Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 1 | /*===---- 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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 31 | #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512vl"))) |
Eric Christopher | 4d185168 | 2015-06-17 07:09:20 +0000 | [diff] [blame] | 32 | |
Craig Topper | 08181f7 | 2016-06-22 06:36:21 +0000 | [diff] [blame] | 33 | /* Doesn't require avx512vl, used in avx512dqintrin.h */ |
| 34 | static __inline __m128i __attribute__((__always_inline__, __nodebug__, __target__("avx512f"))) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 35 | _mm_setzero_di(void) { |
Craig Topper | 08181f7 | 2016-06-22 06:36:21 +0000 | [diff] [blame] | 36 | return (__m128i)(__v2di){ 0LL, 0LL}; |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 37 | } |
| 38 | |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 39 | /* Integer compare */ |
| 40 | |
Craig Topper | fc07498 | 2016-06-13 04:15:11 +0000 | [diff] [blame] | 41 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 42 | _mm_cmpeq_epi32_mask(__m128i __a, __m128i __b) { |
| 43 | return (__mmask8)__builtin_ia32_pcmpeqd128_mask((__v4si)__a, (__v4si)__b, |
| 44 | (__mmask8)-1); |
| 45 | } |
| 46 | |
Craig Topper | fc07498 | 2016-06-13 04:15:11 +0000 | [diff] [blame] | 47 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 48 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 53 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 54 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 59 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 60 | _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 Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 64 | |
Craig Topper | fc07498 | 2016-06-13 04:15:11 +0000 | [diff] [blame] | 65 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 66 | _mm256_cmpeq_epi32_mask(__m256i __a, __m256i __b) { |
| 67 | return (__mmask8)__builtin_ia32_pcmpeqd256_mask((__v8si)__a, (__v8si)__b, |
| 68 | (__mmask8)-1); |
| 69 | } |
| 70 | |
Craig Topper | fc07498 | 2016-06-13 04:15:11 +0000 | [diff] [blame] | 71 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 72 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 77 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 78 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 83 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 84 | _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 | |
Craig Topper | fc07498 | 2016-06-13 04:15:11 +0000 | [diff] [blame] | 89 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 90 | _mm_cmpeq_epi64_mask(__m128i __a, __m128i __b) { |
| 91 | return (__mmask8)__builtin_ia32_pcmpeqq128_mask((__v2di)__a, (__v2di)__b, |
| 92 | (__mmask8)-1); |
| 93 | } |
| 94 | |
Craig Topper | fc07498 | 2016-06-13 04:15:11 +0000 | [diff] [blame] | 95 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 96 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 101 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 102 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 107 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 108 | _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 Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 112 | |
Craig Topper | fc07498 | 2016-06-13 04:15:11 +0000 | [diff] [blame] | 113 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 114 | _mm256_cmpeq_epi64_mask(__m256i __a, __m256i __b) { |
| 115 | return (__mmask8)__builtin_ia32_pcmpeqq256_mask((__v4di)__a, (__v4di)__b, |
| 116 | (__mmask8)-1); |
| 117 | } |
| 118 | |
Craig Topper | fc07498 | 2016-06-13 04:15:11 +0000 | [diff] [blame] | 119 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 120 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 125 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 126 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 131 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 132 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 138 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 139 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 144 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 145 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 150 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 151 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 156 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 157 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 162 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 163 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 168 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 169 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 174 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 175 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 180 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 181 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 186 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 187 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 192 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 193 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 198 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 199 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 204 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 205 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 210 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 211 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 216 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 217 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 222 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 223 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 228 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 229 | _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 | |
Craig Topper | fc07498 | 2016-06-13 04:15:11 +0000 | [diff] [blame] | 234 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 235 | _mm_cmpgt_epi32_mask(__m128i __a, __m128i __b) { |
| 236 | return (__mmask8)__builtin_ia32_pcmpgtd128_mask((__v4si)__a, (__v4si)__b, |
| 237 | (__mmask8)-1); |
| 238 | } |
| 239 | |
Craig Topper | fc07498 | 2016-06-13 04:15:11 +0000 | [diff] [blame] | 240 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 241 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 246 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 247 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 252 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 253 | _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 | |
Craig Topper | fc07498 | 2016-06-13 04:15:11 +0000 | [diff] [blame] | 258 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 259 | _mm256_cmpgt_epi32_mask(__m256i __a, __m256i __b) { |
| 260 | return (__mmask8)__builtin_ia32_pcmpgtd256_mask((__v8si)__a, (__v8si)__b, |
| 261 | (__mmask8)-1); |
| 262 | } |
| 263 | |
Craig Topper | fc07498 | 2016-06-13 04:15:11 +0000 | [diff] [blame] | 264 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 265 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 270 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 271 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 276 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 277 | _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 | |
Craig Topper | fc07498 | 2016-06-13 04:15:11 +0000 | [diff] [blame] | 282 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 283 | _mm_cmpgt_epi64_mask(__m128i __a, __m128i __b) { |
| 284 | return (__mmask8)__builtin_ia32_pcmpgtq128_mask((__v2di)__a, (__v2di)__b, |
| 285 | (__mmask8)-1); |
| 286 | } |
| 287 | |
Craig Topper | fc07498 | 2016-06-13 04:15:11 +0000 | [diff] [blame] | 288 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 289 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 294 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 295 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 300 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 301 | _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 | |
Craig Topper | fc07498 | 2016-06-13 04:15:11 +0000 | [diff] [blame] | 306 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 307 | _mm256_cmpgt_epi64_mask(__m256i __a, __m256i __b) { |
| 308 | return (__mmask8)__builtin_ia32_pcmpgtq256_mask((__v4di)__a, (__v4di)__b, |
| 309 | (__mmask8)-1); |
| 310 | } |
| 311 | |
Craig Topper | fc07498 | 2016-06-13 04:15:11 +0000 | [diff] [blame] | 312 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 313 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 318 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 319 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 324 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 325 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 330 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 331 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 336 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 337 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 342 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 343 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 348 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 349 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 354 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 355 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 360 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 361 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 366 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 367 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 372 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 373 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 378 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 379 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 384 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 385 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 390 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 391 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 396 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 397 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 402 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 403 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 408 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 409 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 414 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 415 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 420 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 421 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 426 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 427 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 432 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 433 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 438 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 439 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 444 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 445 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 450 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 451 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 456 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 457 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 462 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 463 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 468 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 469 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 474 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 475 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 480 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 481 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 486 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 487 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 492 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 493 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 498 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 499 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 504 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 505 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 510 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 511 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 516 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 517 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 522 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 523 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 528 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 529 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 534 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 535 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 540 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 541 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 546 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 547 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 552 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 553 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 558 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 559 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 564 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 565 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 570 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 571 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 576 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 577 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 582 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 583 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 588 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 589 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 594 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 595 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 600 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 601 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 606 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 607 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 612 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 613 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 618 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 619 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 628 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 629 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 638 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 639 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 648 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 649 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 658 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 659 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 668 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 669 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 678 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 679 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 688 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 689 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 698 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 699 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 708 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 709 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 718 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 719 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 728 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 729 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 738 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 739 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 748 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 749 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 758 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 759 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 768 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 769 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 778 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 779 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 787 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 788 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 797 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 798 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 806 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 807 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 816 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 817 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 825 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 826 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 835 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 836 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 844 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 845 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 854 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 855 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 864 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 865 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 873 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 874 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 883 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 884 | _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 Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 892 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 893 | _mm256_mask_and_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 894 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 895 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 896 | (__v8si)_mm256_and_si256(__A, __B), |
| 897 | (__v8si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 898 | } |
| 899 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 900 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 901 | _mm256_maskz_and_epi32(__mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 902 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 903 | return (__m256i)_mm256_mask_and_epi32(_mm256_setzero_si256(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 904 | } |
| 905 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 906 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 907 | _mm_mask_and_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 908 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 909 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 910 | (__v4si)_mm_and_si128(__A, __B), |
| 911 | (__v4si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 912 | } |
| 913 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 914 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 915 | _mm_maskz_and_epi32(__mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 916 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 917 | return (__m128i)_mm_mask_and_epi32(_mm_setzero_si128(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 918 | } |
| 919 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 920 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 921 | _mm256_mask_andnot_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 922 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 923 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 924 | (__v8si)_mm256_andnot_si256(__A, __B), |
| 925 | (__v8si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 926 | } |
| 927 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 928 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 929 | _mm256_maskz_andnot_epi32(__mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 930 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 931 | return (__m256i)_mm256_mask_andnot_epi32(_mm256_setzero_si256(), |
| 932 | __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 933 | } |
| 934 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 935 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 936 | _mm_mask_andnot_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 937 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 938 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 939 | (__v4si)_mm_andnot_si128(__A, __B), |
| 940 | (__v4si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 941 | } |
| 942 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 943 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 944 | _mm_maskz_andnot_epi32 (__mmask8 __U, __m128i __A, __m128i __B) |
| 945 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 946 | return (__m128i)_mm_mask_andnot_epi32(_mm_setzero_si128(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 947 | } |
| 948 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 949 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 950 | _mm256_mask_or_epi32 (__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 951 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 952 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 953 | (__v8si)_mm256_or_si256(__A, __B), |
| 954 | (__v8si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 955 | } |
| 956 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 957 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 958 | _mm256_maskz_or_epi32(__mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 959 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 960 | return (__m256i)_mm256_mask_or_epi32(_mm256_setzero_si256(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 961 | } |
| 962 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 963 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 964 | _mm_mask_or_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 965 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 966 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 967 | (__v4si)_mm_or_si128(__A, __B), |
| 968 | (__v4si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 969 | } |
| 970 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 971 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 972 | _mm_maskz_or_epi32(__mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 973 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 974 | return (__m128i)_mm_mask_or_epi32(_mm_setzero_si128(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 975 | } |
| 976 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 977 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 978 | _mm256_mask_xor_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 979 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 980 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 981 | (__v8si)_mm256_xor_si256(__A, __B), |
| 982 | (__v8si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 983 | } |
| 984 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 985 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 986 | _mm256_maskz_xor_epi32(__mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 987 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 988 | return (__m256i)_mm256_mask_xor_epi32(_mm256_setzero_si256(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 989 | } |
| 990 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 991 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 992 | _mm_mask_xor_epi32(__m128i __W, __mmask8 __U, __m128i __A, |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 993 | __m128i __B) |
| 994 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 995 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 996 | (__v4si)_mm_xor_si128(__A, __B), |
| 997 | (__v4si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 998 | } |
| 999 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1000 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 1001 | _mm_maskz_xor_epi32(__mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1002 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 1003 | return (__m128i)_mm_mask_xor_epi32(_mm_setzero_si128(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1004 | } |
| 1005 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1006 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 1007 | _mm256_mask_and_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1008 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 1009 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 1010 | (__v4di)_mm256_and_si256(__A, __B), |
| 1011 | (__v4di)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1012 | } |
| 1013 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1014 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 1015 | _mm256_maskz_and_epi64(__mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1016 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 1017 | return (__m256i)_mm256_mask_and_epi64(_mm256_setzero_si256(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1018 | } |
| 1019 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1020 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 1021 | _mm_mask_and_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
| 1022 | { |
| 1023 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 1024 | (__v2di)_mm_and_si128(__A, __B), |
| 1025 | (__v2di)__W); |
| 1026 | } |
| 1027 | |
| 1028 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1029 | _mm_maskz_and_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
| 1030 | { |
| 1031 | return (__m128i)_mm_mask_and_epi64(_mm_setzero_si128(), __U, __A, __B); |
| 1032 | } |
| 1033 | |
| 1034 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1035 | _mm256_mask_andnot_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
| 1036 | { |
| 1037 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 1038 | (__v4di)_mm256_andnot_si256(__A, __B), |
| 1039 | (__v4di)__W); |
| 1040 | } |
| 1041 | |
| 1042 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1043 | _mm256_maskz_andnot_epi64(__mmask8 __U, __m256i __A, __m256i __B) |
| 1044 | { |
| 1045 | return (__m256i)_mm256_mask_andnot_epi64(_mm256_setzero_si256(), |
| 1046 | __U, __A, __B); |
| 1047 | } |
| 1048 | |
| 1049 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1050 | _mm_mask_andnot_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
| 1051 | { |
| 1052 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 1053 | (__v2di)_mm_andnot_si128(__A, __B), |
| 1054 | (__v2di)__W); |
| 1055 | } |
| 1056 | |
| 1057 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1058 | _mm_maskz_andnot_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
| 1059 | { |
| 1060 | return (__m128i)_mm_mask_andnot_epi64(_mm_setzero_si128(), __U, __A, __B); |
| 1061 | } |
| 1062 | |
| 1063 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1064 | _mm256_mask_or_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
| 1065 | { |
| 1066 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 1067 | (__v4di)_mm256_or_si256(__A, __B), |
| 1068 | (__v4di)__W); |
| 1069 | } |
| 1070 | |
| 1071 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1072 | _mm256_maskz_or_epi64(__mmask8 __U, __m256i __A, __m256i __B) |
| 1073 | { |
| 1074 | return (__m256i)_mm256_mask_or_epi64(_mm256_setzero_si256(), __U, __A, __B); |
| 1075 | } |
| 1076 | |
| 1077 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1078 | _mm_mask_or_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
| 1079 | { |
| 1080 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 1081 | (__v2di)_mm_or_si128(__A, __B), |
| 1082 | (__v2di)__W); |
| 1083 | } |
| 1084 | |
| 1085 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1086 | _mm_maskz_or_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
| 1087 | { |
| 1088 | return (__m128i)_mm_mask_or_epi64(_mm_setzero_si128(), __U, __A, __B); |
| 1089 | } |
| 1090 | |
| 1091 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1092 | _mm256_mask_xor_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
| 1093 | { |
| 1094 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 1095 | (__v4di)_mm256_xor_si256(__A, __B), |
| 1096 | (__v4di)__W); |
| 1097 | } |
| 1098 | |
| 1099 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1100 | _mm256_maskz_xor_epi64(__mmask8 __U, __m256i __A, __m256i __B) |
| 1101 | { |
| 1102 | return (__m256i)_mm256_mask_xor_epi64(_mm256_setzero_si256(), __U, __A, __B); |
| 1103 | } |
| 1104 | |
| 1105 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1106 | _mm_mask_xor_epi64(__m128i __W, __mmask8 __U, __m128i __A, |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1107 | __m128i __B) |
| 1108 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 1109 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 1110 | (__v2di)_mm_xor_si128(__A, __B), |
| 1111 | (__v2di)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1112 | } |
| 1113 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1114 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 1115 | _mm_maskz_xor_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1116 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame^] | 1117 | return (__m128i)_mm_mask_xor_epi64(_mm_setzero_si128(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1118 | } |
| 1119 | |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1120 | #define _mm_cmp_epi32_mask(a, b, p) __extension__ ({ \ |
| 1121 | (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)(__m128i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1122 | (__v4si)(__m128i)(b), (int)(p), \ |
| 1123 | (__mmask8)-1); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1124 | |
| 1125 | #define _mm_mask_cmp_epi32_mask(m, a, b, p) __extension__ ({ \ |
| 1126 | (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)(__m128i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1127 | (__v4si)(__m128i)(b), (int)(p), \ |
| 1128 | (__mmask8)(m)); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1129 | |
| 1130 | #define _mm_cmp_epu32_mask(a, b, p) __extension__ ({ \ |
| 1131 | (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)(__m128i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1132 | (__v4si)(__m128i)(b), (int)(p), \ |
| 1133 | (__mmask8)-1); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1134 | |
| 1135 | #define _mm_mask_cmp_epu32_mask(m, a, b, p) __extension__ ({ \ |
| 1136 | (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)(__m128i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1137 | (__v4si)(__m128i)(b), (int)(p), \ |
| 1138 | (__mmask8)(m)); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1139 | |
| 1140 | #define _mm256_cmp_epi32_mask(a, b, p) __extension__ ({ \ |
| 1141 | (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)(__m256i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1142 | (__v8si)(__m256i)(b), (int)(p), \ |
| 1143 | (__mmask8)-1); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1144 | |
| 1145 | #define _mm256_mask_cmp_epi32_mask(m, a, b, p) __extension__ ({ \ |
| 1146 | (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)(__m256i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1147 | (__v8si)(__m256i)(b), (int)(p), \ |
| 1148 | (__mmask8)(m)); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1149 | |
| 1150 | #define _mm256_cmp_epu32_mask(a, b, p) __extension__ ({ \ |
| 1151 | (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)(__m256i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1152 | (__v8si)(__m256i)(b), (int)(p), \ |
| 1153 | (__mmask8)-1); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1154 | |
| 1155 | #define _mm256_mask_cmp_epu32_mask(m, a, b, p) __extension__ ({ \ |
| 1156 | (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)(__m256i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1157 | (__v8si)(__m256i)(b), (int)(p), \ |
| 1158 | (__mmask8)(m)); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1159 | |
| 1160 | #define _mm_cmp_epi64_mask(a, b, p) __extension__ ({ \ |
| 1161 | (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)(__m128i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1162 | (__v2di)(__m128i)(b), (int)(p), \ |
| 1163 | (__mmask8)-1); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1164 | |
| 1165 | #define _mm_mask_cmp_epi64_mask(m, a, b, p) __extension__ ({ \ |
| 1166 | (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)(__m128i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1167 | (__v2di)(__m128i)(b), (int)(p), \ |
| 1168 | (__mmask8)(m)); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1169 | |
| 1170 | #define _mm_cmp_epu64_mask(a, b, p) __extension__ ({ \ |
| 1171 | (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)(__m128i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1172 | (__v2di)(__m128i)(b), (int)(p), \ |
| 1173 | (__mmask8)-1); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1174 | |
| 1175 | #define _mm_mask_cmp_epu64_mask(m, a, b, p) __extension__ ({ \ |
| 1176 | (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)(__m128i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1177 | (__v2di)(__m128i)(b), (int)(p), \ |
| 1178 | (__mmask8)(m)); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1179 | |
| 1180 | #define _mm256_cmp_epi64_mask(a, b, p) __extension__ ({ \ |
| 1181 | (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)(__m256i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1182 | (__v4di)(__m256i)(b), (int)(p), \ |
| 1183 | (__mmask8)-1); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1184 | |
| 1185 | #define _mm256_mask_cmp_epi64_mask(m, a, b, p) __extension__ ({ \ |
| 1186 | (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)(__m256i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1187 | (__v4di)(__m256i)(b), (int)(p), \ |
| 1188 | (__mmask8)(m)); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1189 | |
| 1190 | #define _mm256_cmp_epu64_mask(a, b, p) __extension__ ({ \ |
| 1191 | (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)(__m256i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1192 | (__v4di)(__m256i)(b), (int)(p), \ |
| 1193 | (__mmask8)-1); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1194 | |
| 1195 | #define _mm256_mask_cmp_epu64_mask(m, a, b, p) __extension__ ({ \ |
| 1196 | (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)(__m256i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1197 | (__v4di)(__m256i)(b), (int)(p), \ |
| 1198 | (__mmask8)(m)); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1199 | |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 1200 | #define _mm256_cmp_ps_mask(a, b, p) __extension__ ({ \ |
| 1201 | (__mmask8)__builtin_ia32_cmpps256_mask((__v8sf)(__m256)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1202 | (__v8sf)(__m256)(b), (int)(p), \ |
| 1203 | (__mmask8)-1); }) |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 1204 | |
| 1205 | #define _mm256_mask_cmp_ps_mask(m, a, b, p) __extension__ ({ \ |
| 1206 | (__mmask8)__builtin_ia32_cmpps256_mask((__v8sf)(__m256)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1207 | (__v8sf)(__m256)(b), (int)(p), \ |
| 1208 | (__mmask8)(m)); }) |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 1209 | |
| 1210 | #define _mm256_cmp_pd_mask(a, b, p) __extension__ ({ \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1211 | (__mmask8)__builtin_ia32_cmppd256_mask((__v4df)(__m256d)(a), \ |
| 1212 | (__v4df)(__m256d)(b), (int)(p), \ |
| 1213 | (__mmask8)-1); }) |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 1214 | |
| 1215 | #define _mm256_mask_cmp_pd_mask(m, a, b, p) __extension__ ({ \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1216 | (__mmask8)__builtin_ia32_cmppd256_mask((__v4df)(__m256d)(a), \ |
| 1217 | (__v4df)(__m256d)(b), (int)(p), \ |
| 1218 | (__mmask8)(m)); }) |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 1219 | |
Michael Zuckerman | eb5f178 | 2016-05-26 08:10:12 +0000 | [diff] [blame] | 1220 | #define _mm_cmp_ps_mask(a, b, p) __extension__ ({ \ |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 1221 | (__mmask8)__builtin_ia32_cmpps128_mask((__v4sf)(__m128)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1222 | (__v4sf)(__m128)(b), (int)(p), \ |
| 1223 | (__mmask8)-1); }) |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 1224 | |
Michael Zuckerman | eb5f178 | 2016-05-26 08:10:12 +0000 | [diff] [blame] | 1225 | #define _mm_mask_cmp_ps_mask(m, a, b, p) __extension__ ({ \ |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 1226 | (__mmask8)__builtin_ia32_cmpps128_mask((__v4sf)(__m128)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1227 | (__v4sf)(__m128)(b), (int)(p), \ |
| 1228 | (__mmask8)(m)); }) |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 1229 | |
Michael Zuckerman | eb5f178 | 2016-05-26 08:10:12 +0000 | [diff] [blame] | 1230 | #define _mm_cmp_pd_mask(a, b, p) __extension__ ({ \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1231 | (__mmask8)__builtin_ia32_cmppd128_mask((__v2df)(__m128d)(a), \ |
| 1232 | (__v2df)(__m128d)(b), (int)(p), \ |
| 1233 | (__mmask8)-1); }) |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 1234 | |
Michael Zuckerman | eb5f178 | 2016-05-26 08:10:12 +0000 | [diff] [blame] | 1235 | #define _mm_mask_cmp_pd_mask(m, a, b, p) __extension__ ({ \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1236 | (__mmask8)__builtin_ia32_cmppd128_mask((__v2df)(__m128d)(a), \ |
| 1237 | (__v2df)(__m128d)(b), (int)(p), \ |
| 1238 | (__mmask8)(m)); }) |
Eric Christopher | 4d185168 | 2015-06-17 07:09:20 +0000 | [diff] [blame] | 1239 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1240 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1241 | _mm_mask_fmadd_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) |
| 1242 | { |
| 1243 | return (__m128d) __builtin_ia32_vfmaddpd128_mask ((__v2df) __A, |
| 1244 | (__v2df) __B, |
| 1245 | (__v2df) __C, |
| 1246 | (__mmask8) __U); |
| 1247 | } |
| 1248 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1249 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1250 | _mm_mask3_fmadd_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) |
| 1251 | { |
| 1252 | return (__m128d) __builtin_ia32_vfmaddpd128_mask3 ((__v2df) __A, |
| 1253 | (__v2df) __B, |
| 1254 | (__v2df) __C, |
| 1255 | (__mmask8) __U); |
| 1256 | } |
| 1257 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1258 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1259 | _mm_maskz_fmadd_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) |
| 1260 | { |
| 1261 | return (__m128d) __builtin_ia32_vfmaddpd128_maskz ((__v2df) __A, |
| 1262 | (__v2df) __B, |
| 1263 | (__v2df) __C, |
| 1264 | (__mmask8) __U); |
| 1265 | } |
| 1266 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1267 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1268 | _mm_mask_fmsub_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) |
| 1269 | { |
| 1270 | return (__m128d) __builtin_ia32_vfmaddpd128_mask ((__v2df) __A, |
| 1271 | (__v2df) __B, |
| 1272 | -(__v2df) __C, |
| 1273 | (__mmask8) __U); |
| 1274 | } |
| 1275 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1276 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1277 | _mm_maskz_fmsub_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) |
| 1278 | { |
| 1279 | return (__m128d) __builtin_ia32_vfmaddpd128_maskz ((__v2df) __A, |
| 1280 | (__v2df) __B, |
| 1281 | -(__v2df) __C, |
| 1282 | (__mmask8) __U); |
| 1283 | } |
| 1284 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1285 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1286 | _mm_mask3_fnmadd_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) |
| 1287 | { |
| 1288 | return (__m128d) __builtin_ia32_vfmaddpd128_mask3 (-(__v2df) __A, |
| 1289 | (__v2df) __B, |
| 1290 | (__v2df) __C, |
| 1291 | (__mmask8) __U); |
| 1292 | } |
| 1293 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1294 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1295 | _mm_maskz_fnmadd_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) |
| 1296 | { |
| 1297 | return (__m128d) __builtin_ia32_vfmaddpd128_maskz (-(__v2df) __A, |
| 1298 | (__v2df) __B, |
| 1299 | (__v2df) __C, |
| 1300 | (__mmask8) __U); |
| 1301 | } |
| 1302 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1303 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1304 | _mm_maskz_fnmsub_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) |
| 1305 | { |
| 1306 | return (__m128d) __builtin_ia32_vfmaddpd128_maskz (-(__v2df) __A, |
| 1307 | (__v2df) __B, |
| 1308 | -(__v2df) __C, |
| 1309 | (__mmask8) __U); |
| 1310 | } |
| 1311 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1312 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1313 | _mm256_mask_fmadd_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) |
| 1314 | { |
| 1315 | return (__m256d) __builtin_ia32_vfmaddpd256_mask ((__v4df) __A, |
| 1316 | (__v4df) __B, |
| 1317 | (__v4df) __C, |
| 1318 | (__mmask8) __U); |
| 1319 | } |
| 1320 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1321 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1322 | _mm256_mask3_fmadd_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) |
| 1323 | { |
| 1324 | return (__m256d) __builtin_ia32_vfmaddpd256_mask3 ((__v4df) __A, |
| 1325 | (__v4df) __B, |
| 1326 | (__v4df) __C, |
| 1327 | (__mmask8) __U); |
| 1328 | } |
| 1329 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1330 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1331 | _mm256_maskz_fmadd_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) |
| 1332 | { |
| 1333 | return (__m256d) __builtin_ia32_vfmaddpd256_maskz ((__v4df) __A, |
| 1334 | (__v4df) __B, |
| 1335 | (__v4df) __C, |
| 1336 | (__mmask8) __U); |
| 1337 | } |
| 1338 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1339 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1340 | _mm256_mask_fmsub_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) |
| 1341 | { |
| 1342 | return (__m256d) __builtin_ia32_vfmaddpd256_mask ((__v4df) __A, |
| 1343 | (__v4df) __B, |
| 1344 | -(__v4df) __C, |
| 1345 | (__mmask8) __U); |
| 1346 | } |
| 1347 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1348 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1349 | _mm256_maskz_fmsub_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) |
| 1350 | { |
| 1351 | return (__m256d) __builtin_ia32_vfmaddpd256_maskz ((__v4df) __A, |
| 1352 | (__v4df) __B, |
| 1353 | -(__v4df) __C, |
| 1354 | (__mmask8) __U); |
| 1355 | } |
| 1356 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1357 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1358 | _mm256_mask3_fnmadd_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) |
| 1359 | { |
| 1360 | return (__m256d) __builtin_ia32_vfmaddpd256_mask3 (-(__v4df) __A, |
| 1361 | (__v4df) __B, |
| 1362 | (__v4df) __C, |
| 1363 | (__mmask8) __U); |
| 1364 | } |
| 1365 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1366 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1367 | _mm256_maskz_fnmadd_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) |
| 1368 | { |
| 1369 | return (__m256d) __builtin_ia32_vfmaddpd256_maskz (-(__v4df) __A, |
| 1370 | (__v4df) __B, |
| 1371 | (__v4df) __C, |
| 1372 | (__mmask8) __U); |
| 1373 | } |
| 1374 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1375 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1376 | _mm256_maskz_fnmsub_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) |
| 1377 | { |
| 1378 | return (__m256d) __builtin_ia32_vfmaddpd256_maskz (-(__v4df) __A, |
| 1379 | (__v4df) __B, |
| 1380 | -(__v4df) __C, |
| 1381 | (__mmask8) __U); |
| 1382 | } |
| 1383 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1384 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1385 | _mm_mask_fmadd_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) |
| 1386 | { |
| 1387 | return (__m128) __builtin_ia32_vfmaddps128_mask ((__v4sf) __A, |
| 1388 | (__v4sf) __B, |
| 1389 | (__v4sf) __C, |
| 1390 | (__mmask8) __U); |
| 1391 | } |
| 1392 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1393 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1394 | _mm_mask3_fmadd_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) |
| 1395 | { |
| 1396 | return (__m128) __builtin_ia32_vfmaddps128_mask3 ((__v4sf) __A, |
| 1397 | (__v4sf) __B, |
| 1398 | (__v4sf) __C, |
| 1399 | (__mmask8) __U); |
| 1400 | } |
| 1401 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1402 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1403 | _mm_maskz_fmadd_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) |
| 1404 | { |
| 1405 | return (__m128) __builtin_ia32_vfmaddps128_maskz ((__v4sf) __A, |
| 1406 | (__v4sf) __B, |
| 1407 | (__v4sf) __C, |
| 1408 | (__mmask8) __U); |
| 1409 | } |
| 1410 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1411 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1412 | _mm_mask_fmsub_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) |
| 1413 | { |
| 1414 | return (__m128) __builtin_ia32_vfmaddps128_mask ((__v4sf) __A, |
| 1415 | (__v4sf) __B, |
| 1416 | -(__v4sf) __C, |
| 1417 | (__mmask8) __U); |
| 1418 | } |
| 1419 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1420 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1421 | _mm_maskz_fmsub_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) |
| 1422 | { |
| 1423 | return (__m128) __builtin_ia32_vfmaddps128_maskz ((__v4sf) __A, |
| 1424 | (__v4sf) __B, |
| 1425 | -(__v4sf) __C, |
| 1426 | (__mmask8) __U); |
| 1427 | } |
| 1428 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1429 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1430 | _mm_mask3_fnmadd_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) |
| 1431 | { |
| 1432 | return (__m128) __builtin_ia32_vfmaddps128_mask3 (-(__v4sf) __A, |
| 1433 | (__v4sf) __B, |
| 1434 | (__v4sf) __C, |
| 1435 | (__mmask8) __U); |
| 1436 | } |
| 1437 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1438 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1439 | _mm_maskz_fnmadd_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) |
| 1440 | { |
| 1441 | return (__m128) __builtin_ia32_vfmaddps128_maskz (-(__v4sf) __A, |
| 1442 | (__v4sf) __B, |
| 1443 | (__v4sf) __C, |
| 1444 | (__mmask8) __U); |
| 1445 | } |
| 1446 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1447 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1448 | _mm_maskz_fnmsub_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) |
| 1449 | { |
| 1450 | return (__m128) __builtin_ia32_vfmaddps128_maskz (-(__v4sf) __A, |
| 1451 | (__v4sf) __B, |
| 1452 | -(__v4sf) __C, |
| 1453 | (__mmask8) __U); |
| 1454 | } |
| 1455 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1456 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1457 | _mm256_mask_fmadd_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) |
| 1458 | { |
| 1459 | return (__m256) __builtin_ia32_vfmaddps256_mask ((__v8sf) __A, |
| 1460 | (__v8sf) __B, |
| 1461 | (__v8sf) __C, |
| 1462 | (__mmask8) __U); |
| 1463 | } |
| 1464 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1465 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1466 | _mm256_mask3_fmadd_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) |
| 1467 | { |
| 1468 | return (__m256) __builtin_ia32_vfmaddps256_mask3 ((__v8sf) __A, |
| 1469 | (__v8sf) __B, |
| 1470 | (__v8sf) __C, |
| 1471 | (__mmask8) __U); |
| 1472 | } |
| 1473 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1474 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1475 | _mm256_maskz_fmadd_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) |
| 1476 | { |
| 1477 | return (__m256) __builtin_ia32_vfmaddps256_maskz ((__v8sf) __A, |
| 1478 | (__v8sf) __B, |
| 1479 | (__v8sf) __C, |
| 1480 | (__mmask8) __U); |
| 1481 | } |
| 1482 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1483 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1484 | _mm256_mask_fmsub_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) |
| 1485 | { |
| 1486 | return (__m256) __builtin_ia32_vfmaddps256_mask ((__v8sf) __A, |
| 1487 | (__v8sf) __B, |
| 1488 | -(__v8sf) __C, |
| 1489 | (__mmask8) __U); |
| 1490 | } |
| 1491 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1492 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1493 | _mm256_maskz_fmsub_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) |
| 1494 | { |
| 1495 | return (__m256) __builtin_ia32_vfmaddps256_maskz ((__v8sf) __A, |
| 1496 | (__v8sf) __B, |
| 1497 | -(__v8sf) __C, |
| 1498 | (__mmask8) __U); |
| 1499 | } |
| 1500 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1501 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1502 | _mm256_mask3_fnmadd_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) |
| 1503 | { |
| 1504 | return (__m256) __builtin_ia32_vfmaddps256_mask3 (-(__v8sf) __A, |
| 1505 | (__v8sf) __B, |
| 1506 | (__v8sf) __C, |
| 1507 | (__mmask8) __U); |
| 1508 | } |
| 1509 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1510 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1511 | _mm256_maskz_fnmadd_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) |
| 1512 | { |
| 1513 | return (__m256) __builtin_ia32_vfmaddps256_maskz (-(__v8sf) __A, |
| 1514 | (__v8sf) __B, |
| 1515 | (__v8sf) __C, |
| 1516 | (__mmask8) __U); |
| 1517 | } |
| 1518 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1519 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1520 | _mm256_maskz_fnmsub_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) |
| 1521 | { |
| 1522 | return (__m256) __builtin_ia32_vfmaddps256_maskz (-(__v8sf) __A, |
| 1523 | (__v8sf) __B, |
| 1524 | -(__v8sf) __C, |
| 1525 | (__mmask8) __U); |
| 1526 | } |
| 1527 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1528 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1529 | _mm_mask_fmaddsub_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) |
| 1530 | { |
| 1531 | return (__m128d) __builtin_ia32_vfmaddsubpd128_mask ((__v2df) __A, |
| 1532 | (__v2df) __B, |
| 1533 | (__v2df) __C, |
| 1534 | (__mmask8) __U); |
| 1535 | } |
| 1536 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1537 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1538 | _mm_mask3_fmaddsub_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) |
| 1539 | { |
| 1540 | return (__m128d) __builtin_ia32_vfmaddsubpd128_mask3 ((__v2df) __A, |
| 1541 | (__v2df) __B, |
| 1542 | (__v2df) __C, |
| 1543 | (__mmask8) |
| 1544 | __U); |
| 1545 | } |
| 1546 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1547 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1548 | _mm_maskz_fmaddsub_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) |
| 1549 | { |
| 1550 | return (__m128d) __builtin_ia32_vfmaddsubpd128_maskz ((__v2df) __A, |
| 1551 | (__v2df) __B, |
| 1552 | (__v2df) __C, |
| 1553 | (__mmask8) |
| 1554 | __U); |
| 1555 | } |
| 1556 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1557 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1558 | _mm_mask_fmsubadd_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) |
| 1559 | { |
| 1560 | return (__m128d) __builtin_ia32_vfmaddsubpd128_mask ((__v2df) __A, |
| 1561 | (__v2df) __B, |
| 1562 | -(__v2df) __C, |
| 1563 | (__mmask8) __U); |
| 1564 | } |
| 1565 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1566 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1567 | _mm_maskz_fmsubadd_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) |
| 1568 | { |
| 1569 | return (__m128d) __builtin_ia32_vfmaddsubpd128_maskz ((__v2df) __A, |
| 1570 | (__v2df) __B, |
| 1571 | -(__v2df) __C, |
| 1572 | (__mmask8) |
| 1573 | __U); |
| 1574 | } |
| 1575 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1576 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1577 | _mm256_mask_fmaddsub_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) |
| 1578 | { |
| 1579 | return (__m256d) __builtin_ia32_vfmaddsubpd256_mask ((__v4df) __A, |
| 1580 | (__v4df) __B, |
| 1581 | (__v4df) __C, |
| 1582 | (__mmask8) __U); |
| 1583 | } |
| 1584 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1585 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1586 | _mm256_mask3_fmaddsub_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) |
| 1587 | { |
| 1588 | return (__m256d) __builtin_ia32_vfmaddsubpd256_mask3 ((__v4df) __A, |
| 1589 | (__v4df) __B, |
| 1590 | (__v4df) __C, |
| 1591 | (__mmask8) |
| 1592 | __U); |
| 1593 | } |
| 1594 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1595 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1596 | _mm256_maskz_fmaddsub_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) |
| 1597 | { |
| 1598 | return (__m256d) __builtin_ia32_vfmaddsubpd256_maskz ((__v4df) __A, |
| 1599 | (__v4df) __B, |
| 1600 | (__v4df) __C, |
| 1601 | (__mmask8) |
| 1602 | __U); |
| 1603 | } |
| 1604 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1605 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1606 | _mm256_mask_fmsubadd_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) |
| 1607 | { |
| 1608 | return (__m256d) __builtin_ia32_vfmaddsubpd256_mask ((__v4df) __A, |
| 1609 | (__v4df) __B, |
| 1610 | -(__v4df) __C, |
| 1611 | (__mmask8) __U); |
| 1612 | } |
| 1613 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1614 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1615 | _mm256_maskz_fmsubadd_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) |
| 1616 | { |
| 1617 | return (__m256d) __builtin_ia32_vfmaddsubpd256_maskz ((__v4df) __A, |
| 1618 | (__v4df) __B, |
| 1619 | -(__v4df) __C, |
| 1620 | (__mmask8) |
| 1621 | __U); |
| 1622 | } |
| 1623 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1624 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1625 | _mm_mask_fmaddsub_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) |
| 1626 | { |
| 1627 | return (__m128) __builtin_ia32_vfmaddsubps128_mask ((__v4sf) __A, |
| 1628 | (__v4sf) __B, |
| 1629 | (__v4sf) __C, |
| 1630 | (__mmask8) __U); |
| 1631 | } |
| 1632 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1633 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1634 | _mm_mask3_fmaddsub_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) |
| 1635 | { |
| 1636 | return (__m128) __builtin_ia32_vfmaddsubps128_mask3 ((__v4sf) __A, |
| 1637 | (__v4sf) __B, |
| 1638 | (__v4sf) __C, |
| 1639 | (__mmask8) __U); |
| 1640 | } |
| 1641 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1642 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1643 | _mm_maskz_fmaddsub_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) |
| 1644 | { |
| 1645 | return (__m128) __builtin_ia32_vfmaddsubps128_maskz ((__v4sf) __A, |
| 1646 | (__v4sf) __B, |
| 1647 | (__v4sf) __C, |
| 1648 | (__mmask8) __U); |
| 1649 | } |
| 1650 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1651 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1652 | _mm_mask_fmsubadd_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) |
| 1653 | { |
| 1654 | return (__m128) __builtin_ia32_vfmaddsubps128_mask ((__v4sf) __A, |
| 1655 | (__v4sf) __B, |
| 1656 | -(__v4sf) __C, |
| 1657 | (__mmask8) __U); |
| 1658 | } |
| 1659 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1660 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1661 | _mm_maskz_fmsubadd_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) |
| 1662 | { |
| 1663 | return (__m128) __builtin_ia32_vfmaddsubps128_maskz ((__v4sf) __A, |
| 1664 | (__v4sf) __B, |
| 1665 | -(__v4sf) __C, |
| 1666 | (__mmask8) __U); |
| 1667 | } |
| 1668 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1669 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1670 | _mm256_mask_fmaddsub_ps(__m256 __A, __mmask8 __U, __m256 __B, |
| 1671 | __m256 __C) |
| 1672 | { |
| 1673 | return (__m256) __builtin_ia32_vfmaddsubps256_mask ((__v8sf) __A, |
| 1674 | (__v8sf) __B, |
| 1675 | (__v8sf) __C, |
| 1676 | (__mmask8) __U); |
| 1677 | } |
| 1678 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1679 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1680 | _mm256_mask3_fmaddsub_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) |
| 1681 | { |
| 1682 | return (__m256) __builtin_ia32_vfmaddsubps256_mask3 ((__v8sf) __A, |
| 1683 | (__v8sf) __B, |
| 1684 | (__v8sf) __C, |
| 1685 | (__mmask8) __U); |
| 1686 | } |
| 1687 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1688 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1689 | _mm256_maskz_fmaddsub_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) |
| 1690 | { |
| 1691 | return (__m256) __builtin_ia32_vfmaddsubps256_maskz ((__v8sf) __A, |
| 1692 | (__v8sf) __B, |
| 1693 | (__v8sf) __C, |
| 1694 | (__mmask8) __U); |
| 1695 | } |
| 1696 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1697 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1698 | _mm256_mask_fmsubadd_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) |
| 1699 | { |
| 1700 | return (__m256) __builtin_ia32_vfmaddsubps256_mask ((__v8sf) __A, |
| 1701 | (__v8sf) __B, |
| 1702 | -(__v8sf) __C, |
| 1703 | (__mmask8) __U); |
| 1704 | } |
| 1705 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1706 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1707 | _mm256_maskz_fmsubadd_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) |
| 1708 | { |
| 1709 | return (__m256) __builtin_ia32_vfmaddsubps256_maskz ((__v8sf) __A, |
| 1710 | (__v8sf) __B, |
| 1711 | -(__v8sf) __C, |
| 1712 | (__mmask8) __U); |
| 1713 | } |
| 1714 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1715 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1716 | _mm_mask3_fmsub_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) |
| 1717 | { |
| 1718 | return (__m128d) __builtin_ia32_vfmsubpd128_mask3 ((__v2df) __A, |
| 1719 | (__v2df) __B, |
| 1720 | (__v2df) __C, |
| 1721 | (__mmask8) __U); |
| 1722 | } |
| 1723 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1724 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1725 | _mm256_mask3_fmsub_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) |
| 1726 | { |
| 1727 | return (__m256d) __builtin_ia32_vfmsubpd256_mask3 ((__v4df) __A, |
| 1728 | (__v4df) __B, |
| 1729 | (__v4df) __C, |
| 1730 | (__mmask8) __U); |
| 1731 | } |
| 1732 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1733 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1734 | _mm_mask3_fmsub_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) |
| 1735 | { |
| 1736 | return (__m128) __builtin_ia32_vfmsubps128_mask3 ((__v4sf) __A, |
| 1737 | (__v4sf) __B, |
| 1738 | (__v4sf) __C, |
| 1739 | (__mmask8) __U); |
| 1740 | } |
| 1741 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1742 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1743 | _mm256_mask3_fmsub_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) |
| 1744 | { |
| 1745 | return (__m256) __builtin_ia32_vfmsubps256_mask3 ((__v8sf) __A, |
| 1746 | (__v8sf) __B, |
| 1747 | (__v8sf) __C, |
| 1748 | (__mmask8) __U); |
| 1749 | } |
| 1750 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1751 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1752 | _mm_mask3_fmsubadd_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) |
| 1753 | { |
| 1754 | return (__m128d) __builtin_ia32_vfmsubaddpd128_mask3 ((__v2df) __A, |
| 1755 | (__v2df) __B, |
| 1756 | (__v2df) __C, |
| 1757 | (__mmask8) |
| 1758 | __U); |
| 1759 | } |
| 1760 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1761 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1762 | _mm256_mask3_fmsubadd_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) |
| 1763 | { |
| 1764 | return (__m256d) __builtin_ia32_vfmsubaddpd256_mask3 ((__v4df) __A, |
| 1765 | (__v4df) __B, |
| 1766 | (__v4df) __C, |
| 1767 | (__mmask8) |
| 1768 | __U); |
| 1769 | } |
| 1770 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1771 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1772 | _mm_mask3_fmsubadd_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) |
| 1773 | { |
| 1774 | return (__m128) __builtin_ia32_vfmsubaddps128_mask3 ((__v4sf) __A, |
| 1775 | (__v4sf) __B, |
| 1776 | (__v4sf) __C, |
| 1777 | (__mmask8) __U); |
| 1778 | } |
| 1779 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1780 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1781 | _mm256_mask3_fmsubadd_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) |
| 1782 | { |
| 1783 | return (__m256) __builtin_ia32_vfmsubaddps256_mask3 ((__v8sf) __A, |
| 1784 | (__v8sf) __B, |
| 1785 | (__v8sf) __C, |
| 1786 | (__mmask8) __U); |
| 1787 | } |
| 1788 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1789 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1790 | _mm_mask_fnmadd_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) |
| 1791 | { |
| 1792 | return (__m128d) __builtin_ia32_vfnmaddpd128_mask ((__v2df) __A, |
| 1793 | (__v2df) __B, |
| 1794 | (__v2df) __C, |
| 1795 | (__mmask8) __U); |
| 1796 | } |
| 1797 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1798 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1799 | _mm256_mask_fnmadd_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) |
| 1800 | { |
| 1801 | return (__m256d) __builtin_ia32_vfnmaddpd256_mask ((__v4df) __A, |
| 1802 | (__v4df) __B, |
| 1803 | (__v4df) __C, |
| 1804 | (__mmask8) __U); |
| 1805 | } |
| 1806 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1807 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1808 | _mm_mask_fnmadd_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) |
| 1809 | { |
| 1810 | return (__m128) __builtin_ia32_vfnmaddps128_mask ((__v4sf) __A, |
| 1811 | (__v4sf) __B, |
| 1812 | (__v4sf) __C, |
| 1813 | (__mmask8) __U); |
| 1814 | } |
| 1815 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1816 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1817 | _mm256_mask_fnmadd_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) |
| 1818 | { |
| 1819 | return (__m256) __builtin_ia32_vfnmaddps256_mask ((__v8sf) __A, |
| 1820 | (__v8sf) __B, |
| 1821 | (__v8sf) __C, |
| 1822 | (__mmask8) __U); |
| 1823 | } |
| 1824 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1825 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1826 | _mm_mask_fnmsub_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) |
| 1827 | { |
| 1828 | return (__m128d) __builtin_ia32_vfnmsubpd128_mask ((__v2df) __A, |
| 1829 | (__v2df) __B, |
| 1830 | (__v2df) __C, |
| 1831 | (__mmask8) __U); |
| 1832 | } |
| 1833 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1834 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1835 | _mm_mask3_fnmsub_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) |
| 1836 | { |
| 1837 | return (__m128d) __builtin_ia32_vfnmsubpd128_mask3 ((__v2df) __A, |
| 1838 | (__v2df) __B, |
| 1839 | (__v2df) __C, |
| 1840 | (__mmask8) __U); |
| 1841 | } |
| 1842 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1843 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1844 | _mm256_mask_fnmsub_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) |
| 1845 | { |
| 1846 | return (__m256d) __builtin_ia32_vfnmsubpd256_mask ((__v4df) __A, |
| 1847 | (__v4df) __B, |
| 1848 | (__v4df) __C, |
| 1849 | (__mmask8) __U); |
| 1850 | } |
| 1851 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1852 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1853 | _mm256_mask3_fnmsub_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) |
| 1854 | { |
| 1855 | return (__m256d) __builtin_ia32_vfnmsubpd256_mask3 ((__v4df) __A, |
| 1856 | (__v4df) __B, |
| 1857 | (__v4df) __C, |
| 1858 | (__mmask8) __U); |
| 1859 | } |
| 1860 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1861 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1862 | _mm_mask_fnmsub_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) |
| 1863 | { |
| 1864 | return (__m128) __builtin_ia32_vfnmsubps128_mask ((__v4sf) __A, |
| 1865 | (__v4sf) __B, |
| 1866 | (__v4sf) __C, |
| 1867 | (__mmask8) __U); |
| 1868 | } |
| 1869 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1870 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1871 | _mm_mask3_fnmsub_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) |
| 1872 | { |
| 1873 | return (__m128) __builtin_ia32_vfnmsubps128_mask3 ((__v4sf) __A, |
| 1874 | (__v4sf) __B, |
| 1875 | (__v4sf) __C, |
| 1876 | (__mmask8) __U); |
| 1877 | } |
| 1878 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1879 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1880 | _mm256_mask_fnmsub_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) |
| 1881 | { |
| 1882 | return (__m256) __builtin_ia32_vfnmsubps256_mask ((__v8sf) __A, |
| 1883 | (__v8sf) __B, |
| 1884 | (__v8sf) __C, |
| 1885 | (__mmask8) __U); |
| 1886 | } |
| 1887 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1888 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1889 | _mm256_mask3_fnmsub_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) |
| 1890 | { |
| 1891 | return (__m256) __builtin_ia32_vfnmsubps256_mask3 ((__v8sf) __A, |
| 1892 | (__v8sf) __B, |
| 1893 | (__v8sf) __C, |
| 1894 | (__mmask8) __U); |
| 1895 | } |
| 1896 | |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1897 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 1898 | _mm_mask_add_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 1899 | return (__m128d) __builtin_ia32_addpd128_mask ((__v2df) __A, |
| 1900 | (__v2df) __B, |
| 1901 | (__v2df) __W, |
| 1902 | (__mmask8) __U); |
| 1903 | } |
| 1904 | |
| 1905 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 1906 | _mm_maskz_add_pd (__mmask8 __U, __m128d __A, __m128d __B) { |
| 1907 | return (__m128d) __builtin_ia32_addpd128_mask ((__v2df) __A, |
| 1908 | (__v2df) __B, |
| 1909 | (__v2df) |
| 1910 | _mm_setzero_pd (), |
| 1911 | (__mmask8) __U); |
| 1912 | } |
| 1913 | |
| 1914 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 1915 | _mm256_mask_add_pd (__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) { |
| 1916 | return (__m256d) __builtin_ia32_addpd256_mask ((__v4df) __A, |
| 1917 | (__v4df) __B, |
| 1918 | (__v4df) __W, |
| 1919 | (__mmask8) __U); |
| 1920 | } |
| 1921 | |
| 1922 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 1923 | _mm256_maskz_add_pd (__mmask8 __U, __m256d __A, __m256d __B) { |
| 1924 | return (__m256d) __builtin_ia32_addpd256_mask ((__v4df) __A, |
| 1925 | (__v4df) __B, |
| 1926 | (__v4df) |
| 1927 | _mm256_setzero_pd (), |
| 1928 | (__mmask8) __U); |
| 1929 | } |
| 1930 | |
| 1931 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 1932 | _mm_mask_add_ps (__m128 __W, __mmask16 __U, __m128 __A, __m128 __B) { |
| 1933 | return (__m128) __builtin_ia32_addps128_mask ((__v4sf) __A, |
| 1934 | (__v4sf) __B, |
| 1935 | (__v4sf) __W, |
| 1936 | (__mmask8) __U); |
| 1937 | } |
| 1938 | |
| 1939 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 1940 | _mm_maskz_add_ps (__mmask16 __U, __m128 __A, __m128 __B) { |
| 1941 | return (__m128) __builtin_ia32_addps128_mask ((__v4sf) __A, |
| 1942 | (__v4sf) __B, |
| 1943 | (__v4sf) |
| 1944 | _mm_setzero_ps (), |
| 1945 | (__mmask8) __U); |
| 1946 | } |
| 1947 | |
| 1948 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 1949 | _mm256_mask_add_ps (__m256 __W, __mmask16 __U, __m256 __A, __m256 __B) { |
| 1950 | return (__m256) __builtin_ia32_addps256_mask ((__v8sf) __A, |
| 1951 | (__v8sf) __B, |
| 1952 | (__v8sf) __W, |
| 1953 | (__mmask8) __U); |
| 1954 | } |
| 1955 | |
| 1956 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 1957 | _mm256_maskz_add_ps (__mmask16 __U, __m256 __A, __m256 __B) { |
| 1958 | return (__m256) __builtin_ia32_addps256_mask ((__v8sf) __A, |
| 1959 | (__v8sf) __B, |
| 1960 | (__v8sf) |
| 1961 | _mm256_setzero_ps (), |
| 1962 | (__mmask8) __U); |
| 1963 | } |
| 1964 | |
| 1965 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1966 | _mm_mask_blend_epi32 (__mmask8 __U, __m128i __A, __m128i __W) { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1967 | return (__m128i) __builtin_ia32_selectd_128 ((__mmask8) __U, |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1968 | (__v4si) __W, |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1969 | (__v4si) __A); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1970 | } |
| 1971 | |
| 1972 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1973 | _mm256_mask_blend_epi32 (__mmask8 __U, __m256i __A, __m256i __W) { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1974 | return (__m256i) __builtin_ia32_selectd_256 ((__mmask8) __U, |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1975 | (__v8si) __W, |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1976 | (__v8si) __A); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1977 | } |
| 1978 | |
| 1979 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 1980 | _mm_mask_blend_pd (__mmask8 __U, __m128d __A, __m128d __W) { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1981 | return (__m128d) __builtin_ia32_selectpd_128 ((__mmask8) __U, |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1982 | (__v2df) __W, |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1983 | (__v2df) __A); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1984 | } |
| 1985 | |
| 1986 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 1987 | _mm256_mask_blend_pd (__mmask8 __U, __m256d __A, __m256d __W) { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1988 | return (__m256d) __builtin_ia32_selectpd_256 ((__mmask8) __U, |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1989 | (__v4df) __W, |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1990 | (__v4df) __A); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1991 | } |
| 1992 | |
| 1993 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 1994 | _mm_mask_blend_ps (__mmask8 __U, __m128 __A, __m128 __W) { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1995 | return (__m128) __builtin_ia32_selectps_128 ((__mmask8) __U, |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1996 | (__v4sf) __W, |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1997 | (__v4sf) __A); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1998 | } |
| 1999 | |
| 2000 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2001 | _mm256_mask_blend_ps (__mmask8 __U, __m256 __A, __m256 __W) { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 2002 | return (__m256) __builtin_ia32_selectps_256 ((__mmask8) __U, |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2003 | (__v8sf) __W, |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 2004 | (__v8sf) __A); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2005 | } |
| 2006 | |
| 2007 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2008 | _mm_mask_blend_epi64 (__mmask8 __U, __m128i __A, __m128i __W) { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 2009 | return (__m128i) __builtin_ia32_selectq_128 ((__mmask8) __U, |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2010 | (__v2di) __W, |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 2011 | (__v2di) __A); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2012 | } |
| 2013 | |
| 2014 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2015 | _mm256_mask_blend_epi64 (__mmask8 __U, __m256i __A, __m256i __W) { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 2016 | return (__m256i) __builtin_ia32_selectq_256 ((__mmask8) __U, |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2017 | (__v4di) __W, |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 2018 | (__v4di) __A); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2019 | } |
| 2020 | |
| 2021 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2022 | _mm_mask_compress_pd (__m128d __W, __mmask8 __U, __m128d __A) { |
| 2023 | return (__m128d) __builtin_ia32_compressdf128_mask ((__v2df) __A, |
| 2024 | (__v2df) __W, |
| 2025 | (__mmask8) __U); |
| 2026 | } |
| 2027 | |
| 2028 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2029 | _mm_maskz_compress_pd (__mmask8 __U, __m128d __A) { |
| 2030 | return (__m128d) __builtin_ia32_compressdf128_mask ((__v2df) __A, |
| 2031 | (__v2df) |
| 2032 | _mm_setzero_pd (), |
| 2033 | (__mmask8) __U); |
| 2034 | } |
| 2035 | |
| 2036 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2037 | _mm256_mask_compress_pd (__m256d __W, __mmask8 __U, __m256d __A) { |
| 2038 | return (__m256d) __builtin_ia32_compressdf256_mask ((__v4df) __A, |
| 2039 | (__v4df) __W, |
| 2040 | (__mmask8) __U); |
| 2041 | } |
| 2042 | |
| 2043 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2044 | _mm256_maskz_compress_pd (__mmask8 __U, __m256d __A) { |
| 2045 | return (__m256d) __builtin_ia32_compressdf256_mask ((__v4df) __A, |
| 2046 | (__v4df) |
| 2047 | _mm256_setzero_pd (), |
| 2048 | (__mmask8) __U); |
| 2049 | } |
| 2050 | |
| 2051 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2052 | _mm_mask_compress_epi64 (__m128i __W, __mmask8 __U, __m128i __A) { |
| 2053 | return (__m128i) __builtin_ia32_compressdi128_mask ((__v2di) __A, |
| 2054 | (__v2di) __W, |
| 2055 | (__mmask8) __U); |
| 2056 | } |
| 2057 | |
| 2058 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2059 | _mm_maskz_compress_epi64 (__mmask8 __U, __m128i __A) { |
| 2060 | return (__m128i) __builtin_ia32_compressdi128_mask ((__v2di) __A, |
| 2061 | (__v2di) |
| 2062 | _mm_setzero_si128 (), |
| 2063 | (__mmask8) __U); |
| 2064 | } |
| 2065 | |
| 2066 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2067 | _mm256_mask_compress_epi64 (__m256i __W, __mmask8 __U, __m256i __A) { |
| 2068 | return (__m256i) __builtin_ia32_compressdi256_mask ((__v4di) __A, |
| 2069 | (__v4di) __W, |
| 2070 | (__mmask8) __U); |
| 2071 | } |
| 2072 | |
| 2073 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2074 | _mm256_maskz_compress_epi64 (__mmask8 __U, __m256i __A) { |
| 2075 | return (__m256i) __builtin_ia32_compressdi256_mask ((__v4di) __A, |
| 2076 | (__v4di) |
| 2077 | _mm256_setzero_si256 (), |
| 2078 | (__mmask8) __U); |
| 2079 | } |
| 2080 | |
| 2081 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2082 | _mm_mask_compress_ps (__m128 __W, __mmask8 __U, __m128 __A) { |
| 2083 | return (__m128) __builtin_ia32_compresssf128_mask ((__v4sf) __A, |
| 2084 | (__v4sf) __W, |
| 2085 | (__mmask8) __U); |
| 2086 | } |
| 2087 | |
| 2088 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2089 | _mm_maskz_compress_ps (__mmask8 __U, __m128 __A) { |
| 2090 | return (__m128) __builtin_ia32_compresssf128_mask ((__v4sf) __A, |
| 2091 | (__v4sf) |
| 2092 | _mm_setzero_ps (), |
| 2093 | (__mmask8) __U); |
| 2094 | } |
| 2095 | |
| 2096 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2097 | _mm256_mask_compress_ps (__m256 __W, __mmask8 __U, __m256 __A) { |
| 2098 | return (__m256) __builtin_ia32_compresssf256_mask ((__v8sf) __A, |
| 2099 | (__v8sf) __W, |
| 2100 | (__mmask8) __U); |
| 2101 | } |
| 2102 | |
| 2103 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2104 | _mm256_maskz_compress_ps (__mmask8 __U, __m256 __A) { |
| 2105 | return (__m256) __builtin_ia32_compresssf256_mask ((__v8sf) __A, |
| 2106 | (__v8sf) |
| 2107 | _mm256_setzero_ps (), |
| 2108 | (__mmask8) __U); |
| 2109 | } |
| 2110 | |
| 2111 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2112 | _mm_mask_compress_epi32 (__m128i __W, __mmask8 __U, __m128i __A) { |
| 2113 | return (__m128i) __builtin_ia32_compresssi128_mask ((__v4si) __A, |
| 2114 | (__v4si) __W, |
| 2115 | (__mmask8) __U); |
| 2116 | } |
| 2117 | |
| 2118 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2119 | _mm_maskz_compress_epi32 (__mmask8 __U, __m128i __A) { |
| 2120 | return (__m128i) __builtin_ia32_compresssi128_mask ((__v4si) __A, |
| 2121 | (__v4si) |
| 2122 | _mm_setzero_si128 (), |
| 2123 | (__mmask8) __U); |
| 2124 | } |
| 2125 | |
| 2126 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2127 | _mm256_mask_compress_epi32 (__m256i __W, __mmask8 __U, __m256i __A) { |
| 2128 | return (__m256i) __builtin_ia32_compresssi256_mask ((__v8si) __A, |
| 2129 | (__v8si) __W, |
| 2130 | (__mmask8) __U); |
| 2131 | } |
| 2132 | |
| 2133 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2134 | _mm256_maskz_compress_epi32 (__mmask8 __U, __m256i __A) { |
| 2135 | return (__m256i) __builtin_ia32_compresssi256_mask ((__v8si) __A, |
| 2136 | (__v8si) |
| 2137 | _mm256_setzero_si256 (), |
| 2138 | (__mmask8) __U); |
| 2139 | } |
| 2140 | |
| 2141 | static __inline__ void __DEFAULT_FN_ATTRS |
| 2142 | _mm_mask_compressstoreu_pd (void *__P, __mmask8 __U, __m128d __A) { |
| 2143 | __builtin_ia32_compressstoredf128_mask ((__v2df *) __P, |
| 2144 | (__v2df) __A, |
| 2145 | (__mmask8) __U); |
| 2146 | } |
| 2147 | |
| 2148 | static __inline__ void __DEFAULT_FN_ATTRS |
| 2149 | _mm256_mask_compressstoreu_pd (void *__P, __mmask8 __U, __m256d __A) { |
| 2150 | __builtin_ia32_compressstoredf256_mask ((__v4df *) __P, |
| 2151 | (__v4df) __A, |
| 2152 | (__mmask8) __U); |
| 2153 | } |
| 2154 | |
| 2155 | static __inline__ void __DEFAULT_FN_ATTRS |
| 2156 | _mm_mask_compressstoreu_epi64 (void *__P, __mmask8 __U, __m128i __A) { |
| 2157 | __builtin_ia32_compressstoredi128_mask ((__v2di *) __P, |
| 2158 | (__v2di) __A, |
| 2159 | (__mmask8) __U); |
| 2160 | } |
| 2161 | |
| 2162 | static __inline__ void __DEFAULT_FN_ATTRS |
| 2163 | _mm256_mask_compressstoreu_epi64 (void *__P, __mmask8 __U, __m256i __A) { |
| 2164 | __builtin_ia32_compressstoredi256_mask ((__v4di *) __P, |
| 2165 | (__v4di) __A, |
| 2166 | (__mmask8) __U); |
| 2167 | } |
| 2168 | |
| 2169 | static __inline__ void __DEFAULT_FN_ATTRS |
| 2170 | _mm_mask_compressstoreu_ps (void *__P, __mmask8 __U, __m128 __A) { |
| 2171 | __builtin_ia32_compressstoresf128_mask ((__v4sf *) __P, |
| 2172 | (__v4sf) __A, |
| 2173 | (__mmask8) __U); |
| 2174 | } |
| 2175 | |
| 2176 | static __inline__ void __DEFAULT_FN_ATTRS |
| 2177 | _mm256_mask_compressstoreu_ps (void *__P, __mmask8 __U, __m256 __A) { |
| 2178 | __builtin_ia32_compressstoresf256_mask ((__v8sf *) __P, |
| 2179 | (__v8sf) __A, |
| 2180 | (__mmask8) __U); |
| 2181 | } |
| 2182 | |
| 2183 | static __inline__ void __DEFAULT_FN_ATTRS |
| 2184 | _mm_mask_compressstoreu_epi32 (void *__P, __mmask8 __U, __m128i __A) { |
| 2185 | __builtin_ia32_compressstoresi128_mask ((__v4si *) __P, |
| 2186 | (__v4si) __A, |
| 2187 | (__mmask8) __U); |
| 2188 | } |
| 2189 | |
| 2190 | static __inline__ void __DEFAULT_FN_ATTRS |
| 2191 | _mm256_mask_compressstoreu_epi32 (void *__P, __mmask8 __U, __m256i __A) { |
| 2192 | __builtin_ia32_compressstoresi256_mask ((__v8si *) __P, |
| 2193 | (__v8si) __A, |
| 2194 | (__mmask8) __U); |
| 2195 | } |
| 2196 | |
| 2197 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2198 | _mm_mask_cvtepi32_pd (__m128d __W, __mmask8 __U, __m128i __A) { |
| 2199 | return (__m128d) __builtin_ia32_cvtdq2pd128_mask ((__v4si) __A, |
| 2200 | (__v2df) __W, |
| 2201 | (__mmask8) __U); |
| 2202 | } |
| 2203 | |
| 2204 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2205 | _mm_maskz_cvtepi32_pd (__mmask8 __U, __m128i __A) { |
| 2206 | return (__m128d) __builtin_ia32_cvtdq2pd128_mask ((__v4si) __A, |
| 2207 | (__v2df) |
| 2208 | _mm_setzero_pd (), |
| 2209 | (__mmask8) __U); |
| 2210 | } |
| 2211 | |
| 2212 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2213 | _mm256_mask_cvtepi32_pd (__m256d __W, __mmask8 __U, __m128i __A) { |
| 2214 | return (__m256d) __builtin_ia32_cvtdq2pd256_mask ((__v4si) __A, |
| 2215 | (__v4df) __W, |
| 2216 | (__mmask8) __U); |
| 2217 | } |
| 2218 | |
| 2219 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2220 | _mm256_maskz_cvtepi32_pd (__mmask8 __U, __m128i __A) { |
| 2221 | return (__m256d) __builtin_ia32_cvtdq2pd256_mask ((__v4si) __A, |
| 2222 | (__v4df) |
| 2223 | _mm256_setzero_pd (), |
| 2224 | (__mmask8) __U); |
| 2225 | } |
| 2226 | |
| 2227 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2228 | _mm_mask_cvtepi32_ps (__m128 __W, __mmask8 __U, __m128i __A) { |
| 2229 | return (__m128) __builtin_ia32_cvtdq2ps128_mask ((__v4si) __A, |
| 2230 | (__v4sf) __W, |
| 2231 | (__mmask8) __U); |
| 2232 | } |
| 2233 | |
| 2234 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2235 | _mm_maskz_cvtepi32_ps (__mmask16 __U, __m128i __A) { |
| 2236 | return (__m128) __builtin_ia32_cvtdq2ps128_mask ((__v4si) __A, |
| 2237 | (__v4sf) |
| 2238 | _mm_setzero_ps (), |
| 2239 | (__mmask8) __U); |
| 2240 | } |
| 2241 | |
| 2242 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2243 | _mm256_mask_cvtepi32_ps (__m256 __W, __mmask8 __U, __m256i __A) { |
| 2244 | return (__m256) __builtin_ia32_cvtdq2ps256_mask ((__v8si) __A, |
| 2245 | (__v8sf) __W, |
| 2246 | (__mmask8) __U); |
| 2247 | } |
| 2248 | |
| 2249 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2250 | _mm256_maskz_cvtepi32_ps (__mmask16 __U, __m256i __A) { |
| 2251 | return (__m256) __builtin_ia32_cvtdq2ps256_mask ((__v8si) __A, |
| 2252 | (__v8sf) |
| 2253 | _mm256_setzero_ps (), |
| 2254 | (__mmask8) __U); |
| 2255 | } |
| 2256 | |
| 2257 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2258 | _mm_mask_cvtpd_epi32 (__m128i __W, __mmask8 __U, __m128d __A) { |
| 2259 | return (__m128i) __builtin_ia32_cvtpd2dq128_mask ((__v2df) __A, |
| 2260 | (__v4si) __W, |
| 2261 | (__mmask8) __U); |
| 2262 | } |
| 2263 | |
| 2264 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2265 | _mm_maskz_cvtpd_epi32 (__mmask8 __U, __m128d __A) { |
| 2266 | return (__m128i) __builtin_ia32_cvtpd2dq128_mask ((__v2df) __A, |
| 2267 | (__v4si) |
| 2268 | _mm_setzero_si128 (), |
| 2269 | (__mmask8) __U); |
| 2270 | } |
| 2271 | |
| 2272 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2273 | _mm256_mask_cvtpd_epi32 (__m128i __W, __mmask8 __U, __m256d __A) { |
| 2274 | return (__m128i) __builtin_ia32_cvtpd2dq256_mask ((__v4df) __A, |
| 2275 | (__v4si) __W, |
| 2276 | (__mmask8) __U); |
| 2277 | } |
| 2278 | |
| 2279 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2280 | _mm256_maskz_cvtpd_epi32 (__mmask8 __U, __m256d __A) { |
| 2281 | return (__m128i) __builtin_ia32_cvtpd2dq256_mask ((__v4df) __A, |
| 2282 | (__v4si) |
| 2283 | _mm_setzero_si128 (), |
| 2284 | (__mmask8) __U); |
| 2285 | } |
| 2286 | |
| 2287 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2288 | _mm_mask_cvtpd_ps (__m128 __W, __mmask8 __U, __m128d __A) { |
| 2289 | return (__m128) __builtin_ia32_cvtpd2ps_mask ((__v2df) __A, |
| 2290 | (__v4sf) __W, |
| 2291 | (__mmask8) __U); |
| 2292 | } |
| 2293 | |
| 2294 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2295 | _mm_maskz_cvtpd_ps (__mmask8 __U, __m128d __A) { |
| 2296 | return (__m128) __builtin_ia32_cvtpd2ps_mask ((__v2df) __A, |
| 2297 | (__v4sf) |
| 2298 | _mm_setzero_ps (), |
| 2299 | (__mmask8) __U); |
| 2300 | } |
| 2301 | |
| 2302 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2303 | _mm256_mask_cvtpd_ps (__m128 __W, __mmask8 __U, __m256d __A) { |
| 2304 | return (__m128) __builtin_ia32_cvtpd2ps256_mask ((__v4df) __A, |
| 2305 | (__v4sf) __W, |
| 2306 | (__mmask8) __U); |
| 2307 | } |
| 2308 | |
| 2309 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2310 | _mm256_maskz_cvtpd_ps (__mmask8 __U, __m256d __A) { |
| 2311 | return (__m128) __builtin_ia32_cvtpd2ps256_mask ((__v4df) __A, |
| 2312 | (__v4sf) |
| 2313 | _mm_setzero_ps (), |
| 2314 | (__mmask8) __U); |
| 2315 | } |
| 2316 | |
| 2317 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2318 | _mm_cvtpd_epu32 (__m128d __A) { |
| 2319 | return (__m128i) __builtin_ia32_cvtpd2udq128_mask ((__v2df) __A, |
| 2320 | (__v4si) |
| 2321 | _mm_setzero_si128 (), |
| 2322 | (__mmask8) -1); |
| 2323 | } |
| 2324 | |
| 2325 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2326 | _mm_mask_cvtpd_epu32 (__m128i __W, __mmask8 __U, __m128d __A) { |
| 2327 | return (__m128i) __builtin_ia32_cvtpd2udq128_mask ((__v2df) __A, |
| 2328 | (__v4si) __W, |
| 2329 | (__mmask8) __U); |
| 2330 | } |
| 2331 | |
| 2332 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2333 | _mm_maskz_cvtpd_epu32 (__mmask8 __U, __m128d __A) { |
| 2334 | return (__m128i) __builtin_ia32_cvtpd2udq128_mask ((__v2df) __A, |
| 2335 | (__v4si) |
| 2336 | _mm_setzero_si128 (), |
| 2337 | (__mmask8) __U); |
| 2338 | } |
| 2339 | |
| 2340 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2341 | _mm256_cvtpd_epu32 (__m256d __A) { |
| 2342 | return (__m128i) __builtin_ia32_cvtpd2udq256_mask ((__v4df) __A, |
| 2343 | (__v4si) |
| 2344 | _mm_setzero_si128 (), |
| 2345 | (__mmask8) -1); |
| 2346 | } |
| 2347 | |
| 2348 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2349 | _mm256_mask_cvtpd_epu32 (__m128i __W, __mmask8 __U, __m256d __A) { |
| 2350 | return (__m128i) __builtin_ia32_cvtpd2udq256_mask ((__v4df) __A, |
| 2351 | (__v4si) __W, |
| 2352 | (__mmask8) __U); |
| 2353 | } |
| 2354 | |
| 2355 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2356 | _mm256_maskz_cvtpd_epu32 (__mmask8 __U, __m256d __A) { |
| 2357 | return (__m128i) __builtin_ia32_cvtpd2udq256_mask ((__v4df) __A, |
| 2358 | (__v4si) |
| 2359 | _mm_setzero_si128 (), |
| 2360 | (__mmask8) __U); |
| 2361 | } |
| 2362 | |
| 2363 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2364 | _mm_mask_cvtps_epi32 (__m128i __W, __mmask8 __U, __m128 __A) { |
| 2365 | return (__m128i) __builtin_ia32_cvtps2dq128_mask ((__v4sf) __A, |
| 2366 | (__v4si) __W, |
| 2367 | (__mmask8) __U); |
| 2368 | } |
| 2369 | |
| 2370 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2371 | _mm_maskz_cvtps_epi32 (__mmask8 __U, __m128 __A) { |
| 2372 | return (__m128i) __builtin_ia32_cvtps2dq128_mask ((__v4sf) __A, |
| 2373 | (__v4si) |
| 2374 | _mm_setzero_si128 (), |
| 2375 | (__mmask8) __U); |
| 2376 | } |
| 2377 | |
| 2378 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2379 | _mm256_mask_cvtps_epi32 (__m256i __W, __mmask8 __U, __m256 __A) { |
| 2380 | return (__m256i) __builtin_ia32_cvtps2dq256_mask ((__v8sf) __A, |
| 2381 | (__v8si) __W, |
| 2382 | (__mmask8) __U); |
| 2383 | } |
| 2384 | |
| 2385 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2386 | _mm256_maskz_cvtps_epi32 (__mmask8 __U, __m256 __A) { |
| 2387 | return (__m256i) __builtin_ia32_cvtps2dq256_mask ((__v8sf) __A, |
| 2388 | (__v8si) |
| 2389 | _mm256_setzero_si256 (), |
| 2390 | (__mmask8) __U); |
| 2391 | } |
| 2392 | |
| 2393 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2394 | _mm_mask_cvtps_pd (__m128d __W, __mmask8 __U, __m128 __A) { |
| 2395 | return (__m128d) __builtin_ia32_cvtps2pd128_mask ((__v4sf) __A, |
| 2396 | (__v2df) __W, |
| 2397 | (__mmask8) __U); |
| 2398 | } |
| 2399 | |
| 2400 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2401 | _mm_maskz_cvtps_pd (__mmask8 __U, __m128 __A) { |
| 2402 | return (__m128d) __builtin_ia32_cvtps2pd128_mask ((__v4sf) __A, |
| 2403 | (__v2df) |
| 2404 | _mm_setzero_pd (), |
| 2405 | (__mmask8) __U); |
| 2406 | } |
| 2407 | |
| 2408 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2409 | _mm256_mask_cvtps_pd (__m256d __W, __mmask8 __U, __m128 __A) { |
| 2410 | return (__m256d) __builtin_ia32_cvtps2pd256_mask ((__v4sf) __A, |
| 2411 | (__v4df) __W, |
| 2412 | (__mmask8) __U); |
| 2413 | } |
| 2414 | |
| 2415 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2416 | _mm256_maskz_cvtps_pd (__mmask8 __U, __m128 __A) { |
| 2417 | return (__m256d) __builtin_ia32_cvtps2pd256_mask ((__v4sf) __A, |
| 2418 | (__v4df) |
| 2419 | _mm256_setzero_pd (), |
| 2420 | (__mmask8) __U); |
| 2421 | } |
| 2422 | |
| 2423 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2424 | _mm_cvtps_epu32 (__m128 __A) { |
| 2425 | return (__m128i) __builtin_ia32_cvtps2udq128_mask ((__v4sf) __A, |
| 2426 | (__v4si) |
| 2427 | _mm_setzero_si128 (), |
| 2428 | (__mmask8) -1); |
| 2429 | } |
| 2430 | |
| 2431 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2432 | _mm_mask_cvtps_epu32 (__m128i __W, __mmask8 __U, __m128 __A) { |
| 2433 | return (__m128i) __builtin_ia32_cvtps2udq128_mask ((__v4sf) __A, |
| 2434 | (__v4si) __W, |
| 2435 | (__mmask8) __U); |
| 2436 | } |
| 2437 | |
| 2438 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2439 | _mm_maskz_cvtps_epu32 (__mmask8 __U, __m128 __A) { |
| 2440 | return (__m128i) __builtin_ia32_cvtps2udq128_mask ((__v4sf) __A, |
| 2441 | (__v4si) |
| 2442 | _mm_setzero_si128 (), |
| 2443 | (__mmask8) __U); |
| 2444 | } |
| 2445 | |
| 2446 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2447 | _mm256_cvtps_epu32 (__m256 __A) { |
| 2448 | return (__m256i) __builtin_ia32_cvtps2udq256_mask ((__v8sf) __A, |
| 2449 | (__v8si) |
| 2450 | _mm256_setzero_si256 (), |
| 2451 | (__mmask8) -1); |
| 2452 | } |
| 2453 | |
| 2454 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2455 | _mm256_mask_cvtps_epu32 (__m256i __W, __mmask8 __U, __m256 __A) { |
| 2456 | return (__m256i) __builtin_ia32_cvtps2udq256_mask ((__v8sf) __A, |
| 2457 | (__v8si) __W, |
| 2458 | (__mmask8) __U); |
| 2459 | } |
| 2460 | |
| 2461 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2462 | _mm256_maskz_cvtps_epu32 (__mmask8 __U, __m256 __A) { |
| 2463 | return (__m256i) __builtin_ia32_cvtps2udq256_mask ((__v8sf) __A, |
| 2464 | (__v8si) |
| 2465 | _mm256_setzero_si256 (), |
| 2466 | (__mmask8) __U); |
| 2467 | } |
| 2468 | |
| 2469 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2470 | _mm_mask_cvttpd_epi32 (__m128i __W, __mmask8 __U, __m128d __A) { |
| 2471 | return (__m128i) __builtin_ia32_cvttpd2dq128_mask ((__v2df) __A, |
| 2472 | (__v4si) __W, |
| 2473 | (__mmask8) __U); |
| 2474 | } |
| 2475 | |
| 2476 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2477 | _mm_maskz_cvttpd_epi32 (__mmask8 __U, __m128d __A) { |
| 2478 | return (__m128i) __builtin_ia32_cvttpd2dq128_mask ((__v2df) __A, |
| 2479 | (__v4si) |
| 2480 | _mm_setzero_si128 (), |
| 2481 | (__mmask8) __U); |
| 2482 | } |
| 2483 | |
| 2484 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2485 | _mm256_mask_cvttpd_epi32 (__m128i __W, __mmask8 __U, __m256d __A) { |
| 2486 | return (__m128i) __builtin_ia32_cvttpd2dq256_mask ((__v4df) __A, |
| 2487 | (__v4si) __W, |
| 2488 | (__mmask8) __U); |
| 2489 | } |
| 2490 | |
| 2491 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2492 | _mm256_maskz_cvttpd_epi32 (__mmask8 __U, __m256d __A) { |
| 2493 | return (__m128i) __builtin_ia32_cvttpd2dq256_mask ((__v4df) __A, |
| 2494 | (__v4si) |
| 2495 | _mm_setzero_si128 (), |
| 2496 | (__mmask8) __U); |
| 2497 | } |
| 2498 | |
| 2499 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2500 | _mm_cvttpd_epu32 (__m128d __A) { |
| 2501 | return (__m128i) __builtin_ia32_cvttpd2udq128_mask ((__v2df) __A, |
| 2502 | (__v4si) |
| 2503 | _mm_setzero_si128 (), |
| 2504 | (__mmask8) -1); |
| 2505 | } |
| 2506 | |
| 2507 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2508 | _mm_mask_cvttpd_epu32 (__m128i __W, __mmask8 __U, __m128d __A) { |
| 2509 | return (__m128i) __builtin_ia32_cvttpd2udq128_mask ((__v2df) __A, |
| 2510 | (__v4si) __W, |
| 2511 | (__mmask8) __U); |
| 2512 | } |
| 2513 | |
| 2514 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2515 | _mm_maskz_cvttpd_epu32 (__mmask8 __U, __m128d __A) { |
| 2516 | return (__m128i) __builtin_ia32_cvttpd2udq128_mask ((__v2df) __A, |
| 2517 | (__v4si) |
| 2518 | _mm_setzero_si128 (), |
| 2519 | (__mmask8) __U); |
| 2520 | } |
| 2521 | |
| 2522 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2523 | _mm256_cvttpd_epu32 (__m256d __A) { |
| 2524 | return (__m128i) __builtin_ia32_cvttpd2udq256_mask ((__v4df) __A, |
| 2525 | (__v4si) |
| 2526 | _mm_setzero_si128 (), |
| 2527 | (__mmask8) -1); |
| 2528 | } |
| 2529 | |
| 2530 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2531 | _mm256_mask_cvttpd_epu32 (__m128i __W, __mmask8 __U, __m256d __A) { |
| 2532 | return (__m128i) __builtin_ia32_cvttpd2udq256_mask ((__v4df) __A, |
| 2533 | (__v4si) __W, |
| 2534 | (__mmask8) __U); |
| 2535 | } |
| 2536 | |
| 2537 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2538 | _mm256_maskz_cvttpd_epu32 (__mmask8 __U, __m256d __A) { |
| 2539 | return (__m128i) __builtin_ia32_cvttpd2udq256_mask ((__v4df) __A, |
| 2540 | (__v4si) |
| 2541 | _mm_setzero_si128 (), |
| 2542 | (__mmask8) __U); |
| 2543 | } |
| 2544 | |
| 2545 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2546 | _mm_mask_cvttps_epi32 (__m128i __W, __mmask8 __U, __m128 __A) { |
| 2547 | return (__m128i) __builtin_ia32_cvttps2dq128_mask ((__v4sf) __A, |
| 2548 | (__v4si) __W, |
| 2549 | (__mmask8) __U); |
| 2550 | } |
| 2551 | |
| 2552 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2553 | _mm_maskz_cvttps_epi32 (__mmask8 __U, __m128 __A) { |
| 2554 | return (__m128i) __builtin_ia32_cvttps2dq128_mask ((__v4sf) __A, |
| 2555 | (__v4si) |
| 2556 | _mm_setzero_si128 (), |
| 2557 | (__mmask8) __U); |
| 2558 | } |
| 2559 | |
| 2560 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2561 | _mm256_mask_cvttps_epi32 (__m256i __W, __mmask8 __U, __m256 __A) { |
| 2562 | return (__m256i) __builtin_ia32_cvttps2dq256_mask ((__v8sf) __A, |
| 2563 | (__v8si) __W, |
| 2564 | (__mmask8) __U); |
| 2565 | } |
| 2566 | |
| 2567 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2568 | _mm256_maskz_cvttps_epi32 (__mmask8 __U, __m256 __A) { |
| 2569 | return (__m256i) __builtin_ia32_cvttps2dq256_mask ((__v8sf) __A, |
| 2570 | (__v8si) |
| 2571 | _mm256_setzero_si256 (), |
| 2572 | (__mmask8) __U); |
| 2573 | } |
| 2574 | |
| 2575 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2576 | _mm_cvttps_epu32 (__m128 __A) { |
| 2577 | return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A, |
| 2578 | (__v4si) |
| 2579 | _mm_setzero_si128 (), |
| 2580 | (__mmask8) -1); |
| 2581 | } |
| 2582 | |
| 2583 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2584 | _mm_mask_cvttps_epu32 (__m128i __W, __mmask8 __U, __m128 __A) { |
| 2585 | return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A, |
| 2586 | (__v4si) __W, |
| 2587 | (__mmask8) __U); |
| 2588 | } |
| 2589 | |
| 2590 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2591 | _mm_maskz_cvttps_epu32 (__mmask8 __U, __m128 __A) { |
| 2592 | return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A, |
| 2593 | (__v4si) |
| 2594 | _mm_setzero_si128 (), |
| 2595 | (__mmask8) __U); |
| 2596 | } |
| 2597 | |
| 2598 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2599 | _mm256_cvttps_epu32 (__m256 __A) { |
| 2600 | return (__m256i) __builtin_ia32_cvttps2udq256_mask ((__v8sf) __A, |
| 2601 | (__v8si) |
| 2602 | _mm256_setzero_si256 (), |
| 2603 | (__mmask8) -1); |
| 2604 | } |
| 2605 | |
| 2606 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2607 | _mm256_mask_cvttps_epu32 (__m256i __W, __mmask8 __U, __m256 __A) { |
| 2608 | return (__m256i) __builtin_ia32_cvttps2udq256_mask ((__v8sf) __A, |
| 2609 | (__v8si) __W, |
| 2610 | (__mmask8) __U); |
| 2611 | } |
| 2612 | |
| 2613 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2614 | _mm256_maskz_cvttps_epu32 (__mmask8 __U, __m256 __A) { |
| 2615 | return (__m256i) __builtin_ia32_cvttps2udq256_mask ((__v8sf) __A, |
| 2616 | (__v8si) |
| 2617 | _mm256_setzero_si256 (), |
| 2618 | (__mmask8) __U); |
| 2619 | } |
| 2620 | |
| 2621 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2622 | _mm_cvtepu32_pd (__m128i __A) { |
| 2623 | return (__m128d) __builtin_ia32_cvtudq2pd128_mask ((__v4si) __A, |
| 2624 | (__v2df) |
| 2625 | _mm_setzero_pd (), |
| 2626 | (__mmask8) -1); |
| 2627 | } |
| 2628 | |
| 2629 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2630 | _mm_mask_cvtepu32_pd (__m128d __W, __mmask8 __U, __m128i __A) { |
| 2631 | return (__m128d) __builtin_ia32_cvtudq2pd128_mask ((__v4si) __A, |
| 2632 | (__v2df) __W, |
| 2633 | (__mmask8) __U); |
| 2634 | } |
| 2635 | |
| 2636 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2637 | _mm_maskz_cvtepu32_pd (__mmask8 __U, __m128i __A) { |
| 2638 | return (__m128d) __builtin_ia32_cvtudq2pd128_mask ((__v4si) __A, |
| 2639 | (__v2df) |
| 2640 | _mm_setzero_pd (), |
| 2641 | (__mmask8) __U); |
| 2642 | } |
| 2643 | |
| 2644 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2645 | _mm256_cvtepu32_pd (__m128i __A) { |
| 2646 | return (__m256d) __builtin_ia32_cvtudq2pd256_mask ((__v4si) __A, |
| 2647 | (__v4df) |
| 2648 | _mm256_setzero_pd (), |
| 2649 | (__mmask8) -1); |
| 2650 | } |
| 2651 | |
| 2652 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2653 | _mm256_mask_cvtepu32_pd (__m256d __W, __mmask8 __U, __m128i __A) { |
| 2654 | return (__m256d) __builtin_ia32_cvtudq2pd256_mask ((__v4si) __A, |
| 2655 | (__v4df) __W, |
| 2656 | (__mmask8) __U); |
| 2657 | } |
| 2658 | |
| 2659 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2660 | _mm256_maskz_cvtepu32_pd (__mmask8 __U, __m128i __A) { |
| 2661 | return (__m256d) __builtin_ia32_cvtudq2pd256_mask ((__v4si) __A, |
| 2662 | (__v4df) |
| 2663 | _mm256_setzero_pd (), |
| 2664 | (__mmask8) __U); |
| 2665 | } |
| 2666 | |
| 2667 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2668 | _mm_cvtepu32_ps (__m128i __A) { |
| 2669 | return (__m128) __builtin_ia32_cvtudq2ps128_mask ((__v4si) __A, |
| 2670 | (__v4sf) |
| 2671 | _mm_setzero_ps (), |
| 2672 | (__mmask8) -1); |
| 2673 | } |
| 2674 | |
| 2675 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2676 | _mm_mask_cvtepu32_ps (__m128 __W, __mmask8 __U, __m128i __A) { |
| 2677 | return (__m128) __builtin_ia32_cvtudq2ps128_mask ((__v4si) __A, |
| 2678 | (__v4sf) __W, |
| 2679 | (__mmask8) __U); |
| 2680 | } |
| 2681 | |
| 2682 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2683 | _mm_maskz_cvtepu32_ps (__mmask8 __U, __m128i __A) { |
| 2684 | return (__m128) __builtin_ia32_cvtudq2ps128_mask ((__v4si) __A, |
| 2685 | (__v4sf) |
| 2686 | _mm_setzero_ps (), |
| 2687 | (__mmask8) __U); |
| 2688 | } |
| 2689 | |
| 2690 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2691 | _mm256_cvtepu32_ps (__m256i __A) { |
| 2692 | return (__m256) __builtin_ia32_cvtudq2ps256_mask ((__v8si) __A, |
| 2693 | (__v8sf) |
| 2694 | _mm256_setzero_ps (), |
| 2695 | (__mmask8) -1); |
| 2696 | } |
| 2697 | |
| 2698 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2699 | _mm256_mask_cvtepu32_ps (__m256 __W, __mmask8 __U, __m256i __A) { |
| 2700 | return (__m256) __builtin_ia32_cvtudq2ps256_mask ((__v8si) __A, |
| 2701 | (__v8sf) __W, |
| 2702 | (__mmask8) __U); |
| 2703 | } |
| 2704 | |
| 2705 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2706 | _mm256_maskz_cvtepu32_ps (__mmask8 __U, __m256i __A) { |
| 2707 | return (__m256) __builtin_ia32_cvtudq2ps256_mask ((__v8si) __A, |
| 2708 | (__v8sf) |
| 2709 | _mm256_setzero_ps (), |
| 2710 | (__mmask8) __U); |
| 2711 | } |
| 2712 | |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2713 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2714 | _mm_mask_div_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 2715 | return (__m128d) __builtin_ia32_divpd_mask ((__v2df) __A, |
| 2716 | (__v2df) __B, |
| 2717 | (__v2df) __W, |
| 2718 | (__mmask8) __U); |
| 2719 | } |
| 2720 | |
| 2721 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2722 | _mm_maskz_div_pd (__mmask8 __U, __m128d __A, __m128d __B) { |
| 2723 | return (__m128d) __builtin_ia32_divpd_mask ((__v2df) __A, |
| 2724 | (__v2df) __B, |
| 2725 | (__v2df) |
| 2726 | _mm_setzero_pd (), |
| 2727 | (__mmask8) __U); |
| 2728 | } |
| 2729 | |
| 2730 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2731 | _mm256_mask_div_pd (__m256d __W, __mmask8 __U, __m256d __A, |
| 2732 | __m256d __B) { |
| 2733 | return (__m256d) __builtin_ia32_divpd256_mask ((__v4df) __A, |
| 2734 | (__v4df) __B, |
| 2735 | (__v4df) __W, |
| 2736 | (__mmask8) __U); |
| 2737 | } |
| 2738 | |
| 2739 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2740 | _mm256_maskz_div_pd (__mmask8 __U, __m256d __A, __m256d __B) { |
| 2741 | return (__m256d) __builtin_ia32_divpd256_mask ((__v4df) __A, |
| 2742 | (__v4df) __B, |
| 2743 | (__v4df) |
| 2744 | _mm256_setzero_pd (), |
| 2745 | (__mmask8) __U); |
| 2746 | } |
| 2747 | |
| 2748 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2749 | _mm_mask_div_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 2750 | return (__m128) __builtin_ia32_divps_mask ((__v4sf) __A, |
| 2751 | (__v4sf) __B, |
| 2752 | (__v4sf) __W, |
| 2753 | (__mmask8) __U); |
| 2754 | } |
| 2755 | |
| 2756 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2757 | _mm_maskz_div_ps (__mmask8 __U, __m128 __A, __m128 __B) { |
| 2758 | return (__m128) __builtin_ia32_divps_mask ((__v4sf) __A, |
| 2759 | (__v4sf) __B, |
| 2760 | (__v4sf) |
| 2761 | _mm_setzero_ps (), |
| 2762 | (__mmask8) __U); |
| 2763 | } |
| 2764 | |
| 2765 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2766 | _mm256_mask_div_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 2767 | return (__m256) __builtin_ia32_divps256_mask ((__v8sf) __A, |
| 2768 | (__v8sf) __B, |
| 2769 | (__v8sf) __W, |
| 2770 | (__mmask8) __U); |
| 2771 | } |
| 2772 | |
| 2773 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2774 | _mm256_maskz_div_ps (__mmask8 __U, __m256 __A, __m256 __B) { |
| 2775 | return (__m256) __builtin_ia32_divps256_mask ((__v8sf) __A, |
| 2776 | (__v8sf) __B, |
| 2777 | (__v8sf) |
| 2778 | _mm256_setzero_ps (), |
| 2779 | (__mmask8) __U); |
| 2780 | } |
| 2781 | |
| 2782 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2783 | _mm_mask_expand_pd (__m128d __W, __mmask8 __U, __m128d __A) { |
| 2784 | return (__m128d) __builtin_ia32_expanddf128_mask ((__v2df) __A, |
| 2785 | (__v2df) __W, |
| 2786 | (__mmask8) __U); |
| 2787 | } |
| 2788 | |
| 2789 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2790 | _mm_maskz_expand_pd (__mmask8 __U, __m128d __A) { |
| 2791 | return (__m128d) __builtin_ia32_expanddf128_mask ((__v2df) __A, |
| 2792 | (__v2df) |
| 2793 | _mm_setzero_pd (), |
| 2794 | (__mmask8) __U); |
| 2795 | } |
| 2796 | |
| 2797 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2798 | _mm256_mask_expand_pd (__m256d __W, __mmask8 __U, __m256d __A) { |
| 2799 | return (__m256d) __builtin_ia32_expanddf256_mask ((__v4df) __A, |
| 2800 | (__v4df) __W, |
| 2801 | (__mmask8) __U); |
| 2802 | } |
| 2803 | |
| 2804 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2805 | _mm256_maskz_expand_pd (__mmask8 __U, __m256d __A) { |
| 2806 | return (__m256d) __builtin_ia32_expanddf256_mask ((__v4df) __A, |
| 2807 | (__v4df) |
| 2808 | _mm256_setzero_pd (), |
| 2809 | (__mmask8) __U); |
| 2810 | } |
| 2811 | |
| 2812 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2813 | _mm_mask_expand_epi64 (__m128i __W, __mmask8 __U, __m128i __A) { |
| 2814 | return (__m128i) __builtin_ia32_expanddi128_mask ((__v2di) __A, |
| 2815 | (__v2di) __W, |
| 2816 | (__mmask8) __U); |
| 2817 | } |
| 2818 | |
| 2819 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2820 | _mm_maskz_expand_epi64 (__mmask8 __U, __m128i __A) { |
| 2821 | return (__m128i) __builtin_ia32_expanddi128_mask ((__v2di) __A, |
| 2822 | (__v2di) |
| 2823 | _mm_setzero_si128 (), |
| 2824 | (__mmask8) __U); |
| 2825 | } |
| 2826 | |
| 2827 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2828 | _mm256_mask_expand_epi64 (__m256i __W, __mmask8 __U, __m256i __A) { |
| 2829 | return (__m256i) __builtin_ia32_expanddi256_mask ((__v4di) __A, |
| 2830 | (__v4di) __W, |
| 2831 | (__mmask8) __U); |
| 2832 | } |
| 2833 | |
| 2834 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2835 | _mm256_maskz_expand_epi64 (__mmask8 __U, __m256i __A) { |
| 2836 | return (__m256i) __builtin_ia32_expanddi256_mask ((__v4di) __A, |
| 2837 | (__v4di) |
| 2838 | _mm256_setzero_si256 (), |
| 2839 | (__mmask8) __U); |
| 2840 | } |
| 2841 | |
| 2842 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2843 | _mm_mask_expandloadu_pd (__m128d __W, __mmask8 __U, void const *__P) { |
| 2844 | return (__m128d) __builtin_ia32_expandloaddf128_mask ((__v2df *) __P, |
| 2845 | (__v2df) __W, |
| 2846 | (__mmask8) |
| 2847 | __U); |
| 2848 | } |
| 2849 | |
| 2850 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2851 | _mm_maskz_expandloadu_pd (__mmask8 __U, void const *__P) { |
| 2852 | return (__m128d) __builtin_ia32_expandloaddf128_mask ((__v2df *) __P, |
| 2853 | (__v2df) |
| 2854 | _mm_setzero_pd (), |
| 2855 | (__mmask8) |
| 2856 | __U); |
| 2857 | } |
| 2858 | |
| 2859 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2860 | _mm256_mask_expandloadu_pd (__m256d __W, __mmask8 __U, void const *__P) { |
| 2861 | return (__m256d) __builtin_ia32_expandloaddf256_mask ((__v4df *) __P, |
| 2862 | (__v4df) __W, |
| 2863 | (__mmask8) |
| 2864 | __U); |
| 2865 | } |
| 2866 | |
| 2867 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2868 | _mm256_maskz_expandloadu_pd (__mmask8 __U, void const *__P) { |
| 2869 | return (__m256d) __builtin_ia32_expandloaddf256_mask ((__v4df *) __P, |
| 2870 | (__v4df) |
| 2871 | _mm256_setzero_pd (), |
| 2872 | (__mmask8) |
| 2873 | __U); |
| 2874 | } |
| 2875 | |
| 2876 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2877 | _mm_mask_expandloadu_epi64 (__m128i __W, __mmask8 __U, void const *__P) { |
| 2878 | return (__m128i) __builtin_ia32_expandloaddi128_mask ((__v2di *) __P, |
| 2879 | (__v2di) __W, |
| 2880 | (__mmask8) |
| 2881 | __U); |
| 2882 | } |
| 2883 | |
| 2884 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2885 | _mm_maskz_expandloadu_epi64 (__mmask8 __U, void const *__P) { |
| 2886 | return (__m128i) __builtin_ia32_expandloaddi128_mask ((__v2di *) __P, |
| 2887 | (__v2di) |
| 2888 | _mm_setzero_si128 (), |
| 2889 | (__mmask8) |
| 2890 | __U); |
| 2891 | } |
| 2892 | |
| 2893 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2894 | _mm256_mask_expandloadu_epi64 (__m256i __W, __mmask8 __U, |
| 2895 | void const *__P) { |
| 2896 | return (__m256i) __builtin_ia32_expandloaddi256_mask ((__v4di *) __P, |
| 2897 | (__v4di) __W, |
| 2898 | (__mmask8) |
| 2899 | __U); |
| 2900 | } |
| 2901 | |
| 2902 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2903 | _mm256_maskz_expandloadu_epi64 (__mmask8 __U, void const *__P) { |
| 2904 | return (__m256i) __builtin_ia32_expandloaddi256_mask ((__v4di *) __P, |
| 2905 | (__v4di) |
| 2906 | _mm256_setzero_si256 (), |
| 2907 | (__mmask8) |
| 2908 | __U); |
| 2909 | } |
| 2910 | |
| 2911 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2912 | _mm_mask_expandloadu_ps (__m128 __W, __mmask8 __U, void const *__P) { |
| 2913 | return (__m128) __builtin_ia32_expandloadsf128_mask ((__v4sf *) __P, |
| 2914 | (__v4sf) __W, |
| 2915 | (__mmask8) __U); |
| 2916 | } |
| 2917 | |
| 2918 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2919 | _mm_maskz_expandloadu_ps (__mmask8 __U, void const *__P) { |
| 2920 | return (__m128) __builtin_ia32_expandloadsf128_mask ((__v4sf *) __P, |
| 2921 | (__v4sf) |
| 2922 | _mm_setzero_ps (), |
| 2923 | (__mmask8) |
| 2924 | __U); |
| 2925 | } |
| 2926 | |
| 2927 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2928 | _mm256_mask_expandloadu_ps (__m256 __W, __mmask8 __U, void const *__P) { |
| 2929 | return (__m256) __builtin_ia32_expandloadsf256_mask ((__v8sf *) __P, |
| 2930 | (__v8sf) __W, |
| 2931 | (__mmask8) __U); |
| 2932 | } |
| 2933 | |
| 2934 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2935 | _mm256_maskz_expandloadu_ps (__mmask8 __U, void const *__P) { |
| 2936 | return (__m256) __builtin_ia32_expandloadsf256_mask ((__v8sf *) __P, |
| 2937 | (__v8sf) |
| 2938 | _mm256_setzero_ps (), |
| 2939 | (__mmask8) |
| 2940 | __U); |
| 2941 | } |
| 2942 | |
| 2943 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2944 | _mm_mask_expandloadu_epi32 (__m128i __W, __mmask8 __U, void const *__P) { |
| 2945 | return (__m128i) __builtin_ia32_expandloadsi128_mask ((__v4si *) __P, |
| 2946 | (__v4si) __W, |
| 2947 | (__mmask8) |
| 2948 | __U); |
| 2949 | } |
| 2950 | |
| 2951 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2952 | _mm_maskz_expandloadu_epi32 (__mmask8 __U, void const *__P) { |
| 2953 | return (__m128i) __builtin_ia32_expandloadsi128_mask ((__v4si *) __P, |
| 2954 | (__v4si) |
| 2955 | _mm_setzero_si128 (), |
| 2956 | (__mmask8) __U); |
| 2957 | } |
| 2958 | |
| 2959 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2960 | _mm256_mask_expandloadu_epi32 (__m256i __W, __mmask8 __U, |
| 2961 | void const *__P) { |
| 2962 | return (__m256i) __builtin_ia32_expandloadsi256_mask ((__v8si *) __P, |
| 2963 | (__v8si) __W, |
| 2964 | (__mmask8) |
| 2965 | __U); |
| 2966 | } |
| 2967 | |
| 2968 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2969 | _mm256_maskz_expandloadu_epi32 (__mmask8 __U, void const *__P) { |
| 2970 | return (__m256i) __builtin_ia32_expandloadsi256_mask ((__v8si *) __P, |
| 2971 | (__v8si) |
| 2972 | _mm256_setzero_si256 (), |
| 2973 | (__mmask8) |
| 2974 | __U); |
| 2975 | } |
| 2976 | |
| 2977 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2978 | _mm_mask_expand_ps (__m128 __W, __mmask8 __U, __m128 __A) { |
| 2979 | return (__m128) __builtin_ia32_expandsf128_mask ((__v4sf) __A, |
| 2980 | (__v4sf) __W, |
| 2981 | (__mmask8) __U); |
| 2982 | } |
| 2983 | |
| 2984 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2985 | _mm_maskz_expand_ps (__mmask8 __U, __m128 __A) { |
| 2986 | return (__m128) __builtin_ia32_expandsf128_mask ((__v4sf) __A, |
| 2987 | (__v4sf) |
| 2988 | _mm_setzero_ps (), |
| 2989 | (__mmask8) __U); |
| 2990 | } |
| 2991 | |
| 2992 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2993 | _mm256_mask_expand_ps (__m256 __W, __mmask8 __U, __m256 __A) { |
| 2994 | return (__m256) __builtin_ia32_expandsf256_mask ((__v8sf) __A, |
| 2995 | (__v8sf) __W, |
| 2996 | (__mmask8) __U); |
| 2997 | } |
| 2998 | |
| 2999 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3000 | _mm256_maskz_expand_ps (__mmask8 __U, __m256 __A) { |
| 3001 | return (__m256) __builtin_ia32_expandsf256_mask ((__v8sf) __A, |
| 3002 | (__v8sf) |
| 3003 | _mm256_setzero_ps (), |
| 3004 | (__mmask8) __U); |
| 3005 | } |
| 3006 | |
| 3007 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3008 | _mm_mask_expand_epi32 (__m128i __W, __mmask8 __U, __m128i __A) { |
| 3009 | return (__m128i) __builtin_ia32_expandsi128_mask ((__v4si) __A, |
| 3010 | (__v4si) __W, |
| 3011 | (__mmask8) __U); |
| 3012 | } |
| 3013 | |
| 3014 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3015 | _mm_maskz_expand_epi32 (__mmask8 __U, __m128i __A) { |
| 3016 | return (__m128i) __builtin_ia32_expandsi128_mask ((__v4si) __A, |
| 3017 | (__v4si) |
| 3018 | _mm_setzero_si128 (), |
| 3019 | (__mmask8) __U); |
| 3020 | } |
| 3021 | |
| 3022 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3023 | _mm256_mask_expand_epi32 (__m256i __W, __mmask8 __U, __m256i __A) { |
| 3024 | return (__m256i) __builtin_ia32_expandsi256_mask ((__v8si) __A, |
| 3025 | (__v8si) __W, |
| 3026 | (__mmask8) __U); |
| 3027 | } |
| 3028 | |
| 3029 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3030 | _mm256_maskz_expand_epi32 (__mmask8 __U, __m256i __A) { |
| 3031 | return (__m256i) __builtin_ia32_expandsi256_mask ((__v8si) __A, |
| 3032 | (__v8si) |
| 3033 | _mm256_setzero_si256 (), |
| 3034 | (__mmask8) __U); |
| 3035 | } |
| 3036 | |
| 3037 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3038 | _mm_getexp_pd (__m128d __A) { |
| 3039 | return (__m128d) __builtin_ia32_getexppd128_mask ((__v2df) __A, |
| 3040 | (__v2df) |
| 3041 | _mm_setzero_pd (), |
| 3042 | (__mmask8) -1); |
| 3043 | } |
| 3044 | |
| 3045 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3046 | _mm_mask_getexp_pd (__m128d __W, __mmask8 __U, __m128d __A) { |
| 3047 | return (__m128d) __builtin_ia32_getexppd128_mask ((__v2df) __A, |
| 3048 | (__v2df) __W, |
| 3049 | (__mmask8) __U); |
| 3050 | } |
| 3051 | |
| 3052 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3053 | _mm_maskz_getexp_pd (__mmask8 __U, __m128d __A) { |
| 3054 | return (__m128d) __builtin_ia32_getexppd128_mask ((__v2df) __A, |
| 3055 | (__v2df) |
| 3056 | _mm_setzero_pd (), |
| 3057 | (__mmask8) __U); |
| 3058 | } |
| 3059 | |
| 3060 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3061 | _mm256_getexp_pd (__m256d __A) { |
| 3062 | return (__m256d) __builtin_ia32_getexppd256_mask ((__v4df) __A, |
| 3063 | (__v4df) |
| 3064 | _mm256_setzero_pd (), |
| 3065 | (__mmask8) -1); |
| 3066 | } |
| 3067 | |
| 3068 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3069 | _mm256_mask_getexp_pd (__m256d __W, __mmask8 __U, __m256d __A) { |
| 3070 | return (__m256d) __builtin_ia32_getexppd256_mask ((__v4df) __A, |
| 3071 | (__v4df) __W, |
| 3072 | (__mmask8) __U); |
| 3073 | } |
| 3074 | |
| 3075 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3076 | _mm256_maskz_getexp_pd (__mmask8 __U, __m256d __A) { |
| 3077 | return (__m256d) __builtin_ia32_getexppd256_mask ((__v4df) __A, |
| 3078 | (__v4df) |
| 3079 | _mm256_setzero_pd (), |
| 3080 | (__mmask8) __U); |
| 3081 | } |
| 3082 | |
| 3083 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3084 | _mm_getexp_ps (__m128 __A) { |
| 3085 | return (__m128) __builtin_ia32_getexpps128_mask ((__v4sf) __A, |
| 3086 | (__v4sf) |
| 3087 | _mm_setzero_ps (), |
| 3088 | (__mmask8) -1); |
| 3089 | } |
| 3090 | |
| 3091 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3092 | _mm_mask_getexp_ps (__m128 __W, __mmask8 __U, __m128 __A) { |
| 3093 | return (__m128) __builtin_ia32_getexpps128_mask ((__v4sf) __A, |
| 3094 | (__v4sf) __W, |
| 3095 | (__mmask8) __U); |
| 3096 | } |
| 3097 | |
| 3098 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3099 | _mm_maskz_getexp_ps (__mmask8 __U, __m128 __A) { |
| 3100 | return (__m128) __builtin_ia32_getexpps128_mask ((__v4sf) __A, |
| 3101 | (__v4sf) |
| 3102 | _mm_setzero_ps (), |
| 3103 | (__mmask8) __U); |
| 3104 | } |
| 3105 | |
| 3106 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3107 | _mm256_getexp_ps (__m256 __A) { |
| 3108 | return (__m256) __builtin_ia32_getexpps256_mask ((__v8sf) __A, |
| 3109 | (__v8sf) |
| 3110 | _mm256_setzero_ps (), |
| 3111 | (__mmask8) -1); |
| 3112 | } |
| 3113 | |
| 3114 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3115 | _mm256_mask_getexp_ps (__m256 __W, __mmask8 __U, __m256 __A) { |
| 3116 | return (__m256) __builtin_ia32_getexpps256_mask ((__v8sf) __A, |
| 3117 | (__v8sf) __W, |
| 3118 | (__mmask8) __U); |
| 3119 | } |
| 3120 | |
| 3121 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3122 | _mm256_maskz_getexp_ps (__mmask8 __U, __m256 __A) { |
| 3123 | return (__m256) __builtin_ia32_getexpps256_mask ((__v8sf) __A, |
| 3124 | (__v8sf) |
| 3125 | _mm256_setzero_ps (), |
| 3126 | (__mmask8) __U); |
| 3127 | } |
| 3128 | |
| 3129 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3130 | _mm_mask_max_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 3131 | return (__m128d) __builtin_ia32_maxpd_mask ((__v2df) __A, |
| 3132 | (__v2df) __B, |
| 3133 | (__v2df) __W, |
| 3134 | (__mmask8) __U); |
| 3135 | } |
| 3136 | |
| 3137 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3138 | _mm_maskz_max_pd (__mmask8 __U, __m128d __A, __m128d __B) { |
| 3139 | return (__m128d) __builtin_ia32_maxpd_mask ((__v2df) __A, |
| 3140 | (__v2df) __B, |
| 3141 | (__v2df) |
| 3142 | _mm_setzero_pd (), |
| 3143 | (__mmask8) __U); |
| 3144 | } |
| 3145 | |
| 3146 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3147 | _mm256_mask_max_pd (__m256d __W, __mmask8 __U, __m256d __A, |
| 3148 | __m256d __B) { |
| 3149 | return (__m256d) __builtin_ia32_maxpd256_mask ((__v4df) __A, |
| 3150 | (__v4df) __B, |
| 3151 | (__v4df) __W, |
| 3152 | (__mmask8) __U); |
| 3153 | } |
| 3154 | |
| 3155 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3156 | _mm256_maskz_max_pd (__mmask8 __U, __m256d __A, __m256d __B) { |
| 3157 | return (__m256d) __builtin_ia32_maxpd256_mask ((__v4df) __A, |
| 3158 | (__v4df) __B, |
| 3159 | (__v4df) |
| 3160 | _mm256_setzero_pd (), |
| 3161 | (__mmask8) __U); |
| 3162 | } |
| 3163 | |
| 3164 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3165 | _mm_mask_max_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 3166 | return (__m128) __builtin_ia32_maxps_mask ((__v4sf) __A, |
| 3167 | (__v4sf) __B, |
| 3168 | (__v4sf) __W, |
| 3169 | (__mmask8) __U); |
| 3170 | } |
| 3171 | |
| 3172 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3173 | _mm_maskz_max_ps (__mmask8 __U, __m128 __A, __m128 __B) { |
| 3174 | return (__m128) __builtin_ia32_maxps_mask ((__v4sf) __A, |
| 3175 | (__v4sf) __B, |
| 3176 | (__v4sf) |
| 3177 | _mm_setzero_ps (), |
| 3178 | (__mmask8) __U); |
| 3179 | } |
| 3180 | |
| 3181 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3182 | _mm256_mask_max_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 3183 | return (__m256) __builtin_ia32_maxps256_mask ((__v8sf) __A, |
| 3184 | (__v8sf) __B, |
| 3185 | (__v8sf) __W, |
| 3186 | (__mmask8) __U); |
| 3187 | } |
| 3188 | |
| 3189 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3190 | _mm256_maskz_max_ps (__mmask8 __U, __m256 __A, __m256 __B) { |
| 3191 | return (__m256) __builtin_ia32_maxps256_mask ((__v8sf) __A, |
| 3192 | (__v8sf) __B, |
| 3193 | (__v8sf) |
| 3194 | _mm256_setzero_ps (), |
| 3195 | (__mmask8) __U); |
| 3196 | } |
| 3197 | |
| 3198 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3199 | _mm_mask_min_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 3200 | return (__m128d) __builtin_ia32_minpd_mask ((__v2df) __A, |
| 3201 | (__v2df) __B, |
| 3202 | (__v2df) __W, |
| 3203 | (__mmask8) __U); |
| 3204 | } |
| 3205 | |
| 3206 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3207 | _mm_maskz_min_pd (__mmask8 __U, __m128d __A, __m128d __B) { |
| 3208 | return (__m128d) __builtin_ia32_minpd_mask ((__v2df) __A, |
| 3209 | (__v2df) __B, |
| 3210 | (__v2df) |
| 3211 | _mm_setzero_pd (), |
| 3212 | (__mmask8) __U); |
| 3213 | } |
| 3214 | |
| 3215 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3216 | _mm256_mask_min_pd (__m256d __W, __mmask8 __U, __m256d __A, |
| 3217 | __m256d __B) { |
| 3218 | return (__m256d) __builtin_ia32_minpd256_mask ((__v4df) __A, |
| 3219 | (__v4df) __B, |
| 3220 | (__v4df) __W, |
| 3221 | (__mmask8) __U); |
| 3222 | } |
| 3223 | |
| 3224 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3225 | _mm256_maskz_min_pd (__mmask8 __U, __m256d __A, __m256d __B) { |
| 3226 | return (__m256d) __builtin_ia32_minpd256_mask ((__v4df) __A, |
| 3227 | (__v4df) __B, |
| 3228 | (__v4df) |
| 3229 | _mm256_setzero_pd (), |
| 3230 | (__mmask8) __U); |
| 3231 | } |
| 3232 | |
| 3233 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3234 | _mm_mask_min_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 3235 | return (__m128) __builtin_ia32_minps_mask ((__v4sf) __A, |
| 3236 | (__v4sf) __B, |
| 3237 | (__v4sf) __W, |
| 3238 | (__mmask8) __U); |
| 3239 | } |
| 3240 | |
| 3241 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3242 | _mm_maskz_min_ps (__mmask8 __U, __m128 __A, __m128 __B) { |
| 3243 | return (__m128) __builtin_ia32_minps_mask ((__v4sf) __A, |
| 3244 | (__v4sf) __B, |
| 3245 | (__v4sf) |
| 3246 | _mm_setzero_ps (), |
| 3247 | (__mmask8) __U); |
| 3248 | } |
| 3249 | |
| 3250 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3251 | _mm256_mask_min_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 3252 | return (__m256) __builtin_ia32_minps256_mask ((__v8sf) __A, |
| 3253 | (__v8sf) __B, |
| 3254 | (__v8sf) __W, |
| 3255 | (__mmask8) __U); |
| 3256 | } |
| 3257 | |
| 3258 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3259 | _mm256_maskz_min_ps (__mmask8 __U, __m256 __A, __m256 __B) { |
| 3260 | return (__m256) __builtin_ia32_minps256_mask ((__v8sf) __A, |
| 3261 | (__v8sf) __B, |
| 3262 | (__v8sf) |
| 3263 | _mm256_setzero_ps (), |
| 3264 | (__mmask8) __U); |
| 3265 | } |
| 3266 | |
| 3267 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3268 | _mm_mask_mul_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 3269 | return (__m128d) __builtin_ia32_mulpd_mask ((__v2df) __A, |
| 3270 | (__v2df) __B, |
| 3271 | (__v2df) __W, |
| 3272 | (__mmask8) __U); |
| 3273 | } |
| 3274 | |
| 3275 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3276 | _mm_maskz_mul_pd (__mmask8 __U, __m128d __A, __m128d __B) { |
| 3277 | return (__m128d) __builtin_ia32_mulpd_mask ((__v2df) __A, |
| 3278 | (__v2df) __B, |
| 3279 | (__v2df) |
| 3280 | _mm_setzero_pd (), |
| 3281 | (__mmask8) __U); |
| 3282 | } |
| 3283 | |
| 3284 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3285 | _mm256_mask_mul_pd (__m256d __W, __mmask8 __U, __m256d __A, |
| 3286 | __m256d __B) { |
| 3287 | return (__m256d) __builtin_ia32_mulpd256_mask ((__v4df) __A, |
| 3288 | (__v4df) __B, |
| 3289 | (__v4df) __W, |
| 3290 | (__mmask8) __U); |
| 3291 | } |
| 3292 | |
| 3293 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3294 | _mm256_maskz_mul_pd (__mmask8 __U, __m256d __A, __m256d __B) { |
| 3295 | return (__m256d) __builtin_ia32_mulpd256_mask ((__v4df) __A, |
| 3296 | (__v4df) __B, |
| 3297 | (__v4df) |
| 3298 | _mm256_setzero_pd (), |
| 3299 | (__mmask8) __U); |
| 3300 | } |
| 3301 | |
| 3302 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3303 | _mm_mask_mul_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 3304 | return (__m128) __builtin_ia32_mulps_mask ((__v4sf) __A, |
| 3305 | (__v4sf) __B, |
| 3306 | (__v4sf) __W, |
| 3307 | (__mmask8) __U); |
| 3308 | } |
| 3309 | |
| 3310 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3311 | _mm_maskz_mul_ps (__mmask8 __U, __m128 __A, __m128 __B) { |
| 3312 | return (__m128) __builtin_ia32_mulps_mask ((__v4sf) __A, |
| 3313 | (__v4sf) __B, |
| 3314 | (__v4sf) |
| 3315 | _mm_setzero_ps (), |
| 3316 | (__mmask8) __U); |
| 3317 | } |
| 3318 | |
| 3319 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3320 | _mm256_mask_mul_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 3321 | return (__m256) __builtin_ia32_mulps256_mask ((__v8sf) __A, |
| 3322 | (__v8sf) __B, |
| 3323 | (__v8sf) __W, |
| 3324 | (__mmask8) __U); |
| 3325 | } |
| 3326 | |
| 3327 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3328 | _mm256_maskz_mul_ps (__mmask8 __U, __m256 __A, __m256 __B) { |
| 3329 | return (__m256) __builtin_ia32_mulps256_mask ((__v8sf) __A, |
| 3330 | (__v8sf) __B, |
| 3331 | (__v8sf) |
| 3332 | _mm256_setzero_ps (), |
| 3333 | (__mmask8) __U); |
| 3334 | } |
| 3335 | |
| 3336 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3337 | _mm_mask_abs_epi32 (__m128i __W, __mmask8 __U, __m128i __A) { |
| 3338 | return (__m128i) __builtin_ia32_pabsd128_mask ((__v4si) __A, |
| 3339 | (__v4si) __W, |
| 3340 | (__mmask8) __U); |
| 3341 | } |
| 3342 | |
| 3343 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3344 | _mm_maskz_abs_epi32 (__mmask8 __U, __m128i __A) { |
| 3345 | return (__m128i) __builtin_ia32_pabsd128_mask ((__v4si) __A, |
| 3346 | (__v4si) |
| 3347 | _mm_setzero_si128 (), |
| 3348 | (__mmask8) __U); |
| 3349 | } |
| 3350 | |
| 3351 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3352 | _mm256_mask_abs_epi32 (__m256i __W, __mmask8 __U, __m256i __A) { |
| 3353 | return (__m256i) __builtin_ia32_pabsd256_mask ((__v8si) __A, |
| 3354 | (__v8si) __W, |
| 3355 | (__mmask8) __U); |
| 3356 | } |
| 3357 | |
| 3358 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3359 | _mm256_maskz_abs_epi32 (__mmask8 __U, __m256i __A) { |
| 3360 | return (__m256i) __builtin_ia32_pabsd256_mask ((__v8si) __A, |
| 3361 | (__v8si) |
| 3362 | _mm256_setzero_si256 (), |
| 3363 | (__mmask8) __U); |
| 3364 | } |
| 3365 | |
| 3366 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3367 | _mm_abs_epi64 (__m128i __A) { |
| 3368 | return (__m128i) __builtin_ia32_pabsq128_mask ((__v2di) __A, |
| 3369 | (__v2di) |
| 3370 | _mm_setzero_si128 (), |
| 3371 | (__mmask8) -1); |
| 3372 | } |
| 3373 | |
| 3374 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3375 | _mm_mask_abs_epi64 (__m128i __W, __mmask8 __U, __m128i __A) { |
| 3376 | return (__m128i) __builtin_ia32_pabsq128_mask ((__v2di) __A, |
| 3377 | (__v2di) __W, |
| 3378 | (__mmask8) __U); |
| 3379 | } |
| 3380 | |
| 3381 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3382 | _mm_maskz_abs_epi64 (__mmask8 __U, __m128i __A) { |
| 3383 | return (__m128i) __builtin_ia32_pabsq128_mask ((__v2di) __A, |
| 3384 | (__v2di) |
| 3385 | _mm_setzero_si128 (), |
| 3386 | (__mmask8) __U); |
| 3387 | } |
| 3388 | |
| 3389 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3390 | _mm256_abs_epi64 (__m256i __A) { |
| 3391 | return (__m256i) __builtin_ia32_pabsq256_mask ((__v4di) __A, |
| 3392 | (__v4di) |
| 3393 | _mm256_setzero_si256 (), |
| 3394 | (__mmask8) -1); |
| 3395 | } |
| 3396 | |
| 3397 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3398 | _mm256_mask_abs_epi64 (__m256i __W, __mmask8 __U, __m256i __A) { |
| 3399 | return (__m256i) __builtin_ia32_pabsq256_mask ((__v4di) __A, |
| 3400 | (__v4di) __W, |
| 3401 | (__mmask8) __U); |
| 3402 | } |
| 3403 | |
| 3404 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3405 | _mm256_maskz_abs_epi64 (__mmask8 __U, __m256i __A) { |
| 3406 | return (__m256i) __builtin_ia32_pabsq256_mask ((__v4di) __A, |
| 3407 | (__v4di) |
| 3408 | _mm256_setzero_si256 (), |
| 3409 | (__mmask8) __U); |
| 3410 | } |
| 3411 | |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3412 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3413 | _mm_maskz_max_epi32 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3414 | return (__m128i) __builtin_ia32_pmaxsd128_mask ((__v4si) __A, |
| 3415 | (__v4si) __B, |
| 3416 | (__v4si) |
| 3417 | _mm_setzero_si128 (), |
| 3418 | __M); |
| 3419 | } |
| 3420 | |
| 3421 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3422 | _mm_mask_max_epi32 (__m128i __W, __mmask8 __M, __m128i __A, |
| 3423 | __m128i __B) { |
| 3424 | return (__m128i) __builtin_ia32_pmaxsd128_mask ((__v4si) __A, |
| 3425 | (__v4si) __B, |
| 3426 | (__v4si) __W, __M); |
| 3427 | } |
| 3428 | |
| 3429 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3430 | _mm256_maskz_max_epi32 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3431 | return (__m256i) __builtin_ia32_pmaxsd256_mask ((__v8si) __A, |
| 3432 | (__v8si) __B, |
| 3433 | (__v8si) |
| 3434 | _mm256_setzero_si256 (), |
| 3435 | __M); |
| 3436 | } |
| 3437 | |
| 3438 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3439 | _mm256_mask_max_epi32 (__m256i __W, __mmask8 __M, __m256i __A, |
| 3440 | __m256i __B) { |
| 3441 | return (__m256i) __builtin_ia32_pmaxsd256_mask ((__v8si) __A, |
| 3442 | (__v8si) __B, |
| 3443 | (__v8si) __W, __M); |
| 3444 | } |
| 3445 | |
| 3446 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3447 | _mm_maskz_max_epi64 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3448 | return (__m128i) __builtin_ia32_pmaxsq128_mask ((__v2di) __A, |
| 3449 | (__v2di) __B, |
| 3450 | (__v2di) |
| 3451 | _mm_setzero_si128 (), |
| 3452 | __M); |
| 3453 | } |
| 3454 | |
| 3455 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3456 | _mm_mask_max_epi64 (__m128i __W, __mmask8 __M, __m128i __A, |
| 3457 | __m128i __B) { |
| 3458 | return (__m128i) __builtin_ia32_pmaxsq128_mask ((__v2di) __A, |
| 3459 | (__v2di) __B, |
| 3460 | (__v2di) __W, __M); |
| 3461 | } |
| 3462 | |
| 3463 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3464 | _mm_max_epi64 (__m128i __A, __m128i __B) { |
| 3465 | return (__m128i) __builtin_ia32_pmaxsq128_mask ((__v2di) __A, |
| 3466 | (__v2di) __B, |
| 3467 | (__v2di) |
| 3468 | _mm_setzero_si128 (), |
| 3469 | (__mmask8) -1); |
| 3470 | } |
| 3471 | |
| 3472 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3473 | _mm256_maskz_max_epi64 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3474 | return (__m256i) __builtin_ia32_pmaxsq256_mask ((__v4di) __A, |
| 3475 | (__v4di) __B, |
| 3476 | (__v4di) |
| 3477 | _mm256_setzero_si256 (), |
| 3478 | __M); |
| 3479 | } |
| 3480 | |
| 3481 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3482 | _mm256_mask_max_epi64 (__m256i __W, __mmask8 __M, __m256i __A, |
| 3483 | __m256i __B) { |
| 3484 | return (__m256i) __builtin_ia32_pmaxsq256_mask ((__v4di) __A, |
| 3485 | (__v4di) __B, |
| 3486 | (__v4di) __W, __M); |
| 3487 | } |
| 3488 | |
| 3489 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3490 | _mm256_max_epi64 (__m256i __A, __m256i __B) { |
| 3491 | return (__m256i) __builtin_ia32_pmaxsq256_mask ((__v4di) __A, |
| 3492 | (__v4di) __B, |
| 3493 | (__v4di) |
| 3494 | _mm256_setzero_si256 (), |
| 3495 | (__mmask8) -1); |
| 3496 | } |
| 3497 | |
| 3498 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3499 | _mm_maskz_max_epu32 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3500 | return (__m128i) __builtin_ia32_pmaxud128_mask ((__v4si) __A, |
| 3501 | (__v4si) __B, |
| 3502 | (__v4si) |
| 3503 | _mm_setzero_si128 (), |
| 3504 | __M); |
| 3505 | } |
| 3506 | |
| 3507 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3508 | _mm_mask_max_epu32 (__m128i __W, __mmask8 __M, __m128i __A, |
| 3509 | __m128i __B) { |
| 3510 | return (__m128i) __builtin_ia32_pmaxud128_mask ((__v4si) __A, |
| 3511 | (__v4si) __B, |
| 3512 | (__v4si) __W, __M); |
| 3513 | } |
| 3514 | |
| 3515 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3516 | _mm256_maskz_max_epu32 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3517 | return (__m256i) __builtin_ia32_pmaxud256_mask ((__v8si) __A, |
| 3518 | (__v8si) __B, |
| 3519 | (__v8si) |
| 3520 | _mm256_setzero_si256 (), |
| 3521 | __M); |
| 3522 | } |
| 3523 | |
| 3524 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3525 | _mm256_mask_max_epu32 (__m256i __W, __mmask8 __M, __m256i __A, |
| 3526 | __m256i __B) { |
| 3527 | return (__m256i) __builtin_ia32_pmaxud256_mask ((__v8si) __A, |
| 3528 | (__v8si) __B, |
| 3529 | (__v8si) __W, __M); |
| 3530 | } |
| 3531 | |
| 3532 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3533 | _mm_maskz_max_epu64 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3534 | return (__m128i) __builtin_ia32_pmaxuq128_mask ((__v2di) __A, |
| 3535 | (__v2di) __B, |
| 3536 | (__v2di) |
| 3537 | _mm_setzero_si128 (), |
| 3538 | __M); |
| 3539 | } |
| 3540 | |
| 3541 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3542 | _mm_max_epu64 (__m128i __A, __m128i __B) { |
| 3543 | return (__m128i) __builtin_ia32_pmaxuq128_mask ((__v2di) __A, |
| 3544 | (__v2di) __B, |
| 3545 | (__v2di) |
| 3546 | _mm_setzero_si128 (), |
| 3547 | (__mmask8) -1); |
| 3548 | } |
| 3549 | |
| 3550 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3551 | _mm_mask_max_epu64 (__m128i __W, __mmask8 __M, __m128i __A, |
| 3552 | __m128i __B) { |
| 3553 | return (__m128i) __builtin_ia32_pmaxuq128_mask ((__v2di) __A, |
| 3554 | (__v2di) __B, |
| 3555 | (__v2di) __W, __M); |
| 3556 | } |
| 3557 | |
| 3558 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3559 | _mm256_maskz_max_epu64 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3560 | return (__m256i) __builtin_ia32_pmaxuq256_mask ((__v4di) __A, |
| 3561 | (__v4di) __B, |
| 3562 | (__v4di) |
| 3563 | _mm256_setzero_si256 (), |
| 3564 | __M); |
| 3565 | } |
| 3566 | |
| 3567 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3568 | _mm256_max_epu64 (__m256i __A, __m256i __B) { |
| 3569 | return (__m256i) __builtin_ia32_pmaxuq256_mask ((__v4di) __A, |
| 3570 | (__v4di) __B, |
| 3571 | (__v4di) |
| 3572 | _mm256_setzero_si256 (), |
| 3573 | (__mmask8) -1); |
| 3574 | } |
| 3575 | |
| 3576 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3577 | _mm256_mask_max_epu64 (__m256i __W, __mmask8 __M, __m256i __A, |
| 3578 | __m256i __B) { |
| 3579 | return (__m256i) __builtin_ia32_pmaxuq256_mask ((__v4di) __A, |
| 3580 | (__v4di) __B, |
| 3581 | (__v4di) __W, __M); |
| 3582 | } |
| 3583 | |
| 3584 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3585 | _mm_maskz_min_epi32 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3586 | return (__m128i) __builtin_ia32_pminsd128_mask ((__v4si) __A, |
| 3587 | (__v4si) __B, |
| 3588 | (__v4si) |
| 3589 | _mm_setzero_si128 (), |
| 3590 | __M); |
| 3591 | } |
| 3592 | |
| 3593 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3594 | _mm_mask_min_epi32 (__m128i __W, __mmask8 __M, __m128i __A, |
| 3595 | __m128i __B) { |
| 3596 | return (__m128i) __builtin_ia32_pminsd128_mask ((__v4si) __A, |
| 3597 | (__v4si) __B, |
| 3598 | (__v4si) __W, __M); |
| 3599 | } |
| 3600 | |
| 3601 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3602 | _mm256_maskz_min_epi32 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3603 | return (__m256i) __builtin_ia32_pminsd256_mask ((__v8si) __A, |
| 3604 | (__v8si) __B, |
| 3605 | (__v8si) |
| 3606 | _mm256_setzero_si256 (), |
| 3607 | __M); |
| 3608 | } |
| 3609 | |
| 3610 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3611 | _mm256_mask_min_epi32 (__m256i __W, __mmask8 __M, __m256i __A, |
| 3612 | __m256i __B) { |
| 3613 | return (__m256i) __builtin_ia32_pminsd256_mask ((__v8si) __A, |
| 3614 | (__v8si) __B, |
| 3615 | (__v8si) __W, __M); |
| 3616 | } |
| 3617 | |
| 3618 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3619 | _mm_min_epi64 (__m128i __A, __m128i __B) { |
| 3620 | return (__m128i) __builtin_ia32_pminsq128_mask ((__v2di) __A, |
| 3621 | (__v2di) __B, |
| 3622 | (__v2di) |
| 3623 | _mm_setzero_si128 (), |
| 3624 | (__mmask8) -1); |
| 3625 | } |
| 3626 | |
| 3627 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3628 | _mm_mask_min_epi64 (__m128i __W, __mmask8 __M, __m128i __A, |
| 3629 | __m128i __B) { |
| 3630 | return (__m128i) __builtin_ia32_pminsq128_mask ((__v2di) __A, |
| 3631 | (__v2di) __B, |
| 3632 | (__v2di) __W, __M); |
| 3633 | } |
| 3634 | |
| 3635 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3636 | _mm_maskz_min_epi64 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3637 | return (__m128i) __builtin_ia32_pminsq128_mask ((__v2di) __A, |
| 3638 | (__v2di) __B, |
| 3639 | (__v2di) |
| 3640 | _mm_setzero_si128 (), |
| 3641 | __M); |
| 3642 | } |
| 3643 | |
| 3644 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3645 | _mm256_min_epi64 (__m256i __A, __m256i __B) { |
| 3646 | return (__m256i) __builtin_ia32_pminsq256_mask ((__v4di) __A, |
| 3647 | (__v4di) __B, |
| 3648 | (__v4di) |
| 3649 | _mm256_setzero_si256 (), |
| 3650 | (__mmask8) -1); |
| 3651 | } |
| 3652 | |
| 3653 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3654 | _mm256_mask_min_epi64 (__m256i __W, __mmask8 __M, __m256i __A, |
| 3655 | __m256i __B) { |
| 3656 | return (__m256i) __builtin_ia32_pminsq256_mask ((__v4di) __A, |
| 3657 | (__v4di) __B, |
| 3658 | (__v4di) __W, __M); |
| 3659 | } |
| 3660 | |
| 3661 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3662 | _mm256_maskz_min_epi64 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3663 | return (__m256i) __builtin_ia32_pminsq256_mask ((__v4di) __A, |
| 3664 | (__v4di) __B, |
| 3665 | (__v4di) |
| 3666 | _mm256_setzero_si256 (), |
| 3667 | __M); |
| 3668 | } |
| 3669 | |
| 3670 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3671 | _mm_maskz_min_epu32 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3672 | return (__m128i) __builtin_ia32_pminud128_mask ((__v4si) __A, |
| 3673 | (__v4si) __B, |
| 3674 | (__v4si) |
| 3675 | _mm_setzero_si128 (), |
| 3676 | __M); |
| 3677 | } |
| 3678 | |
| 3679 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3680 | _mm_mask_min_epu32 (__m128i __W, __mmask8 __M, __m128i __A, |
| 3681 | __m128i __B) { |
| 3682 | return (__m128i) __builtin_ia32_pminud128_mask ((__v4si) __A, |
| 3683 | (__v4si) __B, |
| 3684 | (__v4si) __W, __M); |
| 3685 | } |
| 3686 | |
| 3687 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3688 | _mm256_maskz_min_epu32 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3689 | return (__m256i) __builtin_ia32_pminud256_mask ((__v8si) __A, |
| 3690 | (__v8si) __B, |
| 3691 | (__v8si) |
| 3692 | _mm256_setzero_si256 (), |
| 3693 | __M); |
| 3694 | } |
| 3695 | |
| 3696 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3697 | _mm256_mask_min_epu32 (__m256i __W, __mmask8 __M, __m256i __A, |
| 3698 | __m256i __B) { |
| 3699 | return (__m256i) __builtin_ia32_pminud256_mask ((__v8si) __A, |
| 3700 | (__v8si) __B, |
| 3701 | (__v8si) __W, __M); |
| 3702 | } |
| 3703 | |
| 3704 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3705 | _mm_min_epu64 (__m128i __A, __m128i __B) { |
| 3706 | return (__m128i) __builtin_ia32_pminuq128_mask ((__v2di) __A, |
| 3707 | (__v2di) __B, |
| 3708 | (__v2di) |
| 3709 | _mm_setzero_si128 (), |
| 3710 | (__mmask8) -1); |
| 3711 | } |
| 3712 | |
| 3713 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3714 | _mm_mask_min_epu64 (__m128i __W, __mmask8 __M, __m128i __A, |
| 3715 | __m128i __B) { |
| 3716 | return (__m128i) __builtin_ia32_pminuq128_mask ((__v2di) __A, |
| 3717 | (__v2di) __B, |
| 3718 | (__v2di) __W, __M); |
| 3719 | } |
| 3720 | |
| 3721 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3722 | _mm_maskz_min_epu64 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3723 | return (__m128i) __builtin_ia32_pminuq128_mask ((__v2di) __A, |
| 3724 | (__v2di) __B, |
| 3725 | (__v2di) |
| 3726 | _mm_setzero_si128 (), |
| 3727 | __M); |
| 3728 | } |
| 3729 | |
| 3730 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3731 | _mm256_min_epu64 (__m256i __A, __m256i __B) { |
| 3732 | return (__m256i) __builtin_ia32_pminuq256_mask ((__v4di) __A, |
| 3733 | (__v4di) __B, |
| 3734 | (__v4di) |
| 3735 | _mm256_setzero_si256 (), |
| 3736 | (__mmask8) -1); |
| 3737 | } |
| 3738 | |
| 3739 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3740 | _mm256_mask_min_epu64 (__m256i __W, __mmask8 __M, __m256i __A, |
| 3741 | __m256i __B) { |
| 3742 | return (__m256i) __builtin_ia32_pminuq256_mask ((__v4di) __A, |
| 3743 | (__v4di) __B, |
| 3744 | (__v4di) __W, __M); |
| 3745 | } |
| 3746 | |
| 3747 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3748 | _mm256_maskz_min_epu64 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3749 | return (__m256i) __builtin_ia32_pminuq256_mask ((__v4di) __A, |
| 3750 | (__v4di) __B, |
| 3751 | (__v4di) |
| 3752 | _mm256_setzero_si256 (), |
| 3753 | __M); |
| 3754 | } |
| 3755 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3756 | #define _mm_roundscale_pd(A, imm) __extension__ ({ \ |
| 3757 | (__m128d)__builtin_ia32_rndscalepd_128_mask((__v2df)(__m128d)(A), \ |
| 3758 | (int)(imm), \ |
| 3759 | (__v2df)_mm_setzero_pd(), \ |
| 3760 | (__mmask8)-1); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3761 | |
| 3762 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3763 | #define _mm_mask_roundscale_pd(W, U, A, imm) __extension__ ({ \ |
| 3764 | (__m128d)__builtin_ia32_rndscalepd_128_mask((__v2df)(__m128d)(A), \ |
| 3765 | (int)(imm), \ |
| 3766 | (__v2df)(__m128d)(W), \ |
| 3767 | (__mmask8)(U)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3768 | |
| 3769 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3770 | #define _mm_maskz_roundscale_pd(U, A, imm) __extension__ ({ \ |
| 3771 | (__m128d)__builtin_ia32_rndscalepd_128_mask((__v2df)(__m128d)(A), \ |
| 3772 | (int)(imm), \ |
| 3773 | (__v2df)_mm_setzero_pd(), \ |
| 3774 | (__mmask8)(U)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3775 | |
| 3776 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3777 | #define _mm256_roundscale_pd(A, imm) __extension__ ({ \ |
| 3778 | (__m256d)__builtin_ia32_rndscalepd_256_mask((__v4df)(__m256d)(A), \ |
| 3779 | (int)(imm), \ |
| 3780 | (__v4df)_mm256_setzero_pd(), \ |
| 3781 | (__mmask8)-1); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3782 | |
| 3783 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3784 | #define _mm256_mask_roundscale_pd(W, U, A, imm) __extension__ ({ \ |
| 3785 | (__m256d)__builtin_ia32_rndscalepd_256_mask((__v4df)(__m256d)(A), \ |
| 3786 | (int)(imm), \ |
| 3787 | (__v4df)(__m256d)(W), \ |
| 3788 | (__mmask8)(U)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3789 | |
| 3790 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3791 | #define _mm256_maskz_roundscale_pd(U, A, imm) __extension__ ({ \ |
| 3792 | (__m256d)__builtin_ia32_rndscalepd_256_mask((__v4df)(__m256d)(A), \ |
| 3793 | (int)(imm), \ |
| 3794 | (__v4df)_mm256_setzero_pd(), \ |
| 3795 | (__mmask8)(U)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3796 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3797 | #define _mm_roundscale_ps(A, imm) __extension__ ({ \ |
| 3798 | (__m128)__builtin_ia32_rndscaleps_128_mask((__v4sf)(__m128)(A), (int)(imm), \ |
| 3799 | (__v4sf)_mm_setzero_ps(), \ |
| 3800 | (__mmask8)-1); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3801 | |
| 3802 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3803 | #define _mm_mask_roundscale_ps(W, U, A, imm) __extension__ ({ \ |
| 3804 | (__m128)__builtin_ia32_rndscaleps_128_mask((__v4sf)(__m128)(A), (int)(imm), \ |
| 3805 | (__v4sf)(__m128)(W), \ |
| 3806 | (__mmask8)(U)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3807 | |
| 3808 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3809 | #define _mm_maskz_roundscale_ps(U, A, imm) __extension__ ({ \ |
| 3810 | (__m128)__builtin_ia32_rndscaleps_128_mask((__v4sf)(__m128)(A), (int)(imm), \ |
| 3811 | (__v4sf)_mm_setzero_ps(), \ |
| 3812 | (__mmask8)(U)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3813 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3814 | #define _mm256_roundscale_ps(A, imm) __extension__ ({ \ |
| 3815 | (__m256)__builtin_ia32_rndscaleps_256_mask((__v8sf)(__m256)(A), (int)(imm), \ |
| 3816 | (__v8sf)_mm256_setzero_ps(), \ |
| 3817 | (__mmask8)-1); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3818 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3819 | #define _mm256_mask_roundscale_ps(W, U, A, imm) __extension__ ({ \ |
| 3820 | (__m256)__builtin_ia32_rndscaleps_256_mask((__v8sf)(__m256)(A), (int)(imm), \ |
| 3821 | (__v8sf)(__m256)(W), \ |
| 3822 | (__mmask8)(U)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3823 | |
| 3824 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3825 | #define _mm256_maskz_roundscale_ps(U, A, imm) __extension__ ({ \ |
| 3826 | (__m256)__builtin_ia32_rndscaleps_256_mask((__v8sf)(__m256)(A), (int)(imm), \ |
| 3827 | (__v8sf)_mm256_setzero_ps(), \ |
| 3828 | (__mmask8)(U)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3829 | |
| 3830 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3831 | _mm_scalef_pd (__m128d __A, __m128d __B) { |
| 3832 | return (__m128d) __builtin_ia32_scalefpd128_mask ((__v2df) __A, |
| 3833 | (__v2df) __B, |
| 3834 | (__v2df) |
| 3835 | _mm_setzero_pd (), |
| 3836 | (__mmask8) -1); |
| 3837 | } |
| 3838 | |
| 3839 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3840 | _mm_mask_scalef_pd (__m128d __W, __mmask8 __U, __m128d __A, |
| 3841 | __m128d __B) { |
| 3842 | return (__m128d) __builtin_ia32_scalefpd128_mask ((__v2df) __A, |
| 3843 | (__v2df) __B, |
| 3844 | (__v2df) __W, |
| 3845 | (__mmask8) __U); |
| 3846 | } |
| 3847 | |
| 3848 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3849 | _mm_maskz_scalef_pd (__mmask8 __U, __m128d __A, __m128d __B) { |
| 3850 | return (__m128d) __builtin_ia32_scalefpd128_mask ((__v2df) __A, |
| 3851 | (__v2df) __B, |
| 3852 | (__v2df) |
| 3853 | _mm_setzero_pd (), |
| 3854 | (__mmask8) __U); |
| 3855 | } |
| 3856 | |
| 3857 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3858 | _mm256_scalef_pd (__m256d __A, __m256d __B) { |
| 3859 | return (__m256d) __builtin_ia32_scalefpd256_mask ((__v4df) __A, |
| 3860 | (__v4df) __B, |
| 3861 | (__v4df) |
| 3862 | _mm256_setzero_pd (), |
| 3863 | (__mmask8) -1); |
| 3864 | } |
| 3865 | |
| 3866 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3867 | _mm256_mask_scalef_pd (__m256d __W, __mmask8 __U, __m256d __A, |
| 3868 | __m256d __B) { |
| 3869 | return (__m256d) __builtin_ia32_scalefpd256_mask ((__v4df) __A, |
| 3870 | (__v4df) __B, |
| 3871 | (__v4df) __W, |
| 3872 | (__mmask8) __U); |
| 3873 | } |
| 3874 | |
| 3875 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3876 | _mm256_maskz_scalef_pd (__mmask8 __U, __m256d __A, __m256d __B) { |
| 3877 | return (__m256d) __builtin_ia32_scalefpd256_mask ((__v4df) __A, |
| 3878 | (__v4df) __B, |
| 3879 | (__v4df) |
| 3880 | _mm256_setzero_pd (), |
| 3881 | (__mmask8) __U); |
| 3882 | } |
| 3883 | |
| 3884 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3885 | _mm_scalef_ps (__m128 __A, __m128 __B) { |
| 3886 | return (__m128) __builtin_ia32_scalefps128_mask ((__v4sf) __A, |
| 3887 | (__v4sf) __B, |
| 3888 | (__v4sf) |
| 3889 | _mm_setzero_ps (), |
| 3890 | (__mmask8) -1); |
| 3891 | } |
| 3892 | |
| 3893 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3894 | _mm_mask_scalef_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 3895 | return (__m128) __builtin_ia32_scalefps128_mask ((__v4sf) __A, |
| 3896 | (__v4sf) __B, |
| 3897 | (__v4sf) __W, |
| 3898 | (__mmask8) __U); |
| 3899 | } |
| 3900 | |
| 3901 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3902 | _mm_maskz_scalef_ps (__mmask8 __U, __m128 __A, __m128 __B) { |
| 3903 | return (__m128) __builtin_ia32_scalefps128_mask ((__v4sf) __A, |
| 3904 | (__v4sf) __B, |
| 3905 | (__v4sf) |
| 3906 | _mm_setzero_ps (), |
| 3907 | (__mmask8) __U); |
| 3908 | } |
| 3909 | |
| 3910 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3911 | _mm256_scalef_ps (__m256 __A, __m256 __B) { |
| 3912 | return (__m256) __builtin_ia32_scalefps256_mask ((__v8sf) __A, |
| 3913 | (__v8sf) __B, |
| 3914 | (__v8sf) |
| 3915 | _mm256_setzero_ps (), |
| 3916 | (__mmask8) -1); |
| 3917 | } |
| 3918 | |
| 3919 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3920 | _mm256_mask_scalef_ps (__m256 __W, __mmask8 __U, __m256 __A, |
| 3921 | __m256 __B) { |
| 3922 | return (__m256) __builtin_ia32_scalefps256_mask ((__v8sf) __A, |
| 3923 | (__v8sf) __B, |
| 3924 | (__v8sf) __W, |
| 3925 | (__mmask8) __U); |
| 3926 | } |
| 3927 | |
| 3928 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3929 | _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) { |
| 3930 | return (__m256) __builtin_ia32_scalefps256_mask ((__v8sf) __A, |
| 3931 | (__v8sf) __B, |
| 3932 | (__v8sf) |
| 3933 | _mm256_setzero_ps (), |
| 3934 | (__mmask8) __U); |
| 3935 | } |
| 3936 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3937 | #define _mm_i64scatter_pd(addr, index, v1, scale) __extension__ ({ \ |
| 3938 | __builtin_ia32_scatterdiv2df((double *)(addr), (__mmask8)-1, \ |
| 3939 | (__v2di)(__m128i)(index), \ |
| 3940 | (__v2df)(__m128d)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3941 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3942 | #define _mm_mask_i64scatter_pd(addr, mask, index, v1, scale) __extension__ ({ \ |
| 3943 | __builtin_ia32_scatterdiv2df((double *)(addr), (__mmask8)(mask), \ |
| 3944 | (__v2di)(__m128i)(index), \ |
| 3945 | (__v2df)(__m128d)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3946 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3947 | #define _mm_i64scatter_epi64(addr, index, v1, scale) __extension__ ({ \ |
| 3948 | __builtin_ia32_scatterdiv2di((long long *)(addr), (__mmask8)-1, \ |
| 3949 | (__v2di)(__m128i)(index), \ |
| 3950 | (__v2di)(__m128i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3951 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3952 | #define _mm_mask_i64scatter_epi64(addr, mask, index, v1, scale) __extension__ ({ \ |
| 3953 | __builtin_ia32_scatterdiv2di((long long *)(addr), (__mmask8)(mask), \ |
| 3954 | (__v2di)(__m128i)(index), \ |
| 3955 | (__v2di)(__m128i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3956 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3957 | #define _mm256_i64scatter_pd(addr, index, v1, scale) __extension__ ({ \ |
| 3958 | __builtin_ia32_scatterdiv4df((double *)(addr), (__mmask8)-1, \ |
| 3959 | (__v4di)(__m256i)(index), \ |
| 3960 | (__v4df)(__m256d)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3961 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3962 | #define _mm256_mask_i64scatter_pd(addr, mask, index, v1, scale) __extension__ ({ \ |
| 3963 | __builtin_ia32_scatterdiv4df((double *)(addr), (__mmask8)(mask), \ |
| 3964 | (__v4di)(__m256i)(index), \ |
| 3965 | (__v4df)(__m256d)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3966 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3967 | #define _mm256_i64scatter_epi64(addr, index, v1, scale) __extension__ ({ \ |
| 3968 | __builtin_ia32_scatterdiv4di((long long *)(addr), (__mmask8)-1, \ |
| 3969 | (__v4di)(__m256i)(index), \ |
| 3970 | (__v4di)(__m256i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3971 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3972 | #define _mm256_mask_i64scatter_epi64(addr, mask, index, v1, scale) __extension__ ({ \ |
| 3973 | __builtin_ia32_scatterdiv4di((long long *)(addr), (__mmask8)(mask), \ |
| 3974 | (__v4di)(__m256i)(index), \ |
| 3975 | (__v4di)(__m256i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3976 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3977 | #define _mm_i64scatter_ps(addr, index, v1, scale) __extension__ ({ \ |
| 3978 | __builtin_ia32_scatterdiv4sf((float *)(addr), (__mmask8)-1, \ |
| 3979 | (__v2di)(__m128i)(index), (__v4sf)(__m128)(v1), \ |
| 3980 | (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3981 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3982 | #define _mm_mask_i64scatter_ps(addr, mask, index, v1, scale) __extension__ ({ \ |
| 3983 | __builtin_ia32_scatterdiv4sf((float *)(addr), (__mmask8)(mask), \ |
| 3984 | (__v2di)(__m128i)(index), (__v4sf)(__m128)(v1), \ |
| 3985 | (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3986 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3987 | #define _mm_i64scatter_epi32(addr, index, v1, scale) __extension__ ({ \ |
| 3988 | __builtin_ia32_scatterdiv4si((int *)(addr), (__mmask8)-1, \ |
| 3989 | (__v2di)(__m128i)(index), \ |
| 3990 | (__v4si)(__m128i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3991 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3992 | #define _mm_mask_i64scatter_epi32(addr, mask, index, v1, scale) __extension__ ({ \ |
| 3993 | __builtin_ia32_scatterdiv4si((int *)(addr), (__mmask8)(mask), \ |
| 3994 | (__v2di)(__m128i)(index), \ |
| 3995 | (__v4si)(__m128i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3996 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3997 | #define _mm256_i64scatter_ps(addr, index, v1, scale) __extension__ ({ \ |
| 3998 | __builtin_ia32_scatterdiv8sf((float *)(addr), (__mmask8)-1, \ |
| 3999 | (__v4di)(__m256i)(index), (__v4sf)(__m128)(v1), \ |
| 4000 | (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4001 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4002 | #define _mm256_mask_i64scatter_ps(addr, mask, index, v1, scale) __extension__ ({ \ |
| 4003 | __builtin_ia32_scatterdiv8sf((float *)(addr), (__mmask8)(mask), \ |
| 4004 | (__v4di)(__m256i)(index), (__v4sf)(__m128)(v1), \ |
| 4005 | (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4006 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4007 | #define _mm256_i64scatter_epi32(addr, index, v1, scale) __extension__ ({ \ |
| 4008 | __builtin_ia32_scatterdiv8si((int *)(addr), (__mmask8)-1, \ |
| 4009 | (__v4di)(__m256i)(index), \ |
| 4010 | (__v4si)(__m128i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4011 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4012 | #define _mm256_mask_i64scatter_epi32(addr, mask, index, v1, scale) __extension__ ({ \ |
| 4013 | __builtin_ia32_scatterdiv8si((int *)(addr), (__mmask8)(mask), \ |
| 4014 | (__v4di)(__m256i)(index), \ |
| 4015 | (__v4si)(__m128i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4016 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4017 | #define _mm_i32scatter_pd(addr, index, v1, scale) __extension__ ({ \ |
| 4018 | __builtin_ia32_scattersiv2df((double *)(addr), (__mmask8)-1, \ |
| 4019 | (__v4si)(__m128i)(index), \ |
| 4020 | (__v2df)(__m128d)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4021 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4022 | #define _mm_mask_i32scatter_pd(addr, mask, index, v1, scale) __extension__ ({ \ |
| 4023 | __builtin_ia32_scattersiv2df((double *)(addr), (__mmask8)(mask), \ |
| 4024 | (__v4si)(__m128i)(index), \ |
| 4025 | (__v2df)(__m128d)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4026 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4027 | #define _mm_i32scatter_epi64(addr, index, v1, scale) __extension__ ({ \ |
| 4028 | __builtin_ia32_scattersiv2di((long long *)(addr), (__mmask8)-1, \ |
| 4029 | (__v4si)(__m128i)(index), \ |
| 4030 | (__v2di)(__m128i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4031 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4032 | #define _mm_mask_i32scatter_epi64(addr, mask, index, v1, scale) __extension__ ({ \ |
| 4033 | __builtin_ia32_scattersiv2di((long long *)(addr), (__mmask8)(mask), \ |
| 4034 | (__v4si)(__m128i)(index), \ |
| 4035 | (__v2di)(__m128i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4036 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4037 | #define _mm256_i32scatter_pd(addr, index, v1, scale) __extension__ ({ \ |
| 4038 | __builtin_ia32_scattersiv4df((double *)(addr), (__mmask8)-1, \ |
| 4039 | (__v4si)(__m128i)(index), \ |
| 4040 | (__v4df)(__m256d)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4041 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4042 | #define _mm256_mask_i32scatter_pd(addr, mask, index, v1, scale) __extension__ ({ \ |
| 4043 | __builtin_ia32_scattersiv4df((double *)(addr), (__mmask8)(mask), \ |
| 4044 | (__v4si)(__m128i)(index), \ |
| 4045 | (__v4df)(__m256d)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4046 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4047 | #define _mm256_i32scatter_epi64(addr, index, v1, scale) __extension__ ({ \ |
| 4048 | __builtin_ia32_scattersiv4di((long long *)(addr), (__mmask8)-1, \ |
| 4049 | (__v4si)(__m128i)(index), \ |
| 4050 | (__v4di)(__m256i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4051 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4052 | #define _mm256_mask_i32scatter_epi64(addr, mask, index, v1, scale) __extension__ ({ \ |
| 4053 | __builtin_ia32_scattersiv4di((long long *)(addr), (__mmask8)(mask), \ |
| 4054 | (__v4si)(__m128i)(index), \ |
| 4055 | (__v4di)(__m256i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4056 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4057 | #define _mm_i32scatter_ps(addr, index, v1, scale) __extension__ ({ \ |
| 4058 | __builtin_ia32_scattersiv4sf((float *)(addr), (__mmask8)-1, \ |
| 4059 | (__v4si)(__m128i)(index), (__v4sf)(__m128)(v1), \ |
| 4060 | (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4061 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4062 | #define _mm_mask_i32scatter_ps(addr, mask, index, v1, scale) __extension__ ({ \ |
| 4063 | __builtin_ia32_scattersiv4sf((float *)(addr), (__mmask8)(mask), \ |
| 4064 | (__v4si)(__m128i)(index), (__v4sf)(__m128)(v1), \ |
| 4065 | (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4066 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4067 | #define _mm_i32scatter_epi32(addr, index, v1, scale) __extension__ ({ \ |
| 4068 | __builtin_ia32_scattersiv4si((int *)(addr), (__mmask8)-1, \ |
| 4069 | (__v4si)(__m128i)(index), \ |
| 4070 | (__v4si)(__m128i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4071 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4072 | #define _mm_mask_i32scatter_epi32(addr, mask, index, v1, scale) __extension__ ({ \ |
| 4073 | __builtin_ia32_scattersiv4si((int *)(addr), (__mmask8)(mask), \ |
| 4074 | (__v4si)(__m128i)(index), \ |
| 4075 | (__v4si)(__m128i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4076 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4077 | #define _mm256_i32scatter_ps(addr, index, v1, scale) __extension__ ({ \ |
| 4078 | __builtin_ia32_scattersiv8sf((float *)(addr), (__mmask8)-1, \ |
| 4079 | (__v8si)(__m256i)(index), (__v8sf)(__m256)(v1), \ |
| 4080 | (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4081 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4082 | #define _mm256_mask_i32scatter_ps(addr, mask, index, v1, scale) __extension__ ({ \ |
| 4083 | __builtin_ia32_scattersiv8sf((float *)(addr), (__mmask8)(mask), \ |
| 4084 | (__v8si)(__m256i)(index), (__v8sf)(__m256)(v1), \ |
| 4085 | (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4086 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4087 | #define _mm256_i32scatter_epi32(addr, index, v1, scale) __extension__ ({ \ |
| 4088 | __builtin_ia32_scattersiv8si((int *)(addr), (__mmask8)-1, \ |
| 4089 | (__v8si)(__m256i)(index), \ |
| 4090 | (__v8si)(__m256i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4091 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4092 | #define _mm256_mask_i32scatter_epi32(addr, mask, index, v1, scale) __extension__ ({ \ |
| 4093 | __builtin_ia32_scattersiv8si((int *)(addr), (__mmask8)(mask), \ |
| 4094 | (__v8si)(__m256i)(index), \ |
| 4095 | (__v8si)(__m256i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4096 | |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 4097 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 4098 | _mm_mask_sqrt_pd (__m128d __W, __mmask8 __U, __m128d __A) { |
| 4099 | return (__m128d) __builtin_ia32_sqrtpd128_mask ((__v2df) __A, |
| 4100 | (__v2df) __W, |
| 4101 | (__mmask8) __U); |
| 4102 | } |
| 4103 | |
| 4104 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 4105 | _mm_maskz_sqrt_pd (__mmask8 __U, __m128d __A) { |
| 4106 | return (__m128d) __builtin_ia32_sqrtpd128_mask ((__v2df) __A, |
| 4107 | (__v2df) |
| 4108 | _mm_setzero_pd (), |
| 4109 | (__mmask8) __U); |
| 4110 | } |
| 4111 | |
| 4112 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 4113 | _mm256_mask_sqrt_pd (__m256d __W, __mmask8 __U, __m256d __A) { |
| 4114 | return (__m256d) __builtin_ia32_sqrtpd256_mask ((__v4df) __A, |
| 4115 | (__v4df) __W, |
| 4116 | (__mmask8) __U); |
| 4117 | } |
| 4118 | |
| 4119 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 4120 | _mm256_maskz_sqrt_pd (__mmask8 __U, __m256d __A) { |
| 4121 | return (__m256d) __builtin_ia32_sqrtpd256_mask ((__v4df) __A, |
| 4122 | (__v4df) |
| 4123 | _mm256_setzero_pd (), |
| 4124 | (__mmask8) __U); |
| 4125 | } |
| 4126 | |
| 4127 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 4128 | _mm_mask_sqrt_ps (__m128 __W, __mmask8 __U, __m128 __A) { |
| 4129 | return (__m128) __builtin_ia32_sqrtps128_mask ((__v4sf) __A, |
| 4130 | (__v4sf) __W, |
| 4131 | (__mmask8) __U); |
| 4132 | } |
| 4133 | |
| 4134 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 4135 | _mm_maskz_sqrt_ps (__mmask8 __U, __m128 __A) { |
| 4136 | return (__m128) __builtin_ia32_sqrtps128_mask ((__v4sf) __A, |
| 4137 | (__v4sf) |
| 4138 | _mm_setzero_ps (), |
| 4139 | (__mmask8) __U); |
| 4140 | } |
| 4141 | |
| 4142 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 4143 | _mm256_mask_sqrt_ps (__m256 __W, __mmask8 __U, __m256 __A) { |
| 4144 | return (__m256) __builtin_ia32_sqrtps256_mask ((__v8sf) __A, |
| 4145 | (__v8sf) __W, |
| 4146 | (__mmask8) __U); |
| 4147 | } |
| 4148 | |
| 4149 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 4150 | _mm256_maskz_sqrt_ps (__mmask8 __U, __m256 __A) { |
| 4151 | return (__m256) __builtin_ia32_sqrtps256_mask ((__v8sf) __A, |
| 4152 | (__v8sf) |
| 4153 | _mm256_setzero_ps (), |
| 4154 | (__mmask8) __U); |
| 4155 | } |
| 4156 | |
| 4157 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 4158 | _mm_mask_sub_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 4159 | return (__m128d) __builtin_ia32_subpd128_mask ((__v2df) __A, |
| 4160 | (__v2df) __B, |
| 4161 | (__v2df) __W, |
| 4162 | (__mmask8) __U); |
| 4163 | } |
| 4164 | |
| 4165 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 4166 | _mm_maskz_sub_pd (__mmask8 __U, __m128d __A, __m128d __B) { |
| 4167 | return (__m128d) __builtin_ia32_subpd128_mask ((__v2df) __A, |
| 4168 | (__v2df) __B, |
| 4169 | (__v2df) |
| 4170 | _mm_setzero_pd (), |
| 4171 | (__mmask8) __U); |
| 4172 | } |
| 4173 | |
| 4174 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 4175 | _mm256_mask_sub_pd (__m256d __W, __mmask8 __U, __m256d __A, |
| 4176 | __m256d __B) { |
| 4177 | return (__m256d) __builtin_ia32_subpd256_mask ((__v4df) __A, |
| 4178 | (__v4df) __B, |
| 4179 | (__v4df) __W, |
| 4180 | (__mmask8) __U); |
| 4181 | } |
| 4182 | |
| 4183 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 4184 | _mm256_maskz_sub_pd (__mmask8 __U, __m256d __A, __m256d __B) { |
| 4185 | return (__m256d) __builtin_ia32_subpd256_mask ((__v4df) __A, |
| 4186 | (__v4df) __B, |
| 4187 | (__v4df) |
| 4188 | _mm256_setzero_pd (), |
| 4189 | (__mmask8) __U); |
| 4190 | } |
| 4191 | |
| 4192 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 4193 | _mm_mask_sub_ps (__m128 __W, __mmask16 __U, __m128 __A, __m128 __B) { |
| 4194 | return (__m128) __builtin_ia32_subps128_mask ((__v4sf) __A, |
| 4195 | (__v4sf) __B, |
| 4196 | (__v4sf) __W, |
| 4197 | (__mmask8) __U); |
| 4198 | } |
| 4199 | |
| 4200 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 4201 | _mm_maskz_sub_ps (__mmask16 __U, __m128 __A, __m128 __B) { |
| 4202 | return (__m128) __builtin_ia32_subps128_mask ((__v4sf) __A, |
| 4203 | (__v4sf) __B, |
| 4204 | (__v4sf) |
| 4205 | _mm_setzero_ps (), |
| 4206 | (__mmask8) __U); |
| 4207 | } |
| 4208 | |
| 4209 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 4210 | _mm256_mask_sub_ps (__m256 __W, __mmask16 __U, __m256 __A, __m256 __B) { |
| 4211 | return (__m256) __builtin_ia32_subps256_mask ((__v8sf) __A, |
| 4212 | (__v8sf) __B, |
| 4213 | (__v8sf) __W, |
| 4214 | (__mmask8) __U); |
| 4215 | } |
| 4216 | |
| 4217 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 4218 | _mm256_maskz_sub_ps (__mmask16 __U, __m256 __A, __m256 __B) { |
| 4219 | return (__m256) __builtin_ia32_subps256_mask ((__v8sf) __A, |
| 4220 | (__v8sf) __B, |
| 4221 | (__v8sf) |
| 4222 | _mm256_setzero_ps (), |
| 4223 | (__mmask8) __U); |
| 4224 | } |
| 4225 | |
| 4226 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4227 | _mm_mask2_permutex2var_epi32 (__m128i __A, __m128i __I, __mmask8 __U, |
| 4228 | __m128i __B) { |
| 4229 | return (__m128i) __builtin_ia32_vpermi2vard128_mask ((__v4si) __A, |
| 4230 | (__v4si) __I |
| 4231 | /* idx */ , |
| 4232 | (__v4si) __B, |
| 4233 | (__mmask8) __U); |
| 4234 | } |
| 4235 | |
| 4236 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4237 | _mm256_mask2_permutex2var_epi32 (__m256i __A, __m256i __I, |
| 4238 | __mmask8 __U, __m256i __B) { |
| 4239 | return (__m256i) __builtin_ia32_vpermi2vard256_mask ((__v8si) __A, |
| 4240 | (__v8si) __I |
| 4241 | /* idx */ , |
| 4242 | (__v8si) __B, |
| 4243 | (__mmask8) __U); |
| 4244 | } |
| 4245 | |
| 4246 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 4247 | _mm_mask2_permutex2var_pd (__m128d __A, __m128i __I, __mmask8 __U, |
| 4248 | __m128d __B) { |
| 4249 | return (__m128d) __builtin_ia32_vpermi2varpd128_mask ((__v2df) __A, |
| 4250 | (__v2di) __I |
| 4251 | /* idx */ , |
| 4252 | (__v2df) __B, |
| 4253 | (__mmask8) |
| 4254 | __U); |
| 4255 | } |
| 4256 | |
| 4257 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 4258 | _mm256_mask2_permutex2var_pd (__m256d __A, __m256i __I, __mmask8 __U, |
| 4259 | __m256d __B) { |
| 4260 | return (__m256d) __builtin_ia32_vpermi2varpd256_mask ((__v4df) __A, |
| 4261 | (__v4di) __I |
| 4262 | /* idx */ , |
| 4263 | (__v4df) __B, |
| 4264 | (__mmask8) |
| 4265 | __U); |
| 4266 | } |
| 4267 | |
| 4268 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 4269 | _mm_mask2_permutex2var_ps (__m128 __A, __m128i __I, __mmask8 __U, |
| 4270 | __m128 __B) { |
| 4271 | return (__m128) __builtin_ia32_vpermi2varps128_mask ((__v4sf) __A, |
| 4272 | (__v4si) __I |
| 4273 | /* idx */ , |
| 4274 | (__v4sf) __B, |
| 4275 | (__mmask8) __U); |
| 4276 | } |
| 4277 | |
| 4278 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 4279 | _mm256_mask2_permutex2var_ps (__m256 __A, __m256i __I, __mmask8 __U, |
| 4280 | __m256 __B) { |
| 4281 | return (__m256) __builtin_ia32_vpermi2varps256_mask ((__v8sf) __A, |
| 4282 | (__v8si) __I |
| 4283 | /* idx */ , |
| 4284 | (__v8sf) __B, |
| 4285 | (__mmask8) __U); |
| 4286 | } |
| 4287 | |
| 4288 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4289 | _mm_mask2_permutex2var_epi64 (__m128i __A, __m128i __I, __mmask8 __U, |
| 4290 | __m128i __B) { |
| 4291 | return (__m128i) __builtin_ia32_vpermi2varq128_mask ((__v2di) __A, |
| 4292 | (__v2di) __I |
| 4293 | /* idx */ , |
| 4294 | (__v2di) __B, |
| 4295 | (__mmask8) __U); |
| 4296 | } |
| 4297 | |
| 4298 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4299 | _mm256_mask2_permutex2var_epi64 (__m256i __A, __m256i __I, |
| 4300 | __mmask8 __U, __m256i __B) { |
| 4301 | return (__m256i) __builtin_ia32_vpermi2varq256_mask ((__v4di) __A, |
| 4302 | (__v4di) __I |
| 4303 | /* idx */ , |
| 4304 | (__v4di) __B, |
| 4305 | (__mmask8) __U); |
| 4306 | } |
| 4307 | |
| 4308 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4309 | _mm_permutex2var_epi32 (__m128i __A, __m128i __I, __m128i __B) { |
| 4310 | return (__m128i) __builtin_ia32_vpermt2vard128_mask ((__v4si) __I |
| 4311 | /* idx */ , |
| 4312 | (__v4si) __A, |
| 4313 | (__v4si) __B, |
| 4314 | (__mmask8) -1); |
| 4315 | } |
| 4316 | |
| 4317 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4318 | _mm_mask_permutex2var_epi32 (__m128i __A, __mmask8 __U, __m128i __I, |
| 4319 | __m128i __B) { |
| 4320 | return (__m128i) __builtin_ia32_vpermt2vard128_mask ((__v4si) __I |
| 4321 | /* idx */ , |
| 4322 | (__v4si) __A, |
| 4323 | (__v4si) __B, |
| 4324 | (__mmask8) __U); |
| 4325 | } |
| 4326 | |
| 4327 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4328 | _mm_maskz_permutex2var_epi32 (__mmask8 __U, __m128i __A, __m128i __I, |
| 4329 | __m128i __B) { |
| 4330 | return (__m128i) __builtin_ia32_vpermt2vard128_maskz ((__v4si) __I |
| 4331 | /* idx */ , |
| 4332 | (__v4si) __A, |
| 4333 | (__v4si) __B, |
| 4334 | (__mmask8) |
| 4335 | __U); |
| 4336 | } |
| 4337 | |
| 4338 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4339 | _mm256_permutex2var_epi32 (__m256i __A, __m256i __I, __m256i __B) { |
| 4340 | return (__m256i) __builtin_ia32_vpermt2vard256_mask ((__v8si) __I |
| 4341 | /* idx */ , |
| 4342 | (__v8si) __A, |
| 4343 | (__v8si) __B, |
| 4344 | (__mmask8) -1); |
| 4345 | } |
| 4346 | |
| 4347 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4348 | _mm256_mask_permutex2var_epi32 (__m256i __A, __mmask8 __U, __m256i __I, |
| 4349 | __m256i __B) { |
| 4350 | return (__m256i) __builtin_ia32_vpermt2vard256_mask ((__v8si) __I |
| 4351 | /* idx */ , |
| 4352 | (__v8si) __A, |
| 4353 | (__v8si) __B, |
| 4354 | (__mmask8) __U); |
| 4355 | } |
| 4356 | |
| 4357 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4358 | _mm256_maskz_permutex2var_epi32 (__mmask8 __U, __m256i __A, |
| 4359 | __m256i __I, __m256i __B) { |
| 4360 | return (__m256i) __builtin_ia32_vpermt2vard256_maskz ((__v8si) __I |
| 4361 | /* idx */ , |
| 4362 | (__v8si) __A, |
| 4363 | (__v8si) __B, |
| 4364 | (__mmask8) |
| 4365 | __U); |
| 4366 | } |
| 4367 | |
| 4368 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 4369 | _mm_permutex2var_pd (__m128d __A, __m128i __I, __m128d __B) { |
| 4370 | return (__m128d) __builtin_ia32_vpermt2varpd128_mask ((__v2di) __I |
| 4371 | /* idx */ , |
| 4372 | (__v2df) __A, |
| 4373 | (__v2df) __B, |
| 4374 | (__mmask8) - |
| 4375 | 1); |
| 4376 | } |
| 4377 | |
| 4378 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 4379 | _mm_mask_permutex2var_pd (__m128d __A, __mmask8 __U, __m128i __I, |
| 4380 | __m128d __B) { |
| 4381 | return (__m128d) __builtin_ia32_vpermt2varpd128_mask ((__v2di) __I |
| 4382 | /* idx */ , |
| 4383 | (__v2df) __A, |
| 4384 | (__v2df) __B, |
| 4385 | (__mmask8) |
| 4386 | __U); |
| 4387 | } |
| 4388 | |
| 4389 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 4390 | _mm_maskz_permutex2var_pd (__mmask8 __U, __m128d __A, __m128i __I, |
| 4391 | __m128d __B) { |
| 4392 | return (__m128d) __builtin_ia32_vpermt2varpd128_maskz ((__v2di) __I |
| 4393 | /* idx */ , |
| 4394 | (__v2df) __A, |
| 4395 | (__v2df) __B, |
| 4396 | (__mmask8) |
| 4397 | __U); |
| 4398 | } |
| 4399 | |
| 4400 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 4401 | _mm256_permutex2var_pd (__m256d __A, __m256i __I, __m256d __B) { |
| 4402 | return (__m256d) __builtin_ia32_vpermt2varpd256_mask ((__v4di) __I |
| 4403 | /* idx */ , |
| 4404 | (__v4df) __A, |
| 4405 | (__v4df) __B, |
| 4406 | (__mmask8) - |
| 4407 | 1); |
| 4408 | } |
| 4409 | |
| 4410 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 4411 | _mm256_mask_permutex2var_pd (__m256d __A, __mmask8 __U, __m256i __I, |
| 4412 | __m256d __B) { |
| 4413 | return (__m256d) __builtin_ia32_vpermt2varpd256_mask ((__v4di) __I |
| 4414 | /* idx */ , |
| 4415 | (__v4df) __A, |
| 4416 | (__v4df) __B, |
| 4417 | (__mmask8) |
| 4418 | __U); |
| 4419 | } |
| 4420 | |
| 4421 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 4422 | _mm256_maskz_permutex2var_pd (__mmask8 __U, __m256d __A, __m256i __I, |
| 4423 | __m256d __B) { |
| 4424 | return (__m256d) __builtin_ia32_vpermt2varpd256_maskz ((__v4di) __I |
| 4425 | /* idx */ , |
| 4426 | (__v4df) __A, |
| 4427 | (__v4df) __B, |
| 4428 | (__mmask8) |
| 4429 | __U); |
| 4430 | } |
| 4431 | |
| 4432 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 4433 | _mm_permutex2var_ps (__m128 __A, __m128i __I, __m128 __B) { |
| 4434 | return (__m128) __builtin_ia32_vpermt2varps128_mask ((__v4si) __I |
| 4435 | /* idx */ , |
| 4436 | (__v4sf) __A, |
| 4437 | (__v4sf) __B, |
| 4438 | (__mmask8) -1); |
| 4439 | } |
| 4440 | |
| 4441 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 4442 | _mm_mask_permutex2var_ps (__m128 __A, __mmask8 __U, __m128i __I, |
| 4443 | __m128 __B) { |
| 4444 | return (__m128) __builtin_ia32_vpermt2varps128_mask ((__v4si) __I |
| 4445 | /* idx */ , |
| 4446 | (__v4sf) __A, |
| 4447 | (__v4sf) __B, |
| 4448 | (__mmask8) __U); |
| 4449 | } |
| 4450 | |
| 4451 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 4452 | _mm_maskz_permutex2var_ps (__mmask8 __U, __m128 __A, __m128i __I, |
| 4453 | __m128 __B) { |
| 4454 | return (__m128) __builtin_ia32_vpermt2varps128_maskz ((__v4si) __I |
| 4455 | /* idx */ , |
| 4456 | (__v4sf) __A, |
| 4457 | (__v4sf) __B, |
| 4458 | (__mmask8) |
| 4459 | __U); |
| 4460 | } |
| 4461 | |
| 4462 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 4463 | _mm256_permutex2var_ps (__m256 __A, __m256i __I, __m256 __B) { |
| 4464 | return (__m256) __builtin_ia32_vpermt2varps256_mask ((__v8si) __I |
| 4465 | /* idx */ , |
| 4466 | (__v8sf) __A, |
| 4467 | (__v8sf) __B, |
| 4468 | (__mmask8) -1); |
| 4469 | } |
| 4470 | |
| 4471 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 4472 | _mm256_mask_permutex2var_ps (__m256 __A, __mmask8 __U, __m256i __I, |
| 4473 | __m256 __B) { |
| 4474 | return (__m256) __builtin_ia32_vpermt2varps256_mask ((__v8si) __I |
| 4475 | /* idx */ , |
| 4476 | (__v8sf) __A, |
| 4477 | (__v8sf) __B, |
| 4478 | (__mmask8) __U); |
| 4479 | } |
| 4480 | |
| 4481 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 4482 | _mm256_maskz_permutex2var_ps (__mmask8 __U, __m256 __A, __m256i __I, |
| 4483 | __m256 __B) { |
| 4484 | return (__m256) __builtin_ia32_vpermt2varps256_maskz ((__v8si) __I |
| 4485 | /* idx */ , |
| 4486 | (__v8sf) __A, |
| 4487 | (__v8sf) __B, |
| 4488 | (__mmask8) |
| 4489 | __U); |
| 4490 | } |
| 4491 | |
| 4492 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4493 | _mm_permutex2var_epi64 (__m128i __A, __m128i __I, __m128i __B) { |
| 4494 | return (__m128i) __builtin_ia32_vpermt2varq128_mask ((__v2di) __I |
| 4495 | /* idx */ , |
| 4496 | (__v2di) __A, |
| 4497 | (__v2di) __B, |
| 4498 | (__mmask8) -1); |
| 4499 | } |
| 4500 | |
| 4501 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4502 | _mm_mask_permutex2var_epi64 (__m128i __A, __mmask8 __U, __m128i __I, |
| 4503 | __m128i __B) { |
| 4504 | return (__m128i) __builtin_ia32_vpermt2varq128_mask ((__v2di) __I |
| 4505 | /* idx */ , |
| 4506 | (__v2di) __A, |
| 4507 | (__v2di) __B, |
| 4508 | (__mmask8) __U); |
| 4509 | } |
| 4510 | |
| 4511 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4512 | _mm_maskz_permutex2var_epi64 (__mmask8 __U, __m128i __A, __m128i __I, |
| 4513 | __m128i __B) { |
| 4514 | return (__m128i) __builtin_ia32_vpermt2varq128_maskz ((__v2di) __I |
| 4515 | /* idx */ , |
| 4516 | (__v2di) __A, |
| 4517 | (__v2di) __B, |
| 4518 | (__mmask8) |
| 4519 | __U); |
| 4520 | } |
| 4521 | |
| 4522 | |
| 4523 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4524 | _mm256_permutex2var_epi64 (__m256i __A, __m256i __I, __m256i __B) { |
| 4525 | return (__m256i) __builtin_ia32_vpermt2varq256_mask ((__v4di) __I |
| 4526 | /* idx */ , |
| 4527 | (__v4di) __A, |
| 4528 | (__v4di) __B, |
| 4529 | (__mmask8) -1); |
| 4530 | } |
| 4531 | |
| 4532 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4533 | _mm256_mask_permutex2var_epi64 (__m256i __A, __mmask8 __U, __m256i __I, |
| 4534 | __m256i __B) { |
| 4535 | return (__m256i) __builtin_ia32_vpermt2varq256_mask ((__v4di) __I |
| 4536 | /* idx */ , |
| 4537 | (__v4di) __A, |
| 4538 | (__v4di) __B, |
| 4539 | (__mmask8) __U); |
| 4540 | } |
| 4541 | |
| 4542 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4543 | _mm256_maskz_permutex2var_epi64 (__mmask8 __U, __m256i __A, |
| 4544 | __m256i __I, __m256i __B) { |
| 4545 | return (__m256i) __builtin_ia32_vpermt2varq256_maskz ((__v4di) __I |
| 4546 | /* idx */ , |
| 4547 | (__v4di) __A, |
| 4548 | (__v4di) __B, |
| 4549 | (__mmask8) |
| 4550 | __U); |
| 4551 | } |
| 4552 | |
Michael Zuckerman | 7cdb72f | 2016-02-18 09:09:34 +0000 | [diff] [blame] | 4553 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4554 | _mm_mask_cvtepi8_epi32 (__m128i __W, __mmask8 __U, __m128i __A) |
| 4555 | { |
| 4556 | return (__m128i) __builtin_ia32_pmovsxbd128_mask ((__v16qi) __A, |
| 4557 | (__v4si) __W, |
| 4558 | (__mmask8) __U); |
| 4559 | } |
| 4560 | |
| 4561 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4562 | _mm_maskz_cvtepi8_epi32 (__mmask8 __U, __m128i __A) |
| 4563 | { |
| 4564 | return (__m128i) __builtin_ia32_pmovsxbd128_mask ((__v16qi) __A, |
| 4565 | (__v4si) |
| 4566 | _mm_setzero_si128 (), |
| 4567 | (__mmask8) __U); |
| 4568 | } |
| 4569 | |
| 4570 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4571 | _mm256_mask_cvtepi8_epi32 (__m256i __W, __mmask8 __U, __m128i __A) |
| 4572 | { |
| 4573 | return (__m256i) __builtin_ia32_pmovsxbd256_mask ((__v16qi) __A, |
| 4574 | (__v8si) __W, |
| 4575 | (__mmask8) __U); |
| 4576 | } |
| 4577 | |
| 4578 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4579 | _mm256_maskz_cvtepi8_epi32 (__mmask8 __U, __m128i __A) |
| 4580 | { |
| 4581 | return (__m256i) __builtin_ia32_pmovsxbd256_mask ((__v16qi) __A, |
| 4582 | (__v8si) |
| 4583 | _mm256_setzero_si256 (), |
| 4584 | (__mmask8) __U); |
| 4585 | } |
| 4586 | |
| 4587 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4588 | _mm_mask_cvtepi8_epi64 (__m128i __W, __mmask8 __U, __m128i __A) |
| 4589 | { |
| 4590 | return (__m128i) __builtin_ia32_pmovsxbq128_mask ((__v16qi) __A, |
| 4591 | (__v2di) __W, |
| 4592 | (__mmask8) __U); |
| 4593 | } |
| 4594 | |
| 4595 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4596 | _mm_maskz_cvtepi8_epi64 (__mmask8 __U, __m128i __A) |
| 4597 | { |
| 4598 | return (__m128i) __builtin_ia32_pmovsxbq128_mask ((__v16qi) __A, |
| 4599 | (__v2di) |
| 4600 | _mm_setzero_si128 (), |
| 4601 | (__mmask8) __U); |
| 4602 | } |
| 4603 | |
| 4604 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4605 | _mm256_mask_cvtepi8_epi64 (__m256i __W, __mmask8 __U, __m128i __A) |
| 4606 | { |
| 4607 | return (__m256i) __builtin_ia32_pmovsxbq256_mask ((__v16qi) __A, |
| 4608 | (__v4di) __W, |
| 4609 | (__mmask8) __U); |
| 4610 | } |
| 4611 | |
| 4612 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4613 | _mm256_maskz_cvtepi8_epi64 (__mmask8 __U, __m128i __A) |
| 4614 | { |
| 4615 | return (__m256i) __builtin_ia32_pmovsxbq256_mask ((__v16qi) __A, |
| 4616 | (__v4di) |
| 4617 | _mm256_setzero_si256 (), |
| 4618 | (__mmask8) __U); |
| 4619 | } |
| 4620 | |
| 4621 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4622 | _mm_mask_cvtepi32_epi64 (__m128i __W, __mmask8 __U, __m128i __X) |
| 4623 | { |
| 4624 | return (__m128i) __builtin_ia32_pmovsxdq128_mask ((__v4si) __X, |
| 4625 | (__v2di) __W, |
| 4626 | (__mmask8) __U); |
| 4627 | } |
| 4628 | |
| 4629 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4630 | _mm_maskz_cvtepi32_epi64 (__mmask8 __U, __m128i __X) |
| 4631 | { |
| 4632 | return (__m128i) __builtin_ia32_pmovsxdq128_mask ((__v4si) __X, |
| 4633 | (__v2di) |
| 4634 | _mm_setzero_si128 (), |
| 4635 | (__mmask8) __U); |
| 4636 | } |
| 4637 | |
| 4638 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4639 | _mm256_mask_cvtepi32_epi64 (__m256i __W, __mmask8 __U, __m128i __X) |
| 4640 | { |
| 4641 | return (__m256i) __builtin_ia32_pmovsxdq256_mask ((__v4si) __X, |
| 4642 | (__v4di) __W, |
| 4643 | (__mmask8) __U); |
| 4644 | } |
| 4645 | |
| 4646 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4647 | _mm256_maskz_cvtepi32_epi64 (__mmask8 __U, __m128i __X) |
| 4648 | { |
| 4649 | return (__m256i) __builtin_ia32_pmovsxdq256_mask ((__v4si) __X, |
| 4650 | (__v4di) |
| 4651 | _mm256_setzero_si256 (), |
| 4652 | (__mmask8) __U); |
| 4653 | } |
| 4654 | |
| 4655 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4656 | _mm_mask_cvtepi16_epi32 (__m128i __W, __mmask8 __U, __m128i __A) |
| 4657 | { |
| 4658 | return (__m128i) __builtin_ia32_pmovsxwd128_mask ((__v8hi) __A, |
| 4659 | (__v4si) __W, |
| 4660 | (__mmask8) __U); |
| 4661 | } |
| 4662 | |
| 4663 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4664 | _mm_maskz_cvtepi16_epi32 (__mmask8 __U, __m128i __A) |
| 4665 | { |
| 4666 | return (__m128i) __builtin_ia32_pmovsxwd128_mask ((__v8hi) __A, |
| 4667 | (__v4si) |
| 4668 | _mm_setzero_si128 (), |
| 4669 | (__mmask8) __U); |
| 4670 | } |
| 4671 | |
| 4672 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4673 | _mm256_mask_cvtepi16_epi32 (__m256i __W, __mmask8 __U, __m128i __A) |
| 4674 | { |
| 4675 | return (__m256i) __builtin_ia32_pmovsxwd256_mask ((__v8hi) __A, |
| 4676 | (__v8si) __W, |
| 4677 | (__mmask8) __U); |
| 4678 | } |
| 4679 | |
| 4680 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4681 | _mm256_maskz_cvtepi16_epi32 (__mmask8 __U, __m128i __A) |
| 4682 | { |
| 4683 | return (__m256i) __builtin_ia32_pmovsxwd256_mask ((__v8hi) __A, |
| 4684 | (__v8si) |
| 4685 | _mm256_setzero_si256 (), |
| 4686 | (__mmask8) __U); |
| 4687 | } |
| 4688 | |
| 4689 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4690 | _mm_mask_cvtepi16_epi64 (__m128i __W, __mmask8 __U, __m128i __A) |
| 4691 | { |
| 4692 | return (__m128i) __builtin_ia32_pmovsxwq128_mask ((__v8hi) __A, |
| 4693 | (__v2di) __W, |
| 4694 | (__mmask8) __U); |
| 4695 | } |
| 4696 | |
| 4697 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4698 | _mm_maskz_cvtepi16_epi64 (__mmask8 __U, __m128i __A) |
| 4699 | { |
| 4700 | return (__m128i) __builtin_ia32_pmovsxwq128_mask ((__v8hi) __A, |
| 4701 | (__v2di) |
| 4702 | _mm_setzero_si128 (), |
| 4703 | (__mmask8) __U); |
| 4704 | } |
| 4705 | |
| 4706 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4707 | _mm256_mask_cvtepi16_epi64 (__m256i __W, __mmask8 __U, __m128i __A) |
| 4708 | { |
| 4709 | return (__m256i) __builtin_ia32_pmovsxwq256_mask ((__v8hi) __A, |
| 4710 | (__v4di) __W, |
| 4711 | (__mmask8) __U); |
| 4712 | } |
| 4713 | |
| 4714 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4715 | _mm256_maskz_cvtepi16_epi64 (__mmask8 __U, __m128i __A) |
| 4716 | { |
| 4717 | return (__m256i) __builtin_ia32_pmovsxwq256_mask ((__v8hi) __A, |
| 4718 | (__v4di) |
| 4719 | _mm256_setzero_si256 (), |
| 4720 | (__mmask8) __U); |
| 4721 | } |
| 4722 | |
| 4723 | |
Michael Zuckerman | 7a33dce | 2016-02-21 14:00:11 +0000 | [diff] [blame] | 4724 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4725 | _mm_mask_cvtepu8_epi32 (__m128i __W, __mmask8 __U, __m128i __A) |
| 4726 | { |
| 4727 | return (__m128i) __builtin_ia32_pmovzxbd128_mask ((__v16qi) __A, |
| 4728 | (__v4si) __W, |
| 4729 | (__mmask8) __U); |
| 4730 | } |
| 4731 | |
| 4732 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4733 | _mm_maskz_cvtepu8_epi32 (__mmask8 __U, __m128i __A) |
| 4734 | { |
| 4735 | return (__m128i) __builtin_ia32_pmovzxbd128_mask ((__v16qi) __A, |
| 4736 | (__v4si) |
| 4737 | _mm_setzero_si128 (), |
| 4738 | (__mmask8) __U); |
| 4739 | } |
| 4740 | |
| 4741 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4742 | _mm256_mask_cvtepu8_epi32 (__m256i __W, __mmask8 __U, __m128i __A) |
| 4743 | { |
| 4744 | return (__m256i) __builtin_ia32_pmovzxbd256_mask ((__v16qi) __A, |
| 4745 | (__v8si) __W, |
| 4746 | (__mmask8) __U); |
| 4747 | } |
| 4748 | |
| 4749 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4750 | _mm256_maskz_cvtepu8_epi32 (__mmask8 __U, __m128i __A) |
| 4751 | { |
| 4752 | return (__m256i) __builtin_ia32_pmovzxbd256_mask ((__v16qi) __A, |
| 4753 | (__v8si) |
| 4754 | _mm256_setzero_si256 (), |
| 4755 | (__mmask8) __U); |
| 4756 | } |
| 4757 | |
| 4758 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4759 | _mm_mask_cvtepu8_epi64 (__m128i __W, __mmask8 __U, __m128i __A) |
| 4760 | { |
| 4761 | return (__m128i) __builtin_ia32_pmovzxbq128_mask ((__v16qi) __A, |
| 4762 | (__v2di) __W, |
| 4763 | (__mmask8) __U); |
| 4764 | } |
| 4765 | |
| 4766 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4767 | _mm_maskz_cvtepu8_epi64 (__mmask8 __U, __m128i __A) |
| 4768 | { |
| 4769 | return (__m128i) __builtin_ia32_pmovzxbq128_mask ((__v16qi) __A, |
| 4770 | (__v2di) |
| 4771 | _mm_setzero_si128 (), |
| 4772 | (__mmask8) __U); |
| 4773 | } |
| 4774 | |
| 4775 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4776 | _mm256_mask_cvtepu8_epi64 (__m256i __W, __mmask8 __U, __m128i __A) |
| 4777 | { |
| 4778 | return (__m256i) __builtin_ia32_pmovzxbq256_mask ((__v16qi) __A, |
| 4779 | (__v4di) __W, |
| 4780 | (__mmask8) __U); |
| 4781 | } |
| 4782 | |
| 4783 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4784 | _mm256_maskz_cvtepu8_epi64 (__mmask8 __U, __m128i __A) |
| 4785 | { |
| 4786 | return (__m256i) __builtin_ia32_pmovzxbq256_mask ((__v16qi) __A, |
| 4787 | (__v4di) |
| 4788 | _mm256_setzero_si256 (), |
| 4789 | (__mmask8) __U); |
| 4790 | } |
| 4791 | |
| 4792 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4793 | _mm_mask_cvtepu32_epi64 (__m128i __W, __mmask8 __U, __m128i __X) |
| 4794 | { |
| 4795 | return (__m128i) __builtin_ia32_pmovzxdq128_mask ((__v4si) __X, |
| 4796 | (__v2di) __W, |
| 4797 | (__mmask8) __U); |
| 4798 | } |
| 4799 | |
| 4800 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4801 | _mm_maskz_cvtepu32_epi64 (__mmask8 __U, __m128i __X) |
| 4802 | { |
| 4803 | return (__m128i) __builtin_ia32_pmovzxdq128_mask ((__v4si) __X, |
| 4804 | (__v2di) |
| 4805 | _mm_setzero_si128 (), |
| 4806 | (__mmask8) __U); |
| 4807 | } |
| 4808 | |
| 4809 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4810 | _mm256_mask_cvtepu32_epi64 (__m256i __W, __mmask8 __U, __m128i __X) |
| 4811 | { |
| 4812 | return (__m256i) __builtin_ia32_pmovzxdq256_mask ((__v4si) __X, |
| 4813 | (__v4di) __W, |
| 4814 | (__mmask8) __U); |
| 4815 | } |
| 4816 | |
| 4817 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4818 | _mm256_maskz_cvtepu32_epi64 (__mmask8 __U, __m128i __X) |
| 4819 | { |
| 4820 | return (__m256i) __builtin_ia32_pmovzxdq256_mask ((__v4si) __X, |
| 4821 | (__v4di) |
| 4822 | _mm256_setzero_si256 (), |
| 4823 | (__mmask8) __U); |
| 4824 | } |
| 4825 | |
| 4826 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4827 | _mm_mask_cvtepu16_epi32 (__m128i __W, __mmask8 __U, __m128i __A) |
| 4828 | { |
| 4829 | return (__m128i) __builtin_ia32_pmovzxwd128_mask ((__v8hi) __A, |
| 4830 | (__v4si) __W, |
| 4831 | (__mmask8) __U); |
| 4832 | } |
| 4833 | |
| 4834 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4835 | _mm_maskz_cvtepu16_epi32 (__mmask8 __U, __m128i __A) |
| 4836 | { |
| 4837 | return (__m128i) __builtin_ia32_pmovzxwd128_mask ((__v8hi) __A, |
| 4838 | (__v4si) |
| 4839 | _mm_setzero_si128 (), |
| 4840 | (__mmask8) __U); |
| 4841 | } |
| 4842 | |
| 4843 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4844 | _mm256_mask_cvtepu16_epi32 (__m256i __W, __mmask8 __U, __m128i __A) |
| 4845 | { |
| 4846 | return (__m256i) __builtin_ia32_pmovzxwd256_mask ((__v8hi) __A, |
| 4847 | (__v8si) __W, |
| 4848 | (__mmask8) __U); |
| 4849 | } |
| 4850 | |
| 4851 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4852 | _mm256_maskz_cvtepu16_epi32 (__mmask8 __U, __m128i __A) |
| 4853 | { |
| 4854 | return (__m256i) __builtin_ia32_pmovzxwd256_mask ((__v8hi) __A, |
| 4855 | (__v8si) |
| 4856 | _mm256_setzero_si256 (), |
| 4857 | (__mmask8) __U); |
| 4858 | } |
| 4859 | |
| 4860 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4861 | _mm_mask_cvtepu16_epi64 (__m128i __W, __mmask8 __U, __m128i __A) |
| 4862 | { |
| 4863 | return (__m128i) __builtin_ia32_pmovzxwq128_mask ((__v8hi) __A, |
| 4864 | (__v2di) __W, |
| 4865 | (__mmask8) __U); |
| 4866 | } |
| 4867 | |
| 4868 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4869 | _mm_maskz_cvtepu16_epi64 (__mmask8 __U, __m128i __A) |
| 4870 | { |
| 4871 | return (__m128i) __builtin_ia32_pmovzxwq128_mask ((__v8hi) __A, |
| 4872 | (__v2di) |
| 4873 | _mm_setzero_si128 (), |
| 4874 | (__mmask8) __U); |
| 4875 | } |
| 4876 | |
| 4877 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4878 | _mm256_mask_cvtepu16_epi64 (__m256i __W, __mmask8 __U, __m128i __A) |
| 4879 | { |
| 4880 | return (__m256i) __builtin_ia32_pmovzxwq256_mask ((__v8hi) __A, |
| 4881 | (__v4di) __W, |
| 4882 | (__mmask8) __U); |
| 4883 | } |
| 4884 | |
| 4885 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4886 | _mm256_maskz_cvtepu16_epi64 (__mmask8 __U, __m128i __A) |
| 4887 | { |
| 4888 | return (__m256i) __builtin_ia32_pmovzxwq256_mask ((__v8hi) __A, |
| 4889 | (__v4di) |
| 4890 | _mm256_setzero_si256 (), |
| 4891 | (__mmask8) __U); |
| 4892 | } |
| 4893 | |
| 4894 | |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4895 | #define _mm_rol_epi32(a, b) __extension__ ({\ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4896 | (__m128i)__builtin_ia32_prold128_mask((__v4si)(__m128i)(a), (int)(b), \ |
| 4897 | (__v4si)_mm_setzero_si128(), \ |
| 4898 | (__mmask8)-1); }) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4899 | |
| 4900 | #define _mm_mask_rol_epi32(w, u, a, b) __extension__ ({\ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4901 | (__m128i)__builtin_ia32_prold128_mask((__v4si)(__m128i)(a), (int)(b), \ |
| 4902 | (__v4si)(__m128i)(w), (__mmask8)(u)); }) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4903 | |
| 4904 | #define _mm_maskz_rol_epi32(u, a, b) __extension__ ({\ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4905 | (__m128i)__builtin_ia32_prold128_mask((__v4si)(__m128i)(a), (int)(b), \ |
| 4906 | (__v4si)_mm_setzero_si128(), \ |
| 4907 | (__mmask8)(u)); }) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4908 | |
| 4909 | #define _mm256_rol_epi32(a, b) __extension__ ({\ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4910 | (__m256i)__builtin_ia32_prold256_mask((__v8si)(__m256i)(a), (int)(b), \ |
| 4911 | (__v8si)_mm256_setzero_si256(), \ |
| 4912 | (__mmask8)-1); }) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4913 | |
| 4914 | #define _mm256_mask_rol_epi32(w, u, a, b) __extension__ ({\ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4915 | (__m256i)__builtin_ia32_prold256_mask((__v8si)(__m256i)(a), (int)(b), \ |
| 4916 | (__v8si)(__m256i)(w), (__mmask8)(u)); }) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4917 | |
| 4918 | #define _mm256_maskz_rol_epi32(u, a, b) __extension__ ({\ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4919 | (__m256i)__builtin_ia32_prold256_mask((__v8si)(__m256i)(a), (int)(b), \ |
| 4920 | (__v8si)_mm256_setzero_si256(), \ |
| 4921 | (__mmask8)(u)); }) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4922 | |
Michael Zuckerman | 477e0a3 | 2016-02-22 09:42:57 +0000 | [diff] [blame] | 4923 | #define _mm_rol_epi64(a, b) __extension__ ({\ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4924 | (__m128i)__builtin_ia32_prolq128_mask((__v2di)(__m128i)(a), (int)(b), \ |
| 4925 | (__v2di)_mm_setzero_di(), \ |
| 4926 | (__mmask8)-1); }) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4927 | |
| 4928 | #define _mm_mask_rol_epi64(w, u, a, b) __extension__ ({\ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4929 | (__m128i)__builtin_ia32_prolq128_mask((__v2di)(__m128i)(a), (int)(b), \ |
| 4930 | (__v2di)(__m128i)(w), (__mmask8)(u)); }) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4931 | |
| 4932 | #define _mm_maskz_rol_epi64(u, a, b) __extension__ ({\ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4933 | (__m128i)__builtin_ia32_prolq128_mask((__v2di)(__m128i)(a), (int)(b), \ |
| 4934 | (__v2di)_mm_setzero_di(), \ |
| 4935 | (__mmask8)(u)); }) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4936 | |
| 4937 | #define _mm256_rol_epi64(a, b) __extension__ ({\ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4938 | (__m256i)__builtin_ia32_prolq256_mask((__v4di)(__m256i)(a), (int)(b), \ |
| 4939 | (__v4di)_mm256_setzero_si256(), \ |
| 4940 | (__mmask8)-1); }) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4941 | |
| 4942 | #define _mm256_mask_rol_epi64(w, u, a, b) __extension__ ({\ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4943 | (__m256i)__builtin_ia32_prolq256_mask((__v4di)(__m256i)(a), (int)(b), \ |
| 4944 | (__v4di)(__m256i)(w), (__mmask8)(u)); }) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4945 | |
| 4946 | #define _mm256_maskz_rol_epi64(u, a, b) __extension__ ({\ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4947 | (__m256i)__builtin_ia32_prolq256_mask((__v4di)(__m256i)(a), (int)(b), \ |
| 4948 | (__v4di)_mm256_setzero_si256(), \ |
| 4949 | (__mmask8)(u)); }) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4950 | |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 4951 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4952 | _mm_rolv_epi32 (__m128i __A, __m128i __B) |
| 4953 | { |
| 4954 | return (__m128i) __builtin_ia32_prolvd128_mask ((__v4si) __A, |
| 4955 | (__v4si) __B, |
| 4956 | (__v4si) |
| 4957 | _mm_setzero_si128 (), |
| 4958 | (__mmask8) -1); |
| 4959 | } |
| 4960 | |
| 4961 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4962 | _mm_mask_rolv_epi32 (__m128i __W, __mmask8 __U, __m128i __A, |
| 4963 | __m128i __B) |
| 4964 | { |
| 4965 | return (__m128i) __builtin_ia32_prolvd128_mask ((__v4si) __A, |
| 4966 | (__v4si) __B, |
| 4967 | (__v4si) __W, |
| 4968 | (__mmask8) __U); |
| 4969 | } |
| 4970 | |
| 4971 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4972 | _mm_maskz_rolv_epi32 (__mmask8 __U, __m128i __A, __m128i __B) |
| 4973 | { |
| 4974 | return (__m128i) __builtin_ia32_prolvd128_mask ((__v4si) __A, |
| 4975 | (__v4si) __B, |
| 4976 | (__v4si) |
| 4977 | _mm_setzero_si128 (), |
| 4978 | (__mmask8) __U); |
| 4979 | } |
| 4980 | |
| 4981 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4982 | _mm256_rolv_epi32 (__m256i __A, __m256i __B) |
| 4983 | { |
| 4984 | return (__m256i) __builtin_ia32_prolvd256_mask ((__v8si) __A, |
| 4985 | (__v8si) __B, |
| 4986 | (__v8si) |
| 4987 | _mm256_setzero_si256 (), |
| 4988 | (__mmask8) -1); |
| 4989 | } |
| 4990 | |
| 4991 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4992 | _mm256_mask_rolv_epi32 (__m256i __W, __mmask8 __U, __m256i __A, |
| 4993 | __m256i __B) |
| 4994 | { |
| 4995 | return (__m256i) __builtin_ia32_prolvd256_mask ((__v8si) __A, |
| 4996 | (__v8si) __B, |
| 4997 | (__v8si) __W, |
| 4998 | (__mmask8) __U); |
| 4999 | } |
| 5000 | |
| 5001 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5002 | _mm256_maskz_rolv_epi32 (__mmask8 __U, __m256i __A, __m256i __B) |
| 5003 | { |
| 5004 | return (__m256i) __builtin_ia32_prolvd256_mask ((__v8si) __A, |
| 5005 | (__v8si) __B, |
| 5006 | (__v8si) |
| 5007 | _mm256_setzero_si256 (), |
| 5008 | (__mmask8) __U); |
| 5009 | } |
| 5010 | |
| 5011 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5012 | _mm_rolv_epi64 (__m128i __A, __m128i __B) |
| 5013 | { |
| 5014 | return (__m128i) __builtin_ia32_prolvq128_mask ((__v2di) __A, |
| 5015 | (__v2di) __B, |
| 5016 | (__v2di) |
| 5017 | _mm_setzero_di (), |
| 5018 | (__mmask8) -1); |
| 5019 | } |
| 5020 | |
| 5021 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5022 | _mm_mask_rolv_epi64 (__m128i __W, __mmask8 __U, __m128i __A, |
| 5023 | __m128i __B) |
| 5024 | { |
| 5025 | return (__m128i) __builtin_ia32_prolvq128_mask ((__v2di) __A, |
| 5026 | (__v2di) __B, |
| 5027 | (__v2di) __W, |
| 5028 | (__mmask8) __U); |
| 5029 | } |
| 5030 | |
| 5031 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5032 | _mm_maskz_rolv_epi64 (__mmask8 __U, __m128i __A, __m128i __B) |
| 5033 | { |
| 5034 | return (__m128i) __builtin_ia32_prolvq128_mask ((__v2di) __A, |
| 5035 | (__v2di) __B, |
| 5036 | (__v2di) |
| 5037 | _mm_setzero_di (), |
| 5038 | (__mmask8) __U); |
| 5039 | } |
| 5040 | |
| 5041 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5042 | _mm256_rolv_epi64 (__m256i __A, __m256i __B) |
| 5043 | { |
| 5044 | return (__m256i) __builtin_ia32_prolvq256_mask ((__v4di) __A, |
| 5045 | (__v4di) __B, |
| 5046 | (__v4di) |
| 5047 | _mm256_setzero_si256 (), |
| 5048 | (__mmask8) -1); |
| 5049 | } |
| 5050 | |
| 5051 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5052 | _mm256_mask_rolv_epi64 (__m256i __W, __mmask8 __U, __m256i __A, |
| 5053 | __m256i __B) |
| 5054 | { |
| 5055 | return (__m256i) __builtin_ia32_prolvq256_mask ((__v4di) __A, |
| 5056 | (__v4di) __B, |
| 5057 | (__v4di) __W, |
| 5058 | (__mmask8) __U); |
| 5059 | } |
| 5060 | |
| 5061 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5062 | _mm256_maskz_rolv_epi64 (__mmask8 __U, __m256i __A, __m256i __B) |
| 5063 | { |
| 5064 | return (__m256i) __builtin_ia32_prolvq256_mask ((__v4di) __A, |
| 5065 | (__v4di) __B, |
| 5066 | (__v4di) |
| 5067 | _mm256_setzero_si256 (), |
| 5068 | (__mmask8) __U); |
| 5069 | } |
| 5070 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5071 | #define _mm_ror_epi32(A, B) __extension__ ({ \ |
| 5072 | (__m128i)__builtin_ia32_prord128_mask((__v4si)(__m128i)(A), (int)(B), \ |
| 5073 | (__v4si)_mm_setzero_si128(), \ |
| 5074 | (__mmask8)-1); }) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 5075 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5076 | #define _mm_mask_ror_epi32(W, U, A, B) __extension__ ({ \ |
| 5077 | (__m128i)__builtin_ia32_prord128_mask((__v4si)(__m128i)(A), (int)(B), \ |
| 5078 | (__v4si)(__m128i)(W), (__mmask8)(U)); }) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 5079 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5080 | #define _mm_maskz_ror_epi32(U, A, B) __extension__ ({ \ |
| 5081 | (__m128i)__builtin_ia32_prord128_mask((__v4si)(__m128i)(A), (int)(B), \ |
| 5082 | (__v4si)_mm_setzero_si128(), \ |
| 5083 | (__mmask8)(U)); }) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 5084 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5085 | #define _mm256_ror_epi32(A, B) __extension__ ({ \ |
| 5086 | (__m256i)__builtin_ia32_prord256_mask((__v8si)(__m256i)(A), (int)(B), \ |
| 5087 | (__v8si)_mm256_setzero_si256(), \ |
| 5088 | (__mmask8)-1); }) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 5089 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5090 | #define _mm256_mask_ror_epi32(W, U, A, B) __extension__ ({ \ |
| 5091 | (__m256i)__builtin_ia32_prord256_mask((__v8si)(__m256i)(A), (int)(B), \ |
| 5092 | (__v8si)(__m256i)(W), (__mmask8)(U)); }) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 5093 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5094 | #define _mm256_maskz_ror_epi32(U, A, B) __extension__ ({ \ |
| 5095 | (__m256i)__builtin_ia32_prord256_mask((__v8si)(__m256i)(A), (int)(B), \ |
| 5096 | (__v8si)_mm256_setzero_si256(), \ |
| 5097 | (__mmask8)(U)); }) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 5098 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5099 | #define _mm_ror_epi64(A, B) __extension__ ({ \ |
| 5100 | (__m128i)__builtin_ia32_prorq128_mask((__v2di)(__m128i)(A), (int)(B), \ |
| 5101 | (__v2di)_mm_setzero_di(), \ |
| 5102 | (__mmask8)-1); }) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 5103 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5104 | #define _mm_mask_ror_epi64(W, U, A, B) __extension__ ({ \ |
| 5105 | (__m128i)__builtin_ia32_prorq128_mask((__v2di)(__m128i)(A), (int)(B), \ |
| 5106 | (__v2di)(__m128i)(W), (__mmask8)(U)); }) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 5107 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5108 | #define _mm_maskz_ror_epi64(U, A, B) __extension__ ({ \ |
| 5109 | (__m128i)__builtin_ia32_prorq128_mask((__v2di)(__m128i)(A), (int)(B), \ |
| 5110 | (__v2di)_mm_setzero_di(), \ |
| 5111 | (__mmask8)(U)); }) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 5112 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5113 | #define _mm256_ror_epi64(A, B) __extension__ ({ \ |
| 5114 | (__m256i)__builtin_ia32_prorq256_mask((__v4di)(__m256i)(A), (int)(B), \ |
| 5115 | (__v4di)_mm256_setzero_si256(), \ |
| 5116 | (__mmask8)-1); }) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 5117 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5118 | #define _mm256_mask_ror_epi64(W, U, A, B) __extension__ ({ \ |
| 5119 | (__m256i)__builtin_ia32_prorq256_mask((__v4di)(__m256i)(A), (int)(B), \ |
| 5120 | (__v4di)(__m256i)(W), (__mmask8)(U)); }) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 5121 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5122 | #define _mm256_maskz_ror_epi64(U, A, B) __extension__ ({ \ |
| 5123 | (__m256i)__builtin_ia32_prorq256_mask((__v4di)(__m256i)(A), (int)(B), \ |
| 5124 | (__v4di)_mm256_setzero_si256(), \ |
| 5125 | (__mmask8)(U)); }) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 5126 | |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 5127 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5128 | _mm_mask_sll_epi32 (__m128i __W, __mmask8 __U, __m128i __A, |
| 5129 | __m128i __B) |
| 5130 | { |
| 5131 | return (__m128i) __builtin_ia32_pslld128_mask ((__v4si) __A, |
| 5132 | (__v4si) __B, |
| 5133 | (__v4si) __W, |
| 5134 | (__mmask8) __U); |
| 5135 | } |
| 5136 | |
| 5137 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5138 | _mm_maskz_sll_epi32 (__mmask8 __U, __m128i __A, __m128i __B) |
| 5139 | { |
| 5140 | return (__m128i) __builtin_ia32_pslld128_mask ((__v4si) __A, |
| 5141 | (__v4si) __B, |
| 5142 | (__v4si) |
| 5143 | _mm_setzero_si128 (), |
| 5144 | (__mmask8) __U); |
| 5145 | } |
| 5146 | |
| 5147 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5148 | _mm256_mask_sll_epi32 (__m256i __W, __mmask8 __U, __m256i __A, |
| 5149 | __m128i __B) |
| 5150 | { |
| 5151 | return (__m256i) __builtin_ia32_pslld256_mask ((__v8si) __A, |
| 5152 | (__v4si) __B, |
| 5153 | (__v8si) __W, |
| 5154 | (__mmask8) __U); |
| 5155 | } |
| 5156 | |
| 5157 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5158 | _mm256_maskz_sll_epi32 (__mmask8 __U, __m256i __A, __m128i __B) |
| 5159 | { |
| 5160 | return (__m256i) __builtin_ia32_pslld256_mask ((__v8si) __A, |
| 5161 | (__v4si) __B, |
| 5162 | (__v8si) |
| 5163 | _mm256_setzero_si256 (), |
| 5164 | (__mmask8) __U); |
| 5165 | } |
| 5166 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5167 | #define _mm_mask_slli_epi32(W, U, A, B) __extension__ ({ \ |
| 5168 | (__m128i)__builtin_ia32_pslldi128_mask((__v4si)(__m128i)(A), (int)(B), \ |
| 5169 | (__v4si)(__m128i)(W), \ |
| 5170 | (__mmask8)(U)); }) |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 5171 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5172 | #define _mm_maskz_slli_epi32(U, A, B) __extension__ ({ \ |
| 5173 | (__m128i)__builtin_ia32_pslldi128_mask((__v4si)(__m128i)(A), (int)(B), \ |
| 5174 | (__v4si)_mm_setzero_si128(), \ |
| 5175 | (__mmask8)(U)); }) |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 5176 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5177 | #define _mm256_mask_slli_epi32(W, U, A, B) __extension__ ({ \ |
| 5178 | (__m256i)__builtin_ia32_pslldi256_mask((__v8si)(__m256i)(A), (int)(B), \ |
| 5179 | (__v8si)(__m256i)(W), \ |
| 5180 | (__mmask8)(U)); }) |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 5181 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5182 | #define _mm256_maskz_slli_epi32(U, A, B) __extension__ ({ \ |
| 5183 | (__m256i)__builtin_ia32_pslldi256_mask((__v8si)(__m256i)(A), (int)(B), \ |
| 5184 | (__v8si)_mm256_setzero_si256(), \ |
| 5185 | (__mmask8)(U)); }) |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 5186 | |
| 5187 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5188 | _mm_mask_sll_epi64 (__m128i __W, __mmask8 __U, __m128i __A, |
| 5189 | __m128i __B) |
| 5190 | { |
| 5191 | return (__m128i) __builtin_ia32_psllq128_mask ((__v2di) __A, |
| 5192 | (__v2di) __B, |
| 5193 | (__v2di) __W, |
| 5194 | (__mmask8) __U); |
| 5195 | } |
| 5196 | |
| 5197 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5198 | _mm_maskz_sll_epi64 (__mmask8 __U, __m128i __A, __m128i __B) |
| 5199 | { |
| 5200 | return (__m128i) __builtin_ia32_psllq128_mask ((__v2di) __A, |
| 5201 | (__v2di) __B, |
| 5202 | (__v2di) |
| 5203 | _mm_setzero_di (), |
| 5204 | (__mmask8) __U); |
| 5205 | } |
| 5206 | |
| 5207 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5208 | _mm256_mask_sll_epi64 (__m256i __W, __mmask8 __U, __m256i __A, |
| 5209 | __m128i __B) |
| 5210 | { |
| 5211 | return (__m256i) __builtin_ia32_psllq256_mask ((__v4di) __A, |
| 5212 | (__v2di) __B, |
| 5213 | (__v4di) __W, |
| 5214 | (__mmask8) __U); |
| 5215 | } |
| 5216 | |
| 5217 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5218 | _mm256_maskz_sll_epi64 (__mmask8 __U, __m256i __A, __m128i __B) |
| 5219 | { |
| 5220 | return (__m256i) __builtin_ia32_psllq256_mask ((__v4di) __A, |
| 5221 | (__v2di) __B, |
| 5222 | (__v4di) |
| 5223 | _mm256_setzero_si256 (), |
| 5224 | (__mmask8) __U); |
| 5225 | } |
| 5226 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5227 | #define _mm_mask_slli_epi64(W, U, A, B) __extension__ ({ \ |
| 5228 | (__m128i)__builtin_ia32_psllqi128_mask((__v2di)(__m128i)(A), (int)(B), \ |
| 5229 | (__v2di)(__m128i)(W), \ |
| 5230 | (__mmask8)(U)); }) |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 5231 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5232 | #define _mm_maskz_slli_epi64(U, A, B) __extension__ ({ \ |
| 5233 | (__m128i)__builtin_ia32_psllqi128_mask((__v2di)(__m128i)(A), (int)(B), \ |
| 5234 | (__v2di)_mm_setzero_di(), \ |
| 5235 | (__mmask8)(U)); }) |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 5236 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5237 | #define _mm256_mask_slli_epi64(W, U, A, B) __extension__ ({ \ |
| 5238 | (__m256i)__builtin_ia32_psllqi256_mask((__v4di)(__m256i)(A), (int)(B), \ |
| 5239 | (__v4di)(__m256i)(W), \ |
| 5240 | (__mmask8)(U)); }) |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 5241 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5242 | #define _mm256_maskz_slli_epi64(U, A, B) __extension__ ({ \ |
| 5243 | (__m256i)__builtin_ia32_psllqi256_mask((__v4di)(__m256i)(A), (int)(B), \ |
| 5244 | (__v4di)_mm256_setzero_si256(), \ |
| 5245 | (__mmask8)(U)); }) |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 5246 | |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 5247 | |
Michael Zuckerman | e98cc74 | 2016-02-23 15:59:47 +0000 | [diff] [blame] | 5248 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5249 | _mm_rorv_epi32 (__m128i __A, __m128i __B) |
| 5250 | { |
| 5251 | return (__m128i) __builtin_ia32_prorvd128_mask ((__v4si) __A, |
| 5252 | (__v4si) __B, |
| 5253 | (__v4si) |
| 5254 | _mm_setzero_si128 (), |
| 5255 | (__mmask8) -1); |
| 5256 | } |
| 5257 | |
| 5258 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5259 | _mm_mask_rorv_epi32 (__m128i __W, __mmask8 __U, __m128i __A, |
| 5260 | __m128i __B) |
| 5261 | { |
| 5262 | return (__m128i) __builtin_ia32_prorvd128_mask ((__v4si) __A, |
| 5263 | (__v4si) __B, |
| 5264 | (__v4si) __W, |
| 5265 | (__mmask8) __U); |
| 5266 | } |
| 5267 | |
| 5268 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5269 | _mm_maskz_rorv_epi32 (__mmask8 __U, __m128i __A, __m128i __B) |
| 5270 | { |
| 5271 | return (__m128i) __builtin_ia32_prorvd128_mask ((__v4si) __A, |
| 5272 | (__v4si) __B, |
| 5273 | (__v4si) |
| 5274 | _mm_setzero_si128 (), |
| 5275 | (__mmask8) __U); |
| 5276 | } |
| 5277 | |
| 5278 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5279 | _mm256_rorv_epi32 (__m256i __A, __m256i __B) |
| 5280 | { |
| 5281 | return (__m256i) __builtin_ia32_prorvd256_mask ((__v8si) __A, |
| 5282 | (__v8si) __B, |
| 5283 | (__v8si) |
| 5284 | _mm256_setzero_si256 (), |
| 5285 | (__mmask8) -1); |
| 5286 | } |
| 5287 | |
| 5288 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5289 | _mm256_mask_rorv_epi32 (__m256i __W, __mmask8 __U, __m256i __A, |
| 5290 | __m256i __B) |
| 5291 | { |
| 5292 | return (__m256i) __builtin_ia32_prorvd256_mask ((__v8si) __A, |
| 5293 | (__v8si) __B, |
| 5294 | (__v8si) __W, |
| 5295 | (__mmask8) __U); |
| 5296 | } |
| 5297 | |
| 5298 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5299 | _mm256_maskz_rorv_epi32 (__mmask8 __U, __m256i __A, __m256i __B) |
| 5300 | { |
| 5301 | return (__m256i) __builtin_ia32_prorvd256_mask ((__v8si) __A, |
| 5302 | (__v8si) __B, |
| 5303 | (__v8si) |
| 5304 | _mm256_setzero_si256 (), |
| 5305 | (__mmask8) __U); |
| 5306 | } |
| 5307 | |
| 5308 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5309 | _mm_rorv_epi64 (__m128i __A, __m128i __B) |
| 5310 | { |
| 5311 | return (__m128i) __builtin_ia32_prorvq128_mask ((__v2di) __A, |
| 5312 | (__v2di) __B, |
| 5313 | (__v2di) |
| 5314 | _mm_setzero_di (), |
| 5315 | (__mmask8) -1); |
| 5316 | } |
| 5317 | |
| 5318 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5319 | _mm_mask_rorv_epi64 (__m128i __W, __mmask8 __U, __m128i __A, |
| 5320 | __m128i __B) |
| 5321 | { |
| 5322 | return (__m128i) __builtin_ia32_prorvq128_mask ((__v2di) __A, |
| 5323 | (__v2di) __B, |
| 5324 | (__v2di) __W, |
| 5325 | (__mmask8) __U); |
| 5326 | } |
| 5327 | |
| 5328 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5329 | _mm_maskz_rorv_epi64 (__mmask8 __U, __m128i __A, __m128i __B) |
| 5330 | { |
| 5331 | return (__m128i) __builtin_ia32_prorvq128_mask ((__v2di) __A, |
| 5332 | (__v2di) __B, |
| 5333 | (__v2di) |
| 5334 | _mm_setzero_di (), |
| 5335 | (__mmask8) __U); |
| 5336 | } |
| 5337 | |
| 5338 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5339 | _mm256_rorv_epi64 (__m256i __A, __m256i __B) |
| 5340 | { |
| 5341 | return (__m256i) __builtin_ia32_prorvq256_mask ((__v4di) __A, |
| 5342 | (__v4di) __B, |
| 5343 | (__v4di) |
| 5344 | _mm256_setzero_si256 (), |
| 5345 | (__mmask8) -1); |
| 5346 | } |
| 5347 | |
| 5348 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5349 | _mm256_mask_rorv_epi64 (__m256i __W, __mmask8 __U, __m256i __A, |
| 5350 | __m256i __B) |
| 5351 | { |
| 5352 | return (__m256i) __builtin_ia32_prorvq256_mask ((__v4di) __A, |
| 5353 | (__v4di) __B, |
| 5354 | (__v4di) __W, |
| 5355 | (__mmask8) __U); |
| 5356 | } |
| 5357 | |
| 5358 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5359 | _mm256_maskz_rorv_epi64 (__mmask8 __U, __m256i __A, __m256i __B) |
| 5360 | { |
| 5361 | return (__m256i) __builtin_ia32_prorvq256_mask ((__v4di) __A, |
| 5362 | (__v4di) __B, |
| 5363 | (__v4di) |
| 5364 | _mm256_setzero_si256 (), |
| 5365 | (__mmask8) __U); |
| 5366 | } |
| 5367 | |
Michael Zuckerman | 431b0e1 | 2016-02-28 07:39:34 +0000 | [diff] [blame] | 5368 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5369 | _mm_mask_sllv_epi64 (__m128i __W, __mmask8 __U, __m128i __X, |
| 5370 | __m128i __Y) |
| 5371 | { |
| 5372 | return (__m128i) __builtin_ia32_psllv2di_mask ((__v2di) __X, |
| 5373 | (__v2di) __Y, |
| 5374 | (__v2di) __W, |
| 5375 | (__mmask8) __U); |
| 5376 | } |
| 5377 | |
| 5378 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5379 | _mm_maskz_sllv_epi64 (__mmask8 __U, __m128i __X, __m128i __Y) |
| 5380 | { |
| 5381 | return (__m128i) __builtin_ia32_psllv2di_mask ((__v2di) __X, |
| 5382 | (__v2di) __Y, |
| 5383 | (__v2di) |
| 5384 | _mm_setzero_di (), |
| 5385 | (__mmask8) __U); |
| 5386 | } |
| 5387 | |
| 5388 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5389 | _mm256_mask_sllv_epi64 (__m256i __W, __mmask8 __U, __m256i __X, |
| 5390 | __m256i __Y) |
| 5391 | { |
| 5392 | return (__m256i) __builtin_ia32_psllv4di_mask ((__v4di) __X, |
| 5393 | (__v4di) __Y, |
| 5394 | (__v4di) __W, |
| 5395 | (__mmask8) __U); |
| 5396 | } |
| 5397 | |
| 5398 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5399 | _mm256_maskz_sllv_epi64 (__mmask8 __U, __m256i __X, __m256i __Y) |
| 5400 | { |
| 5401 | return (__m256i) __builtin_ia32_psllv4di_mask ((__v4di) __X, |
| 5402 | (__v4di) __Y, |
| 5403 | (__v4di) |
| 5404 | _mm256_setzero_si256 (), |
| 5405 | (__mmask8) __U); |
| 5406 | } |
| 5407 | |
| 5408 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5409 | _mm_mask_sllv_epi32 (__m128i __W, __mmask8 __U, __m128i __X, |
| 5410 | __m128i __Y) |
| 5411 | { |
| 5412 | return (__m128i) __builtin_ia32_psllv4si_mask ((__v4si) __X, |
| 5413 | (__v4si) __Y, |
| 5414 | (__v4si) __W, |
| 5415 | (__mmask8) __U); |
| 5416 | } |
| 5417 | |
| 5418 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5419 | _mm_maskz_sllv_epi32 (__mmask8 __U, __m128i __X, __m128i __Y) |
| 5420 | { |
| 5421 | return (__m128i) __builtin_ia32_psllv4si_mask ((__v4si) __X, |
| 5422 | (__v4si) __Y, |
| 5423 | (__v4si) |
| 5424 | _mm_setzero_si128 (), |
| 5425 | (__mmask8) __U); |
| 5426 | } |
| 5427 | |
| 5428 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5429 | _mm256_mask_sllv_epi32 (__m256i __W, __mmask8 __U, __m256i __X, |
| 5430 | __m256i __Y) |
| 5431 | { |
| 5432 | return (__m256i) __builtin_ia32_psllv8si_mask ((__v8si) __X, |
| 5433 | (__v8si) __Y, |
| 5434 | (__v8si) __W, |
| 5435 | (__mmask8) __U); |
| 5436 | } |
| 5437 | |
| 5438 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5439 | _mm256_maskz_sllv_epi32 (__mmask8 __U, __m256i __X, __m256i __Y) |
| 5440 | { |
| 5441 | return (__m256i) __builtin_ia32_psllv8si_mask ((__v8si) __X, |
| 5442 | (__v8si) __Y, |
| 5443 | (__v8si) |
| 5444 | _mm256_setzero_si256 (), |
| 5445 | (__mmask8) __U); |
| 5446 | } |
| 5447 | |
Michael Zuckerman | 0165e76 | 2016-03-01 13:03:45 +0000 | [diff] [blame] | 5448 | |
| 5449 | |
| 5450 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5451 | _mm_mask_srlv_epi64 (__m128i __W, __mmask8 __U, __m128i __X, |
| 5452 | __m128i __Y) |
| 5453 | { |
| 5454 | return (__m128i) __builtin_ia32_psrlv2di_mask ((__v2di) __X, |
| 5455 | (__v2di) __Y, |
| 5456 | (__v2di) __W, |
| 5457 | (__mmask8) __U); |
| 5458 | } |
| 5459 | |
| 5460 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5461 | _mm_maskz_srlv_epi64 (__mmask8 __U, __m128i __X, __m128i __Y) |
| 5462 | { |
| 5463 | return (__m128i) __builtin_ia32_psrlv2di_mask ((__v2di) __X, |
| 5464 | (__v2di) __Y, |
| 5465 | (__v2di) |
| 5466 | _mm_setzero_di (), |
| 5467 | (__mmask8) __U); |
| 5468 | } |
| 5469 | |
| 5470 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5471 | _mm256_mask_srlv_epi64 (__m256i __W, __mmask8 __U, __m256i __X, |
| 5472 | __m256i __Y) |
| 5473 | { |
| 5474 | return (__m256i) __builtin_ia32_psrlv4di_mask ((__v4di) __X, |
| 5475 | (__v4di) __Y, |
| 5476 | (__v4di) __W, |
| 5477 | (__mmask8) __U); |
| 5478 | } |
| 5479 | |
| 5480 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5481 | _mm256_maskz_srlv_epi64 (__mmask8 __U, __m256i __X, __m256i __Y) |
| 5482 | { |
| 5483 | return (__m256i) __builtin_ia32_psrlv4di_mask ((__v4di) __X, |
| 5484 | (__v4di) __Y, |
| 5485 | (__v4di) |
| 5486 | _mm256_setzero_si256 (), |
| 5487 | (__mmask8) __U); |
| 5488 | } |
| 5489 | |
| 5490 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5491 | _mm_mask_srlv_epi32 (__m128i __W, __mmask8 __U, __m128i __X, |
| 5492 | __m128i __Y) |
| 5493 | { |
| 5494 | return (__m128i) __builtin_ia32_psrlv4si_mask ((__v4si) __X, |
| 5495 | (__v4si) __Y, |
| 5496 | (__v4si) __W, |
| 5497 | (__mmask8) __U); |
| 5498 | } |
| 5499 | |
| 5500 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5501 | _mm_maskz_srlv_epi32 (__mmask8 __U, __m128i __X, __m128i __Y) |
| 5502 | { |
| 5503 | return (__m128i) __builtin_ia32_psrlv4si_mask ((__v4si) __X, |
| 5504 | (__v4si) __Y, |
| 5505 | (__v4si) |
| 5506 | _mm_setzero_si128 (), |
| 5507 | (__mmask8) __U); |
| 5508 | } |
| 5509 | |
| 5510 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5511 | _mm256_mask_srlv_epi32 (__m256i __W, __mmask8 __U, __m256i __X, |
| 5512 | __m256i __Y) |
| 5513 | { |
| 5514 | return (__m256i) __builtin_ia32_psrlv8si_mask ((__v8si) __X, |
| 5515 | (__v8si) __Y, |
| 5516 | (__v8si) __W, |
| 5517 | (__mmask8) __U); |
| 5518 | } |
| 5519 | |
| 5520 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5521 | _mm256_maskz_srlv_epi32 (__mmask8 __U, __m256i __X, __m256i __Y) |
| 5522 | { |
| 5523 | return (__m256i) __builtin_ia32_psrlv8si_mask ((__v8si) __X, |
| 5524 | (__v8si) __Y, |
| 5525 | (__v8si) |
| 5526 | _mm256_setzero_si256 (), |
| 5527 | (__mmask8) __U); |
| 5528 | } |
| 5529 | |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 5530 | |
| 5531 | |
| 5532 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5533 | _mm_mask_srl_epi32 (__m128i __W, __mmask8 __U, __m128i __A, |
| 5534 | __m128i __B) |
| 5535 | { |
| 5536 | return (__m128i) __builtin_ia32_psrld128_mask ((__v4si) __A, |
| 5537 | (__v4si) __B, |
| 5538 | (__v4si) __W, |
| 5539 | (__mmask8) __U); |
| 5540 | } |
| 5541 | |
| 5542 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5543 | _mm_maskz_srl_epi32 (__mmask8 __U, __m128i __A, __m128i __B) |
| 5544 | { |
| 5545 | return (__m128i) __builtin_ia32_psrld128_mask ((__v4si) __A, |
| 5546 | (__v4si) __B, |
| 5547 | (__v4si) |
| 5548 | _mm_setzero_si128 (), |
| 5549 | (__mmask8) __U); |
| 5550 | } |
| 5551 | |
| 5552 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5553 | _mm256_mask_srl_epi32 (__m256i __W, __mmask8 __U, __m256i __A, |
| 5554 | __m128i __B) |
| 5555 | { |
| 5556 | return (__m256i) __builtin_ia32_psrld256_mask ((__v8si) __A, |
| 5557 | (__v4si) __B, |
| 5558 | (__v8si) __W, |
| 5559 | (__mmask8) __U); |
| 5560 | } |
| 5561 | |
| 5562 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5563 | _mm256_maskz_srl_epi32 (__mmask8 __U, __m256i __A, __m128i __B) |
| 5564 | { |
| 5565 | return (__m256i) __builtin_ia32_psrld256_mask ((__v8si) __A, |
| 5566 | (__v4si) __B, |
| 5567 | (__v8si) |
| 5568 | _mm256_setzero_si256 (), |
| 5569 | (__mmask8) __U); |
| 5570 | } |
| 5571 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5572 | #define _mm_mask_srli_epi32(W, U, A, imm) __extension__ ({ \ |
| 5573 | (__m128i)__builtin_ia32_psrldi128_mask((__v4si)(__m128i)(A), (int)(imm), \ |
| 5574 | (__v4si)(__m128i)(W), \ |
| 5575 | (__mmask8)(U)); }) |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 5576 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5577 | #define _mm_maskz_srli_epi32(U, A, imm) __extension__ ({ \ |
| 5578 | (__m128i)__builtin_ia32_psrldi128_mask((__v4si)(__m128i)(A), (int)(imm), \ |
| 5579 | (__v4si)_mm_setzero_si128(), \ |
| 5580 | (__mmask8)(U)); }) |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 5581 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5582 | #define _mm256_mask_srli_epi32(W, U, A, imm) __extension__ ({ \ |
| 5583 | (__m256i)__builtin_ia32_psrldi256_mask((__v8si)(__m256i)(A), (int)(imm), \ |
| 5584 | (__v8si)(__m256i)(W), \ |
| 5585 | (__mmask8)(U)); }) |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 5586 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5587 | #define _mm256_maskz_srli_epi32(U, A, imm) __extension__ ({ \ |
| 5588 | (__m256i)__builtin_ia32_psrldi256_mask((__v8si)(__m256i)(A), (int)(imm), \ |
| 5589 | (__v8si)_mm256_setzero_si256(), \ |
| 5590 | (__mmask8)(U)); }) |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 5591 | |
| 5592 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5593 | _mm_mask_srl_epi64 (__m128i __W, __mmask8 __U, __m128i __A, |
| 5594 | __m128i __B) |
| 5595 | { |
| 5596 | return (__m128i) __builtin_ia32_psrlq128_mask ((__v2di) __A, |
| 5597 | (__v2di) __B, |
| 5598 | (__v2di) __W, |
| 5599 | (__mmask8) __U); |
| 5600 | } |
| 5601 | |
| 5602 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5603 | _mm_maskz_srl_epi64 (__mmask8 __U, __m128i __A, __m128i __B) |
| 5604 | { |
| 5605 | return (__m128i) __builtin_ia32_psrlq128_mask ((__v2di) __A, |
| 5606 | (__v2di) __B, |
| 5607 | (__v2di) |
| 5608 | _mm_setzero_di (), |
| 5609 | (__mmask8) __U); |
| 5610 | } |
| 5611 | |
| 5612 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5613 | _mm256_mask_srl_epi64 (__m256i __W, __mmask8 __U, __m256i __A, |
| 5614 | __m128i __B) |
| 5615 | { |
| 5616 | return (__m256i) __builtin_ia32_psrlq256_mask ((__v4di) __A, |
| 5617 | (__v2di) __B, |
| 5618 | (__v4di) __W, |
| 5619 | (__mmask8) __U); |
| 5620 | } |
| 5621 | |
| 5622 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5623 | _mm256_maskz_srl_epi64 (__mmask8 __U, __m256i __A, __m128i __B) |
| 5624 | { |
| 5625 | return (__m256i) __builtin_ia32_psrlq256_mask ((__v4di) __A, |
| 5626 | (__v2di) __B, |
| 5627 | (__v4di) |
| 5628 | _mm256_setzero_si256 (), |
| 5629 | (__mmask8) __U); |
| 5630 | } |
| 5631 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5632 | #define _mm_mask_srli_epi64(W, U, A, imm) __extension__ ({ \ |
| 5633 | (__m128i)__builtin_ia32_psrlqi128_mask((__v2di)(__m128i)(A), (int)(imm), \ |
| 5634 | (__v2di)(__m128i)(W), \ |
| 5635 | (__mmask8)(U)); }) |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 5636 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5637 | #define _mm_maskz_srli_epi64(U, A, imm) __extension__ ({ \ |
| 5638 | (__m128i)__builtin_ia32_psrlqi128_mask((__v2di)(__m128i)(A), (int)(imm), \ |
| 5639 | (__v2di)_mm_setzero_si128(), \ |
| 5640 | (__mmask8)(U)); }) |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 5641 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5642 | #define _mm256_mask_srli_epi64(W, U, A, imm) __extension__ ({ \ |
| 5643 | (__m256i)__builtin_ia32_psrlqi256_mask((__v4di)(__m256i)(A), (int)(imm), \ |
| 5644 | (__v4di)(__m256i)(W), \ |
| 5645 | (__mmask8)(U)); }) |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 5646 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5647 | #define _mm256_maskz_srli_epi64(U, A, imm) __extension__ ({ \ |
| 5648 | (__m256i)__builtin_ia32_psrlqi256_mask((__v4di)(__m256i)(A), (int)(imm), \ |
| 5649 | (__v4di)_mm256_setzero_si256(), \ |
| 5650 | (__mmask8)(U)); }) |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 5651 | |
Michael Zuckerman | d15c95a | 2016-03-02 09:05:46 +0000 | [diff] [blame] | 5652 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5653 | _mm_mask_srav_epi32 (__m128i __W, __mmask8 __U, __m128i __X, |
| 5654 | __m128i __Y) |
| 5655 | { |
| 5656 | return (__m128i) __builtin_ia32_psrav4si_mask ((__v4si) __X, |
| 5657 | (__v4si) __Y, |
| 5658 | (__v4si) __W, |
| 5659 | (__mmask8) __U); |
| 5660 | } |
| 5661 | |
| 5662 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5663 | _mm_maskz_srav_epi32 (__mmask8 __U, __m128i __X, __m128i __Y) |
| 5664 | { |
| 5665 | return (__m128i) __builtin_ia32_psrav4si_mask ((__v4si) __X, |
| 5666 | (__v4si) __Y, |
| 5667 | (__v4si) |
| 5668 | _mm_setzero_si128 (), |
| 5669 | (__mmask8) __U); |
| 5670 | } |
| 5671 | |
| 5672 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5673 | _mm256_mask_srav_epi32 (__m256i __W, __mmask8 __U, __m256i __X, |
| 5674 | __m256i __Y) |
| 5675 | { |
| 5676 | return (__m256i) __builtin_ia32_psrav8si_mask ((__v8si) __X, |
| 5677 | (__v8si) __Y, |
| 5678 | (__v8si) __W, |
| 5679 | (__mmask8) __U); |
| 5680 | } |
| 5681 | |
| 5682 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5683 | _mm256_maskz_srav_epi32 (__mmask8 __U, __m256i __X, __m256i __Y) |
| 5684 | { |
| 5685 | return (__m256i) __builtin_ia32_psrav8si_mask ((__v8si) __X, |
| 5686 | (__v8si) __Y, |
| 5687 | (__v8si) |
| 5688 | _mm256_setzero_si256 (), |
| 5689 | (__mmask8) __U); |
| 5690 | } |
| 5691 | |
| 5692 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5693 | _mm_srav_epi64 (__m128i __X, __m128i __Y) |
| 5694 | { |
| 5695 | return (__m128i) __builtin_ia32_psravq128_mask ((__v2di) __X, |
| 5696 | (__v2di) __Y, |
| 5697 | (__v2di) |
| 5698 | _mm_setzero_di (), |
| 5699 | (__mmask8) -1); |
| 5700 | } |
| 5701 | |
| 5702 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5703 | _mm_mask_srav_epi64 (__m128i __W, __mmask8 __U, __m128i __X, |
| 5704 | __m128i __Y) |
| 5705 | { |
| 5706 | return (__m128i) __builtin_ia32_psravq128_mask ((__v2di) __X, |
| 5707 | (__v2di) __Y, |
| 5708 | (__v2di) __W, |
| 5709 | (__mmask8) __U); |
| 5710 | } |
| 5711 | |
| 5712 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5713 | _mm_maskz_srav_epi64 (__mmask8 __U, __m128i __X, __m128i __Y) |
| 5714 | { |
| 5715 | return (__m128i) __builtin_ia32_psravq128_mask ((__v2di) __X, |
| 5716 | (__v2di) __Y, |
| 5717 | (__v2di) |
| 5718 | _mm_setzero_di (), |
| 5719 | (__mmask8) __U); |
| 5720 | } |
| 5721 | |
| 5722 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5723 | _mm256_srav_epi64 (__m256i __X, __m256i __Y) |
| 5724 | { |
| 5725 | return (__m256i) __builtin_ia32_psravq256_mask ((__v4di) __X, |
| 5726 | (__v4di) __Y, |
| 5727 | (__v4di) |
| 5728 | _mm256_setzero_si256 (), |
| 5729 | (__mmask8) -1); |
| 5730 | } |
| 5731 | |
| 5732 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5733 | _mm256_mask_srav_epi64 (__m256i __W, __mmask8 __U, __m256i __X, |
| 5734 | __m256i __Y) |
| 5735 | { |
| 5736 | return (__m256i) __builtin_ia32_psravq256_mask ((__v4di) __X, |
| 5737 | (__v4di) __Y, |
| 5738 | (__v4di) __W, |
| 5739 | (__mmask8) __U); |
| 5740 | } |
| 5741 | |
| 5742 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5743 | _mm256_maskz_srav_epi64 (__mmask8 __U, __m256i __X, __m256i __Y) |
| 5744 | { |
| 5745 | return (__m256i) __builtin_ia32_psravq256_mask ((__v4di) __X, |
| 5746 | (__v4di) __Y, |
| 5747 | (__v4di) |
| 5748 | _mm256_setzero_si256 (), |
| 5749 | (__mmask8) __U); |
| 5750 | } |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5751 | |
Michael Zuckerman | e654200 | 2016-05-23 08:01:48 +0000 | [diff] [blame] | 5752 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5753 | _mm_mask_mov_epi32 (__m128i __W, __mmask8 __U, __m128i __A) |
| 5754 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 5755 | return (__m128i) __builtin_ia32_selectd_128 ((__mmask8) __U, |
| 5756 | (__v4si) __A, |
| 5757 | (__v4si) __W); |
Michael Zuckerman | e654200 | 2016-05-23 08:01:48 +0000 | [diff] [blame] | 5758 | } |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5759 | |
Michael Zuckerman | e654200 | 2016-05-23 08:01:48 +0000 | [diff] [blame] | 5760 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5761 | _mm_maskz_mov_epi32 (__mmask8 __U, __m128i __A) |
| 5762 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 5763 | return (__m128i) __builtin_ia32_selectd_128 ((__mmask8) __U, |
| 5764 | (__v4si) __A, |
| 5765 | (__v4si) _mm_setzero_si128 ()); |
Michael Zuckerman | e654200 | 2016-05-23 08:01:48 +0000 | [diff] [blame] | 5766 | } |
| 5767 | |
| 5768 | |
| 5769 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5770 | _mm256_mask_mov_epi32 (__m256i __W, __mmask8 __U, __m256i __A) |
| 5771 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 5772 | return (__m256i) __builtin_ia32_selectd_256 ((__mmask8) __U, |
| 5773 | (__v8si) __A, |
| 5774 | (__v8si) __W); |
Michael Zuckerman | e654200 | 2016-05-23 08:01:48 +0000 | [diff] [blame] | 5775 | } |
| 5776 | |
| 5777 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5778 | _mm256_maskz_mov_epi32 (__mmask8 __U, __m256i __A) |
| 5779 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 5780 | return (__m256i) __builtin_ia32_selectd_256 ((__mmask8) __U, |
| 5781 | (__v8si) __A, |
| 5782 | (__v8si) _mm256_setzero_si256 ()); |
Michael Zuckerman | e654200 | 2016-05-23 08:01:48 +0000 | [diff] [blame] | 5783 | } |
| 5784 | |
| 5785 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5786 | _mm_mask_load_epi32 (__m128i __W, __mmask8 __U, void const *__P) |
| 5787 | { |
| 5788 | return (__m128i) __builtin_ia32_movdqa32load128_mask ((__v4si *) __P, |
| 5789 | (__v4si) __W, |
| 5790 | (__mmask8) |
| 5791 | __U); |
| 5792 | } |
| 5793 | |
| 5794 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5795 | _mm_maskz_load_epi32 (__mmask8 __U, void const *__P) |
| 5796 | { |
| 5797 | return (__m128i) __builtin_ia32_movdqa32load128_mask ((__v4si *) __P, |
| 5798 | (__v4si) |
| 5799 | _mm_setzero_si128 (), |
| 5800 | (__mmask8) |
| 5801 | __U); |
| 5802 | } |
| 5803 | |
| 5804 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5805 | _mm256_mask_load_epi32 (__m256i __W, __mmask8 __U, void const *__P) |
| 5806 | { |
| 5807 | return (__m256i) __builtin_ia32_movdqa32load256_mask ((__v8si *) __P, |
| 5808 | (__v8si) __W, |
| 5809 | (__mmask8) |
| 5810 | __U); |
| 5811 | } |
| 5812 | |
| 5813 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5814 | _mm256_maskz_load_epi32 (__mmask8 __U, void const *__P) |
| 5815 | { |
| 5816 | return (__m256i) __builtin_ia32_movdqa32load256_mask ((__v8si *) __P, |
| 5817 | (__v8si) |
| 5818 | _mm256_setzero_si256 (), |
| 5819 | (__mmask8) |
| 5820 | __U); |
| 5821 | } |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5822 | |
| 5823 | static __inline__ void __DEFAULT_FN_ATTRS |
| 5824 | _mm_mask_store_epi32 (void *__P, __mmask8 __U, __m128i __A) |
| 5825 | { |
| 5826 | __builtin_ia32_movdqa32store128_mask ((__v4si *) __P, |
| 5827 | (__v4si) __A, |
| 5828 | (__mmask8) __U); |
| 5829 | } |
| 5830 | |
| 5831 | static __inline__ void __DEFAULT_FN_ATTRS |
| 5832 | _mm256_mask_store_epi32 (void *__P, __mmask8 __U, __m256i __A) |
| 5833 | { |
| 5834 | __builtin_ia32_movdqa32store256_mask ((__v8si *) __P, |
| 5835 | (__v8si) __A, |
| 5836 | (__mmask8) __U); |
| 5837 | } |
| 5838 | |
| 5839 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5840 | _mm_mask_mov_epi64 (__m128i __W, __mmask8 __U, __m128i __A) |
| 5841 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 5842 | return (__m128i) __builtin_ia32_selectq_128 ((__mmask8) __U, |
| 5843 | (__v2di) __A, |
| 5844 | (__v2di) __W); |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5845 | } |
| 5846 | |
| 5847 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5848 | _mm_maskz_mov_epi64 (__mmask8 __U, __m128i __A) |
| 5849 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 5850 | return (__m128i) __builtin_ia32_selectq_128 ((__mmask8) __U, |
| 5851 | (__v2di) __A, |
| 5852 | (__v2di) _mm_setzero_di ()); |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5853 | } |
| 5854 | |
| 5855 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5856 | _mm256_mask_mov_epi64 (__m256i __W, __mmask8 __U, __m256i __A) |
| 5857 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 5858 | return (__m256i) __builtin_ia32_selectq_256 ((__mmask8) __U, |
| 5859 | (__v4di) __A, |
| 5860 | (__v4di) __W); |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5861 | } |
| 5862 | |
| 5863 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5864 | _mm256_maskz_mov_epi64 (__mmask8 __U, __m256i __A) |
| 5865 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 5866 | return (__m256i) __builtin_ia32_selectq_256 ((__mmask8) __U, |
| 5867 | (__v4di) __A, |
| 5868 | (__v4di) _mm256_setzero_si256 ()); |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5869 | } |
| 5870 | |
| 5871 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5872 | _mm_mask_load_epi64 (__m128i __W, __mmask8 __U, void const *__P) |
| 5873 | { |
| 5874 | return (__m128i) __builtin_ia32_movdqa64load128_mask ((__v2di *) __P, |
| 5875 | (__v2di) __W, |
| 5876 | (__mmask8) |
| 5877 | __U); |
| 5878 | } |
| 5879 | |
| 5880 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5881 | _mm_maskz_load_epi64 (__mmask8 __U, void const *__P) |
| 5882 | { |
| 5883 | return (__m128i) __builtin_ia32_movdqa64load128_mask ((__v2di *) __P, |
| 5884 | (__v2di) |
| 5885 | _mm_setzero_di (), |
| 5886 | (__mmask8) |
| 5887 | __U); |
| 5888 | } |
| 5889 | |
| 5890 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5891 | _mm256_mask_load_epi64 (__m256i __W, __mmask8 __U, void const *__P) |
| 5892 | { |
| 5893 | return (__m256i) __builtin_ia32_movdqa64load256_mask ((__v4di *) __P, |
| 5894 | (__v4di) __W, |
| 5895 | (__mmask8) |
| 5896 | __U); |
| 5897 | } |
| 5898 | |
| 5899 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5900 | _mm256_maskz_load_epi64 (__mmask8 __U, void const *__P) |
| 5901 | { |
| 5902 | return (__m256i) __builtin_ia32_movdqa64load256_mask ((__v4di *) __P, |
| 5903 | (__v4di) |
| 5904 | _mm256_setzero_si256 (), |
| 5905 | (__mmask8) |
| 5906 | __U); |
| 5907 | } |
| 5908 | |
| 5909 | static __inline__ void __DEFAULT_FN_ATTRS |
| 5910 | _mm_mask_store_epi64 (void *__P, __mmask8 __U, __m128i __A) |
| 5911 | { |
| 5912 | __builtin_ia32_movdqa64store128_mask ((__v2di *) __P, |
| 5913 | (__v2di) __A, |
| 5914 | (__mmask8) __U); |
| 5915 | } |
| 5916 | |
| 5917 | static __inline__ void __DEFAULT_FN_ATTRS |
| 5918 | _mm256_mask_store_epi64 (void *__P, __mmask8 __U, __m256i __A) |
| 5919 | { |
| 5920 | __builtin_ia32_movdqa64store256_mask ((__v4di *) __P, |
| 5921 | (__v4di) __A, |
| 5922 | (__mmask8) __U); |
| 5923 | } |
| 5924 | |
Michael Zuckerman | 0d67e4b | 2016-03-03 13:43:05 +0000 | [diff] [blame] | 5925 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 5926 | _mm_mask_movedup_pd (__m128d __W, __mmask8 __U, __m128d __A) |
| 5927 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 5928 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 5929 | (__v2df)_mm_movedup_pd(__A), |
| 5930 | (__v2df)__W); |
Michael Zuckerman | 0d67e4b | 2016-03-03 13:43:05 +0000 | [diff] [blame] | 5931 | } |
| 5932 | |
| 5933 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 5934 | _mm_maskz_movedup_pd (__mmask8 __U, __m128d __A) |
| 5935 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 5936 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 5937 | (__v2df)_mm_movedup_pd(__A), |
| 5938 | (__v2df)_mm_setzero_pd()); |
Michael Zuckerman | 0d67e4b | 2016-03-03 13:43:05 +0000 | [diff] [blame] | 5939 | } |
| 5940 | |
| 5941 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 5942 | _mm256_mask_movedup_pd (__m256d __W, __mmask8 __U, __m256d __A) |
| 5943 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 5944 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 5945 | (__v4df)_mm256_movedup_pd(__A), |
| 5946 | (__v4df)__W); |
Michael Zuckerman | 0d67e4b | 2016-03-03 13:43:05 +0000 | [diff] [blame] | 5947 | } |
| 5948 | |
| 5949 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 5950 | _mm256_maskz_movedup_pd (__mmask8 __U, __m256d __A) |
| 5951 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 5952 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 5953 | (__v4df)_mm256_movedup_pd(__A), |
| 5954 | (__v4df)_mm256_setzero_pd()); |
Michael Zuckerman | 0d67e4b | 2016-03-03 13:43:05 +0000 | [diff] [blame] | 5955 | } |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5956 | |
Michael Zuckerman | 912be16 | 2016-03-07 08:29:10 +0000 | [diff] [blame] | 5957 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5958 | #define _mm_mask_set1_epi32(O, M, A) __extension__ ({ \ |
| 5959 | (__m128i)__builtin_ia32_pbroadcastd128_gpr_mask((int)(A), \ |
| 5960 | (__v4si)(__m128i)(O), \ |
| 5961 | (__mmask8)(M)); }) |
Michael Zuckerman | 912be16 | 2016-03-07 08:29:10 +0000 | [diff] [blame] | 5962 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5963 | #define _mm_maskz_set1_epi32(M, A) __extension__ ({ \ |
| 5964 | (__m128i)__builtin_ia32_pbroadcastd128_gpr_mask((int)(A), \ |
| 5965 | (__v4si)_mm_setzero_si128(), \ |
| 5966 | (__mmask8)(M)); }) |
Michael Zuckerman | 912be16 | 2016-03-07 08:29:10 +0000 | [diff] [blame] | 5967 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5968 | #define _mm256_mask_set1_epi32(O, M, A) __extension__ ({ \ |
| 5969 | (__m256i)__builtin_ia32_pbroadcastd256_gpr_mask((int)(A), \ |
| 5970 | (__v8si)(__m256i)(O), \ |
| 5971 | (__mmask8)(M)); }) |
Michael Zuckerman | 912be16 | 2016-03-07 08:29:10 +0000 | [diff] [blame] | 5972 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5973 | #define _mm256_maskz_set1_epi32(M, A) __extension__ ({ \ |
| 5974 | (__m256i)__builtin_ia32_pbroadcastd256_gpr_mask((int)(A), \ |
| 5975 | (__v8si)_mm256_setzero_si256(), \ |
| 5976 | (__mmask8)(M)); }) |
Michael Zuckerman | 912be16 | 2016-03-07 08:29:10 +0000 | [diff] [blame] | 5977 | |
| 5978 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5979 | _mm_mask_set1_epi64 (__m128i __O, __mmask8 __M, long long __A) |
| 5980 | { |
| 5981 | return (__m128i) __builtin_ia32_pbroadcastq128_gpr_mask (__A, (__v2di) __O, |
| 5982 | __M); |
| 5983 | } |
| 5984 | |
| 5985 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5986 | _mm_maskz_set1_epi64 (__mmask8 __M, long long __A) |
| 5987 | { |
| 5988 | return (__m128i) __builtin_ia32_pbroadcastq128_gpr_mask (__A, |
| 5989 | (__v2di) |
| 5990 | _mm_setzero_si128 (), |
| 5991 | __M); |
| 5992 | } |
| 5993 | |
| 5994 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5995 | _mm256_mask_set1_epi64 (__m256i __O, __mmask8 __M, long long __A) |
| 5996 | { |
| 5997 | return (__m256i) __builtin_ia32_pbroadcastq256_gpr_mask (__A, (__v4di) __O, |
| 5998 | __M); |
| 5999 | } |
| 6000 | |
| 6001 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 6002 | _mm256_maskz_set1_epi64 (__mmask8 __M, long long __A) |
| 6003 | { |
| 6004 | return (__m256i) __builtin_ia32_pbroadcastq256_gpr_mask (__A, |
| 6005 | (__v4di) |
| 6006 | _mm256_setzero_si256 (), |
| 6007 | __M); |
| 6008 | } |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 6009 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6010 | #define _mm_fixupimm_pd(A, B, C, imm) __extension__ ({ \ |
| 6011 | (__m128d)__builtin_ia32_fixupimmpd128_mask((__v2df)(__m128d)(A), \ |
| 6012 | (__v2df)(__m128d)(B), \ |
| 6013 | (__v2di)(__m128i)(C), (int)(imm), \ |
| 6014 | (__mmask8)-1); }) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 6015 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6016 | #define _mm_mask_fixupimm_pd(A, U, B, C, imm) __extension__ ({ \ |
| 6017 | (__m128d)__builtin_ia32_fixupimmpd128_mask((__v2df)(__m128d)(A), \ |
| 6018 | (__v2df)(__m128d)(B), \ |
| 6019 | (__v2di)(__m128i)(C), (int)(imm), \ |
| 6020 | (__mmask8)(U)); }) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 6021 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6022 | #define _mm_maskz_fixupimm_pd(U, A, B, C, imm) __extension__ ({ \ |
| 6023 | (__m128d)__builtin_ia32_fixupimmpd128_maskz((__v2df)(__m128d)(A), \ |
| 6024 | (__v2df)(__m128d)(B), \ |
| 6025 | (__v2di)(__m128i)(C), \ |
| 6026 | (int)(imm), (__mmask8)(U)); }) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 6027 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6028 | #define _mm256_fixupimm_pd(A, B, C, imm) __extension__ ({ \ |
| 6029 | (__m256d)__builtin_ia32_fixupimmpd256_mask((__v4df)(__m256d)(A), \ |
| 6030 | (__v4df)(__m256d)(B), \ |
| 6031 | (__v4di)(__m256i)(C), (int)(imm), \ |
| 6032 | (__mmask8)-1); }) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 6033 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6034 | #define _mm256_mask_fixupimm_pd(A, U, B, C, imm) __extension__ ({ \ |
| 6035 | (__m256d)__builtin_ia32_fixupimmpd256_mask((__v4df)(__m256d)(A), \ |
| 6036 | (__v4df)(__m256d)(B), \ |
| 6037 | (__v4di)(__m256i)(C), (int)(imm), \ |
| 6038 | (__mmask8)(U)); }) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 6039 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6040 | #define _mm256_maskz_fixupimm_pd(U, A, B, C, imm) __extension__ ({ \ |
| 6041 | (__m256d)__builtin_ia32_fixupimmpd256_maskz((__v4df)(__m256d)(A), \ |
| 6042 | (__v4df)(__m256d)(B), \ |
| 6043 | (__v4di)(__m256i)(C), \ |
| 6044 | (int)(imm), (__mmask8)(U)); }) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 6045 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6046 | #define _mm_fixupimm_ps(A, B, C, imm) __extension__ ({ \ |
| 6047 | (__m128)__builtin_ia32_fixupimmps128_mask((__v4sf)(__m128)(A), \ |
| 6048 | (__v4sf)(__m128)(B), \ |
| 6049 | (__v4si)(__m128i)(C), (int)(imm), \ |
| 6050 | (__mmask8)-1); }) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 6051 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6052 | #define _mm_mask_fixupimm_ps(A, U, B, C, imm) __extension__ ({ \ |
| 6053 | (__m128)__builtin_ia32_fixupimmps128_mask((__v4sf)(__m128)(A), \ |
| 6054 | (__v4sf)(__m128)(B), \ |
| 6055 | (__v4si)(__m128i)(C), (int)(imm), \ |
| 6056 | (__mmask8)(U)); }) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 6057 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6058 | #define _mm_maskz_fixupimm_ps(U, A, B, C, imm) __extension__ ({ \ |
| 6059 | (__m128)__builtin_ia32_fixupimmps128_maskz((__v4sf)(__m128)(A), \ |
| 6060 | (__v4sf)(__m128)(B), \ |
| 6061 | (__v4si)(__m128i)(C), (int)(imm), \ |
| 6062 | (__mmask8)(U)); }) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 6063 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6064 | #define _mm256_fixupimm_ps(A, B, C, imm) __extension__ ({ \ |
| 6065 | (__m256)__builtin_ia32_fixupimmps256_mask((__v8sf)(__m256)(A), \ |
| 6066 | (__v8sf)(__m256)(B), \ |
| 6067 | (__v8si)(__m256i)(C), (int)(imm), \ |
| 6068 | (__mmask8)-1); }) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 6069 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6070 | #define _mm256_mask_fixupimm_ps(A, U, B, C, imm) __extension__ ({ \ |
| 6071 | (__m256)__builtin_ia32_fixupimmps256_mask((__v8sf)(__m256)(A), \ |
| 6072 | (__v8sf)(__m256)(B), \ |
| 6073 | (__v8si)(__m256i)(C), (int)(imm), \ |
| 6074 | (__mmask8)(U)); }) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 6075 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6076 | #define _mm256_maskz_fixupimm_ps(U, A, B, C, imm) __extension__ ({ \ |
| 6077 | (__m256)__builtin_ia32_fixupimmps256_maskz((__v8sf)(__m256)(A), \ |
| 6078 | (__v8sf)(__m256)(B), \ |
| 6079 | (__v8si)(__m256i)(C), (int)(imm), \ |
| 6080 | (__mmask8)(U)); }) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 6081 | |
| 6082 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 6083 | _mm_mask_load_pd (__m128d __W, __mmask8 __U, void const *__P) |
| 6084 | { |
| 6085 | return (__m128d) __builtin_ia32_loadapd128_mask ((__v2df *) __P, |
| 6086 | (__v2df) __W, |
| 6087 | (__mmask8) __U); |
| 6088 | } |
| 6089 | |
| 6090 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 6091 | _mm_maskz_load_pd (__mmask8 __U, void const *__P) |
| 6092 | { |
| 6093 | return (__m128d) __builtin_ia32_loadapd128_mask ((__v2df *) __P, |
| 6094 | (__v2df) |
| 6095 | _mm_setzero_pd (), |
| 6096 | (__mmask8) __U); |
| 6097 | } |
| 6098 | |
| 6099 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 6100 | _mm256_mask_load_pd (__m256d __W, __mmask8 __U, void const *__P) |
| 6101 | { |
| 6102 | return (__m256d) __builtin_ia32_loadapd256_mask ((__v4df *) __P, |
| 6103 | (__v4df) __W, |
| 6104 | (__mmask8) __U); |
| 6105 | } |
| 6106 | |
| 6107 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 6108 | _mm256_maskz_load_pd (__mmask8 __U, void const *__P) |
| 6109 | { |
| 6110 | return (__m256d) __builtin_ia32_loadapd256_mask ((__v4df *) __P, |
| 6111 | (__v4df) |
| 6112 | _mm256_setzero_pd (), |
| 6113 | (__mmask8) __U); |
| 6114 | } |
| 6115 | |
| 6116 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 6117 | _mm_mask_load_ps (__m128 __W, __mmask8 __U, void const *__P) |
| 6118 | { |
| 6119 | return (__m128) __builtin_ia32_loadaps128_mask ((__v4sf *) __P, |
| 6120 | (__v4sf) __W, |
| 6121 | (__mmask8) __U); |
| 6122 | } |
| 6123 | |
| 6124 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 6125 | _mm_maskz_load_ps (__mmask8 __U, void const *__P) |
| 6126 | { |
| 6127 | return (__m128) __builtin_ia32_loadaps128_mask ((__v4sf *) __P, |
| 6128 | (__v4sf) |
| 6129 | _mm_setzero_ps (), |
| 6130 | (__mmask8) __U); |
| 6131 | } |
| 6132 | |
| 6133 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 6134 | _mm256_mask_load_ps (__m256 __W, __mmask8 __U, void const *__P) |
| 6135 | { |
| 6136 | return (__m256) __builtin_ia32_loadaps256_mask ((__v8sf *) __P, |
| 6137 | (__v8sf) __W, |
| 6138 | (__mmask8) __U); |
| 6139 | } |
| 6140 | |
| 6141 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 6142 | _mm256_maskz_load_ps (__mmask8 __U, void const *__P) |
| 6143 | { |
| 6144 | return (__m256) __builtin_ia32_loadaps256_mask ((__v8sf *) __P, |
| 6145 | (__v8sf) |
| 6146 | _mm256_setzero_ps (), |
| 6147 | (__mmask8) __U); |
| 6148 | } |
| 6149 | |
| 6150 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6151 | _mm_mask_loadu_epi64 (__m128i __W, __mmask8 __U, void const *__P) |
| 6152 | { |
| 6153 | return (__m128i) __builtin_ia32_loaddqudi128_mask ((__v2di *) __P, |
| 6154 | (__v2di) __W, |
| 6155 | (__mmask8) __U); |
| 6156 | } |
| 6157 | |
| 6158 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6159 | _mm_maskz_loadu_epi64 (__mmask8 __U, void const *__P) |
| 6160 | { |
| 6161 | return (__m128i) __builtin_ia32_loaddqudi128_mask ((__v2di *) __P, |
| 6162 | (__v2di) |
| 6163 | _mm_setzero_si128 (), |
| 6164 | (__mmask8) __U); |
| 6165 | } |
| 6166 | |
| 6167 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 6168 | _mm256_mask_loadu_epi64 (__m256i __W, __mmask8 __U, void const *__P) |
| 6169 | { |
| 6170 | return (__m256i) __builtin_ia32_loaddqudi256_mask ((__v4di *) __P, |
| 6171 | (__v4di) __W, |
| 6172 | (__mmask8) __U); |
| 6173 | } |
| 6174 | |
| 6175 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 6176 | _mm256_maskz_loadu_epi64 (__mmask8 __U, void const *__P) |
| 6177 | { |
| 6178 | return (__m256i) __builtin_ia32_loaddqudi256_mask ((__v4di *) __P, |
| 6179 | (__v4di) |
| 6180 | _mm256_setzero_si256 (), |
| 6181 | (__mmask8) __U); |
| 6182 | } |
| 6183 | |
| 6184 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6185 | _mm_mask_loadu_epi32 (__m128i __W, __mmask8 __U, void const *__P) |
| 6186 | { |
| 6187 | return (__m128i) __builtin_ia32_loaddqusi128_mask ((__v4si *) __P, |
| 6188 | (__v4si) __W, |
| 6189 | (__mmask8) __U); |
| 6190 | } |
| 6191 | |
| 6192 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6193 | _mm_maskz_loadu_epi32 (__mmask8 __U, void const *__P) |
| 6194 | { |
| 6195 | return (__m128i) __builtin_ia32_loaddqusi128_mask ((__v4si *) __P, |
| 6196 | (__v4si) |
| 6197 | _mm_setzero_si128 (), |
| 6198 | (__mmask8) __U); |
| 6199 | } |
| 6200 | |
| 6201 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 6202 | _mm256_mask_loadu_epi32 (__m256i __W, __mmask8 __U, void const *__P) |
| 6203 | { |
| 6204 | return (__m256i) __builtin_ia32_loaddqusi256_mask ((__v8si *) __P, |
| 6205 | (__v8si) __W, |
| 6206 | (__mmask8) __U); |
| 6207 | } |
| 6208 | |
| 6209 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 6210 | _mm256_maskz_loadu_epi32 (__mmask8 __U, void const *__P) |
| 6211 | { |
| 6212 | return (__m256i) __builtin_ia32_loaddqusi256_mask ((__v8si *) __P, |
| 6213 | (__v8si) |
| 6214 | _mm256_setzero_si256 (), |
| 6215 | (__mmask8) __U); |
| 6216 | } |
| 6217 | |
| 6218 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 6219 | _mm_mask_loadu_pd (__m128d __W, __mmask8 __U, void const *__P) |
| 6220 | { |
| 6221 | return (__m128d) __builtin_ia32_loadupd128_mask ((__v2df *) __P, |
| 6222 | (__v2df) __W, |
| 6223 | (__mmask8) __U); |
| 6224 | } |
| 6225 | |
| 6226 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 6227 | _mm_maskz_loadu_pd (__mmask8 __U, void const *__P) |
| 6228 | { |
| 6229 | return (__m128d) __builtin_ia32_loadupd128_mask ((__v2df *) __P, |
| 6230 | (__v2df) |
| 6231 | _mm_setzero_pd (), |
| 6232 | (__mmask8) __U); |
| 6233 | } |
| 6234 | |
| 6235 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 6236 | _mm256_mask_loadu_pd (__m256d __W, __mmask8 __U, void const *__P) |
| 6237 | { |
| 6238 | return (__m256d) __builtin_ia32_loadupd256_mask ((__v4df *) __P, |
| 6239 | (__v4df) __W, |
| 6240 | (__mmask8) __U); |
| 6241 | } |
| 6242 | |
| 6243 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 6244 | _mm256_maskz_loadu_pd (__mmask8 __U, void const *__P) |
| 6245 | { |
| 6246 | return (__m256d) __builtin_ia32_loadupd256_mask ((__v4df *) __P, |
| 6247 | (__v4df) |
| 6248 | _mm256_setzero_pd (), |
| 6249 | (__mmask8) __U); |
| 6250 | } |
| 6251 | |
| 6252 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 6253 | _mm_mask_loadu_ps (__m128 __W, __mmask8 __U, void const *__P) |
| 6254 | { |
| 6255 | return (__m128) __builtin_ia32_loadups128_mask ((__v4sf *) __P, |
| 6256 | (__v4sf) __W, |
| 6257 | (__mmask8) __U); |
| 6258 | } |
| 6259 | |
| 6260 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 6261 | _mm_maskz_loadu_ps (__mmask8 __U, void const *__P) |
| 6262 | { |
| 6263 | return (__m128) __builtin_ia32_loadups128_mask ((__v4sf *) __P, |
| 6264 | (__v4sf) |
| 6265 | _mm_setzero_ps (), |
| 6266 | (__mmask8) __U); |
| 6267 | } |
| 6268 | |
| 6269 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 6270 | _mm256_mask_loadu_ps (__m256 __W, __mmask8 __U, void const *__P) |
| 6271 | { |
| 6272 | return (__m256) __builtin_ia32_loadups256_mask ((__v8sf *) __P, |
| 6273 | (__v8sf) __W, |
| 6274 | (__mmask8) __U); |
| 6275 | } |
| 6276 | |
| 6277 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 6278 | _mm256_maskz_loadu_ps (__mmask8 __U, void const *__P) |
| 6279 | { |
| 6280 | return (__m256) __builtin_ia32_loadups256_mask ((__v8sf *) __P, |
| 6281 | (__v8sf) |
| 6282 | _mm256_setzero_ps (), |
| 6283 | (__mmask8) __U); |
| 6284 | } |
Michael Zuckerman | fa7ccc5 | 2016-04-10 10:51:04 +0000 | [diff] [blame] | 6285 | |
| 6286 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6287 | _mm_mask_store_pd (void *__P, __mmask8 __U, __m128d __A) |
| 6288 | { |
| 6289 | __builtin_ia32_storeapd128_mask ((__v2df *) __P, |
| 6290 | (__v2df) __A, |
| 6291 | (__mmask8) __U); |
| 6292 | } |
| 6293 | |
| 6294 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6295 | _mm256_mask_store_pd (void *__P, __mmask8 __U, __m256d __A) |
| 6296 | { |
| 6297 | __builtin_ia32_storeapd256_mask ((__v4df *) __P, |
| 6298 | (__v4df) __A, |
| 6299 | (__mmask8) __U); |
| 6300 | } |
| 6301 | |
| 6302 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6303 | _mm_mask_store_ps (void *__P, __mmask8 __U, __m128 __A) |
| 6304 | { |
| 6305 | __builtin_ia32_storeaps128_mask ((__v4sf *) __P, |
| 6306 | (__v4sf) __A, |
| 6307 | (__mmask8) __U); |
| 6308 | } |
| 6309 | |
| 6310 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6311 | _mm256_mask_store_ps (void *__P, __mmask8 __U, __m256 __A) |
| 6312 | { |
| 6313 | __builtin_ia32_storeaps256_mask ((__v8sf *) __P, |
| 6314 | (__v8sf) __A, |
| 6315 | (__mmask8) __U); |
| 6316 | } |
| 6317 | |
| 6318 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6319 | _mm_mask_storeu_epi64 (void *__P, __mmask8 __U, __m128i __A) |
| 6320 | { |
| 6321 | __builtin_ia32_storedqudi128_mask ((__v2di *) __P, |
| 6322 | (__v2di) __A, |
| 6323 | (__mmask8) __U); |
| 6324 | } |
| 6325 | |
| 6326 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6327 | _mm256_mask_storeu_epi64 (void *__P, __mmask8 __U, __m256i __A) |
| 6328 | { |
| 6329 | __builtin_ia32_storedqudi256_mask ((__v4di *) __P, |
| 6330 | (__v4di) __A, |
| 6331 | (__mmask8) __U); |
| 6332 | } |
| 6333 | |
| 6334 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6335 | _mm_mask_storeu_epi32 (void *__P, __mmask8 __U, __m128i __A) |
| 6336 | { |
| 6337 | __builtin_ia32_storedqusi128_mask ((__v4si *) __P, |
| 6338 | (__v4si) __A, |
| 6339 | (__mmask8) __U); |
| 6340 | } |
| 6341 | |
| 6342 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6343 | _mm256_mask_storeu_epi32 (void *__P, __mmask8 __U, __m256i __A) |
| 6344 | { |
| 6345 | __builtin_ia32_storedqusi256_mask ((__v8si *) __P, |
| 6346 | (__v8si) __A, |
| 6347 | (__mmask8) __U); |
| 6348 | } |
| 6349 | |
| 6350 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6351 | _mm_mask_storeu_pd (void *__P, __mmask8 __U, __m128d __A) |
| 6352 | { |
| 6353 | __builtin_ia32_storeupd128_mask ((__v2df *) __P, |
| 6354 | (__v2df) __A, |
| 6355 | (__mmask8) __U); |
| 6356 | } |
| 6357 | |
| 6358 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6359 | _mm256_mask_storeu_pd (void *__P, __mmask8 __U, __m256d __A) |
| 6360 | { |
| 6361 | __builtin_ia32_storeupd256_mask ((__v4df *) __P, |
| 6362 | (__v4df) __A, |
| 6363 | (__mmask8) __U); |
| 6364 | } |
| 6365 | |
| 6366 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6367 | _mm_mask_storeu_ps (void *__P, __mmask8 __U, __m128 __A) |
| 6368 | { |
| 6369 | __builtin_ia32_storeups128_mask ((__v4sf *) __P, |
| 6370 | (__v4sf) __A, |
| 6371 | (__mmask8) __U); |
| 6372 | } |
| 6373 | |
| 6374 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6375 | _mm256_mask_storeu_ps (void *__P, __mmask8 __U, __m256 __A) |
| 6376 | { |
| 6377 | __builtin_ia32_storeups256_mask ((__v8sf *) __P, |
| 6378 | (__v8sf) __A, |
| 6379 | (__mmask8) __U); |
| 6380 | } |
| 6381 | |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6382 | |
| 6383 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6384 | _mm_mask_unpackhi_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6385 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6386 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 6387 | (__v2df)_mm_unpackhi_pd(__A, __B), |
| 6388 | (__v2df)__W); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6389 | } |
| 6390 | |
| 6391 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6392 | _mm_maskz_unpackhi_pd(__mmask8 __U, __m128d __A, __m128d __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6393 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6394 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 6395 | (__v2df)_mm_unpackhi_pd(__A, __B), |
| 6396 | (__v2df)_mm_setzero_pd()); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6397 | } |
| 6398 | |
| 6399 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6400 | _mm256_mask_unpackhi_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6401 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6402 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 6403 | (__v4df)_mm256_unpackhi_pd(__A, __B), |
| 6404 | (__v4df)__W); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6405 | } |
| 6406 | |
| 6407 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6408 | _mm256_maskz_unpackhi_pd(__mmask8 __U, __m256d __A, __m256d __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6409 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6410 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 6411 | (__v4df)_mm256_unpackhi_pd(__A, __B), |
| 6412 | (__v4df)_mm256_setzero_pd()); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6413 | } |
| 6414 | |
| 6415 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6416 | _mm_mask_unpackhi_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6417 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6418 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 6419 | (__v4sf)_mm_unpackhi_ps(__A, __B), |
| 6420 | (__v4sf)__W); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6421 | } |
| 6422 | |
| 6423 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6424 | _mm_maskz_unpackhi_ps(__mmask8 __U, __m128 __A, __m128 __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6425 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6426 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 6427 | (__v4sf)_mm_unpackhi_ps(__A, __B), |
| 6428 | (__v4sf)_mm_setzero_ps()); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6429 | } |
| 6430 | |
| 6431 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6432 | _mm256_mask_unpackhi_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6433 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6434 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 6435 | (__v8sf)_mm256_unpackhi_ps(__A, __B), |
| 6436 | (__v8sf)__W); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6437 | } |
| 6438 | |
| 6439 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6440 | _mm256_maskz_unpackhi_ps(__mmask8 __U, __m256 __A, __m256 __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6441 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6442 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 6443 | (__v8sf)_mm256_unpackhi_ps(__A, __B), |
| 6444 | (__v8sf)_mm256_setzero_ps()); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6445 | } |
| 6446 | |
| 6447 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6448 | _mm_mask_unpacklo_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6449 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6450 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 6451 | (__v2df)_mm_unpacklo_pd(__A, __B), |
| 6452 | (__v2df)__W); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6453 | } |
| 6454 | |
| 6455 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6456 | _mm_maskz_unpacklo_pd(__mmask8 __U, __m128d __A, __m128d __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6457 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6458 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 6459 | (__v2df)_mm_unpacklo_pd(__A, __B), |
| 6460 | (__v2df)_mm_setzero_pd()); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6461 | } |
| 6462 | |
| 6463 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6464 | _mm256_mask_unpacklo_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6465 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6466 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 6467 | (__v4df)_mm256_unpacklo_pd(__A, __B), |
| 6468 | (__v4df)__W); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6469 | } |
| 6470 | |
| 6471 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6472 | _mm256_maskz_unpacklo_pd(__mmask8 __U, __m256d __A, __m256d __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6473 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6474 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 6475 | (__v4df)_mm256_unpacklo_pd(__A, __B), |
| 6476 | (__v4df)_mm256_setzero_pd()); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6477 | } |
| 6478 | |
| 6479 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6480 | _mm_mask_unpacklo_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6481 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6482 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 6483 | (__v4sf)_mm_unpacklo_ps(__A, __B), |
| 6484 | (__v4sf)__W); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6485 | } |
| 6486 | |
| 6487 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6488 | _mm_maskz_unpacklo_ps(__mmask8 __U, __m128 __A, __m128 __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6489 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6490 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 6491 | (__v4sf)_mm_unpacklo_ps(__A, __B), |
| 6492 | (__v4sf)_mm_setzero_ps()); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6493 | } |
| 6494 | |
| 6495 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6496 | _mm256_mask_unpacklo_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6497 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6498 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 6499 | (__v8sf)_mm256_unpacklo_ps(__A, __B), |
| 6500 | (__v8sf)__W); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6501 | } |
| 6502 | |
| 6503 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6504 | _mm256_maskz_unpacklo_ps(__mmask8 __U, __m256 __A, __m256 __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6505 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6506 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 6507 | (__v8sf)_mm256_unpacklo_ps(__A, __B), |
| 6508 | (__v8sf)_mm256_setzero_ps()); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6509 | } |
| 6510 | |
| 6511 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 6512 | _mm_rcp14_pd (__m128d __A) |
| 6513 | { |
| 6514 | return (__m128d) __builtin_ia32_rcp14pd128_mask ((__v2df) __A, |
| 6515 | (__v2df) |
| 6516 | _mm_setzero_pd (), |
| 6517 | (__mmask8) -1); |
| 6518 | } |
| 6519 | |
| 6520 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 6521 | _mm_mask_rcp14_pd (__m128d __W, __mmask8 __U, __m128d __A) |
| 6522 | { |
| 6523 | return (__m128d) __builtin_ia32_rcp14pd128_mask ((__v2df) __A, |
| 6524 | (__v2df) __W, |
| 6525 | (__mmask8) __U); |
| 6526 | } |
| 6527 | |
| 6528 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 6529 | _mm_maskz_rcp14_pd (__mmask8 __U, __m128d __A) |
| 6530 | { |
| 6531 | return (__m128d) __builtin_ia32_rcp14pd128_mask ((__v2df) __A, |
| 6532 | (__v2df) |
| 6533 | _mm_setzero_pd (), |
| 6534 | (__mmask8) __U); |
| 6535 | } |
| 6536 | |
| 6537 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 6538 | _mm256_rcp14_pd (__m256d __A) |
| 6539 | { |
| 6540 | return (__m256d) __builtin_ia32_rcp14pd256_mask ((__v4df) __A, |
| 6541 | (__v4df) |
| 6542 | _mm256_setzero_pd (), |
| 6543 | (__mmask8) -1); |
| 6544 | } |
| 6545 | |
| 6546 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 6547 | _mm256_mask_rcp14_pd (__m256d __W, __mmask8 __U, __m256d __A) |
| 6548 | { |
| 6549 | return (__m256d) __builtin_ia32_rcp14pd256_mask ((__v4df) __A, |
| 6550 | (__v4df) __W, |
| 6551 | (__mmask8) __U); |
| 6552 | } |
| 6553 | |
| 6554 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 6555 | _mm256_maskz_rcp14_pd (__mmask8 __U, __m256d __A) |
| 6556 | { |
| 6557 | return (__m256d) __builtin_ia32_rcp14pd256_mask ((__v4df) __A, |
| 6558 | (__v4df) |
| 6559 | _mm256_setzero_pd (), |
| 6560 | (__mmask8) __U); |
| 6561 | } |
| 6562 | |
| 6563 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 6564 | _mm_rcp14_ps (__m128 __A) |
| 6565 | { |
| 6566 | return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A, |
| 6567 | (__v4sf) |
| 6568 | _mm_setzero_ps (), |
| 6569 | (__mmask8) -1); |
| 6570 | } |
| 6571 | |
| 6572 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 6573 | _mm_mask_rcp14_ps (__m128 __W, __mmask8 __U, __m128 __A) |
| 6574 | { |
| 6575 | return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A, |
| 6576 | (__v4sf) __W, |
| 6577 | (__mmask8) __U); |
| 6578 | } |
| 6579 | |
| 6580 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 6581 | _mm_maskz_rcp14_ps (__mmask8 __U, __m128 __A) |
| 6582 | { |
| 6583 | return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A, |
| 6584 | (__v4sf) |
| 6585 | _mm_setzero_ps (), |
| 6586 | (__mmask8) __U); |
| 6587 | } |
| 6588 | |
| 6589 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 6590 | _mm256_rcp14_ps (__m256 __A) |
| 6591 | { |
| 6592 | return (__m256) __builtin_ia32_rcp14ps256_mask ((__v8sf) __A, |
| 6593 | (__v8sf) |
| 6594 | _mm256_setzero_ps (), |
| 6595 | (__mmask8) -1); |
| 6596 | } |
| 6597 | |
| 6598 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 6599 | _mm256_mask_rcp14_ps (__m256 __W, __mmask8 __U, __m256 __A) |
| 6600 | { |
| 6601 | return (__m256) __builtin_ia32_rcp14ps256_mask ((__v8sf) __A, |
| 6602 | (__v8sf) __W, |
| 6603 | (__mmask8) __U); |
| 6604 | } |
| 6605 | |
| 6606 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 6607 | _mm256_maskz_rcp14_ps (__mmask8 __U, __m256 __A) |
| 6608 | { |
| 6609 | return (__m256) __builtin_ia32_rcp14ps256_mask ((__v8sf) __A, |
| 6610 | (__v8sf) |
| 6611 | _mm256_setzero_ps (), |
| 6612 | (__mmask8) __U); |
| 6613 | } |
| 6614 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6615 | #define _mm_mask_permute_pd(W, U, X, C) __extension__ ({ \ |
Craig Topper | b3a4477 | 2016-07-02 05:36:43 +0000 | [diff] [blame] | 6616 | (__m128d)__builtin_ia32_selectpd_128((__mmask8)(U), \ |
| 6617 | (__v2df)_mm_permute_pd((X), (C)), \ |
| 6618 | (__v2df)(__m128d)(W)); }) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 6619 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6620 | #define _mm_maskz_permute_pd(U, X, C) __extension__ ({ \ |
Craig Topper | b3a4477 | 2016-07-02 05:36:43 +0000 | [diff] [blame] | 6621 | (__m128d)__builtin_ia32_selectpd_128((__mmask8)(U), \ |
| 6622 | (__v2df)_mm_permute_pd((X), (C)), \ |
| 6623 | (__v2df)_mm_setzero_pd()); }) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 6624 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6625 | #define _mm256_mask_permute_pd(W, U, X, C) __extension__ ({ \ |
Craig Topper | b3a4477 | 2016-07-02 05:36:43 +0000 | [diff] [blame] | 6626 | (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \ |
| 6627 | (__v4df)_mm256_permute_pd((X), (C)), \ |
| 6628 | (__v4df)(__m256d)(W)); }) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 6629 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6630 | #define _mm256_maskz_permute_pd(U, X, C) __extension__ ({ \ |
Craig Topper | b3a4477 | 2016-07-02 05:36:43 +0000 | [diff] [blame] | 6631 | (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \ |
| 6632 | (__v4df)_mm256_permute_pd((X), (C)), \ |
| 6633 | (__v4df)_mm256_setzero_pd()); }) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 6634 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6635 | #define _mm_mask_permute_ps(W, U, X, C) __extension__ ({ \ |
Craig Topper | b3a4477 | 2016-07-02 05:36:43 +0000 | [diff] [blame] | 6636 | (__m128)__builtin_ia32_selectps_128((__mmask8)(U), \ |
| 6637 | (__v4sf)_mm_permute_ps((X), (C)), \ |
| 6638 | (__v4sf)(__m128)(W)); }) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 6639 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6640 | #define _mm_maskz_permute_ps(U, X, C) __extension__ ({ \ |
Craig Topper | b3a4477 | 2016-07-02 05:36:43 +0000 | [diff] [blame] | 6641 | (__m128)__builtin_ia32_selectps_128((__mmask8)(U), \ |
| 6642 | (__v4sf)_mm_permute_ps((X), (C)), \ |
| 6643 | (__v4sf)_mm_setzero_ps()); }) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 6644 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6645 | #define _mm256_mask_permute_ps(W, U, X, C) __extension__ ({ \ |
Craig Topper | b3a4477 | 2016-07-02 05:36:43 +0000 | [diff] [blame] | 6646 | (__m256)__builtin_ia32_selectps_256((__mmask8)(U), \ |
| 6647 | (__v8sf)_mm256_permute_ps((X), (C)), \ |
| 6648 | (__v8sf)(__m256)(W)); }) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 6649 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6650 | #define _mm256_maskz_permute_ps(U, X, C) __extension__ ({ \ |
Craig Topper | b3a4477 | 2016-07-02 05:36:43 +0000 | [diff] [blame] | 6651 | (__m256)__builtin_ia32_selectps_256((__mmask8)(U), \ |
| 6652 | (__v8sf)_mm256_permute_ps((X), (C)), \ |
| 6653 | (__v8sf)_mm256_setzero_ps()); }) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 6654 | |
| 6655 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 6656 | _mm_mask_permutevar_pd (__m128d __W, __mmask8 __U, __m128d __A, |
| 6657 | __m128i __C) |
| 6658 | { |
| 6659 | return (__m128d) __builtin_ia32_vpermilvarpd_mask ((__v2df) __A, |
| 6660 | (__v2di) __C, |
| 6661 | (__v2df) __W, |
| 6662 | (__mmask8) __U); |
| 6663 | } |
| 6664 | |
| 6665 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 6666 | _mm_maskz_permutevar_pd (__mmask8 __U, __m128d __A, __m128i __C) |
| 6667 | { |
| 6668 | return (__m128d) __builtin_ia32_vpermilvarpd_mask ((__v2df) __A, |
| 6669 | (__v2di) __C, |
| 6670 | (__v2df) |
| 6671 | _mm_setzero_pd (), |
| 6672 | (__mmask8) __U); |
| 6673 | } |
| 6674 | |
| 6675 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 6676 | _mm256_mask_permutevar_pd (__m256d __W, __mmask8 __U, __m256d __A, |
| 6677 | __m256i __C) |
| 6678 | { |
| 6679 | return (__m256d) __builtin_ia32_vpermilvarpd256_mask ((__v4df) __A, |
| 6680 | (__v4di) __C, |
| 6681 | (__v4df) __W, |
| 6682 | (__mmask8) |
| 6683 | __U); |
| 6684 | } |
| 6685 | |
| 6686 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 6687 | _mm256_maskz_permutevar_pd (__mmask8 __U, __m256d __A, __m256i __C) |
| 6688 | { |
| 6689 | return (__m256d) __builtin_ia32_vpermilvarpd256_mask ((__v4df) __A, |
| 6690 | (__v4di) __C, |
| 6691 | (__v4df) |
| 6692 | _mm256_setzero_pd (), |
| 6693 | (__mmask8) |
| 6694 | __U); |
| 6695 | } |
| 6696 | |
| 6697 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 6698 | _mm_mask_permutevar_ps (__m128 __W, __mmask8 __U, __m128 __A, |
| 6699 | __m128i __C) |
| 6700 | { |
| 6701 | return (__m128) __builtin_ia32_vpermilvarps_mask ((__v4sf) __A, |
| 6702 | (__v4si) __C, |
| 6703 | (__v4sf) __W, |
| 6704 | (__mmask8) __U); |
| 6705 | } |
| 6706 | |
| 6707 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 6708 | _mm_maskz_permutevar_ps (__mmask8 __U, __m128 __A, __m128i __C) |
| 6709 | { |
| 6710 | return (__m128) __builtin_ia32_vpermilvarps_mask ((__v4sf) __A, |
| 6711 | (__v4si) __C, |
| 6712 | (__v4sf) |
| 6713 | _mm_setzero_ps (), |
| 6714 | (__mmask8) __U); |
| 6715 | } |
| 6716 | |
| 6717 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 6718 | _mm256_mask_permutevar_ps (__m256 __W, __mmask8 __U, __m256 __A, |
| 6719 | __m256i __C) |
| 6720 | { |
| 6721 | return (__m256) __builtin_ia32_vpermilvarps256_mask ((__v8sf) __A, |
| 6722 | (__v8si) __C, |
| 6723 | (__v8sf) __W, |
| 6724 | (__mmask8) __U); |
| 6725 | } |
| 6726 | |
| 6727 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 6728 | _mm256_maskz_permutevar_ps (__mmask8 __U, __m256 __A, __m256i __C) |
| 6729 | { |
| 6730 | return (__m256) __builtin_ia32_vpermilvarps256_mask ((__v8sf) __A, |
| 6731 | (__v8si) __C, |
| 6732 | (__v8sf) |
| 6733 | _mm256_setzero_ps (), |
| 6734 | (__mmask8) __U); |
| 6735 | } |
| 6736 | |
Michael Zuckerman | 0752509 | 2016-04-11 10:22:07 +0000 | [diff] [blame] | 6737 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6738 | _mm_test_epi32_mask (__m128i __A, __m128i __B) |
| 6739 | { |
| 6740 | return (__mmask8) __builtin_ia32_ptestmd128 ((__v4si) __A, |
| 6741 | (__v4si) __B, |
| 6742 | (__mmask8) -1); |
| 6743 | } |
| 6744 | |
| 6745 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6746 | _mm_mask_test_epi32_mask (__mmask8 __U, __m128i __A, __m128i __B) |
| 6747 | { |
| 6748 | return (__mmask8) __builtin_ia32_ptestmd128 ((__v4si) __A, |
| 6749 | (__v4si) __B, __U); |
| 6750 | } |
| 6751 | |
| 6752 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6753 | _mm256_test_epi32_mask (__m256i __A, __m256i __B) |
| 6754 | { |
| 6755 | return (__mmask8) __builtin_ia32_ptestmd256 ((__v8si) __A, |
| 6756 | (__v8si) __B, |
| 6757 | (__mmask8) -1); |
| 6758 | } |
| 6759 | |
| 6760 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6761 | _mm256_mask_test_epi32_mask (__mmask8 __U, __m256i __A, __m256i __B) |
| 6762 | { |
| 6763 | return (__mmask8) __builtin_ia32_ptestmd256 ((__v8si) __A, |
| 6764 | (__v8si) __B, __U); |
| 6765 | } |
| 6766 | |
| 6767 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6768 | _mm_test_epi64_mask (__m128i __A, __m128i __B) |
| 6769 | { |
| 6770 | return (__mmask8) __builtin_ia32_ptestmq128 ((__v2di) __A, |
| 6771 | (__v2di) __B, |
| 6772 | (__mmask8) -1); |
| 6773 | } |
| 6774 | |
| 6775 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6776 | _mm_mask_test_epi64_mask (__mmask8 __U, __m128i __A, __m128i __B) |
| 6777 | { |
| 6778 | return (__mmask8) __builtin_ia32_ptestmq128 ((__v2di) __A, |
| 6779 | (__v2di) __B, __U); |
| 6780 | } |
| 6781 | |
| 6782 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6783 | _mm256_test_epi64_mask (__m256i __A, __m256i __B) |
| 6784 | { |
| 6785 | return (__mmask8) __builtin_ia32_ptestmq256 ((__v4di) __A, |
| 6786 | (__v4di) __B, |
| 6787 | (__mmask8) -1); |
| 6788 | } |
| 6789 | |
| 6790 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6791 | _mm256_mask_test_epi64_mask (__mmask8 __U, __m256i __A, __m256i __B) |
| 6792 | { |
| 6793 | return (__mmask8) __builtin_ia32_ptestmq256 ((__v4di) __A, |
| 6794 | (__v4di) __B, __U); |
| 6795 | } |
| 6796 | |
| 6797 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6798 | _mm_testn_epi32_mask (__m128i __A, __m128i __B) |
| 6799 | { |
| 6800 | return (__mmask8) __builtin_ia32_ptestnmd128 ((__v4si) __A, |
| 6801 | (__v4si) __B, |
| 6802 | (__mmask8) -1); |
| 6803 | } |
| 6804 | |
| 6805 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6806 | _mm_mask_testn_epi32_mask (__mmask8 __U, __m128i __A, __m128i __B) |
| 6807 | { |
| 6808 | return (__mmask8) __builtin_ia32_ptestnmd128 ((__v4si) __A, |
| 6809 | (__v4si) __B, __U); |
| 6810 | } |
| 6811 | |
| 6812 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6813 | _mm256_testn_epi32_mask (__m256i __A, __m256i __B) |
| 6814 | { |
| 6815 | return (__mmask8) __builtin_ia32_ptestnmd256 ((__v8si) __A, |
| 6816 | (__v8si) __B, |
| 6817 | (__mmask8) -1); |
| 6818 | } |
| 6819 | |
| 6820 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6821 | _mm256_mask_testn_epi32_mask (__mmask8 __U, __m256i __A, __m256i __B) |
| 6822 | { |
| 6823 | return (__mmask8) __builtin_ia32_ptestnmd256 ((__v8si) __A, |
| 6824 | (__v8si) __B, __U); |
| 6825 | } |
| 6826 | |
| 6827 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6828 | _mm_testn_epi64_mask (__m128i __A, __m128i __B) |
| 6829 | { |
| 6830 | return (__mmask8) __builtin_ia32_ptestnmq128 ((__v2di) __A, |
| 6831 | (__v2di) __B, |
| 6832 | (__mmask8) -1); |
| 6833 | } |
| 6834 | |
| 6835 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6836 | _mm_mask_testn_epi64_mask (__mmask8 __U, __m128i __A, __m128i __B) |
| 6837 | { |
| 6838 | return (__mmask8) __builtin_ia32_ptestnmq128 ((__v2di) __A, |
| 6839 | (__v2di) __B, __U); |
| 6840 | } |
| 6841 | |
| 6842 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6843 | _mm256_testn_epi64_mask (__m256i __A, __m256i __B) |
| 6844 | { |
| 6845 | return (__mmask8) __builtin_ia32_ptestnmq256 ((__v4di) __A, |
| 6846 | (__v4di) __B, |
| 6847 | (__mmask8) -1); |
| 6848 | } |
| 6849 | |
| 6850 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6851 | _mm256_mask_testn_epi64_mask (__mmask8 __U, __m256i __A, __m256i __B) |
| 6852 | { |
| 6853 | return (__mmask8) __builtin_ia32_ptestnmq256 ((__v4di) __A, |
| 6854 | (__v4di) __B, __U); |
| 6855 | } |
| 6856 | |
| 6857 | |
| 6858 | |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6859 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6860 | _mm_mask_unpackhi_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6861 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6862 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 6863 | (__v4si)_mm_unpackhi_epi32(__A, __B), |
| 6864 | (__v4si)__W); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6865 | } |
| 6866 | |
| 6867 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6868 | _mm_maskz_unpackhi_epi32(__mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6869 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6870 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 6871 | (__v4si)_mm_unpackhi_epi32(__A, __B), |
| 6872 | (__v4si)_mm_setzero_si128()); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6873 | } |
| 6874 | |
| 6875 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6876 | _mm256_mask_unpackhi_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6877 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6878 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 6879 | (__v8si)_mm256_unpackhi_epi32(__A, __B), |
| 6880 | (__v8si)__W); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6881 | } |
| 6882 | |
| 6883 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6884 | _mm256_maskz_unpackhi_epi32(__mmask8 __U, __m256i __A, __m256i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6885 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6886 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 6887 | (__v8si)_mm256_unpackhi_epi32(__A, __B), |
| 6888 | (__v8si)_mm256_setzero_si256()); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6889 | } |
| 6890 | |
| 6891 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6892 | _mm_mask_unpackhi_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6893 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6894 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 6895 | (__v2di)_mm_unpackhi_epi64(__A, __B), |
| 6896 | (__v2di)__W); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6897 | } |
| 6898 | |
| 6899 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6900 | _mm_maskz_unpackhi_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6901 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6902 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 6903 | (__v2di)_mm_unpackhi_epi64(__A, __B), |
| 6904 | (__v2di)_mm_setzero_di()); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6905 | } |
| 6906 | |
| 6907 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6908 | _mm256_mask_unpackhi_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6909 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6910 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 6911 | (__v4di)_mm256_unpackhi_epi64(__A, __B), |
| 6912 | (__v4di)__W); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6913 | } |
| 6914 | |
| 6915 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6916 | _mm256_maskz_unpackhi_epi64(__mmask8 __U, __m256i __A, __m256i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6917 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6918 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 6919 | (__v4di)_mm256_unpackhi_epi64(__A, __B), |
| 6920 | (__v4di)_mm256_setzero_si256()); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6921 | } |
| 6922 | |
| 6923 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6924 | _mm_mask_unpacklo_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6925 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6926 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 6927 | (__v4si)_mm_unpacklo_epi32(__A, __B), |
| 6928 | (__v4si)__W); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6929 | } |
| 6930 | |
| 6931 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6932 | _mm_maskz_unpacklo_epi32(__mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6933 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6934 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 6935 | (__v4si)_mm_unpacklo_epi32(__A, __B), |
| 6936 | (__v4si)_mm_setzero_si128()); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6937 | } |
| 6938 | |
| 6939 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6940 | _mm256_mask_unpacklo_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6941 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6942 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 6943 | (__v8si)_mm256_unpacklo_epi32(__A, __B), |
| 6944 | (__v8si)__W); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6945 | } |
| 6946 | |
| 6947 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6948 | _mm256_maskz_unpacklo_epi32(__mmask8 __U, __m256i __A, __m256i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6949 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6950 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 6951 | (__v8si)_mm256_unpacklo_epi32(__A, __B), |
| 6952 | (__v8si)_mm256_setzero_si256()); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6953 | } |
| 6954 | |
| 6955 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6956 | _mm_mask_unpacklo_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6957 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6958 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 6959 | (__v2di)_mm_unpacklo_epi64(__A, __B), |
| 6960 | (__v2di)__W); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6961 | } |
| 6962 | |
| 6963 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6964 | _mm_maskz_unpacklo_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6965 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6966 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 6967 | (__v2di)_mm_unpacklo_epi64(__A, __B), |
| 6968 | (__v2di)_mm_setzero_di()); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6969 | } |
| 6970 | |
| 6971 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6972 | _mm256_mask_unpacklo_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6973 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6974 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 6975 | (__v4di)_mm256_unpacklo_epi64(__A, __B), |
| 6976 | (__v4di)__W); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6977 | } |
| 6978 | |
| 6979 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6980 | _mm256_maskz_unpacklo_epi64(__mmask8 __U, __m256i __A, __m256i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6981 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6982 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 6983 | (__v4di)_mm256_unpacklo_epi64(__A, __B), |
| 6984 | (__v4di)_mm256_setzero_si256()); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6985 | } |
| 6986 | |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6987 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6988 | _mm_mask_sra_epi32 (__m128i __W, __mmask8 __U, __m128i __A, |
| 6989 | __m128i __B) |
| 6990 | { |
| 6991 | return (__m128i) __builtin_ia32_psrad128_mask ((__v4si) __A, |
| 6992 | (__v4si) __B, |
| 6993 | (__v4si) __W, |
| 6994 | (__mmask8) __U); |
| 6995 | } |
| 6996 | |
| 6997 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6998 | _mm_maskz_sra_epi32 (__mmask8 __U, __m128i __A, __m128i __B) |
| 6999 | { |
| 7000 | return (__m128i) __builtin_ia32_psrad128_mask ((__v4si) __A, |
| 7001 | (__v4si) __B, |
| 7002 | (__v4si) |
| 7003 | _mm_setzero_si128 (), |
| 7004 | (__mmask8) __U); |
| 7005 | } |
| 7006 | |
| 7007 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 7008 | _mm256_mask_sra_epi32 (__m256i __W, __mmask8 __U, __m256i __A, |
| 7009 | __m128i __B) |
| 7010 | { |
| 7011 | return (__m256i) __builtin_ia32_psrad256_mask ((__v8si) __A, |
| 7012 | (__v4si) __B, |
| 7013 | (__v8si) __W, |
| 7014 | (__mmask8) __U); |
| 7015 | } |
| 7016 | |
| 7017 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 7018 | _mm256_maskz_sra_epi32 (__mmask8 __U, __m256i __A, __m128i __B) |
| 7019 | { |
| 7020 | return (__m256i) __builtin_ia32_psrad256_mask ((__v8si) __A, |
| 7021 | (__v4si) __B, |
| 7022 | (__v8si) |
| 7023 | _mm256_setzero_si256 (), |
| 7024 | (__mmask8) __U); |
| 7025 | } |
| 7026 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7027 | #define _mm_mask_srai_epi32(W, U, A, imm) __extension__ ({ \ |
| 7028 | (__m128i)__builtin_ia32_psradi128_mask((__v4si)(__m128i)(A), (int)(imm), \ |
| 7029 | (__v4si)(__m128i)(W), \ |
| 7030 | (__mmask8)(U)); }) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 7031 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7032 | #define _mm_maskz_srai_epi32(U, A, imm) __extension__ ({ \ |
| 7033 | (__m128i)__builtin_ia32_psradi128_mask((__v4si)(__m128i)(A), (int)(imm), \ |
| 7034 | (__v4si)_mm_setzero_si128(), \ |
| 7035 | (__mmask8)(U)); }) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 7036 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7037 | #define _mm256_mask_srai_epi32(W, U, A, imm) __extension__ ({ \ |
| 7038 | (__m256i)__builtin_ia32_psradi256_mask((__v8si)(__m256i)(A), (int)(imm), \ |
| 7039 | (__v8si)(__m256i)(W), \ |
| 7040 | (__mmask8)(U)); }) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 7041 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7042 | #define _mm256_maskz_srai_epi32(U, A, imm) __extension__ ({ \ |
| 7043 | (__m256i)__builtin_ia32_psradi256_mask((__v8si)(__m256i)(A), (int)(imm), \ |
| 7044 | (__v8si)_mm256_setzero_si256(), \ |
| 7045 | (__mmask8)(U)); }) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 7046 | |
| 7047 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7048 | _mm_sra_epi64 (__m128i __A, __m128i __B) |
| 7049 | { |
| 7050 | return (__m128i) __builtin_ia32_psraq128_mask ((__v2di) __A, |
| 7051 | (__v2di) __B, |
| 7052 | (__v2di) |
| 7053 | _mm_setzero_di (), |
| 7054 | (__mmask8) -1); |
| 7055 | } |
| 7056 | |
| 7057 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7058 | _mm_mask_sra_epi64 (__m128i __W, __mmask8 __U, __m128i __A, |
| 7059 | __m128i __B) |
| 7060 | { |
| 7061 | return (__m128i) __builtin_ia32_psraq128_mask ((__v2di) __A, |
| 7062 | (__v2di) __B, |
| 7063 | (__v2di) __W, |
| 7064 | (__mmask8) __U); |
| 7065 | } |
| 7066 | |
| 7067 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7068 | _mm_maskz_sra_epi64 (__mmask8 __U, __m128i __A, __m128i __B) |
| 7069 | { |
| 7070 | return (__m128i) __builtin_ia32_psraq128_mask ((__v2di) __A, |
| 7071 | (__v2di) __B, |
| 7072 | (__v2di) |
| 7073 | _mm_setzero_di (), |
| 7074 | (__mmask8) __U); |
| 7075 | } |
| 7076 | |
| 7077 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 7078 | _mm256_sra_epi64 (__m256i __A, __m128i __B) |
| 7079 | { |
| 7080 | return (__m256i) __builtin_ia32_psraq256_mask ((__v4di) __A, |
| 7081 | (__v2di) __B, |
| 7082 | (__v4di) |
| 7083 | _mm256_setzero_si256 (), |
| 7084 | (__mmask8) -1); |
| 7085 | } |
| 7086 | |
| 7087 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 7088 | _mm256_mask_sra_epi64 (__m256i __W, __mmask8 __U, __m256i __A, |
| 7089 | __m128i __B) |
| 7090 | { |
| 7091 | return (__m256i) __builtin_ia32_psraq256_mask ((__v4di) __A, |
| 7092 | (__v2di) __B, |
| 7093 | (__v4di) __W, |
| 7094 | (__mmask8) __U); |
| 7095 | } |
| 7096 | |
| 7097 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 7098 | _mm256_maskz_sra_epi64 (__mmask8 __U, __m256i __A, __m128i __B) |
| 7099 | { |
| 7100 | return (__m256i) __builtin_ia32_psraq256_mask ((__v4di) __A, |
| 7101 | (__v2di) __B, |
| 7102 | (__v4di) |
| 7103 | _mm256_setzero_si256 (), |
| 7104 | (__mmask8) __U); |
| 7105 | } |
| 7106 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7107 | #define _mm_srai_epi64(A, imm) __extension__ ({ \ |
| 7108 | (__m128i)__builtin_ia32_psraqi128_mask((__v2di)(__m128i)(A), (int)(imm), \ |
| 7109 | (__v2di)_mm_setzero_di(), \ |
| 7110 | (__mmask8)-1); }) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 7111 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7112 | #define _mm_mask_srai_epi64(W, U, A, imm) __extension__ ({ \ |
| 7113 | (__m128i)__builtin_ia32_psraqi128_mask((__v2di)(__m128i)(A), (int)(imm), \ |
| 7114 | (__v2di)(__m128i)(W), \ |
| 7115 | (__mmask8)(U)); }) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 7116 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7117 | #define _mm_maskz_srai_epi64(U, A, imm) __extension__ ({ \ |
| 7118 | (__m128i)__builtin_ia32_psraqi128_mask((__v2di)(__m128i)(A), (int)(imm), \ |
| 7119 | (__v2di)_mm_setzero_si128(), \ |
| 7120 | (__mmask8)(U)); }) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 7121 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7122 | #define _mm256_srai_epi64(A, imm) __extension__ ({ \ |
| 7123 | (__m256i)__builtin_ia32_psraqi256_mask((__v4di)(__m256i)(A), (int)(imm), \ |
| 7124 | (__v4di)_mm256_setzero_si256(), \ |
| 7125 | (__mmask8)-1); }) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 7126 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7127 | #define _mm256_mask_srai_epi64(W, U, A, imm) __extension__ ({ \ |
| 7128 | (__m256i)__builtin_ia32_psraqi256_mask((__v4di)(__m256i)(A), (int)(imm), \ |
| 7129 | (__v4di)(__m256i)(W), \ |
| 7130 | (__mmask8)(U)); }) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 7131 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7132 | #define _mm256_maskz_srai_epi64(U, A, imm) __extension__ ({ \ |
| 7133 | (__m256i)__builtin_ia32_psraqi256_mask((__v4di)(__m256i)(A), (int)(imm), \ |
| 7134 | (__v4di)_mm256_setzero_si256(), \ |
| 7135 | (__mmask8)(U)); }) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 7136 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7137 | #define _mm_ternarylogic_epi32(A, B, C, imm) __extension__ ({ \ |
| 7138 | (__m128i)__builtin_ia32_pternlogd128_mask((__v4si)(__m128i)(A), \ |
| 7139 | (__v4si)(__m128i)(B), \ |
| 7140 | (__v4si)(__m128i)(C), (int)(imm), \ |
| 7141 | (__mmask8)-1); }) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 7142 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7143 | #define _mm_mask_ternarylogic_epi32(A, U, B, C, imm) __extension__ ({ \ |
| 7144 | (__m128i)__builtin_ia32_pternlogd128_mask((__v4si)(__m128i)(A), \ |
| 7145 | (__v4si)(__m128i)(B), \ |
| 7146 | (__v4si)(__m128i)(C), (int)(imm), \ |
| 7147 | (__mmask8)(U)); }) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 7148 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7149 | #define _mm_maskz_ternarylogic_epi32(U, A, B, C, imm) __extension__ ({ \ |
| 7150 | (__m128i)__builtin_ia32_pternlogd128_maskz((__v4si)(__m128i)(A), \ |
| 7151 | (__v4si)(__m128i)(B), \ |
| 7152 | (__v4si)(__m128i)(C), (int)(imm), \ |
| 7153 | (__mmask8)(U)); }) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 7154 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7155 | #define _mm256_ternarylogic_epi32(A, B, C, imm) __extension__ ({ \ |
| 7156 | (__m256i)__builtin_ia32_pternlogd256_mask((__v8si)(__m256i)(A), \ |
| 7157 | (__v8si)(__m256i)(B), \ |
| 7158 | (__v8si)(__m256i)(C), (int)(imm), \ |
| 7159 | (__mmask8)-1); }) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 7160 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7161 | #define _mm256_mask_ternarylogic_epi32(A, U, B, C, imm) __extension__ ({ \ |
| 7162 | (__m256i)__builtin_ia32_pternlogd256_mask((__v8si)(__m256i)(A), \ |
| 7163 | (__v8si)(__m256i)(B), \ |
| 7164 | (__v8si)(__m256i)(C), (int)(imm), \ |
| 7165 | (__mmask8)(U)); }) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 7166 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7167 | #define _mm256_maskz_ternarylogic_epi32(U, A, B, C, imm) __extension__ ({ \ |
| 7168 | (__m256i)__builtin_ia32_pternlogd256_maskz((__v8si)(__m256i)(A), \ |
| 7169 | (__v8si)(__m256i)(B), \ |
| 7170 | (__v8si)(__m256i)(C), (int)(imm), \ |
| 7171 | (__mmask8)(U)); }) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 7172 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7173 | #define _mm_ternarylogic_epi64(A, B, C, imm) __extension__ ({ \ |
| 7174 | (__m128i)__builtin_ia32_pternlogq128_mask((__v2di)(__m128i)(A), \ |
| 7175 | (__v2di)(__m128i)(B), \ |
| 7176 | (__v2di)(__m128i)(C), (int)(imm), \ |
| 7177 | (__mmask8)-1); }) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 7178 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7179 | #define _mm_mask_ternarylogic_epi64(A, U, B, C, imm) __extension__ ({ \ |
| 7180 | (__m128i)__builtin_ia32_pternlogq128_mask((__v2di)(__m128i)(A), \ |
| 7181 | (__v2di)(__m128i)(B), \ |
| 7182 | (__v2di)(__m128i)(C), (int)(imm), \ |
| 7183 | (__mmask8)(U)); }) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 7184 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7185 | #define _mm_maskz_ternarylogic_epi64(U, A, B, C, imm) __extension__ ({ \ |
| 7186 | (__m128i)__builtin_ia32_pternlogq128_maskz((__v2di)(__m128i)(A), \ |
| 7187 | (__v2di)(__m128i)(B), \ |
| 7188 | (__v2di)(__m128i)(C), (int)(imm), \ |
| 7189 | (__mmask8)(U)); }) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 7190 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7191 | #define _mm256_ternarylogic_epi64(A, B, C, imm) __extension__ ({ \ |
| 7192 | (__m256i)__builtin_ia32_pternlogq256_mask((__v4di)(__m256i)(A), \ |
| 7193 | (__v4di)(__m256i)(B), \ |
| 7194 | (__v4di)(__m256i)(C), (int)(imm), \ |
| 7195 | (__mmask8)-1); }) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 7196 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7197 | #define _mm256_mask_ternarylogic_epi64(A, U, B, C, imm) __extension__ ({ \ |
| 7198 | (__m256i)__builtin_ia32_pternlogq256_mask((__v4di)(__m256i)(A), \ |
| 7199 | (__v4di)(__m256i)(B), \ |
| 7200 | (__v4di)(__m256i)(C), (int)(imm), \ |
| 7201 | (__mmask8)(U)); }) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 7202 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7203 | #define _mm256_maskz_ternarylogic_epi64(U, A, B, C, imm) __extension__ ({ \ |
| 7204 | (__m256i)__builtin_ia32_pternlogq256_maskz((__v4di)(__m256i)(A), \ |
| 7205 | (__v4di)(__m256i)(B), \ |
| 7206 | (__v4di)(__m256i)(C), (int)(imm), \ |
| 7207 | (__mmask8)(U)); }) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 7208 | |
| 7209 | |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7210 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7211 | #define _mm256_shuffle_f32x4(A, B, imm) __extension__ ({ \ |
| 7212 | (__m256)__builtin_ia32_shuf_f32x4_256_mask((__v8sf)(__m256)(A), \ |
| 7213 | (__v8sf)(__m256)(B), (int)(imm), \ |
| 7214 | (__v8sf)_mm256_setzero_ps(), \ |
| 7215 | (__mmask8)-1); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7216 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7217 | #define _mm256_mask_shuffle_f32x4(W, U, A, B, imm) __extension__ ({ \ |
| 7218 | (__m256)__builtin_ia32_shuf_f32x4_256_mask((__v8sf)(__m256)(A), \ |
| 7219 | (__v8sf)(__m256)(B), (int)(imm), \ |
| 7220 | (__v8sf)(__m256)(W), \ |
| 7221 | (__mmask8)(U)); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7222 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7223 | #define _mm256_maskz_shuffle_f32x4(U, A, B, imm) __extension__ ({ \ |
| 7224 | (__m256)__builtin_ia32_shuf_f32x4_256_mask((__v8sf)(__m256)(A), \ |
| 7225 | (__v8sf)(__m256)(B), (int)(imm), \ |
| 7226 | (__v8sf)_mm256_setzero_ps(), \ |
| 7227 | (__mmask8)(U)); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7228 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7229 | #define _mm256_shuffle_f64x2(A, B, imm) __extension__ ({ \ |
| 7230 | (__m256d)__builtin_ia32_shuf_f64x2_256_mask((__v4df)(__m256d)(A), \ |
| 7231 | (__v4df)(__m256d)(B), \ |
| 7232 | (int)(imm), \ |
| 7233 | (__v4df)_mm256_setzero_pd(), \ |
| 7234 | (__mmask8)-1); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7235 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7236 | #define _mm256_mask_shuffle_f64x2(W, U, A, B, imm) __extension__ ({ \ |
| 7237 | (__m256d)__builtin_ia32_shuf_f64x2_256_mask((__v4df)(__m256d)(A), \ |
| 7238 | (__v4df)(__m256d)(B), \ |
| 7239 | (int)(imm), \ |
| 7240 | (__v4df)(__m256d)(W), \ |
| 7241 | (__mmask8)(U)); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7242 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7243 | #define _mm256_maskz_shuffle_f64x2(U, A, B, imm) __extension__ ({ \ |
| 7244 | (__m256d)__builtin_ia32_shuf_f64x2_256_mask((__v4df)(__m256d)(A), \ |
| 7245 | (__v4df)(__m256d)(B), \ |
| 7246 | (int)(imm), \ |
| 7247 | (__v4df)_mm256_setzero_pd(), \ |
| 7248 | (__mmask8)(U)); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7249 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7250 | #define _mm256_shuffle_i32x4(A, B, imm) __extension__ ({ \ |
| 7251 | (__m256i)__builtin_ia32_shuf_i32x4_256_mask((__v8si)(__m256i)(A), \ |
| 7252 | (__v8si)(__m256i)(B), \ |
| 7253 | (int)(imm), \ |
| 7254 | (__v8si)_mm256_setzero_si256(), \ |
| 7255 | (__mmask8)-1); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7256 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7257 | #define _mm256_mask_shuffle_i32x4(W, U, A, B, imm) __extension__ ({ \ |
| 7258 | (__m256i)__builtin_ia32_shuf_i32x4_256_mask((__v8si)(__m256i)(A), \ |
| 7259 | (__v8si)(__m256i)(B), \ |
| 7260 | (int)(imm), \ |
| 7261 | (__v8si)(__m256i)(W), \ |
| 7262 | (__mmask8)(U)); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7263 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7264 | #define _mm256_maskz_shuffle_i32x4(U, A, B, imm) __extension__ ({ \ |
| 7265 | (__m256i)__builtin_ia32_shuf_i32x4_256_mask((__v8si)(__m256i)(A), \ |
| 7266 | (__v8si)(__m256i)(B), \ |
| 7267 | (int)(imm), \ |
| 7268 | (__v8si)_mm256_setzero_si256(), \ |
| 7269 | (__mmask8)(U)); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7270 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7271 | #define _mm256_shuffle_i64x2(A, B, imm) __extension__ ({ \ |
| 7272 | (__m256i)__builtin_ia32_shuf_i64x2_256_mask((__v4di)(__m256i)(A), \ |
| 7273 | (__v4di)(__m256i)(B), \ |
| 7274 | (int)(imm), \ |
| 7275 | (__v4di)_mm256_setzero_si256(), \ |
| 7276 | (__mmask8)-1); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7277 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7278 | #define _mm256_mask_shuffle_i64x2(W, U, A, B, imm) __extension__ ({ \ |
| 7279 | (__m256i)__builtin_ia32_shuf_i64x2_256_mask((__v4di)(__m256i)(A), \ |
| 7280 | (__v4di)(__m256i)(B), \ |
| 7281 | (int)(imm), \ |
| 7282 | (__v4di)(__m256i)(W), \ |
| 7283 | (__mmask8)(U)); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7284 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7285 | #define _mm256_maskz_shuffle_i64x2(U, A, B, imm) __extension__ ({ \ |
| 7286 | (__m256i)__builtin_ia32_shuf_i64x2_256_mask((__v4di)(__m256i)(A), \ |
| 7287 | (__v4di)(__m256i)(B), \ |
| 7288 | (int)(imm), \ |
| 7289 | (__v4di)_mm256_setzero_si256(), \ |
| 7290 | (__mmask8)(U)); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7291 | |
Simon Pilgrim | 427154d | 2016-07-04 21:30:47 +0000 | [diff] [blame] | 7292 | #define _mm_mask_shuffle_pd(W, U, A, B, M) __extension__ ({ \ |
| 7293 | (__m128d)__builtin_ia32_selectpd_128((__mmask8)(U), \ |
| 7294 | (__v2df)_mm_shuffle_pd((A), (B), (M)), \ |
| 7295 | (__v2df)(__m128d)(W)); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7296 | |
Simon Pilgrim | 427154d | 2016-07-04 21:30:47 +0000 | [diff] [blame] | 7297 | #define _mm_maskz_shuffle_pd(U, A, B, M) __extension__ ({ \ |
| 7298 | (__m128d)__builtin_ia32_selectpd_128((__mmask8)(U), \ |
| 7299 | (__v2df)_mm_shuffle_pd((A), (B), (M)), \ |
| 7300 | (__v2df)_mm_setzero_pd()); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7301 | |
Simon Pilgrim | 427154d | 2016-07-04 21:30:47 +0000 | [diff] [blame] | 7302 | #define _mm256_mask_shuffle_pd(W, U, A, B, M) __extension__ ({ \ |
| 7303 | (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \ |
| 7304 | (__v4df)_mm256_shuffle_pd((A), (B), (M)), \ |
| 7305 | (__v4df)(__m256d)(W)); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7306 | |
Simon Pilgrim | 427154d | 2016-07-04 21:30:47 +0000 | [diff] [blame] | 7307 | #define _mm256_maskz_shuffle_pd(U, A, B, M) __extension__ ({ \ |
| 7308 | (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \ |
| 7309 | (__v4df)_mm256_shuffle_pd((A), (B), (M)), \ |
| 7310 | (__v4df)_mm256_setzero_pd()); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7311 | |
Simon Pilgrim | 427154d | 2016-07-04 21:30:47 +0000 | [diff] [blame] | 7312 | #define _mm_mask_shuffle_ps(W, U, A, B, M) __extension__ ({ \ |
| 7313 | (__m128)__builtin_ia32_selectps_128((__mmask8)(U), \ |
| 7314 | (__v4sf)_mm_shuffle_ps((A), (B), (M)), \ |
| 7315 | (__v4sf)(__m128)(W)); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7316 | |
Simon Pilgrim | 427154d | 2016-07-04 21:30:47 +0000 | [diff] [blame] | 7317 | #define _mm_maskz_shuffle_ps(U, A, B, M) __extension__ ({ \ |
| 7318 | (__m128)__builtin_ia32_selectps_128((__mmask8)(U), \ |
| 7319 | (__v4sf)_mm_shuffle_ps((A), (B), (M)), \ |
| 7320 | (__v4sf)_mm_setzero_ps()); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7321 | |
Simon Pilgrim | 427154d | 2016-07-04 21:30:47 +0000 | [diff] [blame] | 7322 | #define _mm256_mask_shuffle_ps(W, U, A, B, M) __extension__ ({ \ |
| 7323 | (__m256)__builtin_ia32_selectps_256((__mmask8)(U), \ |
| 7324 | (__v8sf)_mm256_shuffle_ps((A), (B), (M)), \ |
| 7325 | (__v8sf)(__m256)(W)); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7326 | |
Simon Pilgrim | 427154d | 2016-07-04 21:30:47 +0000 | [diff] [blame] | 7327 | #define _mm256_maskz_shuffle_ps(U, A, B, M) __extension__ ({ \ |
| 7328 | (__m256)__builtin_ia32_selectps_256((__mmask8)(U), \ |
| 7329 | (__v8sf)_mm256_shuffle_ps((A), (B), (M)), \ |
| 7330 | (__v8sf)_mm256_setzero_ps()); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7331 | |
| 7332 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 7333 | _mm_rsqrt14_pd (__m128d __A) |
| 7334 | { |
| 7335 | return (__m128d) __builtin_ia32_rsqrt14pd128_mask ((__v2df) __A, |
| 7336 | (__v2df) |
| 7337 | _mm_setzero_pd (), |
| 7338 | (__mmask8) -1); |
| 7339 | } |
| 7340 | |
| 7341 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 7342 | _mm_mask_rsqrt14_pd (__m128d __W, __mmask8 __U, __m128d __A) |
| 7343 | { |
| 7344 | return (__m128d) __builtin_ia32_rsqrt14pd128_mask ((__v2df) __A, |
| 7345 | (__v2df) __W, |
| 7346 | (__mmask8) __U); |
| 7347 | } |
| 7348 | |
| 7349 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 7350 | _mm_maskz_rsqrt14_pd (__mmask8 __U, __m128d __A) |
| 7351 | { |
| 7352 | return (__m128d) __builtin_ia32_rsqrt14pd128_mask ((__v2df) __A, |
| 7353 | (__v2df) |
| 7354 | _mm_setzero_pd (), |
| 7355 | (__mmask8) __U); |
| 7356 | } |
| 7357 | |
| 7358 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 7359 | _mm256_rsqrt14_pd (__m256d __A) |
| 7360 | { |
| 7361 | return (__m256d) __builtin_ia32_rsqrt14pd256_mask ((__v4df) __A, |
| 7362 | (__v4df) |
| 7363 | _mm256_setzero_pd (), |
| 7364 | (__mmask8) -1); |
| 7365 | } |
| 7366 | |
| 7367 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 7368 | _mm256_mask_rsqrt14_pd (__m256d __W, __mmask8 __U, __m256d __A) |
| 7369 | { |
| 7370 | return (__m256d) __builtin_ia32_rsqrt14pd256_mask ((__v4df) __A, |
| 7371 | (__v4df) __W, |
| 7372 | (__mmask8) __U); |
| 7373 | } |
| 7374 | |
| 7375 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 7376 | _mm256_maskz_rsqrt14_pd (__mmask8 __U, __m256d __A) |
| 7377 | { |
| 7378 | return (__m256d) __builtin_ia32_rsqrt14pd256_mask ((__v4df) __A, |
| 7379 | (__v4df) |
| 7380 | _mm256_setzero_pd (), |
| 7381 | (__mmask8) __U); |
| 7382 | } |
| 7383 | |
| 7384 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 7385 | _mm_rsqrt14_ps (__m128 __A) |
| 7386 | { |
| 7387 | return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A, |
| 7388 | (__v4sf) |
| 7389 | _mm_setzero_ps (), |
| 7390 | (__mmask8) -1); |
| 7391 | } |
| 7392 | |
| 7393 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 7394 | _mm_mask_rsqrt14_ps (__m128 __W, __mmask8 __U, __m128 __A) |
| 7395 | { |
| 7396 | return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A, |
| 7397 | (__v4sf) __W, |
| 7398 | (__mmask8) __U); |
| 7399 | } |
| 7400 | |
| 7401 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 7402 | _mm_maskz_rsqrt14_ps (__mmask8 __U, __m128 __A) |
| 7403 | { |
| 7404 | return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A, |
| 7405 | (__v4sf) |
| 7406 | _mm_setzero_ps (), |
| 7407 | (__mmask8) __U); |
| 7408 | } |
| 7409 | |
| 7410 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 7411 | _mm256_rsqrt14_ps (__m256 __A) |
| 7412 | { |
| 7413 | return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A, |
| 7414 | (__v8sf) |
| 7415 | _mm256_setzero_ps (), |
| 7416 | (__mmask8) -1); |
| 7417 | } |
| 7418 | |
| 7419 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 7420 | _mm256_mask_rsqrt14_ps (__m256 __W, __mmask8 __U, __m256 __A) |
| 7421 | { |
| 7422 | return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A, |
| 7423 | (__v8sf) __W, |
| 7424 | (__mmask8) __U); |
| 7425 | } |
| 7426 | |
| 7427 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 7428 | _mm256_maskz_rsqrt14_ps (__mmask8 __U, __m256 __A) |
| 7429 | { |
| 7430 | return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A, |
| 7431 | (__v8sf) |
| 7432 | _mm256_setzero_ps (), |
| 7433 | (__mmask8) __U); |
| 7434 | } |
| 7435 | |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7436 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 7437 | _mm256_broadcast_f32x4 (__m128 __A) |
| 7438 | { |
| 7439 | return (__m256) __builtin_ia32_broadcastf32x4_256_mask ((__v4sf) __A, |
| 7440 | (__v8sf)_mm256_undefined_pd (), |
Craig Topper | 406d5cd | 2016-06-04 05:43:37 +0000 | [diff] [blame] | 7441 | (__mmask8) -1); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7442 | } |
| 7443 | |
| 7444 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 7445 | _mm256_mask_broadcast_f32x4 (__m256 __O, __mmask8 __M, __m128 __A) |
| 7446 | { |
| 7447 | return (__m256) __builtin_ia32_broadcastf32x4_256_mask ((__v4sf) __A, |
| 7448 | (__v8sf) __O, |
| 7449 | __M); |
| 7450 | } |
| 7451 | |
| 7452 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 7453 | _mm256_maskz_broadcast_f32x4 (__mmask8 __M, __m128 __A) |
| 7454 | { |
| 7455 | return (__m256) __builtin_ia32_broadcastf32x4_256_mask ((__v4sf) __A, |
| 7456 | (__v8sf) _mm256_setzero_ps (), |
| 7457 | __M); |
| 7458 | } |
| 7459 | |
| 7460 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 7461 | _mm256_broadcast_i32x4 (__m128i __A) |
| 7462 | { |
| 7463 | return (__m256i) __builtin_ia32_broadcasti32x4_256_mask ((__v4si) __A, |
| 7464 | (__v8si)_mm256_undefined_si256 (), |
Craig Topper | 406d5cd | 2016-06-04 05:43:37 +0000 | [diff] [blame] | 7465 | (__mmask8) -1); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7466 | } |
| 7467 | |
| 7468 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 7469 | _mm256_mask_broadcast_i32x4 (__m256i __O, __mmask8 __M, __m128i __A) |
| 7470 | { |
| 7471 | return (__m256i) __builtin_ia32_broadcasti32x4_256_mask ((__v4si) __A, |
| 7472 | (__v8si) |
| 7473 | __O, __M); |
| 7474 | } |
| 7475 | |
| 7476 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 7477 | _mm256_maskz_broadcast_i32x4 (__mmask8 __M, __m128i __A) |
| 7478 | { |
| 7479 | return (__m256i) __builtin_ia32_broadcasti32x4_256_mask ((__v4si) |
| 7480 | __A, |
| 7481 | (__v8si) _mm256_setzero_si256 (), |
| 7482 | __M); |
| 7483 | } |
| 7484 | |
| 7485 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 7486 | _mm256_mask_broadcastsd_pd (__m256d __O, __mmask8 __M, __m128d __A) |
| 7487 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7488 | return (__m256d)__builtin_ia32_selectpd_256(__M, |
| 7489 | (__v4df) _mm256_broadcastsd_pd(__A), |
| 7490 | (__v4df) __O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7491 | } |
| 7492 | |
| 7493 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 7494 | _mm256_maskz_broadcastsd_pd (__mmask8 __M, __m128d __A) |
| 7495 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7496 | return (__m256d)__builtin_ia32_selectpd_256(__M, |
| 7497 | (__v4df) _mm256_broadcastsd_pd(__A), |
| 7498 | (__v4df) _mm256_setzero_pd()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7499 | } |
| 7500 | |
| 7501 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 7502 | _mm_mask_broadcastss_ps (__m128 __O, __mmask8 __M, __m128 __A) |
| 7503 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7504 | return (__m128)__builtin_ia32_selectps_128(__M, |
| 7505 | (__v4sf) _mm_broadcastss_ps(__A), |
| 7506 | (__v4sf) __O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7507 | } |
| 7508 | |
| 7509 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 7510 | _mm_maskz_broadcastss_ps (__mmask8 __M, __m128 __A) |
| 7511 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7512 | return (__m128)__builtin_ia32_selectps_128(__M, |
| 7513 | (__v4sf) _mm_broadcastss_ps(__A), |
| 7514 | (__v4sf) _mm_setzero_ps()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7515 | } |
| 7516 | |
| 7517 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 7518 | _mm256_mask_broadcastss_ps (__m256 __O, __mmask8 __M, __m128 __A) |
| 7519 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7520 | return (__m256)__builtin_ia32_selectps_256(__M, |
| 7521 | (__v8sf) _mm256_broadcastss_ps(__A), |
| 7522 | (__v8sf) __O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7523 | } |
| 7524 | |
| 7525 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 7526 | _mm256_maskz_broadcastss_ps (__mmask8 __M, __m128 __A) |
| 7527 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7528 | return (__m256)__builtin_ia32_selectps_256(__M, |
| 7529 | (__v8sf) _mm256_broadcastss_ps(__A), |
| 7530 | (__v8sf) _mm256_setzero_ps()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7531 | } |
| 7532 | |
| 7533 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7534 | _mm_mask_broadcastd_epi32 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7535 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7536 | return (__m128i)__builtin_ia32_selectd_128(__M, |
| 7537 | (__v4si) _mm_broadcastd_epi32(__A), |
| 7538 | (__v4si) __O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7539 | } |
| 7540 | |
| 7541 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7542 | _mm_maskz_broadcastd_epi32 (__mmask8 __M, __m128i __A) |
| 7543 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7544 | return (__m128i)__builtin_ia32_selectd_128(__M, |
| 7545 | (__v4si) _mm_broadcastd_epi32(__A), |
| 7546 | (__v4si) _mm_setzero_si128()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7547 | } |
| 7548 | |
| 7549 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 7550 | _mm256_mask_broadcastd_epi32 (__m256i __O, __mmask8 __M, __m128i __A) |
| 7551 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7552 | return (__m256i)__builtin_ia32_selectd_256(__M, |
| 7553 | (__v8si) _mm256_broadcastd_epi32(__A), |
| 7554 | (__v8si) __O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7555 | } |
| 7556 | |
| 7557 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 7558 | _mm256_maskz_broadcastd_epi32 (__mmask8 __M, __m128i __A) |
| 7559 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7560 | return (__m256i)__builtin_ia32_selectd_256(__M, |
| 7561 | (__v8si) _mm256_broadcastd_epi32(__A), |
| 7562 | (__v8si) _mm256_setzero_si256()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7563 | } |
| 7564 | |
| 7565 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7566 | _mm_mask_broadcastq_epi64 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7567 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7568 | return (__m128i)__builtin_ia32_selectq_128(__M, |
| 7569 | (__v2di) _mm_broadcastq_epi64(__A), |
| 7570 | (__v2di) __O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7571 | } |
| 7572 | |
| 7573 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7574 | _mm_maskz_broadcastq_epi64 (__mmask8 __M, __m128i __A) |
| 7575 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7576 | return (__m128i)__builtin_ia32_selectq_128(__M, |
| 7577 | (__v2di) _mm_broadcastq_epi64(__A), |
| 7578 | (__v2di) _mm_setzero_si128()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7579 | } |
| 7580 | |
| 7581 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 7582 | _mm256_mask_broadcastq_epi64 (__m256i __O, __mmask8 __M, __m128i __A) |
| 7583 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7584 | return (__m256i)__builtin_ia32_selectq_256(__M, |
| 7585 | (__v4di) _mm256_broadcastq_epi64(__A), |
| 7586 | (__v4di) __O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7587 | } |
| 7588 | |
| 7589 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 7590 | _mm256_maskz_broadcastq_epi64 (__mmask8 __M, __m128i __A) |
| 7591 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7592 | return (__m256i)__builtin_ia32_selectq_256(__M, |
| 7593 | (__v4di) _mm256_broadcastq_epi64(__A), |
| 7594 | (__v4di) _mm256_setzero_si256()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7595 | } |
| 7596 | |
Michael Zuckerman | e168061 | 2016-04-13 15:02:04 +0000 | [diff] [blame] | 7597 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7598 | _mm_cvtsepi32_epi8 (__m128i __A) |
| 7599 | { |
| 7600 | return (__m128i) __builtin_ia32_pmovsdb128_mask ((__v4si) __A, |
| 7601 | (__v16qi)_mm_undefined_si128(), |
| 7602 | (__mmask8) -1); |
| 7603 | } |
| 7604 | |
| 7605 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7606 | _mm_mask_cvtsepi32_epi8 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7607 | { |
| 7608 | return (__m128i) __builtin_ia32_pmovsdb128_mask ((__v4si) __A, |
| 7609 | (__v16qi) __O, __M); |
| 7610 | } |
| 7611 | |
| 7612 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7613 | _mm_maskz_cvtsepi32_epi8 (__mmask8 __M, __m128i __A) |
| 7614 | { |
| 7615 | return (__m128i) __builtin_ia32_pmovsdb128_mask ((__v4si) __A, |
| 7616 | (__v16qi) _mm_setzero_si128 (), |
| 7617 | __M); |
| 7618 | } |
| 7619 | |
| 7620 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7621 | _mm_mask_cvtsepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) |
| 7622 | { |
| 7623 | __builtin_ia32_pmovsdb128mem_mask ((__v16qi *) __P, (__v4si) __A, __M); |
| 7624 | } |
| 7625 | |
| 7626 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7627 | _mm256_cvtsepi32_epi8 (__m256i __A) |
| 7628 | { |
| 7629 | return (__m128i) __builtin_ia32_pmovsdb256_mask ((__v8si) __A, |
| 7630 | (__v16qi)_mm_undefined_si128(), |
| 7631 | (__mmask8) -1); |
| 7632 | } |
| 7633 | |
| 7634 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7635 | _mm256_mask_cvtsepi32_epi8 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7636 | { |
| 7637 | return (__m128i) __builtin_ia32_pmovsdb256_mask ((__v8si) __A, |
| 7638 | (__v16qi) __O, __M); |
| 7639 | } |
| 7640 | |
| 7641 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7642 | _mm256_maskz_cvtsepi32_epi8 (__mmask8 __M, __m256i __A) |
| 7643 | { |
| 7644 | return (__m128i) __builtin_ia32_pmovsdb256_mask ((__v8si) __A, |
| 7645 | (__v16qi) _mm_setzero_si128 (), |
| 7646 | __M); |
| 7647 | } |
| 7648 | |
| 7649 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7650 | _mm256_mask_cvtsepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) |
| 7651 | { |
| 7652 | __builtin_ia32_pmovsdb256mem_mask ((__v16qi *) __P, (__v8si) __A, __M); |
| 7653 | } |
| 7654 | |
| 7655 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7656 | _mm_cvtsepi32_epi16 (__m128i __A) |
| 7657 | { |
| 7658 | return (__m128i) __builtin_ia32_pmovsdw128_mask ((__v4si) __A, |
| 7659 | (__v8hi)_mm_setzero_si128 (), |
| 7660 | (__mmask8) -1); |
| 7661 | } |
| 7662 | |
| 7663 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7664 | _mm_mask_cvtsepi32_epi16 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7665 | { |
| 7666 | return (__m128i) __builtin_ia32_pmovsdw128_mask ((__v4si) __A, |
| 7667 | (__v8hi)__O, |
| 7668 | __M); |
| 7669 | } |
| 7670 | |
| 7671 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7672 | _mm_maskz_cvtsepi32_epi16 (__mmask8 __M, __m128i __A) |
| 7673 | { |
| 7674 | return (__m128i) __builtin_ia32_pmovsdw128_mask ((__v4si) __A, |
| 7675 | (__v8hi) _mm_setzero_si128 (), |
| 7676 | __M); |
| 7677 | } |
| 7678 | |
| 7679 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7680 | _mm_mask_cvtsepi32_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) |
| 7681 | { |
| 7682 | __builtin_ia32_pmovsdw128mem_mask ((__v8hi *) __P, (__v4si) __A, __M); |
| 7683 | } |
| 7684 | |
| 7685 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7686 | _mm256_cvtsepi32_epi16 (__m256i __A) |
| 7687 | { |
| 7688 | return (__m128i) __builtin_ia32_pmovsdw256_mask ((__v8si) __A, |
| 7689 | (__v8hi)_mm_undefined_si128(), |
| 7690 | (__mmask8) -1); |
| 7691 | } |
| 7692 | |
| 7693 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7694 | _mm256_mask_cvtsepi32_epi16 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7695 | { |
| 7696 | return (__m128i) __builtin_ia32_pmovsdw256_mask ((__v8si) __A, |
| 7697 | (__v8hi) __O, __M); |
| 7698 | } |
| 7699 | |
| 7700 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7701 | _mm256_maskz_cvtsepi32_epi16 (__mmask8 __M, __m256i __A) |
| 7702 | { |
| 7703 | return (__m128i) __builtin_ia32_pmovsdw256_mask ((__v8si) __A, |
| 7704 | (__v8hi) _mm_setzero_si128 (), |
| 7705 | __M); |
| 7706 | } |
| 7707 | |
| 7708 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7709 | _mm256_mask_cvtsepi32_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) |
| 7710 | { |
| 7711 | __builtin_ia32_pmovsdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M); |
| 7712 | } |
| 7713 | |
| 7714 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7715 | _mm_cvtsepi64_epi8 (__m128i __A) |
| 7716 | { |
| 7717 | return (__m128i) __builtin_ia32_pmovsqb128_mask ((__v2di) __A, |
| 7718 | (__v16qi)_mm_undefined_si128(), |
| 7719 | (__mmask8) -1); |
| 7720 | } |
| 7721 | |
| 7722 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7723 | _mm_mask_cvtsepi64_epi8 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7724 | { |
| 7725 | return (__m128i) __builtin_ia32_pmovsqb128_mask ((__v2di) __A, |
| 7726 | (__v16qi) __O, __M); |
| 7727 | } |
| 7728 | |
| 7729 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7730 | _mm_maskz_cvtsepi64_epi8 (__mmask8 __M, __m128i __A) |
| 7731 | { |
| 7732 | return (__m128i) __builtin_ia32_pmovsqb128_mask ((__v2di) __A, |
| 7733 | (__v16qi) _mm_setzero_si128 (), |
| 7734 | __M); |
| 7735 | } |
| 7736 | |
| 7737 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7738 | _mm_mask_cvtsepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) |
| 7739 | { |
| 7740 | __builtin_ia32_pmovsqb128mem_mask ((__v16qi *) __P, (__v2di) __A, __M); |
| 7741 | } |
| 7742 | |
| 7743 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7744 | _mm256_cvtsepi64_epi8 (__m256i __A) |
| 7745 | { |
| 7746 | return (__m128i) __builtin_ia32_pmovsqb256_mask ((__v4di) __A, |
| 7747 | (__v16qi)_mm_undefined_si128(), |
| 7748 | (__mmask8) -1); |
| 7749 | } |
| 7750 | |
| 7751 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7752 | _mm256_mask_cvtsepi64_epi8 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7753 | { |
| 7754 | return (__m128i) __builtin_ia32_pmovsqb256_mask ((__v4di) __A, |
| 7755 | (__v16qi) __O, __M); |
| 7756 | } |
| 7757 | |
| 7758 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7759 | _mm256_maskz_cvtsepi64_epi8 (__mmask8 __M, __m256i __A) |
| 7760 | { |
| 7761 | return (__m128i) __builtin_ia32_pmovsqb256_mask ((__v4di) __A, |
| 7762 | (__v16qi) _mm_setzero_si128 (), |
| 7763 | __M); |
| 7764 | } |
| 7765 | |
| 7766 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7767 | _mm256_mask_cvtsepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) |
| 7768 | { |
| 7769 | __builtin_ia32_pmovsqb256mem_mask ((__v16qi *) __P, (__v4di) __A, __M); |
| 7770 | } |
| 7771 | |
| 7772 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7773 | _mm_cvtsepi64_epi32 (__m128i __A) |
| 7774 | { |
| 7775 | return (__m128i) __builtin_ia32_pmovsqd128_mask ((__v2di) __A, |
| 7776 | (__v4si)_mm_undefined_si128(), |
| 7777 | (__mmask8) -1); |
| 7778 | } |
| 7779 | |
| 7780 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7781 | _mm_mask_cvtsepi64_epi32 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7782 | { |
| 7783 | return (__m128i) __builtin_ia32_pmovsqd128_mask ((__v2di) __A, |
| 7784 | (__v4si) __O, __M); |
| 7785 | } |
| 7786 | |
| 7787 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7788 | _mm_maskz_cvtsepi64_epi32 (__mmask8 __M, __m128i __A) |
| 7789 | { |
| 7790 | return (__m128i) __builtin_ia32_pmovsqd128_mask ((__v2di) __A, |
| 7791 | (__v4si) _mm_setzero_si128 (), |
| 7792 | __M); |
| 7793 | } |
| 7794 | |
| 7795 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7796 | _mm_mask_cvtsepi64_storeu_epi32 (void * __P, __mmask8 __M, __m128i __A) |
| 7797 | { |
| 7798 | __builtin_ia32_pmovsqd128mem_mask ((__v4si *) __P, (__v2di) __A, __M); |
| 7799 | } |
| 7800 | |
| 7801 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7802 | _mm256_cvtsepi64_epi32 (__m256i __A) |
| 7803 | { |
| 7804 | return (__m128i) __builtin_ia32_pmovsqd256_mask ((__v4di) __A, |
| 7805 | (__v4si)_mm_undefined_si128(), |
| 7806 | (__mmask8) -1); |
| 7807 | } |
| 7808 | |
| 7809 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7810 | _mm256_mask_cvtsepi64_epi32 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7811 | { |
| 7812 | return (__m128i) __builtin_ia32_pmovsqd256_mask ((__v4di) __A, |
| 7813 | (__v4si)__O, |
| 7814 | __M); |
| 7815 | } |
| 7816 | |
| 7817 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7818 | _mm256_maskz_cvtsepi64_epi32 (__mmask8 __M, __m256i __A) |
| 7819 | { |
| 7820 | return (__m128i) __builtin_ia32_pmovsqd256_mask ((__v4di) __A, |
| 7821 | (__v4si) _mm_setzero_si128 (), |
| 7822 | __M); |
| 7823 | } |
| 7824 | |
| 7825 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7826 | _mm256_mask_cvtsepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A) |
| 7827 | { |
| 7828 | __builtin_ia32_pmovsqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M); |
| 7829 | } |
| 7830 | |
| 7831 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7832 | _mm_cvtsepi64_epi16 (__m128i __A) |
| 7833 | { |
| 7834 | return (__m128i) __builtin_ia32_pmovsqw128_mask ((__v2di) __A, |
| 7835 | (__v8hi)_mm_undefined_si128(), |
| 7836 | (__mmask8) -1); |
| 7837 | } |
| 7838 | |
| 7839 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7840 | _mm_mask_cvtsepi64_epi16 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7841 | { |
| 7842 | return (__m128i) __builtin_ia32_pmovsqw128_mask ((__v2di) __A, |
| 7843 | (__v8hi) __O, __M); |
| 7844 | } |
| 7845 | |
| 7846 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7847 | _mm_maskz_cvtsepi64_epi16 (__mmask8 __M, __m128i __A) |
| 7848 | { |
| 7849 | return (__m128i) __builtin_ia32_pmovsqw128_mask ((__v2di) __A, |
| 7850 | (__v8hi) _mm_setzero_si128 (), |
| 7851 | __M); |
| 7852 | } |
| 7853 | |
| 7854 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7855 | _mm_mask_cvtsepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) |
| 7856 | { |
| 7857 | __builtin_ia32_pmovsqw128mem_mask ((__v8hi *) __P, (__v2di) __A, __M); |
| 7858 | } |
| 7859 | |
| 7860 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7861 | _mm256_cvtsepi64_epi16 (__m256i __A) |
| 7862 | { |
| 7863 | return (__m128i) __builtin_ia32_pmovsqw256_mask ((__v4di) __A, |
| 7864 | (__v8hi)_mm_undefined_si128(), |
| 7865 | (__mmask8) -1); |
| 7866 | } |
| 7867 | |
| 7868 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7869 | _mm256_mask_cvtsepi64_epi16 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7870 | { |
| 7871 | return (__m128i) __builtin_ia32_pmovsqw256_mask ((__v4di) __A, |
| 7872 | (__v8hi) __O, __M); |
| 7873 | } |
| 7874 | |
| 7875 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7876 | _mm256_maskz_cvtsepi64_epi16 (__mmask8 __M, __m256i __A) |
| 7877 | { |
| 7878 | return (__m128i) __builtin_ia32_pmovsqw256_mask ((__v4di) __A, |
| 7879 | (__v8hi) _mm_setzero_si128 (), |
| 7880 | __M); |
| 7881 | } |
| 7882 | |
| 7883 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7884 | _mm256_mask_cvtsepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) |
| 7885 | { |
| 7886 | __builtin_ia32_pmovsqw256mem_mask ((__v8hi *) __P, (__v4di) __A, __M); |
| 7887 | } |
| 7888 | |
Michael Zuckerman | d871531 | 2016-04-14 06:48:09 +0000 | [diff] [blame] | 7889 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7890 | _mm_cvtusepi32_epi8 (__m128i __A) |
| 7891 | { |
| 7892 | return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A, |
| 7893 | (__v16qi)_mm_undefined_si128(), |
| 7894 | (__mmask8) -1); |
| 7895 | } |
| 7896 | |
| 7897 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7898 | _mm_mask_cvtusepi32_epi8 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7899 | { |
| 7900 | return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A, |
| 7901 | (__v16qi) __O, |
| 7902 | __M); |
| 7903 | } |
| 7904 | |
| 7905 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7906 | _mm_maskz_cvtusepi32_epi8 (__mmask8 __M, __m128i __A) |
| 7907 | { |
| 7908 | return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A, |
| 7909 | (__v16qi) _mm_setzero_si128 (), |
| 7910 | __M); |
| 7911 | } |
| 7912 | |
| 7913 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7914 | _mm_mask_cvtusepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) |
| 7915 | { |
| 7916 | __builtin_ia32_pmovusdb128mem_mask ((__v16qi *) __P, (__v4si) __A, __M); |
| 7917 | } |
| 7918 | |
| 7919 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7920 | _mm256_cvtusepi32_epi8 (__m256i __A) |
| 7921 | { |
| 7922 | return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A, |
| 7923 | (__v16qi)_mm_undefined_si128(), |
| 7924 | (__mmask8) -1); |
| 7925 | } |
| 7926 | |
| 7927 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7928 | _mm256_mask_cvtusepi32_epi8 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7929 | { |
| 7930 | return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A, |
| 7931 | (__v16qi) __O, |
| 7932 | __M); |
| 7933 | } |
| 7934 | |
| 7935 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7936 | _mm256_maskz_cvtusepi32_epi8 (__mmask8 __M, __m256i __A) |
| 7937 | { |
| 7938 | return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A, |
| 7939 | (__v16qi) _mm_setzero_si128 (), |
| 7940 | __M); |
| 7941 | } |
| 7942 | |
| 7943 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7944 | _mm256_mask_cvtusepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) |
| 7945 | { |
| 7946 | __builtin_ia32_pmovusdb256mem_mask ((__v16qi*) __P, (__v8si) __A, __M); |
| 7947 | } |
| 7948 | |
| 7949 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7950 | _mm_cvtusepi32_epi16 (__m128i __A) |
| 7951 | { |
| 7952 | return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A, |
| 7953 | (__v8hi)_mm_undefined_si128(), |
| 7954 | (__mmask8) -1); |
| 7955 | } |
| 7956 | |
| 7957 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7958 | _mm_mask_cvtusepi32_epi16 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7959 | { |
| 7960 | return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A, |
| 7961 | (__v8hi) __O, __M); |
| 7962 | } |
| 7963 | |
| 7964 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7965 | _mm_maskz_cvtusepi32_epi16 (__mmask8 __M, __m128i __A) |
| 7966 | { |
| 7967 | return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A, |
| 7968 | (__v8hi) _mm_setzero_si128 (), |
| 7969 | __M); |
| 7970 | } |
| 7971 | |
| 7972 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7973 | _mm_mask_cvtusepi32_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) |
| 7974 | { |
| 7975 | __builtin_ia32_pmovusdw128mem_mask ((__v8hi *) __P, (__v4si) __A, __M); |
| 7976 | } |
| 7977 | |
| 7978 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7979 | _mm256_cvtusepi32_epi16 (__m256i __A) |
| 7980 | { |
| 7981 | return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A, |
| 7982 | (__v8hi) _mm_undefined_si128(), |
| 7983 | (__mmask8) -1); |
| 7984 | } |
| 7985 | |
| 7986 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7987 | _mm256_mask_cvtusepi32_epi16 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7988 | { |
| 7989 | return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A, |
| 7990 | (__v8hi) __O, __M); |
| 7991 | } |
| 7992 | |
| 7993 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7994 | _mm256_maskz_cvtusepi32_epi16 (__mmask8 __M, __m256i __A) |
| 7995 | { |
| 7996 | return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A, |
| 7997 | (__v8hi) _mm_setzero_si128 (), |
| 7998 | __M); |
| 7999 | } |
| 8000 | |
| 8001 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8002 | _mm256_mask_cvtusepi32_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) |
| 8003 | { |
| 8004 | __builtin_ia32_pmovusdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M); |
| 8005 | } |
| 8006 | |
| 8007 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8008 | _mm_cvtusepi64_epi8 (__m128i __A) |
| 8009 | { |
| 8010 | return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A, |
| 8011 | (__v16qi)_mm_undefined_si128(), |
| 8012 | (__mmask8) -1); |
| 8013 | } |
| 8014 | |
| 8015 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8016 | _mm_mask_cvtusepi64_epi8 (__m128i __O, __mmask8 __M, __m128i __A) |
| 8017 | { |
| 8018 | return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A, |
| 8019 | (__v16qi) __O, |
| 8020 | __M); |
| 8021 | } |
| 8022 | |
| 8023 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8024 | _mm_maskz_cvtusepi64_epi8 (__mmask8 __M, __m128i __A) |
| 8025 | { |
| 8026 | return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A, |
| 8027 | (__v16qi) _mm_setzero_si128 (), |
| 8028 | __M); |
| 8029 | } |
| 8030 | |
| 8031 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8032 | _mm_mask_cvtusepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) |
| 8033 | { |
| 8034 | __builtin_ia32_pmovusqb128mem_mask ((__v16qi *) __P, (__v2di) __A, __M); |
| 8035 | } |
| 8036 | |
| 8037 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8038 | _mm256_cvtusepi64_epi8 (__m256i __A) |
| 8039 | { |
| 8040 | return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A, |
| 8041 | (__v16qi)_mm_undefined_si128(), |
| 8042 | (__mmask8) -1); |
| 8043 | } |
| 8044 | |
| 8045 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8046 | _mm256_mask_cvtusepi64_epi8 (__m128i __O, __mmask8 __M, __m256i __A) |
| 8047 | { |
| 8048 | return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A, |
| 8049 | (__v16qi) __O, |
| 8050 | __M); |
| 8051 | } |
| 8052 | |
| 8053 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8054 | _mm256_maskz_cvtusepi64_epi8 (__mmask8 __M, __m256i __A) |
| 8055 | { |
| 8056 | return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A, |
| 8057 | (__v16qi) _mm_setzero_si128 (), |
| 8058 | __M); |
| 8059 | } |
| 8060 | |
| 8061 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8062 | _mm256_mask_cvtusepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) |
| 8063 | { |
| 8064 | __builtin_ia32_pmovusqb256mem_mask ((__v16qi *) __P, (__v4di) __A, __M); |
| 8065 | } |
| 8066 | |
| 8067 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8068 | _mm_cvtusepi64_epi32 (__m128i __A) |
| 8069 | { |
| 8070 | return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A, |
| 8071 | (__v4si)_mm_undefined_si128(), |
| 8072 | (__mmask8) -1); |
| 8073 | } |
| 8074 | |
| 8075 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8076 | _mm_mask_cvtusepi64_epi32 (__m128i __O, __mmask8 __M, __m128i __A) |
| 8077 | { |
| 8078 | return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A, |
| 8079 | (__v4si) __O, __M); |
| 8080 | } |
| 8081 | |
| 8082 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8083 | _mm_maskz_cvtusepi64_epi32 (__mmask8 __M, __m128i __A) |
| 8084 | { |
| 8085 | return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A, |
| 8086 | (__v4si) _mm_setzero_si128 (), |
| 8087 | __M); |
| 8088 | } |
| 8089 | |
| 8090 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8091 | _mm_mask_cvtusepi64_storeu_epi32 (void * __P, __mmask8 __M, __m128i __A) |
| 8092 | { |
| 8093 | __builtin_ia32_pmovusqd128mem_mask ((__v4si *) __P, (__v2di) __A, __M); |
| 8094 | } |
| 8095 | |
| 8096 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8097 | _mm256_cvtusepi64_epi32 (__m256i __A) |
| 8098 | { |
| 8099 | return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A, |
| 8100 | (__v4si)_mm_undefined_si128(), |
| 8101 | (__mmask8) -1); |
| 8102 | } |
| 8103 | |
| 8104 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8105 | _mm256_mask_cvtusepi64_epi32 (__m128i __O, __mmask8 __M, __m256i __A) |
| 8106 | { |
| 8107 | return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A, |
| 8108 | (__v4si) __O, __M); |
| 8109 | } |
| 8110 | |
| 8111 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8112 | _mm256_maskz_cvtusepi64_epi32 (__mmask8 __M, __m256i __A) |
| 8113 | { |
| 8114 | return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A, |
| 8115 | (__v4si) _mm_setzero_si128 (), |
| 8116 | __M); |
| 8117 | } |
| 8118 | |
| 8119 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8120 | _mm256_mask_cvtusepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A) |
| 8121 | { |
| 8122 | __builtin_ia32_pmovusqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M); |
| 8123 | } |
| 8124 | |
| 8125 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8126 | _mm_cvtusepi64_epi16 (__m128i __A) |
| 8127 | { |
| 8128 | return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A, |
| 8129 | (__v8hi)_mm_undefined_si128(), |
| 8130 | (__mmask8) -1); |
| 8131 | } |
| 8132 | |
| 8133 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8134 | _mm_mask_cvtusepi64_epi16 (__m128i __O, __mmask8 __M, __m128i __A) |
| 8135 | { |
| 8136 | return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A, |
| 8137 | (__v8hi) __O, __M); |
| 8138 | } |
| 8139 | |
| 8140 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8141 | _mm_maskz_cvtusepi64_epi16 (__mmask8 __M, __m128i __A) |
| 8142 | { |
| 8143 | return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A, |
| 8144 | (__v8hi) _mm_setzero_si128 (), |
| 8145 | __M); |
| 8146 | } |
| 8147 | |
| 8148 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8149 | _mm_mask_cvtusepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) |
| 8150 | { |
| 8151 | __builtin_ia32_pmovusqw128mem_mask ((__v8hi *) __P, (__v2di) __A, __M); |
| 8152 | } |
| 8153 | |
| 8154 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8155 | _mm256_cvtusepi64_epi16 (__m256i __A) |
| 8156 | { |
| 8157 | return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A, |
| 8158 | (__v8hi)_mm_undefined_si128(), |
| 8159 | (__mmask8) -1); |
| 8160 | } |
| 8161 | |
| 8162 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8163 | _mm256_mask_cvtusepi64_epi16 (__m128i __O, __mmask8 __M, __m256i __A) |
| 8164 | { |
| 8165 | return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A, |
| 8166 | (__v8hi) __O, __M); |
| 8167 | } |
| 8168 | |
| 8169 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8170 | _mm256_maskz_cvtusepi64_epi16 (__mmask8 __M, __m256i __A) |
| 8171 | { |
| 8172 | return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A, |
| 8173 | (__v8hi) _mm_setzero_si128 (), |
| 8174 | __M); |
| 8175 | } |
| 8176 | |
| 8177 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8178 | _mm256_mask_cvtusepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) |
| 8179 | { |
| 8180 | return __builtin_ia32_pmovusqw256mem_mask ((__v8hi *) __P, (__v4di) __A, __M); |
| 8181 | } |
| 8182 | |
Michael Zuckerman | 0a3508a | 2016-04-14 07:56:51 +0000 | [diff] [blame] | 8183 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8184 | _mm_cvtepi32_epi8 (__m128i __A) |
| 8185 | { |
| 8186 | return (__m128i) __builtin_ia32_pmovdb128_mask ((__v4si) __A, |
| 8187 | (__v16qi)_mm_undefined_si128(), |
| 8188 | (__mmask8) -1); |
| 8189 | } |
| 8190 | |
| 8191 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8192 | _mm_mask_cvtepi32_epi8 (__m128i __O, __mmask8 __M, __m128i __A) |
| 8193 | { |
| 8194 | return (__m128i) __builtin_ia32_pmovdb128_mask ((__v4si) __A, |
| 8195 | (__v16qi) __O, __M); |
| 8196 | } |
| 8197 | |
| 8198 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8199 | _mm_maskz_cvtepi32_epi8 (__mmask8 __M, __m128i __A) |
| 8200 | { |
| 8201 | return (__m128i) __builtin_ia32_pmovdb128_mask ((__v4si) __A, |
| 8202 | (__v16qi) |
| 8203 | _mm_setzero_si128 (), |
| 8204 | __M); |
| 8205 | } |
| 8206 | |
| 8207 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8208 | _mm_mask_cvtepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) |
| 8209 | { |
| 8210 | __builtin_ia32_pmovdb128mem_mask ((__v16qi *) __P, (__v4si) __A, __M); |
| 8211 | } |
| 8212 | |
| 8213 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8214 | _mm256_cvtepi32_epi8 (__m256i __A) |
| 8215 | { |
| 8216 | return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A, |
| 8217 | (__v16qi)_mm_undefined_si128(), |
| 8218 | (__mmask8) -1); |
| 8219 | } |
| 8220 | |
| 8221 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8222 | _mm256_mask_cvtepi32_epi8 (__m128i __O, __mmask8 __M, __m256i __A) |
| 8223 | { |
| 8224 | return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A, |
| 8225 | (__v16qi) __O, __M); |
| 8226 | } |
| 8227 | |
| 8228 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8229 | _mm256_maskz_cvtepi32_epi8 (__mmask8 __M, __m256i __A) |
| 8230 | { |
| 8231 | return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A, |
| 8232 | (__v16qi) _mm_setzero_si128 (), |
| 8233 | __M); |
| 8234 | } |
| 8235 | |
| 8236 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8237 | _mm256_mask_cvtepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) |
| 8238 | { |
| 8239 | __builtin_ia32_pmovdb256mem_mask ((__v16qi *) __P, (__v8si) __A, __M); |
| 8240 | } |
| 8241 | |
| 8242 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8243 | _mm_cvtepi32_epi16 (__m128i __A) |
| 8244 | { |
| 8245 | return (__m128i) __builtin_ia32_pmovdw128_mask ((__v4si) __A, |
| 8246 | (__v8hi) _mm_setzero_si128 (), |
| 8247 | (__mmask8) -1); |
| 8248 | } |
| 8249 | |
| 8250 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8251 | _mm_mask_cvtepi32_epi16 (__m128i __O, __mmask8 __M, __m128i __A) |
| 8252 | { |
| 8253 | return (__m128i) __builtin_ia32_pmovdw128_mask ((__v4si) __A, |
| 8254 | (__v8hi) __O, __M); |
| 8255 | } |
| 8256 | |
| 8257 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8258 | _mm_maskz_cvtepi32_epi16 (__mmask8 __M, __m128i __A) |
| 8259 | { |
| 8260 | return (__m128i) __builtin_ia32_pmovdw128_mask ((__v4si) __A, |
| 8261 | (__v8hi) _mm_setzero_si128 (), |
| 8262 | __M); |
| 8263 | } |
| 8264 | |
| 8265 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8266 | _mm_mask_cvtepi32_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) |
| 8267 | { |
| 8268 | __builtin_ia32_pmovdw128mem_mask ((__v8hi *) __P, (__v4si) __A, __M); |
| 8269 | } |
| 8270 | |
| 8271 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8272 | _mm256_cvtepi32_epi16 (__m256i __A) |
| 8273 | { |
| 8274 | return (__m128i) __builtin_ia32_pmovdw256_mask ((__v8si) __A, |
| 8275 | (__v8hi)_mm_setzero_si128 (), |
| 8276 | (__mmask8) -1); |
| 8277 | } |
| 8278 | |
| 8279 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8280 | _mm256_mask_cvtepi32_epi16 (__m128i __O, __mmask8 __M, __m256i __A) |
| 8281 | { |
| 8282 | return (__m128i) __builtin_ia32_pmovdw256_mask ((__v8si) __A, |
| 8283 | (__v8hi) __O, __M); |
| 8284 | } |
| 8285 | |
| 8286 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8287 | _mm256_maskz_cvtepi32_epi16 (__mmask8 __M, __m256i __A) |
| 8288 | { |
| 8289 | return (__m128i) __builtin_ia32_pmovdw256_mask ((__v8si) __A, |
| 8290 | (__v8hi) _mm_setzero_si128 (), |
| 8291 | __M); |
| 8292 | } |
| 8293 | |
| 8294 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8295 | _mm256_mask_cvtepi32_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) |
| 8296 | { |
| 8297 | __builtin_ia32_pmovdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M); |
| 8298 | } |
| 8299 | |
| 8300 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8301 | _mm_cvtepi64_epi8 (__m128i __A) |
| 8302 | { |
| 8303 | return (__m128i) __builtin_ia32_pmovqb128_mask ((__v2di) __A, |
| 8304 | (__v16qi) _mm_undefined_si128(), |
| 8305 | (__mmask8) -1); |
| 8306 | } |
| 8307 | |
| 8308 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8309 | _mm_mask_cvtepi64_epi8 (__m128i __O, __mmask8 __M, __m128i __A) |
| 8310 | { |
| 8311 | return (__m128i) __builtin_ia32_pmovqb128_mask ((__v2di) __A, |
| 8312 | (__v16qi) __O, __M); |
| 8313 | } |
| 8314 | |
| 8315 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8316 | _mm_maskz_cvtepi64_epi8 (__mmask8 __M, __m128i __A) |
| 8317 | { |
| 8318 | return (__m128i) __builtin_ia32_pmovqb128_mask ((__v2di) __A, |
| 8319 | (__v16qi) _mm_setzero_si128 (), |
| 8320 | __M); |
| 8321 | } |
| 8322 | |
| 8323 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8324 | _mm_mask_cvtepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) |
| 8325 | { |
| 8326 | __builtin_ia32_pmovqb128mem_mask ((__v16qi *) __P, (__v2di) __A, __M); |
| 8327 | } |
| 8328 | |
| 8329 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8330 | _mm256_cvtepi64_epi8 (__m256i __A) |
| 8331 | { |
| 8332 | return (__m128i) __builtin_ia32_pmovqb256_mask ((__v4di) __A, |
| 8333 | (__v16qi) _mm_undefined_si128(), |
| 8334 | (__mmask8) -1); |
| 8335 | } |
| 8336 | |
| 8337 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8338 | _mm256_mask_cvtepi64_epi8 (__m128i __O, __mmask8 __M, __m256i __A) |
| 8339 | { |
| 8340 | return (__m128i) __builtin_ia32_pmovqb256_mask ((__v4di) __A, |
| 8341 | (__v16qi) __O, __M); |
| 8342 | } |
| 8343 | |
| 8344 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8345 | _mm256_maskz_cvtepi64_epi8 (__mmask8 __M, __m256i __A) |
| 8346 | { |
| 8347 | return (__m128i) __builtin_ia32_pmovqb256_mask ((__v4di) __A, |
| 8348 | (__v16qi) _mm_setzero_si128 (), |
| 8349 | __M); |
| 8350 | } |
| 8351 | |
| 8352 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8353 | _mm256_mask_cvtepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) |
| 8354 | { |
| 8355 | __builtin_ia32_pmovqb256mem_mask ((__v16qi *) __P, (__v4di) __A, __M); |
| 8356 | } |
| 8357 | |
| 8358 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8359 | _mm_cvtepi64_epi32 (__m128i __A) |
| 8360 | { |
| 8361 | return (__m128i) __builtin_ia32_pmovqd128_mask ((__v2di) __A, |
| 8362 | (__v4si)_mm_undefined_si128(), |
| 8363 | (__mmask8) -1); |
| 8364 | } |
| 8365 | |
| 8366 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8367 | _mm_mask_cvtepi64_epi32 (__m128i __O, __mmask8 __M, __m128i __A) |
| 8368 | { |
| 8369 | return (__m128i) __builtin_ia32_pmovqd128_mask ((__v2di) __A, |
| 8370 | (__v4si) __O, __M); |
| 8371 | } |
| 8372 | |
| 8373 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8374 | _mm_maskz_cvtepi64_epi32 (__mmask8 __M, __m128i __A) |
| 8375 | { |
| 8376 | return (__m128i) __builtin_ia32_pmovqd128_mask ((__v2di) __A, |
| 8377 | (__v4si) _mm_setzero_si128 (), |
| 8378 | __M); |
| 8379 | } |
| 8380 | |
| 8381 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8382 | _mm_mask_cvtepi64_storeu_epi32 (void * __P, __mmask8 __M, __m128i __A) |
| 8383 | { |
| 8384 | __builtin_ia32_pmovqd128mem_mask ((__v4si *) __P, (__v2di) __A, __M); |
| 8385 | } |
| 8386 | |
| 8387 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8388 | _mm256_cvtepi64_epi32 (__m256i __A) |
| 8389 | { |
| 8390 | return (__m128i) __builtin_ia32_pmovqd256_mask ((__v4di) __A, |
| 8391 | (__v4si) _mm_undefined_si128(), |
| 8392 | (__mmask8) -1); |
| 8393 | } |
| 8394 | |
| 8395 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8396 | _mm256_mask_cvtepi64_epi32 (__m128i __O, __mmask8 __M, __m256i __A) |
| 8397 | { |
| 8398 | return (__m128i) __builtin_ia32_pmovqd256_mask ((__v4di) __A, |
| 8399 | (__v4si) __O, __M); |
| 8400 | } |
| 8401 | |
| 8402 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8403 | _mm256_maskz_cvtepi64_epi32 (__mmask8 __M, __m256i __A) |
| 8404 | { |
| 8405 | return (__m128i) __builtin_ia32_pmovqd256_mask ((__v4di) __A, |
| 8406 | (__v4si) _mm_setzero_si128 (), |
| 8407 | __M); |
| 8408 | } |
| 8409 | |
| 8410 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8411 | _mm256_mask_cvtepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A) |
| 8412 | { |
| 8413 | __builtin_ia32_pmovqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M); |
| 8414 | } |
| 8415 | |
| 8416 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8417 | _mm_cvtepi64_epi16 (__m128i __A) |
| 8418 | { |
| 8419 | return (__m128i) __builtin_ia32_pmovqw128_mask ((__v2di) __A, |
| 8420 | (__v8hi) _mm_undefined_si128(), |
| 8421 | (__mmask8) -1); |
| 8422 | } |
| 8423 | |
| 8424 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8425 | _mm_mask_cvtepi64_epi16 (__m128i __O, __mmask8 __M, __m128i __A) |
| 8426 | { |
| 8427 | return (__m128i) __builtin_ia32_pmovqw128_mask ((__v2di) __A, |
| 8428 | (__v8hi)__O, |
| 8429 | __M); |
| 8430 | } |
| 8431 | |
| 8432 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8433 | _mm_maskz_cvtepi64_epi16 (__mmask8 __M, __m128i __A) |
| 8434 | { |
| 8435 | return (__m128i) __builtin_ia32_pmovqw128_mask ((__v2di) __A, |
| 8436 | (__v8hi) _mm_setzero_si128 (), |
| 8437 | __M); |
| 8438 | } |
| 8439 | |
| 8440 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8441 | _mm_mask_cvtepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) |
| 8442 | { |
| 8443 | __builtin_ia32_pmovqw128mem_mask ((__v8hi *) __P, (__v2di) __A, __M); |
| 8444 | } |
| 8445 | |
| 8446 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8447 | _mm256_cvtepi64_epi16 (__m256i __A) |
| 8448 | { |
| 8449 | return (__m128i) __builtin_ia32_pmovqw256_mask ((__v4di) __A, |
| 8450 | (__v8hi)_mm_undefined_si128(), |
| 8451 | (__mmask8) -1); |
| 8452 | } |
| 8453 | |
| 8454 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8455 | _mm256_mask_cvtepi64_epi16 (__m128i __O, __mmask8 __M, __m256i __A) |
| 8456 | { |
| 8457 | return (__m128i) __builtin_ia32_pmovqw256_mask ((__v4di) __A, |
| 8458 | (__v8hi) __O, __M); |
| 8459 | } |
| 8460 | |
| 8461 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8462 | _mm256_maskz_cvtepi64_epi16 (__mmask8 __M, __m256i __A) |
| 8463 | { |
| 8464 | return (__m128i) __builtin_ia32_pmovqw256_mask ((__v4di) __A, |
| 8465 | (__v8hi) _mm_setzero_si128 (), |
| 8466 | __M); |
| 8467 | } |
| 8468 | |
| 8469 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8470 | _mm256_mask_cvtepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) |
| 8471 | { |
| 8472 | __builtin_ia32_pmovqw256mem_mask ((__v8hi *) __P, (__v4di) __A, __M); |
| 8473 | } |
| 8474 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8475 | #define _mm256_extractf32x4_ps(A, imm) __extension__ ({ \ |
| 8476 | (__m128)__builtin_ia32_extractf32x4_256_mask((__v8sf)(__m256)(A), \ |
| 8477 | (int)(imm), \ |
| 8478 | (__v4sf)_mm_setzero_ps(), \ |
| 8479 | (__mmask8)-1); }) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 8480 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8481 | #define _mm256_mask_extractf32x4_ps(W, U, A, imm) __extension__ ({ \ |
| 8482 | (__m128)__builtin_ia32_extractf32x4_256_mask((__v8sf)(__m256)(A), \ |
| 8483 | (int)(imm), \ |
| 8484 | (__v4sf)(__m128)(W), \ |
| 8485 | (__mmask8)(U)); }) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 8486 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8487 | #define _mm256_maskz_extractf32x4_ps(U, A, imm) __extension__ ({ \ |
| 8488 | (__m128)__builtin_ia32_extractf32x4_256_mask((__v8sf)(__m256)(A), \ |
| 8489 | (int)(imm), \ |
| 8490 | (__v4sf)_mm_setzero_ps(), \ |
| 8491 | (__mmask8)(U)); }) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 8492 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8493 | #define _mm256_extracti32x4_epi32(A, imm) __extension__ ({ \ |
| 8494 | (__m128i)__builtin_ia32_extracti32x4_256_mask((__v8si)(__m256i)(A), \ |
| 8495 | (int)(imm), \ |
| 8496 | (__v4si)_mm_setzero_si128(), \ |
| 8497 | (__mmask8)-1); }) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 8498 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8499 | #define _mm256_mask_extracti32x4_epi32(W, U, A, imm) __extension__ ({ \ |
| 8500 | (__m128i)__builtin_ia32_extracti32x4_256_mask((__v8si)(__m256i)(A), \ |
| 8501 | (int)(imm), \ |
| 8502 | (__v4si)(__m128i)(W), \ |
| 8503 | (__mmask8)(U)); }) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 8504 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8505 | #define _mm256_maskz_extracti32x4_epi32(U, A, imm) __extension__ ({ \ |
| 8506 | (__m128i)__builtin_ia32_extracti32x4_256_mask((__v8si)(__m256i)(A), \ |
| 8507 | (int)(imm), \ |
| 8508 | (__v4si)_mm_setzero_si128(), \ |
| 8509 | (__mmask8)(U)); }) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 8510 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8511 | #define _mm256_insertf32x4(A, B, imm) __extension__ ({ \ |
| 8512 | (__m256)__builtin_ia32_insertf32x4_256_mask((__v8sf)(__m256)(A), \ |
| 8513 | (__v4sf)(__m128)(B), (int)(imm), \ |
| 8514 | (__v8sf)_mm256_setzero_ps(), \ |
| 8515 | (__mmask8)-1); }) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 8516 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8517 | #define _mm256_mask_insertf32x4(W, U, A, B, imm) __extension__ ({ \ |
| 8518 | (__m256)__builtin_ia32_insertf32x4_256_mask((__v8sf)(__m256)(A), \ |
| 8519 | (__v4sf)(__m128)(B), (int)(imm), \ |
| 8520 | (__v8sf)(__m256)(W), \ |
| 8521 | (__mmask8)(U)); }) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 8522 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8523 | #define _mm256_maskz_insertf32x4(U, A, B, imm) __extension__ ({ \ |
| 8524 | (__m256)__builtin_ia32_insertf32x4_256_mask((__v8sf)(__m256)(A), \ |
| 8525 | (__v4sf)(__m128)(B), (int)(imm), \ |
| 8526 | (__v8sf)_mm256_setzero_ps(), \ |
| 8527 | (__mmask8)(U)); }) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 8528 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8529 | #define _mm256_inserti32x4(A, B, imm) __extension__ ({ \ |
| 8530 | (__m256i)__builtin_ia32_inserti32x4_256_mask((__v8si)(__m256i)(A), \ |
| 8531 | (__v4si)(__m128i)(B), \ |
| 8532 | (int)(imm), \ |
| 8533 | (__v8si)_mm256_setzero_si256(), \ |
| 8534 | (__mmask8)-1); }) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 8535 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8536 | #define _mm256_mask_inserti32x4(W, U, A, B, imm) __extension__ ({ \ |
| 8537 | (__m256i)__builtin_ia32_inserti32x4_256_mask((__v8si)(__m256i)(A), \ |
| 8538 | (__v4si)(__m128i)(B), \ |
| 8539 | (int)(imm), \ |
| 8540 | (__v8si)(__m256i)(W), \ |
| 8541 | (__mmask8)(U)); }) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 8542 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8543 | #define _mm256_maskz_inserti32x4(U, A, B, imm) __extension__ ({ \ |
| 8544 | (__m256i)__builtin_ia32_inserti32x4_256_mask((__v8si)(__m256i)(A), \ |
| 8545 | (__v4si)(__m128i)(B), \ |
| 8546 | (int)(imm), \ |
| 8547 | (__v8si)_mm256_setzero_si256(), \ |
| 8548 | (__mmask8)(U)); }) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 8549 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8550 | #define _mm_getmant_pd(A, B, C) __extension__({\ |
| 8551 | (__m128d)__builtin_ia32_getmantpd128_mask((__v2df)(__m128d)(A), \ |
| 8552 | (int)(((C)<<2) | (B)), \ |
| 8553 | (__v2df)_mm_setzero_pd(), \ |
| 8554 | (__mmask8)-1); }) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 8555 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8556 | #define _mm_mask_getmant_pd(W, U, A, B, C) __extension__({\ |
| 8557 | (__m128d)__builtin_ia32_getmantpd128_mask((__v2df)(__m128d)(A), \ |
| 8558 | (int)(((C)<<2) | (B)), \ |
| 8559 | (__v2df)(__m128d)(W), \ |
| 8560 | (__mmask8)(U)); }) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 8561 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8562 | #define _mm_maskz_getmant_pd(U, A, B, C) __extension__({\ |
| 8563 | (__m128d)__builtin_ia32_getmantpd128_mask((__v2df)(__m128d)(A), \ |
| 8564 | (int)(((C)<<2) | (B)), \ |
| 8565 | (__v2df)_mm_setzero_pd(), \ |
| 8566 | (__mmask8)(U)); }) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 8567 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8568 | #define _mm256_getmant_pd(A, B, C) __extension__ ({ \ |
| 8569 | (__m256d)__builtin_ia32_getmantpd256_mask((__v4df)(__m256d)(A), \ |
| 8570 | (int)(((C)<<2) | (B)), \ |
| 8571 | (__v4df)_mm256_setzero_pd(), \ |
| 8572 | (__mmask8)-1); }) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 8573 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8574 | #define _mm256_mask_getmant_pd(W, U, A, B, C) __extension__ ({ \ |
| 8575 | (__m256d)__builtin_ia32_getmantpd256_mask((__v4df)(__m256d)(A), \ |
| 8576 | (int)(((C)<<2) | (B)), \ |
| 8577 | (__v4df)(__m256d)(W), \ |
| 8578 | (__mmask8)(U)); }) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 8579 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8580 | #define _mm256_maskz_getmant_pd(U, A, B, C) __extension__ ({ \ |
| 8581 | (__m256d)__builtin_ia32_getmantpd256_mask((__v4df)(__m256d)(A), \ |
| 8582 | (int)(((C)<<2) | (B)), \ |
| 8583 | (__v4df)_mm256_setzero_pd(), \ |
| 8584 | (__mmask8)(U)); }) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 8585 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8586 | #define _mm_getmant_ps(A, B, C) __extension__ ({ \ |
| 8587 | (__m128)__builtin_ia32_getmantps128_mask((__v4sf)(__m128)(A), \ |
| 8588 | (int)(((C)<<2) | (B)), \ |
| 8589 | (__v4sf)_mm_setzero_ps(), \ |
| 8590 | (__mmask8)-1); }) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 8591 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8592 | #define _mm_mask_getmant_ps(W, U, A, B, C) __extension__ ({ \ |
| 8593 | (__m128)__builtin_ia32_getmantps128_mask((__v4sf)(__m128)(A), \ |
| 8594 | (int)(((C)<<2) | (B)), \ |
| 8595 | (__v4sf)(__m128)(W), \ |
| 8596 | (__mmask8)(U)); }) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 8597 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8598 | #define _mm_maskz_getmant_ps(U, A, B, C) __extension__ ({ \ |
| 8599 | (__m128)__builtin_ia32_getmantps128_mask((__v4sf)(__m128)(A), \ |
| 8600 | (int)(((C)<<2) | (B)), \ |
| 8601 | (__v4sf)_mm_setzero_ps(), \ |
| 8602 | (__mmask8)(U)); }) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 8603 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8604 | #define _mm256_getmant_ps(A, B, C) __extension__ ({ \ |
| 8605 | (__m256)__builtin_ia32_getmantps256_mask((__v8sf)(__m256)(A), \ |
| 8606 | (int)(((C)<<2) | (B)), \ |
| 8607 | (__v8sf)_mm256_setzero_ps(), \ |
| 8608 | (__mmask8)-1); }) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 8609 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8610 | #define _mm256_mask_getmant_ps(W, U, A, B, C) __extension__ ({ \ |
| 8611 | (__m256)__builtin_ia32_getmantps256_mask((__v8sf)(__m256)(A), \ |
| 8612 | (int)(((C)<<2) | (B)), \ |
| 8613 | (__v8sf)(__m256)(W), \ |
| 8614 | (__mmask8)(U)); }) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 8615 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8616 | #define _mm256_maskz_getmant_ps(U, A, B, C) __extension__ ({ \ |
| 8617 | (__m256)__builtin_ia32_getmantps256_mask((__v8sf)(__m256)(A), \ |
| 8618 | (int)(((C)<<2) | (B)), \ |
| 8619 | (__v8sf)_mm256_setzero_ps(), \ |
| 8620 | (__mmask8)(U)); }) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 8621 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8622 | #define _mm_mmask_i64gather_pd(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8623 | (__m128d)__builtin_ia32_gather3div2df((__v2df)(__m128d)(v1_old), \ |
| 8624 | (double const *)(addr), \ |
| 8625 | (__v2di)(__m128i)(index), \ |
| 8626 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8627 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8628 | #define _mm_mmask_i64gather_epi64(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8629 | (__m128i)__builtin_ia32_gather3div2di((__v2di)(__m128i)(v1_old), \ |
| 8630 | (long long const *)(addr), \ |
| 8631 | (__v2di)(__m128i)(index), \ |
| 8632 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8633 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8634 | #define _mm256_mmask_i64gather_pd(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8635 | (__m256d)__builtin_ia32_gather3div4df((__v4df)(__m256d)(v1_old), \ |
| 8636 | (double const *)(addr), \ |
| 8637 | (__v4di)(__m256i)(index), \ |
| 8638 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8639 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8640 | #define _mm256_mmask_i64gather_epi64(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8641 | (__m256i)__builtin_ia32_gather3div4di((__v4di)(__m256i)(v1_old), \ |
| 8642 | (long long const *)(addr), \ |
| 8643 | (__v4di)(__m256i)(index), \ |
| 8644 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8645 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8646 | #define _mm_mmask_i64gather_ps(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8647 | (__m128)__builtin_ia32_gather3div4sf((__v4sf)(__m128)(v1_old), \ |
| 8648 | (float const *)(addr), \ |
| 8649 | (__v2di)(__m128i)(index), \ |
| 8650 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8651 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8652 | #define _mm_mmask_i64gather_epi32(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8653 | (__m128i)__builtin_ia32_gather3div4si((__v4si)(__m128i)(v1_old), \ |
| 8654 | (int const *)(addr), \ |
| 8655 | (__v2di)(__m128i)(index), \ |
| 8656 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8657 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8658 | #define _mm256_mmask_i64gather_ps(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8659 | (__m128)__builtin_ia32_gather3div8sf((__v4sf)(__m128)(v1_old), \ |
| 8660 | (float const *)(addr), \ |
| 8661 | (__v4di)(__m256i)(index), \ |
| 8662 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8663 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8664 | #define _mm256_mmask_i64gather_epi32(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8665 | (__m128i)__builtin_ia32_gather3div8si((__v4si)(__m128i)(v1_old), \ |
| 8666 | (int const *)(addr), \ |
| 8667 | (__v4di)(__m256i)(index), \ |
| 8668 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8669 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8670 | #define _mm_mmask_i32gather_pd(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8671 | (__m128d)__builtin_ia32_gather3siv2df((__v2df)(__m128d)(v1_old), \ |
| 8672 | (double const *)(addr), \ |
| 8673 | (__v4si)(__m128i)(index), \ |
| 8674 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8675 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8676 | #define _mm_mmask_i32gather_epi64(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8677 | (__m128i)__builtin_ia32_gather3siv2di((__v2di)(__m128i)(v1_old), \ |
| 8678 | (long long const *)(addr), \ |
| 8679 | (__v4si)(__m128i)(index), \ |
| 8680 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8681 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8682 | #define _mm256_mmask_i32gather_pd(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8683 | (__m256d)__builtin_ia32_gather3siv4df((__v4df)(__m256d)(v1_old), \ |
| 8684 | (double const *)(addr), \ |
| 8685 | (__v4si)(__m128i)(index), \ |
| 8686 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8687 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8688 | #define _mm256_mmask_i32gather_epi64(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8689 | (__m256i)__builtin_ia32_gather3siv4di((__v4di)(__m256i)(v1_old), \ |
| 8690 | (long long const *)(addr), \ |
| 8691 | (__v4si)(__m128i)(index), \ |
| 8692 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8693 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8694 | #define _mm_mmask_i32gather_ps(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8695 | (__m128)__builtin_ia32_gather3siv4sf((__v4sf)(__m128)(v1_old), \ |
| 8696 | (float const *)(addr), \ |
| 8697 | (__v4si)(__m128i)(index), \ |
| 8698 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8699 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8700 | #define _mm_mmask_i32gather_epi32(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8701 | (__m128i)__builtin_ia32_gather3siv4si((__v4si)(__m128i)(v1_old), \ |
| 8702 | (int const *)(addr), \ |
| 8703 | (__v4si)(__m128i)(index), \ |
| 8704 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8705 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8706 | #define _mm256_mmask_i32gather_ps(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8707 | (__m256)__builtin_ia32_gather3siv8sf((__v8sf)(__m256)(v1_old), \ |
| 8708 | (float const *)(addr), \ |
| 8709 | (__v8si)(__m256i)(index), \ |
| 8710 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8711 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8712 | #define _mm256_mmask_i32gather_epi32(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8713 | (__m256i)__builtin_ia32_gather3siv8si((__v8si)(__m256i)(v1_old), \ |
| 8714 | (int const *)(addr), \ |
| 8715 | (__v8si)(__m256i)(index), \ |
| 8716 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8717 | |
Simon Pilgrim | 30db811 | 2016-07-04 13:34:44 +0000 | [diff] [blame] | 8718 | #define _mm256_permutex_pd(X, C) __extension__ ({ \ |
| 8719 | (__m256d)__builtin_shufflevector((__v4df)(__m256d)(X), \ |
| 8720 | (__v4df)_mm256_undefined_pd(), \ |
Craig Topper | 2a383c9 | 2016-07-04 22:18:01 +0000 | [diff] [blame] | 8721 | ((C) >> 0) & 0x3, ((C) >> 2) & 0x3, \ |
| 8722 | ((C) >> 4) & 0x3, ((C) >> 6) & 0x3); }) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 8723 | |
Simon Pilgrim | 30db811 | 2016-07-04 13:34:44 +0000 | [diff] [blame] | 8724 | #define _mm256_mask_permutex_pd(W, U, X, C) __extension__ ({ \ |
| 8725 | (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \ |
| 8726 | (__v4df)_mm256_permutex_pd((X), (C)), \ |
| 8727 | (__v4df)(__m256d)(W)); }) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 8728 | |
Simon Pilgrim | 30db811 | 2016-07-04 13:34:44 +0000 | [diff] [blame] | 8729 | #define _mm256_maskz_permutex_pd(U, X, C) __extension__ ({ \ |
| 8730 | (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \ |
| 8731 | (__v4df)_mm256_permutex_pd((X), (C)), \ |
| 8732 | (__v4df)_mm256_setzero_pd()); }) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 8733 | |
Simon Pilgrim | 30db811 | 2016-07-04 13:34:44 +0000 | [diff] [blame] | 8734 | #define _mm256_permutex_epi64(X, C) __extension__ ({ \ |
| 8735 | (__m256i)__builtin_shufflevector((__v4di)(__m256i)(X), \ |
| 8736 | (__v4di)_mm256_undefined_si256(), \ |
Craig Topper | 2a383c9 | 2016-07-04 22:18:01 +0000 | [diff] [blame] | 8737 | ((C) >> 0) & 0x3, ((C) >> 2) & 0x3, \ |
| 8738 | ((C) >> 4) & 0x3, ((C) >> 6) & 0x3); }) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 8739 | |
Simon Pilgrim | 30db811 | 2016-07-04 13:34:44 +0000 | [diff] [blame] | 8740 | #define _mm256_mask_permutex_epi64(W, U, X, C) __extension__ ({ \ |
| 8741 | (__m256i)__builtin_ia32_selectq_256((__mmask8)(U), \ |
| 8742 | (__v4di)_mm256_permutex_epi64((X), (C)), \ |
| 8743 | (__v4di)(__m256i)(W)); }) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 8744 | |
Simon Pilgrim | 30db811 | 2016-07-04 13:34:44 +0000 | [diff] [blame] | 8745 | #define _mm256_maskz_permutex_epi64(U, X, C) __extension__ ({ \ |
| 8746 | (__m256i)__builtin_ia32_selectq_256((__mmask8)(U), \ |
| 8747 | (__v4di)_mm256_permutex_epi64((X), (C)), \ |
| 8748 | (__v4di)_mm256_setzero_si256()); }) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 8749 | |
| 8750 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 8751 | _mm256_permutexvar_pd (__m256i __X, __m256d __Y) |
| 8752 | { |
| 8753 | return (__m256d) __builtin_ia32_permvardf256_mask ((__v4df) __Y, |
| 8754 | (__v4di) __X, |
| 8755 | (__v4df) _mm256_undefined_si256 (), |
| 8756 | (__mmask8) -1); |
| 8757 | } |
| 8758 | |
| 8759 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 8760 | _mm256_mask_permutexvar_pd (__m256d __W, __mmask8 __U, __m256i __X, |
| 8761 | __m256d __Y) |
| 8762 | { |
| 8763 | return (__m256d) __builtin_ia32_permvardf256_mask ((__v4df) __Y, |
| 8764 | (__v4di) __X, |
| 8765 | (__v4df) __W, |
| 8766 | (__mmask8) __U); |
| 8767 | } |
| 8768 | |
| 8769 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 8770 | _mm256_maskz_permutexvar_pd (__mmask8 __U, __m256i __X, __m256d __Y) |
| 8771 | { |
| 8772 | return (__m256d) __builtin_ia32_permvardf256_mask ((__v4df) __Y, |
| 8773 | (__v4di) __X, |
| 8774 | (__v4df) _mm256_setzero_pd (), |
| 8775 | (__mmask8) __U); |
| 8776 | } |
| 8777 | |
| 8778 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 8779 | _mm256_maskz_permutexvar_epi64 (__mmask8 __M, __m256i __X, __m256i __Y) |
| 8780 | { |
| 8781 | return (__m256i) __builtin_ia32_permvardi256_mask ((__v4di) __Y, |
| 8782 | (__v4di) __X, |
| 8783 | (__v4di) _mm256_setzero_si256 (), |
| 8784 | (__mmask8) __M); |
| 8785 | } |
| 8786 | |
| 8787 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Michael Zuckerman | a72b49ef | 2016-07-05 11:30:31 +0000 | [diff] [blame] | 8788 | _mm256_permutexvar_epi64 ( __m256i __X, __m256i __Y) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 8789 | { |
| 8790 | return (__m256i) __builtin_ia32_permvardi256_mask ((__v4di) __Y, |
| 8791 | (__v4di) __X, |
| 8792 | (__v4di) _mm256_undefined_si256 (), |
| 8793 | (__mmask8) -1); |
| 8794 | } |
| 8795 | |
| 8796 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 8797 | _mm256_mask_permutexvar_epi64 (__m256i __W, __mmask8 __M, __m256i __X, |
| 8798 | __m256i __Y) |
| 8799 | { |
| 8800 | return (__m256i) __builtin_ia32_permvardi256_mask ((__v4di) __Y, |
| 8801 | (__v4di) __X, |
| 8802 | (__v4di) __W, |
| 8803 | __M); |
| 8804 | } |
| 8805 | |
| 8806 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 8807 | _mm256_mask_permutexvar_ps (__m256 __W, __mmask8 __U, __m256i __X, |
| 8808 | __m256 __Y) |
| 8809 | { |
| 8810 | return (__m256) __builtin_ia32_permvarsf256_mask ((__v8sf) __Y, |
| 8811 | (__v8si) __X, |
| 8812 | (__v8sf) __W, |
| 8813 | (__mmask8) __U); |
| 8814 | } |
| 8815 | |
| 8816 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 8817 | _mm256_maskz_permutexvar_ps (__mmask8 __U, __m256i __X, __m256 __Y) |
| 8818 | { |
| 8819 | return (__m256) __builtin_ia32_permvarsf256_mask ((__v8sf) __Y, |
| 8820 | (__v8si) __X, |
| 8821 | (__v8sf) _mm256_setzero_ps (), |
| 8822 | (__mmask8) __U); |
| 8823 | } |
| 8824 | |
| 8825 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 8826 | _mm256_permutexvar_ps (__m256i __X, __m256 __Y) |
| 8827 | { |
| 8828 | return (__m256) __builtin_ia32_permvarsf256_mask ((__v8sf) __Y, |
| 8829 | (__v8si) __X, |
| 8830 | (__v8sf) _mm256_undefined_si256 (), |
| 8831 | (__mmask8) -1); |
| 8832 | } |
| 8833 | |
| 8834 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 8835 | _mm256_maskz_permutexvar_epi32 (__mmask8 __M, __m256i __X, __m256i __Y) |
| 8836 | { |
| 8837 | return (__m256i) __builtin_ia32_permvarsi256_mask ((__v8si) __Y, |
| 8838 | (__v8si) __X, |
| 8839 | (__v8si) _mm256_setzero_si256 (), |
| 8840 | __M); |
| 8841 | } |
| 8842 | |
| 8843 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 8844 | _mm256_mask_permutexvar_epi32 (__m256i __W, __mmask8 __M, __m256i __X, |
| 8845 | __m256i __Y) |
| 8846 | { |
| 8847 | return (__m256i) __builtin_ia32_permvarsi256_mask ((__v8si) __Y, |
| 8848 | (__v8si) __X, |
| 8849 | (__v8si) __W, |
| 8850 | (__mmask8) __M); |
| 8851 | } |
| 8852 | |
| 8853 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 8854 | _mm256_permutexvar_epi32 (__m256i __X, __m256i __Y) |
| 8855 | { |
| 8856 | return (__m256i) __builtin_ia32_permvarsi256_mask ((__v8si) __Y, |
| 8857 | (__v8si) __X, |
| 8858 | (__v8si) _mm256_undefined_si256(), |
| 8859 | (__mmask8) -1); |
| 8860 | } |
| 8861 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8862 | #define _mm_alignr_epi32(A, B, imm) __extension__ ({ \ |
| 8863 | (__m128i)__builtin_ia32_alignd128_mask((__v4si)(__m128i)(A), \ |
| 8864 | (__v4si)(__m128i)(B), (int)(imm), \ |
| 8865 | (__v4si)_mm_undefined_si128(), \ |
| 8866 | (__mmask8)-1); }) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8867 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8868 | #define _mm_mask_alignr_epi32(W, U, A, B, imm) __extension__ ({ \ |
| 8869 | (__m128i)__builtin_ia32_alignd128_mask((__v4si)(__m128i)(A), \ |
| 8870 | (__v4si)(__m128i)(B), (int)(imm), \ |
| 8871 | (__v4si)(__m128i)(W), \ |
| 8872 | (__mmask8)(U)); }) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8873 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8874 | #define _mm_maskz_alignr_epi32(U, A, B, imm) __extension__ ({ \ |
| 8875 | (__m128i)__builtin_ia32_alignd128_mask((__v4si)(__m128i)(A), \ |
| 8876 | (__v4si)(__m128i)(B), (int)(imm), \ |
| 8877 | (__v4si)_mm_setzero_si128(), \ |
| 8878 | (__mmask8)(U)); }) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8879 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8880 | #define _mm256_alignr_epi32(A, B, imm) __extension__ ({ \ |
| 8881 | (__m256i)__builtin_ia32_alignd256_mask((__v8si)(__m256i)(A), \ |
| 8882 | (__v8si)(__m256i)(B), (int)(imm), \ |
| 8883 | (__v8si)_mm256_undefined_si256(), \ |
| 8884 | (__mmask8)-1); }) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8885 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8886 | #define _mm256_mask_alignr_epi32(W, U, A, B, imm) __extension__ ({ \ |
| 8887 | (__m256i)__builtin_ia32_alignd256_mask((__v8si)(__m256i)(A), \ |
| 8888 | (__v8si)(__m256i)(B), (int)(imm), \ |
| 8889 | (__v8si)(__m256i)(W), \ |
| 8890 | (__mmask8)(U)); }) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8891 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8892 | #define _mm256_maskz_alignr_epi32(U, A, B, imm) __extension__ ({ \ |
| 8893 | (__m256i)__builtin_ia32_alignd256_mask((__v8si)(__m256i)(A), \ |
| 8894 | (__v8si)(__m256i)(B), (int)(imm), \ |
| 8895 | (__v8si)_mm256_setzero_si256(), \ |
| 8896 | (__mmask8)(U)); }) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8897 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8898 | #define _mm_alignr_epi64(A, B, imm) __extension__ ({ \ |
| 8899 | (__m128i)__builtin_ia32_alignq128_mask((__v2di)(__m128i)(A), \ |
| 8900 | (__v2di)(__m128i)(B), (int)(imm), \ |
| 8901 | (__v2di)_mm_setzero_di(), \ |
| 8902 | (__mmask8)-1); }) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8903 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8904 | #define _mm_mask_alignr_epi64(W, U, A, B, imm) __extension__ ({ \ |
| 8905 | (__m128i)__builtin_ia32_alignq128_mask((__v2di)(__m128i)(A), \ |
| 8906 | (__v2di)(__m128i)(B), (int)(imm), \ |
| 8907 | (__v2di)(__m128i)(W), \ |
| 8908 | (__mmask8)(U)); }) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8909 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8910 | #define _mm_maskz_alignr_epi64(U, A, B, imm) __extension__ ({ \ |
| 8911 | (__m128i)__builtin_ia32_alignq128_mask((__v2di)(__m128i)(A), \ |
| 8912 | (__v2di)(__m128i)(B), (int)(imm), \ |
| 8913 | (__v2di)_mm_setzero_di(), \ |
| 8914 | (__mmask8)(U)); }) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8915 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8916 | #define _mm256_alignr_epi64(A, B, imm) __extension__ ({ \ |
| 8917 | (__m256i)__builtin_ia32_alignq256_mask((__v4di)(__m256i)(A), \ |
| 8918 | (__v4di)(__m256i)(B), (int)(imm), \ |
| 8919 | (__v4di)_mm256_undefined_pd(), \ |
| 8920 | (__mmask8)-1); }) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8921 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8922 | #define _mm256_mask_alignr_epi64(W, U, A, B, imm) __extension__ ({ \ |
| 8923 | (__m256i)__builtin_ia32_alignq256_mask((__v4di)(__m256i)(A), \ |
| 8924 | (__v4di)(__m256i)(B), (int)(imm), \ |
| 8925 | (__v4di)(__m256i)(W), \ |
| 8926 | (__mmask8)(U)); }) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8927 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8928 | #define _mm256_maskz_alignr_epi64(U, A, B, imm) __extension__ ({ \ |
| 8929 | (__m256i)__builtin_ia32_alignq256_mask((__v4di)(__m256i)(A), \ |
| 8930 | (__v4di)(__m256i)(B), (int)(imm), \ |
| 8931 | (__v4di)_mm256_setzero_si256(), \ |
| 8932 | (__mmask8)(U)); }) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8933 | |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8934 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 8935 | _mm_mask_movehdup_ps (__m128 __W, __mmask8 __U, __m128 __A) |
| 8936 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 8937 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 8938 | (__v4sf)_mm_movehdup_ps(__A), |
| 8939 | (__v4sf)__W); |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8940 | } |
| 8941 | |
| 8942 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 8943 | _mm_maskz_movehdup_ps (__mmask8 __U, __m128 __A) |
| 8944 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 8945 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 8946 | (__v4sf)_mm_movehdup_ps(__A), |
| 8947 | (__v4sf)_mm_setzero_ps()); |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8948 | } |
| 8949 | |
| 8950 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 8951 | _mm256_mask_movehdup_ps (__m256 __W, __mmask8 __U, __m256 __A) |
| 8952 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 8953 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 8954 | (__v8sf)_mm256_movehdup_ps(__A), |
| 8955 | (__v8sf)__W); |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8956 | } |
| 8957 | |
| 8958 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 8959 | _mm256_maskz_movehdup_ps (__mmask8 __U, __m256 __A) |
| 8960 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 8961 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 8962 | (__v8sf)_mm256_movehdup_ps(__A), |
| 8963 | (__v8sf)_mm256_setzero_ps()); |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8964 | } |
| 8965 | |
| 8966 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 8967 | _mm_mask_moveldup_ps (__m128 __W, __mmask8 __U, __m128 __A) |
| 8968 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 8969 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 8970 | (__v4sf)_mm_moveldup_ps(__A), |
| 8971 | (__v4sf)__W); |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8972 | } |
| 8973 | |
| 8974 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 8975 | _mm_maskz_moveldup_ps (__mmask8 __U, __m128 __A) |
| 8976 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 8977 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 8978 | (__v4sf)_mm_moveldup_ps(__A), |
| 8979 | (__v4sf)_mm_setzero_ps()); |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8980 | } |
| 8981 | |
| 8982 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 8983 | _mm256_mask_moveldup_ps (__m256 __W, __mmask8 __U, __m256 __A) |
| 8984 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 8985 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 8986 | (__v8sf)_mm256_moveldup_ps(__A), |
| 8987 | (__v8sf)__W); |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8988 | } |
| 8989 | |
| 8990 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 8991 | _mm256_maskz_moveldup_ps (__mmask8 __U, __m256 __A) |
| 8992 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 8993 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 8994 | (__v8sf)_mm256_moveldup_ps(__A), |
| 8995 | (__v8sf)_mm256_setzero_ps()); |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8996 | } |
| 8997 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8998 | #define _mm256_mask_shuffle_epi32(W, U, A, I) __extension__({\ |
Craig Topper | 7cc9263 | 2016-06-11 12:50:19 +0000 | [diff] [blame] | 8999 | (__m256i)__builtin_ia32_selectd_256((__mmask8)(U), \ |
| 9000 | (__v8si)_mm256_shuffle_epi32((A), (I)), \ |
| 9001 | (__v8si)(__m256i)(W)); }) |
Michael Zuckerman | c62f27e | 2016-05-02 07:35:27 +0000 | [diff] [blame] | 9002 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 9003 | #define _mm256_maskz_shuffle_epi32(U, A, I) __extension__({\ |
Craig Topper | 7cc9263 | 2016-06-11 12:50:19 +0000 | [diff] [blame] | 9004 | (__m256i)__builtin_ia32_selectd_256((__mmask8)(U), \ |
| 9005 | (__v8si)_mm256_shuffle_epi32((A), (I)), \ |
| 9006 | (__v8si)_mm256_setzero_si256()); }) |
Michael Zuckerman | c62f27e | 2016-05-02 07:35:27 +0000 | [diff] [blame] | 9007 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 9008 | #define _mm_mask_shuffle_epi32(W, U, A, I) __extension__({\ |
Craig Topper | 7cc9263 | 2016-06-11 12:50:19 +0000 | [diff] [blame] | 9009 | (__m128i)__builtin_ia32_selectd_128((__mmask8)(U), \ |
| 9010 | (__v4si)_mm_shuffle_epi32((A), (I)), \ |
| 9011 | (__v4si)(__m128i)(W)); }) |
Michael Zuckerman | c62f27e | 2016-05-02 07:35:27 +0000 | [diff] [blame] | 9012 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 9013 | #define _mm_maskz_shuffle_epi32(U, A, I) __extension__({\ |
Craig Topper | 7cc9263 | 2016-06-11 12:50:19 +0000 | [diff] [blame] | 9014 | (__m128i)__builtin_ia32_selectd_128((__mmask8)(U), \ |
| 9015 | (__v4si)_mm_shuffle_epi32((A), (I)), \ |
| 9016 | (__v4si)_mm_setzero_si128()); }) |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 9017 | |
| 9018 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 9019 | _mm_mask_mov_pd (__m128d __W, __mmask8 __U, __m128d __A) |
| 9020 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 9021 | return (__m128d) __builtin_ia32_selectpd_128 ((__mmask8) __U, |
| 9022 | (__v2df) __A, |
| 9023 | (__v2df) __W); |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 9024 | } |
| 9025 | |
| 9026 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 9027 | _mm_maskz_mov_pd (__mmask8 __U, __m128d __A) |
| 9028 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 9029 | return (__m128d) __builtin_ia32_selectpd_128 ((__mmask8) __U, |
| 9030 | (__v2df) __A, |
| 9031 | (__v2df) _mm_setzero_pd ()); |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 9032 | } |
| 9033 | |
| 9034 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 9035 | _mm256_mask_mov_pd (__m256d __W, __mmask8 __U, __m256d __A) |
| 9036 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 9037 | return (__m256d) __builtin_ia32_selectpd_256 ((__mmask8) __U, |
| 9038 | (__v4df) __A, |
| 9039 | (__v4df) __W); |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 9040 | } |
| 9041 | |
| 9042 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 9043 | _mm256_maskz_mov_pd (__mmask8 __U, __m256d __A) |
| 9044 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 9045 | return (__m256d) __builtin_ia32_selectpd_256 ((__mmask8) __U, |
| 9046 | (__v4df) __A, |
| 9047 | (__v4df) _mm256_setzero_pd ()); |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 9048 | } |
| 9049 | |
| 9050 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 9051 | _mm_mask_mov_ps (__m128 __W, __mmask8 __U, __m128 __A) |
| 9052 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 9053 | return (__m128) __builtin_ia32_selectps_128 ((__mmask8) __U, |
| 9054 | (__v4sf) __A, |
| 9055 | (__v4sf) __W); |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 9056 | } |
| 9057 | |
| 9058 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 9059 | _mm_maskz_mov_ps (__mmask8 __U, __m128 __A) |
| 9060 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 9061 | return (__m128) __builtin_ia32_selectps_128 ((__mmask8) __U, |
| 9062 | (__v4sf) __A, |
| 9063 | (__v4sf) _mm_setzero_ps ()); |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 9064 | } |
| 9065 | |
| 9066 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 9067 | _mm256_mask_mov_ps (__m256 __W, __mmask8 __U, __m256 __A) |
| 9068 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 9069 | return (__m256) __builtin_ia32_selectps_256 ((__mmask8) __U, |
| 9070 | (__v8sf) __A, |
| 9071 | (__v8sf) __W); |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 9072 | } |
| 9073 | |
| 9074 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 9075 | _mm256_maskz_mov_ps (__mmask8 __U, __m256 __A) |
| 9076 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 9077 | return (__m256) __builtin_ia32_selectps_256 ((__mmask8) __U, |
| 9078 | (__v8sf) __A, |
| 9079 | (__v8sf) _mm256_setzero_ps ()); |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 9080 | } |
| 9081 | |
Michael Zuckerman | 8bfb777 | 2016-05-03 12:45:04 +0000 | [diff] [blame] | 9082 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 9083 | _mm_mask_cvtph_ps (__m128 __W, __mmask8 __U, __m128i __A) |
| 9084 | { |
| 9085 | return (__m128) __builtin_ia32_vcvtph2ps_mask ((__v8hi) __A, |
| 9086 | (__v4sf) __W, |
| 9087 | (__mmask8) __U); |
| 9088 | } |
| 9089 | |
| 9090 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 9091 | _mm_maskz_cvtph_ps (__mmask8 __U, __m128i __A) |
| 9092 | { |
| 9093 | return (__m128) __builtin_ia32_vcvtph2ps_mask ((__v8hi) __A, |
| 9094 | (__v4sf) |
| 9095 | _mm_setzero_ps (), |
| 9096 | (__mmask8) __U); |
| 9097 | } |
| 9098 | |
| 9099 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 9100 | _mm256_mask_cvtph_ps (__m256 __W, __mmask8 __U, __m128i __A) |
| 9101 | { |
| 9102 | return (__m256) __builtin_ia32_vcvtph2ps256_mask ((__v8hi) __A, |
| 9103 | (__v8sf) __W, |
| 9104 | (__mmask8) __U); |
| 9105 | } |
| 9106 | |
| 9107 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 9108 | _mm256_maskz_cvtph_ps (__mmask8 __U, __m128i __A) |
| 9109 | { |
| 9110 | return (__m256) __builtin_ia32_vcvtph2ps256_mask ((__v8hi) __A, |
| 9111 | (__v8sf) |
| 9112 | _mm256_setzero_ps (), |
| 9113 | (__mmask8) __U); |
| 9114 | } |
| 9115 | |
Michael Zuckerman | 9e7d0a9 | 2016-06-02 07:44:08 +0000 | [diff] [blame] | 9116 | static __inline __m128i __DEFAULT_FN_ATTRS |
| 9117 | _mm_mask_cvtps_ph (__m128i __W, __mmask8 __U, __m128 __A) |
| 9118 | { |
| 9119 | return (__m128i) __builtin_ia32_vcvtps2ph_mask ((__v4sf) __A, _MM_FROUND_CUR_DIRECTION, |
| 9120 | (__v8hi) __W, |
| 9121 | (__mmask8) __U); |
| 9122 | } |
| 9123 | |
| 9124 | static __inline __m128i __DEFAULT_FN_ATTRS |
| 9125 | _mm_maskz_cvtps_ph (__mmask8 __U, __m128 __A) |
| 9126 | { |
| 9127 | return (__m128i) __builtin_ia32_vcvtps2ph_mask ((__v4sf) __A, _MM_FROUND_CUR_DIRECTION, |
| 9128 | (__v8hi) _mm_setzero_si128 (), |
| 9129 | (__mmask8) __U); |
| 9130 | } |
| 9131 | |
| 9132 | #define _mm_mask_cvt_roundps_ph(W, U, A, I) __extension__ ({ \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 9133 | (__m128i)__builtin_ia32_vcvtps2ph_mask((__v4sf)(__m128)(A), (int)(I), \ |
| 9134 | (__v8hi)(__m128i)(W), \ |
| 9135 | (__mmask8)(U)); }) |
Michael Zuckerman | 8bfb777 | 2016-05-03 12:45:04 +0000 | [diff] [blame] | 9136 | |
Michael Zuckerman | 9e7d0a9 | 2016-06-02 07:44:08 +0000 | [diff] [blame] | 9137 | #define _mm_maskz_cvt_roundps_ph(U, A, I) __extension__ ({ \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 9138 | (__m128i)__builtin_ia32_vcvtps2ph_mask((__v4sf)(__m128)(A), (int)(I), \ |
| 9139 | (__v8hi)_mm_setzero_si128(), \ |
| 9140 | (__mmask8)(U)); }) |
Michael Zuckerman | 8bfb777 | 2016-05-03 12:45:04 +0000 | [diff] [blame] | 9141 | |
Michael Zuckerman | 9e7d0a9 | 2016-06-02 07:44:08 +0000 | [diff] [blame] | 9142 | static __inline __m128i __DEFAULT_FN_ATTRS |
| 9143 | _mm256_mask_cvtps_ph (__m128i __W, __mmask8 __U, __m256 __A) |
| 9144 | { |
| 9145 | return (__m128i) __builtin_ia32_vcvtps2ph256_mask ((__v8sf) __A, _MM_FROUND_CUR_DIRECTION, |
| 9146 | (__v8hi) __W, |
| 9147 | (__mmask8) __U); |
| 9148 | } |
| 9149 | |
| 9150 | static __inline __m128i __DEFAULT_FN_ATTRS |
| 9151 | _mm256_maskz_cvtps_ph ( __mmask8 __U, __m256 __A) |
| 9152 | { |
| 9153 | return (__m128i) __builtin_ia32_vcvtps2ph256_mask ((__v8sf) __A, _MM_FROUND_CUR_DIRECTION, |
| 9154 | (__v8hi) _mm_setzero_si128(), |
| 9155 | (__mmask8) __U); |
| 9156 | } |
| 9157 | #define _mm256_mask_cvt_roundps_ph(W, U, A, I) __extension__ ({ \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 9158 | (__m128i)__builtin_ia32_vcvtps2ph256_mask((__v8sf)(__m256)(A), (int)(I), \ |
| 9159 | (__v8hi)(__m128i)(W), \ |
| 9160 | (__mmask8)(U)); }) |
Michael Zuckerman | 8bfb777 | 2016-05-03 12:45:04 +0000 | [diff] [blame] | 9161 | |
Michael Zuckerman | 9e7d0a9 | 2016-06-02 07:44:08 +0000 | [diff] [blame] | 9162 | #define _mm256_maskz_cvt_roundps_ph(U, A, I) __extension__ ({ \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 9163 | (__m128i)__builtin_ia32_vcvtps2ph256_mask((__v8sf)(__m256)(A), (int)(I), \ |
| 9164 | (__v8hi)_mm_setzero_si128(), \ |
| 9165 | (__mmask8)(U)); }) |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 9166 | |
Michael Zuckerman | 9e7d0a9 | 2016-06-02 07:44:08 +0000 | [diff] [blame] | 9167 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 9168 | #undef __DEFAULT_FN_ATTRS |
Eric Christopher | 4d185168 | 2015-06-17 07:09:20 +0000 | [diff] [blame] | 9169 | |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 9170 | #endif /* __AVX512VLINTRIN_H */ |