Path: blob/master/tools/android-sdk/renderscript/clang-include/intrin.h
496 views
/* ===-------- intrin.h ---------------------------------------------------===1*2* Permission is hereby granted, free of charge, to any person obtaining a copy3* of this software and associated documentation files (the "Software"), to deal4* in the Software without restriction, including without limitation the rights5* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell6* copies of the Software, and to permit persons to whom the Software is7* furnished to do so, subject to the following conditions:8*9* The above copyright notice and this permission notice shall be included in10* all copies or substantial portions of the Software.11*12* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR13* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,14* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE15* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER16* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,17* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN18* THE SOFTWARE.19*20*===-----------------------------------------------------------------------===21*/2223/* Only include this if we're compiling for the windows platform. */24#ifndef _MSC_VER25#include_next <intrin.h>26#else2728#ifndef __INTRIN_H29#define __INTRIN_H3031/* First include the standard intrinsics. */32#if defined(__i386__) || defined(__x86_64__)33#include <x86intrin.h>34#endif3536/* For the definition of jmp_buf. */37#if __STDC_HOSTED__38#include <setjmp.h>39#endif4041/* Define the default attributes for the functions in this file. */42#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__))4344#ifdef __cplusplus45extern "C" {46#endif4748#if defined(__MMX__)49/* And the random ones that aren't in those files. */50__m64 _m_from_float(float);51float _m_to_float(__m64);52#endif5354/* Other assorted instruction intrinsics. */55void __addfsbyte(unsigned long, unsigned char);56void __addfsdword(unsigned long, unsigned long);57void __addfsword(unsigned long, unsigned short);58void __code_seg(const char *);59static __inline__60void __cpuid(int[4], int);61static __inline__62void __cpuidex(int[4], int, int);63void __debugbreak(void);64__int64 __emul(int, int);65unsigned __int64 __emulu(unsigned int, unsigned int);66void __cdecl __fastfail(unsigned int);67unsigned int __getcallerseflags(void);68static __inline__69void __halt(void);70unsigned char __inbyte(unsigned short);71void __inbytestring(unsigned short, unsigned char *, unsigned long);72void __incfsbyte(unsigned long);73void __incfsdword(unsigned long);74void __incfsword(unsigned long);75unsigned long __indword(unsigned short);76void __indwordstring(unsigned short, unsigned long *, unsigned long);77void __int2c(void);78void __invlpg(void *);79unsigned short __inword(unsigned short);80void __inwordstring(unsigned short, unsigned short *, unsigned long);81void __lidt(void *);82unsigned __int64 __ll_lshift(unsigned __int64, int);83__int64 __ll_rshift(__int64, int);84void __llwpcb(void *);85unsigned char __lwpins32(unsigned int, unsigned int, unsigned int);86void __lwpval32(unsigned int, unsigned int, unsigned int);87unsigned int __lzcnt(unsigned int);88unsigned short __lzcnt16(unsigned short);89static __inline__90void __movsb(unsigned char *, unsigned char const *, size_t);91static __inline__92void __movsd(unsigned long *, unsigned long const *, size_t);93static __inline__94void __movsw(unsigned short *, unsigned short const *, size_t);95void __nop(void);96void __nvreg_restore_fence(void);97void __nvreg_save_fence(void);98void __outbyte(unsigned short, unsigned char);99void __outbytestring(unsigned short, unsigned char *, unsigned long);100void __outdword(unsigned short, unsigned long);101void __outdwordstring(unsigned short, unsigned long *, unsigned long);102void __outword(unsigned short, unsigned short);103void __outwordstring(unsigned short, unsigned short *, unsigned long);104static __inline__105unsigned int __popcnt(unsigned int);106static __inline__107unsigned short __popcnt16(unsigned short);108unsigned long __readcr0(void);109unsigned long __readcr2(void);110static __inline__111unsigned long __readcr3(void);112unsigned long __readcr4(void);113unsigned long __readcr8(void);114unsigned int __readdr(unsigned int);115#ifdef __i386__116static __inline__117unsigned char __readfsbyte(unsigned long);118static __inline__119unsigned long __readfsdword(unsigned long);120static __inline__121unsigned __int64 __readfsqword(unsigned long);122static __inline__123unsigned short __readfsword(unsigned long);124#endif125static __inline__126unsigned __int64 __readmsr(unsigned long);127unsigned __int64 __readpmc(unsigned long);128unsigned long __segmentlimit(unsigned long);129void __sidt(void *);130void *__slwpcb(void);131static __inline__132void __stosb(unsigned char *, unsigned char, size_t);133static __inline__134void __stosd(unsigned long *, unsigned long, size_t);135static __inline__136void __stosw(unsigned short *, unsigned short, size_t);137void __svm_clgi(void);138void __svm_invlpga(void *, int);139void __svm_skinit(int);140void __svm_stgi(void);141void __svm_vmload(size_t);142void __svm_vmrun(size_t);143void __svm_vmsave(size_t);144void __ud2(void);145unsigned __int64 __ull_rshift(unsigned __int64, int);146void __vmx_off(void);147void __vmx_vmptrst(unsigned __int64 *);148void __wbinvd(void);149void __writecr0(unsigned int);150static __inline__151void __writecr3(unsigned int);152void __writecr4(unsigned int);153void __writecr8(unsigned int);154void __writedr(unsigned int, unsigned int);155void __writefsbyte(unsigned long, unsigned char);156void __writefsdword(unsigned long, unsigned long);157void __writefsqword(unsigned long, unsigned __int64);158void __writefsword(unsigned long, unsigned short);159void __writemsr(unsigned long, unsigned __int64);160static __inline__161void *_AddressOfReturnAddress(void);162static __inline__163unsigned char _BitScanForward(unsigned long *_Index, unsigned long _Mask);164static __inline__165unsigned char _BitScanReverse(unsigned long *_Index, unsigned long _Mask);166static __inline__167unsigned char _bittest(long const *, long);168static __inline__169unsigned char _bittestandcomplement(long *, long);170static __inline__171unsigned char _bittestandreset(long *, long);172static __inline__173unsigned char _bittestandset(long *, long);174unsigned __int64 __cdecl _byteswap_uint64(unsigned __int64);175unsigned long __cdecl _byteswap_ulong(unsigned long);176unsigned short __cdecl _byteswap_ushort(unsigned short);177void __cdecl _disable(void);178void __cdecl _enable(void);179long _InterlockedAddLargeStatistic(__int64 volatile *_Addend, long _Value);180static __inline__181long _InterlockedAnd(long volatile *_Value, long _Mask);182static __inline__183short _InterlockedAnd16(short volatile *_Value, short _Mask);184static __inline__185char _InterlockedAnd8(char volatile *_Value, char _Mask);186unsigned char _interlockedbittestandreset(long volatile *, long);187static __inline__188unsigned char _interlockedbittestandset(long volatile *, long);189static __inline__190long __cdecl _InterlockedCompareExchange(long volatile *_Destination,191long _Exchange, long _Comparand);192long _InterlockedCompareExchange_HLEAcquire(long volatile *, long, long);193long _InterlockedCompareExchange_HLERelease(long volatile *, long, long);194static __inline__195short _InterlockedCompareExchange16(short volatile *_Destination,196short _Exchange, short _Comparand);197static __inline__198__int64 _InterlockedCompareExchange64(__int64 volatile *_Destination,199__int64 _Exchange, __int64 _Comparand);200__int64 _InterlockedcompareExchange64_HLEAcquire(__int64 volatile *, __int64,201__int64);202__int64 _InterlockedCompareExchange64_HLERelease(__int64 volatile *, __int64,203__int64);204static __inline__205char _InterlockedCompareExchange8(char volatile *_Destination, char _Exchange,206char _Comparand);207void *_InterlockedCompareExchangePointer_HLEAcquire(void *volatile *, void *,208void *);209void *_InterlockedCompareExchangePointer_HLERelease(void *volatile *, void *,210void *);211static __inline__212long __cdecl _InterlockedDecrement(long volatile *_Addend);213static __inline__214short _InterlockedDecrement16(short volatile *_Addend);215long _InterlockedExchange(long volatile *_Target, long _Value);216static __inline__217short _InterlockedExchange16(short volatile *_Target, short _Value);218static __inline__219char _InterlockedExchange8(char volatile *_Target, char _Value);220static __inline__221long __cdecl _InterlockedExchangeAdd(long volatile *_Addend, long _Value);222long _InterlockedExchangeAdd_HLEAcquire(long volatile *, long);223long _InterlockedExchangeAdd_HLERelease(long volatile *, long);224static __inline__225short _InterlockedExchangeAdd16(short volatile *_Addend, short _Value);226__int64 _InterlockedExchangeAdd64_HLEAcquire(__int64 volatile *, __int64);227__int64 _InterlockedExchangeAdd64_HLERelease(__int64 volatile *, __int64);228static __inline__229char _InterlockedExchangeAdd8(char volatile *_Addend, char _Value);230static __inline__231long __cdecl _InterlockedIncrement(long volatile *_Addend);232static __inline__233short _InterlockedIncrement16(short volatile *_Addend);234static __inline__235long _InterlockedOr(long volatile *_Value, long _Mask);236static __inline__237short _InterlockedOr16(short volatile *_Value, short _Mask);238static __inline__239char _InterlockedOr8(char volatile *_Value, char _Mask);240static __inline__241long _InterlockedXor(long volatile *_Value, long _Mask);242static __inline__243short _InterlockedXor16(short volatile *_Value, short _Mask);244static __inline__245char _InterlockedXor8(char volatile *_Value, char _Mask);246void __cdecl _invpcid(unsigned int, void *);247static __inline__248unsigned long __cdecl _lrotl(unsigned long, int);249static __inline__250unsigned long __cdecl _lrotr(unsigned long, int);251static __inline__252void _ReadBarrier(void);253static __inline__254void _ReadWriteBarrier(void);255static __inline__256void *_ReturnAddress(void);257unsigned int _rorx_u32(unsigned int, const unsigned int);258static __inline__259unsigned int __cdecl _rotl(unsigned int _Value, int _Shift);260static __inline__261unsigned short _rotl16(unsigned short _Value, unsigned char _Shift);262static __inline__263unsigned __int64 __cdecl _rotl64(unsigned __int64 _Value, int _Shift);264static __inline__265unsigned char _rotl8(unsigned char _Value, unsigned char _Shift);266static __inline__267unsigned int __cdecl _rotr(unsigned int _Value, int _Shift);268static __inline__269unsigned short _rotr16(unsigned short _Value, unsigned char _Shift);270static __inline__271unsigned __int64 __cdecl _rotr64(unsigned __int64 _Value, int _Shift);272static __inline__273unsigned char _rotr8(unsigned char _Value, unsigned char _Shift);274int _sarx_i32(int, unsigned int);275#if __STDC_HOSTED__276int __cdecl _setjmp(jmp_buf);277#endif278unsigned int _shlx_u32(unsigned int, unsigned int);279unsigned int _shrx_u32(unsigned int, unsigned int);280void _Store_HLERelease(long volatile *, long);281void _Store64_HLERelease(__int64 volatile *, __int64);282void _StorePointer_HLERelease(void *volatile *, void *);283static __inline__284void _WriteBarrier(void);285unsigned __int32 xbegin(void);286void _xend(void);287static __inline__288#define _XCR_XFEATURE_ENABLED_MASK 0289unsigned __int64 __cdecl _xgetbv(unsigned int);290void __cdecl _xsetbv(unsigned int, unsigned __int64);291292/* These additional intrinsics are turned on in x64/amd64/x86_64 mode. */293#ifdef __x86_64__294void __addgsbyte(unsigned long, unsigned char);295void __addgsdword(unsigned long, unsigned long);296void __addgsqword(unsigned long, unsigned __int64);297void __addgsword(unsigned long, unsigned short);298static __inline__299void __faststorefence(void);300void __incgsbyte(unsigned long);301void __incgsdword(unsigned long);302void __incgsqword(unsigned long);303void __incgsword(unsigned long);304unsigned char __lwpins64(unsigned __int64, unsigned int, unsigned int);305void __lwpval64(unsigned __int64, unsigned int, unsigned int);306unsigned __int64 __lzcnt64(unsigned __int64);307static __inline__308void __movsq(unsigned long long *, unsigned long long const *, size_t);309__int64 __mulh(__int64, __int64);310static __inline__311unsigned __int64 __popcnt64(unsigned __int64);312static __inline__313unsigned char __readgsbyte(unsigned long);314static __inline__315unsigned long __readgsdword(unsigned long);316static __inline__317unsigned __int64 __readgsqword(unsigned long);318unsigned short __readgsword(unsigned long);319unsigned __int64 __shiftleft128(unsigned __int64 _LowPart,320unsigned __int64 _HighPart,321unsigned char _Shift);322unsigned __int64 __shiftright128(unsigned __int64 _LowPart,323unsigned __int64 _HighPart,324unsigned char _Shift);325static __inline__326void __stosq(unsigned __int64 *, unsigned __int64, size_t);327unsigned char __vmx_on(unsigned __int64 *);328unsigned char __vmx_vmclear(unsigned __int64 *);329unsigned char __vmx_vmlaunch(void);330unsigned char __vmx_vmptrld(unsigned __int64 *);331unsigned char __vmx_vmread(size_t, size_t *);332unsigned char __vmx_vmresume(void);333unsigned char __vmx_vmwrite(size_t, size_t);334void __writegsbyte(unsigned long, unsigned char);335void __writegsdword(unsigned long, unsigned long);336void __writegsqword(unsigned long, unsigned __int64);337void __writegsword(unsigned long, unsigned short);338static __inline__339unsigned char _BitScanForward64(unsigned long *_Index, unsigned __int64 _Mask);340static __inline__341unsigned char _BitScanReverse64(unsigned long *_Index, unsigned __int64 _Mask);342static __inline__343unsigned char _bittest64(__int64 const *, __int64);344static __inline__345unsigned char _bittestandcomplement64(__int64 *, __int64);346static __inline__347unsigned char _bittestandreset64(__int64 *, __int64);348static __inline__349unsigned char _bittestandset64(__int64 *, __int64);350unsigned __int64 __cdecl _byteswap_uint64(unsigned __int64);351long _InterlockedAnd_np(long volatile *_Value, long _Mask);352short _InterlockedAnd16_np(short volatile *_Value, short _Mask);353__int64 _InterlockedAnd64_np(__int64 volatile *_Value, __int64 _Mask);354char _InterlockedAnd8_np(char volatile *_Value, char _Mask);355unsigned char _interlockedbittestandreset64(__int64 volatile *, __int64);356static __inline__357unsigned char _interlockedbittestandset64(__int64 volatile *, __int64);358long _InterlockedCompareExchange_np(long volatile *_Destination, long _Exchange,359long _Comparand);360unsigned char _InterlockedCompareExchange128(__int64 volatile *_Destination,361__int64 _ExchangeHigh,362__int64 _ExchangeLow,363__int64 *_CompareandResult);364unsigned char _InterlockedCompareExchange128_np(__int64 volatile *_Destination,365__int64 _ExchangeHigh,366__int64 _ExchangeLow,367__int64 *_ComparandResult);368short _InterlockedCompareExchange16_np(short volatile *_Destination,369short _Exchange, short _Comparand);370__int64 _InterlockedCompareExchange64_HLEAcquire(__int64 volatile *, __int64,371__int64);372__int64 _InterlockedCompareExchange64_HLERelease(__int64 volatile *, __int64,373__int64);374__int64 _InterlockedCompareExchange64_np(__int64 volatile *_Destination,375__int64 _Exchange, __int64 _Comparand);376void *_InterlockedCompareExchangePointer(void *volatile *_Destination,377void *_Exchange, void *_Comparand);378void *_InterlockedCompareExchangePointer_np(void *volatile *_Destination,379void *_Exchange, void *_Comparand);380static __inline__381__int64 _InterlockedDecrement64(__int64 volatile *_Addend);382static __inline__383__int64 _InterlockedExchange64(__int64 volatile *_Target, __int64 _Value);384static __inline__385__int64 _InterlockedExchangeAdd64(__int64 volatile *_Addend, __int64 _Value);386void *_InterlockedExchangePointer(void *volatile *_Target, void *_Value);387static __inline__388__int64 _InterlockedIncrement64(__int64 volatile *_Addend);389long _InterlockedOr_np(long volatile *_Value, long _Mask);390short _InterlockedOr16_np(short volatile *_Value, short _Mask);391static __inline__392__int64 _InterlockedOr64(__int64 volatile *_Value, __int64 _Mask);393__int64 _InterlockedOr64_np(__int64 volatile *_Value, __int64 _Mask);394char _InterlockedOr8_np(char volatile *_Value, char _Mask);395long _InterlockedXor_np(long volatile *_Value, long _Mask);396short _InterlockedXor16_np(short volatile *_Value, short _Mask);397static __inline__398__int64 _InterlockedXor64(__int64 volatile *_Value, __int64 _Mask);399__int64 _InterlockedXor64_np(__int64 volatile *_Value, __int64 _Mask);400char _InterlockedXor8_np(char volatile *_Value, char _Mask);401static __inline__402__int64 _mul128(__int64 _Multiplier, __int64 _Multiplicand,403__int64 *_HighProduct);404unsigned __int64 _rorx_u64(unsigned __int64, const unsigned int);405__int64 _sarx_i64(__int64, unsigned int);406#if __STDC_HOSTED__407int __cdecl _setjmpex(jmp_buf);408#endif409unsigned __int64 _shlx_u64(unsigned __int64, unsigned int);410unsigned __int64 _shrx_u64(unsigned __int64, unsigned int);411/*412* Multiply two 64-bit integers and obtain a 64-bit result.413* The low-half is returned directly and the high half is in an out parameter.414*/415static __inline__ unsigned __int64 __DEFAULT_FN_ATTRS416_umul128(unsigned __int64 _Multiplier, unsigned __int64 _Multiplicand,417unsigned __int64 *_HighProduct) {418unsigned __int128 _FullProduct =419(unsigned __int128)_Multiplier * (unsigned __int128)_Multiplicand;420*_HighProduct = _FullProduct >> 64;421return _FullProduct;422}423static __inline__ unsigned __int64 __DEFAULT_FN_ATTRS424__umulh(unsigned __int64 _Multiplier, unsigned __int64 _Multiplicand) {425unsigned __int128 _FullProduct =426(unsigned __int128)_Multiplier * (unsigned __int128)_Multiplicand;427return _FullProduct >> 64;428}429430#endif /* __x86_64__ */431432/*----------------------------------------------------------------------------*\433|* Multiplication434\*----------------------------------------------------------------------------*/435static __inline__ __int64 __DEFAULT_FN_ATTRS436__emul(int __in1, int __in2) {437return (__int64)__in1 * (__int64)__in2;438}439static __inline__ unsigned __int64 __DEFAULT_FN_ATTRS440__emulu(unsigned int __in1, unsigned int __in2) {441return (unsigned __int64)__in1 * (unsigned __int64)__in2;442}443/*----------------------------------------------------------------------------*\444|* Bit Twiddling445\*----------------------------------------------------------------------------*/446static __inline__ unsigned char __DEFAULT_FN_ATTRS447_rotl8(unsigned char _Value, unsigned char _Shift) {448_Shift &= 0x7;449return _Shift ? (_Value << _Shift) | (_Value >> (8 - _Shift)) : _Value;450}451static __inline__ unsigned char __DEFAULT_FN_ATTRS452_rotr8(unsigned char _Value, unsigned char _Shift) {453_Shift &= 0x7;454return _Shift ? (_Value >> _Shift) | (_Value << (8 - _Shift)) : _Value;455}456static __inline__ unsigned short __DEFAULT_FN_ATTRS457_rotl16(unsigned short _Value, unsigned char _Shift) {458_Shift &= 0xf;459return _Shift ? (_Value << _Shift) | (_Value >> (16 - _Shift)) : _Value;460}461static __inline__ unsigned short __DEFAULT_FN_ATTRS462_rotr16(unsigned short _Value, unsigned char _Shift) {463_Shift &= 0xf;464return _Shift ? (_Value >> _Shift) | (_Value << (16 - _Shift)) : _Value;465}466static __inline__ unsigned int __DEFAULT_FN_ATTRS467_rotl(unsigned int _Value, int _Shift) {468_Shift &= 0x1f;469return _Shift ? (_Value << _Shift) | (_Value >> (32 - _Shift)) : _Value;470}471static __inline__ unsigned int __DEFAULT_FN_ATTRS472_rotr(unsigned int _Value, int _Shift) {473_Shift &= 0x1f;474return _Shift ? (_Value >> _Shift) | (_Value << (32 - _Shift)) : _Value;475}476static __inline__ unsigned long __DEFAULT_FN_ATTRS477_lrotl(unsigned long _Value, int _Shift) {478_Shift &= 0x1f;479return _Shift ? (_Value << _Shift) | (_Value >> (32 - _Shift)) : _Value;480}481static __inline__ unsigned long __DEFAULT_FN_ATTRS482_lrotr(unsigned long _Value, int _Shift) {483_Shift &= 0x1f;484return _Shift ? (_Value >> _Shift) | (_Value << (32 - _Shift)) : _Value;485}486static487__inline__ unsigned __int64 __DEFAULT_FN_ATTRS488_rotl64(unsigned __int64 _Value, int _Shift) {489_Shift &= 0x3f;490return _Shift ? (_Value << _Shift) | (_Value >> (64 - _Shift)) : _Value;491}492static493__inline__ unsigned __int64 __DEFAULT_FN_ATTRS494_rotr64(unsigned __int64 _Value, int _Shift) {495_Shift &= 0x3f;496return _Shift ? (_Value >> _Shift) | (_Value << (64 - _Shift)) : _Value;497}498/*----------------------------------------------------------------------------*\499|* Bit Counting and Testing500\*----------------------------------------------------------------------------*/501static __inline__ unsigned char __DEFAULT_FN_ATTRS502_BitScanForward(unsigned long *_Index, unsigned long _Mask) {503if (!_Mask)504return 0;505*_Index = __builtin_ctzl(_Mask);506return 1;507}508static __inline__ unsigned char __DEFAULT_FN_ATTRS509_BitScanReverse(unsigned long *_Index, unsigned long _Mask) {510if (!_Mask)511return 0;512*_Index = 31 - __builtin_clzl(_Mask);513return 1;514}515static __inline__ unsigned short __DEFAULT_FN_ATTRS516__popcnt16(unsigned short _Value) {517return __builtin_popcount((int)_Value);518}519static __inline__ unsigned int __DEFAULT_FN_ATTRS520__popcnt(unsigned int _Value) {521return __builtin_popcount(_Value);522}523static __inline__ unsigned char __DEFAULT_FN_ATTRS524_bittest(long const *_BitBase, long _BitPos) {525return (*_BitBase >> _BitPos) & 1;526}527static __inline__ unsigned char __DEFAULT_FN_ATTRS528_bittestandcomplement(long *_BitBase, long _BitPos) {529unsigned char _Res = (*_BitBase >> _BitPos) & 1;530*_BitBase = *_BitBase ^ (1 << _BitPos);531return _Res;532}533static __inline__ unsigned char __DEFAULT_FN_ATTRS534_bittestandreset(long *_BitBase, long _BitPos) {535unsigned char _Res = (*_BitBase >> _BitPos) & 1;536*_BitBase = *_BitBase & ~(1 << _BitPos);537return _Res;538}539static __inline__ unsigned char __DEFAULT_FN_ATTRS540_bittestandset(long *_BitBase, long _BitPos) {541unsigned char _Res = (*_BitBase >> _BitPos) & 1;542*_BitBase = *_BitBase | (1 << _BitPos);543return _Res;544}545static __inline__ unsigned char __DEFAULT_FN_ATTRS546_interlockedbittestandset(long volatile *_BitBase, long _BitPos) {547long _PrevVal = __atomic_fetch_or(_BitBase, 1l << _BitPos, __ATOMIC_SEQ_CST);548return (_PrevVal >> _BitPos) & 1;549}550#ifdef __x86_64__551static __inline__ unsigned char __DEFAULT_FN_ATTRS552_BitScanForward64(unsigned long *_Index, unsigned __int64 _Mask) {553if (!_Mask)554return 0;555*_Index = __builtin_ctzll(_Mask);556return 1;557}558static __inline__ unsigned char __DEFAULT_FN_ATTRS559_BitScanReverse64(unsigned long *_Index, unsigned __int64 _Mask) {560if (!_Mask)561return 0;562*_Index = 63 - __builtin_clzll(_Mask);563return 1;564}565static __inline__566unsigned __int64 __DEFAULT_FN_ATTRS567__popcnt64(unsigned __int64 _Value) {568return __builtin_popcountll(_Value);569}570static __inline__ unsigned char __DEFAULT_FN_ATTRS571_bittest64(__int64 const *_BitBase, __int64 _BitPos) {572return (*_BitBase >> _BitPos) & 1;573}574static __inline__ unsigned char __DEFAULT_FN_ATTRS575_bittestandcomplement64(__int64 *_BitBase, __int64 _BitPos) {576unsigned char _Res = (*_BitBase >> _BitPos) & 1;577*_BitBase = *_BitBase ^ (1ll << _BitPos);578return _Res;579}580static __inline__ unsigned char __DEFAULT_FN_ATTRS581_bittestandreset64(__int64 *_BitBase, __int64 _BitPos) {582unsigned char _Res = (*_BitBase >> _BitPos) & 1;583*_BitBase = *_BitBase & ~(1ll << _BitPos);584return _Res;585}586static __inline__ unsigned char __DEFAULT_FN_ATTRS587_bittestandset64(__int64 *_BitBase, __int64 _BitPos) {588unsigned char _Res = (*_BitBase >> _BitPos) & 1;589*_BitBase = *_BitBase | (1ll << _BitPos);590return _Res;591}592static __inline__ unsigned char __DEFAULT_FN_ATTRS593_interlockedbittestandset64(__int64 volatile *_BitBase, __int64 _BitPos) {594long long _PrevVal =595__atomic_fetch_or(_BitBase, 1ll << _BitPos, __ATOMIC_SEQ_CST);596return (_PrevVal >> _BitPos) & 1;597}598#endif599/*----------------------------------------------------------------------------*\600|* Interlocked Exchange Add601\*----------------------------------------------------------------------------*/602static __inline__ char __DEFAULT_FN_ATTRS603_InterlockedExchangeAdd8(char volatile *_Addend, char _Value) {604return __atomic_fetch_add(_Addend, _Value, __ATOMIC_SEQ_CST);605}606static __inline__ short __DEFAULT_FN_ATTRS607_InterlockedExchangeAdd16(short volatile *_Addend, short _Value) {608return __atomic_fetch_add(_Addend, _Value, __ATOMIC_SEQ_CST);609}610#ifdef __x86_64__611static __inline__ __int64 __DEFAULT_FN_ATTRS612_InterlockedExchangeAdd64(__int64 volatile *_Addend, __int64 _Value) {613return __atomic_fetch_add(_Addend, _Value, __ATOMIC_SEQ_CST);614}615#endif616/*----------------------------------------------------------------------------*\617|* Interlocked Exchange Sub618\*----------------------------------------------------------------------------*/619static __inline__ char __DEFAULT_FN_ATTRS620_InterlockedExchangeSub8(char volatile *_Subend, char _Value) {621return __atomic_fetch_sub(_Subend, _Value, __ATOMIC_SEQ_CST);622}623static __inline__ short __DEFAULT_FN_ATTRS624_InterlockedExchangeSub16(short volatile *_Subend, short _Value) {625return __atomic_fetch_sub(_Subend, _Value, __ATOMIC_SEQ_CST);626}627static __inline__ long __DEFAULT_FN_ATTRS628_InterlockedExchangeSub(long volatile *_Subend, long _Value) {629return __atomic_fetch_sub(_Subend, _Value, __ATOMIC_SEQ_CST);630}631#ifdef __x86_64__632static __inline__ __int64 __DEFAULT_FN_ATTRS633_InterlockedExchangeSub64(__int64 volatile *_Subend, __int64 _Value) {634return __atomic_fetch_sub(_Subend, _Value, __ATOMIC_SEQ_CST);635}636#endif637/*----------------------------------------------------------------------------*\638|* Interlocked Increment639\*----------------------------------------------------------------------------*/640static __inline__ short __DEFAULT_FN_ATTRS641_InterlockedIncrement16(short volatile *_Value) {642return __atomic_add_fetch(_Value, 1, __ATOMIC_SEQ_CST);643}644#ifdef __x86_64__645static __inline__ __int64 __DEFAULT_FN_ATTRS646_InterlockedIncrement64(__int64 volatile *_Value) {647return __atomic_add_fetch(_Value, 1, __ATOMIC_SEQ_CST);648}649#endif650/*----------------------------------------------------------------------------*\651|* Interlocked Decrement652\*----------------------------------------------------------------------------*/653static __inline__ short __DEFAULT_FN_ATTRS654_InterlockedDecrement16(short volatile *_Value) {655return __atomic_sub_fetch(_Value, 1, __ATOMIC_SEQ_CST);656}657#ifdef __x86_64__658static __inline__ __int64 __DEFAULT_FN_ATTRS659_InterlockedDecrement64(__int64 volatile *_Value) {660return __atomic_sub_fetch(_Value, 1, __ATOMIC_SEQ_CST);661}662#endif663/*----------------------------------------------------------------------------*\664|* Interlocked And665\*----------------------------------------------------------------------------*/666static __inline__ char __DEFAULT_FN_ATTRS667_InterlockedAnd8(char volatile *_Value, char _Mask) {668return __atomic_fetch_and(_Value, _Mask, __ATOMIC_SEQ_CST);669}670static __inline__ short __DEFAULT_FN_ATTRS671_InterlockedAnd16(short volatile *_Value, short _Mask) {672return __atomic_fetch_and(_Value, _Mask, __ATOMIC_SEQ_CST);673}674static __inline__ long __DEFAULT_FN_ATTRS675_InterlockedAnd(long volatile *_Value, long _Mask) {676return __atomic_fetch_and(_Value, _Mask, __ATOMIC_SEQ_CST);677}678#ifdef __x86_64__679static __inline__ __int64 __DEFAULT_FN_ATTRS680_InterlockedAnd64(__int64 volatile *_Value, __int64 _Mask) {681return __atomic_fetch_and(_Value, _Mask, __ATOMIC_SEQ_CST);682}683#endif684/*----------------------------------------------------------------------------*\685|* Interlocked Or686\*----------------------------------------------------------------------------*/687static __inline__ char __DEFAULT_FN_ATTRS688_InterlockedOr8(char volatile *_Value, char _Mask) {689return __atomic_fetch_or(_Value, _Mask, __ATOMIC_SEQ_CST);690}691static __inline__ short __DEFAULT_FN_ATTRS692_InterlockedOr16(short volatile *_Value, short _Mask) {693return __atomic_fetch_or(_Value, _Mask, __ATOMIC_SEQ_CST);694}695static __inline__ long __DEFAULT_FN_ATTRS696_InterlockedOr(long volatile *_Value, long _Mask) {697return __atomic_fetch_or(_Value, _Mask, __ATOMIC_SEQ_CST);698}699#ifdef __x86_64__700static __inline__ __int64 __DEFAULT_FN_ATTRS701_InterlockedOr64(__int64 volatile *_Value, __int64 _Mask) {702return __atomic_fetch_or(_Value, _Mask, __ATOMIC_SEQ_CST);703}704#endif705/*----------------------------------------------------------------------------*\706|* Interlocked Xor707\*----------------------------------------------------------------------------*/708static __inline__ char __DEFAULT_FN_ATTRS709_InterlockedXor8(char volatile *_Value, char _Mask) {710return __atomic_fetch_xor(_Value, _Mask, __ATOMIC_SEQ_CST);711}712static __inline__ short __DEFAULT_FN_ATTRS713_InterlockedXor16(short volatile *_Value, short _Mask) {714return __atomic_fetch_xor(_Value, _Mask, __ATOMIC_SEQ_CST);715}716static __inline__ long __DEFAULT_FN_ATTRS717_InterlockedXor(long volatile *_Value, long _Mask) {718return __atomic_fetch_xor(_Value, _Mask, __ATOMIC_SEQ_CST);719}720#ifdef __x86_64__721static __inline__ __int64 __DEFAULT_FN_ATTRS722_InterlockedXor64(__int64 volatile *_Value, __int64 _Mask) {723return __atomic_fetch_xor(_Value, _Mask, __ATOMIC_SEQ_CST);724}725#endif726/*----------------------------------------------------------------------------*\727|* Interlocked Exchange728\*----------------------------------------------------------------------------*/729static __inline__ char __DEFAULT_FN_ATTRS730_InterlockedExchange8(char volatile *_Target, char _Value) {731__atomic_exchange(_Target, &_Value, &_Value, __ATOMIC_SEQ_CST);732return _Value;733}734static __inline__ short __DEFAULT_FN_ATTRS735_InterlockedExchange16(short volatile *_Target, short _Value) {736__atomic_exchange(_Target, &_Value, &_Value, __ATOMIC_SEQ_CST);737return _Value;738}739#ifdef __x86_64__740static __inline__ __int64 __DEFAULT_FN_ATTRS741_InterlockedExchange64(__int64 volatile *_Target, __int64 _Value) {742__atomic_exchange(_Target, &_Value, &_Value, __ATOMIC_SEQ_CST);743return _Value;744}745#endif746/*----------------------------------------------------------------------------*\747|* Interlocked Compare Exchange748\*----------------------------------------------------------------------------*/749static __inline__ char __DEFAULT_FN_ATTRS750_InterlockedCompareExchange8(char volatile *_Destination,751char _Exchange, char _Comparand) {752__atomic_compare_exchange(_Destination, &_Comparand, &_Exchange, 0,753__ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST);754return _Comparand;755}756static __inline__ short __DEFAULT_FN_ATTRS757_InterlockedCompareExchange16(short volatile *_Destination,758short _Exchange, short _Comparand) {759__atomic_compare_exchange(_Destination, &_Comparand, &_Exchange, 0,760__ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST);761return _Comparand;762}763static __inline__ __int64 __DEFAULT_FN_ATTRS764_InterlockedCompareExchange64(__int64 volatile *_Destination,765__int64 _Exchange, __int64 _Comparand) {766__atomic_compare_exchange(_Destination, &_Comparand, &_Exchange, 0,767__ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST);768return _Comparand;769}770/*----------------------------------------------------------------------------*\771|* Barriers772\*----------------------------------------------------------------------------*/773static __inline__ void __DEFAULT_FN_ATTRS774__attribute__((__deprecated__("use other intrinsics or C++11 atomics instead")))775_ReadWriteBarrier(void) {776__atomic_signal_fence(__ATOMIC_SEQ_CST);777}778static __inline__ void __DEFAULT_FN_ATTRS779__attribute__((__deprecated__("use other intrinsics or C++11 atomics instead")))780_ReadBarrier(void) {781__atomic_signal_fence(__ATOMIC_SEQ_CST);782}783static __inline__ void __DEFAULT_FN_ATTRS784__attribute__((__deprecated__("use other intrinsics or C++11 atomics instead")))785_WriteBarrier(void) {786__atomic_signal_fence(__ATOMIC_SEQ_CST);787}788#ifdef __x86_64__789static __inline__ void __DEFAULT_FN_ATTRS790__faststorefence(void) {791__atomic_thread_fence(__ATOMIC_SEQ_CST);792}793#endif794/*----------------------------------------------------------------------------*\795|* readfs, readgs796|* (Pointers in address space #256 and #257 are relative to the GS and FS797|* segment registers, respectively.)798\*----------------------------------------------------------------------------*/799#define __ptr_to_addr_space(__addr_space_nbr, __type, __offset) \800((volatile __type __attribute__((__address_space__(__addr_space_nbr)))*) \801(__offset))802803#ifdef __i386__804static __inline__ unsigned char __DEFAULT_FN_ATTRS805__readfsbyte(unsigned long __offset) {806return *__ptr_to_addr_space(257, unsigned char, __offset);807}808static __inline__ unsigned short __DEFAULT_FN_ATTRS809__readfsword(unsigned long __offset) {810return *__ptr_to_addr_space(257, unsigned short, __offset);811}812static __inline__ unsigned __int64 __DEFAULT_FN_ATTRS813__readfsqword(unsigned long __offset) {814return *__ptr_to_addr_space(257, unsigned __int64, __offset);815}816#endif817#ifdef __x86_64__818static __inline__ unsigned char __DEFAULT_FN_ATTRS819__readgsbyte(unsigned long __offset) {820return *__ptr_to_addr_space(256, unsigned char, __offset);821}822static __inline__ unsigned short __DEFAULT_FN_ATTRS823__readgsword(unsigned long __offset) {824return *__ptr_to_addr_space(256, unsigned short, __offset);825}826static __inline__ unsigned long __DEFAULT_FN_ATTRS827__readgsdword(unsigned long __offset) {828return *__ptr_to_addr_space(256, unsigned long, __offset);829}830static __inline__ unsigned __int64 __DEFAULT_FN_ATTRS831__readgsqword(unsigned long __offset) {832return *__ptr_to_addr_space(256, unsigned __int64, __offset);833}834#endif835#undef __ptr_to_addr_space836/*----------------------------------------------------------------------------*\837|* movs, stos838\*----------------------------------------------------------------------------*/839#if defined(__i386__) || defined(__x86_64__)840static __inline__ void __DEFAULT_FN_ATTRS841__movsb(unsigned char *__dst, unsigned char const *__src, size_t __n) {842__asm__("rep movsb" : : "D"(__dst), "S"(__src), "c"(__n)843: "%edi", "%esi", "%ecx");844}845static __inline__ void __DEFAULT_FN_ATTRS846__movsd(unsigned long *__dst, unsigned long const *__src, size_t __n) {847__asm__("rep movsl" : : "D"(__dst), "S"(__src), "c"(__n)848: "%edi", "%esi", "%ecx");849}850static __inline__ void __DEFAULT_FN_ATTRS851__movsw(unsigned short *__dst, unsigned short const *__src, size_t __n) {852__asm__("rep movsw" : : "D"(__dst), "S"(__src), "c"(__n)853: "%edi", "%esi", "%ecx");854}855static __inline__ void __DEFAULT_FN_ATTRS856__stosb(unsigned char *__dst, unsigned char __x, size_t __n) {857__asm__("rep stosb" : : "D"(__dst), "a"(__x), "c"(__n)858: "%edi", "%ecx");859}860static __inline__ void __DEFAULT_FN_ATTRS861__stosd(unsigned long *__dst, unsigned long __x, size_t __n) {862__asm__("rep stosl" : : "D"(__dst), "a"(__x), "c"(__n)863: "%edi", "%ecx");864}865static __inline__ void __DEFAULT_FN_ATTRS866__stosw(unsigned short *__dst, unsigned short __x, size_t __n) {867__asm__("rep stosw" : : "D"(__dst), "a"(__x), "c"(__n)868: "%edi", "%ecx");869}870#endif871#ifdef __x86_64__872static __inline__ void __DEFAULT_FN_ATTRS873__movsq(unsigned long long *__dst, unsigned long long const *__src, size_t __n) {874__asm__("rep movsq" : : "D"(__dst), "S"(__src), "c"(__n)875: "%edi", "%esi", "%ecx");876}877static __inline__ void __DEFAULT_FN_ATTRS878__stosq(unsigned __int64 *__dst, unsigned __int64 __x, size_t __n) {879__asm__("rep stosq" : : "D"(__dst), "a"(__x), "c"(__n)880: "%edi", "%ecx");881}882#endif883884/*----------------------------------------------------------------------------*\885|* Misc886\*----------------------------------------------------------------------------*/887static __inline__ void * __DEFAULT_FN_ATTRS888_AddressOfReturnAddress(void) {889return (void*)((char*)__builtin_frame_address(0) + sizeof(void*));890}891static __inline__ void * __DEFAULT_FN_ATTRS892_ReturnAddress(void) {893return __builtin_return_address(0);894}895#if defined(__i386__) || defined(__x86_64__)896static __inline__ void __DEFAULT_FN_ATTRS897__cpuid(int __info[4], int __level) {898__asm__ ("cpuid" : "=a"(__info[0]), "=b" (__info[1]), "=c"(__info[2]), "=d"(__info[3])899: "a"(__level));900}901static __inline__ void __DEFAULT_FN_ATTRS902__cpuidex(int __info[4], int __level, int __ecx) {903__asm__ ("cpuid" : "=a"(__info[0]), "=b" (__info[1]), "=c"(__info[2]), "=d"(__info[3])904: "a"(__level), "c"(__ecx));905}906static __inline__ unsigned __int64 __cdecl __DEFAULT_FN_ATTRS907_xgetbv(unsigned int __xcr_no) {908unsigned int __eax, __edx;909__asm__ ("xgetbv" : "=a" (__eax), "=d" (__edx) : "c" (__xcr_no));910return ((unsigned __int64)__edx << 32) | __eax;911}912static __inline__ void __DEFAULT_FN_ATTRS913__halt(void) {914__asm__ volatile ("hlt");915}916#endif917918/*----------------------------------------------------------------------------*\919|* Privileged intrinsics920\*----------------------------------------------------------------------------*/921#if defined(__i386__) || defined(__x86_64__)922static __inline__ unsigned __int64 __DEFAULT_FN_ATTRS923__readmsr(unsigned long __register) {924// Loads the contents of a 64-bit model specific register (MSR) specified in925// the ECX register into registers EDX:EAX. The EDX register is loaded with926// the high-order 32 bits of the MSR and the EAX register is loaded with the927// low-order 32 bits. If less than 64 bits are implemented in the MSR being928// read, the values returned to EDX:EAX in unimplemented bit locations are929// undefined.930unsigned long __edx;931unsigned long __eax;932__asm__ ("rdmsr" : "=d"(__edx), "=a"(__eax) : "c"(__register));933return (((unsigned __int64)__edx) << 32) | (unsigned __int64)__eax;934}935936static __inline__ unsigned long __DEFAULT_FN_ATTRS937__readcr3(void) {938unsigned long __cr3_val;939__asm__ __volatile__ ("mov %%cr3, %0" : "=q"(__cr3_val) : : "memory");940return __cr3_val;941}942943static __inline__ void __DEFAULT_FN_ATTRS944__writecr3(unsigned int __cr3_val) {945__asm__ ("mov %0, %%cr3" : : "q"(__cr3_val) : "memory");946}947#endif948949#ifdef __cplusplus950}951#endif952953#undef __DEFAULT_FN_ATTRS954955#endif /* __INTRIN_H */956#endif /* _MSC_VER */957958959