1 /*===---- xmmintrin.h - SSE intrinsics -------------------------------------===
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:
10 * The above copyright notice and this permission notice shall be included in
11 * all copies or substantial portions of the Software.
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
21 *===-----------------------------------------------------------------------===
28 #error "SSE instruction set not enabled"
33 typedef int __v4si __attribute__((__vector_size__(16)));
34 typedef float __v4sf __attribute__((__vector_size__(16)));
35 typedef float __m128 __attribute__((__vector_size__(16)));
37 // This header should only be included in a hosted environment as it depends on
38 // a standard library to provide allocation routines.
40 #include <mm_malloc.h>
43 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
44 _mm_add_ss(__m128 __a, __m128 __b)
50 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
51 _mm_add_ps(__m128 __a, __m128 __b)
56 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
57 _mm_sub_ss(__m128 __a, __m128 __b)
63 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
64 _mm_sub_ps(__m128 __a, __m128 __b)
69 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
70 _mm_mul_ss(__m128 __a, __m128 __b)
76 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
77 _mm_mul_ps(__m128 __a, __m128 __b)
82 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
83 _mm_div_ss(__m128 __a, __m128 __b)
89 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
90 _mm_div_ps(__m128 __a, __m128 __b)
95 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
96 _mm_sqrt_ss(__m128 __a)
98 __m128 __c = __builtin_ia32_sqrtss(__a);
99 return (__m128) { __c[0], __a[1], __a[2], __a[3] };
102 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
103 _mm_sqrt_ps(__m128 __a)
105 return __builtin_ia32_sqrtps(__a);
108 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
109 _mm_rcp_ss(__m128 __a)
111 __m128 __c = __builtin_ia32_rcpss(__a);
112 return (__m128) { __c[0], __a[1], __a[2], __a[3] };
115 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
116 _mm_rcp_ps(__m128 __a)
118 return __builtin_ia32_rcpps(__a);
121 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
122 _mm_rsqrt_ss(__m128 __a)
124 __m128 __c = __builtin_ia32_rsqrtss(__a);
125 return (__m128) { __c[0], __a[1], __a[2], __a[3] };
128 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
129 _mm_rsqrt_ps(__m128 __a)
131 return __builtin_ia32_rsqrtps(__a);
134 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
135 _mm_min_ss(__m128 __a, __m128 __b)
137 return __builtin_ia32_minss(__a, __b);
140 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
141 _mm_min_ps(__m128 __a, __m128 __b)
143 return __builtin_ia32_minps(__a, __b);
146 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
147 _mm_max_ss(__m128 __a, __m128 __b)
149 return __builtin_ia32_maxss(__a, __b);
152 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
153 _mm_max_ps(__m128 __a, __m128 __b)
155 return __builtin_ia32_maxps(__a, __b);
158 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
159 _mm_and_ps(__m128 __a, __m128 __b)
161 return (__m128)((__v4si)__a & (__v4si)__b);
164 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
165 _mm_andnot_ps(__m128 __a, __m128 __b)
167 return (__m128)(~(__v4si)__a & (__v4si)__b);
170 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
171 _mm_or_ps(__m128 __a, __m128 __b)
173 return (__m128)((__v4si)__a | (__v4si)__b);
176 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
177 _mm_xor_ps(__m128 __a, __m128 __b)
179 return (__m128)((__v4si)__a ^ (__v4si)__b);
182 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
183 _mm_cmpeq_ss(__m128 __a, __m128 __b)
185 return (__m128)__builtin_ia32_cmpss(__a, __b, 0);
188 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
189 _mm_cmpeq_ps(__m128 __a, __m128 __b)
191 return (__m128)__builtin_ia32_cmpps(__a, __b, 0);
194 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
195 _mm_cmplt_ss(__m128 __a, __m128 __b)
197 return (__m128)__builtin_ia32_cmpss(__a, __b, 1);
200 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
201 _mm_cmplt_ps(__m128 __a, __m128 __b)
203 return (__m128)__builtin_ia32_cmpps(__a, __b, 1);
206 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
207 _mm_cmple_ss(__m128 __a, __m128 __b)
209 return (__m128)__builtin_ia32_cmpss(__a, __b, 2);
212 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
213 _mm_cmple_ps(__m128 __a, __m128 __b)
215 return (__m128)__builtin_ia32_cmpps(__a, __b, 2);
218 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
219 _mm_cmpgt_ss(__m128 __a, __m128 __b)
221 return (__m128)__builtin_ia32_cmpss(__b, __a, 1);
224 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
225 _mm_cmpgt_ps(__m128 __a, __m128 __b)
227 return (__m128)__builtin_ia32_cmpps(__b, __a, 1);
230 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
231 _mm_cmpge_ss(__m128 __a, __m128 __b)
233 return (__m128)__builtin_ia32_cmpss(__b, __a, 2);
236 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
237 _mm_cmpge_ps(__m128 __a, __m128 __b)
239 return (__m128)__builtin_ia32_cmpps(__b, __a, 2);
242 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
243 _mm_cmpneq_ss(__m128 __a, __m128 __b)
245 return (__m128)__builtin_ia32_cmpss(__a, __b, 4);
248 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
249 _mm_cmpneq_ps(__m128 __a, __m128 __b)
251 return (__m128)__builtin_ia32_cmpps(__a, __b, 4);
254 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
255 _mm_cmpnlt_ss(__m128 __a, __m128 __b)
257 return (__m128)__builtin_ia32_cmpss(__a, __b, 5);
260 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
261 _mm_cmpnlt_ps(__m128 __a, __m128 __b)
263 return (__m128)__builtin_ia32_cmpps(__a, __b, 5);
266 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
267 _mm_cmpnle_ss(__m128 __a, __m128 __b)
269 return (__m128)__builtin_ia32_cmpss(__a, __b, 6);
272 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
273 _mm_cmpnle_ps(__m128 __a, __m128 __b)
275 return (__m128)__builtin_ia32_cmpps(__a, __b, 6);
278 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
279 _mm_cmpngt_ss(__m128 __a, __m128 __b)
281 return (__m128)__builtin_ia32_cmpss(__b, __a, 5);
284 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
285 _mm_cmpngt_ps(__m128 __a, __m128 __b)
287 return (__m128)__builtin_ia32_cmpps(__b, __a, 5);
290 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
291 _mm_cmpnge_ss(__m128 __a, __m128 __b)
293 return (__m128)__builtin_ia32_cmpss(__b, __a, 6);
296 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
297 _mm_cmpnge_ps(__m128 __a, __m128 __b)
299 return (__m128)__builtin_ia32_cmpps(__b, __a, 6);
302 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
303 _mm_cmpord_ss(__m128 __a, __m128 __b)
305 return (__m128)__builtin_ia32_cmpss(__a, __b, 7);
308 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
309 _mm_cmpord_ps(__m128 __a, __m128 __b)
311 return (__m128)__builtin_ia32_cmpps(__a, __b, 7);
314 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
315 _mm_cmpunord_ss(__m128 __a, __m128 __b)
317 return (__m128)__builtin_ia32_cmpss(__a, __b, 3);
320 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
321 _mm_cmpunord_ps(__m128 __a, __m128 __b)
323 return (__m128)__builtin_ia32_cmpps(__a, __b, 3);
326 static __inline__ int __attribute__((__always_inline__, __nodebug__))
327 _mm_comieq_ss(__m128 __a, __m128 __b)
329 return __builtin_ia32_comieq(__a, __b);
332 static __inline__ int __attribute__((__always_inline__, __nodebug__))
333 _mm_comilt_ss(__m128 __a, __m128 __b)
335 return __builtin_ia32_comilt(__a, __b);
338 static __inline__ int __attribute__((__always_inline__, __nodebug__))
339 _mm_comile_ss(__m128 __a, __m128 __b)
341 return __builtin_ia32_comile(__a, __b);
344 static __inline__ int __attribute__((__always_inline__, __nodebug__))
345 _mm_comigt_ss(__m128 __a, __m128 __b)
347 return __builtin_ia32_comigt(__a, __b);
350 static __inline__ int __attribute__((__always_inline__, __nodebug__))
351 _mm_comige_ss(__m128 __a, __m128 __b)
353 return __builtin_ia32_comige(__a, __b);
356 static __inline__ int __attribute__((__always_inline__, __nodebug__))
357 _mm_comineq_ss(__m128 __a, __m128 __b)
359 return __builtin_ia32_comineq(__a, __b);
362 static __inline__ int __attribute__((__always_inline__, __nodebug__))
363 _mm_ucomieq_ss(__m128 __a, __m128 __b)
365 return __builtin_ia32_ucomieq(__a, __b);
368 static __inline__ int __attribute__((__always_inline__, __nodebug__))
369 _mm_ucomilt_ss(__m128 __a, __m128 __b)
371 return __builtin_ia32_ucomilt(__a, __b);
374 static __inline__ int __attribute__((__always_inline__, __nodebug__))
375 _mm_ucomile_ss(__m128 __a, __m128 __b)
377 return __builtin_ia32_ucomile(__a, __b);
380 static __inline__ int __attribute__((__always_inline__, __nodebug__))
381 _mm_ucomigt_ss(__m128 __a, __m128 __b)
383 return __builtin_ia32_ucomigt(__a, __b);
386 static __inline__ int __attribute__((__always_inline__, __nodebug__))
387 _mm_ucomige_ss(__m128 __a, __m128 __b)
389 return __builtin_ia32_ucomige(__a, __b);
392 static __inline__ int __attribute__((__always_inline__, __nodebug__))
393 _mm_ucomineq_ss(__m128 __a, __m128 __b)
395 return __builtin_ia32_ucomineq(__a, __b);
398 static __inline__ int __attribute__((__always_inline__, __nodebug__))
399 _mm_cvtss_si32(__m128 __a)
401 return __builtin_ia32_cvtss2si(__a);
404 static __inline__ int __attribute__((__always_inline__, __nodebug__))
405 _mm_cvt_ss2si(__m128 __a)
407 return _mm_cvtss_si32(__a);
412 static __inline__ long long __attribute__((__always_inline__, __nodebug__))
413 _mm_cvtss_si64(__m128 __a)
415 return __builtin_ia32_cvtss2si64(__a);
420 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
421 _mm_cvtps_pi32(__m128 __a)
423 return (__m64)__builtin_ia32_cvtps2pi(__a);
426 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
427 _mm_cvt_ps2pi(__m128 __a)
429 return _mm_cvtps_pi32(__a);
432 static __inline__ int __attribute__((__always_inline__, __nodebug__))
433 _mm_cvttss_si32(__m128 __a)
438 static __inline__ int __attribute__((__always_inline__, __nodebug__))
439 _mm_cvtt_ss2si(__m128 __a)
441 return _mm_cvttss_si32(__a);
444 static __inline__ long long __attribute__((__always_inline__, __nodebug__))
445 _mm_cvttss_si64(__m128 __a)
450 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
451 _mm_cvttps_pi32(__m128 __a)
453 return (__m64)__builtin_ia32_cvttps2pi(__a);
456 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
457 _mm_cvtt_ps2pi(__m128 __a)
459 return _mm_cvttps_pi32(__a);
462 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
463 _mm_cvtsi32_ss(__m128 __a, int __b)
469 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
470 _mm_cvt_si2ss(__m128 __a, int __b)
472 return _mm_cvtsi32_ss(__a, __b);
477 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
478 _mm_cvtsi64_ss(__m128 __a, long long __b)
486 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
487 _mm_cvtpi32_ps(__m128 __a, __m64 __b)
489 return __builtin_ia32_cvtpi2ps(__a, (__v2si)__b);
492 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
493 _mm_cvt_pi2ps(__m128 __a, __m64 __b)
495 return _mm_cvtpi32_ps(__a, __b);
498 static __inline__ float __attribute__((__always_inline__, __nodebug__))
499 _mm_cvtss_f32(__m128 __a)
504 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
505 _mm_loadh_pi(__m128 __a, const __m64 *__p)
507 typedef float __mm_loadh_pi_v2f32 __attribute__((__vector_size__(8)));
508 struct __mm_loadh_pi_struct {
509 __mm_loadh_pi_v2f32 __u;
510 } __attribute__((__packed__, __may_alias__));
511 __mm_loadh_pi_v2f32 __b = ((struct __mm_loadh_pi_struct*)__p)->__u;
512 __m128 __bb = __builtin_shufflevector(__b, __b, 0, 1, 0, 1);
513 return __builtin_shufflevector(__a, __bb, 0, 1, 4, 5);
516 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
517 _mm_loadl_pi(__m128 __a, const __m64 *__p)
519 typedef float __mm_loadl_pi_v2f32 __attribute__((__vector_size__(8)));
520 struct __mm_loadl_pi_struct {
521 __mm_loadl_pi_v2f32 __u;
522 } __attribute__((__packed__, __may_alias__));
523 __mm_loadl_pi_v2f32 __b = ((struct __mm_loadl_pi_struct*)__p)->__u;
524 __m128 __bb = __builtin_shufflevector(__b, __b, 0, 1, 0, 1);
525 return __builtin_shufflevector(__a, __bb, 4, 5, 2, 3);
528 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
529 _mm_load_ss(const float *__p)
531 struct __mm_load_ss_struct {
533 } __attribute__((__packed__, __may_alias__));
534 float __u = ((struct __mm_load_ss_struct*)__p)->__u;
535 return (__m128){ __u, 0, 0, 0 };
538 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
539 _mm_load1_ps(const float *__p)
541 struct __mm_load1_ps_struct {
543 } __attribute__((__packed__, __may_alias__));
544 float __u = ((struct __mm_load1_ps_struct*)__p)->__u;
545 return (__m128){ __u, __u, __u, __u };
548 #define _mm_load_ps1(p) _mm_load1_ps(p)
550 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
551 _mm_load_ps(const float *__p)
553 return *(__m128*)__p;
556 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
557 _mm_loadu_ps(const float *__p)
561 } __attribute__((__packed__, __may_alias__));
562 return ((struct __loadu_ps*)__p)->__v;
565 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
566 _mm_loadr_ps(const float *__p)
568 __m128 __a = _mm_load_ps(__p);
569 return __builtin_shufflevector(__a, __a, 3, 2, 1, 0);
572 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
573 _mm_set_ss(float __w)
575 return (__m128){ __w, 0, 0, 0 };
578 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
579 _mm_set1_ps(float __w)
581 return (__m128){ __w, __w, __w, __w };
584 // Microsoft specific.
585 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
586 _mm_set_ps1(float __w)
588 return _mm_set1_ps(__w);
591 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
592 _mm_set_ps(float __z, float __y, float __x, float __w)
594 return (__m128){ __w, __x, __y, __z };
597 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
598 _mm_setr_ps(float __z, float __y, float __x, float __w)
600 return (__m128){ __z, __y, __x, __w };
603 static __inline__ __m128 __attribute__((__always_inline__))
606 return (__m128){ 0, 0, 0, 0 };
609 static __inline__ void __attribute__((__always_inline__))
610 _mm_storeh_pi(__m64 *__p, __m128 __a)
612 __builtin_ia32_storehps((__v2si *)__p, __a);
615 static __inline__ void __attribute__((__always_inline__))
616 _mm_storel_pi(__m64 *__p, __m128 __a)
618 __builtin_ia32_storelps((__v2si *)__p, __a);
621 static __inline__ void __attribute__((__always_inline__))
622 _mm_store_ss(float *__p, __m128 __a)
624 struct __mm_store_ss_struct {
626 } __attribute__((__packed__, __may_alias__));
627 ((struct __mm_store_ss_struct*)__p)->__u = __a[0];
630 static __inline__ void __attribute__((__always_inline__, __nodebug__))
631 _mm_storeu_ps(float *__p, __m128 __a)
633 __builtin_ia32_storeups(__p, __a);
636 static __inline__ void __attribute__((__always_inline__, __nodebug__))
637 _mm_store1_ps(float *__p, __m128 __a)
639 __a = __builtin_shufflevector(__a, __a, 0, 0, 0, 0);
640 _mm_storeu_ps(__p, __a);
643 static __inline__ void __attribute__((__always_inline__, __nodebug__))
644 _mm_store_ps1(float *__p, __m128 __a)
646 return _mm_store1_ps(__p, __a);
649 static __inline__ void __attribute__((__always_inline__, __nodebug__))
650 _mm_store_ps(float *__p, __m128 __a)
652 *(__m128 *)__p = __a;
655 static __inline__ void __attribute__((__always_inline__, __nodebug__))
656 _mm_storer_ps(float *__p, __m128 __a)
658 __a = __builtin_shufflevector(__a, __a, 3, 2, 1, 0);
659 _mm_store_ps(__p, __a);
662 #define _MM_HINT_T0 3
663 #define _MM_HINT_T1 2
664 #define _MM_HINT_T2 1
665 #define _MM_HINT_NTA 0
667 /* FIXME: We have to #define this because "sel" must be a constant integer, and
668 Sema doesn't do any form of constant propagation yet. */
670 #define _mm_prefetch(a, sel) (__builtin_prefetch((void *)(a), 0, (sel)))
672 static __inline__ void __attribute__((__always_inline__, __nodebug__))
673 _mm_stream_pi(__m64 *__p, __m64 __a)
675 __builtin_ia32_movntq(__p, __a);
678 static __inline__ void __attribute__((__always_inline__, __nodebug__))
679 _mm_stream_ps(float *__p, __m128 __a)
681 __builtin_ia32_movntps(__p, __a);
684 static __inline__ void __attribute__((__always_inline__, __nodebug__))
687 __builtin_ia32_sfence();
690 static __inline__ int __attribute__((__always_inline__, __nodebug__))
691 _mm_extract_pi16(__m64 __a, int __n)
693 __v4hi __b = (__v4hi)__a;
694 return (unsigned short)__b[__n & 3];
697 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
698 _mm_insert_pi16(__m64 __a, int __d, int __n)
700 __v4hi __b = (__v4hi)__a;
705 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
706 _mm_max_pi16(__m64 __a, __m64 __b)
708 return (__m64)__builtin_ia32_pmaxsw((__v4hi)__a, (__v4hi)__b);
711 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
712 _mm_max_pu8(__m64 __a, __m64 __b)
714 return (__m64)__builtin_ia32_pmaxub((__v8qi)__a, (__v8qi)__b);
717 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
718 _mm_min_pi16(__m64 __a, __m64 __b)
720 return (__m64)__builtin_ia32_pminsw((__v4hi)__a, (__v4hi)__b);
723 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
724 _mm_min_pu8(__m64 __a, __m64 __b)
726 return (__m64)__builtin_ia32_pminub((__v8qi)__a, (__v8qi)__b);
729 static __inline__ int __attribute__((__always_inline__, __nodebug__))
730 _mm_movemask_pi8(__m64 __a)
732 return __builtin_ia32_pmovmskb((__v8qi)__a);
735 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
736 _mm_mulhi_pu16(__m64 __a, __m64 __b)
738 return (__m64)__builtin_ia32_pmulhuw((__v4hi)__a, (__v4hi)__b);
741 #define _mm_shuffle_pi16(a, n) __extension__ ({ \
743 (__m64)__builtin_ia32_pshufw((__v4hi)__a, (n)); })
745 static __inline__ void __attribute__((__always_inline__, __nodebug__))
746 _mm_maskmove_si64(__m64 __d, __m64 __n, char *__p)
748 __builtin_ia32_maskmovq((__v8qi)__d, (__v8qi)__n, __p);
751 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
752 _mm_avg_pu8(__m64 __a, __m64 __b)
754 return (__m64)__builtin_ia32_pavgb((__v8qi)__a, (__v8qi)__b);
757 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
758 _mm_avg_pu16(__m64 __a, __m64 __b)
760 return (__m64)__builtin_ia32_pavgw((__v4hi)__a, (__v4hi)__b);
763 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
764 _mm_sad_pu8(__m64 __a, __m64 __b)
766 return (__m64)__builtin_ia32_psadbw((__v8qi)__a, (__v8qi)__b);
769 static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__))
772 return __builtin_ia32_stmxcsr();
775 static __inline__ void __attribute__((__always_inline__, __nodebug__))
776 _mm_setcsr(unsigned int __i)
778 __builtin_ia32_ldmxcsr(__i);
781 #define _mm_shuffle_ps(a, b, mask) __extension__ ({ \
784 (__m128)__builtin_shufflevector((__v4sf)__a, (__v4sf)__b, \
785 (mask) & 0x3, ((mask) & 0xc) >> 2, \
786 (((mask) & 0x30) >> 4) + 4, \
787 (((mask) & 0xc0) >> 6) + 4); })
789 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
790 _mm_unpackhi_ps(__m128 __a, __m128 __b)
792 return __builtin_shufflevector(__a, __b, 2, 6, 3, 7);
795 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
796 _mm_unpacklo_ps(__m128 __a, __m128 __b)
798 return __builtin_shufflevector(__a, __b, 0, 4, 1, 5);
801 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
802 _mm_move_ss(__m128 __a, __m128 __b)
804 return __builtin_shufflevector(__a, __b, 4, 1, 2, 3);
807 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
808 _mm_movehl_ps(__m128 __a, __m128 __b)
810 return __builtin_shufflevector(__a, __b, 6, 7, 2, 3);
813 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
814 _mm_movelh_ps(__m128 __a, __m128 __b)
816 return __builtin_shufflevector(__a, __b, 0, 1, 4, 5);
819 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
820 _mm_cvtpi16_ps(__m64 __a)
825 __b = _mm_setzero_si64();
826 __b = _mm_cmpgt_pi16(__b, __a);
827 __c = _mm_unpackhi_pi16(__a, __b);
828 __r = _mm_setzero_ps();
829 __r = _mm_cvtpi32_ps(__r, __c);
830 __r = _mm_movelh_ps(__r, __r);
831 __c = _mm_unpacklo_pi16(__a, __b);
832 __r = _mm_cvtpi32_ps(__r, __c);
837 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
838 _mm_cvtpu16_ps(__m64 __a)
843 __b = _mm_setzero_si64();
844 __c = _mm_unpackhi_pi16(__a, __b);
845 __r = _mm_setzero_ps();
846 __r = _mm_cvtpi32_ps(__r, __c);
847 __r = _mm_movelh_ps(__r, __r);
848 __c = _mm_unpacklo_pi16(__a, __b);
849 __r = _mm_cvtpi32_ps(__r, __c);
854 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
855 _mm_cvtpi8_ps(__m64 __a)
859 __b = _mm_setzero_si64();
860 __b = _mm_cmpgt_pi8(__b, __a);
861 __b = _mm_unpacklo_pi8(__a, __b);
863 return _mm_cvtpi16_ps(__b);
866 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
867 _mm_cvtpu8_ps(__m64 __a)
871 __b = _mm_setzero_si64();
872 __b = _mm_unpacklo_pi8(__a, __b);
874 return _mm_cvtpi16_ps(__b);
877 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
878 _mm_cvtpi32x2_ps(__m64 __a, __m64 __b)
882 __c = _mm_setzero_ps();
883 __c = _mm_cvtpi32_ps(__c, __b);
884 __c = _mm_movelh_ps(__c, __c);
886 return _mm_cvtpi32_ps(__c, __a);
889 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
890 _mm_cvtps_pi16(__m128 __a)
894 __b = _mm_cvtps_pi32(__a);
895 __a = _mm_movehl_ps(__a, __a);
896 __c = _mm_cvtps_pi32(__a);
898 return _mm_packs_pi16(__b, __c);
901 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
902 _mm_cvtps_pi8(__m128 __a)
906 __b = _mm_cvtps_pi16(__a);
907 __c = _mm_setzero_si64();
909 return _mm_packs_pi16(__b, __c);
912 static __inline__ int __attribute__((__always_inline__, __nodebug__))
913 _mm_movemask_ps(__m128 __a)
915 return __builtin_ia32_movmskps(__a);
918 #define _MM_SHUFFLE(z, y, x, w) (((z) << 6) | ((y) << 4) | ((x) << 2) | (w))
920 #define _MM_EXCEPT_INVALID (0x0001)
921 #define _MM_EXCEPT_DENORM (0x0002)
922 #define _MM_EXCEPT_DIV_ZERO (0x0004)
923 #define _MM_EXCEPT_OVERFLOW (0x0008)
924 #define _MM_EXCEPT_UNDERFLOW (0x0010)
925 #define _MM_EXCEPT_INEXACT (0x0020)
926 #define _MM_EXCEPT_MASK (0x003f)
928 #define _MM_MASK_INVALID (0x0080)
929 #define _MM_MASK_DENORM (0x0100)
930 #define _MM_MASK_DIV_ZERO (0x0200)
931 #define _MM_MASK_OVERFLOW (0x0400)
932 #define _MM_MASK_UNDERFLOW (0x0800)
933 #define _MM_MASK_INEXACT (0x1000)
934 #define _MM_MASK_MASK (0x1f80)
936 #define _MM_ROUND_NEAREST (0x0000)
937 #define _MM_ROUND_DOWN (0x2000)
938 #define _MM_ROUND_UP (0x4000)
939 #define _MM_ROUND_TOWARD_ZERO (0x6000)
940 #define _MM_ROUND_MASK (0x6000)
942 #define _MM_FLUSH_ZERO_MASK (0x8000)
943 #define _MM_FLUSH_ZERO_ON (0x8000)
944 #define _MM_FLUSH_ZERO_OFF (0x0000)
946 #define _MM_GET_EXCEPTION_MASK() (_mm_getcsr() & _MM_MASK_MASK)
947 #define _MM_GET_EXCEPTION_STATE() (_mm_getcsr() & _MM_EXCEPT_MASK)
948 #define _MM_GET_FLUSH_ZERO_MODE() (_mm_getcsr() & _MM_FLUSH_ZERO_MASK)
949 #define _MM_GET_ROUNDING_MODE() (_mm_getcsr() & _MM_ROUND_MASK)
951 #define _MM_SET_EXCEPTION_MASK(x) (_mm_setcsr((_mm_getcsr() & ~_MM_MASK_MASK) | (x)))
952 #define _MM_SET_EXCEPTION_STATE(x) (_mm_setcsr((_mm_getcsr() & ~_MM_EXCEPT_MASK) | (x)))
953 #define _MM_SET_FLUSH_ZERO_MODE(x) (_mm_setcsr((_mm_getcsr() & ~_MM_FLUSH_ZERO_MASK) | (x)))
954 #define _MM_SET_ROUNDING_MODE(x) (_mm_setcsr((_mm_getcsr() & ~_MM_ROUND_MASK) | (x)))
956 #define _MM_TRANSPOSE4_PS(row0, row1, row2, row3) \
958 __m128 tmp3, tmp2, tmp1, tmp0; \
959 tmp0 = _mm_unpacklo_ps((row0), (row1)); \
960 tmp2 = _mm_unpacklo_ps((row2), (row3)); \
961 tmp1 = _mm_unpackhi_ps((row0), (row1)); \
962 tmp3 = _mm_unpackhi_ps((row2), (row3)); \
963 (row0) = _mm_movelh_ps(tmp0, tmp2); \
964 (row1) = _mm_movehl_ps(tmp2, tmp0); \
965 (row2) = _mm_movelh_ps(tmp1, tmp3); \
966 (row3) = _mm_movehl_ps(tmp3, tmp1); \
969 /* Aliases for compatibility. */
970 #define _m_pextrw _mm_extract_pi16
971 #define _m_pinsrw _mm_insert_pi16
972 #define _m_pmaxsw _mm_max_pi16
973 #define _m_pmaxub _mm_max_pu8
974 #define _m_pminsw _mm_min_pi16
975 #define _m_pminub _mm_min_pu8
976 #define _m_pmovmskb _mm_movemask_pi8
977 #define _m_pmulhuw _mm_mulhi_pu16
978 #define _m_pshufw _mm_shuffle_pi16
979 #define _m_maskmovq _mm_maskmove_si64
980 #define _m_pavgb _mm_avg_pu8
981 #define _m_pavgw _mm_avg_pu16
982 #define _m_psadbw _mm_sad_pu8
986 #if !__has_feature(modules)
987 /* Ugly hack for backwards-compatibility (compatible with gcc) */
989 #include <emmintrin.h>
995 #endif /* __XMMINTRIN_H */