1 /* Copyright (C) 2003, 2004, 2005, 2006, 2007 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 # error "SSE2 instruction set not enabled"
37 /* We need definitions from the SSE header files*/
38 #include <xmmintrin.h>
41 typedef double __v2df
__attribute__ ((__vector_size__ (16)));
42 typedef long long __v2di
__attribute__ ((__vector_size__ (16)));
43 typedef int __v4si
__attribute__ ((__vector_size__ (16)));
44 typedef short __v8hi
__attribute__ ((__vector_size__ (16)));
45 typedef char __v16qi
__attribute__ ((__vector_size__ (16)));
47 typedef __v2di __m128i
;
48 typedef __v2df __m128d
;
50 /* Create a selector for use with the SHUFPD instruction. */
51 #define _MM_SHUFFLE2(fp1,fp0) \
52 (((fp1) << 1) | (fp0))
54 /* Create a vector with element 0 as F and the rest zero. */
55 static __inline __m128d
__attribute__((__always_inline__
))
56 _mm_set_sd (double __F
)
58 return __extension__ (__m128d
){ __F
, 0 };
61 /* Create a vector with both elements equal to F. */
62 static __inline __m128d
__attribute__((__always_inline__
))
63 _mm_set1_pd (double __F
)
65 return __extension__ (__m128d
){ __F
, __F
};
68 static __inline __m128d
__attribute__((__always_inline__
))
69 _mm_set_pd1 (double __F
)
71 return _mm_set1_pd (__F
);
74 /* Create a vector with the lower value X and upper value W. */
75 static __inline __m128d
__attribute__((__always_inline__
))
76 _mm_set_pd (double __W
, double __X
)
78 return __extension__ (__m128d
){ __X
, __W
};
81 /* Create a vector with the lower value W and upper value X. */
82 static __inline __m128d
__attribute__((__always_inline__
))
83 _mm_setr_pd (double __W
, double __X
)
85 return __extension__ (__m128d
){ __W
, __X
};
88 /* Create a vector of zeros. */
89 static __inline __m128d
__attribute__((__always_inline__
))
92 return __extension__ (__m128d
){ 0.0, 0.0 };
95 /* Sets the low DPFP value of A from the low value of B. */
96 static __inline __m128d
__attribute__((__always_inline__
))
97 _mm_move_sd (__m128d __A
, __m128d __B
)
99 return (__m128d
) __builtin_ia32_movsd ((__v2df
)__A
, (__v2df
)__B
);
102 /* Load two DPFP values from P. The address must be 16-byte aligned. */
103 static __inline __m128d
__attribute__((__always_inline__
))
104 _mm_load_pd (double const *__P
)
106 return *(__m128d
*)__P
;
109 /* Load two DPFP values from P. The address need not be 16-byte aligned. */
110 static __inline __m128d
__attribute__((__always_inline__
))
111 _mm_loadu_pd (double const *__P
)
113 return __builtin_ia32_loadupd (__P
);
116 /* Create a vector with all two elements equal to *P. */
117 static __inline __m128d
__attribute__((__always_inline__
))
118 _mm_load1_pd (double const *__P
)
120 return _mm_set1_pd (*__P
);
123 /* Create a vector with element 0 as *P and the rest zero. */
124 static __inline __m128d
__attribute__((__always_inline__
))
125 _mm_load_sd (double const *__P
)
127 return _mm_set_sd (*__P
);
130 static __inline __m128d
__attribute__((__always_inline__
))
131 _mm_load_pd1 (double const *__P
)
133 return _mm_load1_pd (__P
);
136 /* Load two DPFP values in reverse order. The address must be aligned. */
137 static __inline __m128d
__attribute__((__always_inline__
))
138 _mm_loadr_pd (double const *__P
)
140 __m128d __tmp
= _mm_load_pd (__P
);
141 return __builtin_ia32_shufpd (__tmp
, __tmp
, _MM_SHUFFLE2 (0,1));
144 /* Store two DPFP values. The address must be 16-byte aligned. */
145 static __inline
void __attribute__((__always_inline__
))
146 _mm_store_pd (double *__P
, __m128d __A
)
148 *(__m128d
*)__P
= __A
;
151 /* Store two DPFP values. The address need not be 16-byte aligned. */
152 static __inline
void __attribute__((__always_inline__
))
153 _mm_storeu_pd (double *__P
, __m128d __A
)
155 __builtin_ia32_storeupd (__P
, __A
);
158 /* Stores the lower DPFP value. */
159 static __inline
void __attribute__((__always_inline__
))
160 _mm_store_sd (double *__P
, __m128d __A
)
162 *__P
= __builtin_ia32_vec_ext_v2df (__A
, 0);
165 static __inline
double __attribute__((__always_inline__
))
166 _mm_cvtsd_f64 (__m128d __A
)
168 return __builtin_ia32_vec_ext_v2df (__A
, 0);
171 static __inline
void __attribute__((__always_inline__
))
172 _mm_storel_pd (double *__P
, __m128d __A
)
174 _mm_store_sd (__P
, __A
);
177 /* Stores the upper DPFP value. */
178 static __inline
void __attribute__((__always_inline__
))
179 _mm_storeh_pd (double *__P
, __m128d __A
)
181 *__P
= __builtin_ia32_vec_ext_v2df (__A
, 1);
184 /* Store the lower DPFP value across two words.
185 The address must be 16-byte aligned. */
186 static __inline
void __attribute__((__always_inline__
))
187 _mm_store1_pd (double *__P
, __m128d __A
)
189 _mm_store_pd (__P
, __builtin_ia32_shufpd (__A
, __A
, _MM_SHUFFLE2 (0,0)));
192 static __inline
void __attribute__((__always_inline__
))
193 _mm_store_pd1 (double *__P
, __m128d __A
)
195 _mm_store1_pd (__P
, __A
);
198 /* Store two DPFP values in reverse order. The address must be aligned. */
199 static __inline
void __attribute__((__always_inline__
))
200 _mm_storer_pd (double *__P
, __m128d __A
)
202 _mm_store_pd (__P
, __builtin_ia32_shufpd (__A
, __A
, _MM_SHUFFLE2 (0,1)));
205 static __inline
int __attribute__((__always_inline__
))
206 _mm_cvtsi128_si32 (__m128i __A
)
208 return __builtin_ia32_vec_ext_v4si ((__v4si
)__A
, 0);
212 /* Intel intrinsic. */
213 static __inline
long long __attribute__((__always_inline__
))
214 _mm_cvtsi128_si64 (__m128i __A
)
216 return __builtin_ia32_vec_ext_v2di ((__v2di
)__A
, 0);
219 /* Microsoft intrinsic. */
220 static __inline
long long __attribute__((__always_inline__
))
221 _mm_cvtsi128_si64x (__m128i __A
)
223 return __builtin_ia32_vec_ext_v2di ((__v2di
)__A
, 0);
227 static __inline __m128d
__attribute__((__always_inline__
))
228 _mm_add_pd (__m128d __A
, __m128d __B
)
230 return (__m128d
)__builtin_ia32_addpd ((__v2df
)__A
, (__v2df
)__B
);
233 static __inline __m128d
__attribute__((__always_inline__
))
234 _mm_add_sd (__m128d __A
, __m128d __B
)
236 return (__m128d
)__builtin_ia32_addsd ((__v2df
)__A
, (__v2df
)__B
);
239 static __inline __m128d
__attribute__((__always_inline__
))
240 _mm_sub_pd (__m128d __A
, __m128d __B
)
242 return (__m128d
)__builtin_ia32_subpd ((__v2df
)__A
, (__v2df
)__B
);
245 static __inline __m128d
__attribute__((__always_inline__
))
246 _mm_sub_sd (__m128d __A
, __m128d __B
)
248 return (__m128d
)__builtin_ia32_subsd ((__v2df
)__A
, (__v2df
)__B
);
251 static __inline __m128d
__attribute__((__always_inline__
))
252 _mm_mul_pd (__m128d __A
, __m128d __B
)
254 return (__m128d
)__builtin_ia32_mulpd ((__v2df
)__A
, (__v2df
)__B
);
257 static __inline __m128d
__attribute__((__always_inline__
))
258 _mm_mul_sd (__m128d __A
, __m128d __B
)
260 return (__m128d
)__builtin_ia32_mulsd ((__v2df
)__A
, (__v2df
)__B
);
263 static __inline __m128d
__attribute__((__always_inline__
))
264 _mm_div_pd (__m128d __A
, __m128d __B
)
266 return (__m128d
)__builtin_ia32_divpd ((__v2df
)__A
, (__v2df
)__B
);
269 static __inline __m128d
__attribute__((__always_inline__
))
270 _mm_div_sd (__m128d __A
, __m128d __B
)
272 return (__m128d
)__builtin_ia32_divsd ((__v2df
)__A
, (__v2df
)__B
);
275 static __inline __m128d
__attribute__((__always_inline__
))
276 _mm_sqrt_pd (__m128d __A
)
278 return (__m128d
)__builtin_ia32_sqrtpd ((__v2df
)__A
);
281 /* Return pair {sqrt (A[0), B[1]}. */
282 static __inline __m128d
__attribute__((__always_inline__
))
283 _mm_sqrt_sd (__m128d __A
, __m128d __B
)
285 __v2df __tmp
= __builtin_ia32_movsd ((__v2df
)__A
, (__v2df
)__B
);
286 return (__m128d
)__builtin_ia32_sqrtsd ((__v2df
)__tmp
);
289 static __inline __m128d
__attribute__((__always_inline__
))
290 _mm_min_pd (__m128d __A
, __m128d __B
)
292 return (__m128d
)__builtin_ia32_minpd ((__v2df
)__A
, (__v2df
)__B
);
295 static __inline __m128d
__attribute__((__always_inline__
))
296 _mm_min_sd (__m128d __A
, __m128d __B
)
298 return (__m128d
)__builtin_ia32_minsd ((__v2df
)__A
, (__v2df
)__B
);
301 static __inline __m128d
__attribute__((__always_inline__
))
302 _mm_max_pd (__m128d __A
, __m128d __B
)
304 return (__m128d
)__builtin_ia32_maxpd ((__v2df
)__A
, (__v2df
)__B
);
307 static __inline __m128d
__attribute__((__always_inline__
))
308 _mm_max_sd (__m128d __A
, __m128d __B
)
310 return (__m128d
)__builtin_ia32_maxsd ((__v2df
)__A
, (__v2df
)__B
);
313 static __inline __m128d
__attribute__((__always_inline__
))
314 _mm_and_pd (__m128d __A
, __m128d __B
)
316 return (__m128d
)__builtin_ia32_andpd ((__v2df
)__A
, (__v2df
)__B
);
319 static __inline __m128d
__attribute__((__always_inline__
))
320 _mm_andnot_pd (__m128d __A
, __m128d __B
)
322 return (__m128d
)__builtin_ia32_andnpd ((__v2df
)__A
, (__v2df
)__B
);
325 static __inline __m128d
__attribute__((__always_inline__
))
326 _mm_or_pd (__m128d __A
, __m128d __B
)
328 return (__m128d
)__builtin_ia32_orpd ((__v2df
)__A
, (__v2df
)__B
);
331 static __inline __m128d
__attribute__((__always_inline__
))
332 _mm_xor_pd (__m128d __A
, __m128d __B
)
334 return (__m128d
)__builtin_ia32_xorpd ((__v2df
)__A
, (__v2df
)__B
);
337 static __inline __m128d
__attribute__((__always_inline__
))
338 _mm_cmpeq_pd (__m128d __A
, __m128d __B
)
340 return (__m128d
)__builtin_ia32_cmpeqpd ((__v2df
)__A
, (__v2df
)__B
);
343 static __inline __m128d
__attribute__((__always_inline__
))
344 _mm_cmplt_pd (__m128d __A
, __m128d __B
)
346 return (__m128d
)__builtin_ia32_cmpltpd ((__v2df
)__A
, (__v2df
)__B
);
349 static __inline __m128d
__attribute__((__always_inline__
))
350 _mm_cmple_pd (__m128d __A
, __m128d __B
)
352 return (__m128d
)__builtin_ia32_cmplepd ((__v2df
)__A
, (__v2df
)__B
);
355 static __inline __m128d
__attribute__((__always_inline__
))
356 _mm_cmpgt_pd (__m128d __A
, __m128d __B
)
358 return (__m128d
)__builtin_ia32_cmpgtpd ((__v2df
)__A
, (__v2df
)__B
);
361 static __inline __m128d
__attribute__((__always_inline__
))
362 _mm_cmpge_pd (__m128d __A
, __m128d __B
)
364 return (__m128d
)__builtin_ia32_cmpgepd ((__v2df
)__A
, (__v2df
)__B
);
367 static __inline __m128d
__attribute__((__always_inline__
))
368 _mm_cmpneq_pd (__m128d __A
, __m128d __B
)
370 return (__m128d
)__builtin_ia32_cmpneqpd ((__v2df
)__A
, (__v2df
)__B
);
373 static __inline __m128d
__attribute__((__always_inline__
))
374 _mm_cmpnlt_pd (__m128d __A
, __m128d __B
)
376 return (__m128d
)__builtin_ia32_cmpnltpd ((__v2df
)__A
, (__v2df
)__B
);
379 static __inline __m128d
__attribute__((__always_inline__
))
380 _mm_cmpnle_pd (__m128d __A
, __m128d __B
)
382 return (__m128d
)__builtin_ia32_cmpnlepd ((__v2df
)__A
, (__v2df
)__B
);
385 static __inline __m128d
__attribute__((__always_inline__
))
386 _mm_cmpngt_pd (__m128d __A
, __m128d __B
)
388 return (__m128d
)__builtin_ia32_cmpngtpd ((__v2df
)__A
, (__v2df
)__B
);
391 static __inline __m128d
__attribute__((__always_inline__
))
392 _mm_cmpnge_pd (__m128d __A
, __m128d __B
)
394 return (__m128d
)__builtin_ia32_cmpngepd ((__v2df
)__A
, (__v2df
)__B
);
397 static __inline __m128d
__attribute__((__always_inline__
))
398 _mm_cmpord_pd (__m128d __A
, __m128d __B
)
400 return (__m128d
)__builtin_ia32_cmpordpd ((__v2df
)__A
, (__v2df
)__B
);
403 static __inline __m128d
__attribute__((__always_inline__
))
404 _mm_cmpunord_pd (__m128d __A
, __m128d __B
)
406 return (__m128d
)__builtin_ia32_cmpunordpd ((__v2df
)__A
, (__v2df
)__B
);
409 static __inline __m128d
__attribute__((__always_inline__
))
410 _mm_cmpeq_sd (__m128d __A
, __m128d __B
)
412 return (__m128d
)__builtin_ia32_cmpeqsd ((__v2df
)__A
, (__v2df
)__B
);
415 static __inline __m128d
__attribute__((__always_inline__
))
416 _mm_cmplt_sd (__m128d __A
, __m128d __B
)
418 return (__m128d
)__builtin_ia32_cmpltsd ((__v2df
)__A
, (__v2df
)__B
);
421 static __inline __m128d
__attribute__((__always_inline__
))
422 _mm_cmple_sd (__m128d __A
, __m128d __B
)
424 return (__m128d
)__builtin_ia32_cmplesd ((__v2df
)__A
, (__v2df
)__B
);
427 static __inline __m128d
__attribute__((__always_inline__
))
428 _mm_cmpgt_sd (__m128d __A
, __m128d __B
)
430 return (__m128d
) __builtin_ia32_movsd ((__v2df
) __A
,
432 __builtin_ia32_cmpltsd ((__v2df
) __B
,
437 static __inline __m128d
__attribute__((__always_inline__
))
438 _mm_cmpge_sd (__m128d __A
, __m128d __B
)
440 return (__m128d
) __builtin_ia32_movsd ((__v2df
) __A
,
442 __builtin_ia32_cmplesd ((__v2df
) __B
,
447 static __inline __m128d
__attribute__((__always_inline__
))
448 _mm_cmpneq_sd (__m128d __A
, __m128d __B
)
450 return (__m128d
)__builtin_ia32_cmpneqsd ((__v2df
)__A
, (__v2df
)__B
);
453 static __inline __m128d
__attribute__((__always_inline__
))
454 _mm_cmpnlt_sd (__m128d __A
, __m128d __B
)
456 return (__m128d
)__builtin_ia32_cmpnltsd ((__v2df
)__A
, (__v2df
)__B
);
459 static __inline __m128d
__attribute__((__always_inline__
))
460 _mm_cmpnle_sd (__m128d __A
, __m128d __B
)
462 return (__m128d
)__builtin_ia32_cmpnlesd ((__v2df
)__A
, (__v2df
)__B
);
465 static __inline __m128d
__attribute__((__always_inline__
))
466 _mm_cmpngt_sd (__m128d __A
, __m128d __B
)
468 return (__m128d
) __builtin_ia32_movsd ((__v2df
) __A
,
470 __builtin_ia32_cmpnltsd ((__v2df
) __B
,
475 static __inline __m128d
__attribute__((__always_inline__
))
476 _mm_cmpnge_sd (__m128d __A
, __m128d __B
)
478 return (__m128d
) __builtin_ia32_movsd ((__v2df
) __A
,
480 __builtin_ia32_cmpnlesd ((__v2df
) __B
,
485 static __inline __m128d
__attribute__((__always_inline__
))
486 _mm_cmpord_sd (__m128d __A
, __m128d __B
)
488 return (__m128d
)__builtin_ia32_cmpordsd ((__v2df
)__A
, (__v2df
)__B
);
491 static __inline __m128d
__attribute__((__always_inline__
))
492 _mm_cmpunord_sd (__m128d __A
, __m128d __B
)
494 return (__m128d
)__builtin_ia32_cmpunordsd ((__v2df
)__A
, (__v2df
)__B
);
497 static __inline
int __attribute__((__always_inline__
))
498 _mm_comieq_sd (__m128d __A
, __m128d __B
)
500 return __builtin_ia32_comisdeq ((__v2df
)__A
, (__v2df
)__B
);
503 static __inline
int __attribute__((__always_inline__
))
504 _mm_comilt_sd (__m128d __A
, __m128d __B
)
506 return __builtin_ia32_comisdlt ((__v2df
)__A
, (__v2df
)__B
);
509 static __inline
int __attribute__((__always_inline__
))
510 _mm_comile_sd (__m128d __A
, __m128d __B
)
512 return __builtin_ia32_comisdle ((__v2df
)__A
, (__v2df
)__B
);
515 static __inline
int __attribute__((__always_inline__
))
516 _mm_comigt_sd (__m128d __A
, __m128d __B
)
518 return __builtin_ia32_comisdgt ((__v2df
)__A
, (__v2df
)__B
);
521 static __inline
int __attribute__((__always_inline__
))
522 _mm_comige_sd (__m128d __A
, __m128d __B
)
524 return __builtin_ia32_comisdge ((__v2df
)__A
, (__v2df
)__B
);
527 static __inline
int __attribute__((__always_inline__
))
528 _mm_comineq_sd (__m128d __A
, __m128d __B
)
530 return __builtin_ia32_comisdneq ((__v2df
)__A
, (__v2df
)__B
);
533 static __inline
int __attribute__((__always_inline__
))
534 _mm_ucomieq_sd (__m128d __A
, __m128d __B
)
536 return __builtin_ia32_ucomisdeq ((__v2df
)__A
, (__v2df
)__B
);
539 static __inline
int __attribute__((__always_inline__
))
540 _mm_ucomilt_sd (__m128d __A
, __m128d __B
)
542 return __builtin_ia32_ucomisdlt ((__v2df
)__A
, (__v2df
)__B
);
545 static __inline
int __attribute__((__always_inline__
))
546 _mm_ucomile_sd (__m128d __A
, __m128d __B
)
548 return __builtin_ia32_ucomisdle ((__v2df
)__A
, (__v2df
)__B
);
551 static __inline
int __attribute__((__always_inline__
))
552 _mm_ucomigt_sd (__m128d __A
, __m128d __B
)
554 return __builtin_ia32_ucomisdgt ((__v2df
)__A
, (__v2df
)__B
);
557 static __inline
int __attribute__((__always_inline__
))
558 _mm_ucomige_sd (__m128d __A
, __m128d __B
)
560 return __builtin_ia32_ucomisdge ((__v2df
)__A
, (__v2df
)__B
);
563 static __inline
int __attribute__((__always_inline__
))
564 _mm_ucomineq_sd (__m128d __A
, __m128d __B
)
566 return __builtin_ia32_ucomisdneq ((__v2df
)__A
, (__v2df
)__B
);
569 /* Create a vector of Qi, where i is the element number. */
571 static __inline __m128i
__attribute__((__always_inline__
))
572 _mm_set_epi64x (long long __q1
, long long __q0
)
574 return __extension__ (__m128i
)(__v2di
){ __q0
, __q1
};
577 static __inline __m128i
__attribute__((__always_inline__
))
578 _mm_set_epi64 (__m64 __q1
, __m64 __q0
)
580 return _mm_set_epi64x ((long long)__q1
, (long long)__q0
);
583 static __inline __m128i
__attribute__((__always_inline__
))
584 _mm_set_epi32 (int __q3
, int __q2
, int __q1
, int __q0
)
586 return __extension__ (__m128i
)(__v4si
){ __q0
, __q1
, __q2
, __q3
};
589 static __inline __m128i
__attribute__((__always_inline__
))
590 _mm_set_epi16 (short __q7
, short __q6
, short __q5
, short __q4
,
591 short __q3
, short __q2
, short __q1
, short __q0
)
593 return __extension__ (__m128i
)(__v8hi
){
594 __q0
, __q1
, __q2
, __q3
, __q4
, __q5
, __q6
, __q7
};
597 static __inline __m128i
__attribute__((__always_inline__
))
598 _mm_set_epi8 (char __q15
, char __q14
, char __q13
, char __q12
,
599 char __q11
, char __q10
, char __q09
, char __q08
,
600 char __q07
, char __q06
, char __q05
, char __q04
,
601 char __q03
, char __q02
, char __q01
, char __q00
)
603 return __extension__ (__m128i
)(__v16qi
){
604 __q00
, __q01
, __q02
, __q03
, __q04
, __q05
, __q06
, __q07
,
605 __q08
, __q09
, __q10
, __q11
, __q12
, __q13
, __q14
, __q15
609 /* Set all of the elements of the vector to A. */
611 static __inline __m128i
__attribute__((__always_inline__
))
612 _mm_set1_epi64x (long long __A
)
614 return _mm_set_epi64x (__A
, __A
);
617 static __inline __m128i
__attribute__((__always_inline__
))
618 _mm_set1_epi64 (__m64 __A
)
620 return _mm_set_epi64 (__A
, __A
);
623 static __inline __m128i
__attribute__((__always_inline__
))
624 _mm_set1_epi32 (int __A
)
626 return _mm_set_epi32 (__A
, __A
, __A
, __A
);
629 static __inline __m128i
__attribute__((__always_inline__
))
630 _mm_set1_epi16 (short __A
)
632 return _mm_set_epi16 (__A
, __A
, __A
, __A
, __A
, __A
, __A
, __A
);
635 static __inline __m128i
__attribute__((__always_inline__
))
636 _mm_set1_epi8 (char __A
)
638 return _mm_set_epi8 (__A
, __A
, __A
, __A
, __A
, __A
, __A
, __A
,
639 __A
, __A
, __A
, __A
, __A
, __A
, __A
, __A
);
642 /* Create a vector of Qi, where i is the element number.
643 The parameter order is reversed from the _mm_set_epi* functions. */
645 static __inline __m128i
__attribute__((__always_inline__
))
646 _mm_setr_epi64 (__m64 __q0
, __m64 __q1
)
648 return _mm_set_epi64 (__q1
, __q0
);
651 static __inline __m128i
__attribute__((__always_inline__
))
652 _mm_setr_epi32 (int __q0
, int __q1
, int __q2
, int __q3
)
654 return _mm_set_epi32 (__q3
, __q2
, __q1
, __q0
);
657 static __inline __m128i
__attribute__((__always_inline__
))
658 _mm_setr_epi16 (short __q0
, short __q1
, short __q2
, short __q3
,
659 short __q4
, short __q5
, short __q6
, short __q7
)
661 return _mm_set_epi16 (__q7
, __q6
, __q5
, __q4
, __q3
, __q2
, __q1
, __q0
);
664 static __inline __m128i
__attribute__((__always_inline__
))
665 _mm_setr_epi8 (char __q00
, char __q01
, char __q02
, char __q03
,
666 char __q04
, char __q05
, char __q06
, char __q07
,
667 char __q08
, char __q09
, char __q10
, char __q11
,
668 char __q12
, char __q13
, char __q14
, char __q15
)
670 return _mm_set_epi8 (__q15
, __q14
, __q13
, __q12
, __q11
, __q10
, __q09
, __q08
,
671 __q07
, __q06
, __q05
, __q04
, __q03
, __q02
, __q01
, __q00
);
674 /* Create a vector with element 0 as *P and the rest zero. */
676 static __inline __m128i
__attribute__((__always_inline__
))
677 _mm_load_si128 (__m128i
const *__P
)
682 static __inline __m128i
__attribute__((__always_inline__
))
683 _mm_loadu_si128 (__m128i
const *__P
)
685 return (__m128i
) __builtin_ia32_loaddqu ((char const *)__P
);
688 static __inline __m128i
__attribute__((__always_inline__
))
689 _mm_loadl_epi64 (__m128i
const *__P
)
691 return _mm_set_epi64 ((__m64
)0LL, *(__m64
*)__P
);
694 static __inline
void __attribute__((__always_inline__
))
695 _mm_store_si128 (__m128i
*__P
, __m128i __B
)
700 static __inline
void __attribute__((__always_inline__
))
701 _mm_storeu_si128 (__m128i
*__P
, __m128i __B
)
703 __builtin_ia32_storedqu ((char *)__P
, (__v16qi
)__B
);
706 static __inline
void __attribute__((__always_inline__
))
707 _mm_storel_epi64 (__m128i
*__P
, __m128i __B
)
709 *(long long *)__P
= __builtin_ia32_vec_ext_v2di ((__v2di
)__B
, 0);
712 static __inline __m64
__attribute__((__always_inline__
))
713 _mm_movepi64_pi64 (__m128i __B
)
715 return (__m64
) __builtin_ia32_vec_ext_v2di ((__v2di
)__B
, 0);
718 static __inline __m128i
__attribute__((__always_inline__
))
719 _mm_movpi64_epi64 (__m64 __A
)
721 return _mm_set_epi64 ((__m64
)0LL, __A
);
724 static __inline __m128i
__attribute__((__always_inline__
))
725 _mm_move_epi64 (__m128i __A
)
727 return _mm_set_epi64 ((__m64
)0LL, _mm_movepi64_pi64 (__A
));
730 /* Create a vector of zeros. */
731 static __inline __m128i
__attribute__((__always_inline__
))
732 _mm_setzero_si128 (void)
734 return __extension__ (__m128i
)(__v4si
){ 0, 0, 0, 0 };
737 static __inline __m128d
__attribute__((__always_inline__
))
738 _mm_cvtepi32_pd (__m128i __A
)
740 return (__m128d
)__builtin_ia32_cvtdq2pd ((__v4si
) __A
);
743 static __inline __m128
__attribute__((__always_inline__
))
744 _mm_cvtepi32_ps (__m128i __A
)
746 return (__m128
)__builtin_ia32_cvtdq2ps ((__v4si
) __A
);
749 static __inline __m128i
__attribute__((__always_inline__
))
750 _mm_cvtpd_epi32 (__m128d __A
)
752 return (__m128i
)__builtin_ia32_cvtpd2dq ((__v2df
) __A
);
755 static __inline __m64
__attribute__((__always_inline__
))
756 _mm_cvtpd_pi32 (__m128d __A
)
758 return (__m64
)__builtin_ia32_cvtpd2pi ((__v2df
) __A
);
761 static __inline __m128
__attribute__((__always_inline__
))
762 _mm_cvtpd_ps (__m128d __A
)
764 return (__m128
)__builtin_ia32_cvtpd2ps ((__v2df
) __A
);
767 static __inline __m128i
__attribute__((__always_inline__
))
768 _mm_cvttpd_epi32 (__m128d __A
)
770 return (__m128i
)__builtin_ia32_cvttpd2dq ((__v2df
) __A
);
773 static __inline __m64
__attribute__((__always_inline__
))
774 _mm_cvttpd_pi32 (__m128d __A
)
776 return (__m64
)__builtin_ia32_cvttpd2pi ((__v2df
) __A
);
779 static __inline __m128d
__attribute__((__always_inline__
))
780 _mm_cvtpi32_pd (__m64 __A
)
782 return (__m128d
)__builtin_ia32_cvtpi2pd ((__v2si
) __A
);
785 static __inline __m128i
__attribute__((__always_inline__
))
786 _mm_cvtps_epi32 (__m128 __A
)
788 return (__m128i
)__builtin_ia32_cvtps2dq ((__v4sf
) __A
);
791 static __inline __m128i
__attribute__((__always_inline__
))
792 _mm_cvttps_epi32 (__m128 __A
)
794 return (__m128i
)__builtin_ia32_cvttps2dq ((__v4sf
) __A
);
797 static __inline __m128d
__attribute__((__always_inline__
))
798 _mm_cvtps_pd (__m128 __A
)
800 return (__m128d
)__builtin_ia32_cvtps2pd ((__v4sf
) __A
);
803 static __inline
int __attribute__((__always_inline__
))
804 _mm_cvtsd_si32 (__m128d __A
)
806 return __builtin_ia32_cvtsd2si ((__v2df
) __A
);
810 /* Intel intrinsic. */
811 static __inline
long long __attribute__((__always_inline__
))
812 _mm_cvtsd_si64 (__m128d __A
)
814 return __builtin_ia32_cvtsd2si64 ((__v2df
) __A
);
817 /* Microsoft intrinsic. */
818 static __inline
long long __attribute__((__always_inline__
))
819 _mm_cvtsd_si64x (__m128d __A
)
821 return __builtin_ia32_cvtsd2si64 ((__v2df
) __A
);
825 static __inline
int __attribute__((__always_inline__
))
826 _mm_cvttsd_si32 (__m128d __A
)
828 return __builtin_ia32_cvttsd2si ((__v2df
) __A
);
832 /* Intel intrinsic. */
833 static __inline
long long __attribute__((__always_inline__
))
834 _mm_cvttsd_si64 (__m128d __A
)
836 return __builtin_ia32_cvttsd2si64 ((__v2df
) __A
);
839 /* Microsoft intrinsic. */
840 static __inline
long long __attribute__((__always_inline__
))
841 _mm_cvttsd_si64x (__m128d __A
)
843 return __builtin_ia32_cvttsd2si64 ((__v2df
) __A
);
847 static __inline __m128
__attribute__((__always_inline__
))
848 _mm_cvtsd_ss (__m128 __A
, __m128d __B
)
850 return (__m128
)__builtin_ia32_cvtsd2ss ((__v4sf
) __A
, (__v2df
) __B
);
853 static __inline __m128d
__attribute__((__always_inline__
))
854 _mm_cvtsi32_sd (__m128d __A
, int __B
)
856 return (__m128d
)__builtin_ia32_cvtsi2sd ((__v2df
) __A
, __B
);
860 /* Intel intrinsic. */
861 static __inline __m128d
__attribute__((__always_inline__
))
862 _mm_cvtsi64_sd (__m128d __A
, long long __B
)
864 return (__m128d
)__builtin_ia32_cvtsi642sd ((__v2df
) __A
, __B
);
867 /* Microsoft intrinsic. */
868 static __inline __m128d
__attribute__((__always_inline__
))
869 _mm_cvtsi64x_sd (__m128d __A
, long long __B
)
871 return (__m128d
)__builtin_ia32_cvtsi642sd ((__v2df
) __A
, __B
);
875 static __inline __m128d
__attribute__((__always_inline__
))
876 _mm_cvtss_sd (__m128d __A
, __m128 __B
)
878 return (__m128d
)__builtin_ia32_cvtss2sd ((__v2df
) __A
, (__v4sf
)__B
);
881 #define _mm_shuffle_pd(__A, __B, __C) ((__m128d)__builtin_ia32_shufpd ((__v2df)__A, (__v2df)__B, (__C)))
883 static __inline __m128d
__attribute__((__always_inline__
))
884 _mm_unpackhi_pd (__m128d __A
, __m128d __B
)
886 return (__m128d
)__builtin_ia32_unpckhpd ((__v2df
)__A
, (__v2df
)__B
);
889 static __inline __m128d
__attribute__((__always_inline__
))
890 _mm_unpacklo_pd (__m128d __A
, __m128d __B
)
892 return (__m128d
)__builtin_ia32_unpcklpd ((__v2df
)__A
, (__v2df
)__B
);
895 static __inline __m128d
__attribute__((__always_inline__
))
896 _mm_loadh_pd (__m128d __A
, double const *__B
)
898 return (__m128d
)__builtin_ia32_loadhpd ((__v2df
)__A
, __B
);
901 static __inline __m128d
__attribute__((__always_inline__
))
902 _mm_loadl_pd (__m128d __A
, double const *__B
)
904 return (__m128d
)__builtin_ia32_loadlpd ((__v2df
)__A
, __B
);
907 static __inline
int __attribute__((__always_inline__
))
908 _mm_movemask_pd (__m128d __A
)
910 return __builtin_ia32_movmskpd ((__v2df
)__A
);
913 static __inline __m128i
__attribute__((__always_inline__
))
914 _mm_packs_epi16 (__m128i __A
, __m128i __B
)
916 return (__m128i
)__builtin_ia32_packsswb128 ((__v8hi
)__A
, (__v8hi
)__B
);
919 static __inline __m128i
__attribute__((__always_inline__
))
920 _mm_packs_epi32 (__m128i __A
, __m128i __B
)
922 return (__m128i
)__builtin_ia32_packssdw128 ((__v4si
)__A
, (__v4si
)__B
);
925 static __inline __m128i
__attribute__((__always_inline__
))
926 _mm_packus_epi16 (__m128i __A
, __m128i __B
)
928 return (__m128i
)__builtin_ia32_packuswb128 ((__v8hi
)__A
, (__v8hi
)__B
);
931 static __inline __m128i
__attribute__((__always_inline__
))
932 _mm_unpackhi_epi8 (__m128i __A
, __m128i __B
)
934 return (__m128i
)__builtin_ia32_punpckhbw128 ((__v16qi
)__A
, (__v16qi
)__B
);
937 static __inline __m128i
__attribute__((__always_inline__
))
938 _mm_unpackhi_epi16 (__m128i __A
, __m128i __B
)
940 return (__m128i
)__builtin_ia32_punpckhwd128 ((__v8hi
)__A
, (__v8hi
)__B
);
943 static __inline __m128i
__attribute__((__always_inline__
))
944 _mm_unpackhi_epi32 (__m128i __A
, __m128i __B
)
946 return (__m128i
)__builtin_ia32_punpckhdq128 ((__v4si
)__A
, (__v4si
)__B
);
949 static __inline __m128i
__attribute__((__always_inline__
))
950 _mm_unpackhi_epi64 (__m128i __A
, __m128i __B
)
952 return (__m128i
)__builtin_ia32_punpckhqdq128 ((__v2di
)__A
, (__v2di
)__B
);
955 static __inline __m128i
__attribute__((__always_inline__
))
956 _mm_unpacklo_epi8 (__m128i __A
, __m128i __B
)
958 return (__m128i
)__builtin_ia32_punpcklbw128 ((__v16qi
)__A
, (__v16qi
)__B
);
961 static __inline __m128i
__attribute__((__always_inline__
))
962 _mm_unpacklo_epi16 (__m128i __A
, __m128i __B
)
964 return (__m128i
)__builtin_ia32_punpcklwd128 ((__v8hi
)__A
, (__v8hi
)__B
);
967 static __inline __m128i
__attribute__((__always_inline__
))
968 _mm_unpacklo_epi32 (__m128i __A
, __m128i __B
)
970 return (__m128i
)__builtin_ia32_punpckldq128 ((__v4si
)__A
, (__v4si
)__B
);
973 static __inline __m128i
__attribute__((__always_inline__
))
974 _mm_unpacklo_epi64 (__m128i __A
, __m128i __B
)
976 return (__m128i
)__builtin_ia32_punpcklqdq128 ((__v2di
)__A
, (__v2di
)__B
);
979 static __inline __m128i
__attribute__((__always_inline__
))
980 _mm_add_epi8 (__m128i __A
, __m128i __B
)
982 return (__m128i
)__builtin_ia32_paddb128 ((__v16qi
)__A
, (__v16qi
)__B
);
985 static __inline __m128i
__attribute__((__always_inline__
))
986 _mm_add_epi16 (__m128i __A
, __m128i __B
)
988 return (__m128i
)__builtin_ia32_paddw128 ((__v8hi
)__A
, (__v8hi
)__B
);
991 static __inline __m128i
__attribute__((__always_inline__
))
992 _mm_add_epi32 (__m128i __A
, __m128i __B
)
994 return (__m128i
)__builtin_ia32_paddd128 ((__v4si
)__A
, (__v4si
)__B
);
997 static __inline __m128i
__attribute__((__always_inline__
))
998 _mm_add_epi64 (__m128i __A
, __m128i __B
)
1000 return (__m128i
)__builtin_ia32_paddq128 ((__v2di
)__A
, (__v2di
)__B
);
1003 static __inline __m128i
__attribute__((__always_inline__
))
1004 _mm_adds_epi8 (__m128i __A
, __m128i __B
)
1006 return (__m128i
)__builtin_ia32_paddsb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1009 static __inline __m128i
__attribute__((__always_inline__
))
1010 _mm_adds_epi16 (__m128i __A
, __m128i __B
)
1012 return (__m128i
)__builtin_ia32_paddsw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1015 static __inline __m128i
__attribute__((__always_inline__
))
1016 _mm_adds_epu8 (__m128i __A
, __m128i __B
)
1018 return (__m128i
)__builtin_ia32_paddusb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1021 static __inline __m128i
__attribute__((__always_inline__
))
1022 _mm_adds_epu16 (__m128i __A
, __m128i __B
)
1024 return (__m128i
)__builtin_ia32_paddusw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1027 static __inline __m128i
__attribute__((__always_inline__
))
1028 _mm_sub_epi8 (__m128i __A
, __m128i __B
)
1030 return (__m128i
)__builtin_ia32_psubb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1033 static __inline __m128i
__attribute__((__always_inline__
))
1034 _mm_sub_epi16 (__m128i __A
, __m128i __B
)
1036 return (__m128i
)__builtin_ia32_psubw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1039 static __inline __m128i
__attribute__((__always_inline__
))
1040 _mm_sub_epi32 (__m128i __A
, __m128i __B
)
1042 return (__m128i
)__builtin_ia32_psubd128 ((__v4si
)__A
, (__v4si
)__B
);
1045 static __inline __m128i
__attribute__((__always_inline__
))
1046 _mm_sub_epi64 (__m128i __A
, __m128i __B
)
1048 return (__m128i
)__builtin_ia32_psubq128 ((__v2di
)__A
, (__v2di
)__B
);
1051 static __inline __m128i
__attribute__((__always_inline__
))
1052 _mm_subs_epi8 (__m128i __A
, __m128i __B
)
1054 return (__m128i
)__builtin_ia32_psubsb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1057 static __inline __m128i
__attribute__((__always_inline__
))
1058 _mm_subs_epi16 (__m128i __A
, __m128i __B
)
1060 return (__m128i
)__builtin_ia32_psubsw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1063 static __inline __m128i
__attribute__((__always_inline__
))
1064 _mm_subs_epu8 (__m128i __A
, __m128i __B
)
1066 return (__m128i
)__builtin_ia32_psubusb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1069 static __inline __m128i
__attribute__((__always_inline__
))
1070 _mm_subs_epu16 (__m128i __A
, __m128i __B
)
1072 return (__m128i
)__builtin_ia32_psubusw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1075 static __inline __m128i
__attribute__((__always_inline__
))
1076 _mm_madd_epi16 (__m128i __A
, __m128i __B
)
1078 return (__m128i
)__builtin_ia32_pmaddwd128 ((__v8hi
)__A
, (__v8hi
)__B
);
1081 static __inline __m128i
__attribute__((__always_inline__
))
1082 _mm_mulhi_epi16 (__m128i __A
, __m128i __B
)
1084 return (__m128i
)__builtin_ia32_pmulhw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1087 static __inline __m128i
__attribute__((__always_inline__
))
1088 _mm_mullo_epi16 (__m128i __A
, __m128i __B
)
1090 return (__m128i
)__builtin_ia32_pmullw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1093 static __inline __m64
__attribute__((__always_inline__
))
1094 _mm_mul_su32 (__m64 __A
, __m64 __B
)
1096 return (__m64
)__builtin_ia32_pmuludq ((__v2si
)__A
, (__v2si
)__B
);
1099 static __inline __m128i
__attribute__((__always_inline__
))
1100 _mm_mul_epu32 (__m128i __A
, __m128i __B
)
1102 return (__m128i
)__builtin_ia32_pmuludq128 ((__v4si
)__A
, (__v4si
)__B
);
1105 static __inline __m128i
__attribute__((__always_inline__
))
1106 _mm_slli_epi16 (__m128i __A
, int __B
)
1108 return (__m128i
)__builtin_ia32_psllwi128 ((__v8hi
)__A
, __B
);
1111 static __inline __m128i
__attribute__((__always_inline__
))
1112 _mm_slli_epi32 (__m128i __A
, int __B
)
1114 return (__m128i
)__builtin_ia32_pslldi128 ((__v4si
)__A
, __B
);
1117 static __inline __m128i
__attribute__((__always_inline__
))
1118 _mm_slli_epi64 (__m128i __A
, int __B
)
1120 return (__m128i
)__builtin_ia32_psllqi128 ((__v2di
)__A
, __B
);
1123 static __inline __m128i
__attribute__((__always_inline__
))
1124 _mm_srai_epi16 (__m128i __A
, int __B
)
1126 return (__m128i
)__builtin_ia32_psrawi128 ((__v8hi
)__A
, __B
);
1129 static __inline __m128i
__attribute__((__always_inline__
))
1130 _mm_srai_epi32 (__m128i __A
, int __B
)
1132 return (__m128i
)__builtin_ia32_psradi128 ((__v4si
)__A
, __B
);
1136 static __m128i
__attribute__((__always_inline__
))
1137 _mm_srli_si128 (__m128i __A
, const int __B
)
1139 return ((__m128i
)__builtin_ia32_psrldqi128 (__A
, __B
))
1142 static __m128i
__attribute__((__always_inline__
))
1143 _mm_srli_si128 (__m128i __A
, const int __B
)
1145 return ((__m128i
)__builtin_ia32_pslldqi128 (__A
, __B
))
1148 #define _mm_srli_si128(__A, __B) \
1149 ((__m128i)__builtin_ia32_psrldqi128 (__A, (__B) * 8))
1150 #define _mm_slli_si128(__A, __B) \
1151 ((__m128i)__builtin_ia32_pslldqi128 (__A, (__B) * 8))
1154 static __inline __m128i
__attribute__((__always_inline__
))
1155 _mm_srli_epi16 (__m128i __A
, int __B
)
1157 return (__m128i
)__builtin_ia32_psrlwi128 ((__v8hi
)__A
, __B
);
1160 static __inline __m128i
__attribute__((__always_inline__
))
1161 _mm_srli_epi32 (__m128i __A
, int __B
)
1163 return (__m128i
)__builtin_ia32_psrldi128 ((__v4si
)__A
, __B
);
1166 static __inline __m128i
__attribute__((__always_inline__
))
1167 _mm_srli_epi64 (__m128i __A
, int __B
)
1169 return (__m128i
)__builtin_ia32_psrlqi128 ((__v2di
)__A
, __B
);
1172 static __inline __m128i
__attribute__((__always_inline__
))
1173 _mm_sll_epi16 (__m128i __A
, __m128i __B
)
1175 return _mm_slli_epi16 (__A
, _mm_cvtsi128_si32 (__B
));
1178 static __inline __m128i
__attribute__((__always_inline__
))
1179 _mm_sll_epi32 (__m128i __A
, __m128i __B
)
1181 return _mm_slli_epi32 (__A
, _mm_cvtsi128_si32 (__B
));
1184 static __inline __m128i
__attribute__((__always_inline__
))
1185 _mm_sll_epi64 (__m128i __A
, __m128i __B
)
1187 return _mm_slli_epi64 (__A
, _mm_cvtsi128_si32 (__B
));
1190 static __inline __m128i
__attribute__((__always_inline__
))
1191 _mm_sra_epi16 (__m128i __A
, __m128i __B
)
1193 return _mm_srai_epi16 (__A
, _mm_cvtsi128_si32 (__B
));
1196 static __inline __m128i
__attribute__((__always_inline__
))
1197 _mm_sra_epi32 (__m128i __A
, __m128i __B
)
1199 return _mm_srai_epi32 (__A
, _mm_cvtsi128_si32 (__B
));
1202 static __inline __m128i
__attribute__((__always_inline__
))
1203 _mm_srl_epi16 (__m128i __A
, __m128i __B
)
1205 return _mm_srli_epi16 (__A
, _mm_cvtsi128_si32 (__B
));
1208 static __inline __m128i
__attribute__((__always_inline__
))
1209 _mm_srl_epi32 (__m128i __A
, __m128i __B
)
1211 return _mm_srli_epi32 (__A
, _mm_cvtsi128_si32 (__B
));
1214 static __inline __m128i
__attribute__((__always_inline__
))
1215 _mm_srl_epi64 (__m128i __A
, __m128i __B
)
1217 return _mm_srli_epi64 (__A
, _mm_cvtsi128_si32 (__B
));
1220 static __inline __m128i
__attribute__((__always_inline__
))
1221 _mm_and_si128 (__m128i __A
, __m128i __B
)
1223 return (__m128i
)__builtin_ia32_pand128 ((__v2di
)__A
, (__v2di
)__B
);
1226 static __inline __m128i
__attribute__((__always_inline__
))
1227 _mm_andnot_si128 (__m128i __A
, __m128i __B
)
1229 return (__m128i
)__builtin_ia32_pandn128 ((__v2di
)__A
, (__v2di
)__B
);
1232 static __inline __m128i
__attribute__((__always_inline__
))
1233 _mm_or_si128 (__m128i __A
, __m128i __B
)
1235 return (__m128i
)__builtin_ia32_por128 ((__v2di
)__A
, (__v2di
)__B
);
1238 static __inline __m128i
__attribute__((__always_inline__
))
1239 _mm_xor_si128 (__m128i __A
, __m128i __B
)
1241 return (__m128i
)__builtin_ia32_pxor128 ((__v2di
)__A
, (__v2di
)__B
);
1244 static __inline __m128i
__attribute__((__always_inline__
))
1245 _mm_cmpeq_epi8 (__m128i __A
, __m128i __B
)
1247 return (__m128i
)__builtin_ia32_pcmpeqb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1250 static __inline __m128i
__attribute__((__always_inline__
))
1251 _mm_cmpeq_epi16 (__m128i __A
, __m128i __B
)
1253 return (__m128i
)__builtin_ia32_pcmpeqw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1256 static __inline __m128i
__attribute__((__always_inline__
))
1257 _mm_cmpeq_epi32 (__m128i __A
, __m128i __B
)
1259 return (__m128i
)__builtin_ia32_pcmpeqd128 ((__v4si
)__A
, (__v4si
)__B
);
1262 static __inline __m128i
__attribute__((__always_inline__
))
1263 _mm_cmplt_epi8 (__m128i __A
, __m128i __B
)
1265 return (__m128i
)__builtin_ia32_pcmpgtb128 ((__v16qi
)__B
, (__v16qi
)__A
);
1268 static __inline __m128i
__attribute__((__always_inline__
))
1269 _mm_cmplt_epi16 (__m128i __A
, __m128i __B
)
1271 return (__m128i
)__builtin_ia32_pcmpgtw128 ((__v8hi
)__B
, (__v8hi
)__A
);
1274 static __inline __m128i
__attribute__((__always_inline__
))
1275 _mm_cmplt_epi32 (__m128i __A
, __m128i __B
)
1277 return (__m128i
)__builtin_ia32_pcmpgtd128 ((__v4si
)__B
, (__v4si
)__A
);
1280 static __inline __m128i
__attribute__((__always_inline__
))
1281 _mm_cmpgt_epi8 (__m128i __A
, __m128i __B
)
1283 return (__m128i
)__builtin_ia32_pcmpgtb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1286 static __inline __m128i
__attribute__((__always_inline__
))
1287 _mm_cmpgt_epi16 (__m128i __A
, __m128i __B
)
1289 return (__m128i
)__builtin_ia32_pcmpgtw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1292 static __inline __m128i
__attribute__((__always_inline__
))
1293 _mm_cmpgt_epi32 (__m128i __A
, __m128i __B
)
1295 return (__m128i
)__builtin_ia32_pcmpgtd128 ((__v4si
)__A
, (__v4si
)__B
);
1299 static __inline
int __attribute__((__always_inline__
))
1300 _mm_extract_epi16 (__m128i
const __A
, int const __N
)
1302 return __builtin_ia32_vec_ext_v8hi ((__v8hi
)__A
, __N
);
1305 static __inline __m128i
__attribute__((__always_inline__
))
1306 _mm_insert_epi16 (__m128i
const __A
, int const __D
, int const __N
)
1308 return (__m128i
) __builtin_ia32_vec_set_v8hi ((__v8hi
)__A
, __D
, __N
);
1311 #define _mm_extract_epi16(A, N) \
1312 ((int) __builtin_ia32_vec_ext_v8hi ((__v8hi)(A), (N)))
1313 #define _mm_insert_epi16(A, D, N) \
1314 ((__m128i) __builtin_ia32_vec_set_v8hi ((__v8hi)(A), (D), (N)))
1317 static __inline __m128i
__attribute__((__always_inline__
))
1318 _mm_max_epi16 (__m128i __A
, __m128i __B
)
1320 return (__m128i
)__builtin_ia32_pmaxsw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1323 static __inline __m128i
__attribute__((__always_inline__
))
1324 _mm_max_epu8 (__m128i __A
, __m128i __B
)
1326 return (__m128i
)__builtin_ia32_pmaxub128 ((__v16qi
)__A
, (__v16qi
)__B
);
1329 static __inline __m128i
__attribute__((__always_inline__
))
1330 _mm_min_epi16 (__m128i __A
, __m128i __B
)
1332 return (__m128i
)__builtin_ia32_pminsw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1335 static __inline __m128i
__attribute__((__always_inline__
))
1336 _mm_min_epu8 (__m128i __A
, __m128i __B
)
1338 return (__m128i
)__builtin_ia32_pminub128 ((__v16qi
)__A
, (__v16qi
)__B
);
1341 static __inline
int __attribute__((__always_inline__
))
1342 _mm_movemask_epi8 (__m128i __A
)
1344 return __builtin_ia32_pmovmskb128 ((__v16qi
)__A
);
1347 static __inline __m128i
__attribute__((__always_inline__
))
1348 _mm_mulhi_epu16 (__m128i __A
, __m128i __B
)
1350 return (__m128i
)__builtin_ia32_pmulhuw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1353 #define _mm_shufflehi_epi16(__A, __B) ((__m128i)__builtin_ia32_pshufhw ((__v8hi)__A, __B))
1354 #define _mm_shufflelo_epi16(__A, __B) ((__m128i)__builtin_ia32_pshuflw ((__v8hi)__A, __B))
1355 #define _mm_shuffle_epi32(__A, __B) ((__m128i)__builtin_ia32_pshufd ((__v4si)__A, __B))
1357 static __inline
void __attribute__((__always_inline__
))
1358 _mm_maskmoveu_si128 (__m128i __A
, __m128i __B
, char *__C
)
1360 __builtin_ia32_maskmovdqu ((__v16qi
)__A
, (__v16qi
)__B
, __C
);
1363 static __inline __m128i
__attribute__((__always_inline__
))
1364 _mm_avg_epu8 (__m128i __A
, __m128i __B
)
1366 return (__m128i
)__builtin_ia32_pavgb128 ((__v16qi
)__A
, (__v16qi
)__B
);
1369 static __inline __m128i
__attribute__((__always_inline__
))
1370 _mm_avg_epu16 (__m128i __A
, __m128i __B
)
1372 return (__m128i
)__builtin_ia32_pavgw128 ((__v8hi
)__A
, (__v8hi
)__B
);
1375 static __inline __m128i
__attribute__((__always_inline__
))
1376 _mm_sad_epu8 (__m128i __A
, __m128i __B
)
1378 return (__m128i
)__builtin_ia32_psadbw128 ((__v16qi
)__A
, (__v16qi
)__B
);
1381 static __inline
void __attribute__((__always_inline__
))
1382 _mm_stream_si32 (int *__A
, int __B
)
1384 __builtin_ia32_movnti (__A
, __B
);
1387 static __inline
void __attribute__((__always_inline__
))
1388 _mm_stream_si128 (__m128i
*__A
, __m128i __B
)
1390 __builtin_ia32_movntdq ((__v2di
*)__A
, (__v2di
)__B
);
1393 static __inline
void __attribute__((__always_inline__
))
1394 _mm_stream_pd (double *__A
, __m128d __B
)
1396 __builtin_ia32_movntpd (__A
, (__v2df
)__B
);
1399 static __inline
void __attribute__((__always_inline__
))
1400 _mm_clflush (void const *__A
)
1402 __builtin_ia32_clflush (__A
);
1405 static __inline
void __attribute__((__always_inline__
))
1408 __builtin_ia32_lfence ();
1411 static __inline
void __attribute__((__always_inline__
))
1414 __builtin_ia32_mfence ();
1417 static __inline __m128i
__attribute__((__always_inline__
))
1418 _mm_cvtsi32_si128 (int __A
)
1420 return _mm_set_epi32 (0, 0, 0, __A
);
1424 /* Intel intrinsic. */
1425 static __inline __m128i
__attribute__((__always_inline__
))
1426 _mm_cvtsi64_si128 (long long __A
)
1428 return _mm_set_epi64x (0, __A
);
1431 /* Microsoft intrinsic. */
1432 static __inline __m128i
__attribute__((__always_inline__
))
1433 _mm_cvtsi64x_si128 (long long __A
)
1435 return _mm_set_epi64x (0, __A
);
1439 /* Casts between various SP, DP, INT vector types. Note that these do no
1440 conversion of values, they just change the type. */
1441 static __inline __m128
__attribute__((__always_inline__
))
1442 _mm_castpd_ps(__m128d __A
)
1444 return (__m128
) __A
;
1447 static __inline __m128i
__attribute__((__always_inline__
))
1448 _mm_castpd_si128(__m128d __A
)
1450 return (__m128i
) __A
;
1453 static __inline __m128d
__attribute__((__always_inline__
))
1454 _mm_castps_pd(__m128 __A
)
1456 return (__m128d
) __A
;
1459 static __inline __m128i
__attribute__((__always_inline__
))
1460 _mm_castps_si128(__m128 __A
)
1462 return (__m128i
) __A
;
1465 static __inline __m128
__attribute__((__always_inline__
))
1466 _mm_castsi128_ps(__m128i __A
)
1468 return (__m128
) __A
;
1471 static __inline __m128d
__attribute__((__always_inline__
))
1472 _mm_castsi128_pd(__m128i __A
)
1474 return (__m128d
) __A
;
1477 #endif /* __SSE2__ */
1479 #endif /* _EMMINTRIN_H_INCLUDED */
