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 |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 619 | _mm256_mask_add_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 620 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 621 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 622 | (__v8si)_mm256_add_epi32(__A, __B), |
| 623 | (__v8si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 624 | } |
| 625 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 626 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 627 | _mm256_maskz_add_epi32(__mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 628 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 629 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 630 | (__v8si)_mm256_add_epi32(__A, __B), |
| 631 | (__v8si)_mm256_setzero_si256()); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 632 | } |
| 633 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 634 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 635 | _mm256_mask_add_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 636 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 637 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 638 | (__v4di)_mm256_add_epi64(__A, __B), |
| 639 | (__v4di)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 640 | } |
| 641 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 642 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 643 | _mm256_maskz_add_epi64(__mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 644 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 645 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 646 | (__v4di)_mm256_add_epi64(__A, __B), |
| 647 | (__v4di)_mm256_setzero_si256()); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 648 | } |
| 649 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 650 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 651 | _mm256_mask_sub_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 652 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 653 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 654 | (__v8si)_mm256_sub_epi32(__A, __B), |
| 655 | (__v8si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 656 | } |
| 657 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 658 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 659 | _mm256_maskz_sub_epi32(__mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 660 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 661 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 662 | (__v8si)_mm256_sub_epi32(__A, __B), |
| 663 | (__v8si)_mm256_setzero_si256()); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 664 | } |
| 665 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 666 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 667 | _mm256_mask_sub_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 668 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 669 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 670 | (__v4di)_mm256_sub_epi64(__A, __B), |
| 671 | (__v4di)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 672 | } |
| 673 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 674 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 675 | _mm256_maskz_sub_epi64(__mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 676 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 677 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 678 | (__v4di)_mm256_sub_epi64(__A, __B), |
| 679 | (__v4di)_mm256_setzero_si256()); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 680 | } |
| 681 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 682 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 683 | _mm_mask_add_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 684 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 685 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 686 | (__v4si)_mm_add_epi32(__A, __B), |
| 687 | (__v4si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 688 | } |
| 689 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 690 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 691 | _mm_maskz_add_epi32(__mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 692 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 693 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 694 | (__v4si)_mm_add_epi32(__A, __B), |
| 695 | (__v4si)_mm_setzero_si128()); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 696 | } |
| 697 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 698 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 699 | _mm_mask_add_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 700 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 701 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 702 | (__v2di)_mm_add_epi64(__A, __B), |
| 703 | (__v2di)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 704 | } |
| 705 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 706 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 707 | _mm_maskz_add_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 708 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 709 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 710 | (__v2di)_mm_add_epi64(__A, __B), |
| 711 | (__v2di)_mm_setzero_si128()); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 712 | } |
| 713 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 714 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 715 | _mm_mask_sub_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 716 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 717 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 718 | (__v4si)_mm_sub_epi32(__A, __B), |
| 719 | (__v4si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 720 | } |
| 721 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 722 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 723 | _mm_maskz_sub_epi32(__mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 724 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 725 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 726 | (__v4si)_mm_sub_epi32(__A, __B), |
| 727 | (__v4si)_mm_setzero_si128()); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 728 | } |
| 729 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 730 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 731 | _mm_mask_sub_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 732 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 733 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 734 | (__v2di)_mm_sub_epi64(__A, __B), |
| 735 | (__v2di)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 736 | } |
| 737 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 738 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 739 | _mm_maskz_sub_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 740 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 741 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 742 | (__v2di)_mm_sub_epi64(__A, __B), |
| 743 | (__v2di)_mm_setzero_si128()); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 744 | } |
| 745 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 746 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 747 | _mm256_mask_mul_epi32 (__m256i __W, __mmask8 __M, __m256i __X, |
| 748 | __m256i __Y) |
| 749 | { |
| 750 | return (__m256i) __builtin_ia32_pmuldq256_mask ((__v8si) __X, |
| 751 | (__v8si) __Y, |
| 752 | (__v4di) __W, __M); |
| 753 | } |
| 754 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 755 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 756 | _mm256_maskz_mul_epi32 (__mmask8 __M, __m256i __X, __m256i __Y) |
| 757 | { |
| 758 | return (__m256i) __builtin_ia32_pmuldq256_mask ((__v8si) __X, |
| 759 | (__v8si) __Y, |
| 760 | (__v4di) |
| 761 | _mm256_setzero_si256 (), |
| 762 | __M); |
| 763 | } |
| 764 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 765 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 766 | _mm_mask_mul_epi32 (__m128i __W, __mmask8 __M, __m128i __X, |
| 767 | __m128i __Y) |
| 768 | { |
| 769 | return (__m128i) __builtin_ia32_pmuldq128_mask ((__v4si) __X, |
| 770 | (__v4si) __Y, |
| 771 | (__v2di) __W, __M); |
| 772 | } |
| 773 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 774 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 775 | _mm_maskz_mul_epi32 (__mmask8 __M, __m128i __X, __m128i __Y) |
| 776 | { |
| 777 | return (__m128i) __builtin_ia32_pmuldq128_mask ((__v4si) __X, |
| 778 | (__v4si) __Y, |
| 779 | (__v2di) |
| 780 | _mm_setzero_si128 (), |
| 781 | __M); |
| 782 | } |
| 783 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 784 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 785 | _mm256_mask_mul_epu32 (__m256i __W, __mmask8 __M, __m256i __X, |
| 786 | __m256i __Y) |
| 787 | { |
| 788 | return (__m256i) __builtin_ia32_pmuludq256_mask ((__v8si) __X, |
| 789 | (__v8si) __Y, |
| 790 | (__v4di) __W, __M); |
| 791 | } |
| 792 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 793 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 794 | _mm256_maskz_mul_epu32 (__mmask8 __M, __m256i __X, __m256i __Y) |
| 795 | { |
| 796 | return (__m256i) __builtin_ia32_pmuludq256_mask ((__v8si) __X, |
| 797 | (__v8si) __Y, |
| 798 | (__v4di) |
| 799 | _mm256_setzero_si256 (), |
| 800 | __M); |
| 801 | } |
| 802 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 803 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 804 | _mm_mask_mul_epu32 (__m128i __W, __mmask8 __M, __m128i __X, |
| 805 | __m128i __Y) |
| 806 | { |
| 807 | return (__m128i) __builtin_ia32_pmuludq128_mask ((__v4si) __X, |
| 808 | (__v4si) __Y, |
| 809 | (__v2di) __W, __M); |
| 810 | } |
| 811 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 812 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 813 | _mm_maskz_mul_epu32 (__mmask8 __M, __m128i __X, __m128i __Y) |
| 814 | { |
| 815 | return (__m128i) __builtin_ia32_pmuludq128_mask ((__v4si) __X, |
| 816 | (__v4si) __Y, |
| 817 | (__v2di) |
| 818 | _mm_setzero_si128 (), |
| 819 | __M); |
| 820 | } |
| 821 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 822 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | f43e4a1 | 2016-09-03 19:19:49 +0000 | [diff] [blame] | 823 | _mm256_maskz_mullo_epi32(__mmask8 __M, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 824 | { |
Craig Topper | f43e4a1 | 2016-09-03 19:19:49 +0000 | [diff] [blame] | 825 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M, |
| 826 | (__v8si)_mm256_mullo_epi32(__A, __B), |
| 827 | (__v8si)_mm256_setzero_si256()); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 828 | } |
| 829 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 830 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | f43e4a1 | 2016-09-03 19:19:49 +0000 | [diff] [blame] | 831 | _mm256_mask_mullo_epi32(__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 832 | { |
Craig Topper | f43e4a1 | 2016-09-03 19:19:49 +0000 | [diff] [blame] | 833 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M, |
| 834 | (__v8si)_mm256_mullo_epi32(__A, __B), |
| 835 | (__v8si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 836 | } |
| 837 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 838 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | f43e4a1 | 2016-09-03 19:19:49 +0000 | [diff] [blame] | 839 | _mm_maskz_mullo_epi32(__mmask8 __M, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 840 | { |
Craig Topper | f43e4a1 | 2016-09-03 19:19:49 +0000 | [diff] [blame] | 841 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M, |
| 842 | (__v4si)_mm_mullo_epi32(__A, __B), |
| 843 | (__v4si)_mm_setzero_si128()); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 844 | } |
| 845 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 846 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | f43e4a1 | 2016-09-03 19:19:49 +0000 | [diff] [blame] | 847 | _mm_mask_mullo_epi32(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 848 | { |
Craig Topper | f43e4a1 | 2016-09-03 19:19:49 +0000 | [diff] [blame] | 849 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M, |
| 850 | (__v4si)_mm_mullo_epi32(__A, __B), |
| 851 | (__v4si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 852 | } |
| 853 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 854 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 855 | _mm256_mask_and_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 856 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 857 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 858 | (__v8si)_mm256_and_si256(__A, __B), |
| 859 | (__v8si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 860 | } |
| 861 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 862 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 863 | _mm256_maskz_and_epi32(__mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 864 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 865 | return (__m256i)_mm256_mask_and_epi32(_mm256_setzero_si256(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 866 | } |
| 867 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 868 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 869 | _mm_mask_and_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 870 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 871 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 872 | (__v4si)_mm_and_si128(__A, __B), |
| 873 | (__v4si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 874 | } |
| 875 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 876 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 877 | _mm_maskz_and_epi32(__mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 878 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 879 | return (__m128i)_mm_mask_and_epi32(_mm_setzero_si128(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 880 | } |
| 881 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 882 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 883 | _mm256_mask_andnot_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 884 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 885 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 886 | (__v8si)_mm256_andnot_si256(__A, __B), |
| 887 | (__v8si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 888 | } |
| 889 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 890 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 891 | _mm256_maskz_andnot_epi32(__mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 892 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 893 | return (__m256i)_mm256_mask_andnot_epi32(_mm256_setzero_si256(), |
| 894 | __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 895 | } |
| 896 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 897 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 898 | _mm_mask_andnot_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 899 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 900 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 901 | (__v4si)_mm_andnot_si128(__A, __B), |
| 902 | (__v4si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 903 | } |
| 904 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 905 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 906 | _mm_maskz_andnot_epi32 (__mmask8 __U, __m128i __A, __m128i __B) |
| 907 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 908 | return (__m128i)_mm_mask_andnot_epi32(_mm_setzero_si128(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 909 | } |
| 910 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 911 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 912 | _mm256_mask_or_epi32 (__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 913 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 914 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 915 | (__v8si)_mm256_or_si256(__A, __B), |
| 916 | (__v8si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 917 | } |
| 918 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 919 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 920 | _mm256_maskz_or_epi32(__mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 921 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 922 | return (__m256i)_mm256_mask_or_epi32(_mm256_setzero_si256(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 923 | } |
| 924 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 925 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 926 | _mm_mask_or_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 927 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 928 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 929 | (__v4si)_mm_or_si128(__A, __B), |
| 930 | (__v4si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 931 | } |
| 932 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 933 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 934 | _mm_maskz_or_epi32(__mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 935 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 936 | return (__m128i)_mm_mask_or_epi32(_mm_setzero_si128(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 937 | } |
| 938 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 939 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 940 | _mm256_mask_xor_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 941 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 942 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 943 | (__v8si)_mm256_xor_si256(__A, __B), |
| 944 | (__v8si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 945 | } |
| 946 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 947 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 948 | _mm256_maskz_xor_epi32(__mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 949 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 950 | return (__m256i)_mm256_mask_xor_epi32(_mm256_setzero_si256(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 951 | } |
| 952 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 953 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 954 | _mm_mask_xor_epi32(__m128i __W, __mmask8 __U, __m128i __A, |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 955 | __m128i __B) |
| 956 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 957 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 958 | (__v4si)_mm_xor_si128(__A, __B), |
| 959 | (__v4si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 960 | } |
| 961 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 962 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 963 | _mm_maskz_xor_epi32(__mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 964 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 965 | return (__m128i)_mm_mask_xor_epi32(_mm_setzero_si128(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 966 | } |
| 967 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 968 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 969 | _mm256_mask_and_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 970 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 971 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 972 | (__v4di)_mm256_and_si256(__A, __B), |
| 973 | (__v4di)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 974 | } |
| 975 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 976 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 977 | _mm256_maskz_and_epi64(__mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 978 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 979 | return (__m256i)_mm256_mask_and_epi64(_mm256_setzero_si256(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 980 | } |
| 981 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 982 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 983 | _mm_mask_and_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
| 984 | { |
| 985 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 986 | (__v2di)_mm_and_si128(__A, __B), |
| 987 | (__v2di)__W); |
| 988 | } |
| 989 | |
| 990 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 991 | _mm_maskz_and_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
| 992 | { |
| 993 | return (__m128i)_mm_mask_and_epi64(_mm_setzero_si128(), __U, __A, __B); |
| 994 | } |
| 995 | |
| 996 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 997 | _mm256_mask_andnot_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
| 998 | { |
| 999 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 1000 | (__v4di)_mm256_andnot_si256(__A, __B), |
| 1001 | (__v4di)__W); |
| 1002 | } |
| 1003 | |
| 1004 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1005 | _mm256_maskz_andnot_epi64(__mmask8 __U, __m256i __A, __m256i __B) |
| 1006 | { |
| 1007 | return (__m256i)_mm256_mask_andnot_epi64(_mm256_setzero_si256(), |
| 1008 | __U, __A, __B); |
| 1009 | } |
| 1010 | |
| 1011 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1012 | _mm_mask_andnot_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
| 1013 | { |
| 1014 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 1015 | (__v2di)_mm_andnot_si128(__A, __B), |
| 1016 | (__v2di)__W); |
| 1017 | } |
| 1018 | |
| 1019 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1020 | _mm_maskz_andnot_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
| 1021 | { |
| 1022 | return (__m128i)_mm_mask_andnot_epi64(_mm_setzero_si128(), __U, __A, __B); |
| 1023 | } |
| 1024 | |
| 1025 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1026 | _mm256_mask_or_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
| 1027 | { |
| 1028 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 1029 | (__v4di)_mm256_or_si256(__A, __B), |
| 1030 | (__v4di)__W); |
| 1031 | } |
| 1032 | |
| 1033 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1034 | _mm256_maskz_or_epi64(__mmask8 __U, __m256i __A, __m256i __B) |
| 1035 | { |
| 1036 | return (__m256i)_mm256_mask_or_epi64(_mm256_setzero_si256(), __U, __A, __B); |
| 1037 | } |
| 1038 | |
| 1039 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1040 | _mm_mask_or_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
| 1041 | { |
| 1042 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 1043 | (__v2di)_mm_or_si128(__A, __B), |
| 1044 | (__v2di)__W); |
| 1045 | } |
| 1046 | |
| 1047 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1048 | _mm_maskz_or_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
| 1049 | { |
| 1050 | return (__m128i)_mm_mask_or_epi64(_mm_setzero_si128(), __U, __A, __B); |
| 1051 | } |
| 1052 | |
| 1053 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1054 | _mm256_mask_xor_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
| 1055 | { |
| 1056 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 1057 | (__v4di)_mm256_xor_si256(__A, __B), |
| 1058 | (__v4di)__W); |
| 1059 | } |
| 1060 | |
| 1061 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1062 | _mm256_maskz_xor_epi64(__mmask8 __U, __m256i __A, __m256i __B) |
| 1063 | { |
| 1064 | return (__m256i)_mm256_mask_xor_epi64(_mm256_setzero_si256(), __U, __A, __B); |
| 1065 | } |
| 1066 | |
| 1067 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1068 | _mm_mask_xor_epi64(__m128i __W, __mmask8 __U, __m128i __A, |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1069 | __m128i __B) |
| 1070 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 1071 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 1072 | (__v2di)_mm_xor_si128(__A, __B), |
| 1073 | (__v2di)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1074 | } |
| 1075 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1076 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 1077 | _mm_maskz_xor_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1078 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 1079 | return (__m128i)_mm_mask_xor_epi64(_mm_setzero_si128(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1080 | } |
| 1081 | |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1082 | #define _mm_cmp_epi32_mask(a, b, p) __extension__ ({ \ |
| 1083 | (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)(__m128i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1084 | (__v4si)(__m128i)(b), (int)(p), \ |
| 1085 | (__mmask8)-1); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1086 | |
| 1087 | #define _mm_mask_cmp_epi32_mask(m, a, b, p) __extension__ ({ \ |
| 1088 | (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)(__m128i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1089 | (__v4si)(__m128i)(b), (int)(p), \ |
| 1090 | (__mmask8)(m)); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1091 | |
| 1092 | #define _mm_cmp_epu32_mask(a, b, p) __extension__ ({ \ |
| 1093 | (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)(__m128i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1094 | (__v4si)(__m128i)(b), (int)(p), \ |
| 1095 | (__mmask8)-1); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1096 | |
| 1097 | #define _mm_mask_cmp_epu32_mask(m, a, b, p) __extension__ ({ \ |
| 1098 | (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)(__m128i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1099 | (__v4si)(__m128i)(b), (int)(p), \ |
| 1100 | (__mmask8)(m)); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1101 | |
| 1102 | #define _mm256_cmp_epi32_mask(a, b, p) __extension__ ({ \ |
| 1103 | (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)(__m256i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1104 | (__v8si)(__m256i)(b), (int)(p), \ |
| 1105 | (__mmask8)-1); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1106 | |
| 1107 | #define _mm256_mask_cmp_epi32_mask(m, a, b, p) __extension__ ({ \ |
| 1108 | (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)(__m256i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1109 | (__v8si)(__m256i)(b), (int)(p), \ |
| 1110 | (__mmask8)(m)); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1111 | |
| 1112 | #define _mm256_cmp_epu32_mask(a, b, p) __extension__ ({ \ |
| 1113 | (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)(__m256i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1114 | (__v8si)(__m256i)(b), (int)(p), \ |
| 1115 | (__mmask8)-1); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1116 | |
| 1117 | #define _mm256_mask_cmp_epu32_mask(m, a, b, p) __extension__ ({ \ |
| 1118 | (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)(__m256i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1119 | (__v8si)(__m256i)(b), (int)(p), \ |
| 1120 | (__mmask8)(m)); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1121 | |
| 1122 | #define _mm_cmp_epi64_mask(a, b, p) __extension__ ({ \ |
| 1123 | (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)(__m128i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1124 | (__v2di)(__m128i)(b), (int)(p), \ |
| 1125 | (__mmask8)-1); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1126 | |
| 1127 | #define _mm_mask_cmp_epi64_mask(m, a, b, p) __extension__ ({ \ |
| 1128 | (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)(__m128i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1129 | (__v2di)(__m128i)(b), (int)(p), \ |
| 1130 | (__mmask8)(m)); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1131 | |
| 1132 | #define _mm_cmp_epu64_mask(a, b, p) __extension__ ({ \ |
| 1133 | (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)(__m128i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1134 | (__v2di)(__m128i)(b), (int)(p), \ |
| 1135 | (__mmask8)-1); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1136 | |
| 1137 | #define _mm_mask_cmp_epu64_mask(m, a, b, p) __extension__ ({ \ |
| 1138 | (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)(__m128i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1139 | (__v2di)(__m128i)(b), (int)(p), \ |
| 1140 | (__mmask8)(m)); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1141 | |
| 1142 | #define _mm256_cmp_epi64_mask(a, b, p) __extension__ ({ \ |
| 1143 | (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)(__m256i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1144 | (__v4di)(__m256i)(b), (int)(p), \ |
| 1145 | (__mmask8)-1); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1146 | |
| 1147 | #define _mm256_mask_cmp_epi64_mask(m, a, b, p) __extension__ ({ \ |
| 1148 | (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)(__m256i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1149 | (__v4di)(__m256i)(b), (int)(p), \ |
| 1150 | (__mmask8)(m)); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1151 | |
| 1152 | #define _mm256_cmp_epu64_mask(a, b, p) __extension__ ({ \ |
| 1153 | (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)(__m256i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1154 | (__v4di)(__m256i)(b), (int)(p), \ |
| 1155 | (__mmask8)-1); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1156 | |
| 1157 | #define _mm256_mask_cmp_epu64_mask(m, a, b, p) __extension__ ({ \ |
| 1158 | (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)(__m256i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1159 | (__v4di)(__m256i)(b), (int)(p), \ |
| 1160 | (__mmask8)(m)); }) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1161 | |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 1162 | #define _mm256_cmp_ps_mask(a, b, p) __extension__ ({ \ |
| 1163 | (__mmask8)__builtin_ia32_cmpps256_mask((__v8sf)(__m256)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1164 | (__v8sf)(__m256)(b), (int)(p), \ |
| 1165 | (__mmask8)-1); }) |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 1166 | |
| 1167 | #define _mm256_mask_cmp_ps_mask(m, a, b, p) __extension__ ({ \ |
| 1168 | (__mmask8)__builtin_ia32_cmpps256_mask((__v8sf)(__m256)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1169 | (__v8sf)(__m256)(b), (int)(p), \ |
| 1170 | (__mmask8)(m)); }) |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 1171 | |
| 1172 | #define _mm256_cmp_pd_mask(a, b, p) __extension__ ({ \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1173 | (__mmask8)__builtin_ia32_cmppd256_mask((__v4df)(__m256d)(a), \ |
| 1174 | (__v4df)(__m256d)(b), (int)(p), \ |
| 1175 | (__mmask8)-1); }) |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 1176 | |
| 1177 | #define _mm256_mask_cmp_pd_mask(m, a, b, p) __extension__ ({ \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1178 | (__mmask8)__builtin_ia32_cmppd256_mask((__v4df)(__m256d)(a), \ |
| 1179 | (__v4df)(__m256d)(b), (int)(p), \ |
| 1180 | (__mmask8)(m)); }) |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 1181 | |
Michael Zuckerman | eb5f178 | 2016-05-26 08:10:12 +0000 | [diff] [blame] | 1182 | #define _mm_cmp_ps_mask(a, b, p) __extension__ ({ \ |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 1183 | (__mmask8)__builtin_ia32_cmpps128_mask((__v4sf)(__m128)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1184 | (__v4sf)(__m128)(b), (int)(p), \ |
| 1185 | (__mmask8)-1); }) |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 1186 | |
Michael Zuckerman | eb5f178 | 2016-05-26 08:10:12 +0000 | [diff] [blame] | 1187 | #define _mm_mask_cmp_ps_mask(m, a, b, p) __extension__ ({ \ |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 1188 | (__mmask8)__builtin_ia32_cmpps128_mask((__v4sf)(__m128)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1189 | (__v4sf)(__m128)(b), (int)(p), \ |
| 1190 | (__mmask8)(m)); }) |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 1191 | |
Michael Zuckerman | eb5f178 | 2016-05-26 08:10:12 +0000 | [diff] [blame] | 1192 | #define _mm_cmp_pd_mask(a, b, p) __extension__ ({ \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1193 | (__mmask8)__builtin_ia32_cmppd128_mask((__v2df)(__m128d)(a), \ |
| 1194 | (__v2df)(__m128d)(b), (int)(p), \ |
| 1195 | (__mmask8)-1); }) |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 1196 | |
Michael Zuckerman | eb5f178 | 2016-05-26 08:10:12 +0000 | [diff] [blame] | 1197 | #define _mm_mask_cmp_pd_mask(m, a, b, p) __extension__ ({ \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 1198 | (__mmask8)__builtin_ia32_cmppd128_mask((__v2df)(__m128d)(a), \ |
| 1199 | (__v2df)(__m128d)(b), (int)(p), \ |
| 1200 | (__mmask8)(m)); }) |
Eric Christopher | 4d185168 | 2015-06-17 07:09:20 +0000 | [diff] [blame] | 1201 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1202 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1203 | _mm_mask_fmadd_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) |
| 1204 | { |
| 1205 | return (__m128d) __builtin_ia32_vfmaddpd128_mask ((__v2df) __A, |
| 1206 | (__v2df) __B, |
| 1207 | (__v2df) __C, |
| 1208 | (__mmask8) __U); |
| 1209 | } |
| 1210 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1211 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1212 | _mm_mask3_fmadd_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) |
| 1213 | { |
| 1214 | return (__m128d) __builtin_ia32_vfmaddpd128_mask3 ((__v2df) __A, |
| 1215 | (__v2df) __B, |
| 1216 | (__v2df) __C, |
| 1217 | (__mmask8) __U); |
| 1218 | } |
| 1219 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1220 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1221 | _mm_maskz_fmadd_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) |
| 1222 | { |
| 1223 | return (__m128d) __builtin_ia32_vfmaddpd128_maskz ((__v2df) __A, |
| 1224 | (__v2df) __B, |
| 1225 | (__v2df) __C, |
| 1226 | (__mmask8) __U); |
| 1227 | } |
| 1228 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1229 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1230 | _mm_mask_fmsub_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) |
| 1231 | { |
| 1232 | return (__m128d) __builtin_ia32_vfmaddpd128_mask ((__v2df) __A, |
| 1233 | (__v2df) __B, |
| 1234 | -(__v2df) __C, |
| 1235 | (__mmask8) __U); |
| 1236 | } |
| 1237 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1238 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1239 | _mm_maskz_fmsub_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) |
| 1240 | { |
| 1241 | return (__m128d) __builtin_ia32_vfmaddpd128_maskz ((__v2df) __A, |
| 1242 | (__v2df) __B, |
| 1243 | -(__v2df) __C, |
| 1244 | (__mmask8) __U); |
| 1245 | } |
| 1246 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1247 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1248 | _mm_mask3_fnmadd_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) |
| 1249 | { |
| 1250 | return (__m128d) __builtin_ia32_vfmaddpd128_mask3 (-(__v2df) __A, |
| 1251 | (__v2df) __B, |
| 1252 | (__v2df) __C, |
| 1253 | (__mmask8) __U); |
| 1254 | } |
| 1255 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1256 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1257 | _mm_maskz_fnmadd_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) |
| 1258 | { |
| 1259 | return (__m128d) __builtin_ia32_vfmaddpd128_maskz (-(__v2df) __A, |
| 1260 | (__v2df) __B, |
| 1261 | (__v2df) __C, |
| 1262 | (__mmask8) __U); |
| 1263 | } |
| 1264 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1265 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1266 | _mm_maskz_fnmsub_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) |
| 1267 | { |
| 1268 | return (__m128d) __builtin_ia32_vfmaddpd128_maskz (-(__v2df) __A, |
| 1269 | (__v2df) __B, |
| 1270 | -(__v2df) __C, |
| 1271 | (__mmask8) __U); |
| 1272 | } |
| 1273 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1274 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1275 | _mm256_mask_fmadd_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) |
| 1276 | { |
| 1277 | return (__m256d) __builtin_ia32_vfmaddpd256_mask ((__v4df) __A, |
| 1278 | (__v4df) __B, |
| 1279 | (__v4df) __C, |
| 1280 | (__mmask8) __U); |
| 1281 | } |
| 1282 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1283 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1284 | _mm256_mask3_fmadd_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) |
| 1285 | { |
| 1286 | return (__m256d) __builtin_ia32_vfmaddpd256_mask3 ((__v4df) __A, |
| 1287 | (__v4df) __B, |
| 1288 | (__v4df) __C, |
| 1289 | (__mmask8) __U); |
| 1290 | } |
| 1291 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1292 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1293 | _mm256_maskz_fmadd_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) |
| 1294 | { |
| 1295 | return (__m256d) __builtin_ia32_vfmaddpd256_maskz ((__v4df) __A, |
| 1296 | (__v4df) __B, |
| 1297 | (__v4df) __C, |
| 1298 | (__mmask8) __U); |
| 1299 | } |
| 1300 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1301 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1302 | _mm256_mask_fmsub_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) |
| 1303 | { |
| 1304 | return (__m256d) __builtin_ia32_vfmaddpd256_mask ((__v4df) __A, |
| 1305 | (__v4df) __B, |
| 1306 | -(__v4df) __C, |
| 1307 | (__mmask8) __U); |
| 1308 | } |
| 1309 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1310 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1311 | _mm256_maskz_fmsub_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) |
| 1312 | { |
| 1313 | return (__m256d) __builtin_ia32_vfmaddpd256_maskz ((__v4df) __A, |
| 1314 | (__v4df) __B, |
| 1315 | -(__v4df) __C, |
| 1316 | (__mmask8) __U); |
| 1317 | } |
| 1318 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1319 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1320 | _mm256_mask3_fnmadd_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) |
| 1321 | { |
| 1322 | return (__m256d) __builtin_ia32_vfmaddpd256_mask3 (-(__v4df) __A, |
| 1323 | (__v4df) __B, |
| 1324 | (__v4df) __C, |
| 1325 | (__mmask8) __U); |
| 1326 | } |
| 1327 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1328 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1329 | _mm256_maskz_fnmadd_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) |
| 1330 | { |
| 1331 | return (__m256d) __builtin_ia32_vfmaddpd256_maskz (-(__v4df) __A, |
| 1332 | (__v4df) __B, |
| 1333 | (__v4df) __C, |
| 1334 | (__mmask8) __U); |
| 1335 | } |
| 1336 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1337 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1338 | _mm256_maskz_fnmsub_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) |
| 1339 | { |
| 1340 | return (__m256d) __builtin_ia32_vfmaddpd256_maskz (-(__v4df) __A, |
| 1341 | (__v4df) __B, |
| 1342 | -(__v4df) __C, |
| 1343 | (__mmask8) __U); |
| 1344 | } |
| 1345 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1346 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1347 | _mm_mask_fmadd_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) |
| 1348 | { |
| 1349 | return (__m128) __builtin_ia32_vfmaddps128_mask ((__v4sf) __A, |
| 1350 | (__v4sf) __B, |
| 1351 | (__v4sf) __C, |
| 1352 | (__mmask8) __U); |
| 1353 | } |
| 1354 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1355 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1356 | _mm_mask3_fmadd_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) |
| 1357 | { |
| 1358 | return (__m128) __builtin_ia32_vfmaddps128_mask3 ((__v4sf) __A, |
| 1359 | (__v4sf) __B, |
| 1360 | (__v4sf) __C, |
| 1361 | (__mmask8) __U); |
| 1362 | } |
| 1363 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1364 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1365 | _mm_maskz_fmadd_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) |
| 1366 | { |
| 1367 | return (__m128) __builtin_ia32_vfmaddps128_maskz ((__v4sf) __A, |
| 1368 | (__v4sf) __B, |
| 1369 | (__v4sf) __C, |
| 1370 | (__mmask8) __U); |
| 1371 | } |
| 1372 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1373 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1374 | _mm_mask_fmsub_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) |
| 1375 | { |
| 1376 | return (__m128) __builtin_ia32_vfmaddps128_mask ((__v4sf) __A, |
| 1377 | (__v4sf) __B, |
| 1378 | -(__v4sf) __C, |
| 1379 | (__mmask8) __U); |
| 1380 | } |
| 1381 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1382 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1383 | _mm_maskz_fmsub_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) |
| 1384 | { |
| 1385 | return (__m128) __builtin_ia32_vfmaddps128_maskz ((__v4sf) __A, |
| 1386 | (__v4sf) __B, |
| 1387 | -(__v4sf) __C, |
| 1388 | (__mmask8) __U); |
| 1389 | } |
| 1390 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1391 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1392 | _mm_mask3_fnmadd_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) |
| 1393 | { |
| 1394 | return (__m128) __builtin_ia32_vfmaddps128_mask3 (-(__v4sf) __A, |
| 1395 | (__v4sf) __B, |
| 1396 | (__v4sf) __C, |
| 1397 | (__mmask8) __U); |
| 1398 | } |
| 1399 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1400 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1401 | _mm_maskz_fnmadd_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) |
| 1402 | { |
| 1403 | return (__m128) __builtin_ia32_vfmaddps128_maskz (-(__v4sf) __A, |
| 1404 | (__v4sf) __B, |
| 1405 | (__v4sf) __C, |
| 1406 | (__mmask8) __U); |
| 1407 | } |
| 1408 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1409 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1410 | _mm_maskz_fnmsub_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) |
| 1411 | { |
| 1412 | return (__m128) __builtin_ia32_vfmaddps128_maskz (-(__v4sf) __A, |
| 1413 | (__v4sf) __B, |
| 1414 | -(__v4sf) __C, |
| 1415 | (__mmask8) __U); |
| 1416 | } |
| 1417 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1418 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1419 | _mm256_mask_fmadd_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) |
| 1420 | { |
| 1421 | return (__m256) __builtin_ia32_vfmaddps256_mask ((__v8sf) __A, |
| 1422 | (__v8sf) __B, |
| 1423 | (__v8sf) __C, |
| 1424 | (__mmask8) __U); |
| 1425 | } |
| 1426 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1427 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1428 | _mm256_mask3_fmadd_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) |
| 1429 | { |
| 1430 | return (__m256) __builtin_ia32_vfmaddps256_mask3 ((__v8sf) __A, |
| 1431 | (__v8sf) __B, |
| 1432 | (__v8sf) __C, |
| 1433 | (__mmask8) __U); |
| 1434 | } |
| 1435 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1436 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1437 | _mm256_maskz_fmadd_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) |
| 1438 | { |
| 1439 | return (__m256) __builtin_ia32_vfmaddps256_maskz ((__v8sf) __A, |
| 1440 | (__v8sf) __B, |
| 1441 | (__v8sf) __C, |
| 1442 | (__mmask8) __U); |
| 1443 | } |
| 1444 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1445 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1446 | _mm256_mask_fmsub_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) |
| 1447 | { |
| 1448 | return (__m256) __builtin_ia32_vfmaddps256_mask ((__v8sf) __A, |
| 1449 | (__v8sf) __B, |
| 1450 | -(__v8sf) __C, |
| 1451 | (__mmask8) __U); |
| 1452 | } |
| 1453 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1454 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1455 | _mm256_maskz_fmsub_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) |
| 1456 | { |
| 1457 | return (__m256) __builtin_ia32_vfmaddps256_maskz ((__v8sf) __A, |
| 1458 | (__v8sf) __B, |
| 1459 | -(__v8sf) __C, |
| 1460 | (__mmask8) __U); |
| 1461 | } |
| 1462 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1463 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1464 | _mm256_mask3_fnmadd_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) |
| 1465 | { |
| 1466 | return (__m256) __builtin_ia32_vfmaddps256_mask3 (-(__v8sf) __A, |
| 1467 | (__v8sf) __B, |
| 1468 | (__v8sf) __C, |
| 1469 | (__mmask8) __U); |
| 1470 | } |
| 1471 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1472 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1473 | _mm256_maskz_fnmadd_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) |
| 1474 | { |
| 1475 | return (__m256) __builtin_ia32_vfmaddps256_maskz (-(__v8sf) __A, |
| 1476 | (__v8sf) __B, |
| 1477 | (__v8sf) __C, |
| 1478 | (__mmask8) __U); |
| 1479 | } |
| 1480 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1481 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1482 | _mm256_maskz_fnmsub_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) |
| 1483 | { |
| 1484 | return (__m256) __builtin_ia32_vfmaddps256_maskz (-(__v8sf) __A, |
| 1485 | (__v8sf) __B, |
| 1486 | -(__v8sf) __C, |
| 1487 | (__mmask8) __U); |
| 1488 | } |
| 1489 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1490 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1491 | _mm_mask_fmaddsub_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) |
| 1492 | { |
| 1493 | return (__m128d) __builtin_ia32_vfmaddsubpd128_mask ((__v2df) __A, |
| 1494 | (__v2df) __B, |
| 1495 | (__v2df) __C, |
| 1496 | (__mmask8) __U); |
| 1497 | } |
| 1498 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1499 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1500 | _mm_mask3_fmaddsub_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) |
| 1501 | { |
| 1502 | return (__m128d) __builtin_ia32_vfmaddsubpd128_mask3 ((__v2df) __A, |
| 1503 | (__v2df) __B, |
| 1504 | (__v2df) __C, |
| 1505 | (__mmask8) |
| 1506 | __U); |
| 1507 | } |
| 1508 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1509 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1510 | _mm_maskz_fmaddsub_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) |
| 1511 | { |
| 1512 | return (__m128d) __builtin_ia32_vfmaddsubpd128_maskz ((__v2df) __A, |
| 1513 | (__v2df) __B, |
| 1514 | (__v2df) __C, |
| 1515 | (__mmask8) |
| 1516 | __U); |
| 1517 | } |
| 1518 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1519 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1520 | _mm_mask_fmsubadd_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) |
| 1521 | { |
| 1522 | return (__m128d) __builtin_ia32_vfmaddsubpd128_mask ((__v2df) __A, |
| 1523 | (__v2df) __B, |
| 1524 | -(__v2df) __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_maskz_fmsubadd_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) |
| 1530 | { |
| 1531 | return (__m128d) __builtin_ia32_vfmaddsubpd128_maskz ((__v2df) __A, |
| 1532 | (__v2df) __B, |
| 1533 | -(__v2df) __C, |
| 1534 | (__mmask8) |
| 1535 | __U); |
| 1536 | } |
| 1537 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1538 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1539 | _mm256_mask_fmaddsub_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) |
| 1540 | { |
| 1541 | return (__m256d) __builtin_ia32_vfmaddsubpd256_mask ((__v4df) __A, |
| 1542 | (__v4df) __B, |
| 1543 | (__v4df) __C, |
| 1544 | (__mmask8) __U); |
| 1545 | } |
| 1546 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1547 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1548 | _mm256_mask3_fmaddsub_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) |
| 1549 | { |
| 1550 | return (__m256d) __builtin_ia32_vfmaddsubpd256_mask3 ((__v4df) __A, |
| 1551 | (__v4df) __B, |
| 1552 | (__v4df) __C, |
| 1553 | (__mmask8) |
| 1554 | __U); |
| 1555 | } |
| 1556 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1557 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1558 | _mm256_maskz_fmaddsub_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) |
| 1559 | { |
| 1560 | return (__m256d) __builtin_ia32_vfmaddsubpd256_maskz ((__v4df) __A, |
| 1561 | (__v4df) __B, |
| 1562 | (__v4df) __C, |
| 1563 | (__mmask8) |
| 1564 | __U); |
| 1565 | } |
| 1566 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1567 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1568 | _mm256_mask_fmsubadd_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) |
| 1569 | { |
| 1570 | return (__m256d) __builtin_ia32_vfmaddsubpd256_mask ((__v4df) __A, |
| 1571 | (__v4df) __B, |
| 1572 | -(__v4df) __C, |
| 1573 | (__mmask8) __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_maskz_fmsubadd_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) |
| 1578 | { |
| 1579 | return (__m256d) __builtin_ia32_vfmaddsubpd256_maskz ((__v4df) __A, |
| 1580 | (__v4df) __B, |
| 1581 | -(__v4df) __C, |
| 1582 | (__mmask8) |
| 1583 | __U); |
| 1584 | } |
| 1585 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1586 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1587 | _mm_mask_fmaddsub_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) |
| 1588 | { |
| 1589 | return (__m128) __builtin_ia32_vfmaddsubps128_mask ((__v4sf) __A, |
| 1590 | (__v4sf) __B, |
| 1591 | (__v4sf) __C, |
| 1592 | (__mmask8) __U); |
| 1593 | } |
| 1594 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1595 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1596 | _mm_mask3_fmaddsub_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) |
| 1597 | { |
| 1598 | return (__m128) __builtin_ia32_vfmaddsubps128_mask3 ((__v4sf) __A, |
| 1599 | (__v4sf) __B, |
| 1600 | (__v4sf) __C, |
| 1601 | (__mmask8) __U); |
| 1602 | } |
| 1603 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1604 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1605 | _mm_maskz_fmaddsub_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) |
| 1606 | { |
| 1607 | return (__m128) __builtin_ia32_vfmaddsubps128_maskz ((__v4sf) __A, |
| 1608 | (__v4sf) __B, |
| 1609 | (__v4sf) __C, |
| 1610 | (__mmask8) __U); |
| 1611 | } |
| 1612 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1613 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1614 | _mm_mask_fmsubadd_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) |
| 1615 | { |
| 1616 | return (__m128) __builtin_ia32_vfmaddsubps128_mask ((__v4sf) __A, |
| 1617 | (__v4sf) __B, |
| 1618 | -(__v4sf) __C, |
| 1619 | (__mmask8) __U); |
| 1620 | } |
| 1621 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1622 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1623 | _mm_maskz_fmsubadd_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) |
| 1624 | { |
| 1625 | return (__m128) __builtin_ia32_vfmaddsubps128_maskz ((__v4sf) __A, |
| 1626 | (__v4sf) __B, |
| 1627 | -(__v4sf) __C, |
| 1628 | (__mmask8) __U); |
| 1629 | } |
| 1630 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1631 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1632 | _mm256_mask_fmaddsub_ps(__m256 __A, __mmask8 __U, __m256 __B, |
| 1633 | __m256 __C) |
| 1634 | { |
| 1635 | return (__m256) __builtin_ia32_vfmaddsubps256_mask ((__v8sf) __A, |
| 1636 | (__v8sf) __B, |
| 1637 | (__v8sf) __C, |
| 1638 | (__mmask8) __U); |
| 1639 | } |
| 1640 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1641 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1642 | _mm256_mask3_fmaddsub_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) |
| 1643 | { |
| 1644 | return (__m256) __builtin_ia32_vfmaddsubps256_mask3 ((__v8sf) __A, |
| 1645 | (__v8sf) __B, |
| 1646 | (__v8sf) __C, |
| 1647 | (__mmask8) __U); |
| 1648 | } |
| 1649 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1650 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1651 | _mm256_maskz_fmaddsub_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) |
| 1652 | { |
| 1653 | return (__m256) __builtin_ia32_vfmaddsubps256_maskz ((__v8sf) __A, |
| 1654 | (__v8sf) __B, |
| 1655 | (__v8sf) __C, |
| 1656 | (__mmask8) __U); |
| 1657 | } |
| 1658 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1659 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1660 | _mm256_mask_fmsubadd_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) |
| 1661 | { |
| 1662 | return (__m256) __builtin_ia32_vfmaddsubps256_mask ((__v8sf) __A, |
| 1663 | (__v8sf) __B, |
| 1664 | -(__v8sf) __C, |
| 1665 | (__mmask8) __U); |
| 1666 | } |
| 1667 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1668 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1669 | _mm256_maskz_fmsubadd_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) |
| 1670 | { |
| 1671 | return (__m256) __builtin_ia32_vfmaddsubps256_maskz ((__v8sf) __A, |
| 1672 | (__v8sf) __B, |
| 1673 | -(__v8sf) __C, |
| 1674 | (__mmask8) __U); |
| 1675 | } |
| 1676 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1677 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1678 | _mm_mask3_fmsub_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) |
| 1679 | { |
| 1680 | return (__m128d) __builtin_ia32_vfmsubpd128_mask3 ((__v2df) __A, |
| 1681 | (__v2df) __B, |
| 1682 | (__v2df) __C, |
| 1683 | (__mmask8) __U); |
| 1684 | } |
| 1685 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1686 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1687 | _mm256_mask3_fmsub_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) |
| 1688 | { |
| 1689 | return (__m256d) __builtin_ia32_vfmsubpd256_mask3 ((__v4df) __A, |
| 1690 | (__v4df) __B, |
| 1691 | (__v4df) __C, |
| 1692 | (__mmask8) __U); |
| 1693 | } |
| 1694 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1695 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1696 | _mm_mask3_fmsub_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) |
| 1697 | { |
| 1698 | return (__m128) __builtin_ia32_vfmsubps128_mask3 ((__v4sf) __A, |
| 1699 | (__v4sf) __B, |
| 1700 | (__v4sf) __C, |
| 1701 | (__mmask8) __U); |
| 1702 | } |
| 1703 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1704 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1705 | _mm256_mask3_fmsub_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) |
| 1706 | { |
| 1707 | return (__m256) __builtin_ia32_vfmsubps256_mask3 ((__v8sf) __A, |
| 1708 | (__v8sf) __B, |
| 1709 | (__v8sf) __C, |
| 1710 | (__mmask8) __U); |
| 1711 | } |
| 1712 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1713 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1714 | _mm_mask3_fmsubadd_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) |
| 1715 | { |
| 1716 | return (__m128d) __builtin_ia32_vfmsubaddpd128_mask3 ((__v2df) __A, |
| 1717 | (__v2df) __B, |
| 1718 | (__v2df) __C, |
| 1719 | (__mmask8) |
| 1720 | __U); |
| 1721 | } |
| 1722 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1723 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1724 | _mm256_mask3_fmsubadd_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) |
| 1725 | { |
| 1726 | return (__m256d) __builtin_ia32_vfmsubaddpd256_mask3 ((__v4df) __A, |
| 1727 | (__v4df) __B, |
| 1728 | (__v4df) __C, |
| 1729 | (__mmask8) |
| 1730 | __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_fmsubadd_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) |
| 1735 | { |
| 1736 | return (__m128) __builtin_ia32_vfmsubaddps128_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_fmsubadd_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) |
| 1744 | { |
| 1745 | return (__m256) __builtin_ia32_vfmsubaddps256_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_mask_fnmadd_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) |
| 1753 | { |
| 1754 | return (__m128d) __builtin_ia32_vfnmaddpd128_mask ((__v2df) __A, |
| 1755 | (__v2df) __B, |
| 1756 | (__v2df) __C, |
| 1757 | (__mmask8) __U); |
| 1758 | } |
| 1759 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1760 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1761 | _mm256_mask_fnmadd_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) |
| 1762 | { |
| 1763 | return (__m256d) __builtin_ia32_vfnmaddpd256_mask ((__v4df) __A, |
| 1764 | (__v4df) __B, |
| 1765 | (__v4df) __C, |
| 1766 | (__mmask8) __U); |
| 1767 | } |
| 1768 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1769 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1770 | _mm_mask_fnmadd_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) |
| 1771 | { |
| 1772 | return (__m128) __builtin_ia32_vfnmaddps128_mask ((__v4sf) __A, |
| 1773 | (__v4sf) __B, |
| 1774 | (__v4sf) __C, |
| 1775 | (__mmask8) __U); |
| 1776 | } |
| 1777 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1778 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1779 | _mm256_mask_fnmadd_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) |
| 1780 | { |
| 1781 | return (__m256) __builtin_ia32_vfnmaddps256_mask ((__v8sf) __A, |
| 1782 | (__v8sf) __B, |
| 1783 | (__v8sf) __C, |
| 1784 | (__mmask8) __U); |
| 1785 | } |
| 1786 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1787 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1788 | _mm_mask_fnmsub_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) |
| 1789 | { |
| 1790 | return (__m128d) __builtin_ia32_vfnmsubpd128_mask ((__v2df) __A, |
| 1791 | (__v2df) __B, |
| 1792 | (__v2df) __C, |
| 1793 | (__mmask8) __U); |
| 1794 | } |
| 1795 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1796 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1797 | _mm_mask3_fnmsub_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) |
| 1798 | { |
| 1799 | return (__m128d) __builtin_ia32_vfnmsubpd128_mask3 ((__v2df) __A, |
| 1800 | (__v2df) __B, |
| 1801 | (__v2df) __C, |
| 1802 | (__mmask8) __U); |
| 1803 | } |
| 1804 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1805 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1806 | _mm256_mask_fnmsub_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) |
| 1807 | { |
| 1808 | return (__m256d) __builtin_ia32_vfnmsubpd256_mask ((__v4df) __A, |
| 1809 | (__v4df) __B, |
| 1810 | (__v4df) __C, |
| 1811 | (__mmask8) __U); |
| 1812 | } |
| 1813 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1814 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1815 | _mm256_mask3_fnmsub_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) |
| 1816 | { |
| 1817 | return (__m256d) __builtin_ia32_vfnmsubpd256_mask3 ((__v4df) __A, |
| 1818 | (__v4df) __B, |
| 1819 | (__v4df) __C, |
| 1820 | (__mmask8) __U); |
| 1821 | } |
| 1822 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1823 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1824 | _mm_mask_fnmsub_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) |
| 1825 | { |
| 1826 | return (__m128) __builtin_ia32_vfnmsubps128_mask ((__v4sf) __A, |
| 1827 | (__v4sf) __B, |
| 1828 | (__v4sf) __C, |
| 1829 | (__mmask8) __U); |
| 1830 | } |
| 1831 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1832 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1833 | _mm_mask3_fnmsub_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) |
| 1834 | { |
| 1835 | return (__m128) __builtin_ia32_vfnmsubps128_mask3 ((__v4sf) __A, |
| 1836 | (__v4sf) __B, |
| 1837 | (__v4sf) __C, |
| 1838 | (__mmask8) __U); |
| 1839 | } |
| 1840 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1841 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1842 | _mm256_mask_fnmsub_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) |
| 1843 | { |
| 1844 | return (__m256) __builtin_ia32_vfnmsubps256_mask ((__v8sf) __A, |
| 1845 | (__v8sf) __B, |
| 1846 | (__v8sf) __C, |
| 1847 | (__mmask8) __U); |
| 1848 | } |
| 1849 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1850 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1851 | _mm256_mask3_fnmsub_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) |
| 1852 | { |
| 1853 | return (__m256) __builtin_ia32_vfnmsubps256_mask3 ((__v8sf) __A, |
| 1854 | (__v8sf) __B, |
| 1855 | (__v8sf) __C, |
| 1856 | (__mmask8) __U); |
| 1857 | } |
| 1858 | |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1859 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 1860 | _mm_mask_add_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 1861 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 1862 | (__v2df)_mm_add_pd(__A, __B), |
| 1863 | (__v2df)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1864 | } |
| 1865 | |
| 1866 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 1867 | _mm_maskz_add_pd(__mmask8 __U, __m128d __A, __m128d __B) { |
| 1868 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 1869 | (__v2df)_mm_add_pd(__A, __B), |
| 1870 | (__v2df)_mm_setzero_pd()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1871 | } |
| 1872 | |
| 1873 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 1874 | _mm256_mask_add_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) { |
| 1875 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 1876 | (__v4df)_mm256_add_pd(__A, __B), |
| 1877 | (__v4df)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1878 | } |
| 1879 | |
| 1880 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 1881 | _mm256_maskz_add_pd(__mmask8 __U, __m256d __A, __m256d __B) { |
| 1882 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 1883 | (__v4df)_mm256_add_pd(__A, __B), |
| 1884 | (__v4df)_mm256_setzero_pd()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1885 | } |
| 1886 | |
| 1887 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 1888 | _mm_mask_add_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 1889 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 1890 | (__v4sf)_mm_add_ps(__A, __B), |
| 1891 | (__v4sf)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1892 | } |
| 1893 | |
| 1894 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 1895 | _mm_maskz_add_ps(__mmask8 __U, __m128 __A, __m128 __B) { |
| 1896 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 1897 | (__v4sf)_mm_add_ps(__A, __B), |
| 1898 | (__v4sf)_mm_setzero_ps()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1899 | } |
| 1900 | |
| 1901 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 1902 | _mm256_mask_add_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 1903 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 1904 | (__v8sf)_mm256_add_ps(__A, __B), |
| 1905 | (__v8sf)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1906 | } |
| 1907 | |
| 1908 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 1909 | _mm256_maskz_add_ps(__mmask8 __U, __m256 __A, __m256 __B) { |
| 1910 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 1911 | (__v8sf)_mm256_add_ps(__A, __B), |
| 1912 | (__v8sf)_mm256_setzero_ps()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1913 | } |
| 1914 | |
| 1915 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1916 | _mm_mask_blend_epi32 (__mmask8 __U, __m128i __A, __m128i __W) { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1917 | return (__m128i) __builtin_ia32_selectd_128 ((__mmask8) __U, |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1918 | (__v4si) __W, |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1919 | (__v4si) __A); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1920 | } |
| 1921 | |
| 1922 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1923 | _mm256_mask_blend_epi32 (__mmask8 __U, __m256i __A, __m256i __W) { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1924 | return (__m256i) __builtin_ia32_selectd_256 ((__mmask8) __U, |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1925 | (__v8si) __W, |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1926 | (__v8si) __A); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1927 | } |
| 1928 | |
| 1929 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 1930 | _mm_mask_blend_pd (__mmask8 __U, __m128d __A, __m128d __W) { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1931 | return (__m128d) __builtin_ia32_selectpd_128 ((__mmask8) __U, |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1932 | (__v2df) __W, |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1933 | (__v2df) __A); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1934 | } |
| 1935 | |
| 1936 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 1937 | _mm256_mask_blend_pd (__mmask8 __U, __m256d __A, __m256d __W) { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1938 | return (__m256d) __builtin_ia32_selectpd_256 ((__mmask8) __U, |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1939 | (__v4df) __W, |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1940 | (__v4df) __A); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1941 | } |
| 1942 | |
| 1943 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 1944 | _mm_mask_blend_ps (__mmask8 __U, __m128 __A, __m128 __W) { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1945 | return (__m128) __builtin_ia32_selectps_128 ((__mmask8) __U, |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1946 | (__v4sf) __W, |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1947 | (__v4sf) __A); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1948 | } |
| 1949 | |
| 1950 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 1951 | _mm256_mask_blend_ps (__mmask8 __U, __m256 __A, __m256 __W) { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1952 | return (__m256) __builtin_ia32_selectps_256 ((__mmask8) __U, |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1953 | (__v8sf) __W, |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1954 | (__v8sf) __A); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1955 | } |
| 1956 | |
| 1957 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1958 | _mm_mask_blend_epi64 (__mmask8 __U, __m128i __A, __m128i __W) { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1959 | return (__m128i) __builtin_ia32_selectq_128 ((__mmask8) __U, |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1960 | (__v2di) __W, |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1961 | (__v2di) __A); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1962 | } |
| 1963 | |
| 1964 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1965 | _mm256_mask_blend_epi64 (__mmask8 __U, __m256i __A, __m256i __W) { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1966 | return (__m256i) __builtin_ia32_selectq_256 ((__mmask8) __U, |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1967 | (__v4di) __W, |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1968 | (__v4di) __A); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1969 | } |
| 1970 | |
| 1971 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 1972 | _mm_mask_compress_pd (__m128d __W, __mmask8 __U, __m128d __A) { |
| 1973 | return (__m128d) __builtin_ia32_compressdf128_mask ((__v2df) __A, |
| 1974 | (__v2df) __W, |
| 1975 | (__mmask8) __U); |
| 1976 | } |
| 1977 | |
| 1978 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 1979 | _mm_maskz_compress_pd (__mmask8 __U, __m128d __A) { |
| 1980 | return (__m128d) __builtin_ia32_compressdf128_mask ((__v2df) __A, |
| 1981 | (__v2df) |
| 1982 | _mm_setzero_pd (), |
| 1983 | (__mmask8) __U); |
| 1984 | } |
| 1985 | |
| 1986 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 1987 | _mm256_mask_compress_pd (__m256d __W, __mmask8 __U, __m256d __A) { |
| 1988 | return (__m256d) __builtin_ia32_compressdf256_mask ((__v4df) __A, |
| 1989 | (__v4df) __W, |
| 1990 | (__mmask8) __U); |
| 1991 | } |
| 1992 | |
| 1993 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 1994 | _mm256_maskz_compress_pd (__mmask8 __U, __m256d __A) { |
| 1995 | return (__m256d) __builtin_ia32_compressdf256_mask ((__v4df) __A, |
| 1996 | (__v4df) |
| 1997 | _mm256_setzero_pd (), |
| 1998 | (__mmask8) __U); |
| 1999 | } |
| 2000 | |
| 2001 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2002 | _mm_mask_compress_epi64 (__m128i __W, __mmask8 __U, __m128i __A) { |
| 2003 | return (__m128i) __builtin_ia32_compressdi128_mask ((__v2di) __A, |
| 2004 | (__v2di) __W, |
| 2005 | (__mmask8) __U); |
| 2006 | } |
| 2007 | |
| 2008 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2009 | _mm_maskz_compress_epi64 (__mmask8 __U, __m128i __A) { |
| 2010 | return (__m128i) __builtin_ia32_compressdi128_mask ((__v2di) __A, |
| 2011 | (__v2di) |
| 2012 | _mm_setzero_si128 (), |
| 2013 | (__mmask8) __U); |
| 2014 | } |
| 2015 | |
| 2016 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2017 | _mm256_mask_compress_epi64 (__m256i __W, __mmask8 __U, __m256i __A) { |
| 2018 | return (__m256i) __builtin_ia32_compressdi256_mask ((__v4di) __A, |
| 2019 | (__v4di) __W, |
| 2020 | (__mmask8) __U); |
| 2021 | } |
| 2022 | |
| 2023 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2024 | _mm256_maskz_compress_epi64 (__mmask8 __U, __m256i __A) { |
| 2025 | return (__m256i) __builtin_ia32_compressdi256_mask ((__v4di) __A, |
| 2026 | (__v4di) |
| 2027 | _mm256_setzero_si256 (), |
| 2028 | (__mmask8) __U); |
| 2029 | } |
| 2030 | |
| 2031 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2032 | _mm_mask_compress_ps (__m128 __W, __mmask8 __U, __m128 __A) { |
| 2033 | return (__m128) __builtin_ia32_compresssf128_mask ((__v4sf) __A, |
| 2034 | (__v4sf) __W, |
| 2035 | (__mmask8) __U); |
| 2036 | } |
| 2037 | |
| 2038 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2039 | _mm_maskz_compress_ps (__mmask8 __U, __m128 __A) { |
| 2040 | return (__m128) __builtin_ia32_compresssf128_mask ((__v4sf) __A, |
| 2041 | (__v4sf) |
| 2042 | _mm_setzero_ps (), |
| 2043 | (__mmask8) __U); |
| 2044 | } |
| 2045 | |
| 2046 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2047 | _mm256_mask_compress_ps (__m256 __W, __mmask8 __U, __m256 __A) { |
| 2048 | return (__m256) __builtin_ia32_compresssf256_mask ((__v8sf) __A, |
| 2049 | (__v8sf) __W, |
| 2050 | (__mmask8) __U); |
| 2051 | } |
| 2052 | |
| 2053 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2054 | _mm256_maskz_compress_ps (__mmask8 __U, __m256 __A) { |
| 2055 | return (__m256) __builtin_ia32_compresssf256_mask ((__v8sf) __A, |
| 2056 | (__v8sf) |
| 2057 | _mm256_setzero_ps (), |
| 2058 | (__mmask8) __U); |
| 2059 | } |
| 2060 | |
| 2061 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2062 | _mm_mask_compress_epi32 (__m128i __W, __mmask8 __U, __m128i __A) { |
| 2063 | return (__m128i) __builtin_ia32_compresssi128_mask ((__v4si) __A, |
| 2064 | (__v4si) __W, |
| 2065 | (__mmask8) __U); |
| 2066 | } |
| 2067 | |
| 2068 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2069 | _mm_maskz_compress_epi32 (__mmask8 __U, __m128i __A) { |
| 2070 | return (__m128i) __builtin_ia32_compresssi128_mask ((__v4si) __A, |
| 2071 | (__v4si) |
| 2072 | _mm_setzero_si128 (), |
| 2073 | (__mmask8) __U); |
| 2074 | } |
| 2075 | |
| 2076 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2077 | _mm256_mask_compress_epi32 (__m256i __W, __mmask8 __U, __m256i __A) { |
| 2078 | return (__m256i) __builtin_ia32_compresssi256_mask ((__v8si) __A, |
| 2079 | (__v8si) __W, |
| 2080 | (__mmask8) __U); |
| 2081 | } |
| 2082 | |
| 2083 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2084 | _mm256_maskz_compress_epi32 (__mmask8 __U, __m256i __A) { |
| 2085 | return (__m256i) __builtin_ia32_compresssi256_mask ((__v8si) __A, |
| 2086 | (__v8si) |
| 2087 | _mm256_setzero_si256 (), |
| 2088 | (__mmask8) __U); |
| 2089 | } |
| 2090 | |
| 2091 | static __inline__ void __DEFAULT_FN_ATTRS |
| 2092 | _mm_mask_compressstoreu_pd (void *__P, __mmask8 __U, __m128d __A) { |
| 2093 | __builtin_ia32_compressstoredf128_mask ((__v2df *) __P, |
| 2094 | (__v2df) __A, |
| 2095 | (__mmask8) __U); |
| 2096 | } |
| 2097 | |
| 2098 | static __inline__ void __DEFAULT_FN_ATTRS |
| 2099 | _mm256_mask_compressstoreu_pd (void *__P, __mmask8 __U, __m256d __A) { |
| 2100 | __builtin_ia32_compressstoredf256_mask ((__v4df *) __P, |
| 2101 | (__v4df) __A, |
| 2102 | (__mmask8) __U); |
| 2103 | } |
| 2104 | |
| 2105 | static __inline__ void __DEFAULT_FN_ATTRS |
| 2106 | _mm_mask_compressstoreu_epi64 (void *__P, __mmask8 __U, __m128i __A) { |
| 2107 | __builtin_ia32_compressstoredi128_mask ((__v2di *) __P, |
| 2108 | (__v2di) __A, |
| 2109 | (__mmask8) __U); |
| 2110 | } |
| 2111 | |
| 2112 | static __inline__ void __DEFAULT_FN_ATTRS |
| 2113 | _mm256_mask_compressstoreu_epi64 (void *__P, __mmask8 __U, __m256i __A) { |
| 2114 | __builtin_ia32_compressstoredi256_mask ((__v4di *) __P, |
| 2115 | (__v4di) __A, |
| 2116 | (__mmask8) __U); |
| 2117 | } |
| 2118 | |
| 2119 | static __inline__ void __DEFAULT_FN_ATTRS |
| 2120 | _mm_mask_compressstoreu_ps (void *__P, __mmask8 __U, __m128 __A) { |
| 2121 | __builtin_ia32_compressstoresf128_mask ((__v4sf *) __P, |
| 2122 | (__v4sf) __A, |
| 2123 | (__mmask8) __U); |
| 2124 | } |
| 2125 | |
| 2126 | static __inline__ void __DEFAULT_FN_ATTRS |
| 2127 | _mm256_mask_compressstoreu_ps (void *__P, __mmask8 __U, __m256 __A) { |
| 2128 | __builtin_ia32_compressstoresf256_mask ((__v8sf *) __P, |
| 2129 | (__v8sf) __A, |
| 2130 | (__mmask8) __U); |
| 2131 | } |
| 2132 | |
| 2133 | static __inline__ void __DEFAULT_FN_ATTRS |
| 2134 | _mm_mask_compressstoreu_epi32 (void *__P, __mmask8 __U, __m128i __A) { |
| 2135 | __builtin_ia32_compressstoresi128_mask ((__v4si *) __P, |
| 2136 | (__v4si) __A, |
| 2137 | (__mmask8) __U); |
| 2138 | } |
| 2139 | |
| 2140 | static __inline__ void __DEFAULT_FN_ATTRS |
| 2141 | _mm256_mask_compressstoreu_epi32 (void *__P, __mmask8 __U, __m256i __A) { |
| 2142 | __builtin_ia32_compressstoresi256_mask ((__v8si *) __P, |
| 2143 | (__v8si) __A, |
| 2144 | (__mmask8) __U); |
| 2145 | } |
| 2146 | |
| 2147 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2148 | _mm_mask_cvtepi32_pd (__m128d __W, __mmask8 __U, __m128i __A) { |
| 2149 | return (__m128d) __builtin_ia32_cvtdq2pd128_mask ((__v4si) __A, |
| 2150 | (__v2df) __W, |
| 2151 | (__mmask8) __U); |
| 2152 | } |
| 2153 | |
| 2154 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2155 | _mm_maskz_cvtepi32_pd (__mmask8 __U, __m128i __A) { |
| 2156 | return (__m128d) __builtin_ia32_cvtdq2pd128_mask ((__v4si) __A, |
| 2157 | (__v2df) |
| 2158 | _mm_setzero_pd (), |
| 2159 | (__mmask8) __U); |
| 2160 | } |
| 2161 | |
| 2162 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2163 | _mm256_mask_cvtepi32_pd (__m256d __W, __mmask8 __U, __m128i __A) { |
| 2164 | return (__m256d) __builtin_ia32_cvtdq2pd256_mask ((__v4si) __A, |
| 2165 | (__v4df) __W, |
| 2166 | (__mmask8) __U); |
| 2167 | } |
| 2168 | |
| 2169 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2170 | _mm256_maskz_cvtepi32_pd (__mmask8 __U, __m128i __A) { |
| 2171 | return (__m256d) __builtin_ia32_cvtdq2pd256_mask ((__v4si) __A, |
| 2172 | (__v4df) |
| 2173 | _mm256_setzero_pd (), |
| 2174 | (__mmask8) __U); |
| 2175 | } |
| 2176 | |
| 2177 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2178 | _mm_mask_cvtepi32_ps (__m128 __W, __mmask8 __U, __m128i __A) { |
| 2179 | return (__m128) __builtin_ia32_cvtdq2ps128_mask ((__v4si) __A, |
| 2180 | (__v4sf) __W, |
| 2181 | (__mmask8) __U); |
| 2182 | } |
| 2183 | |
| 2184 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2185 | _mm_maskz_cvtepi32_ps (__mmask16 __U, __m128i __A) { |
| 2186 | return (__m128) __builtin_ia32_cvtdq2ps128_mask ((__v4si) __A, |
| 2187 | (__v4sf) |
| 2188 | _mm_setzero_ps (), |
| 2189 | (__mmask8) __U); |
| 2190 | } |
| 2191 | |
| 2192 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2193 | _mm256_mask_cvtepi32_ps (__m256 __W, __mmask8 __U, __m256i __A) { |
| 2194 | return (__m256) __builtin_ia32_cvtdq2ps256_mask ((__v8si) __A, |
| 2195 | (__v8sf) __W, |
| 2196 | (__mmask8) __U); |
| 2197 | } |
| 2198 | |
| 2199 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2200 | _mm256_maskz_cvtepi32_ps (__mmask16 __U, __m256i __A) { |
| 2201 | return (__m256) __builtin_ia32_cvtdq2ps256_mask ((__v8si) __A, |
| 2202 | (__v8sf) |
| 2203 | _mm256_setzero_ps (), |
| 2204 | (__mmask8) __U); |
| 2205 | } |
| 2206 | |
| 2207 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2208 | _mm_mask_cvtpd_epi32 (__m128i __W, __mmask8 __U, __m128d __A) { |
| 2209 | return (__m128i) __builtin_ia32_cvtpd2dq128_mask ((__v2df) __A, |
| 2210 | (__v4si) __W, |
| 2211 | (__mmask8) __U); |
| 2212 | } |
| 2213 | |
| 2214 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2215 | _mm_maskz_cvtpd_epi32 (__mmask8 __U, __m128d __A) { |
| 2216 | return (__m128i) __builtin_ia32_cvtpd2dq128_mask ((__v2df) __A, |
| 2217 | (__v4si) |
| 2218 | _mm_setzero_si128 (), |
| 2219 | (__mmask8) __U); |
| 2220 | } |
| 2221 | |
| 2222 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2223 | _mm256_mask_cvtpd_epi32 (__m128i __W, __mmask8 __U, __m256d __A) { |
| 2224 | return (__m128i) __builtin_ia32_cvtpd2dq256_mask ((__v4df) __A, |
| 2225 | (__v4si) __W, |
| 2226 | (__mmask8) __U); |
| 2227 | } |
| 2228 | |
| 2229 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2230 | _mm256_maskz_cvtpd_epi32 (__mmask8 __U, __m256d __A) { |
| 2231 | return (__m128i) __builtin_ia32_cvtpd2dq256_mask ((__v4df) __A, |
| 2232 | (__v4si) |
| 2233 | _mm_setzero_si128 (), |
| 2234 | (__mmask8) __U); |
| 2235 | } |
| 2236 | |
| 2237 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2238 | _mm_mask_cvtpd_ps (__m128 __W, __mmask8 __U, __m128d __A) { |
| 2239 | return (__m128) __builtin_ia32_cvtpd2ps_mask ((__v2df) __A, |
| 2240 | (__v4sf) __W, |
| 2241 | (__mmask8) __U); |
| 2242 | } |
| 2243 | |
| 2244 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2245 | _mm_maskz_cvtpd_ps (__mmask8 __U, __m128d __A) { |
| 2246 | return (__m128) __builtin_ia32_cvtpd2ps_mask ((__v2df) __A, |
| 2247 | (__v4sf) |
| 2248 | _mm_setzero_ps (), |
| 2249 | (__mmask8) __U); |
| 2250 | } |
| 2251 | |
| 2252 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2253 | _mm256_mask_cvtpd_ps (__m128 __W, __mmask8 __U, __m256d __A) { |
| 2254 | return (__m128) __builtin_ia32_cvtpd2ps256_mask ((__v4df) __A, |
| 2255 | (__v4sf) __W, |
| 2256 | (__mmask8) __U); |
| 2257 | } |
| 2258 | |
| 2259 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2260 | _mm256_maskz_cvtpd_ps (__mmask8 __U, __m256d __A) { |
| 2261 | return (__m128) __builtin_ia32_cvtpd2ps256_mask ((__v4df) __A, |
| 2262 | (__v4sf) |
| 2263 | _mm_setzero_ps (), |
| 2264 | (__mmask8) __U); |
| 2265 | } |
| 2266 | |
| 2267 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2268 | _mm_cvtpd_epu32 (__m128d __A) { |
| 2269 | return (__m128i) __builtin_ia32_cvtpd2udq128_mask ((__v2df) __A, |
| 2270 | (__v4si) |
| 2271 | _mm_setzero_si128 (), |
| 2272 | (__mmask8) -1); |
| 2273 | } |
| 2274 | |
| 2275 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2276 | _mm_mask_cvtpd_epu32 (__m128i __W, __mmask8 __U, __m128d __A) { |
| 2277 | return (__m128i) __builtin_ia32_cvtpd2udq128_mask ((__v2df) __A, |
| 2278 | (__v4si) __W, |
| 2279 | (__mmask8) __U); |
| 2280 | } |
| 2281 | |
| 2282 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2283 | _mm_maskz_cvtpd_epu32 (__mmask8 __U, __m128d __A) { |
| 2284 | return (__m128i) __builtin_ia32_cvtpd2udq128_mask ((__v2df) __A, |
| 2285 | (__v4si) |
| 2286 | _mm_setzero_si128 (), |
| 2287 | (__mmask8) __U); |
| 2288 | } |
| 2289 | |
| 2290 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2291 | _mm256_cvtpd_epu32 (__m256d __A) { |
| 2292 | return (__m128i) __builtin_ia32_cvtpd2udq256_mask ((__v4df) __A, |
| 2293 | (__v4si) |
| 2294 | _mm_setzero_si128 (), |
| 2295 | (__mmask8) -1); |
| 2296 | } |
| 2297 | |
| 2298 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2299 | _mm256_mask_cvtpd_epu32 (__m128i __W, __mmask8 __U, __m256d __A) { |
| 2300 | return (__m128i) __builtin_ia32_cvtpd2udq256_mask ((__v4df) __A, |
| 2301 | (__v4si) __W, |
| 2302 | (__mmask8) __U); |
| 2303 | } |
| 2304 | |
| 2305 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2306 | _mm256_maskz_cvtpd_epu32 (__mmask8 __U, __m256d __A) { |
| 2307 | return (__m128i) __builtin_ia32_cvtpd2udq256_mask ((__v4df) __A, |
| 2308 | (__v4si) |
| 2309 | _mm_setzero_si128 (), |
| 2310 | (__mmask8) __U); |
| 2311 | } |
| 2312 | |
| 2313 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2314 | _mm_mask_cvtps_epi32 (__m128i __W, __mmask8 __U, __m128 __A) { |
| 2315 | return (__m128i) __builtin_ia32_cvtps2dq128_mask ((__v4sf) __A, |
| 2316 | (__v4si) __W, |
| 2317 | (__mmask8) __U); |
| 2318 | } |
| 2319 | |
| 2320 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2321 | _mm_maskz_cvtps_epi32 (__mmask8 __U, __m128 __A) { |
| 2322 | return (__m128i) __builtin_ia32_cvtps2dq128_mask ((__v4sf) __A, |
| 2323 | (__v4si) |
| 2324 | _mm_setzero_si128 (), |
| 2325 | (__mmask8) __U); |
| 2326 | } |
| 2327 | |
| 2328 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2329 | _mm256_mask_cvtps_epi32 (__m256i __W, __mmask8 __U, __m256 __A) { |
| 2330 | return (__m256i) __builtin_ia32_cvtps2dq256_mask ((__v8sf) __A, |
| 2331 | (__v8si) __W, |
| 2332 | (__mmask8) __U); |
| 2333 | } |
| 2334 | |
| 2335 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2336 | _mm256_maskz_cvtps_epi32 (__mmask8 __U, __m256 __A) { |
| 2337 | return (__m256i) __builtin_ia32_cvtps2dq256_mask ((__v8sf) __A, |
| 2338 | (__v8si) |
| 2339 | _mm256_setzero_si256 (), |
| 2340 | (__mmask8) __U); |
| 2341 | } |
| 2342 | |
| 2343 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2344 | _mm_mask_cvtps_pd (__m128d __W, __mmask8 __U, __m128 __A) { |
| 2345 | return (__m128d) __builtin_ia32_cvtps2pd128_mask ((__v4sf) __A, |
| 2346 | (__v2df) __W, |
| 2347 | (__mmask8) __U); |
| 2348 | } |
| 2349 | |
| 2350 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2351 | _mm_maskz_cvtps_pd (__mmask8 __U, __m128 __A) { |
| 2352 | return (__m128d) __builtin_ia32_cvtps2pd128_mask ((__v4sf) __A, |
| 2353 | (__v2df) |
| 2354 | _mm_setzero_pd (), |
| 2355 | (__mmask8) __U); |
| 2356 | } |
| 2357 | |
| 2358 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2359 | _mm256_mask_cvtps_pd (__m256d __W, __mmask8 __U, __m128 __A) { |
| 2360 | return (__m256d) __builtin_ia32_cvtps2pd256_mask ((__v4sf) __A, |
| 2361 | (__v4df) __W, |
| 2362 | (__mmask8) __U); |
| 2363 | } |
| 2364 | |
| 2365 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2366 | _mm256_maskz_cvtps_pd (__mmask8 __U, __m128 __A) { |
| 2367 | return (__m256d) __builtin_ia32_cvtps2pd256_mask ((__v4sf) __A, |
| 2368 | (__v4df) |
| 2369 | _mm256_setzero_pd (), |
| 2370 | (__mmask8) __U); |
| 2371 | } |
| 2372 | |
| 2373 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2374 | _mm_cvtps_epu32 (__m128 __A) { |
| 2375 | return (__m128i) __builtin_ia32_cvtps2udq128_mask ((__v4sf) __A, |
| 2376 | (__v4si) |
| 2377 | _mm_setzero_si128 (), |
| 2378 | (__mmask8) -1); |
| 2379 | } |
| 2380 | |
| 2381 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2382 | _mm_mask_cvtps_epu32 (__m128i __W, __mmask8 __U, __m128 __A) { |
| 2383 | return (__m128i) __builtin_ia32_cvtps2udq128_mask ((__v4sf) __A, |
| 2384 | (__v4si) __W, |
| 2385 | (__mmask8) __U); |
| 2386 | } |
| 2387 | |
| 2388 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2389 | _mm_maskz_cvtps_epu32 (__mmask8 __U, __m128 __A) { |
| 2390 | return (__m128i) __builtin_ia32_cvtps2udq128_mask ((__v4sf) __A, |
| 2391 | (__v4si) |
| 2392 | _mm_setzero_si128 (), |
| 2393 | (__mmask8) __U); |
| 2394 | } |
| 2395 | |
| 2396 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2397 | _mm256_cvtps_epu32 (__m256 __A) { |
| 2398 | return (__m256i) __builtin_ia32_cvtps2udq256_mask ((__v8sf) __A, |
| 2399 | (__v8si) |
| 2400 | _mm256_setzero_si256 (), |
| 2401 | (__mmask8) -1); |
| 2402 | } |
| 2403 | |
| 2404 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2405 | _mm256_mask_cvtps_epu32 (__m256i __W, __mmask8 __U, __m256 __A) { |
| 2406 | return (__m256i) __builtin_ia32_cvtps2udq256_mask ((__v8sf) __A, |
| 2407 | (__v8si) __W, |
| 2408 | (__mmask8) __U); |
| 2409 | } |
| 2410 | |
| 2411 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2412 | _mm256_maskz_cvtps_epu32 (__mmask8 __U, __m256 __A) { |
| 2413 | return (__m256i) __builtin_ia32_cvtps2udq256_mask ((__v8sf) __A, |
| 2414 | (__v8si) |
| 2415 | _mm256_setzero_si256 (), |
| 2416 | (__mmask8) __U); |
| 2417 | } |
| 2418 | |
| 2419 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2420 | _mm_mask_cvttpd_epi32 (__m128i __W, __mmask8 __U, __m128d __A) { |
| 2421 | return (__m128i) __builtin_ia32_cvttpd2dq128_mask ((__v2df) __A, |
| 2422 | (__v4si) __W, |
| 2423 | (__mmask8) __U); |
| 2424 | } |
| 2425 | |
| 2426 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2427 | _mm_maskz_cvttpd_epi32 (__mmask8 __U, __m128d __A) { |
| 2428 | return (__m128i) __builtin_ia32_cvttpd2dq128_mask ((__v2df) __A, |
| 2429 | (__v4si) |
| 2430 | _mm_setzero_si128 (), |
| 2431 | (__mmask8) __U); |
| 2432 | } |
| 2433 | |
| 2434 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2435 | _mm256_mask_cvttpd_epi32 (__m128i __W, __mmask8 __U, __m256d __A) { |
| 2436 | return (__m128i) __builtin_ia32_cvttpd2dq256_mask ((__v4df) __A, |
| 2437 | (__v4si) __W, |
| 2438 | (__mmask8) __U); |
| 2439 | } |
| 2440 | |
| 2441 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2442 | _mm256_maskz_cvttpd_epi32 (__mmask8 __U, __m256d __A) { |
| 2443 | return (__m128i) __builtin_ia32_cvttpd2dq256_mask ((__v4df) __A, |
| 2444 | (__v4si) |
| 2445 | _mm_setzero_si128 (), |
| 2446 | (__mmask8) __U); |
| 2447 | } |
| 2448 | |
| 2449 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2450 | _mm_cvttpd_epu32 (__m128d __A) { |
| 2451 | return (__m128i) __builtin_ia32_cvttpd2udq128_mask ((__v2df) __A, |
| 2452 | (__v4si) |
| 2453 | _mm_setzero_si128 (), |
| 2454 | (__mmask8) -1); |
| 2455 | } |
| 2456 | |
| 2457 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2458 | _mm_mask_cvttpd_epu32 (__m128i __W, __mmask8 __U, __m128d __A) { |
| 2459 | return (__m128i) __builtin_ia32_cvttpd2udq128_mask ((__v2df) __A, |
| 2460 | (__v4si) __W, |
| 2461 | (__mmask8) __U); |
| 2462 | } |
| 2463 | |
| 2464 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2465 | _mm_maskz_cvttpd_epu32 (__mmask8 __U, __m128d __A) { |
| 2466 | return (__m128i) __builtin_ia32_cvttpd2udq128_mask ((__v2df) __A, |
| 2467 | (__v4si) |
| 2468 | _mm_setzero_si128 (), |
| 2469 | (__mmask8) __U); |
| 2470 | } |
| 2471 | |
| 2472 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2473 | _mm256_cvttpd_epu32 (__m256d __A) { |
| 2474 | return (__m128i) __builtin_ia32_cvttpd2udq256_mask ((__v4df) __A, |
| 2475 | (__v4si) |
| 2476 | _mm_setzero_si128 (), |
| 2477 | (__mmask8) -1); |
| 2478 | } |
| 2479 | |
| 2480 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2481 | _mm256_mask_cvttpd_epu32 (__m128i __W, __mmask8 __U, __m256d __A) { |
| 2482 | return (__m128i) __builtin_ia32_cvttpd2udq256_mask ((__v4df) __A, |
| 2483 | (__v4si) __W, |
| 2484 | (__mmask8) __U); |
| 2485 | } |
| 2486 | |
| 2487 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2488 | _mm256_maskz_cvttpd_epu32 (__mmask8 __U, __m256d __A) { |
| 2489 | return (__m128i) __builtin_ia32_cvttpd2udq256_mask ((__v4df) __A, |
| 2490 | (__v4si) |
| 2491 | _mm_setzero_si128 (), |
| 2492 | (__mmask8) __U); |
| 2493 | } |
| 2494 | |
| 2495 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2496 | _mm_mask_cvttps_epi32 (__m128i __W, __mmask8 __U, __m128 __A) { |
| 2497 | return (__m128i) __builtin_ia32_cvttps2dq128_mask ((__v4sf) __A, |
| 2498 | (__v4si) __W, |
| 2499 | (__mmask8) __U); |
| 2500 | } |
| 2501 | |
| 2502 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2503 | _mm_maskz_cvttps_epi32 (__mmask8 __U, __m128 __A) { |
| 2504 | return (__m128i) __builtin_ia32_cvttps2dq128_mask ((__v4sf) __A, |
| 2505 | (__v4si) |
| 2506 | _mm_setzero_si128 (), |
| 2507 | (__mmask8) __U); |
| 2508 | } |
| 2509 | |
| 2510 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2511 | _mm256_mask_cvttps_epi32 (__m256i __W, __mmask8 __U, __m256 __A) { |
| 2512 | return (__m256i) __builtin_ia32_cvttps2dq256_mask ((__v8sf) __A, |
| 2513 | (__v8si) __W, |
| 2514 | (__mmask8) __U); |
| 2515 | } |
| 2516 | |
| 2517 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2518 | _mm256_maskz_cvttps_epi32 (__mmask8 __U, __m256 __A) { |
| 2519 | return (__m256i) __builtin_ia32_cvttps2dq256_mask ((__v8sf) __A, |
| 2520 | (__v8si) |
| 2521 | _mm256_setzero_si256 (), |
| 2522 | (__mmask8) __U); |
| 2523 | } |
| 2524 | |
| 2525 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2526 | _mm_cvttps_epu32 (__m128 __A) { |
| 2527 | return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A, |
| 2528 | (__v4si) |
| 2529 | _mm_setzero_si128 (), |
| 2530 | (__mmask8) -1); |
| 2531 | } |
| 2532 | |
| 2533 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2534 | _mm_mask_cvttps_epu32 (__m128i __W, __mmask8 __U, __m128 __A) { |
| 2535 | return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A, |
| 2536 | (__v4si) __W, |
| 2537 | (__mmask8) __U); |
| 2538 | } |
| 2539 | |
| 2540 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2541 | _mm_maskz_cvttps_epu32 (__mmask8 __U, __m128 __A) { |
| 2542 | return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A, |
| 2543 | (__v4si) |
| 2544 | _mm_setzero_si128 (), |
| 2545 | (__mmask8) __U); |
| 2546 | } |
| 2547 | |
| 2548 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2549 | _mm256_cvttps_epu32 (__m256 __A) { |
| 2550 | return (__m256i) __builtin_ia32_cvttps2udq256_mask ((__v8sf) __A, |
| 2551 | (__v8si) |
| 2552 | _mm256_setzero_si256 (), |
| 2553 | (__mmask8) -1); |
| 2554 | } |
| 2555 | |
| 2556 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2557 | _mm256_mask_cvttps_epu32 (__m256i __W, __mmask8 __U, __m256 __A) { |
| 2558 | return (__m256i) __builtin_ia32_cvttps2udq256_mask ((__v8sf) __A, |
| 2559 | (__v8si) __W, |
| 2560 | (__mmask8) __U); |
| 2561 | } |
| 2562 | |
| 2563 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2564 | _mm256_maskz_cvttps_epu32 (__mmask8 __U, __m256 __A) { |
| 2565 | return (__m256i) __builtin_ia32_cvttps2udq256_mask ((__v8sf) __A, |
| 2566 | (__v8si) |
| 2567 | _mm256_setzero_si256 (), |
| 2568 | (__mmask8) __U); |
| 2569 | } |
| 2570 | |
| 2571 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2572 | _mm_cvtepu32_pd (__m128i __A) { |
| 2573 | return (__m128d) __builtin_ia32_cvtudq2pd128_mask ((__v4si) __A, |
| 2574 | (__v2df) |
| 2575 | _mm_setzero_pd (), |
| 2576 | (__mmask8) -1); |
| 2577 | } |
| 2578 | |
| 2579 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2580 | _mm_mask_cvtepu32_pd (__m128d __W, __mmask8 __U, __m128i __A) { |
| 2581 | return (__m128d) __builtin_ia32_cvtudq2pd128_mask ((__v4si) __A, |
| 2582 | (__v2df) __W, |
| 2583 | (__mmask8) __U); |
| 2584 | } |
| 2585 | |
| 2586 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2587 | _mm_maskz_cvtepu32_pd (__mmask8 __U, __m128i __A) { |
| 2588 | return (__m128d) __builtin_ia32_cvtudq2pd128_mask ((__v4si) __A, |
| 2589 | (__v2df) |
| 2590 | _mm_setzero_pd (), |
| 2591 | (__mmask8) __U); |
| 2592 | } |
| 2593 | |
| 2594 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2595 | _mm256_cvtepu32_pd (__m128i __A) { |
| 2596 | return (__m256d) __builtin_ia32_cvtudq2pd256_mask ((__v4si) __A, |
| 2597 | (__v4df) |
| 2598 | _mm256_setzero_pd (), |
| 2599 | (__mmask8) -1); |
| 2600 | } |
| 2601 | |
| 2602 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2603 | _mm256_mask_cvtepu32_pd (__m256d __W, __mmask8 __U, __m128i __A) { |
| 2604 | return (__m256d) __builtin_ia32_cvtudq2pd256_mask ((__v4si) __A, |
| 2605 | (__v4df) __W, |
| 2606 | (__mmask8) __U); |
| 2607 | } |
| 2608 | |
| 2609 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2610 | _mm256_maskz_cvtepu32_pd (__mmask8 __U, __m128i __A) { |
| 2611 | return (__m256d) __builtin_ia32_cvtudq2pd256_mask ((__v4si) __A, |
| 2612 | (__v4df) |
| 2613 | _mm256_setzero_pd (), |
| 2614 | (__mmask8) __U); |
| 2615 | } |
| 2616 | |
| 2617 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2618 | _mm_cvtepu32_ps (__m128i __A) { |
| 2619 | return (__m128) __builtin_ia32_cvtudq2ps128_mask ((__v4si) __A, |
| 2620 | (__v4sf) |
| 2621 | _mm_setzero_ps (), |
| 2622 | (__mmask8) -1); |
| 2623 | } |
| 2624 | |
| 2625 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2626 | _mm_mask_cvtepu32_ps (__m128 __W, __mmask8 __U, __m128i __A) { |
| 2627 | return (__m128) __builtin_ia32_cvtudq2ps128_mask ((__v4si) __A, |
| 2628 | (__v4sf) __W, |
| 2629 | (__mmask8) __U); |
| 2630 | } |
| 2631 | |
| 2632 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2633 | _mm_maskz_cvtepu32_ps (__mmask8 __U, __m128i __A) { |
| 2634 | return (__m128) __builtin_ia32_cvtudq2ps128_mask ((__v4si) __A, |
| 2635 | (__v4sf) |
| 2636 | _mm_setzero_ps (), |
| 2637 | (__mmask8) __U); |
| 2638 | } |
| 2639 | |
| 2640 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2641 | _mm256_cvtepu32_ps (__m256i __A) { |
| 2642 | return (__m256) __builtin_ia32_cvtudq2ps256_mask ((__v8si) __A, |
| 2643 | (__v8sf) |
| 2644 | _mm256_setzero_ps (), |
| 2645 | (__mmask8) -1); |
| 2646 | } |
| 2647 | |
| 2648 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2649 | _mm256_mask_cvtepu32_ps (__m256 __W, __mmask8 __U, __m256i __A) { |
| 2650 | return (__m256) __builtin_ia32_cvtudq2ps256_mask ((__v8si) __A, |
| 2651 | (__v8sf) __W, |
| 2652 | (__mmask8) __U); |
| 2653 | } |
| 2654 | |
| 2655 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2656 | _mm256_maskz_cvtepu32_ps (__mmask8 __U, __m256i __A) { |
| 2657 | return (__m256) __builtin_ia32_cvtudq2ps256_mask ((__v8si) __A, |
| 2658 | (__v8sf) |
| 2659 | _mm256_setzero_ps (), |
| 2660 | (__mmask8) __U); |
| 2661 | } |
| 2662 | |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2663 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 2664 | _mm_mask_div_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 2665 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 2666 | (__v2df)_mm_div_pd(__A, __B), |
| 2667 | (__v2df)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2668 | } |
| 2669 | |
| 2670 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 2671 | _mm_maskz_div_pd(__mmask8 __U, __m128d __A, __m128d __B) { |
| 2672 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 2673 | (__v2df)_mm_div_pd(__A, __B), |
| 2674 | (__v2df)_mm_setzero_pd()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2675 | } |
| 2676 | |
| 2677 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 2678 | _mm256_mask_div_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) { |
| 2679 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 2680 | (__v4df)_mm256_div_pd(__A, __B), |
| 2681 | (__v4df)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2682 | } |
| 2683 | |
| 2684 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 2685 | _mm256_maskz_div_pd(__mmask8 __U, __m256d __A, __m256d __B) { |
| 2686 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 2687 | (__v4df)_mm256_div_pd(__A, __B), |
| 2688 | (__v4df)_mm256_setzero_pd()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2689 | } |
| 2690 | |
| 2691 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 2692 | _mm_mask_div_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 2693 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 2694 | (__v4sf)_mm_div_ps(__A, __B), |
| 2695 | (__v4sf)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2696 | } |
| 2697 | |
| 2698 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 2699 | _mm_maskz_div_ps(__mmask8 __U, __m128 __A, __m128 __B) { |
| 2700 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 2701 | (__v4sf)_mm_div_ps(__A, __B), |
| 2702 | (__v4sf)_mm_setzero_ps()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2703 | } |
| 2704 | |
| 2705 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 2706 | _mm256_mask_div_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 2707 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 2708 | (__v8sf)_mm256_div_ps(__A, __B), |
| 2709 | (__v8sf)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2710 | } |
| 2711 | |
| 2712 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 2713 | _mm256_maskz_div_ps(__mmask8 __U, __m256 __A, __m256 __B) { |
| 2714 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 2715 | (__v8sf)_mm256_div_ps(__A, __B), |
| 2716 | (__v8sf)_mm256_setzero_ps()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2717 | } |
| 2718 | |
| 2719 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2720 | _mm_mask_expand_pd (__m128d __W, __mmask8 __U, __m128d __A) { |
| 2721 | return (__m128d) __builtin_ia32_expanddf128_mask ((__v2df) __A, |
| 2722 | (__v2df) __W, |
| 2723 | (__mmask8) __U); |
| 2724 | } |
| 2725 | |
| 2726 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2727 | _mm_maskz_expand_pd (__mmask8 __U, __m128d __A) { |
| 2728 | return (__m128d) __builtin_ia32_expanddf128_mask ((__v2df) __A, |
| 2729 | (__v2df) |
| 2730 | _mm_setzero_pd (), |
| 2731 | (__mmask8) __U); |
| 2732 | } |
| 2733 | |
| 2734 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2735 | _mm256_mask_expand_pd (__m256d __W, __mmask8 __U, __m256d __A) { |
| 2736 | return (__m256d) __builtin_ia32_expanddf256_mask ((__v4df) __A, |
| 2737 | (__v4df) __W, |
| 2738 | (__mmask8) __U); |
| 2739 | } |
| 2740 | |
| 2741 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2742 | _mm256_maskz_expand_pd (__mmask8 __U, __m256d __A) { |
| 2743 | return (__m256d) __builtin_ia32_expanddf256_mask ((__v4df) __A, |
| 2744 | (__v4df) |
| 2745 | _mm256_setzero_pd (), |
| 2746 | (__mmask8) __U); |
| 2747 | } |
| 2748 | |
| 2749 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2750 | _mm_mask_expand_epi64 (__m128i __W, __mmask8 __U, __m128i __A) { |
| 2751 | return (__m128i) __builtin_ia32_expanddi128_mask ((__v2di) __A, |
| 2752 | (__v2di) __W, |
| 2753 | (__mmask8) __U); |
| 2754 | } |
| 2755 | |
| 2756 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2757 | _mm_maskz_expand_epi64 (__mmask8 __U, __m128i __A) { |
| 2758 | return (__m128i) __builtin_ia32_expanddi128_mask ((__v2di) __A, |
| 2759 | (__v2di) |
| 2760 | _mm_setzero_si128 (), |
| 2761 | (__mmask8) __U); |
| 2762 | } |
| 2763 | |
| 2764 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2765 | _mm256_mask_expand_epi64 (__m256i __W, __mmask8 __U, __m256i __A) { |
| 2766 | return (__m256i) __builtin_ia32_expanddi256_mask ((__v4di) __A, |
| 2767 | (__v4di) __W, |
| 2768 | (__mmask8) __U); |
| 2769 | } |
| 2770 | |
| 2771 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2772 | _mm256_maskz_expand_epi64 (__mmask8 __U, __m256i __A) { |
| 2773 | return (__m256i) __builtin_ia32_expanddi256_mask ((__v4di) __A, |
| 2774 | (__v4di) |
| 2775 | _mm256_setzero_si256 (), |
| 2776 | (__mmask8) __U); |
| 2777 | } |
| 2778 | |
| 2779 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2780 | _mm_mask_expandloadu_pd (__m128d __W, __mmask8 __U, void const *__P) { |
| 2781 | return (__m128d) __builtin_ia32_expandloaddf128_mask ((__v2df *) __P, |
| 2782 | (__v2df) __W, |
| 2783 | (__mmask8) |
| 2784 | __U); |
| 2785 | } |
| 2786 | |
| 2787 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2788 | _mm_maskz_expandloadu_pd (__mmask8 __U, void const *__P) { |
| 2789 | return (__m128d) __builtin_ia32_expandloaddf128_mask ((__v2df *) __P, |
| 2790 | (__v2df) |
| 2791 | _mm_setzero_pd (), |
| 2792 | (__mmask8) |
| 2793 | __U); |
| 2794 | } |
| 2795 | |
| 2796 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2797 | _mm256_mask_expandloadu_pd (__m256d __W, __mmask8 __U, void const *__P) { |
| 2798 | return (__m256d) __builtin_ia32_expandloaddf256_mask ((__v4df *) __P, |
| 2799 | (__v4df) __W, |
| 2800 | (__mmask8) |
| 2801 | __U); |
| 2802 | } |
| 2803 | |
| 2804 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2805 | _mm256_maskz_expandloadu_pd (__mmask8 __U, void const *__P) { |
| 2806 | return (__m256d) __builtin_ia32_expandloaddf256_mask ((__v4df *) __P, |
| 2807 | (__v4df) |
| 2808 | _mm256_setzero_pd (), |
| 2809 | (__mmask8) |
| 2810 | __U); |
| 2811 | } |
| 2812 | |
| 2813 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2814 | _mm_mask_expandloadu_epi64 (__m128i __W, __mmask8 __U, void const *__P) { |
| 2815 | return (__m128i) __builtin_ia32_expandloaddi128_mask ((__v2di *) __P, |
| 2816 | (__v2di) __W, |
| 2817 | (__mmask8) |
| 2818 | __U); |
| 2819 | } |
| 2820 | |
| 2821 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2822 | _mm_maskz_expandloadu_epi64 (__mmask8 __U, void const *__P) { |
| 2823 | return (__m128i) __builtin_ia32_expandloaddi128_mask ((__v2di *) __P, |
| 2824 | (__v2di) |
| 2825 | _mm_setzero_si128 (), |
| 2826 | (__mmask8) |
| 2827 | __U); |
| 2828 | } |
| 2829 | |
| 2830 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2831 | _mm256_mask_expandloadu_epi64 (__m256i __W, __mmask8 __U, |
| 2832 | void const *__P) { |
| 2833 | return (__m256i) __builtin_ia32_expandloaddi256_mask ((__v4di *) __P, |
| 2834 | (__v4di) __W, |
| 2835 | (__mmask8) |
| 2836 | __U); |
| 2837 | } |
| 2838 | |
| 2839 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2840 | _mm256_maskz_expandloadu_epi64 (__mmask8 __U, void const *__P) { |
| 2841 | return (__m256i) __builtin_ia32_expandloaddi256_mask ((__v4di *) __P, |
| 2842 | (__v4di) |
| 2843 | _mm256_setzero_si256 (), |
| 2844 | (__mmask8) |
| 2845 | __U); |
| 2846 | } |
| 2847 | |
| 2848 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2849 | _mm_mask_expandloadu_ps (__m128 __W, __mmask8 __U, void const *__P) { |
| 2850 | return (__m128) __builtin_ia32_expandloadsf128_mask ((__v4sf *) __P, |
| 2851 | (__v4sf) __W, |
| 2852 | (__mmask8) __U); |
| 2853 | } |
| 2854 | |
| 2855 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2856 | _mm_maskz_expandloadu_ps (__mmask8 __U, void const *__P) { |
| 2857 | return (__m128) __builtin_ia32_expandloadsf128_mask ((__v4sf *) __P, |
| 2858 | (__v4sf) |
| 2859 | _mm_setzero_ps (), |
| 2860 | (__mmask8) |
| 2861 | __U); |
| 2862 | } |
| 2863 | |
| 2864 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2865 | _mm256_mask_expandloadu_ps (__m256 __W, __mmask8 __U, void const *__P) { |
| 2866 | return (__m256) __builtin_ia32_expandloadsf256_mask ((__v8sf *) __P, |
| 2867 | (__v8sf) __W, |
| 2868 | (__mmask8) __U); |
| 2869 | } |
| 2870 | |
| 2871 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2872 | _mm256_maskz_expandloadu_ps (__mmask8 __U, void const *__P) { |
| 2873 | return (__m256) __builtin_ia32_expandloadsf256_mask ((__v8sf *) __P, |
| 2874 | (__v8sf) |
| 2875 | _mm256_setzero_ps (), |
| 2876 | (__mmask8) |
| 2877 | __U); |
| 2878 | } |
| 2879 | |
| 2880 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2881 | _mm_mask_expandloadu_epi32 (__m128i __W, __mmask8 __U, void const *__P) { |
| 2882 | return (__m128i) __builtin_ia32_expandloadsi128_mask ((__v4si *) __P, |
| 2883 | (__v4si) __W, |
| 2884 | (__mmask8) |
| 2885 | __U); |
| 2886 | } |
| 2887 | |
| 2888 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2889 | _mm_maskz_expandloadu_epi32 (__mmask8 __U, void const *__P) { |
| 2890 | return (__m128i) __builtin_ia32_expandloadsi128_mask ((__v4si *) __P, |
| 2891 | (__v4si) |
| 2892 | _mm_setzero_si128 (), |
| 2893 | (__mmask8) __U); |
| 2894 | } |
| 2895 | |
| 2896 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2897 | _mm256_mask_expandloadu_epi32 (__m256i __W, __mmask8 __U, |
| 2898 | void const *__P) { |
| 2899 | return (__m256i) __builtin_ia32_expandloadsi256_mask ((__v8si *) __P, |
| 2900 | (__v8si) __W, |
| 2901 | (__mmask8) |
| 2902 | __U); |
| 2903 | } |
| 2904 | |
| 2905 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2906 | _mm256_maskz_expandloadu_epi32 (__mmask8 __U, void const *__P) { |
| 2907 | return (__m256i) __builtin_ia32_expandloadsi256_mask ((__v8si *) __P, |
| 2908 | (__v8si) |
| 2909 | _mm256_setzero_si256 (), |
| 2910 | (__mmask8) |
| 2911 | __U); |
| 2912 | } |
| 2913 | |
| 2914 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2915 | _mm_mask_expand_ps (__m128 __W, __mmask8 __U, __m128 __A) { |
| 2916 | return (__m128) __builtin_ia32_expandsf128_mask ((__v4sf) __A, |
| 2917 | (__v4sf) __W, |
| 2918 | (__mmask8) __U); |
| 2919 | } |
| 2920 | |
| 2921 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2922 | _mm_maskz_expand_ps (__mmask8 __U, __m128 __A) { |
| 2923 | return (__m128) __builtin_ia32_expandsf128_mask ((__v4sf) __A, |
| 2924 | (__v4sf) |
| 2925 | _mm_setzero_ps (), |
| 2926 | (__mmask8) __U); |
| 2927 | } |
| 2928 | |
| 2929 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2930 | _mm256_mask_expand_ps (__m256 __W, __mmask8 __U, __m256 __A) { |
| 2931 | return (__m256) __builtin_ia32_expandsf256_mask ((__v8sf) __A, |
| 2932 | (__v8sf) __W, |
| 2933 | (__mmask8) __U); |
| 2934 | } |
| 2935 | |
| 2936 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2937 | _mm256_maskz_expand_ps (__mmask8 __U, __m256 __A) { |
| 2938 | return (__m256) __builtin_ia32_expandsf256_mask ((__v8sf) __A, |
| 2939 | (__v8sf) |
| 2940 | _mm256_setzero_ps (), |
| 2941 | (__mmask8) __U); |
| 2942 | } |
| 2943 | |
| 2944 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2945 | _mm_mask_expand_epi32 (__m128i __W, __mmask8 __U, __m128i __A) { |
| 2946 | return (__m128i) __builtin_ia32_expandsi128_mask ((__v4si) __A, |
| 2947 | (__v4si) __W, |
| 2948 | (__mmask8) __U); |
| 2949 | } |
| 2950 | |
| 2951 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2952 | _mm_maskz_expand_epi32 (__mmask8 __U, __m128i __A) { |
| 2953 | return (__m128i) __builtin_ia32_expandsi128_mask ((__v4si) __A, |
| 2954 | (__v4si) |
| 2955 | _mm_setzero_si128 (), |
| 2956 | (__mmask8) __U); |
| 2957 | } |
| 2958 | |
| 2959 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2960 | _mm256_mask_expand_epi32 (__m256i __W, __mmask8 __U, __m256i __A) { |
| 2961 | return (__m256i) __builtin_ia32_expandsi256_mask ((__v8si) __A, |
| 2962 | (__v8si) __W, |
| 2963 | (__mmask8) __U); |
| 2964 | } |
| 2965 | |
| 2966 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2967 | _mm256_maskz_expand_epi32 (__mmask8 __U, __m256i __A) { |
| 2968 | return (__m256i) __builtin_ia32_expandsi256_mask ((__v8si) __A, |
| 2969 | (__v8si) |
| 2970 | _mm256_setzero_si256 (), |
| 2971 | (__mmask8) __U); |
| 2972 | } |
| 2973 | |
| 2974 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2975 | _mm_getexp_pd (__m128d __A) { |
| 2976 | return (__m128d) __builtin_ia32_getexppd128_mask ((__v2df) __A, |
| 2977 | (__v2df) |
| 2978 | _mm_setzero_pd (), |
| 2979 | (__mmask8) -1); |
| 2980 | } |
| 2981 | |
| 2982 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2983 | _mm_mask_getexp_pd (__m128d __W, __mmask8 __U, __m128d __A) { |
| 2984 | return (__m128d) __builtin_ia32_getexppd128_mask ((__v2df) __A, |
| 2985 | (__v2df) __W, |
| 2986 | (__mmask8) __U); |
| 2987 | } |
| 2988 | |
| 2989 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2990 | _mm_maskz_getexp_pd (__mmask8 __U, __m128d __A) { |
| 2991 | return (__m128d) __builtin_ia32_getexppd128_mask ((__v2df) __A, |
| 2992 | (__v2df) |
| 2993 | _mm_setzero_pd (), |
| 2994 | (__mmask8) __U); |
| 2995 | } |
| 2996 | |
| 2997 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2998 | _mm256_getexp_pd (__m256d __A) { |
| 2999 | return (__m256d) __builtin_ia32_getexppd256_mask ((__v4df) __A, |
| 3000 | (__v4df) |
| 3001 | _mm256_setzero_pd (), |
| 3002 | (__mmask8) -1); |
| 3003 | } |
| 3004 | |
| 3005 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3006 | _mm256_mask_getexp_pd (__m256d __W, __mmask8 __U, __m256d __A) { |
| 3007 | return (__m256d) __builtin_ia32_getexppd256_mask ((__v4df) __A, |
| 3008 | (__v4df) __W, |
| 3009 | (__mmask8) __U); |
| 3010 | } |
| 3011 | |
| 3012 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3013 | _mm256_maskz_getexp_pd (__mmask8 __U, __m256d __A) { |
| 3014 | return (__m256d) __builtin_ia32_getexppd256_mask ((__v4df) __A, |
| 3015 | (__v4df) |
| 3016 | _mm256_setzero_pd (), |
| 3017 | (__mmask8) __U); |
| 3018 | } |
| 3019 | |
| 3020 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3021 | _mm_getexp_ps (__m128 __A) { |
| 3022 | return (__m128) __builtin_ia32_getexpps128_mask ((__v4sf) __A, |
| 3023 | (__v4sf) |
| 3024 | _mm_setzero_ps (), |
| 3025 | (__mmask8) -1); |
| 3026 | } |
| 3027 | |
| 3028 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3029 | _mm_mask_getexp_ps (__m128 __W, __mmask8 __U, __m128 __A) { |
| 3030 | return (__m128) __builtin_ia32_getexpps128_mask ((__v4sf) __A, |
| 3031 | (__v4sf) __W, |
| 3032 | (__mmask8) __U); |
| 3033 | } |
| 3034 | |
| 3035 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3036 | _mm_maskz_getexp_ps (__mmask8 __U, __m128 __A) { |
| 3037 | return (__m128) __builtin_ia32_getexpps128_mask ((__v4sf) __A, |
| 3038 | (__v4sf) |
| 3039 | _mm_setzero_ps (), |
| 3040 | (__mmask8) __U); |
| 3041 | } |
| 3042 | |
| 3043 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3044 | _mm256_getexp_ps (__m256 __A) { |
| 3045 | return (__m256) __builtin_ia32_getexpps256_mask ((__v8sf) __A, |
| 3046 | (__v8sf) |
| 3047 | _mm256_setzero_ps (), |
| 3048 | (__mmask8) -1); |
| 3049 | } |
| 3050 | |
| 3051 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3052 | _mm256_mask_getexp_ps (__m256 __W, __mmask8 __U, __m256 __A) { |
| 3053 | return (__m256) __builtin_ia32_getexpps256_mask ((__v8sf) __A, |
| 3054 | (__v8sf) __W, |
| 3055 | (__mmask8) __U); |
| 3056 | } |
| 3057 | |
| 3058 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3059 | _mm256_maskz_getexp_ps (__mmask8 __U, __m256 __A) { |
| 3060 | return (__m256) __builtin_ia32_getexpps256_mask ((__v8sf) __A, |
| 3061 | (__v8sf) |
| 3062 | _mm256_setzero_ps (), |
| 3063 | (__mmask8) __U); |
| 3064 | } |
| 3065 | |
| 3066 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3067 | _mm_mask_max_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 3068 | return (__m128d) __builtin_ia32_maxpd_mask ((__v2df) __A, |
| 3069 | (__v2df) __B, |
| 3070 | (__v2df) __W, |
| 3071 | (__mmask8) __U); |
| 3072 | } |
| 3073 | |
| 3074 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3075 | _mm_maskz_max_pd (__mmask8 __U, __m128d __A, __m128d __B) { |
| 3076 | return (__m128d) __builtin_ia32_maxpd_mask ((__v2df) __A, |
| 3077 | (__v2df) __B, |
| 3078 | (__v2df) |
| 3079 | _mm_setzero_pd (), |
| 3080 | (__mmask8) __U); |
| 3081 | } |
| 3082 | |
| 3083 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3084 | _mm256_mask_max_pd (__m256d __W, __mmask8 __U, __m256d __A, |
| 3085 | __m256d __B) { |
| 3086 | return (__m256d) __builtin_ia32_maxpd256_mask ((__v4df) __A, |
| 3087 | (__v4df) __B, |
| 3088 | (__v4df) __W, |
| 3089 | (__mmask8) __U); |
| 3090 | } |
| 3091 | |
| 3092 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3093 | _mm256_maskz_max_pd (__mmask8 __U, __m256d __A, __m256d __B) { |
| 3094 | return (__m256d) __builtin_ia32_maxpd256_mask ((__v4df) __A, |
| 3095 | (__v4df) __B, |
| 3096 | (__v4df) |
| 3097 | _mm256_setzero_pd (), |
| 3098 | (__mmask8) __U); |
| 3099 | } |
| 3100 | |
| 3101 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3102 | _mm_mask_max_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 3103 | return (__m128) __builtin_ia32_maxps_mask ((__v4sf) __A, |
| 3104 | (__v4sf) __B, |
| 3105 | (__v4sf) __W, |
| 3106 | (__mmask8) __U); |
| 3107 | } |
| 3108 | |
| 3109 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3110 | _mm_maskz_max_ps (__mmask8 __U, __m128 __A, __m128 __B) { |
| 3111 | return (__m128) __builtin_ia32_maxps_mask ((__v4sf) __A, |
| 3112 | (__v4sf) __B, |
| 3113 | (__v4sf) |
| 3114 | _mm_setzero_ps (), |
| 3115 | (__mmask8) __U); |
| 3116 | } |
| 3117 | |
| 3118 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3119 | _mm256_mask_max_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 3120 | return (__m256) __builtin_ia32_maxps256_mask ((__v8sf) __A, |
| 3121 | (__v8sf) __B, |
| 3122 | (__v8sf) __W, |
| 3123 | (__mmask8) __U); |
| 3124 | } |
| 3125 | |
| 3126 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3127 | _mm256_maskz_max_ps (__mmask8 __U, __m256 __A, __m256 __B) { |
| 3128 | return (__m256) __builtin_ia32_maxps256_mask ((__v8sf) __A, |
| 3129 | (__v8sf) __B, |
| 3130 | (__v8sf) |
| 3131 | _mm256_setzero_ps (), |
| 3132 | (__mmask8) __U); |
| 3133 | } |
| 3134 | |
| 3135 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3136 | _mm_mask_min_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 3137 | return (__m128d) __builtin_ia32_minpd_mask ((__v2df) __A, |
| 3138 | (__v2df) __B, |
| 3139 | (__v2df) __W, |
| 3140 | (__mmask8) __U); |
| 3141 | } |
| 3142 | |
| 3143 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3144 | _mm_maskz_min_pd (__mmask8 __U, __m128d __A, __m128d __B) { |
| 3145 | return (__m128d) __builtin_ia32_minpd_mask ((__v2df) __A, |
| 3146 | (__v2df) __B, |
| 3147 | (__v2df) |
| 3148 | _mm_setzero_pd (), |
| 3149 | (__mmask8) __U); |
| 3150 | } |
| 3151 | |
| 3152 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3153 | _mm256_mask_min_pd (__m256d __W, __mmask8 __U, __m256d __A, |
| 3154 | __m256d __B) { |
| 3155 | return (__m256d) __builtin_ia32_minpd256_mask ((__v4df) __A, |
| 3156 | (__v4df) __B, |
| 3157 | (__v4df) __W, |
| 3158 | (__mmask8) __U); |
| 3159 | } |
| 3160 | |
| 3161 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3162 | _mm256_maskz_min_pd (__mmask8 __U, __m256d __A, __m256d __B) { |
| 3163 | return (__m256d) __builtin_ia32_minpd256_mask ((__v4df) __A, |
| 3164 | (__v4df) __B, |
| 3165 | (__v4df) |
| 3166 | _mm256_setzero_pd (), |
| 3167 | (__mmask8) __U); |
| 3168 | } |
| 3169 | |
| 3170 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3171 | _mm_mask_min_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 3172 | return (__m128) __builtin_ia32_minps_mask ((__v4sf) __A, |
| 3173 | (__v4sf) __B, |
| 3174 | (__v4sf) __W, |
| 3175 | (__mmask8) __U); |
| 3176 | } |
| 3177 | |
| 3178 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3179 | _mm_maskz_min_ps (__mmask8 __U, __m128 __A, __m128 __B) { |
| 3180 | return (__m128) __builtin_ia32_minps_mask ((__v4sf) __A, |
| 3181 | (__v4sf) __B, |
| 3182 | (__v4sf) |
| 3183 | _mm_setzero_ps (), |
| 3184 | (__mmask8) __U); |
| 3185 | } |
| 3186 | |
| 3187 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3188 | _mm256_mask_min_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 3189 | return (__m256) __builtin_ia32_minps256_mask ((__v8sf) __A, |
| 3190 | (__v8sf) __B, |
| 3191 | (__v8sf) __W, |
| 3192 | (__mmask8) __U); |
| 3193 | } |
| 3194 | |
| 3195 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3196 | _mm256_maskz_min_ps (__mmask8 __U, __m256 __A, __m256 __B) { |
| 3197 | return (__m256) __builtin_ia32_minps256_mask ((__v8sf) __A, |
| 3198 | (__v8sf) __B, |
| 3199 | (__v8sf) |
| 3200 | _mm256_setzero_ps (), |
| 3201 | (__mmask8) __U); |
| 3202 | } |
| 3203 | |
| 3204 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 3205 | _mm_mask_mul_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 3206 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 3207 | (__v2df)_mm_mul_pd(__A, __B), |
| 3208 | (__v2df)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 3209 | } |
| 3210 | |
| 3211 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 3212 | _mm_maskz_mul_pd(__mmask8 __U, __m128d __A, __m128d __B) { |
| 3213 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 3214 | (__v2df)_mm_mul_pd(__A, __B), |
| 3215 | (__v2df)_mm_setzero_pd()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 3216 | } |
| 3217 | |
| 3218 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 3219 | _mm256_mask_mul_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) { |
| 3220 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 3221 | (__v4df)_mm256_mul_pd(__A, __B), |
| 3222 | (__v4df)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 3223 | } |
| 3224 | |
| 3225 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 3226 | _mm256_maskz_mul_pd(__mmask8 __U, __m256d __A, __m256d __B) { |
| 3227 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 3228 | (__v4df)_mm256_mul_pd(__A, __B), |
| 3229 | (__v4df)_mm256_setzero_pd()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 3230 | } |
| 3231 | |
| 3232 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 3233 | _mm_mask_mul_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 3234 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 3235 | (__v4sf)_mm_mul_ps(__A, __B), |
| 3236 | (__v4sf)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 3237 | } |
| 3238 | |
| 3239 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 3240 | _mm_maskz_mul_ps(__mmask8 __U, __m128 __A, __m128 __B) { |
| 3241 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 3242 | (__v4sf)_mm_mul_ps(__A, __B), |
| 3243 | (__v4sf)_mm_setzero_ps()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 3244 | } |
| 3245 | |
| 3246 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 3247 | _mm256_mask_mul_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 3248 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 3249 | (__v8sf)_mm256_mul_ps(__A, __B), |
| 3250 | (__v8sf)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 3251 | } |
| 3252 | |
| 3253 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 3254 | _mm256_maskz_mul_ps(__mmask8 __U, __m256 __A, __m256 __B) { |
| 3255 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 3256 | (__v8sf)_mm256_mul_ps(__A, __B), |
| 3257 | (__v8sf)_mm256_setzero_ps()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 3258 | } |
| 3259 | |
| 3260 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3261 | _mm_mask_abs_epi32 (__m128i __W, __mmask8 __U, __m128i __A) { |
| 3262 | return (__m128i) __builtin_ia32_pabsd128_mask ((__v4si) __A, |
| 3263 | (__v4si) __W, |
| 3264 | (__mmask8) __U); |
| 3265 | } |
| 3266 | |
| 3267 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3268 | _mm_maskz_abs_epi32 (__mmask8 __U, __m128i __A) { |
| 3269 | return (__m128i) __builtin_ia32_pabsd128_mask ((__v4si) __A, |
| 3270 | (__v4si) |
| 3271 | _mm_setzero_si128 (), |
| 3272 | (__mmask8) __U); |
| 3273 | } |
| 3274 | |
| 3275 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3276 | _mm256_mask_abs_epi32 (__m256i __W, __mmask8 __U, __m256i __A) { |
| 3277 | return (__m256i) __builtin_ia32_pabsd256_mask ((__v8si) __A, |
| 3278 | (__v8si) __W, |
| 3279 | (__mmask8) __U); |
| 3280 | } |
| 3281 | |
| 3282 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3283 | _mm256_maskz_abs_epi32 (__mmask8 __U, __m256i __A) { |
| 3284 | return (__m256i) __builtin_ia32_pabsd256_mask ((__v8si) __A, |
| 3285 | (__v8si) |
| 3286 | _mm256_setzero_si256 (), |
| 3287 | (__mmask8) __U); |
| 3288 | } |
| 3289 | |
| 3290 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3291 | _mm_abs_epi64 (__m128i __A) { |
| 3292 | return (__m128i) __builtin_ia32_pabsq128_mask ((__v2di) __A, |
| 3293 | (__v2di) |
| 3294 | _mm_setzero_si128 (), |
| 3295 | (__mmask8) -1); |
| 3296 | } |
| 3297 | |
| 3298 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3299 | _mm_mask_abs_epi64 (__m128i __W, __mmask8 __U, __m128i __A) { |
| 3300 | return (__m128i) __builtin_ia32_pabsq128_mask ((__v2di) __A, |
| 3301 | (__v2di) __W, |
| 3302 | (__mmask8) __U); |
| 3303 | } |
| 3304 | |
| 3305 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3306 | _mm_maskz_abs_epi64 (__mmask8 __U, __m128i __A) { |
| 3307 | return (__m128i) __builtin_ia32_pabsq128_mask ((__v2di) __A, |
| 3308 | (__v2di) |
| 3309 | _mm_setzero_si128 (), |
| 3310 | (__mmask8) __U); |
| 3311 | } |
| 3312 | |
| 3313 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3314 | _mm256_abs_epi64 (__m256i __A) { |
| 3315 | return (__m256i) __builtin_ia32_pabsq256_mask ((__v4di) __A, |
| 3316 | (__v4di) |
| 3317 | _mm256_setzero_si256 (), |
| 3318 | (__mmask8) -1); |
| 3319 | } |
| 3320 | |
| 3321 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3322 | _mm256_mask_abs_epi64 (__m256i __W, __mmask8 __U, __m256i __A) { |
| 3323 | return (__m256i) __builtin_ia32_pabsq256_mask ((__v4di) __A, |
| 3324 | (__v4di) __W, |
| 3325 | (__mmask8) __U); |
| 3326 | } |
| 3327 | |
| 3328 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3329 | _mm256_maskz_abs_epi64 (__mmask8 __U, __m256i __A) { |
| 3330 | return (__m256i) __builtin_ia32_pabsq256_mask ((__v4di) __A, |
| 3331 | (__v4di) |
| 3332 | _mm256_setzero_si256 (), |
| 3333 | (__mmask8) __U); |
| 3334 | } |
| 3335 | |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3336 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3337 | _mm_maskz_max_epi32 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3338 | return (__m128i) __builtin_ia32_pmaxsd128_mask ((__v4si) __A, |
| 3339 | (__v4si) __B, |
| 3340 | (__v4si) |
| 3341 | _mm_setzero_si128 (), |
| 3342 | __M); |
| 3343 | } |
| 3344 | |
| 3345 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3346 | _mm_mask_max_epi32 (__m128i __W, __mmask8 __M, __m128i __A, |
| 3347 | __m128i __B) { |
| 3348 | return (__m128i) __builtin_ia32_pmaxsd128_mask ((__v4si) __A, |
| 3349 | (__v4si) __B, |
| 3350 | (__v4si) __W, __M); |
| 3351 | } |
| 3352 | |
| 3353 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3354 | _mm256_maskz_max_epi32 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3355 | return (__m256i) __builtin_ia32_pmaxsd256_mask ((__v8si) __A, |
| 3356 | (__v8si) __B, |
| 3357 | (__v8si) |
| 3358 | _mm256_setzero_si256 (), |
| 3359 | __M); |
| 3360 | } |
| 3361 | |
| 3362 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3363 | _mm256_mask_max_epi32 (__m256i __W, __mmask8 __M, __m256i __A, |
| 3364 | __m256i __B) { |
| 3365 | return (__m256i) __builtin_ia32_pmaxsd256_mask ((__v8si) __A, |
| 3366 | (__v8si) __B, |
| 3367 | (__v8si) __W, __M); |
| 3368 | } |
| 3369 | |
| 3370 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3371 | _mm_maskz_max_epi64 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3372 | return (__m128i) __builtin_ia32_pmaxsq128_mask ((__v2di) __A, |
| 3373 | (__v2di) __B, |
| 3374 | (__v2di) |
| 3375 | _mm_setzero_si128 (), |
| 3376 | __M); |
| 3377 | } |
| 3378 | |
| 3379 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3380 | _mm_mask_max_epi64 (__m128i __W, __mmask8 __M, __m128i __A, |
| 3381 | __m128i __B) { |
| 3382 | return (__m128i) __builtin_ia32_pmaxsq128_mask ((__v2di) __A, |
| 3383 | (__v2di) __B, |
| 3384 | (__v2di) __W, __M); |
| 3385 | } |
| 3386 | |
| 3387 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3388 | _mm_max_epi64 (__m128i __A, __m128i __B) { |
| 3389 | return (__m128i) __builtin_ia32_pmaxsq128_mask ((__v2di) __A, |
| 3390 | (__v2di) __B, |
| 3391 | (__v2di) |
| 3392 | _mm_setzero_si128 (), |
| 3393 | (__mmask8) -1); |
| 3394 | } |
| 3395 | |
| 3396 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3397 | _mm256_maskz_max_epi64 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3398 | return (__m256i) __builtin_ia32_pmaxsq256_mask ((__v4di) __A, |
| 3399 | (__v4di) __B, |
| 3400 | (__v4di) |
| 3401 | _mm256_setzero_si256 (), |
| 3402 | __M); |
| 3403 | } |
| 3404 | |
| 3405 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3406 | _mm256_mask_max_epi64 (__m256i __W, __mmask8 __M, __m256i __A, |
| 3407 | __m256i __B) { |
| 3408 | return (__m256i) __builtin_ia32_pmaxsq256_mask ((__v4di) __A, |
| 3409 | (__v4di) __B, |
| 3410 | (__v4di) __W, __M); |
| 3411 | } |
| 3412 | |
| 3413 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3414 | _mm256_max_epi64 (__m256i __A, __m256i __B) { |
| 3415 | return (__m256i) __builtin_ia32_pmaxsq256_mask ((__v4di) __A, |
| 3416 | (__v4di) __B, |
| 3417 | (__v4di) |
| 3418 | _mm256_setzero_si256 (), |
| 3419 | (__mmask8) -1); |
| 3420 | } |
| 3421 | |
| 3422 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3423 | _mm_maskz_max_epu32 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3424 | return (__m128i) __builtin_ia32_pmaxud128_mask ((__v4si) __A, |
| 3425 | (__v4si) __B, |
| 3426 | (__v4si) |
| 3427 | _mm_setzero_si128 (), |
| 3428 | __M); |
| 3429 | } |
| 3430 | |
| 3431 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3432 | _mm_mask_max_epu32 (__m128i __W, __mmask8 __M, __m128i __A, |
| 3433 | __m128i __B) { |
| 3434 | return (__m128i) __builtin_ia32_pmaxud128_mask ((__v4si) __A, |
| 3435 | (__v4si) __B, |
| 3436 | (__v4si) __W, __M); |
| 3437 | } |
| 3438 | |
| 3439 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3440 | _mm256_maskz_max_epu32 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3441 | return (__m256i) __builtin_ia32_pmaxud256_mask ((__v8si) __A, |
| 3442 | (__v8si) __B, |
| 3443 | (__v8si) |
| 3444 | _mm256_setzero_si256 (), |
| 3445 | __M); |
| 3446 | } |
| 3447 | |
| 3448 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3449 | _mm256_mask_max_epu32 (__m256i __W, __mmask8 __M, __m256i __A, |
| 3450 | __m256i __B) { |
| 3451 | return (__m256i) __builtin_ia32_pmaxud256_mask ((__v8si) __A, |
| 3452 | (__v8si) __B, |
| 3453 | (__v8si) __W, __M); |
| 3454 | } |
| 3455 | |
| 3456 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3457 | _mm_maskz_max_epu64 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3458 | return (__m128i) __builtin_ia32_pmaxuq128_mask ((__v2di) __A, |
| 3459 | (__v2di) __B, |
| 3460 | (__v2di) |
| 3461 | _mm_setzero_si128 (), |
| 3462 | __M); |
| 3463 | } |
| 3464 | |
| 3465 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3466 | _mm_max_epu64 (__m128i __A, __m128i __B) { |
| 3467 | return (__m128i) __builtin_ia32_pmaxuq128_mask ((__v2di) __A, |
| 3468 | (__v2di) __B, |
| 3469 | (__v2di) |
| 3470 | _mm_setzero_si128 (), |
| 3471 | (__mmask8) -1); |
| 3472 | } |
| 3473 | |
| 3474 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3475 | _mm_mask_max_epu64 (__m128i __W, __mmask8 __M, __m128i __A, |
| 3476 | __m128i __B) { |
| 3477 | return (__m128i) __builtin_ia32_pmaxuq128_mask ((__v2di) __A, |
| 3478 | (__v2di) __B, |
| 3479 | (__v2di) __W, __M); |
| 3480 | } |
| 3481 | |
| 3482 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3483 | _mm256_maskz_max_epu64 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3484 | return (__m256i) __builtin_ia32_pmaxuq256_mask ((__v4di) __A, |
| 3485 | (__v4di) __B, |
| 3486 | (__v4di) |
| 3487 | _mm256_setzero_si256 (), |
| 3488 | __M); |
| 3489 | } |
| 3490 | |
| 3491 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3492 | _mm256_max_epu64 (__m256i __A, __m256i __B) { |
| 3493 | return (__m256i) __builtin_ia32_pmaxuq256_mask ((__v4di) __A, |
| 3494 | (__v4di) __B, |
| 3495 | (__v4di) |
| 3496 | _mm256_setzero_si256 (), |
| 3497 | (__mmask8) -1); |
| 3498 | } |
| 3499 | |
| 3500 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3501 | _mm256_mask_max_epu64 (__m256i __W, __mmask8 __M, __m256i __A, |
| 3502 | __m256i __B) { |
| 3503 | return (__m256i) __builtin_ia32_pmaxuq256_mask ((__v4di) __A, |
| 3504 | (__v4di) __B, |
| 3505 | (__v4di) __W, __M); |
| 3506 | } |
| 3507 | |
| 3508 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3509 | _mm_maskz_min_epi32 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3510 | return (__m128i) __builtin_ia32_pminsd128_mask ((__v4si) __A, |
| 3511 | (__v4si) __B, |
| 3512 | (__v4si) |
| 3513 | _mm_setzero_si128 (), |
| 3514 | __M); |
| 3515 | } |
| 3516 | |
| 3517 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3518 | _mm_mask_min_epi32 (__m128i __W, __mmask8 __M, __m128i __A, |
| 3519 | __m128i __B) { |
| 3520 | return (__m128i) __builtin_ia32_pminsd128_mask ((__v4si) __A, |
| 3521 | (__v4si) __B, |
| 3522 | (__v4si) __W, __M); |
| 3523 | } |
| 3524 | |
| 3525 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3526 | _mm256_maskz_min_epi32 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3527 | return (__m256i) __builtin_ia32_pminsd256_mask ((__v8si) __A, |
| 3528 | (__v8si) __B, |
| 3529 | (__v8si) |
| 3530 | _mm256_setzero_si256 (), |
| 3531 | __M); |
| 3532 | } |
| 3533 | |
| 3534 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3535 | _mm256_mask_min_epi32 (__m256i __W, __mmask8 __M, __m256i __A, |
| 3536 | __m256i __B) { |
| 3537 | return (__m256i) __builtin_ia32_pminsd256_mask ((__v8si) __A, |
| 3538 | (__v8si) __B, |
| 3539 | (__v8si) __W, __M); |
| 3540 | } |
| 3541 | |
| 3542 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3543 | _mm_min_epi64 (__m128i __A, __m128i __B) { |
| 3544 | return (__m128i) __builtin_ia32_pminsq128_mask ((__v2di) __A, |
| 3545 | (__v2di) __B, |
| 3546 | (__v2di) |
| 3547 | _mm_setzero_si128 (), |
| 3548 | (__mmask8) -1); |
| 3549 | } |
| 3550 | |
| 3551 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3552 | _mm_mask_min_epi64 (__m128i __W, __mmask8 __M, __m128i __A, |
| 3553 | __m128i __B) { |
| 3554 | return (__m128i) __builtin_ia32_pminsq128_mask ((__v2di) __A, |
| 3555 | (__v2di) __B, |
| 3556 | (__v2di) __W, __M); |
| 3557 | } |
| 3558 | |
| 3559 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3560 | _mm_maskz_min_epi64 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3561 | return (__m128i) __builtin_ia32_pminsq128_mask ((__v2di) __A, |
| 3562 | (__v2di) __B, |
| 3563 | (__v2di) |
| 3564 | _mm_setzero_si128 (), |
| 3565 | __M); |
| 3566 | } |
| 3567 | |
| 3568 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3569 | _mm256_min_epi64 (__m256i __A, __m256i __B) { |
| 3570 | return (__m256i) __builtin_ia32_pminsq256_mask ((__v4di) __A, |
| 3571 | (__v4di) __B, |
| 3572 | (__v4di) |
| 3573 | _mm256_setzero_si256 (), |
| 3574 | (__mmask8) -1); |
| 3575 | } |
| 3576 | |
| 3577 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3578 | _mm256_mask_min_epi64 (__m256i __W, __mmask8 __M, __m256i __A, |
| 3579 | __m256i __B) { |
| 3580 | return (__m256i) __builtin_ia32_pminsq256_mask ((__v4di) __A, |
| 3581 | (__v4di) __B, |
| 3582 | (__v4di) __W, __M); |
| 3583 | } |
| 3584 | |
| 3585 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3586 | _mm256_maskz_min_epi64 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3587 | return (__m256i) __builtin_ia32_pminsq256_mask ((__v4di) __A, |
| 3588 | (__v4di) __B, |
| 3589 | (__v4di) |
| 3590 | _mm256_setzero_si256 (), |
| 3591 | __M); |
| 3592 | } |
| 3593 | |
| 3594 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3595 | _mm_maskz_min_epu32 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3596 | return (__m128i) __builtin_ia32_pminud128_mask ((__v4si) __A, |
| 3597 | (__v4si) __B, |
| 3598 | (__v4si) |
| 3599 | _mm_setzero_si128 (), |
| 3600 | __M); |
| 3601 | } |
| 3602 | |
| 3603 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3604 | _mm_mask_min_epu32 (__m128i __W, __mmask8 __M, __m128i __A, |
| 3605 | __m128i __B) { |
| 3606 | return (__m128i) __builtin_ia32_pminud128_mask ((__v4si) __A, |
| 3607 | (__v4si) __B, |
| 3608 | (__v4si) __W, __M); |
| 3609 | } |
| 3610 | |
| 3611 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3612 | _mm256_maskz_min_epu32 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3613 | return (__m256i) __builtin_ia32_pminud256_mask ((__v8si) __A, |
| 3614 | (__v8si) __B, |
| 3615 | (__v8si) |
| 3616 | _mm256_setzero_si256 (), |
| 3617 | __M); |
| 3618 | } |
| 3619 | |
| 3620 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3621 | _mm256_mask_min_epu32 (__m256i __W, __mmask8 __M, __m256i __A, |
| 3622 | __m256i __B) { |
| 3623 | return (__m256i) __builtin_ia32_pminud256_mask ((__v8si) __A, |
| 3624 | (__v8si) __B, |
| 3625 | (__v8si) __W, __M); |
| 3626 | } |
| 3627 | |
| 3628 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3629 | _mm_min_epu64 (__m128i __A, __m128i __B) { |
| 3630 | return (__m128i) __builtin_ia32_pminuq128_mask ((__v2di) __A, |
| 3631 | (__v2di) __B, |
| 3632 | (__v2di) |
| 3633 | _mm_setzero_si128 (), |
| 3634 | (__mmask8) -1); |
| 3635 | } |
| 3636 | |
| 3637 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3638 | _mm_mask_min_epu64 (__m128i __W, __mmask8 __M, __m128i __A, |
| 3639 | __m128i __B) { |
| 3640 | return (__m128i) __builtin_ia32_pminuq128_mask ((__v2di) __A, |
| 3641 | (__v2di) __B, |
| 3642 | (__v2di) __W, __M); |
| 3643 | } |
| 3644 | |
| 3645 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3646 | _mm_maskz_min_epu64 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3647 | return (__m128i) __builtin_ia32_pminuq128_mask ((__v2di) __A, |
| 3648 | (__v2di) __B, |
| 3649 | (__v2di) |
| 3650 | _mm_setzero_si128 (), |
| 3651 | __M); |
| 3652 | } |
| 3653 | |
| 3654 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3655 | _mm256_min_epu64 (__m256i __A, __m256i __B) { |
| 3656 | return (__m256i) __builtin_ia32_pminuq256_mask ((__v4di) __A, |
| 3657 | (__v4di) __B, |
| 3658 | (__v4di) |
| 3659 | _mm256_setzero_si256 (), |
| 3660 | (__mmask8) -1); |
| 3661 | } |
| 3662 | |
| 3663 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3664 | _mm256_mask_min_epu64 (__m256i __W, __mmask8 __M, __m256i __A, |
| 3665 | __m256i __B) { |
| 3666 | return (__m256i) __builtin_ia32_pminuq256_mask ((__v4di) __A, |
| 3667 | (__v4di) __B, |
| 3668 | (__v4di) __W, __M); |
| 3669 | } |
| 3670 | |
| 3671 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3672 | _mm256_maskz_min_epu64 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3673 | return (__m256i) __builtin_ia32_pminuq256_mask ((__v4di) __A, |
| 3674 | (__v4di) __B, |
| 3675 | (__v4di) |
| 3676 | _mm256_setzero_si256 (), |
| 3677 | __M); |
| 3678 | } |
| 3679 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3680 | #define _mm_roundscale_pd(A, imm) __extension__ ({ \ |
| 3681 | (__m128d)__builtin_ia32_rndscalepd_128_mask((__v2df)(__m128d)(A), \ |
| 3682 | (int)(imm), \ |
| 3683 | (__v2df)_mm_setzero_pd(), \ |
| 3684 | (__mmask8)-1); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3685 | |
| 3686 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3687 | #define _mm_mask_roundscale_pd(W, U, A, imm) __extension__ ({ \ |
| 3688 | (__m128d)__builtin_ia32_rndscalepd_128_mask((__v2df)(__m128d)(A), \ |
| 3689 | (int)(imm), \ |
| 3690 | (__v2df)(__m128d)(W), \ |
| 3691 | (__mmask8)(U)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3692 | |
| 3693 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3694 | #define _mm_maskz_roundscale_pd(U, A, imm) __extension__ ({ \ |
| 3695 | (__m128d)__builtin_ia32_rndscalepd_128_mask((__v2df)(__m128d)(A), \ |
| 3696 | (int)(imm), \ |
| 3697 | (__v2df)_mm_setzero_pd(), \ |
| 3698 | (__mmask8)(U)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3699 | |
| 3700 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3701 | #define _mm256_roundscale_pd(A, imm) __extension__ ({ \ |
| 3702 | (__m256d)__builtin_ia32_rndscalepd_256_mask((__v4df)(__m256d)(A), \ |
| 3703 | (int)(imm), \ |
| 3704 | (__v4df)_mm256_setzero_pd(), \ |
| 3705 | (__mmask8)-1); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3706 | |
| 3707 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3708 | #define _mm256_mask_roundscale_pd(W, U, A, imm) __extension__ ({ \ |
| 3709 | (__m256d)__builtin_ia32_rndscalepd_256_mask((__v4df)(__m256d)(A), \ |
| 3710 | (int)(imm), \ |
| 3711 | (__v4df)(__m256d)(W), \ |
| 3712 | (__mmask8)(U)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3713 | |
| 3714 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3715 | #define _mm256_maskz_roundscale_pd(U, A, imm) __extension__ ({ \ |
| 3716 | (__m256d)__builtin_ia32_rndscalepd_256_mask((__v4df)(__m256d)(A), \ |
| 3717 | (int)(imm), \ |
| 3718 | (__v4df)_mm256_setzero_pd(), \ |
| 3719 | (__mmask8)(U)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3720 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3721 | #define _mm_roundscale_ps(A, imm) __extension__ ({ \ |
| 3722 | (__m128)__builtin_ia32_rndscaleps_128_mask((__v4sf)(__m128)(A), (int)(imm), \ |
| 3723 | (__v4sf)_mm_setzero_ps(), \ |
| 3724 | (__mmask8)-1); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3725 | |
| 3726 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3727 | #define _mm_mask_roundscale_ps(W, U, A, imm) __extension__ ({ \ |
| 3728 | (__m128)__builtin_ia32_rndscaleps_128_mask((__v4sf)(__m128)(A), (int)(imm), \ |
| 3729 | (__v4sf)(__m128)(W), \ |
| 3730 | (__mmask8)(U)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3731 | |
| 3732 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3733 | #define _mm_maskz_roundscale_ps(U, A, imm) __extension__ ({ \ |
| 3734 | (__m128)__builtin_ia32_rndscaleps_128_mask((__v4sf)(__m128)(A), (int)(imm), \ |
| 3735 | (__v4sf)_mm_setzero_ps(), \ |
| 3736 | (__mmask8)(U)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3737 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3738 | #define _mm256_roundscale_ps(A, imm) __extension__ ({ \ |
| 3739 | (__m256)__builtin_ia32_rndscaleps_256_mask((__v8sf)(__m256)(A), (int)(imm), \ |
| 3740 | (__v8sf)_mm256_setzero_ps(), \ |
| 3741 | (__mmask8)-1); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3742 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3743 | #define _mm256_mask_roundscale_ps(W, U, A, imm) __extension__ ({ \ |
| 3744 | (__m256)__builtin_ia32_rndscaleps_256_mask((__v8sf)(__m256)(A), (int)(imm), \ |
| 3745 | (__v8sf)(__m256)(W), \ |
| 3746 | (__mmask8)(U)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3747 | |
| 3748 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3749 | #define _mm256_maskz_roundscale_ps(U, A, imm) __extension__ ({ \ |
| 3750 | (__m256)__builtin_ia32_rndscaleps_256_mask((__v8sf)(__m256)(A), (int)(imm), \ |
| 3751 | (__v8sf)_mm256_setzero_ps(), \ |
| 3752 | (__mmask8)(U)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3753 | |
| 3754 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3755 | _mm_scalef_pd (__m128d __A, __m128d __B) { |
| 3756 | return (__m128d) __builtin_ia32_scalefpd128_mask ((__v2df) __A, |
| 3757 | (__v2df) __B, |
| 3758 | (__v2df) |
| 3759 | _mm_setzero_pd (), |
| 3760 | (__mmask8) -1); |
| 3761 | } |
| 3762 | |
| 3763 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3764 | _mm_mask_scalef_pd (__m128d __W, __mmask8 __U, __m128d __A, |
| 3765 | __m128d __B) { |
| 3766 | return (__m128d) __builtin_ia32_scalefpd128_mask ((__v2df) __A, |
| 3767 | (__v2df) __B, |
| 3768 | (__v2df) __W, |
| 3769 | (__mmask8) __U); |
| 3770 | } |
| 3771 | |
| 3772 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3773 | _mm_maskz_scalef_pd (__mmask8 __U, __m128d __A, __m128d __B) { |
| 3774 | return (__m128d) __builtin_ia32_scalefpd128_mask ((__v2df) __A, |
| 3775 | (__v2df) __B, |
| 3776 | (__v2df) |
| 3777 | _mm_setzero_pd (), |
| 3778 | (__mmask8) __U); |
| 3779 | } |
| 3780 | |
| 3781 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3782 | _mm256_scalef_pd (__m256d __A, __m256d __B) { |
| 3783 | return (__m256d) __builtin_ia32_scalefpd256_mask ((__v4df) __A, |
| 3784 | (__v4df) __B, |
| 3785 | (__v4df) |
| 3786 | _mm256_setzero_pd (), |
| 3787 | (__mmask8) -1); |
| 3788 | } |
| 3789 | |
| 3790 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3791 | _mm256_mask_scalef_pd (__m256d __W, __mmask8 __U, __m256d __A, |
| 3792 | __m256d __B) { |
| 3793 | return (__m256d) __builtin_ia32_scalefpd256_mask ((__v4df) __A, |
| 3794 | (__v4df) __B, |
| 3795 | (__v4df) __W, |
| 3796 | (__mmask8) __U); |
| 3797 | } |
| 3798 | |
| 3799 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3800 | _mm256_maskz_scalef_pd (__mmask8 __U, __m256d __A, __m256d __B) { |
| 3801 | return (__m256d) __builtin_ia32_scalefpd256_mask ((__v4df) __A, |
| 3802 | (__v4df) __B, |
| 3803 | (__v4df) |
| 3804 | _mm256_setzero_pd (), |
| 3805 | (__mmask8) __U); |
| 3806 | } |
| 3807 | |
| 3808 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3809 | _mm_scalef_ps (__m128 __A, __m128 __B) { |
| 3810 | return (__m128) __builtin_ia32_scalefps128_mask ((__v4sf) __A, |
| 3811 | (__v4sf) __B, |
| 3812 | (__v4sf) |
| 3813 | _mm_setzero_ps (), |
| 3814 | (__mmask8) -1); |
| 3815 | } |
| 3816 | |
| 3817 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3818 | _mm_mask_scalef_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 3819 | return (__m128) __builtin_ia32_scalefps128_mask ((__v4sf) __A, |
| 3820 | (__v4sf) __B, |
| 3821 | (__v4sf) __W, |
| 3822 | (__mmask8) __U); |
| 3823 | } |
| 3824 | |
| 3825 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3826 | _mm_maskz_scalef_ps (__mmask8 __U, __m128 __A, __m128 __B) { |
| 3827 | return (__m128) __builtin_ia32_scalefps128_mask ((__v4sf) __A, |
| 3828 | (__v4sf) __B, |
| 3829 | (__v4sf) |
| 3830 | _mm_setzero_ps (), |
| 3831 | (__mmask8) __U); |
| 3832 | } |
| 3833 | |
| 3834 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3835 | _mm256_scalef_ps (__m256 __A, __m256 __B) { |
| 3836 | return (__m256) __builtin_ia32_scalefps256_mask ((__v8sf) __A, |
| 3837 | (__v8sf) __B, |
| 3838 | (__v8sf) |
| 3839 | _mm256_setzero_ps (), |
| 3840 | (__mmask8) -1); |
| 3841 | } |
| 3842 | |
| 3843 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3844 | _mm256_mask_scalef_ps (__m256 __W, __mmask8 __U, __m256 __A, |
| 3845 | __m256 __B) { |
| 3846 | return (__m256) __builtin_ia32_scalefps256_mask ((__v8sf) __A, |
| 3847 | (__v8sf) __B, |
| 3848 | (__v8sf) __W, |
| 3849 | (__mmask8) __U); |
| 3850 | } |
| 3851 | |
| 3852 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3853 | _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) { |
| 3854 | return (__m256) __builtin_ia32_scalefps256_mask ((__v8sf) __A, |
| 3855 | (__v8sf) __B, |
| 3856 | (__v8sf) |
| 3857 | _mm256_setzero_ps (), |
| 3858 | (__mmask8) __U); |
| 3859 | } |
| 3860 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3861 | #define _mm_i64scatter_pd(addr, index, v1, scale) __extension__ ({ \ |
| 3862 | __builtin_ia32_scatterdiv2df((double *)(addr), (__mmask8)-1, \ |
| 3863 | (__v2di)(__m128i)(index), \ |
| 3864 | (__v2df)(__m128d)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3865 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3866 | #define _mm_mask_i64scatter_pd(addr, mask, index, v1, scale) __extension__ ({ \ |
| 3867 | __builtin_ia32_scatterdiv2df((double *)(addr), (__mmask8)(mask), \ |
| 3868 | (__v2di)(__m128i)(index), \ |
| 3869 | (__v2df)(__m128d)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3870 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3871 | #define _mm_i64scatter_epi64(addr, index, v1, scale) __extension__ ({ \ |
| 3872 | __builtin_ia32_scatterdiv2di((long long *)(addr), (__mmask8)-1, \ |
| 3873 | (__v2di)(__m128i)(index), \ |
| 3874 | (__v2di)(__m128i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3875 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3876 | #define _mm_mask_i64scatter_epi64(addr, mask, index, v1, scale) __extension__ ({ \ |
| 3877 | __builtin_ia32_scatterdiv2di((long long *)(addr), (__mmask8)(mask), \ |
| 3878 | (__v2di)(__m128i)(index), \ |
| 3879 | (__v2di)(__m128i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3880 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3881 | #define _mm256_i64scatter_pd(addr, index, v1, scale) __extension__ ({ \ |
| 3882 | __builtin_ia32_scatterdiv4df((double *)(addr), (__mmask8)-1, \ |
| 3883 | (__v4di)(__m256i)(index), \ |
| 3884 | (__v4df)(__m256d)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3885 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3886 | #define _mm256_mask_i64scatter_pd(addr, mask, index, v1, scale) __extension__ ({ \ |
| 3887 | __builtin_ia32_scatterdiv4df((double *)(addr), (__mmask8)(mask), \ |
| 3888 | (__v4di)(__m256i)(index), \ |
| 3889 | (__v4df)(__m256d)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3890 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3891 | #define _mm256_i64scatter_epi64(addr, index, v1, scale) __extension__ ({ \ |
| 3892 | __builtin_ia32_scatterdiv4di((long long *)(addr), (__mmask8)-1, \ |
| 3893 | (__v4di)(__m256i)(index), \ |
| 3894 | (__v4di)(__m256i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3895 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3896 | #define _mm256_mask_i64scatter_epi64(addr, mask, index, v1, scale) __extension__ ({ \ |
| 3897 | __builtin_ia32_scatterdiv4di((long long *)(addr), (__mmask8)(mask), \ |
| 3898 | (__v4di)(__m256i)(index), \ |
| 3899 | (__v4di)(__m256i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3900 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3901 | #define _mm_i64scatter_ps(addr, index, v1, scale) __extension__ ({ \ |
| 3902 | __builtin_ia32_scatterdiv4sf((float *)(addr), (__mmask8)-1, \ |
| 3903 | (__v2di)(__m128i)(index), (__v4sf)(__m128)(v1), \ |
| 3904 | (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3905 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3906 | #define _mm_mask_i64scatter_ps(addr, mask, index, v1, scale) __extension__ ({ \ |
| 3907 | __builtin_ia32_scatterdiv4sf((float *)(addr), (__mmask8)(mask), \ |
| 3908 | (__v2di)(__m128i)(index), (__v4sf)(__m128)(v1), \ |
| 3909 | (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3910 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3911 | #define _mm_i64scatter_epi32(addr, index, v1, scale) __extension__ ({ \ |
| 3912 | __builtin_ia32_scatterdiv4si((int *)(addr), (__mmask8)-1, \ |
| 3913 | (__v2di)(__m128i)(index), \ |
| 3914 | (__v4si)(__m128i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3915 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3916 | #define _mm_mask_i64scatter_epi32(addr, mask, index, v1, scale) __extension__ ({ \ |
| 3917 | __builtin_ia32_scatterdiv4si((int *)(addr), (__mmask8)(mask), \ |
| 3918 | (__v2di)(__m128i)(index), \ |
| 3919 | (__v4si)(__m128i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3920 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3921 | #define _mm256_i64scatter_ps(addr, index, v1, scale) __extension__ ({ \ |
| 3922 | __builtin_ia32_scatterdiv8sf((float *)(addr), (__mmask8)-1, \ |
| 3923 | (__v4di)(__m256i)(index), (__v4sf)(__m128)(v1), \ |
| 3924 | (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3925 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3926 | #define _mm256_mask_i64scatter_ps(addr, mask, index, v1, scale) __extension__ ({ \ |
| 3927 | __builtin_ia32_scatterdiv8sf((float *)(addr), (__mmask8)(mask), \ |
| 3928 | (__v4di)(__m256i)(index), (__v4sf)(__m128)(v1), \ |
| 3929 | (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3930 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3931 | #define _mm256_i64scatter_epi32(addr, index, v1, scale) __extension__ ({ \ |
| 3932 | __builtin_ia32_scatterdiv8si((int *)(addr), (__mmask8)-1, \ |
| 3933 | (__v4di)(__m256i)(index), \ |
| 3934 | (__v4si)(__m128i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3935 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3936 | #define _mm256_mask_i64scatter_epi32(addr, mask, index, v1, scale) __extension__ ({ \ |
| 3937 | __builtin_ia32_scatterdiv8si((int *)(addr), (__mmask8)(mask), \ |
| 3938 | (__v4di)(__m256i)(index), \ |
| 3939 | (__v4si)(__m128i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3940 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3941 | #define _mm_i32scatter_pd(addr, index, v1, scale) __extension__ ({ \ |
| 3942 | __builtin_ia32_scattersiv2df((double *)(addr), (__mmask8)-1, \ |
| 3943 | (__v4si)(__m128i)(index), \ |
| 3944 | (__v2df)(__m128d)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3945 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3946 | #define _mm_mask_i32scatter_pd(addr, mask, index, v1, scale) __extension__ ({ \ |
| 3947 | __builtin_ia32_scattersiv2df((double *)(addr), (__mmask8)(mask), \ |
| 3948 | (__v4si)(__m128i)(index), \ |
| 3949 | (__v2df)(__m128d)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3950 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3951 | #define _mm_i32scatter_epi64(addr, index, v1, scale) __extension__ ({ \ |
| 3952 | __builtin_ia32_scattersiv2di((long long *)(addr), (__mmask8)-1, \ |
| 3953 | (__v4si)(__m128i)(index), \ |
| 3954 | (__v2di)(__m128i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3955 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3956 | #define _mm_mask_i32scatter_epi64(addr, mask, index, v1, scale) __extension__ ({ \ |
| 3957 | __builtin_ia32_scattersiv2di((long long *)(addr), (__mmask8)(mask), \ |
| 3958 | (__v4si)(__m128i)(index), \ |
| 3959 | (__v2di)(__m128i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3960 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3961 | #define _mm256_i32scatter_pd(addr, index, v1, scale) __extension__ ({ \ |
| 3962 | __builtin_ia32_scattersiv4df((double *)(addr), (__mmask8)-1, \ |
| 3963 | (__v4si)(__m128i)(index), \ |
| 3964 | (__v4df)(__m256d)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3965 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3966 | #define _mm256_mask_i32scatter_pd(addr, mask, index, v1, scale) __extension__ ({ \ |
| 3967 | __builtin_ia32_scattersiv4df((double *)(addr), (__mmask8)(mask), \ |
| 3968 | (__v4si)(__m128i)(index), \ |
| 3969 | (__v4df)(__m256d)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3970 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3971 | #define _mm256_i32scatter_epi64(addr, index, v1, scale) __extension__ ({ \ |
| 3972 | __builtin_ia32_scattersiv4di((long long *)(addr), (__mmask8)-1, \ |
| 3973 | (__v4si)(__m128i)(index), \ |
| 3974 | (__v4di)(__m256i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3975 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3976 | #define _mm256_mask_i32scatter_epi64(addr, mask, index, v1, scale) __extension__ ({ \ |
| 3977 | __builtin_ia32_scattersiv4di((long long *)(addr), (__mmask8)(mask), \ |
| 3978 | (__v4si)(__m128i)(index), \ |
| 3979 | (__v4di)(__m256i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3980 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3981 | #define _mm_i32scatter_ps(addr, index, v1, scale) __extension__ ({ \ |
| 3982 | __builtin_ia32_scattersiv4sf((float *)(addr), (__mmask8)-1, \ |
| 3983 | (__v4si)(__m128i)(index), (__v4sf)(__m128)(v1), \ |
| 3984 | (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3985 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3986 | #define _mm_mask_i32scatter_ps(addr, mask, index, v1, scale) __extension__ ({ \ |
| 3987 | __builtin_ia32_scattersiv4sf((float *)(addr), (__mmask8)(mask), \ |
| 3988 | (__v4si)(__m128i)(index), (__v4sf)(__m128)(v1), \ |
| 3989 | (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3990 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3991 | #define _mm_i32scatter_epi32(addr, index, v1, scale) __extension__ ({ \ |
| 3992 | __builtin_ia32_scattersiv4si((int *)(addr), (__mmask8)-1, \ |
| 3993 | (__v4si)(__m128i)(index), \ |
| 3994 | (__v4si)(__m128i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3995 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3996 | #define _mm_mask_i32scatter_epi32(addr, mask, index, v1, scale) __extension__ ({ \ |
| 3997 | __builtin_ia32_scattersiv4si((int *)(addr), (__mmask8)(mask), \ |
| 3998 | (__v4si)(__m128i)(index), \ |
| 3999 | (__v4si)(__m128i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4000 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4001 | #define _mm256_i32scatter_ps(addr, index, v1, scale) __extension__ ({ \ |
| 4002 | __builtin_ia32_scattersiv8sf((float *)(addr), (__mmask8)-1, \ |
| 4003 | (__v8si)(__m256i)(index), (__v8sf)(__m256)(v1), \ |
| 4004 | (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4005 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4006 | #define _mm256_mask_i32scatter_ps(addr, mask, index, v1, scale) __extension__ ({ \ |
| 4007 | __builtin_ia32_scattersiv8sf((float *)(addr), (__mmask8)(mask), \ |
| 4008 | (__v8si)(__m256i)(index), (__v8sf)(__m256)(v1), \ |
| 4009 | (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4010 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4011 | #define _mm256_i32scatter_epi32(addr, index, v1, scale) __extension__ ({ \ |
| 4012 | __builtin_ia32_scattersiv8si((int *)(addr), (__mmask8)-1, \ |
| 4013 | (__v8si)(__m256i)(index), \ |
| 4014 | (__v8si)(__m256i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4015 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4016 | #define _mm256_mask_i32scatter_epi32(addr, mask, index, v1, scale) __extension__ ({ \ |
| 4017 | __builtin_ia32_scattersiv8si((int *)(addr), (__mmask8)(mask), \ |
| 4018 | (__v8si)(__m256i)(index), \ |
| 4019 | (__v8si)(__m256i)(v1), (int)(scale)); }) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 4020 | |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 4021 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 4022 | _mm_mask_sqrt_pd (__m128d __W, __mmask8 __U, __m128d __A) { |
| 4023 | return (__m128d) __builtin_ia32_sqrtpd128_mask ((__v2df) __A, |
| 4024 | (__v2df) __W, |
| 4025 | (__mmask8) __U); |
| 4026 | } |
| 4027 | |
| 4028 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 4029 | _mm_maskz_sqrt_pd (__mmask8 __U, __m128d __A) { |
| 4030 | return (__m128d) __builtin_ia32_sqrtpd128_mask ((__v2df) __A, |
| 4031 | (__v2df) |
| 4032 | _mm_setzero_pd (), |
| 4033 | (__mmask8) __U); |
| 4034 | } |
| 4035 | |
| 4036 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 4037 | _mm256_mask_sqrt_pd (__m256d __W, __mmask8 __U, __m256d __A) { |
| 4038 | return (__m256d) __builtin_ia32_sqrtpd256_mask ((__v4df) __A, |
| 4039 | (__v4df) __W, |
| 4040 | (__mmask8) __U); |
| 4041 | } |
| 4042 | |
| 4043 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 4044 | _mm256_maskz_sqrt_pd (__mmask8 __U, __m256d __A) { |
| 4045 | return (__m256d) __builtin_ia32_sqrtpd256_mask ((__v4df) __A, |
| 4046 | (__v4df) |
| 4047 | _mm256_setzero_pd (), |
| 4048 | (__mmask8) __U); |
| 4049 | } |
| 4050 | |
| 4051 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 4052 | _mm_mask_sqrt_ps (__m128 __W, __mmask8 __U, __m128 __A) { |
| 4053 | return (__m128) __builtin_ia32_sqrtps128_mask ((__v4sf) __A, |
| 4054 | (__v4sf) __W, |
| 4055 | (__mmask8) __U); |
| 4056 | } |
| 4057 | |
| 4058 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 4059 | _mm_maskz_sqrt_ps (__mmask8 __U, __m128 __A) { |
| 4060 | return (__m128) __builtin_ia32_sqrtps128_mask ((__v4sf) __A, |
| 4061 | (__v4sf) |
| 4062 | _mm_setzero_ps (), |
| 4063 | (__mmask8) __U); |
| 4064 | } |
| 4065 | |
| 4066 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 4067 | _mm256_mask_sqrt_ps (__m256 __W, __mmask8 __U, __m256 __A) { |
| 4068 | return (__m256) __builtin_ia32_sqrtps256_mask ((__v8sf) __A, |
| 4069 | (__v8sf) __W, |
| 4070 | (__mmask8) __U); |
| 4071 | } |
| 4072 | |
| 4073 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 4074 | _mm256_maskz_sqrt_ps (__mmask8 __U, __m256 __A) { |
| 4075 | return (__m256) __builtin_ia32_sqrtps256_mask ((__v8sf) __A, |
| 4076 | (__v8sf) |
| 4077 | _mm256_setzero_ps (), |
| 4078 | (__mmask8) __U); |
| 4079 | } |
| 4080 | |
| 4081 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 4082 | _mm_mask_sub_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 4083 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 4084 | (__v2df)_mm_sub_pd(__A, __B), |
| 4085 | (__v2df)__W); |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 4086 | } |
| 4087 | |
| 4088 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 4089 | _mm_maskz_sub_pd(__mmask8 __U, __m128d __A, __m128d __B) { |
| 4090 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 4091 | (__v2df)_mm_sub_pd(__A, __B), |
| 4092 | (__v2df)_mm_setzero_pd()); |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 4093 | } |
| 4094 | |
| 4095 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 4096 | _mm256_mask_sub_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) { |
| 4097 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 4098 | (__v4df)_mm256_sub_pd(__A, __B), |
| 4099 | (__v4df)__W); |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 4100 | } |
| 4101 | |
| 4102 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 4103 | _mm256_maskz_sub_pd(__mmask8 __U, __m256d __A, __m256d __B) { |
| 4104 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 4105 | (__v4df)_mm256_sub_pd(__A, __B), |
| 4106 | (__v4df)_mm256_setzero_pd()); |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 4107 | } |
| 4108 | |
| 4109 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 4110 | _mm_mask_sub_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 4111 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 4112 | (__v4sf)_mm_sub_ps(__A, __B), |
| 4113 | (__v4sf)__W); |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 4114 | } |
| 4115 | |
| 4116 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 4117 | _mm_maskz_sub_ps(__mmask8 __U, __m128 __A, __m128 __B) { |
| 4118 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 4119 | (__v4sf)_mm_sub_ps(__A, __B), |
| 4120 | (__v4sf)_mm_setzero_ps()); |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 4121 | } |
| 4122 | |
| 4123 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 4124 | _mm256_mask_sub_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 4125 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 4126 | (__v8sf)_mm256_sub_ps(__A, __B), |
| 4127 | (__v8sf)__W); |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 4128 | } |
| 4129 | |
| 4130 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame^] | 4131 | _mm256_maskz_sub_ps(__mmask8 __U, __m256 __A, __m256 __B) { |
| 4132 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 4133 | (__v8sf)_mm256_sub_ps(__A, __B), |
| 4134 | (__v8sf)_mm256_setzero_ps()); |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 4135 | } |
| 4136 | |
| 4137 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4138 | _mm_mask2_permutex2var_epi32 (__m128i __A, __m128i __I, __mmask8 __U, |
| 4139 | __m128i __B) { |
| 4140 | return (__m128i) __builtin_ia32_vpermi2vard128_mask ((__v4si) __A, |
| 4141 | (__v4si) __I |
| 4142 | /* idx */ , |
| 4143 | (__v4si) __B, |
| 4144 | (__mmask8) __U); |
| 4145 | } |
| 4146 | |
| 4147 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4148 | _mm256_mask2_permutex2var_epi32 (__m256i __A, __m256i __I, |
| 4149 | __mmask8 __U, __m256i __B) { |
| 4150 | return (__m256i) __builtin_ia32_vpermi2vard256_mask ((__v8si) __A, |
| 4151 | (__v8si) __I |
| 4152 | /* idx */ , |
| 4153 | (__v8si) __B, |
| 4154 | (__mmask8) __U); |
| 4155 | } |
| 4156 | |
| 4157 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 4158 | _mm_mask2_permutex2var_pd (__m128d __A, __m128i __I, __mmask8 __U, |
| 4159 | __m128d __B) { |
| 4160 | return (__m128d) __builtin_ia32_vpermi2varpd128_mask ((__v2df) __A, |
| 4161 | (__v2di) __I |
| 4162 | /* idx */ , |
| 4163 | (__v2df) __B, |
| 4164 | (__mmask8) |
| 4165 | __U); |
| 4166 | } |
| 4167 | |
| 4168 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 4169 | _mm256_mask2_permutex2var_pd (__m256d __A, __m256i __I, __mmask8 __U, |
| 4170 | __m256d __B) { |
| 4171 | return (__m256d) __builtin_ia32_vpermi2varpd256_mask ((__v4df) __A, |
| 4172 | (__v4di) __I |
| 4173 | /* idx */ , |
| 4174 | (__v4df) __B, |
| 4175 | (__mmask8) |
| 4176 | __U); |
| 4177 | } |
| 4178 | |
| 4179 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 4180 | _mm_mask2_permutex2var_ps (__m128 __A, __m128i __I, __mmask8 __U, |
| 4181 | __m128 __B) { |
| 4182 | return (__m128) __builtin_ia32_vpermi2varps128_mask ((__v4sf) __A, |
| 4183 | (__v4si) __I |
| 4184 | /* idx */ , |
| 4185 | (__v4sf) __B, |
| 4186 | (__mmask8) __U); |
| 4187 | } |
| 4188 | |
| 4189 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 4190 | _mm256_mask2_permutex2var_ps (__m256 __A, __m256i __I, __mmask8 __U, |
| 4191 | __m256 __B) { |
| 4192 | return (__m256) __builtin_ia32_vpermi2varps256_mask ((__v8sf) __A, |
| 4193 | (__v8si) __I |
| 4194 | /* idx */ , |
| 4195 | (__v8sf) __B, |
| 4196 | (__mmask8) __U); |
| 4197 | } |
| 4198 | |
| 4199 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4200 | _mm_mask2_permutex2var_epi64 (__m128i __A, __m128i __I, __mmask8 __U, |
| 4201 | __m128i __B) { |
| 4202 | return (__m128i) __builtin_ia32_vpermi2varq128_mask ((__v2di) __A, |
| 4203 | (__v2di) __I |
| 4204 | /* idx */ , |
| 4205 | (__v2di) __B, |
| 4206 | (__mmask8) __U); |
| 4207 | } |
| 4208 | |
| 4209 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4210 | _mm256_mask2_permutex2var_epi64 (__m256i __A, __m256i __I, |
| 4211 | __mmask8 __U, __m256i __B) { |
| 4212 | return (__m256i) __builtin_ia32_vpermi2varq256_mask ((__v4di) __A, |
| 4213 | (__v4di) __I |
| 4214 | /* idx */ , |
| 4215 | (__v4di) __B, |
| 4216 | (__mmask8) __U); |
| 4217 | } |
| 4218 | |
| 4219 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4220 | _mm_permutex2var_epi32 (__m128i __A, __m128i __I, __m128i __B) { |
| 4221 | return (__m128i) __builtin_ia32_vpermt2vard128_mask ((__v4si) __I |
| 4222 | /* idx */ , |
| 4223 | (__v4si) __A, |
| 4224 | (__v4si) __B, |
| 4225 | (__mmask8) -1); |
| 4226 | } |
| 4227 | |
| 4228 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4229 | _mm_mask_permutex2var_epi32 (__m128i __A, __mmask8 __U, __m128i __I, |
| 4230 | __m128i __B) { |
| 4231 | return (__m128i) __builtin_ia32_vpermt2vard128_mask ((__v4si) __I |
| 4232 | /* idx */ , |
| 4233 | (__v4si) __A, |
| 4234 | (__v4si) __B, |
| 4235 | (__mmask8) __U); |
| 4236 | } |
| 4237 | |
| 4238 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4239 | _mm_maskz_permutex2var_epi32 (__mmask8 __U, __m128i __A, __m128i __I, |
| 4240 | __m128i __B) { |
| 4241 | return (__m128i) __builtin_ia32_vpermt2vard128_maskz ((__v4si) __I |
| 4242 | /* idx */ , |
| 4243 | (__v4si) __A, |
| 4244 | (__v4si) __B, |
| 4245 | (__mmask8) |
| 4246 | __U); |
| 4247 | } |
| 4248 | |
| 4249 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4250 | _mm256_permutex2var_epi32 (__m256i __A, __m256i __I, __m256i __B) { |
| 4251 | return (__m256i) __builtin_ia32_vpermt2vard256_mask ((__v8si) __I |
| 4252 | /* idx */ , |
| 4253 | (__v8si) __A, |
| 4254 | (__v8si) __B, |
| 4255 | (__mmask8) -1); |
| 4256 | } |
| 4257 | |
| 4258 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4259 | _mm256_mask_permutex2var_epi32 (__m256i __A, __mmask8 __U, __m256i __I, |
| 4260 | __m256i __B) { |
| 4261 | return (__m256i) __builtin_ia32_vpermt2vard256_mask ((__v8si) __I |
| 4262 | /* idx */ , |
| 4263 | (__v8si) __A, |
| 4264 | (__v8si) __B, |
| 4265 | (__mmask8) __U); |
| 4266 | } |
| 4267 | |
| 4268 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4269 | _mm256_maskz_permutex2var_epi32 (__mmask8 __U, __m256i __A, |
| 4270 | __m256i __I, __m256i __B) { |
| 4271 | return (__m256i) __builtin_ia32_vpermt2vard256_maskz ((__v8si) __I |
| 4272 | /* idx */ , |
| 4273 | (__v8si) __A, |
| 4274 | (__v8si) __B, |
| 4275 | (__mmask8) |
| 4276 | __U); |
| 4277 | } |
| 4278 | |
| 4279 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 4280 | _mm_permutex2var_pd (__m128d __A, __m128i __I, __m128d __B) { |
| 4281 | return (__m128d) __builtin_ia32_vpermt2varpd128_mask ((__v2di) __I |
| 4282 | /* idx */ , |
| 4283 | (__v2df) __A, |
| 4284 | (__v2df) __B, |
| 4285 | (__mmask8) - |
| 4286 | 1); |
| 4287 | } |
| 4288 | |
| 4289 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 4290 | _mm_mask_permutex2var_pd (__m128d __A, __mmask8 __U, __m128i __I, |
| 4291 | __m128d __B) { |
| 4292 | return (__m128d) __builtin_ia32_vpermt2varpd128_mask ((__v2di) __I |
| 4293 | /* idx */ , |
| 4294 | (__v2df) __A, |
| 4295 | (__v2df) __B, |
| 4296 | (__mmask8) |
| 4297 | __U); |
| 4298 | } |
| 4299 | |
| 4300 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 4301 | _mm_maskz_permutex2var_pd (__mmask8 __U, __m128d __A, __m128i __I, |
| 4302 | __m128d __B) { |
| 4303 | return (__m128d) __builtin_ia32_vpermt2varpd128_maskz ((__v2di) __I |
| 4304 | /* idx */ , |
| 4305 | (__v2df) __A, |
| 4306 | (__v2df) __B, |
| 4307 | (__mmask8) |
| 4308 | __U); |
| 4309 | } |
| 4310 | |
| 4311 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 4312 | _mm256_permutex2var_pd (__m256d __A, __m256i __I, __m256d __B) { |
| 4313 | return (__m256d) __builtin_ia32_vpermt2varpd256_mask ((__v4di) __I |
| 4314 | /* idx */ , |
| 4315 | (__v4df) __A, |
| 4316 | (__v4df) __B, |
| 4317 | (__mmask8) - |
| 4318 | 1); |
| 4319 | } |
| 4320 | |
| 4321 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 4322 | _mm256_mask_permutex2var_pd (__m256d __A, __mmask8 __U, __m256i __I, |
| 4323 | __m256d __B) { |
| 4324 | return (__m256d) __builtin_ia32_vpermt2varpd256_mask ((__v4di) __I |
| 4325 | /* idx */ , |
| 4326 | (__v4df) __A, |
| 4327 | (__v4df) __B, |
| 4328 | (__mmask8) |
| 4329 | __U); |
| 4330 | } |
| 4331 | |
| 4332 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 4333 | _mm256_maskz_permutex2var_pd (__mmask8 __U, __m256d __A, __m256i __I, |
| 4334 | __m256d __B) { |
| 4335 | return (__m256d) __builtin_ia32_vpermt2varpd256_maskz ((__v4di) __I |
| 4336 | /* idx */ , |
| 4337 | (__v4df) __A, |
| 4338 | (__v4df) __B, |
| 4339 | (__mmask8) |
| 4340 | __U); |
| 4341 | } |
| 4342 | |
| 4343 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 4344 | _mm_permutex2var_ps (__m128 __A, __m128i __I, __m128 __B) { |
| 4345 | return (__m128) __builtin_ia32_vpermt2varps128_mask ((__v4si) __I |
| 4346 | /* idx */ , |
| 4347 | (__v4sf) __A, |
| 4348 | (__v4sf) __B, |
| 4349 | (__mmask8) -1); |
| 4350 | } |
| 4351 | |
| 4352 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 4353 | _mm_mask_permutex2var_ps (__m128 __A, __mmask8 __U, __m128i __I, |
| 4354 | __m128 __B) { |
| 4355 | return (__m128) __builtin_ia32_vpermt2varps128_mask ((__v4si) __I |
| 4356 | /* idx */ , |
| 4357 | (__v4sf) __A, |
| 4358 | (__v4sf) __B, |
| 4359 | (__mmask8) __U); |
| 4360 | } |
| 4361 | |
| 4362 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 4363 | _mm_maskz_permutex2var_ps (__mmask8 __U, __m128 __A, __m128i __I, |
| 4364 | __m128 __B) { |
| 4365 | return (__m128) __builtin_ia32_vpermt2varps128_maskz ((__v4si) __I |
| 4366 | /* idx */ , |
| 4367 | (__v4sf) __A, |
| 4368 | (__v4sf) __B, |
| 4369 | (__mmask8) |
| 4370 | __U); |
| 4371 | } |
| 4372 | |
| 4373 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 4374 | _mm256_permutex2var_ps (__m256 __A, __m256i __I, __m256 __B) { |
| 4375 | return (__m256) __builtin_ia32_vpermt2varps256_mask ((__v8si) __I |
| 4376 | /* idx */ , |
| 4377 | (__v8sf) __A, |
| 4378 | (__v8sf) __B, |
| 4379 | (__mmask8) -1); |
| 4380 | } |
| 4381 | |
| 4382 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 4383 | _mm256_mask_permutex2var_ps (__m256 __A, __mmask8 __U, __m256i __I, |
| 4384 | __m256 __B) { |
| 4385 | return (__m256) __builtin_ia32_vpermt2varps256_mask ((__v8si) __I |
| 4386 | /* idx */ , |
| 4387 | (__v8sf) __A, |
| 4388 | (__v8sf) __B, |
| 4389 | (__mmask8) __U); |
| 4390 | } |
| 4391 | |
| 4392 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 4393 | _mm256_maskz_permutex2var_ps (__mmask8 __U, __m256 __A, __m256i __I, |
| 4394 | __m256 __B) { |
| 4395 | return (__m256) __builtin_ia32_vpermt2varps256_maskz ((__v8si) __I |
| 4396 | /* idx */ , |
| 4397 | (__v8sf) __A, |
| 4398 | (__v8sf) __B, |
| 4399 | (__mmask8) |
| 4400 | __U); |
| 4401 | } |
| 4402 | |
| 4403 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4404 | _mm_permutex2var_epi64 (__m128i __A, __m128i __I, __m128i __B) { |
| 4405 | return (__m128i) __builtin_ia32_vpermt2varq128_mask ((__v2di) __I |
| 4406 | /* idx */ , |
| 4407 | (__v2di) __A, |
| 4408 | (__v2di) __B, |
| 4409 | (__mmask8) -1); |
| 4410 | } |
| 4411 | |
| 4412 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4413 | _mm_mask_permutex2var_epi64 (__m128i __A, __mmask8 __U, __m128i __I, |
| 4414 | __m128i __B) { |
| 4415 | return (__m128i) __builtin_ia32_vpermt2varq128_mask ((__v2di) __I |
| 4416 | /* idx */ , |
| 4417 | (__v2di) __A, |
| 4418 | (__v2di) __B, |
| 4419 | (__mmask8) __U); |
| 4420 | } |
| 4421 | |
| 4422 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4423 | _mm_maskz_permutex2var_epi64 (__mmask8 __U, __m128i __A, __m128i __I, |
| 4424 | __m128i __B) { |
| 4425 | return (__m128i) __builtin_ia32_vpermt2varq128_maskz ((__v2di) __I |
| 4426 | /* idx */ , |
| 4427 | (__v2di) __A, |
| 4428 | (__v2di) __B, |
| 4429 | (__mmask8) |
| 4430 | __U); |
| 4431 | } |
| 4432 | |
| 4433 | |
| 4434 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4435 | _mm256_permutex2var_epi64 (__m256i __A, __m256i __I, __m256i __B) { |
| 4436 | return (__m256i) __builtin_ia32_vpermt2varq256_mask ((__v4di) __I |
| 4437 | /* idx */ , |
| 4438 | (__v4di) __A, |
| 4439 | (__v4di) __B, |
| 4440 | (__mmask8) -1); |
| 4441 | } |
| 4442 | |
| 4443 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4444 | _mm256_mask_permutex2var_epi64 (__m256i __A, __mmask8 __U, __m256i __I, |
| 4445 | __m256i __B) { |
| 4446 | return (__m256i) __builtin_ia32_vpermt2varq256_mask ((__v4di) __I |
| 4447 | /* idx */ , |
| 4448 | (__v4di) __A, |
| 4449 | (__v4di) __B, |
| 4450 | (__mmask8) __U); |
| 4451 | } |
| 4452 | |
| 4453 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4454 | _mm256_maskz_permutex2var_epi64 (__mmask8 __U, __m256i __A, |
| 4455 | __m256i __I, __m256i __B) { |
| 4456 | return (__m256i) __builtin_ia32_vpermt2varq256_maskz ((__v4di) __I |
| 4457 | /* idx */ , |
| 4458 | (__v4di) __A, |
| 4459 | (__v4di) __B, |
| 4460 | (__mmask8) |
| 4461 | __U); |
| 4462 | } |
| 4463 | |
Michael Zuckerman | 7cdb72f | 2016-02-18 09:09:34 +0000 | [diff] [blame] | 4464 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4465 | _mm_mask_cvtepi8_epi32 (__m128i __W, __mmask8 __U, __m128i __A) |
| 4466 | { |
| 4467 | return (__m128i) __builtin_ia32_pmovsxbd128_mask ((__v16qi) __A, |
| 4468 | (__v4si) __W, |
| 4469 | (__mmask8) __U); |
| 4470 | } |
| 4471 | |
| 4472 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4473 | _mm_maskz_cvtepi8_epi32 (__mmask8 __U, __m128i __A) |
| 4474 | { |
| 4475 | return (__m128i) __builtin_ia32_pmovsxbd128_mask ((__v16qi) __A, |
| 4476 | (__v4si) |
| 4477 | _mm_setzero_si128 (), |
| 4478 | (__mmask8) __U); |
| 4479 | } |
| 4480 | |
| 4481 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4482 | _mm256_mask_cvtepi8_epi32 (__m256i __W, __mmask8 __U, __m128i __A) |
| 4483 | { |
| 4484 | return (__m256i) __builtin_ia32_pmovsxbd256_mask ((__v16qi) __A, |
| 4485 | (__v8si) __W, |
| 4486 | (__mmask8) __U); |
| 4487 | } |
| 4488 | |
| 4489 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4490 | _mm256_maskz_cvtepi8_epi32 (__mmask8 __U, __m128i __A) |
| 4491 | { |
| 4492 | return (__m256i) __builtin_ia32_pmovsxbd256_mask ((__v16qi) __A, |
| 4493 | (__v8si) |
| 4494 | _mm256_setzero_si256 (), |
| 4495 | (__mmask8) __U); |
| 4496 | } |
| 4497 | |
| 4498 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4499 | _mm_mask_cvtepi8_epi64 (__m128i __W, __mmask8 __U, __m128i __A) |
| 4500 | { |
| 4501 | return (__m128i) __builtin_ia32_pmovsxbq128_mask ((__v16qi) __A, |
| 4502 | (__v2di) __W, |
| 4503 | (__mmask8) __U); |
| 4504 | } |
| 4505 | |
| 4506 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4507 | _mm_maskz_cvtepi8_epi64 (__mmask8 __U, __m128i __A) |
| 4508 | { |
| 4509 | return (__m128i) __builtin_ia32_pmovsxbq128_mask ((__v16qi) __A, |
| 4510 | (__v2di) |
| 4511 | _mm_setzero_si128 (), |
| 4512 | (__mmask8) __U); |
| 4513 | } |
| 4514 | |
| 4515 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4516 | _mm256_mask_cvtepi8_epi64 (__m256i __W, __mmask8 __U, __m128i __A) |
| 4517 | { |
| 4518 | return (__m256i) __builtin_ia32_pmovsxbq256_mask ((__v16qi) __A, |
| 4519 | (__v4di) __W, |
| 4520 | (__mmask8) __U); |
| 4521 | } |
| 4522 | |
| 4523 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4524 | _mm256_maskz_cvtepi8_epi64 (__mmask8 __U, __m128i __A) |
| 4525 | { |
| 4526 | return (__m256i) __builtin_ia32_pmovsxbq256_mask ((__v16qi) __A, |
| 4527 | (__v4di) |
| 4528 | _mm256_setzero_si256 (), |
| 4529 | (__mmask8) __U); |
| 4530 | } |
| 4531 | |
| 4532 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4533 | _mm_mask_cvtepi32_epi64 (__m128i __W, __mmask8 __U, __m128i __X) |
| 4534 | { |
| 4535 | return (__m128i) __builtin_ia32_pmovsxdq128_mask ((__v4si) __X, |
| 4536 | (__v2di) __W, |
| 4537 | (__mmask8) __U); |
| 4538 | } |
| 4539 | |
| 4540 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4541 | _mm_maskz_cvtepi32_epi64 (__mmask8 __U, __m128i __X) |
| 4542 | { |
| 4543 | return (__m128i) __builtin_ia32_pmovsxdq128_mask ((__v4si) __X, |
| 4544 | (__v2di) |
| 4545 | _mm_setzero_si128 (), |
| 4546 | (__mmask8) __U); |
| 4547 | } |
| 4548 | |
| 4549 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4550 | _mm256_mask_cvtepi32_epi64 (__m256i __W, __mmask8 __U, __m128i __X) |
| 4551 | { |
| 4552 | return (__m256i) __builtin_ia32_pmovsxdq256_mask ((__v4si) __X, |
| 4553 | (__v4di) __W, |
| 4554 | (__mmask8) __U); |
| 4555 | } |
| 4556 | |
| 4557 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4558 | _mm256_maskz_cvtepi32_epi64 (__mmask8 __U, __m128i __X) |
| 4559 | { |
| 4560 | return (__m256i) __builtin_ia32_pmovsxdq256_mask ((__v4si) __X, |
| 4561 | (__v4di) |
| 4562 | _mm256_setzero_si256 (), |
| 4563 | (__mmask8) __U); |
| 4564 | } |
| 4565 | |
| 4566 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4567 | _mm_mask_cvtepi16_epi32 (__m128i __W, __mmask8 __U, __m128i __A) |
| 4568 | { |
| 4569 | return (__m128i) __builtin_ia32_pmovsxwd128_mask ((__v8hi) __A, |
| 4570 | (__v4si) __W, |
| 4571 | (__mmask8) __U); |
| 4572 | } |
| 4573 | |
| 4574 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4575 | _mm_maskz_cvtepi16_epi32 (__mmask8 __U, __m128i __A) |
| 4576 | { |
| 4577 | return (__m128i) __builtin_ia32_pmovsxwd128_mask ((__v8hi) __A, |
| 4578 | (__v4si) |
| 4579 | _mm_setzero_si128 (), |
| 4580 | (__mmask8) __U); |
| 4581 | } |
| 4582 | |
| 4583 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4584 | _mm256_mask_cvtepi16_epi32 (__m256i __W, __mmask8 __U, __m128i __A) |
| 4585 | { |
| 4586 | return (__m256i) __builtin_ia32_pmovsxwd256_mask ((__v8hi) __A, |
| 4587 | (__v8si) __W, |
| 4588 | (__mmask8) __U); |
| 4589 | } |
| 4590 | |
| 4591 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4592 | _mm256_maskz_cvtepi16_epi32 (__mmask8 __U, __m128i __A) |
| 4593 | { |
| 4594 | return (__m256i) __builtin_ia32_pmovsxwd256_mask ((__v8hi) __A, |
| 4595 | (__v8si) |
| 4596 | _mm256_setzero_si256 (), |
| 4597 | (__mmask8) __U); |
| 4598 | } |
| 4599 | |
| 4600 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4601 | _mm_mask_cvtepi16_epi64 (__m128i __W, __mmask8 __U, __m128i __A) |
| 4602 | { |
| 4603 | return (__m128i) __builtin_ia32_pmovsxwq128_mask ((__v8hi) __A, |
| 4604 | (__v2di) __W, |
| 4605 | (__mmask8) __U); |
| 4606 | } |
| 4607 | |
| 4608 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4609 | _mm_maskz_cvtepi16_epi64 (__mmask8 __U, __m128i __A) |
| 4610 | { |
| 4611 | return (__m128i) __builtin_ia32_pmovsxwq128_mask ((__v8hi) __A, |
| 4612 | (__v2di) |
| 4613 | _mm_setzero_si128 (), |
| 4614 | (__mmask8) __U); |
| 4615 | } |
| 4616 | |
| 4617 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4618 | _mm256_mask_cvtepi16_epi64 (__m256i __W, __mmask8 __U, __m128i __A) |
| 4619 | { |
| 4620 | return (__m256i) __builtin_ia32_pmovsxwq256_mask ((__v8hi) __A, |
| 4621 | (__v4di) __W, |
| 4622 | (__mmask8) __U); |
| 4623 | } |
| 4624 | |
| 4625 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4626 | _mm256_maskz_cvtepi16_epi64 (__mmask8 __U, __m128i __A) |
| 4627 | { |
| 4628 | return (__m256i) __builtin_ia32_pmovsxwq256_mask ((__v8hi) __A, |
| 4629 | (__v4di) |
| 4630 | _mm256_setzero_si256 (), |
| 4631 | (__mmask8) __U); |
| 4632 | } |
| 4633 | |
| 4634 | |
Michael Zuckerman | 7a33dce | 2016-02-21 14:00:11 +0000 | [diff] [blame] | 4635 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4636 | _mm_mask_cvtepu8_epi32 (__m128i __W, __mmask8 __U, __m128i __A) |
| 4637 | { |
| 4638 | return (__m128i) __builtin_ia32_pmovzxbd128_mask ((__v16qi) __A, |
| 4639 | (__v4si) __W, |
| 4640 | (__mmask8) __U); |
| 4641 | } |
| 4642 | |
| 4643 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4644 | _mm_maskz_cvtepu8_epi32 (__mmask8 __U, __m128i __A) |
| 4645 | { |
| 4646 | return (__m128i) __builtin_ia32_pmovzxbd128_mask ((__v16qi) __A, |
| 4647 | (__v4si) |
| 4648 | _mm_setzero_si128 (), |
| 4649 | (__mmask8) __U); |
| 4650 | } |
| 4651 | |
| 4652 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4653 | _mm256_mask_cvtepu8_epi32 (__m256i __W, __mmask8 __U, __m128i __A) |
| 4654 | { |
| 4655 | return (__m256i) __builtin_ia32_pmovzxbd256_mask ((__v16qi) __A, |
| 4656 | (__v8si) __W, |
| 4657 | (__mmask8) __U); |
| 4658 | } |
| 4659 | |
| 4660 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4661 | _mm256_maskz_cvtepu8_epi32 (__mmask8 __U, __m128i __A) |
| 4662 | { |
| 4663 | return (__m256i) __builtin_ia32_pmovzxbd256_mask ((__v16qi) __A, |
| 4664 | (__v8si) |
| 4665 | _mm256_setzero_si256 (), |
| 4666 | (__mmask8) __U); |
| 4667 | } |
| 4668 | |
| 4669 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4670 | _mm_mask_cvtepu8_epi64 (__m128i __W, __mmask8 __U, __m128i __A) |
| 4671 | { |
| 4672 | return (__m128i) __builtin_ia32_pmovzxbq128_mask ((__v16qi) __A, |
| 4673 | (__v2di) __W, |
| 4674 | (__mmask8) __U); |
| 4675 | } |
| 4676 | |
| 4677 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4678 | _mm_maskz_cvtepu8_epi64 (__mmask8 __U, __m128i __A) |
| 4679 | { |
| 4680 | return (__m128i) __builtin_ia32_pmovzxbq128_mask ((__v16qi) __A, |
| 4681 | (__v2di) |
| 4682 | _mm_setzero_si128 (), |
| 4683 | (__mmask8) __U); |
| 4684 | } |
| 4685 | |
| 4686 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4687 | _mm256_mask_cvtepu8_epi64 (__m256i __W, __mmask8 __U, __m128i __A) |
| 4688 | { |
| 4689 | return (__m256i) __builtin_ia32_pmovzxbq256_mask ((__v16qi) __A, |
| 4690 | (__v4di) __W, |
| 4691 | (__mmask8) __U); |
| 4692 | } |
| 4693 | |
| 4694 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4695 | _mm256_maskz_cvtepu8_epi64 (__mmask8 __U, __m128i __A) |
| 4696 | { |
| 4697 | return (__m256i) __builtin_ia32_pmovzxbq256_mask ((__v16qi) __A, |
| 4698 | (__v4di) |
| 4699 | _mm256_setzero_si256 (), |
| 4700 | (__mmask8) __U); |
| 4701 | } |
| 4702 | |
| 4703 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4704 | _mm_mask_cvtepu32_epi64 (__m128i __W, __mmask8 __U, __m128i __X) |
| 4705 | { |
| 4706 | return (__m128i) __builtin_ia32_pmovzxdq128_mask ((__v4si) __X, |
| 4707 | (__v2di) __W, |
| 4708 | (__mmask8) __U); |
| 4709 | } |
| 4710 | |
| 4711 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4712 | _mm_maskz_cvtepu32_epi64 (__mmask8 __U, __m128i __X) |
| 4713 | { |
| 4714 | return (__m128i) __builtin_ia32_pmovzxdq128_mask ((__v4si) __X, |
| 4715 | (__v2di) |
| 4716 | _mm_setzero_si128 (), |
| 4717 | (__mmask8) __U); |
| 4718 | } |
| 4719 | |
| 4720 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4721 | _mm256_mask_cvtepu32_epi64 (__m256i __W, __mmask8 __U, __m128i __X) |
| 4722 | { |
| 4723 | return (__m256i) __builtin_ia32_pmovzxdq256_mask ((__v4si) __X, |
| 4724 | (__v4di) __W, |
| 4725 | (__mmask8) __U); |
| 4726 | } |
| 4727 | |
| 4728 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4729 | _mm256_maskz_cvtepu32_epi64 (__mmask8 __U, __m128i __X) |
| 4730 | { |
| 4731 | return (__m256i) __builtin_ia32_pmovzxdq256_mask ((__v4si) __X, |
| 4732 | (__v4di) |
| 4733 | _mm256_setzero_si256 (), |
| 4734 | (__mmask8) __U); |
| 4735 | } |
| 4736 | |
| 4737 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4738 | _mm_mask_cvtepu16_epi32 (__m128i __W, __mmask8 __U, __m128i __A) |
| 4739 | { |
| 4740 | return (__m128i) __builtin_ia32_pmovzxwd128_mask ((__v8hi) __A, |
| 4741 | (__v4si) __W, |
| 4742 | (__mmask8) __U); |
| 4743 | } |
| 4744 | |
| 4745 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4746 | _mm_maskz_cvtepu16_epi32 (__mmask8 __U, __m128i __A) |
| 4747 | { |
| 4748 | return (__m128i) __builtin_ia32_pmovzxwd128_mask ((__v8hi) __A, |
| 4749 | (__v4si) |
| 4750 | _mm_setzero_si128 (), |
| 4751 | (__mmask8) __U); |
| 4752 | } |
| 4753 | |
| 4754 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4755 | _mm256_mask_cvtepu16_epi32 (__m256i __W, __mmask8 __U, __m128i __A) |
| 4756 | { |
| 4757 | return (__m256i) __builtin_ia32_pmovzxwd256_mask ((__v8hi) __A, |
| 4758 | (__v8si) __W, |
| 4759 | (__mmask8) __U); |
| 4760 | } |
| 4761 | |
| 4762 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4763 | _mm256_maskz_cvtepu16_epi32 (__mmask8 __U, __m128i __A) |
| 4764 | { |
| 4765 | return (__m256i) __builtin_ia32_pmovzxwd256_mask ((__v8hi) __A, |
| 4766 | (__v8si) |
| 4767 | _mm256_setzero_si256 (), |
| 4768 | (__mmask8) __U); |
| 4769 | } |
| 4770 | |
| 4771 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4772 | _mm_mask_cvtepu16_epi64 (__m128i __W, __mmask8 __U, __m128i __A) |
| 4773 | { |
| 4774 | return (__m128i) __builtin_ia32_pmovzxwq128_mask ((__v8hi) __A, |
| 4775 | (__v2di) __W, |
| 4776 | (__mmask8) __U); |
| 4777 | } |
| 4778 | |
| 4779 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4780 | _mm_maskz_cvtepu16_epi64 (__mmask8 __U, __m128i __A) |
| 4781 | { |
| 4782 | return (__m128i) __builtin_ia32_pmovzxwq128_mask ((__v8hi) __A, |
| 4783 | (__v2di) |
| 4784 | _mm_setzero_si128 (), |
| 4785 | (__mmask8) __U); |
| 4786 | } |
| 4787 | |
| 4788 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4789 | _mm256_mask_cvtepu16_epi64 (__m256i __W, __mmask8 __U, __m128i __A) |
| 4790 | { |
| 4791 | return (__m256i) __builtin_ia32_pmovzxwq256_mask ((__v8hi) __A, |
| 4792 | (__v4di) __W, |
| 4793 | (__mmask8) __U); |
| 4794 | } |
| 4795 | |
| 4796 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4797 | _mm256_maskz_cvtepu16_epi64 (__mmask8 __U, __m128i __A) |
| 4798 | { |
| 4799 | return (__m256i) __builtin_ia32_pmovzxwq256_mask ((__v8hi) __A, |
| 4800 | (__v4di) |
| 4801 | _mm256_setzero_si256 (), |
| 4802 | (__mmask8) __U); |
| 4803 | } |
| 4804 | |
| 4805 | |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4806 | #define _mm_rol_epi32(a, b) __extension__ ({\ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4807 | (__m128i)__builtin_ia32_prold128_mask((__v4si)(__m128i)(a), (int)(b), \ |
| 4808 | (__v4si)_mm_setzero_si128(), \ |
| 4809 | (__mmask8)-1); }) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4810 | |
| 4811 | #define _mm_mask_rol_epi32(w, u, a, b) __extension__ ({\ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4812 | (__m128i)__builtin_ia32_prold128_mask((__v4si)(__m128i)(a), (int)(b), \ |
| 4813 | (__v4si)(__m128i)(w), (__mmask8)(u)); }) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4814 | |
| 4815 | #define _mm_maskz_rol_epi32(u, a, b) __extension__ ({\ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4816 | (__m128i)__builtin_ia32_prold128_mask((__v4si)(__m128i)(a), (int)(b), \ |
| 4817 | (__v4si)_mm_setzero_si128(), \ |
| 4818 | (__mmask8)(u)); }) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4819 | |
| 4820 | #define _mm256_rol_epi32(a, b) __extension__ ({\ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4821 | (__m256i)__builtin_ia32_prold256_mask((__v8si)(__m256i)(a), (int)(b), \ |
| 4822 | (__v8si)_mm256_setzero_si256(), \ |
| 4823 | (__mmask8)-1); }) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4824 | |
| 4825 | #define _mm256_mask_rol_epi32(w, u, a, b) __extension__ ({\ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4826 | (__m256i)__builtin_ia32_prold256_mask((__v8si)(__m256i)(a), (int)(b), \ |
| 4827 | (__v8si)(__m256i)(w), (__mmask8)(u)); }) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4828 | |
| 4829 | #define _mm256_maskz_rol_epi32(u, a, b) __extension__ ({\ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4830 | (__m256i)__builtin_ia32_prold256_mask((__v8si)(__m256i)(a), (int)(b), \ |
| 4831 | (__v8si)_mm256_setzero_si256(), \ |
| 4832 | (__mmask8)(u)); }) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4833 | |
Michael Zuckerman | 477e0a3 | 2016-02-22 09:42:57 +0000 | [diff] [blame] | 4834 | #define _mm_rol_epi64(a, b) __extension__ ({\ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4835 | (__m128i)__builtin_ia32_prolq128_mask((__v2di)(__m128i)(a), (int)(b), \ |
| 4836 | (__v2di)_mm_setzero_di(), \ |
| 4837 | (__mmask8)-1); }) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4838 | |
| 4839 | #define _mm_mask_rol_epi64(w, u, a, b) __extension__ ({\ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4840 | (__m128i)__builtin_ia32_prolq128_mask((__v2di)(__m128i)(a), (int)(b), \ |
| 4841 | (__v2di)(__m128i)(w), (__mmask8)(u)); }) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4842 | |
| 4843 | #define _mm_maskz_rol_epi64(u, a, b) __extension__ ({\ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4844 | (__m128i)__builtin_ia32_prolq128_mask((__v2di)(__m128i)(a), (int)(b), \ |
| 4845 | (__v2di)_mm_setzero_di(), \ |
| 4846 | (__mmask8)(u)); }) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4847 | |
| 4848 | #define _mm256_rol_epi64(a, b) __extension__ ({\ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4849 | (__m256i)__builtin_ia32_prolq256_mask((__v4di)(__m256i)(a), (int)(b), \ |
| 4850 | (__v4di)_mm256_setzero_si256(), \ |
| 4851 | (__mmask8)-1); }) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4852 | |
| 4853 | #define _mm256_mask_rol_epi64(w, u, a, b) __extension__ ({\ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4854 | (__m256i)__builtin_ia32_prolq256_mask((__v4di)(__m256i)(a), (int)(b), \ |
| 4855 | (__v4di)(__m256i)(w), (__mmask8)(u)); }) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4856 | |
| 4857 | #define _mm256_maskz_rol_epi64(u, a, b) __extension__ ({\ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4858 | (__m256i)__builtin_ia32_prolq256_mask((__v4di)(__m256i)(a), (int)(b), \ |
| 4859 | (__v4di)_mm256_setzero_si256(), \ |
| 4860 | (__mmask8)(u)); }) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4861 | |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 4862 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4863 | _mm_rolv_epi32 (__m128i __A, __m128i __B) |
| 4864 | { |
| 4865 | return (__m128i) __builtin_ia32_prolvd128_mask ((__v4si) __A, |
| 4866 | (__v4si) __B, |
| 4867 | (__v4si) |
| 4868 | _mm_setzero_si128 (), |
| 4869 | (__mmask8) -1); |
| 4870 | } |
| 4871 | |
| 4872 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4873 | _mm_mask_rolv_epi32 (__m128i __W, __mmask8 __U, __m128i __A, |
| 4874 | __m128i __B) |
| 4875 | { |
| 4876 | return (__m128i) __builtin_ia32_prolvd128_mask ((__v4si) __A, |
| 4877 | (__v4si) __B, |
| 4878 | (__v4si) __W, |
| 4879 | (__mmask8) __U); |
| 4880 | } |
| 4881 | |
| 4882 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4883 | _mm_maskz_rolv_epi32 (__mmask8 __U, __m128i __A, __m128i __B) |
| 4884 | { |
| 4885 | return (__m128i) __builtin_ia32_prolvd128_mask ((__v4si) __A, |
| 4886 | (__v4si) __B, |
| 4887 | (__v4si) |
| 4888 | _mm_setzero_si128 (), |
| 4889 | (__mmask8) __U); |
| 4890 | } |
| 4891 | |
| 4892 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4893 | _mm256_rolv_epi32 (__m256i __A, __m256i __B) |
| 4894 | { |
| 4895 | return (__m256i) __builtin_ia32_prolvd256_mask ((__v8si) __A, |
| 4896 | (__v8si) __B, |
| 4897 | (__v8si) |
| 4898 | _mm256_setzero_si256 (), |
| 4899 | (__mmask8) -1); |
| 4900 | } |
| 4901 | |
| 4902 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4903 | _mm256_mask_rolv_epi32 (__m256i __W, __mmask8 __U, __m256i __A, |
| 4904 | __m256i __B) |
| 4905 | { |
| 4906 | return (__m256i) __builtin_ia32_prolvd256_mask ((__v8si) __A, |
| 4907 | (__v8si) __B, |
| 4908 | (__v8si) __W, |
| 4909 | (__mmask8) __U); |
| 4910 | } |
| 4911 | |
| 4912 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4913 | _mm256_maskz_rolv_epi32 (__mmask8 __U, __m256i __A, __m256i __B) |
| 4914 | { |
| 4915 | return (__m256i) __builtin_ia32_prolvd256_mask ((__v8si) __A, |
| 4916 | (__v8si) __B, |
| 4917 | (__v8si) |
| 4918 | _mm256_setzero_si256 (), |
| 4919 | (__mmask8) __U); |
| 4920 | } |
| 4921 | |
| 4922 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4923 | _mm_rolv_epi64 (__m128i __A, __m128i __B) |
| 4924 | { |
| 4925 | return (__m128i) __builtin_ia32_prolvq128_mask ((__v2di) __A, |
| 4926 | (__v2di) __B, |
| 4927 | (__v2di) |
| 4928 | _mm_setzero_di (), |
| 4929 | (__mmask8) -1); |
| 4930 | } |
| 4931 | |
| 4932 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4933 | _mm_mask_rolv_epi64 (__m128i __W, __mmask8 __U, __m128i __A, |
| 4934 | __m128i __B) |
| 4935 | { |
| 4936 | return (__m128i) __builtin_ia32_prolvq128_mask ((__v2di) __A, |
| 4937 | (__v2di) __B, |
| 4938 | (__v2di) __W, |
| 4939 | (__mmask8) __U); |
| 4940 | } |
| 4941 | |
| 4942 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4943 | _mm_maskz_rolv_epi64 (__mmask8 __U, __m128i __A, __m128i __B) |
| 4944 | { |
| 4945 | return (__m128i) __builtin_ia32_prolvq128_mask ((__v2di) __A, |
| 4946 | (__v2di) __B, |
| 4947 | (__v2di) |
| 4948 | _mm_setzero_di (), |
| 4949 | (__mmask8) __U); |
| 4950 | } |
| 4951 | |
| 4952 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4953 | _mm256_rolv_epi64 (__m256i __A, __m256i __B) |
| 4954 | { |
| 4955 | return (__m256i) __builtin_ia32_prolvq256_mask ((__v4di) __A, |
| 4956 | (__v4di) __B, |
| 4957 | (__v4di) |
| 4958 | _mm256_setzero_si256 (), |
| 4959 | (__mmask8) -1); |
| 4960 | } |
| 4961 | |
| 4962 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4963 | _mm256_mask_rolv_epi64 (__m256i __W, __mmask8 __U, __m256i __A, |
| 4964 | __m256i __B) |
| 4965 | { |
| 4966 | return (__m256i) __builtin_ia32_prolvq256_mask ((__v4di) __A, |
| 4967 | (__v4di) __B, |
| 4968 | (__v4di) __W, |
| 4969 | (__mmask8) __U); |
| 4970 | } |
| 4971 | |
| 4972 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4973 | _mm256_maskz_rolv_epi64 (__mmask8 __U, __m256i __A, __m256i __B) |
| 4974 | { |
| 4975 | return (__m256i) __builtin_ia32_prolvq256_mask ((__v4di) __A, |
| 4976 | (__v4di) __B, |
| 4977 | (__v4di) |
| 4978 | _mm256_setzero_si256 (), |
| 4979 | (__mmask8) __U); |
| 4980 | } |
| 4981 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4982 | #define _mm_ror_epi32(A, B) __extension__ ({ \ |
| 4983 | (__m128i)__builtin_ia32_prord128_mask((__v4si)(__m128i)(A), (int)(B), \ |
| 4984 | (__v4si)_mm_setzero_si128(), \ |
| 4985 | (__mmask8)-1); }) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 4986 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4987 | #define _mm_mask_ror_epi32(W, U, A, B) __extension__ ({ \ |
| 4988 | (__m128i)__builtin_ia32_prord128_mask((__v4si)(__m128i)(A), (int)(B), \ |
| 4989 | (__v4si)(__m128i)(W), (__mmask8)(U)); }) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 4990 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4991 | #define _mm_maskz_ror_epi32(U, A, B) __extension__ ({ \ |
| 4992 | (__m128i)__builtin_ia32_prord128_mask((__v4si)(__m128i)(A), (int)(B), \ |
| 4993 | (__v4si)_mm_setzero_si128(), \ |
| 4994 | (__mmask8)(U)); }) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 4995 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4996 | #define _mm256_ror_epi32(A, B) __extension__ ({ \ |
| 4997 | (__m256i)__builtin_ia32_prord256_mask((__v8si)(__m256i)(A), (int)(B), \ |
| 4998 | (__v8si)_mm256_setzero_si256(), \ |
| 4999 | (__mmask8)-1); }) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 5000 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5001 | #define _mm256_mask_ror_epi32(W, U, A, B) __extension__ ({ \ |
| 5002 | (__m256i)__builtin_ia32_prord256_mask((__v8si)(__m256i)(A), (int)(B), \ |
| 5003 | (__v8si)(__m256i)(W), (__mmask8)(U)); }) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 5004 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5005 | #define _mm256_maskz_ror_epi32(U, A, B) __extension__ ({ \ |
| 5006 | (__m256i)__builtin_ia32_prord256_mask((__v8si)(__m256i)(A), (int)(B), \ |
| 5007 | (__v8si)_mm256_setzero_si256(), \ |
| 5008 | (__mmask8)(U)); }) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 5009 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5010 | #define _mm_ror_epi64(A, B) __extension__ ({ \ |
| 5011 | (__m128i)__builtin_ia32_prorq128_mask((__v2di)(__m128i)(A), (int)(B), \ |
| 5012 | (__v2di)_mm_setzero_di(), \ |
| 5013 | (__mmask8)-1); }) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 5014 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5015 | #define _mm_mask_ror_epi64(W, U, A, B) __extension__ ({ \ |
| 5016 | (__m128i)__builtin_ia32_prorq128_mask((__v2di)(__m128i)(A), (int)(B), \ |
| 5017 | (__v2di)(__m128i)(W), (__mmask8)(U)); }) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 5018 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5019 | #define _mm_maskz_ror_epi64(U, A, B) __extension__ ({ \ |
| 5020 | (__m128i)__builtin_ia32_prorq128_mask((__v2di)(__m128i)(A), (int)(B), \ |
| 5021 | (__v2di)_mm_setzero_di(), \ |
| 5022 | (__mmask8)(U)); }) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 5023 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5024 | #define _mm256_ror_epi64(A, B) __extension__ ({ \ |
| 5025 | (__m256i)__builtin_ia32_prorq256_mask((__v4di)(__m256i)(A), (int)(B), \ |
| 5026 | (__v4di)_mm256_setzero_si256(), \ |
| 5027 | (__mmask8)-1); }) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 5028 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5029 | #define _mm256_mask_ror_epi64(W, U, A, B) __extension__ ({ \ |
| 5030 | (__m256i)__builtin_ia32_prorq256_mask((__v4di)(__m256i)(A), (int)(B), \ |
| 5031 | (__v4di)(__m256i)(W), (__mmask8)(U)); }) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 5032 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5033 | #define _mm256_maskz_ror_epi64(U, A, B) __extension__ ({ \ |
| 5034 | (__m256i)__builtin_ia32_prorq256_mask((__v4di)(__m256i)(A), (int)(B), \ |
| 5035 | (__v4di)_mm256_setzero_si256(), \ |
| 5036 | (__mmask8)(U)); }) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 5037 | |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 5038 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5039 | _mm_mask_sll_epi32 (__m128i __W, __mmask8 __U, __m128i __A, |
| 5040 | __m128i __B) |
| 5041 | { |
| 5042 | return (__m128i) __builtin_ia32_pslld128_mask ((__v4si) __A, |
| 5043 | (__v4si) __B, |
| 5044 | (__v4si) __W, |
| 5045 | (__mmask8) __U); |
| 5046 | } |
| 5047 | |
| 5048 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5049 | _mm_maskz_sll_epi32 (__mmask8 __U, __m128i __A, __m128i __B) |
| 5050 | { |
| 5051 | return (__m128i) __builtin_ia32_pslld128_mask ((__v4si) __A, |
| 5052 | (__v4si) __B, |
| 5053 | (__v4si) |
| 5054 | _mm_setzero_si128 (), |
| 5055 | (__mmask8) __U); |
| 5056 | } |
| 5057 | |
| 5058 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5059 | _mm256_mask_sll_epi32 (__m256i __W, __mmask8 __U, __m256i __A, |
| 5060 | __m128i __B) |
| 5061 | { |
| 5062 | return (__m256i) __builtin_ia32_pslld256_mask ((__v8si) __A, |
| 5063 | (__v4si) __B, |
| 5064 | (__v8si) __W, |
| 5065 | (__mmask8) __U); |
| 5066 | } |
| 5067 | |
| 5068 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5069 | _mm256_maskz_sll_epi32 (__mmask8 __U, __m256i __A, __m128i __B) |
| 5070 | { |
| 5071 | return (__m256i) __builtin_ia32_pslld256_mask ((__v8si) __A, |
| 5072 | (__v4si) __B, |
| 5073 | (__v8si) |
| 5074 | _mm256_setzero_si256 (), |
| 5075 | (__mmask8) __U); |
| 5076 | } |
| 5077 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5078 | #define _mm_mask_slli_epi32(W, U, A, B) __extension__ ({ \ |
| 5079 | (__m128i)__builtin_ia32_pslldi128_mask((__v4si)(__m128i)(A), (int)(B), \ |
| 5080 | (__v4si)(__m128i)(W), \ |
| 5081 | (__mmask8)(U)); }) |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 5082 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5083 | #define _mm_maskz_slli_epi32(U, A, B) __extension__ ({ \ |
| 5084 | (__m128i)__builtin_ia32_pslldi128_mask((__v4si)(__m128i)(A), (int)(B), \ |
| 5085 | (__v4si)_mm_setzero_si128(), \ |
| 5086 | (__mmask8)(U)); }) |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 5087 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5088 | #define _mm256_mask_slli_epi32(W, U, A, B) __extension__ ({ \ |
| 5089 | (__m256i)__builtin_ia32_pslldi256_mask((__v8si)(__m256i)(A), (int)(B), \ |
| 5090 | (__v8si)(__m256i)(W), \ |
| 5091 | (__mmask8)(U)); }) |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 5092 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5093 | #define _mm256_maskz_slli_epi32(U, A, B) __extension__ ({ \ |
| 5094 | (__m256i)__builtin_ia32_pslldi256_mask((__v8si)(__m256i)(A), (int)(B), \ |
| 5095 | (__v8si)_mm256_setzero_si256(), \ |
| 5096 | (__mmask8)(U)); }) |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 5097 | |
| 5098 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5099 | _mm_mask_sll_epi64 (__m128i __W, __mmask8 __U, __m128i __A, |
| 5100 | __m128i __B) |
| 5101 | { |
| 5102 | return (__m128i) __builtin_ia32_psllq128_mask ((__v2di) __A, |
| 5103 | (__v2di) __B, |
| 5104 | (__v2di) __W, |
| 5105 | (__mmask8) __U); |
| 5106 | } |
| 5107 | |
| 5108 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5109 | _mm_maskz_sll_epi64 (__mmask8 __U, __m128i __A, __m128i __B) |
| 5110 | { |
| 5111 | return (__m128i) __builtin_ia32_psllq128_mask ((__v2di) __A, |
| 5112 | (__v2di) __B, |
| 5113 | (__v2di) |
| 5114 | _mm_setzero_di (), |
| 5115 | (__mmask8) __U); |
| 5116 | } |
| 5117 | |
| 5118 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5119 | _mm256_mask_sll_epi64 (__m256i __W, __mmask8 __U, __m256i __A, |
| 5120 | __m128i __B) |
| 5121 | { |
| 5122 | return (__m256i) __builtin_ia32_psllq256_mask ((__v4di) __A, |
| 5123 | (__v2di) __B, |
| 5124 | (__v4di) __W, |
| 5125 | (__mmask8) __U); |
| 5126 | } |
| 5127 | |
| 5128 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5129 | _mm256_maskz_sll_epi64 (__mmask8 __U, __m256i __A, __m128i __B) |
| 5130 | { |
| 5131 | return (__m256i) __builtin_ia32_psllq256_mask ((__v4di) __A, |
| 5132 | (__v2di) __B, |
| 5133 | (__v4di) |
| 5134 | _mm256_setzero_si256 (), |
| 5135 | (__mmask8) __U); |
| 5136 | } |
| 5137 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5138 | #define _mm_mask_slli_epi64(W, U, A, B) __extension__ ({ \ |
| 5139 | (__m128i)__builtin_ia32_psllqi128_mask((__v2di)(__m128i)(A), (int)(B), \ |
| 5140 | (__v2di)(__m128i)(W), \ |
| 5141 | (__mmask8)(U)); }) |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 5142 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5143 | #define _mm_maskz_slli_epi64(U, A, B) __extension__ ({ \ |
| 5144 | (__m128i)__builtin_ia32_psllqi128_mask((__v2di)(__m128i)(A), (int)(B), \ |
| 5145 | (__v2di)_mm_setzero_di(), \ |
| 5146 | (__mmask8)(U)); }) |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 5147 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5148 | #define _mm256_mask_slli_epi64(W, U, A, B) __extension__ ({ \ |
| 5149 | (__m256i)__builtin_ia32_psllqi256_mask((__v4di)(__m256i)(A), (int)(B), \ |
| 5150 | (__v4di)(__m256i)(W), \ |
| 5151 | (__mmask8)(U)); }) |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 5152 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5153 | #define _mm256_maskz_slli_epi64(U, A, B) __extension__ ({ \ |
| 5154 | (__m256i)__builtin_ia32_psllqi256_mask((__v4di)(__m256i)(A), (int)(B), \ |
| 5155 | (__v4di)_mm256_setzero_si256(), \ |
| 5156 | (__mmask8)(U)); }) |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 5157 | |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 5158 | |
Michael Zuckerman | e98cc74 | 2016-02-23 15:59:47 +0000 | [diff] [blame] | 5159 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5160 | _mm_rorv_epi32 (__m128i __A, __m128i __B) |
| 5161 | { |
| 5162 | return (__m128i) __builtin_ia32_prorvd128_mask ((__v4si) __A, |
| 5163 | (__v4si) __B, |
| 5164 | (__v4si) |
| 5165 | _mm_setzero_si128 (), |
| 5166 | (__mmask8) -1); |
| 5167 | } |
| 5168 | |
| 5169 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5170 | _mm_mask_rorv_epi32 (__m128i __W, __mmask8 __U, __m128i __A, |
| 5171 | __m128i __B) |
| 5172 | { |
| 5173 | return (__m128i) __builtin_ia32_prorvd128_mask ((__v4si) __A, |
| 5174 | (__v4si) __B, |
| 5175 | (__v4si) __W, |
| 5176 | (__mmask8) __U); |
| 5177 | } |
| 5178 | |
| 5179 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5180 | _mm_maskz_rorv_epi32 (__mmask8 __U, __m128i __A, __m128i __B) |
| 5181 | { |
| 5182 | return (__m128i) __builtin_ia32_prorvd128_mask ((__v4si) __A, |
| 5183 | (__v4si) __B, |
| 5184 | (__v4si) |
| 5185 | _mm_setzero_si128 (), |
| 5186 | (__mmask8) __U); |
| 5187 | } |
| 5188 | |
| 5189 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5190 | _mm256_rorv_epi32 (__m256i __A, __m256i __B) |
| 5191 | { |
| 5192 | return (__m256i) __builtin_ia32_prorvd256_mask ((__v8si) __A, |
| 5193 | (__v8si) __B, |
| 5194 | (__v8si) |
| 5195 | _mm256_setzero_si256 (), |
| 5196 | (__mmask8) -1); |
| 5197 | } |
| 5198 | |
| 5199 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5200 | _mm256_mask_rorv_epi32 (__m256i __W, __mmask8 __U, __m256i __A, |
| 5201 | __m256i __B) |
| 5202 | { |
| 5203 | return (__m256i) __builtin_ia32_prorvd256_mask ((__v8si) __A, |
| 5204 | (__v8si) __B, |
| 5205 | (__v8si) __W, |
| 5206 | (__mmask8) __U); |
| 5207 | } |
| 5208 | |
| 5209 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5210 | _mm256_maskz_rorv_epi32 (__mmask8 __U, __m256i __A, __m256i __B) |
| 5211 | { |
| 5212 | return (__m256i) __builtin_ia32_prorvd256_mask ((__v8si) __A, |
| 5213 | (__v8si) __B, |
| 5214 | (__v8si) |
| 5215 | _mm256_setzero_si256 (), |
| 5216 | (__mmask8) __U); |
| 5217 | } |
| 5218 | |
| 5219 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5220 | _mm_rorv_epi64 (__m128i __A, __m128i __B) |
| 5221 | { |
| 5222 | return (__m128i) __builtin_ia32_prorvq128_mask ((__v2di) __A, |
| 5223 | (__v2di) __B, |
| 5224 | (__v2di) |
| 5225 | _mm_setzero_di (), |
| 5226 | (__mmask8) -1); |
| 5227 | } |
| 5228 | |
| 5229 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5230 | _mm_mask_rorv_epi64 (__m128i __W, __mmask8 __U, __m128i __A, |
| 5231 | __m128i __B) |
| 5232 | { |
| 5233 | return (__m128i) __builtin_ia32_prorvq128_mask ((__v2di) __A, |
| 5234 | (__v2di) __B, |
| 5235 | (__v2di) __W, |
| 5236 | (__mmask8) __U); |
| 5237 | } |
| 5238 | |
| 5239 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5240 | _mm_maskz_rorv_epi64 (__mmask8 __U, __m128i __A, __m128i __B) |
| 5241 | { |
| 5242 | return (__m128i) __builtin_ia32_prorvq128_mask ((__v2di) __A, |
| 5243 | (__v2di) __B, |
| 5244 | (__v2di) |
| 5245 | _mm_setzero_di (), |
| 5246 | (__mmask8) __U); |
| 5247 | } |
| 5248 | |
| 5249 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5250 | _mm256_rorv_epi64 (__m256i __A, __m256i __B) |
| 5251 | { |
| 5252 | return (__m256i) __builtin_ia32_prorvq256_mask ((__v4di) __A, |
| 5253 | (__v4di) __B, |
| 5254 | (__v4di) |
| 5255 | _mm256_setzero_si256 (), |
| 5256 | (__mmask8) -1); |
| 5257 | } |
| 5258 | |
| 5259 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5260 | _mm256_mask_rorv_epi64 (__m256i __W, __mmask8 __U, __m256i __A, |
| 5261 | __m256i __B) |
| 5262 | { |
| 5263 | return (__m256i) __builtin_ia32_prorvq256_mask ((__v4di) __A, |
| 5264 | (__v4di) __B, |
| 5265 | (__v4di) __W, |
| 5266 | (__mmask8) __U); |
| 5267 | } |
| 5268 | |
| 5269 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5270 | _mm256_maskz_rorv_epi64 (__mmask8 __U, __m256i __A, __m256i __B) |
| 5271 | { |
| 5272 | return (__m256i) __builtin_ia32_prorvq256_mask ((__v4di) __A, |
| 5273 | (__v4di) __B, |
| 5274 | (__v4di) |
| 5275 | _mm256_setzero_si256 (), |
| 5276 | (__mmask8) __U); |
| 5277 | } |
| 5278 | |
Michael Zuckerman | 431b0e1 | 2016-02-28 07:39:34 +0000 | [diff] [blame] | 5279 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5280 | _mm_mask_sllv_epi64 (__m128i __W, __mmask8 __U, __m128i __X, |
| 5281 | __m128i __Y) |
| 5282 | { |
| 5283 | return (__m128i) __builtin_ia32_psllv2di_mask ((__v2di) __X, |
| 5284 | (__v2di) __Y, |
| 5285 | (__v2di) __W, |
| 5286 | (__mmask8) __U); |
| 5287 | } |
| 5288 | |
| 5289 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5290 | _mm_maskz_sllv_epi64 (__mmask8 __U, __m128i __X, __m128i __Y) |
| 5291 | { |
| 5292 | return (__m128i) __builtin_ia32_psllv2di_mask ((__v2di) __X, |
| 5293 | (__v2di) __Y, |
| 5294 | (__v2di) |
| 5295 | _mm_setzero_di (), |
| 5296 | (__mmask8) __U); |
| 5297 | } |
| 5298 | |
| 5299 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5300 | _mm256_mask_sllv_epi64 (__m256i __W, __mmask8 __U, __m256i __X, |
| 5301 | __m256i __Y) |
| 5302 | { |
| 5303 | return (__m256i) __builtin_ia32_psllv4di_mask ((__v4di) __X, |
| 5304 | (__v4di) __Y, |
| 5305 | (__v4di) __W, |
| 5306 | (__mmask8) __U); |
| 5307 | } |
| 5308 | |
| 5309 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5310 | _mm256_maskz_sllv_epi64 (__mmask8 __U, __m256i __X, __m256i __Y) |
| 5311 | { |
| 5312 | return (__m256i) __builtin_ia32_psllv4di_mask ((__v4di) __X, |
| 5313 | (__v4di) __Y, |
| 5314 | (__v4di) |
| 5315 | _mm256_setzero_si256 (), |
| 5316 | (__mmask8) __U); |
| 5317 | } |
| 5318 | |
| 5319 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5320 | _mm_mask_sllv_epi32 (__m128i __W, __mmask8 __U, __m128i __X, |
| 5321 | __m128i __Y) |
| 5322 | { |
| 5323 | return (__m128i) __builtin_ia32_psllv4si_mask ((__v4si) __X, |
| 5324 | (__v4si) __Y, |
| 5325 | (__v4si) __W, |
| 5326 | (__mmask8) __U); |
| 5327 | } |
| 5328 | |
| 5329 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5330 | _mm_maskz_sllv_epi32 (__mmask8 __U, __m128i __X, __m128i __Y) |
| 5331 | { |
| 5332 | return (__m128i) __builtin_ia32_psllv4si_mask ((__v4si) __X, |
| 5333 | (__v4si) __Y, |
| 5334 | (__v4si) |
| 5335 | _mm_setzero_si128 (), |
| 5336 | (__mmask8) __U); |
| 5337 | } |
| 5338 | |
| 5339 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5340 | _mm256_mask_sllv_epi32 (__m256i __W, __mmask8 __U, __m256i __X, |
| 5341 | __m256i __Y) |
| 5342 | { |
| 5343 | return (__m256i) __builtin_ia32_psllv8si_mask ((__v8si) __X, |
| 5344 | (__v8si) __Y, |
| 5345 | (__v8si) __W, |
| 5346 | (__mmask8) __U); |
| 5347 | } |
| 5348 | |
| 5349 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5350 | _mm256_maskz_sllv_epi32 (__mmask8 __U, __m256i __X, __m256i __Y) |
| 5351 | { |
| 5352 | return (__m256i) __builtin_ia32_psllv8si_mask ((__v8si) __X, |
| 5353 | (__v8si) __Y, |
| 5354 | (__v8si) |
| 5355 | _mm256_setzero_si256 (), |
| 5356 | (__mmask8) __U); |
| 5357 | } |
| 5358 | |
Michael Zuckerman | 0165e76 | 2016-03-01 13:03:45 +0000 | [diff] [blame] | 5359 | |
| 5360 | |
| 5361 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5362 | _mm_mask_srlv_epi64 (__m128i __W, __mmask8 __U, __m128i __X, |
| 5363 | __m128i __Y) |
| 5364 | { |
| 5365 | return (__m128i) __builtin_ia32_psrlv2di_mask ((__v2di) __X, |
| 5366 | (__v2di) __Y, |
| 5367 | (__v2di) __W, |
| 5368 | (__mmask8) __U); |
| 5369 | } |
| 5370 | |
| 5371 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5372 | _mm_maskz_srlv_epi64 (__mmask8 __U, __m128i __X, __m128i __Y) |
| 5373 | { |
| 5374 | return (__m128i) __builtin_ia32_psrlv2di_mask ((__v2di) __X, |
| 5375 | (__v2di) __Y, |
| 5376 | (__v2di) |
| 5377 | _mm_setzero_di (), |
| 5378 | (__mmask8) __U); |
| 5379 | } |
| 5380 | |
| 5381 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5382 | _mm256_mask_srlv_epi64 (__m256i __W, __mmask8 __U, __m256i __X, |
| 5383 | __m256i __Y) |
| 5384 | { |
| 5385 | return (__m256i) __builtin_ia32_psrlv4di_mask ((__v4di) __X, |
| 5386 | (__v4di) __Y, |
| 5387 | (__v4di) __W, |
| 5388 | (__mmask8) __U); |
| 5389 | } |
| 5390 | |
| 5391 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5392 | _mm256_maskz_srlv_epi64 (__mmask8 __U, __m256i __X, __m256i __Y) |
| 5393 | { |
| 5394 | return (__m256i) __builtin_ia32_psrlv4di_mask ((__v4di) __X, |
| 5395 | (__v4di) __Y, |
| 5396 | (__v4di) |
| 5397 | _mm256_setzero_si256 (), |
| 5398 | (__mmask8) __U); |
| 5399 | } |
| 5400 | |
| 5401 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5402 | _mm_mask_srlv_epi32 (__m128i __W, __mmask8 __U, __m128i __X, |
| 5403 | __m128i __Y) |
| 5404 | { |
| 5405 | return (__m128i) __builtin_ia32_psrlv4si_mask ((__v4si) __X, |
| 5406 | (__v4si) __Y, |
| 5407 | (__v4si) __W, |
| 5408 | (__mmask8) __U); |
| 5409 | } |
| 5410 | |
| 5411 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5412 | _mm_maskz_srlv_epi32 (__mmask8 __U, __m128i __X, __m128i __Y) |
| 5413 | { |
| 5414 | return (__m128i) __builtin_ia32_psrlv4si_mask ((__v4si) __X, |
| 5415 | (__v4si) __Y, |
| 5416 | (__v4si) |
| 5417 | _mm_setzero_si128 (), |
| 5418 | (__mmask8) __U); |
| 5419 | } |
| 5420 | |
| 5421 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5422 | _mm256_mask_srlv_epi32 (__m256i __W, __mmask8 __U, __m256i __X, |
| 5423 | __m256i __Y) |
| 5424 | { |
| 5425 | return (__m256i) __builtin_ia32_psrlv8si_mask ((__v8si) __X, |
| 5426 | (__v8si) __Y, |
| 5427 | (__v8si) __W, |
| 5428 | (__mmask8) __U); |
| 5429 | } |
| 5430 | |
| 5431 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5432 | _mm256_maskz_srlv_epi32 (__mmask8 __U, __m256i __X, __m256i __Y) |
| 5433 | { |
| 5434 | return (__m256i) __builtin_ia32_psrlv8si_mask ((__v8si) __X, |
| 5435 | (__v8si) __Y, |
| 5436 | (__v8si) |
| 5437 | _mm256_setzero_si256 (), |
| 5438 | (__mmask8) __U); |
| 5439 | } |
| 5440 | |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 5441 | |
| 5442 | |
| 5443 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5444 | _mm_mask_srl_epi32 (__m128i __W, __mmask8 __U, __m128i __A, |
| 5445 | __m128i __B) |
| 5446 | { |
| 5447 | return (__m128i) __builtin_ia32_psrld128_mask ((__v4si) __A, |
| 5448 | (__v4si) __B, |
| 5449 | (__v4si) __W, |
| 5450 | (__mmask8) __U); |
| 5451 | } |
| 5452 | |
| 5453 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5454 | _mm_maskz_srl_epi32 (__mmask8 __U, __m128i __A, __m128i __B) |
| 5455 | { |
| 5456 | return (__m128i) __builtin_ia32_psrld128_mask ((__v4si) __A, |
| 5457 | (__v4si) __B, |
| 5458 | (__v4si) |
| 5459 | _mm_setzero_si128 (), |
| 5460 | (__mmask8) __U); |
| 5461 | } |
| 5462 | |
| 5463 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5464 | _mm256_mask_srl_epi32 (__m256i __W, __mmask8 __U, __m256i __A, |
| 5465 | __m128i __B) |
| 5466 | { |
| 5467 | return (__m256i) __builtin_ia32_psrld256_mask ((__v8si) __A, |
| 5468 | (__v4si) __B, |
| 5469 | (__v8si) __W, |
| 5470 | (__mmask8) __U); |
| 5471 | } |
| 5472 | |
| 5473 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5474 | _mm256_maskz_srl_epi32 (__mmask8 __U, __m256i __A, __m128i __B) |
| 5475 | { |
| 5476 | return (__m256i) __builtin_ia32_psrld256_mask ((__v8si) __A, |
| 5477 | (__v4si) __B, |
| 5478 | (__v8si) |
| 5479 | _mm256_setzero_si256 (), |
| 5480 | (__mmask8) __U); |
| 5481 | } |
| 5482 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5483 | #define _mm_mask_srli_epi32(W, U, A, imm) __extension__ ({ \ |
| 5484 | (__m128i)__builtin_ia32_psrldi128_mask((__v4si)(__m128i)(A), (int)(imm), \ |
| 5485 | (__v4si)(__m128i)(W), \ |
| 5486 | (__mmask8)(U)); }) |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 5487 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5488 | #define _mm_maskz_srli_epi32(U, A, imm) __extension__ ({ \ |
| 5489 | (__m128i)__builtin_ia32_psrldi128_mask((__v4si)(__m128i)(A), (int)(imm), \ |
| 5490 | (__v4si)_mm_setzero_si128(), \ |
| 5491 | (__mmask8)(U)); }) |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 5492 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5493 | #define _mm256_mask_srli_epi32(W, U, A, imm) __extension__ ({ \ |
| 5494 | (__m256i)__builtin_ia32_psrldi256_mask((__v8si)(__m256i)(A), (int)(imm), \ |
| 5495 | (__v8si)(__m256i)(W), \ |
| 5496 | (__mmask8)(U)); }) |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 5497 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5498 | #define _mm256_maskz_srli_epi32(U, A, imm) __extension__ ({ \ |
| 5499 | (__m256i)__builtin_ia32_psrldi256_mask((__v8si)(__m256i)(A), (int)(imm), \ |
| 5500 | (__v8si)_mm256_setzero_si256(), \ |
| 5501 | (__mmask8)(U)); }) |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 5502 | |
| 5503 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5504 | _mm_mask_srl_epi64 (__m128i __W, __mmask8 __U, __m128i __A, |
| 5505 | __m128i __B) |
| 5506 | { |
| 5507 | return (__m128i) __builtin_ia32_psrlq128_mask ((__v2di) __A, |
| 5508 | (__v2di) __B, |
| 5509 | (__v2di) __W, |
| 5510 | (__mmask8) __U); |
| 5511 | } |
| 5512 | |
| 5513 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5514 | _mm_maskz_srl_epi64 (__mmask8 __U, __m128i __A, __m128i __B) |
| 5515 | { |
| 5516 | return (__m128i) __builtin_ia32_psrlq128_mask ((__v2di) __A, |
| 5517 | (__v2di) __B, |
| 5518 | (__v2di) |
| 5519 | _mm_setzero_di (), |
| 5520 | (__mmask8) __U); |
| 5521 | } |
| 5522 | |
| 5523 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5524 | _mm256_mask_srl_epi64 (__m256i __W, __mmask8 __U, __m256i __A, |
| 5525 | __m128i __B) |
| 5526 | { |
| 5527 | return (__m256i) __builtin_ia32_psrlq256_mask ((__v4di) __A, |
| 5528 | (__v2di) __B, |
| 5529 | (__v4di) __W, |
| 5530 | (__mmask8) __U); |
| 5531 | } |
| 5532 | |
| 5533 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5534 | _mm256_maskz_srl_epi64 (__mmask8 __U, __m256i __A, __m128i __B) |
| 5535 | { |
| 5536 | return (__m256i) __builtin_ia32_psrlq256_mask ((__v4di) __A, |
| 5537 | (__v2di) __B, |
| 5538 | (__v4di) |
| 5539 | _mm256_setzero_si256 (), |
| 5540 | (__mmask8) __U); |
| 5541 | } |
| 5542 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5543 | #define _mm_mask_srli_epi64(W, U, A, imm) __extension__ ({ \ |
| 5544 | (__m128i)__builtin_ia32_psrlqi128_mask((__v2di)(__m128i)(A), (int)(imm), \ |
| 5545 | (__v2di)(__m128i)(W), \ |
| 5546 | (__mmask8)(U)); }) |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 5547 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5548 | #define _mm_maskz_srli_epi64(U, A, imm) __extension__ ({ \ |
| 5549 | (__m128i)__builtin_ia32_psrlqi128_mask((__v2di)(__m128i)(A), (int)(imm), \ |
| 5550 | (__v2di)_mm_setzero_si128(), \ |
| 5551 | (__mmask8)(U)); }) |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 5552 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5553 | #define _mm256_mask_srli_epi64(W, U, A, imm) __extension__ ({ \ |
| 5554 | (__m256i)__builtin_ia32_psrlqi256_mask((__v4di)(__m256i)(A), (int)(imm), \ |
| 5555 | (__v4di)(__m256i)(W), \ |
| 5556 | (__mmask8)(U)); }) |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 5557 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5558 | #define _mm256_maskz_srli_epi64(U, A, imm) __extension__ ({ \ |
| 5559 | (__m256i)__builtin_ia32_psrlqi256_mask((__v4di)(__m256i)(A), (int)(imm), \ |
| 5560 | (__v4di)_mm256_setzero_si256(), \ |
| 5561 | (__mmask8)(U)); }) |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 5562 | |
Michael Zuckerman | d15c95a | 2016-03-02 09:05:46 +0000 | [diff] [blame] | 5563 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5564 | _mm_mask_srav_epi32 (__m128i __W, __mmask8 __U, __m128i __X, |
| 5565 | __m128i __Y) |
| 5566 | { |
| 5567 | return (__m128i) __builtin_ia32_psrav4si_mask ((__v4si) __X, |
| 5568 | (__v4si) __Y, |
| 5569 | (__v4si) __W, |
| 5570 | (__mmask8) __U); |
| 5571 | } |
| 5572 | |
| 5573 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5574 | _mm_maskz_srav_epi32 (__mmask8 __U, __m128i __X, __m128i __Y) |
| 5575 | { |
| 5576 | return (__m128i) __builtin_ia32_psrav4si_mask ((__v4si) __X, |
| 5577 | (__v4si) __Y, |
| 5578 | (__v4si) |
| 5579 | _mm_setzero_si128 (), |
| 5580 | (__mmask8) __U); |
| 5581 | } |
| 5582 | |
| 5583 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5584 | _mm256_mask_srav_epi32 (__m256i __W, __mmask8 __U, __m256i __X, |
| 5585 | __m256i __Y) |
| 5586 | { |
| 5587 | return (__m256i) __builtin_ia32_psrav8si_mask ((__v8si) __X, |
| 5588 | (__v8si) __Y, |
| 5589 | (__v8si) __W, |
| 5590 | (__mmask8) __U); |
| 5591 | } |
| 5592 | |
| 5593 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5594 | _mm256_maskz_srav_epi32 (__mmask8 __U, __m256i __X, __m256i __Y) |
| 5595 | { |
| 5596 | return (__m256i) __builtin_ia32_psrav8si_mask ((__v8si) __X, |
| 5597 | (__v8si) __Y, |
| 5598 | (__v8si) |
| 5599 | _mm256_setzero_si256 (), |
| 5600 | (__mmask8) __U); |
| 5601 | } |
| 5602 | |
| 5603 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5604 | _mm_srav_epi64 (__m128i __X, __m128i __Y) |
| 5605 | { |
| 5606 | return (__m128i) __builtin_ia32_psravq128_mask ((__v2di) __X, |
| 5607 | (__v2di) __Y, |
| 5608 | (__v2di) |
| 5609 | _mm_setzero_di (), |
| 5610 | (__mmask8) -1); |
| 5611 | } |
| 5612 | |
| 5613 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5614 | _mm_mask_srav_epi64 (__m128i __W, __mmask8 __U, __m128i __X, |
| 5615 | __m128i __Y) |
| 5616 | { |
| 5617 | return (__m128i) __builtin_ia32_psravq128_mask ((__v2di) __X, |
| 5618 | (__v2di) __Y, |
| 5619 | (__v2di) __W, |
| 5620 | (__mmask8) __U); |
| 5621 | } |
| 5622 | |
| 5623 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5624 | _mm_maskz_srav_epi64 (__mmask8 __U, __m128i __X, __m128i __Y) |
| 5625 | { |
| 5626 | return (__m128i) __builtin_ia32_psravq128_mask ((__v2di) __X, |
| 5627 | (__v2di) __Y, |
| 5628 | (__v2di) |
| 5629 | _mm_setzero_di (), |
| 5630 | (__mmask8) __U); |
| 5631 | } |
| 5632 | |
| 5633 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5634 | _mm256_srav_epi64 (__m256i __X, __m256i __Y) |
| 5635 | { |
| 5636 | return (__m256i) __builtin_ia32_psravq256_mask ((__v4di) __X, |
| 5637 | (__v4di) __Y, |
| 5638 | (__v4di) |
| 5639 | _mm256_setzero_si256 (), |
| 5640 | (__mmask8) -1); |
| 5641 | } |
| 5642 | |
| 5643 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5644 | _mm256_mask_srav_epi64 (__m256i __W, __mmask8 __U, __m256i __X, |
| 5645 | __m256i __Y) |
| 5646 | { |
| 5647 | return (__m256i) __builtin_ia32_psravq256_mask ((__v4di) __X, |
| 5648 | (__v4di) __Y, |
| 5649 | (__v4di) __W, |
| 5650 | (__mmask8) __U); |
| 5651 | } |
| 5652 | |
| 5653 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5654 | _mm256_maskz_srav_epi64 (__mmask8 __U, __m256i __X, __m256i __Y) |
| 5655 | { |
| 5656 | return (__m256i) __builtin_ia32_psravq256_mask ((__v4di) __X, |
| 5657 | (__v4di) __Y, |
| 5658 | (__v4di) |
| 5659 | _mm256_setzero_si256 (), |
| 5660 | (__mmask8) __U); |
| 5661 | } |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5662 | |
Michael Zuckerman | e654200 | 2016-05-23 08:01:48 +0000 | [diff] [blame] | 5663 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5664 | _mm_mask_mov_epi32 (__m128i __W, __mmask8 __U, __m128i __A) |
| 5665 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 5666 | return (__m128i) __builtin_ia32_selectd_128 ((__mmask8) __U, |
| 5667 | (__v4si) __A, |
| 5668 | (__v4si) __W); |
Michael Zuckerman | e654200 | 2016-05-23 08:01:48 +0000 | [diff] [blame] | 5669 | } |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5670 | |
Michael Zuckerman | e654200 | 2016-05-23 08:01:48 +0000 | [diff] [blame] | 5671 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5672 | _mm_maskz_mov_epi32 (__mmask8 __U, __m128i __A) |
| 5673 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 5674 | return (__m128i) __builtin_ia32_selectd_128 ((__mmask8) __U, |
| 5675 | (__v4si) __A, |
| 5676 | (__v4si) _mm_setzero_si128 ()); |
Michael Zuckerman | e654200 | 2016-05-23 08:01:48 +0000 | [diff] [blame] | 5677 | } |
| 5678 | |
| 5679 | |
| 5680 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5681 | _mm256_mask_mov_epi32 (__m256i __W, __mmask8 __U, __m256i __A) |
| 5682 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 5683 | return (__m256i) __builtin_ia32_selectd_256 ((__mmask8) __U, |
| 5684 | (__v8si) __A, |
| 5685 | (__v8si) __W); |
Michael Zuckerman | e654200 | 2016-05-23 08:01:48 +0000 | [diff] [blame] | 5686 | } |
| 5687 | |
| 5688 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5689 | _mm256_maskz_mov_epi32 (__mmask8 __U, __m256i __A) |
| 5690 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 5691 | return (__m256i) __builtin_ia32_selectd_256 ((__mmask8) __U, |
| 5692 | (__v8si) __A, |
| 5693 | (__v8si) _mm256_setzero_si256 ()); |
Michael Zuckerman | e654200 | 2016-05-23 08:01:48 +0000 | [diff] [blame] | 5694 | } |
| 5695 | |
| 5696 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5697 | _mm_mask_load_epi32 (__m128i __W, __mmask8 __U, void const *__P) |
| 5698 | { |
| 5699 | return (__m128i) __builtin_ia32_movdqa32load128_mask ((__v4si *) __P, |
| 5700 | (__v4si) __W, |
| 5701 | (__mmask8) |
| 5702 | __U); |
| 5703 | } |
| 5704 | |
| 5705 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5706 | _mm_maskz_load_epi32 (__mmask8 __U, void const *__P) |
| 5707 | { |
| 5708 | return (__m128i) __builtin_ia32_movdqa32load128_mask ((__v4si *) __P, |
| 5709 | (__v4si) |
| 5710 | _mm_setzero_si128 (), |
| 5711 | (__mmask8) |
| 5712 | __U); |
| 5713 | } |
| 5714 | |
| 5715 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5716 | _mm256_mask_load_epi32 (__m256i __W, __mmask8 __U, void const *__P) |
| 5717 | { |
| 5718 | return (__m256i) __builtin_ia32_movdqa32load256_mask ((__v8si *) __P, |
| 5719 | (__v8si) __W, |
| 5720 | (__mmask8) |
| 5721 | __U); |
| 5722 | } |
| 5723 | |
| 5724 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5725 | _mm256_maskz_load_epi32 (__mmask8 __U, void const *__P) |
| 5726 | { |
| 5727 | return (__m256i) __builtin_ia32_movdqa32load256_mask ((__v8si *) __P, |
| 5728 | (__v8si) |
| 5729 | _mm256_setzero_si256 (), |
| 5730 | (__mmask8) |
| 5731 | __U); |
| 5732 | } |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5733 | |
| 5734 | static __inline__ void __DEFAULT_FN_ATTRS |
| 5735 | _mm_mask_store_epi32 (void *__P, __mmask8 __U, __m128i __A) |
| 5736 | { |
| 5737 | __builtin_ia32_movdqa32store128_mask ((__v4si *) __P, |
| 5738 | (__v4si) __A, |
| 5739 | (__mmask8) __U); |
| 5740 | } |
| 5741 | |
| 5742 | static __inline__ void __DEFAULT_FN_ATTRS |
| 5743 | _mm256_mask_store_epi32 (void *__P, __mmask8 __U, __m256i __A) |
| 5744 | { |
| 5745 | __builtin_ia32_movdqa32store256_mask ((__v8si *) __P, |
| 5746 | (__v8si) __A, |
| 5747 | (__mmask8) __U); |
| 5748 | } |
| 5749 | |
| 5750 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5751 | _mm_mask_mov_epi64 (__m128i __W, __mmask8 __U, __m128i __A) |
| 5752 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 5753 | return (__m128i) __builtin_ia32_selectq_128 ((__mmask8) __U, |
| 5754 | (__v2di) __A, |
| 5755 | (__v2di) __W); |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5756 | } |
| 5757 | |
| 5758 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5759 | _mm_maskz_mov_epi64 (__mmask8 __U, __m128i __A) |
| 5760 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 5761 | return (__m128i) __builtin_ia32_selectq_128 ((__mmask8) __U, |
| 5762 | (__v2di) __A, |
| 5763 | (__v2di) _mm_setzero_di ()); |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5764 | } |
| 5765 | |
| 5766 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5767 | _mm256_mask_mov_epi64 (__m256i __W, __mmask8 __U, __m256i __A) |
| 5768 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 5769 | return (__m256i) __builtin_ia32_selectq_256 ((__mmask8) __U, |
| 5770 | (__v4di) __A, |
| 5771 | (__v4di) __W); |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5772 | } |
| 5773 | |
| 5774 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5775 | _mm256_maskz_mov_epi64 (__mmask8 __U, __m256i __A) |
| 5776 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 5777 | return (__m256i) __builtin_ia32_selectq_256 ((__mmask8) __U, |
| 5778 | (__v4di) __A, |
| 5779 | (__v4di) _mm256_setzero_si256 ()); |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5780 | } |
| 5781 | |
| 5782 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5783 | _mm_mask_load_epi64 (__m128i __W, __mmask8 __U, void const *__P) |
| 5784 | { |
| 5785 | return (__m128i) __builtin_ia32_movdqa64load128_mask ((__v2di *) __P, |
| 5786 | (__v2di) __W, |
| 5787 | (__mmask8) |
| 5788 | __U); |
| 5789 | } |
| 5790 | |
| 5791 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5792 | _mm_maskz_load_epi64 (__mmask8 __U, void const *__P) |
| 5793 | { |
| 5794 | return (__m128i) __builtin_ia32_movdqa64load128_mask ((__v2di *) __P, |
| 5795 | (__v2di) |
| 5796 | _mm_setzero_di (), |
| 5797 | (__mmask8) |
| 5798 | __U); |
| 5799 | } |
| 5800 | |
| 5801 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5802 | _mm256_mask_load_epi64 (__m256i __W, __mmask8 __U, void const *__P) |
| 5803 | { |
| 5804 | return (__m256i) __builtin_ia32_movdqa64load256_mask ((__v4di *) __P, |
| 5805 | (__v4di) __W, |
| 5806 | (__mmask8) |
| 5807 | __U); |
| 5808 | } |
| 5809 | |
| 5810 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5811 | _mm256_maskz_load_epi64 (__mmask8 __U, void const *__P) |
| 5812 | { |
| 5813 | return (__m256i) __builtin_ia32_movdqa64load256_mask ((__v4di *) __P, |
| 5814 | (__v4di) |
| 5815 | _mm256_setzero_si256 (), |
| 5816 | (__mmask8) |
| 5817 | __U); |
| 5818 | } |
| 5819 | |
| 5820 | static __inline__ void __DEFAULT_FN_ATTRS |
| 5821 | _mm_mask_store_epi64 (void *__P, __mmask8 __U, __m128i __A) |
| 5822 | { |
| 5823 | __builtin_ia32_movdqa64store128_mask ((__v2di *) __P, |
| 5824 | (__v2di) __A, |
| 5825 | (__mmask8) __U); |
| 5826 | } |
| 5827 | |
| 5828 | static __inline__ void __DEFAULT_FN_ATTRS |
| 5829 | _mm256_mask_store_epi64 (void *__P, __mmask8 __U, __m256i __A) |
| 5830 | { |
| 5831 | __builtin_ia32_movdqa64store256_mask ((__v4di *) __P, |
| 5832 | (__v4di) __A, |
| 5833 | (__mmask8) __U); |
| 5834 | } |
| 5835 | |
Michael Zuckerman | 0d67e4b | 2016-03-03 13:43:05 +0000 | [diff] [blame] | 5836 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 5837 | _mm_mask_movedup_pd (__m128d __W, __mmask8 __U, __m128d __A) |
| 5838 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 5839 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 5840 | (__v2df)_mm_movedup_pd(__A), |
| 5841 | (__v2df)__W); |
Michael Zuckerman | 0d67e4b | 2016-03-03 13:43:05 +0000 | [diff] [blame] | 5842 | } |
| 5843 | |
| 5844 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 5845 | _mm_maskz_movedup_pd (__mmask8 __U, __m128d __A) |
| 5846 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 5847 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 5848 | (__v2df)_mm_movedup_pd(__A), |
| 5849 | (__v2df)_mm_setzero_pd()); |
Michael Zuckerman | 0d67e4b | 2016-03-03 13:43:05 +0000 | [diff] [blame] | 5850 | } |
| 5851 | |
| 5852 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 5853 | _mm256_mask_movedup_pd (__m256d __W, __mmask8 __U, __m256d __A) |
| 5854 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 5855 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 5856 | (__v4df)_mm256_movedup_pd(__A), |
| 5857 | (__v4df)__W); |
Michael Zuckerman | 0d67e4b | 2016-03-03 13:43:05 +0000 | [diff] [blame] | 5858 | } |
| 5859 | |
| 5860 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 5861 | _mm256_maskz_movedup_pd (__mmask8 __U, __m256d __A) |
| 5862 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 5863 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 5864 | (__v4df)_mm256_movedup_pd(__A), |
| 5865 | (__v4df)_mm256_setzero_pd()); |
Michael Zuckerman | 0d67e4b | 2016-03-03 13:43:05 +0000 | [diff] [blame] | 5866 | } |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5867 | |
Michael Zuckerman | 912be16 | 2016-03-07 08:29:10 +0000 | [diff] [blame] | 5868 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5869 | #define _mm_mask_set1_epi32(O, M, A) __extension__ ({ \ |
| 5870 | (__m128i)__builtin_ia32_pbroadcastd128_gpr_mask((int)(A), \ |
| 5871 | (__v4si)(__m128i)(O), \ |
| 5872 | (__mmask8)(M)); }) |
Michael Zuckerman | 912be16 | 2016-03-07 08:29:10 +0000 | [diff] [blame] | 5873 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5874 | #define _mm_maskz_set1_epi32(M, A) __extension__ ({ \ |
| 5875 | (__m128i)__builtin_ia32_pbroadcastd128_gpr_mask((int)(A), \ |
| 5876 | (__v4si)_mm_setzero_si128(), \ |
| 5877 | (__mmask8)(M)); }) |
Michael Zuckerman | 912be16 | 2016-03-07 08:29:10 +0000 | [diff] [blame] | 5878 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5879 | #define _mm256_mask_set1_epi32(O, M, A) __extension__ ({ \ |
| 5880 | (__m256i)__builtin_ia32_pbroadcastd256_gpr_mask((int)(A), \ |
| 5881 | (__v8si)(__m256i)(O), \ |
| 5882 | (__mmask8)(M)); }) |
Michael Zuckerman | 912be16 | 2016-03-07 08:29:10 +0000 | [diff] [blame] | 5883 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5884 | #define _mm256_maskz_set1_epi32(M, A) __extension__ ({ \ |
| 5885 | (__m256i)__builtin_ia32_pbroadcastd256_gpr_mask((int)(A), \ |
| 5886 | (__v8si)_mm256_setzero_si256(), \ |
| 5887 | (__mmask8)(M)); }) |
Michael Zuckerman | 912be16 | 2016-03-07 08:29:10 +0000 | [diff] [blame] | 5888 | |
Craig Topper | 45db56c | 2016-07-21 07:38:39 +0000 | [diff] [blame] | 5889 | #ifdef __x86_64__ |
Michael Zuckerman | 912be16 | 2016-03-07 08:29:10 +0000 | [diff] [blame] | 5890 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5891 | _mm_mask_set1_epi64 (__m128i __O, __mmask8 __M, long long __A) |
| 5892 | { |
| 5893 | return (__m128i) __builtin_ia32_pbroadcastq128_gpr_mask (__A, (__v2di) __O, |
| 5894 | __M); |
| 5895 | } |
| 5896 | |
| 5897 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5898 | _mm_maskz_set1_epi64 (__mmask8 __M, long long __A) |
| 5899 | { |
| 5900 | return (__m128i) __builtin_ia32_pbroadcastq128_gpr_mask (__A, |
| 5901 | (__v2di) |
| 5902 | _mm_setzero_si128 (), |
| 5903 | __M); |
| 5904 | } |
| 5905 | |
| 5906 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5907 | _mm256_mask_set1_epi64 (__m256i __O, __mmask8 __M, long long __A) |
| 5908 | { |
| 5909 | return (__m256i) __builtin_ia32_pbroadcastq256_gpr_mask (__A, (__v4di) __O, |
| 5910 | __M); |
| 5911 | } |
| 5912 | |
| 5913 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5914 | _mm256_maskz_set1_epi64 (__mmask8 __M, long long __A) |
| 5915 | { |
| 5916 | return (__m256i) __builtin_ia32_pbroadcastq256_gpr_mask (__A, |
| 5917 | (__v4di) |
| 5918 | _mm256_setzero_si256 (), |
| 5919 | __M); |
| 5920 | } |
Craig Topper | 45db56c | 2016-07-21 07:38:39 +0000 | [diff] [blame] | 5921 | #endif |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5922 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5923 | #define _mm_fixupimm_pd(A, B, C, imm) __extension__ ({ \ |
| 5924 | (__m128d)__builtin_ia32_fixupimmpd128_mask((__v2df)(__m128d)(A), \ |
| 5925 | (__v2df)(__m128d)(B), \ |
| 5926 | (__v2di)(__m128i)(C), (int)(imm), \ |
| 5927 | (__mmask8)-1); }) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5928 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5929 | #define _mm_mask_fixupimm_pd(A, U, B, C, imm) __extension__ ({ \ |
| 5930 | (__m128d)__builtin_ia32_fixupimmpd128_mask((__v2df)(__m128d)(A), \ |
| 5931 | (__v2df)(__m128d)(B), \ |
| 5932 | (__v2di)(__m128i)(C), (int)(imm), \ |
| 5933 | (__mmask8)(U)); }) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5934 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5935 | #define _mm_maskz_fixupimm_pd(U, A, B, C, imm) __extension__ ({ \ |
| 5936 | (__m128d)__builtin_ia32_fixupimmpd128_maskz((__v2df)(__m128d)(A), \ |
| 5937 | (__v2df)(__m128d)(B), \ |
| 5938 | (__v2di)(__m128i)(C), \ |
| 5939 | (int)(imm), (__mmask8)(U)); }) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5940 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5941 | #define _mm256_fixupimm_pd(A, B, C, imm) __extension__ ({ \ |
| 5942 | (__m256d)__builtin_ia32_fixupimmpd256_mask((__v4df)(__m256d)(A), \ |
| 5943 | (__v4df)(__m256d)(B), \ |
| 5944 | (__v4di)(__m256i)(C), (int)(imm), \ |
| 5945 | (__mmask8)-1); }) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5946 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5947 | #define _mm256_mask_fixupimm_pd(A, U, B, C, imm) __extension__ ({ \ |
| 5948 | (__m256d)__builtin_ia32_fixupimmpd256_mask((__v4df)(__m256d)(A), \ |
| 5949 | (__v4df)(__m256d)(B), \ |
| 5950 | (__v4di)(__m256i)(C), (int)(imm), \ |
| 5951 | (__mmask8)(U)); }) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5952 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5953 | #define _mm256_maskz_fixupimm_pd(U, A, B, C, imm) __extension__ ({ \ |
| 5954 | (__m256d)__builtin_ia32_fixupimmpd256_maskz((__v4df)(__m256d)(A), \ |
| 5955 | (__v4df)(__m256d)(B), \ |
| 5956 | (__v4di)(__m256i)(C), \ |
| 5957 | (int)(imm), (__mmask8)(U)); }) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5958 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5959 | #define _mm_fixupimm_ps(A, B, C, imm) __extension__ ({ \ |
| 5960 | (__m128)__builtin_ia32_fixupimmps128_mask((__v4sf)(__m128)(A), \ |
| 5961 | (__v4sf)(__m128)(B), \ |
| 5962 | (__v4si)(__m128i)(C), (int)(imm), \ |
| 5963 | (__mmask8)-1); }) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5964 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5965 | #define _mm_mask_fixupimm_ps(A, U, B, C, imm) __extension__ ({ \ |
| 5966 | (__m128)__builtin_ia32_fixupimmps128_mask((__v4sf)(__m128)(A), \ |
| 5967 | (__v4sf)(__m128)(B), \ |
| 5968 | (__v4si)(__m128i)(C), (int)(imm), \ |
| 5969 | (__mmask8)(U)); }) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5970 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5971 | #define _mm_maskz_fixupimm_ps(U, A, B, C, imm) __extension__ ({ \ |
| 5972 | (__m128)__builtin_ia32_fixupimmps128_maskz((__v4sf)(__m128)(A), \ |
| 5973 | (__v4sf)(__m128)(B), \ |
| 5974 | (__v4si)(__m128i)(C), (int)(imm), \ |
| 5975 | (__mmask8)(U)); }) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5976 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5977 | #define _mm256_fixupimm_ps(A, B, C, imm) __extension__ ({ \ |
| 5978 | (__m256)__builtin_ia32_fixupimmps256_mask((__v8sf)(__m256)(A), \ |
| 5979 | (__v8sf)(__m256)(B), \ |
| 5980 | (__v8si)(__m256i)(C), (int)(imm), \ |
| 5981 | (__mmask8)-1); }) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5982 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5983 | #define _mm256_mask_fixupimm_ps(A, U, B, C, imm) __extension__ ({ \ |
| 5984 | (__m256)__builtin_ia32_fixupimmps256_mask((__v8sf)(__m256)(A), \ |
| 5985 | (__v8sf)(__m256)(B), \ |
| 5986 | (__v8si)(__m256i)(C), (int)(imm), \ |
| 5987 | (__mmask8)(U)); }) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5988 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5989 | #define _mm256_maskz_fixupimm_ps(U, A, B, C, imm) __extension__ ({ \ |
| 5990 | (__m256)__builtin_ia32_fixupimmps256_maskz((__v8sf)(__m256)(A), \ |
| 5991 | (__v8sf)(__m256)(B), \ |
| 5992 | (__v8si)(__m256i)(C), (int)(imm), \ |
| 5993 | (__mmask8)(U)); }) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5994 | |
| 5995 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 5996 | _mm_mask_load_pd (__m128d __W, __mmask8 __U, void const *__P) |
| 5997 | { |
| 5998 | return (__m128d) __builtin_ia32_loadapd128_mask ((__v2df *) __P, |
| 5999 | (__v2df) __W, |
| 6000 | (__mmask8) __U); |
| 6001 | } |
| 6002 | |
| 6003 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 6004 | _mm_maskz_load_pd (__mmask8 __U, void const *__P) |
| 6005 | { |
| 6006 | return (__m128d) __builtin_ia32_loadapd128_mask ((__v2df *) __P, |
| 6007 | (__v2df) |
| 6008 | _mm_setzero_pd (), |
| 6009 | (__mmask8) __U); |
| 6010 | } |
| 6011 | |
| 6012 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 6013 | _mm256_mask_load_pd (__m256d __W, __mmask8 __U, void const *__P) |
| 6014 | { |
| 6015 | return (__m256d) __builtin_ia32_loadapd256_mask ((__v4df *) __P, |
| 6016 | (__v4df) __W, |
| 6017 | (__mmask8) __U); |
| 6018 | } |
| 6019 | |
| 6020 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 6021 | _mm256_maskz_load_pd (__mmask8 __U, void const *__P) |
| 6022 | { |
| 6023 | return (__m256d) __builtin_ia32_loadapd256_mask ((__v4df *) __P, |
| 6024 | (__v4df) |
| 6025 | _mm256_setzero_pd (), |
| 6026 | (__mmask8) __U); |
| 6027 | } |
| 6028 | |
| 6029 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 6030 | _mm_mask_load_ps (__m128 __W, __mmask8 __U, void const *__P) |
| 6031 | { |
| 6032 | return (__m128) __builtin_ia32_loadaps128_mask ((__v4sf *) __P, |
| 6033 | (__v4sf) __W, |
| 6034 | (__mmask8) __U); |
| 6035 | } |
| 6036 | |
| 6037 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 6038 | _mm_maskz_load_ps (__mmask8 __U, void const *__P) |
| 6039 | { |
| 6040 | return (__m128) __builtin_ia32_loadaps128_mask ((__v4sf *) __P, |
| 6041 | (__v4sf) |
| 6042 | _mm_setzero_ps (), |
| 6043 | (__mmask8) __U); |
| 6044 | } |
| 6045 | |
| 6046 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 6047 | _mm256_mask_load_ps (__m256 __W, __mmask8 __U, void const *__P) |
| 6048 | { |
| 6049 | return (__m256) __builtin_ia32_loadaps256_mask ((__v8sf *) __P, |
| 6050 | (__v8sf) __W, |
| 6051 | (__mmask8) __U); |
| 6052 | } |
| 6053 | |
| 6054 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 6055 | _mm256_maskz_load_ps (__mmask8 __U, void const *__P) |
| 6056 | { |
| 6057 | return (__m256) __builtin_ia32_loadaps256_mask ((__v8sf *) __P, |
| 6058 | (__v8sf) |
| 6059 | _mm256_setzero_ps (), |
| 6060 | (__mmask8) __U); |
| 6061 | } |
| 6062 | |
| 6063 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6064 | _mm_mask_loadu_epi64 (__m128i __W, __mmask8 __U, void const *__P) |
| 6065 | { |
| 6066 | return (__m128i) __builtin_ia32_loaddqudi128_mask ((__v2di *) __P, |
| 6067 | (__v2di) __W, |
| 6068 | (__mmask8) __U); |
| 6069 | } |
| 6070 | |
| 6071 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6072 | _mm_maskz_loadu_epi64 (__mmask8 __U, void const *__P) |
| 6073 | { |
| 6074 | return (__m128i) __builtin_ia32_loaddqudi128_mask ((__v2di *) __P, |
| 6075 | (__v2di) |
| 6076 | _mm_setzero_si128 (), |
| 6077 | (__mmask8) __U); |
| 6078 | } |
| 6079 | |
| 6080 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 6081 | _mm256_mask_loadu_epi64 (__m256i __W, __mmask8 __U, void const *__P) |
| 6082 | { |
| 6083 | return (__m256i) __builtin_ia32_loaddqudi256_mask ((__v4di *) __P, |
| 6084 | (__v4di) __W, |
| 6085 | (__mmask8) __U); |
| 6086 | } |
| 6087 | |
| 6088 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 6089 | _mm256_maskz_loadu_epi64 (__mmask8 __U, void const *__P) |
| 6090 | { |
| 6091 | return (__m256i) __builtin_ia32_loaddqudi256_mask ((__v4di *) __P, |
| 6092 | (__v4di) |
| 6093 | _mm256_setzero_si256 (), |
| 6094 | (__mmask8) __U); |
| 6095 | } |
| 6096 | |
| 6097 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6098 | _mm_mask_loadu_epi32 (__m128i __W, __mmask8 __U, void const *__P) |
| 6099 | { |
| 6100 | return (__m128i) __builtin_ia32_loaddqusi128_mask ((__v4si *) __P, |
| 6101 | (__v4si) __W, |
| 6102 | (__mmask8) __U); |
| 6103 | } |
| 6104 | |
| 6105 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6106 | _mm_maskz_loadu_epi32 (__mmask8 __U, void const *__P) |
| 6107 | { |
| 6108 | return (__m128i) __builtin_ia32_loaddqusi128_mask ((__v4si *) __P, |
| 6109 | (__v4si) |
| 6110 | _mm_setzero_si128 (), |
| 6111 | (__mmask8) __U); |
| 6112 | } |
| 6113 | |
| 6114 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 6115 | _mm256_mask_loadu_epi32 (__m256i __W, __mmask8 __U, void const *__P) |
| 6116 | { |
| 6117 | return (__m256i) __builtin_ia32_loaddqusi256_mask ((__v8si *) __P, |
| 6118 | (__v8si) __W, |
| 6119 | (__mmask8) __U); |
| 6120 | } |
| 6121 | |
| 6122 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 6123 | _mm256_maskz_loadu_epi32 (__mmask8 __U, void const *__P) |
| 6124 | { |
| 6125 | return (__m256i) __builtin_ia32_loaddqusi256_mask ((__v8si *) __P, |
| 6126 | (__v8si) |
| 6127 | _mm256_setzero_si256 (), |
| 6128 | (__mmask8) __U); |
| 6129 | } |
| 6130 | |
| 6131 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 6132 | _mm_mask_loadu_pd (__m128d __W, __mmask8 __U, void const *__P) |
| 6133 | { |
| 6134 | return (__m128d) __builtin_ia32_loadupd128_mask ((__v2df *) __P, |
| 6135 | (__v2df) __W, |
| 6136 | (__mmask8) __U); |
| 6137 | } |
| 6138 | |
| 6139 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 6140 | _mm_maskz_loadu_pd (__mmask8 __U, void const *__P) |
| 6141 | { |
| 6142 | return (__m128d) __builtin_ia32_loadupd128_mask ((__v2df *) __P, |
| 6143 | (__v2df) |
| 6144 | _mm_setzero_pd (), |
| 6145 | (__mmask8) __U); |
| 6146 | } |
| 6147 | |
| 6148 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 6149 | _mm256_mask_loadu_pd (__m256d __W, __mmask8 __U, void const *__P) |
| 6150 | { |
| 6151 | return (__m256d) __builtin_ia32_loadupd256_mask ((__v4df *) __P, |
| 6152 | (__v4df) __W, |
| 6153 | (__mmask8) __U); |
| 6154 | } |
| 6155 | |
| 6156 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 6157 | _mm256_maskz_loadu_pd (__mmask8 __U, void const *__P) |
| 6158 | { |
| 6159 | return (__m256d) __builtin_ia32_loadupd256_mask ((__v4df *) __P, |
| 6160 | (__v4df) |
| 6161 | _mm256_setzero_pd (), |
| 6162 | (__mmask8) __U); |
| 6163 | } |
| 6164 | |
| 6165 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 6166 | _mm_mask_loadu_ps (__m128 __W, __mmask8 __U, void const *__P) |
| 6167 | { |
| 6168 | return (__m128) __builtin_ia32_loadups128_mask ((__v4sf *) __P, |
| 6169 | (__v4sf) __W, |
| 6170 | (__mmask8) __U); |
| 6171 | } |
| 6172 | |
| 6173 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 6174 | _mm_maskz_loadu_ps (__mmask8 __U, void const *__P) |
| 6175 | { |
| 6176 | return (__m128) __builtin_ia32_loadups128_mask ((__v4sf *) __P, |
| 6177 | (__v4sf) |
| 6178 | _mm_setzero_ps (), |
| 6179 | (__mmask8) __U); |
| 6180 | } |
| 6181 | |
| 6182 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 6183 | _mm256_mask_loadu_ps (__m256 __W, __mmask8 __U, void const *__P) |
| 6184 | { |
| 6185 | return (__m256) __builtin_ia32_loadups256_mask ((__v8sf *) __P, |
| 6186 | (__v8sf) __W, |
| 6187 | (__mmask8) __U); |
| 6188 | } |
| 6189 | |
| 6190 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 6191 | _mm256_maskz_loadu_ps (__mmask8 __U, void const *__P) |
| 6192 | { |
| 6193 | return (__m256) __builtin_ia32_loadups256_mask ((__v8sf *) __P, |
| 6194 | (__v8sf) |
| 6195 | _mm256_setzero_ps (), |
| 6196 | (__mmask8) __U); |
| 6197 | } |
Michael Zuckerman | fa7ccc5 | 2016-04-10 10:51:04 +0000 | [diff] [blame] | 6198 | |
| 6199 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6200 | _mm_mask_store_pd (void *__P, __mmask8 __U, __m128d __A) |
| 6201 | { |
| 6202 | __builtin_ia32_storeapd128_mask ((__v2df *) __P, |
| 6203 | (__v2df) __A, |
| 6204 | (__mmask8) __U); |
| 6205 | } |
| 6206 | |
| 6207 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6208 | _mm256_mask_store_pd (void *__P, __mmask8 __U, __m256d __A) |
| 6209 | { |
| 6210 | __builtin_ia32_storeapd256_mask ((__v4df *) __P, |
| 6211 | (__v4df) __A, |
| 6212 | (__mmask8) __U); |
| 6213 | } |
| 6214 | |
| 6215 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6216 | _mm_mask_store_ps (void *__P, __mmask8 __U, __m128 __A) |
| 6217 | { |
| 6218 | __builtin_ia32_storeaps128_mask ((__v4sf *) __P, |
| 6219 | (__v4sf) __A, |
| 6220 | (__mmask8) __U); |
| 6221 | } |
| 6222 | |
| 6223 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6224 | _mm256_mask_store_ps (void *__P, __mmask8 __U, __m256 __A) |
| 6225 | { |
| 6226 | __builtin_ia32_storeaps256_mask ((__v8sf *) __P, |
| 6227 | (__v8sf) __A, |
| 6228 | (__mmask8) __U); |
| 6229 | } |
| 6230 | |
| 6231 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6232 | _mm_mask_storeu_epi64 (void *__P, __mmask8 __U, __m128i __A) |
| 6233 | { |
| 6234 | __builtin_ia32_storedqudi128_mask ((__v2di *) __P, |
| 6235 | (__v2di) __A, |
| 6236 | (__mmask8) __U); |
| 6237 | } |
| 6238 | |
| 6239 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6240 | _mm256_mask_storeu_epi64 (void *__P, __mmask8 __U, __m256i __A) |
| 6241 | { |
| 6242 | __builtin_ia32_storedqudi256_mask ((__v4di *) __P, |
| 6243 | (__v4di) __A, |
| 6244 | (__mmask8) __U); |
| 6245 | } |
| 6246 | |
| 6247 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6248 | _mm_mask_storeu_epi32 (void *__P, __mmask8 __U, __m128i __A) |
| 6249 | { |
| 6250 | __builtin_ia32_storedqusi128_mask ((__v4si *) __P, |
| 6251 | (__v4si) __A, |
| 6252 | (__mmask8) __U); |
| 6253 | } |
| 6254 | |
| 6255 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6256 | _mm256_mask_storeu_epi32 (void *__P, __mmask8 __U, __m256i __A) |
| 6257 | { |
| 6258 | __builtin_ia32_storedqusi256_mask ((__v8si *) __P, |
| 6259 | (__v8si) __A, |
| 6260 | (__mmask8) __U); |
| 6261 | } |
| 6262 | |
| 6263 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6264 | _mm_mask_storeu_pd (void *__P, __mmask8 __U, __m128d __A) |
| 6265 | { |
| 6266 | __builtin_ia32_storeupd128_mask ((__v2df *) __P, |
| 6267 | (__v2df) __A, |
| 6268 | (__mmask8) __U); |
| 6269 | } |
| 6270 | |
| 6271 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6272 | _mm256_mask_storeu_pd (void *__P, __mmask8 __U, __m256d __A) |
| 6273 | { |
| 6274 | __builtin_ia32_storeupd256_mask ((__v4df *) __P, |
| 6275 | (__v4df) __A, |
| 6276 | (__mmask8) __U); |
| 6277 | } |
| 6278 | |
| 6279 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6280 | _mm_mask_storeu_ps (void *__P, __mmask8 __U, __m128 __A) |
| 6281 | { |
| 6282 | __builtin_ia32_storeups128_mask ((__v4sf *) __P, |
| 6283 | (__v4sf) __A, |
| 6284 | (__mmask8) __U); |
| 6285 | } |
| 6286 | |
| 6287 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6288 | _mm256_mask_storeu_ps (void *__P, __mmask8 __U, __m256 __A) |
| 6289 | { |
| 6290 | __builtin_ia32_storeups256_mask ((__v8sf *) __P, |
| 6291 | (__v8sf) __A, |
| 6292 | (__mmask8) __U); |
| 6293 | } |
| 6294 | |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6295 | |
| 6296 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6297 | _mm_mask_unpackhi_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6298 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6299 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 6300 | (__v2df)_mm_unpackhi_pd(__A, __B), |
| 6301 | (__v2df)__W); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6302 | } |
| 6303 | |
| 6304 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6305 | _mm_maskz_unpackhi_pd(__mmask8 __U, __m128d __A, __m128d __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6306 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6307 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 6308 | (__v2df)_mm_unpackhi_pd(__A, __B), |
| 6309 | (__v2df)_mm_setzero_pd()); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6310 | } |
| 6311 | |
| 6312 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6313 | _mm256_mask_unpackhi_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6314 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6315 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 6316 | (__v4df)_mm256_unpackhi_pd(__A, __B), |
| 6317 | (__v4df)__W); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6318 | } |
| 6319 | |
| 6320 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6321 | _mm256_maskz_unpackhi_pd(__mmask8 __U, __m256d __A, __m256d __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6322 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6323 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 6324 | (__v4df)_mm256_unpackhi_pd(__A, __B), |
| 6325 | (__v4df)_mm256_setzero_pd()); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6326 | } |
| 6327 | |
| 6328 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6329 | _mm_mask_unpackhi_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6330 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6331 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 6332 | (__v4sf)_mm_unpackhi_ps(__A, __B), |
| 6333 | (__v4sf)__W); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6334 | } |
| 6335 | |
| 6336 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6337 | _mm_maskz_unpackhi_ps(__mmask8 __U, __m128 __A, __m128 __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6338 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6339 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 6340 | (__v4sf)_mm_unpackhi_ps(__A, __B), |
| 6341 | (__v4sf)_mm_setzero_ps()); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6342 | } |
| 6343 | |
| 6344 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6345 | _mm256_mask_unpackhi_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6346 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6347 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 6348 | (__v8sf)_mm256_unpackhi_ps(__A, __B), |
| 6349 | (__v8sf)__W); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6350 | } |
| 6351 | |
| 6352 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6353 | _mm256_maskz_unpackhi_ps(__mmask8 __U, __m256 __A, __m256 __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6354 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6355 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 6356 | (__v8sf)_mm256_unpackhi_ps(__A, __B), |
| 6357 | (__v8sf)_mm256_setzero_ps()); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6358 | } |
| 6359 | |
| 6360 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6361 | _mm_mask_unpacklo_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6362 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6363 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 6364 | (__v2df)_mm_unpacklo_pd(__A, __B), |
| 6365 | (__v2df)__W); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6366 | } |
| 6367 | |
| 6368 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6369 | _mm_maskz_unpacklo_pd(__mmask8 __U, __m128d __A, __m128d __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6370 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6371 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 6372 | (__v2df)_mm_unpacklo_pd(__A, __B), |
| 6373 | (__v2df)_mm_setzero_pd()); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6374 | } |
| 6375 | |
| 6376 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6377 | _mm256_mask_unpacklo_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6378 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6379 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 6380 | (__v4df)_mm256_unpacklo_pd(__A, __B), |
| 6381 | (__v4df)__W); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6382 | } |
| 6383 | |
| 6384 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6385 | _mm256_maskz_unpacklo_pd(__mmask8 __U, __m256d __A, __m256d __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6386 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6387 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 6388 | (__v4df)_mm256_unpacklo_pd(__A, __B), |
| 6389 | (__v4df)_mm256_setzero_pd()); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6390 | } |
| 6391 | |
| 6392 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6393 | _mm_mask_unpacklo_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6394 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6395 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 6396 | (__v4sf)_mm_unpacklo_ps(__A, __B), |
| 6397 | (__v4sf)__W); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6398 | } |
| 6399 | |
| 6400 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6401 | _mm_maskz_unpacklo_ps(__mmask8 __U, __m128 __A, __m128 __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6402 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6403 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 6404 | (__v4sf)_mm_unpacklo_ps(__A, __B), |
| 6405 | (__v4sf)_mm_setzero_ps()); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6406 | } |
| 6407 | |
| 6408 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6409 | _mm256_mask_unpacklo_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6410 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6411 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 6412 | (__v8sf)_mm256_unpacklo_ps(__A, __B), |
| 6413 | (__v8sf)__W); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6414 | } |
| 6415 | |
| 6416 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6417 | _mm256_maskz_unpacklo_ps(__mmask8 __U, __m256 __A, __m256 __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6418 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6419 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 6420 | (__v8sf)_mm256_unpacklo_ps(__A, __B), |
| 6421 | (__v8sf)_mm256_setzero_ps()); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 6422 | } |
| 6423 | |
| 6424 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 6425 | _mm_rcp14_pd (__m128d __A) |
| 6426 | { |
| 6427 | return (__m128d) __builtin_ia32_rcp14pd128_mask ((__v2df) __A, |
| 6428 | (__v2df) |
| 6429 | _mm_setzero_pd (), |
| 6430 | (__mmask8) -1); |
| 6431 | } |
| 6432 | |
| 6433 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 6434 | _mm_mask_rcp14_pd (__m128d __W, __mmask8 __U, __m128d __A) |
| 6435 | { |
| 6436 | return (__m128d) __builtin_ia32_rcp14pd128_mask ((__v2df) __A, |
| 6437 | (__v2df) __W, |
| 6438 | (__mmask8) __U); |
| 6439 | } |
| 6440 | |
| 6441 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 6442 | _mm_maskz_rcp14_pd (__mmask8 __U, __m128d __A) |
| 6443 | { |
| 6444 | return (__m128d) __builtin_ia32_rcp14pd128_mask ((__v2df) __A, |
| 6445 | (__v2df) |
| 6446 | _mm_setzero_pd (), |
| 6447 | (__mmask8) __U); |
| 6448 | } |
| 6449 | |
| 6450 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 6451 | _mm256_rcp14_pd (__m256d __A) |
| 6452 | { |
| 6453 | return (__m256d) __builtin_ia32_rcp14pd256_mask ((__v4df) __A, |
| 6454 | (__v4df) |
| 6455 | _mm256_setzero_pd (), |
| 6456 | (__mmask8) -1); |
| 6457 | } |
| 6458 | |
| 6459 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 6460 | _mm256_mask_rcp14_pd (__m256d __W, __mmask8 __U, __m256d __A) |
| 6461 | { |
| 6462 | return (__m256d) __builtin_ia32_rcp14pd256_mask ((__v4df) __A, |
| 6463 | (__v4df) __W, |
| 6464 | (__mmask8) __U); |
| 6465 | } |
| 6466 | |
| 6467 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 6468 | _mm256_maskz_rcp14_pd (__mmask8 __U, __m256d __A) |
| 6469 | { |
| 6470 | return (__m256d) __builtin_ia32_rcp14pd256_mask ((__v4df) __A, |
| 6471 | (__v4df) |
| 6472 | _mm256_setzero_pd (), |
| 6473 | (__mmask8) __U); |
| 6474 | } |
| 6475 | |
| 6476 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 6477 | _mm_rcp14_ps (__m128 __A) |
| 6478 | { |
| 6479 | return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A, |
| 6480 | (__v4sf) |
| 6481 | _mm_setzero_ps (), |
| 6482 | (__mmask8) -1); |
| 6483 | } |
| 6484 | |
| 6485 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 6486 | _mm_mask_rcp14_ps (__m128 __W, __mmask8 __U, __m128 __A) |
| 6487 | { |
| 6488 | return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A, |
| 6489 | (__v4sf) __W, |
| 6490 | (__mmask8) __U); |
| 6491 | } |
| 6492 | |
| 6493 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 6494 | _mm_maskz_rcp14_ps (__mmask8 __U, __m128 __A) |
| 6495 | { |
| 6496 | return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A, |
| 6497 | (__v4sf) |
| 6498 | _mm_setzero_ps (), |
| 6499 | (__mmask8) __U); |
| 6500 | } |
| 6501 | |
| 6502 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 6503 | _mm256_rcp14_ps (__m256 __A) |
| 6504 | { |
| 6505 | return (__m256) __builtin_ia32_rcp14ps256_mask ((__v8sf) __A, |
| 6506 | (__v8sf) |
| 6507 | _mm256_setzero_ps (), |
| 6508 | (__mmask8) -1); |
| 6509 | } |
| 6510 | |
| 6511 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 6512 | _mm256_mask_rcp14_ps (__m256 __W, __mmask8 __U, __m256 __A) |
| 6513 | { |
| 6514 | return (__m256) __builtin_ia32_rcp14ps256_mask ((__v8sf) __A, |
| 6515 | (__v8sf) __W, |
| 6516 | (__mmask8) __U); |
| 6517 | } |
| 6518 | |
| 6519 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 6520 | _mm256_maskz_rcp14_ps (__mmask8 __U, __m256 __A) |
| 6521 | { |
| 6522 | return (__m256) __builtin_ia32_rcp14ps256_mask ((__v8sf) __A, |
| 6523 | (__v8sf) |
| 6524 | _mm256_setzero_ps (), |
| 6525 | (__mmask8) __U); |
| 6526 | } |
| 6527 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6528 | #define _mm_mask_permute_pd(W, U, X, C) __extension__ ({ \ |
Craig Topper | b3a4477 | 2016-07-02 05:36:43 +0000 | [diff] [blame] | 6529 | (__m128d)__builtin_ia32_selectpd_128((__mmask8)(U), \ |
| 6530 | (__v2df)_mm_permute_pd((X), (C)), \ |
| 6531 | (__v2df)(__m128d)(W)); }) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 6532 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6533 | #define _mm_maskz_permute_pd(U, X, C) __extension__ ({ \ |
Craig Topper | b3a4477 | 2016-07-02 05:36:43 +0000 | [diff] [blame] | 6534 | (__m128d)__builtin_ia32_selectpd_128((__mmask8)(U), \ |
| 6535 | (__v2df)_mm_permute_pd((X), (C)), \ |
| 6536 | (__v2df)_mm_setzero_pd()); }) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 6537 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6538 | #define _mm256_mask_permute_pd(W, U, X, C) __extension__ ({ \ |
Craig Topper | b3a4477 | 2016-07-02 05:36:43 +0000 | [diff] [blame] | 6539 | (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \ |
| 6540 | (__v4df)_mm256_permute_pd((X), (C)), \ |
| 6541 | (__v4df)(__m256d)(W)); }) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 6542 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6543 | #define _mm256_maskz_permute_pd(U, X, C) __extension__ ({ \ |
Craig Topper | b3a4477 | 2016-07-02 05:36:43 +0000 | [diff] [blame] | 6544 | (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \ |
| 6545 | (__v4df)_mm256_permute_pd((X), (C)), \ |
| 6546 | (__v4df)_mm256_setzero_pd()); }) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 6547 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6548 | #define _mm_mask_permute_ps(W, U, X, C) __extension__ ({ \ |
Craig Topper | b3a4477 | 2016-07-02 05:36:43 +0000 | [diff] [blame] | 6549 | (__m128)__builtin_ia32_selectps_128((__mmask8)(U), \ |
| 6550 | (__v4sf)_mm_permute_ps((X), (C)), \ |
| 6551 | (__v4sf)(__m128)(W)); }) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 6552 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6553 | #define _mm_maskz_permute_ps(U, X, C) __extension__ ({ \ |
Craig Topper | b3a4477 | 2016-07-02 05:36:43 +0000 | [diff] [blame] | 6554 | (__m128)__builtin_ia32_selectps_128((__mmask8)(U), \ |
| 6555 | (__v4sf)_mm_permute_ps((X), (C)), \ |
| 6556 | (__v4sf)_mm_setzero_ps()); }) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 6557 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6558 | #define _mm256_mask_permute_ps(W, U, X, C) __extension__ ({ \ |
Craig Topper | b3a4477 | 2016-07-02 05:36:43 +0000 | [diff] [blame] | 6559 | (__m256)__builtin_ia32_selectps_256((__mmask8)(U), \ |
| 6560 | (__v8sf)_mm256_permute_ps((X), (C)), \ |
| 6561 | (__v8sf)(__m256)(W)); }) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 6562 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6563 | #define _mm256_maskz_permute_ps(U, X, C) __extension__ ({ \ |
Craig Topper | b3a4477 | 2016-07-02 05:36:43 +0000 | [diff] [blame] | 6564 | (__m256)__builtin_ia32_selectps_256((__mmask8)(U), \ |
| 6565 | (__v8sf)_mm256_permute_ps((X), (C)), \ |
| 6566 | (__v8sf)_mm256_setzero_ps()); }) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 6567 | |
| 6568 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 6569 | _mm_mask_permutevar_pd (__m128d __W, __mmask8 __U, __m128d __A, |
| 6570 | __m128i __C) |
| 6571 | { |
| 6572 | return (__m128d) __builtin_ia32_vpermilvarpd_mask ((__v2df) __A, |
| 6573 | (__v2di) __C, |
| 6574 | (__v2df) __W, |
| 6575 | (__mmask8) __U); |
| 6576 | } |
| 6577 | |
| 6578 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 6579 | _mm_maskz_permutevar_pd (__mmask8 __U, __m128d __A, __m128i __C) |
| 6580 | { |
| 6581 | return (__m128d) __builtin_ia32_vpermilvarpd_mask ((__v2df) __A, |
| 6582 | (__v2di) __C, |
| 6583 | (__v2df) |
| 6584 | _mm_setzero_pd (), |
| 6585 | (__mmask8) __U); |
| 6586 | } |
| 6587 | |
| 6588 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 6589 | _mm256_mask_permutevar_pd (__m256d __W, __mmask8 __U, __m256d __A, |
| 6590 | __m256i __C) |
| 6591 | { |
| 6592 | return (__m256d) __builtin_ia32_vpermilvarpd256_mask ((__v4df) __A, |
| 6593 | (__v4di) __C, |
| 6594 | (__v4df) __W, |
| 6595 | (__mmask8) |
| 6596 | __U); |
| 6597 | } |
| 6598 | |
| 6599 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 6600 | _mm256_maskz_permutevar_pd (__mmask8 __U, __m256d __A, __m256i __C) |
| 6601 | { |
| 6602 | return (__m256d) __builtin_ia32_vpermilvarpd256_mask ((__v4df) __A, |
| 6603 | (__v4di) __C, |
| 6604 | (__v4df) |
| 6605 | _mm256_setzero_pd (), |
| 6606 | (__mmask8) |
| 6607 | __U); |
| 6608 | } |
| 6609 | |
| 6610 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 6611 | _mm_mask_permutevar_ps (__m128 __W, __mmask8 __U, __m128 __A, |
| 6612 | __m128i __C) |
| 6613 | { |
| 6614 | return (__m128) __builtin_ia32_vpermilvarps_mask ((__v4sf) __A, |
| 6615 | (__v4si) __C, |
| 6616 | (__v4sf) __W, |
| 6617 | (__mmask8) __U); |
| 6618 | } |
| 6619 | |
| 6620 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 6621 | _mm_maskz_permutevar_ps (__mmask8 __U, __m128 __A, __m128i __C) |
| 6622 | { |
| 6623 | return (__m128) __builtin_ia32_vpermilvarps_mask ((__v4sf) __A, |
| 6624 | (__v4si) __C, |
| 6625 | (__v4sf) |
| 6626 | _mm_setzero_ps (), |
| 6627 | (__mmask8) __U); |
| 6628 | } |
| 6629 | |
| 6630 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 6631 | _mm256_mask_permutevar_ps (__m256 __W, __mmask8 __U, __m256 __A, |
| 6632 | __m256i __C) |
| 6633 | { |
| 6634 | return (__m256) __builtin_ia32_vpermilvarps256_mask ((__v8sf) __A, |
| 6635 | (__v8si) __C, |
| 6636 | (__v8sf) __W, |
| 6637 | (__mmask8) __U); |
| 6638 | } |
| 6639 | |
| 6640 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 6641 | _mm256_maskz_permutevar_ps (__mmask8 __U, __m256 __A, __m256i __C) |
| 6642 | { |
| 6643 | return (__m256) __builtin_ia32_vpermilvarps256_mask ((__v8sf) __A, |
| 6644 | (__v8si) __C, |
| 6645 | (__v8sf) |
| 6646 | _mm256_setzero_ps (), |
| 6647 | (__mmask8) __U); |
| 6648 | } |
| 6649 | |
Michael Zuckerman | 0752509 | 2016-04-11 10:22:07 +0000 | [diff] [blame] | 6650 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6651 | _mm_test_epi32_mask (__m128i __A, __m128i __B) |
| 6652 | { |
| 6653 | return (__mmask8) __builtin_ia32_ptestmd128 ((__v4si) __A, |
| 6654 | (__v4si) __B, |
| 6655 | (__mmask8) -1); |
| 6656 | } |
| 6657 | |
| 6658 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6659 | _mm_mask_test_epi32_mask (__mmask8 __U, __m128i __A, __m128i __B) |
| 6660 | { |
| 6661 | return (__mmask8) __builtin_ia32_ptestmd128 ((__v4si) __A, |
| 6662 | (__v4si) __B, __U); |
| 6663 | } |
| 6664 | |
| 6665 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6666 | _mm256_test_epi32_mask (__m256i __A, __m256i __B) |
| 6667 | { |
| 6668 | return (__mmask8) __builtin_ia32_ptestmd256 ((__v8si) __A, |
| 6669 | (__v8si) __B, |
| 6670 | (__mmask8) -1); |
| 6671 | } |
| 6672 | |
| 6673 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6674 | _mm256_mask_test_epi32_mask (__mmask8 __U, __m256i __A, __m256i __B) |
| 6675 | { |
| 6676 | return (__mmask8) __builtin_ia32_ptestmd256 ((__v8si) __A, |
| 6677 | (__v8si) __B, __U); |
| 6678 | } |
| 6679 | |
| 6680 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6681 | _mm_test_epi64_mask (__m128i __A, __m128i __B) |
| 6682 | { |
| 6683 | return (__mmask8) __builtin_ia32_ptestmq128 ((__v2di) __A, |
| 6684 | (__v2di) __B, |
| 6685 | (__mmask8) -1); |
| 6686 | } |
| 6687 | |
| 6688 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6689 | _mm_mask_test_epi64_mask (__mmask8 __U, __m128i __A, __m128i __B) |
| 6690 | { |
| 6691 | return (__mmask8) __builtin_ia32_ptestmq128 ((__v2di) __A, |
| 6692 | (__v2di) __B, __U); |
| 6693 | } |
| 6694 | |
| 6695 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6696 | _mm256_test_epi64_mask (__m256i __A, __m256i __B) |
| 6697 | { |
| 6698 | return (__mmask8) __builtin_ia32_ptestmq256 ((__v4di) __A, |
| 6699 | (__v4di) __B, |
| 6700 | (__mmask8) -1); |
| 6701 | } |
| 6702 | |
| 6703 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6704 | _mm256_mask_test_epi64_mask (__mmask8 __U, __m256i __A, __m256i __B) |
| 6705 | { |
| 6706 | return (__mmask8) __builtin_ia32_ptestmq256 ((__v4di) __A, |
| 6707 | (__v4di) __B, __U); |
| 6708 | } |
| 6709 | |
| 6710 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6711 | _mm_testn_epi32_mask (__m128i __A, __m128i __B) |
| 6712 | { |
| 6713 | return (__mmask8) __builtin_ia32_ptestnmd128 ((__v4si) __A, |
| 6714 | (__v4si) __B, |
| 6715 | (__mmask8) -1); |
| 6716 | } |
| 6717 | |
| 6718 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6719 | _mm_mask_testn_epi32_mask (__mmask8 __U, __m128i __A, __m128i __B) |
| 6720 | { |
| 6721 | return (__mmask8) __builtin_ia32_ptestnmd128 ((__v4si) __A, |
| 6722 | (__v4si) __B, __U); |
| 6723 | } |
| 6724 | |
| 6725 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6726 | _mm256_testn_epi32_mask (__m256i __A, __m256i __B) |
| 6727 | { |
| 6728 | return (__mmask8) __builtin_ia32_ptestnmd256 ((__v8si) __A, |
| 6729 | (__v8si) __B, |
| 6730 | (__mmask8) -1); |
| 6731 | } |
| 6732 | |
| 6733 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6734 | _mm256_mask_testn_epi32_mask (__mmask8 __U, __m256i __A, __m256i __B) |
| 6735 | { |
| 6736 | return (__mmask8) __builtin_ia32_ptestnmd256 ((__v8si) __A, |
| 6737 | (__v8si) __B, __U); |
| 6738 | } |
| 6739 | |
| 6740 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6741 | _mm_testn_epi64_mask (__m128i __A, __m128i __B) |
| 6742 | { |
| 6743 | return (__mmask8) __builtin_ia32_ptestnmq128 ((__v2di) __A, |
| 6744 | (__v2di) __B, |
| 6745 | (__mmask8) -1); |
| 6746 | } |
| 6747 | |
| 6748 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6749 | _mm_mask_testn_epi64_mask (__mmask8 __U, __m128i __A, __m128i __B) |
| 6750 | { |
| 6751 | return (__mmask8) __builtin_ia32_ptestnmq128 ((__v2di) __A, |
| 6752 | (__v2di) __B, __U); |
| 6753 | } |
| 6754 | |
| 6755 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6756 | _mm256_testn_epi64_mask (__m256i __A, __m256i __B) |
| 6757 | { |
| 6758 | return (__mmask8) __builtin_ia32_ptestnmq256 ((__v4di) __A, |
| 6759 | (__v4di) __B, |
| 6760 | (__mmask8) -1); |
| 6761 | } |
| 6762 | |
| 6763 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6764 | _mm256_mask_testn_epi64_mask (__mmask8 __U, __m256i __A, __m256i __B) |
| 6765 | { |
| 6766 | return (__mmask8) __builtin_ia32_ptestnmq256 ((__v4di) __A, |
| 6767 | (__v4di) __B, __U); |
| 6768 | } |
| 6769 | |
| 6770 | |
| 6771 | |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6772 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6773 | _mm_mask_unpackhi_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6774 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6775 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 6776 | (__v4si)_mm_unpackhi_epi32(__A, __B), |
| 6777 | (__v4si)__W); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6778 | } |
| 6779 | |
| 6780 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6781 | _mm_maskz_unpackhi_epi32(__mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6782 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6783 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 6784 | (__v4si)_mm_unpackhi_epi32(__A, __B), |
| 6785 | (__v4si)_mm_setzero_si128()); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6786 | } |
| 6787 | |
| 6788 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6789 | _mm256_mask_unpackhi_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6790 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6791 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 6792 | (__v8si)_mm256_unpackhi_epi32(__A, __B), |
| 6793 | (__v8si)__W); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6794 | } |
| 6795 | |
| 6796 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6797 | _mm256_maskz_unpackhi_epi32(__mmask8 __U, __m256i __A, __m256i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6798 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6799 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 6800 | (__v8si)_mm256_unpackhi_epi32(__A, __B), |
| 6801 | (__v8si)_mm256_setzero_si256()); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6802 | } |
| 6803 | |
| 6804 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6805 | _mm_mask_unpackhi_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6806 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6807 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 6808 | (__v2di)_mm_unpackhi_epi64(__A, __B), |
| 6809 | (__v2di)__W); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6810 | } |
| 6811 | |
| 6812 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6813 | _mm_maskz_unpackhi_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6814 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6815 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 6816 | (__v2di)_mm_unpackhi_epi64(__A, __B), |
| 6817 | (__v2di)_mm_setzero_di()); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6818 | } |
| 6819 | |
| 6820 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6821 | _mm256_mask_unpackhi_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6822 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6823 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 6824 | (__v4di)_mm256_unpackhi_epi64(__A, __B), |
| 6825 | (__v4di)__W); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6826 | } |
| 6827 | |
| 6828 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6829 | _mm256_maskz_unpackhi_epi64(__mmask8 __U, __m256i __A, __m256i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6830 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6831 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 6832 | (__v4di)_mm256_unpackhi_epi64(__A, __B), |
| 6833 | (__v4di)_mm256_setzero_si256()); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6834 | } |
| 6835 | |
| 6836 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6837 | _mm_mask_unpacklo_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6838 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6839 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 6840 | (__v4si)_mm_unpacklo_epi32(__A, __B), |
| 6841 | (__v4si)__W); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6842 | } |
| 6843 | |
| 6844 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6845 | _mm_maskz_unpacklo_epi32(__mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6846 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6847 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 6848 | (__v4si)_mm_unpacklo_epi32(__A, __B), |
| 6849 | (__v4si)_mm_setzero_si128()); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6850 | } |
| 6851 | |
| 6852 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6853 | _mm256_mask_unpacklo_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6854 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6855 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 6856 | (__v8si)_mm256_unpacklo_epi32(__A, __B), |
| 6857 | (__v8si)__W); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6858 | } |
| 6859 | |
| 6860 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6861 | _mm256_maskz_unpacklo_epi32(__mmask8 __U, __m256i __A, __m256i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6862 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6863 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 6864 | (__v8si)_mm256_unpacklo_epi32(__A, __B), |
| 6865 | (__v8si)_mm256_setzero_si256()); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6866 | } |
| 6867 | |
| 6868 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6869 | _mm_mask_unpacklo_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6870 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6871 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 6872 | (__v2di)_mm_unpacklo_epi64(__A, __B), |
| 6873 | (__v2di)__W); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6874 | } |
| 6875 | |
| 6876 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6877 | _mm_maskz_unpacklo_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6878 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6879 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 6880 | (__v2di)_mm_unpacklo_epi64(__A, __B), |
| 6881 | (__v2di)_mm_setzero_di()); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6882 | } |
| 6883 | |
| 6884 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6885 | _mm256_mask_unpacklo_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6886 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6887 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 6888 | (__v4di)_mm256_unpacklo_epi64(__A, __B), |
| 6889 | (__v4di)__W); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6890 | } |
| 6891 | |
| 6892 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6893 | _mm256_maskz_unpacklo_epi64(__mmask8 __U, __m256i __A, __m256i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6894 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6895 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 6896 | (__v4di)_mm256_unpacklo_epi64(__A, __B), |
| 6897 | (__v4di)_mm256_setzero_si256()); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6898 | } |
| 6899 | |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6900 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6901 | _mm_mask_sra_epi32 (__m128i __W, __mmask8 __U, __m128i __A, |
| 6902 | __m128i __B) |
| 6903 | { |
| 6904 | return (__m128i) __builtin_ia32_psrad128_mask ((__v4si) __A, |
| 6905 | (__v4si) __B, |
| 6906 | (__v4si) __W, |
| 6907 | (__mmask8) __U); |
| 6908 | } |
| 6909 | |
| 6910 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6911 | _mm_maskz_sra_epi32 (__mmask8 __U, __m128i __A, __m128i __B) |
| 6912 | { |
| 6913 | return (__m128i) __builtin_ia32_psrad128_mask ((__v4si) __A, |
| 6914 | (__v4si) __B, |
| 6915 | (__v4si) |
| 6916 | _mm_setzero_si128 (), |
| 6917 | (__mmask8) __U); |
| 6918 | } |
| 6919 | |
| 6920 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 6921 | _mm256_mask_sra_epi32 (__m256i __W, __mmask8 __U, __m256i __A, |
| 6922 | __m128i __B) |
| 6923 | { |
| 6924 | return (__m256i) __builtin_ia32_psrad256_mask ((__v8si) __A, |
| 6925 | (__v4si) __B, |
| 6926 | (__v8si) __W, |
| 6927 | (__mmask8) __U); |
| 6928 | } |
| 6929 | |
| 6930 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 6931 | _mm256_maskz_sra_epi32 (__mmask8 __U, __m256i __A, __m128i __B) |
| 6932 | { |
| 6933 | return (__m256i) __builtin_ia32_psrad256_mask ((__v8si) __A, |
| 6934 | (__v4si) __B, |
| 6935 | (__v8si) |
| 6936 | _mm256_setzero_si256 (), |
| 6937 | (__mmask8) __U); |
| 6938 | } |
| 6939 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6940 | #define _mm_mask_srai_epi32(W, U, A, imm) __extension__ ({ \ |
| 6941 | (__m128i)__builtin_ia32_psradi128_mask((__v4si)(__m128i)(A), (int)(imm), \ |
| 6942 | (__v4si)(__m128i)(W), \ |
| 6943 | (__mmask8)(U)); }) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6944 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6945 | #define _mm_maskz_srai_epi32(U, A, imm) __extension__ ({ \ |
| 6946 | (__m128i)__builtin_ia32_psradi128_mask((__v4si)(__m128i)(A), (int)(imm), \ |
| 6947 | (__v4si)_mm_setzero_si128(), \ |
| 6948 | (__mmask8)(U)); }) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6949 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6950 | #define _mm256_mask_srai_epi32(W, U, A, imm) __extension__ ({ \ |
| 6951 | (__m256i)__builtin_ia32_psradi256_mask((__v8si)(__m256i)(A), (int)(imm), \ |
| 6952 | (__v8si)(__m256i)(W), \ |
| 6953 | (__mmask8)(U)); }) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6954 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6955 | #define _mm256_maskz_srai_epi32(U, A, imm) __extension__ ({ \ |
| 6956 | (__m256i)__builtin_ia32_psradi256_mask((__v8si)(__m256i)(A), (int)(imm), \ |
| 6957 | (__v8si)_mm256_setzero_si256(), \ |
| 6958 | (__mmask8)(U)); }) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6959 | |
| 6960 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6961 | _mm_sra_epi64 (__m128i __A, __m128i __B) |
| 6962 | { |
| 6963 | return (__m128i) __builtin_ia32_psraq128_mask ((__v2di) __A, |
| 6964 | (__v2di) __B, |
| 6965 | (__v2di) |
| 6966 | _mm_setzero_di (), |
| 6967 | (__mmask8) -1); |
| 6968 | } |
| 6969 | |
| 6970 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6971 | _mm_mask_sra_epi64 (__m128i __W, __mmask8 __U, __m128i __A, |
| 6972 | __m128i __B) |
| 6973 | { |
| 6974 | return (__m128i) __builtin_ia32_psraq128_mask ((__v2di) __A, |
| 6975 | (__v2di) __B, |
| 6976 | (__v2di) __W, |
| 6977 | (__mmask8) __U); |
| 6978 | } |
| 6979 | |
| 6980 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6981 | _mm_maskz_sra_epi64 (__mmask8 __U, __m128i __A, __m128i __B) |
| 6982 | { |
| 6983 | return (__m128i) __builtin_ia32_psraq128_mask ((__v2di) __A, |
| 6984 | (__v2di) __B, |
| 6985 | (__v2di) |
| 6986 | _mm_setzero_di (), |
| 6987 | (__mmask8) __U); |
| 6988 | } |
| 6989 | |
| 6990 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 6991 | _mm256_sra_epi64 (__m256i __A, __m128i __B) |
| 6992 | { |
| 6993 | return (__m256i) __builtin_ia32_psraq256_mask ((__v4di) __A, |
| 6994 | (__v2di) __B, |
| 6995 | (__v4di) |
| 6996 | _mm256_setzero_si256 (), |
| 6997 | (__mmask8) -1); |
| 6998 | } |
| 6999 | |
| 7000 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 7001 | _mm256_mask_sra_epi64 (__m256i __W, __mmask8 __U, __m256i __A, |
| 7002 | __m128i __B) |
| 7003 | { |
| 7004 | return (__m256i) __builtin_ia32_psraq256_mask ((__v4di) __A, |
| 7005 | (__v2di) __B, |
| 7006 | (__v4di) __W, |
| 7007 | (__mmask8) __U); |
| 7008 | } |
| 7009 | |
| 7010 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 7011 | _mm256_maskz_sra_epi64 (__mmask8 __U, __m256i __A, __m128i __B) |
| 7012 | { |
| 7013 | return (__m256i) __builtin_ia32_psraq256_mask ((__v4di) __A, |
| 7014 | (__v2di) __B, |
| 7015 | (__v4di) |
| 7016 | _mm256_setzero_si256 (), |
| 7017 | (__mmask8) __U); |
| 7018 | } |
| 7019 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7020 | #define _mm_srai_epi64(A, imm) __extension__ ({ \ |
| 7021 | (__m128i)__builtin_ia32_psraqi128_mask((__v2di)(__m128i)(A), (int)(imm), \ |
| 7022 | (__v2di)_mm_setzero_di(), \ |
| 7023 | (__mmask8)-1); }) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 7024 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7025 | #define _mm_mask_srai_epi64(W, U, A, imm) __extension__ ({ \ |
| 7026 | (__m128i)__builtin_ia32_psraqi128_mask((__v2di)(__m128i)(A), (int)(imm), \ |
| 7027 | (__v2di)(__m128i)(W), \ |
| 7028 | (__mmask8)(U)); }) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 7029 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7030 | #define _mm_maskz_srai_epi64(U, A, imm) __extension__ ({ \ |
| 7031 | (__m128i)__builtin_ia32_psraqi128_mask((__v2di)(__m128i)(A), (int)(imm), \ |
| 7032 | (__v2di)_mm_setzero_si128(), \ |
| 7033 | (__mmask8)(U)); }) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 7034 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7035 | #define _mm256_srai_epi64(A, imm) __extension__ ({ \ |
| 7036 | (__m256i)__builtin_ia32_psraqi256_mask((__v4di)(__m256i)(A), (int)(imm), \ |
| 7037 | (__v4di)_mm256_setzero_si256(), \ |
| 7038 | (__mmask8)-1); }) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 7039 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7040 | #define _mm256_mask_srai_epi64(W, U, A, imm) __extension__ ({ \ |
| 7041 | (__m256i)__builtin_ia32_psraqi256_mask((__v4di)(__m256i)(A), (int)(imm), \ |
| 7042 | (__v4di)(__m256i)(W), \ |
| 7043 | (__mmask8)(U)); }) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 7044 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7045 | #define _mm256_maskz_srai_epi64(U, A, imm) __extension__ ({ \ |
| 7046 | (__m256i)__builtin_ia32_psraqi256_mask((__v4di)(__m256i)(A), (int)(imm), \ |
| 7047 | (__v4di)_mm256_setzero_si256(), \ |
| 7048 | (__mmask8)(U)); }) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 7049 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7050 | #define _mm_ternarylogic_epi32(A, B, C, imm) __extension__ ({ \ |
| 7051 | (__m128i)__builtin_ia32_pternlogd128_mask((__v4si)(__m128i)(A), \ |
| 7052 | (__v4si)(__m128i)(B), \ |
| 7053 | (__v4si)(__m128i)(C), (int)(imm), \ |
| 7054 | (__mmask8)-1); }) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 7055 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7056 | #define _mm_mask_ternarylogic_epi32(A, U, B, C, imm) __extension__ ({ \ |
| 7057 | (__m128i)__builtin_ia32_pternlogd128_mask((__v4si)(__m128i)(A), \ |
| 7058 | (__v4si)(__m128i)(B), \ |
| 7059 | (__v4si)(__m128i)(C), (int)(imm), \ |
| 7060 | (__mmask8)(U)); }) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 7061 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7062 | #define _mm_maskz_ternarylogic_epi32(U, A, B, C, imm) __extension__ ({ \ |
| 7063 | (__m128i)__builtin_ia32_pternlogd128_maskz((__v4si)(__m128i)(A), \ |
| 7064 | (__v4si)(__m128i)(B), \ |
| 7065 | (__v4si)(__m128i)(C), (int)(imm), \ |
| 7066 | (__mmask8)(U)); }) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 7067 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7068 | #define _mm256_ternarylogic_epi32(A, B, C, imm) __extension__ ({ \ |
| 7069 | (__m256i)__builtin_ia32_pternlogd256_mask((__v8si)(__m256i)(A), \ |
| 7070 | (__v8si)(__m256i)(B), \ |
| 7071 | (__v8si)(__m256i)(C), (int)(imm), \ |
| 7072 | (__mmask8)-1); }) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 7073 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7074 | #define _mm256_mask_ternarylogic_epi32(A, U, B, C, imm) __extension__ ({ \ |
| 7075 | (__m256i)__builtin_ia32_pternlogd256_mask((__v8si)(__m256i)(A), \ |
| 7076 | (__v8si)(__m256i)(B), \ |
| 7077 | (__v8si)(__m256i)(C), (int)(imm), \ |
| 7078 | (__mmask8)(U)); }) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 7079 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7080 | #define _mm256_maskz_ternarylogic_epi32(U, A, B, C, imm) __extension__ ({ \ |
| 7081 | (__m256i)__builtin_ia32_pternlogd256_maskz((__v8si)(__m256i)(A), \ |
| 7082 | (__v8si)(__m256i)(B), \ |
| 7083 | (__v8si)(__m256i)(C), (int)(imm), \ |
| 7084 | (__mmask8)(U)); }) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 7085 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7086 | #define _mm_ternarylogic_epi64(A, B, C, imm) __extension__ ({ \ |
| 7087 | (__m128i)__builtin_ia32_pternlogq128_mask((__v2di)(__m128i)(A), \ |
| 7088 | (__v2di)(__m128i)(B), \ |
| 7089 | (__v2di)(__m128i)(C), (int)(imm), \ |
| 7090 | (__mmask8)-1); }) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 7091 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7092 | #define _mm_mask_ternarylogic_epi64(A, U, B, C, imm) __extension__ ({ \ |
| 7093 | (__m128i)__builtin_ia32_pternlogq128_mask((__v2di)(__m128i)(A), \ |
| 7094 | (__v2di)(__m128i)(B), \ |
| 7095 | (__v2di)(__m128i)(C), (int)(imm), \ |
| 7096 | (__mmask8)(U)); }) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 7097 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7098 | #define _mm_maskz_ternarylogic_epi64(U, A, B, C, imm) __extension__ ({ \ |
| 7099 | (__m128i)__builtin_ia32_pternlogq128_maskz((__v2di)(__m128i)(A), \ |
| 7100 | (__v2di)(__m128i)(B), \ |
| 7101 | (__v2di)(__m128i)(C), (int)(imm), \ |
| 7102 | (__mmask8)(U)); }) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 7103 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7104 | #define _mm256_ternarylogic_epi64(A, B, C, imm) __extension__ ({ \ |
| 7105 | (__m256i)__builtin_ia32_pternlogq256_mask((__v4di)(__m256i)(A), \ |
| 7106 | (__v4di)(__m256i)(B), \ |
| 7107 | (__v4di)(__m256i)(C), (int)(imm), \ |
| 7108 | (__mmask8)-1); }) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 7109 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7110 | #define _mm256_mask_ternarylogic_epi64(A, U, B, C, imm) __extension__ ({ \ |
| 7111 | (__m256i)__builtin_ia32_pternlogq256_mask((__v4di)(__m256i)(A), \ |
| 7112 | (__v4di)(__m256i)(B), \ |
| 7113 | (__v4di)(__m256i)(C), (int)(imm), \ |
| 7114 | (__mmask8)(U)); }) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 7115 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7116 | #define _mm256_maskz_ternarylogic_epi64(U, A, B, C, imm) __extension__ ({ \ |
| 7117 | (__m256i)__builtin_ia32_pternlogq256_maskz((__v4di)(__m256i)(A), \ |
| 7118 | (__v4di)(__m256i)(B), \ |
| 7119 | (__v4di)(__m256i)(C), (int)(imm), \ |
| 7120 | (__mmask8)(U)); }) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 7121 | |
| 7122 | |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7123 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7124 | #define _mm256_shuffle_f32x4(A, B, imm) __extension__ ({ \ |
| 7125 | (__m256)__builtin_ia32_shuf_f32x4_256_mask((__v8sf)(__m256)(A), \ |
| 7126 | (__v8sf)(__m256)(B), (int)(imm), \ |
| 7127 | (__v8sf)_mm256_setzero_ps(), \ |
| 7128 | (__mmask8)-1); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7129 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7130 | #define _mm256_mask_shuffle_f32x4(W, U, A, B, imm) __extension__ ({ \ |
| 7131 | (__m256)__builtin_ia32_shuf_f32x4_256_mask((__v8sf)(__m256)(A), \ |
| 7132 | (__v8sf)(__m256)(B), (int)(imm), \ |
| 7133 | (__v8sf)(__m256)(W), \ |
| 7134 | (__mmask8)(U)); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7135 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7136 | #define _mm256_maskz_shuffle_f32x4(U, A, B, imm) __extension__ ({ \ |
| 7137 | (__m256)__builtin_ia32_shuf_f32x4_256_mask((__v8sf)(__m256)(A), \ |
| 7138 | (__v8sf)(__m256)(B), (int)(imm), \ |
| 7139 | (__v8sf)_mm256_setzero_ps(), \ |
| 7140 | (__mmask8)(U)); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7141 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7142 | #define _mm256_shuffle_f64x2(A, B, imm) __extension__ ({ \ |
| 7143 | (__m256d)__builtin_ia32_shuf_f64x2_256_mask((__v4df)(__m256d)(A), \ |
| 7144 | (__v4df)(__m256d)(B), \ |
| 7145 | (int)(imm), \ |
| 7146 | (__v4df)_mm256_setzero_pd(), \ |
| 7147 | (__mmask8)-1); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7148 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7149 | #define _mm256_mask_shuffle_f64x2(W, U, A, B, imm) __extension__ ({ \ |
| 7150 | (__m256d)__builtin_ia32_shuf_f64x2_256_mask((__v4df)(__m256d)(A), \ |
| 7151 | (__v4df)(__m256d)(B), \ |
| 7152 | (int)(imm), \ |
| 7153 | (__v4df)(__m256d)(W), \ |
| 7154 | (__mmask8)(U)); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7155 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7156 | #define _mm256_maskz_shuffle_f64x2(U, A, B, imm) __extension__ ({ \ |
| 7157 | (__m256d)__builtin_ia32_shuf_f64x2_256_mask((__v4df)(__m256d)(A), \ |
| 7158 | (__v4df)(__m256d)(B), \ |
| 7159 | (int)(imm), \ |
| 7160 | (__v4df)_mm256_setzero_pd(), \ |
| 7161 | (__mmask8)(U)); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7162 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7163 | #define _mm256_shuffle_i32x4(A, B, imm) __extension__ ({ \ |
| 7164 | (__m256i)__builtin_ia32_shuf_i32x4_256_mask((__v8si)(__m256i)(A), \ |
| 7165 | (__v8si)(__m256i)(B), \ |
| 7166 | (int)(imm), \ |
| 7167 | (__v8si)_mm256_setzero_si256(), \ |
| 7168 | (__mmask8)-1); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7169 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7170 | #define _mm256_mask_shuffle_i32x4(W, U, A, B, imm) __extension__ ({ \ |
| 7171 | (__m256i)__builtin_ia32_shuf_i32x4_256_mask((__v8si)(__m256i)(A), \ |
| 7172 | (__v8si)(__m256i)(B), \ |
| 7173 | (int)(imm), \ |
| 7174 | (__v8si)(__m256i)(W), \ |
| 7175 | (__mmask8)(U)); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7176 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7177 | #define _mm256_maskz_shuffle_i32x4(U, A, B, imm) __extension__ ({ \ |
| 7178 | (__m256i)__builtin_ia32_shuf_i32x4_256_mask((__v8si)(__m256i)(A), \ |
| 7179 | (__v8si)(__m256i)(B), \ |
| 7180 | (int)(imm), \ |
| 7181 | (__v8si)_mm256_setzero_si256(), \ |
| 7182 | (__mmask8)(U)); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7183 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7184 | #define _mm256_shuffle_i64x2(A, B, imm) __extension__ ({ \ |
| 7185 | (__m256i)__builtin_ia32_shuf_i64x2_256_mask((__v4di)(__m256i)(A), \ |
| 7186 | (__v4di)(__m256i)(B), \ |
| 7187 | (int)(imm), \ |
| 7188 | (__v4di)_mm256_setzero_si256(), \ |
| 7189 | (__mmask8)-1); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7190 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7191 | #define _mm256_mask_shuffle_i64x2(W, U, A, B, imm) __extension__ ({ \ |
| 7192 | (__m256i)__builtin_ia32_shuf_i64x2_256_mask((__v4di)(__m256i)(A), \ |
| 7193 | (__v4di)(__m256i)(B), \ |
| 7194 | (int)(imm), \ |
| 7195 | (__v4di)(__m256i)(W), \ |
| 7196 | (__mmask8)(U)); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7197 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7198 | #define _mm256_maskz_shuffle_i64x2(U, A, B, imm) __extension__ ({ \ |
| 7199 | (__m256i)__builtin_ia32_shuf_i64x2_256_mask((__v4di)(__m256i)(A), \ |
| 7200 | (__v4di)(__m256i)(B), \ |
| 7201 | (int)(imm), \ |
| 7202 | (__v4di)_mm256_setzero_si256(), \ |
| 7203 | (__mmask8)(U)); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7204 | |
Simon Pilgrim | 427154d | 2016-07-04 21:30:47 +0000 | [diff] [blame] | 7205 | #define _mm_mask_shuffle_pd(W, U, A, B, M) __extension__ ({ \ |
| 7206 | (__m128d)__builtin_ia32_selectpd_128((__mmask8)(U), \ |
| 7207 | (__v2df)_mm_shuffle_pd((A), (B), (M)), \ |
| 7208 | (__v2df)(__m128d)(W)); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7209 | |
Simon Pilgrim | 427154d | 2016-07-04 21:30:47 +0000 | [diff] [blame] | 7210 | #define _mm_maskz_shuffle_pd(U, A, B, M) __extension__ ({ \ |
| 7211 | (__m128d)__builtin_ia32_selectpd_128((__mmask8)(U), \ |
| 7212 | (__v2df)_mm_shuffle_pd((A), (B), (M)), \ |
| 7213 | (__v2df)_mm_setzero_pd()); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7214 | |
Simon Pilgrim | 427154d | 2016-07-04 21:30:47 +0000 | [diff] [blame] | 7215 | #define _mm256_mask_shuffle_pd(W, U, A, B, M) __extension__ ({ \ |
| 7216 | (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \ |
| 7217 | (__v4df)_mm256_shuffle_pd((A), (B), (M)), \ |
| 7218 | (__v4df)(__m256d)(W)); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7219 | |
Simon Pilgrim | 427154d | 2016-07-04 21:30:47 +0000 | [diff] [blame] | 7220 | #define _mm256_maskz_shuffle_pd(U, A, B, M) __extension__ ({ \ |
| 7221 | (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \ |
| 7222 | (__v4df)_mm256_shuffle_pd((A), (B), (M)), \ |
| 7223 | (__v4df)_mm256_setzero_pd()); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7224 | |
Simon Pilgrim | 427154d | 2016-07-04 21:30:47 +0000 | [diff] [blame] | 7225 | #define _mm_mask_shuffle_ps(W, U, A, B, M) __extension__ ({ \ |
| 7226 | (__m128)__builtin_ia32_selectps_128((__mmask8)(U), \ |
| 7227 | (__v4sf)_mm_shuffle_ps((A), (B), (M)), \ |
| 7228 | (__v4sf)(__m128)(W)); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7229 | |
Simon Pilgrim | 427154d | 2016-07-04 21:30:47 +0000 | [diff] [blame] | 7230 | #define _mm_maskz_shuffle_ps(U, A, B, M) __extension__ ({ \ |
| 7231 | (__m128)__builtin_ia32_selectps_128((__mmask8)(U), \ |
| 7232 | (__v4sf)_mm_shuffle_ps((A), (B), (M)), \ |
| 7233 | (__v4sf)_mm_setzero_ps()); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7234 | |
Simon Pilgrim | 427154d | 2016-07-04 21:30:47 +0000 | [diff] [blame] | 7235 | #define _mm256_mask_shuffle_ps(W, U, A, B, M) __extension__ ({ \ |
| 7236 | (__m256)__builtin_ia32_selectps_256((__mmask8)(U), \ |
| 7237 | (__v8sf)_mm256_shuffle_ps((A), (B), (M)), \ |
| 7238 | (__v8sf)(__m256)(W)); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7239 | |
Simon Pilgrim | 427154d | 2016-07-04 21:30:47 +0000 | [diff] [blame] | 7240 | #define _mm256_maskz_shuffle_ps(U, A, B, M) __extension__ ({ \ |
| 7241 | (__m256)__builtin_ia32_selectps_256((__mmask8)(U), \ |
| 7242 | (__v8sf)_mm256_shuffle_ps((A), (B), (M)), \ |
| 7243 | (__v8sf)_mm256_setzero_ps()); }) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 7244 | |
| 7245 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 7246 | _mm_rsqrt14_pd (__m128d __A) |
| 7247 | { |
| 7248 | return (__m128d) __builtin_ia32_rsqrt14pd128_mask ((__v2df) __A, |
| 7249 | (__v2df) |
| 7250 | _mm_setzero_pd (), |
| 7251 | (__mmask8) -1); |
| 7252 | } |
| 7253 | |
| 7254 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 7255 | _mm_mask_rsqrt14_pd (__m128d __W, __mmask8 __U, __m128d __A) |
| 7256 | { |
| 7257 | return (__m128d) __builtin_ia32_rsqrt14pd128_mask ((__v2df) __A, |
| 7258 | (__v2df) __W, |
| 7259 | (__mmask8) __U); |
| 7260 | } |
| 7261 | |
| 7262 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 7263 | _mm_maskz_rsqrt14_pd (__mmask8 __U, __m128d __A) |
| 7264 | { |
| 7265 | return (__m128d) __builtin_ia32_rsqrt14pd128_mask ((__v2df) __A, |
| 7266 | (__v2df) |
| 7267 | _mm_setzero_pd (), |
| 7268 | (__mmask8) __U); |
| 7269 | } |
| 7270 | |
| 7271 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 7272 | _mm256_rsqrt14_pd (__m256d __A) |
| 7273 | { |
| 7274 | return (__m256d) __builtin_ia32_rsqrt14pd256_mask ((__v4df) __A, |
| 7275 | (__v4df) |
| 7276 | _mm256_setzero_pd (), |
| 7277 | (__mmask8) -1); |
| 7278 | } |
| 7279 | |
| 7280 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 7281 | _mm256_mask_rsqrt14_pd (__m256d __W, __mmask8 __U, __m256d __A) |
| 7282 | { |
| 7283 | return (__m256d) __builtin_ia32_rsqrt14pd256_mask ((__v4df) __A, |
| 7284 | (__v4df) __W, |
| 7285 | (__mmask8) __U); |
| 7286 | } |
| 7287 | |
| 7288 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 7289 | _mm256_maskz_rsqrt14_pd (__mmask8 __U, __m256d __A) |
| 7290 | { |
| 7291 | return (__m256d) __builtin_ia32_rsqrt14pd256_mask ((__v4df) __A, |
| 7292 | (__v4df) |
| 7293 | _mm256_setzero_pd (), |
| 7294 | (__mmask8) __U); |
| 7295 | } |
| 7296 | |
| 7297 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 7298 | _mm_rsqrt14_ps (__m128 __A) |
| 7299 | { |
| 7300 | return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A, |
| 7301 | (__v4sf) |
| 7302 | _mm_setzero_ps (), |
| 7303 | (__mmask8) -1); |
| 7304 | } |
| 7305 | |
| 7306 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 7307 | _mm_mask_rsqrt14_ps (__m128 __W, __mmask8 __U, __m128 __A) |
| 7308 | { |
| 7309 | return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A, |
| 7310 | (__v4sf) __W, |
| 7311 | (__mmask8) __U); |
| 7312 | } |
| 7313 | |
| 7314 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 7315 | _mm_maskz_rsqrt14_ps (__mmask8 __U, __m128 __A) |
| 7316 | { |
| 7317 | return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A, |
| 7318 | (__v4sf) |
| 7319 | _mm_setzero_ps (), |
| 7320 | (__mmask8) __U); |
| 7321 | } |
| 7322 | |
| 7323 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 7324 | _mm256_rsqrt14_ps (__m256 __A) |
| 7325 | { |
| 7326 | return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A, |
| 7327 | (__v8sf) |
| 7328 | _mm256_setzero_ps (), |
| 7329 | (__mmask8) -1); |
| 7330 | } |
| 7331 | |
| 7332 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 7333 | _mm256_mask_rsqrt14_ps (__m256 __W, __mmask8 __U, __m256 __A) |
| 7334 | { |
| 7335 | return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A, |
| 7336 | (__v8sf) __W, |
| 7337 | (__mmask8) __U); |
| 7338 | } |
| 7339 | |
| 7340 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 7341 | _mm256_maskz_rsqrt14_ps (__mmask8 __U, __m256 __A) |
| 7342 | { |
| 7343 | return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A, |
| 7344 | (__v8sf) |
| 7345 | _mm256_setzero_ps (), |
| 7346 | (__mmask8) __U); |
| 7347 | } |
| 7348 | |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7349 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 7350 | _mm256_broadcast_f32x4 (__m128 __A) |
| 7351 | { |
| 7352 | return (__m256) __builtin_ia32_broadcastf32x4_256_mask ((__v4sf) __A, |
| 7353 | (__v8sf)_mm256_undefined_pd (), |
Craig Topper | 406d5cd | 2016-06-04 05:43:37 +0000 | [diff] [blame] | 7354 | (__mmask8) -1); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7355 | } |
| 7356 | |
| 7357 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 7358 | _mm256_mask_broadcast_f32x4 (__m256 __O, __mmask8 __M, __m128 __A) |
| 7359 | { |
| 7360 | return (__m256) __builtin_ia32_broadcastf32x4_256_mask ((__v4sf) __A, |
| 7361 | (__v8sf) __O, |
| 7362 | __M); |
| 7363 | } |
| 7364 | |
| 7365 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 7366 | _mm256_maskz_broadcast_f32x4 (__mmask8 __M, __m128 __A) |
| 7367 | { |
| 7368 | return (__m256) __builtin_ia32_broadcastf32x4_256_mask ((__v4sf) __A, |
| 7369 | (__v8sf) _mm256_setzero_ps (), |
| 7370 | __M); |
| 7371 | } |
| 7372 | |
| 7373 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 7374 | _mm256_broadcast_i32x4 (__m128i __A) |
| 7375 | { |
| 7376 | return (__m256i) __builtin_ia32_broadcasti32x4_256_mask ((__v4si) __A, |
| 7377 | (__v8si)_mm256_undefined_si256 (), |
Craig Topper | 406d5cd | 2016-06-04 05:43:37 +0000 | [diff] [blame] | 7378 | (__mmask8) -1); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7379 | } |
| 7380 | |
| 7381 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 7382 | _mm256_mask_broadcast_i32x4 (__m256i __O, __mmask8 __M, __m128i __A) |
| 7383 | { |
| 7384 | return (__m256i) __builtin_ia32_broadcasti32x4_256_mask ((__v4si) __A, |
| 7385 | (__v8si) |
| 7386 | __O, __M); |
| 7387 | } |
| 7388 | |
| 7389 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 7390 | _mm256_maskz_broadcast_i32x4 (__mmask8 __M, __m128i __A) |
| 7391 | { |
| 7392 | return (__m256i) __builtin_ia32_broadcasti32x4_256_mask ((__v4si) |
| 7393 | __A, |
| 7394 | (__v8si) _mm256_setzero_si256 (), |
| 7395 | __M); |
| 7396 | } |
| 7397 | |
| 7398 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 7399 | _mm256_mask_broadcastsd_pd (__m256d __O, __mmask8 __M, __m128d __A) |
| 7400 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7401 | return (__m256d)__builtin_ia32_selectpd_256(__M, |
| 7402 | (__v4df) _mm256_broadcastsd_pd(__A), |
| 7403 | (__v4df) __O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7404 | } |
| 7405 | |
| 7406 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 7407 | _mm256_maskz_broadcastsd_pd (__mmask8 __M, __m128d __A) |
| 7408 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7409 | return (__m256d)__builtin_ia32_selectpd_256(__M, |
| 7410 | (__v4df) _mm256_broadcastsd_pd(__A), |
| 7411 | (__v4df) _mm256_setzero_pd()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7412 | } |
| 7413 | |
| 7414 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 7415 | _mm_mask_broadcastss_ps (__m128 __O, __mmask8 __M, __m128 __A) |
| 7416 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7417 | return (__m128)__builtin_ia32_selectps_128(__M, |
| 7418 | (__v4sf) _mm_broadcastss_ps(__A), |
| 7419 | (__v4sf) __O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7420 | } |
| 7421 | |
| 7422 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 7423 | _mm_maskz_broadcastss_ps (__mmask8 __M, __m128 __A) |
| 7424 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7425 | return (__m128)__builtin_ia32_selectps_128(__M, |
| 7426 | (__v4sf) _mm_broadcastss_ps(__A), |
| 7427 | (__v4sf) _mm_setzero_ps()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7428 | } |
| 7429 | |
| 7430 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 7431 | _mm256_mask_broadcastss_ps (__m256 __O, __mmask8 __M, __m128 __A) |
| 7432 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7433 | return (__m256)__builtin_ia32_selectps_256(__M, |
| 7434 | (__v8sf) _mm256_broadcastss_ps(__A), |
| 7435 | (__v8sf) __O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7436 | } |
| 7437 | |
| 7438 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 7439 | _mm256_maskz_broadcastss_ps (__mmask8 __M, __m128 __A) |
| 7440 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7441 | return (__m256)__builtin_ia32_selectps_256(__M, |
| 7442 | (__v8sf) _mm256_broadcastss_ps(__A), |
| 7443 | (__v8sf) _mm256_setzero_ps()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7444 | } |
| 7445 | |
| 7446 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7447 | _mm_mask_broadcastd_epi32 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7448 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7449 | return (__m128i)__builtin_ia32_selectd_128(__M, |
| 7450 | (__v4si) _mm_broadcastd_epi32(__A), |
| 7451 | (__v4si) __O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7452 | } |
| 7453 | |
| 7454 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7455 | _mm_maskz_broadcastd_epi32 (__mmask8 __M, __m128i __A) |
| 7456 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7457 | return (__m128i)__builtin_ia32_selectd_128(__M, |
| 7458 | (__v4si) _mm_broadcastd_epi32(__A), |
| 7459 | (__v4si) _mm_setzero_si128()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7460 | } |
| 7461 | |
| 7462 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 7463 | _mm256_mask_broadcastd_epi32 (__m256i __O, __mmask8 __M, __m128i __A) |
| 7464 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7465 | return (__m256i)__builtin_ia32_selectd_256(__M, |
| 7466 | (__v8si) _mm256_broadcastd_epi32(__A), |
| 7467 | (__v8si) __O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7468 | } |
| 7469 | |
| 7470 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 7471 | _mm256_maskz_broadcastd_epi32 (__mmask8 __M, __m128i __A) |
| 7472 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7473 | return (__m256i)__builtin_ia32_selectd_256(__M, |
| 7474 | (__v8si) _mm256_broadcastd_epi32(__A), |
| 7475 | (__v8si) _mm256_setzero_si256()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7476 | } |
| 7477 | |
| 7478 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7479 | _mm_mask_broadcastq_epi64 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7480 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7481 | return (__m128i)__builtin_ia32_selectq_128(__M, |
| 7482 | (__v2di) _mm_broadcastq_epi64(__A), |
| 7483 | (__v2di) __O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7484 | } |
| 7485 | |
| 7486 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7487 | _mm_maskz_broadcastq_epi64 (__mmask8 __M, __m128i __A) |
| 7488 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7489 | return (__m128i)__builtin_ia32_selectq_128(__M, |
| 7490 | (__v2di) _mm_broadcastq_epi64(__A), |
| 7491 | (__v2di) _mm_setzero_si128()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7492 | } |
| 7493 | |
| 7494 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 7495 | _mm256_mask_broadcastq_epi64 (__m256i __O, __mmask8 __M, __m128i __A) |
| 7496 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7497 | return (__m256i)__builtin_ia32_selectq_256(__M, |
| 7498 | (__v4di) _mm256_broadcastq_epi64(__A), |
| 7499 | (__v4di) __O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7500 | } |
| 7501 | |
| 7502 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 7503 | _mm256_maskz_broadcastq_epi64 (__mmask8 __M, __m128i __A) |
| 7504 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 7505 | return (__m256i)__builtin_ia32_selectq_256(__M, |
| 7506 | (__v4di) _mm256_broadcastq_epi64(__A), |
| 7507 | (__v4di) _mm256_setzero_si256()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 7508 | } |
| 7509 | |
Michael Zuckerman | e168061 | 2016-04-13 15:02:04 +0000 | [diff] [blame] | 7510 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7511 | _mm_cvtsepi32_epi8 (__m128i __A) |
| 7512 | { |
| 7513 | return (__m128i) __builtin_ia32_pmovsdb128_mask ((__v4si) __A, |
| 7514 | (__v16qi)_mm_undefined_si128(), |
| 7515 | (__mmask8) -1); |
| 7516 | } |
| 7517 | |
| 7518 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7519 | _mm_mask_cvtsepi32_epi8 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7520 | { |
| 7521 | return (__m128i) __builtin_ia32_pmovsdb128_mask ((__v4si) __A, |
| 7522 | (__v16qi) __O, __M); |
| 7523 | } |
| 7524 | |
| 7525 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7526 | _mm_maskz_cvtsepi32_epi8 (__mmask8 __M, __m128i __A) |
| 7527 | { |
| 7528 | return (__m128i) __builtin_ia32_pmovsdb128_mask ((__v4si) __A, |
| 7529 | (__v16qi) _mm_setzero_si128 (), |
| 7530 | __M); |
| 7531 | } |
| 7532 | |
| 7533 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7534 | _mm_mask_cvtsepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) |
| 7535 | { |
| 7536 | __builtin_ia32_pmovsdb128mem_mask ((__v16qi *) __P, (__v4si) __A, __M); |
| 7537 | } |
| 7538 | |
| 7539 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7540 | _mm256_cvtsepi32_epi8 (__m256i __A) |
| 7541 | { |
| 7542 | return (__m128i) __builtin_ia32_pmovsdb256_mask ((__v8si) __A, |
| 7543 | (__v16qi)_mm_undefined_si128(), |
| 7544 | (__mmask8) -1); |
| 7545 | } |
| 7546 | |
| 7547 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7548 | _mm256_mask_cvtsepi32_epi8 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7549 | { |
| 7550 | return (__m128i) __builtin_ia32_pmovsdb256_mask ((__v8si) __A, |
| 7551 | (__v16qi) __O, __M); |
| 7552 | } |
| 7553 | |
| 7554 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7555 | _mm256_maskz_cvtsepi32_epi8 (__mmask8 __M, __m256i __A) |
| 7556 | { |
| 7557 | return (__m128i) __builtin_ia32_pmovsdb256_mask ((__v8si) __A, |
| 7558 | (__v16qi) _mm_setzero_si128 (), |
| 7559 | __M); |
| 7560 | } |
| 7561 | |
| 7562 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7563 | _mm256_mask_cvtsepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) |
| 7564 | { |
| 7565 | __builtin_ia32_pmovsdb256mem_mask ((__v16qi *) __P, (__v8si) __A, __M); |
| 7566 | } |
| 7567 | |
| 7568 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7569 | _mm_cvtsepi32_epi16 (__m128i __A) |
| 7570 | { |
| 7571 | return (__m128i) __builtin_ia32_pmovsdw128_mask ((__v4si) __A, |
| 7572 | (__v8hi)_mm_setzero_si128 (), |
| 7573 | (__mmask8) -1); |
| 7574 | } |
| 7575 | |
| 7576 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7577 | _mm_mask_cvtsepi32_epi16 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7578 | { |
| 7579 | return (__m128i) __builtin_ia32_pmovsdw128_mask ((__v4si) __A, |
| 7580 | (__v8hi)__O, |
| 7581 | __M); |
| 7582 | } |
| 7583 | |
| 7584 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7585 | _mm_maskz_cvtsepi32_epi16 (__mmask8 __M, __m128i __A) |
| 7586 | { |
| 7587 | return (__m128i) __builtin_ia32_pmovsdw128_mask ((__v4si) __A, |
| 7588 | (__v8hi) _mm_setzero_si128 (), |
| 7589 | __M); |
| 7590 | } |
| 7591 | |
| 7592 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7593 | _mm_mask_cvtsepi32_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) |
| 7594 | { |
| 7595 | __builtin_ia32_pmovsdw128mem_mask ((__v8hi *) __P, (__v4si) __A, __M); |
| 7596 | } |
| 7597 | |
| 7598 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7599 | _mm256_cvtsepi32_epi16 (__m256i __A) |
| 7600 | { |
| 7601 | return (__m128i) __builtin_ia32_pmovsdw256_mask ((__v8si) __A, |
| 7602 | (__v8hi)_mm_undefined_si128(), |
| 7603 | (__mmask8) -1); |
| 7604 | } |
| 7605 | |
| 7606 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7607 | _mm256_mask_cvtsepi32_epi16 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7608 | { |
| 7609 | return (__m128i) __builtin_ia32_pmovsdw256_mask ((__v8si) __A, |
| 7610 | (__v8hi) __O, __M); |
| 7611 | } |
| 7612 | |
| 7613 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7614 | _mm256_maskz_cvtsepi32_epi16 (__mmask8 __M, __m256i __A) |
| 7615 | { |
| 7616 | return (__m128i) __builtin_ia32_pmovsdw256_mask ((__v8si) __A, |
| 7617 | (__v8hi) _mm_setzero_si128 (), |
| 7618 | __M); |
| 7619 | } |
| 7620 | |
| 7621 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7622 | _mm256_mask_cvtsepi32_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) |
| 7623 | { |
| 7624 | __builtin_ia32_pmovsdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M); |
| 7625 | } |
| 7626 | |
| 7627 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7628 | _mm_cvtsepi64_epi8 (__m128i __A) |
| 7629 | { |
| 7630 | return (__m128i) __builtin_ia32_pmovsqb128_mask ((__v2di) __A, |
| 7631 | (__v16qi)_mm_undefined_si128(), |
| 7632 | (__mmask8) -1); |
| 7633 | } |
| 7634 | |
| 7635 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7636 | _mm_mask_cvtsepi64_epi8 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7637 | { |
| 7638 | return (__m128i) __builtin_ia32_pmovsqb128_mask ((__v2di) __A, |
| 7639 | (__v16qi) __O, __M); |
| 7640 | } |
| 7641 | |
| 7642 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7643 | _mm_maskz_cvtsepi64_epi8 (__mmask8 __M, __m128i __A) |
| 7644 | { |
| 7645 | return (__m128i) __builtin_ia32_pmovsqb128_mask ((__v2di) __A, |
| 7646 | (__v16qi) _mm_setzero_si128 (), |
| 7647 | __M); |
| 7648 | } |
| 7649 | |
| 7650 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7651 | _mm_mask_cvtsepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) |
| 7652 | { |
| 7653 | __builtin_ia32_pmovsqb128mem_mask ((__v16qi *) __P, (__v2di) __A, __M); |
| 7654 | } |
| 7655 | |
| 7656 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7657 | _mm256_cvtsepi64_epi8 (__m256i __A) |
| 7658 | { |
| 7659 | return (__m128i) __builtin_ia32_pmovsqb256_mask ((__v4di) __A, |
| 7660 | (__v16qi)_mm_undefined_si128(), |
| 7661 | (__mmask8) -1); |
| 7662 | } |
| 7663 | |
| 7664 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7665 | _mm256_mask_cvtsepi64_epi8 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7666 | { |
| 7667 | return (__m128i) __builtin_ia32_pmovsqb256_mask ((__v4di) __A, |
| 7668 | (__v16qi) __O, __M); |
| 7669 | } |
| 7670 | |
| 7671 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7672 | _mm256_maskz_cvtsepi64_epi8 (__mmask8 __M, __m256i __A) |
| 7673 | { |
| 7674 | return (__m128i) __builtin_ia32_pmovsqb256_mask ((__v4di) __A, |
| 7675 | (__v16qi) _mm_setzero_si128 (), |
| 7676 | __M); |
| 7677 | } |
| 7678 | |
| 7679 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7680 | _mm256_mask_cvtsepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) |
| 7681 | { |
| 7682 | __builtin_ia32_pmovsqb256mem_mask ((__v16qi *) __P, (__v4di) __A, __M); |
| 7683 | } |
| 7684 | |
| 7685 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7686 | _mm_cvtsepi64_epi32 (__m128i __A) |
| 7687 | { |
| 7688 | return (__m128i) __builtin_ia32_pmovsqd128_mask ((__v2di) __A, |
| 7689 | (__v4si)_mm_undefined_si128(), |
| 7690 | (__mmask8) -1); |
| 7691 | } |
| 7692 | |
| 7693 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7694 | _mm_mask_cvtsepi64_epi32 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7695 | { |
| 7696 | return (__m128i) __builtin_ia32_pmovsqd128_mask ((__v2di) __A, |
| 7697 | (__v4si) __O, __M); |
| 7698 | } |
| 7699 | |
| 7700 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7701 | _mm_maskz_cvtsepi64_epi32 (__mmask8 __M, __m128i __A) |
| 7702 | { |
| 7703 | return (__m128i) __builtin_ia32_pmovsqd128_mask ((__v2di) __A, |
| 7704 | (__v4si) _mm_setzero_si128 (), |
| 7705 | __M); |
| 7706 | } |
| 7707 | |
| 7708 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7709 | _mm_mask_cvtsepi64_storeu_epi32 (void * __P, __mmask8 __M, __m128i __A) |
| 7710 | { |
| 7711 | __builtin_ia32_pmovsqd128mem_mask ((__v4si *) __P, (__v2di) __A, __M); |
| 7712 | } |
| 7713 | |
| 7714 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7715 | _mm256_cvtsepi64_epi32 (__m256i __A) |
| 7716 | { |
| 7717 | return (__m128i) __builtin_ia32_pmovsqd256_mask ((__v4di) __A, |
| 7718 | (__v4si)_mm_undefined_si128(), |
| 7719 | (__mmask8) -1); |
| 7720 | } |
| 7721 | |
| 7722 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7723 | _mm256_mask_cvtsepi64_epi32 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7724 | { |
| 7725 | return (__m128i) __builtin_ia32_pmovsqd256_mask ((__v4di) __A, |
| 7726 | (__v4si)__O, |
| 7727 | __M); |
| 7728 | } |
| 7729 | |
| 7730 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7731 | _mm256_maskz_cvtsepi64_epi32 (__mmask8 __M, __m256i __A) |
| 7732 | { |
| 7733 | return (__m128i) __builtin_ia32_pmovsqd256_mask ((__v4di) __A, |
| 7734 | (__v4si) _mm_setzero_si128 (), |
| 7735 | __M); |
| 7736 | } |
| 7737 | |
| 7738 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7739 | _mm256_mask_cvtsepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A) |
| 7740 | { |
| 7741 | __builtin_ia32_pmovsqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M); |
| 7742 | } |
| 7743 | |
| 7744 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7745 | _mm_cvtsepi64_epi16 (__m128i __A) |
| 7746 | { |
| 7747 | return (__m128i) __builtin_ia32_pmovsqw128_mask ((__v2di) __A, |
| 7748 | (__v8hi)_mm_undefined_si128(), |
| 7749 | (__mmask8) -1); |
| 7750 | } |
| 7751 | |
| 7752 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7753 | _mm_mask_cvtsepi64_epi16 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7754 | { |
| 7755 | return (__m128i) __builtin_ia32_pmovsqw128_mask ((__v2di) __A, |
| 7756 | (__v8hi) __O, __M); |
| 7757 | } |
| 7758 | |
| 7759 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7760 | _mm_maskz_cvtsepi64_epi16 (__mmask8 __M, __m128i __A) |
| 7761 | { |
| 7762 | return (__m128i) __builtin_ia32_pmovsqw128_mask ((__v2di) __A, |
| 7763 | (__v8hi) _mm_setzero_si128 (), |
| 7764 | __M); |
| 7765 | } |
| 7766 | |
| 7767 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7768 | _mm_mask_cvtsepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) |
| 7769 | { |
| 7770 | __builtin_ia32_pmovsqw128mem_mask ((__v8hi *) __P, (__v2di) __A, __M); |
| 7771 | } |
| 7772 | |
| 7773 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7774 | _mm256_cvtsepi64_epi16 (__m256i __A) |
| 7775 | { |
| 7776 | return (__m128i) __builtin_ia32_pmovsqw256_mask ((__v4di) __A, |
| 7777 | (__v8hi)_mm_undefined_si128(), |
| 7778 | (__mmask8) -1); |
| 7779 | } |
| 7780 | |
| 7781 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7782 | _mm256_mask_cvtsepi64_epi16 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7783 | { |
| 7784 | return (__m128i) __builtin_ia32_pmovsqw256_mask ((__v4di) __A, |
| 7785 | (__v8hi) __O, __M); |
| 7786 | } |
| 7787 | |
| 7788 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7789 | _mm256_maskz_cvtsepi64_epi16 (__mmask8 __M, __m256i __A) |
| 7790 | { |
| 7791 | return (__m128i) __builtin_ia32_pmovsqw256_mask ((__v4di) __A, |
| 7792 | (__v8hi) _mm_setzero_si128 (), |
| 7793 | __M); |
| 7794 | } |
| 7795 | |
| 7796 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7797 | _mm256_mask_cvtsepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) |
| 7798 | { |
| 7799 | __builtin_ia32_pmovsqw256mem_mask ((__v8hi *) __P, (__v4di) __A, __M); |
| 7800 | } |
| 7801 | |
Michael Zuckerman | d871531 | 2016-04-14 06:48:09 +0000 | [diff] [blame] | 7802 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7803 | _mm_cvtusepi32_epi8 (__m128i __A) |
| 7804 | { |
| 7805 | return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A, |
| 7806 | (__v16qi)_mm_undefined_si128(), |
| 7807 | (__mmask8) -1); |
| 7808 | } |
| 7809 | |
| 7810 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7811 | _mm_mask_cvtusepi32_epi8 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7812 | { |
| 7813 | return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A, |
| 7814 | (__v16qi) __O, |
| 7815 | __M); |
| 7816 | } |
| 7817 | |
| 7818 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7819 | _mm_maskz_cvtusepi32_epi8 (__mmask8 __M, __m128i __A) |
| 7820 | { |
| 7821 | return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A, |
| 7822 | (__v16qi) _mm_setzero_si128 (), |
| 7823 | __M); |
| 7824 | } |
| 7825 | |
| 7826 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7827 | _mm_mask_cvtusepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) |
| 7828 | { |
| 7829 | __builtin_ia32_pmovusdb128mem_mask ((__v16qi *) __P, (__v4si) __A, __M); |
| 7830 | } |
| 7831 | |
| 7832 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7833 | _mm256_cvtusepi32_epi8 (__m256i __A) |
| 7834 | { |
| 7835 | return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A, |
| 7836 | (__v16qi)_mm_undefined_si128(), |
| 7837 | (__mmask8) -1); |
| 7838 | } |
| 7839 | |
| 7840 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7841 | _mm256_mask_cvtusepi32_epi8 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7842 | { |
| 7843 | return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A, |
| 7844 | (__v16qi) __O, |
| 7845 | __M); |
| 7846 | } |
| 7847 | |
| 7848 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7849 | _mm256_maskz_cvtusepi32_epi8 (__mmask8 __M, __m256i __A) |
| 7850 | { |
| 7851 | return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A, |
| 7852 | (__v16qi) _mm_setzero_si128 (), |
| 7853 | __M); |
| 7854 | } |
| 7855 | |
| 7856 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7857 | _mm256_mask_cvtusepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) |
| 7858 | { |
| 7859 | __builtin_ia32_pmovusdb256mem_mask ((__v16qi*) __P, (__v8si) __A, __M); |
| 7860 | } |
| 7861 | |
| 7862 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7863 | _mm_cvtusepi32_epi16 (__m128i __A) |
| 7864 | { |
| 7865 | return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A, |
| 7866 | (__v8hi)_mm_undefined_si128(), |
| 7867 | (__mmask8) -1); |
| 7868 | } |
| 7869 | |
| 7870 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7871 | _mm_mask_cvtusepi32_epi16 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7872 | { |
| 7873 | return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A, |
| 7874 | (__v8hi) __O, __M); |
| 7875 | } |
| 7876 | |
| 7877 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7878 | _mm_maskz_cvtusepi32_epi16 (__mmask8 __M, __m128i __A) |
| 7879 | { |
| 7880 | return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A, |
| 7881 | (__v8hi) _mm_setzero_si128 (), |
| 7882 | __M); |
| 7883 | } |
| 7884 | |
| 7885 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7886 | _mm_mask_cvtusepi32_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) |
| 7887 | { |
| 7888 | __builtin_ia32_pmovusdw128mem_mask ((__v8hi *) __P, (__v4si) __A, __M); |
| 7889 | } |
| 7890 | |
| 7891 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7892 | _mm256_cvtusepi32_epi16 (__m256i __A) |
| 7893 | { |
| 7894 | return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A, |
| 7895 | (__v8hi) _mm_undefined_si128(), |
| 7896 | (__mmask8) -1); |
| 7897 | } |
| 7898 | |
| 7899 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7900 | _mm256_mask_cvtusepi32_epi16 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7901 | { |
| 7902 | return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A, |
| 7903 | (__v8hi) __O, __M); |
| 7904 | } |
| 7905 | |
| 7906 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7907 | _mm256_maskz_cvtusepi32_epi16 (__mmask8 __M, __m256i __A) |
| 7908 | { |
| 7909 | return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A, |
| 7910 | (__v8hi) _mm_setzero_si128 (), |
| 7911 | __M); |
| 7912 | } |
| 7913 | |
| 7914 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7915 | _mm256_mask_cvtusepi32_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) |
| 7916 | { |
| 7917 | __builtin_ia32_pmovusdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M); |
| 7918 | } |
| 7919 | |
| 7920 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7921 | _mm_cvtusepi64_epi8 (__m128i __A) |
| 7922 | { |
| 7923 | return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A, |
| 7924 | (__v16qi)_mm_undefined_si128(), |
| 7925 | (__mmask8) -1); |
| 7926 | } |
| 7927 | |
| 7928 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7929 | _mm_mask_cvtusepi64_epi8 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7930 | { |
| 7931 | return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A, |
| 7932 | (__v16qi) __O, |
| 7933 | __M); |
| 7934 | } |
| 7935 | |
| 7936 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7937 | _mm_maskz_cvtusepi64_epi8 (__mmask8 __M, __m128i __A) |
| 7938 | { |
| 7939 | return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A, |
| 7940 | (__v16qi) _mm_setzero_si128 (), |
| 7941 | __M); |
| 7942 | } |
| 7943 | |
| 7944 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7945 | _mm_mask_cvtusepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) |
| 7946 | { |
| 7947 | __builtin_ia32_pmovusqb128mem_mask ((__v16qi *) __P, (__v2di) __A, __M); |
| 7948 | } |
| 7949 | |
| 7950 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7951 | _mm256_cvtusepi64_epi8 (__m256i __A) |
| 7952 | { |
| 7953 | return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A, |
| 7954 | (__v16qi)_mm_undefined_si128(), |
| 7955 | (__mmask8) -1); |
| 7956 | } |
| 7957 | |
| 7958 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7959 | _mm256_mask_cvtusepi64_epi8 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7960 | { |
| 7961 | return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A, |
| 7962 | (__v16qi) __O, |
| 7963 | __M); |
| 7964 | } |
| 7965 | |
| 7966 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7967 | _mm256_maskz_cvtusepi64_epi8 (__mmask8 __M, __m256i __A) |
| 7968 | { |
| 7969 | return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A, |
| 7970 | (__v16qi) _mm_setzero_si128 (), |
| 7971 | __M); |
| 7972 | } |
| 7973 | |
| 7974 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7975 | _mm256_mask_cvtusepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) |
| 7976 | { |
| 7977 | __builtin_ia32_pmovusqb256mem_mask ((__v16qi *) __P, (__v4di) __A, __M); |
| 7978 | } |
| 7979 | |
| 7980 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7981 | _mm_cvtusepi64_epi32 (__m128i __A) |
| 7982 | { |
| 7983 | return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A, |
| 7984 | (__v4si)_mm_undefined_si128(), |
| 7985 | (__mmask8) -1); |
| 7986 | } |
| 7987 | |
| 7988 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7989 | _mm_mask_cvtusepi64_epi32 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7990 | { |
| 7991 | return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A, |
| 7992 | (__v4si) __O, __M); |
| 7993 | } |
| 7994 | |
| 7995 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7996 | _mm_maskz_cvtusepi64_epi32 (__mmask8 __M, __m128i __A) |
| 7997 | { |
| 7998 | return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A, |
| 7999 | (__v4si) _mm_setzero_si128 (), |
| 8000 | __M); |
| 8001 | } |
| 8002 | |
| 8003 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8004 | _mm_mask_cvtusepi64_storeu_epi32 (void * __P, __mmask8 __M, __m128i __A) |
| 8005 | { |
| 8006 | __builtin_ia32_pmovusqd128mem_mask ((__v4si *) __P, (__v2di) __A, __M); |
| 8007 | } |
| 8008 | |
| 8009 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8010 | _mm256_cvtusepi64_epi32 (__m256i __A) |
| 8011 | { |
| 8012 | return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A, |
| 8013 | (__v4si)_mm_undefined_si128(), |
| 8014 | (__mmask8) -1); |
| 8015 | } |
| 8016 | |
| 8017 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8018 | _mm256_mask_cvtusepi64_epi32 (__m128i __O, __mmask8 __M, __m256i __A) |
| 8019 | { |
| 8020 | return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A, |
| 8021 | (__v4si) __O, __M); |
| 8022 | } |
| 8023 | |
| 8024 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8025 | _mm256_maskz_cvtusepi64_epi32 (__mmask8 __M, __m256i __A) |
| 8026 | { |
| 8027 | return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A, |
| 8028 | (__v4si) _mm_setzero_si128 (), |
| 8029 | __M); |
| 8030 | } |
| 8031 | |
| 8032 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8033 | _mm256_mask_cvtusepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A) |
| 8034 | { |
| 8035 | __builtin_ia32_pmovusqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M); |
| 8036 | } |
| 8037 | |
| 8038 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8039 | _mm_cvtusepi64_epi16 (__m128i __A) |
| 8040 | { |
| 8041 | return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A, |
| 8042 | (__v8hi)_mm_undefined_si128(), |
| 8043 | (__mmask8) -1); |
| 8044 | } |
| 8045 | |
| 8046 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8047 | _mm_mask_cvtusepi64_epi16 (__m128i __O, __mmask8 __M, __m128i __A) |
| 8048 | { |
| 8049 | return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A, |
| 8050 | (__v8hi) __O, __M); |
| 8051 | } |
| 8052 | |
| 8053 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8054 | _mm_maskz_cvtusepi64_epi16 (__mmask8 __M, __m128i __A) |
| 8055 | { |
| 8056 | return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A, |
| 8057 | (__v8hi) _mm_setzero_si128 (), |
| 8058 | __M); |
| 8059 | } |
| 8060 | |
| 8061 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8062 | _mm_mask_cvtusepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) |
| 8063 | { |
| 8064 | __builtin_ia32_pmovusqw128mem_mask ((__v8hi *) __P, (__v2di) __A, __M); |
| 8065 | } |
| 8066 | |
| 8067 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8068 | _mm256_cvtusepi64_epi16 (__m256i __A) |
| 8069 | { |
| 8070 | return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A, |
| 8071 | (__v8hi)_mm_undefined_si128(), |
| 8072 | (__mmask8) -1); |
| 8073 | } |
| 8074 | |
| 8075 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8076 | _mm256_mask_cvtusepi64_epi16 (__m128i __O, __mmask8 __M, __m256i __A) |
| 8077 | { |
| 8078 | return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A, |
| 8079 | (__v8hi) __O, __M); |
| 8080 | } |
| 8081 | |
| 8082 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8083 | _mm256_maskz_cvtusepi64_epi16 (__mmask8 __M, __m256i __A) |
| 8084 | { |
| 8085 | return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A, |
| 8086 | (__v8hi) _mm_setzero_si128 (), |
| 8087 | __M); |
| 8088 | } |
| 8089 | |
| 8090 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8091 | _mm256_mask_cvtusepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) |
| 8092 | { |
| 8093 | return __builtin_ia32_pmovusqw256mem_mask ((__v8hi *) __P, (__v4di) __A, __M); |
| 8094 | } |
| 8095 | |
Michael Zuckerman | 0a3508a | 2016-04-14 07:56:51 +0000 | [diff] [blame] | 8096 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8097 | _mm_cvtepi32_epi8 (__m128i __A) |
| 8098 | { |
| 8099 | return (__m128i) __builtin_ia32_pmovdb128_mask ((__v4si) __A, |
| 8100 | (__v16qi)_mm_undefined_si128(), |
| 8101 | (__mmask8) -1); |
| 8102 | } |
| 8103 | |
| 8104 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8105 | _mm_mask_cvtepi32_epi8 (__m128i __O, __mmask8 __M, __m128i __A) |
| 8106 | { |
| 8107 | return (__m128i) __builtin_ia32_pmovdb128_mask ((__v4si) __A, |
| 8108 | (__v16qi) __O, __M); |
| 8109 | } |
| 8110 | |
| 8111 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8112 | _mm_maskz_cvtepi32_epi8 (__mmask8 __M, __m128i __A) |
| 8113 | { |
| 8114 | return (__m128i) __builtin_ia32_pmovdb128_mask ((__v4si) __A, |
| 8115 | (__v16qi) |
| 8116 | _mm_setzero_si128 (), |
| 8117 | __M); |
| 8118 | } |
| 8119 | |
| 8120 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8121 | _mm_mask_cvtepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) |
| 8122 | { |
| 8123 | __builtin_ia32_pmovdb128mem_mask ((__v16qi *) __P, (__v4si) __A, __M); |
| 8124 | } |
| 8125 | |
| 8126 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8127 | _mm256_cvtepi32_epi8 (__m256i __A) |
| 8128 | { |
| 8129 | return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A, |
| 8130 | (__v16qi)_mm_undefined_si128(), |
| 8131 | (__mmask8) -1); |
| 8132 | } |
| 8133 | |
| 8134 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8135 | _mm256_mask_cvtepi32_epi8 (__m128i __O, __mmask8 __M, __m256i __A) |
| 8136 | { |
| 8137 | return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A, |
| 8138 | (__v16qi) __O, __M); |
| 8139 | } |
| 8140 | |
| 8141 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8142 | _mm256_maskz_cvtepi32_epi8 (__mmask8 __M, __m256i __A) |
| 8143 | { |
| 8144 | return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A, |
| 8145 | (__v16qi) _mm_setzero_si128 (), |
| 8146 | __M); |
| 8147 | } |
| 8148 | |
| 8149 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8150 | _mm256_mask_cvtepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) |
| 8151 | { |
| 8152 | __builtin_ia32_pmovdb256mem_mask ((__v16qi *) __P, (__v8si) __A, __M); |
| 8153 | } |
| 8154 | |
| 8155 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8156 | _mm_cvtepi32_epi16 (__m128i __A) |
| 8157 | { |
| 8158 | return (__m128i) __builtin_ia32_pmovdw128_mask ((__v4si) __A, |
| 8159 | (__v8hi) _mm_setzero_si128 (), |
| 8160 | (__mmask8) -1); |
| 8161 | } |
| 8162 | |
| 8163 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8164 | _mm_mask_cvtepi32_epi16 (__m128i __O, __mmask8 __M, __m128i __A) |
| 8165 | { |
| 8166 | return (__m128i) __builtin_ia32_pmovdw128_mask ((__v4si) __A, |
| 8167 | (__v8hi) __O, __M); |
| 8168 | } |
| 8169 | |
| 8170 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8171 | _mm_maskz_cvtepi32_epi16 (__mmask8 __M, __m128i __A) |
| 8172 | { |
| 8173 | return (__m128i) __builtin_ia32_pmovdw128_mask ((__v4si) __A, |
| 8174 | (__v8hi) _mm_setzero_si128 (), |
| 8175 | __M); |
| 8176 | } |
| 8177 | |
| 8178 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8179 | _mm_mask_cvtepi32_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) |
| 8180 | { |
| 8181 | __builtin_ia32_pmovdw128mem_mask ((__v8hi *) __P, (__v4si) __A, __M); |
| 8182 | } |
| 8183 | |
| 8184 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8185 | _mm256_cvtepi32_epi16 (__m256i __A) |
| 8186 | { |
| 8187 | return (__m128i) __builtin_ia32_pmovdw256_mask ((__v8si) __A, |
| 8188 | (__v8hi)_mm_setzero_si128 (), |
| 8189 | (__mmask8) -1); |
| 8190 | } |
| 8191 | |
| 8192 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8193 | _mm256_mask_cvtepi32_epi16 (__m128i __O, __mmask8 __M, __m256i __A) |
| 8194 | { |
| 8195 | return (__m128i) __builtin_ia32_pmovdw256_mask ((__v8si) __A, |
| 8196 | (__v8hi) __O, __M); |
| 8197 | } |
| 8198 | |
| 8199 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8200 | _mm256_maskz_cvtepi32_epi16 (__mmask8 __M, __m256i __A) |
| 8201 | { |
| 8202 | return (__m128i) __builtin_ia32_pmovdw256_mask ((__v8si) __A, |
| 8203 | (__v8hi) _mm_setzero_si128 (), |
| 8204 | __M); |
| 8205 | } |
| 8206 | |
| 8207 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8208 | _mm256_mask_cvtepi32_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) |
| 8209 | { |
| 8210 | __builtin_ia32_pmovdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M); |
| 8211 | } |
| 8212 | |
| 8213 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8214 | _mm_cvtepi64_epi8 (__m128i __A) |
| 8215 | { |
| 8216 | return (__m128i) __builtin_ia32_pmovqb128_mask ((__v2di) __A, |
| 8217 | (__v16qi) _mm_undefined_si128(), |
| 8218 | (__mmask8) -1); |
| 8219 | } |
| 8220 | |
| 8221 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8222 | _mm_mask_cvtepi64_epi8 (__m128i __O, __mmask8 __M, __m128i __A) |
| 8223 | { |
| 8224 | return (__m128i) __builtin_ia32_pmovqb128_mask ((__v2di) __A, |
| 8225 | (__v16qi) __O, __M); |
| 8226 | } |
| 8227 | |
| 8228 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8229 | _mm_maskz_cvtepi64_epi8 (__mmask8 __M, __m128i __A) |
| 8230 | { |
| 8231 | return (__m128i) __builtin_ia32_pmovqb128_mask ((__v2di) __A, |
| 8232 | (__v16qi) _mm_setzero_si128 (), |
| 8233 | __M); |
| 8234 | } |
| 8235 | |
| 8236 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8237 | _mm_mask_cvtepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) |
| 8238 | { |
| 8239 | __builtin_ia32_pmovqb128mem_mask ((__v16qi *) __P, (__v2di) __A, __M); |
| 8240 | } |
| 8241 | |
| 8242 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8243 | _mm256_cvtepi64_epi8 (__m256i __A) |
| 8244 | { |
| 8245 | return (__m128i) __builtin_ia32_pmovqb256_mask ((__v4di) __A, |
| 8246 | (__v16qi) _mm_undefined_si128(), |
| 8247 | (__mmask8) -1); |
| 8248 | } |
| 8249 | |
| 8250 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8251 | _mm256_mask_cvtepi64_epi8 (__m128i __O, __mmask8 __M, __m256i __A) |
| 8252 | { |
| 8253 | return (__m128i) __builtin_ia32_pmovqb256_mask ((__v4di) __A, |
| 8254 | (__v16qi) __O, __M); |
| 8255 | } |
| 8256 | |
| 8257 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8258 | _mm256_maskz_cvtepi64_epi8 (__mmask8 __M, __m256i __A) |
| 8259 | { |
| 8260 | return (__m128i) __builtin_ia32_pmovqb256_mask ((__v4di) __A, |
| 8261 | (__v16qi) _mm_setzero_si128 (), |
| 8262 | __M); |
| 8263 | } |
| 8264 | |
| 8265 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8266 | _mm256_mask_cvtepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) |
| 8267 | { |
| 8268 | __builtin_ia32_pmovqb256mem_mask ((__v16qi *) __P, (__v4di) __A, __M); |
| 8269 | } |
| 8270 | |
| 8271 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8272 | _mm_cvtepi64_epi32 (__m128i __A) |
| 8273 | { |
| 8274 | return (__m128i) __builtin_ia32_pmovqd128_mask ((__v2di) __A, |
| 8275 | (__v4si)_mm_undefined_si128(), |
| 8276 | (__mmask8) -1); |
| 8277 | } |
| 8278 | |
| 8279 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8280 | _mm_mask_cvtepi64_epi32 (__m128i __O, __mmask8 __M, __m128i __A) |
| 8281 | { |
| 8282 | return (__m128i) __builtin_ia32_pmovqd128_mask ((__v2di) __A, |
| 8283 | (__v4si) __O, __M); |
| 8284 | } |
| 8285 | |
| 8286 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8287 | _mm_maskz_cvtepi64_epi32 (__mmask8 __M, __m128i __A) |
| 8288 | { |
| 8289 | return (__m128i) __builtin_ia32_pmovqd128_mask ((__v2di) __A, |
| 8290 | (__v4si) _mm_setzero_si128 (), |
| 8291 | __M); |
| 8292 | } |
| 8293 | |
| 8294 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8295 | _mm_mask_cvtepi64_storeu_epi32 (void * __P, __mmask8 __M, __m128i __A) |
| 8296 | { |
| 8297 | __builtin_ia32_pmovqd128mem_mask ((__v4si *) __P, (__v2di) __A, __M); |
| 8298 | } |
| 8299 | |
| 8300 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8301 | _mm256_cvtepi64_epi32 (__m256i __A) |
| 8302 | { |
| 8303 | return (__m128i) __builtin_ia32_pmovqd256_mask ((__v4di) __A, |
| 8304 | (__v4si) _mm_undefined_si128(), |
| 8305 | (__mmask8) -1); |
| 8306 | } |
| 8307 | |
| 8308 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8309 | _mm256_mask_cvtepi64_epi32 (__m128i __O, __mmask8 __M, __m256i __A) |
| 8310 | { |
| 8311 | return (__m128i) __builtin_ia32_pmovqd256_mask ((__v4di) __A, |
| 8312 | (__v4si) __O, __M); |
| 8313 | } |
| 8314 | |
| 8315 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8316 | _mm256_maskz_cvtepi64_epi32 (__mmask8 __M, __m256i __A) |
| 8317 | { |
| 8318 | return (__m128i) __builtin_ia32_pmovqd256_mask ((__v4di) __A, |
| 8319 | (__v4si) _mm_setzero_si128 (), |
| 8320 | __M); |
| 8321 | } |
| 8322 | |
| 8323 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8324 | _mm256_mask_cvtepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A) |
| 8325 | { |
| 8326 | __builtin_ia32_pmovqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M); |
| 8327 | } |
| 8328 | |
| 8329 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8330 | _mm_cvtepi64_epi16 (__m128i __A) |
| 8331 | { |
| 8332 | return (__m128i) __builtin_ia32_pmovqw128_mask ((__v2di) __A, |
| 8333 | (__v8hi) _mm_undefined_si128(), |
| 8334 | (__mmask8) -1); |
| 8335 | } |
| 8336 | |
| 8337 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8338 | _mm_mask_cvtepi64_epi16 (__m128i __O, __mmask8 __M, __m128i __A) |
| 8339 | { |
| 8340 | return (__m128i) __builtin_ia32_pmovqw128_mask ((__v2di) __A, |
| 8341 | (__v8hi)__O, |
| 8342 | __M); |
| 8343 | } |
| 8344 | |
| 8345 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8346 | _mm_maskz_cvtepi64_epi16 (__mmask8 __M, __m128i __A) |
| 8347 | { |
| 8348 | return (__m128i) __builtin_ia32_pmovqw128_mask ((__v2di) __A, |
| 8349 | (__v8hi) _mm_setzero_si128 (), |
| 8350 | __M); |
| 8351 | } |
| 8352 | |
| 8353 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8354 | _mm_mask_cvtepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) |
| 8355 | { |
| 8356 | __builtin_ia32_pmovqw128mem_mask ((__v8hi *) __P, (__v2di) __A, __M); |
| 8357 | } |
| 8358 | |
| 8359 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8360 | _mm256_cvtepi64_epi16 (__m256i __A) |
| 8361 | { |
| 8362 | return (__m128i) __builtin_ia32_pmovqw256_mask ((__v4di) __A, |
| 8363 | (__v8hi)_mm_undefined_si128(), |
| 8364 | (__mmask8) -1); |
| 8365 | } |
| 8366 | |
| 8367 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8368 | _mm256_mask_cvtepi64_epi16 (__m128i __O, __mmask8 __M, __m256i __A) |
| 8369 | { |
| 8370 | return (__m128i) __builtin_ia32_pmovqw256_mask ((__v4di) __A, |
| 8371 | (__v8hi) __O, __M); |
| 8372 | } |
| 8373 | |
| 8374 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 8375 | _mm256_maskz_cvtepi64_epi16 (__mmask8 __M, __m256i __A) |
| 8376 | { |
| 8377 | return (__m128i) __builtin_ia32_pmovqw256_mask ((__v4di) __A, |
| 8378 | (__v8hi) _mm_setzero_si128 (), |
| 8379 | __M); |
| 8380 | } |
| 8381 | |
| 8382 | static __inline__ void __DEFAULT_FN_ATTRS |
| 8383 | _mm256_mask_cvtepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) |
| 8384 | { |
| 8385 | __builtin_ia32_pmovqw256mem_mask ((__v8hi *) __P, (__v4di) __A, __M); |
| 8386 | } |
| 8387 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8388 | #define _mm256_extractf32x4_ps(A, imm) __extension__ ({ \ |
| 8389 | (__m128)__builtin_ia32_extractf32x4_256_mask((__v8sf)(__m256)(A), \ |
| 8390 | (int)(imm), \ |
| 8391 | (__v4sf)_mm_setzero_ps(), \ |
| 8392 | (__mmask8)-1); }) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 8393 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8394 | #define _mm256_mask_extractf32x4_ps(W, U, A, imm) __extension__ ({ \ |
| 8395 | (__m128)__builtin_ia32_extractf32x4_256_mask((__v8sf)(__m256)(A), \ |
| 8396 | (int)(imm), \ |
| 8397 | (__v4sf)(__m128)(W), \ |
| 8398 | (__mmask8)(U)); }) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 8399 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8400 | #define _mm256_maskz_extractf32x4_ps(U, A, imm) __extension__ ({ \ |
| 8401 | (__m128)__builtin_ia32_extractf32x4_256_mask((__v8sf)(__m256)(A), \ |
| 8402 | (int)(imm), \ |
| 8403 | (__v4sf)_mm_setzero_ps(), \ |
| 8404 | (__mmask8)(U)); }) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 8405 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8406 | #define _mm256_extracti32x4_epi32(A, imm) __extension__ ({ \ |
| 8407 | (__m128i)__builtin_ia32_extracti32x4_256_mask((__v8si)(__m256i)(A), \ |
| 8408 | (int)(imm), \ |
| 8409 | (__v4si)_mm_setzero_si128(), \ |
| 8410 | (__mmask8)-1); }) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 8411 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8412 | #define _mm256_mask_extracti32x4_epi32(W, U, A, imm) __extension__ ({ \ |
| 8413 | (__m128i)__builtin_ia32_extracti32x4_256_mask((__v8si)(__m256i)(A), \ |
| 8414 | (int)(imm), \ |
| 8415 | (__v4si)(__m128i)(W), \ |
| 8416 | (__mmask8)(U)); }) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 8417 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8418 | #define _mm256_maskz_extracti32x4_epi32(U, A, imm) __extension__ ({ \ |
| 8419 | (__m128i)__builtin_ia32_extracti32x4_256_mask((__v8si)(__m256i)(A), \ |
| 8420 | (int)(imm), \ |
| 8421 | (__v4si)_mm_setzero_si128(), \ |
| 8422 | (__mmask8)(U)); }) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 8423 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8424 | #define _mm256_insertf32x4(A, B, imm) __extension__ ({ \ |
| 8425 | (__m256)__builtin_ia32_insertf32x4_256_mask((__v8sf)(__m256)(A), \ |
| 8426 | (__v4sf)(__m128)(B), (int)(imm), \ |
| 8427 | (__v8sf)_mm256_setzero_ps(), \ |
| 8428 | (__mmask8)-1); }) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 8429 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8430 | #define _mm256_mask_insertf32x4(W, U, A, B, imm) __extension__ ({ \ |
| 8431 | (__m256)__builtin_ia32_insertf32x4_256_mask((__v8sf)(__m256)(A), \ |
| 8432 | (__v4sf)(__m128)(B), (int)(imm), \ |
| 8433 | (__v8sf)(__m256)(W), \ |
| 8434 | (__mmask8)(U)); }) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 8435 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8436 | #define _mm256_maskz_insertf32x4(U, A, B, imm) __extension__ ({ \ |
| 8437 | (__m256)__builtin_ia32_insertf32x4_256_mask((__v8sf)(__m256)(A), \ |
| 8438 | (__v4sf)(__m128)(B), (int)(imm), \ |
| 8439 | (__v8sf)_mm256_setzero_ps(), \ |
| 8440 | (__mmask8)(U)); }) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 8441 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8442 | #define _mm256_inserti32x4(A, B, imm) __extension__ ({ \ |
| 8443 | (__m256i)__builtin_ia32_inserti32x4_256_mask((__v8si)(__m256i)(A), \ |
| 8444 | (__v4si)(__m128i)(B), \ |
| 8445 | (int)(imm), \ |
| 8446 | (__v8si)_mm256_setzero_si256(), \ |
| 8447 | (__mmask8)-1); }) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 8448 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8449 | #define _mm256_mask_inserti32x4(W, U, A, B, imm) __extension__ ({ \ |
| 8450 | (__m256i)__builtin_ia32_inserti32x4_256_mask((__v8si)(__m256i)(A), \ |
| 8451 | (__v4si)(__m128i)(B), \ |
| 8452 | (int)(imm), \ |
| 8453 | (__v8si)(__m256i)(W), \ |
| 8454 | (__mmask8)(U)); }) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 8455 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8456 | #define _mm256_maskz_inserti32x4(U, A, B, imm) __extension__ ({ \ |
| 8457 | (__m256i)__builtin_ia32_inserti32x4_256_mask((__v8si)(__m256i)(A), \ |
| 8458 | (__v4si)(__m128i)(B), \ |
| 8459 | (int)(imm), \ |
| 8460 | (__v8si)_mm256_setzero_si256(), \ |
| 8461 | (__mmask8)(U)); }) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 8462 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8463 | #define _mm_getmant_pd(A, B, C) __extension__({\ |
| 8464 | (__m128d)__builtin_ia32_getmantpd128_mask((__v2df)(__m128d)(A), \ |
| 8465 | (int)(((C)<<2) | (B)), \ |
| 8466 | (__v2df)_mm_setzero_pd(), \ |
| 8467 | (__mmask8)-1); }) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 8468 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8469 | #define _mm_mask_getmant_pd(W, U, A, B, C) __extension__({\ |
| 8470 | (__m128d)__builtin_ia32_getmantpd128_mask((__v2df)(__m128d)(A), \ |
| 8471 | (int)(((C)<<2) | (B)), \ |
| 8472 | (__v2df)(__m128d)(W), \ |
| 8473 | (__mmask8)(U)); }) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 8474 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8475 | #define _mm_maskz_getmant_pd(U, A, B, C) __extension__({\ |
| 8476 | (__m128d)__builtin_ia32_getmantpd128_mask((__v2df)(__m128d)(A), \ |
| 8477 | (int)(((C)<<2) | (B)), \ |
| 8478 | (__v2df)_mm_setzero_pd(), \ |
| 8479 | (__mmask8)(U)); }) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 8480 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8481 | #define _mm256_getmant_pd(A, B, C) __extension__ ({ \ |
| 8482 | (__m256d)__builtin_ia32_getmantpd256_mask((__v4df)(__m256d)(A), \ |
| 8483 | (int)(((C)<<2) | (B)), \ |
| 8484 | (__v4df)_mm256_setzero_pd(), \ |
| 8485 | (__mmask8)-1); }) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 8486 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8487 | #define _mm256_mask_getmant_pd(W, U, A, B, C) __extension__ ({ \ |
| 8488 | (__m256d)__builtin_ia32_getmantpd256_mask((__v4df)(__m256d)(A), \ |
| 8489 | (int)(((C)<<2) | (B)), \ |
| 8490 | (__v4df)(__m256d)(W), \ |
| 8491 | (__mmask8)(U)); }) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 8492 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8493 | #define _mm256_maskz_getmant_pd(U, A, B, C) __extension__ ({ \ |
| 8494 | (__m256d)__builtin_ia32_getmantpd256_mask((__v4df)(__m256d)(A), \ |
| 8495 | (int)(((C)<<2) | (B)), \ |
| 8496 | (__v4df)_mm256_setzero_pd(), \ |
| 8497 | (__mmask8)(U)); }) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 8498 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8499 | #define _mm_getmant_ps(A, B, C) __extension__ ({ \ |
| 8500 | (__m128)__builtin_ia32_getmantps128_mask((__v4sf)(__m128)(A), \ |
| 8501 | (int)(((C)<<2) | (B)), \ |
| 8502 | (__v4sf)_mm_setzero_ps(), \ |
| 8503 | (__mmask8)-1); }) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 8504 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8505 | #define _mm_mask_getmant_ps(W, U, A, B, C) __extension__ ({ \ |
| 8506 | (__m128)__builtin_ia32_getmantps128_mask((__v4sf)(__m128)(A), \ |
| 8507 | (int)(((C)<<2) | (B)), \ |
| 8508 | (__v4sf)(__m128)(W), \ |
| 8509 | (__mmask8)(U)); }) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 8510 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8511 | #define _mm_maskz_getmant_ps(U, A, B, C) __extension__ ({ \ |
| 8512 | (__m128)__builtin_ia32_getmantps128_mask((__v4sf)(__m128)(A), \ |
| 8513 | (int)(((C)<<2) | (B)), \ |
| 8514 | (__v4sf)_mm_setzero_ps(), \ |
| 8515 | (__mmask8)(U)); }) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 8516 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8517 | #define _mm256_getmant_ps(A, B, C) __extension__ ({ \ |
| 8518 | (__m256)__builtin_ia32_getmantps256_mask((__v8sf)(__m256)(A), \ |
| 8519 | (int)(((C)<<2) | (B)), \ |
| 8520 | (__v8sf)_mm256_setzero_ps(), \ |
| 8521 | (__mmask8)-1); }) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 8522 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8523 | #define _mm256_mask_getmant_ps(W, U, A, B, C) __extension__ ({ \ |
| 8524 | (__m256)__builtin_ia32_getmantps256_mask((__v8sf)(__m256)(A), \ |
| 8525 | (int)(((C)<<2) | (B)), \ |
| 8526 | (__v8sf)(__m256)(W), \ |
| 8527 | (__mmask8)(U)); }) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 8528 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8529 | #define _mm256_maskz_getmant_ps(U, A, B, C) __extension__ ({ \ |
| 8530 | (__m256)__builtin_ia32_getmantps256_mask((__v8sf)(__m256)(A), \ |
| 8531 | (int)(((C)<<2) | (B)), \ |
| 8532 | (__v8sf)_mm256_setzero_ps(), \ |
| 8533 | (__mmask8)(U)); }) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 8534 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8535 | #define _mm_mmask_i64gather_pd(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8536 | (__m128d)__builtin_ia32_gather3div2df((__v2df)(__m128d)(v1_old), \ |
| 8537 | (double const *)(addr), \ |
| 8538 | (__v2di)(__m128i)(index), \ |
| 8539 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8540 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8541 | #define _mm_mmask_i64gather_epi64(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8542 | (__m128i)__builtin_ia32_gather3div2di((__v2di)(__m128i)(v1_old), \ |
| 8543 | (long long const *)(addr), \ |
| 8544 | (__v2di)(__m128i)(index), \ |
| 8545 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8546 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8547 | #define _mm256_mmask_i64gather_pd(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8548 | (__m256d)__builtin_ia32_gather3div4df((__v4df)(__m256d)(v1_old), \ |
| 8549 | (double const *)(addr), \ |
| 8550 | (__v4di)(__m256i)(index), \ |
| 8551 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8552 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8553 | #define _mm256_mmask_i64gather_epi64(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8554 | (__m256i)__builtin_ia32_gather3div4di((__v4di)(__m256i)(v1_old), \ |
| 8555 | (long long const *)(addr), \ |
| 8556 | (__v4di)(__m256i)(index), \ |
| 8557 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8558 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8559 | #define _mm_mmask_i64gather_ps(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8560 | (__m128)__builtin_ia32_gather3div4sf((__v4sf)(__m128)(v1_old), \ |
| 8561 | (float const *)(addr), \ |
| 8562 | (__v2di)(__m128i)(index), \ |
| 8563 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8564 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8565 | #define _mm_mmask_i64gather_epi32(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8566 | (__m128i)__builtin_ia32_gather3div4si((__v4si)(__m128i)(v1_old), \ |
| 8567 | (int const *)(addr), \ |
| 8568 | (__v2di)(__m128i)(index), \ |
| 8569 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8570 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8571 | #define _mm256_mmask_i64gather_ps(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8572 | (__m128)__builtin_ia32_gather3div8sf((__v4sf)(__m128)(v1_old), \ |
| 8573 | (float const *)(addr), \ |
| 8574 | (__v4di)(__m256i)(index), \ |
| 8575 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8576 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8577 | #define _mm256_mmask_i64gather_epi32(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8578 | (__m128i)__builtin_ia32_gather3div8si((__v4si)(__m128i)(v1_old), \ |
| 8579 | (int const *)(addr), \ |
| 8580 | (__v4di)(__m256i)(index), \ |
| 8581 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8582 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8583 | #define _mm_mmask_i32gather_pd(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8584 | (__m128d)__builtin_ia32_gather3siv2df((__v2df)(__m128d)(v1_old), \ |
| 8585 | (double const *)(addr), \ |
| 8586 | (__v4si)(__m128i)(index), \ |
| 8587 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8588 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8589 | #define _mm_mmask_i32gather_epi64(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8590 | (__m128i)__builtin_ia32_gather3siv2di((__v2di)(__m128i)(v1_old), \ |
| 8591 | (long long const *)(addr), \ |
| 8592 | (__v4si)(__m128i)(index), \ |
| 8593 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8594 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8595 | #define _mm256_mmask_i32gather_pd(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8596 | (__m256d)__builtin_ia32_gather3siv4df((__v4df)(__m256d)(v1_old), \ |
| 8597 | (double const *)(addr), \ |
| 8598 | (__v4si)(__m128i)(index), \ |
| 8599 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8600 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8601 | #define _mm256_mmask_i32gather_epi64(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8602 | (__m256i)__builtin_ia32_gather3siv4di((__v4di)(__m256i)(v1_old), \ |
| 8603 | (long long const *)(addr), \ |
| 8604 | (__v4si)(__m128i)(index), \ |
| 8605 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8606 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8607 | #define _mm_mmask_i32gather_ps(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8608 | (__m128)__builtin_ia32_gather3siv4sf((__v4sf)(__m128)(v1_old), \ |
| 8609 | (float const *)(addr), \ |
| 8610 | (__v4si)(__m128i)(index), \ |
| 8611 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8612 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8613 | #define _mm_mmask_i32gather_epi32(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8614 | (__m128i)__builtin_ia32_gather3siv4si((__v4si)(__m128i)(v1_old), \ |
| 8615 | (int const *)(addr), \ |
| 8616 | (__v4si)(__m128i)(index), \ |
| 8617 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8618 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8619 | #define _mm256_mmask_i32gather_ps(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8620 | (__m256)__builtin_ia32_gather3siv8sf((__v8sf)(__m256)(v1_old), \ |
| 8621 | (float const *)(addr), \ |
| 8622 | (__v8si)(__m256i)(index), \ |
| 8623 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8624 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8625 | #define _mm256_mmask_i32gather_epi32(v1_old, mask, index, addr, scale) __extension__ ({\ |
| 8626 | (__m256i)__builtin_ia32_gather3siv8si((__v8si)(__m256i)(v1_old), \ |
| 8627 | (int const *)(addr), \ |
| 8628 | (__v8si)(__m256i)(index), \ |
| 8629 | (__mmask8)(mask), (int)(scale)); }) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 8630 | |
Simon Pilgrim | 30db811 | 2016-07-04 13:34:44 +0000 | [diff] [blame] | 8631 | #define _mm256_permutex_pd(X, C) __extension__ ({ \ |
| 8632 | (__m256d)__builtin_shufflevector((__v4df)(__m256d)(X), \ |
| 8633 | (__v4df)_mm256_undefined_pd(), \ |
Craig Topper | 2a383c9 | 2016-07-04 22:18:01 +0000 | [diff] [blame] | 8634 | ((C) >> 0) & 0x3, ((C) >> 2) & 0x3, \ |
| 8635 | ((C) >> 4) & 0x3, ((C) >> 6) & 0x3); }) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 8636 | |
Simon Pilgrim | 30db811 | 2016-07-04 13:34:44 +0000 | [diff] [blame] | 8637 | #define _mm256_mask_permutex_pd(W, U, X, C) __extension__ ({ \ |
| 8638 | (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \ |
| 8639 | (__v4df)_mm256_permutex_pd((X), (C)), \ |
| 8640 | (__v4df)(__m256d)(W)); }) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 8641 | |
Simon Pilgrim | 30db811 | 2016-07-04 13:34:44 +0000 | [diff] [blame] | 8642 | #define _mm256_maskz_permutex_pd(U, X, C) __extension__ ({ \ |
| 8643 | (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \ |
| 8644 | (__v4df)_mm256_permutex_pd((X), (C)), \ |
| 8645 | (__v4df)_mm256_setzero_pd()); }) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 8646 | |
Simon Pilgrim | 30db811 | 2016-07-04 13:34:44 +0000 | [diff] [blame] | 8647 | #define _mm256_permutex_epi64(X, C) __extension__ ({ \ |
| 8648 | (__m256i)__builtin_shufflevector((__v4di)(__m256i)(X), \ |
| 8649 | (__v4di)_mm256_undefined_si256(), \ |
Craig Topper | 2a383c9 | 2016-07-04 22:18:01 +0000 | [diff] [blame] | 8650 | ((C) >> 0) & 0x3, ((C) >> 2) & 0x3, \ |
| 8651 | ((C) >> 4) & 0x3, ((C) >> 6) & 0x3); }) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 8652 | |
Simon Pilgrim | 30db811 | 2016-07-04 13:34:44 +0000 | [diff] [blame] | 8653 | #define _mm256_mask_permutex_epi64(W, U, X, C) __extension__ ({ \ |
| 8654 | (__m256i)__builtin_ia32_selectq_256((__mmask8)(U), \ |
| 8655 | (__v4di)_mm256_permutex_epi64((X), (C)), \ |
| 8656 | (__v4di)(__m256i)(W)); }) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 8657 | |
Simon Pilgrim | 30db811 | 2016-07-04 13:34:44 +0000 | [diff] [blame] | 8658 | #define _mm256_maskz_permutex_epi64(U, X, C) __extension__ ({ \ |
| 8659 | (__m256i)__builtin_ia32_selectq_256((__mmask8)(U), \ |
| 8660 | (__v4di)_mm256_permutex_epi64((X), (C)), \ |
| 8661 | (__v4di)_mm256_setzero_si256()); }) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 8662 | |
| 8663 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 8664 | _mm256_permutexvar_pd (__m256i __X, __m256d __Y) |
| 8665 | { |
| 8666 | return (__m256d) __builtin_ia32_permvardf256_mask ((__v4df) __Y, |
| 8667 | (__v4di) __X, |
| 8668 | (__v4df) _mm256_undefined_si256 (), |
| 8669 | (__mmask8) -1); |
| 8670 | } |
| 8671 | |
| 8672 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 8673 | _mm256_mask_permutexvar_pd (__m256d __W, __mmask8 __U, __m256i __X, |
| 8674 | __m256d __Y) |
| 8675 | { |
| 8676 | return (__m256d) __builtin_ia32_permvardf256_mask ((__v4df) __Y, |
| 8677 | (__v4di) __X, |
| 8678 | (__v4df) __W, |
| 8679 | (__mmask8) __U); |
| 8680 | } |
| 8681 | |
| 8682 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 8683 | _mm256_maskz_permutexvar_pd (__mmask8 __U, __m256i __X, __m256d __Y) |
| 8684 | { |
| 8685 | return (__m256d) __builtin_ia32_permvardf256_mask ((__v4df) __Y, |
| 8686 | (__v4di) __X, |
| 8687 | (__v4df) _mm256_setzero_pd (), |
| 8688 | (__mmask8) __U); |
| 8689 | } |
| 8690 | |
| 8691 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 8692 | _mm256_maskz_permutexvar_epi64 (__mmask8 __M, __m256i __X, __m256i __Y) |
| 8693 | { |
| 8694 | return (__m256i) __builtin_ia32_permvardi256_mask ((__v4di) __Y, |
| 8695 | (__v4di) __X, |
| 8696 | (__v4di) _mm256_setzero_si256 (), |
| 8697 | (__mmask8) __M); |
| 8698 | } |
| 8699 | |
| 8700 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Michael Zuckerman | a72b49ef | 2016-07-05 11:30:31 +0000 | [diff] [blame] | 8701 | _mm256_permutexvar_epi64 ( __m256i __X, __m256i __Y) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 8702 | { |
| 8703 | return (__m256i) __builtin_ia32_permvardi256_mask ((__v4di) __Y, |
| 8704 | (__v4di) __X, |
| 8705 | (__v4di) _mm256_undefined_si256 (), |
| 8706 | (__mmask8) -1); |
| 8707 | } |
| 8708 | |
| 8709 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 8710 | _mm256_mask_permutexvar_epi64 (__m256i __W, __mmask8 __M, __m256i __X, |
| 8711 | __m256i __Y) |
| 8712 | { |
| 8713 | return (__m256i) __builtin_ia32_permvardi256_mask ((__v4di) __Y, |
| 8714 | (__v4di) __X, |
| 8715 | (__v4di) __W, |
| 8716 | __M); |
| 8717 | } |
| 8718 | |
| 8719 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 8720 | _mm256_mask_permutexvar_ps (__m256 __W, __mmask8 __U, __m256i __X, |
| 8721 | __m256 __Y) |
| 8722 | { |
| 8723 | return (__m256) __builtin_ia32_permvarsf256_mask ((__v8sf) __Y, |
| 8724 | (__v8si) __X, |
| 8725 | (__v8sf) __W, |
| 8726 | (__mmask8) __U); |
| 8727 | } |
| 8728 | |
| 8729 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 8730 | _mm256_maskz_permutexvar_ps (__mmask8 __U, __m256i __X, __m256 __Y) |
| 8731 | { |
| 8732 | return (__m256) __builtin_ia32_permvarsf256_mask ((__v8sf) __Y, |
| 8733 | (__v8si) __X, |
| 8734 | (__v8sf) _mm256_setzero_ps (), |
| 8735 | (__mmask8) __U); |
| 8736 | } |
| 8737 | |
| 8738 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 8739 | _mm256_permutexvar_ps (__m256i __X, __m256 __Y) |
| 8740 | { |
| 8741 | return (__m256) __builtin_ia32_permvarsf256_mask ((__v8sf) __Y, |
| 8742 | (__v8si) __X, |
| 8743 | (__v8sf) _mm256_undefined_si256 (), |
| 8744 | (__mmask8) -1); |
| 8745 | } |
| 8746 | |
| 8747 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 8748 | _mm256_maskz_permutexvar_epi32 (__mmask8 __M, __m256i __X, __m256i __Y) |
| 8749 | { |
| 8750 | return (__m256i) __builtin_ia32_permvarsi256_mask ((__v8si) __Y, |
| 8751 | (__v8si) __X, |
| 8752 | (__v8si) _mm256_setzero_si256 (), |
| 8753 | __M); |
| 8754 | } |
| 8755 | |
| 8756 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 8757 | _mm256_mask_permutexvar_epi32 (__m256i __W, __mmask8 __M, __m256i __X, |
| 8758 | __m256i __Y) |
| 8759 | { |
| 8760 | return (__m256i) __builtin_ia32_permvarsi256_mask ((__v8si) __Y, |
| 8761 | (__v8si) __X, |
| 8762 | (__v8si) __W, |
| 8763 | (__mmask8) __M); |
| 8764 | } |
| 8765 | |
| 8766 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 8767 | _mm256_permutexvar_epi32 (__m256i __X, __m256i __Y) |
| 8768 | { |
| 8769 | return (__m256i) __builtin_ia32_permvarsi256_mask ((__v8si) __Y, |
| 8770 | (__v8si) __X, |
| 8771 | (__v8si) _mm256_undefined_si256(), |
| 8772 | (__mmask8) -1); |
| 8773 | } |
| 8774 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8775 | #define _mm_alignr_epi32(A, B, imm) __extension__ ({ \ |
| 8776 | (__m128i)__builtin_ia32_alignd128_mask((__v4si)(__m128i)(A), \ |
| 8777 | (__v4si)(__m128i)(B), (int)(imm), \ |
| 8778 | (__v4si)_mm_undefined_si128(), \ |
| 8779 | (__mmask8)-1); }) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8780 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8781 | #define _mm_mask_alignr_epi32(W, U, A, B, imm) __extension__ ({ \ |
| 8782 | (__m128i)__builtin_ia32_alignd128_mask((__v4si)(__m128i)(A), \ |
| 8783 | (__v4si)(__m128i)(B), (int)(imm), \ |
| 8784 | (__v4si)(__m128i)(W), \ |
| 8785 | (__mmask8)(U)); }) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8786 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8787 | #define _mm_maskz_alignr_epi32(U, A, B, imm) __extension__ ({ \ |
| 8788 | (__m128i)__builtin_ia32_alignd128_mask((__v4si)(__m128i)(A), \ |
| 8789 | (__v4si)(__m128i)(B), (int)(imm), \ |
| 8790 | (__v4si)_mm_setzero_si128(), \ |
| 8791 | (__mmask8)(U)); }) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8792 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8793 | #define _mm256_alignr_epi32(A, B, imm) __extension__ ({ \ |
| 8794 | (__m256i)__builtin_ia32_alignd256_mask((__v8si)(__m256i)(A), \ |
| 8795 | (__v8si)(__m256i)(B), (int)(imm), \ |
| 8796 | (__v8si)_mm256_undefined_si256(), \ |
| 8797 | (__mmask8)-1); }) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8798 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8799 | #define _mm256_mask_alignr_epi32(W, U, A, B, imm) __extension__ ({ \ |
| 8800 | (__m256i)__builtin_ia32_alignd256_mask((__v8si)(__m256i)(A), \ |
| 8801 | (__v8si)(__m256i)(B), (int)(imm), \ |
| 8802 | (__v8si)(__m256i)(W), \ |
| 8803 | (__mmask8)(U)); }) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8804 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8805 | #define _mm256_maskz_alignr_epi32(U, A, B, imm) __extension__ ({ \ |
| 8806 | (__m256i)__builtin_ia32_alignd256_mask((__v8si)(__m256i)(A), \ |
| 8807 | (__v8si)(__m256i)(B), (int)(imm), \ |
| 8808 | (__v8si)_mm256_setzero_si256(), \ |
| 8809 | (__mmask8)(U)); }) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8810 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8811 | #define _mm_alignr_epi64(A, B, imm) __extension__ ({ \ |
| 8812 | (__m128i)__builtin_ia32_alignq128_mask((__v2di)(__m128i)(A), \ |
| 8813 | (__v2di)(__m128i)(B), (int)(imm), \ |
| 8814 | (__v2di)_mm_setzero_di(), \ |
| 8815 | (__mmask8)-1); }) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8816 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8817 | #define _mm_mask_alignr_epi64(W, U, A, B, imm) __extension__ ({ \ |
| 8818 | (__m128i)__builtin_ia32_alignq128_mask((__v2di)(__m128i)(A), \ |
| 8819 | (__v2di)(__m128i)(B), (int)(imm), \ |
| 8820 | (__v2di)(__m128i)(W), \ |
| 8821 | (__mmask8)(U)); }) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8822 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8823 | #define _mm_maskz_alignr_epi64(U, A, B, imm) __extension__ ({ \ |
| 8824 | (__m128i)__builtin_ia32_alignq128_mask((__v2di)(__m128i)(A), \ |
| 8825 | (__v2di)(__m128i)(B), (int)(imm), \ |
| 8826 | (__v2di)_mm_setzero_di(), \ |
| 8827 | (__mmask8)(U)); }) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8828 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8829 | #define _mm256_alignr_epi64(A, B, imm) __extension__ ({ \ |
| 8830 | (__m256i)__builtin_ia32_alignq256_mask((__v4di)(__m256i)(A), \ |
| 8831 | (__v4di)(__m256i)(B), (int)(imm), \ |
| 8832 | (__v4di)_mm256_undefined_pd(), \ |
| 8833 | (__mmask8)-1); }) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8834 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8835 | #define _mm256_mask_alignr_epi64(W, U, A, B, imm) __extension__ ({ \ |
| 8836 | (__m256i)__builtin_ia32_alignq256_mask((__v4di)(__m256i)(A), \ |
| 8837 | (__v4di)(__m256i)(B), (int)(imm), \ |
| 8838 | (__v4di)(__m256i)(W), \ |
| 8839 | (__mmask8)(U)); }) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8840 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8841 | #define _mm256_maskz_alignr_epi64(U, A, B, imm) __extension__ ({ \ |
| 8842 | (__m256i)__builtin_ia32_alignq256_mask((__v4di)(__m256i)(A), \ |
| 8843 | (__v4di)(__m256i)(B), (int)(imm), \ |
| 8844 | (__v4di)_mm256_setzero_si256(), \ |
| 8845 | (__mmask8)(U)); }) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8846 | |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8847 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 8848 | _mm_mask_movehdup_ps (__m128 __W, __mmask8 __U, __m128 __A) |
| 8849 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 8850 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 8851 | (__v4sf)_mm_movehdup_ps(__A), |
| 8852 | (__v4sf)__W); |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8853 | } |
| 8854 | |
| 8855 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 8856 | _mm_maskz_movehdup_ps (__mmask8 __U, __m128 __A) |
| 8857 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 8858 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 8859 | (__v4sf)_mm_movehdup_ps(__A), |
| 8860 | (__v4sf)_mm_setzero_ps()); |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8861 | } |
| 8862 | |
| 8863 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 8864 | _mm256_mask_movehdup_ps (__m256 __W, __mmask8 __U, __m256 __A) |
| 8865 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 8866 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 8867 | (__v8sf)_mm256_movehdup_ps(__A), |
| 8868 | (__v8sf)__W); |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8869 | } |
| 8870 | |
| 8871 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 8872 | _mm256_maskz_movehdup_ps (__mmask8 __U, __m256 __A) |
| 8873 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 8874 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 8875 | (__v8sf)_mm256_movehdup_ps(__A), |
| 8876 | (__v8sf)_mm256_setzero_ps()); |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8877 | } |
| 8878 | |
| 8879 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 8880 | _mm_mask_moveldup_ps (__m128 __W, __mmask8 __U, __m128 __A) |
| 8881 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 8882 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 8883 | (__v4sf)_mm_moveldup_ps(__A), |
| 8884 | (__v4sf)__W); |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8885 | } |
| 8886 | |
| 8887 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 8888 | _mm_maskz_moveldup_ps (__mmask8 __U, __m128 __A) |
| 8889 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 8890 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 8891 | (__v4sf)_mm_moveldup_ps(__A), |
| 8892 | (__v4sf)_mm_setzero_ps()); |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8893 | } |
| 8894 | |
| 8895 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 8896 | _mm256_mask_moveldup_ps (__m256 __W, __mmask8 __U, __m256 __A) |
| 8897 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 8898 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 8899 | (__v8sf)_mm256_moveldup_ps(__A), |
| 8900 | (__v8sf)__W); |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8901 | } |
| 8902 | |
| 8903 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 8904 | _mm256_maskz_moveldup_ps (__mmask8 __U, __m256 __A) |
| 8905 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 8906 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 8907 | (__v8sf)_mm256_moveldup_ps(__A), |
| 8908 | (__v8sf)_mm256_setzero_ps()); |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8909 | } |
| 8910 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8911 | #define _mm256_mask_shuffle_epi32(W, U, A, I) __extension__({\ |
Craig Topper | 7cc9263 | 2016-06-11 12:50:19 +0000 | [diff] [blame] | 8912 | (__m256i)__builtin_ia32_selectd_256((__mmask8)(U), \ |
| 8913 | (__v8si)_mm256_shuffle_epi32((A), (I)), \ |
| 8914 | (__v8si)(__m256i)(W)); }) |
Michael Zuckerman | c62f27e | 2016-05-02 07:35:27 +0000 | [diff] [blame] | 8915 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8916 | #define _mm256_maskz_shuffle_epi32(U, A, I) __extension__({\ |
Craig Topper | 7cc9263 | 2016-06-11 12:50:19 +0000 | [diff] [blame] | 8917 | (__m256i)__builtin_ia32_selectd_256((__mmask8)(U), \ |
| 8918 | (__v8si)_mm256_shuffle_epi32((A), (I)), \ |
| 8919 | (__v8si)_mm256_setzero_si256()); }) |
Michael Zuckerman | c62f27e | 2016-05-02 07:35:27 +0000 | [diff] [blame] | 8920 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8921 | #define _mm_mask_shuffle_epi32(W, U, A, I) __extension__({\ |
Craig Topper | 7cc9263 | 2016-06-11 12:50:19 +0000 | [diff] [blame] | 8922 | (__m128i)__builtin_ia32_selectd_128((__mmask8)(U), \ |
| 8923 | (__v4si)_mm_shuffle_epi32((A), (I)), \ |
| 8924 | (__v4si)(__m128i)(W)); }) |
Michael Zuckerman | c62f27e | 2016-05-02 07:35:27 +0000 | [diff] [blame] | 8925 | |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8926 | #define _mm_maskz_shuffle_epi32(U, A, I) __extension__({\ |
Craig Topper | 7cc9263 | 2016-06-11 12:50:19 +0000 | [diff] [blame] | 8927 | (__m128i)__builtin_ia32_selectd_128((__mmask8)(U), \ |
| 8928 | (__v4si)_mm_shuffle_epi32((A), (I)), \ |
| 8929 | (__v4si)_mm_setzero_si128()); }) |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 8930 | |
| 8931 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 8932 | _mm_mask_mov_pd (__m128d __W, __mmask8 __U, __m128d __A) |
| 8933 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 8934 | return (__m128d) __builtin_ia32_selectpd_128 ((__mmask8) __U, |
| 8935 | (__v2df) __A, |
| 8936 | (__v2df) __W); |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 8937 | } |
| 8938 | |
| 8939 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 8940 | _mm_maskz_mov_pd (__mmask8 __U, __m128d __A) |
| 8941 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 8942 | return (__m128d) __builtin_ia32_selectpd_128 ((__mmask8) __U, |
| 8943 | (__v2df) __A, |
| 8944 | (__v2df) _mm_setzero_pd ()); |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 8945 | } |
| 8946 | |
| 8947 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 8948 | _mm256_mask_mov_pd (__m256d __W, __mmask8 __U, __m256d __A) |
| 8949 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 8950 | return (__m256d) __builtin_ia32_selectpd_256 ((__mmask8) __U, |
| 8951 | (__v4df) __A, |
| 8952 | (__v4df) __W); |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 8953 | } |
| 8954 | |
| 8955 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 8956 | _mm256_maskz_mov_pd (__mmask8 __U, __m256d __A) |
| 8957 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 8958 | return (__m256d) __builtin_ia32_selectpd_256 ((__mmask8) __U, |
| 8959 | (__v4df) __A, |
| 8960 | (__v4df) _mm256_setzero_pd ()); |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 8961 | } |
| 8962 | |
| 8963 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 8964 | _mm_mask_mov_ps (__m128 __W, __mmask8 __U, __m128 __A) |
| 8965 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 8966 | return (__m128) __builtin_ia32_selectps_128 ((__mmask8) __U, |
| 8967 | (__v4sf) __A, |
| 8968 | (__v4sf) __W); |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 8969 | } |
| 8970 | |
| 8971 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 8972 | _mm_maskz_mov_ps (__mmask8 __U, __m128 __A) |
| 8973 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 8974 | return (__m128) __builtin_ia32_selectps_128 ((__mmask8) __U, |
| 8975 | (__v4sf) __A, |
| 8976 | (__v4sf) _mm_setzero_ps ()); |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 8977 | } |
| 8978 | |
| 8979 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 8980 | _mm256_mask_mov_ps (__m256 __W, __mmask8 __U, __m256 __A) |
| 8981 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 8982 | return (__m256) __builtin_ia32_selectps_256 ((__mmask8) __U, |
| 8983 | (__v8sf) __A, |
| 8984 | (__v8sf) __W); |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 8985 | } |
| 8986 | |
| 8987 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 8988 | _mm256_maskz_mov_ps (__mmask8 __U, __m256 __A) |
| 8989 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 8990 | return (__m256) __builtin_ia32_selectps_256 ((__mmask8) __U, |
| 8991 | (__v8sf) __A, |
| 8992 | (__v8sf) _mm256_setzero_ps ()); |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 8993 | } |
| 8994 | |
Michael Zuckerman | 8bfb777 | 2016-05-03 12:45:04 +0000 | [diff] [blame] | 8995 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 8996 | _mm_mask_cvtph_ps (__m128 __W, __mmask8 __U, __m128i __A) |
| 8997 | { |
| 8998 | return (__m128) __builtin_ia32_vcvtph2ps_mask ((__v8hi) __A, |
| 8999 | (__v4sf) __W, |
| 9000 | (__mmask8) __U); |
| 9001 | } |
| 9002 | |
| 9003 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 9004 | _mm_maskz_cvtph_ps (__mmask8 __U, __m128i __A) |
| 9005 | { |
| 9006 | return (__m128) __builtin_ia32_vcvtph2ps_mask ((__v8hi) __A, |
| 9007 | (__v4sf) |
| 9008 | _mm_setzero_ps (), |
| 9009 | (__mmask8) __U); |
| 9010 | } |
| 9011 | |
| 9012 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 9013 | _mm256_mask_cvtph_ps (__m256 __W, __mmask8 __U, __m128i __A) |
| 9014 | { |
| 9015 | return (__m256) __builtin_ia32_vcvtph2ps256_mask ((__v8hi) __A, |
| 9016 | (__v8sf) __W, |
| 9017 | (__mmask8) __U); |
| 9018 | } |
| 9019 | |
| 9020 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 9021 | _mm256_maskz_cvtph_ps (__mmask8 __U, __m128i __A) |
| 9022 | { |
| 9023 | return (__m256) __builtin_ia32_vcvtph2ps256_mask ((__v8hi) __A, |
| 9024 | (__v8sf) |
| 9025 | _mm256_setzero_ps (), |
| 9026 | (__mmask8) __U); |
| 9027 | } |
| 9028 | |
Michael Zuckerman | 9e7d0a9 | 2016-06-02 07:44:08 +0000 | [diff] [blame] | 9029 | static __inline __m128i __DEFAULT_FN_ATTRS |
| 9030 | _mm_mask_cvtps_ph (__m128i __W, __mmask8 __U, __m128 __A) |
| 9031 | { |
| 9032 | return (__m128i) __builtin_ia32_vcvtps2ph_mask ((__v4sf) __A, _MM_FROUND_CUR_DIRECTION, |
| 9033 | (__v8hi) __W, |
| 9034 | (__mmask8) __U); |
| 9035 | } |
| 9036 | |
| 9037 | static __inline __m128i __DEFAULT_FN_ATTRS |
| 9038 | _mm_maskz_cvtps_ph (__mmask8 __U, __m128 __A) |
| 9039 | { |
| 9040 | return (__m128i) __builtin_ia32_vcvtps2ph_mask ((__v4sf) __A, _MM_FROUND_CUR_DIRECTION, |
| 9041 | (__v8hi) _mm_setzero_si128 (), |
| 9042 | (__mmask8) __U); |
| 9043 | } |
| 9044 | |
| 9045 | #define _mm_mask_cvt_roundps_ph(W, U, A, I) __extension__ ({ \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 9046 | (__m128i)__builtin_ia32_vcvtps2ph_mask((__v4sf)(__m128)(A), (int)(I), \ |
| 9047 | (__v8hi)(__m128i)(W), \ |
| 9048 | (__mmask8)(U)); }) |
Michael Zuckerman | 8bfb777 | 2016-05-03 12:45:04 +0000 | [diff] [blame] | 9049 | |
Michael Zuckerman | 9e7d0a9 | 2016-06-02 07:44:08 +0000 | [diff] [blame] | 9050 | #define _mm_maskz_cvt_roundps_ph(U, A, I) __extension__ ({ \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 9051 | (__m128i)__builtin_ia32_vcvtps2ph_mask((__v4sf)(__m128)(A), (int)(I), \ |
| 9052 | (__v8hi)_mm_setzero_si128(), \ |
| 9053 | (__mmask8)(U)); }) |
Michael Zuckerman | 8bfb777 | 2016-05-03 12:45:04 +0000 | [diff] [blame] | 9054 | |
Michael Zuckerman | 9e7d0a9 | 2016-06-02 07:44:08 +0000 | [diff] [blame] | 9055 | static __inline __m128i __DEFAULT_FN_ATTRS |
| 9056 | _mm256_mask_cvtps_ph (__m128i __W, __mmask8 __U, __m256 __A) |
| 9057 | { |
| 9058 | return (__m128i) __builtin_ia32_vcvtps2ph256_mask ((__v8sf) __A, _MM_FROUND_CUR_DIRECTION, |
| 9059 | (__v8hi) __W, |
| 9060 | (__mmask8) __U); |
| 9061 | } |
| 9062 | |
| 9063 | static __inline __m128i __DEFAULT_FN_ATTRS |
| 9064 | _mm256_maskz_cvtps_ph ( __mmask8 __U, __m256 __A) |
| 9065 | { |
| 9066 | return (__m128i) __builtin_ia32_vcvtps2ph256_mask ((__v8sf) __A, _MM_FROUND_CUR_DIRECTION, |
| 9067 | (__v8hi) _mm_setzero_si128(), |
| 9068 | (__mmask8) __U); |
| 9069 | } |
| 9070 | #define _mm256_mask_cvt_roundps_ph(W, U, A, I) __extension__ ({ \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 9071 | (__m128i)__builtin_ia32_vcvtps2ph256_mask((__v8sf)(__m256)(A), (int)(I), \ |
| 9072 | (__v8hi)(__m128i)(W), \ |
| 9073 | (__mmask8)(U)); }) |
Michael Zuckerman | 8bfb777 | 2016-05-03 12:45:04 +0000 | [diff] [blame] | 9074 | |
Michael Zuckerman | 9e7d0a9 | 2016-06-02 07:44:08 +0000 | [diff] [blame] | 9075 | #define _mm256_maskz_cvt_roundps_ph(U, A, I) __extension__ ({ \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 9076 | (__m128i)__builtin_ia32_vcvtps2ph256_mask((__v8sf)(__m256)(A), (int)(I), \ |
| 9077 | (__v8hi)_mm_setzero_si128(), \ |
| 9078 | (__mmask8)(U)); }) |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 9079 | |
Michael Zuckerman | 9e7d0a9 | 2016-06-02 07:44:08 +0000 | [diff] [blame] | 9080 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 9081 | #undef __DEFAULT_FN_ATTRS |
Eric Christopher | 4d185168 | 2015-06-17 07:09:20 +0000 | [diff] [blame] | 9082 | |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 9083 | #endif /* __AVX512VLINTRIN_H */ |