xmmintrin.h revision 62005c16bc3a416c1ecfe0ceaeda9a8dd3e5b0be 
1/*===---- xmmintrin.h - SSE 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 __XMMINTRIN_H
25#define __XMMINTRIN_H
26
27#ifndef __SSE__
28#error "SSE instruction set not enabled"
29#else
30
31#include <mmintrin.h>
32
33typedef float __v4sf __attribute__((__vector_size__(16)));
34typedef float __m128 __attribute__((__vector_size__(16)));
35
36#include <mm_malloc.h>
37
38static inline __m128 __attribute__((__always_inline__)) _mm_add_ss(__m128 a, __m128 b)
39{
40  return __builtin_ia32_addss(a, b);
41}
42
43static inline __m128 __attribute__((__always_inline__)) _mm_add_ps(__m128 a, __m128 b)
44{
45  return a + b;
46}
47
48static inline __m128 __attribute__((__always_inline__)) _mm_sub_ss(__m128 a, __m128 b)
49{
50  return __builtin_ia32_subss(a, b);
51}
52
53static inline __m128 __attribute__((__always_inline__)) _mm_sub_ps(__m128 a, __m128 b)
54{
55  return a - b;
56}
57
58static inline __m128 __attribute__((__always_inline__)) _mm_mul_ss(__m128 a, __m128 b)
59{
60  return __builtin_ia32_mulss(a, b);
61}
62
63static inline __m128 __attribute__((__always_inline__)) _mm_mul_ps(__m128 a, __m128 b)
64{
65  return a * b;
66}
67
68static inline __m128 __attribute__((__always_inline__)) _mm_div_ss(__m128 a, __m128 b)
69{
70  return __builtin_ia32_divss(a, b);
71}
72
73static inline __m128 __attribute__((__always_inline__)) _mm_div_ps(__m128 a, __m128 b)
74{
75  return a / b;
76}
77
78static inline __m128 __attribute__((__always_inline__)) _mm_sqrt_ss(__m128 a)
79{
80  return __builtin_ia32_sqrtss(a);
81}
82
83static inline __m128 __attribute__((__always_inline__)) _mm_sqrt_ps(__m128 a)
84{
85  return __builtin_ia32_sqrtps(a);
86}
87
88static inline __m128 __attribute__((__always_inline__)) _mm_rcp_ss(__m128 a)
89{
90  return __builtin_ia32_rcpss(a);
91}
92
93static inline __m128 __attribute__((__always_inline__)) _mm_rcp_ps(__m128 a)
94{
95  return __builtin_ia32_rcpps(a);
96}
97
98static inline __m128 __attribute__((__always_inline__)) _mm_rsqrt_ss(__m128 a)
99{
100  return __builtin_ia32_rsqrtss(a);
101}
102
103static inline __m128 __attribute__((__always_inline__)) _mm_rsqrt_ps(__m128 a)
104{
105  return __builtin_ia32_rsqrtps(a);
106}
107
108static inline __m128 __attribute__((__always_inline__)) _mm_min_ss(__m128 a, __m128 b)
109{
110  return __builtin_ia32_minss(a, b);
111}
112
113static inline __m128 __attribute__((__always_inline__)) _mm_min_ps(__m128 a, __m128 b)
114{
115  return __builtin_ia32_minps(a, b);
116}
117
118static inline __m128 __attribute__((__always_inline__)) _mm_max_ss(__m128 a, __m128 b)
119{
120  return __builtin_ia32_maxss(a, b);
121}
122
123static inline __m128 __attribute__((__always_inline__)) _mm_max_ps(__m128 a, __m128 b)
124{
125  return __builtin_ia32_maxps(a, b);
126}
127
128static inline __m128 __attribute__((__always_inline__)) _mm_and_ps(__m128 a, __m128 b)
129{
130  return __builtin_ia32_andps(a, b);
131}
132
133static inline __m128 __attribute__((__always_inline__)) _mm_andnot_ps(__m128 a, __m128 b)
134{
135  return __builtin_ia32_andnps(a, b);
136}
137
138static inline __m128 __attribute__((__always_inline__)) _mm_or_ps(__m128 a, __m128 b)
139{
140  return __builtin_ia32_orps(a, b);
141}
142
143static inline __m128 __attribute__((__always_inline__)) _mm_xor_ps(__m128 a, __m128 b)
144{
145  return __builtin_ia32_xorps(a, b);
146}
147
148static inline __m128 __attribute__((__always_inline__)) _mm_cmpeq_ss(__m128 a, __m128 b)
149{
150  return (__m128)__builtin_ia32_cmpeqss(a, b);
151}
152
153static inline __m128 __attribute__((__always_inline__)) _mm_cmpeq_ps(__m128 a, __m128 b)
154{
155  return (__m128)__builtin_ia32_cmpeqps(a, b);
156}
157
158static inline __m128 __attribute__((__always_inline__)) _mm_cmplt_ss(__m128 a, __m128 b)
159{
160  return (__m128)__builtin_ia32_cmpltss(a, b);
161}
162
163static inline __m128 __attribute__((__always_inline__)) _mm_cmplt_ps(__m128 a, __m128 b)
164{
165  return (__m128)__builtin_ia32_cmpltps(a, b);
166}
167
168static inline __m128 __attribute__((__always_inline__)) _mm_cmple_ss(__m128 a, __m128 b)
169{
170  return (__m128)__builtin_ia32_cmpless(a, b);
171}
172
173static inline __m128 __attribute__((__always_inline__)) _mm_cmple_ps(__m128 a, __m128 b)
174{
175  return (__m128)__builtin_ia32_cmpleps(a, b);
176}
177
178static inline __m128 __attribute__((__always_inline__)) _mm_cmpgt_ss(__m128 a, __m128 b)
179{
180  return (__m128)__builtin_ia32_cmpltss(b, a);
181}
182
183static inline __m128 __attribute__((__always_inline__)) _mm_cmpgt_ps(__m128 a, __m128 b)
184{
185  return (__m128)__builtin_ia32_cmpltps(b, a);
186}
187
188static inline __m128 __attribute__((__always_inline__)) _mm_cmpge_ss(__m128 a, __m128 b)
189{
190  return (__m128)__builtin_ia32_cmpless(b, a);
191}
192
193static inline __m128 __attribute__((__always_inline__)) _mm_cmpge_ps(__m128 a, __m128 b)
194{
195  return (__m128)__builtin_ia32_cmpleps(b, a);
196}
197
198static inline __m128 __attribute__((__always_inline__)) _mm_cmpneq_ss(__m128 a, __m128 b)
199{
200  return (__m128)__builtin_ia32_cmpneqss(a, b);
201}
202
203static inline __m128 __attribute__((__always_inline__)) _mm_cmpneq_ps(__m128 a, __m128 b)
204{
205  return (__m128)__builtin_ia32_cmpneqps(a, b);
206}
207
208static inline __m128 __attribute__((__always_inline__)) _mm_cmpnlt_ss(__m128 a, __m128 b)
209{
210  return (__m128)__builtin_ia32_cmpnltss(a, b);
211}
212
213static inline __m128 __attribute__((__always_inline__)) _mm_cmpnlt_ps(__m128 a, __m128 b)
214{
215  return (__m128)__builtin_ia32_cmpnltps(a, b);
216}
217
218static inline __m128 __attribute__((__always_inline__)) _mm_cmpnle_ss(__m128 a, __m128 b)
219{
220  return (__m128)__builtin_ia32_cmpnless(a, b);
221}
222
223static inline __m128 __attribute__((__always_inline__)) _mm_cmpnle_ps(__m128 a, __m128 b)
224{
225  return (__m128)__builtin_ia32_cmpnleps(a, b);
226}
227
228static inline __m128 __attribute__((__always_inline__)) _mm_cmpngt_ss(__m128 a, __m128 b)
229{
230  return (__m128)__builtin_ia32_cmpnltss(b, a);
231}
232
233static inline __m128 __attribute__((__always_inline__)) _mm_cmpngt_ps(__m128 a, __m128 b)
234{
235  return (__m128)__builtin_ia32_cmpnltps(b, a);
236}
237
238static inline __m128 __attribute__((__always_inline__)) _mm_cmpnge_ss(__m128 a, __m128 b)
239{
240  return (__m128)__builtin_ia32_cmpnless(b, a);
241}
242
243static inline __m128 __attribute__((__always_inline__)) _mm_cmpnge_ps(__m128 a, __m128 b)
244{
245  return (__m128)__builtin_ia32_cmpnleps(b, a);
246}
247
248static inline __m128 __attribute__((__always_inline__)) _mm_cmpord_ss(__m128 a, __m128 b)
249{
250  return (__m128)__builtin_ia32_cmpordss(a, b);
251}
252
253static inline __m128 __attribute__((__always_inline__)) _mm_cmpord_ps(__m128 a, __m128 b)
254{
255  return (__m128)__builtin_ia32_cmpordps(a, b);
256}
257
258static inline __m128 __attribute__((__always_inline__)) _mm_cmpunord_ss(__m128 a, __m128 b)
259{
260  return (__m128)__builtin_ia32_cmpunordss(a, b);
261}
262
263static inline __m128 __attribute__((__always_inline__)) _mm_cmpunord_ps(__m128 a, __m128 b)
264{
265  return (__m128)__builtin_ia32_cmpunordps(a, b);
266}
267
268static inline int __attribute__((__always_inline__)) _mm_comieq_ss(__m128 a, __m128 b)
269{
270  return __builtin_ia32_comieq(a, b);
271}
272
273static inline int __attribute__((__always_inline__)) _mm_comilt_ss(__m128 a, __m128 b)
274{
275  return __builtin_ia32_comilt(a, b);
276}
277
278static inline int __attribute__((__always_inline__)) _mm_comile_ss(__m128 a, __m128 b)
279{
280  return __builtin_ia32_comile(a, b);
281}
282
283static inline int __attribute__((__always_inline__)) _mm_comigt_ss(__m128 a, __m128 b)
284{
285  return __builtin_ia32_comigt(a, b);
286}
287
288static inline int __attribute__((__always_inline__)) _mm_comige_ss(__m128 a, __m128 b)
289{
290  return __builtin_ia32_comige(a, b);
291}
292
293static inline int __attribute__((__always_inline__)) _mm_comineq_ss(__m128 a, __m128 b)
294{
295  return __builtin_ia32_comineq(a, b);
296}
297
298static inline int __attribute__((__always_inline__)) _mm_ucomieq_ss(__m128 a, __m128 b)
299{
300  return __builtin_ia32_ucomieq(a, b);
301}
302
303static inline int __attribute__((__always_inline__)) _mm_ucomilt_ss(__m128 a, __m128 b)
304{
305  return __builtin_ia32_ucomilt(a, b);
306}
307
308static inline int __attribute__((__always_inline__)) _mm_ucomile_ss(__m128 a, __m128 b)
309{
310  return __builtin_ia32_ucomile(a, b);
311}
312
313static inline int __attribute__((__always_inline__)) _mm_ucomigt_ss(__m128 a, __m128 b)
314{
315  return __builtin_ia32_ucomigt(a, b);
316}
317
318static inline int __attribute__((__always_inline__)) _mm_ucomige_ss(__m128 a, __m128 b)
319{
320  return __builtin_ia32_ucomige(a, b);
321}
322
323static inline int __attribute__((__always_inline__)) _mm_ucomineq_ss(__m128 a, __m128 b)
324{
325  return __builtin_ia32_ucomineq(a, b);
326}
327
328static inline int __attribute__((__always_inline__)) _mm_cvtss_si32(__m128 a)
329{
330  return __builtin_ia32_cvtss2si(a);
331}
332
333static inline long long __attribute__((__always_inline__)) _mm_cvtss_si64(__m128 a)
334{
335  return __builtin_ia32_cvtss2si64(a);
336}
337
338static inline __m64 __attribute__((__always_inline__)) _mm_cvtps_pi32(__m128 a)
339{
340  return (__m64)__builtin_ia32_cvtps2pi(a);
341}
342
343static inline int __attribute__((__always_inline__)) _mm_cvttss_si32(__m128 a)
344{
345  return __builtin_ia32_cvttss2si(a);
346}
347
348static inline long long __attribute__((__always_inline__)) _mm_cvttss_si64(__m128 a)
349{
350  return __builtin_ia32_cvttss2si64(a);
351}
352
353static inline __m64 __attribute__((__always_inline__)) _mm_cvttps_pi32(__m128 a)
354{
355  return (__m64)__builtin_ia32_cvttps2pi(a);
356}
357
358static inline __m128 __attribute__((__always_inline__)) _mm_cvtsi32_ss(__m128 a, int b)
359{
360  return __builtin_ia32_cvtsi2ss(a, b);
361}
362
363#ifdef __x86_64__
364
365static inline __m128 __attribute__((__always_inline__)) _mm_cvtsi64_ss(__m128 a, long long b)
366{
367  return __builtin_ia32_cvtsi642ss(a, b);
368}
369
370#endif
371
372static inline __m128 __attribute__((__always_inline__)) _mm_cvtpi32_ps(__m128 a, __m64 b)
373{
374  return __builtin_ia32_cvtpi2ps(a, (__v2si)b);
375}
376
377static inline float __attribute__((__always_inline__)) _mm_cvtss_f32(__m128 a)
378{
379  return a[0];
380}
381
382static inline __m128 __attribute__((__always_inline__)) _mm_loadh_pi(__m128 a, __m64 const *p)
383{
384  return __builtin_ia32_loadhps(a, (__v2si *)p);
385}
386
387static inline __m128 __attribute__((__always_inline__)) _mm_loadl_pi(__m128 a, __m64 const *p)
388{
389  return __builtin_ia32_loadlps(a, (__v2si *)p);
390}
391
392static inline __m128 __attribute__((__always_inline__)) _mm_load_ss(float *p)
393{
394  return (__m128){ *p, 0, 0, 0 };
395}
396
397static inline __m128 __attribute__((__always_inline__)) _mm_load1_ps(float *p)
398{
399  return (__m128){ *p, *p, *p, *p };
400}
401
402static inline __m128 __attribute__((__always_inline__)) _mm_load_ps(float *p)
403{
404  return *(__m128*)p;
405}
406
407static inline __m128 __attribute__((__always_inline__)) _mm_loadu_ps(float *p)
408{
409  return __builtin_ia32_loadups(p);
410}
411
412static inline __m128 __attribute__((__always_inline__)) _mm_loadr_ps(float *p)
413{
414  __m128 a = _mm_load_ps(p);
415  return __builtin_shufflevector(a, a, 3, 2, 1, 0);
416}
417
418static inline __m128 __attribute__((__always_inline__)) _mm_set_ss(float w)
419{
420  return (__m128){ w, 0, 0, 0 };
421}
422
423static inline __m128 __attribute__((__always_inline__)) _mm_set1_ps(float w)
424{
425  return (__m128){ w, w, w, w };
426}
427
428// Microsoft specific.
429static inline __m128 __attribute__((__always_inline__)) _mm_set_ps1(float w)
430{
431    return _mm_set1_ps(w);
432}
433
434static inline __m128 __attribute__((__always_inline__)) _mm_set_ps(float z, float y, float x, float w)
435{
436  return (__m128){ w, x, y, z };
437}
438
439static inline __m128 __attribute__((__always_inline__)) _mm_setr_ps(float z, float y, float x, float w)
440{
441  return (__m128){ z, y, x, w };
442}
443
444static inline __m128 __attribute__((__always__inline__)) _mm_setzero_ps(void)
445{
446  return (__m128){ 0, 0, 0, 0 };
447}
448
449static inline void __attribute__((__always__inline__)) _mm_storeh_pi(__m64 *p, __m128 a)
450{
451  __builtin_ia32_storehps((__v2si *)p, a);
452}
453
454static inline void __attribute__((__always__inline__)) _mm_storel_pi(__m64 *p, __m128 a)
455{
456  __builtin_ia32_storelps((__v2si *)p, a);
457}
458
459static inline void __attribute__((__always__inline__)) _mm_store_ss(float *p, __m128 a)
460{
461  *p = a[0];
462}
463
464static inline void __attribute__((__always_inline__)) _mm_storeu_ps(float *p, __m128 a)
465{
466  __builtin_ia32_storeups(p, a);
467}
468
469static inline void __attribute__((__always_inline__)) _mm_store1_ps(float *p, __m128 a)
470{
471  a = __builtin_shufflevector(a, a, 0, 0, 0, 0);
472  _mm_storeu_ps(p, a);
473}
474
475static inline void __attribute__((__always_inline__)) _mm_store_ps(float *p, __m128 a)
476{
477  *(__m128 *)p = a;
478}
479
480static inline void __attribute__((__always_inline__)) _mm_storer_ps(float *p, __m128 a)
481{
482  a = __builtin_shufflevector(a, a, 3, 2, 1, 0);
483  _mm_store_ps(p, a);
484}
485
486#define _MM_HINT_T0 1
487#define _MM_HINT_T1 2
488#define _MM_HINT_T2 3
489#define _MM_HINT_NTA 0
490
491/* FIXME: We have to #define this because "sel" must be a constant integer, and
492   Sema doesn't do any form of constant propagation yet. */
493
494#define _mm_prefetch(a, sel) (__builtin_prefetch((void *)a, 0, sel))
495
496static inline void __attribute__((__always_inline__)) _mm_stream_pi(__m64 *p, __m64 a)
497{
498  __builtin_ia32_movntq(p, a);
499}
500
501static inline void __attribute__((__always_inline__)) _mm_stream_ps(float *p, __m128 a)
502{
503  __builtin_ia32_movntps(p, a);
504}
505
506static inline void __attribute__((__always_inline__)) _mm_sfence(void)
507{
508  __builtin_ia32_sfence();
509}
510
511static inline int __attribute__((__always_inline__)) _mm_extract_pi16(__m64 a, int n)
512{
513  /* FIXME:
514   * This should force n to be an immediate.
515   * This does not use the PEXTRW instruction. From looking at the LLVM source, the
516     instruction doesn't seem to be hooked up.
517   * The code could probably be made better :)
518   */
519  __v4hi b = (__v4hi)a;
520  return b[(n == 0) ? 0 : (n == 1 ? 1 : (n == 2 ? 2 : 3))];
521}
522
523/* FIXME: Implement this. We could add a __builtin_insertelement function that's similar to
524   the already existing __builtin_shufflevector.
525*/
526/*
527static inline __m64 __attribute__((__always_inline__)) _mm_insert_pi16(__m64 a, int d, int n)
528{
529   return (__m64){ 0LL };
530}
531*/
532
533static inline __m64 __attribute__((__always_inline__)) _mm_max_pi16(__m64 a, __m64 b)
534{
535  return (__m64)__builtin_ia32_pmaxsw((__v4hi)a, (__v4hi)b);
536}
537
538static inline __m64 __attribute__((__always_inline__)) _mm_max_pu8(__m64 a, __m64 b)
539{
540  return (__m64)__builtin_ia32_pmaxub((__v8qi)a, (__v8qi)b);
541}
542
543static inline __m64 __attribute__((__always_inline__)) _mm_min_pi16(__m64 a, __m64 b)
544{
545  return (__m64)__builtin_ia32_pminsw((__v4hi)a, (__v4hi)b);
546}
547
548static inline __m64 __attribute__((__always_inline__)) _mm_min_pu8(__m64 a, __m64 b)
549{
550  return (__m64)__builtin_ia32_pminub((__v8qi)a, (__v8qi)b);
551}
552
553static inline int __attribute__((__always_inline__)) _mm_movemask_pi8(__m64 a)
554{
555  return __builtin_ia32_pmovmskb((__v8qi)a);
556}
557
558static inline __m64 __attribute__((__always_inline__)) _mm_mulhi_pu16(__m64 a, __m64 b)
559{
560  return (__m64)__builtin_ia32_pmulhuw((__v4hi)a, (__v4hi)b);
561}
562
563#define _mm_shuffle_pi16(a, n) ((__m64)__builtin_ia32_pshufw((__v4hi)a, n))
564
565static inline void __attribute__((__always_inline__)) _mm_maskmove_si64(__m64 d, __m64 n, char *p)
566{
567  __builtin_ia32_maskmovq((__v8qi)d, (__v8qi)n, p);
568}
569
570static inline __m64 __attribute__((__always_inline__)) _mm_avg_pu8(__m64 a, __m64 b)
571{
572  return (__m64)__builtin_ia32_pavgb((__v8qi)a, (__v8qi)b);
573}
574
575static inline __m64 __attribute__((__always_inline__)) _mm_avg_pu16(__m64 a, __m64 b)
576{
577  return (__m64)__builtin_ia32_pavgw((__v4hi)a, (__v4hi)b);
578}
579
580static inline __m64 __attribute__((__always_inline___)) _mm_sad_pu8(__m64 a, __m64 b)
581{
582  return (__m64)__builtin_ia32_psadbw((__v8qi)a, (__v8qi)b);
583}
584
585static inline unsigned int __attribute__((__always_inline___)) _mm_getcsr(void)
586{
587  return __builtin_ia32_stmxcsr();
588}
589
590static inline void __attribute__((__always_inline__)) _mm_setcsr(unsigned int i)
591{
592  __builtin_ia32_ldmxcsr(i);
593}
594
595#define _mm_shuffle_ps(a, b, mask) (__builtin_ia32_shufps(a, b, mask))
596
597static inline __m128 __attribute__((__always_inline__)) _mm_unpackhi_ps(__m128 a, __m128 b)
598{
599  return __builtin_shufflevector(a, b, 2, 6, 3, 7);
600}
601
602static inline __m128 __attribute__((__always_inline__)) _mm_unpacklo_ps(__m128 a, __m128 b)
603{
604  return __builtin_shufflevector(a, b, 0, 4, 1, 5);
605}
606
607static inline __m128 __attribute__((__always_inline__)) _mm_move_ss(__m128 a, __m128 b)
608{
609  return __builtin_shufflevector(a, b, 4, 1, 2, 3);
610}
611
612static inline __m128 __attribute__((__always_inline__)) _mm_movehl_ps(__m128 a, __m128 b)
613{
614  return __builtin_shufflevector(a, b, 6, 7, 2, 3);
615}
616
617static inline __m128 __attribute__((__always_inline__)) _mm_movelh_ps(__m128 a, __m128 b)
618{
619  return __builtin_shufflevector(a, b, 0, 1, 4, 5);
620}
621
622static inline __m128 __attribute__((__always_inline__)) _mm_cvtpi16_ps(__m64 a)
623{
624  __m64 b, c;
625  __m128 r;
626
627  b = _mm_setzero_si64();
628  b = _mm_cmpgt_pi16(b, a);
629  c = _mm_unpackhi_pi16(a, b);
630  r = _mm_setzero_ps();
631  r = _mm_cvtpi32_ps(r, c);
632  r = _mm_movelh_ps(r, r);
633  c = _mm_unpacklo_pi16(a, b);
634  r = _mm_cvtpi32_ps(r, c);
635
636  return r;
637}
638
639static inline __m128 __attribute__((__always_inline__)) _mm_cvtpu16_ps(__m64 a)
640{
641  __m64 b, c;
642  __m128 r;
643
644  b = _mm_setzero_si64();
645  c = _mm_unpackhi_pi16(a, b);
646  r = _mm_setzero_ps();
647  r = _mm_cvtpi32_ps(r, c);
648  r = _mm_movelh_ps(r, r);
649  c = _mm_unpacklo_pi16(a, b);
650  r = _mm_cvtpi32_ps(r, c);
651
652  return r;
653}
654
655static inline __m128 __attribute__((__always_inline__)) _mm_cvtpi8_ps(__m64 a)
656{
657  __m64 b;
658
659  b = _mm_setzero_si64();
660  b = _mm_cmpgt_pi8(b, a);
661  b = _mm_unpacklo_pi8(a, b);
662
663  return _mm_cvtpi16_ps(b);
664}
665
666static inline __m128 __attribute__((__always_inline__)) _mm_cvtpu8_ps(__m64 a)
667{
668  __m64 b;
669
670  b = _mm_setzero_si64();
671  b = _mm_unpacklo_pi8(a, b);
672
673  return _mm_cvtpi16_ps(b);
674}
675
676static inline __m128 __attribute__((__always_inline__)) _mm_cvtpi32x2_ps(__m64 a, __m64 b)
677{
678  __m128 c;
679
680  c = _mm_setzero_ps();
681  c = _mm_cvtpi32_ps(c, b);
682  c = _mm_movelh_ps(c, c);
683
684  return _mm_cvtpi32_ps(c, a);
685}
686
687static inline __m64 __attribute__((__always_inline__)) _mm_cvtps_pi16(__m128 a)
688{
689  __m64 b, c;
690
691  b = _mm_cvtps_pi32(a);
692  a = _mm_movehl_ps(a, a);
693  c = _mm_cvtps_pi32(a);
694
695  return _mm_packs_pi16(b, c);
696}
697
698static inline __m64 __attribute__((__always_inline__)) _mm_cvtps_pi8(__m128 a)
699{
700  __m64 b, c;
701
702  b = _mm_cvtps_pi16(a);
703  c = _mm_setzero_si64();
704
705  return _mm_packs_pi16(b, c);
706}
707
708static inline int __attribute__((__always_inline__)) _mm_movemask_ps(__m128 a)
709{
710  return __builtin_ia32_movmskps(a);
711}
712
713#define _MM_SHUFFLE(z, y, x, w) (((z) << 6) | ((y) << 4) | ((x) << 2) | (w))
714
715#define _MM_MASK_MASK (0x1f80)
716#define _MM_EXCEPT_MASK (0x003f)
717#define _MM_FLUSH_ZERO_MASK (0x8000)
718#define _MM_ROUND_MASK (0x6000)
719
720#define _MM_GET_EXCEPTION_MASK() (_mm_getcsr() & _MM_MASK_MASK)
721#define _MM_GET_EXCEPTION_STATE() (_mm_getcsr() & _MM_EXCEPT_MASK)
722#define _MM_GET_FLUSH_ZERO_MODE() (_mm_getcsr() & _MM_FLUSH_ZERO_MASK)
723#define _MM_GET_ROUNDING_MODE() (_mm_getcsr() & _MM_ROUND_MASK)
724
725#define _MM_SET_EXCEPTION_MASK(x) (_mm_setcsr((_mm_getcsr() & ~_MM_MASK_MASK) | (x)))
726#define _MM_SET_EXCEPTION_STATE(x) (_mm_setcsr((_mm_getcsr() & ~_MM_EXCEPT_MASK) | (x)))
727#define _MM_SET_FLUSH_ZERO_MODE(x) (_mm_setcsr((_mm_getcsr() & ~_MM_FLUSH_ZERO_MASK) | (x)))
728#define _MM_SET_ROUNDING_MODE(x) (_mm_setcsr((_mm_getcsr() & ~_MM_ROUND_MASK) | (x)))
729
730#define _MM_TRANSPOSE4_PS(row0, row1, row2, row3) \
731do { \
732  __m128 tmp3, tmp2, tmp1, tmp0; \
733  tmp0 = _mm_unpacklo_ps((row0), (row1)); \
734  tmp2 = _mm_unpacklo_ps((row2), (row3)); \
735  tmp1 = _mm_unpackhi_ps((row0), (row1)); \
736  tmp3 = _mm_unpackhi_ps((row2), (row3)); \
737  (row0) = _mm_movelh_ps(tmp0, tmp2); \
738  (row1) = _mm_movehl_ps(tmp2, tmp0); \
739  (row2) = _mm_movelh_ps(tmp1, tmp3); \
740  (row3) = _mm_movelh_ps(tmp3, tmp1); \
741} while (0)
742
743#include <emmintrin.h>
744
745#endif /* __SSE__ */
746
747#endif /* __XMMINTRIN_H */
748