Home | History | Annotate | Download | only in Headers 
      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 
     33 typedef int __v4si __attribute__((__vector_size__(16)));
     34 typedef float __v4sf __attribute__((__vector_size__(16)));
     35 typedef float __m128 __attribute__((__vector_size__(16)));
     36 
     37 /* This header should only be included in a hosted environment as it depends on
     38  * a standard library to provide allocation routines. */
     39 #if __STDC_HOSTED__
     40 #include <mm_malloc.h>
     41 #endif
     42 
     43 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
     44 _mm_add_ss(__m128 __a, __m128 __b)
     45 {
     46   __a[0] += __b[0];
     47   return __a;
     48 }
     49 
     50 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
     51 _mm_add_ps(__m128 __a, __m128 __b)
     52 {
     53   return __a + __b;
     54 }
     55 
     56 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
     57 _mm_sub_ss(__m128 __a, __m128 __b)
     58 {
     59   __a[0] -= __b[0];
     60   return __a;
     61 }
     62 
     63 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
     64 _mm_sub_ps(__m128 __a, __m128 __b)
     65 {
     66   return __a - __b;
     67 }
     68 
     69 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
     70 _mm_mul_ss(__m128 __a, __m128 __b)
     71 {
     72   __a[0] *= __b[0];
     73   return __a;
     74 }
     75 
     76 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
     77 _mm_mul_ps(__m128 __a, __m128 __b)
     78 {
     79   return __a * __b;
     80 }
     81 
     82 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
     83 _mm_div_ss(__m128 __a, __m128 __b)
     84 {
     85   __a[0] /= __b[0];
     86   return __a;
     87 }
     88 
     89 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
     90 _mm_div_ps(__m128 __a, __m128 __b)
     91 {
     92   return __a / __b;
     93 }
     94 
     95 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
     96 _mm_sqrt_ss(__m128 __a)
     97 {
     98   __m128 __c = __builtin_ia32_sqrtss(__a);
     99   return (__m128) { __c[0], __a[1], __a[2], __a[3] };
    100 }
    101 
    102 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    103 _mm_sqrt_ps(__m128 __a)
    104 {
    105   return __builtin_ia32_sqrtps(__a);
    106 }
    107 
    108 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    109 _mm_rcp_ss(__m128 __a)
    110 {
    111   __m128 __c = __builtin_ia32_rcpss(__a);
    112   return (__m128) { __c[0], __a[1], __a[2], __a[3] };
    113 }
    114 
    115 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    116 _mm_rcp_ps(__m128 __a)
    117 {
    118   return __builtin_ia32_rcpps(__a);
    119 }
    120 
    121 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    122 _mm_rsqrt_ss(__m128 __a)
    123 {
    124   __m128 __c = __builtin_ia32_rsqrtss(__a);
    125   return (__m128) { __c[0], __a[1], __a[2], __a[3] };
    126 }
    127 
    128 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    129 _mm_rsqrt_ps(__m128 __a)
    130 {
    131   return __builtin_ia32_rsqrtps(__a);
    132 }
    133 
    134 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    135 _mm_min_ss(__m128 __a, __m128 __b)
    136 {
    137   return __builtin_ia32_minss(__a, __b);
    138 }
    139 
    140 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    141 _mm_min_ps(__m128 __a, __m128 __b)
    142 {
    143   return __builtin_ia32_minps(__a, __b);
    144 }
    145 
    146 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    147 _mm_max_ss(__m128 __a, __m128 __b)
    148 {
    149   return __builtin_ia32_maxss(__a, __b);
    150 }
    151 
    152 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    153 _mm_max_ps(__m128 __a, __m128 __b)
    154 {
    155   return __builtin_ia32_maxps(__a, __b);
    156 }
    157 
    158 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    159 _mm_and_ps(__m128 __a, __m128 __b)
    160 {
    161   return (__m128)((__v4si)__a & (__v4si)__b);
    162 }
    163 
    164 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    165 _mm_andnot_ps(__m128 __a, __m128 __b)
    166 {
    167   return (__m128)(~(__v4si)__a & (__v4si)__b);
    168 }
    169 
    170 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    171 _mm_or_ps(__m128 __a, __m128 __b)
    172 {
    173   return (__m128)((__v4si)__a | (__v4si)__b);
    174 }
    175 
    176 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    177 _mm_xor_ps(__m128 __a, __m128 __b)
    178 {
    179   return (__m128)((__v4si)__a ^ (__v4si)__b);
    180 }
    181 
    182 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    183 _mm_cmpeq_ss(__m128 __a, __m128 __b)
    184 {
    185   return (__m128)__builtin_ia32_cmpss(__a, __b, 0);
    186 }
    187 
    188 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    189 _mm_cmpeq_ps(__m128 __a, __m128 __b)
    190 {
    191   return (__m128)__builtin_ia32_cmpps(__a, __b, 0);
    192 }
    193 
    194 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    195 _mm_cmplt_ss(__m128 __a, __m128 __b)
    196 {
    197   return (__m128)__builtin_ia32_cmpss(__a, __b, 1);
    198 }
    199 
    200 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    201 _mm_cmplt_ps(__m128 __a, __m128 __b)
    202 {
    203   return (__m128)__builtin_ia32_cmpps(__a, __b, 1);
    204 }
    205 
    206 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    207 _mm_cmple_ss(__m128 __a, __m128 __b)
    208 {
    209   return (__m128)__builtin_ia32_cmpss(__a, __b, 2);
    210 }
    211 
    212 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    213 _mm_cmple_ps(__m128 __a, __m128 __b)
    214 {
    215   return (__m128)__builtin_ia32_cmpps(__a, __b, 2);
    216 }
    217 
    218 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    219 _mm_cmpgt_ss(__m128 __a, __m128 __b)
    220 {
    221   return (__m128)__builtin_shufflevector(__a,
    222                                          __builtin_ia32_cmpss(__b, __a, 1),
    223                                          4, 1, 2, 3);
    224 }
    225 
    226 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    227 _mm_cmpgt_ps(__m128 __a, __m128 __b)
    228 {
    229   return (__m128)__builtin_ia32_cmpps(__b, __a, 1);
    230 }
    231 
    232 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    233 _mm_cmpge_ss(__m128 __a, __m128 __b)
    234 {
    235   return (__m128)__builtin_shufflevector(__a,
    236                                          __builtin_ia32_cmpss(__b, __a, 2),
    237                                          4, 1, 2, 3);
    238 }
    239 
    240 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    241 _mm_cmpge_ps(__m128 __a, __m128 __b)
    242 {
    243   return (__m128)__builtin_ia32_cmpps(__b, __a, 2);
    244 }
    245 
    246 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    247 _mm_cmpneq_ss(__m128 __a, __m128 __b)
    248 {
    249   return (__m128)__builtin_ia32_cmpss(__a, __b, 4);
    250 }
    251 
    252 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    253 _mm_cmpneq_ps(__m128 __a, __m128 __b)
    254 {
    255   return (__m128)__builtin_ia32_cmpps(__a, __b, 4);
    256 }
    257 
    258 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    259 _mm_cmpnlt_ss(__m128 __a, __m128 __b)
    260 {
    261   return (__m128)__builtin_ia32_cmpss(__a, __b, 5);
    262 }
    263 
    264 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    265 _mm_cmpnlt_ps(__m128 __a, __m128 __b)
    266 {
    267   return (__m128)__builtin_ia32_cmpps(__a, __b, 5);
    268 }
    269 
    270 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    271 _mm_cmpnle_ss(__m128 __a, __m128 __b)
    272 {
    273   return (__m128)__builtin_ia32_cmpss(__a, __b, 6);
    274 }
    275 
    276 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    277 _mm_cmpnle_ps(__m128 __a, __m128 __b)
    278 {
    279   return (__m128)__builtin_ia32_cmpps(__a, __b, 6);
    280 }
    281 
    282 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    283 _mm_cmpngt_ss(__m128 __a, __m128 __b)
    284 {
    285   return (__m128)__builtin_shufflevector(__a,
    286                                          __builtin_ia32_cmpss(__b, __a, 5),
    287                                          4, 1, 2, 3);
    288 }
    289 
    290 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    291 _mm_cmpngt_ps(__m128 __a, __m128 __b)
    292 {
    293   return (__m128)__builtin_ia32_cmpps(__b, __a, 5);
    294 }
    295 
    296 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    297 _mm_cmpnge_ss(__m128 __a, __m128 __b)
    298 {
    299   return (__m128)__builtin_shufflevector(__a,
    300                                          __builtin_ia32_cmpss(__b, __a, 6),
    301                                          4, 1, 2, 3);
    302 }
    303 
    304 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    305 _mm_cmpnge_ps(__m128 __a, __m128 __b)
    306 {
    307   return (__m128)__builtin_ia32_cmpps(__b, __a, 6);
    308 }
    309 
    310 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    311 _mm_cmpord_ss(__m128 __a, __m128 __b)
    312 {
    313   return (__m128)__builtin_ia32_cmpss(__a, __b, 7);
    314 }
    315 
    316 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    317 _mm_cmpord_ps(__m128 __a, __m128 __b)
    318 {
    319   return (__m128)__builtin_ia32_cmpps(__a, __b, 7);
    320 }
    321 
    322 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    323 _mm_cmpunord_ss(__m128 __a, __m128 __b)
    324 {
    325   return (__m128)__builtin_ia32_cmpss(__a, __b, 3);
    326 }
    327 
    328 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    329 _mm_cmpunord_ps(__m128 __a, __m128 __b)
    330 {
    331   return (__m128)__builtin_ia32_cmpps(__a, __b, 3);
    332 }
    333 
    334 static __inline__ int __attribute__((__always_inline__, __nodebug__))
    335 _mm_comieq_ss(__m128 __a, __m128 __b)
    336 {
    337   return __builtin_ia32_comieq(__a, __b);
    338 }
    339 
    340 static __inline__ int __attribute__((__always_inline__, __nodebug__))
    341 _mm_comilt_ss(__m128 __a, __m128 __b)
    342 {
    343   return __builtin_ia32_comilt(__a, __b);
    344 }
    345 
    346 static __inline__ int __attribute__((__always_inline__, __nodebug__))
    347 _mm_comile_ss(__m128 __a, __m128 __b)
    348 {
    349   return __builtin_ia32_comile(__a, __b);
    350 }
    351 
    352 static __inline__ int __attribute__((__always_inline__, __nodebug__))
    353 _mm_comigt_ss(__m128 __a, __m128 __b)
    354 {
    355   return __builtin_ia32_comigt(__a, __b);
    356 }
    357 
    358 static __inline__ int __attribute__((__always_inline__, __nodebug__))
    359 _mm_comige_ss(__m128 __a, __m128 __b)
    360 {
    361   return __builtin_ia32_comige(__a, __b);
    362 }
    363 
    364 static __inline__ int __attribute__((__always_inline__, __nodebug__))
    365 _mm_comineq_ss(__m128 __a, __m128 __b)
    366 {
    367   return __builtin_ia32_comineq(__a, __b);
    368 }
    369 
    370 static __inline__ int __attribute__((__always_inline__, __nodebug__))
    371 _mm_ucomieq_ss(__m128 __a, __m128 __b)
    372 {
    373   return __builtin_ia32_ucomieq(__a, __b);
    374 }
    375 
    376 static __inline__ int __attribute__((__always_inline__, __nodebug__))
    377 _mm_ucomilt_ss(__m128 __a, __m128 __b)
    378 {
    379   return __builtin_ia32_ucomilt(__a, __b);
    380 }
    381 
    382 static __inline__ int __attribute__((__always_inline__, __nodebug__))
    383 _mm_ucomile_ss(__m128 __a, __m128 __b)
    384 {
    385   return __builtin_ia32_ucomile(__a, __b);
    386 }
    387 
    388 static __inline__ int __attribute__((__always_inline__, __nodebug__))
    389 _mm_ucomigt_ss(__m128 __a, __m128 __b)
    390 {
    391   return __builtin_ia32_ucomigt(__a, __b);
    392 }
    393 
    394 static __inline__ int __attribute__((__always_inline__, __nodebug__))
    395 _mm_ucomige_ss(__m128 __a, __m128 __b)
    396 {
    397   return __builtin_ia32_ucomige(__a, __b);
    398 }
    399 
    400 static __inline__ int __attribute__((__always_inline__, __nodebug__))
    401 _mm_ucomineq_ss(__m128 __a, __m128 __b)
    402 {
    403   return __builtin_ia32_ucomineq(__a, __b);
    404 }
    405 
    406 static __inline__ int __attribute__((__always_inline__, __nodebug__))
    407 _mm_cvtss_si32(__m128 __a)
    408 {
    409   return __builtin_ia32_cvtss2si(__a);
    410 }
    411 
    412 static __inline__ int __attribute__((__always_inline__, __nodebug__))
    413 _mm_cvt_ss2si(__m128 __a)
    414 {
    415   return _mm_cvtss_si32(__a);
    416 }
    417 
    418 #ifdef __x86_64__
    419 
    420 static __inline__ long long __attribute__((__always_inline__, __nodebug__))
    421 _mm_cvtss_si64(__m128 __a)
    422 {
    423   return __builtin_ia32_cvtss2si64(__a);
    424 }
    425 
    426 #endif
    427 
    428 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
    429 _mm_cvtps_pi32(__m128 __a)
    430 {
    431   return (__m64)__builtin_ia32_cvtps2pi(__a);
    432 }
    433 
    434 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
    435 _mm_cvt_ps2pi(__m128 __a)
    436 {
    437   return _mm_cvtps_pi32(__a);
    438 }
    439 
    440 static __inline__ int __attribute__((__always_inline__, __nodebug__))
    441 _mm_cvttss_si32(__m128 __a)
    442 {
    443   return __a[0];
    444 }
    445 
    446 static __inline__ int __attribute__((__always_inline__, __nodebug__))
    447 _mm_cvtt_ss2si(__m128 __a)
    448 {
    449   return _mm_cvttss_si32(__a);
    450 }
    451 
    452 static __inline__ long long __attribute__((__always_inline__, __nodebug__))
    453 _mm_cvttss_si64(__m128 __a)
    454 {
    455   return __a[0];
    456 }
    457 
    458 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
    459 _mm_cvttps_pi32(__m128 __a)
    460 {
    461   return (__m64)__builtin_ia32_cvttps2pi(__a);
    462 }
    463 
    464 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
    465 _mm_cvtt_ps2pi(__m128 __a)
    466 {
    467   return _mm_cvttps_pi32(__a);
    468 }
    469 
    470 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    471 _mm_cvtsi32_ss(__m128 __a, int __b)
    472 {
    473   __a[0] = __b;
    474   return __a;
    475 }
    476 
    477 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    478 _mm_cvt_si2ss(__m128 __a, int __b)
    479 {
    480   return _mm_cvtsi32_ss(__a, __b);
    481 }
    482 
    483 #ifdef __x86_64__
    484 
    485 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    486 _mm_cvtsi64_ss(__m128 __a, long long __b)
    487 {
    488   __a[0] = __b;
    489   return __a;
    490 }
    491 
    492 #endif
    493 
    494 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    495 _mm_cvtpi32_ps(__m128 __a, __m64 __b)
    496 {
    497   return __builtin_ia32_cvtpi2ps(__a, (__v2si)__b);
    498 }
    499 
    500 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    501 _mm_cvt_pi2ps(__m128 __a, __m64 __b)
    502 {
    503   return _mm_cvtpi32_ps(__a, __b);
    504 }
    505 
    506 static __inline__ float __attribute__((__always_inline__, __nodebug__))
    507 _mm_cvtss_f32(__m128 __a)
    508 {
    509   return __a[0];
    510 }
    511 
    512 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    513 _mm_loadh_pi(__m128 __a, const __m64 *__p)
    514 {
    515   typedef float __mm_loadh_pi_v2f32 __attribute__((__vector_size__(8)));
    516   struct __mm_loadh_pi_struct {
    517     __mm_loadh_pi_v2f32 __u;
    518   } __attribute__((__packed__, __may_alias__));
    519   __mm_loadh_pi_v2f32 __b = ((struct __mm_loadh_pi_struct*)__p)->__u;
    520   __m128 __bb = __builtin_shufflevector(__b, __b, 0, 1, 0, 1);
    521   return __builtin_shufflevector(__a, __bb, 0, 1, 4, 5);
    522 }
    523 
    524 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    525 _mm_loadl_pi(__m128 __a, const __m64 *__p)
    526 {
    527   typedef float __mm_loadl_pi_v2f32 __attribute__((__vector_size__(8)));
    528   struct __mm_loadl_pi_struct {
    529     __mm_loadl_pi_v2f32 __u;
    530   } __attribute__((__packed__, __may_alias__));
    531   __mm_loadl_pi_v2f32 __b = ((struct __mm_loadl_pi_struct*)__p)->__u;
    532   __m128 __bb = __builtin_shufflevector(__b, __b, 0, 1, 0, 1);
    533   return __builtin_shufflevector(__a, __bb, 4, 5, 2, 3);
    534 }
    535 
    536 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    537 _mm_load_ss(const float *__p)
    538 {
    539   struct __mm_load_ss_struct {
    540     float __u;
    541   } __attribute__((__packed__, __may_alias__));
    542   float __u = ((struct __mm_load_ss_struct*)__p)->__u;
    543   return (__m128){ __u, 0, 0, 0 };
    544 }
    545 
    546 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    547 _mm_load1_ps(const float *__p)
    548 {
    549   struct __mm_load1_ps_struct {
    550     float __u;
    551   } __attribute__((__packed__, __may_alias__));
    552   float __u = ((struct __mm_load1_ps_struct*)__p)->__u;
    553   return (__m128){ __u, __u, __u, __u };
    554 }
    555 
    556 #define        _mm_load_ps1(p) _mm_load1_ps(p)
    557 
    558 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    559 _mm_load_ps(const float *__p)
    560 {
    561   return *(__m128*)__p;
    562 }
    563 
    564 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    565 _mm_loadu_ps(const float *__p)
    566 {
    567   struct __loadu_ps {
    568     __m128 __v;
    569   } __attribute__((__packed__, __may_alias__));
    570   return ((struct __loadu_ps*)__p)->__v;
    571 }
    572 
    573 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    574 _mm_loadr_ps(const float *__p)
    575 {
    576   __m128 __a = _mm_load_ps(__p);
    577   return __builtin_shufflevector(__a, __a, 3, 2, 1, 0);
    578 }
    579 
    580 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    581 _mm_set_ss(float __w)
    582 {
    583   return (__m128){ __w, 0, 0, 0 };
    584 }
    585 
    586 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    587 _mm_set1_ps(float __w)
    588 {
    589   return (__m128){ __w, __w, __w, __w };
    590 }
    591 
    592 /* Microsoft specific. */
    593 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    594 _mm_set_ps1(float __w)
    595 {
    596     return _mm_set1_ps(__w);
    597 }
    598 
    599 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    600 _mm_set_ps(float __z, float __y, float __x, float __w)
    601 {
    602   return (__m128){ __w, __x, __y, __z };
    603 }
    604 
    605 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    606 _mm_setr_ps(float __z, float __y, float __x, float __w)
    607 {
    608   return (__m128){ __z, __y, __x, __w };
    609 }
    610 
    611 static __inline__ __m128 __attribute__((__always_inline__))
    612 _mm_setzero_ps(void)
    613 {
    614   return (__m128){ 0, 0, 0, 0 };
    615 }
    616 
    617 static __inline__ void __attribute__((__always_inline__))
    618 _mm_storeh_pi(__m64 *__p, __m128 __a)
    619 {
    620   __builtin_ia32_storehps((__v2si *)__p, __a);
    621 }
    622 
    623 static __inline__ void __attribute__((__always_inline__))
    624 _mm_storel_pi(__m64 *__p, __m128 __a)
    625 {
    626   __builtin_ia32_storelps((__v2si *)__p, __a);
    627 }
    628 
    629 static __inline__ void __attribute__((__always_inline__))
    630 _mm_store_ss(float *__p, __m128 __a)
    631 {
    632   struct __mm_store_ss_struct {
    633     float __u;
    634   } __attribute__((__packed__, __may_alias__));
    635   ((struct __mm_store_ss_struct*)__p)->__u = __a[0];
    636 }
    637 
    638 static __inline__ void __attribute__((__always_inline__, __nodebug__))
    639 _mm_storeu_ps(float *__p, __m128 __a)
    640 {
    641   __builtin_ia32_storeups(__p, __a);
    642 }
    643 
    644 static __inline__ void __attribute__((__always_inline__, __nodebug__))
    645 _mm_store1_ps(float *__p, __m128 __a)
    646 {
    647   __a = __builtin_shufflevector(__a, __a, 0, 0, 0, 0);
    648   _mm_storeu_ps(__p, __a);
    649 }
    650 
    651 static __inline__ void __attribute__((__always_inline__, __nodebug__))
    652 _mm_store_ps1(float *__p, __m128 __a)
    653 {
    654     return _mm_store1_ps(__p, __a);
    655 }
    656 
    657 static __inline__ void __attribute__((__always_inline__, __nodebug__))
    658 _mm_store_ps(float *__p, __m128 __a)
    659 {
    660   *(__m128 *)__p = __a;
    661 }
    662 
    663 static __inline__ void __attribute__((__always_inline__, __nodebug__))
    664 _mm_storer_ps(float *__p, __m128 __a)
    665 {
    666   __a = __builtin_shufflevector(__a, __a, 3, 2, 1, 0);
    667   _mm_store_ps(__p, __a);
    668 }
    669 
    670 #define _MM_HINT_T0 3
    671 #define _MM_HINT_T1 2
    672 #define _MM_HINT_T2 1
    673 #define _MM_HINT_NTA 0
    674 
    675 #ifndef _MSC_VER
    676 /* FIXME: We have to #define this because "sel" must be a constant integer, and
    677    Sema doesn't do any form of constant propagation yet. */
    678 
    679 #define _mm_prefetch(a, sel) (__builtin_prefetch((void *)(a), 0, (sel)))
    680 #endif
    681 
    682 static __inline__ void __attribute__((__always_inline__, __nodebug__))
    683 _mm_stream_pi(__m64 *__p, __m64 __a)
    684 {
    685   __builtin_ia32_movntq(__p, __a);
    686 }
    687 
    688 static __inline__ void __attribute__((__always_inline__, __nodebug__))
    689 _mm_stream_ps(float *__p, __m128 __a)
    690 {
    691   __builtin_ia32_movntps(__p, __a);
    692 }
    693 
    694 static __inline__ void __attribute__((__always_inline__, __nodebug__))
    695 _mm_sfence(void)
    696 {
    697   __builtin_ia32_sfence();
    698 }
    699 
    700 static __inline__ int __attribute__((__always_inline__, __nodebug__))
    701 _mm_extract_pi16(__m64 __a, int __n)
    702 {
    703   __v4hi __b = (__v4hi)__a;
    704   return (unsigned short)__b[__n & 3];
    705 }
    706 
    707 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
    708 _mm_insert_pi16(__m64 __a, int __d, int __n)
    709 {
    710    __v4hi __b = (__v4hi)__a;
    711    __b[__n & 3] = __d;
    712    return (__m64)__b;
    713 }
    714 
    715 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
    716 _mm_max_pi16(__m64 __a, __m64 __b)
    717 {
    718   return (__m64)__builtin_ia32_pmaxsw((__v4hi)__a, (__v4hi)__b);
    719 }
    720 
    721 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
    722 _mm_max_pu8(__m64 __a, __m64 __b)
    723 {
    724   return (__m64)__builtin_ia32_pmaxub((__v8qi)__a, (__v8qi)__b);
    725 }
    726 
    727 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
    728 _mm_min_pi16(__m64 __a, __m64 __b)
    729 {
    730   return (__m64)__builtin_ia32_pminsw((__v4hi)__a, (__v4hi)__b);
    731 }
    732 
    733 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
    734 _mm_min_pu8(__m64 __a, __m64 __b)
    735 {
    736   return (__m64)__builtin_ia32_pminub((__v8qi)__a, (__v8qi)__b);
    737 }
    738 
    739 static __inline__ int __attribute__((__always_inline__, __nodebug__))
    740 _mm_movemask_pi8(__m64 __a)
    741 {
    742   return __builtin_ia32_pmovmskb((__v8qi)__a);
    743 }
    744 
    745 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
    746 _mm_mulhi_pu16(__m64 __a, __m64 __b)
    747 {
    748   return (__m64)__builtin_ia32_pmulhuw((__v4hi)__a, (__v4hi)__b);
    749 }
    750 
    751 #define _mm_shuffle_pi16(a, n) __extension__ ({ \
    752   __m64 __a = (a); \
    753   (__m64)__builtin_ia32_pshufw((__v4hi)__a, (n)); })
    754 
    755 static __inline__ void __attribute__((__always_inline__, __nodebug__))
    756 _mm_maskmove_si64(__m64 __d, __m64 __n, char *__p)
    757 {
    758   __builtin_ia32_maskmovq((__v8qi)__d, (__v8qi)__n, __p);
    759 }
    760 
    761 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
    762 _mm_avg_pu8(__m64 __a, __m64 __b)
    763 {
    764   return (__m64)__builtin_ia32_pavgb((__v8qi)__a, (__v8qi)__b);
    765 }
    766 
    767 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
    768 _mm_avg_pu16(__m64 __a, __m64 __b)
    769 {
    770   return (__m64)__builtin_ia32_pavgw((__v4hi)__a, (__v4hi)__b);
    771 }
    772 
    773 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
    774 _mm_sad_pu8(__m64 __a, __m64 __b)
    775 {
    776   return (__m64)__builtin_ia32_psadbw((__v8qi)__a, (__v8qi)__b);
    777 }
    778 
    779 static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__))
    780 _mm_getcsr(void)
    781 {
    782   return __builtin_ia32_stmxcsr();
    783 }
    784 
    785 static __inline__ void __attribute__((__always_inline__, __nodebug__))
    786 _mm_setcsr(unsigned int __i)
    787 {
    788   __builtin_ia32_ldmxcsr(__i);
    789 }
    790 
    791 #define _mm_shuffle_ps(a, b, mask) __extension__ ({ \
    792   __m128 __a = (a); \
    793   __m128 __b = (b); \
    794   (__m128)__builtin_shufflevector((__v4sf)__a, (__v4sf)__b, \
    795                                   (mask) & 0x3, ((mask) & 0xc) >> 2, \
    796                                   (((mask) & 0x30) >> 4) + 4, \
    797                                   (((mask) & 0xc0) >> 6) + 4); })
    798 
    799 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    800 _mm_unpackhi_ps(__m128 __a, __m128 __b)
    801 {
    802   return __builtin_shufflevector(__a, __b, 2, 6, 3, 7);
    803 }
    804 
    805 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    806 _mm_unpacklo_ps(__m128 __a, __m128 __b)
    807 {
    808   return __builtin_shufflevector(__a, __b, 0, 4, 1, 5);
    809 }
    810 
    811 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    812 _mm_move_ss(__m128 __a, __m128 __b)
    813 {
    814   return __builtin_shufflevector(__a, __b, 4, 1, 2, 3);
    815 }
    816 
    817 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    818 _mm_movehl_ps(__m128 __a, __m128 __b)
    819 {
    820   return __builtin_shufflevector(__a, __b, 6, 7, 2, 3);
    821 }
    822 
    823 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    824 _mm_movelh_ps(__m128 __a, __m128 __b)
    825 {
    826   return __builtin_shufflevector(__a, __b, 0, 1, 4, 5);
    827 }
    828 
    829 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    830 _mm_cvtpi16_ps(__m64 __a)
    831 {
    832   __m64 __b, __c;
    833   __m128 __r;
    834 
    835   __b = _mm_setzero_si64();
    836   __b = _mm_cmpgt_pi16(__b, __a);
    837   __c = _mm_unpackhi_pi16(__a, __b);
    838   __r = _mm_setzero_ps();
    839   __r = _mm_cvtpi32_ps(__r, __c);
    840   __r = _mm_movelh_ps(__r, __r);
    841   __c = _mm_unpacklo_pi16(__a, __b);
    842   __r = _mm_cvtpi32_ps(__r, __c);
    843 
    844   return __r;
    845 }
    846 
    847 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    848 _mm_cvtpu16_ps(__m64 __a)
    849 {
    850   __m64 __b, __c;
    851   __m128 __r;
    852 
    853   __b = _mm_setzero_si64();
    854   __c = _mm_unpackhi_pi16(__a, __b);
    855   __r = _mm_setzero_ps();
    856   __r = _mm_cvtpi32_ps(__r, __c);
    857   __r = _mm_movelh_ps(__r, __r);
    858   __c = _mm_unpacklo_pi16(__a, __b);
    859   __r = _mm_cvtpi32_ps(__r, __c);
    860 
    861   return __r;
    862 }
    863 
    864 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    865 _mm_cvtpi8_ps(__m64 __a)
    866 {
    867   __m64 __b;
    868 
    869   __b = _mm_setzero_si64();
    870   __b = _mm_cmpgt_pi8(__b, __a);
    871   __b = _mm_unpacklo_pi8(__a, __b);
    872 
    873   return _mm_cvtpi16_ps(__b);
    874 }
    875 
    876 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    877 _mm_cvtpu8_ps(__m64 __a)
    878 {
    879   __m64 __b;
    880 
    881   __b = _mm_setzero_si64();
    882   __b = _mm_unpacklo_pi8(__a, __b);
    883 
    884   return _mm_cvtpi16_ps(__b);
    885 }
    886 
    887 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
    888 _mm_cvtpi32x2_ps(__m64 __a, __m64 __b)
    889 {
    890   __m128 __c;
    891 
    892   __c = _mm_setzero_ps();
    893   __c = _mm_cvtpi32_ps(__c, __b);
    894   __c = _mm_movelh_ps(__c, __c);
    895 
    896   return _mm_cvtpi32_ps(__c, __a);
    897 }
    898 
    899 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
    900 _mm_cvtps_pi16(__m128 __a)
    901 {
    902   __m64 __b, __c;
    903 
    904   __b = _mm_cvtps_pi32(__a);
    905   __a = _mm_movehl_ps(__a, __a);
    906   __c = _mm_cvtps_pi32(__a);
    907 
    908   return _mm_packs_pi32(__b, __c);
    909 }
    910 
    911 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
    912 _mm_cvtps_pi8(__m128 __a)
    913 {
    914   __m64 __b, __c;
    915 
    916   __b = _mm_cvtps_pi16(__a);
    917   __c = _mm_setzero_si64();
    918 
    919   return _mm_packs_pi16(__b, __c);
    920 }
    921 
    922 static __inline__ int __attribute__((__always_inline__, __nodebug__))
    923 _mm_movemask_ps(__m128 __a)
    924 {
    925   return __builtin_ia32_movmskps(__a);
    926 }
    927 
    928 #define _MM_SHUFFLE(z, y, x, w) (((z) << 6) | ((y) << 4) | ((x) << 2) | (w))
    929 
    930 #define _MM_EXCEPT_INVALID    (0x0001)
    931 #define _MM_EXCEPT_DENORM     (0x0002)
    932 #define _MM_EXCEPT_DIV_ZERO   (0x0004)
    933 #define _MM_EXCEPT_OVERFLOW   (0x0008)
    934 #define _MM_EXCEPT_UNDERFLOW  (0x0010)
    935 #define _MM_EXCEPT_INEXACT    (0x0020)
    936 #define _MM_EXCEPT_MASK       (0x003f)
    937 
    938 #define _MM_MASK_INVALID      (0x0080)
    939 #define _MM_MASK_DENORM       (0x0100)
    940 #define _MM_MASK_DIV_ZERO     (0x0200)
    941 #define _MM_MASK_OVERFLOW     (0x0400)
    942 #define _MM_MASK_UNDERFLOW    (0x0800)
    943 #define _MM_MASK_INEXACT      (0x1000)
    944 #define _MM_MASK_MASK         (0x1f80)
    945 
    946 #define _MM_ROUND_NEAREST     (0x0000)
    947 #define _MM_ROUND_DOWN        (0x2000)
    948 #define _MM_ROUND_UP          (0x4000)
    949 #define _MM_ROUND_TOWARD_ZERO (0x6000)
    950 #define _MM_ROUND_MASK        (0x6000)
    951 
    952 #define _MM_FLUSH_ZERO_MASK   (0x8000)
    953 #define _MM_FLUSH_ZERO_ON     (0x8000)
    954 #define _MM_FLUSH_ZERO_OFF    (0x0000)
    955 
    956 #define _MM_GET_EXCEPTION_MASK() (_mm_getcsr() & _MM_MASK_MASK)
    957 #define _MM_GET_EXCEPTION_STATE() (_mm_getcsr() & _MM_EXCEPT_MASK)
    958 #define _MM_GET_FLUSH_ZERO_MODE() (_mm_getcsr() & _MM_FLUSH_ZERO_MASK)
    959 #define _MM_GET_ROUNDING_MODE() (_mm_getcsr() & _MM_ROUND_MASK)
    960 
    961 #define _MM_SET_EXCEPTION_MASK(x) (_mm_setcsr((_mm_getcsr() & ~_MM_MASK_MASK) | (x)))
    962 #define _MM_SET_EXCEPTION_STATE(x) (_mm_setcsr((_mm_getcsr() & ~_MM_EXCEPT_MASK) | (x)))
    963 #define _MM_SET_FLUSH_ZERO_MODE(x) (_mm_setcsr((_mm_getcsr() & ~_MM_FLUSH_ZERO_MASK) | (x)))
    964 #define _MM_SET_ROUNDING_MODE(x) (_mm_setcsr((_mm_getcsr() & ~_MM_ROUND_MASK) | (x)))
    965 
    966 #define _MM_TRANSPOSE4_PS(row0, row1, row2, row3) \
    967 do { \
    968   __m128 tmp3, tmp2, tmp1, tmp0; \
    969   tmp0 = _mm_unpacklo_ps((row0), (row1)); \
    970   tmp2 = _mm_unpacklo_ps((row2), (row3)); \
    971   tmp1 = _mm_unpackhi_ps((row0), (row1)); \
    972   tmp3 = _mm_unpackhi_ps((row2), (row3)); \
    973   (row0) = _mm_movelh_ps(tmp0, tmp2); \
    974   (row1) = _mm_movehl_ps(tmp2, tmp0); \
    975   (row2) = _mm_movelh_ps(tmp1, tmp3); \
    976   (row3) = _mm_movehl_ps(tmp3, tmp1); \
    977 } while (0)
    978 
    979 /* Aliases for compatibility. */
    980 #define _m_pextrw _mm_extract_pi16
    981 #define _m_pinsrw _mm_insert_pi16
    982 #define _m_pmaxsw _mm_max_pi16
    983 #define _m_pmaxub _mm_max_pu8
    984 #define _m_pminsw _mm_min_pi16
    985 #define _m_pminub _mm_min_pu8
    986 #define _m_pmovmskb _mm_movemask_pi8
    987 #define _m_pmulhuw _mm_mulhi_pu16
    988 #define _m_pshufw _mm_shuffle_pi16
    989 #define _m_maskmovq _mm_maskmove_si64
    990 #define _m_pavgb _mm_avg_pu8
    991 #define _m_pavgw _mm_avg_pu16
    992 #define _m_psadbw _mm_sad_pu8
    993 #define _m_ _mm_
    994 #define _m_ _mm_
    995 
    996 /* Ugly hack for backwards-compatibility (compatible with gcc) */
    997 #ifdef __SSE2__
    998 #include <emmintrin.h>
    999 #endif
   1000 
   1001 #endif /* __SSE__ */
   1002 
   1003 #endif /* __XMMINTRIN_H */
   1004