1 /* ===-------- Intrin.h ---------------------------------------------------=== 2 * 3 * Permission is hereby granted, free of charge, to any person obtaining a copy 4 * of this software and associated documentation files (the "Software"), to deal 5 * in the Software without restriction, including without limitation the rights 6 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 7 * copies of the Software, and to permit persons to whom the Software is 8 * furnished to do so, subject to the following conditions: 9 * 10 * The above copyright notice and this permission notice shall be included in 11 * all copies or substantial portions of the Software. 12 * 13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 15 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 16 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 17 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 18 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN 19 * THE SOFTWARE. 20 * 21 *===-----------------------------------------------------------------------=== 22 */ 23 24 /* Only include this if we're compiling for the windows platform. */ 25 #ifndef _MSC_VER 26 #include_next <Intrin.h> 27 #else 28 29 #ifndef __INTRIN_H 30 #define __INTRIN_H 31 32 /* First include the standard intrinsics. */ 33 #if defined(__i386__) || defined(__x86_64__) 34 #include <x86intrin.h> 35 #endif 36 37 /* For the definition of jmp_buf. */ 38 #if __STDC_HOSTED__ 39 #include <setjmp.h> 40 #endif 41 42 #ifdef __cplusplus 43 extern "C" { 44 #endif 45 46 #if defined(__MMX__) 47 /* And the random ones that aren't in those files. */ 48 __m64 _m_from_float(float); 49 __m64 _m_from_int(int _l); 50 void _m_prefetch(void *); 51 float _m_to_float(__m64); 52 int _m_to_int(__m64 _M); 53 #endif 54 55 /* Other assorted instruction intrinsics. */ 56 void __addfsbyte(unsigned long, unsigned char); 57 void __addfsdword(unsigned long, unsigned long); 58 void __addfsword(unsigned long, unsigned short); 59 void __code_seg(const char *); 60 static __inline__ 61 void __cpuid(int[4], int); 62 static __inline__ 63 void __cpuidex(int[4], int, int); 64 void __debugbreak(void); 65 __int64 __emul(int, int); 66 unsigned __int64 __emulu(unsigned int, unsigned int); 67 void __cdecl __fastfail(unsigned int); 68 unsigned int __getcallerseflags(void); 69 static __inline__ 70 void __halt(void); 71 unsigned char __inbyte(unsigned short); 72 void __inbytestring(unsigned short, unsigned char *, unsigned long); 73 void __incfsbyte(unsigned long); 74 void __incfsdword(unsigned long); 75 void __incfsword(unsigned long); 76 unsigned long __indword(unsigned short); 77 void __indwordstring(unsigned short, unsigned long *, unsigned long); 78 void __int2c(void); 79 void __invlpg(void *); 80 unsigned short __inword(unsigned short); 81 void __inwordstring(unsigned short, unsigned short *, unsigned long); 82 void __lidt(void *); 83 unsigned __int64 __ll_lshift(unsigned __int64, int); 84 __int64 __ll_rshift(__int64, int); 85 void __llwpcb(void *); 86 unsigned char __lwpins32(unsigned int, unsigned int, unsigned int); 87 void __lwpval32(unsigned int, unsigned int, unsigned int); 88 unsigned int __lzcnt(unsigned int); 89 unsigned short __lzcnt16(unsigned short); 90 static __inline__ 91 void __movsb(unsigned char *, unsigned char const *, size_t); 92 static __inline__ 93 void __movsd(unsigned long *, unsigned long const *, size_t); 94 static __inline__ 95 void __movsw(unsigned short *, unsigned short const *, size_t); 96 void __nop(void); 97 void __nvreg_restore_fence(void); 98 void __nvreg_save_fence(void); 99 void __outbyte(unsigned short, unsigned char); 100 void __outbytestring(unsigned short, unsigned char *, unsigned long); 101 void __outdword(unsigned short, unsigned long); 102 void __outdwordstring(unsigned short, unsigned long *, unsigned long); 103 void __outword(unsigned short, unsigned short); 104 void __outwordstring(unsigned short, unsigned short *, unsigned long); 105 static __inline__ 106 unsigned int __popcnt(unsigned int); 107 static __inline__ 108 unsigned short __popcnt16(unsigned short); 109 unsigned long __readcr0(void); 110 unsigned long __readcr2(void); 111 static __inline__ 112 unsigned long __readcr3(void); 113 unsigned long __readcr4(void); 114 unsigned long __readcr8(void); 115 unsigned int __readdr(unsigned int); 116 #ifdef __i386__ 117 static __inline__ 118 unsigned char __readfsbyte(unsigned long); 119 static __inline__ 120 unsigned long __readfsdword(unsigned long); 121 static __inline__ 122 unsigned __int64 __readfsqword(unsigned long); 123 static __inline__ 124 unsigned short __readfsword(unsigned long); 125 #endif 126 static __inline__ 127 unsigned __int64 __readmsr(unsigned long); 128 unsigned __int64 __readpmc(unsigned long); 129 unsigned long __segmentlimit(unsigned long); 130 void __sidt(void *); 131 void *__slwpcb(void); 132 static __inline__ 133 void __stosb(unsigned char *, unsigned char, size_t); 134 static __inline__ 135 void __stosd(unsigned long *, unsigned long, size_t); 136 static __inline__ 137 void __stosw(unsigned short *, unsigned short, size_t); 138 void __svm_clgi(void); 139 void __svm_invlpga(void *, int); 140 void __svm_skinit(int); 141 void __svm_stgi(void); 142 void __svm_vmload(size_t); 143 void __svm_vmrun(size_t); 144 void __svm_vmsave(size_t); 145 void __ud2(void); 146 unsigned __int64 __ull_rshift(unsigned __int64, int); 147 void __vmx_off(void); 148 void __vmx_vmptrst(unsigned __int64 *); 149 void __wbinvd(void); 150 void __writecr0(unsigned int); 151 static __inline__ 152 void __writecr3(unsigned int); 153 void __writecr4(unsigned int); 154 void __writecr8(unsigned int); 155 void __writedr(unsigned int, unsigned int); 156 void __writefsbyte(unsigned long, unsigned char); 157 void __writefsdword(unsigned long, unsigned long); 158 void __writefsqword(unsigned long, unsigned __int64); 159 void __writefsword(unsigned long, unsigned short); 160 void __writemsr(unsigned long, unsigned __int64); 161 static __inline__ 162 void *_AddressOfReturnAddress(void); 163 static __inline__ 164 unsigned char _BitScanForward(unsigned long *_Index, unsigned long _Mask); 165 static __inline__ 166 unsigned char _BitScanReverse(unsigned long *_Index, unsigned long _Mask); 167 static __inline__ 168 unsigned char _bittest(long const *, long); 169 static __inline__ 170 unsigned char _bittestandcomplement(long *, long); 171 static __inline__ 172 unsigned char _bittestandreset(long *, long); 173 static __inline__ 174 unsigned char _bittestandset(long *, long); 175 unsigned __int64 __cdecl _byteswap_uint64(unsigned __int64); 176 unsigned long __cdecl _byteswap_ulong(unsigned long); 177 unsigned short __cdecl _byteswap_ushort(unsigned short); 178 void __cdecl _disable(void); 179 void __cdecl _enable(void); 180 void __cdecl _fxrstor(void const *); 181 void __cdecl _fxsave(void *); 182 long _InterlockedAddLargeStatistic(__int64 volatile *_Addend, long _Value); 183 static __inline__ 184 long _InterlockedAnd(long volatile *_Value, long _Mask); 185 static __inline__ 186 short _InterlockedAnd16(short volatile *_Value, short _Mask); 187 static __inline__ 188 char _InterlockedAnd8(char volatile *_Value, char _Mask); 189 unsigned char _interlockedbittestandreset(long volatile *, long); 190 static __inline__ 191 unsigned char _interlockedbittestandset(long volatile *, long); 192 static __inline__ 193 long __cdecl _InterlockedCompareExchange(long volatile *_Destination, 194 long _Exchange, long _Comparand); 195 long _InterlockedCompareExchange_HLEAcquire(long volatile *, long, long); 196 long _InterlockedCompareExchange_HLERelease(long volatile *, long, long); 197 static __inline__ 198 short _InterlockedCompareExchange16(short volatile *_Destination, 199 short _Exchange, short _Comparand); 200 static __inline__ 201 __int64 _InterlockedCompareExchange64(__int64 volatile *_Destination, 202 __int64 _Exchange, __int64 _Comparand); 203 __int64 _InterlockedcompareExchange64_HLEAcquire(__int64 volatile *, __int64, 204 __int64); 205 __int64 _InterlockedCompareExchange64_HLERelease(__int64 volatile *, __int64, 206 __int64); 207 static __inline__ 208 char _InterlockedCompareExchange8(char volatile *_Destination, char _Exchange, 209 char _Comparand); 210 void *_InterlockedCompareExchangePointer_HLEAcquire(void *volatile *, void *, 211 void *); 212 void *_InterlockedCompareExchangePointer_HLERelease(void *volatile *, void *, 213 void *); 214 static __inline__ 215 long __cdecl _InterlockedDecrement(long volatile *_Addend); 216 static __inline__ 217 short _InterlockedDecrement16(short volatile *_Addend); 218 long _InterlockedExchange(long volatile *_Target, long _Value); 219 static __inline__ 220 short _InterlockedExchange16(short volatile *_Target, short _Value); 221 static __inline__ 222 char _InterlockedExchange8(char volatile *_Target, char _Value); 223 static __inline__ 224 long __cdecl _InterlockedExchangeAdd(long volatile *_Addend, long _Value); 225 long _InterlockedExchangeAdd_HLEAcquire(long volatile *, long); 226 long _InterlockedExchangeAdd_HLERelease(long volatile *, long); 227 static __inline__ 228 short _InterlockedExchangeAdd16(short volatile *_Addend, short _Value); 229 __int64 _InterlockedExchangeAdd64_HLEAcquire(__int64 volatile *, __int64); 230 __int64 _InterlockedExchangeAdd64_HLERelease(__int64 volatile *, __int64); 231 static __inline__ 232 char _InterlockedExchangeAdd8(char volatile *_Addend, char _Value); 233 static __inline__ 234 long __cdecl _InterlockedIncrement(long volatile *_Addend); 235 static __inline__ 236 short _InterlockedIncrement16(short volatile *_Addend); 237 static __inline__ 238 long _InterlockedOr(long volatile *_Value, long _Mask); 239 static __inline__ 240 short _InterlockedOr16(short volatile *_Value, short _Mask); 241 static __inline__ 242 char _InterlockedOr8(char volatile *_Value, char _Mask); 243 static __inline__ 244 long _InterlockedXor(long volatile *_Value, long _Mask); 245 static __inline__ 246 short _InterlockedXor16(short volatile *_Value, short _Mask); 247 static __inline__ 248 char _InterlockedXor8(char volatile *_Value, char _Mask); 249 void __cdecl _invpcid(unsigned int, void *); 250 static __inline__ 251 unsigned long __cdecl _lrotl(unsigned long, int); 252 static __inline__ 253 unsigned long __cdecl _lrotr(unsigned long, int); 254 static __inline__ 255 static __inline__ 256 void _ReadBarrier(void); 257 static __inline__ 258 void _ReadWriteBarrier(void); 259 static __inline__ 260 void *_ReturnAddress(void); 261 unsigned int _rorx_u32(unsigned int, const unsigned int); 262 static __inline__ 263 unsigned int __cdecl _rotl(unsigned int _Value, int _Shift); 264 static __inline__ 265 unsigned short _rotl16(unsigned short _Value, unsigned char _Shift); 266 static __inline__ 267 unsigned __int64 __cdecl _rotl64(unsigned __int64 _Value, int _Shift); 268 static __inline__ 269 unsigned char _rotl8(unsigned char _Value, unsigned char _Shift); 270 static __inline__ 271 unsigned int __cdecl _rotr(unsigned int _Value, int _Shift); 272 static __inline__ 273 unsigned short _rotr16(unsigned short _Value, unsigned char _Shift); 274 static __inline__ 275 unsigned __int64 __cdecl _rotr64(unsigned __int64 _Value, int _Shift); 276 static __inline__ 277 unsigned char _rotr8(unsigned char _Value, unsigned char _Shift); 278 int _sarx_i32(int, unsigned int); 279 #if __STDC_HOSTED__ 280 int __cdecl _setjmp(jmp_buf); 281 #endif 282 unsigned int _shlx_u32(unsigned int, unsigned int); 283 unsigned int _shrx_u32(unsigned int, unsigned int); 284 void _Store_HLERelease(long volatile *, long); 285 void _Store64_HLERelease(__int64 volatile *, __int64); 286 void _StorePointer_HLERelease(void *volatile *, void *); 287 static __inline__ 288 void _WriteBarrier(void); 289 unsigned __int32 xbegin(void); 290 void _xend(void); 291 static __inline__ 292 #define _XCR_XFEATURE_ENABLED_MASK 0 293 unsigned __int64 __cdecl _xgetbv(unsigned int); 294 void __cdecl _xrstor(void const *, unsigned __int64); 295 void __cdecl _xsave(void *, unsigned __int64); 296 void __cdecl _xsaveopt(void *, unsigned __int64); 297 void __cdecl _xsetbv(unsigned int, unsigned __int64); 298 299 /* These additional intrinsics are turned on in x64/amd64/x86_64 mode. */ 300 #ifdef __x86_64__ 301 void __addgsbyte(unsigned long, unsigned char); 302 void __addgsdword(unsigned long, unsigned long); 303 void __addgsqword(unsigned long, unsigned __int64); 304 void __addgsword(unsigned long, unsigned short); 305 static __inline__ 306 void __faststorefence(void); 307 void __incgsbyte(unsigned long); 308 void __incgsdword(unsigned long); 309 void __incgsqword(unsigned long); 310 void __incgsword(unsigned long); 311 unsigned char __lwpins64(unsigned __int64, unsigned int, unsigned int); 312 void __lwpval64(unsigned __int64, unsigned int, unsigned int); 313 unsigned __int64 __lzcnt64(unsigned __int64); 314 static __inline__ 315 void __movsq(unsigned long long *, unsigned long long const *, size_t); 316 __int64 __mulh(__int64, __int64); 317 static __inline__ 318 unsigned __int64 __popcnt64(unsigned __int64); 319 static __inline__ 320 unsigned char __readgsbyte(unsigned long); 321 static __inline__ 322 unsigned long __readgsdword(unsigned long); 323 static __inline__ 324 unsigned __int64 __readgsqword(unsigned long); 325 unsigned short __readgsword(unsigned long); 326 unsigned __int64 __shiftleft128(unsigned __int64 _LowPart, 327 unsigned __int64 _HighPart, 328 unsigned char _Shift); 329 unsigned __int64 __shiftright128(unsigned __int64 _LowPart, 330 unsigned __int64 _HighPart, 331 unsigned char _Shift); 332 static __inline__ 333 void __stosq(unsigned __int64 *, unsigned __int64, size_t); 334 unsigned char __vmx_on(unsigned __int64 *); 335 unsigned char __vmx_vmclear(unsigned __int64 *); 336 unsigned char __vmx_vmlaunch(void); 337 unsigned char __vmx_vmptrld(unsigned __int64 *); 338 unsigned char __vmx_vmread(size_t, size_t *); 339 unsigned char __vmx_vmresume(void); 340 unsigned char __vmx_vmwrite(size_t, size_t); 341 void __writegsbyte(unsigned long, unsigned char); 342 void __writegsdword(unsigned long, unsigned long); 343 void __writegsqword(unsigned long, unsigned __int64); 344 void __writegsword(unsigned long, unsigned short); 345 static __inline__ 346 unsigned char _BitScanForward64(unsigned long *_Index, unsigned __int64 _Mask); 347 static __inline__ 348 unsigned char _BitScanReverse64(unsigned long *_Index, unsigned __int64 _Mask); 349 static __inline__ 350 unsigned char _bittest64(__int64 const *, __int64); 351 static __inline__ 352 unsigned char _bittestandcomplement64(__int64 *, __int64); 353 static __inline__ 354 unsigned char _bittestandreset64(__int64 *, __int64); 355 static __inline__ 356 unsigned char _bittestandset64(__int64 *, __int64); 357 unsigned __int64 __cdecl _byteswap_uint64(unsigned __int64); 358 void __cdecl _fxrstor64(void const *); 359 void __cdecl _fxsave64(void *); 360 long _InterlockedAnd_np(long volatile *_Value, long _Mask); 361 short _InterlockedAnd16_np(short volatile *_Value, short _Mask); 362 __int64 _InterlockedAnd64_np(__int64 volatile *_Value, __int64 _Mask); 363 char _InterlockedAnd8_np(char volatile *_Value, char _Mask); 364 unsigned char _interlockedbittestandreset64(__int64 volatile *, __int64); 365 static __inline__ 366 unsigned char _interlockedbittestandset64(__int64 volatile *, __int64); 367 long _InterlockedCompareExchange_np(long volatile *_Destination, long _Exchange, 368 long _Comparand); 369 unsigned char _InterlockedCompareExchange128(__int64 volatile *_Destination, 370 __int64 _ExchangeHigh, 371 __int64 _ExchangeLow, 372 __int64 *_CompareandResult); 373 unsigned char _InterlockedCompareExchange128_np(__int64 volatile *_Destination, 374 __int64 _ExchangeHigh, 375 __int64 _ExchangeLow, 376 __int64 *_ComparandResult); 377 short _InterlockedCompareExchange16_np(short volatile *_Destination, 378 short _Exchange, short _Comparand); 379 __int64 _InterlockedCompareExchange64_HLEAcquire(__int64 volatile *, __int64, 380 __int64); 381 __int64 _InterlockedCompareExchange64_HLERelease(__int64 volatile *, __int64, 382 __int64); 383 __int64 _InterlockedCompareExchange64_np(__int64 volatile *_Destination, 384 __int64 _Exchange, __int64 _Comparand); 385 void *_InterlockedCompareExchangePointer(void *volatile *_Destination, 386 void *_Exchange, void *_Comparand); 387 void *_InterlockedCompareExchangePointer_np(void *volatile *_Destination, 388 void *_Exchange, void *_Comparand); 389 static __inline__ 390 __int64 _InterlockedDecrement64(__int64 volatile *_Addend); 391 static __inline__ 392 __int64 _InterlockedExchange64(__int64 volatile *_Target, __int64 _Value); 393 static __inline__ 394 __int64 _InterlockedExchangeAdd64(__int64 volatile *_Addend, __int64 _Value); 395 void *_InterlockedExchangePointer(void *volatile *_Target, void *_Value); 396 static __inline__ 397 __int64 _InterlockedIncrement64(__int64 volatile *_Addend); 398 long _InterlockedOr_np(long volatile *_Value, long _Mask); 399 short _InterlockedOr16_np(short volatile *_Value, short _Mask); 400 static __inline__ 401 __int64 _InterlockedOr64(__int64 volatile *_Value, __int64 _Mask); 402 __int64 _InterlockedOr64_np(__int64 volatile *_Value, __int64 _Mask); 403 char _InterlockedOr8_np(char volatile *_Value, char _Mask); 404 long _InterlockedXor_np(long volatile *_Value, long _Mask); 405 short _InterlockedXor16_np(short volatile *_Value, short _Mask); 406 static __inline__ 407 __int64 _InterlockedXor64(__int64 volatile *_Value, __int64 _Mask); 408 __int64 _InterlockedXor64_np(__int64 volatile *_Value, __int64 _Mask); 409 char _InterlockedXor8_np(char volatile *_Value, char _Mask); 410 static __inline__ 411 __int64 _mul128(__int64 _Multiplier, __int64 _Multiplicand, 412 __int64 *_HighProduct); 413 unsigned __int64 _rorx_u64(unsigned __int64, const unsigned int); 414 __int64 _sarx_i64(__int64, unsigned int); 415 #if __STDC_HOSTED__ 416 int __cdecl _setjmpex(jmp_buf); 417 #endif 418 unsigned __int64 _shlx_u64(unsigned __int64, unsigned int); 419 unsigned __int64 _shrx_u64(unsigned __int64, unsigned int); 420 /* 421 * Multiply two 64-bit integers and obtain a 64-bit result. 422 * The low-half is returned directly and the high half is in an out parameter. 423 */ 424 static __inline__ unsigned __int64 __attribute__((__always_inline__, __nodebug__)) 425 _umul128(unsigned __int64 _Multiplier, unsigned __int64 _Multiplicand, 426 unsigned __int64 *_HighProduct) { 427 unsigned __int128 _FullProduct = 428 (unsigned __int128)_Multiplier * (unsigned __int128)_Multiplicand; 429 *_HighProduct = _FullProduct >> 64; 430 return _FullProduct; 431 } 432 static __inline__ unsigned __int64 __attribute__((__always_inline__, __nodebug__)) 433 __umulh(unsigned __int64 _Multiplier, unsigned __int64 _Multiplicand) { 434 unsigned __int128 _FullProduct = 435 (unsigned __int128)_Multiplier * (unsigned __int128)_Multiplicand; 436 return _FullProduct >> 64; 437 } 438 void __cdecl _xrstor64(void const *, unsigned __int64); 439 void __cdecl _xsave64(void *, unsigned __int64); 440 void __cdecl _xsaveopt64(void *, unsigned __int64); 441 442 #endif /* __x86_64__ */ 443 444 /*----------------------------------------------------------------------------*\ 445 |* Bit Twiddling 446 \*----------------------------------------------------------------------------*/ 447 static __inline__ unsigned char __attribute__((__always_inline__, __nodebug__)) 448 _rotl8(unsigned char _Value, unsigned char _Shift) { 449 _Shift &= 0x7; 450 return _Shift ? (_Value << _Shift) | (_Value >> (8 - _Shift)) : _Value; 451 } 452 static __inline__ unsigned char __attribute__((__always_inline__, __nodebug__)) 453 _rotr8(unsigned char _Value, unsigned char _Shift) { 454 _Shift &= 0x7; 455 return _Shift ? (_Value >> _Shift) | (_Value << (8 - _Shift)) : _Value; 456 } 457 static __inline__ unsigned short __attribute__((__always_inline__, __nodebug__)) 458 _rotl16(unsigned short _Value, unsigned char _Shift) { 459 _Shift &= 0xf; 460 return _Shift ? (_Value << _Shift) | (_Value >> (16 - _Shift)) : _Value; 461 } 462 static __inline__ unsigned short __attribute__((__always_inline__, __nodebug__)) 463 _rotr16(unsigned short _Value, unsigned char _Shift) { 464 _Shift &= 0xf; 465 return _Shift ? (_Value >> _Shift) | (_Value << (16 - _Shift)) : _Value; 466 } 467 static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__)) 468 _rotl(unsigned int _Value, int _Shift) { 469 _Shift &= 0x1f; 470 return _Shift ? (_Value << _Shift) | (_Value >> (32 - _Shift)) : _Value; 471 } 472 static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__)) 473 _rotr(unsigned int _Value, int _Shift) { 474 _Shift &= 0x1f; 475 return _Shift ? (_Value >> _Shift) | (_Value << (32 - _Shift)) : _Value; 476 } 477 static __inline__ unsigned long __attribute__((__always_inline__, __nodebug__)) 478 _lrotl(unsigned long _Value, int _Shift) { 479 _Shift &= 0x1f; 480 return _Shift ? (_Value << _Shift) | (_Value >> (32 - _Shift)) : _Value; 481 } 482 static __inline__ unsigned long __attribute__((__always_inline__, __nodebug__)) 483 _lrotr(unsigned long _Value, int _Shift) { 484 _Shift &= 0x1f; 485 return _Shift ? (_Value >> _Shift) | (_Value << (32 - _Shift)) : _Value; 486 } 487 static 488 __inline__ unsigned __int64 __attribute__((__always_inline__, __nodebug__)) 489 _rotl64(unsigned __int64 _Value, int _Shift) { 490 _Shift &= 0x3f; 491 return _Shift ? (_Value << _Shift) | (_Value >> (64 - _Shift)) : _Value; 492 } 493 static 494 __inline__ unsigned __int64 __attribute__((__always_inline__, __nodebug__)) 495 _rotr64(unsigned __int64 _Value, int _Shift) { 496 _Shift &= 0x3f; 497 return _Shift ? (_Value >> _Shift) | (_Value << (64 - _Shift)) : _Value; 498 } 499 /*----------------------------------------------------------------------------*\ 500 |* Bit Counting and Testing 501 \*----------------------------------------------------------------------------*/ 502 static __inline__ unsigned char __attribute__((__always_inline__, __nodebug__)) 503 _BitScanForward(unsigned long *_Index, unsigned long _Mask) { 504 if (!_Mask) 505 return 0; 506 *_Index = __builtin_ctzl(_Mask); 507 return 1; 508 } 509 static __inline__ unsigned char __attribute__((__always_inline__, __nodebug__)) 510 _BitScanReverse(unsigned long *_Index, unsigned long _Mask) { 511 if (!_Mask) 512 return 0; 513 *_Index = 31 - __builtin_clzl(_Mask); 514 return 1; 515 } 516 static __inline__ unsigned short __attribute__((__always_inline__, __nodebug__)) 517 __popcnt16(unsigned short value) { 518 return __builtin_popcount((int)value); 519 } 520 static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__)) 521 __popcnt(unsigned int value) { 522 return __builtin_popcount(value); 523 } 524 static __inline__ unsigned char __attribute__((__always_inline__, __nodebug__)) 525 _bittest(long const *a, long b) { 526 return (*a >> b) & 1; 527 } 528 static __inline__ unsigned char __attribute__((__always_inline__, __nodebug__)) 529 _bittestandcomplement(long *a, long b) { 530 unsigned char x = (*a >> b) & 1; 531 *a = *a ^ (1 << b); 532 return x; 533 } 534 static __inline__ unsigned char __attribute__((__always_inline__, __nodebug__)) 535 _bittestandreset(long *a, long b) { 536 unsigned char x = (*a >> b) & 1; 537 *a = *a & ~(1 << b); 538 return x; 539 } 540 static __inline__ unsigned char __attribute__((__always_inline__, __nodebug__)) 541 _bittestandset(long *a, long b) { 542 unsigned char x = (*a >> b) & 1; 543 *a = *a | (1 << b); 544 return x; 545 } 546 #if defined(__i386__) || defined(__x86_64__) 547 static __inline__ unsigned char __attribute__((__always_inline__, __nodebug__)) 548 _interlockedbittestandset(long volatile *__BitBase, long __BitPos) { 549 unsigned char __Res; 550 __asm__ ("xor %0, %0\n" 551 "lock bts %2, %1\n" 552 "setc %0\n" 553 : "=r" (__Res), "+m"(*__BitBase) 554 : "Ir"(__BitPos)); 555 return __Res; 556 } 557 #endif 558 #ifdef __x86_64__ 559 static __inline__ unsigned char __attribute__((__always_inline__, __nodebug__)) 560 _BitScanForward64(unsigned long *_Index, unsigned __int64 _Mask) { 561 if (!_Mask) 562 return 0; 563 *_Index = __builtin_ctzll(_Mask); 564 return 1; 565 } 566 static __inline__ unsigned char __attribute__((__always_inline__, __nodebug__)) 567 _BitScanReverse64(unsigned long *_Index, unsigned __int64 _Mask) { 568 if (!_Mask) 569 return 0; 570 *_Index = 63 - __builtin_clzll(_Mask); 571 return 1; 572 } 573 static __inline__ 574 unsigned __int64 __attribute__((__always_inline__, __nodebug__)) 575 __popcnt64(unsigned __int64 value) { 576 return __builtin_popcountll(value); 577 } 578 static __inline__ unsigned char __attribute__((__always_inline__, __nodebug__)) 579 _bittest64(__int64 const *a, __int64 b) { 580 return (*a >> b) & 1; 581 } 582 static __inline__ unsigned char __attribute__((__always_inline__, __nodebug__)) 583 _bittestandcomplement64(__int64 *a, __int64 b) { 584 unsigned char x = (*a >> b) & 1; 585 *a = *a ^ (1ll << b); 586 return x; 587 } 588 static __inline__ unsigned char __attribute__((__always_inline__, __nodebug__)) 589 _bittestandreset64(__int64 *a, __int64 b) { 590 unsigned char x = (*a >> b) & 1; 591 *a = *a & ~(1ll << b); 592 return x; 593 } 594 static __inline__ unsigned char __attribute__((__always_inline__, __nodebug__)) 595 _bittestandset64(__int64 *a, __int64 b) { 596 unsigned char x = (*a >> b) & 1; 597 *a = *a | (1ll << b); 598 return x; 599 } 600 static __inline__ unsigned char __attribute__((__always_inline__, __nodebug__)) 601 _interlockedbittestandset64(__int64 volatile *__BitBase, __int64 __BitPos) { 602 unsigned char __Res; 603 __asm__ ("xor %0, %0\n" 604 "lock bts %2, %1\n" 605 "setc %0\n" 606 : "=r" (__Res), "+m"(*__BitBase) 607 : "Ir"(__BitPos)); 608 return __Res; 609 } 610 #endif 611 /*----------------------------------------------------------------------------*\ 612 |* Interlocked Exchange Add 613 \*----------------------------------------------------------------------------*/ 614 static __inline__ char __attribute__((__always_inline__, __nodebug__)) 615 _InterlockedExchangeAdd8(char volatile *_Addend, char _Value) { 616 return __atomic_add_fetch(_Addend, _Value, 0) - _Value; 617 } 618 static __inline__ short __attribute__((__always_inline__, __nodebug__)) 619 _InterlockedExchangeAdd16(short volatile *_Addend, short _Value) { 620 return __atomic_add_fetch(_Addend, _Value, 0) - _Value; 621 } 622 #ifdef __x86_64__ 623 static __inline__ __int64 __attribute__((__always_inline__, __nodebug__)) 624 _InterlockedExchangeAdd64(__int64 volatile *_Addend, __int64 _Value) { 625 return __atomic_add_fetch(_Addend, _Value, 0) - _Value; 626 } 627 #endif 628 /*----------------------------------------------------------------------------*\ 629 |* Interlocked Exchange Sub 630 \*----------------------------------------------------------------------------*/ 631 static __inline__ char __attribute__((__always_inline__, __nodebug__)) 632 _InterlockedExchangeSub8(char volatile *_Subend, char _Value) { 633 return __atomic_sub_fetch(_Subend, _Value, 0) + _Value; 634 } 635 static __inline__ short __attribute__((__always_inline__, __nodebug__)) 636 _InterlockedExchangeSub16(short volatile *_Subend, short _Value) { 637 return __atomic_sub_fetch(_Subend, _Value, 0) + _Value; 638 } 639 static __inline__ long __attribute__((__always_inline__, __nodebug__)) 640 _InterlockedExchangeSub(long volatile *_Subend, long _Value) { 641 return __atomic_sub_fetch(_Subend, _Value, 0) + _Value; 642 } 643 #ifdef __x86_64__ 644 static __inline__ __int64 __attribute__((__always_inline__, __nodebug__)) 645 _InterlockedExchangeSub64(__int64 volatile *_Subend, __int64 _Value) { 646 return __atomic_sub_fetch(_Subend, _Value, 0) + _Value; 647 } 648 #endif 649 /*----------------------------------------------------------------------------*\ 650 |* Interlocked Increment 651 \*----------------------------------------------------------------------------*/ 652 static __inline__ short __attribute__((__always_inline__, __nodebug__)) 653 _InterlockedIncrement16(short volatile *_Value) { 654 return __atomic_add_fetch(_Value, 1, 0); 655 } 656 #ifdef __x86_64__ 657 static __inline__ __int64 __attribute__((__always_inline__, __nodebug__)) 658 _InterlockedIncrement64(__int64 volatile *_Value) { 659 return __atomic_add_fetch(_Value, 1, 0); 660 } 661 #endif 662 /*----------------------------------------------------------------------------*\ 663 |* Interlocked Decrement 664 \*----------------------------------------------------------------------------*/ 665 static __inline__ short __attribute__((__always_inline__, __nodebug__)) 666 _InterlockedDecrement16(short volatile *_Value) { 667 return __atomic_sub_fetch(_Value, 1, 0); 668 } 669 #ifdef __x86_64__ 670 static __inline__ __int64 __attribute__((__always_inline__, __nodebug__)) 671 _InterlockedDecrement64(__int64 volatile *_Value) { 672 return __atomic_sub_fetch(_Value, 1, 0); 673 } 674 #endif 675 /*----------------------------------------------------------------------------*\ 676 |* Interlocked And 677 \*----------------------------------------------------------------------------*/ 678 static __inline__ char __attribute__((__always_inline__, __nodebug__)) 679 _InterlockedAnd8(char volatile *_Value, char _Mask) { 680 return __atomic_and_fetch(_Value, _Mask, 0); 681 } 682 static __inline__ short __attribute__((__always_inline__, __nodebug__)) 683 _InterlockedAnd16(short volatile *_Value, short _Mask) { 684 return __atomic_and_fetch(_Value, _Mask, 0); 685 } 686 static __inline__ long __attribute__((__always_inline__, __nodebug__)) 687 _InterlockedAnd(long volatile *_Value, long _Mask) { 688 return __atomic_and_fetch(_Value, _Mask, 0); 689 } 690 #ifdef __x86_64__ 691 static __inline__ __int64 __attribute__((__always_inline__, __nodebug__)) 692 _InterlockedAnd64(__int64 volatile *_Value, __int64 _Mask) { 693 return __atomic_and_fetch(_Value, _Mask, 0); 694 } 695 #endif 696 /*----------------------------------------------------------------------------*\ 697 |* Interlocked Or 698 \*----------------------------------------------------------------------------*/ 699 static __inline__ char __attribute__((__always_inline__, __nodebug__)) 700 _InterlockedOr8(char volatile *_Value, char _Mask) { 701 return __atomic_or_fetch(_Value, _Mask, 0); 702 } 703 static __inline__ short __attribute__((__always_inline__, __nodebug__)) 704 _InterlockedOr16(short volatile *_Value, short _Mask) { 705 return __atomic_or_fetch(_Value, _Mask, 0); 706 } 707 static __inline__ long __attribute__((__always_inline__, __nodebug__)) 708 _InterlockedOr(long volatile *_Value, long _Mask) { 709 return __atomic_or_fetch(_Value, _Mask, 0); 710 } 711 #ifdef __x86_64__ 712 static __inline__ __int64 __attribute__((__always_inline__, __nodebug__)) 713 _InterlockedOr64(__int64 volatile *_Value, __int64 _Mask) { 714 return __atomic_or_fetch(_Value, _Mask, 0); 715 } 716 #endif 717 /*----------------------------------------------------------------------------*\ 718 |* Interlocked Xor 719 \*----------------------------------------------------------------------------*/ 720 static __inline__ char __attribute__((__always_inline__, __nodebug__)) 721 _InterlockedXor8(char volatile *_Value, char _Mask) { 722 return __atomic_xor_fetch(_Value, _Mask, 0); 723 } 724 static __inline__ short __attribute__((__always_inline__, __nodebug__)) 725 _InterlockedXor16(short volatile *_Value, short _Mask) { 726 return __atomic_xor_fetch(_Value, _Mask, 0); 727 } 728 static __inline__ long __attribute__((__always_inline__, __nodebug__)) 729 _InterlockedXor(long volatile *_Value, long _Mask) { 730 return __atomic_xor_fetch(_Value, _Mask, 0); 731 } 732 #ifdef __x86_64__ 733 static __inline__ __int64 __attribute__((__always_inline__, __nodebug__)) 734 _InterlockedXor64(__int64 volatile *_Value, __int64 _Mask) { 735 return __atomic_xor_fetch(_Value, _Mask, 0); 736 } 737 #endif 738 /*----------------------------------------------------------------------------*\ 739 |* Interlocked Exchange 740 \*----------------------------------------------------------------------------*/ 741 static __inline__ char __attribute__((__always_inline__, __nodebug__)) 742 _InterlockedExchange8(char volatile *_Target, char _Value) { 743 __atomic_exchange(_Target, &_Value, &_Value, 0); 744 return _Value; 745 } 746 static __inline__ short __attribute__((__always_inline__, __nodebug__)) 747 _InterlockedExchange16(short volatile *_Target, short _Value) { 748 __atomic_exchange(_Target, &_Value, &_Value, 0); 749 return _Value; 750 } 751 #ifdef __x86_64__ 752 static __inline__ __int64 __attribute__((__always_inline__, __nodebug__)) 753 _InterlockedExchange64(__int64 volatile *_Target, __int64 _Value) { 754 __atomic_exchange(_Target, &_Value, &_Value, 0); 755 return _Value; 756 } 757 #endif 758 /*----------------------------------------------------------------------------*\ 759 |* Interlocked Compare Exchange 760 \*----------------------------------------------------------------------------*/ 761 static __inline__ char __attribute__((__always_inline__, __nodebug__)) 762 _InterlockedCompareExchange8(char volatile *_Destination, 763 char _Exchange, char _Comparand) { 764 __atomic_compare_exchange(_Destination, &_Comparand, &_Exchange, 0, 0, 0); 765 return _Comparand; 766 } 767 static __inline__ short __attribute__((__always_inline__, __nodebug__)) 768 _InterlockedCompareExchange16(short volatile *_Destination, 769 short _Exchange, short _Comparand) { 770 __atomic_compare_exchange(_Destination, &_Comparand, &_Exchange, 0, 0, 0); 771 return _Comparand; 772 } 773 static __inline__ __int64 __attribute__((__always_inline__, __nodebug__)) 774 _InterlockedCompareExchange64(__int64 volatile *_Destination, 775 __int64 _Exchange, __int64 _Comparand) { 776 __atomic_compare_exchange(_Destination, &_Comparand, &_Exchange, 0, 0, 0); 777 return _Comparand; 778 } 779 /*----------------------------------------------------------------------------*\ 780 |* Barriers 781 \*----------------------------------------------------------------------------*/ 782 #if defined(__i386__) || defined(__x86_64__) 783 static __inline__ void __attribute__((__always_inline__, __nodebug__)) 784 __attribute__((__deprecated__("use other intrinsics or C++11 atomics instead"))) 785 _ReadWriteBarrier(void) { 786 __asm__ volatile ("" : : : "memory"); 787 } 788 static __inline__ void __attribute__((__always_inline__, __nodebug__)) 789 __attribute__((__deprecated__("use other intrinsics or C++11 atomics instead"))) 790 _ReadBarrier(void) { 791 __asm__ volatile ("" : : : "memory"); 792 } 793 static __inline__ void __attribute__((__always_inline__, __nodebug__)) 794 __attribute__((__deprecated__("use other intrinsics or C++11 atomics instead"))) 795 _WriteBarrier(void) { 796 __asm__ volatile ("" : : : "memory"); 797 } 798 #endif 799 #ifdef __x86_64__ 800 static __inline__ void __attribute__((__always_inline__, __nodebug__)) 801 __faststorefence(void) { 802 __asm__ volatile("lock orq $0, (%%rsp)" : : : "memory"); 803 } 804 #endif 805 /*----------------------------------------------------------------------------*\ 806 |* readfs, readgs 807 |* (Pointers in address space #256 and #257 are relative to the GS and FS 808 |* segment registers, respectively.) 809 \*----------------------------------------------------------------------------*/ 810 #define __ptr_to_addr_space(__addr_space_nbr, __type, __offset) \ 811 ((volatile __type __attribute__((__address_space__(__addr_space_nbr)))*) \ 812 (__offset)) 813 814 #ifdef __i386__ 815 static __inline__ unsigned char __attribute__((__always_inline__, __nodebug__)) 816 __readfsbyte(unsigned long __offset) { 817 return *__ptr_to_addr_space(257, unsigned char, __offset); 818 } 819 static __inline__ unsigned __int64 __attribute__((__always_inline__, __nodebug__)) 820 __readfsqword(unsigned long __offset) { 821 return *__ptr_to_addr_space(257, unsigned __int64, __offset); 822 } 823 static __inline__ unsigned short __attribute__((__always_inline__, __nodebug__)) 824 __readfsword(unsigned long __offset) { 825 return *__ptr_to_addr_space(257, unsigned short, __offset); 826 } 827 #endif 828 #ifdef __x86_64__ 829 static __inline__ unsigned char __attribute__((__always_inline__, __nodebug__)) 830 __readgsbyte(unsigned long __offset) { 831 return *__ptr_to_addr_space(256, unsigned char, __offset); 832 } 833 static __inline__ unsigned long __attribute__((__always_inline__, __nodebug__)) 834 __readgsdword(unsigned long __offset) { 835 return *__ptr_to_addr_space(256, unsigned long, __offset); 836 } 837 static __inline__ unsigned __int64 __attribute__((__always_inline__, __nodebug__)) 838 __readgsqword(unsigned long __offset) { 839 return *__ptr_to_addr_space(256, unsigned __int64, __offset); 840 } 841 static __inline__ unsigned short __attribute__((__always_inline__, __nodebug__)) 842 __readgsword(unsigned long __offset) { 843 return *__ptr_to_addr_space(256, unsigned short, __offset); 844 } 845 #endif 846 #undef __ptr_to_addr_space 847 /*----------------------------------------------------------------------------*\ 848 |* movs, stos 849 \*----------------------------------------------------------------------------*/ 850 #if defined(__i386__) || defined(__x86_64__) 851 static __inline__ void __attribute__((__always_inline__, __nodebug__)) 852 __movsb(unsigned char *__dst, unsigned char const *__src, size_t __n) { 853 __asm__("rep movsb" : : "D"(__dst), "S"(__src), "c"(__n) 854 : "%edi", "%esi", "%ecx"); 855 } 856 static __inline__ void __attribute__((__always_inline__, __nodebug__)) 857 __movsd(unsigned long *__dst, unsigned long const *__src, size_t __n) { 858 __asm__("rep movsl" : : "D"(__dst), "S"(__src), "c"(__n) 859 : "%edi", "%esi", "%ecx"); 860 } 861 static __inline__ void __attribute__((__always_inline__, __nodebug__)) 862 __movsw(unsigned short *__dst, unsigned short const *__src, size_t __n) { 863 __asm__("rep movsh" : : "D"(__dst), "S"(__src), "c"(__n) 864 : "%edi", "%esi", "%ecx"); 865 } 866 static __inline__ void __attribute__((__always_inline__, __nodebug__)) 867 __stosb(unsigned char *__dst, unsigned char __x, size_t __n) { 868 __asm__("rep stosb" : : "D"(__dst), "a"(__x), "c"(__n) 869 : "%edi", "%ecx"); 870 } 871 static __inline__ void __attribute__((__always_inline__, __nodebug__)) 872 __stosd(unsigned long *__dst, unsigned long __x, size_t __n) { 873 __asm__("rep stosl" : : "D"(__dst), "a"(__x), "c"(__n) 874 : "%edi", "%ecx"); 875 } 876 static __inline__ void __attribute__((__always_inline__, __nodebug__)) 877 __stosw(unsigned short *__dst, unsigned short __x, size_t __n) { 878 __asm__("rep stosh" : : "D"(__dst), "a"(__x), "c"(__n) 879 : "%edi", "%ecx"); 880 } 881 #endif 882 #ifdef __x86_64__ 883 static __inline__ void __attribute__((__always_inline__, __nodebug__)) 884 __movsq(unsigned long long *__dst, unsigned long long const *__src, size_t __n) { 885 __asm__("rep movsq" : : "D"(__dst), "S"(__src), "c"(__n) 886 : "%edi", "%esi", "%ecx"); 887 } 888 static __inline__ void __attribute__((__always_inline__, __nodebug__)) 889 __stosq(unsigned __int64 *__dst, unsigned __int64 __x, size_t __n) { 890 __asm__("rep stosq" : : "D"(__dst), "a"(__x), "c"(__n) 891 : "%edi", "%ecx"); 892 } 893 #endif 894 895 /*----------------------------------------------------------------------------*\ 896 |* Misc 897 \*----------------------------------------------------------------------------*/ 898 static __inline__ void * __attribute__((__always_inline__, __nodebug__)) 899 _AddressOfReturnAddress(void) { 900 return (void*)((char*)__builtin_frame_address(0) + sizeof(void*)); 901 } 902 static __inline__ void * __attribute__((__always_inline__, __nodebug__)) 903 _ReturnAddress(void) { 904 return __builtin_return_address(0); 905 } 906 #if defined(__i386__) || defined(__x86_64__) 907 static __inline__ void __attribute__((__always_inline__, __nodebug__)) 908 __cpuid(int __info[4], int __level) { 909 __asm__ ("cpuid" : "=a"(__info[0]), "=b" (__info[1]), "=c"(__info[2]), "=d"(__info[3]) 910 : "a"(__level)); 911 } 912 static __inline__ void __attribute__((__always_inline__, __nodebug__)) 913 __cpuidex(int __info[4], int __level, int __ecx) { 914 __asm__ ("cpuid" : "=a"(__info[0]), "=b" (__info[1]), "=c"(__info[2]), "=d"(__info[3]) 915 : "a"(__level), "c"(__ecx)); 916 } 917 static __inline__ unsigned __int64 __cdecl __attribute__((__always_inline__, __nodebug__)) 918 _xgetbv(unsigned int __xcr_no) { 919 unsigned int __eax, __edx; 920 __asm__ ("xgetbv" : "=a" (__eax), "=d" (__edx) : "c" (__xcr_no)); 921 return ((unsigned __int64)__edx << 32) | __eax; 922 } 923 static __inline__ void __attribute__((__always_inline__, __nodebug__)) 924 __halt(void) { 925 __asm__ volatile ("hlt"); 926 } 927 #endif 928 929 /*----------------------------------------------------------------------------*\ 930 |* Privileged intrinsics 931 \*----------------------------------------------------------------------------*/ 932 #if defined(__i386__) || defined(__x86_64__) 933 static __inline__ unsigned __int64 __attribute__((__always_inline__, __nodebug__)) 934 __readmsr(unsigned long __register) { 935 // Loads the contents of a 64-bit model specific register (MSR) specified in 936 // the ECX register into registers EDX:EAX. The EDX register is loaded with 937 // the high-order 32 bits of the MSR and the EAX register is loaded with the 938 // low-order 32 bits. If less than 64 bits are implemented in the MSR being 939 // read, the values returned to EDX:EAX in unimplemented bit locations are 940 // undefined. 941 unsigned long __edx; 942 unsigned long __eax; 943 __asm__ ("rdmsr" : "=d"(__edx), "=a"(__eax) : "c"(__register)); 944 return (((unsigned __int64)__edx) << 32) | (unsigned __int64)__eax; 945 } 946 947 static __inline__ unsigned long __attribute__((__always_inline__, __nodebug__)) 948 __readcr3(void) { 949 unsigned long __cr3_val; 950 __asm__ __volatile__ ("mov %%cr3, %0" : "=q"(__cr3_val) : : "memory"); 951 return __cr3_val; 952 } 953 954 static __inline__ void __attribute__((__always_inline__, __nodebug__)) 955 __writecr3(unsigned int __cr3_val) { 956 __asm__ ("mov %0, %%cr3" : : "q"(__cr3_val) : "memory"); 957 } 958 #endif 959 960 #ifdef __cplusplus 961 } 962 #endif 963 964 #endif /* __INTRIN_H */ 965 #endif /* _MSC_VER */ 966
