Details | Last modification | View Log | RSS feed
| Rev | Author | Line No. | Line |
|---|---|---|---|
| 14 | pmbaty | 1 | /*===---- immintrin.h - Intel intrinsics -----------------------------------=== |
| 2 | * |
||
| 3 | * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
||
| 4 | * See https://llvm.org/LICENSE.txt for license information. |
||
| 5 | * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
||
| 6 | * |
||
| 7 | *===-----------------------------------------------------------------------=== |
||
| 8 | */ |
||
| 9 | |||
| 10 | #ifndef __IMMINTRIN_H |
||
| 11 | #define __IMMINTRIN_H |
||
| 12 | |||
| 13 | #if !defined(__i386__) && !defined(__x86_64__) |
||
| 14 | #error "This header is only meant to be used on x86 and x64 architecture" |
||
| 15 | #endif |
||
| 16 | |||
| 17 | #include <x86gprintrin.h> |
||
| 18 | |||
| 19 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 20 | defined(__MMX__) |
||
| 21 | #include <mmintrin.h> |
||
| 22 | #endif |
||
| 23 | |||
| 24 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 25 | defined(__SSE__) |
||
| 26 | #include <xmmintrin.h> |
||
| 27 | #endif |
||
| 28 | |||
| 29 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 30 | defined(__SSE2__) |
||
| 31 | #include <emmintrin.h> |
||
| 32 | #endif |
||
| 33 | |||
| 34 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 35 | defined(__SSE3__) |
||
| 36 | #include <pmmintrin.h> |
||
| 37 | #endif |
||
| 38 | |||
| 39 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 40 | defined(__SSSE3__) |
||
| 41 | #include <tmmintrin.h> |
||
| 42 | #endif |
||
| 43 | |||
| 44 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 45 | (defined(__SSE4_2__) || defined(__SSE4_1__)) |
||
| 46 | #include <smmintrin.h> |
||
| 47 | #endif |
||
| 48 | |||
| 49 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 50 | (defined(__AES__) || defined(__PCLMUL__)) |
||
| 51 | #include <wmmintrin.h> |
||
| 52 | #endif |
||
| 53 | |||
| 54 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 55 | defined(__CLFLUSHOPT__) |
||
| 56 | #include <clflushoptintrin.h> |
||
| 57 | #endif |
||
| 58 | |||
| 59 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 60 | defined(__CLWB__) |
||
| 61 | #include <clwbintrin.h> |
||
| 62 | #endif |
||
| 63 | |||
| 64 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 65 | defined(__AVX__) |
||
| 66 | #include <avxintrin.h> |
||
| 67 | #endif |
||
| 68 | |||
| 69 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 70 | defined(__AVX2__) |
||
| 71 | #include <avx2intrin.h> |
||
| 72 | #endif |
||
| 73 | |||
| 74 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 75 | defined(__F16C__) |
||
| 76 | #include <f16cintrin.h> |
||
| 77 | #endif |
||
| 78 | |||
| 79 | /* No feature check desired due to internal checks */ |
||
| 80 | #include <bmiintrin.h> |
||
| 81 | |||
| 82 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 83 | defined(__BMI2__) |
||
| 84 | #include <bmi2intrin.h> |
||
| 85 | #endif |
||
| 86 | |||
| 87 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 88 | defined(__LZCNT__) |
||
| 89 | #include <lzcntintrin.h> |
||
| 90 | #endif |
||
| 91 | |||
| 92 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 93 | defined(__POPCNT__) |
||
| 94 | #include <popcntintrin.h> |
||
| 95 | #endif |
||
| 96 | |||
| 97 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 98 | defined(__FMA__) |
||
| 99 | #include <fmaintrin.h> |
||
| 100 | #endif |
||
| 101 | |||
| 102 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 103 | defined(__AVX512F__) |
||
| 104 | #include <avx512fintrin.h> |
||
| 105 | #endif |
||
| 106 | |||
| 107 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 108 | defined(__AVX512VL__) |
||
| 109 | #include <avx512vlintrin.h> |
||
| 110 | #endif |
||
| 111 | |||
| 112 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 113 | defined(__AVX512BW__) |
||
| 114 | #include <avx512bwintrin.h> |
||
| 115 | #endif |
||
| 116 | |||
| 117 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 118 | defined(__AVX512BITALG__) |
||
| 119 | #include <avx512bitalgintrin.h> |
||
| 120 | #endif |
||
| 121 | |||
| 122 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 123 | defined(__AVX512CD__) |
||
| 124 | #include <avx512cdintrin.h> |
||
| 125 | #endif |
||
| 126 | |||
| 127 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 128 | defined(__AVX512VPOPCNTDQ__) |
||
| 129 | #include <avx512vpopcntdqintrin.h> |
||
| 130 | #endif |
||
| 131 | |||
| 132 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 133 | (defined(__AVX512VL__) && defined(__AVX512VPOPCNTDQ__)) |
||
| 134 | #include <avx512vpopcntdqvlintrin.h> |
||
| 135 | #endif |
||
| 136 | |||
| 137 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 138 | defined(__AVX512VNNI__) |
||
| 139 | #include <avx512vnniintrin.h> |
||
| 140 | #endif |
||
| 141 | |||
| 142 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 143 | (defined(__AVX512VL__) && defined(__AVX512VNNI__)) |
||
| 144 | #include <avx512vlvnniintrin.h> |
||
| 145 | #endif |
||
| 146 | |||
| 147 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 148 | defined(__AVXVNNI__) |
||
| 149 | #include <avxvnniintrin.h> |
||
| 150 | #endif |
||
| 151 | |||
| 152 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 153 | defined(__AVX512DQ__) |
||
| 154 | #include <avx512dqintrin.h> |
||
| 155 | #endif |
||
| 156 | |||
| 157 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 158 | (defined(__AVX512VL__) && defined(__AVX512BITALG__)) |
||
| 159 | #include <avx512vlbitalgintrin.h> |
||
| 160 | #endif |
||
| 161 | |||
| 162 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 163 | (defined(__AVX512VL__) && defined(__AVX512BW__)) |
||
| 164 | #include <avx512vlbwintrin.h> |
||
| 165 | #endif |
||
| 166 | |||
| 167 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 168 | (defined(__AVX512VL__) && defined(__AVX512CD__)) |
||
| 169 | #include <avx512vlcdintrin.h> |
||
| 170 | #endif |
||
| 171 | |||
| 172 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 173 | (defined(__AVX512VL__) && defined(__AVX512DQ__)) |
||
| 174 | #include <avx512vldqintrin.h> |
||
| 175 | #endif |
||
| 176 | |||
| 177 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 178 | defined(__AVX512ER__) |
||
| 179 | #include <avx512erintrin.h> |
||
| 180 | #endif |
||
| 181 | |||
| 182 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 183 | defined(__AVX512IFMA__) |
||
| 184 | #include <avx512ifmaintrin.h> |
||
| 185 | #endif |
||
| 186 | |||
| 187 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 188 | (defined(__AVX512IFMA__) && defined(__AVX512VL__)) |
||
| 189 | #include <avx512ifmavlintrin.h> |
||
| 190 | #endif |
||
| 191 | |||
| 192 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 193 | defined(__AVXIFMA__) |
||
| 194 | #include <avxifmaintrin.h> |
||
| 195 | #endif |
||
| 196 | |||
| 197 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 198 | defined(__AVX512VBMI__) |
||
| 199 | #include <avx512vbmiintrin.h> |
||
| 200 | #endif |
||
| 201 | |||
| 202 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 203 | (defined(__AVX512VBMI__) && defined(__AVX512VL__)) |
||
| 204 | #include <avx512vbmivlintrin.h> |
||
| 205 | #endif |
||
| 206 | |||
| 207 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 208 | defined(__AVX512VBMI2__) |
||
| 209 | #include <avx512vbmi2intrin.h> |
||
| 210 | #endif |
||
| 211 | |||
| 212 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 213 | (defined(__AVX512VBMI2__) && defined(__AVX512VL__)) |
||
| 214 | #include <avx512vlvbmi2intrin.h> |
||
| 215 | #endif |
||
| 216 | |||
| 217 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 218 | defined(__AVX512PF__) |
||
| 219 | #include <avx512pfintrin.h> |
||
| 220 | #endif |
||
| 221 | |||
| 222 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 223 | defined(__AVX512FP16__) |
||
| 224 | #include <avx512fp16intrin.h> |
||
| 225 | #endif |
||
| 226 | |||
| 227 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 228 | (defined(__AVX512VL__) && defined(__AVX512FP16__)) |
||
| 229 | #include <avx512vlfp16intrin.h> |
||
| 230 | #endif |
||
| 231 | |||
| 232 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 233 | defined(__AVX512BF16__) |
||
| 234 | #include <avx512bf16intrin.h> |
||
| 235 | #endif |
||
| 236 | |||
| 237 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 238 | (defined(__AVX512VL__) && defined(__AVX512BF16__)) |
||
| 239 | #include <avx512vlbf16intrin.h> |
||
| 240 | #endif |
||
| 241 | |||
| 242 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 243 | defined(__PKU__) |
||
| 244 | #include <pkuintrin.h> |
||
| 245 | #endif |
||
| 246 | |||
| 247 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 248 | defined(__VPCLMULQDQ__) |
||
| 249 | #include <vpclmulqdqintrin.h> |
||
| 250 | #endif |
||
| 251 | |||
| 252 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 253 | defined(__VAES__) |
||
| 254 | #include <vaesintrin.h> |
||
| 255 | #endif |
||
| 256 | |||
| 257 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 258 | defined(__GFNI__) |
||
| 259 | #include <gfniintrin.h> |
||
| 260 | #endif |
||
| 261 | |||
| 262 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 263 | defined(__AVXVNNIINT8__) |
||
| 264 | #include <avxvnniint8intrin.h> |
||
| 265 | #endif |
||
| 266 | |||
| 267 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 268 | defined(__AVXNECONVERT__) |
||
| 269 | #include <avxneconvertintrin.h> |
||
| 270 | #endif |
||
| 271 | |||
| 272 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 273 | defined(__RDPID__) |
||
| 274 | /// Returns the value of the IA32_TSC_AUX MSR (0xc0000103). |
||
| 275 | /// |
||
| 276 | /// \headerfile <immintrin.h> |
||
| 277 | /// |
||
| 278 | /// This intrinsic corresponds to the <c> RDPID </c> instruction. |
||
| 279 | static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__, __target__("rdpid"))) |
||
| 280 | _rdpid_u32(void) { |
||
| 281 | return __builtin_ia32_rdpid(); |
||
| 282 | } |
||
| 283 | #endif // __RDPID__ |
||
| 284 | |||
| 285 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 286 | defined(__RDRND__) |
||
| 287 | static __inline__ int __attribute__((__always_inline__, __nodebug__, __target__("rdrnd"))) |
||
| 288 | _rdrand16_step(unsigned short *__p) |
||
| 289 | { |
||
| 290 | return (int)__builtin_ia32_rdrand16_step(__p); |
||
| 291 | } |
||
| 292 | |||
| 293 | static __inline__ int __attribute__((__always_inline__, __nodebug__, __target__("rdrnd"))) |
||
| 294 | _rdrand32_step(unsigned int *__p) |
||
| 295 | { |
||
| 296 | return (int)__builtin_ia32_rdrand32_step(__p); |
||
| 297 | } |
||
| 298 | |||
| 299 | #ifdef __x86_64__ |
||
| 300 | static __inline__ int __attribute__((__always_inline__, __nodebug__, __target__("rdrnd"))) |
||
| 301 | _rdrand64_step(unsigned long long *__p) |
||
| 302 | { |
||
| 303 | return (int)__builtin_ia32_rdrand64_step(__p); |
||
| 304 | } |
||
| 305 | #else |
||
| 306 | // We need to emulate the functionality of 64-bit rdrand with 2 32-bit |
||
| 307 | // rdrand instructions. |
||
| 308 | static __inline__ int __attribute__((__always_inline__, __nodebug__, __target__("rdrnd"))) |
||
| 309 | _rdrand64_step(unsigned long long *__p) |
||
| 310 | { |
||
| 311 | unsigned int __lo, __hi; |
||
| 312 | unsigned int __res_lo = __builtin_ia32_rdrand32_step(&__lo); |
||
| 313 | unsigned int __res_hi = __builtin_ia32_rdrand32_step(&__hi); |
||
| 314 | if (__res_lo && __res_hi) { |
||
| 315 | *__p = ((unsigned long long)__hi << 32) | (unsigned long long)__lo; |
||
| 316 | return 1; |
||
| 317 | } else { |
||
| 318 | *__p = 0; |
||
| 319 | return 0; |
||
| 320 | } |
||
| 321 | } |
||
| 322 | #endif |
||
| 323 | #endif /* __RDRND__ */ |
||
| 324 | |||
| 325 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 326 | defined(__FSGSBASE__) |
||
| 327 | #ifdef __x86_64__ |
||
| 328 | static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase"))) |
||
| 329 | _readfsbase_u32(void) |
||
| 330 | { |
||
| 331 | return __builtin_ia32_rdfsbase32(); |
||
| 332 | } |
||
| 333 | |||
| 334 | static __inline__ unsigned long long __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase"))) |
||
| 335 | _readfsbase_u64(void) |
||
| 336 | { |
||
| 337 | return __builtin_ia32_rdfsbase64(); |
||
| 338 | } |
||
| 339 | |||
| 340 | static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase"))) |
||
| 341 | _readgsbase_u32(void) |
||
| 342 | { |
||
| 343 | return __builtin_ia32_rdgsbase32(); |
||
| 344 | } |
||
| 345 | |||
| 346 | static __inline__ unsigned long long __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase"))) |
||
| 347 | _readgsbase_u64(void) |
||
| 348 | { |
||
| 349 | return __builtin_ia32_rdgsbase64(); |
||
| 350 | } |
||
| 351 | |||
| 352 | static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase"))) |
||
| 353 | _writefsbase_u32(unsigned int __V) |
||
| 354 | { |
||
| 355 | __builtin_ia32_wrfsbase32(__V); |
||
| 356 | } |
||
| 357 | |||
| 358 | static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase"))) |
||
| 359 | _writefsbase_u64(unsigned long long __V) |
||
| 360 | { |
||
| 361 | __builtin_ia32_wrfsbase64(__V); |
||
| 362 | } |
||
| 363 | |||
| 364 | static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase"))) |
||
| 365 | _writegsbase_u32(unsigned int __V) |
||
| 366 | { |
||
| 367 | __builtin_ia32_wrgsbase32(__V); |
||
| 368 | } |
||
| 369 | |||
| 370 | static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase"))) |
||
| 371 | _writegsbase_u64(unsigned long long __V) |
||
| 372 | { |
||
| 373 | __builtin_ia32_wrgsbase64(__V); |
||
| 374 | } |
||
| 375 | |||
| 376 | #endif |
||
| 377 | #endif /* __FSGSBASE__ */ |
||
| 378 | |||
| 379 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 380 | defined(__MOVBE__) |
||
| 381 | |||
| 382 | /* The structs used below are to force the load/store to be unaligned. This |
||
| 383 | * is accomplished with the __packed__ attribute. The __may_alias__ prevents |
||
| 384 | * tbaa metadata from being generated based on the struct and the type of the |
||
| 385 | * field inside of it. |
||
| 386 | */ |
||
| 387 | |||
| 388 | static __inline__ short __attribute__((__always_inline__, __nodebug__, __target__("movbe"))) |
||
| 389 | _loadbe_i16(void const * __P) { |
||
| 390 | struct __loadu_i16 { |
||
| 391 | unsigned short __v; |
||
| 392 | } __attribute__((__packed__, __may_alias__)); |
||
| 393 | return (short)__builtin_bswap16(((const struct __loadu_i16*)__P)->__v); |
||
| 394 | } |
||
| 395 | |||
| 396 | static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("movbe"))) |
||
| 397 | _storebe_i16(void * __P, short __D) { |
||
| 398 | struct __storeu_i16 { |
||
| 399 | unsigned short __v; |
||
| 400 | } __attribute__((__packed__, __may_alias__)); |
||
| 401 | ((struct __storeu_i16*)__P)->__v = __builtin_bswap16((unsigned short)__D); |
||
| 402 | } |
||
| 403 | |||
| 404 | static __inline__ int __attribute__((__always_inline__, __nodebug__, __target__("movbe"))) |
||
| 405 | _loadbe_i32(void const * __P) { |
||
| 406 | struct __loadu_i32 { |
||
| 407 | unsigned int __v; |
||
| 408 | } __attribute__((__packed__, __may_alias__)); |
||
| 409 | return (int)__builtin_bswap32(((const struct __loadu_i32*)__P)->__v); |
||
| 410 | } |
||
| 411 | |||
| 412 | static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("movbe"))) |
||
| 413 | _storebe_i32(void * __P, int __D) { |
||
| 414 | struct __storeu_i32 { |
||
| 415 | unsigned int __v; |
||
| 416 | } __attribute__((__packed__, __may_alias__)); |
||
| 417 | ((struct __storeu_i32*)__P)->__v = __builtin_bswap32((unsigned int)__D); |
||
| 418 | } |
||
| 419 | |||
| 420 | #ifdef __x86_64__ |
||
| 421 | static __inline__ long long __attribute__((__always_inline__, __nodebug__, __target__("movbe"))) |
||
| 422 | _loadbe_i64(void const * __P) { |
||
| 423 | struct __loadu_i64 { |
||
| 424 | unsigned long long __v; |
||
| 425 | } __attribute__((__packed__, __may_alias__)); |
||
| 426 | return (long long)__builtin_bswap64(((const struct __loadu_i64*)__P)->__v); |
||
| 427 | } |
||
| 428 | |||
| 429 | static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("movbe"))) |
||
| 430 | _storebe_i64(void * __P, long long __D) { |
||
| 431 | struct __storeu_i64 { |
||
| 432 | unsigned long long __v; |
||
| 433 | } __attribute__((__packed__, __may_alias__)); |
||
| 434 | ((struct __storeu_i64*)__P)->__v = __builtin_bswap64((unsigned long long)__D); |
||
| 435 | } |
||
| 436 | #endif |
||
| 437 | #endif /* __MOVBE */ |
||
| 438 | |||
| 439 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 440 | defined(__RTM__) |
||
| 441 | #include <rtmintrin.h> |
||
| 442 | #include <xtestintrin.h> |
||
| 443 | #endif |
||
| 444 | |||
| 445 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 446 | defined(__SHA__) |
||
| 447 | #include <shaintrin.h> |
||
| 448 | #endif |
||
| 449 | |||
| 450 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 451 | defined(__FXSR__) |
||
| 452 | #include <fxsrintrin.h> |
||
| 453 | #endif |
||
| 454 | |||
| 455 | /* No feature check desired due to internal MSC_VER checks */ |
||
| 456 | #include <xsaveintrin.h> |
||
| 457 | |||
| 458 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 459 | defined(__XSAVEOPT__) |
||
| 460 | #include <xsaveoptintrin.h> |
||
| 461 | #endif |
||
| 462 | |||
| 463 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 464 | defined(__XSAVEC__) |
||
| 465 | #include <xsavecintrin.h> |
||
| 466 | #endif |
||
| 467 | |||
| 468 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 469 | defined(__XSAVES__) |
||
| 470 | #include <xsavesintrin.h> |
||
| 471 | #endif |
||
| 472 | |||
| 473 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 474 | defined(__SHSTK__) |
||
| 475 | #include <cetintrin.h> |
||
| 476 | #endif |
||
| 477 | |||
| 478 | /* Some intrinsics inside adxintrin.h are available only on processors with ADX, |
||
| 479 | * whereas others are also available at all times. */ |
||
| 480 | #include <adxintrin.h> |
||
| 481 | |||
| 482 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 483 | defined(__RDSEED__) |
||
| 484 | #include <rdseedintrin.h> |
||
| 485 | #endif |
||
| 486 | |||
| 487 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 488 | defined(__WBNOINVD__) |
||
| 489 | #include <wbnoinvdintrin.h> |
||
| 490 | #endif |
||
| 491 | |||
| 492 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 493 | defined(__CLDEMOTE__) |
||
| 494 | #include <cldemoteintrin.h> |
||
| 495 | #endif |
||
| 496 | |||
| 497 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 498 | defined(__WAITPKG__) |
||
| 499 | #include <waitpkgintrin.h> |
||
| 500 | #endif |
||
| 501 | |||
| 502 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 503 | defined(__MOVDIRI__) || defined(__MOVDIR64B__) |
||
| 504 | #include <movdirintrin.h> |
||
| 505 | #endif |
||
| 506 | |||
| 507 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 508 | defined(__PCONFIG__) |
||
| 509 | #include <pconfigintrin.h> |
||
| 510 | #endif |
||
| 511 | |||
| 512 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 513 | defined(__SGX__) |
||
| 514 | #include <sgxintrin.h> |
||
| 515 | #endif |
||
| 516 | |||
| 517 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 518 | defined(__PTWRITE__) |
||
| 519 | #include <ptwriteintrin.h> |
||
| 520 | #endif |
||
| 521 | |||
| 522 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 523 | defined(__INVPCID__) |
||
| 524 | #include <invpcidintrin.h> |
||
| 525 | #endif |
||
| 526 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 527 | defined(__AMX_FP16__) |
||
| 528 | #include <amxfp16intrin.h> |
||
| 529 | #endif |
||
| 530 | |||
| 531 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 532 | defined(__KL__) || defined(__WIDEKL__) |
||
| 533 | #include <keylockerintrin.h> |
||
| 534 | #endif |
||
| 535 | |||
| 536 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 537 | defined(__AMX_TILE__) || defined(__AMX_INT8__) || defined(__AMX_BF16__) |
||
| 538 | #include <amxintrin.h> |
||
| 539 | #endif |
||
| 540 | |||
| 541 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 542 | defined(__AVX512VP2INTERSECT__) |
||
| 543 | #include <avx512vp2intersectintrin.h> |
||
| 544 | #endif |
||
| 545 | |||
| 546 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 547 | (defined(__AVX512VL__) && defined(__AVX512VP2INTERSECT__)) |
||
| 548 | #include <avx512vlvp2intersectintrin.h> |
||
| 549 | #endif |
||
| 550 | |||
| 551 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 552 | defined(__ENQCMD__) |
||
| 553 | #include <enqcmdintrin.h> |
||
| 554 | #endif |
||
| 555 | |||
| 556 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 557 | defined(__SERIALIZE__) |
||
| 558 | #include <serializeintrin.h> |
||
| 559 | #endif |
||
| 560 | |||
| 561 | #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) || \ |
||
| 562 | defined(__TSXLDTRK__) |
||
| 563 | #include <tsxldtrkintrin.h> |
||
| 564 | #endif |
||
| 565 | |||
| 566 | #if defined(_MSC_VER) && __has_extension(gnu_asm) |
||
| 567 | /* Define the default attributes for these intrinsics */ |
||
| 568 | #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__)) |
||
| 569 | #ifdef __cplusplus |
||
| 570 | extern "C" { |
||
| 571 | #endif |
||
| 572 | /*----------------------------------------------------------------------------*\ |
||
| 573 | |* Interlocked Exchange HLE |
||
| 574 | \*----------------------------------------------------------------------------*/ |
||
| 575 | #if defined(__i386__) || defined(__x86_64__) |
||
| 576 | static __inline__ long __DEFAULT_FN_ATTRS |
||
| 577 | _InterlockedExchange_HLEAcquire(long volatile *_Target, long _Value) { |
||
| 578 | __asm__ __volatile__(".byte 0xf2 ; lock ; xchg {%0, %1|%1, %0}" |
||
| 579 | : "+r" (_Value), "+m" (*_Target) :: "memory"); |
||
| 580 | return _Value; |
||
| 581 | } |
||
| 582 | static __inline__ long __DEFAULT_FN_ATTRS |
||
| 583 | _InterlockedExchange_HLERelease(long volatile *_Target, long _Value) { |
||
| 584 | __asm__ __volatile__(".byte 0xf3 ; lock ; xchg {%0, %1|%1, %0}" |
||
| 585 | : "+r" (_Value), "+m" (*_Target) :: "memory"); |
||
| 586 | return _Value; |
||
| 587 | } |
||
| 588 | #endif |
||
| 589 | #if defined(__x86_64__) |
||
| 590 | static __inline__ __int64 __DEFAULT_FN_ATTRS |
||
| 591 | _InterlockedExchange64_HLEAcquire(__int64 volatile *_Target, __int64 _Value) { |
||
| 592 | __asm__ __volatile__(".byte 0xf2 ; lock ; xchg {%0, %1|%1, %0}" |
||
| 593 | : "+r" (_Value), "+m" (*_Target) :: "memory"); |
||
| 594 | return _Value; |
||
| 595 | } |
||
| 596 | static __inline__ __int64 __DEFAULT_FN_ATTRS |
||
| 597 | _InterlockedExchange64_HLERelease(__int64 volatile *_Target, __int64 _Value) { |
||
| 598 | __asm__ __volatile__(".byte 0xf3 ; lock ; xchg {%0, %1|%1, %0}" |
||
| 599 | : "+r" (_Value), "+m" (*_Target) :: "memory"); |
||
| 600 | return _Value; |
||
| 601 | } |
||
| 602 | #endif |
||
| 603 | /*----------------------------------------------------------------------------*\ |
||
| 604 | |* Interlocked Compare Exchange HLE |
||
| 605 | \*----------------------------------------------------------------------------*/ |
||
| 606 | #if defined(__i386__) || defined(__x86_64__) |
||
| 607 | static __inline__ long __DEFAULT_FN_ATTRS |
||
| 608 | _InterlockedCompareExchange_HLEAcquire(long volatile *_Destination, |
||
| 609 | long _Exchange, long _Comparand) { |
||
| 610 | __asm__ __volatile__(".byte 0xf2 ; lock ; cmpxchg {%2, %1|%1, %2}" |
||
| 611 | : "+a" (_Comparand), "+m" (*_Destination) |
||
| 612 | : "r" (_Exchange) : "memory"); |
||
| 613 | return _Comparand; |
||
| 614 | } |
||
| 615 | static __inline__ long __DEFAULT_FN_ATTRS |
||
| 616 | _InterlockedCompareExchange_HLERelease(long volatile *_Destination, |
||
| 617 | long _Exchange, long _Comparand) { |
||
| 618 | __asm__ __volatile__(".byte 0xf3 ; lock ; cmpxchg {%2, %1|%1, %2}" |
||
| 619 | : "+a" (_Comparand), "+m" (*_Destination) |
||
| 620 | : "r" (_Exchange) : "memory"); |
||
| 621 | return _Comparand; |
||
| 622 | } |
||
| 623 | #endif |
||
| 624 | #if defined(__x86_64__) |
||
| 625 | static __inline__ __int64 __DEFAULT_FN_ATTRS |
||
| 626 | _InterlockedCompareExchange64_HLEAcquire(__int64 volatile *_Destination, |
||
| 627 | __int64 _Exchange, __int64 _Comparand) { |
||
| 628 | __asm__ __volatile__(".byte 0xf2 ; lock ; cmpxchg {%2, %1|%1, %2}" |
||
| 629 | : "+a" (_Comparand), "+m" (*_Destination) |
||
| 630 | : "r" (_Exchange) : "memory"); |
||
| 631 | return _Comparand; |
||
| 632 | } |
||
| 633 | static __inline__ __int64 __DEFAULT_FN_ATTRS |
||
| 634 | _InterlockedCompareExchange64_HLERelease(__int64 volatile *_Destination, |
||
| 635 | __int64 _Exchange, __int64 _Comparand) { |
||
| 636 | __asm__ __volatile__(".byte 0xf3 ; lock ; cmpxchg {%2, %1|%1, %2}" |
||
| 637 | : "+a" (_Comparand), "+m" (*_Destination) |
||
| 638 | : "r" (_Exchange) : "memory"); |
||
| 639 | return _Comparand; |
||
| 640 | } |
||
| 641 | #endif |
||
| 642 | #ifdef __cplusplus |
||
| 643 | } |
||
| 644 | #endif |
||
| 645 | |||
| 646 | #undef __DEFAULT_FN_ATTRS |
||
| 647 | |||
| 648 | #endif /* defined(_MSC_VER) && __has_extension(gnu_asm) */ |
||
| 649 | |||
| 650 | #endif /* __IMMINTRIN_H */ |