f867e530a0b5e98f53f899ecf2aff34d5d55dc32
1 /* Copyright (C) 2003, 2004, 2005, 2006 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, 51 Franklin Street, Fifth Floor,
18 Boston, MA 02110-1301, 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 9.0. */
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
__attribute__((__always_inline__
))
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
__attribute__((__always_inline__
))
59 _mm_set1_pd (double __F
)
61 return __extension__ (__m128d
){ __F
, __F
};
64 static __inline __m128d
__attribute__((__always_inline__
))
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
__attribute__((__always_inline__
))
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
__attribute__((__always_inline__
))
79 _mm_setr_pd (double __W
, double __X
)
81 return __extension__ (__m128d
){ __W
, __X
};
84 /* Create a vector of zeros. */
85 static __inline __m128d
__attribute__((__always_inline__
))
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
__attribute__((__always_inline__
))
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
__attribute__((__always_inline__
))
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
__attribute__((__always_inline__
))
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
__attribute__((__always_inline__
))
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
__attribute__((__always_inline__
))
121 _mm_load_sd (double const *__P
)
123 return _mm_set_sd (*__P
);
126 static __inline __m128d
__attribute__((__always_inline__
))
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
__attribute__((__always_inline__
))
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. */
141 static __inline
void __attribute__((__always_inline__
))
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. */
148 static __inline
void __attribute__((__always_inline__
))
149 _mm_storeu_pd (double *__P
, __m128d __A
)
151 __builtin_ia32_storeupd (__P
, __A
);
154 /* Stores the lower DPFP value. */
155 static __inline
void __attribute__((__always_inline__
))
156 _mm_store_sd (double *__P
, __m128d __A
)
158 *__P
= __builtin_ia32_vec_ext_v2df (__A
, 0);
161 static __inline
double __attribute__((__always_inline__
))
162 _mm_cvtsd_f64 (__m128d __A
)
164 return __builtin_ia32_vec_ext_v2df (__A
, 0);
167 static __inline
void __attribute__((__always_inline__
))
168 _mm_storel_pd (double *__P
, __m128d __A
)
170 _mm_store_sd (__P
, __A
);
173 /* Stores the upper DPFP value. */
174 static __inline
void __attribute__((__always_inline__
))
175 _mm_storeh_pd (double *__P
, __m128d __A
)
177 *__P
= __builtin_ia32_vec_ext_v2df (__A
, 1);
180 /* Store the lower DPFP value across two words.
181 The address must be 16-byte aligned. */
182 static __inline
void __attribute__((__always_inline__
))
183 _mm_store1_pd (double *__P
, __m128d __A
)
185 _mm_store_pd (__P
, __builtin_ia32_shufpd (__A
, __A
, _MM_SHUFFLE2 (0,0)));
188 static __inline
void __attribute__((__always_inline__
))
189 _mm_store_pd1 (double *__P
, __m128d __A
)
191 _mm_store1_pd (__P
, __A
);
194 /* Store two DPFP values in reverse order. The address must be aligned. */
195 static __inline
void __attribute__((__always_inline__
))
196 _mm_storer_pd (double *__P
, __m128d __A
)
198 _mm_store_pd (__P
, __builtin_ia32_shufpd (__A
, __A
, _MM_SHUFFLE2 (0,1)));
201 static __inline
int __attribute__((__always_inline__
))
202 _mm_cvtsi128_si32 (__m128i __A
)
204 return __builtin_ia32_vec_ext_v4si ((__v4si
)__A
, 0);
208 /* Intel intrinsic. */
209 static __inline
long long __attribute__((__always_inline__
))
210 _mm_cvtsi128_si64 (__m128i __A
)
212 return __builtin_ia32_vec_ext_v2di ((__v2di
)__A
, 0);
215 /* Microsoft intrinsic. */
216 static __inline
long long __attribute__((__always_inline__
))
217 _mm_cvtsi128_si64x (__m128i __A
)
219 return __builtin_ia32_vec_ext_v2di ((__v2di
)__A
, 0);
223 static __inline __m128d
__attribute__((__always_inline__
))
224 _mm_add_pd (__m128d __A
, __m128d __B
)
226 return (__m128d
)__builtin_ia32_addpd ((__v2df
)__A
, (__v2df
)__B
);
229 static __inline __m128d
__attribute__((__always_inline__
))
230 _mm_add_sd (__m128d __A
, __m128d __B
)
232 return (__m128d
)__builtin_ia32_addsd ((__v2df
)__A
, (__v2df
)__B
);
235 static __inline __m128d
__attribute__((__always_inline__
))
236 _mm_sub_pd (__m128d __A
, __m128d __B
)
238 return (__m128d
)__builtin_ia32_subpd ((__v2df
)__A
, (__v2df
)__B
);
241 static __inline __m128d
__attribute__((__always_inline__
))
242 _mm_sub_sd (__m128d __A
, __m128d __B
)
244 return (__m128d
)__builtin_ia32_subsd ((__v2df
)__A
, (__v2df
)__B
);
247 static __inline __m128d
__attribute__((__always_inline__
))
248 _mm_mul_pd (__m128d __A
, __m128d __B
)
250 return (__m128d
)__builtin_ia32_mulpd ((__v2df
)__A
, (__v2df
)__B
);
253 static __inline __m128d
__attribute__((__always_inline__
))
254 _mm_mul_sd (__m128d __A
, __m128d __B
)
256 return (__m128d
)__builtin_ia32_mulsd ((__v2df
)__A
, (__v2df
)__B
);
259 static __inline __m128d
__attribute__((__always_inline__
))
260 _mm_div_pd (__m128d __A
, __m128d __B
)
262 return (__m128d
)__builtin_ia32_divpd ((__v2df
)__A
, (__v2df
)__B
);
265 static __inline __m128d
__attribute__((__always_inline__
))
266 _mm_div_sd (__m128d __A
, __m128d __B
)
268 return (__m128d
)__builtin_ia32_divsd ((__v2df
)__A
, (__v2df
)__B
);
271 static __inline __m128d
__attribute__((__always_inline__
))
272 _mm_sqrt_pd (__m128d __A
)
274 return (__m128d
)__builtin_ia32_sqrtpd ((__v2df
)__A
);
277 /* Return pair {sqrt (A[0), B[1]}. */
278 static __inline __m128d
__attribute__((__always_inline__
))
279 _mm_sqrt_sd (__m128d __A
, __m128d __B
)
281 __v2df __tmp
= __builtin_ia32_movsd ((__v2df
)__A
, (__v2df
)__B
);
282 return (__m128d
)__builtin_ia32_sqrtsd ((__v2df
)__tmp
);
285 static __inline __m128d
__attribute__((__always_inline__
))
286 _mm_min_pd (__m128d __A
, __m128d __B
)
288 return (__m128d
)__builtin_ia32_minpd ((__v2df
)__A
, (__v2df
)__B
);
291 static __inline __m128d
__attribute__((__always_inline__
))
292 _mm_min_sd (__m128d __A
, __m128d __B
)
294 return (__m128d
)__builtin_ia32_minsd ((__v2df
)__A
, (__v2df
)__B
);
297 static __inline __m128d
__attribute__((__always_inline__
))
298 _mm_max_pd (__m128d __A
, __m128d __B
)
300 return (__m128d
)__builtin_ia32_maxpd ((__v2df
)__A
, (__v2df
)__B
);
303 static __inline __m128d
__attribute__((__always_inline__
))
304 _mm_max_sd (__m128d __A
, __m128d __B
)
306 return (__m128d
)__builtin_ia32_maxsd ((__v2df
)__A
, (__v2df
)__B
);
309 static __inline __m128d
__attribute__((__always_inline__
))
310 _mm_and_pd (__m128d __A
, __m128d __B
)
312 return (__m128d
)__builtin_ia32_andpd ((__v2df
)__A
, (__v2df
)__B
);
315 static __inline __m128d
__attribute__((__always_inline__
))
316 _mm_andnot_pd (__m128d __A
, __m128d __B
)
318 return (__m128d
)__builtin_ia32_andnpd ((__v2df
)__A
, (__v2df
)__B
);
321 static __inline __m128d
__attribute__((__always_inline__
))
322 _mm_or_pd (__m128d __A
, __m128d __B
)
324 return (__m128d
)__builtin_ia32_orpd ((__v2df
)__A
, (__v2df
)__B
);
327 static __inline __m128d
__attribute__((__always_inline__
))
328 _mm_xor_pd (__m128d __A
, __m128d __B
)
330 return (__m128d
)__builtin_ia32_xorpd ((__v2df
)__A
, (__v2df
)__B
);
333 static __inline __m128d
__attribute__((__always_inline__
))
334 _mm_cmpeq_pd (__m128d __A
, __m128d __B
)
336 return (__m128d
)__builtin_ia32_cmpeqpd ((__v2df
)__A
, (__v2df
)__B
);
339 static __inline __m128d
__attribute__((__always_inline__
))
340 _mm_cmplt_pd (__m128d __A
, __m128d __B
)
342 return (__m128d
)__builtin_ia32_cmpltpd ((__v2df
)__A
, (__v2df
)__B
);
345 static __inline __m128d
__attribute__((__always_inline__
))
346 _mm_cmple_pd (__m128d __A
, __m128d __B
)
348 return (__m128d
)__builtin_ia32_cmplepd ((__v2df
)__A
, (__v2df
)__B
);
351 static __inline __m128d
__attribute__((__always_inline__
))
352 _mm_cmpgt_pd (__m128d __A
, __m128d __B
)
354 return (__m128d
)__builtin_ia32_cmpgtpd ((__v2df
)__A
, (__v2df
)__B
);
357 static __inline __m128d
__attribute__((__always_inline__
))
358 _mm_cmpge_pd (__m128d __A
, __m128d __B
)
360 return (__m128d
)__builtin_ia32_cmpgepd ((__v2df
)__A
, (__v2df
)__B
);
363 static __inline __m128d
__attribute__((__always_inline__
))
364 _mm_cmpneq_pd (__m128d __A
, __m128d __B
)
366 return (__m128d
)__builtin_ia32_cmpneqpd ((__v2df
)__A
, (__v2df
)__B
);
369 static __inline __m128d
__attribute__((__always_inline__
))
370 _mm_cmpnlt_pd (__m128d __A
, __m128d __B
)
372 return (__m128d
)__builtin_ia32_cmpnltpd ((__v2df
)__A
, (__v2df
)__B
);
375 static __inline __m128d
__attribute__((__always_inline__
))
376 _mm_cmpnle_pd (__m128d __A
, __m128d __B
)
378 return (__m128d
)__builtin_ia32_cmpnlepd ((__v2df
)__A
, (__v2df
)__B
);
381 static __inline __m128d
__attribute__((__always_inline__
))
382 _mm_cmpngt_pd (__m128d __A
, __m128d __B
)
384 return (__m128d
)__builtin_ia32_cmpngtpd ((__v2df
)__A
, (__v2df
)__B
);
387 static __inline __m128d
__attribute__((__always_inline__
))
388 _mm_cmpnge_pd (__m128d __A
, __m128d __B
)
390 return (__m128d
)__builtin_ia32_cmpngepd ((__v2df
)__A
, (__v2df
)__B
);
393 static __inline __m128d
__attribute__((__always_inline__
))
394 _mm_cmpord_pd (__m128d __A
, __m128d __B
)
396 return (__m128d
)__builtin_ia32_cmpordpd ((__v2df
)__A
, (__v2df
)__B
);
399 static __inline __m128d
__attribute__((__always_inline__
))
400 _mm_cmpunord_pd (__m128d __A
, __m128d __B
)
402 return (__m128d
)__builtin_ia32_cmpunordpd ((__v2df
)__A
, (__v2df
)__B
);
405 static __inline __m128d
__attribute__((__always_inline__
))
406 _mm_cmpeq_sd (__m128d __A
, __m128d __B
)
408 return (__m128d
)__builtin_ia32_cmpeqsd ((__v2df
)__A
, (__v2df
)__B
);
411 static __inline __m128d
__attribute__((__always_inline__
))
412 _mm_cmplt_sd (__m128d __A
, __m128d __B
)
414 return (__m128d
)__builtin_ia32_cmpltsd ((__v2df
)__A
, (__v2df
)__B
);
417 static __inline __m128d
__attribute__((__always_inline__
))
418 _mm_cmple_sd (__m128d __A
, __m128d __B
)
420 return (__m128d
)__builtin_ia32_cmplesd ((__v2df
)__A
, (__v2df
)__B
);
423 static __inline __m128d
__attribute__((__always_inline__
))
424 _mm_cmpgt_sd (__m128d __A
, __m128d __B
)
426 return (__m128d
) __builtin_ia32_movsd ((__v2df
) __A
,
428 __builtin_ia32_cmpltsd ((__v2df
) __B
,
433 static __inline __m128d
__attribute__((__always_inline__
))
434 _mm_cmpge_sd (__m128d __A
, __m128d __B
)
436 return (__m128d
) __builtin_ia32_movsd ((__v2df
) __A
,
438 __builtin_ia32_cmplesd ((__v2df
) __B
,
443 static __inline __m128d
__attribute__((__always_inline__
))
444 _mm_cmpneq_sd (__m128d __A
, __m128d __B
)
446 return (__m128d
)__builtin_ia32_cmpneqsd ((__v2df
)__A
, (__v2df
)__B
);
449 static __inline __m128d
__attribute__((__always_inline__
))
450 _mm_cmpnlt_sd (__m128d __A
, __m128d __B
)
452 return (__m128d
)__builtin_ia32_cmpnltsd ((__v2df
)__A
, (__v2df
)__B
);
455 static __inline __m128d
__attribute__((__always_inline__
))
456 _mm_cmpnle_sd (__m128d __A
, __m128d __B
)
458 return (__m128d
)__builtin_ia32_cmpnlesd ((__v2df
)__A
, (__v2df
)__B
);
461 static __inline __m128d
__attribute__((__always_inline__
))
462 _mm_cmpngt_sd (__m128d __A
, __m128d __B
)
464 return (__m128d
) __builtin_ia32_movsd ((__v2df
) __A
,
466 __builtin_ia32_cmpnltsd ((__v2df
) __B
,
471 static __inline __m128d
__attribute__((__always_inline__
))
472 _mm_cmpnge_sd (__m128d __A
, __m128d __B
)
474 return (__m128d
) __builtin_ia32_movsd ((__v2df
) __A
,
476 __builtin_ia32_cmpnlesd ((__v2df
) __B
,
481 static __inline __m128d
__attribute__((__always_inline__
))
482 _mm_cmpord_sd (__m128d __A
, __m128d __B
)
484 return (__m128d
)__builtin_ia32_cmpordsd ((__v2df
)__A
, (__v2df
)__B
);
487 static __inline __m128d
__attribute__((__always_inline__
))
488 _mm_cmpunord_sd (__m128d __A
, __m128d __B
)
490 return (__m128d
)__builtin_ia32_cmpunordsd ((__v2df
)__A
, (__v2df
)__B
);
493 static __inline
int __attribute__((__always_inline__
))
494 _mm_comieq_sd (__m128d __A
, __m128d __B
)
496 return __builtin_ia32_comisdeq ((__v2df
)__A
, (__v2df
)__B
);
499 static __inline
int __attribute__((__always_inline__
))
500 _mm_comilt_sd (__m128d __A
, __m128d __B
)
502 return __builtin_ia32_comisdlt ((__v2df
)__A
, (__v2df
)__B
);
505 static __inline
int __attribute__((__always_inline__
))
506 _mm_comile_sd (__m128d __A
, __m128d __B
)
508 return __builtin_ia32_comisdle ((__v2df
)__A
, (__v2df
)__B
);
511 static __inline
int __attribute__((__always_inline__
))
512 _mm_comigt_sd (__m128d __A
, __m128d __B
)
514 return __builtin_ia32_comisdgt ((__v2df
)__A
, (__v2df
)__B
);
517 static __inline
int __attribute__((__always_inline__
))
518 _mm_comige_sd (__m128d __A
, __m128d __B
)
520 return __builtin_ia32_comisdge ((__v2df
)__A
, (__v2df
)__B
);
523 static __inline
int __attribute__((__always_inline__
))
524 _mm_comineq_sd (__m128d __A
, __m128d __B
)
526 return __builtin_ia32_comisdneq ((__v2df
)__A
, (__v2df
)__B
);
529 static __inline
int __attribute__((__always_inline__
))
530 _mm_ucomieq_sd (__m128d __A
, __m128d __B
)
532 return __builtin_ia32_ucomisdeq ((__v2df
)__A
, (__v2df
)__B
);
535 static __inline
int __attribute__((__always_inline__
))
536 _mm_ucomilt_sd (__m128d __A
, __m128d __B
)
538 return __builtin_ia32_ucomisdlt ((__v2df
)__A
, (__v2df
)__B
);
541 static __inline
int __attribute__((__always_inline__
))
542 _mm_ucomile_sd (__m128d __A
, __m128d __B
)
544 return __builtin_ia32_ucomisdle ((__v2df
)__A
, (__v2df
)__B
);
547 static __inline
int __attribute__((__always_inline__
))
548 _mm_ucomigt_sd (__m128d __A
, __m128d __B
)
550 return __builtin_ia32_ucomisdgt ((__v2df
)__A
, (__v2df
)__B
);
553 static __inline
int __attribute__((__always_inline__
))
554 _mm_ucomige_sd (__m128d __A
, __m128d __B
)
556 return __builtin_ia32_ucomisdge ((__v2df
)__A
, (__v2df
)__B
);
559 static __inline
int __attribute__((__always_inline__
))
560 _mm_ucomineq_sd (__m128d __A
, __m128d __B
)
562 return __builtin_ia32_ucomisdneq ((__v2df
)__A
, (__v2df
)__B
);
565 /* Create a vector of Qi, where i is the element number. */
567 static __inline __m128i
__attribute__((__always_inline__
))
568 _mm_set_epi64x (long long __q1
, long long __q0
)
570 return __extension__ (__m128i
)(__v2di
){ __q0
, __q1
};
573 static __inline __m128i
__attribute__((__always_inline__
))
574 _mm_set_epi64 (__m64 __q1
, __m64 __q0
)
576 return _mm_set_epi64x ((long long)__q1
, (long long)__q0
);
579 static __inline __m128i
__attribute__((__always_inline__
))
580 _mm_set_epi32 (int __q3
, int __q2
, int __q1
, int __q0
)
582 return __extension__ (__m128i
)(__v4si
){ __q0
, __q1
, __q2
, __q3
};
585 static __inline __m128i
__attribute__((__always_inline__
))
586 _mm_set_epi16 (short __q7
, short __q6
, short __q5
, short __q4
,
587 short __q3
, short __q2
, short __q1
, short __q0
)
589 return __extension__ (__m128i
)(__v8hi
){
590 __q0
, __q1
, __q2
, __q3
, __q4
, __q5
, __q6
, __q7
};
593 static __inline __m128i
__attribute__((__always_inline__
))
594 _mm_set_epi8 (char __q15
, char __q14
, char __q13
, char __q12
,
595 char __q11
, char __q10
, char __q09
, char __q08
,
596 char __q07
, char __q06
, char __q05
, char __q04
,
597 char __q03
, char __q02
, char __q01
, char __q00
)
599 return __extension__ (__m128i
)(__v16qi
){
600 __q00
, __q01
, __q02
, __q03
, __q04
, __q05
, __q06
, __q07
,
601 __q08
, __q09
, __q10
, __q11
, __q12
, __q13
, __q14
, __q15
605 /* Set all of the elements of the vector to A. */
607 static __inline __m128i
__attribute__((__always_inline__
))
608 _mm_set1_epi64x (long long __A
)
610 return _mm_set_epi64x (__A
, __A
);
613 static __inline __m128i
__attribute__((__always_inline__
))
614 _mm_set1_epi64 (__m64 __A
)
616 return _mm_set_epi64 (__A
, __A
);
619 static __inline __m128i
__attribute__((__always_inline__
))
620 _mm_set1_epi32 (int __A
)
622 return _mm_set_epi32 (__A
, __A
, __A
, __A
);
625 static __inline __m128i
__attribute__((__always_inline__
))
626 _mm_set1_epi16 (short __A
)
628 return _mm_set_epi16 (__A
, __A
, __A
, __A
, __A
, __A
, __A
, __A
);
631 static __inline __m128i
__attribute__((__always_inline__
))
632 _mm_set1_epi8 (char __A
)
634 return _mm_set_epi8 (__A
, __A
, __A
, __A
, __A
, __A
, __A
, __A
,
635 __A
, __A
, __A
, __A
, __A
, __A
, __A
, __A
);
638 /* Create a vector of Qi, where i is the element number.
639 The parameter order is reversed from the _mm_set_epi* functions. */
641 static __inline __m128i
__attribute__((__always_inline__
))
642 _mm_setr_epi64 (__m64 __q0
, __m64 __q1
)
644 return _mm_set_epi64 (__q1
, __q0
);
647 static __inline __m128i
__attribute__((__always_inline__
))
648 _mm_setr_epi32 (int __q0
, int __q1
, int __q2
, int __q3
)
650 return _mm_set_epi32 (__q3
, __q2
, __q1
, __q0
);
653 static __inline __m128i
__attribute__((__always_inline__
))
654 _mm_setr_epi16 (short __q0
, short __q1
, short __q2
, short __q3
,
655 short __q4
, short __q5
, short __q6
, short __q7
)
657 return _mm_set_epi16 (__q7
, __q6
, __q5
, __q4
, __q3
, __q2
, __q1
, __q0
);
660 static __inline __m128i
__attribute__((__always_inline__
))
661 _mm_setr_epi8 (char __q00
, char __q01
, char __q02
, char __q03
,
662 char __q04
, char __q05
, char __q06
, char __q07
,
663 char __q08
, char __q09
, char __q10
, char __q11
,
664 char __q12
, char __q13
, char __q14
, char __q15
)
666 return _mm_set_epi8 (__q15
, __q14
, __q13
, __q12
, __q11
, __q10
, __q09
, __q08
,
667 __q07
, __q06
, __q05
, __q04
, __q03
, __q02
, __q01
, __q00
);
670 /* Create a vector with element 0 as *P and the rest zero. */
672 static __inline __m128i
__attribute__((__always_inline__
))
673 _mm_load_si128 (__m128i
const *__P
)
678 static __inline __m128i
__attribute__((__always_inline__
))
679 _mm_loadu_si128 (__m128i
const *__P
)
681 return (__m128i
) __builtin_ia32_loaddqu ((char const *)__P
);
684 static __inline __m128i
__attribute__((__always_inline__
))
685 _mm_loadl_epi64 (__m128i
const *__P
)
687 return _mm_set_epi64 ((__m64
)0LL, *(__m64
*)__P
);
690 static __inline
void __attribute__((__always_inline__
))
691 _mm_store_si128 (__m128i
*__P
, __m128i __B
)
696 static __inline
void __attribute__((__always_inline__
))
697 _mm_storeu_si128 (__m128i
*__P
, __m128i __B
)
699 __builtin_ia32_storedqu ((char *)__P
, (__v16qi
)__B
);
702 static __inline
void __attribute__((__always_inline__
))
703 _mm_storel_epi64 (__m128i
*__P
, __m128i __B
)
705 *(long long *)__P
= __builtin_ia32_vec_ext_v2di ((__v2di
)__B
, 0);
708 static __inline __m64
__attribute__((__always_inline__
))
709 _mm_movepi64_pi64 (__m128i __B
)
711 return (__m64
) __builtin_ia32_vec_ext_v2di ((__v2di
)__B
, 0);
714 static __inline __m128i
__attribute__((__always_inline__
))
715 _mm_movpi64_epi64 (__m64 __A
)
717 return _mm_set_epi64 ((__m64
)0LL, __A
);
720 static __inline __m128i
__attribute__((__always_inline__
))
721 _mm_move_epi64 (__m128i __A
)
723 return _mm_set_epi64 ((__m64
)0LL, _mm_movepi64_pi64 (__A
));
726 /* Create a vector of zeros. */
727 static __inline __m128i
__attribute__((__always_inline__
))
728 _mm_setzero_si128 (void)
730 return __extension__ (__m128i
)(__v4si
){ 0, 0, 0, 0 };
733 static __inline __m128d
__attribute__((__always_inline__
))
734 _mm_cvtepi32_pd (__m128i __A
)
736 return (__m128d
)__builtin_ia32_cvtdq2pd ((__v4si
) __A
);
739 static __inline __m128
__attribute__((__always_inline__
))
740 _mm_cvtepi32_ps (__m128i __A
)
742 return (__m128
)__builtin_ia32_cvtdq2ps ((__v4si
) __A
);
745 static __inline __m128i
__attribute__((__always_inline__
))
746 _mm_cvtpd_epi32 (__m128d __A
)
748 return (__m128i
)__builtin_ia32_cvtpd2dq ((__v2df
) __A
);
751 static __inline __m64
__attribute__((__always_inline__
))
752 _mm_cvtpd_pi32 (__m128d __A
)
754 return (__m64
)__builtin_ia32_cvtpd2pi ((__v2df
) __A
);
757 static __inline __m128
__attribute__((__always_inline__
))
758 _mm_cvtpd_ps (__m128d __A
)
760 return (__m128
)__builtin_ia32_cvtpd2ps ((__v2df
) __A
);
763 static __inline __m128i
__attribute__((__always_inline__
))
764 _mm_cvttpd_epi32 (__m128d __A
)
766 return (__m128i
)__builtin_ia32_cvttpd2dq ((__v2df
) __A
);
769 static __inline __m64
__attribute__((__always_inline__
))
770 _mm_cvttpd_pi32 (__m128d __A
)
772 return (__m64
)__builtin_ia32_cvttpd2pi ((__v2df
) __A
);
775 static __inline __m128d
__attribute__((__always_inline__
))
776 _mm_cvtpi32_pd (__m64 __A
)
778 return (__m128d
)__builtin_ia32_cvtpi2pd ((__v2si
) __A
);
781 static __inline __m128i
__attribute__((__always_inline__
))
782 _mm_cvtps_epi32 (__m128 __A
)
784 return (__m128i
)__builtin_ia32_cvtps2dq ((__v4sf
) __A
);
787 static __inline __m128i
__attribute__((__always_inline__
))
788 _mm_cvttps_epi32 (__m128 __A
)
790 return (__m128i
)__builtin_ia32_cvttps2dq ((__v4sf
) __A
);
793 static __inline __m128d
__attribute__((__always_inline__
))
794 _mm_cvtps_pd (__m128 __A
)
796 return (__m128d
)__builtin_ia32_cvtps2pd ((__v4sf
) __A
);
799 static __inline
int __attribute__((__always_inline__
))
800 _mm_cvtsd_si32 (__m128d __A
)
802 return __builtin_ia32_cvtsd2si ((__v2df
) __A
);
806 /* Intel intrinsic. */
807 static __inline
long long __attribute__((__always_inline__
))
808 _mm_cvtsd_si64 (__m128d __A
)
810 return __builtin_ia32_cvtsd2si64 ((__v2df
) __A
);
813 /* Microsoft intrinsic. */
814 static __inline
long long __attribute__((__always_inline__
))
815 _mm_cvtsd_si64x (__m128d __A
)
817 return __builtin_ia32_cvtsd2si64 ((__v2df
) __A
);
821 static __inline
int __attribute__((__always_inline__
))
822 _mm_cvttsd_si32 (__m128d __A
)
824 return __builtin_ia32_cvttsd2si ((__v2df
) __A
);
828 /* Intel intrinsic. */
829 static __inline
long long __attribute__((__always_inline__
))
830 _mm_cvttsd_si64 (__m128d __A
)
832 return __builtin_ia32_cvttsd2si64 ((__v2df
) __A
);
835 /* Microsoft intrinsic. */
836 static __inline
long long __attribute__((__always_inline__
))
837 _mm_cvttsd_si64x (__m128d __A
)
839 return __builtin_ia32_cvttsd2si64 ((__v2df
) __A
);
843 static __inline __m128
__attribute__((__always_inline__
))
844 _mm_cvtsd_ss (__m128 __A
, __m128d __B
)
846 return (__m128
)__builtin_ia32_cvtsd2ss ((__v4sf
) __A
, (__v2df
) __B
);
849 static __inline __m128d
__attribute__((__always_inline__
))
850 _mm_cvtsi32_sd (__m128d __A
, int __B
)
852 return (__m128d
)__builtin_ia32_cvtsi2sd ((__v2df
) __A
, __B
);
856 /* Intel intrinsic. */
857 static __inline __m128d
__attribute__((__always_inline__
))
858 _mm_cvtsi64_sd (__m128d __A
, long long __B
)
860 return (__m128d
)__builtin_ia32_cvtsi642sd ((__v2df
) __A
, __B
);
863 /* Microsoft intrinsic. */
864 static __inline __m128d
__attribute__((__always_inline__
))
865 _mm_cvtsi64x_sd (__m128d __A
, long long __B
)
867 return (__m128d
)__builtin_ia32_cvtsi642sd ((__v2df
) __A
, __B
);
871 static __inline __m128d
__attribute__((__always_inline__
))
872 _mm_cvtss_sd (__m128d __A
, __m128 __B
)
874 return (__m128d
)__builtin_ia32_cvtss2sd ((__v2df
) __A
, (__v4sf
)__B
);
877 #define _mm_shuffle_pd(__A, __B, __C) ((__m128d)__builtin_ia32_shufpd ((__v2df)__A, (__v2df)__B, (__C)))
879 static __inline __m128d
__attribute__((__always_inline__
))
880 _mm_unpackhi_pd (__m128d __A
, __m128d __B
)
882 return (__m128d
)__builtin_ia32_unpckhpd ((__v2df
)__A
, (__v2df
)__B
);
885 static __inline __m128d
__attribute__((__always_inline__
))
886 _mm_unpacklo_pd (__m128d __A
, __m128d __B
)
888 return (__m128d
)__builtin_ia32_unpcklpd ((__v2df
)__A
, (__v2df
)__B
);
891 static __inline __m128d
__attribute__((__always_inline__
))
892 _mm_loadh_pd (__m128d __A
, double const *__B
)
894 return (__m128d
)__builtin_ia32_loadhpd ((__v2df
)__A
, __B
);
897 static __inline __m128d
__attribute__((__always_inline__
))
898 _mm_loadl_pd (__m128d __A
, double const *__B
)
900 return (__m128d
)__builtin_ia32_loadlpd ((__v2df
)__A
, __B
);
903 static __inline
int __attribute__((__always_inline__
))
904 _mm_movemask_pd (__m128d __A
)
906 return __builtin_ia32_movmskpd ((__v2df
)__A
);
909 static __inline __m128i
__attribute__((__always_inline__
))
910 _mm_packs_epi16 (__m128i __A
, __m128i __B
)
912 return (__m128i
)__builtin_ia32_packsswb128 ((__v8hi
)__A
, (__v8hi
)__B
);
915 static __inline __m128i
__attribute__((__always_inline__
))
916 _mm_packs_epi32 (__m128i __A
, __m128i __B
)
918 return (__m128i
)__builtin_ia32_packssdw128 ((__v4si
)__A
, (__v4si
)__B
);
921 static __inline __m128i
__attribute__((__always_inline__
))
922 _mm_packus_epi16 (__m128i __A
, __m128i __B
)
924 return (__m128i
)__builtin_ia32_packuswb128 ((__v8hi
)__A
, (__v8hi
)__B
);
927 static __inline __m128i
__attribute__((__always_inline__
))
928 _mm_unpackhi_epi8 (__m128i __A
, __m128i __B
)
930 return (__m128i
)__builtin_ia32_punpckhbw128 ((__v16qi
)__A
, (__v16qi
)__B
);
933 static __inline __m128i
__attribute__((__always_inline__
))
934 _mm_unpackhi_epi16 (__m128i __A
, __m128i __B
)
936 return (__m128i
)__builtin_ia32_punpckhwd128 ((__v8hi
)__A
, (__v8hi
)__B
);
939 static __inline __m128i
__attribute__((__always_inline__
))
940 _mm_unpackhi_epi32 (__m128i __A
, __m128i __B
)
942 return (__m128i
)__builtin_ia32_punpckhdq128 ((__v4si
)__A
, (__v4si
)__B
);
945 static __inline __m128i
__attribute__((__always_inline__
))
946 _mm_unpackhi_epi64 (__m128i __A
, __m128i __B
)
948 return (__m128i
)__builtin_ia32_punpckhqdq128 ((__v2di
)__A
, (__v2di
)__B
);
951 static __inline __m128i
__attribute__((__always_inline__
))
952 _mm_unpacklo_epi8 (__m128i __A
, __m128i __B
)
954 return (__m128i
)__builtin_ia32_punpcklbw128 ((__v16qi
)__A
, (__v16qi
)__B
);
957 static __inline __m128i
__attribute__((__always_inline__
))
958 _mm_unpacklo_epi16 (__m128i __A
, __m128i __B
)
960 return (__m128i
)__builtin_ia32_punpcklwd128 ((__v8hi
)__A
, (__v8hi
)__B
);
963 static __inline __m128i
__attribute__((__always_inline__
))
964 _mm_unpacklo_epi32 (__m128i __A
, __m128i __B
)
966 return (__m128i
)__builtin_ia32_punpckldq128 ((__v4si
)__A
, (__v4si
)__B
);
969 static __inline __m128i
__attribute__((__always_inline__
))
970 _mm_unpacklo_epi64 (__m128i __A
, __m128i __B
)
972 return (__m128i
)__builtin_ia32_punpcklqdq128 ((__v2di
)__A
, (__v2di
)__B
);
975 static __inline __m128i
__attribute__((__always_inline__
))
976 _mm_add_epi8 (__m128i __A
, __m128i __B
)
978 return (__m128i
)__builtin_ia32_paddb128 ((__v16qi
)__A
, (__v16qi
)__B
);
981 static __inline __m128i
__attribute__((__always_inline__
))
982 _mm_add_epi16 (__m128i __A
, __m128i __B
)
984 return (__m128i
)__builtin_ia32_paddw128 ((__v8hi
)__A
, (__v8hi
)__B
);
987 static __inline __m128i
__attribute__((__always_inline__
))
988 _mm_add_epi32 (__m128i __A
, __m128i __B
)
990 return (__m128i
)__builtin_ia32_paddd128 ((__v4si
)__A
, (__v4si
)__B
);
993 static __inline __m128i
__attribute__((__always_inline__
))
994 _mm_add_epi64 (__m128i __A
, __m128i __B
)
996 return (__m128i
)__builtin_ia32_paddq128 ((__v2di
)__A
, (__v2di
)__B
);
999 static __inline __m128i
__attribute__((__always_inline__
))
1000 _mm_adds_epi8 (__m128i __A
, __m128i __B
)
1002 return (__m128i
)__builtin_ia32_paddsb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1005 static __inline __m128i
__attribute__((__always_inline__
))
1006 _mm_adds_epi16 (__m128i __A
, __m128i __B
)
1008 return (__m128i
)__builtin_ia32_paddsw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1011 static __inline __m128i
__attribute__((__always_inline__
))
1012 _mm_adds_epu8 (__m128i __A
, __m128i __B
)
1014 return (__m128i
)__builtin_ia32_paddusb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1017 static __inline __m128i
__attribute__((__always_inline__
))
1018 _mm_adds_epu16 (__m128i __A
, __m128i __B
)
1020 return (__m128i
)__builtin_ia32_paddusw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1023 static __inline __m128i
__attribute__((__always_inline__
))
1024 _mm_sub_epi8 (__m128i __A
, __m128i __B
)
1026 return (__m128i
)__builtin_ia32_psubb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1029 static __inline __m128i
__attribute__((__always_inline__
))
1030 _mm_sub_epi16 (__m128i __A
, __m128i __B
)
1032 return (__m128i
)__builtin_ia32_psubw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1035 static __inline __m128i
__attribute__((__always_inline__
))
1036 _mm_sub_epi32 (__m128i __A
, __m128i __B
)
1038 return (__m128i
)__builtin_ia32_psubd128 ((__v4si
)__A
, (__v4si
)__B
);
1041 static __inline __m128i
__attribute__((__always_inline__
))
1042 _mm_sub_epi64 (__m128i __A
, __m128i __B
)
1044 return (__m128i
)__builtin_ia32_psubq128 ((__v2di
)__A
, (__v2di
)__B
);
1047 static __inline __m128i
__attribute__((__always_inline__
))
1048 _mm_subs_epi8 (__m128i __A
, __m128i __B
)
1050 return (__m128i
)__builtin_ia32_psubsb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1053 static __inline __m128i
__attribute__((__always_inline__
))
1054 _mm_subs_epi16 (__m128i __A
, __m128i __B
)
1056 return (__m128i
)__builtin_ia32_psubsw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1059 static __inline __m128i
__attribute__((__always_inline__
))
1060 _mm_subs_epu8 (__m128i __A
, __m128i __B
)
1062 return (__m128i
)__builtin_ia32_psubusb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1065 static __inline __m128i
__attribute__((__always_inline__
))
1066 _mm_subs_epu16 (__m128i __A
, __m128i __B
)
1068 return (__m128i
)__builtin_ia32_psubusw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1071 static __inline __m128i
__attribute__((__always_inline__
))
1072 _mm_madd_epi16 (__m128i __A
, __m128i __B
)
1074 return (__m128i
)__builtin_ia32_pmaddwd128 ((__v8hi
)__A
, (__v8hi
)__B
);
1077 static __inline __m128i
__attribute__((__always_inline__
))
1078 _mm_mulhi_epi16 (__m128i __A
, __m128i __B
)
1080 return (__m128i
)__builtin_ia32_pmulhw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1083 static __inline __m128i
__attribute__((__always_inline__
))
1084 _mm_mullo_epi16 (__m128i __A
, __m128i __B
)
1086 return (__m128i
)__builtin_ia32_pmullw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1089 static __inline __m64
__attribute__((__always_inline__
))
1090 _mm_mul_su32 (__m64 __A
, __m64 __B
)
1092 return (__m64
)__builtin_ia32_pmuludq ((__v2si
)__A
, (__v2si
)__B
);
1095 static __inline __m128i
__attribute__((__always_inline__
))
1096 _mm_mul_epu32 (__m128i __A
, __m128i __B
)
1098 return (__m128i
)__builtin_ia32_pmuludq128 ((__v4si
)__A
, (__v4si
)__B
);
1101 static __inline __m128i
__attribute__((__always_inline__
))
1102 _mm_slli_epi16 (__m128i __A
, int __B
)
1104 return (__m128i
)__builtin_ia32_psllwi128 ((__v8hi
)__A
, __B
);
1107 static __inline __m128i
__attribute__((__always_inline__
))
1108 _mm_slli_epi32 (__m128i __A
, int __B
)
1110 return (__m128i
)__builtin_ia32_pslldi128 ((__v4si
)__A
, __B
);
1113 static __inline __m128i
__attribute__((__always_inline__
))
1114 _mm_slli_epi64 (__m128i __A
, int __B
)
1116 return (__m128i
)__builtin_ia32_psllqi128 ((__v2di
)__A
, __B
);
1119 static __inline __m128i
__attribute__((__always_inline__
))
1120 _mm_srai_epi16 (__m128i __A
, int __B
)
1122 return (__m128i
)__builtin_ia32_psrawi128 ((__v8hi
)__A
, __B
);
1125 static __inline __m128i
__attribute__((__always_inline__
))
1126 _mm_srai_epi32 (__m128i __A
, int __B
)
1128 return (__m128i
)__builtin_ia32_psradi128 ((__v4si
)__A
, __B
);
1132 static __m128i
__attribute__((__always_inline__
))
1133 _mm_srli_si128 (__m128i __A
, const int __B
)
1135 return ((__m128i
)__builtin_ia32_psrldqi128 (__A
, __B
))
1138 static __m128i
__attribute__((__always_inline__
))
1139 _mm_srli_si128 (__m128i __A
, const int __B
)
1141 return ((__m128i
)__builtin_ia32_pslldqi128 (__A
, __B
))
1144 #define _mm_srli_si128(__A, __B) \
1145 ((__m128i)__builtin_ia32_psrldqi128 (__A, (__B) * 8))
1146 #define _mm_slli_si128(__A, __B) \
1147 ((__m128i)__builtin_ia32_pslldqi128 (__A, (__B) * 8))
1150 static __inline __m128i
__attribute__((__always_inline__
))
1151 _mm_srli_epi16 (__m128i __A
, int __B
)
1153 return (__m128i
)__builtin_ia32_psrlwi128 ((__v8hi
)__A
, __B
);
1156 static __inline __m128i
__attribute__((__always_inline__
))
1157 _mm_srli_epi32 (__m128i __A
, int __B
)
1159 return (__m128i
)__builtin_ia32_psrldi128 ((__v4si
)__A
, __B
);
1162 static __inline __m128i
__attribute__((__always_inline__
))
1163 _mm_srli_epi64 (__m128i __A
, int __B
)
1165 return (__m128i
)__builtin_ia32_psrlqi128 ((__v2di
)__A
, __B
);
1168 static __inline __m128i
__attribute__((__always_inline__
))
1169 _mm_sll_epi16 (__m128i __A
, __m128i __B
)
1171 return _mm_slli_epi16 (__A
, _mm_cvtsi128_si32 (__B
));
1174 static __inline __m128i
__attribute__((__always_inline__
))
1175 _mm_sll_epi32 (__m128i __A
, __m128i __B
)
1177 return _mm_slli_epi32 (__A
, _mm_cvtsi128_si32 (__B
));
1180 static __inline __m128i
__attribute__((__always_inline__
))
1181 _mm_sll_epi64 (__m128i __A
, __m128i __B
)
1183 return _mm_slli_epi64 (__A
, _mm_cvtsi128_si32 (__B
));
1186 static __inline __m128i
__attribute__((__always_inline__
))
1187 _mm_sra_epi16 (__m128i __A
, __m128i __B
)
1189 return _mm_srai_epi16 (__A
, _mm_cvtsi128_si32 (__B
));
1192 static __inline __m128i
__attribute__((__always_inline__
))
1193 _mm_sra_epi32 (__m128i __A
, __m128i __B
)
1195 return _mm_srai_epi32 (__A
, _mm_cvtsi128_si32 (__B
));
1198 static __inline __m128i
__attribute__((__always_inline__
))
1199 _mm_srl_epi16 (__m128i __A
, __m128i __B
)
1201 return _mm_srli_epi16 (__A
, _mm_cvtsi128_si32 (__B
));
1204 static __inline __m128i
__attribute__((__always_inline__
))
1205 _mm_srl_epi32 (__m128i __A
, __m128i __B
)
1207 return _mm_srli_epi32 (__A
, _mm_cvtsi128_si32 (__B
));
1210 static __inline __m128i
__attribute__((__always_inline__
))
1211 _mm_srl_epi64 (__m128i __A
, __m128i __B
)
1213 return _mm_srli_epi64 (__A
, _mm_cvtsi128_si32 (__B
));
1216 static __inline __m128i
__attribute__((__always_inline__
))
1217 _mm_and_si128 (__m128i __A
, __m128i __B
)
1219 return (__m128i
)__builtin_ia32_pand128 ((__v2di
)__A
, (__v2di
)__B
);
1222 static __inline __m128i
__attribute__((__always_inline__
))
1223 _mm_andnot_si128 (__m128i __A
, __m128i __B
)
1225 return (__m128i
)__builtin_ia32_pandn128 ((__v2di
)__A
, (__v2di
)__B
);
1228 static __inline __m128i
__attribute__((__always_inline__
))
1229 _mm_or_si128 (__m128i __A
, __m128i __B
)
1231 return (__m128i
)__builtin_ia32_por128 ((__v2di
)__A
, (__v2di
)__B
);
1234 static __inline __m128i
__attribute__((__always_inline__
))
1235 _mm_xor_si128 (__m128i __A
, __m128i __B
)
1237 return (__m128i
)__builtin_ia32_pxor128 ((__v2di
)__A
, (__v2di
)__B
);
1240 static __inline __m128i
__attribute__((__always_inline__
))
1241 _mm_cmpeq_epi8 (__m128i __A
, __m128i __B
)
1243 return (__m128i
)__builtin_ia32_pcmpeqb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1246 static __inline __m128i
__attribute__((__always_inline__
))
1247 _mm_cmpeq_epi16 (__m128i __A
, __m128i __B
)
1249 return (__m128i
)__builtin_ia32_pcmpeqw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1252 static __inline __m128i
__attribute__((__always_inline__
))
1253 _mm_cmpeq_epi32 (__m128i __A
, __m128i __B
)
1255 return (__m128i
)__builtin_ia32_pcmpeqd128 ((__v4si
)__A
, (__v4si
)__B
);
1258 static __inline __m128i
__attribute__((__always_inline__
))
1259 _mm_cmplt_epi8 (__m128i __A
, __m128i __B
)
1261 return (__m128i
)__builtin_ia32_pcmpgtb128 ((__v16qi
)__B
, (__v16qi
)__A
);
1264 static __inline __m128i
__attribute__((__always_inline__
))
1265 _mm_cmplt_epi16 (__m128i __A
, __m128i __B
)
1267 return (__m128i
)__builtin_ia32_pcmpgtw128 ((__v8hi
)__B
, (__v8hi
)__A
);
1270 static __inline __m128i
__attribute__((__always_inline__
))
1271 _mm_cmplt_epi32 (__m128i __A
, __m128i __B
)
1273 return (__m128i
)__builtin_ia32_pcmpgtd128 ((__v4si
)__B
, (__v4si
)__A
);
1276 static __inline __m128i
__attribute__((__always_inline__
))
1277 _mm_cmpgt_epi8 (__m128i __A
, __m128i __B
)
1279 return (__m128i
)__builtin_ia32_pcmpgtb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1282 static __inline __m128i
__attribute__((__always_inline__
))
1283 _mm_cmpgt_epi16 (__m128i __A
, __m128i __B
)
1285 return (__m128i
)__builtin_ia32_pcmpgtw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1288 static __inline __m128i
__attribute__((__always_inline__
))
1289 _mm_cmpgt_epi32 (__m128i __A
, __m128i __B
)
1291 return (__m128i
)__builtin_ia32_pcmpgtd128 ((__v4si
)__A
, (__v4si
)__B
);
1295 static __inline
int __attribute__((__always_inline__
))
1296 _mm_extract_epi16 (__m128i
const __A
, int const __N
)
1298 return __builtin_ia32_vec_ext_v8hi ((__v8hi
)__A
, __N
);
1301 static __inline __m128i
__attribute__((__always_inline__
))
1302 _mm_insert_epi16 (__m128i
const __A
, int const __D
, int const __N
)
1304 return (__m128i
) __builtin_ia32_vec_set_v8hi ((__v8hi
)__A
, __D
, __N
);
1307 #define _mm_extract_epi16(A, N) \
1308 ((int) __builtin_ia32_vec_ext_v8hi ((__v8hi)(A), (N)))
1309 #define _mm_insert_epi16(A, D, N) \
1310 ((__m128i) __builtin_ia32_vec_set_v8hi ((__v8hi)(A), (D), (N)))
1313 static __inline __m128i
__attribute__((__always_inline__
))
1314 _mm_max_epi16 (__m128i __A
, __m128i __B
)
1316 return (__m128i
)__builtin_ia32_pmaxsw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1319 static __inline __m128i
__attribute__((__always_inline__
))
1320 _mm_max_epu8 (__m128i __A
, __m128i __B
)
1322 return (__m128i
)__builtin_ia32_pmaxub128 ((__v16qi
)__A
, (__v16qi
)__B
);
1325 static __inline __m128i
__attribute__((__always_inline__
))
1326 _mm_min_epi16 (__m128i __A
, __m128i __B
)
1328 return (__m128i
)__builtin_ia32_pminsw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1331 static __inline __m128i
__attribute__((__always_inline__
))
1332 _mm_min_epu8 (__m128i __A
, __m128i __B
)
1334 return (__m128i
)__builtin_ia32_pminub128 ((__v16qi
)__A
, (__v16qi
)__B
);
1337 static __inline
int __attribute__((__always_inline__
))
1338 _mm_movemask_epi8 (__m128i __A
)
1340 return __builtin_ia32_pmovmskb128 ((__v16qi
)__A
);
1343 static __inline __m128i
__attribute__((__always_inline__
))
1344 _mm_mulhi_epu16 (__m128i __A
, __m128i __B
)
1346 return (__m128i
)__builtin_ia32_pmulhuw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1349 #define _mm_shufflehi_epi16(__A, __B) ((__m128i)__builtin_ia32_pshufhw ((__v8hi)__A, __B))
1350 #define _mm_shufflelo_epi16(__A, __B) ((__m128i)__builtin_ia32_pshuflw ((__v8hi)__A, __B))
1351 #define _mm_shuffle_epi32(__A, __B) ((__m128i)__builtin_ia32_pshufd ((__v4si)__A, __B))
1353 static __inline
void __attribute__((__always_inline__
))
1354 _mm_maskmoveu_si128 (__m128i __A
, __m128i __B
, char *__C
)
1356 __builtin_ia32_maskmovdqu ((__v16qi
)__A
, (__v16qi
)__B
, __C
);
1359 static __inline __m128i
__attribute__((__always_inline__
))
1360 _mm_avg_epu8 (__m128i __A
, __m128i __B
)
1362 return (__m128i
)__builtin_ia32_pavgb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1365 static __inline __m128i
__attribute__((__always_inline__
))
1366 _mm_avg_epu16 (__m128i __A
, __m128i __B
)
1368 return (__m128i
)__builtin_ia32_pavgw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1371 static __inline __m128i
__attribute__((__always_inline__
))
1372 _mm_sad_epu8 (__m128i __A
, __m128i __B
)
1374 return (__m128i
)__builtin_ia32_psadbw128 ((__v16qi
)__A
, (__v16qi
)__B
);
1377 static __inline
void __attribute__((__always_inline__
))
1378 _mm_stream_si32 (int *__A
, int __B
)
1380 __builtin_ia32_movnti (__A
, __B
);
1383 static __inline
void __attribute__((__always_inline__
))
1384 _mm_stream_si128 (__m128i
*__A
, __m128i __B
)
1386 __builtin_ia32_movntdq ((__v2di
*)__A
, (__v2di
)__B
);
1389 static __inline
void __attribute__((__always_inline__
))
1390 _mm_stream_pd (double *__A
, __m128d __B
)
1392 __builtin_ia32_movntpd (__A
, (__v2df
)__B
);
1395 static __inline
void __attribute__((__always_inline__
))
1396 _mm_clflush (void const *__A
)
1398 __builtin_ia32_clflush (__A
);
1401 static __inline
void __attribute__((__always_inline__
))
1404 __builtin_ia32_lfence ();
1407 static __inline
void __attribute__((__always_inline__
))
1410 __builtin_ia32_mfence ();
1413 static __inline __m128i
__attribute__((__always_inline__
))
1414 _mm_cvtsi32_si128 (int __A
)
1416 return _mm_set_epi32 (0, 0, 0, __A
);
1420 /* Intel intrinsic. */
1421 static __inline __m128i
__attribute__((__always_inline__
))
1422 _mm_cvtsi64_si128 (long long __A
)
1424 return _mm_set_epi64x (0, __A
);
1427 /* Microsoft intrinsic. */
1428 static __inline __m128i
__attribute__((__always_inline__
))
1429 _mm_cvtsi64x_si128 (long long __A
)
1431 return _mm_set_epi64x (0, __A
);
1435 /* Casts between various SP, DP, INT vector types. Note that these do no
1436 conversion of values, they just change the type. */
1437 static __inline __m128
__attribute__((__always_inline__
))
1438 _mm_castpd_ps(__m128d __A
)
1440 return (__m128
) __A
;
1443 static __inline __m128i
__attribute__((__always_inline__
))
1444 _mm_castpd_si128(__m128d __A
)
1446 return (__m128i
) __A
;
1449 static __inline __m128d
__attribute__((__always_inline__
))
1450 _mm_castps_pd(__m128 __A
)
1452 return (__m128d
) __A
;
1455 static __inline __m128i
__attribute__((__always_inline__
))
1456 _mm_castps_si128(__m128 __A
)
1458 return (__m128i
) __A
;
1461 static __inline __m128
__attribute__((__always_inline__
))
1462 _mm_castsi128_ps(__m128i __A
)
1464 return (__m128
) __A
;
1467 static __inline __m128d
__attribute__((__always_inline__
))
1468 _mm_castsi128_pd(__m128i __A
)
1470 return (__m128d
) __A
;
1473 #endif /* __SSE2__ */
1475 #endif /* _EMMINTRIN_H_INCLUDED */
projects / gcc.git / blob