0
|
1 /* Copyright (C) 2007, 2008, 2009 Free Software Foundation, Inc.
|
|
2
|
|
3 This file is part of GCC.
|
|
4
|
|
5 GCC is free software; you can redistribute it and/or modify
|
|
6 it under the terms of the GNU General Public License as published by
|
|
7 the Free Software Foundation; either version 3, or (at your option)
|
|
8 any later version.
|
|
9
|
|
10 GCC is distributed in the hope that it will be useful,
|
|
11 but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
|
13 GNU General Public License for more details.
|
|
14
|
|
15 Under Section 7 of GPL version 3, you are granted additional
|
|
16 permissions described in the GCC Runtime Library Exception, version
|
|
17 3.1, as published by the Free Software Foundation.
|
|
18
|
|
19 You should have received a copy of the GNU General Public License and
|
|
20 a copy of the GCC Runtime Library Exception along with this program;
|
|
21 see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
|
|
22 <http://www.gnu.org/licenses/>. */
|
|
23
|
|
24 #ifndef _BMMINTRIN_H_INCLUDED
|
|
25 #define _BMMINTRIN_H_INCLUDED
|
|
26
|
|
27 #ifndef __SSE5__
|
|
28 # error "SSE5 instruction set not enabled"
|
|
29 #else
|
|
30
|
|
31 /* We need definitions from the SSE4A, SSE3, SSE2 and SSE header files. */
|
|
32 #include <ammintrin.h>
|
|
33 #include <mmintrin-common.h>
|
|
34
|
|
35 /* Floating point multiply/add type instructions */
|
|
36 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
37 _mm_macc_ps(__m128 __A, __m128 __B, __m128 __C)
|
|
38 {
|
|
39 return (__m128) __builtin_ia32_fmaddps ((__v4sf)__A, (__v4sf)__B, (__v4sf)__C);
|
|
40 }
|
|
41
|
|
42 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
43 _mm_macc_pd(__m128d __A, __m128d __B, __m128d __C)
|
|
44 {
|
|
45 return (__m128d) __builtin_ia32_fmaddpd ((__v2df)__A, (__v2df)__B, (__v2df)__C);
|
|
46 }
|
|
47
|
|
48 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
49 _mm_macc_ss(__m128 __A, __m128 __B, __m128 __C)
|
|
50 {
|
|
51 return (__m128) __builtin_ia32_fmaddss ((__v4sf)__A, (__v4sf)__B, (__v4sf)__C);
|
|
52 }
|
|
53
|
|
54 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
55 _mm_macc_sd(__m128d __A, __m128d __B, __m128d __C)
|
|
56 {
|
|
57 return (__m128d) __builtin_ia32_fmaddsd ((__v2df)__A, (__v2df)__B, (__v2df)__C);
|
|
58 }
|
|
59
|
|
60 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
61 _mm_msub_ps(__m128 __A, __m128 __B, __m128 __C)
|
|
62 {
|
|
63 return (__m128) __builtin_ia32_fmsubps ((__v4sf)__A, (__v4sf)__B, (__v4sf)__C);
|
|
64 }
|
|
65
|
|
66 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
67 _mm_msub_pd(__m128d __A, __m128d __B, __m128d __C)
|
|
68 {
|
|
69 return (__m128d) __builtin_ia32_fmsubpd ((__v2df)__A, (__v2df)__B, (__v2df)__C);
|
|
70 }
|
|
71
|
|
72 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
73 _mm_msub_ss(__m128 __A, __m128 __B, __m128 __C)
|
|
74 {
|
|
75 return (__m128) __builtin_ia32_fmsubss ((__v4sf)__A, (__v4sf)__B, (__v4sf)__C);
|
|
76 }
|
|
77
|
|
78 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
79 _mm_msub_sd(__m128d __A, __m128d __B, __m128d __C)
|
|
80 {
|
|
81 return (__m128d) __builtin_ia32_fmsubsd ((__v2df)__A, (__v2df)__B, (__v2df)__C);
|
|
82 }
|
|
83
|
|
84 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
85 _mm_nmacc_ps(__m128 __A, __m128 __B, __m128 __C)
|
|
86 {
|
|
87 return (__m128) __builtin_ia32_fnmaddps ((__v4sf)__A, (__v4sf)__B, (__v4sf)__C);
|
|
88 }
|
|
89
|
|
90 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
91 _mm_nmacc_pd(__m128d __A, __m128d __B, __m128d __C)
|
|
92 {
|
|
93 return (__m128d) __builtin_ia32_fnmaddpd ((__v2df)__A, (__v2df)__B, (__v2df)__C);
|
|
94 }
|
|
95
|
|
96 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
97 _mm_nmacc_ss(__m128 __A, __m128 __B, __m128 __C)
|
|
98 {
|
|
99 return (__m128) __builtin_ia32_fnmaddss ((__v4sf)__A, (__v4sf)__B, (__v4sf)__C);
|
|
100 }
|
|
101
|
|
102 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
103 _mm_nmacc_sd(__m128d __A, __m128d __B, __m128d __C)
|
|
104 {
|
|
105 return (__m128d) __builtin_ia32_fnmaddsd ((__v2df)__A, (__v2df)__B, (__v2df)__C);
|
|
106 }
|
|
107
|
|
108 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
109 _mm_nmsub_ps(__m128 __A, __m128 __B, __m128 __C)
|
|
110 {
|
|
111 return (__m128) __builtin_ia32_fnmsubps ((__v4sf)__A, (__v4sf)__B, (__v4sf)__C);
|
|
112 }
|
|
113
|
|
114 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
115 _mm_nmsub_pd(__m128d __A, __m128d __B, __m128d __C)
|
|
116 {
|
|
117 return (__m128d) __builtin_ia32_fnmsubpd ((__v2df)__A, (__v2df)__B, (__v2df)__C);
|
|
118 }
|
|
119
|
|
120 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
121 _mm_nmsub_ss(__m128 __A, __m128 __B, __m128 __C)
|
|
122 {
|
|
123 return (__m128) __builtin_ia32_fnmsubss ((__v4sf)__A, (__v4sf)__B, (__v4sf)__C);
|
|
124 }
|
|
125
|
|
126 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
127 _mm_nmsub_sd(__m128d __A, __m128d __B, __m128d __C)
|
|
128 {
|
|
129 return (__m128d) __builtin_ia32_fnmsubsd ((__v2df)__A, (__v2df)__B, (__v2df)__C);
|
|
130 }
|
|
131
|
|
132 /* Integer multiply/add intructions. */
|
|
133 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
134 _mm_maccs_epi16(__m128i __A, __m128i __B, __m128i __C)
|
|
135 {
|
|
136 return (__m128i) __builtin_ia32_pmacssww ((__v8hi)__A,(__v8hi)__B, (__v8hi)__C);
|
|
137 }
|
|
138
|
|
139 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
140 _mm_macc_epi16(__m128i __A, __m128i __B, __m128i __C)
|
|
141 {
|
|
142 return (__m128i) __builtin_ia32_pmacsww ((__v8hi)__A, (__v8hi)__B, (__v8hi)__C);
|
|
143 }
|
|
144
|
|
145 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
146 _mm_maccsd_epi16(__m128i __A, __m128i __B, __m128i __C)
|
|
147 {
|
|
148 return (__m128i) __builtin_ia32_pmacsswd ((__v8hi)__A, (__v8hi)__B, (__v4si)__C);
|
|
149 }
|
|
150
|
|
151 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
152 _mm_maccd_epi16(__m128i __A, __m128i __B, __m128i __C)
|
|
153 {
|
|
154 return (__m128i) __builtin_ia32_pmacswd ((__v8hi)__A, (__v8hi)__B, (__v4si)__C);
|
|
155 }
|
|
156
|
|
157 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
158 _mm_maccs_epi32(__m128i __A, __m128i __B, __m128i __C)
|
|
159 {
|
|
160 return (__m128i) __builtin_ia32_pmacssdd ((__v4si)__A, (__v4si)__B, (__v4si)__C);
|
|
161 }
|
|
162
|
|
163 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
164 _mm_macc_epi32(__m128i __A, __m128i __B, __m128i __C)
|
|
165 {
|
|
166 return (__m128i) __builtin_ia32_pmacsdd ((__v4si)__A, (__v4si)__B, (__v4si)__C);
|
|
167 }
|
|
168
|
|
169 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
170 _mm_maccslo_epi32(__m128i __A, __m128i __B, __m128i __C)
|
|
171 {
|
|
172 return (__m128i) __builtin_ia32_pmacssdql ((__v4si)__A, (__v4si)__B, (__v2di)__C);
|
|
173 }
|
|
174
|
|
175 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
176 _mm_macclo_epi32(__m128i __A, __m128i __B, __m128i __C)
|
|
177 {
|
|
178 return (__m128i) __builtin_ia32_pmacsdql ((__v4si)__A, (__v4si)__B, (__v2di)__C);
|
|
179 }
|
|
180
|
|
181 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
182 _mm_maccshi_epi32(__m128i __A, __m128i __B, __m128i __C)
|
|
183 {
|
|
184 return (__m128i) __builtin_ia32_pmacssdqh ((__v4si)__A, (__v4si)__B, (__v2di)__C);
|
|
185 }
|
|
186
|
|
187 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
188 _mm_macchi_epi32(__m128i __A, __m128i __B, __m128i __C)
|
|
189 {
|
|
190 return (__m128i) __builtin_ia32_pmacsdqh ((__v4si)__A, (__v4si)__B, (__v2di)__C);
|
|
191 }
|
|
192
|
|
193 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
194 _mm_maddsd_epi16(__m128i __A, __m128i __B, __m128i __C)
|
|
195 {
|
|
196 return (__m128i) __builtin_ia32_pmadcsswd ((__v8hi)__A,(__v8hi)__B,(__v4si)__C);
|
|
197 }
|
|
198
|
|
199 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
200 _mm_maddd_epi16(__m128i __A, __m128i __B, __m128i __C)
|
|
201 {
|
|
202 return (__m128i) __builtin_ia32_pmadcswd ((__v8hi)__A,(__v8hi)__B,(__v4si)__C);
|
|
203 }
|
|
204
|
|
205 /* Packed Integer Horizontal Add and Subtract */
|
|
206 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
207 _mm_haddw_epi8(__m128i __A)
|
|
208 {
|
|
209 return (__m128i) __builtin_ia32_phaddbw ((__v16qi)__A);
|
|
210 }
|
|
211
|
|
212 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
213 _mm_haddd_epi8(__m128i __A)
|
|
214 {
|
|
215 return (__m128i) __builtin_ia32_phaddbd ((__v16qi)__A);
|
|
216 }
|
|
217
|
|
218 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
219 _mm_haddq_epi8(__m128i __A)
|
|
220 {
|
|
221 return (__m128i) __builtin_ia32_phaddbq ((__v16qi)__A);
|
|
222 }
|
|
223
|
|
224 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
225 _mm_haddd_epi16(__m128i __A)
|
|
226 {
|
|
227 return (__m128i) __builtin_ia32_phaddwd ((__v8hi)__A);
|
|
228 }
|
|
229
|
|
230 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
231 _mm_haddq_epi16(__m128i __A)
|
|
232 {
|
|
233 return (__m128i) __builtin_ia32_phaddwq ((__v8hi)__A);
|
|
234 }
|
|
235
|
|
236 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
237 _mm_haddq_epi32(__m128i __A)
|
|
238 {
|
|
239 return (__m128i) __builtin_ia32_phadddq ((__v4si)__A);
|
|
240 }
|
|
241
|
|
242 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
243 _mm_haddw_epu8(__m128i __A)
|
|
244 {
|
|
245 return (__m128i) __builtin_ia32_phaddubw ((__v16qi)__A);
|
|
246 }
|
|
247
|
|
248 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
249 _mm_haddd_epu8(__m128i __A)
|
|
250 {
|
|
251 return (__m128i) __builtin_ia32_phaddubd ((__v16qi)__A);
|
|
252 }
|
|
253
|
|
254 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
255 _mm_haddq_epu8(__m128i __A)
|
|
256 {
|
|
257 return (__m128i) __builtin_ia32_phaddubq ((__v16qi)__A);
|
|
258 }
|
|
259
|
|
260 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
261 _mm_haddd_epu16(__m128i __A)
|
|
262 {
|
|
263 return (__m128i) __builtin_ia32_phadduwd ((__v8hi)__A);
|
|
264 }
|
|
265
|
|
266 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
267 _mm_haddq_epu16(__m128i __A)
|
|
268 {
|
|
269 return (__m128i) __builtin_ia32_phadduwq ((__v8hi)__A);
|
|
270 }
|
|
271
|
|
272 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
273 _mm_haddq_epu32(__m128i __A)
|
|
274 {
|
|
275 return (__m128i) __builtin_ia32_phaddudq ((__v4si)__A);
|
|
276 }
|
|
277
|
|
278 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
279 _mm_hsubw_epi8(__m128i __A)
|
|
280 {
|
|
281 return (__m128i) __builtin_ia32_phsubbw ((__v16qi)__A);
|
|
282 }
|
|
283
|
|
284 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
285 _mm_hsubd_epi16(__m128i __A)
|
|
286 {
|
|
287 return (__m128i) __builtin_ia32_phsubwd ((__v8hi)__A);
|
|
288 }
|
|
289
|
|
290 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
291 _mm_hsubq_epi32(__m128i __A)
|
|
292 {
|
|
293 return (__m128i) __builtin_ia32_phsubdq ((__v4si)__A);
|
|
294 }
|
|
295
|
|
296 /* Vector conditional move and permute */
|
|
297 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
298 _mm_cmov_si128(__m128i __A, __m128i __B, __m128i __C)
|
|
299 {
|
|
300 return (__m128i) __builtin_ia32_pcmov (__A, __B, __C);
|
|
301 }
|
|
302
|
|
303 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
304 _mm_perm_epi8(__m128i __A, __m128i __B, __m128i __C)
|
|
305 {
|
|
306 return (__m128i) __builtin_ia32_pperm ((__v16qi)__A, (__v16qi)__B, (__v16qi)__C);
|
|
307 }
|
|
308
|
|
309 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
310 _mm_perm_ps(__m128 __A, __m128 __B, __m128i __C)
|
|
311 {
|
|
312 return (__m128) __builtin_ia32_permps ((__m128)__A, (__m128)__B, (__v16qi)__C);
|
|
313 }
|
|
314
|
|
315 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
316 _mm_perm_pd(__m128d __A, __m128d __B, __m128i __C)
|
|
317 {
|
|
318 return (__m128d) __builtin_ia32_permpd ((__m128d)__A, (__m128d)__B, (__v16qi)__C);
|
|
319 }
|
|
320
|
|
321 /* Packed Integer Rotates and Shifts */
|
|
322
|
|
323 /* Rotates - Non-Immediate form */
|
|
324 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
325 _mm_rot_epi8(__m128i __A, __m128i __B)
|
|
326 {
|
|
327 return (__m128i) __builtin_ia32_protb ((__v16qi)__A, (__v16qi)__B);
|
|
328 }
|
|
329
|
|
330 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
331 _mm_rot_epi16(__m128i __A, __m128i __B)
|
|
332 {
|
|
333 return (__m128i) __builtin_ia32_protw ((__v8hi)__A, (__v8hi)__B);
|
|
334 }
|
|
335
|
|
336 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
337 _mm_rot_epi32(__m128i __A, __m128i __B)
|
|
338 {
|
|
339 return (__m128i) __builtin_ia32_protd ((__v4si)__A, (__v4si)__B);
|
|
340 }
|
|
341
|
|
342 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
343 _mm_rot_epi64(__m128i __A, __m128i __B)
|
|
344 {
|
|
345 return (__m128i) __builtin_ia32_protq ((__v2di)__A, (__v2di)__B);
|
|
346 }
|
|
347
|
|
348
|
|
349 /* Rotates - Immediate form */
|
|
350 #ifdef __OPTIMIZE__
|
|
351 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
352 _mm_roti_epi8(__m128i __A, const int __B)
|
|
353 {
|
|
354 return (__m128i) __builtin_ia32_protbi ((__v16qi)__A, __B);
|
|
355 }
|
|
356
|
|
357 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
358 _mm_roti_epi16(__m128i __A, const int __B)
|
|
359 {
|
|
360 return (__m128i) __builtin_ia32_protwi ((__v8hi)__A, __B);
|
|
361 }
|
|
362
|
|
363 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
364 _mm_roti_epi32(__m128i __A, const int __B)
|
|
365 {
|
|
366 return (__m128i) __builtin_ia32_protdi ((__v4si)__A, __B);
|
|
367 }
|
|
368
|
|
369 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
370 _mm_roti_epi64(__m128i __A, const int __B)
|
|
371 {
|
|
372 return (__m128i) __builtin_ia32_protqi ((__v2di)__A, __B);
|
|
373 }
|
|
374 #else
|
|
375 #define _mm_roti_epi8(A, N) \
|
|
376 ((__m128i) __builtin_ia32_protbi ((__v16qi)(__m128i)(A), (int)(N)))
|
|
377 #define _mm_roti_epi16(A, N) \
|
|
378 ((__m128i) __builtin_ia32_protwi ((__v8hi)(__m128i)(A), (int)(N)))
|
|
379 #define _mm_roti_epi32(A, N) \
|
|
380 ((__m128i) __builtin_ia32_protdi ((__v4si)(__m128i)(A), (int)(N)))
|
|
381 #define _mm_roti_epi64(A, N) \
|
|
382 ((__m128i) __builtin_ia32_protqi ((__v2di)(__m128i)(A), (int)(N)))
|
|
383 #endif
|
|
384
|
|
385 /* pshl */
|
|
386
|
|
387 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
388 _mm_shl_epi8(__m128i __A, __m128i __B)
|
|
389 {
|
|
390 return (__m128i) __builtin_ia32_pshlb ((__v16qi)__A, (__v16qi)__B);
|
|
391 }
|
|
392
|
|
393 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
394 _mm_shl_epi16(__m128i __A, __m128i __B)
|
|
395 {
|
|
396 return (__m128i) __builtin_ia32_pshlw ((__v8hi)__A, (__v8hi)__B);
|
|
397 }
|
|
398
|
|
399 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
400 _mm_shl_epi32(__m128i __A, __m128i __B)
|
|
401 {
|
|
402 return (__m128i) __builtin_ia32_pshld ((__v4si)__A, (__v4si)__B);
|
|
403 }
|
|
404
|
|
405 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
406 _mm_shl_epi64(__m128i __A, __m128i __B)
|
|
407 {
|
|
408 return (__m128i) __builtin_ia32_pshlq ((__v2di)__A, (__v2di)__B);
|
|
409 }
|
|
410
|
|
411 /* psha */
|
|
412 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
413 _mm_sha_epi8(__m128i __A, __m128i __B)
|
|
414 {
|
|
415 return (__m128i) __builtin_ia32_pshab ((__v16qi)__A, (__v16qi)__B);
|
|
416 }
|
|
417
|
|
418 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
419 _mm_sha_epi16(__m128i __A, __m128i __B)
|
|
420 {
|
|
421 return (__m128i) __builtin_ia32_pshaw ((__v8hi)__A, (__v8hi)__B);
|
|
422 }
|
|
423
|
|
424 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
425 _mm_sha_epi32(__m128i __A, __m128i __B)
|
|
426 {
|
|
427 return (__m128i) __builtin_ia32_pshad ((__v4si)__A, (__v4si)__B);
|
|
428 }
|
|
429
|
|
430 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
431 _mm_sha_epi64(__m128i __A, __m128i __B)
|
|
432 {
|
|
433 return (__m128i) __builtin_ia32_pshaq ((__v2di)__A, (__v2di)__B);
|
|
434 }
|
|
435
|
|
436 /* Compare and Predicate Generation */
|
|
437
|
|
438 /* com (floating point, packed single) */
|
|
439 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
440 _mm_comeq_ps(__m128 __A, __m128 __B)
|
|
441 {
|
|
442 return (__m128) __builtin_ia32_comeqps ((__v4sf)__A, (__v4sf)__B);
|
|
443 }
|
|
444
|
|
445 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
446 _mm_comlt_ps(__m128 __A, __m128 __B)
|
|
447 {
|
|
448 return (__m128) __builtin_ia32_comltps ((__v4sf)__A, (__v4sf)__B);
|
|
449 }
|
|
450
|
|
451 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
452 _mm_comle_ps(__m128 __A, __m128 __B)
|
|
453 {
|
|
454 return (__m128) __builtin_ia32_comleps ((__v4sf)__A, (__v4sf)__B);
|
|
455 }
|
|
456
|
|
457 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
458 _mm_comunord_ps(__m128 __A, __m128 __B)
|
|
459 {
|
|
460 return (__m128) __builtin_ia32_comunordps ((__v4sf)__A, (__v4sf)__B);
|
|
461 }
|
|
462
|
|
463 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
464 _mm_comneq_ps(__m128 __A, __m128 __B)
|
|
465 {
|
|
466 return (__m128) __builtin_ia32_comuneqps ((__v4sf)__A, (__v4sf)__B);
|
|
467 }
|
|
468
|
|
469 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
470 _mm_comnlt_ps(__m128 __A, __m128 __B)
|
|
471 {
|
|
472 return (__m128) __builtin_ia32_comunltps ((__v4sf)__A, (__v4sf)__B);
|
|
473 }
|
|
474
|
|
475 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
476 _mm_comnle_ps(__m128 __A, __m128 __B)
|
|
477 {
|
|
478 return (__m128) __builtin_ia32_comunleps ((__v4sf)__A, (__v4sf)__B);
|
|
479 }
|
|
480
|
|
481
|
|
482 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
483 _mm_comord_ps(__m128 __A, __m128 __B)
|
|
484 {
|
|
485 return (__m128) __builtin_ia32_comordps ((__v4sf)__A, (__v4sf)__B);
|
|
486 }
|
|
487
|
|
488
|
|
489 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
490 _mm_comueq_ps(__m128 __A, __m128 __B)
|
|
491 {
|
|
492 return (__m128) __builtin_ia32_comueqps ((__v4sf)__A, (__v4sf)__B);
|
|
493 }
|
|
494
|
|
495 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
496 _mm_comnge_ps(__m128 __A, __m128 __B)
|
|
497 {
|
|
498 return (__m128) __builtin_ia32_comungeps ((__v4sf)__A, (__v4sf)__B);
|
|
499 }
|
|
500
|
|
501 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
502 _mm_comngt_ps(__m128 __A, __m128 __B)
|
|
503 {
|
|
504 return (__m128) __builtin_ia32_comungtps ((__v4sf)__A, (__v4sf)__B);
|
|
505 }
|
|
506
|
|
507 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
508 _mm_comfalse_ps(__m128 __A, __m128 __B)
|
|
509 {
|
|
510 return (__m128) __builtin_ia32_comfalseps ((__v4sf)__A, (__v4sf)__B);
|
|
511 }
|
|
512
|
|
513 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
514 _mm_comoneq_ps(__m128 __A, __m128 __B)
|
|
515 {
|
|
516 return (__m128) __builtin_ia32_comneqps ((__v4sf)__A, (__v4sf)__B);
|
|
517 }
|
|
518
|
|
519 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
520 _mm_comge_ps(__m128 __A, __m128 __B)
|
|
521 {
|
|
522 return (__m128) __builtin_ia32_comgeps ((__v4sf)__A, (__v4sf)__B);
|
|
523 }
|
|
524
|
|
525 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
526 _mm_comgt_ps(__m128 __A, __m128 __B)
|
|
527 {
|
|
528 return (__m128) __builtin_ia32_comgtps ((__v4sf)__A, (__v4sf)__B);
|
|
529 }
|
|
530
|
|
531 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
532 _mm_comtrue_ps(__m128 __A, __m128 __B)
|
|
533 {
|
|
534 return (__m128) __builtin_ia32_comtrueps ((__v4sf)__A, (__v4sf)__B);
|
|
535 }
|
|
536
|
|
537 /* com (floating point, packed double) */
|
|
538
|
|
539 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
540 _mm_comeq_pd(__m128d __A, __m128d __B)
|
|
541 {
|
|
542 return (__m128d) __builtin_ia32_comeqpd ((__v2df)__A, (__v2df)__B);
|
|
543 }
|
|
544
|
|
545 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
546 _mm_comlt_pd(__m128d __A, __m128d __B)
|
|
547 {
|
|
548 return (__m128d) __builtin_ia32_comltpd ((__v2df)__A, (__v2df)__B);
|
|
549 }
|
|
550
|
|
551 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
552 _mm_comle_pd(__m128d __A, __m128d __B)
|
|
553 {
|
|
554 return (__m128d) __builtin_ia32_comlepd ((__v2df)__A, (__v2df)__B);
|
|
555 }
|
|
556
|
|
557 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
558 _mm_comunord_pd(__m128d __A, __m128d __B)
|
|
559 {
|
|
560 return (__m128d) __builtin_ia32_comunordpd ((__v2df)__A, (__v2df)__B);
|
|
561 }
|
|
562
|
|
563 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
564 _mm_comneq_pd(__m128d __A, __m128d __B)
|
|
565 {
|
|
566 return (__m128d) __builtin_ia32_comuneqpd ((__v2df)__A, (__v2df)__B);
|
|
567 }
|
|
568
|
|
569 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
570 _mm_comnlt_pd(__m128d __A, __m128d __B)
|
|
571 {
|
|
572 return (__m128d) __builtin_ia32_comunltpd ((__v2df)__A, (__v2df)__B);
|
|
573 }
|
|
574
|
|
575 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
576 _mm_comnle_pd(__m128d __A, __m128d __B)
|
|
577 {
|
|
578 return (__m128d) __builtin_ia32_comunlepd ((__v2df)__A, (__v2df)__B);
|
|
579 }
|
|
580
|
|
581
|
|
582 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
583 _mm_comord_pd(__m128d __A, __m128d __B)
|
|
584 {
|
|
585 return (__m128d) __builtin_ia32_comordpd ((__v2df)__A, (__v2df)__B);
|
|
586 }
|
|
587
|
|
588 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
589 _mm_comueq_pd(__m128d __A, __m128d __B)
|
|
590 {
|
|
591 return (__m128d) __builtin_ia32_comueqpd ((__v2df)__A, (__v2df)__B);
|
|
592 }
|
|
593
|
|
594 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
595 _mm_comnge_pd(__m128d __A, __m128d __B)
|
|
596 {
|
|
597 return (__m128d) __builtin_ia32_comungepd ((__v2df)__A, (__v2df)__B);
|
|
598 }
|
|
599
|
|
600 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
601 _mm_comngt_pd(__m128d __A, __m128d __B)
|
|
602 {
|
|
603 return (__m128d) __builtin_ia32_comungtpd ((__v2df)__A, (__v2df)__B);
|
|
604 }
|
|
605
|
|
606 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
607 _mm_comfalse_pd(__m128d __A, __m128d __B)
|
|
608 {
|
|
609 return (__m128d) __builtin_ia32_comfalsepd ((__v2df)__A, (__v2df)__B);
|
|
610 }
|
|
611
|
|
612 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
613 _mm_comoneq_pd(__m128d __A, __m128d __B)
|
|
614 {
|
|
615 return (__m128d) __builtin_ia32_comneqpd ((__v2df)__A, (__v2df)__B);
|
|
616 }
|
|
617
|
|
618 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
619 _mm_comge_pd(__m128d __A, __m128d __B)
|
|
620 {
|
|
621 return (__m128d) __builtin_ia32_comgepd ((__v2df)__A, (__v2df)__B);
|
|
622 }
|
|
623
|
|
624 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
625 _mm_comgt_pd(__m128d __A, __m128d __B)
|
|
626 {
|
|
627 return (__m128d) __builtin_ia32_comgtpd ((__v2df)__A, (__v2df)__B);
|
|
628 }
|
|
629
|
|
630 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
631 _mm_comtrue_pd(__m128d __A, __m128d __B)
|
|
632 {
|
|
633 return (__m128d) __builtin_ia32_comtruepd ((__v2df)__A, (__v2df)__B);
|
|
634 }
|
|
635
|
|
636 /* com (floating point, scalar single) */
|
|
637 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
638 _mm_comeq_ss(__m128 __A, __m128 __B)
|
|
639 {
|
|
640 return (__m128) __builtin_ia32_comeqss ((__v4sf)__A, (__v4sf)__B);
|
|
641 }
|
|
642
|
|
643 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
644 _mm_comlt_ss(__m128 __A, __m128 __B)
|
|
645 {
|
|
646 return (__m128) __builtin_ia32_comltss ((__v4sf)__A, (__v4sf)__B);
|
|
647 }
|
|
648
|
|
649 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
650 _mm_comle_ss(__m128 __A, __m128 __B)
|
|
651 {
|
|
652 return (__m128) __builtin_ia32_comless ((__v4sf)__A, (__v4sf)__B);
|
|
653 }
|
|
654
|
|
655 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
656 _mm_comunord_ss(__m128 __A, __m128 __B)
|
|
657 {
|
|
658 return (__m128) __builtin_ia32_comunordss ((__v4sf)__A, (__v4sf)__B);
|
|
659 }
|
|
660
|
|
661 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
662 _mm_comneq_ss(__m128 __A, __m128 __B)
|
|
663 {
|
|
664 return (__m128) __builtin_ia32_comuneqss ((__v4sf)__A, (__v4sf)__B);
|
|
665 }
|
|
666
|
|
667 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
668 _mm_comnlt_ss(__m128 __A, __m128 __B)
|
|
669 {
|
|
670 return (__m128) __builtin_ia32_comunltss ((__v4sf)__A, (__v4sf)__B);
|
|
671 }
|
|
672
|
|
673 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
674 _mm_comnle_ss(__m128 __A, __m128 __B)
|
|
675 {
|
|
676 return (__m128) __builtin_ia32_comunless ((__v4sf)__A, (__v4sf)__B);
|
|
677 }
|
|
678
|
|
679
|
|
680 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
681 _mm_comord_ss(__m128 __A, __m128 __B)
|
|
682 {
|
|
683 return (__m128) __builtin_ia32_comordss ((__v4sf)__A, (__v4sf)__B);
|
|
684 }
|
|
685
|
|
686 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
687 _mm_comueq_ss(__m128 __A, __m128 __B)
|
|
688 {
|
|
689 return (__m128) __builtin_ia32_comueqss ((__v4sf)__A, (__v4sf)__B);
|
|
690 }
|
|
691
|
|
692 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
693 _mm_comnge_ss(__m128 __A, __m128 __B)
|
|
694 {
|
|
695 return (__m128) __builtin_ia32_comungess ((__v4sf)__A, (__v4sf)__B);
|
|
696 }
|
|
697
|
|
698 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
699 _mm_comngt_ss(__m128 __A, __m128 __B)
|
|
700 {
|
|
701 return (__m128) __builtin_ia32_comungtss ((__v4sf)__A, (__v4sf)__B);
|
|
702 }
|
|
703
|
|
704 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
705 _mm_comfalse_ss(__m128 __A, __m128 __B)
|
|
706 {
|
|
707 return (__m128) __builtin_ia32_comfalsess ((__v4sf)__A, (__v4sf)__B);
|
|
708 }
|
|
709
|
|
710 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
711 _mm_comoneq_ss(__m128 __A, __m128 __B)
|
|
712 {
|
|
713 return (__m128) __builtin_ia32_comneqss ((__v4sf)__A, (__v4sf)__B);
|
|
714 }
|
|
715
|
|
716 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
717 _mm_comge_ss(__m128 __A, __m128 __B)
|
|
718 {
|
|
719 return (__m128) __builtin_ia32_comgess ((__v4sf)__A, (__v4sf)__B);
|
|
720 }
|
|
721
|
|
722 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
723 _mm_comgt_ss(__m128 __A, __m128 __B)
|
|
724 {
|
|
725 return (__m128) __builtin_ia32_comgtss ((__v4sf)__A, (__v4sf)__B);
|
|
726 }
|
|
727
|
|
728 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
729 _mm_comtrue_ss(__m128 __A, __m128 __B)
|
|
730 {
|
|
731 return (__m128) __builtin_ia32_comtruess ((__v4sf)__A, (__v4sf)__B);
|
|
732 }
|
|
733
|
|
734 /* com (floating point, scalar double) */
|
|
735
|
|
736 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
737 _mm_comeq_sd(__m128d __A, __m128d __B)
|
|
738 {
|
|
739 return (__m128d) __builtin_ia32_comeqsd ((__v2df)__A, (__v2df)__B);
|
|
740 }
|
|
741
|
|
742 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
743 _mm_comlt_sd(__m128d __A, __m128d __B)
|
|
744 {
|
|
745 return (__m128d) __builtin_ia32_comltsd ((__v2df)__A, (__v2df)__B);
|
|
746 }
|
|
747
|
|
748 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
749 _mm_comle_sd(__m128d __A, __m128d __B)
|
|
750 {
|
|
751 return (__m128d) __builtin_ia32_comlesd ((__v2df)__A, (__v2df)__B);
|
|
752 }
|
|
753
|
|
754 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
755 _mm_comunord_sd(__m128d __A, __m128d __B)
|
|
756 {
|
|
757 return (__m128d) __builtin_ia32_comunordsd ((__v2df)__A, (__v2df)__B);
|
|
758 }
|
|
759
|
|
760 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
761 _mm_comneq_sd(__m128d __A, __m128d __B)
|
|
762 {
|
|
763 return (__m128d) __builtin_ia32_comuneqsd ((__v2df)__A, (__v2df)__B);
|
|
764 }
|
|
765
|
|
766 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
767 _mm_comnlt_sd(__m128d __A, __m128d __B)
|
|
768 {
|
|
769 return (__m128d) __builtin_ia32_comunltsd ((__v2df)__A, (__v2df)__B);
|
|
770 }
|
|
771
|
|
772 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
773 _mm_comnle_sd(__m128d __A, __m128d __B)
|
|
774 {
|
|
775 return (__m128d) __builtin_ia32_comunlesd ((__v2df)__A, (__v2df)__B);
|
|
776 }
|
|
777
|
|
778
|
|
779 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
780 _mm_comord_sd(__m128d __A, __m128d __B)
|
|
781 {
|
|
782 return (__m128d) __builtin_ia32_comordsd ((__v2df)__A, (__v2df)__B);
|
|
783 }
|
|
784
|
|
785 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
786 _mm_comueq_sd(__m128d __A, __m128d __B)
|
|
787 {
|
|
788 return (__m128d) __builtin_ia32_comueqsd ((__v2df)__A, (__v2df)__B);
|
|
789 }
|
|
790
|
|
791 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
792 _mm_comnge_sd(__m128d __A, __m128d __B)
|
|
793 {
|
|
794 return (__m128d) __builtin_ia32_comungesd ((__v2df)__A, (__v2df)__B);
|
|
795 }
|
|
796
|
|
797 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
798 _mm_comngt_sd(__m128d __A, __m128d __B)
|
|
799 {
|
|
800 return (__m128d) __builtin_ia32_comungtsd ((__v2df)__A, (__v2df)__B);
|
|
801 }
|
|
802
|
|
803 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
804 _mm_comfalse_sd(__m128d __A, __m128d __B)
|
|
805 {
|
|
806 return (__m128d) __builtin_ia32_comfalsesd ((__v2df)__A, (__v2df)__B);
|
|
807 }
|
|
808
|
|
809 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
810 _mm_comoneq_sd(__m128d __A, __m128d __B)
|
|
811 {
|
|
812 return (__m128d) __builtin_ia32_comneqsd ((__v2df)__A, (__v2df)__B);
|
|
813 }
|
|
814
|
|
815 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
816 _mm_comge_sd(__m128d __A, __m128d __B)
|
|
817 {
|
|
818 return (__m128d) __builtin_ia32_comgesd ((__v2df)__A, (__v2df)__B);
|
|
819 }
|
|
820
|
|
821 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
822 _mm_comgt_sd(__m128d __A, __m128d __B)
|
|
823 {
|
|
824 return (__m128d) __builtin_ia32_comgtsd ((__v2df)__A, (__v2df)__B);
|
|
825 }
|
|
826
|
|
827 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
828 _mm_comtrue_sd(__m128d __A, __m128d __B)
|
|
829 {
|
|
830 return (__m128d) __builtin_ia32_comtruesd ((__v2df)__A, (__v2df)__B);
|
|
831 }
|
|
832
|
|
833
|
|
834 /*pcom (integer, unsinged bytes) */
|
|
835
|
|
836 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
837 _mm_comlt_epu8(__m128i __A, __m128i __B)
|
|
838 {
|
|
839 return (__m128i) __builtin_ia32_pcomltub ((__v16qi)__A, (__v16qi)__B);
|
|
840 }
|
|
841
|
|
842 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
843 _mm_comle_epu8(__m128i __A, __m128i __B)
|
|
844 {
|
|
845 return (__m128i) __builtin_ia32_pcomleub ((__v16qi)__A, (__v16qi)__B);
|
|
846 }
|
|
847
|
|
848 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
849 _mm_comgt_epu8(__m128i __A, __m128i __B)
|
|
850 {
|
|
851 return (__m128i) __builtin_ia32_pcomgtub ((__v16qi)__A, (__v16qi)__B);
|
|
852 }
|
|
853
|
|
854 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
855 _mm_comge_epu8(__m128i __A, __m128i __B)
|
|
856 {
|
|
857 return (__m128i) __builtin_ia32_pcomgeub ((__v16qi)__A, (__v16qi)__B);
|
|
858 }
|
|
859
|
|
860 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
861 _mm_comeq_epu8(__m128i __A, __m128i __B)
|
|
862 {
|
|
863 return (__m128i) __builtin_ia32_pcomequb ((__v16qi)__A, (__v16qi)__B);
|
|
864 }
|
|
865
|
|
866 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
867 _mm_comneq_epu8(__m128i __A, __m128i __B)
|
|
868 {
|
|
869 return (__m128i) __builtin_ia32_pcomnequb ((__v16qi)__A, (__v16qi)__B);
|
|
870 }
|
|
871
|
|
872 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
873 _mm_comfalse_epu8(__m128i __A, __m128i __B)
|
|
874 {
|
|
875 return (__m128i) __builtin_ia32_pcomfalseub ((__v16qi)__A, (__v16qi)__B);
|
|
876 }
|
|
877
|
|
878 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
879 _mm_comtrue_epu8(__m128i __A, __m128i __B)
|
|
880 {
|
|
881 return (__m128i) __builtin_ia32_pcomtrueub ((__v16qi)__A, (__v16qi)__B);
|
|
882 }
|
|
883
|
|
884 /*pcom (integer, unsinged words) */
|
|
885
|
|
886 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
887 _mm_comlt_epu16(__m128i __A, __m128i __B)
|
|
888 {
|
|
889 return (__m128i) __builtin_ia32_pcomltuw ((__v8hi)__A, (__v8hi)__B);
|
|
890 }
|
|
891
|
|
892 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
893 _mm_comle_epu16(__m128i __A, __m128i __B)
|
|
894 {
|
|
895 return (__m128i) __builtin_ia32_pcomleuw ((__v8hi)__A, (__v8hi)__B);
|
|
896 }
|
|
897
|
|
898 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
899 _mm_comgt_epu16(__m128i __A, __m128i __B)
|
|
900 {
|
|
901 return (__m128i) __builtin_ia32_pcomgtuw ((__v8hi)__A, (__v8hi)__B);
|
|
902 }
|
|
903
|
|
904 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
905 _mm_comge_epu16(__m128i __A, __m128i __B)
|
|
906 {
|
|
907 return (__m128i) __builtin_ia32_pcomgeuw ((__v8hi)__A, (__v8hi)__B);
|
|
908 }
|
|
909
|
|
910 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
911 _mm_comeq_epu16(__m128i __A, __m128i __B)
|
|
912 {
|
|
913 return (__m128i) __builtin_ia32_pcomequw ((__v8hi)__A, (__v8hi)__B);
|
|
914 }
|
|
915
|
|
916 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
917 _mm_comneq_epu16(__m128i __A, __m128i __B)
|
|
918 {
|
|
919 return (__m128i) __builtin_ia32_pcomnequw ((__v8hi)__A, (__v8hi)__B);
|
|
920 }
|
|
921
|
|
922 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
923 _mm_comfalse_epu16(__m128i __A, __m128i __B)
|
|
924 {
|
|
925 return (__m128i) __builtin_ia32_pcomfalseuw ((__v8hi)__A, (__v8hi)__B);
|
|
926 }
|
|
927
|
|
928 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
929 _mm_comtrue_epu16(__m128i __A, __m128i __B)
|
|
930 {
|
|
931 return (__m128i) __builtin_ia32_pcomtrueuw ((__v8hi)__A, (__v8hi)__B);
|
|
932 }
|
|
933
|
|
934 /*pcom (integer, unsinged double words) */
|
|
935
|
|
936 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
937 _mm_comlt_epu32(__m128i __A, __m128i __B)
|
|
938 {
|
|
939 return (__m128i) __builtin_ia32_pcomltud ((__v4si)__A, (__v4si)__B);
|
|
940 }
|
|
941
|
|
942 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
943 _mm_comle_epu32(__m128i __A, __m128i __B)
|
|
944 {
|
|
945 return (__m128i) __builtin_ia32_pcomleud ((__v4si)__A, (__v4si)__B);
|
|
946 }
|
|
947
|
|
948 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
949 _mm_comgt_epu32(__m128i __A, __m128i __B)
|
|
950 {
|
|
951 return (__m128i) __builtin_ia32_pcomgtud ((__v4si)__A, (__v4si)__B);
|
|
952 }
|
|
953
|
|
954 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
955 _mm_comge_epu32(__m128i __A, __m128i __B)
|
|
956 {
|
|
957 return (__m128i) __builtin_ia32_pcomgeud ((__v4si)__A, (__v4si)__B);
|
|
958 }
|
|
959
|
|
960 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
961 _mm_comeq_epu32(__m128i __A, __m128i __B)
|
|
962 {
|
|
963 return (__m128i) __builtin_ia32_pcomequd ((__v4si)__A, (__v4si)__B);
|
|
964 }
|
|
965
|
|
966 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
967 _mm_comneq_epu32(__m128i __A, __m128i __B)
|
|
968 {
|
|
969 return (__m128i) __builtin_ia32_pcomnequd ((__v4si)__A, (__v4si)__B);
|
|
970 }
|
|
971
|
|
972 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
973 _mm_comfalse_epu32(__m128i __A, __m128i __B)
|
|
974 {
|
|
975 return (__m128i) __builtin_ia32_pcomfalseud ((__v4si)__A, (__v4si)__B);
|
|
976 }
|
|
977
|
|
978 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
979 _mm_comtrue_epu32(__m128i __A, __m128i __B)
|
|
980 {
|
|
981 return (__m128i) __builtin_ia32_pcomtrueud ((__v4si)__A, (__v4si)__B);
|
|
982 }
|
|
983
|
|
984 /*pcom (integer, unsinged quad words) */
|
|
985
|
|
986 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
987 _mm_comlt_epu64(__m128i __A, __m128i __B)
|
|
988 {
|
|
989 return (__m128i) __builtin_ia32_pcomltuq ((__v2di)__A, (__v2di)__B);
|
|
990 }
|
|
991
|
|
992 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
993 _mm_comle_epu64(__m128i __A, __m128i __B)
|
|
994 {
|
|
995 return (__m128i) __builtin_ia32_pcomleuq ((__v2di)__A, (__v2di)__B);
|
|
996 }
|
|
997
|
|
998 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
999 _mm_comgt_epu64(__m128i __A, __m128i __B)
|
|
1000 {
|
|
1001 return (__m128i) __builtin_ia32_pcomgtuq ((__v2di)__A, (__v2di)__B);
|
|
1002 }
|
|
1003
|
|
1004 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1005 _mm_comge_epu64(__m128i __A, __m128i __B)
|
|
1006 {
|
|
1007 return (__m128i) __builtin_ia32_pcomgeuq ((__v2di)__A, (__v2di)__B);
|
|
1008 }
|
|
1009
|
|
1010 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1011 _mm_comeq_epu64(__m128i __A, __m128i __B)
|
|
1012 {
|
|
1013 return (__m128i) __builtin_ia32_pcomequq ((__v2di)__A, (__v2di)__B);
|
|
1014 }
|
|
1015
|
|
1016 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1017 _mm_comneq_epu64(__m128i __A, __m128i __B)
|
|
1018 {
|
|
1019 return (__m128i) __builtin_ia32_pcomnequq ((__v2di)__A, (__v2di)__B);
|
|
1020 }
|
|
1021
|
|
1022 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1023 _mm_comfalse_epu64(__m128i __A, __m128i __B)
|
|
1024 {
|
|
1025 return (__m128i) __builtin_ia32_pcomfalseuq ((__v2di)__A, (__v2di)__B);
|
|
1026 }
|
|
1027
|
|
1028 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1029 _mm_comtrue_epu64(__m128i __A, __m128i __B)
|
|
1030 {
|
|
1031 return (__m128i) __builtin_ia32_pcomtrueuq ((__v2di)__A, (__v2di)__B);
|
|
1032 }
|
|
1033
|
|
1034 /*pcom (integer, signed bytes) */
|
|
1035
|
|
1036 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1037 _mm_comlt_epi8(__m128i __A, __m128i __B)
|
|
1038 {
|
|
1039 return (__m128i) __builtin_ia32_pcomltb ((__v16qi)__A, (__v16qi)__B);
|
|
1040 }
|
|
1041
|
|
1042 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1043 _mm_comle_epi8(__m128i __A, __m128i __B)
|
|
1044 {
|
|
1045 return (__m128i) __builtin_ia32_pcomleb ((__v16qi)__A, (__v16qi)__B);
|
|
1046 }
|
|
1047
|
|
1048 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1049 _mm_comgt_epi8(__m128i __A, __m128i __B)
|
|
1050 {
|
|
1051 return (__m128i) __builtin_ia32_pcomgtb ((__v16qi)__A, (__v16qi)__B);
|
|
1052 }
|
|
1053
|
|
1054 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1055 _mm_comge_epi8(__m128i __A, __m128i __B)
|
|
1056 {
|
|
1057 return (__m128i) __builtin_ia32_pcomgeb ((__v16qi)__A, (__v16qi)__B);
|
|
1058 }
|
|
1059
|
|
1060 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1061 _mm_comeq_epi8(__m128i __A, __m128i __B)
|
|
1062 {
|
|
1063 return (__m128i) __builtin_ia32_pcomeqb ((__v16qi)__A, (__v16qi)__B);
|
|
1064 }
|
|
1065
|
|
1066 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1067 _mm_comneq_epi8(__m128i __A, __m128i __B)
|
|
1068 {
|
|
1069 return (__m128i) __builtin_ia32_pcomneqb ((__v16qi)__A, (__v16qi)__B);
|
|
1070 }
|
|
1071
|
|
1072 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1073 _mm_comfalse_epi8(__m128i __A, __m128i __B)
|
|
1074 {
|
|
1075 return (__m128i) __builtin_ia32_pcomfalseb ((__v16qi)__A, (__v16qi)__B);
|
|
1076 }
|
|
1077
|
|
1078 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1079 _mm_comtrue_epi8(__m128i __A, __m128i __B)
|
|
1080 {
|
|
1081 return (__m128i) __builtin_ia32_pcomtrueb ((__v16qi)__A, (__v16qi)__B);
|
|
1082 }
|
|
1083
|
|
1084 /*pcom (integer, signed words) */
|
|
1085
|
|
1086 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1087 _mm_comlt_epi16(__m128i __A, __m128i __B)
|
|
1088 {
|
|
1089 return (__m128i) __builtin_ia32_pcomltw ((__v8hi)__A, (__v8hi)__B);
|
|
1090 }
|
|
1091
|
|
1092 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1093 _mm_comle_epi16(__m128i __A, __m128i __B)
|
|
1094 {
|
|
1095 return (__m128i) __builtin_ia32_pcomlew ((__v8hi)__A, (__v8hi)__B);
|
|
1096 }
|
|
1097
|
|
1098 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1099 _mm_comgt_epi16(__m128i __A, __m128i __B)
|
|
1100 {
|
|
1101 return (__m128i) __builtin_ia32_pcomgtw ((__v8hi)__A, (__v8hi)__B);
|
|
1102 }
|
|
1103
|
|
1104 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1105 _mm_comge_epi16(__m128i __A, __m128i __B)
|
|
1106 {
|
|
1107 return (__m128i) __builtin_ia32_pcomgew ((__v8hi)__A, (__v8hi)__B);
|
|
1108 }
|
|
1109
|
|
1110 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1111 _mm_comeq_epi16(__m128i __A, __m128i __B)
|
|
1112 {
|
|
1113 return (__m128i) __builtin_ia32_pcomeqw ((__v8hi)__A, (__v8hi)__B);
|
|
1114 }
|
|
1115
|
|
1116 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1117 _mm_comneq_epi16(__m128i __A, __m128i __B)
|
|
1118 {
|
|
1119 return (__m128i) __builtin_ia32_pcomneqw ((__v8hi)__A, (__v8hi)__B);
|
|
1120 }
|
|
1121
|
|
1122 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1123 _mm_comfalse_epi16(__m128i __A, __m128i __B)
|
|
1124 {
|
|
1125 return (__m128i) __builtin_ia32_pcomfalsew ((__v8hi)__A, (__v8hi)__B);
|
|
1126 }
|
|
1127
|
|
1128 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1129 _mm_comtrue_epi16(__m128i __A, __m128i __B)
|
|
1130 {
|
|
1131 return (__m128i) __builtin_ia32_pcomtruew ((__v8hi)__A, (__v8hi)__B);
|
|
1132 }
|
|
1133
|
|
1134 /*pcom (integer, signed double words) */
|
|
1135
|
|
1136 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1137 _mm_comlt_epi32(__m128i __A, __m128i __B)
|
|
1138 {
|
|
1139 return (__m128i) __builtin_ia32_pcomltd ((__v4si)__A, (__v4si)__B);
|
|
1140 }
|
|
1141
|
|
1142 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1143 _mm_comle_epi32(__m128i __A, __m128i __B)
|
|
1144 {
|
|
1145 return (__m128i) __builtin_ia32_pcomled ((__v4si)__A, (__v4si)__B);
|
|
1146 }
|
|
1147
|
|
1148 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1149 _mm_comgt_epi32(__m128i __A, __m128i __B)
|
|
1150 {
|
|
1151 return (__m128i) __builtin_ia32_pcomgtd ((__v4si)__A, (__v4si)__B);
|
|
1152 }
|
|
1153
|
|
1154 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1155 _mm_comge_epi32(__m128i __A, __m128i __B)
|
|
1156 {
|
|
1157 return (__m128i) __builtin_ia32_pcomged ((__v4si)__A, (__v4si)__B);
|
|
1158 }
|
|
1159
|
|
1160 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1161 _mm_comeq_epi32(__m128i __A, __m128i __B)
|
|
1162 {
|
|
1163 return (__m128i) __builtin_ia32_pcomeqd ((__v4si)__A, (__v4si)__B);
|
|
1164 }
|
|
1165
|
|
1166 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1167 _mm_comneq_epi32(__m128i __A, __m128i __B)
|
|
1168 {
|
|
1169 return (__m128i) __builtin_ia32_pcomneqd ((__v4si)__A, (__v4si)__B);
|
|
1170 }
|
|
1171
|
|
1172 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1173 _mm_comfalse_epi32(__m128i __A, __m128i __B)
|
|
1174 {
|
|
1175 return (__m128i) __builtin_ia32_pcomfalsed ((__v4si)__A, (__v4si)__B);
|
|
1176 }
|
|
1177
|
|
1178 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1179 _mm_comtrue_epi32(__m128i __A, __m128i __B)
|
|
1180 {
|
|
1181 return (__m128i) __builtin_ia32_pcomtrued ((__v4si)__A, (__v4si)__B);
|
|
1182 }
|
|
1183
|
|
1184 /*pcom (integer, signed quad words) */
|
|
1185
|
|
1186 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1187 _mm_comlt_epi64(__m128i __A, __m128i __B)
|
|
1188 {
|
|
1189 return (__m128i) __builtin_ia32_pcomltq ((__v2di)__A, (__v2di)__B);
|
|
1190 }
|
|
1191
|
|
1192 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1193 _mm_comle_epi64(__m128i __A, __m128i __B)
|
|
1194 {
|
|
1195 return (__m128i) __builtin_ia32_pcomleq ((__v2di)__A, (__v2di)__B);
|
|
1196 }
|
|
1197
|
|
1198 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1199 _mm_comgt_epi64(__m128i __A, __m128i __B)
|
|
1200 {
|
|
1201 return (__m128i) __builtin_ia32_pcomgtq ((__v2di)__A, (__v2di)__B);
|
|
1202 }
|
|
1203
|
|
1204 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1205 _mm_comge_epi64(__m128i __A, __m128i __B)
|
|
1206 {
|
|
1207 return (__m128i) __builtin_ia32_pcomgeq ((__v2di)__A, (__v2di)__B);
|
|
1208 }
|
|
1209
|
|
1210 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1211 _mm_comeq_epi64(__m128i __A, __m128i __B)
|
|
1212 {
|
|
1213 return (__m128i) __builtin_ia32_pcomeqq ((__v2di)__A, (__v2di)__B);
|
|
1214 }
|
|
1215
|
|
1216 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1217 _mm_comneq_epi64(__m128i __A, __m128i __B)
|
|
1218 {
|
|
1219 return (__m128i) __builtin_ia32_pcomneqq ((__v2di)__A, (__v2di)__B);
|
|
1220 }
|
|
1221
|
|
1222 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1223 _mm_comfalse_epi64(__m128i __A, __m128i __B)
|
|
1224 {
|
|
1225 return (__m128i) __builtin_ia32_pcomfalseq ((__v2di)__A, (__v2di)__B);
|
|
1226 }
|
|
1227
|
|
1228 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1229 _mm_comtrue_epi64(__m128i __A, __m128i __B)
|
|
1230 {
|
|
1231 return (__m128i) __builtin_ia32_pcomtrueq ((__v2di)__A, (__v2di)__B);
|
|
1232 }
|
|
1233
|
|
1234 /* FRCZ */
|
|
1235 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1236 _mm_frcz_ps (__m128 __A)
|
|
1237 {
|
|
1238 return (__m128) __builtin_ia32_frczps ((__v4sf)__A);
|
|
1239 }
|
|
1240
|
|
1241 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1242 _mm_frcz_pd (__m128d __A)
|
|
1243 {
|
|
1244 return (__m128d) __builtin_ia32_frczpd ((__v2df)__A);
|
|
1245 }
|
|
1246
|
|
1247 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1248 _mm_frcz_ss (__m128 __A, __m128 __B)
|
|
1249 {
|
|
1250 return (__m128) __builtin_ia32_frczss ((__v4sf)__A, (__v4sf)__B);
|
|
1251 }
|
|
1252
|
|
1253 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|
1254 _mm_frcz_sd (__m128d __A, __m128d __B)
|
|
1255 {
|
|
1256 return (__m128d) __builtin_ia32_frczsd ((__v2df)__A, (__v2df)__B);
|
|
1257 }
|
|
1258
|
|
1259 #endif /* __SSE5__ */
|
|
1260
|
|
1261 #endif /* _BMMINTRIN_H_INCLUDED */
|