Craig Topper | 880f60b | 2015-11-11 08:00:39 +0000 | [diff] [blame] | 1 | /*===---- avx512vldqintrin.h - AVX512VL and AVX512DQ intrinsics ------------=== |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 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 | */ |
Sean Silva | e4c3760 | 2015-09-12 02:55:19 +0000 | [diff] [blame] | 23 | |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 24 | #ifndef __IMMINTRIN_H |
| 25 | #error "Never use <avx512vldqintrin.h> directly; include <immintrin.h> instead." |
| 26 | #endif |
| 27 | |
| 28 | #ifndef __AVX512VLDQINTRIN_H |
| 29 | #define __AVX512VLDQINTRIN_H |
| 30 | |
Eric Christopher | 4d185168 | 2015-06-17 07:09:20 +0000 | [diff] [blame] | 31 | /* Define the default attributes for the functions in this file. */ |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 32 | #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512vl,avx512dq"))) |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 33 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 34 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 35 | _mm256_mullo_epi64 (__m256i __A, __m256i __B) { |
Craig Topper | 6a77b62 | 2016-06-04 05:43:41 +0000 | [diff] [blame] | 36 | return (__m256i) ((__v4du) __A * (__v4du) __B); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 37 | } |
| 38 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 39 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | f43e4a1 | 2016-09-03 19:19:49 +0000 | [diff] [blame] | 40 | _mm256_mask_mullo_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) { |
| 41 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 42 | (__v4di)_mm256_mullo_epi64(__A, __B), |
| 43 | (__v4di)__W); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 44 | } |
| 45 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 46 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | f43e4a1 | 2016-09-03 19:19:49 +0000 | [diff] [blame] | 47 | _mm256_maskz_mullo_epi64(__mmask8 __U, __m256i __A, __m256i __B) { |
| 48 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 49 | (__v4di)_mm256_mullo_epi64(__A, __B), |
| 50 | (__v4di)_mm256_setzero_si256()); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 51 | } |
| 52 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 53 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 54 | _mm_mullo_epi64 (__m128i __A, __m128i __B) { |
Craig Topper | 6a77b62 | 2016-06-04 05:43:41 +0000 | [diff] [blame] | 55 | return (__m128i) ((__v2du) __A * (__v2du) __B); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 56 | } |
| 57 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 58 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | f43e4a1 | 2016-09-03 19:19:49 +0000 | [diff] [blame] | 59 | _mm_mask_mullo_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) { |
| 60 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 61 | (__v2di)_mm_mullo_epi64(__A, __B), |
| 62 | (__v2di)__W); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 63 | } |
| 64 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 65 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | f43e4a1 | 2016-09-03 19:19:49 +0000 | [diff] [blame] | 66 | _mm_maskz_mullo_epi64(__mmask8 __U, __m128i __A, __m128i __B) { |
| 67 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 68 | (__v2di)_mm_mullo_epi64(__A, __B), |
| 69 | (__v2di)_mm_setzero_si128()); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 70 | } |
| 71 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 72 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 73 | _mm256_mask_andnot_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) { |
| 74 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 75 | (__v4df)_mm256_andnot_pd(__A, __B), |
| 76 | (__v4df)__W); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 77 | } |
| 78 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 79 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 80 | _mm256_maskz_andnot_pd(__mmask8 __U, __m256d __A, __m256d __B) { |
| 81 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 82 | (__v4df)_mm256_andnot_pd(__A, __B), |
| 83 | (__v4df)_mm256_setzero_pd()); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 84 | } |
| 85 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 86 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 87 | _mm_mask_andnot_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 88 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 89 | (__v2df)_mm_andnot_pd(__A, __B), |
| 90 | (__v2df)__W); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 91 | } |
| 92 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 93 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 94 | _mm_maskz_andnot_pd(__mmask8 __U, __m128d __A, __m128d __B) { |
| 95 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 96 | (__v2df)_mm_andnot_pd(__A, __B), |
| 97 | (__v2df)_mm_setzero_pd()); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 98 | } |
| 99 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 100 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 101 | _mm256_mask_andnot_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 102 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 103 | (__v8sf)_mm256_andnot_ps(__A, __B), |
| 104 | (__v8sf)__W); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 105 | } |
| 106 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 107 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 108 | _mm256_maskz_andnot_ps(__mmask8 __U, __m256 __A, __m256 __B) { |
| 109 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 110 | (__v8sf)_mm256_andnot_ps(__A, __B), |
| 111 | (__v8sf)_mm256_setzero_ps()); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 112 | } |
| 113 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 114 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 115 | _mm_mask_andnot_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 116 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 117 | (__v4sf)_mm_andnot_ps(__A, __B), |
| 118 | (__v4sf)__W); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 119 | } |
| 120 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 121 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 122 | _mm_maskz_andnot_ps(__mmask8 __U, __m128 __A, __m128 __B) { |
| 123 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 124 | (__v4sf)_mm_andnot_ps(__A, __B), |
| 125 | (__v4sf)_mm_setzero_ps()); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 126 | } |
| 127 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 128 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 129 | _mm256_mask_and_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) { |
| 130 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 131 | (__v4df)_mm256_and_pd(__A, __B), |
| 132 | (__v4df)__W); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 133 | } |
| 134 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 135 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 136 | _mm256_maskz_and_pd(__mmask8 __U, __m256d __A, __m256d __B) { |
| 137 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 138 | (__v4df)_mm256_and_pd(__A, __B), |
| 139 | (__v4df)_mm256_setzero_pd()); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 140 | } |
| 141 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 142 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 143 | _mm_mask_and_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 144 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 145 | (__v2df)_mm_and_pd(__A, __B), |
| 146 | (__v2df)__W); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 147 | } |
| 148 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 149 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 150 | _mm_maskz_and_pd(__mmask8 __U, __m128d __A, __m128d __B) { |
| 151 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 152 | (__v2df)_mm_and_pd(__A, __B), |
| 153 | (__v2df)_mm_setzero_pd()); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 154 | } |
| 155 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 156 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 157 | _mm256_mask_and_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 158 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 159 | (__v8sf)_mm256_and_ps(__A, __B), |
| 160 | (__v8sf)__W); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 161 | } |
| 162 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 163 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 164 | _mm256_maskz_and_ps(__mmask8 __U, __m256 __A, __m256 __B) { |
| 165 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 166 | (__v8sf)_mm256_and_ps(__A, __B), |
| 167 | (__v8sf)_mm256_setzero_ps()); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 168 | } |
| 169 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 170 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 171 | _mm_mask_and_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 172 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 173 | (__v4sf)_mm_and_ps(__A, __B), |
| 174 | (__v4sf)__W); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 175 | } |
| 176 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 177 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 178 | _mm_maskz_and_ps(__mmask8 __U, __m128 __A, __m128 __B) { |
| 179 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 180 | (__v4sf)_mm_and_ps(__A, __B), |
| 181 | (__v4sf)_mm_setzero_ps()); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 182 | } |
| 183 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 184 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 185 | _mm256_mask_xor_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) { |
| 186 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 187 | (__v4df)_mm256_xor_pd(__A, __B), |
| 188 | (__v4df)__W); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 189 | } |
| 190 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 191 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 192 | _mm256_maskz_xor_pd(__mmask8 __U, __m256d __A, __m256d __B) { |
| 193 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 194 | (__v4df)_mm256_xor_pd(__A, __B), |
| 195 | (__v4df)_mm256_setzero_pd()); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 196 | } |
| 197 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 198 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 199 | _mm_mask_xor_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 200 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 201 | (__v2df)_mm_xor_pd(__A, __B), |
| 202 | (__v2df)__W); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 203 | } |
| 204 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 205 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 206 | _mm_maskz_xor_pd (__mmask8 __U, __m128d __A, __m128d __B) { |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 207 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 208 | (__v2df)_mm_xor_pd(__A, __B), |
| 209 | (__v2df)_mm_setzero_pd()); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 210 | } |
| 211 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 212 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 213 | _mm256_mask_xor_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 214 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 215 | (__v8sf)_mm256_xor_ps(__A, __B), |
| 216 | (__v8sf)__W); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 217 | } |
| 218 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 219 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 220 | _mm256_maskz_xor_ps(__mmask8 __U, __m256 __A, __m256 __B) { |
| 221 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 222 | (__v8sf)_mm256_xor_ps(__A, __B), |
| 223 | (__v8sf)_mm256_setzero_ps()); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 224 | } |
| 225 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 226 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 227 | _mm_mask_xor_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 228 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 229 | (__v4sf)_mm_xor_ps(__A, __B), |
| 230 | (__v4sf)__W); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 231 | } |
| 232 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 233 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 234 | _mm_maskz_xor_ps(__mmask8 __U, __m128 __A, __m128 __B) { |
| 235 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 236 | (__v4sf)_mm_xor_ps(__A, __B), |
| 237 | (__v4sf)_mm_setzero_ps()); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 238 | } |
| 239 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 240 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 241 | _mm256_mask_or_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) { |
| 242 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 243 | (__v4df)_mm256_or_pd(__A, __B), |
| 244 | (__v4df)__W); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 245 | } |
| 246 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 247 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 248 | _mm256_maskz_or_pd(__mmask8 __U, __m256d __A, __m256d __B) { |
| 249 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 250 | (__v4df)_mm256_or_pd(__A, __B), |
| 251 | (__v4df)_mm256_setzero_pd()); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 252 | } |
| 253 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 254 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 255 | _mm_mask_or_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 256 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 257 | (__v2df)_mm_or_pd(__A, __B), |
| 258 | (__v2df)__W); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 259 | } |
| 260 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 261 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 262 | _mm_maskz_or_pd(__mmask8 __U, __m128d __A, __m128d __B) { |
| 263 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 264 | (__v2df)_mm_or_pd(__A, __B), |
| 265 | (__v2df)_mm_setzero_pd()); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 266 | } |
| 267 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 268 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 269 | _mm256_mask_or_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 270 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 271 | (__v8sf)_mm256_or_ps(__A, __B), |
| 272 | (__v8sf)__W); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 273 | } |
| 274 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 275 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 276 | _mm256_maskz_or_ps(__mmask8 __U, __m256 __A, __m256 __B) { |
| 277 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 278 | (__v8sf)_mm256_or_ps(__A, __B), |
| 279 | (__v8sf)_mm256_setzero_ps()); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 280 | } |
| 281 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 282 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 283 | _mm_mask_or_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 284 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 285 | (__v4sf)_mm_or_ps(__A, __B), |
| 286 | (__v4sf)__W); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 287 | } |
| 288 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 289 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | a815f48 | 2016-08-31 05:38:58 +0000 | [diff] [blame] | 290 | _mm_maskz_or_ps(__mmask8 __U, __m128 __A, __m128 __B) { |
| 291 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 292 | (__v4sf)_mm_or_ps(__A, __B), |
| 293 | (__v4sf)_mm_setzero_ps()); |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 294 | } |
| 295 | |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 296 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 297 | _mm_cvtpd_epi64 (__m128d __A) { |
| 298 | return (__m128i) __builtin_ia32_cvtpd2qq128_mask ((__v2df) __A, |
| 299 | (__v2di) _mm_setzero_si128(), |
| 300 | (__mmask8) -1); |
| 301 | } |
| 302 | |
| 303 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 304 | _mm_mask_cvtpd_epi64 (__m128i __W, __mmask8 __U, __m128d __A) { |
| 305 | return (__m128i) __builtin_ia32_cvtpd2qq128_mask ((__v2df) __A, |
| 306 | (__v2di) __W, |
| 307 | (__mmask8) __U); |
| 308 | } |
| 309 | |
| 310 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 311 | _mm_maskz_cvtpd_epi64 (__mmask8 __U, __m128d __A) { |
| 312 | return (__m128i) __builtin_ia32_cvtpd2qq128_mask ((__v2df) __A, |
| 313 | (__v2di) _mm_setzero_si128(), |
| 314 | (__mmask8) __U); |
| 315 | } |
| 316 | |
| 317 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 318 | _mm256_cvtpd_epi64 (__m256d __A) { |
| 319 | return (__m256i) __builtin_ia32_cvtpd2qq256_mask ((__v4df) __A, |
| 320 | (__v4di) _mm256_setzero_si256(), |
| 321 | (__mmask8) -1); |
| 322 | } |
| 323 | |
| 324 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 325 | _mm256_mask_cvtpd_epi64 (__m256i __W, __mmask8 __U, __m256d __A) { |
| 326 | return (__m256i) __builtin_ia32_cvtpd2qq256_mask ((__v4df) __A, |
| 327 | (__v4di) __W, |
| 328 | (__mmask8) __U); |
| 329 | } |
| 330 | |
| 331 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 332 | _mm256_maskz_cvtpd_epi64 (__mmask8 __U, __m256d __A) { |
| 333 | return (__m256i) __builtin_ia32_cvtpd2qq256_mask ((__v4df) __A, |
| 334 | (__v4di) _mm256_setzero_si256(), |
| 335 | (__mmask8) __U); |
| 336 | } |
| 337 | |
| 338 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 339 | _mm_cvtpd_epu64 (__m128d __A) { |
| 340 | return (__m128i) __builtin_ia32_cvtpd2uqq128_mask ((__v2df) __A, |
| 341 | (__v2di) _mm_setzero_si128(), |
| 342 | (__mmask8) -1); |
| 343 | } |
| 344 | |
| 345 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 346 | _mm_mask_cvtpd_epu64 (__m128i __W, __mmask8 __U, __m128d __A) { |
| 347 | return (__m128i) __builtin_ia32_cvtpd2uqq128_mask ((__v2df) __A, |
| 348 | (__v2di) __W, |
| 349 | (__mmask8) __U); |
| 350 | } |
| 351 | |
| 352 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 353 | _mm_maskz_cvtpd_epu64 (__mmask8 __U, __m128d __A) { |
| 354 | return (__m128i) __builtin_ia32_cvtpd2uqq128_mask ((__v2df) __A, |
| 355 | (__v2di) _mm_setzero_si128(), |
| 356 | (__mmask8) __U); |
| 357 | } |
| 358 | |
| 359 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 360 | _mm256_cvtpd_epu64 (__m256d __A) { |
| 361 | return (__m256i) __builtin_ia32_cvtpd2uqq256_mask ((__v4df) __A, |
| 362 | (__v4di) _mm256_setzero_si256(), |
| 363 | (__mmask8) -1); |
| 364 | } |
| 365 | |
| 366 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 367 | _mm256_mask_cvtpd_epu64 (__m256i __W, __mmask8 __U, __m256d __A) { |
| 368 | return (__m256i) __builtin_ia32_cvtpd2uqq256_mask ((__v4df) __A, |
| 369 | (__v4di) __W, |
| 370 | (__mmask8) __U); |
| 371 | } |
| 372 | |
| 373 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 374 | _mm256_maskz_cvtpd_epu64 (__mmask8 __U, __m256d __A) { |
| 375 | return (__m256i) __builtin_ia32_cvtpd2uqq256_mask ((__v4df) __A, |
| 376 | (__v4di) _mm256_setzero_si256(), |
| 377 | (__mmask8) __U); |
| 378 | } |
| 379 | |
| 380 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 381 | _mm_cvtps_epi64 (__m128 __A) { |
| 382 | return (__m128i) __builtin_ia32_cvtps2qq128_mask ((__v4sf) __A, |
| 383 | (__v2di) _mm_setzero_si128(), |
| 384 | (__mmask8) -1); |
| 385 | } |
| 386 | |
| 387 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 388 | _mm_mask_cvtps_epi64 (__m128i __W, __mmask8 __U, __m128 __A) { |
| 389 | return (__m128i) __builtin_ia32_cvtps2qq128_mask ((__v4sf) __A, |
| 390 | (__v2di) __W, |
| 391 | (__mmask8) __U); |
| 392 | } |
| 393 | |
| 394 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 395 | _mm_maskz_cvtps_epi64 (__mmask8 __U, __m128 __A) { |
| 396 | return (__m128i) __builtin_ia32_cvtps2qq128_mask ((__v4sf) __A, |
| 397 | (__v2di) _mm_setzero_si128(), |
| 398 | (__mmask8) __U); |
| 399 | } |
| 400 | |
| 401 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 402 | _mm256_cvtps_epi64 (__m128 __A) { |
| 403 | return (__m256i) __builtin_ia32_cvtps2qq256_mask ((__v4sf) __A, |
| 404 | (__v4di) _mm256_setzero_si256(), |
| 405 | (__mmask8) -1); |
| 406 | } |
| 407 | |
| 408 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 409 | _mm256_mask_cvtps_epi64 (__m256i __W, __mmask8 __U, __m128 __A) { |
| 410 | return (__m256i) __builtin_ia32_cvtps2qq256_mask ((__v4sf) __A, |
| 411 | (__v4di) __W, |
| 412 | (__mmask8) __U); |
| 413 | } |
| 414 | |
| 415 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 416 | _mm256_maskz_cvtps_epi64 (__mmask8 __U, __m128 __A) { |
| 417 | return (__m256i) __builtin_ia32_cvtps2qq256_mask ((__v4sf) __A, |
| 418 | (__v4di) _mm256_setzero_si256(), |
| 419 | (__mmask8) __U); |
| 420 | } |
| 421 | |
| 422 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 423 | _mm_cvtps_epu64 (__m128 __A) { |
| 424 | return (__m128i) __builtin_ia32_cvtps2uqq128_mask ((__v4sf) __A, |
| 425 | (__v2di) _mm_setzero_si128(), |
| 426 | (__mmask8) -1); |
| 427 | } |
| 428 | |
| 429 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 430 | _mm_mask_cvtps_epu64 (__m128i __W, __mmask8 __U, __m128 __A) { |
| 431 | return (__m128i) __builtin_ia32_cvtps2uqq128_mask ((__v4sf) __A, |
| 432 | (__v2di) __W, |
| 433 | (__mmask8) __U); |
| 434 | } |
| 435 | |
| 436 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 437 | _mm_maskz_cvtps_epu64 (__mmask8 __U, __m128 __A) { |
| 438 | return (__m128i) __builtin_ia32_cvtps2uqq128_mask ((__v4sf) __A, |
| 439 | (__v2di) _mm_setzero_si128(), |
| 440 | (__mmask8) __U); |
| 441 | } |
| 442 | |
| 443 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 444 | _mm256_cvtps_epu64 (__m128 __A) { |
| 445 | return (__m256i) __builtin_ia32_cvtps2uqq256_mask ((__v4sf) __A, |
| 446 | (__v4di) _mm256_setzero_si256(), |
| 447 | (__mmask8) -1); |
| 448 | } |
| 449 | |
| 450 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 451 | _mm256_mask_cvtps_epu64 (__m256i __W, __mmask8 __U, __m128 __A) { |
| 452 | return (__m256i) __builtin_ia32_cvtps2uqq256_mask ((__v4sf) __A, |
| 453 | (__v4di) __W, |
| 454 | (__mmask8) __U); |
| 455 | } |
| 456 | |
| 457 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 458 | _mm256_maskz_cvtps_epu64 (__mmask8 __U, __m128 __A) { |
| 459 | return (__m256i) __builtin_ia32_cvtps2uqq256_mask ((__v4sf) __A, |
| 460 | (__v4di) _mm256_setzero_si256(), |
| 461 | (__mmask8) __U); |
| 462 | } |
| 463 | |
| 464 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 465 | _mm_cvtepi64_pd (__m128i __A) { |
Craig Topper | 842171d | 2018-05-21 20:19:17 +0000 | [diff] [blame] | 466 | return (__m128d)__builtin_convertvector((__v2di)__A, __v2df); |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 467 | } |
| 468 | |
| 469 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 470 | _mm_mask_cvtepi64_pd (__m128d __W, __mmask8 __U, __m128i __A) { |
Craig Topper | 842171d | 2018-05-21 20:19:17 +0000 | [diff] [blame] | 471 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 472 | (__v2df)_mm_cvtepi64_pd(__A), |
| 473 | (__v2df)__W); |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 474 | } |
| 475 | |
| 476 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 477 | _mm_maskz_cvtepi64_pd (__mmask8 __U, __m128i __A) { |
Craig Topper | 842171d | 2018-05-21 20:19:17 +0000 | [diff] [blame] | 478 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 479 | (__v2df)_mm_cvtepi64_pd(__A), |
| 480 | (__v2df)_mm_setzero_pd()); |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 481 | } |
| 482 | |
| 483 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 484 | _mm256_cvtepi64_pd (__m256i __A) { |
Craig Topper | 842171d | 2018-05-21 20:19:17 +0000 | [diff] [blame] | 485 | return (__m256d)__builtin_convertvector((__v4di)__A, __v4df); |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 486 | } |
| 487 | |
| 488 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 489 | _mm256_mask_cvtepi64_pd (__m256d __W, __mmask8 __U, __m256i __A) { |
Craig Topper | 842171d | 2018-05-21 20:19:17 +0000 | [diff] [blame] | 490 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 491 | (__v4df)_mm256_cvtepi64_pd(__A), |
| 492 | (__v4df)__W); |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 493 | } |
| 494 | |
| 495 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 496 | _mm256_maskz_cvtepi64_pd (__mmask8 __U, __m256i __A) { |
Craig Topper | 842171d | 2018-05-21 20:19:17 +0000 | [diff] [blame] | 497 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 498 | (__v4df)_mm256_cvtepi64_pd(__A), |
| 499 | (__v4df)_mm256_setzero_pd()); |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 500 | } |
| 501 | |
| 502 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 503 | _mm_cvtepi64_ps (__m128i __A) { |
| 504 | return (__m128) __builtin_ia32_cvtqq2ps128_mask ((__v2di) __A, |
| 505 | (__v4sf) _mm_setzero_ps(), |
| 506 | (__mmask8) -1); |
| 507 | } |
| 508 | |
| 509 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 510 | _mm_mask_cvtepi64_ps (__m128 __W, __mmask8 __U, __m128i __A) { |
| 511 | return (__m128) __builtin_ia32_cvtqq2ps128_mask ((__v2di) __A, |
| 512 | (__v4sf) __W, |
| 513 | (__mmask8) __U); |
| 514 | } |
| 515 | |
| 516 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 517 | _mm_maskz_cvtepi64_ps (__mmask8 __U, __m128i __A) { |
| 518 | return (__m128) __builtin_ia32_cvtqq2ps128_mask ((__v2di) __A, |
| 519 | (__v4sf) _mm_setzero_ps(), |
| 520 | (__mmask8) __U); |
| 521 | } |
| 522 | |
| 523 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 524 | _mm256_cvtepi64_ps (__m256i __A) { |
| 525 | return (__m128) __builtin_ia32_cvtqq2ps256_mask ((__v4di) __A, |
| 526 | (__v4sf) _mm_setzero_ps(), |
| 527 | (__mmask8) -1); |
| 528 | } |
| 529 | |
| 530 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 531 | _mm256_mask_cvtepi64_ps (__m128 __W, __mmask8 __U, __m256i __A) { |
| 532 | return (__m128) __builtin_ia32_cvtqq2ps256_mask ((__v4di) __A, |
| 533 | (__v4sf) __W, |
| 534 | (__mmask8) __U); |
| 535 | } |
| 536 | |
| 537 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 538 | _mm256_maskz_cvtepi64_ps (__mmask8 __U, __m256i __A) { |
| 539 | return (__m128) __builtin_ia32_cvtqq2ps256_mask ((__v4di) __A, |
| 540 | (__v4sf) _mm_setzero_ps(), |
| 541 | (__mmask8) __U); |
| 542 | } |
| 543 | |
| 544 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 545 | _mm_cvttpd_epi64 (__m128d __A) { |
| 546 | return (__m128i) __builtin_ia32_cvttpd2qq128_mask ((__v2df) __A, |
| 547 | (__v2di) _mm_setzero_si128(), |
| 548 | (__mmask8) -1); |
| 549 | } |
| 550 | |
| 551 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 552 | _mm_mask_cvttpd_epi64 (__m128i __W, __mmask8 __U, __m128d __A) { |
| 553 | return (__m128i) __builtin_ia32_cvttpd2qq128_mask ((__v2df) __A, |
| 554 | (__v2di) __W, |
| 555 | (__mmask8) __U); |
| 556 | } |
| 557 | |
| 558 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 559 | _mm_maskz_cvttpd_epi64 (__mmask8 __U, __m128d __A) { |
| 560 | return (__m128i) __builtin_ia32_cvttpd2qq128_mask ((__v2df) __A, |
| 561 | (__v2di) _mm_setzero_si128(), |
| 562 | (__mmask8) __U); |
| 563 | } |
| 564 | |
| 565 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 566 | _mm256_cvttpd_epi64 (__m256d __A) { |
| 567 | return (__m256i) __builtin_ia32_cvttpd2qq256_mask ((__v4df) __A, |
| 568 | (__v4di) _mm256_setzero_si256(), |
| 569 | (__mmask8) -1); |
| 570 | } |
| 571 | |
| 572 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 573 | _mm256_mask_cvttpd_epi64 (__m256i __W, __mmask8 __U, __m256d __A) { |
| 574 | return (__m256i) __builtin_ia32_cvttpd2qq256_mask ((__v4df) __A, |
| 575 | (__v4di) __W, |
| 576 | (__mmask8) __U); |
| 577 | } |
| 578 | |
| 579 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 580 | _mm256_maskz_cvttpd_epi64 (__mmask8 __U, __m256d __A) { |
| 581 | return (__m256i) __builtin_ia32_cvttpd2qq256_mask ((__v4df) __A, |
| 582 | (__v4di) _mm256_setzero_si256(), |
| 583 | (__mmask8) __U); |
| 584 | } |
| 585 | |
| 586 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 587 | _mm_cvttpd_epu64 (__m128d __A) { |
| 588 | return (__m128i) __builtin_ia32_cvttpd2uqq128_mask ((__v2df) __A, |
| 589 | (__v2di) _mm_setzero_si128(), |
| 590 | (__mmask8) -1); |
| 591 | } |
| 592 | |
| 593 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 594 | _mm_mask_cvttpd_epu64 (__m128i __W, __mmask8 __U, __m128d __A) { |
| 595 | return (__m128i) __builtin_ia32_cvttpd2uqq128_mask ((__v2df) __A, |
| 596 | (__v2di) __W, |
| 597 | (__mmask8) __U); |
| 598 | } |
| 599 | |
| 600 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 601 | _mm_maskz_cvttpd_epu64 (__mmask8 __U, __m128d __A) { |
| 602 | return (__m128i) __builtin_ia32_cvttpd2uqq128_mask ((__v2df) __A, |
| 603 | (__v2di) _mm_setzero_si128(), |
| 604 | (__mmask8) __U); |
| 605 | } |
| 606 | |
| 607 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 608 | _mm256_cvttpd_epu64 (__m256d __A) { |
| 609 | return (__m256i) __builtin_ia32_cvttpd2uqq256_mask ((__v4df) __A, |
| 610 | (__v4di) _mm256_setzero_si256(), |
| 611 | (__mmask8) -1); |
| 612 | } |
| 613 | |
| 614 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 615 | _mm256_mask_cvttpd_epu64 (__m256i __W, __mmask8 __U, __m256d __A) { |
| 616 | return (__m256i) __builtin_ia32_cvttpd2uqq256_mask ((__v4df) __A, |
| 617 | (__v4di) __W, |
| 618 | (__mmask8) __U); |
| 619 | } |
| 620 | |
| 621 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 622 | _mm256_maskz_cvttpd_epu64 (__mmask8 __U, __m256d __A) { |
| 623 | return (__m256i) __builtin_ia32_cvttpd2uqq256_mask ((__v4df) __A, |
| 624 | (__v4di) _mm256_setzero_si256(), |
| 625 | (__mmask8) __U); |
| 626 | } |
| 627 | |
| 628 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 629 | _mm_cvttps_epi64 (__m128 __A) { |
| 630 | return (__m128i) __builtin_ia32_cvttps2qq128_mask ((__v4sf) __A, |
| 631 | (__v2di) _mm_setzero_si128(), |
| 632 | (__mmask8) -1); |
| 633 | } |
| 634 | |
| 635 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 636 | _mm_mask_cvttps_epi64 (__m128i __W, __mmask8 __U, __m128 __A) { |
| 637 | return (__m128i) __builtin_ia32_cvttps2qq128_mask ((__v4sf) __A, |
| 638 | (__v2di) __W, |
| 639 | (__mmask8) __U); |
| 640 | } |
| 641 | |
| 642 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 643 | _mm_maskz_cvttps_epi64 (__mmask8 __U, __m128 __A) { |
| 644 | return (__m128i) __builtin_ia32_cvttps2qq128_mask ((__v4sf) __A, |
| 645 | (__v2di) _mm_setzero_si128(), |
| 646 | (__mmask8) __U); |
| 647 | } |
| 648 | |
| 649 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 650 | _mm256_cvttps_epi64 (__m128 __A) { |
| 651 | return (__m256i) __builtin_ia32_cvttps2qq256_mask ((__v4sf) __A, |
| 652 | (__v4di) _mm256_setzero_si256(), |
| 653 | (__mmask8) -1); |
| 654 | } |
| 655 | |
| 656 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 657 | _mm256_mask_cvttps_epi64 (__m256i __W, __mmask8 __U, __m128 __A) { |
| 658 | return (__m256i) __builtin_ia32_cvttps2qq256_mask ((__v4sf) __A, |
| 659 | (__v4di) __W, |
| 660 | (__mmask8) __U); |
| 661 | } |
| 662 | |
| 663 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 664 | _mm256_maskz_cvttps_epi64 (__mmask8 __U, __m128 __A) { |
| 665 | return (__m256i) __builtin_ia32_cvttps2qq256_mask ((__v4sf) __A, |
| 666 | (__v4di) _mm256_setzero_si256(), |
| 667 | (__mmask8) __U); |
| 668 | } |
| 669 | |
| 670 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 671 | _mm_cvttps_epu64 (__m128 __A) { |
| 672 | return (__m128i) __builtin_ia32_cvttps2uqq128_mask ((__v4sf) __A, |
| 673 | (__v2di) _mm_setzero_si128(), |
| 674 | (__mmask8) -1); |
| 675 | } |
| 676 | |
| 677 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 678 | _mm_mask_cvttps_epu64 (__m128i __W, __mmask8 __U, __m128 __A) { |
| 679 | return (__m128i) __builtin_ia32_cvttps2uqq128_mask ((__v4sf) __A, |
| 680 | (__v2di) __W, |
| 681 | (__mmask8) __U); |
| 682 | } |
| 683 | |
| 684 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 685 | _mm_maskz_cvttps_epu64 (__mmask8 __U, __m128 __A) { |
| 686 | return (__m128i) __builtin_ia32_cvttps2uqq128_mask ((__v4sf) __A, |
| 687 | (__v2di) _mm_setzero_si128(), |
| 688 | (__mmask8) __U); |
| 689 | } |
| 690 | |
| 691 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 692 | _mm256_cvttps_epu64 (__m128 __A) { |
| 693 | return (__m256i) __builtin_ia32_cvttps2uqq256_mask ((__v4sf) __A, |
| 694 | (__v4di) _mm256_setzero_si256(), |
| 695 | (__mmask8) -1); |
| 696 | } |
| 697 | |
| 698 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 699 | _mm256_mask_cvttps_epu64 (__m256i __W, __mmask8 __U, __m128 __A) { |
| 700 | return (__m256i) __builtin_ia32_cvttps2uqq256_mask ((__v4sf) __A, |
| 701 | (__v4di) __W, |
| 702 | (__mmask8) __U); |
| 703 | } |
| 704 | |
| 705 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 706 | _mm256_maskz_cvttps_epu64 (__mmask8 __U, __m128 __A) { |
| 707 | return (__m256i) __builtin_ia32_cvttps2uqq256_mask ((__v4sf) __A, |
| 708 | (__v4di) _mm256_setzero_si256(), |
| 709 | (__mmask8) __U); |
| 710 | } |
| 711 | |
| 712 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 713 | _mm_cvtepu64_pd (__m128i __A) { |
Craig Topper | 842171d | 2018-05-21 20:19:17 +0000 | [diff] [blame] | 714 | return (__m128d)__builtin_convertvector((__v2du)__A, __v2df); |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 715 | } |
| 716 | |
| 717 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 718 | _mm_mask_cvtepu64_pd (__m128d __W, __mmask8 __U, __m128i __A) { |
Craig Topper | 842171d | 2018-05-21 20:19:17 +0000 | [diff] [blame] | 719 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 720 | (__v2df)_mm_cvtepu64_pd(__A), |
| 721 | (__v2df)__W); |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 722 | } |
| 723 | |
| 724 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 725 | _mm_maskz_cvtepu64_pd (__mmask8 __U, __m128i __A) { |
Craig Topper | 842171d | 2018-05-21 20:19:17 +0000 | [diff] [blame] | 726 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 727 | (__v2df)_mm_cvtepu64_pd(__A), |
| 728 | (__v2df)_mm_setzero_pd()); |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 729 | } |
| 730 | |
| 731 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 732 | _mm256_cvtepu64_pd (__m256i __A) { |
Craig Topper | 842171d | 2018-05-21 20:19:17 +0000 | [diff] [blame] | 733 | return (__m256d)__builtin_convertvector((__v4du)__A, __v4df); |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 734 | } |
| 735 | |
| 736 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 737 | _mm256_mask_cvtepu64_pd (__m256d __W, __mmask8 __U, __m256i __A) { |
Craig Topper | 842171d | 2018-05-21 20:19:17 +0000 | [diff] [blame] | 738 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 739 | (__v4df)_mm256_cvtepu64_pd(__A), |
| 740 | (__v4df)__W); |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 741 | } |
| 742 | |
| 743 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 744 | _mm256_maskz_cvtepu64_pd (__mmask8 __U, __m256i __A) { |
Craig Topper | 842171d | 2018-05-21 20:19:17 +0000 | [diff] [blame] | 745 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 746 | (__v4df)_mm256_cvtepu64_pd(__A), |
| 747 | (__v4df)_mm256_setzero_pd()); |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 748 | } |
| 749 | |
| 750 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 751 | _mm_cvtepu64_ps (__m128i __A) { |
| 752 | return (__m128) __builtin_ia32_cvtuqq2ps128_mask ((__v2di) __A, |
| 753 | (__v4sf) _mm_setzero_ps(), |
| 754 | (__mmask8) -1); |
| 755 | } |
| 756 | |
| 757 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 758 | _mm_mask_cvtepu64_ps (__m128 __W, __mmask8 __U, __m128i __A) { |
| 759 | return (__m128) __builtin_ia32_cvtuqq2ps128_mask ((__v2di) __A, |
| 760 | (__v4sf) __W, |
| 761 | (__mmask8) __U); |
| 762 | } |
| 763 | |
| 764 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 765 | _mm_maskz_cvtepu64_ps (__mmask8 __U, __m128i __A) { |
| 766 | return (__m128) __builtin_ia32_cvtuqq2ps128_mask ((__v2di) __A, |
| 767 | (__v4sf) _mm_setzero_ps(), |
| 768 | (__mmask8) __U); |
| 769 | } |
| 770 | |
| 771 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 772 | _mm256_cvtepu64_ps (__m256i __A) { |
| 773 | return (__m128) __builtin_ia32_cvtuqq2ps256_mask ((__v4di) __A, |
| 774 | (__v4sf) _mm_setzero_ps(), |
| 775 | (__mmask8) -1); |
| 776 | } |
| 777 | |
| 778 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 779 | _mm256_mask_cvtepu64_ps (__m128 __W, __mmask8 __U, __m256i __A) { |
| 780 | return (__m128) __builtin_ia32_cvtuqq2ps256_mask ((__v4di) __A, |
| 781 | (__v4sf) __W, |
| 782 | (__mmask8) __U); |
| 783 | } |
| 784 | |
| 785 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 786 | _mm256_maskz_cvtepu64_ps (__mmask8 __U, __m256i __A) { |
| 787 | return (__m128) __builtin_ia32_cvtuqq2ps256_mask ((__v4di) __A, |
| 788 | (__v4sf) _mm_setzero_ps(), |
| 789 | (__mmask8) __U); |
| 790 | } |
| 791 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 792 | #define _mm_range_pd(A, B, C) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 793 | (__m128d)__builtin_ia32_rangepd128_mask((__v2df)(__m128d)(A), \ |
| 794 | (__v2df)(__m128d)(B), (int)(C), \ |
| 795 | (__v2df)_mm_setzero_pd(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 796 | (__mmask8)-1) |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 797 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 798 | #define _mm_mask_range_pd(W, U, A, B, C) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 799 | (__m128d)__builtin_ia32_rangepd128_mask((__v2df)(__m128d)(A), \ |
| 800 | (__v2df)(__m128d)(B), (int)(C), \ |
| 801 | (__v2df)(__m128d)(W), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 802 | (__mmask8)(U)) |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 803 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 804 | #define _mm_maskz_range_pd(U, A, B, C) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 805 | (__m128d)__builtin_ia32_rangepd128_mask((__v2df)(__m128d)(A), \ |
| 806 | (__v2df)(__m128d)(B), (int)(C), \ |
| 807 | (__v2df)_mm_setzero_pd(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 808 | (__mmask8)(U)) |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 809 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 810 | #define _mm256_range_pd(A, B, C) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 811 | (__m256d)__builtin_ia32_rangepd256_mask((__v4df)(__m256d)(A), \ |
| 812 | (__v4df)(__m256d)(B), (int)(C), \ |
| 813 | (__v4df)_mm256_setzero_pd(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 814 | (__mmask8)-1) |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 815 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 816 | #define _mm256_mask_range_pd(W, U, A, B, C) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 817 | (__m256d)__builtin_ia32_rangepd256_mask((__v4df)(__m256d)(A), \ |
| 818 | (__v4df)(__m256d)(B), (int)(C), \ |
| 819 | (__v4df)(__m256d)(W), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 820 | (__mmask8)(U)) |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 821 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 822 | #define _mm256_maskz_range_pd(U, A, B, C) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 823 | (__m256d)__builtin_ia32_rangepd256_mask((__v4df)(__m256d)(A), \ |
| 824 | (__v4df)(__m256d)(B), (int)(C), \ |
| 825 | (__v4df)_mm256_setzero_pd(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 826 | (__mmask8)(U)) |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 827 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 828 | #define _mm_range_ps(A, B, C) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 829 | (__m128)__builtin_ia32_rangeps128_mask((__v4sf)(__m128)(A), \ |
| 830 | (__v4sf)(__m128)(B), (int)(C), \ |
| 831 | (__v4sf)_mm_setzero_ps(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 832 | (__mmask8)-1) |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 833 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 834 | #define _mm_mask_range_ps(W, U, A, B, C) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 835 | (__m128)__builtin_ia32_rangeps128_mask((__v4sf)(__m128)(A), \ |
| 836 | (__v4sf)(__m128)(B), (int)(C), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 837 | (__v4sf)(__m128)(W), (__mmask8)(U)) |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 838 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 839 | #define _mm_maskz_range_ps(U, A, B, C) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 840 | (__m128)__builtin_ia32_rangeps128_mask((__v4sf)(__m128)(A), \ |
| 841 | (__v4sf)(__m128)(B), (int)(C), \ |
| 842 | (__v4sf)_mm_setzero_ps(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 843 | (__mmask8)(U)) |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 844 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 845 | #define _mm256_range_ps(A, B, C) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 846 | (__m256)__builtin_ia32_rangeps256_mask((__v8sf)(__m256)(A), \ |
| 847 | (__v8sf)(__m256)(B), (int)(C), \ |
| 848 | (__v8sf)_mm256_setzero_ps(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 849 | (__mmask8)-1) |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 850 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 851 | #define _mm256_mask_range_ps(W, U, A, B, C) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 852 | (__m256)__builtin_ia32_rangeps256_mask((__v8sf)(__m256)(A), \ |
| 853 | (__v8sf)(__m256)(B), (int)(C), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 854 | (__v8sf)(__m256)(W), (__mmask8)(U)) |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 855 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 856 | #define _mm256_maskz_range_ps(U, A, B, C) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 857 | (__m256)__builtin_ia32_rangeps256_mask((__v8sf)(__m256)(A), \ |
| 858 | (__v8sf)(__m256)(B), (int)(C), \ |
| 859 | (__v8sf)_mm256_setzero_ps(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 860 | (__mmask8)(U)) |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 861 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 862 | #define _mm_reduce_pd(A, B) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 863 | (__m128d)__builtin_ia32_reducepd128_mask((__v2df)(__m128d)(A), (int)(B), \ |
| 864 | (__v2df)_mm_setzero_pd(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 865 | (__mmask8)-1) |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 866 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 867 | #define _mm_mask_reduce_pd(W, U, A, B) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 868 | (__m128d)__builtin_ia32_reducepd128_mask((__v2df)(__m128d)(A), (int)(B), \ |
| 869 | (__v2df)(__m128d)(W), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 870 | (__mmask8)(U)) |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 871 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 872 | #define _mm_maskz_reduce_pd(U, A, B) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 873 | (__m128d)__builtin_ia32_reducepd128_mask((__v2df)(__m128d)(A), (int)(B), \ |
| 874 | (__v2df)_mm_setzero_pd(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 875 | (__mmask8)(U)) |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 876 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 877 | #define _mm256_reduce_pd(A, B) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 878 | (__m256d)__builtin_ia32_reducepd256_mask((__v4df)(__m256d)(A), (int)(B), \ |
| 879 | (__v4df)_mm256_setzero_pd(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 880 | (__mmask8)-1) |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 881 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 882 | #define _mm256_mask_reduce_pd(W, U, A, B) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 883 | (__m256d)__builtin_ia32_reducepd256_mask((__v4df)(__m256d)(A), (int)(B), \ |
| 884 | (__v4df)(__m256d)(W), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 885 | (__mmask8)(U)) |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 886 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 887 | #define _mm256_maskz_reduce_pd(U, A, B) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 888 | (__m256d)__builtin_ia32_reducepd256_mask((__v4df)(__m256d)(A), (int)(B), \ |
| 889 | (__v4df)_mm256_setzero_pd(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 890 | (__mmask8)(U)) |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 891 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 892 | #define _mm_reduce_ps(A, B) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 893 | (__m128)__builtin_ia32_reduceps128_mask((__v4sf)(__m128)(A), (int)(B), \ |
| 894 | (__v4sf)_mm_setzero_ps(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 895 | (__mmask8)-1) |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 896 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 897 | #define _mm_mask_reduce_ps(W, U, A, B) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 898 | (__m128)__builtin_ia32_reduceps128_mask((__v4sf)(__m128)(A), (int)(B), \ |
| 899 | (__v4sf)(__m128)(W), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 900 | (__mmask8)(U)) |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 901 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 902 | #define _mm_maskz_reduce_ps(U, A, B) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 903 | (__m128)__builtin_ia32_reduceps128_mask((__v4sf)(__m128)(A), (int)(B), \ |
| 904 | (__v4sf)_mm_setzero_ps(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 905 | (__mmask8)(U)) |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 906 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 907 | #define _mm256_reduce_ps(A, B) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 908 | (__m256)__builtin_ia32_reduceps256_mask((__v8sf)(__m256)(A), (int)(B), \ |
| 909 | (__v8sf)_mm256_setzero_ps(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 910 | (__mmask8)-1) |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 911 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 912 | #define _mm256_mask_reduce_ps(W, U, A, B) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 913 | (__m256)__builtin_ia32_reduceps256_mask((__v8sf)(__m256)(A), (int)(B), \ |
| 914 | (__v8sf)(__m256)(W), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 915 | (__mmask8)(U)) |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 916 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 917 | #define _mm256_maskz_reduce_ps(U, A, B) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 918 | (__m256)__builtin_ia32_reduceps256_mask((__v8sf)(__m256)(A), (int)(B), \ |
| 919 | (__v8sf)_mm256_setzero_ps(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 920 | (__mmask8)(U)) |
Asaf Badouh | 73b639f | 2015-08-02 12:43:08 +0000 | [diff] [blame] | 921 | |
Michael Zuckerman | 074edd7 | 2016-04-13 10:49:37 +0000 | [diff] [blame] | 922 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 923 | _mm_movepi32_mask (__m128i __A) |
| 924 | { |
| 925 | return (__mmask8) __builtin_ia32_cvtd2mask128 ((__v4si) __A); |
| 926 | } |
| 927 | |
| 928 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 929 | _mm256_movepi32_mask (__m256i __A) |
| 930 | { |
| 931 | return (__mmask8) __builtin_ia32_cvtd2mask256 ((__v8si) __A); |
| 932 | } |
| 933 | |
| 934 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 935 | _mm_movm_epi32 (__mmask8 __A) |
| 936 | { |
| 937 | return (__m128i) __builtin_ia32_cvtmask2d128 (__A); |
| 938 | } |
| 939 | |
| 940 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 941 | _mm256_movm_epi32 (__mmask8 __A) |
| 942 | { |
| 943 | return (__m256i) __builtin_ia32_cvtmask2d256 (__A); |
| 944 | } |
| 945 | |
| 946 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 947 | _mm_movm_epi64 (__mmask8 __A) |
| 948 | { |
| 949 | return (__m128i) __builtin_ia32_cvtmask2q128 (__A); |
| 950 | } |
| 951 | |
| 952 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 953 | _mm256_movm_epi64 (__mmask8 __A) |
| 954 | { |
| 955 | return (__m256i) __builtin_ia32_cvtmask2q256 (__A); |
| 956 | } |
| 957 | |
| 958 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 959 | _mm_movepi64_mask (__m128i __A) |
| 960 | { |
| 961 | return (__mmask8) __builtin_ia32_cvtq2mask128 ((__v2di) __A); |
| 962 | } |
| 963 | |
| 964 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 965 | _mm256_movepi64_mask (__m256i __A) |
| 966 | { |
| 967 | return (__mmask8) __builtin_ia32_cvtq2mask256 ((__v4di) __A); |
| 968 | } |
| 969 | |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 970 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 971 | _mm256_broadcast_f32x2 (__m128 __A) |
| 972 | { |
Craig Topper | 95ed88a | 2018-06-04 19:28:09 +0000 | [diff] [blame] | 973 | return (__m256)__builtin_shufflevector((__v4sf)__A, (__v4sf)__A, |
Craig Topper | 5ece4cf | 2017-08-30 16:15:12 +0000 | [diff] [blame] | 974 | 0, 1, 0, 1, 0, 1, 0, 1); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 975 | } |
| 976 | |
| 977 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 978 | _mm256_mask_broadcast_f32x2 (__m256 __O, __mmask8 __M, __m128 __A) |
| 979 | { |
Craig Topper | 5ece4cf | 2017-08-30 16:15:12 +0000 | [diff] [blame] | 980 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__M, |
| 981 | (__v8sf)_mm256_broadcast_f32x2(__A), |
| 982 | (__v8sf)__O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 983 | } |
| 984 | |
| 985 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 986 | _mm256_maskz_broadcast_f32x2 (__mmask8 __M, __m128 __A) |
| 987 | { |
Craig Topper | 5ece4cf | 2017-08-30 16:15:12 +0000 | [diff] [blame] | 988 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__M, |
| 989 | (__v8sf)_mm256_broadcast_f32x2(__A), |
| 990 | (__v8sf)_mm256_setzero_ps()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 991 | } |
| 992 | |
| 993 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 367c86d | 2017-01-18 02:17:10 +0000 | [diff] [blame] | 994 | _mm256_broadcast_f64x2(__m128d __A) |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 995 | { |
Craig Topper | 367c86d | 2017-01-18 02:17:10 +0000 | [diff] [blame] | 996 | return (__m256d)__builtin_shufflevector((__v2df)__A, (__v2df)__A, |
| 997 | 0, 1, 0, 1); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 998 | } |
| 999 | |
| 1000 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 367c86d | 2017-01-18 02:17:10 +0000 | [diff] [blame] | 1001 | _mm256_mask_broadcast_f64x2(__m256d __O, __mmask8 __M, __m128d __A) |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 1002 | { |
Craig Topper | 367c86d | 2017-01-18 02:17:10 +0000 | [diff] [blame] | 1003 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__M, |
| 1004 | (__v4df)_mm256_broadcast_f64x2(__A), |
| 1005 | (__v4df)__O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 1006 | } |
| 1007 | |
| 1008 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 1009 | _mm256_maskz_broadcast_f64x2 (__mmask8 __M, __m128d __A) |
| 1010 | { |
Craig Topper | 367c86d | 2017-01-18 02:17:10 +0000 | [diff] [blame] | 1011 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__M, |
| 1012 | (__v4df)_mm256_broadcast_f64x2(__A), |
| 1013 | (__v4df)_mm256_setzero_pd()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 1014 | } |
| 1015 | |
| 1016 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1017 | _mm_broadcast_i32x2 (__m128i __A) |
| 1018 | { |
Craig Topper | 95ed88a | 2018-06-04 19:28:09 +0000 | [diff] [blame] | 1019 | return (__m128i)__builtin_shufflevector((__v4si)__A, (__v4si)__A, |
Craig Topper | 5ece4cf | 2017-08-30 16:15:12 +0000 | [diff] [blame] | 1020 | 0, 1, 0, 1); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 1021 | } |
| 1022 | |
| 1023 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1024 | _mm_mask_broadcast_i32x2 (__m128i __O, __mmask8 __M, __m128i __A) |
| 1025 | { |
Craig Topper | 5ece4cf | 2017-08-30 16:15:12 +0000 | [diff] [blame] | 1026 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M, |
| 1027 | (__v4si)_mm_broadcast_i32x2(__A), |
| 1028 | (__v4si)__O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 1029 | } |
| 1030 | |
| 1031 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1032 | _mm_maskz_broadcast_i32x2 (__mmask8 __M, __m128i __A) |
| 1033 | { |
Craig Topper | 5ece4cf | 2017-08-30 16:15:12 +0000 | [diff] [blame] | 1034 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M, |
| 1035 | (__v4si)_mm_broadcast_i32x2(__A), |
| 1036 | (__v4si)_mm_setzero_si128()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 1037 | } |
| 1038 | |
| 1039 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1040 | _mm256_broadcast_i32x2 (__m128i __A) |
| 1041 | { |
Craig Topper | 95ed88a | 2018-06-04 19:28:09 +0000 | [diff] [blame] | 1042 | return (__m256i)__builtin_shufflevector((__v4si)__A, (__v4si)__A, |
Craig Topper | 5ece4cf | 2017-08-30 16:15:12 +0000 | [diff] [blame] | 1043 | 0, 1, 0, 1, 0, 1, 0, 1); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 1044 | } |
| 1045 | |
| 1046 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1047 | _mm256_mask_broadcast_i32x2 (__m256i __O, __mmask8 __M, __m128i __A) |
| 1048 | { |
Craig Topper | 5ece4cf | 2017-08-30 16:15:12 +0000 | [diff] [blame] | 1049 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M, |
| 1050 | (__v8si)_mm256_broadcast_i32x2(__A), |
| 1051 | (__v8si)__O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 1052 | } |
| 1053 | |
| 1054 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1055 | _mm256_maskz_broadcast_i32x2 (__mmask8 __M, __m128i __A) |
| 1056 | { |
Craig Topper | 5ece4cf | 2017-08-30 16:15:12 +0000 | [diff] [blame] | 1057 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M, |
| 1058 | (__v8si)_mm256_broadcast_i32x2(__A), |
| 1059 | (__v8si)_mm256_setzero_si256()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 1060 | } |
| 1061 | |
| 1062 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 367c86d | 2017-01-18 02:17:10 +0000 | [diff] [blame] | 1063 | _mm256_broadcast_i64x2(__m128i __A) |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 1064 | { |
Craig Topper | 367c86d | 2017-01-18 02:17:10 +0000 | [diff] [blame] | 1065 | return (__m256i)__builtin_shufflevector((__v2di)__A, (__v2di)__A, |
| 1066 | 0, 1, 0, 1); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 1067 | } |
| 1068 | |
| 1069 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 367c86d | 2017-01-18 02:17:10 +0000 | [diff] [blame] | 1070 | _mm256_mask_broadcast_i64x2(__m256i __O, __mmask8 __M, __m128i __A) |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 1071 | { |
Craig Topper | 367c86d | 2017-01-18 02:17:10 +0000 | [diff] [blame] | 1072 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, |
| 1073 | (__v4di)_mm256_broadcast_i64x2(__A), |
| 1074 | (__v4di)__O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 1075 | } |
| 1076 | |
| 1077 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1078 | _mm256_maskz_broadcast_i64x2 (__mmask8 __M, __m128i __A) |
| 1079 | { |
Craig Topper | 367c86d | 2017-01-18 02:17:10 +0000 | [diff] [blame] | 1080 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, |
| 1081 | (__v4di)_mm256_broadcast_i64x2(__A), |
| 1082 | (__v4di)_mm256_setzero_si256()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 1083 | } |
| 1084 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1085 | #define _mm256_extractf64x2_pd(A, imm) \ |
Craig Topper | 5f50f338 | 2018-06-08 21:50:07 +0000 | [diff] [blame^] | 1086 | (__m128d)__builtin_ia32_extractf64x2_256_mask((__v4df)(__m256d)(A), \ |
| 1087 | (int)(imm), \ |
| 1088 | (__v2df)_mm_undefined_pd(), \ |
| 1089 | (__mmask8)-1) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 1090 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1091 | #define _mm256_mask_extractf64x2_pd(W, U, A, imm) \ |
Craig Topper | 5f50f338 | 2018-06-08 21:50:07 +0000 | [diff] [blame^] | 1092 | (__m128d)__builtin_ia32_extractf64x2_256_mask((__v4df)(__m256d)(A), \ |
| 1093 | (int)(imm), \ |
| 1094 | (__v2df)(__m128d)(W), \ |
| 1095 | (__mmask8)(U)) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 1096 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1097 | #define _mm256_maskz_extractf64x2_pd(U, A, imm) \ |
Craig Topper | 5f50f338 | 2018-06-08 21:50:07 +0000 | [diff] [blame^] | 1098 | (__m128d)__builtin_ia32_extractf64x2_256_mask((__v4df)(__m256d)(A), \ |
| 1099 | (int)(imm), \ |
| 1100 | (__v2df)_mm_setzero_pd(), \ |
| 1101 | (__mmask8)(U)) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 1102 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1103 | #define _mm256_extracti64x2_epi64(A, imm) \ |
Craig Topper | 5f50f338 | 2018-06-08 21:50:07 +0000 | [diff] [blame^] | 1104 | (__m128i)__builtin_ia32_extracti64x2_256_mask((__v4di)(__m256i)(A), \ |
| 1105 | (int)(imm), \ |
| 1106 | (__v2di)_mm_undefined_si128(), \ |
| 1107 | (__mmask8)-1) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 1108 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1109 | #define _mm256_mask_extracti64x2_epi64(W, U, A, imm) \ |
Craig Topper | 5f50f338 | 2018-06-08 21:50:07 +0000 | [diff] [blame^] | 1110 | (__m128i)__builtin_ia32_extracti64x2_256_mask((__v4di)(__m256i)(A), \ |
| 1111 | (int)(imm), \ |
| 1112 | (__v2di)(__m128i)(W), \ |
| 1113 | (__mmask8)(U)) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 1114 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1115 | #define _mm256_maskz_extracti64x2_epi64(U, A, imm) \ |
Craig Topper | 5f50f338 | 2018-06-08 21:50:07 +0000 | [diff] [blame^] | 1116 | (__m128i)__builtin_ia32_extracti64x2_256_mask((__v4di)(__m256i)(A), \ |
| 1117 | (int)(imm), \ |
| 1118 | (__v2di)_mm_setzero_si128(), \ |
| 1119 | (__mmask8)(U)) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 1120 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1121 | #define _mm256_insertf64x2(A, B, imm) \ |
Craig Topper | 3428bee | 2018-06-08 03:24:47 +0000 | [diff] [blame] | 1122 | (__m256d)__builtin_ia32_insertf64x2_256((__v4df)(__m256d)(A), \ |
| 1123 | (__v2df)(__m128d)(B), (int)(imm)) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 1124 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1125 | #define _mm256_mask_insertf64x2(W, U, A, B, imm) \ |
Craig Topper | 08bf53f | 2016-11-01 05:47:56 +0000 | [diff] [blame] | 1126 | (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \ |
| 1127 | (__v4df)_mm256_insertf64x2((A), (B), (imm)), \ |
Craig Topper | cbf3929 | 2018-05-31 01:24:40 +0000 | [diff] [blame] | 1128 | (__v4df)(__m256d)(W)) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 1129 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1130 | #define _mm256_maskz_insertf64x2(U, A, B, imm) \ |
Craig Topper | 08bf53f | 2016-11-01 05:47:56 +0000 | [diff] [blame] | 1131 | (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \ |
| 1132 | (__v4df)_mm256_insertf64x2((A), (B), (imm)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1133 | (__v4df)_mm256_setzero_pd()) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 1134 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1135 | #define _mm256_inserti64x2(A, B, imm) \ |
Craig Topper | 3428bee | 2018-06-08 03:24:47 +0000 | [diff] [blame] | 1136 | (__m256i)__builtin_ia32_inserti64x2_256((__v4di)(__m256i)(A), \ |
| 1137 | (__v2di)(__m128i)(B), (int)(imm)) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 1138 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1139 | #define _mm256_mask_inserti64x2(W, U, A, B, imm) \ |
Craig Topper | 08bf53f | 2016-11-01 05:47:56 +0000 | [diff] [blame] | 1140 | (__m256i)__builtin_ia32_selectq_256((__mmask8)(U), \ |
| 1141 | (__v4di)_mm256_inserti64x2((A), (B), (imm)), \ |
Craig Topper | cbf3929 | 2018-05-31 01:24:40 +0000 | [diff] [blame] | 1142 | (__v4di)(__m256i)(W)) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 1143 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1144 | #define _mm256_maskz_inserti64x2(U, A, B, imm) \ |
Craig Topper | 08bf53f | 2016-11-01 05:47:56 +0000 | [diff] [blame] | 1145 | (__m256i)__builtin_ia32_selectq_256((__mmask8)(U), \ |
| 1146 | (__v4di)_mm256_inserti64x2((A), (B), (imm)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1147 | (__v4di)_mm256_setzero_si256()) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 1148 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1149 | #define _mm_mask_fpclass_pd_mask(U, A, imm) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 1150 | (__mmask8)__builtin_ia32_fpclasspd128_mask((__v2df)(__m128d)(A), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1151 | (__mmask8)(U)) |
Michael Zuckerman | edc82fe | 2016-04-25 14:48:23 +0000 | [diff] [blame] | 1152 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1153 | #define _mm_fpclass_pd_mask(A, imm) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 1154 | (__mmask8)__builtin_ia32_fpclasspd128_mask((__v2df)(__m128d)(A), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1155 | (__mmask8)-1) |
Michael Zuckerman | edc82fe | 2016-04-25 14:48:23 +0000 | [diff] [blame] | 1156 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1157 | #define _mm256_mask_fpclass_pd_mask(U, A, imm) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 1158 | (__mmask8)__builtin_ia32_fpclasspd256_mask((__v4df)(__m256d)(A), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1159 | (__mmask8)(U)) |
Michael Zuckerman | edc82fe | 2016-04-25 14:48:23 +0000 | [diff] [blame] | 1160 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1161 | #define _mm256_fpclass_pd_mask(A, imm) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 1162 | (__mmask8)__builtin_ia32_fpclasspd256_mask((__v4df)(__m256d)(A), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1163 | (__mmask8)-1) |
Michael Zuckerman | edc82fe | 2016-04-25 14:48:23 +0000 | [diff] [blame] | 1164 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1165 | #define _mm_mask_fpclass_ps_mask(U, A, imm) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 1166 | (__mmask8)__builtin_ia32_fpclassps128_mask((__v4sf)(__m128)(A), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1167 | (__mmask8)(U)) |
Michael Zuckerman | edc82fe | 2016-04-25 14:48:23 +0000 | [diff] [blame] | 1168 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1169 | #define _mm_fpclass_ps_mask(A, imm) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 1170 | (__mmask8)__builtin_ia32_fpclassps128_mask((__v4sf)(__m128)(A), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1171 | (__mmask8)-1) |
Michael Zuckerman | edc82fe | 2016-04-25 14:48:23 +0000 | [diff] [blame] | 1172 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1173 | #define _mm256_mask_fpclass_ps_mask(U, A, imm) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 1174 | (__mmask8)__builtin_ia32_fpclassps256_mask((__v8sf)(__m256)(A), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1175 | (__mmask8)(U)) |
Michael Zuckerman | edc82fe | 2016-04-25 14:48:23 +0000 | [diff] [blame] | 1176 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1177 | #define _mm256_fpclass_ps_mask(A, imm) \ |
Craig Topper | f2e67a0 | 2016-05-17 04:41:46 +0000 | [diff] [blame] | 1178 | (__mmask8)__builtin_ia32_fpclassps256_mask((__v8sf)(__m256)(A), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 1179 | (__mmask8)-1) |
Michael Zuckerman | edc82fe | 2016-04-25 14:48:23 +0000 | [diff] [blame] | 1180 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1181 | #undef __DEFAULT_FN_ATTRS |
Eric Christopher | 4d185168 | 2015-06-17 07:09:20 +0000 | [diff] [blame] | 1182 | |
Elena Demikhovsky | e7d4c2e | 2015-04-30 09:24:29 +0000 | [diff] [blame] | 1183 | #endif |