Home | History | Annotate | Download | only in include
      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 #if defined(__arm__)
     38 #include <armintr.h>
     39 #endif
     40 
     41 /* For the definition of jmp_buf. */
     42 #if __STDC_HOSTED__
     43 #include <setjmp.h>
     44 #endif
     45 
     46 /* Define the default attributes for the functions in this file. */
     47 #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__))
     48 
     49 #ifdef __cplusplus
     50 extern "C" {
     51 #endif
     52 
     53 #if defined(__MMX__)
     54 /* And the random ones that aren't in those files. */
     55 __m64 _m_from_float(float);
     56 float _m_to_float(__m64);
     57 #endif
     58 
     59 /* Other assorted instruction intrinsics. */
     60 void __addfsbyte(unsigned long, unsigned char);
     61 void __addfsdword(unsigned long, unsigned long);
     62 void __addfsword(unsigned long, unsigned short);
     63 void __code_seg(const char *);
     64 static __inline__
     65 void __cpuid(int[4], int);
     66 static __inline__
     67 void __cpuidex(int[4], int, int);
     68 static __inline__
     69 __int64 __emul(int, int);
     70 static __inline__
     71 unsigned __int64 __emulu(unsigned int, unsigned int);
     72 unsigned int __getcallerseflags(void);
     73 static __inline__
     74 void __halt(void);
     75 unsigned char __inbyte(unsigned short);
     76 void __inbytestring(unsigned short, unsigned char *, unsigned long);
     77 void __incfsbyte(unsigned long);
     78 void __incfsdword(unsigned long);
     79 void __incfsword(unsigned long);
     80 unsigned long __indword(unsigned short);
     81 void __indwordstring(unsigned short, unsigned long *, unsigned long);
     82 void __invlpg(void *);
     83 unsigned short __inword(unsigned short);
     84 void __inwordstring(unsigned short, unsigned short *, unsigned long);
     85 void __lidt(void *);
     86 unsigned __int64 __ll_lshift(unsigned __int64, int);
     87 __int64 __ll_rshift(__int64, int);
     88 void __llwpcb(void *);
     89 unsigned char __lwpins32(unsigned int, unsigned int, unsigned int);
     90 void __lwpval32(unsigned int, unsigned int, unsigned int);
     91 unsigned int __lzcnt(unsigned int);
     92 unsigned short __lzcnt16(unsigned short);
     93 static __inline__
     94 void __movsb(unsigned char *, unsigned char const *, size_t);
     95 static __inline__
     96 void __movsd(unsigned long *, unsigned long const *, size_t);
     97 static __inline__
     98 void __movsw(unsigned short *, unsigned short const *, size_t);
     99 static __inline__
    100 void __nop(void);
    101 void __nvreg_restore_fence(void);
    102 void __nvreg_save_fence(void);
    103 void __outbyte(unsigned short, unsigned char);
    104 void __outbytestring(unsigned short, unsigned char *, unsigned long);
    105 void __outdword(unsigned short, unsigned long);
    106 void __outdwordstring(unsigned short, unsigned long *, unsigned long);
    107 void __outword(unsigned short, unsigned short);
    108 void __outwordstring(unsigned short, unsigned short *, unsigned long);
    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 __int64 __readfsqword(unsigned long);
    121 static __inline__
    122 unsigned short __readfsword(unsigned long);
    123 #endif
    124 static __inline__
    125 unsigned __int64 __readmsr(unsigned long);
    126 unsigned __int64 __readpmc(unsigned long);
    127 unsigned long __segmentlimit(unsigned long);
    128 void __sidt(void *);
    129 void *__slwpcb(void);
    130 static __inline__
    131 void __stosb(unsigned char *, unsigned char, size_t);
    132 static __inline__
    133 void __stosd(unsigned long *, unsigned long, size_t);
    134 static __inline__
    135 void __stosw(unsigned short *, unsigned short, size_t);
    136 void __svm_clgi(void);
    137 void __svm_invlpga(void *, int);
    138 void __svm_skinit(int);
    139 void __svm_stgi(void);
    140 void __svm_vmload(size_t);
    141 void __svm_vmrun(size_t);
    142 void __svm_vmsave(size_t);
    143 unsigned __int64 __ull_rshift(unsigned __int64, int);
    144 void __vmx_off(void);
    145 void __vmx_vmptrst(unsigned __int64 *);
    146 void __wbinvd(void);
    147 void __writecr0(unsigned int);
    148 static __inline__
    149 void __writecr3(unsigned int);
    150 void __writecr4(unsigned int);
    151 void __writecr8(unsigned int);
    152 void __writedr(unsigned int, unsigned int);
    153 void __writefsbyte(unsigned long, unsigned char);
    154 void __writefsdword(unsigned long, unsigned long);
    155 void __writefsqword(unsigned long, unsigned __int64);
    156 void __writefsword(unsigned long, unsigned short);
    157 void __writemsr(unsigned long, unsigned __int64);
    158 static __inline__
    159 void *_AddressOfReturnAddress(void);
    160 static __inline__
    161 unsigned char _BitScanForward(unsigned long *_Index, unsigned long _Mask);
    162 static __inline__
    163 unsigned char _BitScanReverse(unsigned long *_Index, unsigned long _Mask);
    164 static __inline__
    165 unsigned char _bittest(long const *, long);
    166 static __inline__
    167 unsigned char _bittestandcomplement(long *, long);
    168 static __inline__
    169 unsigned char _bittestandreset(long *, long);
    170 static __inline__
    171 unsigned char _bittestandset(long *, long);
    172 void __cdecl _disable(void);
    173 void __cdecl _enable(void);
    174 long _InterlockedAddLargeStatistic(__int64 volatile *_Addend, long _Value);
    175 unsigned char _interlockedbittestandreset(long volatile *, long);
    176 unsigned char _interlockedbittestandset(long volatile *, long);
    177 long _InterlockedCompareExchange_HLEAcquire(long volatile *, long, long);
    178 long _InterlockedCompareExchange_HLERelease(long volatile *, long, long);
    179 __int64 _InterlockedcompareExchange64_HLEAcquire(__int64 volatile *, __int64,
    180                                                  __int64);
    181 __int64 _InterlockedCompareExchange64_HLERelease(__int64 volatile *, __int64,
    182                                                  __int64);
    183 void *_InterlockedCompareExchangePointer_HLEAcquire(void *volatile *, void *,
    184                                                     void *);
    185 void *_InterlockedCompareExchangePointer_HLERelease(void *volatile *, void *,
    186                                                     void *);
    187 long _InterlockedExchangeAdd_HLEAcquire(long volatile *, long);
    188 long _InterlockedExchangeAdd_HLERelease(long volatile *, long);
    189 __int64 _InterlockedExchangeAdd64_HLEAcquire(__int64 volatile *, __int64);
    190 __int64 _InterlockedExchangeAdd64_HLERelease(__int64 volatile *, __int64);
    191 void __cdecl _invpcid(unsigned int, void *);
    192 static __inline__ void
    193 __attribute__((__deprecated__("use other intrinsics or C++11 atomics instead")))
    194 _ReadBarrier(void);
    195 static __inline__ void
    196 __attribute__((__deprecated__("use other intrinsics or C++11 atomics instead")))
    197 _ReadWriteBarrier(void);
    198 unsigned int _rorx_u32(unsigned int, const unsigned int);
    199 int _sarx_i32(int, unsigned int);
    200 #if __STDC_HOSTED__
    201 int __cdecl _setjmp(jmp_buf);
    202 #endif
    203 unsigned int _shlx_u32(unsigned int, unsigned int);
    204 unsigned int _shrx_u32(unsigned int, unsigned int);
    205 void _Store_HLERelease(long volatile *, long);
    206 void _Store64_HLERelease(__int64 volatile *, __int64);
    207 void _StorePointer_HLERelease(void *volatile *, void *);
    208 static __inline__ void
    209 __attribute__((__deprecated__("use other intrinsics or C++11 atomics instead")))
    210 _WriteBarrier(void);
    211 unsigned __int32 xbegin(void);
    212 void _xend(void);
    213 static __inline__
    214 #define _XCR_XFEATURE_ENABLED_MASK 0
    215 unsigned __int64 __cdecl _xgetbv(unsigned int);
    216 void __cdecl _xsetbv(unsigned int, unsigned __int64);
    217 
    218 /* These additional intrinsics are turned on in x64/amd64/x86_64 mode. */
    219 #ifdef __x86_64__
    220 void __addgsbyte(unsigned long, unsigned char);
    221 void __addgsdword(unsigned long, unsigned long);
    222 void __addgsqword(unsigned long, unsigned __int64);
    223 void __addgsword(unsigned long, unsigned short);
    224 static __inline__
    225 void __faststorefence(void);
    226 void __incgsbyte(unsigned long);
    227 void __incgsdword(unsigned long);
    228 void __incgsqword(unsigned long);
    229 void __incgsword(unsigned long);
    230 unsigned char __lwpins64(unsigned __int64, unsigned int, unsigned int);
    231 void __lwpval64(unsigned __int64, unsigned int, unsigned int);
    232 unsigned __int64 __lzcnt64(unsigned __int64);
    233 static __inline__
    234 void __movsq(unsigned long long *, unsigned long long const *, size_t);
    235 static __inline__
    236 unsigned char __readgsbyte(unsigned long);
    237 static __inline__
    238 unsigned long __readgsdword(unsigned long);
    239 static __inline__
    240 unsigned __int64 __readgsqword(unsigned long);
    241 unsigned short __readgsword(unsigned long);
    242 unsigned __int64 __shiftleft128(unsigned __int64 _LowPart,
    243                                 unsigned __int64 _HighPart,
    244                                 unsigned char _Shift);
    245 unsigned __int64 __shiftright128(unsigned __int64 _LowPart,
    246                                  unsigned __int64 _HighPart,
    247                                  unsigned char _Shift);
    248 static __inline__
    249 void __stosq(unsigned __int64 *, unsigned __int64, size_t);
    250 unsigned char __vmx_on(unsigned __int64 *);
    251 unsigned char __vmx_vmclear(unsigned __int64 *);
    252 unsigned char __vmx_vmlaunch(void);
    253 unsigned char __vmx_vmptrld(unsigned __int64 *);
    254 unsigned char __vmx_vmread(size_t, size_t *);
    255 unsigned char __vmx_vmresume(void);
    256 unsigned char __vmx_vmwrite(size_t, size_t);
    257 void __writegsbyte(unsigned long, unsigned char);
    258 void __writegsdword(unsigned long, unsigned long);
    259 void __writegsqword(unsigned long, unsigned __int64);
    260 void __writegsword(unsigned long, unsigned short);
    261 static __inline__
    262 unsigned char _BitScanForward64(unsigned long *_Index, unsigned __int64 _Mask);
    263 static __inline__
    264 unsigned char _BitScanReverse64(unsigned long *_Index, unsigned __int64 _Mask);
    265 static __inline__
    266 unsigned char _bittest64(__int64 const *, __int64);
    267 static __inline__
    268 unsigned char _bittestandcomplement64(__int64 *, __int64);
    269 static __inline__
    270 unsigned char _bittestandreset64(__int64 *, __int64);
    271 static __inline__
    272 unsigned char _bittestandset64(__int64 *, __int64);
    273 long _InterlockedAnd_np(long volatile *_Value, long _Mask);
    274 short _InterlockedAnd16_np(short volatile *_Value, short _Mask);
    275 __int64 _InterlockedAnd64_np(__int64 volatile *_Value, __int64 _Mask);
    276 char _InterlockedAnd8_np(char volatile *_Value, char _Mask);
    277 unsigned char _interlockedbittestandreset64(__int64 volatile *, __int64);
    278 static __inline__
    279 unsigned char _interlockedbittestandset64(__int64 volatile *, __int64);
    280 long _InterlockedCompareExchange_np(long volatile *_Destination, long _Exchange,
    281                                     long _Comparand);
    282 unsigned char _InterlockedCompareExchange128(__int64 volatile *_Destination,
    283                                              __int64 _ExchangeHigh,
    284                                              __int64 _ExchangeLow,
    285                                              __int64 *_CompareandResult);
    286 unsigned char _InterlockedCompareExchange128_np(__int64 volatile *_Destination,
    287                                                 __int64 _ExchangeHigh,
    288                                                 __int64 _ExchangeLow,
    289                                                 __int64 *_ComparandResult);
    290 short _InterlockedCompareExchange16_np(short volatile *_Destination,
    291                                        short _Exchange, short _Comparand);
    292 __int64 _InterlockedCompareExchange64_HLEAcquire(__int64 volatile *, __int64,
    293                                                  __int64);
    294 __int64 _InterlockedCompareExchange64_HLERelease(__int64 volatile *, __int64,
    295                                                  __int64);
    296 __int64 _InterlockedCompareExchange64_np(__int64 volatile *_Destination,
    297                                          __int64 _Exchange, __int64 _Comparand);
    298 void *_InterlockedCompareExchangePointer_np(void *volatile *_Destination,
    299                                             void *_Exchange, void *_Comparand);
    300 long _InterlockedOr_np(long volatile *_Value, long _Mask);
    301 short _InterlockedOr16_np(short volatile *_Value, short _Mask);
    302 __int64 _InterlockedOr64_np(__int64 volatile *_Value, __int64 _Mask);
    303 char _InterlockedOr8_np(char volatile *_Value, char _Mask);
    304 long _InterlockedXor_np(long volatile *_Value, long _Mask);
    305 short _InterlockedXor16_np(short volatile *_Value, short _Mask);
    306 __int64 _InterlockedXor64_np(__int64 volatile *_Value, __int64 _Mask);
    307 char _InterlockedXor8_np(char volatile *_Value, char _Mask);
    308 unsigned __int64 _rorx_u64(unsigned __int64, const unsigned int);
    309 __int64 _sarx_i64(__int64, unsigned int);
    310 unsigned __int64 _shlx_u64(unsigned __int64, unsigned int);
    311 unsigned __int64 _shrx_u64(unsigned __int64, unsigned int);
    312 static __inline__
    313 __int64 __mulh(__int64, __int64);
    314 static __inline__
    315 unsigned __int64 __umulh(unsigned __int64, unsigned __int64);
    316 static __inline__
    317 __int64 _mul128(__int64, __int64, __int64*);
    318 static __inline__
    319 unsigned __int64 _umul128(unsigned __int64,
    320                           unsigned __int64,
    321                           unsigned __int64*);
    322 
    323 #endif /* __x86_64__ */
    324 
    325 #if defined(__x86_64__) || defined(__arm__)
    326 
    327 static __inline__
    328 __int64 _InterlockedDecrement64(__int64 volatile *_Addend);
    329 static __inline__
    330 __int64 _InterlockedExchange64(__int64 volatile *_Target, __int64 _Value);
    331 static __inline__
    332 __int64 _InterlockedExchangeAdd64(__int64 volatile *_Addend, __int64 _Value);
    333 static __inline__
    334 __int64 _InterlockedExchangeSub64(__int64 volatile *_Subend, __int64 _Value);
    335 static __inline__
    336 __int64 _InterlockedIncrement64(__int64 volatile *_Addend);
    337 static __inline__
    338 __int64 _InterlockedOr64(__int64 volatile *_Value, __int64 _Mask);
    339 static __inline__
    340 __int64 _InterlockedXor64(__int64 volatile *_Value, __int64 _Mask);
    341 static __inline__
    342 __int64 _InterlockedAnd64(__int64 volatile *_Value, __int64 _Mask);
    343 
    344 #endif
    345 
    346 /*----------------------------------------------------------------------------*\
    347 |* Bit Counting and Testing
    348 \*----------------------------------------------------------------------------*/
    349 static __inline__ unsigned char __DEFAULT_FN_ATTRS
    350 _bittest(long const *_BitBase, long _BitPos) {
    351   return (*_BitBase >> _BitPos) & 1;
    352 }
    353 static __inline__ unsigned char __DEFAULT_FN_ATTRS
    354 _bittestandcomplement(long *_BitBase, long _BitPos) {
    355   unsigned char _Res = (*_BitBase >> _BitPos) & 1;
    356   *_BitBase = *_BitBase ^ (1 << _BitPos);
    357   return _Res;
    358 }
    359 static __inline__ unsigned char __DEFAULT_FN_ATTRS
    360 _bittestandreset(long *_BitBase, long _BitPos) {
    361   unsigned char _Res = (*_BitBase >> _BitPos) & 1;
    362   *_BitBase = *_BitBase & ~(1 << _BitPos);
    363   return _Res;
    364 }
    365 static __inline__ unsigned char __DEFAULT_FN_ATTRS
    366 _bittestandset(long *_BitBase, long _BitPos) {
    367   unsigned char _Res = (*_BitBase >> _BitPos) & 1;
    368   *_BitBase = *_BitBase | (1 << _BitPos);
    369   return _Res;
    370 }
    371 #if defined(__arm__) || defined(__aarch64__)
    372 static __inline__ unsigned char __DEFAULT_FN_ATTRS
    373 _interlockedbittestandset_acq(long volatile *_BitBase, long _BitPos) {
    374   long _PrevVal = __atomic_fetch_or(_BitBase, 1l << _BitPos, __ATOMIC_ACQUIRE);
    375   return (_PrevVal >> _BitPos) & 1;
    376 }
    377 static __inline__ unsigned char __DEFAULT_FN_ATTRS
    378 _interlockedbittestandset_nf(long volatile *_BitBase, long _BitPos) {
    379   long _PrevVal = __atomic_fetch_or(_BitBase, 1l << _BitPos, __ATOMIC_RELAXED);
    380   return (_PrevVal >> _BitPos) & 1;
    381 }
    382 static __inline__ unsigned char __DEFAULT_FN_ATTRS
    383 _interlockedbittestandset_rel(long volatile *_BitBase, long _BitPos) {
    384   long _PrevVal = __atomic_fetch_or(_BitBase, 1l << _BitPos, __ATOMIC_RELEASE);
    385   return (_PrevVal >> _BitPos) & 1;
    386 }
    387 #endif
    388 #ifdef __x86_64__
    389 static __inline__ unsigned char __DEFAULT_FN_ATTRS
    390 _bittest64(__int64 const *_BitBase, __int64 _BitPos) {
    391   return (*_BitBase >> _BitPos) & 1;
    392 }
    393 static __inline__ unsigned char __DEFAULT_FN_ATTRS
    394 _bittestandcomplement64(__int64 *_BitBase, __int64 _BitPos) {
    395   unsigned char _Res = (*_BitBase >> _BitPos) & 1;
    396   *_BitBase = *_BitBase ^ (1ll << _BitPos);
    397   return _Res;
    398 }
    399 static __inline__ unsigned char __DEFAULT_FN_ATTRS
    400 _bittestandreset64(__int64 *_BitBase, __int64 _BitPos) {
    401   unsigned char _Res = (*_BitBase >> _BitPos) & 1;
    402   *_BitBase = *_BitBase & ~(1ll << _BitPos);
    403   return _Res;
    404 }
    405 static __inline__ unsigned char __DEFAULT_FN_ATTRS
    406 _bittestandset64(__int64 *_BitBase, __int64 _BitPos) {
    407   unsigned char _Res = (*_BitBase >> _BitPos) & 1;
    408   *_BitBase = *_BitBase | (1ll << _BitPos);
    409   return _Res;
    410 }
    411 static __inline__ unsigned char __DEFAULT_FN_ATTRS
    412 _interlockedbittestandset64(__int64 volatile *_BitBase, __int64 _BitPos) {
    413   long long _PrevVal =
    414       __atomic_fetch_or(_BitBase, 1ll << _BitPos, __ATOMIC_SEQ_CST);
    415   return (_PrevVal >> _BitPos) & 1;
    416 }
    417 #endif
    418 /*----------------------------------------------------------------------------*\
    419 |* Interlocked Exchange Add
    420 \*----------------------------------------------------------------------------*/
    421 #if defined(__arm__) || defined(__aarch64__)
    422 static __inline__ char __DEFAULT_FN_ATTRS
    423 _InterlockedExchangeAdd8_acq(char volatile *_Addend, char _Value) {
    424   return __atomic_fetch_add(_Addend, _Value, __ATOMIC_ACQUIRE);
    425 }
    426 static __inline__ char __DEFAULT_FN_ATTRS
    427 _InterlockedExchangeAdd8_nf(char volatile *_Addend, char _Value) {
    428   return __atomic_fetch_add(_Addend, _Value, __ATOMIC_RELAXED);
    429 }
    430 static __inline__ char __DEFAULT_FN_ATTRS
    431 _InterlockedExchangeAdd8_rel(char volatile *_Addend, char _Value) {
    432   return __atomic_fetch_add(_Addend, _Value, __ATOMIC_RELAXED);
    433 }
    434 static __inline__ short __DEFAULT_FN_ATTRS
    435 _InterlockedExchangeAdd16_acq(short volatile *_Addend, short _Value) {
    436   return __atomic_fetch_add(_Addend, _Value, __ATOMIC_ACQUIRE);
    437 }
    438 static __inline__ short __DEFAULT_FN_ATTRS
    439 _InterlockedExchangeAdd16_nf(short volatile *_Addend, short _Value) {
    440   return __atomic_fetch_add(_Addend, _Value, __ATOMIC_RELAXED);
    441 }
    442 static __inline__ short __DEFAULT_FN_ATTRS
    443 _InterlockedExchangeAdd16_rel(short volatile *_Addend, short _Value) {
    444   return __atomic_fetch_add(_Addend, _Value, __ATOMIC_RELEASE);
    445 }
    446 static __inline__ long __DEFAULT_FN_ATTRS
    447 _InterlockedExchangeAdd_acq(long volatile *_Addend, long _Value) {
    448   return __atomic_fetch_add(_Addend, _Value, __ATOMIC_ACQUIRE);
    449 }
    450 static __inline__ long __DEFAULT_FN_ATTRS
    451 _InterlockedExchangeAdd_nf(long volatile *_Addend, long _Value) {
    452   return __atomic_fetch_add(_Addend, _Value, __ATOMIC_RELAXED);
    453 }
    454 static __inline__ long __DEFAULT_FN_ATTRS
    455 _InterlockedExchangeAdd_rel(long volatile *_Addend, long _Value) {
    456   return __atomic_fetch_add(_Addend, _Value, __ATOMIC_RELEASE);
    457 }
    458 static __inline__ __int64 __DEFAULT_FN_ATTRS
    459 _InterlockedExchangeAdd64_acq(__int64 volatile *_Addend, __int64 _Value) {
    460   return __atomic_fetch_add(_Addend, _Value, __ATOMIC_ACQUIRE);
    461 }
    462 static __inline__ __int64 __DEFAULT_FN_ATTRS
    463 _InterlockedExchangeAdd64_nf(__int64 volatile *_Addend, __int64 _Value) {
    464   return __atomic_fetch_add(_Addend, _Value, __ATOMIC_RELAXED);
    465 }
    466 static __inline__ __int64 __DEFAULT_FN_ATTRS
    467 _InterlockedExchangeAdd64_rel(__int64 volatile *_Addend, __int64 _Value) {
    468   return __atomic_fetch_add(_Addend, _Value, __ATOMIC_RELEASE);
    469 }
    470 #endif
    471 /*----------------------------------------------------------------------------*\
    472 |* Interlocked Increment
    473 \*----------------------------------------------------------------------------*/
    474 #if defined(__arm__) || defined(__aarch64__)
    475 static __inline__ short __DEFAULT_FN_ATTRS
    476 _InterlockedIncrement16_acq(short volatile *_Value) {
    477   return __atomic_add_fetch(_Value, 1, __ATOMIC_ACQUIRE);
    478 }
    479 static __inline__ short __DEFAULT_FN_ATTRS
    480 _InterlockedIncrement16_nf(short volatile *_Value) {
    481   return __atomic_add_fetch(_Value, 1, __ATOMIC_RELAXED);
    482 }
    483 static __inline__ short __DEFAULT_FN_ATTRS
    484 _InterlockedIncrement16_rel(short volatile *_Value) {
    485   return __atomic_add_fetch(_Value, 1, __ATOMIC_RELEASE);
    486 }
    487 static __inline__ long __DEFAULT_FN_ATTRS
    488 _InterlockedIncrement_acq(long volatile *_Value) {
    489   return __atomic_add_fetch(_Value, 1, __ATOMIC_ACQUIRE);
    490 }
    491 static __inline__ long __DEFAULT_FN_ATTRS
    492 _InterlockedIncrement_nf(long volatile *_Value) {
    493   return __atomic_add_fetch(_Value, 1, __ATOMIC_RELAXED);
    494 }
    495 static __inline__ long __DEFAULT_FN_ATTRS
    496 _InterlockedIncrement_rel(long volatile *_Value) {
    497   return __atomic_add_fetch(_Value, 1, __ATOMIC_RELEASE);
    498 }
    499 static __inline__ __int64 __DEFAULT_FN_ATTRS
    500 _InterlockedIncrement64_acq(__int64 volatile *_Value) {
    501   return __atomic_add_fetch(_Value, 1, __ATOMIC_ACQUIRE);
    502 }
    503 static __inline__ __int64 __DEFAULT_FN_ATTRS
    504 _InterlockedIncrement64_nf(__int64 volatile *_Value) {
    505   return __atomic_add_fetch(_Value, 1, __ATOMIC_RELAXED);
    506 }
    507 static __inline__ __int64 __DEFAULT_FN_ATTRS
    508 _InterlockedIncrement64_rel(__int64 volatile *_Value) {
    509   return __atomic_add_fetch(_Value, 1, __ATOMIC_RELEASE);
    510 }
    511 #endif
    512 /*----------------------------------------------------------------------------*\
    513 |* Interlocked Decrement
    514 \*----------------------------------------------------------------------------*/
    515 #if defined(__arm__) || defined(__aarch64__)
    516 static __inline__ short __DEFAULT_FN_ATTRS
    517 _InterlockedDecrement16_acq(short volatile *_Value) {
    518   return __atomic_sub_fetch(_Value, 1, __ATOMIC_ACQUIRE);
    519 }
    520 static __inline__ short __DEFAULT_FN_ATTRS
    521 _InterlockedDecrement16_nf(short volatile *_Value) {
    522   return __atomic_sub_fetch(_Value, 1, __ATOMIC_RELAXED);
    523 }
    524 static __inline__ short __DEFAULT_FN_ATTRS
    525 _InterlockedDecrement16_rel(short volatile *_Value) {
    526   return __atomic_sub_fetch(_Value, 1, __ATOMIC_RELEASE);
    527 }
    528 static __inline__ long __DEFAULT_FN_ATTRS
    529 _InterlockedDecrement_acq(long volatile *_Value) {
    530   return __atomic_sub_fetch(_Value, 1, __ATOMIC_ACQUIRE);
    531 }
    532 static __inline__ long __DEFAULT_FN_ATTRS
    533 _InterlockedDecrement_nf(long volatile *_Value) {
    534   return __atomic_sub_fetch(_Value, 1, __ATOMIC_RELAXED);
    535 }
    536 static __inline__ long __DEFAULT_FN_ATTRS
    537 _InterlockedDecrement_rel(long volatile *_Value) {
    538   return __atomic_sub_fetch(_Value, 1, __ATOMIC_RELEASE);
    539 }
    540 static __inline__ __int64 __DEFAULT_FN_ATTRS
    541 _InterlockedDecrement64_acq(__int64 volatile *_Value) {
    542   return __atomic_sub_fetch(_Value, 1, __ATOMIC_ACQUIRE);
    543 }
    544 static __inline__ __int64 __DEFAULT_FN_ATTRS
    545 _InterlockedDecrement64_nf(__int64 volatile *_Value) {
    546   return __atomic_sub_fetch(_Value, 1, __ATOMIC_RELAXED);
    547 }
    548 static __inline__ __int64 __DEFAULT_FN_ATTRS
    549 _InterlockedDecrement64_rel(__int64 volatile *_Value) {
    550   return __atomic_sub_fetch(_Value, 1, __ATOMIC_RELEASE);
    551 }
    552 #endif
    553 /*----------------------------------------------------------------------------*\
    554 |* Interlocked And
    555 \*----------------------------------------------------------------------------*/
    556 #if defined(__arm__) || defined(__aarch64__)
    557 static __inline__ char __DEFAULT_FN_ATTRS
    558 _InterlockedAnd8_acq(char volatile *_Value, char _Mask) {
    559   return __atomic_fetch_and(_Value, _Mask, __ATOMIC_ACQUIRE);
    560 }
    561 static __inline__ char __DEFAULT_FN_ATTRS
    562 _InterlockedAnd8_nf(char volatile *_Value, char _Mask) {
    563   return __atomic_fetch_and(_Value, _Mask, __ATOMIC_RELAXED);
    564 }
    565 static __inline__ char __DEFAULT_FN_ATTRS
    566 _InterlockedAnd8_rel(char volatile *_Value, char _Mask) {
    567   return __atomic_fetch_and(_Value, _Mask, __ATOMIC_RELEASE);
    568 }
    569 static __inline__ short __DEFAULT_FN_ATTRS
    570 _InterlockedAnd16_acq(short volatile *_Value, short _Mask) {
    571   return __atomic_fetch_and(_Value, _Mask, __ATOMIC_ACQUIRE);
    572 }
    573 static __inline__ short __DEFAULT_FN_ATTRS
    574 _InterlockedAnd16_nf(short volatile *_Value, short _Mask) {
    575   return __atomic_fetch_and(_Value, _Mask, __ATOMIC_RELAXED);
    576 }
    577 static __inline__ short __DEFAULT_FN_ATTRS
    578 _InterlockedAnd16_rel(short volatile *_Value, short _Mask) {
    579   return __atomic_fetch_and(_Value, _Mask, __ATOMIC_RELEASE);
    580 }
    581 static __inline__ long __DEFAULT_FN_ATTRS
    582 _InterlockedAnd_acq(long volatile *_Value, long _Mask) {
    583   return __atomic_fetch_and(_Value, _Mask, __ATOMIC_ACQUIRE);
    584 }
    585 static __inline__ long __DEFAULT_FN_ATTRS
    586 _InterlockedAnd_nf(long volatile *_Value, long _Mask) {
    587   return __atomic_fetch_and(_Value, _Mask, __ATOMIC_RELAXED);
    588 }
    589 static __inline__ long __DEFAULT_FN_ATTRS
    590 _InterlockedAnd_rel(long volatile *_Value, long _Mask) {
    591   return __atomic_fetch_and(_Value, _Mask, __ATOMIC_RELEASE);
    592 }
    593 static __inline__ __int64 __DEFAULT_FN_ATTRS
    594 _InterlockedAnd64_acq(__int64 volatile *_Value, __int64 _Mask) {
    595   return __atomic_fetch_and(_Value, _Mask, __ATOMIC_ACQUIRE);
    596 }
    597 static __inline__ __int64 __DEFAULT_FN_ATTRS
    598 _InterlockedAnd64_nf(__int64 volatile *_Value, __int64 _Mask) {
    599   return __atomic_fetch_and(_Value, _Mask, __ATOMIC_RELAXED);
    600 }
    601 static __inline__ __int64 __DEFAULT_FN_ATTRS
    602 _InterlockedAnd64_rel(__int64 volatile *_Value, __int64 _Mask) {
    603   return __atomic_fetch_and(_Value, _Mask, __ATOMIC_RELEASE);
    604 }
    605 #endif
    606 /*----------------------------------------------------------------------------*\
    607 |* Interlocked Or
    608 \*----------------------------------------------------------------------------*/
    609 #if defined(__arm__) || defined(__aarch64__)
    610 static __inline__ char __DEFAULT_FN_ATTRS
    611 _InterlockedOr8_acq(char volatile *_Value, char _Mask) {
    612   return __atomic_fetch_or(_Value, _Mask, __ATOMIC_ACQUIRE);
    613 }
    614 static __inline__ char __DEFAULT_FN_ATTRS
    615 _InterlockedOr8_nf(char volatile *_Value, char _Mask) {
    616   return __atomic_fetch_or(_Value, _Mask, __ATOMIC_RELAXED);
    617 }
    618 static __inline__ char __DEFAULT_FN_ATTRS
    619 _InterlockedOr8_rel(char volatile *_Value, char _Mask) {
    620   return __atomic_fetch_or(_Value, _Mask, __ATOMIC_RELEASE);
    621 }
    622 static __inline__ short __DEFAULT_FN_ATTRS
    623 _InterlockedOr16_acq(short volatile *_Value, short _Mask) {
    624   return __atomic_fetch_or(_Value, _Mask, __ATOMIC_ACQUIRE);
    625 }
    626 static __inline__ short __DEFAULT_FN_ATTRS
    627 _InterlockedOr16_nf(short volatile *_Value, short _Mask) {
    628   return __atomic_fetch_or(_Value, _Mask, __ATOMIC_RELAXED);
    629 }
    630 static __inline__ short __DEFAULT_FN_ATTRS
    631 _InterlockedOr16_rel(short volatile *_Value, short _Mask) {
    632   return __atomic_fetch_or(_Value, _Mask, __ATOMIC_RELEASE);
    633 }
    634 static __inline__ long __DEFAULT_FN_ATTRS
    635 _InterlockedOr_acq(long volatile *_Value, long _Mask) {
    636   return __atomic_fetch_or(_Value, _Mask, __ATOMIC_ACQUIRE);
    637 }
    638 static __inline__ long __DEFAULT_FN_ATTRS
    639 _InterlockedOr_nf(long volatile *_Value, long _Mask) {
    640   return __atomic_fetch_or(_Value, _Mask, __ATOMIC_RELAXED);
    641 }
    642 static __inline__ long __DEFAULT_FN_ATTRS
    643 _InterlockedOr_rel(long volatile *_Value, long _Mask) {
    644   return __atomic_fetch_or(_Value, _Mask, __ATOMIC_RELEASE);
    645 }
    646 static __inline__ __int64 __DEFAULT_FN_ATTRS
    647 _InterlockedOr64_acq(__int64 volatile *_Value, __int64 _Mask) {
    648   return __atomic_fetch_or(_Value, _Mask, __ATOMIC_ACQUIRE);
    649 }
    650 static __inline__ __int64 __DEFAULT_FN_ATTRS
    651 _InterlockedOr64_nf(__int64 volatile *_Value, __int64 _Mask) {
    652   return __atomic_fetch_or(_Value, _Mask, __ATOMIC_RELAXED);
    653 }
    654 static __inline__ __int64 __DEFAULT_FN_ATTRS
    655 _InterlockedOr64_rel(__int64 volatile *_Value, __int64 _Mask) {
    656   return __atomic_fetch_or(_Value, _Mask, __ATOMIC_RELEASE);
    657 }
    658 #endif
    659 /*----------------------------------------------------------------------------*\
    660 |* Interlocked Xor
    661 \*----------------------------------------------------------------------------*/
    662 #if defined(__arm__) || defined(__aarch64__)
    663 static __inline__ char __DEFAULT_FN_ATTRS
    664 _InterlockedXor8_acq(char volatile *_Value, char _Mask) {
    665   return __atomic_fetch_xor(_Value, _Mask, __ATOMIC_ACQUIRE);
    666 }
    667 static __inline__ char __DEFAULT_FN_ATTRS
    668 _InterlockedXor8_nf(char volatile *_Value, char _Mask) {
    669   return __atomic_fetch_xor(_Value, _Mask, __ATOMIC_RELAXED);
    670 }
    671 static __inline__ char __DEFAULT_FN_ATTRS
    672 _InterlockedXor8_rel(char volatile *_Value, char _Mask) {
    673   return __atomic_fetch_xor(_Value, _Mask, __ATOMIC_RELEASE);
    674 }
    675 static __inline__ short __DEFAULT_FN_ATTRS
    676 _InterlockedXor16_acq(short volatile *_Value, short _Mask) {
    677   return __atomic_fetch_xor(_Value, _Mask, __ATOMIC_ACQUIRE);
    678 }
    679 static __inline__ short __DEFAULT_FN_ATTRS
    680 _InterlockedXor16_nf(short volatile *_Value, short _Mask) {
    681   return __atomic_fetch_xor(_Value, _Mask, __ATOMIC_RELAXED);
    682 }
    683 static __inline__ short __DEFAULT_FN_ATTRS
    684 _InterlockedXor16_rel(short volatile *_Value, short _Mask) {
    685   return __atomic_fetch_xor(_Value, _Mask, __ATOMIC_RELEASE);
    686 }
    687 static __inline__ long __DEFAULT_FN_ATTRS
    688 _InterlockedXor_acq(long volatile *_Value, long _Mask) {
    689   return __atomic_fetch_xor(_Value, _Mask, __ATOMIC_ACQUIRE);
    690 }
    691 static __inline__ long __DEFAULT_FN_ATTRS
    692 _InterlockedXor_nf(long volatile *_Value, long _Mask) {
    693   return __atomic_fetch_xor(_Value, _Mask, __ATOMIC_RELAXED);
    694 }
    695 static __inline__ long __DEFAULT_FN_ATTRS
    696 _InterlockedXor_rel(long volatile *_Value, long _Mask) {
    697   return __atomic_fetch_xor(_Value, _Mask, __ATOMIC_RELEASE);
    698 }
    699 static __inline__ __int64 __DEFAULT_FN_ATTRS
    700 _InterlockedXor64_acq(__int64 volatile *_Value, __int64 _Mask) {
    701   return __atomic_fetch_xor(_Value, _Mask, __ATOMIC_ACQUIRE);
    702 }
    703 static __inline__ __int64 __DEFAULT_FN_ATTRS
    704 _InterlockedXor64_nf(__int64 volatile *_Value, __int64 _Mask) {
    705   return __atomic_fetch_xor(_Value, _Mask, __ATOMIC_RELAXED);
    706 }
    707 static __inline__ __int64 __DEFAULT_FN_ATTRS
    708 _InterlockedXor64_rel(__int64 volatile *_Value, __int64 _Mask) {
    709   return __atomic_fetch_xor(_Value, _Mask, __ATOMIC_RELEASE);
    710 }
    711 #endif
    712 /*----------------------------------------------------------------------------*\
    713 |* Interlocked Exchange
    714 \*----------------------------------------------------------------------------*/
    715 #if defined(__arm__) || defined(__aarch64__)
    716 static __inline__ char __DEFAULT_FN_ATTRS
    717 _InterlockedExchange8_acq(char volatile *_Target, char _Value) {
    718   __atomic_exchange(_Target, &_Value, &_Value, __ATOMIC_ACQUIRE);
    719   return _Value;
    720 }
    721 static __inline__ char __DEFAULT_FN_ATTRS
    722 _InterlockedExchange8_nf(char volatile *_Target, char _Value) {
    723   __atomic_exchange(_Target, &_Value, &_Value, __ATOMIC_RELAXED);
    724   return _Value;
    725 }
    726 static __inline__ char __DEFAULT_FN_ATTRS
    727 _InterlockedExchange8_rel(char volatile *_Target, char _Value) {
    728   __atomic_exchange(_Target, &_Value, &_Value, __ATOMIC_RELEASE);
    729   return _Value;
    730 }
    731 static __inline__ short __DEFAULT_FN_ATTRS
    732 _InterlockedExchange16_acq(short volatile *_Target, short _Value) {
    733   __atomic_exchange(_Target, &_Value, &_Value, __ATOMIC_ACQUIRE);
    734   return _Value;
    735 }
    736 static __inline__ short __DEFAULT_FN_ATTRS
    737 _InterlockedExchange16_nf(short volatile *_Target, short _Value) {
    738   __atomic_exchange(_Target, &_Value, &_Value, __ATOMIC_RELAXED);
    739   return _Value;
    740 }
    741 static __inline__ short __DEFAULT_FN_ATTRS
    742 _InterlockedExchange16_rel(short volatile *_Target, short _Value) {
    743   __atomic_exchange(_Target, &_Value, &_Value, __ATOMIC_RELEASE);
    744   return _Value;
    745 }
    746 static __inline__ long __DEFAULT_FN_ATTRS
    747 _InterlockedExchange_acq(long volatile *_Target, long _Value) {
    748   __atomic_exchange(_Target, &_Value, &_Value, __ATOMIC_ACQUIRE);
    749   return _Value;
    750 }
    751 static __inline__ long __DEFAULT_FN_ATTRS
    752 _InterlockedExchange_nf(long volatile *_Target, long _Value) {
    753   __atomic_exchange(_Target, &_Value, &_Value, __ATOMIC_RELAXED);
    754   return _Value;
    755 }
    756 static __inline__ long __DEFAULT_FN_ATTRS
    757 _InterlockedExchange_rel(long volatile *_Target, long _Value) {
    758   __atomic_exchange(_Target, &_Value, &_Value, __ATOMIC_RELEASE);
    759   return _Value;
    760 }
    761 static __inline__ __int64 __DEFAULT_FN_ATTRS
    762 _InterlockedExchange64_acq(__int64 volatile *_Target, __int64 _Value) {
    763   __atomic_exchange(_Target, &_Value, &_Value, __ATOMIC_ACQUIRE);
    764   return _Value;
    765 }
    766 static __inline__ __int64 __DEFAULT_FN_ATTRS
    767 _InterlockedExchange64_nf(__int64 volatile *_Target, __int64 _Value) {
    768   __atomic_exchange(_Target, &_Value, &_Value, __ATOMIC_RELAXED);
    769   return _Value;
    770 }
    771 static __inline__ __int64 __DEFAULT_FN_ATTRS
    772 _InterlockedExchange64_rel(__int64 volatile *_Target, __int64 _Value) {
    773   __atomic_exchange(_Target, &_Value, &_Value, __ATOMIC_RELEASE);
    774   return _Value;
    775 }
    776 #endif
    777 /*----------------------------------------------------------------------------*\
    778 |* Interlocked Compare Exchange
    779 \*----------------------------------------------------------------------------*/
    780 #if defined(__arm__) || defined(__aarch64__)
    781 static __inline__ char __DEFAULT_FN_ATTRS
    782 _InterlockedCompareExchange8_acq(char volatile *_Destination,
    783                              char _Exchange, char _Comparand) {
    784   __atomic_compare_exchange(_Destination, &_Comparand, &_Exchange, 0,
    785                             __ATOMIC_SEQ_CST, __ATOMIC_ACQUIRE);
    786   return _Comparand;
    787 }
    788 static __inline__ char __DEFAULT_FN_ATTRS
    789 _InterlockedCompareExchange8_nf(char volatile *_Destination,
    790                              char _Exchange, char _Comparand) {
    791   __atomic_compare_exchange(_Destination, &_Comparand, &_Exchange, 0,
    792                             __ATOMIC_SEQ_CST, __ATOMIC_RELAXED);
    793   return _Comparand;
    794 }
    795 static __inline__ char __DEFAULT_FN_ATTRS
    796 _InterlockedCompareExchange8_rel(char volatile *_Destination,
    797                              char _Exchange, char _Comparand) {
    798   __atomic_compare_exchange(_Destination, &_Comparand, &_Exchange, 0,
    799                             __ATOMIC_SEQ_CST, __ATOMIC_RELEASE);
    800   return _Comparand;
    801 }
    802 static __inline__ short __DEFAULT_FN_ATTRS
    803 _InterlockedCompareExchange16_acq(short volatile *_Destination,
    804                               short _Exchange, short _Comparand) {
    805   __atomic_compare_exchange(_Destination, &_Comparand, &_Exchange, 0,
    806                             __ATOMIC_SEQ_CST, __ATOMIC_ACQUIRE);
    807   return _Comparand;
    808 }
    809 static __inline__ short __DEFAULT_FN_ATTRS
    810 _InterlockedCompareExchange16_nf(short volatile *_Destination,
    811                               short _Exchange, short _Comparand) {
    812   __atomic_compare_exchange(_Destination, &_Comparand, &_Exchange, 0,
    813                             __ATOMIC_SEQ_CST, __ATOMIC_RELAXED);
    814   return _Comparand;
    815 }
    816 static __inline__ short __DEFAULT_FN_ATTRS
    817 _InterlockedCompareExchange16_rel(short volatile *_Destination,
    818                               short _Exchange, short _Comparand) {
    819   __atomic_compare_exchange(_Destination, &_Comparand, &_Exchange, 0,
    820                             __ATOMIC_SEQ_CST, __ATOMIC_RELEASE);
    821   return _Comparand;
    822 }
    823 static __inline__ long __DEFAULT_FN_ATTRS
    824 _InterlockedCompareExchange_acq(long volatile *_Destination,
    825                               long _Exchange, long _Comparand) {
    826   __atomic_compare_exchange(_Destination, &_Comparand, &_Exchange, 0,
    827                             __ATOMIC_SEQ_CST, __ATOMIC_ACQUIRE);
    828   return _Comparand;
    829 }
    830 static __inline__ long __DEFAULT_FN_ATTRS
    831 _InterlockedCompareExchange_nf(long volatile *_Destination,
    832                               long _Exchange, long _Comparand) {
    833   __atomic_compare_exchange(_Destination, &_Comparand, &_Exchange, 0,
    834                             __ATOMIC_SEQ_CST, __ATOMIC_RELAXED);
    835   return _Comparand;
    836 }
    837 static __inline__ short __DEFAULT_FN_ATTRS
    838 _InterlockedCompareExchange_rel(long volatile *_Destination,
    839                               long _Exchange, long _Comparand) {
    840   __atomic_compare_exchange(_Destination, &_Comparand, &_Exchange, 0,
    841                             __ATOMIC_SEQ_CST, __ATOMIC_RELEASE);
    842   return _Comparand;
    843 }
    844 static __inline__ __int64 __DEFAULT_FN_ATTRS
    845 _InterlockedCompareExchange64_acq(__int64 volatile *_Destination,
    846                               __int64 _Exchange, __int64 _Comparand) {
    847   __atomic_compare_exchange(_Destination, &_Comparand, &_Exchange, 0,
    848                             __ATOMIC_SEQ_CST, __ATOMIC_ACQUIRE);
    849   return _Comparand;
    850 }
    851 static __inline__ __int64 __DEFAULT_FN_ATTRS
    852 _InterlockedCompareExchange64_nf(__int64 volatile *_Destination,
    853                               __int64 _Exchange, __int64 _Comparand) {
    854   __atomic_compare_exchange(_Destination, &_Comparand, &_Exchange, 0,
    855                             __ATOMIC_SEQ_CST, __ATOMIC_RELAXED);
    856   return _Comparand;
    857 }
    858 static __inline__ __int64 __DEFAULT_FN_ATTRS
    859 _InterlockedCompareExchange64_rel(__int64 volatile *_Destination,
    860                               __int64 _Exchange, __int64 _Comparand) {
    861   __atomic_compare_exchange(_Destination, &_Comparand, &_Exchange, 0,
    862                             __ATOMIC_SEQ_CST, __ATOMIC_RELEASE);
    863   return _Comparand;
    864 }
    865 #endif
    866 
    867 /*----------------------------------------------------------------------------*\
    868 |* movs, stos
    869 \*----------------------------------------------------------------------------*/
    870 #if defined(__i386__) || defined(__x86_64__)
    871 static __inline__ void __DEFAULT_FN_ATTRS
    872 __movsb(unsigned char *__dst, unsigned char const *__src, size_t __n) {
    873   __asm__("rep movsb" : : "D"(__dst), "S"(__src), "c"(__n));
    874 }
    875 static __inline__ void __DEFAULT_FN_ATTRS
    876 __movsd(unsigned long *__dst, unsigned long const *__src, size_t __n) {
    877   __asm__("rep movsl" : : "D"(__dst), "S"(__src), "c"(__n));
    878 }
    879 static __inline__ void __DEFAULT_FN_ATTRS
    880 __movsw(unsigned short *__dst, unsigned short const *__src, size_t __n) {
    881   __asm__("rep movsw" : : "D"(__dst), "S"(__src), "c"(__n));
    882 }
    883 static __inline__ void __DEFAULT_FN_ATTRS
    884 __stosd(unsigned long *__dst, unsigned long __x, size_t __n) {
    885   __asm__("rep stosl" : : "D"(__dst), "a"(__x), "c"(__n));
    886 }
    887 static __inline__ void __DEFAULT_FN_ATTRS
    888 __stosw(unsigned short *__dst, unsigned short __x, size_t __n) {
    889   __asm__("rep stosw" : : "D"(__dst), "a"(__x), "c"(__n));
    890 }
    891 #endif
    892 #ifdef __x86_64__
    893 static __inline__ void __DEFAULT_FN_ATTRS
    894 __movsq(unsigned long long *__dst, unsigned long long const *__src, size_t __n) {
    895   __asm__("rep movsq" : : "D"(__dst), "S"(__src), "c"(__n));
    896 }
    897 static __inline__ void __DEFAULT_FN_ATTRS
    898 __stosq(unsigned __int64 *__dst, unsigned __int64 __x, size_t __n) {
    899   __asm__("rep stosq" : : "D"(__dst), "a"(__x), "c"(__n));
    900 }
    901 #endif
    902 
    903 /*----------------------------------------------------------------------------*\
    904 |* Misc
    905 \*----------------------------------------------------------------------------*/
    906 #if defined(__i386__) || defined(__x86_64__)
    907 static __inline__ void __DEFAULT_FN_ATTRS
    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 __DEFAULT_FN_ATTRS
    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 __DEFAULT_FN_ATTRS
    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 __DEFAULT_FN_ATTRS
    924 __halt(void) {
    925   __asm__ volatile ("hlt");
    926 }
    927 static __inline__ void __DEFAULT_FN_ATTRS
    928 __nop(void) {
    929   __asm__ volatile ("nop");
    930 }
    931 #endif
    932 
    933 /*----------------------------------------------------------------------------*\
    934 |* Privileged intrinsics
    935 \*----------------------------------------------------------------------------*/
    936 #if defined(__i386__) || defined(__x86_64__)
    937 static __inline__ unsigned __int64 __DEFAULT_FN_ATTRS
    938 __readmsr(unsigned long __register) {
    939   // Loads the contents of a 64-bit model specific register (MSR) specified in
    940   // the ECX register into registers EDX:EAX. The EDX register is loaded with
    941   // the high-order 32 bits of the MSR and the EAX register is loaded with the
    942   // low-order 32 bits. If less than 64 bits are implemented in the MSR being
    943   // read, the values returned to EDX:EAX in unimplemented bit locations are
    944   // undefined.
    945   unsigned long __edx;
    946   unsigned long __eax;
    947   __asm__ ("rdmsr" : "=d"(__edx), "=a"(__eax) : "c"(__register));
    948   return (((unsigned __int64)__edx) << 32) | (unsigned __int64)__eax;
    949 }
    950 
    951 static __inline__ unsigned long __DEFAULT_FN_ATTRS
    952 __readcr3(void) {
    953   unsigned long __cr3_val;
    954   __asm__ __volatile__ ("mov %%cr3, %0" : "=q"(__cr3_val) : : "memory");
    955   return __cr3_val;
    956 }
    957 
    958 static __inline__ void __DEFAULT_FN_ATTRS
    959 __writecr3(unsigned int __cr3_val) {
    960   __asm__ ("mov %0, %%cr3" : : "q"(__cr3_val) : "memory");
    961 }
    962 #endif
    963 
    964 #ifdef __cplusplus
    965 }
    966 #endif
    967 
    968 #undef __DEFAULT_FN_ATTRS
    969 
    970 #endif /* __INTRIN_H */
    971 #endif /* _MSC_VER */
    972