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 (__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 (__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 (__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 (__m128d
){ __W
, __X
};
84 /* Create a vector of zeros. */
85 static __inline __m128d
88 return (__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 (__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 (__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 (__m128i
)(__v8hi
){ __q0
, __q1
, __q2
, __q3
, __q4
, __q5
, __q6
, __q7
};
578 static __inline __m128i
579 _mm_set_epi8 (char __q15
, char __q14
, char __q13
, char __q12
,
580 char __q11
, char __q10
, char __q09
, char __q08
,
581 char __q07
, char __q06
, char __q05
, char __q04
,
582 char __q03
, char __q02
, char __q01
, char __q00
)
584 return (__m128i
)(__v16qi
){
585 __q00
, __q01
, __q02
, __q03
, __q04
, __q05
, __q06
, __q07
,
586 __q08
, __q09
, __q10
, __q11
, __q12
, __q13
, __q14
, __q15
590 /* Set all of the elements of the vector to A. */
592 static __inline __m128i
593 _mm_set1_epi64x (long long __A
)
595 return _mm_set_epi64x (__A
, __A
);
598 static __inline __m128i
599 _mm_set1_epi64 (__m64 __A
)
601 return _mm_set_epi64 (__A
, __A
);
604 static __inline __m128i
605 _mm_set1_epi32 (int __A
)
607 return _mm_set_epi32 (__A
, __A
, __A
, __A
);
610 static __inline __m128i
611 _mm_set1_epi16 (short __A
)
613 return _mm_set_epi16 (__A
, __A
, __A
, __A
, __A
, __A
, __A
, __A
);
616 static __inline __m128i
617 _mm_set1_epi8 (char __A
)
619 return _mm_set_epi8 (__A
, __A
, __A
, __A
, __A
, __A
, __A
, __A
,
620 __A
, __A
, __A
, __A
, __A
, __A
, __A
, __A
);
623 /* Create a vector of Qi, where i is the element number.
624 The parameter order is reversed from the _mm_set_epi* functions. */
626 static __inline __m128i
627 _mm_setr_epi64 (__m64 __q0
, __m64 __q1
)
629 return _mm_set_epi64 (__q1
, __q0
);
632 static __inline __m128i
633 _mm_setr_epi32 (int __q0
, int __q1
, int __q2
, int __q3
)
635 return _mm_set_epi32 (__q3
, __q2
, __q1
, __q0
);
638 static __inline __m128i
639 _mm_setr_epi16 (short __q0
, short __q1
, short __q2
, short __q3
,
640 short __q4
, short __q5
, short __q6
, short __q7
)
642 return _mm_set_epi16 (__q7
, __q6
, __q5
, __q4
, __q3
, __q2
, __q1
, __q0
);
645 static __inline __m128i
646 _mm_setr_epi8 (char __q00
, char __q01
, char __q02
, char __q03
,
647 char __q04
, char __q05
, char __q06
, char __q07
,
648 char __q08
, char __q09
, char __q10
, char __q11
,
649 char __q12
, char __q13
, char __q14
, char __q15
)
651 return _mm_set_epi8 (__q15
, __q14
, __q13
, __q12
, __q11
, __q10
, __q09
, __q08
,
652 __q07
, __q06
, __q05
, __q04
, __q03
, __q02
, __q01
, __q00
);
655 /* Create a vector with element 0 as *P and the rest zero. */
657 static __inline __m128i
658 _mm_load_si128 (__m128i
const *__P
)
663 static __inline __m128i
664 _mm_loadu_si128 (__m128i
const *__P
)
666 return (__m128i
) __builtin_ia32_loaddqu ((char const *)__P
);
669 static __inline __m128i
670 _mm_loadl_epi64 (__m128i
const *__P
)
672 return _mm_set_epi64 ((__m64
)0LL, *(__m64
*)__P
);
676 _mm_store_si128 (__m128i
*__P
, __m128i __B
)
682 _mm_storeu_si128 (__m128i
*__P
, __m128i __B
)
684 __builtin_ia32_storedqu ((char *)__P
, (__v16qi
)__B
);
688 _mm_storel_epi64 (__m128i
*__P
, __m128i __B
)
690 *(long long *)__P
= __builtin_ia32_vec_ext_v2di ((__v2di
)__B
, 0);
693 static __inline __m64
694 _mm_movepi64_pi64 (__m128i __B
)
696 return (__m64
) __builtin_ia32_vec_ext_v2di ((__v2di
)__B
, 0);
699 static __inline __m128i
700 _mm_movpi64_epi64 (__m64 __A
)
702 return _mm_set_epi64 ((__m64
)0LL, __A
);
705 static __inline __m128i
706 _mm_move_epi64 (__m128i __A
)
708 return _mm_set_epi64 ((__m64
)0LL, _mm_movepi64_pi64 (__A
));
711 /* Create a vector of zeros. */
712 static __inline __m128i
713 _mm_setzero_si128 (void)
715 return (__m128i
)(__v4si
){ 0, 0, 0, 0 };
718 static __inline __m128d
719 _mm_cvtepi32_pd (__m128i __A
)
721 return (__m128d
)__builtin_ia32_cvtdq2pd ((__v4si
) __A
);
724 static __inline __m128
725 _mm_cvtepi32_ps (__m128i __A
)
727 return (__m128
)__builtin_ia32_cvtdq2ps ((__v4si
) __A
);
730 static __inline __m128i
731 _mm_cvtpd_epi32 (__m128d __A
)
733 return (__m128i
)__builtin_ia32_cvtpd2dq ((__v2df
) __A
);
736 static __inline __m64
737 _mm_cvtpd_pi32 (__m128d __A
)
739 return (__m64
)__builtin_ia32_cvtpd2pi ((__v2df
) __A
);
742 static __inline __m128
743 _mm_cvtpd_ps (__m128d __A
)
745 return (__m128
)__builtin_ia32_cvtpd2ps ((__v2df
) __A
);
748 static __inline __m128i
749 _mm_cvttpd_epi32 (__m128d __A
)
751 return (__m128i
)__builtin_ia32_cvttpd2dq ((__v2df
) __A
);
754 static __inline __m64
755 _mm_cvttpd_pi32 (__m128d __A
)
757 return (__m64
)__builtin_ia32_cvttpd2pi ((__v2df
) __A
);
760 static __inline __m128d
761 _mm_cvtpi32_pd (__m64 __A
)
763 return (__m128d
)__builtin_ia32_cvtpi2pd ((__v2si
) __A
);
766 static __inline __m128i
767 _mm_cvtps_epi32 (__m128 __A
)
769 return (__m128i
)__builtin_ia32_cvtps2dq ((__v4sf
) __A
);
772 static __inline __m128i
773 _mm_cvttps_epi32 (__m128 __A
)
775 return (__m128i
)__builtin_ia32_cvttps2dq ((__v4sf
) __A
);
778 static __inline __m128d
779 _mm_cvtps_pd (__m128 __A
)
781 return (__m128d
)__builtin_ia32_cvtps2pd ((__v4sf
) __A
);
785 _mm_cvtsd_si32 (__m128d __A
)
787 return __builtin_ia32_cvtsd2si ((__v2df
) __A
);
791 static __inline
long long
792 _mm_cvtsd_si64x (__m128d __A
)
794 return __builtin_ia32_cvtsd2si64 ((__v2df
) __A
);
799 _mm_cvttsd_si32 (__m128d __A
)
801 return __builtin_ia32_cvttsd2si ((__v2df
) __A
);
805 static __inline
long long
806 _mm_cvttsd_si64x (__m128d __A
)
808 return __builtin_ia32_cvttsd2si64 ((__v2df
) __A
);
812 static __inline __m128
813 _mm_cvtsd_ss (__m128 __A
, __m128d __B
)
815 return (__m128
)__builtin_ia32_cvtsd2ss ((__v4sf
) __A
, (__v2df
) __B
);
818 static __inline __m128d
819 _mm_cvtsi32_sd (__m128d __A
, int __B
)
821 return (__m128d
)__builtin_ia32_cvtsi2sd ((__v2df
) __A
, __B
);
825 static __inline __m128d
826 _mm_cvtsi64x_sd (__m128d __A
, long long __B
)
828 return (__m128d
)__builtin_ia32_cvtsi642sd ((__v2df
) __A
, __B
);
832 static __inline __m128d
833 _mm_cvtss_sd (__m128d __A
, __m128 __B
)
835 return (__m128d
)__builtin_ia32_cvtss2sd ((__v2df
) __A
, (__v4sf
)__B
);
838 #define _mm_shuffle_pd(__A, __B, __C) ((__m128d)__builtin_ia32_shufpd ((__v2df)__A, (__v2df)__B, (__C)))
840 static __inline __m128d
841 _mm_unpackhi_pd (__m128d __A
, __m128d __B
)
843 return (__m128d
)__builtin_ia32_unpckhpd ((__v2df
)__A
, (__v2df
)__B
);
846 static __inline __m128d
847 _mm_unpacklo_pd (__m128d __A
, __m128d __B
)
849 return (__m128d
)__builtin_ia32_unpcklpd ((__v2df
)__A
, (__v2df
)__B
);
852 static __inline __m128d
853 _mm_loadh_pd (__m128d __A
, double const *__B
)
855 return (__m128d
)__builtin_ia32_loadhpd ((__v2df
)__A
, __B
);
858 static __inline __m128d
859 _mm_loadl_pd (__m128d __A
, double const *__B
)
861 return (__m128d
)__builtin_ia32_loadlpd ((__v2df
)__A
, __B
);
865 _mm_movemask_pd (__m128d __A
)
867 return __builtin_ia32_movmskpd ((__v2df
)__A
);
870 static __inline __m128i
871 _mm_packs_epi16 (__m128i __A
, __m128i __B
)
873 return (__m128i
)__builtin_ia32_packsswb128 ((__v8hi
)__A
, (__v8hi
)__B
);
876 static __inline __m128i
877 _mm_packs_epi32 (__m128i __A
, __m128i __B
)
879 return (__m128i
)__builtin_ia32_packssdw128 ((__v4si
)__A
, (__v4si
)__B
);
882 static __inline __m128i
883 _mm_packus_epi16 (__m128i __A
, __m128i __B
)
885 return (__m128i
)__builtin_ia32_packuswb128 ((__v8hi
)__A
, (__v8hi
)__B
);
888 static __inline __m128i
889 _mm_unpackhi_epi8 (__m128i __A
, __m128i __B
)
891 return (__m128i
)__builtin_ia32_punpckhbw128 ((__v16qi
)__A
, (__v16qi
)__B
);
894 static __inline __m128i
895 _mm_unpackhi_epi16 (__m128i __A
, __m128i __B
)
897 return (__m128i
)__builtin_ia32_punpckhwd128 ((__v8hi
)__A
, (__v8hi
)__B
);
900 static __inline __m128i
901 _mm_unpackhi_epi32 (__m128i __A
, __m128i __B
)
903 return (__m128i
)__builtin_ia32_punpckhdq128 ((__v4si
)__A
, (__v4si
)__B
);
906 static __inline __m128i
907 _mm_unpackhi_epi64 (__m128i __A
, __m128i __B
)
909 return (__m128i
)__builtin_ia32_punpckhqdq128 ((__v2di
)__A
, (__v2di
)__B
);
912 static __inline __m128i
913 _mm_unpacklo_epi8 (__m128i __A
, __m128i __B
)
915 return (__m128i
)__builtin_ia32_punpcklbw128 ((__v16qi
)__A
, (__v16qi
)__B
);
918 static __inline __m128i
919 _mm_unpacklo_epi16 (__m128i __A
, __m128i __B
)
921 return (__m128i
)__builtin_ia32_punpcklwd128 ((__v8hi
)__A
, (__v8hi
)__B
);
924 static __inline __m128i
925 _mm_unpacklo_epi32 (__m128i __A
, __m128i __B
)
927 return (__m128i
)__builtin_ia32_punpckldq128 ((__v4si
)__A
, (__v4si
)__B
);
930 static __inline __m128i
931 _mm_unpacklo_epi64 (__m128i __A
, __m128i __B
)
933 return (__m128i
)__builtin_ia32_punpcklqdq128 ((__v2di
)__A
, (__v2di
)__B
);
936 static __inline __m128i
937 _mm_add_epi8 (__m128i __A
, __m128i __B
)
939 return (__m128i
)__builtin_ia32_paddb128 ((__v16qi
)__A
, (__v16qi
)__B
);
942 static __inline __m128i
943 _mm_add_epi16 (__m128i __A
, __m128i __B
)
945 return (__m128i
)__builtin_ia32_paddw128 ((__v8hi
)__A
, (__v8hi
)__B
);
948 static __inline __m128i
949 _mm_add_epi32 (__m128i __A
, __m128i __B
)
951 return (__m128i
)__builtin_ia32_paddd128 ((__v4si
)__A
, (__v4si
)__B
);
954 static __inline __m128i
955 _mm_add_epi64 (__m128i __A
, __m128i __B
)
957 return (__m128i
)__builtin_ia32_paddq128 ((__v2di
)__A
, (__v2di
)__B
);
960 static __inline __m128i
961 _mm_adds_epi8 (__m128i __A
, __m128i __B
)
963 return (__m128i
)__builtin_ia32_paddsb128 ((__v16qi
)__A
, (__v16qi
)__B
);
966 static __inline __m128i
967 _mm_adds_epi16 (__m128i __A
, __m128i __B
)
969 return (__m128i
)__builtin_ia32_paddsw128 ((__v8hi
)__A
, (__v8hi
)__B
);
972 static __inline __m128i
973 _mm_adds_epu8 (__m128i __A
, __m128i __B
)
975 return (__m128i
)__builtin_ia32_paddusb128 ((__v16qi
)__A
, (__v16qi
)__B
);
978 static __inline __m128i
979 _mm_adds_epu16 (__m128i __A
, __m128i __B
)
981 return (__m128i
)__builtin_ia32_paddusw128 ((__v8hi
)__A
, (__v8hi
)__B
);
984 static __inline __m128i
985 _mm_sub_epi8 (__m128i __A
, __m128i __B
)
987 return (__m128i
)__builtin_ia32_psubb128 ((__v16qi
)__A
, (__v16qi
)__B
);
990 static __inline __m128i
991 _mm_sub_epi16 (__m128i __A
, __m128i __B
)
993 return (__m128i
)__builtin_ia32_psubw128 ((__v8hi
)__A
, (__v8hi
)__B
);
996 static __inline __m128i
997 _mm_sub_epi32 (__m128i __A
, __m128i __B
)
999 return (__m128i
)__builtin_ia32_psubd128 ((__v4si
)__A
, (__v4si
)__B
);
1002 static __inline __m128i
1003 _mm_sub_epi64 (__m128i __A
, __m128i __B
)
1005 return (__m128i
)__builtin_ia32_psubq128 ((__v2di
)__A
, (__v2di
)__B
);
1008 static __inline __m128i
1009 _mm_subs_epi8 (__m128i __A
, __m128i __B
)
1011 return (__m128i
)__builtin_ia32_psubsb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1014 static __inline __m128i
1015 _mm_subs_epi16 (__m128i __A
, __m128i __B
)
1017 return (__m128i
)__builtin_ia32_psubsw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1020 static __inline __m128i
1021 _mm_subs_epu8 (__m128i __A
, __m128i __B
)
1023 return (__m128i
)__builtin_ia32_psubusb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1026 static __inline __m128i
1027 _mm_subs_epu16 (__m128i __A
, __m128i __B
)
1029 return (__m128i
)__builtin_ia32_psubusw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1032 static __inline __m128i
1033 _mm_madd_epi16 (__m128i __A
, __m128i __B
)
1035 return (__m128i
)__builtin_ia32_pmaddwd128 ((__v8hi
)__A
, (__v8hi
)__B
);
1038 static __inline __m128i
1039 _mm_mulhi_epi16 (__m128i __A
, __m128i __B
)
1041 return (__m128i
)__builtin_ia32_pmulhw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1044 static __inline __m128i
1045 _mm_mullo_epi16 (__m128i __A
, __m128i __B
)
1047 return (__m128i
)__builtin_ia32_pmullw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1050 static __inline __m64
1051 _mm_mul_su32 (__m64 __A
, __m64 __B
)
1053 return (__m64
)__builtin_ia32_pmuludq ((__v2si
)__A
, (__v2si
)__B
);
1056 static __inline __m128i
1057 _mm_mul_epu32 (__m128i __A
, __m128i __B
)
1059 return (__m128i
)__builtin_ia32_pmuludq128 ((__v4si
)__A
, (__v4si
)__B
);
1062 static __inline __m128i
1063 _mm_slli_epi16 (__m128i __A
, int __B
)
1065 return (__m128i
)__builtin_ia32_psllwi128 ((__v8hi
)__A
, __B
);
1068 static __inline __m128i
1069 _mm_slli_epi32 (__m128i __A
, int __B
)
1071 return (__m128i
)__builtin_ia32_pslldi128 ((__v4si
)__A
, __B
);
1074 static __inline __m128i
1075 _mm_slli_epi64 (__m128i __A
, int __B
)
1077 return (__m128i
)__builtin_ia32_psllqi128 ((__v2di
)__A
, __B
);
1080 static __inline __m128i
1081 _mm_srai_epi16 (__m128i __A
, int __B
)
1083 return (__m128i
)__builtin_ia32_psrawi128 ((__v8hi
)__A
, __B
);
1086 static __inline __m128i
1087 _mm_srai_epi32 (__m128i __A
, int __B
)
1089 return (__m128i
)__builtin_ia32_psradi128 ((__v4si
)__A
, __B
);
1093 static __m128i
__attribute__((__always_inline__
))
1094 _mm_srli_si128 (__m128i __A
, const int __B
)
1096 return ((__m128i
)__builtin_ia32_psrldqi128 (__A
, __B
))
1099 static __m128i
__attribute__((__always_inline__
))
1100 _mm_srli_si128 (__m128i __A
, const int __B
)
1102 return ((__m128i
)__builtin_ia32_pslldqi128 (__A
, __B
))
1105 #define _mm_srli_si128(__A, __B) \
1106 ((__m128i)__builtin_ia32_psrldqi128 (__A, (__B) * 8))
1107 #define _mm_slli_si128(__A, __B) \
1108 ((__m128i)__builtin_ia32_pslldqi128 (__A, (__B) * 8))
1111 static __inline __m128i
1112 _mm_srli_epi16 (__m128i __A
, int __B
)
1114 return (__m128i
)__builtin_ia32_psrlwi128 ((__v8hi
)__A
, __B
);
1117 static __inline __m128i
1118 _mm_srli_epi32 (__m128i __A
, int __B
)
1120 return (__m128i
)__builtin_ia32_psrldi128 ((__v4si
)__A
, __B
);
1123 static __inline __m128i
1124 _mm_srli_epi64 (__m128i __A
, int __B
)
1126 return (__m128i
)__builtin_ia32_psrlqi128 ((__v2di
)__A
, __B
);
1129 static __inline __m128i
1130 _mm_sll_epi16 (__m128i __A
, __m128i __B
)
1132 return _mm_slli_epi16 (__A
, _mm_cvtsi128_si32 (__B
));
1135 static __inline __m128i
1136 _mm_sll_epi32 (__m128i __A
, __m128i __B
)
1138 return _mm_slli_epi32 (__A
, _mm_cvtsi128_si32 (__B
));
1141 static __inline __m128i
1142 _mm_sll_epi64 (__m128i __A
, __m128i __B
)
1144 return _mm_slli_epi64 (__A
, _mm_cvtsi128_si32 (__B
));
1147 static __inline __m128i
1148 _mm_sra_epi16 (__m128i __A
, __m128i __B
)
1150 return _mm_srai_epi16 (__A
, _mm_cvtsi128_si32 (__B
));
1153 static __inline __m128i
1154 _mm_sra_epi32 (__m128i __A
, __m128i __B
)
1156 return _mm_srai_epi32 (__A
, _mm_cvtsi128_si32 (__B
));
1159 static __inline __m128i
1160 _mm_srl_epi16 (__m128i __A
, __m128i __B
)
1162 return _mm_srli_epi16 (__A
, _mm_cvtsi128_si32 (__B
));
1165 static __inline __m128i
1166 _mm_srl_epi32 (__m128i __A
, __m128i __B
)
1168 return _mm_srli_epi32 (__A
, _mm_cvtsi128_si32 (__B
));
1171 static __inline __m128i
1172 _mm_srl_epi64 (__m128i __A
, __m128i __B
)
1174 return _mm_srli_epi64 (__A
, _mm_cvtsi128_si32 (__B
));
1177 static __inline __m128i
1178 _mm_and_si128 (__m128i __A
, __m128i __B
)
1180 return (__m128i
)__builtin_ia32_pand128 ((__v2di
)__A
, (__v2di
)__B
);
1183 static __inline __m128i
1184 _mm_andnot_si128 (__m128i __A
, __m128i __B
)
1186 return (__m128i
)__builtin_ia32_pandn128 ((__v2di
)__A
, (__v2di
)__B
);
1189 static __inline __m128i
1190 _mm_or_si128 (__m128i __A
, __m128i __B
)
1192 return (__m128i
)__builtin_ia32_por128 ((__v2di
)__A
, (__v2di
)__B
);
1195 static __inline __m128i
1196 _mm_xor_si128 (__m128i __A
, __m128i __B
)
1198 return (__m128i
)__builtin_ia32_pxor128 ((__v2di
)__A
, (__v2di
)__B
);
1201 static __inline __m128i
1202 _mm_cmpeq_epi8 (__m128i __A
, __m128i __B
)
1204 return (__m128i
)__builtin_ia32_pcmpeqb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1207 static __inline __m128i
1208 _mm_cmpeq_epi16 (__m128i __A
, __m128i __B
)
1210 return (__m128i
)__builtin_ia32_pcmpeqw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1213 static __inline __m128i
1214 _mm_cmpeq_epi32 (__m128i __A
, __m128i __B
)
1216 return (__m128i
)__builtin_ia32_pcmpeqd128 ((__v4si
)__A
, (__v4si
)__B
);
1219 static __inline __m128i
1220 _mm_cmplt_epi8 (__m128i __A
, __m128i __B
)
1222 return (__m128i
)__builtin_ia32_pcmpgtb128 ((__v16qi
)__B
, (__v16qi
)__A
);
1225 static __inline __m128i
1226 _mm_cmplt_epi16 (__m128i __A
, __m128i __B
)
1228 return (__m128i
)__builtin_ia32_pcmpgtw128 ((__v8hi
)__B
, (__v8hi
)__A
);
1231 static __inline __m128i
1232 _mm_cmplt_epi32 (__m128i __A
, __m128i __B
)
1234 return (__m128i
)__builtin_ia32_pcmpgtd128 ((__v4si
)__B
, (__v4si
)__A
);
1237 static __inline __m128i
1238 _mm_cmpgt_epi8 (__m128i __A
, __m128i __B
)
1240 return (__m128i
)__builtin_ia32_pcmpgtb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1243 static __inline __m128i
1244 _mm_cmpgt_epi16 (__m128i __A
, __m128i __B
)
1246 return (__m128i
)__builtin_ia32_pcmpgtw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1249 static __inline __m128i
1250 _mm_cmpgt_epi32 (__m128i __A
, __m128i __B
)
1252 return (__m128i
)__builtin_ia32_pcmpgtd128 ((__v4si
)__A
, (__v4si
)__B
);
1256 static __inline
int __attribute__((__always_inline__
))
1257 _mm_extract_epi16 (__m128i
const __A
, int const __N
)
1259 return __builtin_ia32_vec_ext_v8hi ((__v8hi
)__A
, __N
);
1262 static __inline __m128i
__attribute__((__always_inline__
))
1263 _mm_insert_epi16 (__m128i
const __A
, int const __D
, int const __N
)
1265 return (__m128i
) __builtin_ia32_vec_set_v8hi ((__v8hi
)__A
, __D
, __N
);
1268 #define _mm_extract_epi16(A, N) \
1269 ((int) __builtin_ia32_vec_ext_v8hi ((__v8hi)(A), (N))
1270 #define _mm_insert_epi16(A, D, N) \
1271 ((__m128i) __builtin_ia32_vec_set_v8hi ((__v8hi)(A), (D), (N)))
1274 static __inline __m128i
1275 _mm_max_epi16 (__m128i __A
, __m128i __B
)
1277 return (__m128i
)__builtin_ia32_pmaxsw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1280 static __inline __m128i
1281 _mm_max_epu8 (__m128i __A
, __m128i __B
)
1283 return (__m128i
)__builtin_ia32_pmaxub128 ((__v16qi
)__A
, (__v16qi
)__B
);
1286 static __inline __m128i
1287 _mm_min_epi16 (__m128i __A
, __m128i __B
)
1289 return (__m128i
)__builtin_ia32_pminsw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1292 static __inline __m128i
1293 _mm_min_epu8 (__m128i __A
, __m128i __B
)
1295 return (__m128i
)__builtin_ia32_pminub128 ((__v16qi
)__A
, (__v16qi
)__B
);
1299 _mm_movemask_epi8 (__m128i __A
)
1301 return __builtin_ia32_pmovmskb128 ((__v16qi
)__A
);
1304 static __inline __m128i
1305 _mm_mulhi_epu16 (__m128i __A
, __m128i __B
)
1307 return (__m128i
)__builtin_ia32_pmulhuw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1310 #define _mm_shufflehi_epi16(__A, __B) ((__m128i)__builtin_ia32_pshufhw ((__v8hi)__A, __B))
1311 #define _mm_shufflelo_epi16(__A, __B) ((__m128i)__builtin_ia32_pshuflw ((__v8hi)__A, __B))
1312 #define _mm_shuffle_epi32(__A, __B) ((__m128i)__builtin_ia32_pshufd ((__v4si)__A, __B))
1314 static __inline
void
1315 _mm_maskmoveu_si128 (__m128i __A
, __m128i __B
, char *__C
)
1317 __builtin_ia32_maskmovdqu ((__v16qi
)__A
, (__v16qi
)__B
, __C
);
1320 static __inline __m128i
1321 _mm_avg_epu8 (__m128i __A
, __m128i __B
)
1323 return (__m128i
)__builtin_ia32_pavgb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1326 static __inline __m128i
1327 _mm_avg_epu16 (__m128i __A
, __m128i __B
)
1329 return (__m128i
)__builtin_ia32_pavgw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1332 static __inline __m128i
1333 _mm_sad_epu8 (__m128i __A
, __m128i __B
)
1335 return (__m128i
)__builtin_ia32_psadbw128 ((__v16qi
)__A
, (__v16qi
)__B
);
1338 static __inline
void
1339 _mm_stream_si32 (int *__A
, int __B
)
1341 __builtin_ia32_movnti (__A
, __B
);
1344 static __inline
void
1345 _mm_stream_si128 (__m128i
*__A
, __m128i __B
)
1347 __builtin_ia32_movntdq ((__v2di
*)__A
, (__v2di
)__B
);
1350 static __inline
void
1351 _mm_stream_pd (double *__A
, __m128d __B
)
1353 __builtin_ia32_movntpd (__A
, (__v2df
)__B
);
1356 static __inline
void
1357 _mm_clflush (void const *__A
)
1359 return __builtin_ia32_clflush (__A
);
1362 static __inline
void
1365 __builtin_ia32_lfence ();
1368 static __inline
void
1371 __builtin_ia32_mfence ();
1374 static __inline __m128i
1375 _mm_cvtsi32_si128 (int __A
)
1377 return _mm_set_epi32 (0, 0, 0, __A
);
1381 static __inline __m128i
1382 _mm_cvtsi64x_si128 (long long __A
)
1384 return _mm_set_epi64x (0, __A
);
1388 /* Casts between various SP, DP, INT vector types. Note that these do no
1389 conversion of values, they just change the type. */
1390 static inline __m128
1391 _mm_castpd_ps(__m128d __A
)
1393 return (__m128
) __A
;
1396 static inline __m128i
1397 _mm_castpd_si128(__m128d __A
)
1399 return (__m128i
) __A
;
1402 static inline __m128d
1403 _mm_castps_pd(__m128 __A
)
1405 return (__m128d
) __A
;
1408 static inline __m128i
1409 _mm_castps_si128(__m128 __A
)
1411 return (__m128i
) __A
;
1414 static inline __m128
1415 _mm_castsi128_ps(__m128i __A
)
1417 return (__m128
) __A
;
1420 static inline __m128d
1421 _mm_castsi128_pd(__m128i __A
)
1423 return (__m128d
) __A
;
1426 #endif /* __SSE2__ */
1428 #endif /* _EMMINTRIN_H_INCLUDED */
