1 /* Copyright (C) 2003, 2004 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
)
199 __builtin_ia32_stored (&__tmp
, (__v4si
)__A
);
204 static __inline
long long
205 _mm_cvtsi128_si64x (__m128i __A
)
207 return __builtin_ia32_movdq2q ((__v2di
)__A
);
212 static __inline __m128d
213 _mm_add_pd (__m128d __A
, __m128d __B
)
215 return (__m128d
)__builtin_ia32_addpd ((__v2df
)__A
, (__v2df
)__B
);
218 static __inline __m128d
219 _mm_add_sd (__m128d __A
, __m128d __B
)
221 return (__m128d
)__builtin_ia32_addsd ((__v2df
)__A
, (__v2df
)__B
);
224 static __inline __m128d
225 _mm_sub_pd (__m128d __A
, __m128d __B
)
227 return (__m128d
)__builtin_ia32_subpd ((__v2df
)__A
, (__v2df
)__B
);
230 static __inline __m128d
231 _mm_sub_sd (__m128d __A
, __m128d __B
)
233 return (__m128d
)__builtin_ia32_subsd ((__v2df
)__A
, (__v2df
)__B
);
236 static __inline __m128d
237 _mm_mul_pd (__m128d __A
, __m128d __B
)
239 return (__m128d
)__builtin_ia32_mulpd ((__v2df
)__A
, (__v2df
)__B
);
242 static __inline __m128d
243 _mm_mul_sd (__m128d __A
, __m128d __B
)
245 return (__m128d
)__builtin_ia32_mulsd ((__v2df
)__A
, (__v2df
)__B
);
248 static __inline __m128d
249 _mm_div_pd (__m128d __A
, __m128d __B
)
251 return (__m128d
)__builtin_ia32_divpd ((__v2df
)__A
, (__v2df
)__B
);
254 static __inline __m128d
255 _mm_div_sd (__m128d __A
, __m128d __B
)
257 return (__m128d
)__builtin_ia32_divsd ((__v2df
)__A
, (__v2df
)__B
);
260 static __inline __m128d
261 _mm_sqrt_pd (__m128d __A
)
263 return (__m128d
)__builtin_ia32_sqrtpd ((__v2df
)__A
);
266 /* Return pair {sqrt (A[0), B[1]}. */
267 static __inline __m128d
268 _mm_sqrt_sd (__m128d __A
, __m128d __B
)
270 __v2df __tmp
= __builtin_ia32_movsd ((__v2df
)__A
, (__v2df
)__B
);
271 return (__m128d
)__builtin_ia32_sqrtsd ((__v2df
)__tmp
);
274 static __inline __m128d
275 _mm_min_pd (__m128d __A
, __m128d __B
)
277 return (__m128d
)__builtin_ia32_minpd ((__v2df
)__A
, (__v2df
)__B
);
280 static __inline __m128d
281 _mm_min_sd (__m128d __A
, __m128d __B
)
283 return (__m128d
)__builtin_ia32_minsd ((__v2df
)__A
, (__v2df
)__B
);
286 static __inline __m128d
287 _mm_max_pd (__m128d __A
, __m128d __B
)
289 return (__m128d
)__builtin_ia32_maxpd ((__v2df
)__A
, (__v2df
)__B
);
292 static __inline __m128d
293 _mm_max_sd (__m128d __A
, __m128d __B
)
295 return (__m128d
)__builtin_ia32_maxsd ((__v2df
)__A
, (__v2df
)__B
);
298 static __inline __m128d
299 _mm_and_pd (__m128d __A
, __m128d __B
)
301 return (__m128d
)__builtin_ia32_andpd ((__v2df
)__A
, (__v2df
)__B
);
304 static __inline __m128d
305 _mm_andnot_pd (__m128d __A
, __m128d __B
)
307 return (__m128d
)__builtin_ia32_andnpd ((__v2df
)__A
, (__v2df
)__B
);
310 static __inline __m128d
311 _mm_or_pd (__m128d __A
, __m128d __B
)
313 return (__m128d
)__builtin_ia32_orpd ((__v2df
)__A
, (__v2df
)__B
);
316 static __inline __m128d
317 _mm_xor_pd (__m128d __A
, __m128d __B
)
319 return (__m128d
)__builtin_ia32_xorpd ((__v2df
)__A
, (__v2df
)__B
);
322 static __inline __m128d
323 _mm_cmpeq_pd (__m128d __A
, __m128d __B
)
325 return (__m128d
)__builtin_ia32_cmpeqpd ((__v2df
)__A
, (__v2df
)__B
);
328 static __inline __m128d
329 _mm_cmplt_pd (__m128d __A
, __m128d __B
)
331 return (__m128d
)__builtin_ia32_cmpltpd ((__v2df
)__A
, (__v2df
)__B
);
334 static __inline __m128d
335 _mm_cmple_pd (__m128d __A
, __m128d __B
)
337 return (__m128d
)__builtin_ia32_cmplepd ((__v2df
)__A
, (__v2df
)__B
);
340 static __inline __m128d
341 _mm_cmpgt_pd (__m128d __A
, __m128d __B
)
343 return (__m128d
)__builtin_ia32_cmpgtpd ((__v2df
)__A
, (__v2df
)__B
);
346 static __inline __m128d
347 _mm_cmpge_pd (__m128d __A
, __m128d __B
)
349 return (__m128d
)__builtin_ia32_cmpgepd ((__v2df
)__A
, (__v2df
)__B
);
352 static __inline __m128d
353 _mm_cmpneq_pd (__m128d __A
, __m128d __B
)
355 return (__m128d
)__builtin_ia32_cmpneqpd ((__v2df
)__A
, (__v2df
)__B
);
358 static __inline __m128d
359 _mm_cmpnlt_pd (__m128d __A
, __m128d __B
)
361 return (__m128d
)__builtin_ia32_cmpnltpd ((__v2df
)__A
, (__v2df
)__B
);
364 static __inline __m128d
365 _mm_cmpnle_pd (__m128d __A
, __m128d __B
)
367 return (__m128d
)__builtin_ia32_cmpnlepd ((__v2df
)__A
, (__v2df
)__B
);
370 static __inline __m128d
371 _mm_cmpngt_pd (__m128d __A
, __m128d __B
)
373 return (__m128d
)__builtin_ia32_cmpngtpd ((__v2df
)__A
, (__v2df
)__B
);
376 static __inline __m128d
377 _mm_cmpnge_pd (__m128d __A
, __m128d __B
)
379 return (__m128d
)__builtin_ia32_cmpngepd ((__v2df
)__A
, (__v2df
)__B
);
382 static __inline __m128d
383 _mm_cmpord_pd (__m128d __A
, __m128d __B
)
385 return (__m128d
)__builtin_ia32_cmpordpd ((__v2df
)__A
, (__v2df
)__B
);
388 static __inline __m128d
389 _mm_cmpunord_pd (__m128d __A
, __m128d __B
)
391 return (__m128d
)__builtin_ia32_cmpunordpd ((__v2df
)__A
, (__v2df
)__B
);
394 static __inline __m128d
395 _mm_cmpeq_sd (__m128d __A
, __m128d __B
)
397 return (__m128d
)__builtin_ia32_cmpeqsd ((__v2df
)__A
, (__v2df
)__B
);
400 static __inline __m128d
401 _mm_cmplt_sd (__m128d __A
, __m128d __B
)
403 return (__m128d
)__builtin_ia32_cmpltsd ((__v2df
)__A
, (__v2df
)__B
);
406 static __inline __m128d
407 _mm_cmple_sd (__m128d __A
, __m128d __B
)
409 return (__m128d
)__builtin_ia32_cmplesd ((__v2df
)__A
, (__v2df
)__B
);
412 static __inline __m128d
413 _mm_cmpgt_sd (__m128d __A
, __m128d __B
)
415 return (__m128d
) __builtin_ia32_movsd ((__v2df
) __A
,
417 __builtin_ia32_cmpltsd ((__v2df
) __B
,
422 static __inline __m128d
423 _mm_cmpge_sd (__m128d __A
, __m128d __B
)
425 return (__m128d
) __builtin_ia32_movsd ((__v2df
) __A
,
427 __builtin_ia32_cmplesd ((__v2df
) __B
,
432 static __inline __m128d
433 _mm_cmpneq_sd (__m128d __A
, __m128d __B
)
435 return (__m128d
)__builtin_ia32_cmpneqsd ((__v2df
)__A
, (__v2df
)__B
);
438 static __inline __m128d
439 _mm_cmpnlt_sd (__m128d __A
, __m128d __B
)
441 return (__m128d
)__builtin_ia32_cmpnltsd ((__v2df
)__A
, (__v2df
)__B
);
444 static __inline __m128d
445 _mm_cmpnle_sd (__m128d __A
, __m128d __B
)
447 return (__m128d
)__builtin_ia32_cmpnlesd ((__v2df
)__A
, (__v2df
)__B
);
450 static __inline __m128d
451 _mm_cmpngt_sd (__m128d __A
, __m128d __B
)
453 return (__m128d
) __builtin_ia32_movsd ((__v2df
) __A
,
455 __builtin_ia32_cmpnltsd ((__v2df
) __B
,
460 static __inline __m128d
461 _mm_cmpnge_sd (__m128d __A
, __m128d __B
)
463 return (__m128d
) __builtin_ia32_movsd ((__v2df
) __A
,
465 __builtin_ia32_cmpnlesd ((__v2df
) __B
,
470 static __inline __m128d
471 _mm_cmpord_sd (__m128d __A
, __m128d __B
)
473 return (__m128d
)__builtin_ia32_cmpordsd ((__v2df
)__A
, (__v2df
)__B
);
476 static __inline __m128d
477 _mm_cmpunord_sd (__m128d __A
, __m128d __B
)
479 return (__m128d
)__builtin_ia32_cmpunordsd ((__v2df
)__A
, (__v2df
)__B
);
483 _mm_comieq_sd (__m128d __A
, __m128d __B
)
485 return __builtin_ia32_comisdeq ((__v2df
)__A
, (__v2df
)__B
);
489 _mm_comilt_sd (__m128d __A
, __m128d __B
)
491 return __builtin_ia32_comisdlt ((__v2df
)__A
, (__v2df
)__B
);
495 _mm_comile_sd (__m128d __A
, __m128d __B
)
497 return __builtin_ia32_comisdle ((__v2df
)__A
, (__v2df
)__B
);
501 _mm_comigt_sd (__m128d __A
, __m128d __B
)
503 return __builtin_ia32_comisdgt ((__v2df
)__A
, (__v2df
)__B
);
507 _mm_comige_sd (__m128d __A
, __m128d __B
)
509 return __builtin_ia32_comisdge ((__v2df
)__A
, (__v2df
)__B
);
513 _mm_comineq_sd (__m128d __A
, __m128d __B
)
515 return __builtin_ia32_comisdneq ((__v2df
)__A
, (__v2df
)__B
);
519 _mm_ucomieq_sd (__m128d __A
, __m128d __B
)
521 return __builtin_ia32_ucomisdeq ((__v2df
)__A
, (__v2df
)__B
);
525 _mm_ucomilt_sd (__m128d __A
, __m128d __B
)
527 return __builtin_ia32_ucomisdlt ((__v2df
)__A
, (__v2df
)__B
);
531 _mm_ucomile_sd (__m128d __A
, __m128d __B
)
533 return __builtin_ia32_ucomisdle ((__v2df
)__A
, (__v2df
)__B
);
537 _mm_ucomigt_sd (__m128d __A
, __m128d __B
)
539 return __builtin_ia32_ucomisdgt ((__v2df
)__A
, (__v2df
)__B
);
543 _mm_ucomige_sd (__m128d __A
, __m128d __B
)
545 return __builtin_ia32_ucomisdge ((__v2df
)__A
, (__v2df
)__B
);
549 _mm_ucomineq_sd (__m128d __A
, __m128d __B
)
551 return __builtin_ia32_ucomisdneq ((__v2df
)__A
, (__v2df
)__B
);
554 /* Create a vector of Qi, where i is the element number. */
556 static __inline __m128i
557 _mm_set_epi64x (long long __q1
, long long __q0
)
559 return (__m128i
)(__v2di
){ __q0
, __q1
};
562 static __inline __m128i
563 _mm_set_epi64 (__m64 __q1
, __m64 __q0
)
565 return _mm_set_epi64x ((long long)__q1
, (long long)__q0
);
568 static __inline __m128i
569 _mm_set_epi32 (int __q3
, int __q2
, int __q1
, int __q0
)
571 return (__m128i
)(__v4si
){ __q0
, __q1
, __q2
, __q3
};
574 static __inline __m128i
575 _mm_set_epi16 (short __q7
, short __q6
, short __q5
, short __q4
,
576 short __q3
, short __q2
, short __q1
, short __q0
)
578 return (__m128i
)(__v8hi
){ __q0
, __q1
, __q2
, __q3
, __q4
, __q5
, __q6
, __q7
};
581 static __inline __m128i
582 _mm_set_epi8 (char __q15
, char __q14
, char __q13
, char __q12
,
583 char __q11
, char __q10
, char __q09
, char __q08
,
584 char __q07
, char __q06
, char __q05
, char __q04
,
585 char __q03
, char __q02
, char __q01
, char __q00
)
587 return (__m128i
)(__v16qi
){
588 __q00
, __q01
, __q02
, __q03
, __q04
, __q05
, __q06
, __q07
,
589 __q08
, __q09
, __q10
, __q11
, __q12
, __q13
, __q14
, __q15
593 /* Set all of the elements of the vector to A. */
595 static __inline __m128i
596 _mm_set1_epi64x (long long __A
)
598 return _mm_set_epi64x (__A
, __A
);
601 static __inline __m128i
602 _mm_set1_epi64 (__m64 __A
)
604 return _mm_set_epi64 (__A
, __A
);
607 static __inline __m128i
608 _mm_set1_epi32 (int __A
)
610 return _mm_set_epi32 (__A
, __A
, __A
, __A
);
613 static __inline __m128i
614 _mm_set1_epi16 (short __A
)
616 return _mm_set_epi16 (__A
, __A
, __A
, __A
, __A
, __A
, __A
, __A
);
619 static __inline __m128i
620 _mm_set1_epi8 (char __A
)
622 return _mm_set_epi8 (__A
, __A
, __A
, __A
, __A
, __A
, __A
, __A
,
623 __A
, __A
, __A
, __A
, __A
, __A
, __A
, __A
);
626 /* Create a vector of Qi, where i is the element number.
627 The parameter order is reversed from the _mm_set_epi* functions. */
629 static __inline __m128i
630 _mm_setr_epi64 (__m64 __q0
, __m64 __q1
)
632 return _mm_set_epi64 (__q1
, __q0
);
635 static __inline __m128i
636 _mm_setr_epi32 (int __q0
, int __q1
, int __q2
, int __q3
)
638 return _mm_set_epi32 (__q3
, __q2
, __q1
, __q0
);
641 static __inline __m128i
642 _mm_setr_epi16 (short __q0
, short __q1
, short __q2
, short __q3
,
643 short __q4
, short __q5
, short __q6
, short __q7
)
645 return _mm_set_epi16 (__q7
, __q6
, __q5
, __q4
, __q3
, __q2
, __q1
, __q0
);
648 static __inline __m128i
649 _mm_setr_epi8 (char __q00
, char __q01
, char __q02
, char __q03
,
650 char __q04
, char __q05
, char __q06
, char __q07
,
651 char __q08
, char __q09
, char __q10
, char __q11
,
652 char __q12
, char __q13
, char __q14
, char __q15
)
654 return _mm_set_epi8 (__q15
, __q14
, __q13
, __q12
, __q11
, __q10
, __q09
, __q08
,
655 __q07
, __q06
, __q05
, __q04
, __q03
, __q02
, __q01
, __q00
);
658 /* Create a vector with element 0 as *P and the rest zero. */
660 static __inline __m128i
661 _mm_load_si128 (__m128i
const *__P
)
666 static __inline __m128i
667 _mm_loadu_si128 (__m128i
const *__P
)
669 return (__m128i
) __builtin_ia32_loaddqu ((char const *)__P
);
672 static __inline __m128i
673 _mm_loadl_epi64 (__m128i
const *__P
)
675 return _mm_set_epi64 ((__m64
)0LL, *(__m64
*)__P
);
679 _mm_store_si128 (__m128i
*__P
, __m128i __B
)
685 _mm_storeu_si128 (__m128i
*__P
, __m128i __B
)
687 __builtin_ia32_storedqu ((char *)__P
, (__v16qi
)__B
);
691 _mm_storel_epi64 (__m128i
*__P
, __m128i __B
)
693 *(long long *)__P
= __builtin_ia32_vec_ext_v2di ((__v2di
)__B
, 0);
696 static __inline __m64
697 _mm_movepi64_pi64 (__m128i __B
)
699 return (__m64
) __builtin_ia32_vec_ext_v2di ((__v2di
)__B
, 0);
702 static __inline __m128i
703 _mm_movpi64_epi64 (__m64 __A
)
705 return _mm_set_epi64 ((__m64
)0LL, __A
);
708 static __inline __m128i
709 _mm_move_epi64 (__m128i __A
)
711 return _mm_set_epi64 ((__m64
)0LL, _mm_movepi64_pi64 (__A
));
714 /* Create a vector of zeros. */
715 static __inline __m128i
716 _mm_setzero_si128 (void)
718 return (__m128i
)(__v4si
){ 0, 0, 0, 0 };
721 static __inline __m128d
722 _mm_cvtepi32_pd (__m128i __A
)
724 return (__m128d
)__builtin_ia32_cvtdq2pd ((__v4si
) __A
);
727 static __inline __m128
728 _mm_cvtepi32_ps (__m128i __A
)
730 return (__m128
)__builtin_ia32_cvtdq2ps ((__v4si
) __A
);
733 static __inline __m128i
734 _mm_cvtpd_epi32 (__m128d __A
)
736 return (__m128i
)__builtin_ia32_cvtpd2dq ((__v2df
) __A
);
739 static __inline __m64
740 _mm_cvtpd_pi32 (__m128d __A
)
742 return (__m64
)__builtin_ia32_cvtpd2pi ((__v2df
) __A
);
745 static __inline __m128
746 _mm_cvtpd_ps (__m128d __A
)
748 return (__m128
)__builtin_ia32_cvtpd2ps ((__v2df
) __A
);
751 static __inline __m128i
752 _mm_cvttpd_epi32 (__m128d __A
)
754 return (__m128i
)__builtin_ia32_cvttpd2dq ((__v2df
) __A
);
757 static __inline __m64
758 _mm_cvttpd_pi32 (__m128d __A
)
760 return (__m64
)__builtin_ia32_cvttpd2pi ((__v2df
) __A
);
763 static __inline __m128d
764 _mm_cvtpi32_pd (__m64 __A
)
766 return (__m128d
)__builtin_ia32_cvtpi2pd ((__v2si
) __A
);
769 static __inline __m128i
770 _mm_cvtps_epi32 (__m128 __A
)
772 return (__m128i
)__builtin_ia32_cvtps2dq ((__v4sf
) __A
);
775 static __inline __m128i
776 _mm_cvttps_epi32 (__m128 __A
)
778 return (__m128i
)__builtin_ia32_cvttps2dq ((__v4sf
) __A
);
781 static __inline __m128d
782 _mm_cvtps_pd (__m128 __A
)
784 return (__m128d
)__builtin_ia32_cvtps2pd ((__v4sf
) __A
);
788 _mm_cvtsd_si32 (__m128d __A
)
790 return __builtin_ia32_cvtsd2si ((__v2df
) __A
);
794 static __inline
long long
795 _mm_cvtsd_si64x (__m128d __A
)
797 return __builtin_ia32_cvtsd2si64 ((__v2df
) __A
);
802 _mm_cvttsd_si32 (__m128d __A
)
804 return __builtin_ia32_cvttsd2si ((__v2df
) __A
);
808 static __inline
long long
809 _mm_cvttsd_si64x (__m128d __A
)
811 return __builtin_ia32_cvttsd2si64 ((__v2df
) __A
);
815 static __inline __m128
816 _mm_cvtsd_ss (__m128 __A
, __m128d __B
)
818 return (__m128
)__builtin_ia32_cvtsd2ss ((__v4sf
) __A
, (__v2df
) __B
);
821 static __inline __m128d
822 _mm_cvtsi32_sd (__m128d __A
, int __B
)
824 return (__m128d
)__builtin_ia32_cvtsi2sd ((__v2df
) __A
, __B
);
828 static __inline __m128d
829 _mm_cvtsi64x_sd (__m128d __A
, long long __B
)
831 return (__m128d
)__builtin_ia32_cvtsi642sd ((__v2df
) __A
, __B
);
835 static __inline __m128d
836 _mm_cvtss_sd (__m128d __A
, __m128 __B
)
838 return (__m128d
)__builtin_ia32_cvtss2sd ((__v2df
) __A
, (__v4sf
)__B
);
841 #define _mm_shuffle_pd(__A, __B, __C) ((__m128d)__builtin_ia32_shufpd ((__v2df)__A, (__v2df)__B, (__C)))
843 static __inline __m128d
844 _mm_unpackhi_pd (__m128d __A
, __m128d __B
)
846 return (__m128d
)__builtin_ia32_unpckhpd ((__v2df
)__A
, (__v2df
)__B
);
849 static __inline __m128d
850 _mm_unpacklo_pd (__m128d __A
, __m128d __B
)
852 return (__m128d
)__builtin_ia32_unpcklpd ((__v2df
)__A
, (__v2df
)__B
);
855 static __inline __m128d
856 _mm_loadh_pd (__m128d __A
, double const *__B
)
858 return (__m128d
)__builtin_ia32_loadhpd ((__v2df
)__A
, __B
);
861 static __inline __m128d
862 _mm_loadl_pd (__m128d __A
, double const *__B
)
864 return (__m128d
)__builtin_ia32_loadlpd ((__v2df
)__A
, __B
);
868 _mm_movemask_pd (__m128d __A
)
870 return __builtin_ia32_movmskpd ((__v2df
)__A
);
873 static __inline __m128i
874 _mm_packs_epi16 (__m128i __A
, __m128i __B
)
876 return (__m128i
)__builtin_ia32_packsswb128 ((__v8hi
)__A
, (__v8hi
)__B
);
879 static __inline __m128i
880 _mm_packs_epi32 (__m128i __A
, __m128i __B
)
882 return (__m128i
)__builtin_ia32_packssdw128 ((__v4si
)__A
, (__v4si
)__B
);
885 static __inline __m128i
886 _mm_packus_epi16 (__m128i __A
, __m128i __B
)
888 return (__m128i
)__builtin_ia32_packuswb128 ((__v8hi
)__A
, (__v8hi
)__B
);
891 static __inline __m128i
892 _mm_unpackhi_epi8 (__m128i __A
, __m128i __B
)
894 return (__m128i
)__builtin_ia32_punpckhbw128 ((__v16qi
)__A
, (__v16qi
)__B
);
897 static __inline __m128i
898 _mm_unpackhi_epi16 (__m128i __A
, __m128i __B
)
900 return (__m128i
)__builtin_ia32_punpckhwd128 ((__v8hi
)__A
, (__v8hi
)__B
);
903 static __inline __m128i
904 _mm_unpackhi_epi32 (__m128i __A
, __m128i __B
)
906 return (__m128i
)__builtin_ia32_punpckhdq128 ((__v4si
)__A
, (__v4si
)__B
);
909 static __inline __m128i
910 _mm_unpackhi_epi64 (__m128i __A
, __m128i __B
)
912 return (__m128i
)__builtin_ia32_punpckhqdq128 ((__v2di
)__A
, (__v2di
)__B
);
915 static __inline __m128i
916 _mm_unpacklo_epi8 (__m128i __A
, __m128i __B
)
918 return (__m128i
)__builtin_ia32_punpcklbw128 ((__v16qi
)__A
, (__v16qi
)__B
);
921 static __inline __m128i
922 _mm_unpacklo_epi16 (__m128i __A
, __m128i __B
)
924 return (__m128i
)__builtin_ia32_punpcklwd128 ((__v8hi
)__A
, (__v8hi
)__B
);
927 static __inline __m128i
928 _mm_unpacklo_epi32 (__m128i __A
, __m128i __B
)
930 return (__m128i
)__builtin_ia32_punpckldq128 ((__v4si
)__A
, (__v4si
)__B
);
933 static __inline __m128i
934 _mm_unpacklo_epi64 (__m128i __A
, __m128i __B
)
936 return (__m128i
)__builtin_ia32_punpcklqdq128 ((__v2di
)__A
, (__v2di
)__B
);
939 static __inline __m128i
940 _mm_add_epi8 (__m128i __A
, __m128i __B
)
942 return (__m128i
)__builtin_ia32_paddb128 ((__v16qi
)__A
, (__v16qi
)__B
);
945 static __inline __m128i
946 _mm_add_epi16 (__m128i __A
, __m128i __B
)
948 return (__m128i
)__builtin_ia32_paddw128 ((__v8hi
)__A
, (__v8hi
)__B
);
951 static __inline __m128i
952 _mm_add_epi32 (__m128i __A
, __m128i __B
)
954 return (__m128i
)__builtin_ia32_paddd128 ((__v4si
)__A
, (__v4si
)__B
);
957 static __inline __m128i
958 _mm_add_epi64 (__m128i __A
, __m128i __B
)
960 return (__m128i
)__builtin_ia32_paddq128 ((__v2di
)__A
, (__v2di
)__B
);
963 static __inline __m128i
964 _mm_adds_epi8 (__m128i __A
, __m128i __B
)
966 return (__m128i
)__builtin_ia32_paddsb128 ((__v16qi
)__A
, (__v16qi
)__B
);
969 static __inline __m128i
970 _mm_adds_epi16 (__m128i __A
, __m128i __B
)
972 return (__m128i
)__builtin_ia32_paddsw128 ((__v8hi
)__A
, (__v8hi
)__B
);
975 static __inline __m128i
976 _mm_adds_epu8 (__m128i __A
, __m128i __B
)
978 return (__m128i
)__builtin_ia32_paddusb128 ((__v16qi
)__A
, (__v16qi
)__B
);
981 static __inline __m128i
982 _mm_adds_epu16 (__m128i __A
, __m128i __B
)
984 return (__m128i
)__builtin_ia32_paddusw128 ((__v8hi
)__A
, (__v8hi
)__B
);
987 static __inline __m128i
988 _mm_sub_epi8 (__m128i __A
, __m128i __B
)
990 return (__m128i
)__builtin_ia32_psubb128 ((__v16qi
)__A
, (__v16qi
)__B
);
993 static __inline __m128i
994 _mm_sub_epi16 (__m128i __A
, __m128i __B
)
996 return (__m128i
)__builtin_ia32_psubw128 ((__v8hi
)__A
, (__v8hi
)__B
);
999 static __inline __m128i
1000 _mm_sub_epi32 (__m128i __A
, __m128i __B
)
1002 return (__m128i
)__builtin_ia32_psubd128 ((__v4si
)__A
, (__v4si
)__B
);
1005 static __inline __m128i
1006 _mm_sub_epi64 (__m128i __A
, __m128i __B
)
1008 return (__m128i
)__builtin_ia32_psubq128 ((__v2di
)__A
, (__v2di
)__B
);
1011 static __inline __m128i
1012 _mm_subs_epi8 (__m128i __A
, __m128i __B
)
1014 return (__m128i
)__builtin_ia32_psubsb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1017 static __inline __m128i
1018 _mm_subs_epi16 (__m128i __A
, __m128i __B
)
1020 return (__m128i
)__builtin_ia32_psubsw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1023 static __inline __m128i
1024 _mm_subs_epu8 (__m128i __A
, __m128i __B
)
1026 return (__m128i
)__builtin_ia32_psubusb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1029 static __inline __m128i
1030 _mm_subs_epu16 (__m128i __A
, __m128i __B
)
1032 return (__m128i
)__builtin_ia32_psubusw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1035 static __inline __m128i
1036 _mm_madd_epi16 (__m128i __A
, __m128i __B
)
1038 return (__m128i
)__builtin_ia32_pmaddwd128 ((__v8hi
)__A
, (__v8hi
)__B
);
1041 static __inline __m128i
1042 _mm_mulhi_epi16 (__m128i __A
, __m128i __B
)
1044 return (__m128i
)__builtin_ia32_pmulhw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1047 static __inline __m128i
1048 _mm_mullo_epi16 (__m128i __A
, __m128i __B
)
1050 return (__m128i
)__builtin_ia32_pmullw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1053 static __inline __m64
1054 _mm_mul_su32 (__m64 __A
, __m64 __B
)
1056 return (__m64
)__builtin_ia32_pmuludq ((__v2si
)__A
, (__v2si
)__B
);
1059 static __inline __m128i
1060 _mm_mul_epu32 (__m128i __A
, __m128i __B
)
1062 return (__m128i
)__builtin_ia32_pmuludq128 ((__v4si
)__A
, (__v4si
)__B
);
1065 static __inline __m128i
1066 _mm_slli_epi16 (__m128i __A
, int __B
)
1068 return (__m128i
)__builtin_ia32_psllwi128 ((__v8hi
)__A
, __B
);
1071 static __inline __m128i
1072 _mm_slli_epi32 (__m128i __A
, int __B
)
1074 return (__m128i
)__builtin_ia32_pslldi128 ((__v4si
)__A
, __B
);
1077 static __inline __m128i
1078 _mm_slli_epi64 (__m128i __A
, int __B
)
1080 return (__m128i
)__builtin_ia32_psllqi128 ((__v2di
)__A
, __B
);
1083 static __inline __m128i
1084 _mm_srai_epi16 (__m128i __A
, int __B
)
1086 return (__m128i
)__builtin_ia32_psrawi128 ((__v8hi
)__A
, __B
);
1089 static __inline __m128i
1090 _mm_srai_epi32 (__m128i __A
, int __B
)
1092 return (__m128i
)__builtin_ia32_psradi128 ((__v4si
)__A
, __B
);
1096 static __m128i
__attribute__((__always_inline__
))
1097 _mm_srli_si128 (__m128i __A
, const int __B
)
1099 return ((__m128i
)__builtin_ia32_psrldqi128 (__A
, __B
))
1102 static __m128i
__attribute__((__always_inline__
))
1103 _mm_srli_si128 (__m128i __A
, const int __B
)
1105 return ((__m128i
)__builtin_ia32_pslldqi128 (__A
, __B
))
1108 #define _mm_srli_si128(__A, __B) \
1109 ((__m128i)__builtin_ia32_psrldqi128 (__A, (__B) * 8))
1110 #define _mm_slli_si128(__A, __B) \
1111 ((__m128i)__builtin_ia32_pslldqi128 (__A, (__B) * 8))
1114 static __inline __m128i
1115 _mm_srli_epi16 (__m128i __A
, int __B
)
1117 return (__m128i
)__builtin_ia32_psrlwi128 ((__v8hi
)__A
, __B
);
1120 static __inline __m128i
1121 _mm_srli_epi32 (__m128i __A
, int __B
)
1123 return (__m128i
)__builtin_ia32_psrldi128 ((__v4si
)__A
, __B
);
1126 static __inline __m128i
1127 _mm_srli_epi64 (__m128i __A
, int __B
)
1129 return (__m128i
)__builtin_ia32_psrlqi128 ((__v2di
)__A
, __B
);
1132 static __inline __m128i
1133 _mm_sll_epi16 (__m128i __A
, __m128i __B
)
1135 return _mm_slli_epi16 (__A
, _mm_cvtsi128_si32 (__B
));
1138 static __inline __m128i
1139 _mm_sll_epi32 (__m128i __A
, __m128i __B
)
1141 return _mm_slli_epi32 (__A
, _mm_cvtsi128_si32 (__B
));
1144 static __inline __m128i
1145 _mm_sll_epi64 (__m128i __A
, __m128i __B
)
1147 return _mm_slli_epi64 (__A
, _mm_cvtsi128_si32 (__B
));
1150 static __inline __m128i
1151 _mm_sra_epi16 (__m128i __A
, __m128i __B
)
1153 return _mm_srai_epi16 (__A
, _mm_cvtsi128_si32 (__B
));
1156 static __inline __m128i
1157 _mm_sra_epi32 (__m128i __A
, __m128i __B
)
1159 return _mm_srai_epi32 (__A
, _mm_cvtsi128_si32 (__B
));
1162 static __inline __m128i
1163 _mm_srl_epi16 (__m128i __A
, __m128i __B
)
1165 return _mm_srli_epi16 (__A
, _mm_cvtsi128_si32 (__B
));
1168 static __inline __m128i
1169 _mm_srl_epi32 (__m128i __A
, __m128i __B
)
1171 return _mm_srli_epi32 (__A
, _mm_cvtsi128_si32 (__B
));
1174 static __inline __m128i
1175 _mm_srl_epi64 (__m128i __A
, __m128i __B
)
1177 return _mm_srli_epi64 (__A
, _mm_cvtsi128_si32 (__B
));
1180 static __inline __m128i
1181 _mm_and_si128 (__m128i __A
, __m128i __B
)
1183 return (__m128i
)__builtin_ia32_pand128 ((__v2di
)__A
, (__v2di
)__B
);
1186 static __inline __m128i
1187 _mm_andnot_si128 (__m128i __A
, __m128i __B
)
1189 return (__m128i
)__builtin_ia32_pandn128 ((__v2di
)__A
, (__v2di
)__B
);
1192 static __inline __m128i
1193 _mm_or_si128 (__m128i __A
, __m128i __B
)
1195 return (__m128i
)__builtin_ia32_por128 ((__v2di
)__A
, (__v2di
)__B
);
1198 static __inline __m128i
1199 _mm_xor_si128 (__m128i __A
, __m128i __B
)
1201 return (__m128i
)__builtin_ia32_pxor128 ((__v2di
)__A
, (__v2di
)__B
);
1204 static __inline __m128i
1205 _mm_cmpeq_epi8 (__m128i __A
, __m128i __B
)
1207 return (__m128i
)__builtin_ia32_pcmpeqb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1210 static __inline __m128i
1211 _mm_cmpeq_epi16 (__m128i __A
, __m128i __B
)
1213 return (__m128i
)__builtin_ia32_pcmpeqw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1216 static __inline __m128i
1217 _mm_cmpeq_epi32 (__m128i __A
, __m128i __B
)
1219 return (__m128i
)__builtin_ia32_pcmpeqd128 ((__v4si
)__A
, (__v4si
)__B
);
1222 static __inline __m128i
1223 _mm_cmplt_epi8 (__m128i __A
, __m128i __B
)
1225 return (__m128i
)__builtin_ia32_pcmpgtb128 ((__v16qi
)__B
, (__v16qi
)__A
);
1228 static __inline __m128i
1229 _mm_cmplt_epi16 (__m128i __A
, __m128i __B
)
1231 return (__m128i
)__builtin_ia32_pcmpgtw128 ((__v8hi
)__B
, (__v8hi
)__A
);
1234 static __inline __m128i
1235 _mm_cmplt_epi32 (__m128i __A
, __m128i __B
)
1237 return (__m128i
)__builtin_ia32_pcmpgtd128 ((__v4si
)__B
, (__v4si
)__A
);
1240 static __inline __m128i
1241 _mm_cmpgt_epi8 (__m128i __A
, __m128i __B
)
1243 return (__m128i
)__builtin_ia32_pcmpgtb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1246 static __inline __m128i
1247 _mm_cmpgt_epi16 (__m128i __A
, __m128i __B
)
1249 return (__m128i
)__builtin_ia32_pcmpgtw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1252 static __inline __m128i
1253 _mm_cmpgt_epi32 (__m128i __A
, __m128i __B
)
1255 return (__m128i
)__builtin_ia32_pcmpgtd128 ((__v4si
)__A
, (__v4si
)__B
);
1259 static __inline
int __attribute__((__always_inline__
))
1260 _mm_extract_epi16 (__m128i
const __A
, int const __N
)
1262 return __builtin_ia32_vec_ext_v8hi ((__v8hi
)__A
, __N
);
1265 static __inline __m128i
__attribute__((__always_inline__
))
1266 _mm_insert_epi16 (__m128i
const __A
, int const __D
, int const __N
)
1268 return (__m128i
) __builtin_ia32_vec_set_v8hi ((__v8hi
)__A
, __D
, __N
);
1271 #define _mm_extract_epi16(A, N) \
1272 ((int) __builtin_ia32_vec_ext_v8hi ((__v8hi)(A), (N))
1273 #define _mm_insert_epi16(A, D, N) \
1274 ((__m128i) __builtin_ia32_vec_set_v8hi ((__v8hi)(A), (D), (N)))
1277 static __inline __m128i
1278 _mm_max_epi16 (__m128i __A
, __m128i __B
)
1280 return (__m128i
)__builtin_ia32_pmaxsw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1283 static __inline __m128i
1284 _mm_max_epu8 (__m128i __A
, __m128i __B
)
1286 return (__m128i
)__builtin_ia32_pmaxub128 ((__v16qi
)__A
, (__v16qi
)__B
);
1289 static __inline __m128i
1290 _mm_min_epi16 (__m128i __A
, __m128i __B
)
1292 return (__m128i
)__builtin_ia32_pminsw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1295 static __inline __m128i
1296 _mm_min_epu8 (__m128i __A
, __m128i __B
)
1298 return (__m128i
)__builtin_ia32_pminub128 ((__v16qi
)__A
, (__v16qi
)__B
);
1302 _mm_movemask_epi8 (__m128i __A
)
1304 return __builtin_ia32_pmovmskb128 ((__v16qi
)__A
);
1307 static __inline __m128i
1308 _mm_mulhi_epu16 (__m128i __A
, __m128i __B
)
1310 return (__m128i
)__builtin_ia32_pmulhuw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1313 #define _mm_shufflehi_epi16(__A, __B) ((__m128i)__builtin_ia32_pshufhw ((__v8hi)__A, __B))
1314 #define _mm_shufflelo_epi16(__A, __B) ((__m128i)__builtin_ia32_pshuflw ((__v8hi)__A, __B))
1315 #define _mm_shuffle_epi32(__A, __B) ((__m128i)__builtin_ia32_pshufd ((__v4si)__A, __B))
1317 static __inline
void
1318 _mm_maskmoveu_si128 (__m128i __A
, __m128i __B
, char *__C
)
1320 __builtin_ia32_maskmovdqu ((__v16qi
)__A
, (__v16qi
)__B
, __C
);
1323 static __inline __m128i
1324 _mm_avg_epu8 (__m128i __A
, __m128i __B
)
1326 return (__m128i
)__builtin_ia32_pavgb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1329 static __inline __m128i
1330 _mm_avg_epu16 (__m128i __A
, __m128i __B
)
1332 return (__m128i
)__builtin_ia32_pavgw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1335 static __inline __m128i
1336 _mm_sad_epu8 (__m128i __A
, __m128i __B
)
1338 return (__m128i
)__builtin_ia32_psadbw128 ((__v16qi
)__A
, (__v16qi
)__B
);
1341 static __inline
void
1342 _mm_stream_si32 (int *__A
, int __B
)
1344 __builtin_ia32_movnti (__A
, __B
);
1347 static __inline
void
1348 _mm_stream_si128 (__m128i
*__A
, __m128i __B
)
1350 __builtin_ia32_movntdq ((__v2di
*)__A
, (__v2di
)__B
);
1353 static __inline
void
1354 _mm_stream_pd (double *__A
, __m128d __B
)
1356 __builtin_ia32_movntpd (__A
, (__v2df
)__B
);
1359 static __inline
void
1360 _mm_clflush (void const *__A
)
1362 return __builtin_ia32_clflush (__A
);
1365 static __inline
void
1368 __builtin_ia32_lfence ();
1371 static __inline
void
1374 __builtin_ia32_mfence ();
1377 static __inline __m128i
1378 _mm_cvtsi32_si128 (int __A
)
1380 return (__m128i
) __builtin_ia32_loadd (&__A
);
1384 static __inline __m128i
1385 _mm_cvtsi64x_si128 (long long __A
)
1387 return (__m128i
) __builtin_ia32_movq2dq (__A
);
1391 #endif /* __SSE2__ */
1393 #endif /* _EMMINTRIN_H_INCLUDED */
projects / gcc.git / blob