blob: fd18e7177b778fd09675ec7aaa254c5f90639238 [file] [log] [blame]
Robert Khasanovb9f3a912014-10-08 17:18:13 +00001/*===---- avx512vlintrin.h - AVX512VL 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 __IMMINTRIN_H
25#error "Never use <avx512vlintrin.h> directly; include <immintrin.h> instead."
26#endif
27
28#ifndef __AVX512VLINTRIN_H
29#define __AVX512VLINTRIN_H
30
Michael Kupersteine45af542015-06-30 13:36:19 +000031#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512vl")))
Eric Christopher99af5b22015-10-27 06:11:03 +000032#define __DEFAULT_FN_ATTRS_BOTH __attribute__((__always_inline__, __nodebug__, __target__("avx512vl, avx512bw")))
Eric Christopher4d1851682015-06-17 07:09:20 +000033
Michael Zuckerman38a27272016-02-22 09:05:41 +000034static __inline __v2di __DEFAULT_FN_ATTRS
35_mm_setzero_di(void) {
36 return (__v2di){ 0, 0};
37}
38
Robert Khasanovb9f3a912014-10-08 17:18:13 +000039/* Integer compare */
40
Eric Christopher99af5b22015-10-27 06:11:03 +000041static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
Robert Khasanovb9f3a912014-10-08 17:18:13 +000042_mm_cmpeq_epi32_mask(__m128i __a, __m128i __b) {
43 return (__mmask8)__builtin_ia32_pcmpeqd128_mask((__v4si)__a, (__v4si)__b,
44 (__mmask8)-1);
45}
46
Eric Christopher99af5b22015-10-27 06:11:03 +000047static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
Robert Khasanovb9f3a912014-10-08 17:18:13 +000048_mm_mask_cmpeq_epi32_mask(__mmask8 __u, __m128i __a, __m128i __b) {
49 return (__mmask8)__builtin_ia32_pcmpeqd128_mask((__v4si)__a, (__v4si)__b,
50 __u);
51}
52
Michael Kupersteine45af542015-06-30 13:36:19 +000053static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +000054_mm_cmpeq_epu32_mask(__m128i __a, __m128i __b) {
55 return (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)__a, (__v4si)__b, 0,
56 (__mmask8)-1);
57}
58
Michael Kupersteine45af542015-06-30 13:36:19 +000059static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +000060_mm_mask_cmpeq_epu32_mask(__mmask8 __u, __m128i __a, __m128i __b) {
61 return (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)__a, (__v4si)__b, 0,
62 __u);
63}
Robert Khasanovb9f3a912014-10-08 17:18:13 +000064
Eric Christopher99af5b22015-10-27 06:11:03 +000065static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
Robert Khasanovb9f3a912014-10-08 17:18:13 +000066_mm256_cmpeq_epi32_mask(__m256i __a, __m256i __b) {
67 return (__mmask8)__builtin_ia32_pcmpeqd256_mask((__v8si)__a, (__v8si)__b,
68 (__mmask8)-1);
69}
70
Eric Christopher99af5b22015-10-27 06:11:03 +000071static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
Robert Khasanovb9f3a912014-10-08 17:18:13 +000072_mm256_mask_cmpeq_epi32_mask(__mmask8 __u, __m256i __a, __m256i __b) {
73 return (__mmask8)__builtin_ia32_pcmpeqd256_mask((__v8si)__a, (__v8si)__b,
74 __u);
75}
76
Michael Kupersteine45af542015-06-30 13:36:19 +000077static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +000078_mm256_cmpeq_epu32_mask(__m256i __a, __m256i __b) {
79 return (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)__a, (__v8si)__b, 0,
80 (__mmask8)-1);
81}
82
Michael Kupersteine45af542015-06-30 13:36:19 +000083static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +000084_mm256_mask_cmpeq_epu32_mask(__mmask8 __u, __m256i __a, __m256i __b) {
85 return (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)__a, (__v8si)__b, 0,
86 __u);
87}
88
Eric Christopher99af5b22015-10-27 06:11:03 +000089static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
Robert Khasanovb9f3a912014-10-08 17:18:13 +000090_mm_cmpeq_epi64_mask(__m128i __a, __m128i __b) {
91 return (__mmask8)__builtin_ia32_pcmpeqq128_mask((__v2di)__a, (__v2di)__b,
92 (__mmask8)-1);
93}
94
Eric Christopher99af5b22015-10-27 06:11:03 +000095static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
Robert Khasanovb9f3a912014-10-08 17:18:13 +000096_mm_mask_cmpeq_epi64_mask(__mmask8 __u, __m128i __a, __m128i __b) {
97 return (__mmask8)__builtin_ia32_pcmpeqq128_mask((__v2di)__a, (__v2di)__b,
98 __u);
99}
100
Michael Kupersteine45af542015-06-30 13:36:19 +0000101static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000102_mm_cmpeq_epu64_mask(__m128i __a, __m128i __b) {
103 return (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)__a, (__v2di)__b, 0,
104 (__mmask8)-1);
105}
106
Michael Kupersteine45af542015-06-30 13:36:19 +0000107static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000108_mm_mask_cmpeq_epu64_mask(__mmask8 __u, __m128i __a, __m128i __b) {
109 return (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)__a, (__v2di)__b, 0,
110 __u);
111}
Robert Khasanovb9f3a912014-10-08 17:18:13 +0000112
Eric Christopher99af5b22015-10-27 06:11:03 +0000113static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
Robert Khasanovb9f3a912014-10-08 17:18:13 +0000114_mm256_cmpeq_epi64_mask(__m256i __a, __m256i __b) {
115 return (__mmask8)__builtin_ia32_pcmpeqq256_mask((__v4di)__a, (__v4di)__b,
116 (__mmask8)-1);
117}
118
Eric Christopher99af5b22015-10-27 06:11:03 +0000119static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
Robert Khasanovb9f3a912014-10-08 17:18:13 +0000120_mm256_mask_cmpeq_epi64_mask(__mmask8 __u, __m256i __a, __m256i __b) {
121 return (__mmask8)__builtin_ia32_pcmpeqq256_mask((__v4di)__a, (__v4di)__b,
122 __u);
123}
124
Michael Kupersteine45af542015-06-30 13:36:19 +0000125static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000126_mm256_cmpeq_epu64_mask(__m256i __a, __m256i __b) {
127 return (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)__a, (__v4di)__b, 0,
128 (__mmask8)-1);
129}
130
Michael Kupersteine45af542015-06-30 13:36:19 +0000131static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000132_mm256_mask_cmpeq_epu64_mask(__mmask8 __u, __m256i __a, __m256i __b) {
133 return (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)__a, (__v4di)__b, 0,
134 __u);
135}
136
137
Michael Kupersteine45af542015-06-30 13:36:19 +0000138static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000139_mm_cmpge_epi32_mask(__m128i __a, __m128i __b) {
140 return (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)__a, (__v4si)__b, 5,
141 (__mmask8)-1);
142}
143
Michael Kupersteine45af542015-06-30 13:36:19 +0000144static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000145_mm_mask_cmpge_epi32_mask(__mmask8 __u, __m128i __a, __m128i __b) {
146 return (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)__a, (__v4si)__b, 5,
147 __u);
148}
149
Michael Kupersteine45af542015-06-30 13:36:19 +0000150static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000151_mm_cmpge_epu32_mask(__m128i __a, __m128i __b) {
152 return (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)__a, (__v4si)__b, 5,
153 (__mmask8)-1);
154}
155
Michael Kupersteine45af542015-06-30 13:36:19 +0000156static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000157_mm_mask_cmpge_epu32_mask(__mmask8 __u, __m128i __a, __m128i __b) {
158 return (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)__a, (__v4si)__b, 5,
159 __u);
160}
161
Michael Kupersteine45af542015-06-30 13:36:19 +0000162static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000163_mm256_cmpge_epi32_mask(__m256i __a, __m256i __b) {
164 return (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)__a, (__v8si)__b, 5,
165 (__mmask8)-1);
166}
167
Michael Kupersteine45af542015-06-30 13:36:19 +0000168static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000169_mm256_mask_cmpge_epi32_mask(__mmask8 __u, __m256i __a, __m256i __b) {
170 return (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)__a, (__v8si)__b, 5,
171 __u);
172}
173
Michael Kupersteine45af542015-06-30 13:36:19 +0000174static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000175_mm256_cmpge_epu32_mask(__m256i __a, __m256i __b) {
176 return (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)__a, (__v8si)__b, 5,
177 (__mmask8)-1);
178}
179
Michael Kupersteine45af542015-06-30 13:36:19 +0000180static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000181_mm256_mask_cmpge_epu32_mask(__mmask8 __u, __m256i __a, __m256i __b) {
182 return (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)__a, (__v8si)__b, 5,
183 __u);
184}
185
Michael Kupersteine45af542015-06-30 13:36:19 +0000186static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000187_mm_cmpge_epi64_mask(__m128i __a, __m128i __b) {
188 return (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)__a, (__v2di)__b, 5,
189 (__mmask8)-1);
190}
191
Michael Kupersteine45af542015-06-30 13:36:19 +0000192static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000193_mm_mask_cmpge_epi64_mask(__mmask8 __u, __m128i __a, __m128i __b) {
194 return (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)__a, (__v2di)__b, 5,
195 __u);
196}
197
Michael Kupersteine45af542015-06-30 13:36:19 +0000198static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000199_mm_cmpge_epu64_mask(__m128i __a, __m128i __b) {
200 return (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)__a, (__v2di)__b, 5,
201 (__mmask8)-1);
202}
203
Michael Kupersteine45af542015-06-30 13:36:19 +0000204static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000205_mm_mask_cmpge_epu64_mask(__mmask8 __u, __m128i __a, __m128i __b) {
206 return (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)__a, (__v2di)__b, 5,
207 __u);
208}
209
Michael Kupersteine45af542015-06-30 13:36:19 +0000210static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000211_mm256_cmpge_epi64_mask(__m256i __a, __m256i __b) {
212 return (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)__a, (__v4di)__b, 5,
213 (__mmask8)-1);
214}
215
Michael Kupersteine45af542015-06-30 13:36:19 +0000216static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000217_mm256_mask_cmpge_epi64_mask(__mmask8 __u, __m256i __a, __m256i __b) {
218 return (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)__a, (__v4di)__b, 5,
219 __u);
220}
221
Michael Kupersteine45af542015-06-30 13:36:19 +0000222static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000223_mm256_cmpge_epu64_mask(__m256i __a, __m256i __b) {
224 return (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)__a, (__v4di)__b, 5,
225 (__mmask8)-1);
226}
227
Michael Kupersteine45af542015-06-30 13:36:19 +0000228static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000229_mm256_mask_cmpge_epu64_mask(__mmask8 __u, __m256i __a, __m256i __b) {
230 return (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)__a, (__v4di)__b, 5,
231 __u);
232}
233
Eric Christopher99af5b22015-10-27 06:11:03 +0000234static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
Craig Topper2f25a5a2015-01-26 08:11:49 +0000235_mm_cmpgt_epi32_mask(__m128i __a, __m128i __b) {
236 return (__mmask8)__builtin_ia32_pcmpgtd128_mask((__v4si)__a, (__v4si)__b,
237 (__mmask8)-1);
238}
239
Eric Christopher99af5b22015-10-27 06:11:03 +0000240static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
Craig Topper2f25a5a2015-01-26 08:11:49 +0000241_mm_mask_cmpgt_epi32_mask(__mmask8 __u, __m128i __a, __m128i __b) {
242 return (__mmask8)__builtin_ia32_pcmpgtd128_mask((__v4si)__a, (__v4si)__b,
243 __u);
244}
245
Michael Kupersteine45af542015-06-30 13:36:19 +0000246static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000247_mm_cmpgt_epu32_mask(__m128i __a, __m128i __b) {
248 return (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)__a, (__v4si)__b, 6,
249 (__mmask8)-1);
250}
251
Michael Kupersteine45af542015-06-30 13:36:19 +0000252static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000253_mm_mask_cmpgt_epu32_mask(__mmask8 __u, __m128i __a, __m128i __b) {
254 return (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)__a, (__v4si)__b, 6,
255 __u);
256}
257
Eric Christopher99af5b22015-10-27 06:11:03 +0000258static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
Craig Topper2f25a5a2015-01-26 08:11:49 +0000259_mm256_cmpgt_epi32_mask(__m256i __a, __m256i __b) {
260 return (__mmask8)__builtin_ia32_pcmpgtd256_mask((__v8si)__a, (__v8si)__b,
261 (__mmask8)-1);
262}
263
Eric Christopher99af5b22015-10-27 06:11:03 +0000264static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
Craig Topper2f25a5a2015-01-26 08:11:49 +0000265_mm256_mask_cmpgt_epi32_mask(__mmask8 __u, __m256i __a, __m256i __b) {
266 return (__mmask8)__builtin_ia32_pcmpgtd256_mask((__v8si)__a, (__v8si)__b,
267 __u);
268}
269
Michael Kupersteine45af542015-06-30 13:36:19 +0000270static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000271_mm256_cmpgt_epu32_mask(__m256i __a, __m256i __b) {
272 return (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)__a, (__v8si)__b, 6,
273 (__mmask8)-1);
274}
275
Michael Kupersteine45af542015-06-30 13:36:19 +0000276static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000277_mm256_mask_cmpgt_epu32_mask(__mmask8 __u, __m256i __a, __m256i __b) {
278 return (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)__a, (__v8si)__b, 6,
279 __u);
280}
281
Eric Christopher99af5b22015-10-27 06:11:03 +0000282static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
Craig Topper2f25a5a2015-01-26 08:11:49 +0000283_mm_cmpgt_epi64_mask(__m128i __a, __m128i __b) {
284 return (__mmask8)__builtin_ia32_pcmpgtq128_mask((__v2di)__a, (__v2di)__b,
285 (__mmask8)-1);
286}
287
Eric Christopher99af5b22015-10-27 06:11:03 +0000288static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
Craig Topper2f25a5a2015-01-26 08:11:49 +0000289_mm_mask_cmpgt_epi64_mask(__mmask8 __u, __m128i __a, __m128i __b) {
290 return (__mmask8)__builtin_ia32_pcmpgtq128_mask((__v2di)__a, (__v2di)__b,
291 __u);
292}
293
Michael Kupersteine45af542015-06-30 13:36:19 +0000294static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000295_mm_cmpgt_epu64_mask(__m128i __a, __m128i __b) {
296 return (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)__a, (__v2di)__b, 6,
297 (__mmask8)-1);
298}
299
Michael Kupersteine45af542015-06-30 13:36:19 +0000300static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000301_mm_mask_cmpgt_epu64_mask(__mmask8 __u, __m128i __a, __m128i __b) {
302 return (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)__a, (__v2di)__b, 6,
303 __u);
304}
305
Eric Christopher99af5b22015-10-27 06:11:03 +0000306static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
Craig Topper2f25a5a2015-01-26 08:11:49 +0000307_mm256_cmpgt_epi64_mask(__m256i __a, __m256i __b) {
308 return (__mmask8)__builtin_ia32_pcmpgtq256_mask((__v4di)__a, (__v4di)__b,
309 (__mmask8)-1);
310}
311
Eric Christopher99af5b22015-10-27 06:11:03 +0000312static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
Craig Topper2f25a5a2015-01-26 08:11:49 +0000313_mm256_mask_cmpgt_epi64_mask(__mmask8 __u, __m256i __a, __m256i __b) {
314 return (__mmask8)__builtin_ia32_pcmpgtq256_mask((__v4di)__a, (__v4di)__b,
315 __u);
316}
317
Michael Kupersteine45af542015-06-30 13:36:19 +0000318static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000319_mm256_cmpgt_epu64_mask(__m256i __a, __m256i __b) {
320 return (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)__a, (__v4di)__b, 6,
321 (__mmask8)-1);
322}
323
Michael Kupersteine45af542015-06-30 13:36:19 +0000324static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000325_mm256_mask_cmpgt_epu64_mask(__mmask8 __u, __m256i __a, __m256i __b) {
326 return (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)__a, (__v4di)__b, 6,
327 __u);
328}
329
Michael Kupersteine45af542015-06-30 13:36:19 +0000330static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000331_mm_cmple_epi32_mask(__m128i __a, __m128i __b) {
332 return (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)__a, (__v4si)__b, 2,
333 (__mmask8)-1);
334}
335
Michael Kupersteine45af542015-06-30 13:36:19 +0000336static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000337_mm_mask_cmple_epi32_mask(__mmask8 __u, __m128i __a, __m128i __b) {
338 return (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)__a, (__v4si)__b, 2,
339 __u);
340}
341
Michael Kupersteine45af542015-06-30 13:36:19 +0000342static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000343_mm_cmple_epu32_mask(__m128i __a, __m128i __b) {
344 return (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)__a, (__v4si)__b, 2,
345 (__mmask8)-1);
346}
347
Michael Kupersteine45af542015-06-30 13:36:19 +0000348static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000349_mm_mask_cmple_epu32_mask(__mmask8 __u, __m128i __a, __m128i __b) {
350 return (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)__a, (__v4si)__b, 2,
351 __u);
352}
353
Michael Kupersteine45af542015-06-30 13:36:19 +0000354static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000355_mm256_cmple_epi32_mask(__m256i __a, __m256i __b) {
356 return (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)__a, (__v8si)__b, 2,
357 (__mmask8)-1);
358}
359
Michael Kupersteine45af542015-06-30 13:36:19 +0000360static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000361_mm256_mask_cmple_epi32_mask(__mmask8 __u, __m256i __a, __m256i __b) {
362 return (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)__a, (__v8si)__b, 2,
363 __u);
364}
365
Michael Kupersteine45af542015-06-30 13:36:19 +0000366static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000367_mm256_cmple_epu32_mask(__m256i __a, __m256i __b) {
368 return (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)__a, (__v8si)__b, 2,
369 (__mmask8)-1);
370}
371
Michael Kupersteine45af542015-06-30 13:36:19 +0000372static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000373_mm256_mask_cmple_epu32_mask(__mmask8 __u, __m256i __a, __m256i __b) {
374 return (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)__a, (__v8si)__b, 2,
375 __u);
376}
377
Michael Kupersteine45af542015-06-30 13:36:19 +0000378static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000379_mm_cmple_epi64_mask(__m128i __a, __m128i __b) {
380 return (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)__a, (__v2di)__b, 2,
381 (__mmask8)-1);
382}
383
Michael Kupersteine45af542015-06-30 13:36:19 +0000384static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000385_mm_mask_cmple_epi64_mask(__mmask8 __u, __m128i __a, __m128i __b) {
386 return (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)__a, (__v2di)__b, 2,
387 __u);
388}
389
Michael Kupersteine45af542015-06-30 13:36:19 +0000390static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000391_mm_cmple_epu64_mask(__m128i __a, __m128i __b) {
392 return (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)__a, (__v2di)__b, 2,
393 (__mmask8)-1);
394}
395
Michael Kupersteine45af542015-06-30 13:36:19 +0000396static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000397_mm_mask_cmple_epu64_mask(__mmask8 __u, __m128i __a, __m128i __b) {
398 return (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)__a, (__v2di)__b, 2,
399 __u);
400}
401
Michael Kupersteine45af542015-06-30 13:36:19 +0000402static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000403_mm256_cmple_epi64_mask(__m256i __a, __m256i __b) {
404 return (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)__a, (__v4di)__b, 2,
405 (__mmask8)-1);
406}
407
Michael Kupersteine45af542015-06-30 13:36:19 +0000408static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000409_mm256_mask_cmple_epi64_mask(__mmask8 __u, __m256i __a, __m256i __b) {
410 return (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)__a, (__v4di)__b, 2,
411 __u);
412}
413
Michael Kupersteine45af542015-06-30 13:36:19 +0000414static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000415_mm256_cmple_epu64_mask(__m256i __a, __m256i __b) {
416 return (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)__a, (__v4di)__b, 2,
417 (__mmask8)-1);
418}
419
Michael Kupersteine45af542015-06-30 13:36:19 +0000420static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000421_mm256_mask_cmple_epu64_mask(__mmask8 __u, __m256i __a, __m256i __b) {
422 return (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)__a, (__v4di)__b, 2,
423 __u);
424}
425
Michael Kupersteine45af542015-06-30 13:36:19 +0000426static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000427_mm_cmplt_epi32_mask(__m128i __a, __m128i __b) {
428 return (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)__a, (__v4si)__b, 1,
429 (__mmask8)-1);
430}
431
Michael Kupersteine45af542015-06-30 13:36:19 +0000432static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000433_mm_mask_cmplt_epi32_mask(__mmask8 __u, __m128i __a, __m128i __b) {
434 return (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)__a, (__v4si)__b, 1,
435 __u);
436}
437
Michael Kupersteine45af542015-06-30 13:36:19 +0000438static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000439_mm_cmplt_epu32_mask(__m128i __a, __m128i __b) {
440 return (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)__a, (__v4si)__b, 1,
441 (__mmask8)-1);
442}
443
Michael Kupersteine45af542015-06-30 13:36:19 +0000444static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000445_mm_mask_cmplt_epu32_mask(__mmask8 __u, __m128i __a, __m128i __b) {
446 return (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)__a, (__v4si)__b, 1,
447 __u);
448}
449
Michael Kupersteine45af542015-06-30 13:36:19 +0000450static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000451_mm256_cmplt_epi32_mask(__m256i __a, __m256i __b) {
452 return (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)__a, (__v8si)__b, 1,
453 (__mmask8)-1);
454}
455
Michael Kupersteine45af542015-06-30 13:36:19 +0000456static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000457_mm256_mask_cmplt_epi32_mask(__mmask8 __u, __m256i __a, __m256i __b) {
458 return (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)__a, (__v8si)__b, 1,
459 __u);
460}
461
Michael Kupersteine45af542015-06-30 13:36:19 +0000462static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000463_mm256_cmplt_epu32_mask(__m256i __a, __m256i __b) {
464 return (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)__a, (__v8si)__b, 1,
465 (__mmask8)-1);
466}
467
Michael Kupersteine45af542015-06-30 13:36:19 +0000468static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000469_mm256_mask_cmplt_epu32_mask(__mmask8 __u, __m256i __a, __m256i __b) {
470 return (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)__a, (__v8si)__b, 1,
471 __u);
472}
473
Michael Kupersteine45af542015-06-30 13:36:19 +0000474static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000475_mm_cmplt_epi64_mask(__m128i __a, __m128i __b) {
476 return (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)__a, (__v2di)__b, 1,
477 (__mmask8)-1);
478}
479
Michael Kupersteine45af542015-06-30 13:36:19 +0000480static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000481_mm_mask_cmplt_epi64_mask(__mmask8 __u, __m128i __a, __m128i __b) {
482 return (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)__a, (__v2di)__b, 1,
483 __u);
484}
485
Michael Kupersteine45af542015-06-30 13:36:19 +0000486static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000487_mm_cmplt_epu64_mask(__m128i __a, __m128i __b) {
488 return (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)__a, (__v2di)__b, 1,
489 (__mmask8)-1);
490}
491
Michael Kupersteine45af542015-06-30 13:36:19 +0000492static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000493_mm_mask_cmplt_epu64_mask(__mmask8 __u, __m128i __a, __m128i __b) {
494 return (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)__a, (__v2di)__b, 1,
495 __u);
496}
497
Michael Kupersteine45af542015-06-30 13:36:19 +0000498static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000499_mm256_cmplt_epi64_mask(__m256i __a, __m256i __b) {
500 return (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)__a, (__v4di)__b, 1,
501 (__mmask8)-1);
502}
503
Michael Kupersteine45af542015-06-30 13:36:19 +0000504static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000505_mm256_mask_cmplt_epi64_mask(__mmask8 __u, __m256i __a, __m256i __b) {
506 return (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)__a, (__v4di)__b, 1,
507 __u);
508}
509
Michael Kupersteine45af542015-06-30 13:36:19 +0000510static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000511_mm256_cmplt_epu64_mask(__m256i __a, __m256i __b) {
512 return (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)__a, (__v4di)__b, 1,
513 (__mmask8)-1);
514}
515
Michael Kupersteine45af542015-06-30 13:36:19 +0000516static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000517_mm256_mask_cmplt_epu64_mask(__mmask8 __u, __m256i __a, __m256i __b) {
518 return (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)__a, (__v4di)__b, 1,
519 __u);
520}
521
Michael Kupersteine45af542015-06-30 13:36:19 +0000522static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000523_mm_cmpneq_epi32_mask(__m128i __a, __m128i __b) {
524 return (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)__a, (__v4si)__b, 4,
525 (__mmask8)-1);
526}
527
Michael Kupersteine45af542015-06-30 13:36:19 +0000528static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000529_mm_mask_cmpneq_epi32_mask(__mmask8 __u, __m128i __a, __m128i __b) {
530 return (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)__a, (__v4si)__b, 4,
531 __u);
532}
533
Michael Kupersteine45af542015-06-30 13:36:19 +0000534static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000535_mm_cmpneq_epu32_mask(__m128i __a, __m128i __b) {
536 return (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)__a, (__v4si)__b, 4,
537 (__mmask8)-1);
538}
539
Michael Kupersteine45af542015-06-30 13:36:19 +0000540static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000541_mm_mask_cmpneq_epu32_mask(__mmask8 __u, __m128i __a, __m128i __b) {
542 return (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)__a, (__v4si)__b, 4,
543 __u);
544}
545
Michael Kupersteine45af542015-06-30 13:36:19 +0000546static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000547_mm256_cmpneq_epi32_mask(__m256i __a, __m256i __b) {
548 return (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)__a, (__v8si)__b, 4,
549 (__mmask8)-1);
550}
551
Michael Kupersteine45af542015-06-30 13:36:19 +0000552static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000553_mm256_mask_cmpneq_epi32_mask(__mmask8 __u, __m256i __a, __m256i __b) {
554 return (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)__a, (__v8si)__b, 4,
555 __u);
556}
557
Michael Kupersteine45af542015-06-30 13:36:19 +0000558static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000559_mm256_cmpneq_epu32_mask(__m256i __a, __m256i __b) {
560 return (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)__a, (__v8si)__b, 4,
561 (__mmask8)-1);
562}
563
Michael Kupersteine45af542015-06-30 13:36:19 +0000564static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000565_mm256_mask_cmpneq_epu32_mask(__mmask8 __u, __m256i __a, __m256i __b) {
566 return (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)__a, (__v8si)__b, 4,
567 __u);
568}
569
Michael Kupersteine45af542015-06-30 13:36:19 +0000570static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000571_mm_cmpneq_epi64_mask(__m128i __a, __m128i __b) {
572 return (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)__a, (__v2di)__b, 4,
573 (__mmask8)-1);
574}
575
Michael Kupersteine45af542015-06-30 13:36:19 +0000576static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000577_mm_mask_cmpneq_epi64_mask(__mmask8 __u, __m128i __a, __m128i __b) {
578 return (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)__a, (__v2di)__b, 4,
579 __u);
580}
581
Michael Kupersteine45af542015-06-30 13:36:19 +0000582static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000583_mm_cmpneq_epu64_mask(__m128i __a, __m128i __b) {
584 return (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)__a, (__v2di)__b, 4,
585 (__mmask8)-1);
586}
587
Michael Kupersteine45af542015-06-30 13:36:19 +0000588static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000589_mm_mask_cmpneq_epu64_mask(__mmask8 __u, __m128i __a, __m128i __b) {
590 return (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)__a, (__v2di)__b, 4,
591 __u);
592}
593
Michael Kupersteine45af542015-06-30 13:36:19 +0000594static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000595_mm256_cmpneq_epi64_mask(__m256i __a, __m256i __b) {
596 return (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)__a, (__v4di)__b, 4,
597 (__mmask8)-1);
598}
599
Michael Kupersteine45af542015-06-30 13:36:19 +0000600static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000601_mm256_mask_cmpneq_epi64_mask(__mmask8 __u, __m256i __a, __m256i __b) {
602 return (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)__a, (__v4di)__b, 4,
603 __u);
604}
605
Michael Kupersteine45af542015-06-30 13:36:19 +0000606static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000607_mm256_cmpneq_epu64_mask(__m256i __a, __m256i __b) {
608 return (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)__a, (__v4di)__b, 4,
609 (__mmask8)-1);
610}
611
Michael Kupersteine45af542015-06-30 13:36:19 +0000612static __inline__ __mmask8 __DEFAULT_FN_ATTRS
Craig Topper2f25a5a2015-01-26 08:11:49 +0000613_mm256_mask_cmpneq_epu64_mask(__mmask8 __u, __m256i __a, __m256i __b) {
614 return (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)__a, (__v4di)__b, 4,
615 __u);
616}
617
Michael Kupersteine45af542015-06-30 13:36:19 +0000618static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000619_mm256_mask_add_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
620 __m256i __B)
621{
622 return (__m256i) __builtin_ia32_paddd256_mask ((__v8si) __A,
623 (__v8si) __B,
624 (__v8si) __W,
625 (__mmask8) __U);
626}
627
Michael Kupersteine45af542015-06-30 13:36:19 +0000628static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000629_mm256_maskz_add_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
630{
631 return (__m256i) __builtin_ia32_paddd256_mask ((__v8si) __A,
632 (__v8si) __B,
633 (__v8si)
634 _mm256_setzero_si256 (),
635 (__mmask8) __U);
636}
637
Michael Kupersteine45af542015-06-30 13:36:19 +0000638static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000639_mm256_mask_add_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
640 __m256i __B)
641{
642 return (__m256i) __builtin_ia32_paddq256_mask ((__v4di) __A,
643 (__v4di) __B,
644 (__v4di) __W,
645 (__mmask8) __U);
646}
647
Michael Kupersteine45af542015-06-30 13:36:19 +0000648static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000649_mm256_maskz_add_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
650{
651 return (__m256i) __builtin_ia32_paddq256_mask ((__v4di) __A,
652 (__v4di) __B,
653 (__v4di)
654 _mm256_setzero_si256 (),
655 (__mmask8) __U);
656}
657
Michael Kupersteine45af542015-06-30 13:36:19 +0000658static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000659_mm256_mask_sub_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
660 __m256i __B)
661{
662 return (__m256i) __builtin_ia32_psubd256_mask ((__v8si) __A,
663 (__v8si) __B,
664 (__v8si) __W,
665 (__mmask8) __U);
666}
667
Michael Kupersteine45af542015-06-30 13:36:19 +0000668static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000669_mm256_maskz_sub_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
670{
671 return (__m256i) __builtin_ia32_psubd256_mask ((__v8si) __A,
672 (__v8si) __B,
673 (__v8si)
674 _mm256_setzero_si256 (),
675 (__mmask8) __U);
676}
677
Michael Kupersteine45af542015-06-30 13:36:19 +0000678static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000679_mm256_mask_sub_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
680 __m256i __B)
681{
682 return (__m256i) __builtin_ia32_psubq256_mask ((__v4di) __A,
683 (__v4di) __B,
684 (__v4di) __W,
685 (__mmask8) __U);
686}
687
Michael Kupersteine45af542015-06-30 13:36:19 +0000688static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000689_mm256_maskz_sub_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
690{
691 return (__m256i) __builtin_ia32_psubq256_mask ((__v4di) __A,
692 (__v4di) __B,
693 (__v4di)
694 _mm256_setzero_si256 (),
695 (__mmask8) __U);
696}
697
Michael Kupersteine45af542015-06-30 13:36:19 +0000698static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000699_mm_mask_add_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
700 __m128i __B)
701{
702 return (__m128i) __builtin_ia32_paddd128_mask ((__v4si) __A,
703 (__v4si) __B,
704 (__v4si) __W,
705 (__mmask8) __U);
706}
707
Michael Kupersteine45af542015-06-30 13:36:19 +0000708static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000709_mm_maskz_add_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
710{
711 return (__m128i) __builtin_ia32_paddd128_mask ((__v4si) __A,
712 (__v4si) __B,
713 (__v4si)
714 _mm_setzero_si128 (),
715 (__mmask8) __U);
716}
717
Michael Kupersteine45af542015-06-30 13:36:19 +0000718static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000719_mm_mask_add_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
720 __m128i __B)
721{
722 return (__m128i) __builtin_ia32_paddq128_mask ((__v2di) __A,
723 (__v2di) __B,
724 (__v2di) __W,
725 (__mmask8) __U);
726}
727
Michael Kupersteine45af542015-06-30 13:36:19 +0000728static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000729_mm_maskz_add_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
730{
731 return (__m128i) __builtin_ia32_paddq128_mask ((__v2di) __A,
732 (__v2di) __B,
733 (__v2di)
734 _mm_setzero_si128 (),
735 (__mmask8) __U);
736}
737
Michael Kupersteine45af542015-06-30 13:36:19 +0000738static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000739_mm_mask_sub_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
740 __m128i __B)
741{
742 return (__m128i) __builtin_ia32_psubd128_mask ((__v4si) __A,
743 (__v4si) __B,
744 (__v4si) __W,
745 (__mmask8) __U);
746}
747
Michael Kupersteine45af542015-06-30 13:36:19 +0000748static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000749_mm_maskz_sub_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
750{
751 return (__m128i) __builtin_ia32_psubd128_mask ((__v4si) __A,
752 (__v4si) __B,
753 (__v4si)
754 _mm_setzero_si128 (),
755 (__mmask8) __U);
756}
757
Michael Kupersteine45af542015-06-30 13:36:19 +0000758static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000759_mm_mask_sub_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
760 __m128i __B)
761{
762 return (__m128i) __builtin_ia32_psubq128_mask ((__v2di) __A,
763 (__v2di) __B,
764 (__v2di) __W,
765 (__mmask8) __U);
766}
767
Michael Kupersteine45af542015-06-30 13:36:19 +0000768static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000769_mm_maskz_sub_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
770{
771 return (__m128i) __builtin_ia32_psubq128_mask ((__v2di) __A,
772 (__v2di) __B,
773 (__v2di)
774 _mm_setzero_si128 (),
775 (__mmask8) __U);
776}
777
Michael Kupersteine45af542015-06-30 13:36:19 +0000778static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000779_mm256_mask_mul_epi32 (__m256i __W, __mmask8 __M, __m256i __X,
780 __m256i __Y)
781{
782 return (__m256i) __builtin_ia32_pmuldq256_mask ((__v8si) __X,
783 (__v8si) __Y,
784 (__v4di) __W, __M);
785}
786
Michael Kupersteine45af542015-06-30 13:36:19 +0000787static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000788_mm256_maskz_mul_epi32 (__mmask8 __M, __m256i __X, __m256i __Y)
789{
790 return (__m256i) __builtin_ia32_pmuldq256_mask ((__v8si) __X,
791 (__v8si) __Y,
792 (__v4di)
793 _mm256_setzero_si256 (),
794 __M);
795}
796
Michael Kupersteine45af542015-06-30 13:36:19 +0000797static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000798_mm_mask_mul_epi32 (__m128i __W, __mmask8 __M, __m128i __X,
799 __m128i __Y)
800{
801 return (__m128i) __builtin_ia32_pmuldq128_mask ((__v4si) __X,
802 (__v4si) __Y,
803 (__v2di) __W, __M);
804}
805
Michael Kupersteine45af542015-06-30 13:36:19 +0000806static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000807_mm_maskz_mul_epi32 (__mmask8 __M, __m128i __X, __m128i __Y)
808{
809 return (__m128i) __builtin_ia32_pmuldq128_mask ((__v4si) __X,
810 (__v4si) __Y,
811 (__v2di)
812 _mm_setzero_si128 (),
813 __M);
814}
815
Michael Kupersteine45af542015-06-30 13:36:19 +0000816static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000817_mm256_mask_mul_epu32 (__m256i __W, __mmask8 __M, __m256i __X,
818 __m256i __Y)
819{
820 return (__m256i) __builtin_ia32_pmuludq256_mask ((__v8si) __X,
821 (__v8si) __Y,
822 (__v4di) __W, __M);
823}
824
Michael Kupersteine45af542015-06-30 13:36:19 +0000825static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000826_mm256_maskz_mul_epu32 (__mmask8 __M, __m256i __X, __m256i __Y)
827{
828 return (__m256i) __builtin_ia32_pmuludq256_mask ((__v8si) __X,
829 (__v8si) __Y,
830 (__v4di)
831 _mm256_setzero_si256 (),
832 __M);
833}
834
Michael Kupersteine45af542015-06-30 13:36:19 +0000835static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000836_mm_mask_mul_epu32 (__m128i __W, __mmask8 __M, __m128i __X,
837 __m128i __Y)
838{
839 return (__m128i) __builtin_ia32_pmuludq128_mask ((__v4si) __X,
840 (__v4si) __Y,
841 (__v2di) __W, __M);
842}
843
Michael Kupersteine45af542015-06-30 13:36:19 +0000844static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000845_mm_maskz_mul_epu32 (__mmask8 __M, __m128i __X, __m128i __Y)
846{
847 return (__m128i) __builtin_ia32_pmuludq128_mask ((__v4si) __X,
848 (__v4si) __Y,
849 (__v2di)
850 _mm_setzero_si128 (),
851 __M);
852}
853
Michael Kupersteine45af542015-06-30 13:36:19 +0000854static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000855_mm256_maskz_mullo_epi32 (__mmask8 __M, __m256i __A, __m256i __B)
856{
857 return (__m256i) __builtin_ia32_pmulld256_mask ((__v8si) __A,
858 (__v8si) __B,
859 (__v8si)
860 _mm256_setzero_si256 (),
861 __M);
862}
863
Michael Kupersteine45af542015-06-30 13:36:19 +0000864static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000865_mm256_mask_mullo_epi32 (__m256i __W, __mmask8 __M, __m256i __A,
866 __m256i __B)
867{
868 return (__m256i) __builtin_ia32_pmulld256_mask ((__v8si) __A,
869 (__v8si) __B,
870 (__v8si) __W, __M);
871}
872
Michael Kupersteine45af542015-06-30 13:36:19 +0000873static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000874_mm_maskz_mullo_epi32 (__mmask8 __M, __m128i __A, __m128i __B)
875{
876 return (__m128i) __builtin_ia32_pmulld128_mask ((__v4si) __A,
877 (__v4si) __B,
878 (__v4si)
879 _mm_setzero_si128 (),
880 __M);
881}
882
Michael Kupersteine45af542015-06-30 13:36:19 +0000883static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000884_mm_mask_mullo_epi32 (__m128i __W, __mmask16 __M, __m128i __A,
885 __m128i __B)
886{
887 return (__m128i) __builtin_ia32_pmulld128_mask ((__v4si) __A,
888 (__v4si) __B,
889 (__v4si) __W, __M);
890}
891
Michael Kupersteine45af542015-06-30 13:36:19 +0000892static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000893_mm256_mask_and_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
894 __m256i __B)
895{
896 return (__m256i) __builtin_ia32_pandd256_mask ((__v8si) __A,
897 (__v8si) __B,
898 (__v8si) __W,
899 (__mmask8) __U);
900}
901
Michael Kupersteine45af542015-06-30 13:36:19 +0000902static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000903_mm256_maskz_and_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
904{
905 return (__m256i) __builtin_ia32_pandd256_mask ((__v8si) __A,
906 (__v8si) __B,
907 (__v8si)
908 _mm256_setzero_si256 (),
909 (__mmask8) __U);
910}
911
Michael Kupersteine45af542015-06-30 13:36:19 +0000912static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000913_mm_mask_and_epi32 (__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
914{
915 return (__m128i) __builtin_ia32_pandd128_mask ((__v4si) __A,
916 (__v4si) __B,
917 (__v4si) __W,
918 (__mmask8) __U);
919}
920
Michael Kupersteine45af542015-06-30 13:36:19 +0000921static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000922_mm_maskz_and_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
923{
924 return (__m128i) __builtin_ia32_pandd128_mask ((__v4si) __A,
925 (__v4si) __B,
926 (__v4si)
927 _mm_setzero_si128 (),
928 (__mmask8) __U);
929}
930
Michael Kupersteine45af542015-06-30 13:36:19 +0000931static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000932_mm256_mask_andnot_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
933 __m256i __B)
934{
935 return (__m256i) __builtin_ia32_pandnd256_mask ((__v8si) __A,
936 (__v8si) __B,
937 (__v8si) __W,
938 (__mmask8) __U);
939}
940
Michael Kupersteine45af542015-06-30 13:36:19 +0000941static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000942_mm256_maskz_andnot_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
943{
944 return (__m256i) __builtin_ia32_pandnd256_mask ((__v8si) __A,
945 (__v8si) __B,
946 (__v8si)
947 _mm256_setzero_si256 (),
948 (__mmask8) __U);
949}
950
Michael Kupersteine45af542015-06-30 13:36:19 +0000951static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000952_mm_mask_andnot_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
953 __m128i __B)
954{
955 return (__m128i) __builtin_ia32_pandnd128_mask ((__v4si) __A,
956 (__v4si) __B,
957 (__v4si) __W,
958 (__mmask8) __U);
959}
960
Michael Kupersteine45af542015-06-30 13:36:19 +0000961static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000962_mm_maskz_andnot_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
963{
964 return (__m128i) __builtin_ia32_pandnd128_mask ((__v4si) __A,
965 (__v4si) __B,
966 (__v4si)
967 _mm_setzero_si128 (),
968 (__mmask8) __U);
969}
970
Michael Kupersteine45af542015-06-30 13:36:19 +0000971static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000972_mm256_mask_or_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
973 __m256i __B)
974{
975 return (__m256i) __builtin_ia32_pord256_mask ((__v8si) __A,
976 (__v8si) __B,
977 (__v8si) __W,
978 (__mmask8) __U);
979}
980
Michael Kupersteine45af542015-06-30 13:36:19 +0000981static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000982_mm256_maskz_or_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
983{
984 return (__m256i) __builtin_ia32_pord256_mask ((__v8si) __A,
985 (__v8si) __B,
986 (__v8si)
987 _mm256_setzero_si256 (),
988 (__mmask8) __U);
989}
990
Michael Kupersteine45af542015-06-30 13:36:19 +0000991static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +0000992_mm_mask_or_epi32 (__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
993{
994 return (__m128i) __builtin_ia32_pord128_mask ((__v4si) __A,
995 (__v4si) __B,
996 (__v4si) __W,
997 (__mmask8) __U);
998}
999
Michael Kupersteine45af542015-06-30 13:36:19 +00001000static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +00001001_mm_maskz_or_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
1002{
1003 return (__m128i) __builtin_ia32_pord128_mask ((__v4si) __A,
1004 (__v4si) __B,
1005 (__v4si)
1006 _mm_setzero_si128 (),
1007 (__mmask8) __U);
1008}
1009
Michael Kupersteine45af542015-06-30 13:36:19 +00001010static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +00001011_mm256_mask_xor_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
1012 __m256i __B)
1013{
1014 return (__m256i) __builtin_ia32_pxord256_mask ((__v8si) __A,
1015 (__v8si) __B,
1016 (__v8si) __W,
1017 (__mmask8) __U);
1018}
1019
Michael Kupersteine45af542015-06-30 13:36:19 +00001020static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +00001021_mm256_maskz_xor_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
1022{
1023 return (__m256i) __builtin_ia32_pxord256_mask ((__v8si) __A,
1024 (__v8si) __B,
1025 (__v8si)
1026 _mm256_setzero_si256 (),
1027 (__mmask8) __U);
1028}
1029
Michael Kupersteine45af542015-06-30 13:36:19 +00001030static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +00001031_mm_mask_xor_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
1032 __m128i __B)
1033{
1034 return (__m128i) __builtin_ia32_pxord128_mask ((__v4si) __A,
1035 (__v4si) __B,
1036 (__v4si) __W,
1037 (__mmask8) __U);
1038}
1039
Michael Kupersteine45af542015-06-30 13:36:19 +00001040static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +00001041_mm_maskz_xor_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
1042{
1043 return (__m128i) __builtin_ia32_pxord128_mask ((__v4si) __A,
1044 (__v4si) __B,
1045 (__v4si)
1046 _mm_setzero_si128 (),
1047 (__mmask8) __U);
1048}
1049
Michael Kupersteine45af542015-06-30 13:36:19 +00001050static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +00001051_mm256_mask_and_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
1052 __m256i __B)
1053{
1054 return (__m256i) __builtin_ia32_pandq256_mask ((__v4di) __A,
1055 (__v4di) __B,
1056 (__v4di) __W, __U);
1057}
1058
Michael Kupersteine45af542015-06-30 13:36:19 +00001059static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +00001060_mm256_maskz_and_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
1061{
1062 return (__m256i) __builtin_ia32_pandq256_mask ((__v4di) __A,
1063 (__v4di) __B,
1064 (__v4di)
1065 _mm256_setzero_pd (),
1066 __U);
1067}
1068
Michael Kupersteine45af542015-06-30 13:36:19 +00001069static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +00001070_mm_mask_and_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
1071 __m128i __B)
1072{
1073 return (__m128i) __builtin_ia32_pandq128_mask ((__v2di) __A,
1074 (__v2di) __B,
1075 (__v2di) __W, __U);
1076}
1077
Michael Kupersteine45af542015-06-30 13:36:19 +00001078static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +00001079_mm_maskz_and_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
1080{
1081 return (__m128i) __builtin_ia32_pandq128_mask ((__v2di) __A,
1082 (__v2di) __B,
1083 (__v2di)
1084 _mm_setzero_pd (),
1085 __U);
1086}
1087
Michael Kupersteine45af542015-06-30 13:36:19 +00001088static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +00001089_mm256_mask_andnot_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
1090 __m256i __B)
1091{
1092 return (__m256i) __builtin_ia32_pandnq256_mask ((__v4di) __A,
1093 (__v4di) __B,
1094 (__v4di) __W, __U);
1095}
1096
Michael Kupersteine45af542015-06-30 13:36:19 +00001097static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +00001098_mm256_maskz_andnot_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
1099{
1100 return (__m256i) __builtin_ia32_pandnq256_mask ((__v4di) __A,
1101 (__v4di) __B,
1102 (__v4di)
1103 _mm256_setzero_pd (),
1104 __U);
1105}
1106
Michael Kupersteine45af542015-06-30 13:36:19 +00001107static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +00001108_mm_mask_andnot_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
1109 __m128i __B)
1110{
1111 return (__m128i) __builtin_ia32_pandnq128_mask ((__v2di) __A,
1112 (__v2di) __B,
1113 (__v2di) __W, __U);
1114}
1115
Michael Kupersteine45af542015-06-30 13:36:19 +00001116static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +00001117_mm_maskz_andnot_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
1118{
1119 return (__m128i) __builtin_ia32_pandnq128_mask ((__v2di) __A,
1120 (__v2di) __B,
1121 (__v2di)
1122 _mm_setzero_pd (),
1123 __U);
1124}
1125
Michael Kupersteine45af542015-06-30 13:36:19 +00001126static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +00001127_mm256_mask_or_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
1128 __m256i __B)
1129{
1130 return (__m256i) __builtin_ia32_porq256_mask ((__v4di) __A,
1131 (__v4di) __B,
1132 (__v4di) __W,
1133 (__mmask8) __U);
1134}
1135
Michael Kupersteine45af542015-06-30 13:36:19 +00001136static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +00001137_mm256_maskz_or_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
1138{
1139 return (__m256i) __builtin_ia32_porq256_mask ((__v4di) __A,
1140 (__v4di) __B,
1141 (__v4di)
1142 _mm256_setzero_si256 (),
1143 (__mmask8) __U);
1144}
1145
Michael Kupersteine45af542015-06-30 13:36:19 +00001146static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +00001147_mm_mask_or_epi64 (__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
1148{
1149 return (__m128i) __builtin_ia32_porq128_mask ((__v2di) __A,
1150 (__v2di) __B,
1151 (__v2di) __W,
1152 (__mmask8) __U);
1153}
1154
Michael Kupersteine45af542015-06-30 13:36:19 +00001155static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +00001156_mm_maskz_or_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
1157{
1158 return (__m128i) __builtin_ia32_porq128_mask ((__v2di) __A,
1159 (__v2di) __B,
1160 (__v2di)
1161 _mm_setzero_si128 (),
1162 (__mmask8) __U);
1163}
1164
Michael Kupersteine45af542015-06-30 13:36:19 +00001165static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +00001166_mm256_mask_xor_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
1167 __m256i __B)
1168{
1169 return (__m256i) __builtin_ia32_pxorq256_mask ((__v4di) __A,
1170 (__v4di) __B,
1171 (__v4di) __W,
1172 (__mmask8) __U);
1173}
1174
Michael Kupersteine45af542015-06-30 13:36:19 +00001175static __inline__ __m256i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +00001176_mm256_maskz_xor_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
1177{
1178 return (__m256i) __builtin_ia32_pxorq256_mask ((__v4di) __A,
1179 (__v4di) __B,
1180 (__v4di)
1181 _mm256_setzero_si256 (),
1182 (__mmask8) __U);
1183}
1184
Michael Kupersteine45af542015-06-30 13:36:19 +00001185static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +00001186_mm_mask_xor_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
1187 __m128i __B)
1188{
1189 return (__m128i) __builtin_ia32_pxorq128_mask ((__v2di) __A,
1190 (__v2di) __B,
1191 (__v2di) __W,
1192 (__mmask8) __U);
1193}
1194
Michael Kupersteine45af542015-06-30 13:36:19 +00001195static __inline__ __m128i __DEFAULT_FN_ATTRS
Elena Demikhovsky35dc8c02015-04-28 13:28:01 +00001196_mm_maskz_xor_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
1197{
1198 return (__m128i) __builtin_ia32_pxorq128_mask ((__v2di) __A,
1199 (__v2di) __B,
1200 (__v2di)
1201 _mm_setzero_si128 (),
1202 (__mmask8) __U);
1203}
1204
Craig Topper2f25a5a2015-01-26 08:11:49 +00001205#define _mm_cmp_epi32_mask(a, b, p) __extension__ ({ \
1206 (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)(__m128i)(a), \
1207 (__v4si)(__m128i)(b), \
1208 (p), (__mmask8)-1); })
1209
1210#define _mm_mask_cmp_epi32_mask(m, a, b, p) __extension__ ({ \
1211 (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)(__m128i)(a), \
1212 (__v4si)(__m128i)(b), \
1213 (p), (__mmask8)(m)); })
1214
1215#define _mm_cmp_epu32_mask(a, b, p) __extension__ ({ \
1216 (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)(__m128i)(a), \
1217 (__v4si)(__m128i)(b), \
1218 (p), (__mmask8)-1); })
1219
1220#define _mm_mask_cmp_epu32_mask(m, a, b, p) __extension__ ({ \
1221 (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)(__m128i)(a), \
1222 (__v4si)(__m128i)(b), \
1223 (p), (__mmask8)(m)); })
1224
1225#define _mm256_cmp_epi32_mask(a, b, p) __extension__ ({ \
1226 (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)(__m256i)(a), \
1227 (__v8si)(__m256i)(b), \
1228 (p), (__mmask8)-1); })
1229
1230#define _mm256_mask_cmp_epi32_mask(m, a, b, p) __extension__ ({ \
1231 (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)(__m256i)(a), \
1232 (__v8si)(__m256i)(b), \
1233 (p), (__mmask8)(m)); })
1234
1235#define _mm256_cmp_epu32_mask(a, b, p) __extension__ ({ \
1236 (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)(__m256i)(a), \
1237 (__v8si)(__m256i)(b), \
1238 (p), (__mmask8)-1); })
1239
1240#define _mm256_mask_cmp_epu32_mask(m, a, b, p) __extension__ ({ \
1241 (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)(__m256i)(a), \
1242 (__v8si)(__m256i)(b), \
1243 (p), (__mmask8)(m)); })
1244
1245#define _mm_cmp_epi64_mask(a, b, p) __extension__ ({ \
1246 (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)(__m128i)(a), \
1247 (__v2di)(__m128i)(b), \
1248 (p), (__mmask8)-1); })
1249
1250#define _mm_mask_cmp_epi64_mask(m, a, b, p) __extension__ ({ \
1251 (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)(__m128i)(a), \
1252 (__v2di)(__m128i)(b), \
1253 (p), (__mmask8)(m)); })
1254
1255#define _mm_cmp_epu64_mask(a, b, p) __extension__ ({ \
1256 (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)(__m128i)(a), \
1257 (__v2di)(__m128i)(b), \
1258 (p), (__mmask8)-1); })
1259
1260#define _mm_mask_cmp_epu64_mask(m, a, b, p) __extension__ ({ \
1261 (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)(__m128i)(a), \
1262 (__v2di)(__m128i)(b), \
1263 (p), (__mmask8)(m)); })
1264
1265#define _mm256_cmp_epi64_mask(a, b, p) __extension__ ({ \
1266 (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)(__m256i)(a), \
1267 (__v4di)(__m256i)(b), \
1268 (p), (__mmask8)-1); })
1269
1270#define _mm256_mask_cmp_epi64_mask(m, a, b, p) __extension__ ({ \
1271 (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)(__m256i)(a), \
1272 (__v4di)(__m256i)(b), \
1273 (p), (__mmask8)(m)); })
1274
1275#define _mm256_cmp_epu64_mask(a, b, p) __extension__ ({ \
1276 (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)(__m256i)(a), \
1277 (__v4di)(__m256i)(b), \
1278 (p), (__mmask8)-1); })
1279
1280#define _mm256_mask_cmp_epu64_mask(m, a, b, p) __extension__ ({ \
1281 (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)(__m256i)(a), \
1282 (__v4di)(__m256i)(b), \
1283 (p), (__mmask8)(m)); })
1284
Elena Demikhovskybd5c8b92015-05-07 11:26:36 +00001285#define _mm256_cmp_ps_mask(a, b, p) __extension__ ({ \
1286 (__mmask8)__builtin_ia32_cmpps256_mask((__v8sf)(__m256)(a), \
1287 (__v8sf)(__m256)(b), \
1288 (p), (__mmask8)-1); })
1289
1290#define _mm256_mask_cmp_ps_mask(m, a, b, p) __extension__ ({ \
1291 (__mmask8)__builtin_ia32_cmpps256_mask((__v8sf)(__m256)(a), \
1292 (__v8sf)(__m256)(b), \
1293 (p), (__mmask8)(m)); })
1294
1295#define _mm256_cmp_pd_mask(a, b, p) __extension__ ({ \
1296 (__mmask8)__builtin_ia32_cmppd256_mask((__v4df)(__m256)(a), \
1297 (__v4df)(__m256)(b), \
1298 (p), (__mmask8)-1); })
1299
1300#define _mm256_mask_cmp_pd_mask(m, a, b, p) __extension__ ({ \
1301 (__mmask8)__builtin_ia32_cmppd256_mask((__v4df)(__m256)(a), \
1302 (__v4df)(__m256)(b), \
1303 (p), (__mmask8)(m)); })
1304
1305#define _mm128_cmp_ps_mask(a, b, p) __extension__ ({ \
1306 (__mmask8)__builtin_ia32_cmpps128_mask((__v4sf)(__m128)(a), \
1307 (__v4sf)(__m128)(b), \
1308 (p), (__mmask8)-1); })
1309
1310#define _mm128_mask_cmp_ps_mask(m, a, b, p) __extension__ ({ \
1311 (__mmask8)__builtin_ia32_cmpps128_mask((__v4sf)(__m128)(a), \
1312 (__v4sf)(__m128)(b), \
1313 (p), (__mmask8)(m)); })
1314
1315#define _mm128_cmp_pd_mask(a, b, p) __extension__ ({ \
1316 (__mmask8)__builtin_ia32_cmppd128_mask((__v2df)(__m128)(a), \
1317 (__v2df)(__m128)(b), \
1318 (p), (__mmask8)-1); })
1319
1320#define _mm128_mask_cmp_pd_mask(m, a, b, p) __extension__ ({ \
1321 (__mmask8)__builtin_ia32_cmppd128_mask((__v2df)(__m128)(a), \
1322 (__v2df)(__m128)(b), \
1323 (p), (__mmask8)(m)); })
Eric Christopher4d1851682015-06-17 07:09:20 +00001324
Michael Kupersteine45af542015-06-30 13:36:19 +00001325static __inline__ __m128d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001326_mm_mask_fmadd_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C)
1327{
1328 return (__m128d) __builtin_ia32_vfmaddpd128_mask ((__v2df) __A,
1329 (__v2df) __B,
1330 (__v2df) __C,
1331 (__mmask8) __U);
1332}
1333
Michael Kupersteine45af542015-06-30 13:36:19 +00001334static __inline__ __m128d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001335_mm_mask3_fmadd_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U)
1336{
1337 return (__m128d) __builtin_ia32_vfmaddpd128_mask3 ((__v2df) __A,
1338 (__v2df) __B,
1339 (__v2df) __C,
1340 (__mmask8) __U);
1341}
1342
Michael Kupersteine45af542015-06-30 13:36:19 +00001343static __inline__ __m128d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001344_mm_maskz_fmadd_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C)
1345{
1346 return (__m128d) __builtin_ia32_vfmaddpd128_maskz ((__v2df) __A,
1347 (__v2df) __B,
1348 (__v2df) __C,
1349 (__mmask8) __U);
1350}
1351
Michael Kupersteine45af542015-06-30 13:36:19 +00001352static __inline__ __m128d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001353_mm_mask_fmsub_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C)
1354{
1355 return (__m128d) __builtin_ia32_vfmaddpd128_mask ((__v2df) __A,
1356 (__v2df) __B,
1357 -(__v2df) __C,
1358 (__mmask8) __U);
1359}
1360
Michael Kupersteine45af542015-06-30 13:36:19 +00001361static __inline__ __m128d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001362_mm_maskz_fmsub_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C)
1363{
1364 return (__m128d) __builtin_ia32_vfmaddpd128_maskz ((__v2df) __A,
1365 (__v2df) __B,
1366 -(__v2df) __C,
1367 (__mmask8) __U);
1368}
1369
Michael Kupersteine45af542015-06-30 13:36:19 +00001370static __inline__ __m128d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001371_mm_mask3_fnmadd_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U)
1372{
1373 return (__m128d) __builtin_ia32_vfmaddpd128_mask3 (-(__v2df) __A,
1374 (__v2df) __B,
1375 (__v2df) __C,
1376 (__mmask8) __U);
1377}
1378
Michael Kupersteine45af542015-06-30 13:36:19 +00001379static __inline__ __m128d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001380_mm_maskz_fnmadd_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C)
1381{
1382 return (__m128d) __builtin_ia32_vfmaddpd128_maskz (-(__v2df) __A,
1383 (__v2df) __B,
1384 (__v2df) __C,
1385 (__mmask8) __U);
1386}
1387
Michael Kupersteine45af542015-06-30 13:36:19 +00001388static __inline__ __m128d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001389_mm_maskz_fnmsub_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C)
1390{
1391 return (__m128d) __builtin_ia32_vfmaddpd128_maskz (-(__v2df) __A,
1392 (__v2df) __B,
1393 -(__v2df) __C,
1394 (__mmask8) __U);
1395}
1396
Michael Kupersteine45af542015-06-30 13:36:19 +00001397static __inline__ __m256d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001398_mm256_mask_fmadd_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C)
1399{
1400 return (__m256d) __builtin_ia32_vfmaddpd256_mask ((__v4df) __A,
1401 (__v4df) __B,
1402 (__v4df) __C,
1403 (__mmask8) __U);
1404}
1405
Michael Kupersteine45af542015-06-30 13:36:19 +00001406static __inline__ __m256d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001407_mm256_mask3_fmadd_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U)
1408{
1409 return (__m256d) __builtin_ia32_vfmaddpd256_mask3 ((__v4df) __A,
1410 (__v4df) __B,
1411 (__v4df) __C,
1412 (__mmask8) __U);
1413}
1414
Michael Kupersteine45af542015-06-30 13:36:19 +00001415static __inline__ __m256d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001416_mm256_maskz_fmadd_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C)
1417{
1418 return (__m256d) __builtin_ia32_vfmaddpd256_maskz ((__v4df) __A,
1419 (__v4df) __B,
1420 (__v4df) __C,
1421 (__mmask8) __U);
1422}
1423
Michael Kupersteine45af542015-06-30 13:36:19 +00001424static __inline__ __m256d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001425_mm256_mask_fmsub_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C)
1426{
1427 return (__m256d) __builtin_ia32_vfmaddpd256_mask ((__v4df) __A,
1428 (__v4df) __B,
1429 -(__v4df) __C,
1430 (__mmask8) __U);
1431}
1432
Michael Kupersteine45af542015-06-30 13:36:19 +00001433static __inline__ __m256d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001434_mm256_maskz_fmsub_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C)
1435{
1436 return (__m256d) __builtin_ia32_vfmaddpd256_maskz ((__v4df) __A,
1437 (__v4df) __B,
1438 -(__v4df) __C,
1439 (__mmask8) __U);
1440}
1441
Michael Kupersteine45af542015-06-30 13:36:19 +00001442static __inline__ __m256d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001443_mm256_mask3_fnmadd_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U)
1444{
1445 return (__m256d) __builtin_ia32_vfmaddpd256_mask3 (-(__v4df) __A,
1446 (__v4df) __B,
1447 (__v4df) __C,
1448 (__mmask8) __U);
1449}
1450
Michael Kupersteine45af542015-06-30 13:36:19 +00001451static __inline__ __m256d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001452_mm256_maskz_fnmadd_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C)
1453{
1454 return (__m256d) __builtin_ia32_vfmaddpd256_maskz (-(__v4df) __A,
1455 (__v4df) __B,
1456 (__v4df) __C,
1457 (__mmask8) __U);
1458}
1459
Michael Kupersteine45af542015-06-30 13:36:19 +00001460static __inline__ __m256d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001461_mm256_maskz_fnmsub_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C)
1462{
1463 return (__m256d) __builtin_ia32_vfmaddpd256_maskz (-(__v4df) __A,
1464 (__v4df) __B,
1465 -(__v4df) __C,
1466 (__mmask8) __U);
1467}
1468
Michael Kupersteine45af542015-06-30 13:36:19 +00001469static __inline__ __m128 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001470_mm_mask_fmadd_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C)
1471{
1472 return (__m128) __builtin_ia32_vfmaddps128_mask ((__v4sf) __A,
1473 (__v4sf) __B,
1474 (__v4sf) __C,
1475 (__mmask8) __U);
1476}
1477
Michael Kupersteine45af542015-06-30 13:36:19 +00001478static __inline__ __m128 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001479_mm_mask3_fmadd_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U)
1480{
1481 return (__m128) __builtin_ia32_vfmaddps128_mask3 ((__v4sf) __A,
1482 (__v4sf) __B,
1483 (__v4sf) __C,
1484 (__mmask8) __U);
1485}
1486
Michael Kupersteine45af542015-06-30 13:36:19 +00001487static __inline__ __m128 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001488_mm_maskz_fmadd_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C)
1489{
1490 return (__m128) __builtin_ia32_vfmaddps128_maskz ((__v4sf) __A,
1491 (__v4sf) __B,
1492 (__v4sf) __C,
1493 (__mmask8) __U);
1494}
1495
Michael Kupersteine45af542015-06-30 13:36:19 +00001496static __inline__ __m128 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001497_mm_mask_fmsub_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C)
1498{
1499 return (__m128) __builtin_ia32_vfmaddps128_mask ((__v4sf) __A,
1500 (__v4sf) __B,
1501 -(__v4sf) __C,
1502 (__mmask8) __U);
1503}
1504
Michael Kupersteine45af542015-06-30 13:36:19 +00001505static __inline__ __m128 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001506_mm_maskz_fmsub_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C)
1507{
1508 return (__m128) __builtin_ia32_vfmaddps128_maskz ((__v4sf) __A,
1509 (__v4sf) __B,
1510 -(__v4sf) __C,
1511 (__mmask8) __U);
1512}
1513
Michael Kupersteine45af542015-06-30 13:36:19 +00001514static __inline__ __m128 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001515_mm_mask3_fnmadd_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U)
1516{
1517 return (__m128) __builtin_ia32_vfmaddps128_mask3 (-(__v4sf) __A,
1518 (__v4sf) __B,
1519 (__v4sf) __C,
1520 (__mmask8) __U);
1521}
1522
Michael Kupersteine45af542015-06-30 13:36:19 +00001523static __inline__ __m128 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001524_mm_maskz_fnmadd_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C)
1525{
1526 return (__m128) __builtin_ia32_vfmaddps128_maskz (-(__v4sf) __A,
1527 (__v4sf) __B,
1528 (__v4sf) __C,
1529 (__mmask8) __U);
1530}
1531
Michael Kupersteine45af542015-06-30 13:36:19 +00001532static __inline__ __m128 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001533_mm_maskz_fnmsub_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C)
1534{
1535 return (__m128) __builtin_ia32_vfmaddps128_maskz (-(__v4sf) __A,
1536 (__v4sf) __B,
1537 -(__v4sf) __C,
1538 (__mmask8) __U);
1539}
1540
Michael Kupersteine45af542015-06-30 13:36:19 +00001541static __inline__ __m256 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001542_mm256_mask_fmadd_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C)
1543{
1544 return (__m256) __builtin_ia32_vfmaddps256_mask ((__v8sf) __A,
1545 (__v8sf) __B,
1546 (__v8sf) __C,
1547 (__mmask8) __U);
1548}
1549
Michael Kupersteine45af542015-06-30 13:36:19 +00001550static __inline__ __m256 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001551_mm256_mask3_fmadd_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U)
1552{
1553 return (__m256) __builtin_ia32_vfmaddps256_mask3 ((__v8sf) __A,
1554 (__v8sf) __B,
1555 (__v8sf) __C,
1556 (__mmask8) __U);
1557}
1558
Michael Kupersteine45af542015-06-30 13:36:19 +00001559static __inline__ __m256 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001560_mm256_maskz_fmadd_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C)
1561{
1562 return (__m256) __builtin_ia32_vfmaddps256_maskz ((__v8sf) __A,
1563 (__v8sf) __B,
1564 (__v8sf) __C,
1565 (__mmask8) __U);
1566}
1567
Michael Kupersteine45af542015-06-30 13:36:19 +00001568static __inline__ __m256 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001569_mm256_mask_fmsub_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C)
1570{
1571 return (__m256) __builtin_ia32_vfmaddps256_mask ((__v8sf) __A,
1572 (__v8sf) __B,
1573 -(__v8sf) __C,
1574 (__mmask8) __U);
1575}
1576
Michael Kupersteine45af542015-06-30 13:36:19 +00001577static __inline__ __m256 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001578_mm256_maskz_fmsub_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C)
1579{
1580 return (__m256) __builtin_ia32_vfmaddps256_maskz ((__v8sf) __A,
1581 (__v8sf) __B,
1582 -(__v8sf) __C,
1583 (__mmask8) __U);
1584}
1585
Michael Kupersteine45af542015-06-30 13:36:19 +00001586static __inline__ __m256 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001587_mm256_mask3_fnmadd_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U)
1588{
1589 return (__m256) __builtin_ia32_vfmaddps256_mask3 (-(__v8sf) __A,
1590 (__v8sf) __B,
1591 (__v8sf) __C,
1592 (__mmask8) __U);
1593}
1594
Michael Kupersteine45af542015-06-30 13:36:19 +00001595static __inline__ __m256 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001596_mm256_maskz_fnmadd_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C)
1597{
1598 return (__m256) __builtin_ia32_vfmaddps256_maskz (-(__v8sf) __A,
1599 (__v8sf) __B,
1600 (__v8sf) __C,
1601 (__mmask8) __U);
1602}
1603
Michael Kupersteine45af542015-06-30 13:36:19 +00001604static __inline__ __m256 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001605_mm256_maskz_fnmsub_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C)
1606{
1607 return (__m256) __builtin_ia32_vfmaddps256_maskz (-(__v8sf) __A,
1608 (__v8sf) __B,
1609 -(__v8sf) __C,
1610 (__mmask8) __U);
1611}
1612
Michael Kupersteine45af542015-06-30 13:36:19 +00001613static __inline__ __m128d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001614_mm_mask_fmaddsub_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C)
1615{
1616 return (__m128d) __builtin_ia32_vfmaddsubpd128_mask ((__v2df) __A,
1617 (__v2df) __B,
1618 (__v2df) __C,
1619 (__mmask8) __U);
1620}
1621
Michael Kupersteine45af542015-06-30 13:36:19 +00001622static __inline__ __m128d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001623_mm_mask3_fmaddsub_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U)
1624{
1625 return (__m128d) __builtin_ia32_vfmaddsubpd128_mask3 ((__v2df) __A,
1626 (__v2df) __B,
1627 (__v2df) __C,
1628 (__mmask8)
1629 __U);
1630}
1631
Michael Kupersteine45af542015-06-30 13:36:19 +00001632static __inline__ __m128d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001633_mm_maskz_fmaddsub_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C)
1634{
1635 return (__m128d) __builtin_ia32_vfmaddsubpd128_maskz ((__v2df) __A,
1636 (__v2df) __B,
1637 (__v2df) __C,
1638 (__mmask8)
1639 __U);
1640}
1641
Michael Kupersteine45af542015-06-30 13:36:19 +00001642static __inline__ __m128d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001643_mm_mask_fmsubadd_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C)
1644{
1645 return (__m128d) __builtin_ia32_vfmaddsubpd128_mask ((__v2df) __A,
1646 (__v2df) __B,
1647 -(__v2df) __C,
1648 (__mmask8) __U);
1649}
1650
Michael Kupersteine45af542015-06-30 13:36:19 +00001651static __inline__ __m128d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001652_mm_maskz_fmsubadd_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C)
1653{
1654 return (__m128d) __builtin_ia32_vfmaddsubpd128_maskz ((__v2df) __A,
1655 (__v2df) __B,
1656 -(__v2df) __C,
1657 (__mmask8)
1658 __U);
1659}
1660
Michael Kupersteine45af542015-06-30 13:36:19 +00001661static __inline__ __m256d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001662_mm256_mask_fmaddsub_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C)
1663{
1664 return (__m256d) __builtin_ia32_vfmaddsubpd256_mask ((__v4df) __A,
1665 (__v4df) __B,
1666 (__v4df) __C,
1667 (__mmask8) __U);
1668}
1669
Michael Kupersteine45af542015-06-30 13:36:19 +00001670static __inline__ __m256d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001671_mm256_mask3_fmaddsub_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U)
1672{
1673 return (__m256d) __builtin_ia32_vfmaddsubpd256_mask3 ((__v4df) __A,
1674 (__v4df) __B,
1675 (__v4df) __C,
1676 (__mmask8)
1677 __U);
1678}
1679
Michael Kupersteine45af542015-06-30 13:36:19 +00001680static __inline__ __m256d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001681_mm256_maskz_fmaddsub_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C)
1682{
1683 return (__m256d) __builtin_ia32_vfmaddsubpd256_maskz ((__v4df) __A,
1684 (__v4df) __B,
1685 (__v4df) __C,
1686 (__mmask8)
1687 __U);
1688}
1689
Michael Kupersteine45af542015-06-30 13:36:19 +00001690static __inline__ __m256d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001691_mm256_mask_fmsubadd_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C)
1692{
1693 return (__m256d) __builtin_ia32_vfmaddsubpd256_mask ((__v4df) __A,
1694 (__v4df) __B,
1695 -(__v4df) __C,
1696 (__mmask8) __U);
1697}
1698
Michael Kupersteine45af542015-06-30 13:36:19 +00001699static __inline__ __m256d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001700_mm256_maskz_fmsubadd_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C)
1701{
1702 return (__m256d) __builtin_ia32_vfmaddsubpd256_maskz ((__v4df) __A,
1703 (__v4df) __B,
1704 -(__v4df) __C,
1705 (__mmask8)
1706 __U);
1707}
1708
Michael Kupersteine45af542015-06-30 13:36:19 +00001709static __inline__ __m128 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001710_mm_mask_fmaddsub_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C)
1711{
1712 return (__m128) __builtin_ia32_vfmaddsubps128_mask ((__v4sf) __A,
1713 (__v4sf) __B,
1714 (__v4sf) __C,
1715 (__mmask8) __U);
1716}
1717
Michael Kupersteine45af542015-06-30 13:36:19 +00001718static __inline__ __m128 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001719_mm_mask3_fmaddsub_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U)
1720{
1721 return (__m128) __builtin_ia32_vfmaddsubps128_mask3 ((__v4sf) __A,
1722 (__v4sf) __B,
1723 (__v4sf) __C,
1724 (__mmask8) __U);
1725}
1726
Michael Kupersteine45af542015-06-30 13:36:19 +00001727static __inline__ __m128 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001728_mm_maskz_fmaddsub_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C)
1729{
1730 return (__m128) __builtin_ia32_vfmaddsubps128_maskz ((__v4sf) __A,
1731 (__v4sf) __B,
1732 (__v4sf) __C,
1733 (__mmask8) __U);
1734}
1735
Michael Kupersteine45af542015-06-30 13:36:19 +00001736static __inline__ __m128 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001737_mm_mask_fmsubadd_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C)
1738{
1739 return (__m128) __builtin_ia32_vfmaddsubps128_mask ((__v4sf) __A,
1740 (__v4sf) __B,
1741 -(__v4sf) __C,
1742 (__mmask8) __U);
1743}
1744
Michael Kupersteine45af542015-06-30 13:36:19 +00001745static __inline__ __m128 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001746_mm_maskz_fmsubadd_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C)
1747{
1748 return (__m128) __builtin_ia32_vfmaddsubps128_maskz ((__v4sf) __A,
1749 (__v4sf) __B,
1750 -(__v4sf) __C,
1751 (__mmask8) __U);
1752}
1753
Michael Kupersteine45af542015-06-30 13:36:19 +00001754static __inline__ __m256 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001755_mm256_mask_fmaddsub_ps(__m256 __A, __mmask8 __U, __m256 __B,
1756 __m256 __C)
1757{
1758 return (__m256) __builtin_ia32_vfmaddsubps256_mask ((__v8sf) __A,
1759 (__v8sf) __B,
1760 (__v8sf) __C,
1761 (__mmask8) __U);
1762}
1763
Michael Kupersteine45af542015-06-30 13:36:19 +00001764static __inline__ __m256 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001765_mm256_mask3_fmaddsub_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U)
1766{
1767 return (__m256) __builtin_ia32_vfmaddsubps256_mask3 ((__v8sf) __A,
1768 (__v8sf) __B,
1769 (__v8sf) __C,
1770 (__mmask8) __U);
1771}
1772
Michael Kupersteine45af542015-06-30 13:36:19 +00001773static __inline__ __m256 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001774_mm256_maskz_fmaddsub_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C)
1775{
1776 return (__m256) __builtin_ia32_vfmaddsubps256_maskz ((__v8sf) __A,
1777 (__v8sf) __B,
1778 (__v8sf) __C,
1779 (__mmask8) __U);
1780}
1781
Michael Kupersteine45af542015-06-30 13:36:19 +00001782static __inline__ __m256 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001783_mm256_mask_fmsubadd_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C)
1784{
1785 return (__m256) __builtin_ia32_vfmaddsubps256_mask ((__v8sf) __A,
1786 (__v8sf) __B,
1787 -(__v8sf) __C,
1788 (__mmask8) __U);
1789}
1790
Michael Kupersteine45af542015-06-30 13:36:19 +00001791static __inline__ __m256 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001792_mm256_maskz_fmsubadd_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C)
1793{
1794 return (__m256) __builtin_ia32_vfmaddsubps256_maskz ((__v8sf) __A,
1795 (__v8sf) __B,
1796 -(__v8sf) __C,
1797 (__mmask8) __U);
1798}
1799
Michael Kupersteine45af542015-06-30 13:36:19 +00001800static __inline__ __m128d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001801_mm_mask3_fmsub_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U)
1802{
1803 return (__m128d) __builtin_ia32_vfmsubpd128_mask3 ((__v2df) __A,
1804 (__v2df) __B,
1805 (__v2df) __C,
1806 (__mmask8) __U);
1807}
1808
Michael Kupersteine45af542015-06-30 13:36:19 +00001809static __inline__ __m256d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001810_mm256_mask3_fmsub_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U)
1811{
1812 return (__m256d) __builtin_ia32_vfmsubpd256_mask3 ((__v4df) __A,
1813 (__v4df) __B,
1814 (__v4df) __C,
1815 (__mmask8) __U);
1816}
1817
Michael Kupersteine45af542015-06-30 13:36:19 +00001818static __inline__ __m128 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001819_mm_mask3_fmsub_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U)
1820{
1821 return (__m128) __builtin_ia32_vfmsubps128_mask3 ((__v4sf) __A,
1822 (__v4sf) __B,
1823 (__v4sf) __C,
1824 (__mmask8) __U);
1825}
1826
Michael Kupersteine45af542015-06-30 13:36:19 +00001827static __inline__ __m256 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001828_mm256_mask3_fmsub_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U)
1829{
1830 return (__m256) __builtin_ia32_vfmsubps256_mask3 ((__v8sf) __A,
1831 (__v8sf) __B,
1832 (__v8sf) __C,
1833 (__mmask8) __U);
1834}
1835
Michael Kupersteine45af542015-06-30 13:36:19 +00001836static __inline__ __m128d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001837_mm_mask3_fmsubadd_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U)
1838{
1839 return (__m128d) __builtin_ia32_vfmsubaddpd128_mask3 ((__v2df) __A,
1840 (__v2df) __B,
1841 (__v2df) __C,
1842 (__mmask8)
1843 __U);
1844}
1845
Michael Kupersteine45af542015-06-30 13:36:19 +00001846static __inline__ __m256d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001847_mm256_mask3_fmsubadd_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U)
1848{
1849 return (__m256d) __builtin_ia32_vfmsubaddpd256_mask3 ((__v4df) __A,
1850 (__v4df) __B,
1851 (__v4df) __C,
1852 (__mmask8)
1853 __U);
1854}
1855
Michael Kupersteine45af542015-06-30 13:36:19 +00001856static __inline__ __m128 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001857_mm_mask3_fmsubadd_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U)
1858{
1859 return (__m128) __builtin_ia32_vfmsubaddps128_mask3 ((__v4sf) __A,
1860 (__v4sf) __B,
1861 (__v4sf) __C,
1862 (__mmask8) __U);
1863}
1864
Michael Kupersteine45af542015-06-30 13:36:19 +00001865static __inline__ __m256 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001866_mm256_mask3_fmsubadd_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U)
1867{
1868 return (__m256) __builtin_ia32_vfmsubaddps256_mask3 ((__v8sf) __A,
1869 (__v8sf) __B,
1870 (__v8sf) __C,
1871 (__mmask8) __U);
1872}
1873
Michael Kupersteine45af542015-06-30 13:36:19 +00001874static __inline__ __m128d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001875_mm_mask_fnmadd_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C)
1876{
1877 return (__m128d) __builtin_ia32_vfnmaddpd128_mask ((__v2df) __A,
1878 (__v2df) __B,
1879 (__v2df) __C,
1880 (__mmask8) __U);
1881}
1882
Michael Kupersteine45af542015-06-30 13:36:19 +00001883static __inline__ __m256d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001884_mm256_mask_fnmadd_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C)
1885{
1886 return (__m256d) __builtin_ia32_vfnmaddpd256_mask ((__v4df) __A,
1887 (__v4df) __B,
1888 (__v4df) __C,
1889 (__mmask8) __U);
1890}
1891
Michael Kupersteine45af542015-06-30 13:36:19 +00001892static __inline__ __m128 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001893_mm_mask_fnmadd_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C)
1894{
1895 return (__m128) __builtin_ia32_vfnmaddps128_mask ((__v4sf) __A,
1896 (__v4sf) __B,
1897 (__v4sf) __C,
1898 (__mmask8) __U);
1899}
1900
Michael Kupersteine45af542015-06-30 13:36:19 +00001901static __inline__ __m256 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001902_mm256_mask_fnmadd_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C)
1903{
1904 return (__m256) __builtin_ia32_vfnmaddps256_mask ((__v8sf) __A,
1905 (__v8sf) __B,
1906 (__v8sf) __C,
1907 (__mmask8) __U);
1908}
1909
Michael Kupersteine45af542015-06-30 13:36:19 +00001910static __inline__ __m128d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001911_mm_mask_fnmsub_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C)
1912{
1913 return (__m128d) __builtin_ia32_vfnmsubpd128_mask ((__v2df) __A,
1914 (__v2df) __B,
1915 (__v2df) __C,
1916 (__mmask8) __U);
1917}
1918
Michael Kupersteine45af542015-06-30 13:36:19 +00001919static __inline__ __m128d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001920_mm_mask3_fnmsub_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U)
1921{
1922 return (__m128d) __builtin_ia32_vfnmsubpd128_mask3 ((__v2df) __A,
1923 (__v2df) __B,
1924 (__v2df) __C,
1925 (__mmask8) __U);
1926}
1927
Michael Kupersteine45af542015-06-30 13:36:19 +00001928static __inline__ __m256d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001929_mm256_mask_fnmsub_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C)
1930{
1931 return (__m256d) __builtin_ia32_vfnmsubpd256_mask ((__v4df) __A,
1932 (__v4df) __B,
1933 (__v4df) __C,
1934 (__mmask8) __U);
1935}
1936
Michael Kupersteine45af542015-06-30 13:36:19 +00001937static __inline__ __m256d __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001938_mm256_mask3_fnmsub_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U)
1939{
1940 return (__m256d) __builtin_ia32_vfnmsubpd256_mask3 ((__v4df) __A,
1941 (__v4df) __B,
1942 (__v4df) __C,
1943 (__mmask8) __U);
1944}
1945
Michael Kupersteine45af542015-06-30 13:36:19 +00001946static __inline__ __m128 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001947_mm_mask_fnmsub_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C)
1948{
1949 return (__m128) __builtin_ia32_vfnmsubps128_mask ((__v4sf) __A,
1950 (__v4sf) __B,
1951 (__v4sf) __C,
1952 (__mmask8) __U);
1953}
1954
Michael Kupersteine45af542015-06-30 13:36:19 +00001955static __inline__ __m128 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001956_mm_mask3_fnmsub_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U)
1957{
1958 return (__m128) __builtin_ia32_vfnmsubps128_mask3 ((__v4sf) __A,
1959 (__v4sf) __B,
1960 (__v4sf) __C,
1961 (__mmask8) __U);
1962}
1963
Michael Kupersteine45af542015-06-30 13:36:19 +00001964static __inline__ __m256 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001965_mm256_mask_fnmsub_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C)
1966{
1967 return (__m256) __builtin_ia32_vfnmsubps256_mask ((__v8sf) __A,
1968 (__v8sf) __B,
1969 (__v8sf) __C,
1970 (__mmask8) __U);
1971}
1972
Michael Kupersteine45af542015-06-30 13:36:19 +00001973static __inline__ __m256 __DEFAULT_FN_ATTRS
Elena Demikhovskyc563c2c2015-06-29 09:20:57 +00001974_mm256_mask3_fnmsub_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U)
1975{
1976 return (__m256) __builtin_ia32_vfnmsubps256_mask3 ((__v8sf) __A,
1977 (__v8sf) __B,
1978 (__v8sf) __C,
1979 (__mmask8) __U);
1980}
1981
Asaf Badouh74da3872015-07-28 08:26:14 +00001982static __inline__ __m128d __DEFAULT_FN_ATTRS
1983_mm_mask_add_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
1984 return (__m128d) __builtin_ia32_addpd128_mask ((__v2df) __A,
1985 (__v2df) __B,
1986 (__v2df) __W,
1987 (__mmask8) __U);
1988}
1989
1990static __inline__ __m128d __DEFAULT_FN_ATTRS
1991_mm_maskz_add_pd (__mmask8 __U, __m128d __A, __m128d __B) {
1992 return (__m128d) __builtin_ia32_addpd128_mask ((__v2df) __A,
1993 (__v2df) __B,
1994 (__v2df)
1995 _mm_setzero_pd (),
1996 (__mmask8) __U);
1997}
1998
1999static __inline__ __m256d __DEFAULT_FN_ATTRS
2000_mm256_mask_add_pd (__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) {
2001 return (__m256d) __builtin_ia32_addpd256_mask ((__v4df) __A,
2002 (__v4df) __B,
2003 (__v4df) __W,
2004 (__mmask8) __U);
2005}
2006
2007static __inline__ __m256d __DEFAULT_FN_ATTRS
2008_mm256_maskz_add_pd (__mmask8 __U, __m256d __A, __m256d __B) {
2009 return (__m256d) __builtin_ia32_addpd256_mask ((__v4df) __A,
2010 (__v4df) __B,
2011 (__v4df)
2012 _mm256_setzero_pd (),
2013 (__mmask8) __U);
2014}
2015
2016static __inline__ __m128 __DEFAULT_FN_ATTRS
2017_mm_mask_add_ps (__m128 __W, __mmask16 __U, __m128 __A, __m128 __B) {
2018 return (__m128) __builtin_ia32_addps128_mask ((__v4sf) __A,
2019 (__v4sf) __B,
2020 (__v4sf) __W,
2021 (__mmask8) __U);
2022}
2023
2024static __inline__ __m128 __DEFAULT_FN_ATTRS
2025_mm_maskz_add_ps (__mmask16 __U, __m128 __A, __m128 __B) {
2026 return (__m128) __builtin_ia32_addps128_mask ((__v4sf) __A,
2027 (__v4sf) __B,
2028 (__v4sf)
2029 _mm_setzero_ps (),
2030 (__mmask8) __U);
2031}
2032
2033static __inline__ __m256 __DEFAULT_FN_ATTRS
2034_mm256_mask_add_ps (__m256 __W, __mmask16 __U, __m256 __A, __m256 __B) {
2035 return (__m256) __builtin_ia32_addps256_mask ((__v8sf) __A,
2036 (__v8sf) __B,
2037 (__v8sf) __W,
2038 (__mmask8) __U);
2039}
2040
2041static __inline__ __m256 __DEFAULT_FN_ATTRS
2042_mm256_maskz_add_ps (__mmask16 __U, __m256 __A, __m256 __B) {
2043 return (__m256) __builtin_ia32_addps256_mask ((__v8sf) __A,
2044 (__v8sf) __B,
2045 (__v8sf)
2046 _mm256_setzero_ps (),
2047 (__mmask8) __U);
2048}
2049
2050static __inline__ __m128i __DEFAULT_FN_ATTRS
2051_mm_mask_blend_epi32 (__mmask8 __U, __m128i __A, __m128i __W) {
2052 return (__m128i) __builtin_ia32_blendmd_128_mask ((__v4si) __A,
2053 (__v4si) __W,
2054 (__mmask8) __U);
2055}
2056
2057static __inline__ __m256i __DEFAULT_FN_ATTRS
2058_mm256_mask_blend_epi32 (__mmask8 __U, __m256i __A, __m256i __W) {
2059 return (__m256i) __builtin_ia32_blendmd_256_mask ((__v8si) __A,
2060 (__v8si) __W,
2061 (__mmask8) __U);
2062}
2063
2064static __inline__ __m128d __DEFAULT_FN_ATTRS
2065_mm_mask_blend_pd (__mmask8 __U, __m128d __A, __m128d __W) {
2066 return (__m128d) __builtin_ia32_blendmpd_128_mask ((__v2df) __A,
2067 (__v2df) __W,
2068 (__mmask8) __U);
2069}
2070
2071static __inline__ __m256d __DEFAULT_FN_ATTRS
2072_mm256_mask_blend_pd (__mmask8 __U, __m256d __A, __m256d __W) {
2073 return (__m256d) __builtin_ia32_blendmpd_256_mask ((__v4df) __A,
2074 (__v4df) __W,
2075 (__mmask8) __U);
2076}
2077
2078static __inline__ __m128 __DEFAULT_FN_ATTRS
2079_mm_mask_blend_ps (__mmask8 __U, __m128 __A, __m128 __W) {
2080 return (__m128) __builtin_ia32_blendmps_128_mask ((__v4sf) __A,
2081 (__v4sf) __W,
2082 (__mmask8) __U);
2083}
2084
2085static __inline__ __m256 __DEFAULT_FN_ATTRS
2086_mm256_mask_blend_ps (__mmask8 __U, __m256 __A, __m256 __W) {
2087 return (__m256) __builtin_ia32_blendmps_256_mask ((__v8sf) __A,
2088 (__v8sf) __W,
2089 (__mmask8) __U);
2090}
2091
2092static __inline__ __m128i __DEFAULT_FN_ATTRS
2093_mm_mask_blend_epi64 (__mmask8 __U, __m128i __A, __m128i __W) {
2094 return (__m128i) __builtin_ia32_blendmq_128_mask ((__v2di) __A,
2095 (__v2di) __W,
2096 (__mmask8) __U);
2097}
2098
2099static __inline__ __m256i __DEFAULT_FN_ATTRS
2100_mm256_mask_blend_epi64 (__mmask8 __U, __m256i __A, __m256i __W) {
2101 return (__m256i) __builtin_ia32_blendmq_256_mask ((__v4di) __A,
2102 (__v4di) __W,
2103 (__mmask8) __U);
2104}
2105
2106static __inline__ __m128d __DEFAULT_FN_ATTRS
2107_mm_mask_compress_pd (__m128d __W, __mmask8 __U, __m128d __A) {
2108 return (__m128d) __builtin_ia32_compressdf128_mask ((__v2df) __A,
2109 (__v2df) __W,
2110 (__mmask8) __U);
2111}
2112
2113static __inline__ __m128d __DEFAULT_FN_ATTRS
2114_mm_maskz_compress_pd (__mmask8 __U, __m128d __A) {
2115 return (__m128d) __builtin_ia32_compressdf128_mask ((__v2df) __A,
2116 (__v2df)
2117 _mm_setzero_pd (),
2118 (__mmask8) __U);
2119}
2120
2121static __inline__ __m256d __DEFAULT_FN_ATTRS
2122_mm256_mask_compress_pd (__m256d __W, __mmask8 __U, __m256d __A) {
2123 return (__m256d) __builtin_ia32_compressdf256_mask ((__v4df) __A,
2124 (__v4df) __W,
2125 (__mmask8) __U);
2126}
2127
2128static __inline__ __m256d __DEFAULT_FN_ATTRS
2129_mm256_maskz_compress_pd (__mmask8 __U, __m256d __A) {
2130 return (__m256d) __builtin_ia32_compressdf256_mask ((__v4df) __A,
2131 (__v4df)
2132 _mm256_setzero_pd (),
2133 (__mmask8) __U);
2134}
2135
2136static __inline__ __m128i __DEFAULT_FN_ATTRS
2137_mm_mask_compress_epi64 (__m128i __W, __mmask8 __U, __m128i __A) {
2138 return (__m128i) __builtin_ia32_compressdi128_mask ((__v2di) __A,
2139 (__v2di) __W,
2140 (__mmask8) __U);
2141}
2142
2143static __inline__ __m128i __DEFAULT_FN_ATTRS
2144_mm_maskz_compress_epi64 (__mmask8 __U, __m128i __A) {
2145 return (__m128i) __builtin_ia32_compressdi128_mask ((__v2di) __A,
2146 (__v2di)
2147 _mm_setzero_si128 (),
2148 (__mmask8) __U);
2149}
2150
2151static __inline__ __m256i __DEFAULT_FN_ATTRS
2152_mm256_mask_compress_epi64 (__m256i __W, __mmask8 __U, __m256i __A) {
2153 return (__m256i) __builtin_ia32_compressdi256_mask ((__v4di) __A,
2154 (__v4di) __W,
2155 (__mmask8) __U);
2156}
2157
2158static __inline__ __m256i __DEFAULT_FN_ATTRS
2159_mm256_maskz_compress_epi64 (__mmask8 __U, __m256i __A) {
2160 return (__m256i) __builtin_ia32_compressdi256_mask ((__v4di) __A,
2161 (__v4di)
2162 _mm256_setzero_si256 (),
2163 (__mmask8) __U);
2164}
2165
2166static __inline__ __m128 __DEFAULT_FN_ATTRS
2167_mm_mask_compress_ps (__m128 __W, __mmask8 __U, __m128 __A) {
2168 return (__m128) __builtin_ia32_compresssf128_mask ((__v4sf) __A,
2169 (__v4sf) __W,
2170 (__mmask8) __U);
2171}
2172
2173static __inline__ __m128 __DEFAULT_FN_ATTRS
2174_mm_maskz_compress_ps (__mmask8 __U, __m128 __A) {
2175 return (__m128) __builtin_ia32_compresssf128_mask ((__v4sf) __A,
2176 (__v4sf)
2177 _mm_setzero_ps (),
2178 (__mmask8) __U);
2179}
2180
2181static __inline__ __m256 __DEFAULT_FN_ATTRS
2182_mm256_mask_compress_ps (__m256 __W, __mmask8 __U, __m256 __A) {
2183 return (__m256) __builtin_ia32_compresssf256_mask ((__v8sf) __A,
2184 (__v8sf) __W,
2185 (__mmask8) __U);
2186}
2187
2188static __inline__ __m256 __DEFAULT_FN_ATTRS
2189_mm256_maskz_compress_ps (__mmask8 __U, __m256 __A) {
2190 return (__m256) __builtin_ia32_compresssf256_mask ((__v8sf) __A,
2191 (__v8sf)
2192 _mm256_setzero_ps (),
2193 (__mmask8) __U);
2194}
2195
2196static __inline__ __m128i __DEFAULT_FN_ATTRS
2197_mm_mask_compress_epi32 (__m128i __W, __mmask8 __U, __m128i __A) {
2198 return (__m128i) __builtin_ia32_compresssi128_mask ((__v4si) __A,
2199 (__v4si) __W,
2200 (__mmask8) __U);
2201}
2202
2203static __inline__ __m128i __DEFAULT_FN_ATTRS
2204_mm_maskz_compress_epi32 (__mmask8 __U, __m128i __A) {
2205 return (__m128i) __builtin_ia32_compresssi128_mask ((__v4si) __A,
2206 (__v4si)
2207 _mm_setzero_si128 (),
2208 (__mmask8) __U);
2209}
2210
2211static __inline__ __m256i __DEFAULT_FN_ATTRS
2212_mm256_mask_compress_epi32 (__m256i __W, __mmask8 __U, __m256i __A) {
2213 return (__m256i) __builtin_ia32_compresssi256_mask ((__v8si) __A,
2214 (__v8si) __W,
2215 (__mmask8) __U);
2216}
2217
2218static __inline__ __m256i __DEFAULT_FN_ATTRS
2219_mm256_maskz_compress_epi32 (__mmask8 __U, __m256i __A) {
2220 return (__m256i) __builtin_ia32_compresssi256_mask ((__v8si) __A,
2221 (__v8si)
2222 _mm256_setzero_si256 (),
2223 (__mmask8) __U);
2224}
2225
2226static __inline__ void __DEFAULT_FN_ATTRS
2227_mm_mask_compressstoreu_pd (void *__P, __mmask8 __U, __m128d __A) {
2228 __builtin_ia32_compressstoredf128_mask ((__v2df *) __P,
2229 (__v2df) __A,
2230 (__mmask8) __U);
2231}
2232
2233static __inline__ void __DEFAULT_FN_ATTRS
2234_mm256_mask_compressstoreu_pd (void *__P, __mmask8 __U, __m256d __A) {
2235 __builtin_ia32_compressstoredf256_mask ((__v4df *) __P,
2236 (__v4df) __A,
2237 (__mmask8) __U);
2238}
2239
2240static __inline__ void __DEFAULT_FN_ATTRS
2241_mm_mask_compressstoreu_epi64 (void *__P, __mmask8 __U, __m128i __A) {
2242 __builtin_ia32_compressstoredi128_mask ((__v2di *) __P,
2243 (__v2di) __A,
2244 (__mmask8) __U);
2245}
2246
2247static __inline__ void __DEFAULT_FN_ATTRS
2248_mm256_mask_compressstoreu_epi64 (void *__P, __mmask8 __U, __m256i __A) {
2249 __builtin_ia32_compressstoredi256_mask ((__v4di *) __P,
2250 (__v4di) __A,
2251 (__mmask8) __U);
2252}
2253
2254static __inline__ void __DEFAULT_FN_ATTRS
2255_mm_mask_compressstoreu_ps (void *__P, __mmask8 __U, __m128 __A) {
2256 __builtin_ia32_compressstoresf128_mask ((__v4sf *) __P,
2257 (__v4sf) __A,
2258 (__mmask8) __U);
2259}
2260
2261static __inline__ void __DEFAULT_FN_ATTRS
2262_mm256_mask_compressstoreu_ps (void *__P, __mmask8 __U, __m256 __A) {
2263 __builtin_ia32_compressstoresf256_mask ((__v8sf *) __P,
2264 (__v8sf) __A,
2265 (__mmask8) __U);
2266}
2267
2268static __inline__ void __DEFAULT_FN_ATTRS
2269_mm_mask_compressstoreu_epi32 (void *__P, __mmask8 __U, __m128i __A) {
2270 __builtin_ia32_compressstoresi128_mask ((__v4si *) __P,
2271 (__v4si) __A,
2272 (__mmask8) __U);
2273}
2274
2275static __inline__ void __DEFAULT_FN_ATTRS
2276_mm256_mask_compressstoreu_epi32 (void *__P, __mmask8 __U, __m256i __A) {
2277 __builtin_ia32_compressstoresi256_mask ((__v8si *) __P,
2278 (__v8si) __A,
2279 (__mmask8) __U);
2280}
2281
2282static __inline__ __m128d __DEFAULT_FN_ATTRS
2283_mm_mask_cvtepi32_pd (__m128d __W, __mmask8 __U, __m128i __A) {
2284 return (__m128d) __builtin_ia32_cvtdq2pd128_mask ((__v4si) __A,
2285 (__v2df) __W,
2286 (__mmask8) __U);
2287}
2288
2289static __inline__ __m128d __DEFAULT_FN_ATTRS
2290_mm_maskz_cvtepi32_pd (__mmask8 __U, __m128i __A) {
2291 return (__m128d) __builtin_ia32_cvtdq2pd128_mask ((__v4si) __A,
2292 (__v2df)
2293 _mm_setzero_pd (),
2294 (__mmask8) __U);
2295}
2296
2297static __inline__ __m256d __DEFAULT_FN_ATTRS
2298_mm256_mask_cvtepi32_pd (__m256d __W, __mmask8 __U, __m128i __A) {
2299 return (__m256d) __builtin_ia32_cvtdq2pd256_mask ((__v4si) __A,
2300 (__v4df) __W,
2301 (__mmask8) __U);
2302}
2303
2304static __inline__ __m256d __DEFAULT_FN_ATTRS
2305_mm256_maskz_cvtepi32_pd (__mmask8 __U, __m128i __A) {
2306 return (__m256d) __builtin_ia32_cvtdq2pd256_mask ((__v4si) __A,
2307 (__v4df)
2308 _mm256_setzero_pd (),
2309 (__mmask8) __U);
2310}
2311
2312static __inline__ __m128 __DEFAULT_FN_ATTRS
2313_mm_mask_cvtepi32_ps (__m128 __W, __mmask8 __U, __m128i __A) {
2314 return (__m128) __builtin_ia32_cvtdq2ps128_mask ((__v4si) __A,
2315 (__v4sf) __W,
2316 (__mmask8) __U);
2317}
2318
2319static __inline__ __m128 __DEFAULT_FN_ATTRS
2320_mm_maskz_cvtepi32_ps (__mmask16 __U, __m128i __A) {
2321 return (__m128) __builtin_ia32_cvtdq2ps128_mask ((__v4si) __A,
2322 (__v4sf)
2323 _mm_setzero_ps (),
2324 (__mmask8) __U);
2325}
2326
2327static __inline__ __m256 __DEFAULT_FN_ATTRS
2328_mm256_mask_cvtepi32_ps (__m256 __W, __mmask8 __U, __m256i __A) {
2329 return (__m256) __builtin_ia32_cvtdq2ps256_mask ((__v8si) __A,
2330 (__v8sf) __W,
2331 (__mmask8) __U);
2332}
2333
2334static __inline__ __m256 __DEFAULT_FN_ATTRS
2335_mm256_maskz_cvtepi32_ps (__mmask16 __U, __m256i __A) {
2336 return (__m256) __builtin_ia32_cvtdq2ps256_mask ((__v8si) __A,
2337 (__v8sf)
2338 _mm256_setzero_ps (),
2339 (__mmask8) __U);
2340}
2341
2342static __inline__ __m128i __DEFAULT_FN_ATTRS
2343_mm_mask_cvtpd_epi32 (__m128i __W, __mmask8 __U, __m128d __A) {
2344 return (__m128i) __builtin_ia32_cvtpd2dq128_mask ((__v2df) __A,
2345 (__v4si) __W,
2346 (__mmask8) __U);
2347}
2348
2349static __inline__ __m128i __DEFAULT_FN_ATTRS
2350_mm_maskz_cvtpd_epi32 (__mmask8 __U, __m128d __A) {
2351 return (__m128i) __builtin_ia32_cvtpd2dq128_mask ((__v2df) __A,
2352 (__v4si)
2353 _mm_setzero_si128 (),
2354 (__mmask8) __U);
2355}
2356
2357static __inline__ __m128i __DEFAULT_FN_ATTRS
2358_mm256_mask_cvtpd_epi32 (__m128i __W, __mmask8 __U, __m256d __A) {
2359 return (__m128i) __builtin_ia32_cvtpd2dq256_mask ((__v4df) __A,
2360 (__v4si) __W,
2361 (__mmask8) __U);
2362}
2363
2364static __inline__ __m128i __DEFAULT_FN_ATTRS
2365_mm256_maskz_cvtpd_epi32 (__mmask8 __U, __m256d __A) {
2366 return (__m128i) __builtin_ia32_cvtpd2dq256_mask ((__v4df) __A,
2367 (__v4si)
2368 _mm_setzero_si128 (),
2369 (__mmask8) __U);
2370}
2371
2372static __inline__ __m128 __DEFAULT_FN_ATTRS
2373_mm_mask_cvtpd_ps (__m128 __W, __mmask8 __U, __m128d __A) {
2374 return (__m128) __builtin_ia32_cvtpd2ps_mask ((__v2df) __A,
2375 (__v4sf) __W,
2376 (__mmask8) __U);
2377}
2378
2379static __inline__ __m128 __DEFAULT_FN_ATTRS
2380_mm_maskz_cvtpd_ps (__mmask8 __U, __m128d __A) {
2381 return (__m128) __builtin_ia32_cvtpd2ps_mask ((__v2df) __A,
2382 (__v4sf)
2383 _mm_setzero_ps (),
2384 (__mmask8) __U);
2385}
2386
2387static __inline__ __m128 __DEFAULT_FN_ATTRS
2388_mm256_mask_cvtpd_ps (__m128 __W, __mmask8 __U, __m256d __A) {
2389 return (__m128) __builtin_ia32_cvtpd2ps256_mask ((__v4df) __A,
2390 (__v4sf) __W,
2391 (__mmask8) __U);
2392}
2393
2394static __inline__ __m128 __DEFAULT_FN_ATTRS
2395_mm256_maskz_cvtpd_ps (__mmask8 __U, __m256d __A) {
2396 return (__m128) __builtin_ia32_cvtpd2ps256_mask ((__v4df) __A,
2397 (__v4sf)
2398 _mm_setzero_ps (),
2399 (__mmask8) __U);
2400}
2401
2402static __inline__ __m128i __DEFAULT_FN_ATTRS
2403_mm_cvtpd_epu32 (__m128d __A) {
2404 return (__m128i) __builtin_ia32_cvtpd2udq128_mask ((__v2df) __A,
2405 (__v4si)
2406 _mm_setzero_si128 (),
2407 (__mmask8) -1);
2408}
2409
2410static __inline__ __m128i __DEFAULT_FN_ATTRS
2411_mm_mask_cvtpd_epu32 (__m128i __W, __mmask8 __U, __m128d __A) {
2412 return (__m128i) __builtin_ia32_cvtpd2udq128_mask ((__v2df) __A,
2413 (__v4si) __W,
2414 (__mmask8) __U);
2415}
2416
2417static __inline__ __m128i __DEFAULT_FN_ATTRS
2418_mm_maskz_cvtpd_epu32 (__mmask8 __U, __m128d __A) {
2419 return (__m128i) __builtin_ia32_cvtpd2udq128_mask ((__v2df) __A,
2420 (__v4si)
2421 _mm_setzero_si128 (),
2422 (__mmask8) __U);
2423}
2424
2425static __inline__ __m128i __DEFAULT_FN_ATTRS
2426_mm256_cvtpd_epu32 (__m256d __A) {
2427 return (__m128i) __builtin_ia32_cvtpd2udq256_mask ((__v4df) __A,
2428 (__v4si)
2429 _mm_setzero_si128 (),
2430 (__mmask8) -1);
2431}
2432
2433static __inline__ __m128i __DEFAULT_FN_ATTRS
2434_mm256_mask_cvtpd_epu32 (__m128i __W, __mmask8 __U, __m256d __A) {
2435 return (__m128i) __builtin_ia32_cvtpd2udq256_mask ((__v4df) __A,
2436 (__v4si) __W,
2437 (__mmask8) __U);
2438}
2439
2440static __inline__ __m128i __DEFAULT_FN_ATTRS
2441_mm256_maskz_cvtpd_epu32 (__mmask8 __U, __m256d __A) {
2442 return (__m128i) __builtin_ia32_cvtpd2udq256_mask ((__v4df) __A,
2443 (__v4si)
2444 _mm_setzero_si128 (),
2445 (__mmask8) __U);
2446}
2447
2448static __inline__ __m128i __DEFAULT_FN_ATTRS
2449_mm_mask_cvtps_epi32 (__m128i __W, __mmask8 __U, __m128 __A) {
2450 return (__m128i) __builtin_ia32_cvtps2dq128_mask ((__v4sf) __A,
2451 (__v4si) __W,
2452 (__mmask8) __U);
2453}
2454
2455static __inline__ __m128i __DEFAULT_FN_ATTRS
2456_mm_maskz_cvtps_epi32 (__mmask8 __U, __m128 __A) {
2457 return (__m128i) __builtin_ia32_cvtps2dq128_mask ((__v4sf) __A,
2458 (__v4si)
2459 _mm_setzero_si128 (),
2460 (__mmask8) __U);
2461}
2462
2463static __inline__ __m256i __DEFAULT_FN_ATTRS
2464_mm256_mask_cvtps_epi32 (__m256i __W, __mmask8 __U, __m256 __A) {
2465 return (__m256i) __builtin_ia32_cvtps2dq256_mask ((__v8sf) __A,
2466 (__v8si) __W,
2467 (__mmask8) __U);
2468}
2469
2470static __inline__ __m256i __DEFAULT_FN_ATTRS
2471_mm256_maskz_cvtps_epi32 (__mmask8 __U, __m256 __A) {
2472 return (__m256i) __builtin_ia32_cvtps2dq256_mask ((__v8sf) __A,
2473 (__v8si)
2474 _mm256_setzero_si256 (),
2475 (__mmask8) __U);
2476}
2477
2478static __inline__ __m128d __DEFAULT_FN_ATTRS
2479_mm_mask_cvtps_pd (__m128d __W, __mmask8 __U, __m128 __A) {
2480 return (__m128d) __builtin_ia32_cvtps2pd128_mask ((__v4sf) __A,
2481 (__v2df) __W,
2482 (__mmask8) __U);
2483}
2484
2485static __inline__ __m128d __DEFAULT_FN_ATTRS
2486_mm_maskz_cvtps_pd (__mmask8 __U, __m128 __A) {
2487 return (__m128d) __builtin_ia32_cvtps2pd128_mask ((__v4sf) __A,
2488 (__v2df)
2489 _mm_setzero_pd (),
2490 (__mmask8) __U);
2491}
2492
2493static __inline__ __m256d __DEFAULT_FN_ATTRS
2494_mm256_mask_cvtps_pd (__m256d __W, __mmask8 __U, __m128 __A) {
2495 return (__m256d) __builtin_ia32_cvtps2pd256_mask ((__v4sf) __A,
2496 (__v4df) __W,
2497 (__mmask8) __U);
2498}
2499
2500static __inline__ __m256d __DEFAULT_FN_ATTRS
2501_mm256_maskz_cvtps_pd (__mmask8 __U, __m128 __A) {
2502 return (__m256d) __builtin_ia32_cvtps2pd256_mask ((__v4sf) __A,
2503 (__v4df)
2504 _mm256_setzero_pd (),
2505 (__mmask8) __U);
2506}
2507
2508static __inline__ __m128i __DEFAULT_FN_ATTRS
2509_mm_cvtps_epu32 (__m128 __A) {
2510 return (__m128i) __builtin_ia32_cvtps2udq128_mask ((__v4sf) __A,
2511 (__v4si)
2512 _mm_setzero_si128 (),
2513 (__mmask8) -1);
2514}
2515
2516static __inline__ __m128i __DEFAULT_FN_ATTRS
2517_mm_mask_cvtps_epu32 (__m128i __W, __mmask8 __U, __m128 __A) {
2518 return (__m128i) __builtin_ia32_cvtps2udq128_mask ((__v4sf) __A,
2519 (__v4si) __W,
2520 (__mmask8) __U);
2521}
2522
2523static __inline__ __m128i __DEFAULT_FN_ATTRS
2524_mm_maskz_cvtps_epu32 (__mmask8 __U, __m128 __A) {
2525 return (__m128i) __builtin_ia32_cvtps2udq128_mask ((__v4sf) __A,
2526 (__v4si)
2527 _mm_setzero_si128 (),
2528 (__mmask8) __U);
2529}
2530
2531static __inline__ __m256i __DEFAULT_FN_ATTRS
2532_mm256_cvtps_epu32 (__m256 __A) {
2533 return (__m256i) __builtin_ia32_cvtps2udq256_mask ((__v8sf) __A,
2534 (__v8si)
2535 _mm256_setzero_si256 (),
2536 (__mmask8) -1);
2537}
2538
2539static __inline__ __m256i __DEFAULT_FN_ATTRS
2540_mm256_mask_cvtps_epu32 (__m256i __W, __mmask8 __U, __m256 __A) {
2541 return (__m256i) __builtin_ia32_cvtps2udq256_mask ((__v8sf) __A,
2542 (__v8si) __W,
2543 (__mmask8) __U);
2544}
2545
2546static __inline__ __m256i __DEFAULT_FN_ATTRS
2547_mm256_maskz_cvtps_epu32 (__mmask8 __U, __m256 __A) {
2548 return (__m256i) __builtin_ia32_cvtps2udq256_mask ((__v8sf) __A,
2549 (__v8si)
2550 _mm256_setzero_si256 (),
2551 (__mmask8) __U);
2552}
2553
2554static __inline__ __m128i __DEFAULT_FN_ATTRS
2555_mm_mask_cvttpd_epi32 (__m128i __W, __mmask8 __U, __m128d __A) {
2556 return (__m128i) __builtin_ia32_cvttpd2dq128_mask ((__v2df) __A,
2557 (__v4si) __W,
2558 (__mmask8) __U);
2559}
2560
2561static __inline__ __m128i __DEFAULT_FN_ATTRS
2562_mm_maskz_cvttpd_epi32 (__mmask8 __U, __m128d __A) {
2563 return (__m128i) __builtin_ia32_cvttpd2dq128_mask ((__v2df) __A,
2564 (__v4si)
2565 _mm_setzero_si128 (),
2566 (__mmask8) __U);
2567}
2568
2569static __inline__ __m128i __DEFAULT_FN_ATTRS
2570_mm256_mask_cvttpd_epi32 (__m128i __W, __mmask8 __U, __m256d __A) {
2571 return (__m128i) __builtin_ia32_cvttpd2dq256_mask ((__v4df) __A,
2572 (__v4si) __W,
2573 (__mmask8) __U);
2574}
2575
2576static __inline__ __m128i __DEFAULT_FN_ATTRS
2577_mm256_maskz_cvttpd_epi32 (__mmask8 __U, __m256d __A) {
2578 return (__m128i) __builtin_ia32_cvttpd2dq256_mask ((__v4df) __A,
2579 (__v4si)
2580 _mm_setzero_si128 (),
2581 (__mmask8) __U);
2582}
2583
2584static __inline__ __m128i __DEFAULT_FN_ATTRS
2585_mm_cvttpd_epu32 (__m128d __A) {
2586 return (__m128i) __builtin_ia32_cvttpd2udq128_mask ((__v2df) __A,
2587 (__v4si)
2588 _mm_setzero_si128 (),
2589 (__mmask8) -1);
2590}
2591
2592static __inline__ __m128i __DEFAULT_FN_ATTRS
2593_mm_mask_cvttpd_epu32 (__m128i __W, __mmask8 __U, __m128d __A) {
2594 return (__m128i) __builtin_ia32_cvttpd2udq128_mask ((__v2df) __A,
2595 (__v4si) __W,
2596 (__mmask8) __U);
2597}
2598
2599static __inline__ __m128i __DEFAULT_FN_ATTRS
2600_mm_maskz_cvttpd_epu32 (__mmask8 __U, __m128d __A) {
2601 return (__m128i) __builtin_ia32_cvttpd2udq128_mask ((__v2df) __A,
2602 (__v4si)
2603 _mm_setzero_si128 (),
2604 (__mmask8) __U);
2605}
2606
2607static __inline__ __m128i __DEFAULT_FN_ATTRS
2608_mm256_cvttpd_epu32 (__m256d __A) {
2609 return (__m128i) __builtin_ia32_cvttpd2udq256_mask ((__v4df) __A,
2610 (__v4si)
2611 _mm_setzero_si128 (),
2612 (__mmask8) -1);
2613}
2614
2615static __inline__ __m128i __DEFAULT_FN_ATTRS
2616_mm256_mask_cvttpd_epu32 (__m128i __W, __mmask8 __U, __m256d __A) {
2617 return (__m128i) __builtin_ia32_cvttpd2udq256_mask ((__v4df) __A,
2618 (__v4si) __W,
2619 (__mmask8) __U);
2620}
2621
2622static __inline__ __m128i __DEFAULT_FN_ATTRS
2623_mm256_maskz_cvttpd_epu32 (__mmask8 __U, __m256d __A) {
2624 return (__m128i) __builtin_ia32_cvttpd2udq256_mask ((__v4df) __A,
2625 (__v4si)
2626 _mm_setzero_si128 (),
2627 (__mmask8) __U);
2628}
2629
2630static __inline__ __m128i __DEFAULT_FN_ATTRS
2631_mm_mask_cvttps_epi32 (__m128i __W, __mmask8 __U, __m128 __A) {
2632 return (__m128i) __builtin_ia32_cvttps2dq128_mask ((__v4sf) __A,
2633 (__v4si) __W,
2634 (__mmask8) __U);
2635}
2636
2637static __inline__ __m128i __DEFAULT_FN_ATTRS
2638_mm_maskz_cvttps_epi32 (__mmask8 __U, __m128 __A) {
2639 return (__m128i) __builtin_ia32_cvttps2dq128_mask ((__v4sf) __A,
2640 (__v4si)
2641 _mm_setzero_si128 (),
2642 (__mmask8) __U);
2643}
2644
2645static __inline__ __m256i __DEFAULT_FN_ATTRS
2646_mm256_mask_cvttps_epi32 (__m256i __W, __mmask8 __U, __m256 __A) {
2647 return (__m256i) __builtin_ia32_cvttps2dq256_mask ((__v8sf) __A,
2648 (__v8si) __W,
2649 (__mmask8) __U);
2650}
2651
2652static __inline__ __m256i __DEFAULT_FN_ATTRS
2653_mm256_maskz_cvttps_epi32 (__mmask8 __U, __m256 __A) {
2654 return (__m256i) __builtin_ia32_cvttps2dq256_mask ((__v8sf) __A,
2655 (__v8si)
2656 _mm256_setzero_si256 (),
2657 (__mmask8) __U);
2658}
2659
2660static __inline__ __m128i __DEFAULT_FN_ATTRS
2661_mm_cvttps_epu32 (__m128 __A) {
2662 return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A,
2663 (__v4si)
2664 _mm_setzero_si128 (),
2665 (__mmask8) -1);
2666}
2667
2668static __inline__ __m128i __DEFAULT_FN_ATTRS
2669_mm_mask_cvttps_epu32 (__m128i __W, __mmask8 __U, __m128 __A) {
2670 return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A,
2671 (__v4si) __W,
2672 (__mmask8) __U);
2673}
2674
2675static __inline__ __m128i __DEFAULT_FN_ATTRS
2676_mm_maskz_cvttps_epu32 (__mmask8 __U, __m128 __A) {
2677 return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A,
2678 (__v4si)
2679 _mm_setzero_si128 (),
2680 (__mmask8) __U);
2681}
2682
2683static __inline__ __m256i __DEFAULT_FN_ATTRS
2684_mm256_cvttps_epu32 (__m256 __A) {
2685 return (__m256i) __builtin_ia32_cvttps2udq256_mask ((__v8sf) __A,
2686 (__v8si)
2687 _mm256_setzero_si256 (),
2688 (__mmask8) -1);
2689}
2690
2691static __inline__ __m256i __DEFAULT_FN_ATTRS
2692_mm256_mask_cvttps_epu32 (__m256i __W, __mmask8 __U, __m256 __A) {
2693 return (__m256i) __builtin_ia32_cvttps2udq256_mask ((__v8sf) __A,
2694 (__v8si) __W,
2695 (__mmask8) __U);
2696}
2697
2698static __inline__ __m256i __DEFAULT_FN_ATTRS
2699_mm256_maskz_cvttps_epu32 (__mmask8 __U, __m256 __A) {
2700 return (__m256i) __builtin_ia32_cvttps2udq256_mask ((__v8sf) __A,
2701 (__v8si)
2702 _mm256_setzero_si256 (),
2703 (__mmask8) __U);
2704}
2705
2706static __inline__ __m128d __DEFAULT_FN_ATTRS
2707_mm_cvtepu32_pd (__m128i __A) {
2708 return (__m128d) __builtin_ia32_cvtudq2pd128_mask ((__v4si) __A,
2709 (__v2df)
2710 _mm_setzero_pd (),
2711 (__mmask8) -1);
2712}
2713
2714static __inline__ __m128d __DEFAULT_FN_ATTRS
2715_mm_mask_cvtepu32_pd (__m128d __W, __mmask8 __U, __m128i __A) {
2716 return (__m128d) __builtin_ia32_cvtudq2pd128_mask ((__v4si) __A,
2717 (__v2df) __W,
2718 (__mmask8) __U);
2719}
2720
2721static __inline__ __m128d __DEFAULT_FN_ATTRS
2722_mm_maskz_cvtepu32_pd (__mmask8 __U, __m128i __A) {
2723 return (__m128d) __builtin_ia32_cvtudq2pd128_mask ((__v4si) __A,
2724 (__v2df)
2725 _mm_setzero_pd (),
2726 (__mmask8) __U);
2727}
2728
2729static __inline__ __m256d __DEFAULT_FN_ATTRS
2730_mm256_cvtepu32_pd (__m128i __A) {
2731 return (__m256d) __builtin_ia32_cvtudq2pd256_mask ((__v4si) __A,
2732 (__v4df)
2733 _mm256_setzero_pd (),
2734 (__mmask8) -1);
2735}
2736
2737static __inline__ __m256d __DEFAULT_FN_ATTRS
2738_mm256_mask_cvtepu32_pd (__m256d __W, __mmask8 __U, __m128i __A) {
2739 return (__m256d) __builtin_ia32_cvtudq2pd256_mask ((__v4si) __A,
2740 (__v4df) __W,
2741 (__mmask8) __U);
2742}
2743
2744static __inline__ __m256d __DEFAULT_FN_ATTRS
2745_mm256_maskz_cvtepu32_pd (__mmask8 __U, __m128i __A) {
2746 return (__m256d) __builtin_ia32_cvtudq2pd256_mask ((__v4si) __A,
2747 (__v4df)
2748 _mm256_setzero_pd (),
2749 (__mmask8) __U);
2750}
2751
2752static __inline__ __m128 __DEFAULT_FN_ATTRS
2753_mm_cvtepu32_ps (__m128i __A) {
2754 return (__m128) __builtin_ia32_cvtudq2ps128_mask ((__v4si) __A,
2755 (__v4sf)
2756 _mm_setzero_ps (),
2757 (__mmask8) -1);
2758}
2759
2760static __inline__ __m128 __DEFAULT_FN_ATTRS
2761_mm_mask_cvtepu32_ps (__m128 __W, __mmask8 __U, __m128i __A) {
2762 return (__m128) __builtin_ia32_cvtudq2ps128_mask ((__v4si) __A,
2763 (__v4sf) __W,
2764 (__mmask8) __U);
2765}
2766
2767static __inline__ __m128 __DEFAULT_FN_ATTRS
2768_mm_maskz_cvtepu32_ps (__mmask8 __U, __m128i __A) {
2769 return (__m128) __builtin_ia32_cvtudq2ps128_mask ((__v4si) __A,
2770 (__v4sf)
2771 _mm_setzero_ps (),
2772 (__mmask8) __U);
2773}
2774
2775static __inline__ __m256 __DEFAULT_FN_ATTRS
2776_mm256_cvtepu32_ps (__m256i __A) {
2777 return (__m256) __builtin_ia32_cvtudq2ps256_mask ((__v8si) __A,
2778 (__v8sf)
2779 _mm256_setzero_ps (),
2780 (__mmask8) -1);
2781}
2782
2783static __inline__ __m256 __DEFAULT_FN_ATTRS
2784_mm256_mask_cvtepu32_ps (__m256 __W, __mmask8 __U, __m256i __A) {
2785 return (__m256) __builtin_ia32_cvtudq2ps256_mask ((__v8si) __A,
2786 (__v8sf) __W,
2787 (__mmask8) __U);
2788}
2789
2790static __inline__ __m256 __DEFAULT_FN_ATTRS
2791_mm256_maskz_cvtepu32_ps (__mmask8 __U, __m256i __A) {
2792 return (__m256) __builtin_ia32_cvtudq2ps256_mask ((__v8si) __A,
2793 (__v8sf)
2794 _mm256_setzero_ps (),
2795 (__mmask8) __U);
2796}
2797
Asaf Badouh78ee5cc2015-07-28 10:30:56 +00002798static __inline__ __m128d __DEFAULT_FN_ATTRS
2799_mm_mask_div_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
2800 return (__m128d) __builtin_ia32_divpd_mask ((__v2df) __A,
2801 (__v2df) __B,
2802 (__v2df) __W,
2803 (__mmask8) __U);
2804}
2805
2806static __inline__ __m128d __DEFAULT_FN_ATTRS
2807_mm_maskz_div_pd (__mmask8 __U, __m128d __A, __m128d __B) {
2808 return (__m128d) __builtin_ia32_divpd_mask ((__v2df) __A,
2809 (__v2df) __B,
2810 (__v2df)
2811 _mm_setzero_pd (),
2812 (__mmask8) __U);
2813}
2814
2815static __inline__ __m256d __DEFAULT_FN_ATTRS
2816_mm256_mask_div_pd (__m256d __W, __mmask8 __U, __m256d __A,
2817 __m256d __B) {
2818 return (__m256d) __builtin_ia32_divpd256_mask ((__v4df) __A,
2819 (__v4df) __B,
2820 (__v4df) __W,
2821 (__mmask8) __U);
2822}
2823
2824static __inline__ __m256d __DEFAULT_FN_ATTRS
2825_mm256_maskz_div_pd (__mmask8 __U, __m256d __A, __m256d __B) {
2826 return (__m256d) __builtin_ia32_divpd256_mask ((__v4df) __A,
2827 (__v4df) __B,
2828 (__v4df)
2829 _mm256_setzero_pd (),
2830 (__mmask8) __U);
2831}
2832
2833static __inline__ __m128 __DEFAULT_FN_ATTRS
2834_mm_mask_div_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
2835 return (__m128) __builtin_ia32_divps_mask ((__v4sf) __A,
2836 (__v4sf) __B,
2837 (__v4sf) __W,
2838 (__mmask8) __U);
2839}
2840
2841static __inline__ __m128 __DEFAULT_FN_ATTRS
2842_mm_maskz_div_ps (__mmask8 __U, __m128 __A, __m128 __B) {
2843 return (__m128) __builtin_ia32_divps_mask ((__v4sf) __A,
2844 (__v4sf) __B,
2845 (__v4sf)
2846 _mm_setzero_ps (),
2847 (__mmask8) __U);
2848}
2849
2850static __inline__ __m256 __DEFAULT_FN_ATTRS
2851_mm256_mask_div_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) {
2852 return (__m256) __builtin_ia32_divps256_mask ((__v8sf) __A,
2853 (__v8sf) __B,
2854 (__v8sf) __W,
2855 (__mmask8) __U);
2856}
2857
2858static __inline__ __m256 __DEFAULT_FN_ATTRS
2859_mm256_maskz_div_ps (__mmask8 __U, __m256 __A, __m256 __B) {
2860 return (__m256) __builtin_ia32_divps256_mask ((__v8sf) __A,
2861 (__v8sf) __B,
2862 (__v8sf)
2863 _mm256_setzero_ps (),
2864 (__mmask8) __U);
2865}
2866
2867static __inline__ __m128d __DEFAULT_FN_ATTRS
2868_mm_mask_expand_pd (__m128d __W, __mmask8 __U, __m128d __A) {
2869 return (__m128d) __builtin_ia32_expanddf128_mask ((__v2df) __A,
2870 (__v2df) __W,
2871 (__mmask8) __U);
2872}
2873
2874static __inline__ __m128d __DEFAULT_FN_ATTRS
2875_mm_maskz_expand_pd (__mmask8 __U, __m128d __A) {
2876 return (__m128d) __builtin_ia32_expanddf128_mask ((__v2df) __A,
2877 (__v2df)
2878 _mm_setzero_pd (),
2879 (__mmask8) __U);
2880}
2881
2882static __inline__ __m256d __DEFAULT_FN_ATTRS
2883_mm256_mask_expand_pd (__m256d __W, __mmask8 __U, __m256d __A) {
2884 return (__m256d) __builtin_ia32_expanddf256_mask ((__v4df) __A,
2885 (__v4df) __W,
2886 (__mmask8) __U);
2887}
2888
2889static __inline__ __m256d __DEFAULT_FN_ATTRS
2890_mm256_maskz_expand_pd (__mmask8 __U, __m256d __A) {
2891 return (__m256d) __builtin_ia32_expanddf256_mask ((__v4df) __A,
2892 (__v4df)
2893 _mm256_setzero_pd (),
2894 (__mmask8) __U);
2895}
2896
2897static __inline__ __m128i __DEFAULT_FN_ATTRS
2898_mm_mask_expand_epi64 (__m128i __W, __mmask8 __U, __m128i __A) {
2899 return (__m128i) __builtin_ia32_expanddi128_mask ((__v2di) __A,
2900 (__v2di) __W,
2901 (__mmask8) __U);
2902}
2903
2904static __inline__ __m128i __DEFAULT_FN_ATTRS
2905_mm_maskz_expand_epi64 (__mmask8 __U, __m128i __A) {
2906 return (__m128i) __builtin_ia32_expanddi128_mask ((__v2di) __A,
2907 (__v2di)
2908 _mm_setzero_si128 (),
2909 (__mmask8) __U);
2910}
2911
2912static __inline__ __m256i __DEFAULT_FN_ATTRS
2913_mm256_mask_expand_epi64 (__m256i __W, __mmask8 __U, __m256i __A) {
2914 return (__m256i) __builtin_ia32_expanddi256_mask ((__v4di) __A,
2915 (__v4di) __W,
2916 (__mmask8) __U);
2917}
2918
2919static __inline__ __m256i __DEFAULT_FN_ATTRS
2920_mm256_maskz_expand_epi64 (__mmask8 __U, __m256i __A) {
2921 return (__m256i) __builtin_ia32_expanddi256_mask ((__v4di) __A,
2922 (__v4di)
2923 _mm256_setzero_si256 (),
2924 (__mmask8) __U);
2925}
2926
2927static __inline__ __m128d __DEFAULT_FN_ATTRS
2928_mm_mask_expandloadu_pd (__m128d __W, __mmask8 __U, void const *__P) {
2929 return (__m128d) __builtin_ia32_expandloaddf128_mask ((__v2df *) __P,
2930 (__v2df) __W,
2931 (__mmask8)
2932 __U);
2933}
2934
2935static __inline__ __m128d __DEFAULT_FN_ATTRS
2936_mm_maskz_expandloadu_pd (__mmask8 __U, void const *__P) {
2937 return (__m128d) __builtin_ia32_expandloaddf128_mask ((__v2df *) __P,
2938 (__v2df)
2939 _mm_setzero_pd (),
2940 (__mmask8)
2941 __U);
2942}
2943
2944static __inline__ __m256d __DEFAULT_FN_ATTRS
2945_mm256_mask_expandloadu_pd (__m256d __W, __mmask8 __U, void const *__P) {
2946 return (__m256d) __builtin_ia32_expandloaddf256_mask ((__v4df *) __P,
2947 (__v4df) __W,
2948 (__mmask8)
2949 __U);
2950}
2951
2952static __inline__ __m256d __DEFAULT_FN_ATTRS
2953_mm256_maskz_expandloadu_pd (__mmask8 __U, void const *__P) {
2954 return (__m256d) __builtin_ia32_expandloaddf256_mask ((__v4df *) __P,
2955 (__v4df)
2956 _mm256_setzero_pd (),
2957 (__mmask8)
2958 __U);
2959}
2960
2961static __inline__ __m128i __DEFAULT_FN_ATTRS
2962_mm_mask_expandloadu_epi64 (__m128i __W, __mmask8 __U, void const *__P) {
2963 return (__m128i) __builtin_ia32_expandloaddi128_mask ((__v2di *) __P,
2964 (__v2di) __W,
2965 (__mmask8)
2966 __U);
2967}
2968
2969static __inline__ __m128i __DEFAULT_FN_ATTRS
2970_mm_maskz_expandloadu_epi64 (__mmask8 __U, void const *__P) {
2971 return (__m128i) __builtin_ia32_expandloaddi128_mask ((__v2di *) __P,
2972 (__v2di)
2973 _mm_setzero_si128 (),
2974 (__mmask8)
2975 __U);
2976}
2977
2978static __inline__ __m256i __DEFAULT_FN_ATTRS
2979_mm256_mask_expandloadu_epi64 (__m256i __W, __mmask8 __U,
2980 void const *__P) {
2981 return (__m256i) __builtin_ia32_expandloaddi256_mask ((__v4di *) __P,
2982 (__v4di) __W,
2983 (__mmask8)
2984 __U);
2985}
2986
2987static __inline__ __m256i __DEFAULT_FN_ATTRS
2988_mm256_maskz_expandloadu_epi64 (__mmask8 __U, void const *__P) {
2989 return (__m256i) __builtin_ia32_expandloaddi256_mask ((__v4di *) __P,
2990 (__v4di)
2991 _mm256_setzero_si256 (),
2992 (__mmask8)
2993 __U);
2994}
2995
2996static __inline__ __m128 __DEFAULT_FN_ATTRS
2997_mm_mask_expandloadu_ps (__m128 __W, __mmask8 __U, void const *__P) {
2998 return (__m128) __builtin_ia32_expandloadsf128_mask ((__v4sf *) __P,
2999 (__v4sf) __W,
3000 (__mmask8) __U);
3001}
3002
3003static __inline__ __m128 __DEFAULT_FN_ATTRS
3004_mm_maskz_expandloadu_ps (__mmask8 __U, void const *__P) {
3005 return (__m128) __builtin_ia32_expandloadsf128_mask ((__v4sf *) __P,
3006 (__v4sf)
3007 _mm_setzero_ps (),
3008 (__mmask8)
3009 __U);
3010}
3011
3012static __inline__ __m256 __DEFAULT_FN_ATTRS
3013_mm256_mask_expandloadu_ps (__m256 __W, __mmask8 __U, void const *__P) {
3014 return (__m256) __builtin_ia32_expandloadsf256_mask ((__v8sf *) __P,
3015 (__v8sf) __W,
3016 (__mmask8) __U);
3017}
3018
3019static __inline__ __m256 __DEFAULT_FN_ATTRS
3020_mm256_maskz_expandloadu_ps (__mmask8 __U, void const *__P) {
3021 return (__m256) __builtin_ia32_expandloadsf256_mask ((__v8sf *) __P,
3022 (__v8sf)
3023 _mm256_setzero_ps (),
3024 (__mmask8)
3025 __U);
3026}
3027
3028static __inline__ __m128i __DEFAULT_FN_ATTRS
3029_mm_mask_expandloadu_epi32 (__m128i __W, __mmask8 __U, void const *__P) {
3030 return (__m128i) __builtin_ia32_expandloadsi128_mask ((__v4si *) __P,
3031 (__v4si) __W,
3032 (__mmask8)
3033 __U);
3034}
3035
3036static __inline__ __m128i __DEFAULT_FN_ATTRS
3037_mm_maskz_expandloadu_epi32 (__mmask8 __U, void const *__P) {
3038 return (__m128i) __builtin_ia32_expandloadsi128_mask ((__v4si *) __P,
3039 (__v4si)
3040 _mm_setzero_si128 (),
3041 (__mmask8) __U);
3042}
3043
3044static __inline__ __m256i __DEFAULT_FN_ATTRS
3045_mm256_mask_expandloadu_epi32 (__m256i __W, __mmask8 __U,
3046 void const *__P) {
3047 return (__m256i) __builtin_ia32_expandloadsi256_mask ((__v8si *) __P,
3048 (__v8si) __W,
3049 (__mmask8)
3050 __U);
3051}
3052
3053static __inline__ __m256i __DEFAULT_FN_ATTRS
3054_mm256_maskz_expandloadu_epi32 (__mmask8 __U, void const *__P) {
3055 return (__m256i) __builtin_ia32_expandloadsi256_mask ((__v8si *) __P,
3056 (__v8si)
3057 _mm256_setzero_si256 (),
3058 (__mmask8)
3059 __U);
3060}
3061
3062static __inline__ __m128 __DEFAULT_FN_ATTRS
3063_mm_mask_expand_ps (__m128 __W, __mmask8 __U, __m128 __A) {
3064 return (__m128) __builtin_ia32_expandsf128_mask ((__v4sf) __A,
3065 (__v4sf) __W,
3066 (__mmask8) __U);
3067}
3068
3069static __inline__ __m128 __DEFAULT_FN_ATTRS
3070_mm_maskz_expand_ps (__mmask8 __U, __m128 __A) {
3071 return (__m128) __builtin_ia32_expandsf128_mask ((__v4sf) __A,
3072 (__v4sf)
3073 _mm_setzero_ps (),
3074 (__mmask8) __U);
3075}
3076
3077static __inline__ __m256 __DEFAULT_FN_ATTRS
3078_mm256_mask_expand_ps (__m256 __W, __mmask8 __U, __m256 __A) {
3079 return (__m256) __builtin_ia32_expandsf256_mask ((__v8sf) __A,
3080 (__v8sf) __W,
3081 (__mmask8) __U);
3082}
3083
3084static __inline__ __m256 __DEFAULT_FN_ATTRS
3085_mm256_maskz_expand_ps (__mmask8 __U, __m256 __A) {
3086 return (__m256) __builtin_ia32_expandsf256_mask ((__v8sf) __A,
3087 (__v8sf)
3088 _mm256_setzero_ps (),
3089 (__mmask8) __U);
3090}
3091
3092static __inline__ __m128i __DEFAULT_FN_ATTRS
3093_mm_mask_expand_epi32 (__m128i __W, __mmask8 __U, __m128i __A) {
3094 return (__m128i) __builtin_ia32_expandsi128_mask ((__v4si) __A,
3095 (__v4si) __W,
3096 (__mmask8) __U);
3097}
3098
3099static __inline__ __m128i __DEFAULT_FN_ATTRS
3100_mm_maskz_expand_epi32 (__mmask8 __U, __m128i __A) {
3101 return (__m128i) __builtin_ia32_expandsi128_mask ((__v4si) __A,
3102 (__v4si)
3103 _mm_setzero_si128 (),
3104 (__mmask8) __U);
3105}
3106
3107static __inline__ __m256i __DEFAULT_FN_ATTRS
3108_mm256_mask_expand_epi32 (__m256i __W, __mmask8 __U, __m256i __A) {
3109 return (__m256i) __builtin_ia32_expandsi256_mask ((__v8si) __A,
3110 (__v8si) __W,
3111 (__mmask8) __U);
3112}
3113
3114static __inline__ __m256i __DEFAULT_FN_ATTRS
3115_mm256_maskz_expand_epi32 (__mmask8 __U, __m256i __A) {
3116 return (__m256i) __builtin_ia32_expandsi256_mask ((__v8si) __A,
3117 (__v8si)
3118 _mm256_setzero_si256 (),
3119 (__mmask8) __U);
3120}
3121
3122static __inline__ __m128d __DEFAULT_FN_ATTRS
3123_mm_getexp_pd (__m128d __A) {
3124 return (__m128d) __builtin_ia32_getexppd128_mask ((__v2df) __A,
3125 (__v2df)
3126 _mm_setzero_pd (),
3127 (__mmask8) -1);
3128}
3129
3130static __inline__ __m128d __DEFAULT_FN_ATTRS
3131_mm_mask_getexp_pd (__m128d __W, __mmask8 __U, __m128d __A) {
3132 return (__m128d) __builtin_ia32_getexppd128_mask ((__v2df) __A,
3133 (__v2df) __W,
3134 (__mmask8) __U);
3135}
3136
3137static __inline__ __m128d __DEFAULT_FN_ATTRS
3138_mm_maskz_getexp_pd (__mmask8 __U, __m128d __A) {
3139 return (__m128d) __builtin_ia32_getexppd128_mask ((__v2df) __A,
3140 (__v2df)
3141 _mm_setzero_pd (),
3142 (__mmask8) __U);
3143}
3144
3145static __inline__ __m256d __DEFAULT_FN_ATTRS
3146_mm256_getexp_pd (__m256d __A) {
3147 return (__m256d) __builtin_ia32_getexppd256_mask ((__v4df) __A,
3148 (__v4df)
3149 _mm256_setzero_pd (),
3150 (__mmask8) -1);
3151}
3152
3153static __inline__ __m256d __DEFAULT_FN_ATTRS
3154_mm256_mask_getexp_pd (__m256d __W, __mmask8 __U, __m256d __A) {
3155 return (__m256d) __builtin_ia32_getexppd256_mask ((__v4df) __A,
3156 (__v4df) __W,
3157 (__mmask8) __U);
3158}
3159
3160static __inline__ __m256d __DEFAULT_FN_ATTRS
3161_mm256_maskz_getexp_pd (__mmask8 __U, __m256d __A) {
3162 return (__m256d) __builtin_ia32_getexppd256_mask ((__v4df) __A,
3163 (__v4df)
3164 _mm256_setzero_pd (),
3165 (__mmask8) __U);
3166}
3167
3168static __inline__ __m128 __DEFAULT_FN_ATTRS
3169_mm_getexp_ps (__m128 __A) {
3170 return (__m128) __builtin_ia32_getexpps128_mask ((__v4sf) __A,
3171 (__v4sf)
3172 _mm_setzero_ps (),
3173 (__mmask8) -1);
3174}
3175
3176static __inline__ __m128 __DEFAULT_FN_ATTRS
3177_mm_mask_getexp_ps (__m128 __W, __mmask8 __U, __m128 __A) {
3178 return (__m128) __builtin_ia32_getexpps128_mask ((__v4sf) __A,
3179 (__v4sf) __W,
3180 (__mmask8) __U);
3181}
3182
3183static __inline__ __m128 __DEFAULT_FN_ATTRS
3184_mm_maskz_getexp_ps (__mmask8 __U, __m128 __A) {
3185 return (__m128) __builtin_ia32_getexpps128_mask ((__v4sf) __A,
3186 (__v4sf)
3187 _mm_setzero_ps (),
3188 (__mmask8) __U);
3189}
3190
3191static __inline__ __m256 __DEFAULT_FN_ATTRS
3192_mm256_getexp_ps (__m256 __A) {
3193 return (__m256) __builtin_ia32_getexpps256_mask ((__v8sf) __A,
3194 (__v8sf)
3195 _mm256_setzero_ps (),
3196 (__mmask8) -1);
3197}
3198
3199static __inline__ __m256 __DEFAULT_FN_ATTRS
3200_mm256_mask_getexp_ps (__m256 __W, __mmask8 __U, __m256 __A) {
3201 return (__m256) __builtin_ia32_getexpps256_mask ((__v8sf) __A,
3202 (__v8sf) __W,
3203 (__mmask8) __U);
3204}
3205
3206static __inline__ __m256 __DEFAULT_FN_ATTRS
3207_mm256_maskz_getexp_ps (__mmask8 __U, __m256 __A) {
3208 return (__m256) __builtin_ia32_getexpps256_mask ((__v8sf) __A,
3209 (__v8sf)
3210 _mm256_setzero_ps (),
3211 (__mmask8) __U);
3212}
3213
3214static __inline__ __m128d __DEFAULT_FN_ATTRS
3215_mm_mask_max_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
3216 return (__m128d) __builtin_ia32_maxpd_mask ((__v2df) __A,
3217 (__v2df) __B,
3218 (__v2df) __W,
3219 (__mmask8) __U);
3220}
3221
3222static __inline__ __m128d __DEFAULT_FN_ATTRS
3223_mm_maskz_max_pd (__mmask8 __U, __m128d __A, __m128d __B) {
3224 return (__m128d) __builtin_ia32_maxpd_mask ((__v2df) __A,
3225 (__v2df) __B,
3226 (__v2df)
3227 _mm_setzero_pd (),
3228 (__mmask8) __U);
3229}
3230
3231static __inline__ __m256d __DEFAULT_FN_ATTRS
3232_mm256_mask_max_pd (__m256d __W, __mmask8 __U, __m256d __A,
3233 __m256d __B) {
3234 return (__m256d) __builtin_ia32_maxpd256_mask ((__v4df) __A,
3235 (__v4df) __B,
3236 (__v4df) __W,
3237 (__mmask8) __U);
3238}
3239
3240static __inline__ __m256d __DEFAULT_FN_ATTRS
3241_mm256_maskz_max_pd (__mmask8 __U, __m256d __A, __m256d __B) {
3242 return (__m256d) __builtin_ia32_maxpd256_mask ((__v4df) __A,
3243 (__v4df) __B,
3244 (__v4df)
3245 _mm256_setzero_pd (),
3246 (__mmask8) __U);
3247}
3248
3249static __inline__ __m128 __DEFAULT_FN_ATTRS
3250_mm_mask_max_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
3251 return (__m128) __builtin_ia32_maxps_mask ((__v4sf) __A,
3252 (__v4sf) __B,
3253 (__v4sf) __W,
3254 (__mmask8) __U);
3255}
3256
3257static __inline__ __m128 __DEFAULT_FN_ATTRS
3258_mm_maskz_max_ps (__mmask8 __U, __m128 __A, __m128 __B) {
3259 return (__m128) __builtin_ia32_maxps_mask ((__v4sf) __A,
3260 (__v4sf) __B,
3261 (__v4sf)
3262 _mm_setzero_ps (),
3263 (__mmask8) __U);
3264}
3265
3266static __inline__ __m256 __DEFAULT_FN_ATTRS
3267_mm256_mask_max_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) {
3268 return (__m256) __builtin_ia32_maxps256_mask ((__v8sf) __A,
3269 (__v8sf) __B,
3270 (__v8sf) __W,
3271 (__mmask8) __U);
3272}
3273
3274static __inline__ __m256 __DEFAULT_FN_ATTRS
3275_mm256_maskz_max_ps (__mmask8 __U, __m256 __A, __m256 __B) {
3276 return (__m256) __builtin_ia32_maxps256_mask ((__v8sf) __A,
3277 (__v8sf) __B,
3278 (__v8sf)
3279 _mm256_setzero_ps (),
3280 (__mmask8) __U);
3281}
3282
3283static __inline__ __m128d __DEFAULT_FN_ATTRS
3284_mm_mask_min_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
3285 return (__m128d) __builtin_ia32_minpd_mask ((__v2df) __A,
3286 (__v2df) __B,
3287 (__v2df) __W,
3288 (__mmask8) __U);
3289}
3290
3291static __inline__ __m128d __DEFAULT_FN_ATTRS
3292_mm_maskz_min_pd (__mmask8 __U, __m128d __A, __m128d __B) {
3293 return (__m128d) __builtin_ia32_minpd_mask ((__v2df) __A,
3294 (__v2df) __B,
3295 (__v2df)
3296 _mm_setzero_pd (),
3297 (__mmask8) __U);
3298}
3299
3300static __inline__ __m256d __DEFAULT_FN_ATTRS
3301_mm256_mask_min_pd (__m256d __W, __mmask8 __U, __m256d __A,
3302 __m256d __B) {
3303 return (__m256d) __builtin_ia32_minpd256_mask ((__v4df) __A,
3304 (__v4df) __B,
3305 (__v4df) __W,
3306 (__mmask8) __U);
3307}
3308
3309static __inline__ __m256d __DEFAULT_FN_ATTRS
3310_mm256_maskz_min_pd (__mmask8 __U, __m256d __A, __m256d __B) {
3311 return (__m256d) __builtin_ia32_minpd256_mask ((__v4df) __A,
3312 (__v4df) __B,
3313 (__v4df)
3314 _mm256_setzero_pd (),
3315 (__mmask8) __U);
3316}
3317
3318static __inline__ __m128 __DEFAULT_FN_ATTRS
3319_mm_mask_min_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
3320 return (__m128) __builtin_ia32_minps_mask ((__v4sf) __A,
3321 (__v4sf) __B,
3322 (__v4sf) __W,
3323 (__mmask8) __U);
3324}
3325
3326static __inline__ __m128 __DEFAULT_FN_ATTRS
3327_mm_maskz_min_ps (__mmask8 __U, __m128 __A, __m128 __B) {
3328 return (__m128) __builtin_ia32_minps_mask ((__v4sf) __A,
3329 (__v4sf) __B,
3330 (__v4sf)
3331 _mm_setzero_ps (),
3332 (__mmask8) __U);
3333}
3334
3335static __inline__ __m256 __DEFAULT_FN_ATTRS
3336_mm256_mask_min_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) {
3337 return (__m256) __builtin_ia32_minps256_mask ((__v8sf) __A,
3338 (__v8sf) __B,
3339 (__v8sf) __W,
3340 (__mmask8) __U);
3341}
3342
3343static __inline__ __m256 __DEFAULT_FN_ATTRS
3344_mm256_maskz_min_ps (__mmask8 __U, __m256 __A, __m256 __B) {
3345 return (__m256) __builtin_ia32_minps256_mask ((__v8sf) __A,
3346 (__v8sf) __B,
3347 (__v8sf)
3348 _mm256_setzero_ps (),
3349 (__mmask8) __U);
3350}
3351
3352static __inline__ __m128d __DEFAULT_FN_ATTRS
3353_mm_mask_mul_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
3354 return (__m128d) __builtin_ia32_mulpd_mask ((__v2df) __A,
3355 (__v2df) __B,
3356 (__v2df) __W,
3357 (__mmask8) __U);
3358}
3359
3360static __inline__ __m128d __DEFAULT_FN_ATTRS
3361_mm_maskz_mul_pd (__mmask8 __U, __m128d __A, __m128d __B) {
3362 return (__m128d) __builtin_ia32_mulpd_mask ((__v2df) __A,
3363 (__v2df) __B,
3364 (__v2df)
3365 _mm_setzero_pd (),
3366 (__mmask8) __U);
3367}
3368
3369static __inline__ __m256d __DEFAULT_FN_ATTRS
3370_mm256_mask_mul_pd (__m256d __W, __mmask8 __U, __m256d __A,
3371 __m256d __B) {
3372 return (__m256d) __builtin_ia32_mulpd256_mask ((__v4df) __A,
3373 (__v4df) __B,
3374 (__v4df) __W,
3375 (__mmask8) __U);
3376}
3377
3378static __inline__ __m256d __DEFAULT_FN_ATTRS
3379_mm256_maskz_mul_pd (__mmask8 __U, __m256d __A, __m256d __B) {
3380 return (__m256d) __builtin_ia32_mulpd256_mask ((__v4df) __A,
3381 (__v4df) __B,
3382 (__v4df)
3383 _mm256_setzero_pd (),
3384 (__mmask8) __U);
3385}
3386
3387static __inline__ __m128 __DEFAULT_FN_ATTRS
3388_mm_mask_mul_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
3389 return (__m128) __builtin_ia32_mulps_mask ((__v4sf) __A,
3390 (__v4sf) __B,
3391 (__v4sf) __W,
3392 (__mmask8) __U);
3393}
3394
3395static __inline__ __m128 __DEFAULT_FN_ATTRS
3396_mm_maskz_mul_ps (__mmask8 __U, __m128 __A, __m128 __B) {
3397 return (__m128) __builtin_ia32_mulps_mask ((__v4sf) __A,
3398 (__v4sf) __B,
3399 (__v4sf)
3400 _mm_setzero_ps (),
3401 (__mmask8) __U);
3402}
3403
3404static __inline__ __m256 __DEFAULT_FN_ATTRS
3405_mm256_mask_mul_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) {
3406 return (__m256) __builtin_ia32_mulps256_mask ((__v8sf) __A,
3407 (__v8sf) __B,
3408 (__v8sf) __W,
3409 (__mmask8) __U);
3410}
3411
3412static __inline__ __m256 __DEFAULT_FN_ATTRS
3413_mm256_maskz_mul_ps (__mmask8 __U, __m256 __A, __m256 __B) {
3414 return (__m256) __builtin_ia32_mulps256_mask ((__v8sf) __A,
3415 (__v8sf) __B,
3416 (__v8sf)
3417 _mm256_setzero_ps (),
3418 (__mmask8) __U);
3419}
3420
3421static __inline__ __m128i __DEFAULT_FN_ATTRS
3422_mm_mask_abs_epi32 (__m128i __W, __mmask8 __U, __m128i __A) {
3423 return (__m128i) __builtin_ia32_pabsd128_mask ((__v4si) __A,
3424 (__v4si) __W,
3425 (__mmask8) __U);
3426}
3427
3428static __inline__ __m128i __DEFAULT_FN_ATTRS
3429_mm_maskz_abs_epi32 (__mmask8 __U, __m128i __A) {
3430 return (__m128i) __builtin_ia32_pabsd128_mask ((__v4si) __A,
3431 (__v4si)
3432 _mm_setzero_si128 (),
3433 (__mmask8) __U);
3434}
3435
3436static __inline__ __m256i __DEFAULT_FN_ATTRS
3437_mm256_mask_abs_epi32 (__m256i __W, __mmask8 __U, __m256i __A) {
3438 return (__m256i) __builtin_ia32_pabsd256_mask ((__v8si) __A,
3439 (__v8si) __W,
3440 (__mmask8) __U);
3441}
3442
3443static __inline__ __m256i __DEFAULT_FN_ATTRS
3444_mm256_maskz_abs_epi32 (__mmask8 __U, __m256i __A) {
3445 return (__m256i) __builtin_ia32_pabsd256_mask ((__v8si) __A,
3446 (__v8si)
3447 _mm256_setzero_si256 (),
3448 (__mmask8) __U);
3449}
3450
3451static __inline__ __m128i __DEFAULT_FN_ATTRS
3452_mm_abs_epi64 (__m128i __A) {
3453 return (__m128i) __builtin_ia32_pabsq128_mask ((__v2di) __A,
3454 (__v2di)
3455 _mm_setzero_si128 (),
3456 (__mmask8) -1);
3457}
3458
3459static __inline__ __m128i __DEFAULT_FN_ATTRS
3460_mm_mask_abs_epi64 (__m128i __W, __mmask8 __U, __m128i __A) {
3461 return (__m128i) __builtin_ia32_pabsq128_mask ((__v2di) __A,
3462 (__v2di) __W,
3463 (__mmask8) __U);
3464}
3465
3466static __inline__ __m128i __DEFAULT_FN_ATTRS
3467_mm_maskz_abs_epi64 (__mmask8 __U, __m128i __A) {
3468 return (__m128i) __builtin_ia32_pabsq128_mask ((__v2di) __A,
3469 (__v2di)
3470 _mm_setzero_si128 (),
3471 (__mmask8) __U);
3472}
3473
3474static __inline__ __m256i __DEFAULT_FN_ATTRS
3475_mm256_abs_epi64 (__m256i __A) {
3476 return (__m256i) __builtin_ia32_pabsq256_mask ((__v4di) __A,
3477 (__v4di)
3478 _mm256_setzero_si256 (),
3479 (__mmask8) -1);
3480}
3481
3482static __inline__ __m256i __DEFAULT_FN_ATTRS
3483_mm256_mask_abs_epi64 (__m256i __W, __mmask8 __U, __m256i __A) {
3484 return (__m256i) __builtin_ia32_pabsq256_mask ((__v4di) __A,
3485 (__v4di) __W,
3486 (__mmask8) __U);
3487}
3488
3489static __inline__ __m256i __DEFAULT_FN_ATTRS
3490_mm256_maskz_abs_epi64 (__mmask8 __U, __m256i __A) {
3491 return (__m256i) __builtin_ia32_pabsq256_mask ((__v4di) __A,
3492 (__v4di)
3493 _mm256_setzero_si256 (),
3494 (__mmask8) __U);
3495}
3496
Asaf Badouhb7cf71b2015-07-28 11:14:09 +00003497static __inline__ __m128i __DEFAULT_FN_ATTRS
3498_mm_maskz_max_epi32 (__mmask8 __M, __m128i __A, __m128i __B) {
3499 return (__m128i) __builtin_ia32_pmaxsd128_mask ((__v4si) __A,
3500 (__v4si) __B,
3501 (__v4si)
3502 _mm_setzero_si128 (),
3503 __M);
3504}
3505
3506static __inline__ __m128i __DEFAULT_FN_ATTRS
3507_mm_mask_max_epi32 (__m128i __W, __mmask8 __M, __m128i __A,
3508 __m128i __B) {
3509 return (__m128i) __builtin_ia32_pmaxsd128_mask ((__v4si) __A,
3510 (__v4si) __B,
3511 (__v4si) __W, __M);
3512}
3513
3514static __inline__ __m256i __DEFAULT_FN_ATTRS
3515_mm256_maskz_max_epi32 (__mmask8 __M, __m256i __A, __m256i __B) {
3516 return (__m256i) __builtin_ia32_pmaxsd256_mask ((__v8si) __A,
3517 (__v8si) __B,
3518 (__v8si)
3519 _mm256_setzero_si256 (),
3520 __M);
3521}
3522
3523static __inline__ __m256i __DEFAULT_FN_ATTRS
3524_mm256_mask_max_epi32 (__m256i __W, __mmask8 __M, __m256i __A,
3525 __m256i __B) {
3526 return (__m256i) __builtin_ia32_pmaxsd256_mask ((__v8si) __A,
3527 (__v8si) __B,
3528 (__v8si) __W, __M);
3529}
3530
3531static __inline__ __m128i __DEFAULT_FN_ATTRS
3532_mm_maskz_max_epi64 (__mmask8 __M, __m128i __A, __m128i __B) {
3533 return (__m128i) __builtin_ia32_pmaxsq128_mask ((__v2di) __A,
3534 (__v2di) __B,
3535 (__v2di)
3536 _mm_setzero_si128 (),
3537 __M);
3538}
3539
3540static __inline__ __m128i __DEFAULT_FN_ATTRS
3541_mm_mask_max_epi64 (__m128i __W, __mmask8 __M, __m128i __A,
3542 __m128i __B) {
3543 return (__m128i) __builtin_ia32_pmaxsq128_mask ((__v2di) __A,
3544 (__v2di) __B,
3545 (__v2di) __W, __M);
3546}
3547
3548static __inline__ __m128i __DEFAULT_FN_ATTRS
3549_mm_max_epi64 (__m128i __A, __m128i __B) {
3550 return (__m128i) __builtin_ia32_pmaxsq128_mask ((__v2di) __A,
3551 (__v2di) __B,
3552 (__v2di)
3553 _mm_setzero_si128 (),
3554 (__mmask8) -1);
3555}
3556
3557static __inline__ __m256i __DEFAULT_FN_ATTRS
3558_mm256_maskz_max_epi64 (__mmask8 __M, __m256i __A, __m256i __B) {
3559 return (__m256i) __builtin_ia32_pmaxsq256_mask ((__v4di) __A,
3560 (__v4di) __B,
3561 (__v4di)
3562 _mm256_setzero_si256 (),
3563 __M);
3564}
3565
3566static __inline__ __m256i __DEFAULT_FN_ATTRS
3567_mm256_mask_max_epi64 (__m256i __W, __mmask8 __M, __m256i __A,
3568 __m256i __B) {
3569 return (__m256i) __builtin_ia32_pmaxsq256_mask ((__v4di) __A,
3570 (__v4di) __B,
3571 (__v4di) __W, __M);
3572}
3573
3574static __inline__ __m256i __DEFAULT_FN_ATTRS
3575_mm256_max_epi64 (__m256i __A, __m256i __B) {
3576 return (__m256i) __builtin_ia32_pmaxsq256_mask ((__v4di) __A,
3577 (__v4di) __B,
3578 (__v4di)
3579 _mm256_setzero_si256 (),
3580 (__mmask8) -1);
3581}
3582
3583static __inline__ __m128i __DEFAULT_FN_ATTRS
3584_mm_maskz_max_epu32 (__mmask8 __M, __m128i __A, __m128i __B) {
3585 return (__m128i) __builtin_ia32_pmaxud128_mask ((__v4si) __A,
3586 (__v4si) __B,
3587 (__v4si)
3588 _mm_setzero_si128 (),
3589 __M);
3590}
3591
3592static __inline__ __m128i __DEFAULT_FN_ATTRS
3593_mm_mask_max_epu32 (__m128i __W, __mmask8 __M, __m128i __A,
3594 __m128i __B) {
3595 return (__m128i) __builtin_ia32_pmaxud128_mask ((__v4si) __A,
3596 (__v4si) __B,
3597 (__v4si) __W, __M);
3598}
3599
3600static __inline__ __m256i __DEFAULT_FN_ATTRS
3601_mm256_maskz_max_epu32 (__mmask8 __M, __m256i __A, __m256i __B) {
3602 return (__m256i) __builtin_ia32_pmaxud256_mask ((__v8si) __A,
3603 (__v8si) __B,
3604 (__v8si)
3605 _mm256_setzero_si256 (),
3606 __M);
3607}
3608
3609static __inline__ __m256i __DEFAULT_FN_ATTRS
3610_mm256_mask_max_epu32 (__m256i __W, __mmask8 __M, __m256i __A,
3611 __m256i __B) {
3612 return (__m256i) __builtin_ia32_pmaxud256_mask ((__v8si) __A,
3613 (__v8si) __B,
3614 (__v8si) __W, __M);
3615}
3616
3617static __inline__ __m128i __DEFAULT_FN_ATTRS
3618_mm_maskz_max_epu64 (__mmask8 __M, __m128i __A, __m128i __B) {
3619 return (__m128i) __builtin_ia32_pmaxuq128_mask ((__v2di) __A,
3620 (__v2di) __B,
3621 (__v2di)
3622 _mm_setzero_si128 (),
3623 __M);
3624}
3625
3626static __inline__ __m128i __DEFAULT_FN_ATTRS
3627_mm_max_epu64 (__m128i __A, __m128i __B) {
3628 return (__m128i) __builtin_ia32_pmaxuq128_mask ((__v2di) __A,
3629 (__v2di) __B,
3630 (__v2di)
3631 _mm_setzero_si128 (),
3632 (__mmask8) -1);
3633}
3634
3635static __inline__ __m128i __DEFAULT_FN_ATTRS
3636_mm_mask_max_epu64 (__m128i __W, __mmask8 __M, __m128i __A,
3637 __m128i __B) {
3638 return (__m128i) __builtin_ia32_pmaxuq128_mask ((__v2di) __A,
3639 (__v2di) __B,
3640 (__v2di) __W, __M);
3641}
3642
3643static __inline__ __m256i __DEFAULT_FN_ATTRS
3644_mm256_maskz_max_epu64 (__mmask8 __M, __m256i __A, __m256i __B) {
3645 return (__m256i) __builtin_ia32_pmaxuq256_mask ((__v4di) __A,
3646 (__v4di) __B,
3647 (__v4di)
3648 _mm256_setzero_si256 (),
3649 __M);
3650}
3651
3652static __inline__ __m256i __DEFAULT_FN_ATTRS
3653_mm256_max_epu64 (__m256i __A, __m256i __B) {
3654 return (__m256i) __builtin_ia32_pmaxuq256_mask ((__v4di) __A,
3655 (__v4di) __B,
3656 (__v4di)
3657 _mm256_setzero_si256 (),
3658 (__mmask8) -1);
3659}
3660
3661static __inline__ __m256i __DEFAULT_FN_ATTRS
3662_mm256_mask_max_epu64 (__m256i __W, __mmask8 __M, __m256i __A,
3663 __m256i __B) {
3664 return (__m256i) __builtin_ia32_pmaxuq256_mask ((__v4di) __A,
3665 (__v4di) __B,
3666 (__v4di) __W, __M);
3667}
3668
3669static __inline__ __m128i __DEFAULT_FN_ATTRS
3670_mm_maskz_min_epi32 (__mmask8 __M, __m128i __A, __m128i __B) {
3671 return (__m128i) __builtin_ia32_pminsd128_mask ((__v4si) __A,
3672 (__v4si) __B,
3673 (__v4si)
3674 _mm_setzero_si128 (),
3675 __M);
3676}
3677
3678static __inline__ __m128i __DEFAULT_FN_ATTRS
3679_mm_mask_min_epi32 (__m128i __W, __mmask8 __M, __m128i __A,
3680 __m128i __B) {
3681 return (__m128i) __builtin_ia32_pminsd128_mask ((__v4si) __A,
3682 (__v4si) __B,
3683 (__v4si) __W, __M);
3684}
3685
3686static __inline__ __m256i __DEFAULT_FN_ATTRS
3687_mm256_maskz_min_epi32 (__mmask8 __M, __m256i __A, __m256i __B) {
3688 return (__m256i) __builtin_ia32_pminsd256_mask ((__v8si) __A,
3689 (__v8si) __B,
3690 (__v8si)
3691 _mm256_setzero_si256 (),
3692 __M);
3693}
3694
3695static __inline__ __m256i __DEFAULT_FN_ATTRS
3696_mm256_mask_min_epi32 (__m256i __W, __mmask8 __M, __m256i __A,
3697 __m256i __B) {
3698 return (__m256i) __builtin_ia32_pminsd256_mask ((__v8si) __A,
3699 (__v8si) __B,
3700 (__v8si) __W, __M);
3701}
3702
3703static __inline__ __m128i __DEFAULT_FN_ATTRS
3704_mm_min_epi64 (__m128i __A, __m128i __B) {
3705 return (__m128i) __builtin_ia32_pminsq128_mask ((__v2di) __A,
3706 (__v2di) __B,
3707 (__v2di)
3708 _mm_setzero_si128 (),
3709 (__mmask8) -1);
3710}
3711
3712static __inline__ __m128i __DEFAULT_FN_ATTRS
3713_mm_mask_min_epi64 (__m128i __W, __mmask8 __M, __m128i __A,
3714 __m128i __B) {
3715 return (__m128i) __builtin_ia32_pminsq128_mask ((__v2di) __A,
3716 (__v2di) __B,
3717 (__v2di) __W, __M);
3718}
3719
3720static __inline__ __m128i __DEFAULT_FN_ATTRS
3721_mm_maskz_min_epi64 (__mmask8 __M, __m128i __A, __m128i __B) {
3722 return (__m128i) __builtin_ia32_pminsq128_mask ((__v2di) __A,
3723 (__v2di) __B,
3724 (__v2di)
3725 _mm_setzero_si128 (),
3726 __M);
3727}
3728
3729static __inline__ __m256i __DEFAULT_FN_ATTRS
3730_mm256_min_epi64 (__m256i __A, __m256i __B) {
3731 return (__m256i) __builtin_ia32_pminsq256_mask ((__v4di) __A,
3732 (__v4di) __B,
3733 (__v4di)
3734 _mm256_setzero_si256 (),
3735 (__mmask8) -1);
3736}
3737
3738static __inline__ __m256i __DEFAULT_FN_ATTRS
3739_mm256_mask_min_epi64 (__m256i __W, __mmask8 __M, __m256i __A,
3740 __m256i __B) {
3741 return (__m256i) __builtin_ia32_pminsq256_mask ((__v4di) __A,
3742 (__v4di) __B,
3743 (__v4di) __W, __M);
3744}
3745
3746static __inline__ __m256i __DEFAULT_FN_ATTRS
3747_mm256_maskz_min_epi64 (__mmask8 __M, __m256i __A, __m256i __B) {
3748 return (__m256i) __builtin_ia32_pminsq256_mask ((__v4di) __A,
3749 (__v4di) __B,
3750 (__v4di)
3751 _mm256_setzero_si256 (),
3752 __M);
3753}
3754
3755static __inline__ __m128i __DEFAULT_FN_ATTRS
3756_mm_maskz_min_epu32 (__mmask8 __M, __m128i __A, __m128i __B) {
3757 return (__m128i) __builtin_ia32_pminud128_mask ((__v4si) __A,
3758 (__v4si) __B,
3759 (__v4si)
3760 _mm_setzero_si128 (),
3761 __M);
3762}
3763
3764static __inline__ __m128i __DEFAULT_FN_ATTRS
3765_mm_mask_min_epu32 (__m128i __W, __mmask8 __M, __m128i __A,
3766 __m128i __B) {
3767 return (__m128i) __builtin_ia32_pminud128_mask ((__v4si) __A,
3768 (__v4si) __B,
3769 (__v4si) __W, __M);
3770}
3771
3772static __inline__ __m256i __DEFAULT_FN_ATTRS
3773_mm256_maskz_min_epu32 (__mmask8 __M, __m256i __A, __m256i __B) {
3774 return (__m256i) __builtin_ia32_pminud256_mask ((__v8si) __A,
3775 (__v8si) __B,
3776 (__v8si)
3777 _mm256_setzero_si256 (),
3778 __M);
3779}
3780
3781static __inline__ __m256i __DEFAULT_FN_ATTRS
3782_mm256_mask_min_epu32 (__m256i __W, __mmask8 __M, __m256i __A,
3783 __m256i __B) {
3784 return (__m256i) __builtin_ia32_pminud256_mask ((__v8si) __A,
3785 (__v8si) __B,
3786 (__v8si) __W, __M);
3787}
3788
3789static __inline__ __m128i __DEFAULT_FN_ATTRS
3790_mm_min_epu64 (__m128i __A, __m128i __B) {
3791 return (__m128i) __builtin_ia32_pminuq128_mask ((__v2di) __A,
3792 (__v2di) __B,
3793 (__v2di)
3794 _mm_setzero_si128 (),
3795 (__mmask8) -1);
3796}
3797
3798static __inline__ __m128i __DEFAULT_FN_ATTRS
3799_mm_mask_min_epu64 (__m128i __W, __mmask8 __M, __m128i __A,
3800 __m128i __B) {
3801 return (__m128i) __builtin_ia32_pminuq128_mask ((__v2di) __A,
3802 (__v2di) __B,
3803 (__v2di) __W, __M);
3804}
3805
3806static __inline__ __m128i __DEFAULT_FN_ATTRS
3807_mm_maskz_min_epu64 (__mmask8 __M, __m128i __A, __m128i __B) {
3808 return (__m128i) __builtin_ia32_pminuq128_mask ((__v2di) __A,
3809 (__v2di) __B,
3810 (__v2di)
3811 _mm_setzero_si128 (),
3812 __M);
3813}
3814
3815static __inline__ __m256i __DEFAULT_FN_ATTRS
3816_mm256_min_epu64 (__m256i __A, __m256i __B) {
3817 return (__m256i) __builtin_ia32_pminuq256_mask ((__v4di) __A,
3818 (__v4di) __B,
3819 (__v4di)
3820 _mm256_setzero_si256 (),
3821 (__mmask8) -1);
3822}
3823
3824static __inline__ __m256i __DEFAULT_FN_ATTRS
3825_mm256_mask_min_epu64 (__m256i __W, __mmask8 __M, __m256i __A,
3826 __m256i __B) {
3827 return (__m256i) __builtin_ia32_pminuq256_mask ((__v4di) __A,
3828 (__v4di) __B,
3829 (__v4di) __W, __M);
3830}
3831
3832static __inline__ __m256i __DEFAULT_FN_ATTRS
3833_mm256_maskz_min_epu64 (__mmask8 __M, __m256i __A, __m256i __B) {
3834 return (__m256i) __builtin_ia32_pminuq256_mask ((__v4di) __A,
3835 (__v4di) __B,
3836 (__v4di)
3837 _mm256_setzero_si256 (),
3838 __M);
3839}
3840
3841#define _mm_roundscale_pd(__A, __imm) __extension__ ({ \
3842 (__m128d) __builtin_ia32_rndscalepd_128_mask ((__v2df) __A, \
3843 __imm, (__v2df) _mm_setzero_pd (), (__mmask8) -1); })
3844
3845
3846#define _mm_mask_roundscale_pd(__W, __U, __A, __imm) __extension__ ({ \
3847 (__m128d) __builtin_ia32_rndscalepd_128_mask ((__v2df) __A, __imm, \
3848 (__v2df) __W, (__mmask8) __U); })
3849
3850
3851#define _mm_maskz_roundscale_pd(__U, __A, __imm) __extension__ ({ \
3852 (__m128d) __builtin_ia32_rndscalepd_128_mask ((__v2df) __A, __imm, \
3853 (__v2df) _mm_setzero_pd (), (__mmask8) __U); })
3854
3855
3856#define _mm256_roundscale_pd(__A, __imm) __extension__ ({ \
3857 (__m256d) __builtin_ia32_rndscalepd_256_mask ((__v4df) __A, __imm, \
3858 (__v4df) _mm256_setzero_pd (), (__mmask8) -1); })
3859
3860
3861#define _mm256_mask_roundscale_pd(__W, __U, __A, __imm) __extension__ ({ \
3862 (__m256d) __builtin_ia32_rndscalepd_256_mask ((__v4df) __A, __imm, \
3863 (__v4df) __W, (__mmask8) __U); })
3864
3865
3866#define _mm256_maskz_roundscale_pd(__U, __A, __imm) __extension__ ({ \
3867 (__m256d) __builtin_ia32_rndscalepd_256_mask ((__v4df) __A, __imm, \
3868 (__v4df) _mm256_setzero_pd(), (__mmask8) __U); })
3869
3870#define _mm_roundscale_ps(__A, __imm) __extension__ ({ \
3871 (__m128) __builtin_ia32_rndscaleps_128_mask ((__v4sf) __A, __imm, \
3872 (__v4sf) _mm_setzero_ps(), (__mmask8) -1); })
3873
3874
3875#define _mm_mask_roundscale_ps(__W, __U, __A, __imm) __extension__ ({ \
3876 (__m128) __builtin_ia32_rndscaleps_128_mask ((__v4sf) __A, __imm, \
3877 (__v4sf) __W, (__mmask8) __U); })
3878
3879
3880#define _mm_maskz_roundscale_ps(__U, __A, __imm) __extension__ ({ \
3881 (__m128) __builtin_ia32_rndscaleps_128_mask ((__v4sf) __A, __imm, \
3882 (__v4sf) _mm_setzero_ps(), (__mmask8) __U); })
3883
3884#define _mm256_roundscale_ps(__A, __imm) __extension__ ({ \
3885 (__m256) __builtin_ia32_rndscaleps_256_mask ((__v8sf) __A,__imm, \
3886 (__v8sf) _mm256_setzero_ps(), (__mmask8) -1); })
3887
3888#define _mm256_mask_roundscale_ps(__W, __U, __A,__imm) __extension__ ({ \
3889 (__m256) __builtin_ia32_rndscaleps_256_mask ((__v8sf) __A, __imm, \
3890 (__v8sf) __W, (__mmask8) __U); })
3891
3892
3893#define _mm256_maskz_roundscale_ps(__U, __A, __imm) __extension__ ({ \
3894 (__m256) __builtin_ia32_rndscaleps_256_mask ((__v8sf) __A, __imm, \
3895 (__v8sf) _mm256_setzero_ps(), (__mmask8) __U); })
3896
3897static __inline__ __m128d __DEFAULT_FN_ATTRS
3898_mm_scalef_pd (__m128d __A, __m128d __B) {
3899 return (__m128d) __builtin_ia32_scalefpd128_mask ((__v2df) __A,
3900 (__v2df) __B,
3901 (__v2df)
3902 _mm_setzero_pd (),
3903 (__mmask8) -1);
3904}
3905
3906static __inline__ __m128d __DEFAULT_FN_ATTRS
3907_mm_mask_scalef_pd (__m128d __W, __mmask8 __U, __m128d __A,
3908 __m128d __B) {
3909 return (__m128d) __builtin_ia32_scalefpd128_mask ((__v2df) __A,
3910 (__v2df) __B,
3911 (__v2df) __W,
3912 (__mmask8) __U);
3913}
3914
3915static __inline__ __m128d __DEFAULT_FN_ATTRS
3916_mm_maskz_scalef_pd (__mmask8 __U, __m128d __A, __m128d __B) {
3917 return (__m128d) __builtin_ia32_scalefpd128_mask ((__v2df) __A,
3918 (__v2df) __B,
3919 (__v2df)
3920 _mm_setzero_pd (),
3921 (__mmask8) __U);
3922}
3923
3924static __inline__ __m256d __DEFAULT_FN_ATTRS
3925_mm256_scalef_pd (__m256d __A, __m256d __B) {
3926 return (__m256d) __builtin_ia32_scalefpd256_mask ((__v4df) __A,
3927 (__v4df) __B,
3928 (__v4df)
3929 _mm256_setzero_pd (),
3930 (__mmask8) -1);
3931}
3932
3933static __inline__ __m256d __DEFAULT_FN_ATTRS
3934_mm256_mask_scalef_pd (__m256d __W, __mmask8 __U, __m256d __A,
3935 __m256d __B) {
3936 return (__m256d) __builtin_ia32_scalefpd256_mask ((__v4df) __A,
3937 (__v4df) __B,
3938 (__v4df) __W,
3939 (__mmask8) __U);
3940}
3941
3942static __inline__ __m256d __DEFAULT_FN_ATTRS
3943_mm256_maskz_scalef_pd (__mmask8 __U, __m256d __A, __m256d __B) {
3944 return (__m256d) __builtin_ia32_scalefpd256_mask ((__v4df) __A,
3945 (__v4df) __B,
3946 (__v4df)
3947 _mm256_setzero_pd (),
3948 (__mmask8) __U);
3949}
3950
3951static __inline__ __m128 __DEFAULT_FN_ATTRS
3952_mm_scalef_ps (__m128 __A, __m128 __B) {
3953 return (__m128) __builtin_ia32_scalefps128_mask ((__v4sf) __A,
3954 (__v4sf) __B,
3955 (__v4sf)
3956 _mm_setzero_ps (),
3957 (__mmask8) -1);
3958}
3959
3960static __inline__ __m128 __DEFAULT_FN_ATTRS
3961_mm_mask_scalef_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
3962 return (__m128) __builtin_ia32_scalefps128_mask ((__v4sf) __A,
3963 (__v4sf) __B,
3964 (__v4sf) __W,
3965 (__mmask8) __U);
3966}
3967
3968static __inline__ __m128 __DEFAULT_FN_ATTRS
3969_mm_maskz_scalef_ps (__mmask8 __U, __m128 __A, __m128 __B) {
3970 return (__m128) __builtin_ia32_scalefps128_mask ((__v4sf) __A,
3971 (__v4sf) __B,
3972 (__v4sf)
3973 _mm_setzero_ps (),
3974 (__mmask8) __U);
3975}
3976
3977static __inline__ __m256 __DEFAULT_FN_ATTRS
3978_mm256_scalef_ps (__m256 __A, __m256 __B) {
3979 return (__m256) __builtin_ia32_scalefps256_mask ((__v8sf) __A,
3980 (__v8sf) __B,
3981 (__v8sf)
3982 _mm256_setzero_ps (),
3983 (__mmask8) -1);
3984}
3985
3986static __inline__ __m256 __DEFAULT_FN_ATTRS
3987_mm256_mask_scalef_ps (__m256 __W, __mmask8 __U, __m256 __A,
3988 __m256 __B) {
3989 return (__m256) __builtin_ia32_scalefps256_mask ((__v8sf) __A,
3990 (__v8sf) __B,
3991 (__v8sf) __W,
3992 (__mmask8) __U);
3993}
3994
3995static __inline__ __m256 __DEFAULT_FN_ATTRS
3996_mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) {
3997 return (__m256) __builtin_ia32_scalefps256_mask ((__v8sf) __A,
3998 (__v8sf) __B,
3999 (__v8sf)
4000 _mm256_setzero_ps (),
4001 (__mmask8) __U);
4002}
4003
4004#define _mm_i64scatter_pd(__addr,__index, __v1, __scale) __extension__ ({ \
4005 __builtin_ia32_scatterdiv2df(__addr, (__mmask8) 0xFF, (__v2di) __index, \
4006 (__v2df) __v1, __scale); })
4007
4008#define _mm_mask_i64scatter_pd(__addr, __mask, __index, __v1, \
4009 __scale) __extension__ ({ \
4010 __builtin_ia32_scatterdiv2df (__addr, __mask, (__v2di) __index, \
4011 (__v2df) __v1, __scale); })
4012
4013
4014#define _mm_i64scatter_epi64(__addr, __index, __v1, __scale) __extension__ ({ \
4015 __builtin_ia32_scatterdiv2di (__addr, (__mmask8) 0xFF, \
4016 (__v2di) __index, (__v2di) __v1, __scale); })
4017
4018#define _mm_mask_i64scatter_epi64(__addr, __mask, __index, __v1,\
4019 __scale) __extension__ ({ \
4020 __builtin_ia32_scatterdiv2di (__addr, __mask, (__v2di) __index,\
4021 (__v2di) __v1, __scale); })
4022
4023#define _mm256_i64scatter_pd(__addr, __index, __v1, __scale) __extension__ ({ \
4024 __builtin_ia32_scatterdiv4df (__addr, (__mmask8) 0xFF,\
4025 (__v4di) __index, (__v4df) __v1, __scale); })
4026
4027#define _mm256_mask_i64scatter_pd(__addr, __mask, __index, __v1,\
4028 __scale) __extension__ ({ \
4029 __builtin_ia32_scatterdiv4df (__addr, __mask, (__v4di) __index,\
4030 (__v4df) __v1, __scale); })
4031
4032#define _mm256_i64scatter_epi64(__addr, __index, __v1, __scale) __extension__ ({ \
4033 __builtin_ia32_scatterdiv4di (__addr, (__mmask8) 0xFF, (__v4di) __index,\
4034 (__v4di) __v1, __scale); })
4035
4036#define _mm256_mask_i64scatter_epi64(__addr, __mask, __index, __v1,\
4037 __scale) __extension__ ({ \
4038 __builtin_ia32_scatterdiv4di (__addr, __mask, (__v4di) __index,\
4039 (__v4di) __v1, __scale); })
4040
4041#define _mm_i64scatter_ps(__addr, __index, __v1, __scale) __extension__ ({ \
4042 __builtin_ia32_scatterdiv4sf (__addr, (__mmask8) 0xFF,\
4043 (__v2di) __index, (__v4sf) __v1, __scale); })
4044
4045#define _mm_mask_i64scatter_ps(__addr, __mask, __index, __v1, \
4046 __scale) __extension__ ({ \
4047 __builtin_ia32_scatterdiv4sf (__addr, __mask, (__v2di) __index,\
4048 (__v4sf) __v1, __scale); })
4049
4050#define _mm_i64scatter_epi32(__addr, __index, __v1, \
4051 __scale) __extension__ ({ \
4052 __builtin_ia32_scatterdiv4si (__addr, (__mmask8) 0xFF,\
4053 (__v2di) __index, (__v4si) __v1, __scale); })
4054
4055#define _mm_mask_i64scatter_epi32(__addr, __mask, __index, __v1,\
4056 __scale) __extension__ ({ \
4057 __builtin_ia32_scatterdiv4si (__addr, __mask, (__v2di) __index,\
4058 (__v4si) __v1, __scale); })
4059
4060#define _mm256_i64scatter_ps(__addr, __index, __v1, __scale) __extension__ ({ \
4061 __builtin_ia32_scatterdiv8sf (__addr, (__mmask8) 0xFF, (__v4di) __index, \
4062 (__v4sf) __v1, __scale); })
4063
4064#define _mm256_mask_i64scatter_ps(__addr, __mask, __index, __v1, \
4065 __scale) __extension__ ({ \
4066 __builtin_ia32_scatterdiv8sf (__addr, __mask, (__v4di) __index, \
4067 (__v4sf) __v1, __scale); })
4068
4069#define _mm256_i64scatter_epi32(__addr, __index, __v1, __scale) __extension__ ({ \
4070 __builtin_ia32_scatterdiv8si (__addr, (__mmask8) 0xFF, \
4071 (__v4di) __index, (__v4si) __v1, __scale); })
4072
4073#define _mm256_mask_i64scatter_epi32(__addr, __mask, __index, __v1, \
4074 __scale) __extension__ ({ \
4075 __builtin_ia32_scatterdiv8si(__addr, __mask, (__v4di) __index, \
4076 (__v4si) __v1, __scale); })
4077
4078#define _mm_i32scatter_pd(__addr, __index, __v1, \
4079 __scale) __extension__ ({ \
4080 __builtin_ia32_scattersiv2df (__addr, (__mmask8) 0xFF, \
4081 (__v4si) __index, (__v2df) __v1, __scale); })
4082
4083#define _mm_mask_i32scatter_pd(__addr, __mask, __index, __v1, \
4084 __scale) __extension__ ({ \
4085 __builtin_ia32_scattersiv2df (__addr, __mask, (__v4si) __index,\
4086 (__v2df) __v1, __scale); })
4087
4088#define _mm_i32scatter_epi64(__addr, __index, __v1, __scale) __extension__ ({ \
4089 __builtin_ia32_scattersiv2di (__addr, (__mmask8) 0xFF, \
4090 (__v4si) __index, (__v2di) __v1, __scale); })
4091
4092#define _mm_mask_i32scatter_epi64(__addr, __mask, __index, __v1, \
4093 __scale) __extension__ ({ \
4094 __builtin_ia32_scattersiv2di (__addr, __mask, (__v4si) __index, \
4095 (__v2di) __v1, __scale); })
4096
4097#define _mm256_i32scatter_pd(__addr, __index, __v1, __scale) __extension__ ({ \
4098 __builtin_ia32_scattersiv4df (__addr, (__mmask8) 0xFF, \
4099 (__v4si) __index, (__v4df) __v1, __scale); })
4100
4101#define _mm256_mask_i32scatter_pd(__addr, __mask, __index, __v1, \
4102 __scale) __extension__ ({ \
4103 __builtin_ia32_scattersiv4df (__addr, __mask, (__v4si) __index, \
4104 (__v4df) __v1, __scale); })
4105
4106#define _mm256_i32scatter_epi64(__addr, __index, __v1, \
4107 __scale) __extension__ ({ \
4108 __builtin_ia32_scattersiv4di (__addr, (__mmask8) 0xFF, \
4109 (__v4si) __index, (__v4di) __v1, __scale); })
4110
4111#define _mm256_mask_i32scatter_epi64(__addr, __mask, __index, __v1, \
4112 __scale) __extension__ ({ \
4113 __builtin_ia32_scattersiv4di (__addr, __mask, (__v4si) __index, \
4114 (__v4di) __v1, __scale); })
4115
4116#define _mm_i32scatter_ps(__addr, __index, __v1, __scale) __extension__ ({ \
4117 __builtin_ia32_scattersiv4sf (__addr, (__mmask8) 0xFF, \
4118 (__v4si) __index, (__v4sf) __v1, __scale); })
4119
4120#define _mm_mask_i32scatter_ps(__addr, __mask, __index, __v1, \
4121 __scale) __extension__ ({ \
4122 __builtin_ia32_scattersiv4sf (__addr, __mask, (__v4si) __index, \
4123 (__v4sf) __v1, __scale); })
4124
4125#define _mm_i32scatter_epi32(__addr, __index, __v1, __scale) __extension__ ({ \
4126 __builtin_ia32_scattersiv4si (__addr, (__mmask8) 0xFF, \
4127 (__v4si) __index, (__v4si) __v1, __scale); })
4128
4129#define _mm_mask_i32scatter_epi32(__addr, __mask, __index, __v1, \
4130 __scale) __extension__ ({ \
4131 __builtin_ia32_scattersiv4si (__addr, __mask, (__v4si) __index,\
4132 (__v4si) __v1, __scale); })
4133
4134#define _mm256_i32scatter_ps(__addr, __index, __v1, __scale) __extension__ ({ \
4135 __builtin_ia32_scattersiv8sf (__addr, (__mmask8) 0xFF, \
4136 (__v8si) __index, (__v8sf) __v1, __scale); })
4137
4138#define _mm256_mask_i32scatter_ps(__addr, __mask, __index, __v1, \
4139 __scale) __extension__ ({ \
4140 __builtin_ia32_scattersiv8sf (__addr, __mask, (__v8si) __index,\
4141 (__v8sf) __v1, __scale); })
4142
4143#define _mm256_i32scatter_epi32(__addr, __index, __v1, __scale) __extension__ ({ \
4144 __builtin_ia32_scattersiv8si (__addr, (__mmask8) 0xFF, \
4145 (__v8si) __index, (__v8si) __v1, __scale); })
4146
4147#define _mm256_mask_i32scatter_epi32(__addr, __mask, __index, __v1, \
4148 __scale) __extension__ ({ \
4149 __builtin_ia32_scattersiv8si (__addr, __mask, (__v8si) __index, \
4150 (__v8si) __v1, __scale); })
4151
Asaf Badouh93aa4c82015-07-28 12:04:40 +00004152static __inline__ __m128d __DEFAULT_FN_ATTRS
4153_mm_mask_sqrt_pd (__m128d __W, __mmask8 __U, __m128d __A) {
4154 return (__m128d) __builtin_ia32_sqrtpd128_mask ((__v2df) __A,
4155 (__v2df) __W,
4156 (__mmask8) __U);
4157}
4158
4159static __inline__ __m128d __DEFAULT_FN_ATTRS
4160_mm_maskz_sqrt_pd (__mmask8 __U, __m128d __A) {
4161 return (__m128d) __builtin_ia32_sqrtpd128_mask ((__v2df) __A,
4162 (__v2df)
4163 _mm_setzero_pd (),
4164 (__mmask8) __U);
4165}
4166
4167static __inline__ __m256d __DEFAULT_FN_ATTRS
4168_mm256_mask_sqrt_pd (__m256d __W, __mmask8 __U, __m256d __A) {
4169 return (__m256d) __builtin_ia32_sqrtpd256_mask ((__v4df) __A,
4170 (__v4df) __W,
4171 (__mmask8) __U);
4172}
4173
4174static __inline__ __m256d __DEFAULT_FN_ATTRS
4175_mm256_maskz_sqrt_pd (__mmask8 __U, __m256d __A) {
4176 return (__m256d) __builtin_ia32_sqrtpd256_mask ((__v4df) __A,
4177 (__v4df)
4178 _mm256_setzero_pd (),
4179 (__mmask8) __U);
4180}
4181
4182static __inline__ __m128 __DEFAULT_FN_ATTRS
4183_mm_mask_sqrt_ps (__m128 __W, __mmask8 __U, __m128 __A) {
4184 return (__m128) __builtin_ia32_sqrtps128_mask ((__v4sf) __A,
4185 (__v4sf) __W,
4186 (__mmask8) __U);
4187}
4188
4189static __inline__ __m128 __DEFAULT_FN_ATTRS
4190_mm_maskz_sqrt_ps (__mmask8 __U, __m128 __A) {
4191 return (__m128) __builtin_ia32_sqrtps128_mask ((__v4sf) __A,
4192 (__v4sf)
4193 _mm_setzero_ps (),
4194 (__mmask8) __U);
4195}
4196
4197static __inline__ __m256 __DEFAULT_FN_ATTRS
4198_mm256_mask_sqrt_ps (__m256 __W, __mmask8 __U, __m256 __A) {
4199 return (__m256) __builtin_ia32_sqrtps256_mask ((__v8sf) __A,
4200 (__v8sf) __W,
4201 (__mmask8) __U);
4202}
4203
4204static __inline__ __m256 __DEFAULT_FN_ATTRS
4205_mm256_maskz_sqrt_ps (__mmask8 __U, __m256 __A) {
4206 return (__m256) __builtin_ia32_sqrtps256_mask ((__v8sf) __A,
4207 (__v8sf)
4208 _mm256_setzero_ps (),
4209 (__mmask8) __U);
4210}
4211
4212static __inline__ __m128d __DEFAULT_FN_ATTRS
4213_mm_mask_sub_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
4214 return (__m128d) __builtin_ia32_subpd128_mask ((__v2df) __A,
4215 (__v2df) __B,
4216 (__v2df) __W,
4217 (__mmask8) __U);
4218}
4219
4220static __inline__ __m128d __DEFAULT_FN_ATTRS
4221_mm_maskz_sub_pd (__mmask8 __U, __m128d __A, __m128d __B) {
4222 return (__m128d) __builtin_ia32_subpd128_mask ((__v2df) __A,
4223 (__v2df) __B,
4224 (__v2df)
4225 _mm_setzero_pd (),
4226 (__mmask8) __U);
4227}
4228
4229static __inline__ __m256d __DEFAULT_FN_ATTRS
4230_mm256_mask_sub_pd (__m256d __W, __mmask8 __U, __m256d __A,
4231 __m256d __B) {
4232 return (__m256d) __builtin_ia32_subpd256_mask ((__v4df) __A,
4233 (__v4df) __B,
4234 (__v4df) __W,
4235 (__mmask8) __U);
4236}
4237
4238static __inline__ __m256d __DEFAULT_FN_ATTRS
4239_mm256_maskz_sub_pd (__mmask8 __U, __m256d __A, __m256d __B) {
4240 return (__m256d) __builtin_ia32_subpd256_mask ((__v4df) __A,
4241 (__v4df) __B,
4242 (__v4df)
4243 _mm256_setzero_pd (),
4244 (__mmask8) __U);
4245}
4246
4247static __inline__ __m128 __DEFAULT_FN_ATTRS
4248_mm_mask_sub_ps (__m128 __W, __mmask16 __U, __m128 __A, __m128 __B) {
4249 return (__m128) __builtin_ia32_subps128_mask ((__v4sf) __A,
4250 (__v4sf) __B,
4251 (__v4sf) __W,
4252 (__mmask8) __U);
4253}
4254
4255static __inline__ __m128 __DEFAULT_FN_ATTRS
4256_mm_maskz_sub_ps (__mmask16 __U, __m128 __A, __m128 __B) {
4257 return (__m128) __builtin_ia32_subps128_mask ((__v4sf) __A,
4258 (__v4sf) __B,
4259 (__v4sf)
4260 _mm_setzero_ps (),
4261 (__mmask8) __U);
4262}
4263
4264static __inline__ __m256 __DEFAULT_FN_ATTRS
4265_mm256_mask_sub_ps (__m256 __W, __mmask16 __U, __m256 __A, __m256 __B) {
4266 return (__m256) __builtin_ia32_subps256_mask ((__v8sf) __A,
4267 (__v8sf) __B,
4268 (__v8sf) __W,
4269 (__mmask8) __U);
4270}
4271
4272static __inline__ __m256 __DEFAULT_FN_ATTRS
4273_mm256_maskz_sub_ps (__mmask16 __U, __m256 __A, __m256 __B) {
4274 return (__m256) __builtin_ia32_subps256_mask ((__v8sf) __A,
4275 (__v8sf) __B,
4276 (__v8sf)
4277 _mm256_setzero_ps (),
4278 (__mmask8) __U);
4279}
4280
4281static __inline__ __m128i __DEFAULT_FN_ATTRS
4282_mm_mask2_permutex2var_epi32 (__m128i __A, __m128i __I, __mmask8 __U,
4283 __m128i __B) {
4284 return (__m128i) __builtin_ia32_vpermi2vard128_mask ((__v4si) __A,
4285 (__v4si) __I
4286 /* idx */ ,
4287 (__v4si) __B,
4288 (__mmask8) __U);
4289}
4290
4291static __inline__ __m256i __DEFAULT_FN_ATTRS
4292_mm256_mask2_permutex2var_epi32 (__m256i __A, __m256i __I,
4293 __mmask8 __U, __m256i __B) {
4294 return (__m256i) __builtin_ia32_vpermi2vard256_mask ((__v8si) __A,
4295 (__v8si) __I
4296 /* idx */ ,
4297 (__v8si) __B,
4298 (__mmask8) __U);
4299}
4300
4301static __inline__ __m128d __DEFAULT_FN_ATTRS
4302_mm_mask2_permutex2var_pd (__m128d __A, __m128i __I, __mmask8 __U,
4303 __m128d __B) {
4304 return (__m128d) __builtin_ia32_vpermi2varpd128_mask ((__v2df) __A,
4305 (__v2di) __I
4306 /* idx */ ,
4307 (__v2df) __B,
4308 (__mmask8)
4309 __U);
4310}
4311
4312static __inline__ __m256d __DEFAULT_FN_ATTRS
4313_mm256_mask2_permutex2var_pd (__m256d __A, __m256i __I, __mmask8 __U,
4314 __m256d __B) {
4315 return (__m256d) __builtin_ia32_vpermi2varpd256_mask ((__v4df) __A,
4316 (__v4di) __I
4317 /* idx */ ,
4318 (__v4df) __B,
4319 (__mmask8)
4320 __U);
4321}
4322
4323static __inline__ __m128 __DEFAULT_FN_ATTRS
4324_mm_mask2_permutex2var_ps (__m128 __A, __m128i __I, __mmask8 __U,
4325 __m128 __B) {
4326 return (__m128) __builtin_ia32_vpermi2varps128_mask ((__v4sf) __A,
4327 (__v4si) __I
4328 /* idx */ ,
4329 (__v4sf) __B,
4330 (__mmask8) __U);
4331}
4332
4333static __inline__ __m256 __DEFAULT_FN_ATTRS
4334_mm256_mask2_permutex2var_ps (__m256 __A, __m256i __I, __mmask8 __U,
4335 __m256 __B) {
4336 return (__m256) __builtin_ia32_vpermi2varps256_mask ((__v8sf) __A,
4337 (__v8si) __I
4338 /* idx */ ,
4339 (__v8sf) __B,
4340 (__mmask8) __U);
4341}
4342
4343static __inline__ __m128i __DEFAULT_FN_ATTRS
4344_mm_mask2_permutex2var_epi64 (__m128i __A, __m128i __I, __mmask8 __U,
4345 __m128i __B) {
4346 return (__m128i) __builtin_ia32_vpermi2varq128_mask ((__v2di) __A,
4347 (__v2di) __I
4348 /* idx */ ,
4349 (__v2di) __B,
4350 (__mmask8) __U);
4351}
4352
4353static __inline__ __m256i __DEFAULT_FN_ATTRS
4354_mm256_mask2_permutex2var_epi64 (__m256i __A, __m256i __I,
4355 __mmask8 __U, __m256i __B) {
4356 return (__m256i) __builtin_ia32_vpermi2varq256_mask ((__v4di) __A,
4357 (__v4di) __I
4358 /* idx */ ,
4359 (__v4di) __B,
4360 (__mmask8) __U);
4361}
4362
4363static __inline__ __m128i __DEFAULT_FN_ATTRS
4364_mm_permutex2var_epi32 (__m128i __A, __m128i __I, __m128i __B) {
4365 return (__m128i) __builtin_ia32_vpermt2vard128_mask ((__v4si) __I
4366 /* idx */ ,
4367 (__v4si) __A,
4368 (__v4si) __B,
4369 (__mmask8) -1);
4370}
4371
4372static __inline__ __m128i __DEFAULT_FN_ATTRS
4373_mm_mask_permutex2var_epi32 (__m128i __A, __mmask8 __U, __m128i __I,
4374 __m128i __B) {
4375 return (__m128i) __builtin_ia32_vpermt2vard128_mask ((__v4si) __I
4376 /* idx */ ,
4377 (__v4si) __A,
4378 (__v4si) __B,
4379 (__mmask8) __U);
4380}
4381
4382static __inline__ __m128i __DEFAULT_FN_ATTRS
4383_mm_maskz_permutex2var_epi32 (__mmask8 __U, __m128i __A, __m128i __I,
4384 __m128i __B) {
4385 return (__m128i) __builtin_ia32_vpermt2vard128_maskz ((__v4si) __I
4386 /* idx */ ,
4387 (__v4si) __A,
4388 (__v4si) __B,
4389 (__mmask8)
4390 __U);
4391}
4392
4393static __inline__ __m256i __DEFAULT_FN_ATTRS
4394_mm256_permutex2var_epi32 (__m256i __A, __m256i __I, __m256i __B) {
4395 return (__m256i) __builtin_ia32_vpermt2vard256_mask ((__v8si) __I
4396 /* idx */ ,
4397 (__v8si) __A,
4398 (__v8si) __B,
4399 (__mmask8) -1);
4400}
4401
4402static __inline__ __m256i __DEFAULT_FN_ATTRS
4403_mm256_mask_permutex2var_epi32 (__m256i __A, __mmask8 __U, __m256i __I,
4404 __m256i __B) {
4405 return (__m256i) __builtin_ia32_vpermt2vard256_mask ((__v8si) __I
4406 /* idx */ ,
4407 (__v8si) __A,
4408 (__v8si) __B,
4409 (__mmask8) __U);
4410}
4411
4412static __inline__ __m256i __DEFAULT_FN_ATTRS
4413_mm256_maskz_permutex2var_epi32 (__mmask8 __U, __m256i __A,
4414 __m256i __I, __m256i __B) {
4415 return (__m256i) __builtin_ia32_vpermt2vard256_maskz ((__v8si) __I
4416 /* idx */ ,
4417 (__v8si) __A,
4418 (__v8si) __B,
4419 (__mmask8)
4420 __U);
4421}
4422
4423static __inline__ __m128d __DEFAULT_FN_ATTRS
4424_mm_permutex2var_pd (__m128d __A, __m128i __I, __m128d __B) {
4425 return (__m128d) __builtin_ia32_vpermt2varpd128_mask ((__v2di) __I
4426 /* idx */ ,
4427 (__v2df) __A,
4428 (__v2df) __B,
4429 (__mmask8) -
4430 1);
4431}
4432
4433static __inline__ __m128d __DEFAULT_FN_ATTRS
4434_mm_mask_permutex2var_pd (__m128d __A, __mmask8 __U, __m128i __I,
4435 __m128d __B) {
4436 return (__m128d) __builtin_ia32_vpermt2varpd128_mask ((__v2di) __I
4437 /* idx */ ,
4438 (__v2df) __A,
4439 (__v2df) __B,
4440 (__mmask8)
4441 __U);
4442}
4443
4444static __inline__ __m128d __DEFAULT_FN_ATTRS
4445_mm_maskz_permutex2var_pd (__mmask8 __U, __m128d __A, __m128i __I,
4446 __m128d __B) {
4447 return (__m128d) __builtin_ia32_vpermt2varpd128_maskz ((__v2di) __I
4448 /* idx */ ,
4449 (__v2df) __A,
4450 (__v2df) __B,
4451 (__mmask8)
4452 __U);
4453}
4454
4455static __inline__ __m256d __DEFAULT_FN_ATTRS
4456_mm256_permutex2var_pd (__m256d __A, __m256i __I, __m256d __B) {
4457 return (__m256d) __builtin_ia32_vpermt2varpd256_mask ((__v4di) __I
4458 /* idx */ ,
4459 (__v4df) __A,
4460 (__v4df) __B,
4461 (__mmask8) -
4462 1);
4463}
4464
4465static __inline__ __m256d __DEFAULT_FN_ATTRS
4466_mm256_mask_permutex2var_pd (__m256d __A, __mmask8 __U, __m256i __I,
4467 __m256d __B) {
4468 return (__m256d) __builtin_ia32_vpermt2varpd256_mask ((__v4di) __I
4469 /* idx */ ,
4470 (__v4df) __A,
4471 (__v4df) __B,
4472 (__mmask8)
4473 __U);
4474}
4475
4476static __inline__ __m256d __DEFAULT_FN_ATTRS
4477_mm256_maskz_permutex2var_pd (__mmask8 __U, __m256d __A, __m256i __I,
4478 __m256d __B) {
4479 return (__m256d) __builtin_ia32_vpermt2varpd256_maskz ((__v4di) __I
4480 /* idx */ ,
4481 (__v4df) __A,
4482 (__v4df) __B,
4483 (__mmask8)
4484 __U);
4485}
4486
4487static __inline__ __m128 __DEFAULT_FN_ATTRS
4488_mm_permutex2var_ps (__m128 __A, __m128i __I, __m128 __B) {
4489 return (__m128) __builtin_ia32_vpermt2varps128_mask ((__v4si) __I
4490 /* idx */ ,
4491 (__v4sf) __A,
4492 (__v4sf) __B,
4493 (__mmask8) -1);
4494}
4495
4496static __inline__ __m128 __DEFAULT_FN_ATTRS
4497_mm_mask_permutex2var_ps (__m128 __A, __mmask8 __U, __m128i __I,
4498 __m128 __B) {
4499 return (__m128) __builtin_ia32_vpermt2varps128_mask ((__v4si) __I
4500 /* idx */ ,
4501 (__v4sf) __A,
4502 (__v4sf) __B,
4503 (__mmask8) __U);
4504}
4505
4506static __inline__ __m128 __DEFAULT_FN_ATTRS
4507_mm_maskz_permutex2var_ps (__mmask8 __U, __m128 __A, __m128i __I,
4508 __m128 __B) {
4509 return (__m128) __builtin_ia32_vpermt2varps128_maskz ((__v4si) __I
4510 /* idx */ ,
4511 (__v4sf) __A,
4512 (__v4sf) __B,
4513 (__mmask8)
4514 __U);
4515}
4516
4517static __inline__ __m256 __DEFAULT_FN_ATTRS
4518_mm256_permutex2var_ps (__m256 __A, __m256i __I, __m256 __B) {
4519 return (__m256) __builtin_ia32_vpermt2varps256_mask ((__v8si) __I
4520 /* idx */ ,
4521 (__v8sf) __A,
4522 (__v8sf) __B,
4523 (__mmask8) -1);
4524}
4525
4526static __inline__ __m256 __DEFAULT_FN_ATTRS
4527_mm256_mask_permutex2var_ps (__m256 __A, __mmask8 __U, __m256i __I,
4528 __m256 __B) {
4529 return (__m256) __builtin_ia32_vpermt2varps256_mask ((__v8si) __I
4530 /* idx */ ,
4531 (__v8sf) __A,
4532 (__v8sf) __B,
4533 (__mmask8) __U);
4534}
4535
4536static __inline__ __m256 __DEFAULT_FN_ATTRS
4537_mm256_maskz_permutex2var_ps (__mmask8 __U, __m256 __A, __m256i __I,
4538 __m256 __B) {
4539 return (__m256) __builtin_ia32_vpermt2varps256_maskz ((__v8si) __I
4540 /* idx */ ,
4541 (__v8sf) __A,
4542 (__v8sf) __B,
4543 (__mmask8)
4544 __U);
4545}
4546
4547static __inline__ __m128i __DEFAULT_FN_ATTRS
4548_mm_permutex2var_epi64 (__m128i __A, __m128i __I, __m128i __B) {
4549 return (__m128i) __builtin_ia32_vpermt2varq128_mask ((__v2di) __I
4550 /* idx */ ,
4551 (__v2di) __A,
4552 (__v2di) __B,
4553 (__mmask8) -1);
4554}
4555
4556static __inline__ __m128i __DEFAULT_FN_ATTRS
4557_mm_mask_permutex2var_epi64 (__m128i __A, __mmask8 __U, __m128i __I,
4558 __m128i __B) {
4559 return (__m128i) __builtin_ia32_vpermt2varq128_mask ((__v2di) __I
4560 /* idx */ ,
4561 (__v2di) __A,
4562 (__v2di) __B,
4563 (__mmask8) __U);
4564}
4565
4566static __inline__ __m128i __DEFAULT_FN_ATTRS
4567_mm_maskz_permutex2var_epi64 (__mmask8 __U, __m128i __A, __m128i __I,
4568 __m128i __B) {
4569 return (__m128i) __builtin_ia32_vpermt2varq128_maskz ((__v2di) __I
4570 /* idx */ ,
4571 (__v2di) __A,
4572 (__v2di) __B,
4573 (__mmask8)
4574 __U);
4575}
4576
4577
4578static __inline__ __m256i __DEFAULT_FN_ATTRS
4579_mm256_permutex2var_epi64 (__m256i __A, __m256i __I, __m256i __B) {
4580 return (__m256i) __builtin_ia32_vpermt2varq256_mask ((__v4di) __I
4581 /* idx */ ,
4582 (__v4di) __A,
4583 (__v4di) __B,
4584 (__mmask8) -1);
4585}
4586
4587static __inline__ __m256i __DEFAULT_FN_ATTRS
4588_mm256_mask_permutex2var_epi64 (__m256i __A, __mmask8 __U, __m256i __I,
4589 __m256i __B) {
4590 return (__m256i) __builtin_ia32_vpermt2varq256_mask ((__v4di) __I
4591 /* idx */ ,
4592 (__v4di) __A,
4593 (__v4di) __B,
4594 (__mmask8) __U);
4595}
4596
4597static __inline__ __m256i __DEFAULT_FN_ATTRS
4598_mm256_maskz_permutex2var_epi64 (__mmask8 __U, __m256i __A,
4599 __m256i __I, __m256i __B) {
4600 return (__m256i) __builtin_ia32_vpermt2varq256_maskz ((__v4di) __I
4601 /* idx */ ,
4602 (__v4di) __A,
4603 (__v4di) __B,
4604 (__mmask8)
4605 __U);
4606}
4607
Michael Zuckerman7cdb72f2016-02-18 09:09:34 +00004608static __inline__ __m128i __DEFAULT_FN_ATTRS
4609_mm_mask_cvtepi8_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
4610{
4611 return (__m128i) __builtin_ia32_pmovsxbd128_mask ((__v16qi) __A,
4612 (__v4si) __W,
4613 (__mmask8) __U);
4614}
4615
4616static __inline__ __m128i __DEFAULT_FN_ATTRS
4617_mm_maskz_cvtepi8_epi32 (__mmask8 __U, __m128i __A)
4618{
4619 return (__m128i) __builtin_ia32_pmovsxbd128_mask ((__v16qi) __A,
4620 (__v4si)
4621 _mm_setzero_si128 (),
4622 (__mmask8) __U);
4623}
4624
4625static __inline__ __m256i __DEFAULT_FN_ATTRS
4626_mm256_mask_cvtepi8_epi32 (__m256i __W, __mmask8 __U, __m128i __A)
4627{
4628 return (__m256i) __builtin_ia32_pmovsxbd256_mask ((__v16qi) __A,
4629 (__v8si) __W,
4630 (__mmask8) __U);
4631}
4632
4633static __inline__ __m256i __DEFAULT_FN_ATTRS
4634_mm256_maskz_cvtepi8_epi32 (__mmask8 __U, __m128i __A)
4635{
4636 return (__m256i) __builtin_ia32_pmovsxbd256_mask ((__v16qi) __A,
4637 (__v8si)
4638 _mm256_setzero_si256 (),
4639 (__mmask8) __U);
4640}
4641
4642static __inline__ __m128i __DEFAULT_FN_ATTRS
4643_mm_mask_cvtepi8_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
4644{
4645 return (__m128i) __builtin_ia32_pmovsxbq128_mask ((__v16qi) __A,
4646 (__v2di) __W,
4647 (__mmask8) __U);
4648}
4649
4650static __inline__ __m128i __DEFAULT_FN_ATTRS
4651_mm_maskz_cvtepi8_epi64 (__mmask8 __U, __m128i __A)
4652{
4653 return (__m128i) __builtin_ia32_pmovsxbq128_mask ((__v16qi) __A,
4654 (__v2di)
4655 _mm_setzero_si128 (),
4656 (__mmask8) __U);
4657}
4658
4659static __inline__ __m256i __DEFAULT_FN_ATTRS
4660_mm256_mask_cvtepi8_epi64 (__m256i __W, __mmask8 __U, __m128i __A)
4661{
4662 return (__m256i) __builtin_ia32_pmovsxbq256_mask ((__v16qi) __A,
4663 (__v4di) __W,
4664 (__mmask8) __U);
4665}
4666
4667static __inline__ __m256i __DEFAULT_FN_ATTRS
4668_mm256_maskz_cvtepi8_epi64 (__mmask8 __U, __m128i __A)
4669{
4670 return (__m256i) __builtin_ia32_pmovsxbq256_mask ((__v16qi) __A,
4671 (__v4di)
4672 _mm256_setzero_si256 (),
4673 (__mmask8) __U);
4674}
4675
4676static __inline__ __m128i __DEFAULT_FN_ATTRS
4677_mm_mask_cvtepi32_epi64 (__m128i __W, __mmask8 __U, __m128i __X)
4678{
4679 return (__m128i) __builtin_ia32_pmovsxdq128_mask ((__v4si) __X,
4680 (__v2di) __W,
4681 (__mmask8) __U);
4682}
4683
4684static __inline__ __m128i __DEFAULT_FN_ATTRS
4685_mm_maskz_cvtepi32_epi64 (__mmask8 __U, __m128i __X)
4686{
4687 return (__m128i) __builtin_ia32_pmovsxdq128_mask ((__v4si) __X,
4688 (__v2di)
4689 _mm_setzero_si128 (),
4690 (__mmask8) __U);
4691}
4692
4693static __inline__ __m256i __DEFAULT_FN_ATTRS
4694_mm256_mask_cvtepi32_epi64 (__m256i __W, __mmask8 __U, __m128i __X)
4695{
4696 return (__m256i) __builtin_ia32_pmovsxdq256_mask ((__v4si) __X,
4697 (__v4di) __W,
4698 (__mmask8) __U);
4699}
4700
4701static __inline__ __m256i __DEFAULT_FN_ATTRS
4702_mm256_maskz_cvtepi32_epi64 (__mmask8 __U, __m128i __X)
4703{
4704 return (__m256i) __builtin_ia32_pmovsxdq256_mask ((__v4si) __X,
4705 (__v4di)
4706 _mm256_setzero_si256 (),
4707 (__mmask8) __U);
4708}
4709
4710static __inline__ __m128i __DEFAULT_FN_ATTRS
4711_mm_mask_cvtepi16_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
4712{
4713 return (__m128i) __builtin_ia32_pmovsxwd128_mask ((__v8hi) __A,
4714 (__v4si) __W,
4715 (__mmask8) __U);
4716}
4717
4718static __inline__ __m128i __DEFAULT_FN_ATTRS
4719_mm_maskz_cvtepi16_epi32 (__mmask8 __U, __m128i __A)
4720{
4721 return (__m128i) __builtin_ia32_pmovsxwd128_mask ((__v8hi) __A,
4722 (__v4si)
4723 _mm_setzero_si128 (),
4724 (__mmask8) __U);
4725}
4726
4727static __inline__ __m256i __DEFAULT_FN_ATTRS
4728_mm256_mask_cvtepi16_epi32 (__m256i __W, __mmask8 __U, __m128i __A)
4729{
4730 return (__m256i) __builtin_ia32_pmovsxwd256_mask ((__v8hi) __A,
4731 (__v8si) __W,
4732 (__mmask8) __U);
4733}
4734
4735static __inline__ __m256i __DEFAULT_FN_ATTRS
4736_mm256_maskz_cvtepi16_epi32 (__mmask8 __U, __m128i __A)
4737{
4738 return (__m256i) __builtin_ia32_pmovsxwd256_mask ((__v8hi) __A,
4739 (__v8si)
4740 _mm256_setzero_si256 (),
4741 (__mmask8) __U);
4742}
4743
4744static __inline__ __m128i __DEFAULT_FN_ATTRS
4745_mm_mask_cvtepi16_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
4746{
4747 return (__m128i) __builtin_ia32_pmovsxwq128_mask ((__v8hi) __A,
4748 (__v2di) __W,
4749 (__mmask8) __U);
4750}
4751
4752static __inline__ __m128i __DEFAULT_FN_ATTRS
4753_mm_maskz_cvtepi16_epi64 (__mmask8 __U, __m128i __A)
4754{
4755 return (__m128i) __builtin_ia32_pmovsxwq128_mask ((__v8hi) __A,
4756 (__v2di)
4757 _mm_setzero_si128 (),
4758 (__mmask8) __U);
4759}
4760
4761static __inline__ __m256i __DEFAULT_FN_ATTRS
4762_mm256_mask_cvtepi16_epi64 (__m256i __W, __mmask8 __U, __m128i __A)
4763{
4764 return (__m256i) __builtin_ia32_pmovsxwq256_mask ((__v8hi) __A,
4765 (__v4di) __W,
4766 (__mmask8) __U);
4767}
4768
4769static __inline__ __m256i __DEFAULT_FN_ATTRS
4770_mm256_maskz_cvtepi16_epi64 (__mmask8 __U, __m128i __A)
4771{
4772 return (__m256i) __builtin_ia32_pmovsxwq256_mask ((__v8hi) __A,
4773 (__v4di)
4774 _mm256_setzero_si256 (),
4775 (__mmask8) __U);
4776}
4777
4778
Michael Zuckerman7a33dce2016-02-21 14:00:11 +00004779static __inline__ __m128i __DEFAULT_FN_ATTRS
4780_mm_mask_cvtepu8_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
4781{
4782 return (__m128i) __builtin_ia32_pmovzxbd128_mask ((__v16qi) __A,
4783 (__v4si) __W,
4784 (__mmask8) __U);
4785}
4786
4787static __inline__ __m128i __DEFAULT_FN_ATTRS
4788_mm_maskz_cvtepu8_epi32 (__mmask8 __U, __m128i __A)
4789{
4790 return (__m128i) __builtin_ia32_pmovzxbd128_mask ((__v16qi) __A,
4791 (__v4si)
4792 _mm_setzero_si128 (),
4793 (__mmask8) __U);
4794}
4795
4796static __inline__ __m256i __DEFAULT_FN_ATTRS
4797_mm256_mask_cvtepu8_epi32 (__m256i __W, __mmask8 __U, __m128i __A)
4798{
4799 return (__m256i) __builtin_ia32_pmovzxbd256_mask ((__v16qi) __A,
4800 (__v8si) __W,
4801 (__mmask8) __U);
4802}
4803
4804static __inline__ __m256i __DEFAULT_FN_ATTRS
4805_mm256_maskz_cvtepu8_epi32 (__mmask8 __U, __m128i __A)
4806{
4807 return (__m256i) __builtin_ia32_pmovzxbd256_mask ((__v16qi) __A,
4808 (__v8si)
4809 _mm256_setzero_si256 (),
4810 (__mmask8) __U);
4811}
4812
4813static __inline__ __m128i __DEFAULT_FN_ATTRS
4814_mm_mask_cvtepu8_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
4815{
4816 return (__m128i) __builtin_ia32_pmovzxbq128_mask ((__v16qi) __A,
4817 (__v2di) __W,
4818 (__mmask8) __U);
4819}
4820
4821static __inline__ __m128i __DEFAULT_FN_ATTRS
4822_mm_maskz_cvtepu8_epi64 (__mmask8 __U, __m128i __A)
4823{
4824 return (__m128i) __builtin_ia32_pmovzxbq128_mask ((__v16qi) __A,
4825 (__v2di)
4826 _mm_setzero_si128 (),
4827 (__mmask8) __U);
4828}
4829
4830static __inline__ __m256i __DEFAULT_FN_ATTRS
4831_mm256_mask_cvtepu8_epi64 (__m256i __W, __mmask8 __U, __m128i __A)
4832{
4833 return (__m256i) __builtin_ia32_pmovzxbq256_mask ((__v16qi) __A,
4834 (__v4di) __W,
4835 (__mmask8) __U);
4836}
4837
4838static __inline__ __m256i __DEFAULT_FN_ATTRS
4839_mm256_maskz_cvtepu8_epi64 (__mmask8 __U, __m128i __A)
4840{
4841 return (__m256i) __builtin_ia32_pmovzxbq256_mask ((__v16qi) __A,
4842 (__v4di)
4843 _mm256_setzero_si256 (),
4844 (__mmask8) __U);
4845}
4846
4847static __inline__ __m128i __DEFAULT_FN_ATTRS
4848_mm_mask_cvtepu32_epi64 (__m128i __W, __mmask8 __U, __m128i __X)
4849{
4850 return (__m128i) __builtin_ia32_pmovzxdq128_mask ((__v4si) __X,
4851 (__v2di) __W,
4852 (__mmask8) __U);
4853}
4854
4855static __inline__ __m128i __DEFAULT_FN_ATTRS
4856_mm_maskz_cvtepu32_epi64 (__mmask8 __U, __m128i __X)
4857{
4858 return (__m128i) __builtin_ia32_pmovzxdq128_mask ((__v4si) __X,
4859 (__v2di)
4860 _mm_setzero_si128 (),
4861 (__mmask8) __U);
4862}
4863
4864static __inline__ __m256i __DEFAULT_FN_ATTRS
4865_mm256_mask_cvtepu32_epi64 (__m256i __W, __mmask8 __U, __m128i __X)
4866{
4867 return (__m256i) __builtin_ia32_pmovzxdq256_mask ((__v4si) __X,
4868 (__v4di) __W,
4869 (__mmask8) __U);
4870}
4871
4872static __inline__ __m256i __DEFAULT_FN_ATTRS
4873_mm256_maskz_cvtepu32_epi64 (__mmask8 __U, __m128i __X)
4874{
4875 return (__m256i) __builtin_ia32_pmovzxdq256_mask ((__v4si) __X,
4876 (__v4di)
4877 _mm256_setzero_si256 (),
4878 (__mmask8) __U);
4879}
4880
4881static __inline__ __m128i __DEFAULT_FN_ATTRS
4882_mm_mask_cvtepu16_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
4883{
4884 return (__m128i) __builtin_ia32_pmovzxwd128_mask ((__v8hi) __A,
4885 (__v4si) __W,
4886 (__mmask8) __U);
4887}
4888
4889static __inline__ __m128i __DEFAULT_FN_ATTRS
4890_mm_maskz_cvtepu16_epi32 (__mmask8 __U, __m128i __A)
4891{
4892 return (__m128i) __builtin_ia32_pmovzxwd128_mask ((__v8hi) __A,
4893 (__v4si)
4894 _mm_setzero_si128 (),
4895 (__mmask8) __U);
4896}
4897
4898static __inline__ __m256i __DEFAULT_FN_ATTRS
4899_mm256_mask_cvtepu16_epi32 (__m256i __W, __mmask8 __U, __m128i __A)
4900{
4901 return (__m256i) __builtin_ia32_pmovzxwd256_mask ((__v8hi) __A,
4902 (__v8si) __W,
4903 (__mmask8) __U);
4904}
4905
4906static __inline__ __m256i __DEFAULT_FN_ATTRS
4907_mm256_maskz_cvtepu16_epi32 (__mmask8 __U, __m128i __A)
4908{
4909 return (__m256i) __builtin_ia32_pmovzxwd256_mask ((__v8hi) __A,
4910 (__v8si)
4911 _mm256_setzero_si256 (),
4912 (__mmask8) __U);
4913}
4914
4915static __inline__ __m128i __DEFAULT_FN_ATTRS
4916_mm_mask_cvtepu16_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
4917{
4918 return (__m128i) __builtin_ia32_pmovzxwq128_mask ((__v8hi) __A,
4919 (__v2di) __W,
4920 (__mmask8) __U);
4921}
4922
4923static __inline__ __m128i __DEFAULT_FN_ATTRS
4924_mm_maskz_cvtepu16_epi64 (__mmask8 __U, __m128i __A)
4925{
4926 return (__m128i) __builtin_ia32_pmovzxwq128_mask ((__v8hi) __A,
4927 (__v2di)
4928 _mm_setzero_si128 (),
4929 (__mmask8) __U);
4930}
4931
4932static __inline__ __m256i __DEFAULT_FN_ATTRS
4933_mm256_mask_cvtepu16_epi64 (__m256i __W, __mmask8 __U, __m128i __A)
4934{
4935 return (__m256i) __builtin_ia32_pmovzxwq256_mask ((__v8hi) __A,
4936 (__v4di) __W,
4937 (__mmask8) __U);
4938}
4939
4940static __inline__ __m256i __DEFAULT_FN_ATTRS
4941_mm256_maskz_cvtepu16_epi64 (__mmask8 __U, __m128i __A)
4942{
4943 return (__m256i) __builtin_ia32_pmovzxwq256_mask ((__v8hi) __A,
4944 (__v4di)
4945 _mm256_setzero_si256 (),
4946 (__mmask8) __U);
4947}
4948
4949
Michael Zuckerman38a27272016-02-22 09:05:41 +00004950#define _mm_rol_epi32(a, b) __extension__ ({\
4951 (__m128i)__builtin_ia32_prold128_mask((__v4si) (a), (b),\
4952 (__v4si)\
4953 _mm_setzero_si128 (),\
4954 (__mmask8) -1); })
4955
4956#define _mm_mask_rol_epi32(w, u, a, b) __extension__ ({\
4957 (__m128i)__builtin_ia32_prold128_mask((__v4si) (a), (b),\
4958 (__v4si) (w),\
4959 (__mmask8) (u)); })
4960
4961#define _mm_maskz_rol_epi32(u, a, b) __extension__ ({\
4962 (__m128i)__builtin_ia32_prold128_mask((__v4si) (a), (b),\
4963 (__v4si)\
4964 _mm_setzero_si128 (),\
4965 (__mmask8) (u)); })
4966
4967#define _mm256_rol_epi32(a, b) __extension__ ({\
4968 (__m256i)__builtin_ia32_prold256_mask((__v8si) (a), (b),\
4969 (__v8si)\
4970 _mm256_setzero_si256 (),\
4971 (__mmask8) -1); })
4972
4973#define _mm256_mask_rol_epi32(w, u, a, b) __extension__ ({\
4974 (__m256i)__builtin_ia32_prold256_mask((__v8si) (a), (b),\
4975 (__v8si) (w),\
4976 (__mmask8) (u)); })
4977
4978#define _mm256_maskz_rol_epi32(u, a, b) __extension__ ({\
4979 (__m256i)__builtin_ia32_prold256_mask((__v8si) (a), (b),\
4980 (__v8si)\
4981 _mm256_setzero_si256 (),\
4982 (__mmask8) (u)); })
4983
Michael Zuckerman477e0a32016-02-22 09:42:57 +00004984#define _mm_rol_epi64(a, b) __extension__ ({\
Michael Zuckerman38a27272016-02-22 09:05:41 +00004985 (__m128i)__builtin_ia32_prolq128_mask((__v2di) (a), (b),\
4986 (__v2di)\
4987 _mm_setzero_di (),\
4988 (__mmask8) -1); })
4989
4990#define _mm_mask_rol_epi64(w, u, a, b) __extension__ ({\
4991 (__m128i)__builtin_ia32_prolq128_mask((__v2di) (a), (b),\
4992 (__v2di) (w),\
4993 (__mmask8) (u)); })
4994
4995#define _mm_maskz_rol_epi64(u, a, b) __extension__ ({\
4996 (__m128i)__builtin_ia32_prolq128_mask((__v2di) (a), (b),\
4997 (__v2di)\
4998 _mm_setzero_di(),\
4999 (__mmask8) (u)); })
5000
5001#define _mm256_rol_epi64(a, b) __extension__ ({\
5002 (__m256i)__builtin_ia32_prolq256_mask((__v4di) (a), (b),\
5003 (__v4di)\
5004 _mm256_setzero_si256 (),\
5005 (__mmask8) -1); })
5006
5007#define _mm256_mask_rol_epi64(w, u, a, b) __extension__ ({\
5008 (__m256i)__builtin_ia32_prolq256_mask((__v4di) (a), (b),\
5009 (__v4di) (w),\
5010 (__mmask8) (u)); })
5011
5012#define _mm256_maskz_rol_epi64(u, a, b) __extension__ ({\
5013 (__m256i)__builtin_ia32_prolq256_mask((__v4di) (a), (b),\
5014 (__v4di)\
5015 _mm256_setzero_si256 (),\
5016 (__mmask8) (u)); })
5017
Michael Zuckerman0231f162016-02-23 13:41:13 +00005018static __inline__ __m128i __DEFAULT_FN_ATTRS
5019_mm_rolv_epi32 (__m128i __A, __m128i __B)
5020{
5021 return (__m128i) __builtin_ia32_prolvd128_mask ((__v4si) __A,
5022 (__v4si) __B,
5023 (__v4si)
5024 _mm_setzero_si128 (),
5025 (__mmask8) -1);
5026}
5027
5028static __inline__ __m128i __DEFAULT_FN_ATTRS
5029_mm_mask_rolv_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
5030 __m128i __B)
5031{
5032 return (__m128i) __builtin_ia32_prolvd128_mask ((__v4si) __A,
5033 (__v4si) __B,
5034 (__v4si) __W,
5035 (__mmask8) __U);
5036}
5037
5038static __inline__ __m128i __DEFAULT_FN_ATTRS
5039_mm_maskz_rolv_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
5040{
5041 return (__m128i) __builtin_ia32_prolvd128_mask ((__v4si) __A,
5042 (__v4si) __B,
5043 (__v4si)
5044 _mm_setzero_si128 (),
5045 (__mmask8) __U);
5046}
5047
5048static __inline__ __m256i __DEFAULT_FN_ATTRS
5049_mm256_rolv_epi32 (__m256i __A, __m256i __B)
5050{
5051 return (__m256i) __builtin_ia32_prolvd256_mask ((__v8si) __A,
5052 (__v8si) __B,
5053 (__v8si)
5054 _mm256_setzero_si256 (),
5055 (__mmask8) -1);
5056}
5057
5058static __inline__ __m256i __DEFAULT_FN_ATTRS
5059_mm256_mask_rolv_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
5060 __m256i __B)
5061{
5062 return (__m256i) __builtin_ia32_prolvd256_mask ((__v8si) __A,
5063 (__v8si) __B,
5064 (__v8si) __W,
5065 (__mmask8) __U);
5066}
5067
5068static __inline__ __m256i __DEFAULT_FN_ATTRS
5069_mm256_maskz_rolv_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
5070{
5071 return (__m256i) __builtin_ia32_prolvd256_mask ((__v8si) __A,
5072 (__v8si) __B,
5073 (__v8si)
5074 _mm256_setzero_si256 (),
5075 (__mmask8) __U);
5076}
5077
5078static __inline__ __m128i __DEFAULT_FN_ATTRS
5079_mm_rolv_epi64 (__m128i __A, __m128i __B)
5080{
5081 return (__m128i) __builtin_ia32_prolvq128_mask ((__v2di) __A,
5082 (__v2di) __B,
5083 (__v2di)
5084 _mm_setzero_di (),
5085 (__mmask8) -1);
5086}
5087
5088static __inline__ __m128i __DEFAULT_FN_ATTRS
5089_mm_mask_rolv_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
5090 __m128i __B)
5091{
5092 return (__m128i) __builtin_ia32_prolvq128_mask ((__v2di) __A,
5093 (__v2di) __B,
5094 (__v2di) __W,
5095 (__mmask8) __U);
5096}
5097
5098static __inline__ __m128i __DEFAULT_FN_ATTRS
5099_mm_maskz_rolv_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
5100{
5101 return (__m128i) __builtin_ia32_prolvq128_mask ((__v2di) __A,
5102 (__v2di) __B,
5103 (__v2di)
5104 _mm_setzero_di (),
5105 (__mmask8) __U);
5106}
5107
5108static __inline__ __m256i __DEFAULT_FN_ATTRS
5109_mm256_rolv_epi64 (__m256i __A, __m256i __B)
5110{
5111 return (__m256i) __builtin_ia32_prolvq256_mask ((__v4di) __A,
5112 (__v4di) __B,
5113 (__v4di)
5114 _mm256_setzero_si256 (),
5115 (__mmask8) -1);
5116}
5117
5118static __inline__ __m256i __DEFAULT_FN_ATTRS
5119_mm256_mask_rolv_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
5120 __m256i __B)
5121{
5122 return (__m256i) __builtin_ia32_prolvq256_mask ((__v4di) __A,
5123 (__v4di) __B,
5124 (__v4di) __W,
5125 (__mmask8) __U);
5126}
5127
5128static __inline__ __m256i __DEFAULT_FN_ATTRS
5129_mm256_maskz_rolv_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
5130{
5131 return (__m256i) __builtin_ia32_prolvq256_mask ((__v4di) __A,
5132 (__v4di) __B,
5133 (__v4di)
5134 _mm256_setzero_si256 (),
5135 (__mmask8) __U);
5136}
5137
5138#define _mm_ror_epi32( __A, __B) __extension__ ({ \
5139__builtin_ia32_prord128_mask ((__v4si)( __A),( __B),\
5140 (__v4si)\
5141 _mm_setzero_si128 (),\
5142 (__mmask8) -1);\
Michael Zuckerman4924c7a2016-02-23 14:23:53 +00005143})
Michael Zuckerman0231f162016-02-23 13:41:13 +00005144
5145#define _mm_mask_ror_epi32( __W, __U, __A ,__B) __extension__ ({ \
5146__builtin_ia32_prord128_mask ((__v4si) __A, __B,\
5147 (__v4si)( __W),\
5148 (__mmask8)( __U));\
Michael Zuckerman4924c7a2016-02-23 14:23:53 +00005149})
Michael Zuckerman0231f162016-02-23 13:41:13 +00005150
5151#define _mm_maskz_ror_epi32( __U, __A, __B) __extension__ ({ \
5152__builtin_ia32_prord128_mask ((__v4si)( __A),( __B),\
5153 (__v4si)\
5154 _mm_setzero_si128 (),\
5155 (__mmask8)( __U));\
Michael Zuckerman4924c7a2016-02-23 14:23:53 +00005156})
Michael Zuckerman0231f162016-02-23 13:41:13 +00005157
5158#define _mm256_ror_epi32( __A, __B) __extension__ ({ \
5159__builtin_ia32_prord256_mask ((__v8si)( __A),( __B),\
5160 (__v8si)\
5161 _mm256_setzero_si256 (),\
5162 (__mmask8) -1);\
Michael Zuckerman4924c7a2016-02-23 14:23:53 +00005163})
Michael Zuckerman0231f162016-02-23 13:41:13 +00005164
5165#define _mm256_mask_ror_epi32( __W, __U, __A ,__B) __extension__ ({ \
5166__builtin_ia32_prord256_mask ((__v8si) __A, __B,\
5167 (__v8si)( __W),\
5168 (__mmask8)( __U));\
Michael Zuckerman4924c7a2016-02-23 14:23:53 +00005169})
Michael Zuckerman0231f162016-02-23 13:41:13 +00005170
5171#define _mm256_maskz_ror_epi32( __U, __A, __B) __extension__ ({ \
5172__builtin_ia32_prord256_mask ((__v8si)( __A),( __B),\
5173 (__v8si)\
5174 _mm256_setzero_si256 (),\
5175 (__mmask8)( __U));\
Michael Zuckerman4924c7a2016-02-23 14:23:53 +00005176})
Michael Zuckerman0231f162016-02-23 13:41:13 +00005177
5178#define _mm_ror_epi64( __A, __B) __extension__ ({ \
5179__builtin_ia32_prorq128_mask ((__v2di)( __A),( __B),\
5180 (__v2di)\
5181 _mm_setzero_di (),\
5182 (__mmask8) -1);\
Michael Zuckerman4924c7a2016-02-23 14:23:53 +00005183})
Michael Zuckerman0231f162016-02-23 13:41:13 +00005184
5185#define _mm_mask_ror_epi64( __W, __U, __A ,__B) __extension__ ({ \
5186__builtin_ia32_prorq128_mask ((__v2di) __A, __B,\
5187 (__v2di)( __W),\
5188 (__mmask8)( __U));\
Michael Zuckerman4924c7a2016-02-23 14:23:53 +00005189})
Michael Zuckerman0231f162016-02-23 13:41:13 +00005190
5191#define _mm_maskz_ror_epi64( __U, __A, __B) __extension__ ({ \
5192__builtin_ia32_prorq128_mask ((__v2di)( __A),( __B),\
5193 (__v2di)\
5194 _mm_setzero_di (),\
5195 (__mmask8)( __U));\
Michael Zuckerman4924c7a2016-02-23 14:23:53 +00005196})
Michael Zuckerman0231f162016-02-23 13:41:13 +00005197
5198#define _mm256_ror_epi64( __A, __B) __extension__ ({ \
5199__builtin_ia32_prorq256_mask ((__v4di)( __A),( __B),\
5200 (__v4di)\
5201 _mm256_setzero_si256 (),\
5202 (__mmask8) -1);\
Michael Zuckerman4924c7a2016-02-23 14:23:53 +00005203})
Michael Zuckerman0231f162016-02-23 13:41:13 +00005204
5205#define _mm256_mask_ror_epi64( __W, __U, __A ,__B) __extension__ ({ \
5206__builtin_ia32_prorq256_mask ((__v4di) __A, __B,\
5207 (__v4di)( __W),\
5208 (__mmask8)( __U));\
Michael Zuckerman4924c7a2016-02-23 14:23:53 +00005209})
Michael Zuckerman0231f162016-02-23 13:41:13 +00005210
5211#define _mm256_maskz_ror_epi64( __U, __A, __B) __extension__ ({ \
5212__builtin_ia32_prorq256_mask ((__v4di)( __A),( __B),\
5213 (__v4di)\
5214 _mm256_setzero_si256 (),\
5215 (__mmask8)( __U));\
Michael Zuckerman4924c7a2016-02-23 14:23:53 +00005216})
Michael Zuckerman0231f162016-02-23 13:41:13 +00005217
Michael Zuckerman1ac360c2016-03-01 11:38:16 +00005218static __inline__ __m128i __DEFAULT_FN_ATTRS
5219_mm_mask_sll_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
5220 __m128i __B)
5221{
5222 return (__m128i) __builtin_ia32_pslld128_mask ((__v4si) __A,
5223 (__v4si) __B,
5224 (__v4si) __W,
5225 (__mmask8) __U);
5226}
5227
5228static __inline__ __m128i __DEFAULT_FN_ATTRS
5229_mm_maskz_sll_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
5230{
5231 return (__m128i) __builtin_ia32_pslld128_mask ((__v4si) __A,
5232 (__v4si) __B,
5233 (__v4si)
5234 _mm_setzero_si128 (),
5235 (__mmask8) __U);
5236}
5237
5238static __inline__ __m256i __DEFAULT_FN_ATTRS
5239_mm256_mask_sll_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
5240 __m128i __B)
5241{
5242 return (__m256i) __builtin_ia32_pslld256_mask ((__v8si) __A,
5243 (__v4si) __B,
5244 (__v8si) __W,
5245 (__mmask8) __U);
5246}
5247
5248static __inline__ __m256i __DEFAULT_FN_ATTRS
5249_mm256_maskz_sll_epi32 (__mmask8 __U, __m256i __A, __m128i __B)
5250{
5251 return (__m256i) __builtin_ia32_pslld256_mask ((__v8si) __A,
5252 (__v4si) __B,
5253 (__v8si)
5254 _mm256_setzero_si256 (),
5255 (__mmask8) __U);
5256}
5257
5258#define _mm_mask_slli_epi32( __W, __U, __A, __B) __extension__ ({ \
5259__builtin_ia32_pslldi128_mask ((__v4si)( __A),( __B),\
5260 (__v4si)( __W),\
5261 (__mmask8)( __U));\
5262})
5263
5264#define _mm_maskz_slli_epi32( __U, __A, __B) __extension__ ({ \
5265__builtin_ia32_pslldi128_mask ((__v4si)( __A),( __B),\
5266 (__v4si)\
5267 _mm_setzero_si128 (),\
5268 (__mmask8)( __U));\
5269})
5270
5271#define _mm256_mask_slli_epi32( __W, __U, __A ,__B) __extension__ ({ \
5272__builtin_ia32_pslldi256_mask ((__v8si) (__A), (__B),\
5273 (__v8si)( __W),\
5274 (__mmask8)( __U));\
5275})
5276
5277#define _mm256_maskz_slli_epi32( __U, __A, __B) __extension__ ({ \
5278__builtin_ia32_pslldi256_mask ((__v8si)( __A),( __B),\
5279 (__v8si)\
5280 _mm256_setzero_si256 (),\
5281 (__mmask8)( __U));\
5282})
5283
5284static __inline__ __m128i __DEFAULT_FN_ATTRS
5285_mm_mask_sll_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
5286 __m128i __B)
5287{
5288 return (__m128i) __builtin_ia32_psllq128_mask ((__v2di) __A,
5289 (__v2di) __B,
5290 (__v2di) __W,
5291 (__mmask8) __U);
5292}
5293
5294static __inline__ __m128i __DEFAULT_FN_ATTRS
5295_mm_maskz_sll_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
5296{
5297 return (__m128i) __builtin_ia32_psllq128_mask ((__v2di) __A,
5298 (__v2di) __B,
5299 (__v2di)
5300 _mm_setzero_di (),
5301 (__mmask8) __U);
5302}
5303
5304static __inline__ __m256i __DEFAULT_FN_ATTRS
5305_mm256_mask_sll_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
5306 __m128i __B)
5307{
5308 return (__m256i) __builtin_ia32_psllq256_mask ((__v4di) __A,
5309 (__v2di) __B,
5310 (__v4di) __W,
5311 (__mmask8) __U);
5312}
5313
5314static __inline__ __m256i __DEFAULT_FN_ATTRS
5315_mm256_maskz_sll_epi64 (__mmask8 __U, __m256i __A, __m128i __B)
5316{
5317 return (__m256i) __builtin_ia32_psllq256_mask ((__v4di) __A,
5318 (__v2di) __B,
5319 (__v4di)
5320 _mm256_setzero_si256 (),
5321 (__mmask8) __U);
5322}
5323
5324#define _mm_mask_slli_epi64( __W, __U, __A, __B) __extension__ ({ \
5325__builtin_ia32_psllqi128_mask ((__v2di)( __A),( __B),\
5326 (__v2di)( __W),\
5327 (__mmask8)( __U));\
5328})
5329
5330#define _mm_maskz_slli_epi64( __U, __A, __B) __extension__ ({ \
5331__builtin_ia32_psllqi128_mask ((__v2di)( __A),( __B),\
5332 (__v2di)\
5333 _mm_setzero_di (),\
5334 (__mmask8)( __U));\
5335})
5336
5337#define _mm256_mask_slli_epi64( __W, __U, __A ,__B) __extension__ ({ \
5338__builtin_ia32_psllqi256_mask ((__v4di) (__A), (__B),\
5339 (__v4di)( __W),\
5340 (__mmask8)( __U));\
5341})
5342
5343#define _mm256_maskz_slli_epi64( __U, __A, __B) __extension__ ({ \
5344__builtin_ia32_psllqi256_mask ((__v4di)( __A),( __B),\
5345 (__v4di)\
5346 _mm256_setzero_si256 (),\
5347 (__mmask8)( __U));\
5348})
5349
Michael Zuckerman0231f162016-02-23 13:41:13 +00005350
Michael Zuckermane98cc742016-02-23 15:59:47 +00005351static __inline__ __m128i __DEFAULT_FN_ATTRS
5352_mm_rorv_epi32 (__m128i __A, __m128i __B)
5353{
5354 return (__m128i) __builtin_ia32_prorvd128_mask ((__v4si) __A,
5355 (__v4si) __B,
5356 (__v4si)
5357 _mm_setzero_si128 (),
5358 (__mmask8) -1);
5359}
5360
5361static __inline__ __m128i __DEFAULT_FN_ATTRS
5362_mm_mask_rorv_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
5363 __m128i __B)
5364{
5365 return (__m128i) __builtin_ia32_prorvd128_mask ((__v4si) __A,
5366 (__v4si) __B,
5367 (__v4si) __W,
5368 (__mmask8) __U);
5369}
5370
5371static __inline__ __m128i __DEFAULT_FN_ATTRS
5372_mm_maskz_rorv_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
5373{
5374 return (__m128i) __builtin_ia32_prorvd128_mask ((__v4si) __A,
5375 (__v4si) __B,
5376 (__v4si)
5377 _mm_setzero_si128 (),
5378 (__mmask8) __U);
5379}
5380
5381static __inline__ __m256i __DEFAULT_FN_ATTRS
5382_mm256_rorv_epi32 (__m256i __A, __m256i __B)
5383{
5384 return (__m256i) __builtin_ia32_prorvd256_mask ((__v8si) __A,
5385 (__v8si) __B,
5386 (__v8si)
5387 _mm256_setzero_si256 (),
5388 (__mmask8) -1);
5389}
5390
5391static __inline__ __m256i __DEFAULT_FN_ATTRS
5392_mm256_mask_rorv_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
5393 __m256i __B)
5394{
5395 return (__m256i) __builtin_ia32_prorvd256_mask ((__v8si) __A,
5396 (__v8si) __B,
5397 (__v8si) __W,
5398 (__mmask8) __U);
5399}
5400
5401static __inline__ __m256i __DEFAULT_FN_ATTRS
5402_mm256_maskz_rorv_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
5403{
5404 return (__m256i) __builtin_ia32_prorvd256_mask ((__v8si) __A,
5405 (__v8si) __B,
5406 (__v8si)
5407 _mm256_setzero_si256 (),
5408 (__mmask8) __U);
5409}
5410
5411static __inline__ __m128i __DEFAULT_FN_ATTRS
5412_mm_rorv_epi64 (__m128i __A, __m128i __B)
5413{
5414 return (__m128i) __builtin_ia32_prorvq128_mask ((__v2di) __A,
5415 (__v2di) __B,
5416 (__v2di)
5417 _mm_setzero_di (),
5418 (__mmask8) -1);
5419}
5420
5421static __inline__ __m128i __DEFAULT_FN_ATTRS
5422_mm_mask_rorv_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
5423 __m128i __B)
5424{
5425 return (__m128i) __builtin_ia32_prorvq128_mask ((__v2di) __A,
5426 (__v2di) __B,
5427 (__v2di) __W,
5428 (__mmask8) __U);
5429}
5430
5431static __inline__ __m128i __DEFAULT_FN_ATTRS
5432_mm_maskz_rorv_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
5433{
5434 return (__m128i) __builtin_ia32_prorvq128_mask ((__v2di) __A,
5435 (__v2di) __B,
5436 (__v2di)
5437 _mm_setzero_di (),
5438 (__mmask8) __U);
5439}
5440
5441static __inline__ __m256i __DEFAULT_FN_ATTRS
5442_mm256_rorv_epi64 (__m256i __A, __m256i __B)
5443{
5444 return (__m256i) __builtin_ia32_prorvq256_mask ((__v4di) __A,
5445 (__v4di) __B,
5446 (__v4di)
5447 _mm256_setzero_si256 (),
5448 (__mmask8) -1);
5449}
5450
5451static __inline__ __m256i __DEFAULT_FN_ATTRS
5452_mm256_mask_rorv_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
5453 __m256i __B)
5454{
5455 return (__m256i) __builtin_ia32_prorvq256_mask ((__v4di) __A,
5456 (__v4di) __B,
5457 (__v4di) __W,
5458 (__mmask8) __U);
5459}
5460
5461static __inline__ __m256i __DEFAULT_FN_ATTRS
5462_mm256_maskz_rorv_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
5463{
5464 return (__m256i) __builtin_ia32_prorvq256_mask ((__v4di) __A,
5465 (__v4di) __B,
5466 (__v4di)
5467 _mm256_setzero_si256 (),
5468 (__mmask8) __U);
5469}
5470
Michael Zuckerman431b0e12016-02-28 07:39:34 +00005471static __inline__ __m128i __DEFAULT_FN_ATTRS
5472_mm_mask_sllv_epi64 (__m128i __W, __mmask8 __U, __m128i __X,
5473 __m128i __Y)
5474{
5475 return (__m128i) __builtin_ia32_psllv2di_mask ((__v2di) __X,
5476 (__v2di) __Y,
5477 (__v2di) __W,
5478 (__mmask8) __U);
5479}
5480
5481static __inline__ __m128i __DEFAULT_FN_ATTRS
5482_mm_maskz_sllv_epi64 (__mmask8 __U, __m128i __X, __m128i __Y)
5483{
5484 return (__m128i) __builtin_ia32_psllv2di_mask ((__v2di) __X,
5485 (__v2di) __Y,
5486 (__v2di)
5487 _mm_setzero_di (),
5488 (__mmask8) __U);
5489}
5490
5491static __inline__ __m256i __DEFAULT_FN_ATTRS
5492_mm256_mask_sllv_epi64 (__m256i __W, __mmask8 __U, __m256i __X,
5493 __m256i __Y)
5494{
5495 return (__m256i) __builtin_ia32_psllv4di_mask ((__v4di) __X,
5496 (__v4di) __Y,
5497 (__v4di) __W,
5498 (__mmask8) __U);
5499}
5500
5501static __inline__ __m256i __DEFAULT_FN_ATTRS
5502_mm256_maskz_sllv_epi64 (__mmask8 __U, __m256i __X, __m256i __Y)
5503{
5504 return (__m256i) __builtin_ia32_psllv4di_mask ((__v4di) __X,
5505 (__v4di) __Y,
5506 (__v4di)
5507 _mm256_setzero_si256 (),
5508 (__mmask8) __U);
5509}
5510
5511static __inline__ __m128i __DEFAULT_FN_ATTRS
5512_mm_mask_sllv_epi32 (__m128i __W, __mmask8 __U, __m128i __X,
5513 __m128i __Y)
5514{
5515 return (__m128i) __builtin_ia32_psllv4si_mask ((__v4si) __X,
5516 (__v4si) __Y,
5517 (__v4si) __W,
5518 (__mmask8) __U);
5519}
5520
5521static __inline__ __m128i __DEFAULT_FN_ATTRS
5522_mm_maskz_sllv_epi32 (__mmask8 __U, __m128i __X, __m128i __Y)
5523{
5524 return (__m128i) __builtin_ia32_psllv4si_mask ((__v4si) __X,
5525 (__v4si) __Y,
5526 (__v4si)
5527 _mm_setzero_si128 (),
5528 (__mmask8) __U);
5529}
5530
5531static __inline__ __m256i __DEFAULT_FN_ATTRS
5532_mm256_mask_sllv_epi32 (__m256i __W, __mmask8 __U, __m256i __X,
5533 __m256i __Y)
5534{
5535 return (__m256i) __builtin_ia32_psllv8si_mask ((__v8si) __X,
5536 (__v8si) __Y,
5537 (__v8si) __W,
5538 (__mmask8) __U);
5539}
5540
5541static __inline__ __m256i __DEFAULT_FN_ATTRS
5542_mm256_maskz_sllv_epi32 (__mmask8 __U, __m256i __X, __m256i __Y)
5543{
5544 return (__m256i) __builtin_ia32_psllv8si_mask ((__v8si) __X,
5545 (__v8si) __Y,
5546 (__v8si)
5547 _mm256_setzero_si256 (),
5548 (__mmask8) __U);
5549}
5550
Michael Zuckerman0165e762016-03-01 13:03:45 +00005551
5552
5553static __inline__ __m128i __DEFAULT_FN_ATTRS
5554_mm_mask_srlv_epi64 (__m128i __W, __mmask8 __U, __m128i __X,
5555 __m128i __Y)
5556{
5557 return (__m128i) __builtin_ia32_psrlv2di_mask ((__v2di) __X,
5558 (__v2di) __Y,
5559 (__v2di) __W,
5560 (__mmask8) __U);
5561}
5562
5563static __inline__ __m128i __DEFAULT_FN_ATTRS
5564_mm_maskz_srlv_epi64 (__mmask8 __U, __m128i __X, __m128i __Y)
5565{
5566 return (__m128i) __builtin_ia32_psrlv2di_mask ((__v2di) __X,
5567 (__v2di) __Y,
5568 (__v2di)
5569 _mm_setzero_di (),
5570 (__mmask8) __U);
5571}
5572
5573static __inline__ __m256i __DEFAULT_FN_ATTRS
5574_mm256_mask_srlv_epi64 (__m256i __W, __mmask8 __U, __m256i __X,
5575 __m256i __Y)
5576{
5577 return (__m256i) __builtin_ia32_psrlv4di_mask ((__v4di) __X,
5578 (__v4di) __Y,
5579 (__v4di) __W,
5580 (__mmask8) __U);
5581}
5582
5583static __inline__ __m256i __DEFAULT_FN_ATTRS
5584_mm256_maskz_srlv_epi64 (__mmask8 __U, __m256i __X, __m256i __Y)
5585{
5586 return (__m256i) __builtin_ia32_psrlv4di_mask ((__v4di) __X,
5587 (__v4di) __Y,
5588 (__v4di)
5589 _mm256_setzero_si256 (),
5590 (__mmask8) __U);
5591}
5592
5593static __inline__ __m128i __DEFAULT_FN_ATTRS
5594_mm_mask_srlv_epi32 (__m128i __W, __mmask8 __U, __m128i __X,
5595 __m128i __Y)
5596{
5597 return (__m128i) __builtin_ia32_psrlv4si_mask ((__v4si) __X,
5598 (__v4si) __Y,
5599 (__v4si) __W,
5600 (__mmask8) __U);
5601}
5602
5603static __inline__ __m128i __DEFAULT_FN_ATTRS
5604_mm_maskz_srlv_epi32 (__mmask8 __U, __m128i __X, __m128i __Y)
5605{
5606 return (__m128i) __builtin_ia32_psrlv4si_mask ((__v4si) __X,
5607 (__v4si) __Y,
5608 (__v4si)
5609 _mm_setzero_si128 (),
5610 (__mmask8) __U);
5611}
5612
5613static __inline__ __m256i __DEFAULT_FN_ATTRS
5614_mm256_mask_srlv_epi32 (__m256i __W, __mmask8 __U, __m256i __X,
5615 __m256i __Y)
5616{
5617 return (__m256i) __builtin_ia32_psrlv8si_mask ((__v8si) __X,
5618 (__v8si) __Y,
5619 (__v8si) __W,
5620 (__mmask8) __U);
5621}
5622
5623static __inline__ __m256i __DEFAULT_FN_ATTRS
5624_mm256_maskz_srlv_epi32 (__mmask8 __U, __m256i __X, __m256i __Y)
5625{
5626 return (__m256i) __builtin_ia32_psrlv8si_mask ((__v8si) __X,
5627 (__v8si) __Y,
5628 (__v8si)
5629 _mm256_setzero_si256 (),
5630 (__mmask8) __U);
5631}
5632
Michael Zuckermand176d742016-03-01 17:49:03 +00005633
5634
5635static __inline__ __m128i __DEFAULT_FN_ATTRS
5636_mm_mask_srl_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
5637 __m128i __B)
5638{
5639 return (__m128i) __builtin_ia32_psrld128_mask ((__v4si) __A,
5640 (__v4si) __B,
5641 (__v4si) __W,
5642 (__mmask8) __U);
5643}
5644
5645static __inline__ __m128i __DEFAULT_FN_ATTRS
5646_mm_maskz_srl_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
5647{
5648 return (__m128i) __builtin_ia32_psrld128_mask ((__v4si) __A,
5649 (__v4si) __B,
5650 (__v4si)
5651 _mm_setzero_si128 (),
5652 (__mmask8) __U);
5653}
5654
5655static __inline__ __m256i __DEFAULT_FN_ATTRS
5656_mm256_mask_srl_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
5657 __m128i __B)
5658{
5659 return (__m256i) __builtin_ia32_psrld256_mask ((__v8si) __A,
5660 (__v4si) __B,
5661 (__v8si) __W,
5662 (__mmask8) __U);
5663}
5664
5665static __inline__ __m256i __DEFAULT_FN_ATTRS
5666_mm256_maskz_srl_epi32 (__mmask8 __U, __m256i __A, __m128i __B)
5667{
5668 return (__m256i) __builtin_ia32_psrld256_mask ((__v8si) __A,
5669 (__v4si) __B,
5670 (__v8si)
5671 _mm256_setzero_si256 (),
5672 (__mmask8) __U);
5673}
5674
5675#define _mm_mask_srli_epi32( __W, __U, __A, __imm) __extension__ ({ \
5676__builtin_ia32_psrldi128_mask ((__v4si)( __A),( __imm),\
5677 (__v4si)( __W),\
5678 (__mmask8)( __U));\
5679})
5680
5681#define _mm_maskz_srli_epi32( __U, __A, __imm) __extension__ ({ \
5682__builtin_ia32_psrldi128_mask ((__v4si)( __A),( __imm),\
5683 (__v4si)\
5684 _mm_setzero_si128 (),\
5685 (__mmask8)( __U));\
5686})
5687
5688#define _mm256_mask_srli_epi32( __W, __U, __A, __imm) __extension__ ({ \
5689__builtin_ia32_psrldi256_mask ((__v8si)( __A),( __imm),\
5690 (__v8si)( __W),\
5691 (__mmask8)( __U));\
5692})
5693
5694#define _mm256_maskz_srli_epi32( __U, __A, __imm) __extension__ ({ \
5695__builtin_ia32_psrldi256_mask ((__v8si)( __A),( __imm),\
5696 (__v8si)\
5697 _mm256_setzero_si256 (),\
5698 (__mmask8)( __U));\
5699})
5700
5701static __inline__ __m128i __DEFAULT_FN_ATTRS
5702_mm_mask_srl_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
5703 __m128i __B)
5704{
5705 return (__m128i) __builtin_ia32_psrlq128_mask ((__v2di) __A,
5706 (__v2di) __B,
5707 (__v2di) __W,
5708 (__mmask8) __U);
5709}
5710
5711static __inline__ __m128i __DEFAULT_FN_ATTRS
5712_mm_maskz_srl_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
5713{
5714 return (__m128i) __builtin_ia32_psrlq128_mask ((__v2di) __A,
5715 (__v2di) __B,
5716 (__v2di)
5717 _mm_setzero_di (),
5718 (__mmask8) __U);
5719}
5720
5721static __inline__ __m256i __DEFAULT_FN_ATTRS
5722_mm256_mask_srl_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
5723 __m128i __B)
5724{
5725 return (__m256i) __builtin_ia32_psrlq256_mask ((__v4di) __A,
5726 (__v2di) __B,
5727 (__v4di) __W,
5728 (__mmask8) __U);
5729}
5730
5731static __inline__ __m256i __DEFAULT_FN_ATTRS
5732_mm256_maskz_srl_epi64 (__mmask8 __U, __m256i __A, __m128i __B)
5733{
5734 return (__m256i) __builtin_ia32_psrlq256_mask ((__v4di) __A,
5735 (__v2di) __B,
5736 (__v4di)
5737 _mm256_setzero_si256 (),
5738 (__mmask8) __U);
5739}
5740
5741#define _mm_mask_srli_epi64( __W, __U, __A, __imm) __extension__ ({ \
5742__builtin_ia32_psrlqi128_mask ((__v2di)( __A),( __imm),\
5743 (__v2di)( __W),\
5744 (__mmask8)( __U));\
5745})
5746
5747#define _mm_maskz_srli_epi64( __U, __A, __imm) __extension__ ({ \
5748__builtin_ia32_psrlqi128_mask ((__v2di)( __A),( __imm),\
5749 (__v2di)\
5750 _mm_setzero_si128 (),\
5751 (__mmask8)( __U));\
5752})
5753
5754#define _mm256_mask_srli_epi64( __W, __U, __A, __imm) __extension__ ({ \
5755__builtin_ia32_psrlqi256_mask ((__v4di)( __A),( __imm),\
5756 (__v4di)( __W),\
5757 (__mmask8)( __U));\
5758})
5759
5760#define _mm256_maskz_srli_epi64( __U, __A, __imm) __extension__ ({ \
5761__builtin_ia32_psrlqi256_mask ((__v4di)( __A),( __imm),\
5762 (__v4di)\
5763 _mm256_setzero_si256 (),\
5764 (__mmask8)( __U));\
5765})
5766
Michael Zuckermand15c95a2016-03-02 09:05:46 +00005767static __inline__ __m128i __DEFAULT_FN_ATTRS
5768_mm_mask_srav_epi32 (__m128i __W, __mmask8 __U, __m128i __X,
5769 __m128i __Y)
5770{
5771 return (__m128i) __builtin_ia32_psrav4si_mask ((__v4si) __X,
5772 (__v4si) __Y,
5773 (__v4si) __W,
5774 (__mmask8) __U);
5775}
5776
5777static __inline__ __m128i __DEFAULT_FN_ATTRS
5778_mm_maskz_srav_epi32 (__mmask8 __U, __m128i __X, __m128i __Y)
5779{
5780 return (__m128i) __builtin_ia32_psrav4si_mask ((__v4si) __X,
5781 (__v4si) __Y,
5782 (__v4si)
5783 _mm_setzero_si128 (),
5784 (__mmask8) __U);
5785}
5786
5787static __inline__ __m256i __DEFAULT_FN_ATTRS
5788_mm256_mask_srav_epi32 (__m256i __W, __mmask8 __U, __m256i __X,
5789 __m256i __Y)
5790{
5791 return (__m256i) __builtin_ia32_psrav8si_mask ((__v8si) __X,
5792 (__v8si) __Y,
5793 (__v8si) __W,
5794 (__mmask8) __U);
5795}
5796
5797static __inline__ __m256i __DEFAULT_FN_ATTRS
5798_mm256_maskz_srav_epi32 (__mmask8 __U, __m256i __X, __m256i __Y)
5799{
5800 return (__m256i) __builtin_ia32_psrav8si_mask ((__v8si) __X,
5801 (__v8si) __Y,
5802 (__v8si)
5803 _mm256_setzero_si256 (),
5804 (__mmask8) __U);
5805}
5806
5807static __inline__ __m128i __DEFAULT_FN_ATTRS
5808_mm_srav_epi64 (__m128i __X, __m128i __Y)
5809{
5810 return (__m128i) __builtin_ia32_psravq128_mask ((__v2di) __X,
5811 (__v2di) __Y,
5812 (__v2di)
5813 _mm_setzero_di (),
5814 (__mmask8) -1);
5815}
5816
5817static __inline__ __m128i __DEFAULT_FN_ATTRS
5818_mm_mask_srav_epi64 (__m128i __W, __mmask8 __U, __m128i __X,
5819 __m128i __Y)
5820{
5821 return (__m128i) __builtin_ia32_psravq128_mask ((__v2di) __X,
5822 (__v2di) __Y,
5823 (__v2di) __W,
5824 (__mmask8) __U);
5825}
5826
5827static __inline__ __m128i __DEFAULT_FN_ATTRS
5828_mm_maskz_srav_epi64 (__mmask8 __U, __m128i __X, __m128i __Y)
5829{
5830 return (__m128i) __builtin_ia32_psravq128_mask ((__v2di) __X,
5831 (__v2di) __Y,
5832 (__v2di)
5833 _mm_setzero_di (),
5834 (__mmask8) __U);
5835}
5836
5837static __inline__ __m256i __DEFAULT_FN_ATTRS
5838_mm256_srav_epi64 (__m256i __X, __m256i __Y)
5839{
5840 return (__m256i) __builtin_ia32_psravq256_mask ((__v4di) __X,
5841 (__v4di) __Y,
5842 (__v4di)
5843 _mm256_setzero_si256 (),
5844 (__mmask8) -1);
5845}
5846
5847static __inline__ __m256i __DEFAULT_FN_ATTRS
5848_mm256_mask_srav_epi64 (__m256i __W, __mmask8 __U, __m256i __X,
5849 __m256i __Y)
5850{
5851 return (__m256i) __builtin_ia32_psravq256_mask ((__v4di) __X,
5852 (__v4di) __Y,
5853 (__v4di) __W,
5854 (__mmask8) __U);
5855}
5856
5857static __inline__ __m256i __DEFAULT_FN_ATTRS
5858_mm256_maskz_srav_epi64 (__mmask8 __U, __m256i __X, __m256i __Y)
5859{
5860 return (__m256i) __builtin_ia32_psravq256_mask ((__v4di) __X,
5861 (__v4di) __Y,
5862 (__v4di)
5863 _mm256_setzero_si256 (),
5864 (__mmask8) __U);
5865}
Michael Zuckermanffbb67a2016-03-03 09:26:01 +00005866
5867
5868
5869static __inline__ void __DEFAULT_FN_ATTRS
5870_mm_mask_store_epi32 (void *__P, __mmask8 __U, __m128i __A)
5871{
5872 __builtin_ia32_movdqa32store128_mask ((__v4si *) __P,
5873 (__v4si) __A,
5874 (__mmask8) __U);
5875}
5876
5877static __inline__ void __DEFAULT_FN_ATTRS
5878_mm256_mask_store_epi32 (void *__P, __mmask8 __U, __m256i __A)
5879{
5880 __builtin_ia32_movdqa32store256_mask ((__v8si *) __P,
5881 (__v8si) __A,
5882 (__mmask8) __U);
5883}
5884
5885static __inline__ __m128i __DEFAULT_FN_ATTRS
5886_mm_mask_mov_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
5887{
5888 return (__m128i) __builtin_ia32_movdqa64_128_mask ((__v2di) __A,
5889 (__v2di) __W,
5890 (__mmask8) __U);
5891}
5892
5893static __inline__ __m128i __DEFAULT_FN_ATTRS
5894_mm_maskz_mov_epi64 (__mmask8 __U, __m128i __A)
5895{
5896 return (__m128i) __builtin_ia32_movdqa64_128_mask ((__v2di) __A,
5897 (__v2di)
5898 _mm_setzero_di (),
5899 (__mmask8) __U);
5900}
5901
5902static __inline__ __m256i __DEFAULT_FN_ATTRS
5903_mm256_mask_mov_epi64 (__m256i __W, __mmask8 __U, __m256i __A)
5904{
5905 return (__m256i) __builtin_ia32_movdqa64_256_mask ((__v4di) __A,
5906 (__v4di) __W,
5907 (__mmask8) __U);
5908}
5909
5910static __inline__ __m256i __DEFAULT_FN_ATTRS
5911_mm256_maskz_mov_epi64 (__mmask8 __U, __m256i __A)
5912{
5913 return (__m256i) __builtin_ia32_movdqa64_256_mask ((__v4di) __A,
5914 (__v4di)
5915 _mm256_setzero_si256 (),
5916 (__mmask8) __U);
5917}
5918
5919static __inline__ __m128i __DEFAULT_FN_ATTRS
5920_mm_mask_load_epi64 (__m128i __W, __mmask8 __U, void const *__P)
5921{
5922 return (__m128i) __builtin_ia32_movdqa64load128_mask ((__v2di *) __P,
5923 (__v2di) __W,
5924 (__mmask8)
5925 __U);
5926}
5927
5928static __inline__ __m128i __DEFAULT_FN_ATTRS
5929_mm_maskz_load_epi64 (__mmask8 __U, void const *__P)
5930{
5931 return (__m128i) __builtin_ia32_movdqa64load128_mask ((__v2di *) __P,
5932 (__v2di)
5933 _mm_setzero_di (),
5934 (__mmask8)
5935 __U);
5936}
5937
5938static __inline__ __m256i __DEFAULT_FN_ATTRS
5939_mm256_mask_load_epi64 (__m256i __W, __mmask8 __U, void const *__P)
5940{
5941 return (__m256i) __builtin_ia32_movdqa64load256_mask ((__v4di *) __P,
5942 (__v4di) __W,
5943 (__mmask8)
5944 __U);
5945}
5946
5947static __inline__ __m256i __DEFAULT_FN_ATTRS
5948_mm256_maskz_load_epi64 (__mmask8 __U, void const *__P)
5949{
5950 return (__m256i) __builtin_ia32_movdqa64load256_mask ((__v4di *) __P,
5951 (__v4di)
5952 _mm256_setzero_si256 (),
5953 (__mmask8)
5954 __U);
5955}
5956
5957static __inline__ void __DEFAULT_FN_ATTRS
5958_mm_mask_store_epi64 (void *__P, __mmask8 __U, __m128i __A)
5959{
5960 __builtin_ia32_movdqa64store128_mask ((__v2di *) __P,
5961 (__v2di) __A,
5962 (__mmask8) __U);
5963}
5964
5965static __inline__ void __DEFAULT_FN_ATTRS
5966_mm256_mask_store_epi64 (void *__P, __mmask8 __U, __m256i __A)
5967{
5968 __builtin_ia32_movdqa64store256_mask ((__v4di *) __P,
5969 (__v4di) __A,
5970 (__mmask8) __U);
5971}
5972
Michael Zuckerman0d67e4b2016-03-03 13:43:05 +00005973static __inline__ __m128d __DEFAULT_FN_ATTRS
5974_mm_mask_movedup_pd (__m128d __W, __mmask8 __U, __m128d __A)
5975{
5976 return (__m128d) __builtin_ia32_movddup128_mask ((__v2df) __A,
5977 (__v2df) __W,
5978 (__mmask8) __U);
5979}
5980
5981static __inline__ __m128d __DEFAULT_FN_ATTRS
5982_mm_maskz_movedup_pd (__mmask8 __U, __m128d __A)
5983{
5984 return (__m128d) __builtin_ia32_movddup128_mask ((__v2df) __A,
5985 (__v2df)
5986 _mm_setzero_pd (),
5987 (__mmask8) __U);
5988}
5989
5990static __inline__ __m256d __DEFAULT_FN_ATTRS
5991_mm256_mask_movedup_pd (__m256d __W, __mmask8 __U, __m256d __A)
5992{
5993 return (__m256d) __builtin_ia32_movddup256_mask ((__v4df) __A,
5994 (__v4df) __W,
5995 (__mmask8) __U);
5996}
5997
5998static __inline__ __m256d __DEFAULT_FN_ATTRS
5999_mm256_maskz_movedup_pd (__mmask8 __U, __m256d __A)
6000{
6001 return (__m256d) __builtin_ia32_movddup256_mask ((__v4df) __A,
6002 (__v4df)
6003 _mm256_setzero_pd (),
6004 (__mmask8) __U);
6005}
Michael Zuckermanffbb67a2016-03-03 09:26:01 +00006006
Michael Zuckerman912be162016-03-07 08:29:10 +00006007
6008#define _mm_mask_set1_epi32( __O, __M, __A) __extension__ ({ \
6009__builtin_ia32_pbroadcastd128_gpr_mask (__A, (__v4si)( __O),\
6010 ( __M));\
6011})
6012
6013#define _mm_maskz_set1_epi32( __M, __A) __extension__ ({ \
6014__builtin_ia32_pbroadcastd128_gpr_mask (__A,\
6015 (__v4si)\
6016 _mm_setzero_si128 (),\
6017 ( __M));\
6018})
6019
6020#define _mm256_mask_set1_epi32( __O, __M, __A) __extension__ ({ \
6021__builtin_ia32_pbroadcastd256_gpr_mask (__A, (__v8si)( __O),\
6022 ( __M));\
6023})
6024
6025#define _mm256_maskz_set1_epi32( __M, __A) __extension__ ({ \
6026__builtin_ia32_pbroadcastd256_gpr_mask (__A,\
6027 (__v8si)\
6028 _mm256_setzero_si256 (),\
6029 ( __M));\
6030})
6031
6032static __inline__ __m128i __DEFAULT_FN_ATTRS
6033_mm_mask_set1_epi64 (__m128i __O, __mmask8 __M, long long __A)
6034{
6035 return (__m128i) __builtin_ia32_pbroadcastq128_gpr_mask (__A, (__v2di) __O,
6036 __M);
6037}
6038
6039static __inline__ __m128i __DEFAULT_FN_ATTRS
6040_mm_maskz_set1_epi64 (__mmask8 __M, long long __A)
6041{
6042 return (__m128i) __builtin_ia32_pbroadcastq128_gpr_mask (__A,
6043 (__v2di)
6044 _mm_setzero_si128 (),
6045 __M);
6046}
6047
6048static __inline__ __m256i __DEFAULT_FN_ATTRS
6049_mm256_mask_set1_epi64 (__m256i __O, __mmask8 __M, long long __A)
6050{
6051 return (__m256i) __builtin_ia32_pbroadcastq256_gpr_mask (__A, (__v4di) __O,
6052 __M);
6053}
6054
6055static __inline__ __m256i __DEFAULT_FN_ATTRS
6056_mm256_maskz_set1_epi64 (__mmask8 __M, long long __A)
6057{
6058 return (__m256i) __builtin_ia32_pbroadcastq256_gpr_mask (__A,
6059 (__v4di)
6060 _mm256_setzero_si256 (),
6061 __M);
6062}
Michael Zuckermandef78752016-03-28 12:23:09 +00006063
6064#define _mm_fixupimm_pd( __A, __B, __C, __imm) __extension__ ({ \
6065__builtin_ia32_fixupimmpd128_mask ((__v2df)( __A),\
6066 (__v2df)( __B),\
6067 (__v2di)( __C),\
6068 ( __imm),\
6069 (__mmask8) -1);\
6070})
6071
6072#define _mm_mask_fixupimm_pd( __A, __U, __B, __C, __imm) __extension__ ({ \
6073__builtin_ia32_fixupimmpd128_mask ((__v2df)( __A),\
6074 (__v2df)( __B),\
6075 (__v2di)( __C),\
6076 ( __imm),\
6077 (__mmask8)( __U));\
6078})
6079
6080#define _mm_maskz_fixupimm_pd( __U, __A, __B, __C, __imm) __extension__ ({ \
6081__builtin_ia32_fixupimmpd128_maskz ((__v2df)( __A),\
6082 (__v2df)( __B),\
6083 (__v2di)( __C),\
6084 ( __imm),\
6085 (__mmask8)( __U));\
6086})
6087
6088#define _mm256_fixupimm_pd( __A, __B, __C, __imm) __extension__ ({ \
6089__builtin_ia32_fixupimmpd256_mask ((__v4df)( __A),\
6090 (__v4df)( __B),\
6091 (__v4di)( __C),\
6092 ( __imm),\
6093 (__mmask8) -1);\
6094})
6095
6096#define _mm256_mask_fixupimm_pd( __A, __U, __B, __C, __imm) __extension__ ({ \
6097__builtin_ia32_fixupimmpd256_mask ((__v4df)( __A),\
6098 (__v4df)( __B),\
6099 (__v4di)( __C),\
6100 ( __imm),\
6101 (__mmask8)( __U));\
6102})
6103
6104#define _mm256_maskz_fixupimm_pd( __U, __A, __B, __C, __imm) __extension__ ({ \
6105__builtin_ia32_fixupimmpd256_maskz ((__v4df)( __A),\
6106 (__v4df)( __B),\
6107 (__v4di)( __C),\
6108 ( __imm),\
6109 (__mmask8)( __U));\
6110})
6111
6112#define _mm_fixupimm_ps( __A, __B, __C, __imm) __extension__ ({ \
6113__builtin_ia32_fixupimmps128_mask ((__v4sf)( __A),\
6114 (__v4sf)( __B),\
6115 (__v4si)( __C),\
6116 ( __imm),\
6117 (__mmask8) -1);\
6118})
6119
6120#define _mm_mask_fixupimm_ps( __A, __U, __B, __C, __imm) __extension__ ({ \
6121__builtin_ia32_fixupimmps128_mask ((__v4sf)( __A),\
6122 (__v4sf)( __B),\
6123 (__v4si)( __C),\
6124 ( __imm),\
6125 (__mmask8)( __U));\
6126})
6127
6128#define _mm_maskz_fixupimm_ps( __U, __A, __B, __C, __imm) __extension__ ({ \
6129__builtin_ia32_fixupimmps128_maskz ((__v4sf)( __A),\
6130 (__v4sf)( __B),\
6131 (__v4si)( __C),\
6132 ( __imm),\
6133 (__mmask8)( __U));\
6134})
6135
6136#define _mm256_fixupimm_ps( __A, __B, __C, __imm) __extension__ ({ \
6137__builtin_ia32_fixupimmps256_mask ((__v8sf)( __A),\
6138 (__v8sf)( __B),\
6139 (__v8si)( __C),\
6140 ( __imm),\
6141 (__mmask8) -1);\
6142})
6143
6144#define _mm256_mask_fixupimm_ps( __A, __U, __B, __C, __imm) __extension__ ({ \
6145__builtin_ia32_fixupimmps256_mask ((__v8sf)( __A),\
6146 (__v8sf)( __B),\
6147 (__v8si)( __C),\
6148 ( __imm),\
6149 (__mmask8)( __U));\
6150})
6151
6152#define _mm256_maskz_fixupimm_ps( __U, __A, __B, __C, __imm) __extension__ ({ \
6153__builtin_ia32_fixupimmps256_maskz ((__v8sf)( __A),\
6154 (__v8sf)( __B),\
6155 (__v8si)( __C),\
6156 ( __imm),\
6157 (__mmask8)( __U));\
6158})
6159
6160static __inline__ __m128d __DEFAULT_FN_ATTRS
6161_mm_mask_load_pd (__m128d __W, __mmask8 __U, void const *__P)
6162{
6163 return (__m128d) __builtin_ia32_loadapd128_mask ((__v2df *) __P,
6164 (__v2df) __W,
6165 (__mmask8) __U);
6166}
6167
6168static __inline__ __m128d __DEFAULT_FN_ATTRS
6169_mm_maskz_load_pd (__mmask8 __U, void const *__P)
6170{
6171 return (__m128d) __builtin_ia32_loadapd128_mask ((__v2df *) __P,
6172 (__v2df)
6173 _mm_setzero_pd (),
6174 (__mmask8) __U);
6175}
6176
6177static __inline__ __m256d __DEFAULT_FN_ATTRS
6178_mm256_mask_load_pd (__m256d __W, __mmask8 __U, void const *__P)
6179{
6180 return (__m256d) __builtin_ia32_loadapd256_mask ((__v4df *) __P,
6181 (__v4df) __W,
6182 (__mmask8) __U);
6183}
6184
6185static __inline__ __m256d __DEFAULT_FN_ATTRS
6186_mm256_maskz_load_pd (__mmask8 __U, void const *__P)
6187{
6188 return (__m256d) __builtin_ia32_loadapd256_mask ((__v4df *) __P,
6189 (__v4df)
6190 _mm256_setzero_pd (),
6191 (__mmask8) __U);
6192}
6193
6194static __inline__ __m128 __DEFAULT_FN_ATTRS
6195_mm_mask_load_ps (__m128 __W, __mmask8 __U, void const *__P)
6196{
6197 return (__m128) __builtin_ia32_loadaps128_mask ((__v4sf *) __P,
6198 (__v4sf) __W,
6199 (__mmask8) __U);
6200}
6201
6202static __inline__ __m128 __DEFAULT_FN_ATTRS
6203_mm_maskz_load_ps (__mmask8 __U, void const *__P)
6204{
6205 return (__m128) __builtin_ia32_loadaps128_mask ((__v4sf *) __P,
6206 (__v4sf)
6207 _mm_setzero_ps (),
6208 (__mmask8) __U);
6209}
6210
6211static __inline__ __m256 __DEFAULT_FN_ATTRS
6212_mm256_mask_load_ps (__m256 __W, __mmask8 __U, void const *__P)
6213{
6214 return (__m256) __builtin_ia32_loadaps256_mask ((__v8sf *) __P,
6215 (__v8sf) __W,
6216 (__mmask8) __U);
6217}
6218
6219static __inline__ __m256 __DEFAULT_FN_ATTRS
6220_mm256_maskz_load_ps (__mmask8 __U, void const *__P)
6221{
6222 return (__m256) __builtin_ia32_loadaps256_mask ((__v8sf *) __P,
6223 (__v8sf)
6224 _mm256_setzero_ps (),
6225 (__mmask8) __U);
6226}
6227
6228static __inline__ __m128i __DEFAULT_FN_ATTRS
6229_mm_mask_loadu_epi64 (__m128i __W, __mmask8 __U, void const *__P)
6230{
6231 return (__m128i) __builtin_ia32_loaddqudi128_mask ((__v2di *) __P,
6232 (__v2di) __W,
6233 (__mmask8) __U);
6234}
6235
6236static __inline__ __m128i __DEFAULT_FN_ATTRS
6237_mm_maskz_loadu_epi64 (__mmask8 __U, void const *__P)
6238{
6239 return (__m128i) __builtin_ia32_loaddqudi128_mask ((__v2di *) __P,
6240 (__v2di)
6241 _mm_setzero_si128 (),
6242 (__mmask8) __U);
6243}
6244
6245static __inline__ __m256i __DEFAULT_FN_ATTRS
6246_mm256_mask_loadu_epi64 (__m256i __W, __mmask8 __U, void const *__P)
6247{
6248 return (__m256i) __builtin_ia32_loaddqudi256_mask ((__v4di *) __P,
6249 (__v4di) __W,
6250 (__mmask8) __U);
6251}
6252
6253static __inline__ __m256i __DEFAULT_FN_ATTRS
6254_mm256_maskz_loadu_epi64 (__mmask8 __U, void const *__P)
6255{
6256 return (__m256i) __builtin_ia32_loaddqudi256_mask ((__v4di *) __P,
6257 (__v4di)
6258 _mm256_setzero_si256 (),
6259 (__mmask8) __U);
6260}
6261
6262static __inline__ __m128i __DEFAULT_FN_ATTRS
6263_mm_mask_loadu_epi32 (__m128i __W, __mmask8 __U, void const *__P)
6264{
6265 return (__m128i) __builtin_ia32_loaddqusi128_mask ((__v4si *) __P,
6266 (__v4si) __W,
6267 (__mmask8) __U);
6268}
6269
6270static __inline__ __m128i __DEFAULT_FN_ATTRS
6271_mm_maskz_loadu_epi32 (__mmask8 __U, void const *__P)
6272{
6273 return (__m128i) __builtin_ia32_loaddqusi128_mask ((__v4si *) __P,
6274 (__v4si)
6275 _mm_setzero_si128 (),
6276 (__mmask8) __U);
6277}
6278
6279static __inline__ __m256i __DEFAULT_FN_ATTRS
6280_mm256_mask_loadu_epi32 (__m256i __W, __mmask8 __U, void const *__P)
6281{
6282 return (__m256i) __builtin_ia32_loaddqusi256_mask ((__v8si *) __P,
6283 (__v8si) __W,
6284 (__mmask8) __U);
6285}
6286
6287static __inline__ __m256i __DEFAULT_FN_ATTRS
6288_mm256_maskz_loadu_epi32 (__mmask8 __U, void const *__P)
6289{
6290 return (__m256i) __builtin_ia32_loaddqusi256_mask ((__v8si *) __P,
6291 (__v8si)
6292 _mm256_setzero_si256 (),
6293 (__mmask8) __U);
6294}
6295
6296static __inline__ __m128d __DEFAULT_FN_ATTRS
6297_mm_mask_loadu_pd (__m128d __W, __mmask8 __U, void const *__P)
6298{
6299 return (__m128d) __builtin_ia32_loadupd128_mask ((__v2df *) __P,
6300 (__v2df) __W,
6301 (__mmask8) __U);
6302}
6303
6304static __inline__ __m128d __DEFAULT_FN_ATTRS
6305_mm_maskz_loadu_pd (__mmask8 __U, void const *__P)
6306{
6307 return (__m128d) __builtin_ia32_loadupd128_mask ((__v2df *) __P,
6308 (__v2df)
6309 _mm_setzero_pd (),
6310 (__mmask8) __U);
6311}
6312
6313static __inline__ __m256d __DEFAULT_FN_ATTRS
6314_mm256_mask_loadu_pd (__m256d __W, __mmask8 __U, void const *__P)
6315{
6316 return (__m256d) __builtin_ia32_loadupd256_mask ((__v4df *) __P,
6317 (__v4df) __W,
6318 (__mmask8) __U);
6319}
6320
6321static __inline__ __m256d __DEFAULT_FN_ATTRS
6322_mm256_maskz_loadu_pd (__mmask8 __U, void const *__P)
6323{
6324 return (__m256d) __builtin_ia32_loadupd256_mask ((__v4df *) __P,
6325 (__v4df)
6326 _mm256_setzero_pd (),
6327 (__mmask8) __U);
6328}
6329
6330static __inline__ __m128 __DEFAULT_FN_ATTRS
6331_mm_mask_loadu_ps (__m128 __W, __mmask8 __U, void const *__P)
6332{
6333 return (__m128) __builtin_ia32_loadups128_mask ((__v4sf *) __P,
6334 (__v4sf) __W,
6335 (__mmask8) __U);
6336}
6337
6338static __inline__ __m128 __DEFAULT_FN_ATTRS
6339_mm_maskz_loadu_ps (__mmask8 __U, void const *__P)
6340{
6341 return (__m128) __builtin_ia32_loadups128_mask ((__v4sf *) __P,
6342 (__v4sf)
6343 _mm_setzero_ps (),
6344 (__mmask8) __U);
6345}
6346
6347static __inline__ __m256 __DEFAULT_FN_ATTRS
6348_mm256_mask_loadu_ps (__m256 __W, __mmask8 __U, void const *__P)
6349{
6350 return (__m256) __builtin_ia32_loadups256_mask ((__v8sf *) __P,
6351 (__v8sf) __W,
6352 (__mmask8) __U);
6353}
6354
6355static __inline__ __m256 __DEFAULT_FN_ATTRS
6356_mm256_maskz_loadu_ps (__mmask8 __U, void const *__P)
6357{
6358 return (__m256) __builtin_ia32_loadups256_mask ((__v8sf *) __P,
6359 (__v8sf)
6360 _mm256_setzero_ps (),
6361 (__mmask8) __U);
6362}
Michael Zuckermanfa7ccc52016-04-10 10:51:04 +00006363
6364static __inline__ void __DEFAULT_FN_ATTRS
6365_mm_mask_store_pd (void *__P, __mmask8 __U, __m128d __A)
6366{
6367 __builtin_ia32_storeapd128_mask ((__v2df *) __P,
6368 (__v2df) __A,
6369 (__mmask8) __U);
6370}
6371
6372static __inline__ void __DEFAULT_FN_ATTRS
6373_mm256_mask_store_pd (void *__P, __mmask8 __U, __m256d __A)
6374{
6375 __builtin_ia32_storeapd256_mask ((__v4df *) __P,
6376 (__v4df) __A,
6377 (__mmask8) __U);
6378}
6379
6380static __inline__ void __DEFAULT_FN_ATTRS
6381_mm_mask_store_ps (void *__P, __mmask8 __U, __m128 __A)
6382{
6383 __builtin_ia32_storeaps128_mask ((__v4sf *) __P,
6384 (__v4sf) __A,
6385 (__mmask8) __U);
6386}
6387
6388static __inline__ void __DEFAULT_FN_ATTRS
6389_mm256_mask_store_ps (void *__P, __mmask8 __U, __m256 __A)
6390{
6391 __builtin_ia32_storeaps256_mask ((__v8sf *) __P,
6392 (__v8sf) __A,
6393 (__mmask8) __U);
6394}
6395
6396static __inline__ void __DEFAULT_FN_ATTRS
6397_mm_mask_storeu_epi64 (void *__P, __mmask8 __U, __m128i __A)
6398{
6399 __builtin_ia32_storedqudi128_mask ((__v2di *) __P,
6400 (__v2di) __A,
6401 (__mmask8) __U);
6402}
6403
6404static __inline__ void __DEFAULT_FN_ATTRS
6405_mm256_mask_storeu_epi64 (void *__P, __mmask8 __U, __m256i __A)
6406{
6407 __builtin_ia32_storedqudi256_mask ((__v4di *) __P,
6408 (__v4di) __A,
6409 (__mmask8) __U);
6410}
6411
6412static __inline__ void __DEFAULT_FN_ATTRS
6413_mm_mask_storeu_epi32 (void *__P, __mmask8 __U, __m128i __A)
6414{
6415 __builtin_ia32_storedqusi128_mask ((__v4si *) __P,
6416 (__v4si) __A,
6417 (__mmask8) __U);
6418}
6419
6420static __inline__ void __DEFAULT_FN_ATTRS
6421_mm256_mask_storeu_epi32 (void *__P, __mmask8 __U, __m256i __A)
6422{
6423 __builtin_ia32_storedqusi256_mask ((__v8si *) __P,
6424 (__v8si) __A,
6425 (__mmask8) __U);
6426}
6427
6428static __inline__ void __DEFAULT_FN_ATTRS
6429_mm_mask_storeu_pd (void *__P, __mmask8 __U, __m128d __A)
6430{
6431 __builtin_ia32_storeupd128_mask ((__v2df *) __P,
6432 (__v2df) __A,
6433 (__mmask8) __U);
6434}
6435
6436static __inline__ void __DEFAULT_FN_ATTRS
6437_mm256_mask_storeu_pd (void *__P, __mmask8 __U, __m256d __A)
6438{
6439 __builtin_ia32_storeupd256_mask ((__v4df *) __P,
6440 (__v4df) __A,
6441 (__mmask8) __U);
6442}
6443
6444static __inline__ void __DEFAULT_FN_ATTRS
6445_mm_mask_storeu_ps (void *__P, __mmask8 __U, __m128 __A)
6446{
6447 __builtin_ia32_storeups128_mask ((__v4sf *) __P,
6448 (__v4sf) __A,
6449 (__mmask8) __U);
6450}
6451
6452static __inline__ void __DEFAULT_FN_ATTRS
6453_mm256_mask_storeu_ps (void *__P, __mmask8 __U, __m256 __A)
6454{
6455 __builtin_ia32_storeups256_mask ((__v8sf *) __P,
6456 (__v8sf) __A,
6457 (__mmask8) __U);
6458}
6459
Michael Zuckermancdd54c82016-04-10 12:54:23 +00006460
6461static __inline__ __m128d __DEFAULT_FN_ATTRS
6462_mm_mask_unpackhi_pd (__m128d __W, __mmask8 __U, __m128d __A,
6463 __m128d __B)
6464{
6465 return (__m128d) __builtin_ia32_unpckhpd128_mask ((__v2df) __A,
6466 (__v2df) __B,
6467 (__v2df) __W,
6468 (__mmask8) __U);
6469}
6470
6471static __inline__ __m128d __DEFAULT_FN_ATTRS
6472_mm_maskz_unpackhi_pd (__mmask8 __U, __m128d __A, __m128d __B)
6473{
6474 return (__m128d) __builtin_ia32_unpckhpd128_mask ((__v2df) __A,
6475 (__v2df) __B,
6476 (__v2df)
6477 _mm_setzero_pd (),
6478 (__mmask8) __U);
6479}
6480
6481static __inline__ __m256d __DEFAULT_FN_ATTRS
6482_mm256_mask_unpackhi_pd (__m256d __W, __mmask8 __U, __m256d __A,
6483 __m256d __B)
6484{
6485 return (__m256d) __builtin_ia32_unpckhpd256_mask ((__v4df) __A,
6486 (__v4df) __B,
6487 (__v4df) __W,
6488 (__mmask8) __U);
6489}
6490
6491static __inline__ __m256d __DEFAULT_FN_ATTRS
6492_mm256_maskz_unpackhi_pd (__mmask8 __U, __m256d __A, __m256d __B)
6493{
6494 return (__m256d) __builtin_ia32_unpckhpd256_mask ((__v4df) __A,
6495 (__v4df) __B,
6496 (__v4df)
6497 _mm256_setzero_pd (),
6498 (__mmask8) __U);
6499}
6500
6501static __inline__ __m128 __DEFAULT_FN_ATTRS
6502_mm_mask_unpackhi_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B)
6503{
6504 return (__m128) __builtin_ia32_unpckhps128_mask ((__v4sf) __A,
6505 (__v4sf) __B,
6506 (__v4sf) __W,
6507 (__mmask8) __U);
6508}
6509
6510static __inline__ __m128 __DEFAULT_FN_ATTRS
6511_mm_maskz_unpackhi_ps (__mmask8 __U, __m128 __A, __m128 __B)
6512{
6513 return (__m128) __builtin_ia32_unpckhps128_mask ((__v4sf) __A,
6514 (__v4sf) __B,
6515 (__v4sf)
6516 _mm_setzero_ps (),
6517 (__mmask8) __U);
6518}
6519
6520static __inline__ __m256 __DEFAULT_FN_ATTRS
6521_mm256_mask_unpackhi_ps (__m256 __W, __mmask8 __U, __m256 __A,
6522 __m256 __B)
6523{
6524 return (__m256) __builtin_ia32_unpckhps256_mask ((__v8sf) __A,
6525 (__v8sf) __B,
6526 (__v8sf) __W,
6527 (__mmask8) __U);
6528}
6529
6530static __inline__ __m256 __DEFAULT_FN_ATTRS
6531_mm256_maskz_unpackhi_ps (__mmask8 __U, __m256 __A, __m256 __B)
6532{
6533 return (__m256) __builtin_ia32_unpckhps256_mask ((__v8sf) __A,
6534 (__v8sf) __B,
6535 (__v8sf)
6536 _mm256_setzero_ps (),
6537 (__mmask8) __U);
6538}
6539
6540static __inline__ __m128d __DEFAULT_FN_ATTRS
6541_mm_mask_unpacklo_pd (__m128d __W, __mmask8 __U, __m128d __A,
6542 __m128d __B)
6543{
6544 return (__m128d) __builtin_ia32_unpcklpd128_mask ((__v2df) __A,
6545 (__v2df) __B,
6546 (__v2df) __W,
6547 (__mmask8) __U);
6548}
6549
6550static __inline__ __m128d __DEFAULT_FN_ATTRS
6551_mm_maskz_unpacklo_pd (__mmask8 __U, __m128d __A, __m128d __B)
6552{
6553 return (__m128d) __builtin_ia32_unpcklpd128_mask ((__v2df) __A,
6554 (__v2df) __B,
6555 (__v2df)
6556 _mm_setzero_pd (),
6557 (__mmask8) __U);
6558}
6559
6560static __inline__ __m256d __DEFAULT_FN_ATTRS
6561_mm256_mask_unpacklo_pd (__m256d __W, __mmask8 __U, __m256d __A,
6562 __m256d __B)
6563{
6564 return (__m256d) __builtin_ia32_unpcklpd256_mask ((__v4df) __A,
6565 (__v4df) __B,
6566 (__v4df) __W,
6567 (__mmask8) __U);
6568}
6569
6570static __inline__ __m256d __DEFAULT_FN_ATTRS
6571_mm256_maskz_unpacklo_pd (__mmask8 __U, __m256d __A, __m256d __B)
6572{
6573 return (__m256d) __builtin_ia32_unpcklpd256_mask ((__v4df) __A,
6574 (__v4df) __B,
6575 (__v4df)
6576 _mm256_setzero_pd (),
6577 (__mmask8) __U);
6578}
6579
6580static __inline__ __m128 __DEFAULT_FN_ATTRS
6581_mm_mask_unpacklo_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B)
6582{
6583 return (__m128) __builtin_ia32_unpcklps128_mask ((__v4sf) __A,
6584 (__v4sf) __B,
6585 (__v4sf) __W,
6586 (__mmask8) __U);
6587}
6588
6589static __inline__ __m128 __DEFAULT_FN_ATTRS
6590_mm_maskz_unpacklo_ps (__mmask8 __U, __m128 __A, __m128 __B)
6591{
6592 return (__m128) __builtin_ia32_unpcklps128_mask ((__v4sf) __A,
6593 (__v4sf) __B,
6594 (__v4sf)
6595 _mm_setzero_ps (),
6596 (__mmask8) __U);
6597}
6598
6599static __inline__ __m256 __DEFAULT_FN_ATTRS
6600_mm256_mask_unpacklo_ps (__m256 __W, __mmask8 __U, __m256 __A,
6601 __m256 __B)
6602{
6603 return (__m256) __builtin_ia32_unpcklps256_mask ((__v8sf) __A,
6604 (__v8sf) __B,
6605 (__v8sf) __W,
6606 (__mmask8) __U);
6607}
6608
6609static __inline__ __m256 __DEFAULT_FN_ATTRS
6610_mm256_maskz_unpacklo_ps (__mmask8 __U, __m256 __A, __m256 __B)
6611{
6612 return (__m256) __builtin_ia32_unpcklps256_mask ((__v8sf) __A,
6613 (__v8sf) __B,
6614 (__v8sf)
6615 _mm256_setzero_ps (),
6616 (__mmask8) __U);
6617}
6618
6619static __inline__ __m128d __DEFAULT_FN_ATTRS
6620_mm_rcp14_pd (__m128d __A)
6621{
6622 return (__m128d) __builtin_ia32_rcp14pd128_mask ((__v2df) __A,
6623 (__v2df)
6624 _mm_setzero_pd (),
6625 (__mmask8) -1);
6626}
6627
6628static __inline__ __m128d __DEFAULT_FN_ATTRS
6629_mm_mask_rcp14_pd (__m128d __W, __mmask8 __U, __m128d __A)
6630{
6631 return (__m128d) __builtin_ia32_rcp14pd128_mask ((__v2df) __A,
6632 (__v2df) __W,
6633 (__mmask8) __U);
6634}
6635
6636static __inline__ __m128d __DEFAULT_FN_ATTRS
6637_mm_maskz_rcp14_pd (__mmask8 __U, __m128d __A)
6638{
6639 return (__m128d) __builtin_ia32_rcp14pd128_mask ((__v2df) __A,
6640 (__v2df)
6641 _mm_setzero_pd (),
6642 (__mmask8) __U);
6643}
6644
6645static __inline__ __m256d __DEFAULT_FN_ATTRS
6646_mm256_rcp14_pd (__m256d __A)
6647{
6648 return (__m256d) __builtin_ia32_rcp14pd256_mask ((__v4df) __A,
6649 (__v4df)
6650 _mm256_setzero_pd (),
6651 (__mmask8) -1);
6652}
6653
6654static __inline__ __m256d __DEFAULT_FN_ATTRS
6655_mm256_mask_rcp14_pd (__m256d __W, __mmask8 __U, __m256d __A)
6656{
6657 return (__m256d) __builtin_ia32_rcp14pd256_mask ((__v4df) __A,
6658 (__v4df) __W,
6659 (__mmask8) __U);
6660}
6661
6662static __inline__ __m256d __DEFAULT_FN_ATTRS
6663_mm256_maskz_rcp14_pd (__mmask8 __U, __m256d __A)
6664{
6665 return (__m256d) __builtin_ia32_rcp14pd256_mask ((__v4df) __A,
6666 (__v4df)
6667 _mm256_setzero_pd (),
6668 (__mmask8) __U);
6669}
6670
6671static __inline__ __m128 __DEFAULT_FN_ATTRS
6672_mm_rcp14_ps (__m128 __A)
6673{
6674 return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A,
6675 (__v4sf)
6676 _mm_setzero_ps (),
6677 (__mmask8) -1);
6678}
6679
6680static __inline__ __m128 __DEFAULT_FN_ATTRS
6681_mm_mask_rcp14_ps (__m128 __W, __mmask8 __U, __m128 __A)
6682{
6683 return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A,
6684 (__v4sf) __W,
6685 (__mmask8) __U);
6686}
6687
6688static __inline__ __m128 __DEFAULT_FN_ATTRS
6689_mm_maskz_rcp14_ps (__mmask8 __U, __m128 __A)
6690{
6691 return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A,
6692 (__v4sf)
6693 _mm_setzero_ps (),
6694 (__mmask8) __U);
6695}
6696
6697static __inline__ __m256 __DEFAULT_FN_ATTRS
6698_mm256_rcp14_ps (__m256 __A)
6699{
6700 return (__m256) __builtin_ia32_rcp14ps256_mask ((__v8sf) __A,
6701 (__v8sf)
6702 _mm256_setzero_ps (),
6703 (__mmask8) -1);
6704}
6705
6706static __inline__ __m256 __DEFAULT_FN_ATTRS
6707_mm256_mask_rcp14_ps (__m256 __W, __mmask8 __U, __m256 __A)
6708{
6709 return (__m256) __builtin_ia32_rcp14ps256_mask ((__v8sf) __A,
6710 (__v8sf) __W,
6711 (__mmask8) __U);
6712}
6713
6714static __inline__ __m256 __DEFAULT_FN_ATTRS
6715_mm256_maskz_rcp14_ps (__mmask8 __U, __m256 __A)
6716{
6717 return (__m256) __builtin_ia32_rcp14ps256_mask ((__v8sf) __A,
6718 (__v8sf)
6719 _mm256_setzero_ps (),
6720 (__mmask8) __U);
6721}
6722
6723static __inline__ __m128i __DEFAULT_FN_ATTRS
6724_mm_lzcnt_epi32 (__m128i __A)
6725{
6726 return (__m128i) __builtin_ia32_vplzcntd_128_mask ((__v4si) __A,
6727 (__v4si)
6728 _mm_setzero_si128 (),
6729 (__mmask8) -1);
6730}
6731
6732static __inline__ __m128i __DEFAULT_FN_ATTRS
6733_mm_mask_lzcnt_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
6734{
6735 return (__m128i) __builtin_ia32_vplzcntd_128_mask ((__v4si) __A,
6736 (__v4si) __W,
6737 (__mmask8) __U);
6738}
6739
6740static __inline__ __m128i __DEFAULT_FN_ATTRS
6741_mm_maskz_lzcnt_epi32 (__mmask8 __U, __m128i __A)
6742{
6743 return (__m128i) __builtin_ia32_vplzcntd_128_mask ((__v4si) __A,
6744 (__v4si)
6745 _mm_setzero_si128 (),
6746 (__mmask8) __U);
6747}
6748
6749static __inline__ __m256i __DEFAULT_FN_ATTRS
6750_mm256_lzcnt_epi32 (__m256i __A)
6751{
6752 return (__m256i) __builtin_ia32_vplzcntd_256_mask ((__v8si) __A,
6753 (__v8si)
6754 _mm256_setzero_si256 (),
6755 (__mmask8) -1);
6756}
6757
6758static __inline__ __m256i __DEFAULT_FN_ATTRS
6759_mm256_mask_lzcnt_epi32 (__m256i __W, __mmask8 __U, __m256i __A)
6760{
6761 return (__m256i) __builtin_ia32_vplzcntd_256_mask ((__v8si) __A,
6762 (__v8si) __W,
6763 (__mmask8) __U);
6764}
6765
6766static __inline__ __m256i __DEFAULT_FN_ATTRS
6767_mm256_maskz_lzcnt_epi32 (__mmask8 __U, __m256i __A)
6768{
6769 return (__m256i) __builtin_ia32_vplzcntd_256_mask ((__v8si) __A,
6770 (__v8si)
6771 _mm256_setzero_si256 (),
6772 (__mmask8) __U);
6773}
6774
6775static __inline__ __m128i __DEFAULT_FN_ATTRS
6776_mm_lzcnt_epi64 (__m128i __A)
6777{
6778 return (__m128i) __builtin_ia32_vplzcntq_128_mask ((__v2di) __A,
6779 (__v2di)
6780 _mm_setzero_di (),
6781 (__mmask8) -1);
6782}
6783
6784static __inline__ __m128i __DEFAULT_FN_ATTRS
6785_mm_mask_lzcnt_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
6786{
6787 return (__m128i) __builtin_ia32_vplzcntq_128_mask ((__v2di) __A,
6788 (__v2di) __W,
6789 (__mmask8) __U);
6790}
6791
6792static __inline__ __m128i __DEFAULT_FN_ATTRS
6793_mm_maskz_lzcnt_epi64 (__mmask8 __U, __m128i __A)
6794{
6795 return (__m128i) __builtin_ia32_vplzcntq_128_mask ((__v2di) __A,
6796 (__v2di)
6797 _mm_setzero_di (),
6798 (__mmask8) __U);
6799}
6800
6801static __inline__ __m256i __DEFAULT_FN_ATTRS
6802_mm256_lzcnt_epi64 (__m256i __A)
6803{
6804 return (__m256i) __builtin_ia32_vplzcntq_256_mask ((__v4di) __A,
6805 (__v4di)
6806 _mm256_setzero_si256 (),
6807 (__mmask8) -1);
6808}
6809
6810static __inline__ __m256i __DEFAULT_FN_ATTRS
6811_mm256_mask_lzcnt_epi64 (__m256i __W, __mmask8 __U, __m256i __A)
6812{
6813 return (__m256i) __builtin_ia32_vplzcntq_256_mask ((__v4di) __A,
6814 (__v4di) __W,
6815 (__mmask8) __U);
6816}
6817
6818static __inline__ __m256i __DEFAULT_FN_ATTRS
6819_mm256_maskz_lzcnt_epi64 (__mmask8 __U, __m256i __A)
6820{
6821 return (__m256i) __builtin_ia32_vplzcntq_256_mask ((__v4di) __A,
6822 (__v4di)
6823 _mm256_setzero_si256 (),
6824 (__mmask8) __U);
6825}
6826
Michael Kupersteine45af542015-06-30 13:36:19 +00006827#undef __DEFAULT_FN_ATTRS
Eric Christopher99af5b22015-10-27 06:11:03 +00006828#undef __DEFAULT_FN_ATTRS_BOTH
Eric Christopher4d1851682015-06-17 07:09:20 +00006829
Robert Khasanovb9f3a912014-10-08 17:18:13 +00006830#endif /* __AVX512VLINTRIN_H */