i
1010 _mm_subs_epi8 (__m128i __A
, __m128i __B
)
1012 return (__m128i
)__builtin_ia32_psubsb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1015 static __inline __m128i
1016 _mm_subs_epi16 (__m128i __A
, __m128i __B
)
1018 return (__m128i
)__builtin_ia32_psubsw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1021 static __inline __m128i
1022 _mm_subs_epu8 (__m128i __A
, __m128i __B
)
1024 return (__m128i
)__builtin_ia32_psubusb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1027 static __inline __m128i
1028 _mm_subs_epu16 (__m128i __A
, __m128i __B
)
1030 return (__m128i
)__builtin_ia32_psubusw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1033 static __inline __m128i
1034 _mm_madd_epi16 (__m128i __A
, __m128i __B
)
1036 return (__m128i
)__builtin_ia32_pmaddwd128 ((__v8hi
)__A
, (__v8hi
)__B
);
1039 static __inline __m128i
1040 _mm_mulhi_epi16 (__m128i __A
, __m128i __B
)
1042 return (__m128i
)__builtin_ia32_pmulhw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1045 static __inline __m128i
1046 _mm_mullo_epi16 (__m128i __A
, __m128i __B
)
1048 return (__m128i
)__builtin_ia32_pmullw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1051 static __inline __m64
1052 _mm_mul_su32 (__m64 __A
, __m64 __B
)
1054 return (__m64
)__builtin_ia32_pmuludq ((__v2si
)__A
, (__v2si
)__B
);
1057 static __inline __m128i
1058 _mm_mul_epu32 (__m128i __A
, __m128i __B
)
1060 return (__m128i
)__builtin_ia32_pmuludq128 ((__v4si
)__A
, (__v4si
)__B
);
1063 static __inline __m128i
1064 _mm_slli_epi16 (__m128i __A
, int __B
)
1066 return (__m128i
)__builtin_ia32_psllwi128 ((__v8hi
)__A
, __B
);
1069 static __inline __m128i
1070 _mm_slli_epi32 (__m128i __A
, int __B
)
1072 return (__m128i
)__builtin_ia32_pslldi128 ((__v4si
)__A
, __B
);
1075 static __inline __m128i
1076 _mm_slli_epi64 (__m128i __A
, int __B
)
1078 return (__m128i
)__builtin_ia32_psllqi128 ((__v2di
)__A
, __B
);
1081 static __inline __m128i
1082 _mm_srai_epi16 (__m128i __A
, int __B
)
1084 return (__m128i
)__builtin_ia32_psrawi128 ((__v8hi
)__A
, __B
);
1087 static __inline __m128i
1088 _mm_srai_epi32 (__m128i __A
, int __B
)
1090 return (__m128i
)__builtin_ia32_psradi128 ((__v4si
)__A
, __B
);
1094 static __m128i
__attribute__((__always_inline__
))
1095 _mm_srli_si128 (__m128i __A
, const int __B
)
1097 return ((__m128i
)__builtin_ia32_psrldqi128 (__A
, __B
))
1100 static __m128i
__attribute__((__always_inline__
))
1101 _mm_srli_si128 (__m128i __A
, const int __B
)
1103 return ((__m128i
)__builtin_ia32_pslldqi128 (__A
, __B
))
1106 #define _mm_srli_si128(__A, __B) \
1107 ((__m128i)__builtin_ia32_psrldqi128 (__A, (__B) * 8))
1108 #define _mm_slli_si128(__A, __B) \
1109 ((__m128i)__builtin_ia32_pslldqi128 (__A, (__B) * 8))
1112 static __inline __m128i
1113 _mm_srli_epi16 (__m128i __A
, int __B
)
1115 return (__m128i
)__builtin_ia32_psrlwi128 ((__v8hi
)__A
, __B
);
1118 static __inline __m128i
1119 _mm_srli_epi32 (__m128i __A
, int __B
)
1121 return (__m128i
)__builtin_ia32_psrldi128 ((__v4si
)__A
, __B
);
1124 static __inline __m128i
1125 _mm_srli_epi64 (__m128i __A
, int __B
)
1127 return (__m128i
)__builtin_ia32_psrlqi128 ((__v2di
)__A
, __B
);
1130 static __inline __m128i
1131 _mm_sll_epi16 (__m128i __A
, __m128i __B
)
1133 return _mm_slli_epi16 (__A
, _mm_cvtsi128_si32 (__B
));
1136 static __inline __m128i
1137 _mm_sll_epi32 (__m128i __A
, __m128i __B
)
1139 return _mm_slli_epi32 (__A
, _mm_cvtsi128_si32 (__B
));
1142 static __inline __m128i
1143 _mm_sll_epi64 (__m128i __A
, __m128i __B
)
1145 return _mm_slli_epi64 (__A
, _mm_cvtsi128_si32 (__B
));
1148 static __inline __m128i
1149 _mm_sra_epi16 (__m128i __A
, __m128i __B
)
1151 return _mm_srai_epi16 (__A
, _mm_cvtsi128_si32 (__B
));
1154 static __inline __m128i
1155 _mm_sra_epi32 (__m128i __A
, __m128i __B
)
1157 return _mm_srai_epi32 (__A
, _mm_cvtsi128_si32 (__B
));
1160 static __inline __m128i
1161 _mm_srl_epi16 (__m128i __A
, __m128i __B
)
1163 return _mm_srli_epi16 (__A
, _mm_cvtsi128_si32 (__B
));
1166 static __inline __m128i
1167 _mm_srl_epi32 (__m128i __A
, __m128i __B
)
1169 return _mm_srli_epi32 (__A
, _mm_cvtsi128_si32 (__B
));
1172 static __inline __m128i
1173 _mm_srl_epi64 (__m128i __A
, __m128i __B
)
1175 return _mm_srli_epi64 (__A
, _mm_cvtsi128_si32 (__B
));
1178 static __inline __m128i
1179 _mm_and_si128 (__m128i __A
, __m128i __B
)
1181 return (__m128i
)__builtin_ia32_pand128 ((__v2di
)__A
, (__v2di
)__B
);
1184 static __inline __m128i
1185 _mm_andnot_si128 (__m128i __A
, __m128i __B
)
1187 return (__m128i
)__builtin_ia32_pandn128 ((__v2di
)__A
, (__v2di
)__B
);
1190 static __inline __m128i
1191 _mm_or_si128 (__m128i __A
, __m128i __B
)
1193 return (__m128i
)__builtin_ia32_por128 ((__v2di
)__A
, (__v2di
)__B
);
1196 static __inline __m128i
1197 _mm_xor_si128 (__m128i __A
, __m128i __B
)
1199 return (__m128i
)__builtin_ia32_pxor128 ((__v2di
)__A
, (__v2di
)__B
);
1202 static __inline __m128i
1203 _mm_cmpeq_epi8 (__m128i __A
, __m128i __B
)
1205 return (__m128i
)__builtin_ia32_pcmpeqb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1208 static __inline __m128i
1209 _mm_cmpeq_epi16 (__m128i __A
, __m128i __B
)
1211 return (__m128i
)__builtin_ia32_pcmpeqw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1214 static __inline __m128i
1215 _mm_cmpeq_epi32 (__m128i __A
, __m128i __B
)
1217 return (__m128i
)__builtin_ia32_pcmpeqd128 ((__v4si
)__A
, (__v4si
)__B
);
1220 static __inline __m128i
1221 _mm_cmplt_epi8 (__m128i __A
, __m128i __B
)
1223 return (__m128i
)__builtin_ia32_pcmpgtb128 ((__v16qi
)__B
, (__v16qi
)__A
);
1226 static __inline __m128i
1227 _mm_cmplt_epi16 (__m128i __A
, __m128i __B
)
1229 return (__m128i
)__builtin_ia32_pcmpgtw128 ((__v8hi
)__B
, (__v8hi
)__A
);
1232 static __inline __m128i
1233 _mm_cmplt_epi32 (__m128i __A
, __m128i __B
)
1235 return (__m128i
)__builtin_ia32_pcmpgtd128 ((__v4si
)__B
, (__v4si
)__A
);
1238 static __inline __m128i
1239 _mm_cmpgt_epi8 (__m128i __A
, __m128i __B
)
1241 return (__m128i
)__builtin_ia32_pcmpgtb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1244 static __inline __m128i
1245 _mm_cmpgt_epi16 (__m128i __A
, __m128i __B
)
1247 return (__m128i
)__builtin_ia32_pcmpgtw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1250 static __inline __m128i
1251 _mm_cmpgt_epi32 (__m128i __A
, __m128i __B
)
1253 return (__m128i
)__builtin_ia32_pcmpgtd128 ((__v4si
)__A
, (__v4si
)__B
);
1257 static __inline
int __attribute__((__always_inline__
))
1258 _mm_extract_epi16 (__m128i
const __A
, int const __N
)
1260 return __builtin_ia32_vec_ext_v8hi ((__v8hi
)__A
, __N
);
1263 static __inline __m128i
__attribute__((__always_inline__
))
1264 _mm_insert_epi16 (__m128i
const __A
, int const __D
, int const __N
)
1266 return (__m128i
) __builtin_ia32_vec_set_v8hi ((__v8hi
)__A
, __D
, __N
);
1269 #define _mm_extract_epi16(A, N) \
1270 ((int) __builtin_ia32_vec_ext_v8hi ((__v8hi)(A), (N)))
1271 #define _mm_insert_epi16(A, D, N) \
1272 ((__m128i) __builtin_ia32_vec_set_v8hi ((__v8hi)(A), (D), (N)))
1275 static __inline __m128i
1276 _mm_max_epi16 (__m128i __A
, __m128i __B
)
1278 return (__m128i
)__builtin_ia32_pmaxsw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1281 static __inline __m128i
1282 _mm_max_epu8 (__m128i __A
, __m128i __B
)
1284 return (__m128i
)__builtin_ia32_pmaxub128 ((__v16qi
)__A
, (__v16qi
)__B
);
1287 static __inline __m128i
1288 _mm_min_epi16 (__m128i __A
, __m128i __B
)
1290 return (__m128i
)__builtin_ia32_pminsw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1293 static __inline __m128i
1294 _mm_min_epu8 (__m128i __A
, __m128i __B
)
1296 return (__m128i
)__builtin_ia32_pminub128 ((__v16qi
)__A
, (__v16qi
)__B
);
1300 _mm_movemask_epi8 (__m128i __A
)
1302 return __builtin_ia32_pmovmskb128 ((__v16qi
)__A
);
1305 static __inline __m128i
1306 _mm_mulhi_epu16 (__m128i __A
, __m128i __B
)
1308 return (__m128i
)__builtin_ia32_pmulhuw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1311 #define _mm_shufflehi_epi16(__A, __B) ((__m128i)__builtin_ia32_pshufhw ((__v8hi)__A, __B))
1312 #define _mm_shufflelo_epi16(__A, __B) ((__m128i)__builtin_ia32_pshuflw ((__v8hi)__A, __B))
1313 #define _mm_shuffle_epi32(__A, __B) ((__m128i)__builtin_ia32_pshufd ((__v4si)__A, __B))
1315 static __inline
void
1316 _mm_maskmoveu_si128 (__m128i __A
, __m128i __B
, char *__C
)
1318 __builtin_ia32_maskmovdqu ((__v16qi
)__A
, (__v16qi
)__B
, __C
);
1321 static __inline __m128i
1322 _mm_avg_epu8 (__m128i __A
, __m128i __B
)
1324 return (__m128i
)__builtin_ia32_pavgb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1327 static __inline __m128i
1328 _mm_avg_epu16 (__m128i __A
, __m128i __B
)
1330 return (__m128i
)__builtin_ia32_pavgw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1333 static __inline __m128i
1334 _mm_sad_epu8 (__m128i __A
, __m128i __B
)
1336 return (__m128i
)__builtin_ia32_psadbw128 ((__v16qi
)__A
, (__v16qi
)__B
);
1339 static __inline
void
1340 _mm_stream_si32 (int *__A
, int __B
)
1342 __builtin_ia32_movnti (__A
, __B
);
1345 static __inline
void
1346 _mm_stream_si128 (__m128i
*__A
, __m128i __B
)
1348 __builtin_ia32_movntdq ((__v2di
*)__A
, (__v2di
)__B
);
1351 static __inline
void
1352 _mm_stream_pd (double *__A
, __m128d __B
)
1354 __builtin_ia32_movntpd (__A
, (__v2df
)__B
);
1357 static __inline
void
1358 _mm_clflush (void const *__A
)
1360 __builtin_ia32_clflush (__A
);
1363 static __inline
void
1366 __builtin_ia32_lfence ();
1369 static __inline
void
1372 __builtin_ia32_mfence ();
1375 static __inline __m128i
1376 _mm_cvtsi32_si128 (int __A
)
1378 return _mm_set_epi32 (0, 0, 0, __A
);
1382 static __inline __m128i
1383 _mm_cvtsi64x_si128 (long long __A
)
1385 return _mm_set_epi64x (0, __A
);
1389 /* Casts between various SP, DP, INT vector types. Note that these do no
1390 conversion of values, they just change the type. */
1391 static __inline __m128
1392 _mm_castpd_ps(__m128d __A
)
1394 return (__m128
) __A
;
1397 static __inline __m128i
1398 _mm_castpd_si128(__m128d __A
)
1400 return (__m128i
) __A
;
1403 static __inline __m128d
1404 _mm_castps_pd(__m128 __A
)
1406 return (__m128d
) __A
;
1409 static __inline __m128i
1410 _mm_castps_si128(__m128 __A
)
1412 return (__m128i
) __A
;
1415 static __inline __m128
1416 _mm_castsi128_ps(__m128i __A
)
1418 return (__m128
) __A
;
1421 static __inline __m128d
1422 _mm_castsi128_pd(__m128i __A
)
1424 return (__m128d
) __A
;
1427 #endif /* __SSE2__ */
1429 #endif /* _EMMINTRIN_H_INCLUDED */
