1/*===---- emmintrin.h - SSE2 intrinsics ------------------------------------===
2 *
3 * Permission is hereby granted, free of charge, to any person obtaining a copy
4 * of this software and associated documentation files (the "Software"), to deal
5 * in the Software without restriction, including without limitation the rights
6 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
7 * copies of the Software, and to permit persons to whom the Software is
8 * furnished to do so, subject to the following conditions:
9 *
10 * The above copyright notice and this permission notice shall be included in
11 * all copies or substantial portions of the Software.
12 *
13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
16 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
17 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
18 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
19 * THE SOFTWARE.
20 *
21 *===-----------------------------------------------------------------------===
22 */
23
24#ifndef __EMMINTRIN_H
25#define __EMMINTRIN_H
26
27#ifndef __SSE2__
28#error "SSE2 instruction set not enabled"
29#else
30
31#include <xmmintrin.h>
32
33typedef double __m128d __attribute__((__vector_size__(16)));
34typedef long long __m128i __attribute__((__vector_size__(16)));
35
36/* Type defines.  */
37typedef double __v2df __attribute__ ((__vector_size__ (16)));
38typedef long long __v2di __attribute__ ((__vector_size__ (16)));
39typedef short __v8hi __attribute__((__vector_size__(16)));
40typedef char __v16qi __attribute__((__vector_size__(16)));
41
42static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
43_mm_add_sd(__m128d __a, __m128d __b)
44{
45  __a[0] += __b[0];
46  return __a;
47}
48
49static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
50_mm_add_pd(__m128d __a, __m128d __b)
51{
52  return __a + __b;
53}
54
55static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
56_mm_sub_sd(__m128d __a, __m128d __b)
57{
58  __a[0] -= __b[0];
59  return __a;
60}
61
62static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
63_mm_sub_pd(__m128d __a, __m128d __b)
64{
65  return __a - __b;
66}
67
68static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
69_mm_mul_sd(__m128d __a, __m128d __b)
70{
71  __a[0] *= __b[0];
72  return __a;
73}
74
75static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
76_mm_mul_pd(__m128d __a, __m128d __b)
77{
78  return __a * __b;
79}
80
81static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
82_mm_div_sd(__m128d __a, __m128d __b)
83{
84  __a[0] /= __b[0];
85  return __a;
86}
87
88static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
89_mm_div_pd(__m128d __a, __m128d __b)
90{
91  return __a / __b;
92}
93
94static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
95_mm_sqrt_sd(__m128d __a, __m128d __b)
96{
97  __m128d __c = __builtin_ia32_sqrtsd(__b);
98  return (__m128d) { __c[0], __a[1] };
99}
100
101static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
102_mm_sqrt_pd(__m128d __a)
103{
104  return __builtin_ia32_sqrtpd(__a);
105}
106
107static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
108_mm_min_sd(__m128d __a, __m128d __b)
109{
110  return __builtin_ia32_minsd(__a, __b);
111}
112
113static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
114_mm_min_pd(__m128d __a, __m128d __b)
115{
116  return __builtin_ia32_minpd(__a, __b);
117}
118
119static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
120_mm_max_sd(__m128d __a, __m128d __b)
121{
122  return __builtin_ia32_maxsd(__a, __b);
123}
124
125static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
126_mm_max_pd(__m128d __a, __m128d __b)
127{
128  return __builtin_ia32_maxpd(__a, __b);
129}
130
131static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
132_mm_and_pd(__m128d __a, __m128d __b)
133{
134  return (__m128d)((__v4si)__a & (__v4si)__b);
135}
136
137static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
138_mm_andnot_pd(__m128d __a, __m128d __b)
139{
140  return (__m128d)(~(__v4si)__a & (__v4si)__b);
141}
142
143static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
144_mm_or_pd(__m128d __a, __m128d __b)
145{
146  return (__m128d)((__v4si)__a | (__v4si)__b);
147}
148
149static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
150_mm_xor_pd(__m128d __a, __m128d __b)
151{
152  return (__m128d)((__v4si)__a ^ (__v4si)__b);
153}
154
155static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
156_mm_cmpeq_pd(__m128d __a, __m128d __b)
157{
158  return (__m128d)__builtin_ia32_cmppd(__a, __b, 0);
159}
160
161static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
162_mm_cmplt_pd(__m128d __a, __m128d __b)
163{
164  return (__m128d)__builtin_ia32_cmppd(__a, __b, 1);
165}
166
167static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
168_mm_cmple_pd(__m128d __a, __m128d __b)
169{
170  return (__m128d)__builtin_ia32_cmppd(__a, __b, 2);
171}
172
173static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
174_mm_cmpgt_pd(__m128d __a, __m128d __b)
175{
176  return (__m128d)__builtin_ia32_cmppd(__b, __a, 1);
177}
178
179static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
180_mm_cmpge_pd(__m128d __a, __m128d __b)
181{
182  return (__m128d)__builtin_ia32_cmppd(__b, __a, 2);
183}
184
185static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
186_mm_cmpord_pd(__m128d __a, __m128d __b)
187{
188  return (__m128d)__builtin_ia32_cmppd(__a, __b, 7);
189}
190
191static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
192_mm_cmpunord_pd(__m128d __a, __m128d __b)
193{
194  return (__m128d)__builtin_ia32_cmppd(__a, __b, 3);
195}
196
197static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
198_mm_cmpneq_pd(__m128d __a, __m128d __b)
199{
200  return (__m128d)__builtin_ia32_cmppd(__a, __b, 4);
201}
202
203static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
204_mm_cmpnlt_pd(__m128d __a, __m128d __b)
205{
206  return (__m128d)__builtin_ia32_cmppd(__a, __b, 5);
207}
208
209static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
210_mm_cmpnle_pd(__m128d __a, __m128d __b)
211{
212  return (__m128d)__builtin_ia32_cmppd(__a, __b, 6);
213}
214
215static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
216_mm_cmpngt_pd(__m128d __a, __m128d __b)
217{
218  return (__m128d)__builtin_ia32_cmppd(__b, __a, 5);
219}
220
221static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
222_mm_cmpnge_pd(__m128d __a, __m128d __b)
223{
224  return (__m128d)__builtin_ia32_cmppd(__b, __a, 6);
225}
226
227static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
228_mm_cmpeq_sd(__m128d __a, __m128d __b)
229{
230  return (__m128d)__builtin_ia32_cmpsd(__a, __b, 0);
231}
232
233static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
234_mm_cmplt_sd(__m128d __a, __m128d __b)
235{
236  return (__m128d)__builtin_ia32_cmpsd(__a, __b, 1);
237}
238
239static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
240_mm_cmple_sd(__m128d __a, __m128d __b)
241{
242  return (__m128d)__builtin_ia32_cmpsd(__a, __b, 2);
243}
244
245static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
246_mm_cmpgt_sd(__m128d __a, __m128d __b)
247{
248  __m128d __c = __builtin_ia32_cmpsd(__b, __a, 1);
249  return (__m128d) { __c[0], __a[1] };
250}
251
252static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
253_mm_cmpge_sd(__m128d __a, __m128d __b)
254{
255  __m128d __c = __builtin_ia32_cmpsd(__b, __a, 2);
256  return (__m128d) { __c[0], __a[1] };
257}
258
259static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
260_mm_cmpord_sd(__m128d __a, __m128d __b)
261{
262  return (__m128d)__builtin_ia32_cmpsd(__a, __b, 7);
263}
264
265static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
266_mm_cmpunord_sd(__m128d __a, __m128d __b)
267{
268  return (__m128d)__builtin_ia32_cmpsd(__a, __b, 3);
269}
270
271static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
272_mm_cmpneq_sd(__m128d __a, __m128d __b)
273{
274  return (__m128d)__builtin_ia32_cmpsd(__a, __b, 4);
275}
276
277static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
278_mm_cmpnlt_sd(__m128d __a, __m128d __b)
279{
280  return (__m128d)__builtin_ia32_cmpsd(__a, __b, 5);
281}
282
283static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
284_mm_cmpnle_sd(__m128d __a, __m128d __b)
285{
286  return (__m128d)__builtin_ia32_cmpsd(__a, __b, 6);
287}
288
289static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
290_mm_cmpngt_sd(__m128d __a, __m128d __b)
291{
292  __m128d __c = __builtin_ia32_cmpsd(__b, __a, 5);
293  return (__m128d) { __c[0], __a[1] };
294}
295
296static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
297_mm_cmpnge_sd(__m128d __a, __m128d __b)
298{
299  __m128d __c = __builtin_ia32_cmpsd(__b, __a, 6);
300  return (__m128d) { __c[0], __a[1] };
301}
302
303static __inline__ int __attribute__((__always_inline__, __nodebug__))
304_mm_comieq_sd(__m128d __a, __m128d __b)
305{
306  return __builtin_ia32_comisdeq(__a, __b);
307}
308
309static __inline__ int __attribute__((__always_inline__, __nodebug__))
310_mm_comilt_sd(__m128d __a, __m128d __b)
311{
312  return __builtin_ia32_comisdlt(__a, __b);
313}
314
315static __inline__ int __attribute__((__always_inline__, __nodebug__))
316_mm_comile_sd(__m128d __a, __m128d __b)
317{
318  return __builtin_ia32_comisdle(__a, __b);
319}
320
321static __inline__ int __attribute__((__always_inline__, __nodebug__))
322_mm_comigt_sd(__m128d __a, __m128d __b)
323{
324  return __builtin_ia32_comisdgt(__a, __b);
325}
326
327static __inline__ int __attribute__((__always_inline__, __nodebug__))
328_mm_comige_sd(__m128d __a, __m128d __b)
329{
330  return __builtin_ia32_comisdge(__a, __b);
331}
332
333static __inline__ int __attribute__((__always_inline__, __nodebug__))
334_mm_comineq_sd(__m128d __a, __m128d __b)
335{
336  return __builtin_ia32_comisdneq(__a, __b);
337}
338
339static __inline__ int __attribute__((__always_inline__, __nodebug__))
340_mm_ucomieq_sd(__m128d __a, __m128d __b)
341{
342  return __builtin_ia32_ucomisdeq(__a, __b);
343}
344
345static __inline__ int __attribute__((__always_inline__, __nodebug__))
346_mm_ucomilt_sd(__m128d __a, __m128d __b)
347{
348  return __builtin_ia32_ucomisdlt(__a, __b);
349}
350
351static __inline__ int __attribute__((__always_inline__, __nodebug__))
352_mm_ucomile_sd(__m128d __a, __m128d __b)
353{
354  return __builtin_ia32_ucomisdle(__a, __b);
355}
356
357static __inline__ int __attribute__((__always_inline__, __nodebug__))
358_mm_ucomigt_sd(__m128d __a, __m128d __b)
359{
360  return __builtin_ia32_ucomisdgt(__a, __b);
361}
362
363static __inline__ int __attribute__((__always_inline__, __nodebug__))
364_mm_ucomige_sd(__m128d __a, __m128d __b)
365{
366  return __builtin_ia32_ucomisdge(__a, __b);
367}
368
369static __inline__ int __attribute__((__always_inline__, __nodebug__))
370_mm_ucomineq_sd(__m128d __a, __m128d __b)
371{
372  return __builtin_ia32_ucomisdneq(__a, __b);
373}
374
375static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
376_mm_cvtpd_ps(__m128d __a)
377{
378  return __builtin_ia32_cvtpd2ps(__a);
379}
380
381static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
382_mm_cvtps_pd(__m128 __a)
383{
384  return __builtin_ia32_cvtps2pd(__a);
385}
386
387static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
388_mm_cvtepi32_pd(__m128i __a)
389{
390  return __builtin_ia32_cvtdq2pd((__v4si)__a);
391}
392
393static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
394_mm_cvtpd_epi32(__m128d __a)
395{
396  return __builtin_ia32_cvtpd2dq(__a);
397}
398
399static __inline__ int __attribute__((__always_inline__, __nodebug__))
400_mm_cvtsd_si32(__m128d __a)
401{
402  return __builtin_ia32_cvtsd2si(__a);
403}
404
405static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
406_mm_cvtsd_ss(__m128 __a, __m128d __b)
407{
408  __a[0] = __b[0];
409  return __a;
410}
411
412static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
413_mm_cvtsi32_sd(__m128d __a, int __b)
414{
415  __a[0] = __b;
416  return __a;
417}
418
419static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
420_mm_cvtss_sd(__m128d __a, __m128 __b)
421{
422  __a[0] = __b[0];
423  return __a;
424}
425
426static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
427_mm_cvttpd_epi32(__m128d __a)
428{
429  return (__m128i)__builtin_ia32_cvttpd2dq(__a);
430}
431
432static __inline__ int __attribute__((__always_inline__, __nodebug__))
433_mm_cvttsd_si32(__m128d __a)
434{
435  return __a[0];
436}
437
438static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
439_mm_cvtpd_pi32(__m128d __a)
440{
441  return (__m64)__builtin_ia32_cvtpd2pi(__a);
442}
443
444static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
445_mm_cvttpd_pi32(__m128d __a)
446{
447  return (__m64)__builtin_ia32_cvttpd2pi(__a);
448}
449
450static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
451_mm_cvtpi32_pd(__m64 __a)
452{
453  return __builtin_ia32_cvtpi2pd((__v2si)__a);
454}
455
456static __inline__ double __attribute__((__always_inline__, __nodebug__))
457_mm_cvtsd_f64(__m128d __a)
458{
459  return __a[0];
460}
461
462static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
463_mm_load_pd(double const *__dp)
464{
465  return *(__m128d*)__dp;
466}
467
468static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
469_mm_load1_pd(double const *__dp)
470{
471  struct __mm_load1_pd_struct {
472    double __u;
473  } __attribute__((__packed__, __may_alias__));
474  double __u = ((struct __mm_load1_pd_struct*)__dp)->__u;
475  return (__m128d){ __u, __u };
476}
477
478#define        _mm_load_pd1(dp)        _mm_load1_pd(dp)
479
480static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
481_mm_loadr_pd(double const *__dp)
482{
483  __m128d __u = *(__m128d*)__dp;
484  return __builtin_shufflevector(__u, __u, 1, 0);
485}
486
487static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
488_mm_loadu_pd(double const *__dp)
489{
490  struct __loadu_pd {
491    __m128d __v;
492  } __attribute__((packed, may_alias));
493  return ((struct __loadu_pd*)__dp)->__v;
494}
495
496static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
497_mm_load_sd(double const *__dp)
498{
499  struct __mm_load_sd_struct {
500    double __u;
501  } __attribute__((__packed__, __may_alias__));
502  double __u = ((struct __mm_load_sd_struct*)__dp)->__u;
503  return (__m128d){ __u, 0 };
504}
505
506static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
507_mm_loadh_pd(__m128d __a, double const *__dp)
508{
509  struct __mm_loadh_pd_struct {
510    double __u;
511  } __attribute__((__packed__, __may_alias__));
512  double __u = ((struct __mm_loadh_pd_struct*)__dp)->__u;
513  return (__m128d){ __a[0], __u };
514}
515
516static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
517_mm_loadl_pd(__m128d __a, double const *__dp)
518{
519  struct __mm_loadl_pd_struct {
520    double __u;
521  } __attribute__((__packed__, __may_alias__));
522  double __u = ((struct __mm_loadl_pd_struct*)__dp)->__u;
523  return (__m128d){ __u, __a[1] };
524}
525
526static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
527_mm_set_sd(double __w)
528{
529  return (__m128d){ __w, 0 };
530}
531
532static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
533_mm_set1_pd(double __w)
534{
535  return (__m128d){ __w, __w };
536}
537
538static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
539_mm_set_pd(double __w, double __x)
540{
541  return (__m128d){ __x, __w };
542}
543
544static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
545_mm_setr_pd(double __w, double __x)
546{
547  return (__m128d){ __w, __x };
548}
549
550static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
551_mm_setzero_pd(void)
552{
553  return (__m128d){ 0, 0 };
554}
555
556static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
557_mm_move_sd(__m128d __a, __m128d __b)
558{
559  return (__m128d){ __b[0], __a[1] };
560}
561
562static __inline__ void __attribute__((__always_inline__, __nodebug__))
563_mm_store_sd(double *__dp, __m128d __a)
564{
565  struct __mm_store_sd_struct {
566    double __u;
567  } __attribute__((__packed__, __may_alias__));
568  ((struct __mm_store_sd_struct*)__dp)->__u = __a[0];
569}
570
571static __inline__ void __attribute__((__always_inline__, __nodebug__))
572_mm_store1_pd(double *__dp, __m128d __a)
573{
574  struct __mm_store1_pd_struct {
575    double __u[2];
576  } __attribute__((__packed__, __may_alias__));
577  ((struct __mm_store1_pd_struct*)__dp)->__u[0] = __a[0];
578  ((struct __mm_store1_pd_struct*)__dp)->__u[1] = __a[0];
579}
580
581static __inline__ void __attribute__((__always_inline__, __nodebug__))
582_mm_store_pd(double *__dp, __m128d __a)
583{
584  *(__m128d *)__dp = __a;
585}
586
587static __inline__ void __attribute__((__always_inline__, __nodebug__))
588_mm_storeu_pd(double *__dp, __m128d __a)
589{
590  __builtin_ia32_storeupd(__dp, __a);
591}
592
593static __inline__ void __attribute__((__always_inline__, __nodebug__))
594_mm_storer_pd(double *__dp, __m128d __a)
595{
596  __a = __builtin_shufflevector(__a, __a, 1, 0);
597  *(__m128d *)__dp = __a;
598}
599
600static __inline__ void __attribute__((__always_inline__, __nodebug__))
601_mm_storeh_pd(double *__dp, __m128d __a)
602{
603  struct __mm_storeh_pd_struct {
604    double __u;
605  } __attribute__((__packed__, __may_alias__));
606  ((struct __mm_storeh_pd_struct*)__dp)->__u = __a[1];
607}
608
609static __inline__ void __attribute__((__always_inline__, __nodebug__))
610_mm_storel_pd(double *__dp, __m128d __a)
611{
612  struct __mm_storeh_pd_struct {
613    double __u;
614  } __attribute__((__packed__, __may_alias__));
615  ((struct __mm_storeh_pd_struct*)__dp)->__u = __a[0];
616}
617
618static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
619_mm_add_epi8(__m128i __a, __m128i __b)
620{
621  return (__m128i)((__v16qi)__a + (__v16qi)__b);
622}
623
624static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
625_mm_add_epi16(__m128i __a, __m128i __b)
626{
627  return (__m128i)((__v8hi)__a + (__v8hi)__b);
628}
629
630static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
631_mm_add_epi32(__m128i __a, __m128i __b)
632{
633  return (__m128i)((__v4si)__a + (__v4si)__b);
634}
635
636static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
637_mm_add_si64(__m64 __a, __m64 __b)
638{
639  return __a + __b;
640}
641
642static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
643_mm_add_epi64(__m128i __a, __m128i __b)
644{
645  return __a + __b;
646}
647
648static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
649_mm_adds_epi8(__m128i __a, __m128i __b)
650{
651  return (__m128i)__builtin_ia32_paddsb128((__v16qi)__a, (__v16qi)__b);
652}
653
654static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
655_mm_adds_epi16(__m128i __a, __m128i __b)
656{
657  return (__m128i)__builtin_ia32_paddsw128((__v8hi)__a, (__v8hi)__b);
658}
659
660static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
661_mm_adds_epu8(__m128i __a, __m128i __b)
662{
663  return (__m128i)__builtin_ia32_paddusb128((__v16qi)__a, (__v16qi)__b);
664}
665
666static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
667_mm_adds_epu16(__m128i __a, __m128i __b)
668{
669  return (__m128i)__builtin_ia32_paddusw128((__v8hi)__a, (__v8hi)__b);
670}
671
672static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
673_mm_avg_epu8(__m128i __a, __m128i __b)
674{
675  return (__m128i)__builtin_ia32_pavgb128((__v16qi)__a, (__v16qi)__b);
676}
677
678static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
679_mm_avg_epu16(__m128i __a, __m128i __b)
680{
681  return (__m128i)__builtin_ia32_pavgw128((__v8hi)__a, (__v8hi)__b);
682}
683
684static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
685_mm_madd_epi16(__m128i __a, __m128i __b)
686{
687  return (__m128i)__builtin_ia32_pmaddwd128((__v8hi)__a, (__v8hi)__b);
688}
689
690static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
691_mm_max_epi16(__m128i __a, __m128i __b)
692{
693  return (__m128i)__builtin_ia32_pmaxsw128((__v8hi)__a, (__v8hi)__b);
694}
695
696static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
697_mm_max_epu8(__m128i __a, __m128i __b)
698{
699  return (__m128i)__builtin_ia32_pmaxub128((__v16qi)__a, (__v16qi)__b);
700}
701
702static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
703_mm_min_epi16(__m128i __a, __m128i __b)
704{
705  return (__m128i)__builtin_ia32_pminsw128((__v8hi)__a, (__v8hi)__b);
706}
707
708static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
709_mm_min_epu8(__m128i __a, __m128i __b)
710{
711  return (__m128i)__builtin_ia32_pminub128((__v16qi)__a, (__v16qi)__b);
712}
713
714static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
715_mm_mulhi_epi16(__m128i __a, __m128i __b)
716{
717  return (__m128i)__builtin_ia32_pmulhw128((__v8hi)__a, (__v8hi)__b);
718}
719
720static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
721_mm_mulhi_epu16(__m128i __a, __m128i __b)
722{
723  return (__m128i)__builtin_ia32_pmulhuw128((__v8hi)__a, (__v8hi)__b);
724}
725
726static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
727_mm_mullo_epi16(__m128i __a, __m128i __b)
728{
729  return (__m128i)((__v8hi)__a * (__v8hi)__b);
730}
731
732static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
733_mm_mul_su32(__m64 __a, __m64 __b)
734{
735  return __builtin_ia32_pmuludq((__v2si)__a, (__v2si)__b);
736}
737
738static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
739_mm_mul_epu32(__m128i __a, __m128i __b)
740{
741  return __builtin_ia32_pmuludq128((__v4si)__a, (__v4si)__b);
742}
743
744static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
745_mm_sad_epu8(__m128i __a, __m128i __b)
746{
747  return __builtin_ia32_psadbw128((__v16qi)__a, (__v16qi)__b);
748}
749
750static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
751_mm_sub_epi8(__m128i __a, __m128i __b)
752{
753  return (__m128i)((__v16qi)__a - (__v16qi)__b);
754}
755
756static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
757_mm_sub_epi16(__m128i __a, __m128i __b)
758{
759  return (__m128i)((__v8hi)__a - (__v8hi)__b);
760}
761
762static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
763_mm_sub_epi32(__m128i __a, __m128i __b)
764{
765  return (__m128i)((__v4si)__a - (__v4si)__b);
766}
767
768static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
769_mm_sub_si64(__m64 __a, __m64 __b)
770{
771  return __a - __b;
772}
773
774static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
775_mm_sub_epi64(__m128i __a, __m128i __b)
776{
777  return __a - __b;
778}
779
780static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
781_mm_subs_epi8(__m128i __a, __m128i __b)
782{
783  return (__m128i)__builtin_ia32_psubsb128((__v16qi)__a, (__v16qi)__b);
784}
785
786static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
787_mm_subs_epi16(__m128i __a, __m128i __b)
788{
789  return (__m128i)__builtin_ia32_psubsw128((__v8hi)__a, (__v8hi)__b);
790}
791
792static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
793_mm_subs_epu8(__m128i __a, __m128i __b)
794{
795  return (__m128i)__builtin_ia32_psubusb128((__v16qi)__a, (__v16qi)__b);
796}
797
798static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
799_mm_subs_epu16(__m128i __a, __m128i __b)
800{
801  return (__m128i)__builtin_ia32_psubusw128((__v8hi)__a, (__v8hi)__b);
802}
803
804static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
805_mm_and_si128(__m128i __a, __m128i __b)
806{
807  return __a & __b;
808}
809
810static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
811_mm_andnot_si128(__m128i __a, __m128i __b)
812{
813  return ~__a & __b;
814}
815
816static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
817_mm_or_si128(__m128i __a, __m128i __b)
818{
819  return __a | __b;
820}
821
822static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
823_mm_xor_si128(__m128i __a, __m128i __b)
824{
825  return __a ^ __b;
826}
827
828#define _mm_slli_si128(a, count) __extension__ ({ \
829  __m128i __a = (a); \
830  (__m128i)__builtin_ia32_pslldqi128(__a, (count)*8); })
831
832static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
833_mm_slli_epi16(__m128i __a, int __count)
834{
835  return (__m128i)__builtin_ia32_psllwi128((__v8hi)__a, __count);
836}
837
838static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
839_mm_sll_epi16(__m128i __a, __m128i __count)
840{
841  return (__m128i)__builtin_ia32_psllw128((__v8hi)__a, (__v8hi)__count);
842}
843
844static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
845_mm_slli_epi32(__m128i __a, int __count)
846{
847  return (__m128i)__builtin_ia32_pslldi128((__v4si)__a, __count);
848}
849
850static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
851_mm_sll_epi32(__m128i __a, __m128i __count)
852{
853  return (__m128i)__builtin_ia32_pslld128((__v4si)__a, (__v4si)__count);
854}
855
856static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
857_mm_slli_epi64(__m128i __a, int __count)
858{
859  return __builtin_ia32_psllqi128(__a, __count);
860}
861
862static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
863_mm_sll_epi64(__m128i __a, __m128i __count)
864{
865  return __builtin_ia32_psllq128(__a, __count);
866}
867
868static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
869_mm_srai_epi16(__m128i __a, int __count)
870{
871  return (__m128i)__builtin_ia32_psrawi128((__v8hi)__a, __count);
872}
873
874static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
875_mm_sra_epi16(__m128i __a, __m128i __count)
876{
877  return (__m128i)__builtin_ia32_psraw128((__v8hi)__a, (__v8hi)__count);
878}
879
880static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
881_mm_srai_epi32(__m128i __a, int __count)
882{
883  return (__m128i)__builtin_ia32_psradi128((__v4si)__a, __count);
884}
885
886static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
887_mm_sra_epi32(__m128i __a, __m128i __count)
888{
889  return (__m128i)__builtin_ia32_psrad128((__v4si)__a, (__v4si)__count);
890}
891
892
893#define _mm_srli_si128(a, count) __extension__ ({ \
894  __m128i __a = (a); \
895  (__m128i)__builtin_ia32_psrldqi128(__a, (count)*8); })
896
897static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
898_mm_srli_epi16(__m128i __a, int __count)
899{
900  return (__m128i)__builtin_ia32_psrlwi128((__v8hi)__a, __count);
901}
902
903static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
904_mm_srl_epi16(__m128i __a, __m128i __count)
905{
906  return (__m128i)__builtin_ia32_psrlw128((__v8hi)__a, (__v8hi)__count);
907}
908
909static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
910_mm_srli_epi32(__m128i __a, int __count)
911{
912  return (__m128i)__builtin_ia32_psrldi128((__v4si)__a, __count);
913}
914
915static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
916_mm_srl_epi32(__m128i __a, __m128i __count)
917{
918  return (__m128i)__builtin_ia32_psrld128((__v4si)__a, (__v4si)__count);
919}
920
921static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
922_mm_srli_epi64(__m128i __a, int __count)
923{
924  return __builtin_ia32_psrlqi128(__a, __count);
925}
926
927static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
928_mm_srl_epi64(__m128i __a, __m128i __count)
929{
930  return __builtin_ia32_psrlq128(__a, __count);
931}
932
933static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
934_mm_cmpeq_epi8(__m128i __a, __m128i __b)
935{
936  return (__m128i)((__v16qi)__a == (__v16qi)__b);
937}
938
939static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
940_mm_cmpeq_epi16(__m128i __a, __m128i __b)
941{
942  return (__m128i)((__v8hi)__a == (__v8hi)__b);
943}
944
945static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
946_mm_cmpeq_epi32(__m128i __a, __m128i __b)
947{
948  return (__m128i)((__v4si)__a == (__v4si)__b);
949}
950
951static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
952_mm_cmpgt_epi8(__m128i __a, __m128i __b)
953{
954  /* This function always performs a signed comparison, but __v16qi is a char
955     which may be signed or unsigned. */
956  typedef signed char __v16qs __attribute__((__vector_size__(16)));
957  return (__m128i)((__v16qs)__a > (__v16qs)__b);
958}
959
960static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
961_mm_cmpgt_epi16(__m128i __a, __m128i __b)
962{
963  return (__m128i)((__v8hi)__a > (__v8hi)__b);
964}
965
966static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
967_mm_cmpgt_epi32(__m128i __a, __m128i __b)
968{
969  return (__m128i)((__v4si)__a > (__v4si)__b);
970}
971
972static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
973_mm_cmplt_epi8(__m128i __a, __m128i __b)
974{
975  return _mm_cmpgt_epi8(__b, __a);
976}
977
978static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
979_mm_cmplt_epi16(__m128i __a, __m128i __b)
980{
981  return _mm_cmpgt_epi16(__b, __a);
982}
983
984static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
985_mm_cmplt_epi32(__m128i __a, __m128i __b)
986{
987  return _mm_cmpgt_epi32(__b, __a);
988}
989
990#ifdef __x86_64__
991static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
992_mm_cvtsi64_sd(__m128d __a, long long __b)
993{
994  __a[0] = __b;
995  return __a;
996}
997
998static __inline__ long long __attribute__((__always_inline__, __nodebug__))
999_mm_cvtsd_si64(__m128d __a)
1000{
1001  return __builtin_ia32_cvtsd2si64(__a);
1002}
1003
1004static __inline__ long long __attribute__((__always_inline__, __nodebug__))
1005_mm_cvttsd_si64(__m128d __a)
1006{
1007  return __a[0];
1008}
1009#endif
1010
1011static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
1012_mm_cvtepi32_ps(__m128i __a)
1013{
1014  return __builtin_ia32_cvtdq2ps((__v4si)__a);
1015}
1016
1017static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1018_mm_cvtps_epi32(__m128 __a)
1019{
1020  return (__m128i)__builtin_ia32_cvtps2dq(__a);
1021}
1022
1023static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1024_mm_cvttps_epi32(__m128 __a)
1025{
1026  return (__m128i)__builtin_ia32_cvttps2dq(__a);
1027}
1028
1029static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1030_mm_cvtsi32_si128(int __a)
1031{
1032  return (__m128i)(__v4si){ __a, 0, 0, 0 };
1033}
1034
1035#ifdef __x86_64__
1036static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1037_mm_cvtsi64_si128(long long __a)
1038{
1039  return (__m128i){ __a, 0 };
1040}
1041#endif
1042
1043static __inline__ int __attribute__((__always_inline__, __nodebug__))
1044_mm_cvtsi128_si32(__m128i __a)
1045{
1046  __v4si __b = (__v4si)__a;
1047  return __b[0];
1048}
1049
1050#ifdef __x86_64__
1051static __inline__ long long __attribute__((__always_inline__, __nodebug__))
1052_mm_cvtsi128_si64(__m128i __a)
1053{
1054  return __a[0];
1055}
1056#endif
1057
1058static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1059_mm_load_si128(__m128i const *__p)
1060{
1061  return *__p;
1062}
1063
1064static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1065_mm_loadu_si128(__m128i const *__p)
1066{
1067  struct __loadu_si128 {
1068    __m128i __v;
1069  } __attribute__((packed, may_alias));
1070  return ((struct __loadu_si128*)__p)->__v;
1071}
1072
1073static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1074_mm_loadl_epi64(__m128i const *__p)
1075{
1076  struct __mm_loadl_epi64_struct {
1077    long long __u;
1078  } __attribute__((__packed__, __may_alias__));
1079  return (__m128i) { ((struct __mm_loadl_epi64_struct*)__p)->__u, 0};
1080}
1081
1082static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1083_mm_set_epi64x(long long q1, long long q0)
1084{
1085  return (__m128i){ q0, q1 };
1086}
1087
1088static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1089_mm_set_epi64(__m64 q1, __m64 q0)
1090{
1091  return (__m128i){ (long long)q0, (long long)q1 };
1092}
1093
1094static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1095_mm_set_epi32(int i3, int i2, int i1, int i0)
1096{
1097  return (__m128i)(__v4si){ i0, i1, i2, i3};
1098}
1099
1100static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1101_mm_set_epi16(short w7, short w6, short w5, short w4, short w3, short w2, short w1, short w0)
1102{
1103  return (__m128i)(__v8hi){ w0, w1, w2, w3, w4, w5, w6, w7 };
1104}
1105
1106static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1107_mm_set_epi8(char b15, char b14, char b13, char b12, char b11, char b10, char b9, char b8, char b7, char b6, char b5, char b4, char b3, char b2, char b1, char b0)
1108{
1109  return (__m128i)(__v16qi){ b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, b10, b11, b12, b13, b14, b15 };
1110}
1111
1112static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1113_mm_set1_epi64x(long long __q)
1114{
1115  return (__m128i){ __q, __q };
1116}
1117
1118static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1119_mm_set1_epi64(__m64 __q)
1120{
1121  return (__m128i){ (long long)__q, (long long)__q };
1122}
1123
1124static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1125_mm_set1_epi32(int __i)
1126{
1127  return (__m128i)(__v4si){ __i, __i, __i, __i };
1128}
1129
1130static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1131_mm_set1_epi16(short __w)
1132{
1133  return (__m128i)(__v8hi){ __w, __w, __w, __w, __w, __w, __w, __w };
1134}
1135
1136static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1137_mm_set1_epi8(char __b)
1138{
1139  return (__m128i)(__v16qi){ __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b };
1140}
1141
1142static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1143_mm_setr_epi64(__m64 q0, __m64 q1)
1144{
1145  return (__m128i){ (long long)q0, (long long)q1 };
1146}
1147
1148static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1149_mm_setr_epi32(int i0, int i1, int i2, int i3)
1150{
1151  return (__m128i)(__v4si){ i0, i1, i2, i3};
1152}
1153
1154static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1155_mm_setr_epi16(short w0, short w1, short w2, short w3, short w4, short w5, short w6, short w7)
1156{
1157  return (__m128i)(__v8hi){ w0, w1, w2, w3, w4, w5, w6, w7 };
1158}
1159
1160static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1161_mm_setr_epi8(char b0, char b1, char b2, char b3, char b4, char b5, char b6, char b7, char b8, char b9, char b10, char b11, char b12, char b13, char b14, char b15)
1162{
1163  return (__m128i)(__v16qi){ b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, b10, b11, b12, b13, b14, b15 };
1164}
1165
1166static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1167_mm_setzero_si128(void)
1168{
1169  return (__m128i){ 0LL, 0LL };
1170}
1171
1172static __inline__ void __attribute__((__always_inline__, __nodebug__))
1173_mm_store_si128(__m128i *__p, __m128i __b)
1174{
1175  *__p = __b;
1176}
1177
1178static __inline__ void __attribute__((__always_inline__, __nodebug__))
1179_mm_storeu_si128(__m128i *__p, __m128i __b)
1180{
1181  __builtin_ia32_storedqu((char *)__p, (__v16qi)__b);
1182}
1183
1184static __inline__ void __attribute__((__always_inline__, __nodebug__))
1185_mm_maskmoveu_si128(__m128i __d, __m128i __n, char *__p)
1186{
1187  __builtin_ia32_maskmovdqu((__v16qi)__d, (__v16qi)__n, __p);
1188}
1189
1190static __inline__ void __attribute__((__always_inline__, __nodebug__))
1191_mm_storel_epi64(__m128i *__p, __m128i __a)
1192{
1193  struct __mm_storel_epi64_struct {
1194    long long __u;
1195  } __attribute__((__packed__, __may_alias__));
1196  ((struct __mm_storel_epi64_struct*)__p)->__u = __a[0];
1197}
1198
1199static __inline__ void __attribute__((__always_inline__, __nodebug__))
1200_mm_stream_pd(double *__p, __m128d __a)
1201{
1202  __builtin_ia32_movntpd(__p, __a);
1203}
1204
1205static __inline__ void __attribute__((__always_inline__, __nodebug__))
1206_mm_stream_si128(__m128i *__p, __m128i __a)
1207{
1208  __builtin_ia32_movntdq(__p, __a);
1209}
1210
1211static __inline__ void __attribute__((__always_inline__, __nodebug__))
1212_mm_stream_si32(int *__p, int __a)
1213{
1214  __builtin_ia32_movnti(__p, __a);
1215}
1216
1217static __inline__ void __attribute__((__always_inline__, __nodebug__))
1218_mm_clflush(void const *__p)
1219{
1220  __builtin_ia32_clflush(__p);
1221}
1222
1223static __inline__ void __attribute__((__always_inline__, __nodebug__))
1224_mm_lfence(void)
1225{
1226  __builtin_ia32_lfence();
1227}
1228
1229static __inline__ void __attribute__((__always_inline__, __nodebug__))
1230_mm_mfence(void)
1231{
1232  __builtin_ia32_mfence();
1233}
1234
1235static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1236_mm_packs_epi16(__m128i __a, __m128i __b)
1237{
1238  return (__m128i)__builtin_ia32_packsswb128((__v8hi)__a, (__v8hi)__b);
1239}
1240
1241static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1242_mm_packs_epi32(__m128i __a, __m128i __b)
1243{
1244  return (__m128i)__builtin_ia32_packssdw128((__v4si)__a, (__v4si)__b);
1245}
1246
1247static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1248_mm_packus_epi16(__m128i __a, __m128i __b)
1249{
1250  return (__m128i)__builtin_ia32_packuswb128((__v8hi)__a, (__v8hi)__b);
1251}
1252
1253static __inline__ int __attribute__((__always_inline__, __nodebug__))
1254_mm_extract_epi16(__m128i __a, int __imm)
1255{
1256  __v8hi __b = (__v8hi)__a;
1257  return (unsigned short)__b[__imm];
1258}
1259
1260static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1261_mm_insert_epi16(__m128i __a, int __b, int __imm)
1262{
1263  __v8hi __c = (__v8hi)__a;
1264  __c[__imm & 7] = __b;
1265  return (__m128i)__c;
1266}
1267
1268static __inline__ int __attribute__((__always_inline__, __nodebug__))
1269_mm_movemask_epi8(__m128i __a)
1270{
1271  return __builtin_ia32_pmovmskb128((__v16qi)__a);
1272}
1273
1274#define _mm_shuffle_epi32(a, imm) __extension__ ({ \
1275  __m128i __a = (a); \
1276  (__m128i)__builtin_shufflevector((__v4si)__a, (__v4si) _mm_set1_epi32(0), \
1277                                   (imm) & 0x3, ((imm) & 0xc) >> 2, \
1278                                   ((imm) & 0x30) >> 4, ((imm) & 0xc0) >> 6); })
1279
1280#define _mm_shufflelo_epi16(a, imm) __extension__ ({ \
1281  __m128i __a = (a); \
1282  (__m128i)__builtin_shufflevector((__v8hi)__a, (__v8hi) _mm_set1_epi16(0), \
1283                                   (imm) & 0x3, ((imm) & 0xc) >> 2, \
1284                                   ((imm) & 0x30) >> 4, ((imm) & 0xc0) >> 6, \
1285                                   4, 5, 6, 7); })
1286
1287#define _mm_shufflehi_epi16(a, imm) __extension__ ({ \
1288  __m128i __a = (a); \
1289  (__m128i)__builtin_shufflevector((__v8hi)__a, (__v8hi) _mm_set1_epi16(0), \
1290                                   0, 1, 2, 3, \
1291                                   4 + (((imm) & 0x03) >> 0), \
1292                                   4 + (((imm) & 0x0c) >> 2), \
1293                                   4 + (((imm) & 0x30) >> 4), \
1294                                   4 + (((imm) & 0xc0) >> 6)); })
1295
1296static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1297_mm_unpackhi_epi8(__m128i __a, __m128i __b)
1298{
1299  return (__m128i)__builtin_shufflevector((__v16qi)__a, (__v16qi)__b, 8, 16+8, 9, 16+9, 10, 16+10, 11, 16+11, 12, 16+12, 13, 16+13, 14, 16+14, 15, 16+15);
1300}
1301
1302static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1303_mm_unpackhi_epi16(__m128i __a, __m128i __b)
1304{
1305  return (__m128i)__builtin_shufflevector((__v8hi)__a, (__v8hi)__b, 4, 8+4, 5, 8+5, 6, 8+6, 7, 8+7);
1306}
1307
1308static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1309_mm_unpackhi_epi32(__m128i __a, __m128i __b)
1310{
1311  return (__m128i)__builtin_shufflevector((__v4si)__a, (__v4si)__b, 2, 4+2, 3, 4+3);
1312}
1313
1314static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1315_mm_unpackhi_epi64(__m128i __a, __m128i __b)
1316{
1317  return (__m128i)__builtin_shufflevector(__a, __b, 1, 2+1);
1318}
1319
1320static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1321_mm_unpacklo_epi8(__m128i __a, __m128i __b)
1322{
1323  return (__m128i)__builtin_shufflevector((__v16qi)__a, (__v16qi)__b, 0, 16+0, 1, 16+1, 2, 16+2, 3, 16+3, 4, 16+4, 5, 16+5, 6, 16+6, 7, 16+7);
1324}
1325
1326static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1327_mm_unpacklo_epi16(__m128i __a, __m128i __b)
1328{
1329  return (__m128i)__builtin_shufflevector((__v8hi)__a, (__v8hi)__b, 0, 8+0, 1, 8+1, 2, 8+2, 3, 8+3);
1330}
1331
1332static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1333_mm_unpacklo_epi32(__m128i __a, __m128i __b)
1334{
1335  return (__m128i)__builtin_shufflevector((__v4si)__a, (__v4si)__b, 0, 4+0, 1, 4+1);
1336}
1337
1338static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1339_mm_unpacklo_epi64(__m128i __a, __m128i __b)
1340{
1341  return (__m128i)__builtin_shufflevector(__a, __b, 0, 2+0);
1342}
1343
1344static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
1345_mm_movepi64_pi64(__m128i __a)
1346{
1347  return (__m64)__a[0];
1348}
1349
1350static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1351_mm_movpi64_pi64(__m64 __a)
1352{
1353  return (__m128i){ (long long)__a, 0 };
1354}
1355
1356static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1357_mm_move_epi64(__m128i __a)
1358{
1359  return __builtin_shufflevector(__a, (__m128i){ 0 }, 0, 2);
1360}
1361
1362static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
1363_mm_unpackhi_pd(__m128d __a, __m128d __b)
1364{
1365  return __builtin_shufflevector(__a, __b, 1, 2+1);
1366}
1367
1368static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
1369_mm_unpacklo_pd(__m128d __a, __m128d __b)
1370{
1371  return __builtin_shufflevector(__a, __b, 0, 2+0);
1372}
1373
1374static __inline__ int __attribute__((__always_inline__, __nodebug__))
1375_mm_movemask_pd(__m128d __a)
1376{
1377  return __builtin_ia32_movmskpd(__a);
1378}
1379
1380#define _mm_shuffle_pd(a, b, i) __extension__ ({ \
1381  __m128d __a = (a); \
1382  __m128d __b = (b); \
1383  __builtin_shufflevector(__a, __b, (i) & 1, (((i) & 2) >> 1) + 2); })
1384
1385static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
1386_mm_castpd_ps(__m128d __a)
1387{
1388  return (__m128)__a;
1389}
1390
1391static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1392_mm_castpd_si128(__m128d __a)
1393{
1394  return (__m128i)__a;
1395}
1396
1397static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
1398_mm_castps_pd(__m128 __a)
1399{
1400  return (__m128d)__a;
1401}
1402
1403static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1404_mm_castps_si128(__m128 __a)
1405{
1406  return (__m128i)__a;
1407}
1408
1409static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
1410_mm_castsi128_ps(__m128i __a)
1411{
1412  return (__m128)__a;
1413}
1414
1415static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
1416_mm_castsi128_pd(__m128i __a)
1417{
1418  return (__m128d)__a;
1419}
1420
1421static __inline__ void __attribute__((__always_inline__, __nodebug__))
1422_mm_pause(void)
1423{
1424  __asm__ volatile ("pause");
1425}
1426
1427#define _MM_SHUFFLE2(x, y) (((x) << 1) | (y))
1428
1429#endif /* __SSE2__ */
1430
1431#endif /* __EMMINTRIN_H */
1432