concrt.h

Go to the documentation of this file.

/***
* ==++==
*
* Copyright (c) Microsoft Corporation. All rights reserved.
*
* ==--==
* =+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+
*
* concrt.h
*
* Main public header file for ConcRT. This is the only header file a C++ program must include to use the core concurrency runtime features.
*
* The Agents And Message Blocks Library and the Parallel Patterns Library (PPL) are defined in separate header files.
* =-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
****/

#pragma once

#include <crtdefs.h>

#if defined(BUILD_WINDOWS)
#error Must not be included during CRT build with _CRT_WINDOWS flag enabled
#endif

#if !(defined (_M_X64) || defined (_M_IX86) || defined (_M_ARM) || defined (_M_ARM64))
    #error ERROR: Concurrency Runtime is supported only on X64, X86, ARM, and ARM64 architectures.
#endif  /* !(defined (_M_X64) || defined (_M_IX86) || defined (_M_ARM) || defined (_M_ARM64)) */

#if defined (_M_CEE)
    #error ERROR: Concurrency Runtime is not supported when compiling /clr.
#endif  /* defined (_M_CEE) */

#ifndef __cplusplus
    #error ERROR: Concurrency Runtime is supported only for C++.
#endif  /* __cplusplus */

#define _CONCRT_H

#include <exception>
#include <sal.h>
#include <limits.h>
#include <crtdbg.h>
#include <guiddef.h>
#include <intrin.h>
#include <new>

#include <pplinterface.h>

#pragma pack(push,_CRT_PACKING)
#pragma push_macro("new")
#undef new

// Forward declare structs needed from Windows header files

struct _SECURITY_ATTRIBUTES;
typedef _SECURITY_ATTRIBUTES* LPSECURITY_ATTRIBUTES;

struct _GROUP_AFFINITY;
typedef _GROUP_AFFINITY* PGROUP_AFFINITY;

// Define essential types needed from Windows header files

typedef unsigned long DWORD;
#ifndef _HRESULT_DEFINED
#define _HRESULT_DEFINED
#ifdef __midl
typedef LONG HRESULT;
#else  /* __midl */
typedef _Return_type_success_(return >= 0) long HRESULT;
#endif  /* __midl */
#endif  /* _HRESULT_DEFINED */
typedef void * HANDLE;

#pragma push_macro("_YieldProcessor")
#undef _YieldProcessor

#if (defined (_M_IX86) || defined (_M_X64))
#define _YieldProcessor _mm_pause
#else  /* (defined (_M_IX86) || defined (_M_X64)) */
inline void _YieldProcessor() {}
#endif  /* (defined (_M_IX86) || defined (_M_X64)) */

#if (defined (_M_IX86) || defined (_M_ARM))

#define _InterlockedIncrementSizeT(_Target) static_cast<size_t>(_InterlockedIncrement(reinterpret_cast<long volatile *>(_Target)))
#define _InterlockedDecrementSizeT(_Target) static_cast<size_t>(_InterlockedDecrement(reinterpret_cast<long volatile *>(_Target)))
#define _InterlockedCompareExchangeSizeT(_Target, _Exchange, _Comparand) static_cast<size_t>(_InterlockedCompareExchange( \
    reinterpret_cast<long volatile *>(_Target), \
    static_cast<long>(_Exchange), \
    static_cast<long>(_Comparand)))

typedef unsigned long DWORD_PTR, *PDWORD_PTR;

#else  /* (defined (_M_IX86) || defined (_M_ARM)) */

#define _InterlockedIncrementSizeT(_Target) static_cast<size_t>(_InterlockedIncrement64(reinterpret_cast<__int64 volatile *>(_Target)))
#define _InterlockedDecrementSizeT(_Target) static_cast<size_t>(_InterlockedDecrement64(reinterpret_cast<__int64 volatile *>(_Target)))
#define _InterlockedCompareExchangeSizeT(_Target, _Exchange, _Comparand) static_cast<size_t>(_InterlockedCompareExchange64( \
    reinterpret_cast<__int64 volatile *>(_Target), \
    static_cast<__int64>(_Exchange), \
    static_cast<__int64>(_Comparand)))

typedef unsigned __int64 DWORD_PTR, *PDWORD_PTR;

#endif  /* (defined (_M_IX86) || defined (_M_ARM)) */

#ifdef _DEBUG
#ifdef _MSC_VER
// Turn off compiler warnings that are exacerbated by constructs in this
// file's definitions:

// Warning C4127: conditional expression is constant. This is caused by
//      the macros with "do { ... } while (false)" syntax. The syntax is
//      a good way to ensure that a statement-like macro can be used in all
//      contexts (specifically if statements), but the compiler warns about
//      the "while (false)" part.

#define _CONCRT_ASSERT(x)   __pragma (warning (suppress: 4127)) do {_ASSERTE(x); __assume(x);} while(false)
#else  /* _MSC_VER */
#define _CONCRT_ASSERT(x)   do {_ASSERTE(x); __assume(x);} while(false)
#endif  /* _MSC_VER */
#else  /* _DEBUG */
#define _CONCRT_ASSERT(x)   __assume(x)
#endif  /* _DEBUG */

// Used internally to represent the smallest unit in which to allocate hidden types


typedef void * _CONCRT_BUFFER;
#define _LISTENTRY_SIZE ((2 * sizeof(void *) + sizeof(_CONCRT_BUFFER) - 1) / sizeof(_CONCRT_BUFFER))
#define _SAFERWLIST_SIZE ((3 * sizeof(void *) + 2 * sizeof(long) + sizeof(_CONCRT_BUFFER) - 1) / sizeof(_CONCRT_BUFFER))


namespace Concurrency
{

_CONCRTIMP void __cdecl wait(unsigned int _Milliseconds);


_CONCRTIMP void * __cdecl Alloc(size_t _NumBytes);


_CONCRTIMP void __cdecl Free(_Pre_maybenull_ _Post_invalid_ void * _PAllocation);



#ifdef _CRT_USE_WINAPI_FAMILY_DESKTOP_APP


_CONCRTIMP void __cdecl set_task_execution_resources(DWORD_PTR _ProcessAffinityMask);


_CONCRTIMP void __cdecl set_task_execution_resources(unsigned short _Count, PGROUP_AFFINITY _PGroupAffinity);

#endif  /* _CRT_USE_WINAPI_FAMILY_DESKTOP_APP */


typedef void (__cdecl * TaskProc)(void *);

//
// Forward declarations:
//
class Scheduler;
class ScheduleGroup;
class Context;

namespace details
{
    //
    // Forward declarations:
    //
    class ContextBase;
    class _TaskCollectionBase;

    //
    // A utility to hide operator delete from certain objects while still allowing the runtime to delete them internally.
    //
    template<class _Ty>
    void _InternalDeleteHelper(_Ty * _PObject)
    {
        delete _PObject;
    }

    // The purpose of the class is solely to direct allocations of ConcRT classes
    // through a single point, using an internal allocator.
    struct _AllocBase
    {
        // Standard operator new
        void * operator new(size_t _Size)
        {
            return ::Concurrency::Alloc(_Size);
        }

        // Standard operator delete
        void operator delete(void * _Ptr) throw()
        {
            ::Concurrency::Free(_Ptr);
        }

        // Standard operator new, no-throw version
        void * operator new(size_t _Size, const std::nothrow_t&) throw()
        {
            void * _Ptr;

            try
            {
                _Ptr = ::Concurrency::Alloc(_Size);
            }
            catch(...)
            {
                _Ptr = NULL;
            }

            return (_Ptr);
        }

        // Standard operator delete, no-throw version
        void operator delete(void * _Ptr, const std::nothrow_t&) throw()
        {
            operator delete(_Ptr);
        }

        // Standard operator new array
        void * operator new[](size_t _Size)
        {
            return operator new(_Size);
        }

        // Standard operator delete array
        void operator delete[](void * _Ptr) throw()
        {
            operator delete(_Ptr);
        }

        // Standard operator new array, no-throw version
        void * operator new[](size_t _Size, const std::nothrow_t& _No_throw) throw ()
        {
            return operator new(_Size, _No_throw);
        }

        // Standard operator delete array, no-throw version
        void operator delete[](void * _Ptr, const std::nothrow_t& _No_throw) throw()
        {
            operator delete(_Ptr, _No_throw);
        }

        // Standard operator new with void* placement
        void * operator new(size_t, void * _Location) throw()
        {
            return _Location;
        }

        // Standard operator delete with void* placement
        void operator delete(void *, void *) throw()
        {
        }

        // Standard operator new array with void* placement
        void * __cdecl operator new[](size_t, void * _Location) throw()
        {
            return _Location;
        }

        // Standard operator delete array with void* placement
        void __cdecl operator delete[](void *, void *) throw()
        {
        }
    };

    // Stubs to allow the header files to access runtime functionality for WINAPI_PARTITION apps.
    class _Context
    {
    public:
        _CONCRTIMP _Context(::Concurrency::Context * _PContext = NULL) : _M_pContext(_PContext) {}
        _CONCRTIMP static _Context __cdecl _CurrentContext();
        _CONCRTIMP static void __cdecl _Yield();
        _CONCRTIMP static void __cdecl _Oversubscribe(bool _BeginOversubscription);
        _CONCRTIMP bool _IsSynchronouslyBlocked() const;
    private:
        ::Concurrency::Context * _M_pContext;
    };

    class _Scheduler
    {
    public:
        _CONCRTIMP _Scheduler(::Concurrency::Scheduler * _PScheduler = NULL) : _M_pScheduler(_PScheduler) {}
        _CONCRTIMP unsigned int _Reference();
        _CONCRTIMP unsigned int _Release();
        _CONCRTIMP ::Concurrency::Scheduler * _GetScheduler() { return _M_pScheduler; }

    private:
        ::Concurrency::Scheduler * _M_pScheduler;
    };

    class _CurrentScheduler
    {
    public:
        _CONCRTIMP static void __cdecl _ScheduleTask(TaskProc _Proc, void * _Data);
        _CONCRTIMP static unsigned int __cdecl _Id();
        _CONCRTIMP static unsigned int __cdecl _GetNumberOfVirtualProcessors();
        _CONCRTIMP static _Scheduler __cdecl _Get();
    };

    //
    // Wrappers for atomic access
    //
    template <size_t _Size>
    struct _Subatomic_impl { };

    template<>
    struct _Subatomic_impl<4> {
        template <typename _Ty>
        static void _StoreWithRelease(volatile _Ty& _Location, _Ty _Rhs) {
            // For the compiler, a volatile write has release semantics. In addition, on ARM,
            // the volatile write will emit a data memory barrier before the write.
            _Location = _Rhs;
        }

        template <typename _Ty>
        static _Ty _LoadWithAquire(volatile _Ty& _Location) {
            // For the compiler, a volatile read has acquire semantics. In addition, on ARM,
            // the volatile read will emit a data memory barrier after the read.
            return _Location;
        }

        template <typename _Ty>
        static _Ty _CompareAndSwap(volatile _Ty& _Location, _Ty _NewValue, _Ty _Comperand) {
            return (_Ty)_InterlockedCompareExchange((volatile long*)&_Location, (long)_NewValue, (long)_Comperand);
        }

        template <typename _Ty>
        static _Ty _FetchAndAdd(volatile _Ty& _Location, _Ty _Addend) {
            return (_Ty)_InterlockedExchangeAdd((volatile long*)&_Location, (long)_Addend);
        }

        template <typename _Ty>
        static _Ty _Increment(volatile _Ty& _Location) {
            return (_Ty)_InterlockedIncrement((volatile long*)&_Location);
        }

        template <typename _Ty>
        static _Ty _Decrement(volatile _Ty& _Location) {
            return (_Ty)_InterlockedDecrement((volatile long*)&_Location);
        }
    };

#if defined (_WIN64)
    template<>
    struct _Subatomic_impl<8> {
        template <typename _Ty>
        static void _StoreWithRelease(volatile _Ty& _Location, _Ty _Rhs) {
            // For the compiler, a volatile write has release semantics.
            _Location = _Rhs;
        }

        template <typename _Ty>
        static _Ty _LoadWithAquire(volatile _Ty& _Location) {
            // For the compiler, a volatile read has acquire semantics.
            return _Location;
        }

        template <typename _Ty>
        static _Ty _CompareAndSwap(volatile _Ty& _Location, _Ty _NewValue, _Ty _Comperand) {
            return (_Ty)_InterlockedCompareExchange64((volatile __int64*)&_Location, (__int64)_NewValue, (__int64)_Comperand);
        }

        template <typename _Ty>
        static _Ty _FetchAndAdd(volatile _Ty& _Location, _Ty _Addend) {
            return (_Ty)_InterlockedExchangeAdd64((volatile __int64*)&_Location, (__int64)_Addend);
        }

        template <typename _Ty>
        static _Ty _Increment(volatile _Ty& _Location) {
            return (_Ty)_InterlockedIncrement64((volatile __int64*)&_Location);
        }

        template <typename _Ty>
        static _Ty _Decrement(volatile _Ty& _Location) {
            return (_Ty)_InterlockedDecrement64((volatile __int64*)&_Location);
        }
    };
#endif  /* defined (_M_X64) */


    //
    // Wrapper for atomic access. Only works for 4-byte or 8-byte types (for example, int, long, long long, size_t, pointer).
    // Anything else might fail to compile.
    //
    template <typename _Ty>
    class _Subatomic {
    private:
        volatile _Ty _M_value;

    public:
        operator _Ty() const volatile {
            return _Subatomic_impl<sizeof(_Ty)>::_LoadWithAquire(_M_value);
        }

        _Ty operator=(_Ty _Rhs) {
            _Subatomic_impl<sizeof(_Ty)>::_StoreWithRelease(_M_value, _Rhs);
            return _Rhs;
        }

        _Ty _CompareAndSwap(_Ty _NewValue, _Ty _Comperand) {
            return _Subatomic_impl<sizeof(_Ty)>::_CompareAndSwap(_M_value, _NewValue, _Comperand);
        }

        _Ty _FetchAndAdd(_Ty _Addend) {
            return _Subatomic_impl<sizeof(_Ty)>::_FetchAndAdd(_M_value, _Addend);
        }

        _Ty operator++() {
            return _Subatomic_impl<sizeof(_Ty)>::_Increment(_M_value);
        }

        _Ty operator++(int) {
            return _Subatomic_impl<sizeof(_Ty)>::_Increment(_M_value) - 1;
        }

        _Ty operator--() {
            return _Subatomic_impl<sizeof(_Ty)>::_Decrement(_M_value);
        }

        _Ty operator--(int) {
            return _Subatomic_impl<sizeof(_Ty)>::_Decrement(_M_value) + 1;
        }

        _Ty operator+=(_Ty _Addend) {
            return _FetchAndAdd(_Addend) + _Addend;
        }
    };

    //
    // An RAII class that spin-waits on a "rented" flag.
    //
    class _SpinLock
    {
    private:
        volatile long& _M_flag;

    public:
        _CONCRTIMP _SpinLock(volatile long& _Flag);
        _CONCRTIMP ~_SpinLock();

    private:
        _SpinLock(const _SpinLock&);
        void operator=(const _SpinLock&);
    };

    //
    // A class that holds the count used for spinning and is dependent
    // on the number of hardware threads
    //
    struct _SpinCount
    {
        // Initializes the spinCount to either 0 or SPIN_COUNT, depending on
        // the number of hardware threads.
        static void __cdecl _Initialize();

        // Returns the current value of s_spinCount
        _CONCRTIMP static unsigned int __cdecl _Value();

        // The number of iterations used for spinning
        static unsigned int _S_spinCount;
    };

    
    void _CONCRTIMP __cdecl _UnderlyingYield();

    
    unsigned int _CONCRTIMP __cdecl _GetConcurrency();

    
    template<unsigned int _YieldCount = 1>
    class _CONCRTIMP _SpinWait
    {
    public:

        typedef void (__cdecl *_YieldFunction)();

        
        _SpinWait(_YieldFunction _YieldMethod = _UnderlyingYield)
            : _M_yieldFunction(_YieldMethod), _M_state(_StateInitial)
        {
            // Defer initialization of other fields to _SpinOnce().
        }

        
        void _SetSpinCount(unsigned int _Count)
        {
            _CONCRT_ASSERT(_M_state == _StateInitial);
            if (_Count == 0)
            {
                // Specify a count of 0 if we are on a single proc.
                _M_state = _StateSingle;
            }
            else
            {
                _M_currentSpin = _Count;
                _M_currentYield = _YieldCount;
                _M_state = _StateSpin;
            }
        }

        
        bool _SpinOnce()
        {
            switch (_M_state)
            {
            case _StateSpin:
            {
                unsigned long _Count = _NumberOfSpins();

                for (unsigned long _I = 0; _I < _Count; _I++)
                {
                    _YieldProcessor();
                }

                if (!_ShouldSpinAgain())
                {
                    _M_state = (_M_currentYield == 0) ? _StateBlock : _StateYield;
                }

                return true;
            }

            case _StateYield:
                _CONCRT_ASSERT(_M_currentYield > 0);
                if (--_M_currentYield == 0)
                {
                    _M_state = _StateBlock;
                }

                // Execute the yield
                _DoYield();
                return true;

            case _StateBlock:
                // Reset to defaults if client does not block
                _Reset();
                return false;

            case _StateSingle:
                // No need to spin on a single processor: just execute the yield
                _DoYield();
                return false;

            case _StateInitial:
                // Reset counters to their default value and Spin once.
                _Reset();
                return _SpinOnce();
            default:
                // Unreached
                return false;
            };
        }

    protected:

        
        enum _SpinState
        {
            _StateInitial,
            _StateSpin,
            _StateYield,
            _StateBlock,
            _StateSingle
        };

        
        void _DoYield()
        {
#pragma warning ( push )
#pragma warning ( disable : 6326 )  // potential comparison of a constant with another constant
            bool _ShouldYield = (_YieldCount != 0);
#pragma warning ( pop )
            if (_ShouldYield)
            {
                _CONCRT_ASSERT(_M_yieldFunction != NULL);
                _M_yieldFunction();
            }
            else
            {
                _YieldProcessor();
            }
        }

        
        void _Reset()
        {
            _M_state = _StateInitial;

            // Reset to the default spin value. The value specified
            // by the client is ignored on a reset.
            _SetSpinCount(_SpinCount::_Value());

            _CONCRT_ASSERT(_M_state != _StateInitial);
        }

        
        unsigned long _NumberOfSpins()
        {
            return 1;
        }

        
        bool _ShouldSpinAgain()
        {
            return (--_M_currentSpin > 0);
        }

        unsigned long  _M_currentSpin;
        unsigned long  _M_currentYield;
        _SpinState     _M_state;
        _YieldFunction _M_yieldFunction;
    };

    typedef _SpinWait<>   _SpinWaitBackoffNone;
    typedef _SpinWait<0>  _SpinWaitNoYield;

    //
    // This reentrant lock uses CRITICAL_SECTION and is intended for use when kernel blocking
    // is desirable and where it is either known that the lock will be taken recursively in
    // the same thread, or not known that a non-reentrant lock can be used safely.
    //
    class _ReentrantBlockingLock
    {
    public:
        // Constructor for _ReentrantBlockingLock
        _CONCRTIMP _ReentrantBlockingLock();

        // Destructor for _ReentrantBlockingLock
        _CONCRTIMP ~_ReentrantBlockingLock();

        // Acquire the lock, spin if necessary
        _CONCRTIMP void _Acquire();

        // Tries to acquire the lock, does not spin.
        // Returns true if the acquisition worked, false otherwise
        _CONCRTIMP bool _TryAcquire();

        // Releases the lock
        _CONCRTIMP void _Release();


        // An exception safe RAII wrapper.
        class _Scoped_lock
        {
        public:
            // Constructs a holder and acquires the specified lock
            explicit _Scoped_lock(_ReentrantBlockingLock& _Lock) : _M_lock(_Lock)
            {
                _M_lock._Acquire();
            }

            // Destroys the holder and releases the lock
            ~_Scoped_lock()
            {
                _M_lock._Release();
            }
        private:
            _ReentrantBlockingLock& _M_lock;

            _Scoped_lock(const _Scoped_lock&);                    // no copy constructor
            _Scoped_lock const & operator=(const _Scoped_lock&);  // no assignment operator
        };

    private:
        // Critical section requires windows.h. Hide the implementation so that
        // user code need not include windows.
        _CONCRT_BUFFER _M_criticalSection[(4 * sizeof(void *) + 2 * sizeof(long) + sizeof(_CONCRT_BUFFER) - 1) / sizeof(_CONCRT_BUFFER)];
    };

    //
    // This reentrant lock is a pure spin lock and is intended for use when kernel blocking
    // is desirable and where it is either known that the lock will be taken recursively in
    // the same thread, or not known that a non-reentrant lock can be used safely.
    //
    class _ReentrantLock
    {
    public:
        // Constructor for _ReentrantLock
        _CONCRTIMP _ReentrantLock();

        // Acquire the lock, spin if necessary
        _CONCRTIMP void _Acquire();

        // Tries to acquire the lock, does not spin
        // Returns true if the acquisition worked, false otherwise
        _CONCRTIMP bool _TryAcquire();

        // Releases the lock
        _CONCRTIMP void _Release();

        // An exception safe RAII wrapper.
        class _Scoped_lock
        {
        public:
            // Constructs a holder and acquires the specified lock
            explicit _Scoped_lock(_ReentrantLock& _Lock) : _M_lock(_Lock)
            {
                _M_lock._Acquire();
            }

            // Destroys the holder and releases the lock
            ~_Scoped_lock()
            {
                _M_lock._Release();
            }
        private:
            _ReentrantLock& _M_lock;

            _Scoped_lock(const _Scoped_lock&);                    // no copy constructor
            _Scoped_lock const & operator=(const _Scoped_lock&);  // no assignment operator
        };

    private:
        long _M_recursionCount;
        volatile long _M_owner;
    };

    //
    // This non-reentrant lock uses CRITICAL_SECTION and is intended for use in situations
    // where it is known that the lock will not be taken recursively, and can be more
    // efficiently implemented.
    //
    class _NonReentrantBlockingLock
    {
    public:
        // Constructor for _NonReentrantBlockingLock
        //
        // The constructor is exported because _NonReentrantLock is
        // included in DevUnitTests.
        _CONCRTIMP _NonReentrantBlockingLock();

        // Constructor for _NonReentrantBlockingLock
        _CONCRTIMP ~_NonReentrantBlockingLock();

        // Acquire the lock, spin if necessary
        _CONCRTIMP void _Acquire();

        // Tries to acquire the lock, does not spin
        // Returns true if the lock is taken, false otherwise
        _CONCRTIMP bool _TryAcquire();

        // Releases the lock
        _CONCRTIMP void _Release();

        // An exception safe RAII wrapper.
        class _Scoped_lock
        {
        public:
            // Constructs a holder and acquires the specified lock
            explicit _Scoped_lock(_NonReentrantBlockingLock& _Lock) : _M_lock(_Lock)
            {
                _M_lock._Acquire();
            }

            // Destroys the holder and releases the lock
            ~_Scoped_lock()
            {
                _M_lock._Release();
            }
        private:
            _NonReentrantBlockingLock& _M_lock;

            _Scoped_lock(const _Scoped_lock&);                    // no copy constructor
            _Scoped_lock const & operator=(const _Scoped_lock&);  // no assignment operator
        };

    private:
        // Critical section requires windows.h. Hide the implementation so that
        // user code need not include windows.h
        _CONCRT_BUFFER _M_criticalSection[(4 * sizeof(void *) + 2 * sizeof(long) + sizeof(_CONCRT_BUFFER) - 1) / sizeof(_CONCRT_BUFFER)];
    };

    //
    // A Reader-Writer Lock is intended for use in situations with many readers and rare
    // writers.
    //
    // A writer request immediately blocks future readers and then waits until all current
    // readers drain. A reader request does not block future writers and must wait until
    // all writers are done, even those that cut in front of it. In any race between a
    // reader and a writer, the writer always wins.
    //
    class _ReaderWriterLock
    {
    public:
        // Constructor for _ReaderWriterLock
        //
        // The constructor and destructor are exported because _ReaderWriterLock is
        // included in DevUnitTests.
        _CONCRTIMP _ReaderWriterLock();

        // Acquire lock for reading. Spins until all writers finish, new writers
        // can cut in front of a waiting reader.
        _CONCRTIMP void _AcquireRead();

        // Release lock for reading. The last reader changes m_state to State.kFree
        _CONCRTIMP void _ReleaseRead();

        // Acquire lock for writing. Spin until no readers exist, then acquire lock
        // and prevent new readers.
        _CONCRTIMP void _AcquireWrite();

        // Release lock for writing.
        _CONCRTIMP void _ReleaseWrite();

        // Try to acquire the write lock, do not spin if unable to acquire.
        // Returns true if the acquisition worked, false otherwise
        _CONCRTIMP bool _TryAcquireWrite();

        // Returns true if it is in write state, false otherwise
        bool _HasWriteLock() const
        {
            return (_M_state == _Write);
        }

        // Guarantees that all writers are out of the lock. This does nothing if there are no pending writers.
        void _FlushWriteOwners();

        // An exception safe RAII wrapper.
        class _Scoped_lock
        {
        public:
            // Constructs a holder and acquires the writer lock
            explicit _Scoped_lock(_ReaderWriterLock& _Lock) : _M_lock(_Lock)
            {
                _M_lock._AcquireWrite();
            }

            // Destroys the holder and releases the writer lock
            ~_Scoped_lock()
            {
                _M_lock._ReleaseWrite();
            }

        private:

            _ReaderWriterLock& _M_lock;

            _Scoped_lock(const _Scoped_lock&);                    // no copy constructor
            _Scoped_lock const & operator=(const _Scoped_lock&);  // no assignment operator
        };

        // An exception safe RAII wrapper for reads.
        class _Scoped_lock_read
        {
        public:
            // Constructs a holder and acquires the reader lock
            explicit _Scoped_lock_read(_ReaderWriterLock& _Lock) : _M_lock(_Lock)
            {
                _M_lock._AcquireRead();
            }

            // Destroys the holder and releases the reader lock
            ~_Scoped_lock_read()
            {
                _M_lock._ReleaseRead();
            }

        private:

            _ReaderWriterLock& _M_lock;

            _Scoped_lock_read(const _Scoped_lock_read&);                    // no copy constructor
            _Scoped_lock_read const & operator=(const _Scoped_lock_read&);  // no assignment operator
        };

    private:
        // State enum where:
        // -1    --> write mode
        // 0     --> free
        // n > 0 --> n readers have locked in read mode.
        enum _State
        {
          _Write = -1,
          _Free  = 0,
          _Read  = 1
        };

        // The current state of the lock, mapping to the State enum. This is also
        // an indicator of the number of readers holding the lock, for any number > 0.
        volatile long _M_state;

        // A writer increments this as soon as it wants to lock and decrements this
        // after releasing the lock. To prevent writers from starving, a reader will
        // wait until this counter is zero, and only then will try to obtain the lock.
        volatile long _M_numberOfWriters;

        // Spin-Wait-Until variant
        static void __cdecl _WaitEquals(volatile const long& _Location, long _Value, long _Mask = 0xFFFFFFFF);
    };

    //
    // Exception safe RAII wrappers for _malloca()
    //

    //
    // _MallocaArrayHolder is used when the allocation size is known up front, and the memory must be allocated in a contiguous space
    //
    template<typename _ElemType>
    class _MallocaArrayHolder
    {
    public:

        _MallocaArrayHolder() : _M_ElemArray(NULL), _M_ElemsConstructed(0) {}

        // _Initialize takes the pointer to the memory allocated by the user using _malloca
        void _Initialize(_ElemType * _Elem)
        {
            // The object must be initialized exactly once
            _CONCRT_ASSERT(_M_ElemArray == NULL && _M_ElemsConstructed == 0);
            _M_ElemArray = _Elem;
            _M_ElemsConstructed = 0;
        }

        // _InitOnRawMalloca take the raw pointer returned by _malloca directly
        // It will initialize itself with that pointer and return a strong typed pointer.
        // To be noted that the constructor will NOT be called.
        _ElemType * _InitOnRawMalloca(void * _MallocaRet)
        {
            if (_MallocaRet == nullptr)
                throw std::bad_alloc();
            _Initialize(static_cast<_ElemType *>(_MallocaRet));
            return static_cast<_ElemType *>(_MallocaRet);
        }

        // Register the next slot for destruction. Because we only keep the index of the last slot to be destructed,
        // this method must be called sequentially from 0 to N where N < _ElemCount.
        void _IncrementConstructedElemsCount()
        {
            _CONCRT_ASSERT(_M_ElemArray != NULL); // must already be initialized
            _M_ElemsConstructed++;
        }

        virtual ~_MallocaArrayHolder()
        {
            for( size_t _I=0; _I < _M_ElemsConstructed; ++_I )
            {
                _M_ElemArray[_I].~_ElemType();
            }
            // Works even when object was not initialized, that is, _M_ElemArray == NULL
            _freea(_M_ElemArray);
        }
    private:
        _ElemType * _M_ElemArray;
        size_t     _M_ElemsConstructed;

        // Copy construction and assignment are not supported.
        _MallocaArrayHolder(const _MallocaArrayHolder & );
        _MallocaArrayHolder&  operator = (const _MallocaArrayHolder & );
    };

    //
    // _MallocaListHolder is used when the allocation size is not known up front, and the elements are added to the list dynamically
    //
    template<typename _ElemType>
    class _MallocaListHolder
    {
    public:
        // Returns the size required to allocate the payload itself and the pointer to the next element
        size_t _GetAllocationSize() const
        {
            return sizeof(_ElemNodeType);
        }

        _MallocaListHolder() : _M_FirstNode(NULL)
        {
        }

        // Add the next element to the list. The memory is allocated in the caller's frame by _malloca
        void _AddNode(_ElemType * _Elem)
        {
            _ElemNodeType * _Node = reinterpret_cast<_ElemNodeType *>(_Elem);
            _Node->_M_Next = _M_FirstNode;
            _M_FirstNode = reinterpret_cast<_ElemNodeType *>(_Elem);
        }

        // _AddRawMallocaNode take the raw pointer returned by _malloca directly
        // It will add that bucket of memory to the list and return a strong typed pointer.
        // To be noted that the constructor will NOT be called.
        _ElemType * _AddRawMallocaNode(void * _MallocaRet)
        {
            if (_MallocaRet == nullptr)
                throw std::bad_alloc();
            _AddNode(static_cast<_ElemType *>(_MallocaRet));
            return static_cast<_ElemType *>(_MallocaRet);
        }

        // Walk the list and destruct, then free each element
        _At_(this->_M_FirstNode, _Pre_valid_) virtual ~_MallocaListHolder()
        {
            for( _ElemNodeType * _Node = _M_FirstNode; _Node != NULL; )
            {
                auto _M_Next = _Node->_M_Next;
                _Node->_M_Elem._ElemType::~_ElemType();
                 _freea(_Node);
                _Node = _M_Next;
            }
        }

    private:

        class _ElemNodeType
        {
            friend class _MallocaListHolder;
            _ElemType _M_Elem;
            _ElemNodeType * _M_Next;
            // Always instantiated using malloc, so default constructor and destructor are not needed.
            _ElemNodeType();
            ~_ElemNodeType();
            // Copy construction and assignment are not supported.
            _ElemNodeType(const _ElemNodeType & );
            _ElemNodeType &  operator = (const _ElemNodeType & );
        };

        _ElemNodeType* _M_FirstNode;

        // Copy construction and assignment are not supported.
        _MallocaListHolder(const _MallocaListHolder & );
        _MallocaListHolder &  operator = (const _MallocaListHolder & );
    };

    // Forward declarations
    class _StructuredTaskCollection;
    class _TaskCollection;
    class _UnrealizedChore;
} // namespace details

//**************************************************************************
// Public Namespace:
//
// Anything in the Concurrency namespace is intended for direct client consumption.
//
//**************************************************************************


class scheduler_resource_allocation_error : public std::exception
{
public:
    
    _CONCRTIMP scheduler_resource_allocation_error(_In_z_ const char * _Message, HRESULT _Hresult) throw();

    
    explicit _CONCRTIMP scheduler_resource_allocation_error(HRESULT _Hresult) throw();

    
    _CONCRTIMP HRESULT get_error_code() const throw();

private:
    HRESULT _Hresult;
};


class scheduler_worker_creation_error : public scheduler_resource_allocation_error
{
public:
    
    _CONCRTIMP scheduler_worker_creation_error(_In_z_ const char * _Message, HRESULT _Hresult) throw();

    
    explicit _CONCRTIMP scheduler_worker_creation_error(HRESULT _Hresult) throw();
};


class unsupported_os  : public std::exception
{
public:
    
    explicit _CONCRTIMP unsupported_os(_In_z_ const char * _Message) throw();

    
    _CONCRTIMP unsupported_os() throw();
};


class scheduler_not_attached  : public std::exception
{
public:
    
    explicit _CONCRTIMP scheduler_not_attached(_In_z_ const char * _Message) throw();

    
    _CONCRTIMP scheduler_not_attached() throw();
};


class improper_scheduler_attach : public std::exception
{
public:
    
    explicit _CONCRTIMP improper_scheduler_attach(_In_z_ const char * _Message) throw();

    
    _CONCRTIMP improper_scheduler_attach() throw();
};


class improper_scheduler_detach : public std::exception
{
public:

    
    explicit _CONCRTIMP improper_scheduler_detach(_In_z_ const char * _Message) throw();

    
    _CONCRTIMP improper_scheduler_detach() throw();
};


class improper_scheduler_reference : public std::exception
{
public:

    
    explicit _CONCRTIMP improper_scheduler_reference(_In_z_ const char* _Message) throw();

    
    _CONCRTIMP improper_scheduler_reference() throw();
};


class default_scheduler_exists : public std::exception
{
public:
    
    explicit _CONCRTIMP default_scheduler_exists(_In_z_ const char * _Message) throw();

    
    _CONCRTIMP default_scheduler_exists() throw();
};


class context_unblock_unbalanced : public std::exception
{
public:
    
    explicit _CONCRTIMP context_unblock_unbalanced(_In_z_ const char * _Message) throw();

    
    _CONCRTIMP context_unblock_unbalanced() throw();
};


class context_self_unblock : public std::exception
{
public:
    
    explicit _CONCRTIMP context_self_unblock(_In_z_ const char * _Message) throw();

    
    _CONCRTIMP context_self_unblock() throw();
};


class missing_wait : public std::exception
{
public:
    
    explicit _CONCRTIMP missing_wait(_In_z_ const char * _Message) throw();

    
    _CONCRTIMP missing_wait() throw();
};


class bad_target : public std::exception
{
public:
    
    explicit _CONCRTIMP bad_target(_In_z_ const char * _Message) throw();

    
    _CONCRTIMP bad_target() throw();
};


class message_not_found : public std::exception
{
public:
    
    explicit _CONCRTIMP message_not_found(_In_z_ const char * _Message) throw();

    
    _CONCRTIMP message_not_found() throw();
};


class invalid_link_target : public std::exception
{
public:
    
    explicit _CONCRTIMP invalid_link_target(_In_z_ const char * _Message) throw();

    
    _CONCRTIMP invalid_link_target() throw();
};


class invalid_scheduler_policy_key : public std::exception
{
public:
    
    explicit _CONCRTIMP invalid_scheduler_policy_key(_In_z_ const char * _Message) throw();

    
    _CONCRTIMP invalid_scheduler_policy_key() throw();
};


class invalid_scheduler_policy_value : public std::exception
{
public:
    
    explicit _CONCRTIMP invalid_scheduler_policy_value(_In_z_ const char * _Message) throw();

    
    _CONCRTIMP invalid_scheduler_policy_value() throw();
};


class invalid_scheduler_policy_thread_specification : public std::exception
{
public:
    
    explicit _CONCRTIMP invalid_scheduler_policy_thread_specification(_In_z_ const char * _Message) throw();

    
    _CONCRTIMP invalid_scheduler_policy_thread_specification() throw();
};


class nested_scheduler_missing_detach : public std::exception
{
public:
    
    explicit _CONCRTIMP nested_scheduler_missing_detach(_In_z_ const char * _Message) throw();

    
    _CONCRTIMP nested_scheduler_missing_detach() throw();
};


class operation_timed_out : public std::exception
{
public:
    
    explicit _CONCRTIMP operation_timed_out(_In_z_ const char * _Message) throw();

    
    _CONCRTIMP operation_timed_out() throw();
};


class invalid_multiple_scheduling : public std::exception
{
public:
    
    explicit _CONCRTIMP invalid_multiple_scheduling(_In_z_ const char * _Message) throw();

    
    _CONCRTIMP invalid_multiple_scheduling() throw();
};


class invalid_oversubscribe_operation : public std::exception
{
public:
    
    explicit _CONCRTIMP invalid_oversubscribe_operation(_In_z_ const char * _Message) throw();

    
    _CONCRTIMP invalid_oversubscribe_operation() throw();
};


class improper_lock : public std::exception
{
public:

    
    explicit _CONCRTIMP improper_lock(_In_z_ const char * _Message) throw();

    
    _CONCRTIMP improper_lock() throw();
};


class location
{
public:

    
    location() :
        _M_type(_System),
        _M_reserved(0),
        _M_bindingId(0),
        _M_pBinding(NULL),
        _M_ptr(NULL)
    {
    }

    
    location(const location& _Src)
    {
        _Assign(_Src);
    }

#ifdef _CRT_USE_WINAPI_FAMILY_DESKTOP_APP

    
    _CONCRTIMP static location __cdecl from_numa_node(unsigned short _NumaNodeNumber);

#endif  /* _CRT_USE_WINAPI_FAMILY_DESKTOP_APP */

    
    _CONCRTIMP static location __cdecl current();

    
    location& operator=(const location& _Rhs)
    {
        _Assign(_Rhs);
        return *this;
    }

    
    ~location()
    {
    }

    
    bool operator==(const location& _Rhs) const
    {
        return (_M_type == _Rhs._M_type && _M_ptr == _Rhs._M_ptr);
    }

    
    bool operator!=(const location& _Rhs) const
    {
        return !operator==(_Rhs);
    }

    //**************************************************
    //
    // Runtime internal public pieces of location. No code outside the core of ConcRT can depend on anything
    // below. It is internal implementation detail:
    //

    
    _CONCRTIMP static location __cdecl _Current_node();

    
    enum _Type
    {
        _System, // _M_id is meaningless

        _NumaNode, // _M_id is the Windows NUMA node number

        _SchedulingNode, // _M_id is the unique identifier for the scheduling node

        _ExecutionResource, // _M_id is the unique identifier for the execution resource
    };

    
    location(_Type _LocationType, unsigned int _Id, unsigned int _BindingId = 0, _Inout_opt_ void *_PBinding = NULL);

    
    bool _FastVPIntersects(const location& _Rhs) const;

    
    bool _FastNodeIntersects(const location& _Rhs) const;

    
    void _Assign(const location& _Rhs)
    {
        _M_type = _Rhs._M_type;
        _M_reserved = _Rhs._M_reserved;

        _M_ptr = _Rhs._M_ptr;

        _M_bindingId = _Rhs._M_bindingId;
        _M_pBinding = _Rhs._M_pBinding;
    }

    
    bool _Is_system() const
    {
        return (_Type)_M_type == _System;
    }

    
    template<typename _Ty>
    _Ty* _As() const
    {
        return reinterpret_cast<_Ty *>(_M_pBinding);
    }

    
    unsigned int _GetId() const
    {
        return _M_id;
    }

    
    _Type _GetType() const
    {
        return (_Type)_M_type;
    }

    
    unsigned int _GetBindingId() const
    {
        return _M_bindingId;
    }

private:

    // Indicates the type of location (as _Type)
    unsigned int _M_type : 28;

    // Flags on the location. Reserved for future use.
    unsigned int _M_reserved : 4;

    // If the location has a tight binding, this is the unique identifier of the scheduler to which the binding has specific meaning.
    unsigned int _M_bindingId;

    // Defines the agnostic (abstract hardware) binding of the location.
    union
    {
        // The identifier for the binding (NUMA node number, scheduler node ID, execution resource ID)
        unsigned int _M_id;

        // Pointer binding.
        void *_M_ptr;
    };

    // The specific binding to a scheduler. (For example, a specific virtual processor for something like location::current() )
    // This will be NULL if there is no tight binding.
    void *_M_pBinding;
};

#ifdef _CRT_USE_WINAPI_FAMILY_DESKTOP_APP


class ScheduleGroup
{
public:

    
    virtual void ScheduleTask(TaskProc _Proc, _Inout_opt_ void * _Data) =0;

    
    virtual unsigned int Id() const =0;

    
    virtual unsigned int Reference() =0;

    
    virtual unsigned int Release() =0;

protected:

    //
    // Privatize operator delete. Clients should utilize Release to relinquish a schedule group.
    //
    template<class _Ty> friend void ::Concurrency::details::_InternalDeleteHelper(_Ty * _PObject);

    virtual ~ScheduleGroup() {};
};


const unsigned int MaxExecutionResources = 0xFFFFFFFF;


const unsigned int INHERIT_THREAD_PRIORITY = 0x0000F000;


enum PolicyElementKey
{
    
    SchedulerKind,

    
    MaxConcurrency,

    
    MinConcurrency,

    
    TargetOversubscriptionFactor,

    
    LocalContextCacheSize,

    
    ContextStackSize,

    
    ContextPriority,

    
    SchedulingProtocol,

    
    DynamicProgressFeedback,

    
    WinRTInitialization,

    
    MaxPolicyElementKey
};


enum SchedulerType
{
    
    ThreadScheduler,

    
    UmsThreadDefault = ThreadScheduler
};

#pragma deprecated(UmsThreadDefault)


enum SchedulingProtocolType
{
    
    EnhanceScheduleGroupLocality,

    
    EnhanceForwardProgress
};


enum DynamicProgressFeedbackType
{
    
    ProgressFeedbackDisabled,

    
    ProgressFeedbackEnabled
};


enum WinRTInitializationType
{
    
    InitializeWinRTAsMTA,

    
    DoNotInitializeWinRT
};


class SchedulerPolicy
{
public:

    
    _CONCRTIMP SchedulerPolicy();

    
    _CONCRTIMP SchedulerPolicy(size_t _PolicyKeyCount, ...);

    
    _CONCRTIMP SchedulerPolicy(const SchedulerPolicy& _SrcPolicy);

    
    _CONCRTIMP SchedulerPolicy& operator=(const SchedulerPolicy& _RhsPolicy);

    
    _CONCRTIMP ~SchedulerPolicy();

    
    _CONCRTIMP unsigned int GetPolicyValue(PolicyElementKey _Key) const;

    
    _CONCRTIMP unsigned int SetPolicyValue(PolicyElementKey _Key, unsigned int _Value);

    
    _CONCRTIMP void SetConcurrencyLimits(unsigned int _MinConcurrency, unsigned int _MaxConcurrency = MaxExecutionResources);

    
    void _ValidateConcRTPolicy() const;

private:

    struct _PolicyBag
    {
        union
        {
            unsigned int _M_pPolicyBag[MaxPolicyElementKey];
            struct
            {
                SchedulerType _M_schedulerKind;
                unsigned int _M_maxConcurrency;
                unsigned int _M_minConcurrency;
                unsigned int _M_targetOversubscriptionFactor;
                unsigned int _M_localContextCacheSize;
                unsigned int _M_contextStackSize;
                unsigned int _M_contextPriority;
                SchedulingProtocolType _M_schedulingProtocol;
                DynamicProgressFeedbackType _M_dynamicProgressFeedback;
                WinRTInitializationType _M_WinRTInitialization;
            } _M_specificValues;
        } _M_values;
    } *_M_pPolicyBag;

    
    void _Initialize(size_t _PolicyKeyCount, va_list * _PArgs);

    
    void _Assign(const SchedulerPolicy& _SrcPolicy);

    
    static bool __cdecl _ValidPolicyKey(PolicyElementKey _Key);

    
    static bool __cdecl _ValidPolicyValue(PolicyElementKey _Key, unsigned int _Value);

    
    static bool __cdecl _AreConcurrencyLimitsValid(unsigned int _MinConcurrency, unsigned int _MaxConcurrency);
    bool _AreConcurrencyLimitsValid() const;

    
    bool _ArePolicyCombinationsValid() const;

    
    void _ResolvePolicyValues();

    
    static char * __cdecl _StringFromPolicyKey(unsigned int _Index);
};


class CurrentScheduler
{
private:
    CurrentScheduler() {}

public:
    
    _CONCRTIMP static unsigned int __cdecl Id();

    
    _CONCRTIMP static SchedulerPolicy __cdecl GetPolicy();

    
    _CONCRTIMP static Scheduler * __cdecl Get();

    
    _CONCRTIMP static unsigned int __cdecl GetNumberOfVirtualProcessors();

    
    _CONCRTIMP static void __cdecl Create(const SchedulerPolicy& _Policy);

    
    _CONCRTIMP static void __cdecl Detach();

    
    _CONCRTIMP static void __cdecl RegisterShutdownEvent(HANDLE _ShutdownEvent);

    
    _CONCRTIMP static ScheduleGroup * __cdecl CreateScheduleGroup();

    
    _CONCRTIMP static ScheduleGroup * __cdecl CreateScheduleGroup(location& _Placement);

    
    _CONCRTIMP static void __cdecl ScheduleTask(TaskProc _Proc, _Inout_opt_ void * _Data);

    
    _CONCRTIMP static void __cdecl ScheduleTask(TaskProc _Proc, _Inout_opt_ void * _Data, location& _Placement);

    
    _CONCRTIMP static bool __cdecl IsAvailableLocation(const location& _Placement);
};


class Scheduler
{
protected:
    
    Scheduler() {}

    
    virtual ~Scheduler() {}

public:

    
    _CONCRTIMP static Scheduler * __cdecl Create(const SchedulerPolicy& _Policy);

    
    virtual unsigned int Id() const =0;

    
    virtual unsigned int GetNumberOfVirtualProcessors() const =0;


    
    virtual SchedulerPolicy GetPolicy() const =0;

    
    virtual unsigned int Reference() =0 ;

    
    virtual unsigned int Release() =0;

    
    virtual void RegisterShutdownEvent(HANDLE _Event) =0;

    
    virtual void Attach() =0;

    
    _CONCRTIMP static void __cdecl SetDefaultSchedulerPolicy(const SchedulerPolicy& _Policy);

    
    _CONCRTIMP static void __cdecl ResetDefaultSchedulerPolicy();

    
    virtual ScheduleGroup * CreateScheduleGroup() =0;

    
    virtual ScheduleGroup * CreateScheduleGroup(location& _Placement) =0;

    
    virtual void ScheduleTask(TaskProc _Proc, _Inout_opt_ void * _Data) =0;

    
    virtual void ScheduleTask(TaskProc _Proc, _Inout_opt_ void * _Data, location& _Placement) =0;

    
    virtual bool IsAvailableLocation(const location& _Placement) const =0;
};


class Context
{
public:
    
    virtual unsigned int GetId() const =0;

    
    virtual unsigned int GetVirtualProcessorId() const =0;

    
    virtual unsigned int GetScheduleGroupId() const =0;

    
    _CONCRTIMP static unsigned int __cdecl Id();

    
    _CONCRTIMP static unsigned int __cdecl VirtualProcessorId();

    
    _CONCRTIMP static unsigned int __cdecl ScheduleGroupId();

    
    _CONCRTIMP static void __cdecl Block();

    
    virtual void Unblock() =0;

    
    virtual bool IsSynchronouslyBlocked() const =0;

    
    _CONCRTIMP static void __cdecl _SpinYield();

    
    _CONCRTIMP static void (__cdecl Yield)();

    
    static inline void __cdecl YieldExecution()
    {
        (Yield)();
    }

    
    _CONCRTIMP static bool __cdecl IsCurrentTaskCollectionCanceling();

    
    _CONCRTIMP static Context * __cdecl CurrentContext();

    
    _CONCRTIMP static void __cdecl Oversubscribe(bool _BeginOversubscription);

protected:

    //
    // Privatize operator delete. The scheduler internally manages contexts.
    //
    template<class _Ty> friend void ::Concurrency::details::_InternalDeleteHelper(_Ty * _PObject);

    virtual ~Context() {};
};

#endif  /* _CRT_USE_WINAPI_FAMILY_DESKTOP_APP */


const size_t COOPERATIVE_WAIT_TIMEOUT = SIZE_MAX;


const unsigned int COOPERATIVE_TIMEOUT_INFINITE = (unsigned int)-1;


class critical_section
{
public:

    
    _CONCRTIMP critical_section();

    
    _CONCRTIMP ~critical_section();

    
    _CONCRTIMP void lock();

    
    _CONCRTIMP bool try_lock();

    
    _CONCRTIMP bool try_lock_for(unsigned int _Timeout);

    
    _CONCRTIMP void unlock();

    
    typedef critical_section& native_handle_type;

    
    _CONCRTIMP native_handle_type native_handle();

    
    void _Flush_current_owner();

    
    bool _Acquire_lock(void * _PLockingNode, bool _FHasExternalNode);

    
    class scoped_lock
    {
    public:

        
        explicit _CONCRTIMP scoped_lock(critical_section& _Critical_section);

        
        _CONCRTIMP ~scoped_lock();

    private:

        critical_section& _M_critical_section;
        _CONCRT_BUFFER _M_node[(4 * sizeof(void *) + 2 * sizeof(unsigned int) + sizeof(_CONCRT_BUFFER) - 1) / sizeof(_CONCRT_BUFFER)];

        scoped_lock(const scoped_lock&);                    // no copy constructor
        scoped_lock const & operator=(const scoped_lock&);  // no assignment operator
    };

private:
    
    void _Switch_to_active(void * _PLockingNode);

    _CONCRT_BUFFER  _M_activeNode[(4 * sizeof(void *) + 2 * sizeof(unsigned int) + sizeof(_CONCRT_BUFFER) - 1) / sizeof(_CONCRT_BUFFER)];
    void * volatile _M_pHead;
    void * volatile _M_pTail;

    
    critical_section(const critical_section&);

    
    critical_section& operator=(const critical_section&);
};


class reader_writer_lock
{
public:

    
    _CONCRTIMP reader_writer_lock();

    
    _CONCRTIMP ~reader_writer_lock();

    
    _CONCRTIMP void lock();

    
    _CONCRTIMP bool try_lock();

    
    _CONCRTIMP void lock_read();

    
    _CONCRTIMP bool try_lock_read();

    
    _CONCRTIMP void unlock();

    
    void _Acquire_lock(void * _PLockingNode, bool _FHasExternalNode);

    
    class scoped_lock
    {
    public:
        
        explicit _CONCRTIMP scoped_lock(reader_writer_lock& _Reader_writer_lock);

        
        _CONCRTIMP ~scoped_lock();

    private:

        reader_writer_lock& _M_reader_writer_lock;
        _CONCRT_BUFFER _M_writerNode[(4 * sizeof(void *) + 2 * sizeof(unsigned int) + sizeof(_CONCRT_BUFFER) - 1) / sizeof(_CONCRT_BUFFER)];

        scoped_lock(const scoped_lock&);                    // no copy constructor
        scoped_lock const & operator=(const scoped_lock&);  // no assignment operator
    };

    
    class scoped_lock_read
    {
    public:
        
        explicit _CONCRTIMP scoped_lock_read(reader_writer_lock& _Reader_writer_lock);

        
        _CONCRTIMP ~scoped_lock_read();

    private:

        reader_writer_lock& _M_reader_writer_lock;

        scoped_lock_read(const scoped_lock_read&);                    // no copy constructor
        scoped_lock_read const & operator=(const scoped_lock_read&);  // no assignment operator
    };

private:

    
    bool _Set_next_writer(void * _PWriter);

    
    void * _Get_reader_convoy();

    
    void _Unlock_writer();

    
    void _Unlock_reader();

    
    void _Remove_last_writer(void * _PWriter);

    
    void _Switch_to_active(void * _PWriter);

    _CONCRT_BUFFER _M_activeWriter[(4 * sizeof(void *) + 2 * sizeof(unsigned int) + sizeof(_CONCRT_BUFFER) - 1) / sizeof(_CONCRT_BUFFER)];
    void *         _M_pReaderHead;
    void *         _M_pWriterHead;
    void *         _M_pWriterTail;
    volatile long  _M_lockState;

    
    reader_writer_lock (const reader_writer_lock& _Lock);

    
    reader_writer_lock& operator=(const reader_writer_lock& _Lock);
};


class event
{
public:

    
    _CONCRTIMP event();

    
    _CONCRTIMP ~event();

    
    _CONCRTIMP size_t wait(unsigned int _Timeout = COOPERATIVE_TIMEOUT_INFINITE);

    
    _CONCRTIMP void set();

    
    _CONCRTIMP void reset();

    
    _CONCRTIMP static size_t __cdecl wait_for_multiple(_In_reads_(_Count) event ** _PPEvents, size_t _Count, bool _FWaitAll, unsigned int _Timeout = COOPERATIVE_TIMEOUT_INFINITE);


    static const unsigned int timeout_infinite = COOPERATIVE_TIMEOUT_INFINITE;
private:

    // Prevent bad usage of copy-constructor and copy-assignment
    event(const event& _Event);
    event& operator=(const event& _Event);

    void * volatile _M_pWaitChain;
    void * _M_pResetChain;
    ::Concurrency::critical_section _M_lock;
};

namespace details
{
    // Base class for all reference counted objects
    class _RefCounterBase
    {
    public:

        virtual ~_RefCounterBase()
        {
            _CONCRT_ASSERT(_M_refCount == 0);
        }

        // Acquires a reference
        // Returns the new reference count.
        long _Reference()
        {
            long _Refcount = _InterlockedIncrement(&_M_refCount);

            // 0 - 1 transition is illegal
            _CONCRT_ASSERT(_Refcount > 1);
            return _Refcount;
        }

        // Releases the reference
        // Returns the new reference count
        long _Release()
        {
            long _Refcount = _InterlockedDecrement(&_M_refCount);
            _CONCRT_ASSERT(_Refcount >= 0);

            if (_Refcount == 0)
            {
                _Destroy();
            }

            return _Refcount;
        }

    protected:

        // Allow derived classes to provide their own deleter
        virtual void _Destroy()
        {
            delete this;
        }

        // Only allow instantiation through derived class
        _RefCounterBase(long _InitialCount = 1) : _M_refCount(_InitialCount)
        {
            _CONCRT_ASSERT(_M_refCount > 0);
        }

        // Reference count
        volatile long _M_refCount;
    };

    class _CancellationTokenState;
    class _CancellationTokenRegistration;

    // This is a non-reentrant lock wrapper around the ConcRT critical-section
    // and used by agents/messaging
    class _NonReentrantPPLLock
    {
    public:

        // Constructor for _NonReentrantPPLLock
        _CONCRTIMP _NonReentrantPPLLock();
        _NonReentrantPPLLock(const _NonReentrantPPLLock&) = delete;

        _NonReentrantPPLLock& operator=(const _NonReentrantPPLLock&) = delete;
        // Acquire the lock, spin if necessary
        _CONCRTIMP void _Acquire(void * _Lock_node);

        // Releases the lock
        _CONCRTIMP void _Release();

        // An exception safe RAII wrapper.
        class _Scoped_lock
        {
        public:
            // Constructs a holder and acquires the specified lock
            _CONCRTIMP explicit _Scoped_lock(_NonReentrantPPLLock& _Lock);

            _Scoped_lock(const _Scoped_lock&) = delete;
            _Scoped_lock& operator=(const _Scoped_lock&) = delete;

            // Destroys the holder and releases the lock
            _CONCRTIMP ~_Scoped_lock();

        private:
            _NonReentrantPPLLock& _M_lock;
            _CONCRT_BUFFER  _M_lockNode[(4 * sizeof(void *) + 2 * sizeof(unsigned int) + sizeof(_CONCRT_BUFFER) - 1) / sizeof(_CONCRT_BUFFER)];
        };

    private:
        // critical_section
        ::Concurrency::critical_section _M_criticalSection;
    };

    // This is a reentrant lock implemented using the ConcRT critical section
    class _ReentrantPPLLock
    {
    public:
        // Constructor for _ReentrantPPLLock
        _CONCRTIMP _ReentrantPPLLock();
        _ReentrantPPLLock(const _ReentrantPPLLock&) = delete;

        _ReentrantPPLLock& operator=(const _ReentrantPPLLock&) = delete;

        // Acquire the lock, spin if necessary
        _CONCRTIMP void _Acquire(void * _Lock_node);

        // Releases the lock
        _CONCRTIMP void _Release();

        // An exception safe RAII wrapper.
        class _Scoped_lock
        {
        public:
            // Constructs a holder and acquires the specified lock
            _CONCRTIMP explicit _Scoped_lock(_ReentrantPPLLock& _Lock);

            // Destroys the holder and releases the lock
            _CONCRTIMP ~_Scoped_lock();

        private:
            _ReentrantPPLLock& _M_lock;
            _CONCRT_BUFFER  _M_lockNode[(4 * sizeof(void *) + 2 * sizeof(unsigned int) + sizeof(_CONCRT_BUFFER) - 1) / sizeof(_CONCRT_BUFFER)];

            _Scoped_lock(const _Scoped_lock&);                    // no copy constructor
            _Scoped_lock const & operator=(const _Scoped_lock&);  // no assignment operator
        };

    private:
        // critical_section
        ::Concurrency::critical_section _M_criticalSection;

        // The number of times this lock has been taken recursively
        long _M_recursionCount;

        // The current owner of the lock
        volatile long _M_owner;
    };

    struct _Chore
    {
    protected:
        // Constructors.
        explicit _Chore(TaskProc _PFunction) : m_pFunction(_PFunction)
        {
        }

        _Chore()
        {
        }

        virtual ~_Chore()
        {
        }

    public:

        // The function which invokes the work of the chore.
        TaskProc m_pFunction;
    };

    // _UnrealizedChore represents an unrealized chore -- a unit of work that scheduled in a work
    // stealing capacity. Some higher level construct (language or library) will map atop this to provide
    // a usable abstraction to clients.
    class _UnrealizedChore : public _Chore, public _AllocBase
    {
    public:
        // Constructor for an unrealized chore.
        _UnrealizedChore() :
            _M_pTaskCollection(NULL)
        {
        }
        virtual ~_UnrealizedChore() {}


        // Method that executes the unrealized chore.
        void _Invoke()
        {
            _M_pChoreFunction(this);
        }

        // Sets the attachment state of the chore at the time of stealing.
        void _SetDetached(bool _FDetached);

        // Returns the owning collection of the chore.
        ::Concurrency::details::_TaskCollectionBase* _OwningCollection() const
        {
            return _M_pTaskCollection;
        }

        // Set flag that indicates whether the scheduler owns the lifetime of the object and is responsible for freeing it.
        // The flag is ignored by _StructuredTaskCollection
        void _SetRuntimeOwnsLifetime(bool _FValue)
        {
            _M_fRuntimeOwnsLifetime = _FValue;
        }

        // Returns the flag that indicates whether the scheduler owns the lifetime of the object and is responsible for freeing it.
        // The flag is ignored by _StructuredTaskCollection
        bool _GetRuntimeOwnsLifetime() const
        {
            return _M_fRuntimeOwnsLifetime;
        }

        // Allocator to be used when runtime owns lifetime.
        template <typename _ChoreType, typename _Function>
        static _ChoreType * _InternalAlloc(const _Function& _Func)
        {
            // This is always invoked from the PPL layer by the user and can never be attached to the default scheduler. Therefore '_concrt_new' is not required here
            _ChoreType * _Chore = new _ChoreType(_Func);
            _Chore->_M_fRuntimeOwnsLifetime = true;
            return _Chore;
        }

        // Internal helper routine to prepare for execution as a stolen chore.
        void _PrepareSteal(ContextBase *_PContext);

    protected:
        // Invocation bridge between the _UnrealizedChore and PPL.
        template <typename _ChoreType>
        static void __cdecl _InvokeBridge(void * _PContext)
        {
            auto _PChore = static_cast<_ChoreType *>(_PContext);
            (*_PChore)();
        }

        // Place associated task collection in a safe state.
        _CONCRTIMP void _CheckTaskCollection();

    private:

        friend class _StructuredTaskCollection;
        friend class _TaskCollection;
        typedef void (__cdecl * CHOREFUNC)(_UnrealizedChore * _PChore);

        // The collection of work to which this particular chore belongs.
        ::Concurrency::details::_TaskCollectionBase * _M_pTaskCollection;

        // Internal invocation inside the scheduler.
        CHOREFUNC _M_pChoreFunction;

        // Indicates whether the scheduler owns the lifetime of the object and is responsible for freeing it.
        // This flag is ignored by _StructuredTaskCollection
        bool _M_fRuntimeOwnsLifetime;

        // An indication of whether the chore (if stolen) was detached.
        bool _M_fDetached;

        // Helper routines
        void _PrepareStealStructured(ContextBase *_PContext);
        void _PrepareStealUnstructured(ContextBase *_PContext);

        // The internal wrapper around invocation of stolen structured chores.
        __declspec(noinline)
        static void __cdecl _StructuredChoreWrapper(_UnrealizedChore * _PChore);

        // The internal wrapper around invocation of stolen unstructured chores.
        __declspec(noinline)
        static void __cdecl _UnstructuredChoreWrapper(_UnrealizedChore * _PChore);

        // To free memory allocated with _InternalAlloc.
        static void _InternalFree(_UnrealizedChore * _PChore);

        // Cancellation via token to a stolen chore
        static void __cdecl _CancelViaToken(::Concurrency::details::ContextBase *_PContext);
    };

    // Represents possible results of waiting on a task collection.
    enum _TaskCollectionStatus
    {
        _NotComplete,
        _Completed,
        _Canceled
    };

    // _TaskCollectionBase represents an abstract set of work and provides shared waiting semantics for stolen work.
    class _TaskCollectionBase
    {
    public:
        // Constructs a new task collection.
        _TaskCollectionBase() :
            _M_pTokenState(NULL),
            _M_completedStolenChores(_CollectionNotInitialized),
            _M_unpoppedChores(0),
            _M_pException(NULL),
            _M_inliningDepth(_S_notInlined)
        {
        }
        _TaskCollectionBase(const _TaskCollectionBase&) = delete;

        // Constructs a new task collection based on a given cancellation token.
        _TaskCollectionBase(_CancellationTokenState *_PTokenState) :
            _M_pTokenState(_PTokenState),
            _M_completedStolenChores(_CollectionNotInitialized),
            _M_unpoppedChores(0),
            _M_pException(NULL),
            _M_inliningDepth(_S_notInlined)
        {
        }

        _TaskCollectionBase& operator=(const _TaskCollectionBase&) = delete;

        // Returns the owning context of the task collection.
        void * _OwningContext() const
        {
            return _M_pOwningContext;
        }

        // Returns the inlining depth.
        int _InliningDepth() const
        {
            return _M_inliningDepth;
        }

        // Tells if the task collection is inlined - some thread somewhere is currently invoking wait on it.
        bool _IsCurrentlyInlined() const
        {
            return (_M_inliningDepth != _S_notInlined);
        }

        // Returns whether this is a structured collection or not.
        bool _IsStructured()
        {
            return (_M_inlineFlags & _S_structured) != 0;
        }

        // Returns the token state associated with this task collection
        _CancellationTokenState *_GetTokenState(_CancellationTokenRegistration **_PRegistration = NULL);

    protected:

        friend class ::Concurrency::details::_UnrealizedChore;
        friend class ::Concurrency::details::ContextBase;

        enum _TaskCollectionBaseState
        {
            _CollectionNotInitialized = LONG_MIN,
            _CollectionInitializationInProgress = LONG_MIN+1,
            _CollectionInitialized = 0
        };

        // Returns the exception portion of _M_pException.
        std::exception_ptr * _Exception() const
        {
            return (std::exception_ptr *) ((size_t)_M_pException & ~_S_cancelBitsMask);
        }

        // Indicates whether or not this task collection has an abnormal exit.
        bool _IsAbnormalExit() const
        {
            return _M_pException != NULL;
        }

        // Returns the cancel flags.
        size_t _CancelState() const
        {
            return (size_t) _M_pException & _S_cancelBitsMask;
        }

        // Returns whether or not the collection is marked for cancellation.
        bool _IsMarkedForCancellation() const
        {
            return (_CancelState() & _S_cancelBitsMask) != 0;
        }

        // Returns whether an inline cancellation was performed.
        bool _PerformedInlineCancel() const
        {
            _CONCRT_ASSERT(_CancelState() != _S_cancelStarted);
            return _CancelState() == _S_cancelShotdownOwner;
        }

        bool _PerformedPendingCancel() const
        {
            _CONCRT_ASSERT(_CancelState() != _S_cancelStarted);
            return _CancelState() == _S_cancelDeferredShootdownOwner;
        }

        // Returns the parent collection safely.
        _TaskCollectionBase *_SafeGetParent()
        {
            return ((_M_inliningDepth != _S_notInlined) ? _M_pParent : NULL);
        }

        // Called in order to determine whether this task collection will interrupt for a pending cancellation at or above it.
        bool _WillInterruptForPendingCancel();

        // Called when an exception is raised on a chore on a given task collection, this makes a determination of what to do with the exception
        // and saves it for potential transport back to the thread performing a join on a chore collection.
        void _RaisedException();

        // Potentially rethrows the exception which was set with _RaisedException. The caller has responsibility to ensure that _RaisedException
        //  was called prior to calling this and that _M_pException has progressed beyond the _S_nonNull state.
        void _RethrowException();

        // Marks the collection for cancellation and returns whether the collection was marked.
        bool _MarkCancellation();

        // Finishes the cancellation state (changing from _S_cancelStarted to one of the other states). Note that only the
        // thread which successfully marked cancellation can call this.
        void _FinishCancelState(size_t _NewCancelState);

        // Called when a cancellation is raised on a chore on a given task collection. This makes a determination of what to do with the exception
        // and saves it for potential transport back to the thread performing a join on a chore collection. Note that every other exception
        // has precedence over a cancellation.
        void _RaisedCancel();

        // Tracks the parent collection. (For example, A task collection B created during execution of a chore C on task collection A is
        // considered a child of A).
        _TaskCollectionBase * _M_pParent;

        // Tracks the inlining depth of this collection for cancellation purposes and packs a series of definition bits.
        int _M_inliningDepth : 28;
        int _M_inlineFlags : 4;

        // The cancellation token for the task collection.
        _CancellationTokenState *_M_pTokenState;

        // The context which owns the task collection. This is the context where the collection is created.
        void * _M_pOwningContext;

        // The number of unpopped chores associated with the task collection (set by the derived
        // class during chore association.
        long _M_unpoppedChores;

        // The number of stolen chores executed so far.
        volatile long _M_completedStolenChores;

        // The stored exception which has been marshaled from the thread a stolen chore ran upon to the thread that is waiting on the
        // task collection.
        //
        // The lower two bits of _M_pException are utilized for the cancellation state machine. The upper 30 are the exception pointer. This implies
        // that the exception pointer must be 4-byte aligned. Because of intermediate states, the exception pointer cannot be between 0x8 and 0xF. The heap should
        // not be allocating such...
        //
        std::exception_ptr * _M_pException;

        // Cancellation states
        static const size_t _S_cancelBitsMask = 0x3;
        static const size_t _S_cancelNone = 0x0;
        static const size_t _S_cancelStarted = 0x1;
        static const size_t _S_cancelDeferredShootdownOwner = 0x2;
        static const size_t _S_cancelShotdownOwner = 0x3;

        // Intermediate exceptions.
        static const size_t _S_nonNull = 0x8;
        static const size_t _S_cancelException = 0xC;

        // initialization state for inlining depth.
        static const int _S_notInlined = -1;

        // Inline flags.
        static const int _S_structured = 0x00000001;
        static const int _S_localCancel = 0x00000002;
        static const int _S_reserved = 0x0000000C;
    };

    
    class _StructuredTaskCollection : public _TaskCollectionBase
    {
    public:

        
        _StructuredTaskCollection()
        {
            _Construct();
            _M_pTokenState = NULL;
        }

        _StructuredTaskCollection(const _StructuredTaskCollection&) = delete;
        _StructuredTaskCollection& operator=(const _StructuredTaskCollection&) = delete;
        
        _CONCRTIMP _StructuredTaskCollection(_CancellationTokenState *_PTokenState);

        
        _CONCRTIMP ~_StructuredTaskCollection();

        
        _CONCRTIMP void _Schedule(_UnrealizedChore * _PChore, location * _PLocation);

        
        _CONCRTIMP void _Schedule(_UnrealizedChore * _PChore);

        
        _CONCRTIMP void _Cancel();

        
        _CONCRTIMP bool _IsCanceling();

        
        _CONCRTIMP _TaskCollectionStatus __stdcall _RunAndWait(_UnrealizedChore * _PChore = NULL);

        
        _TaskCollectionStatus _Wait()
        {
            return _RunAndWait();
        }

        
        void _CancelStolenContexts();

    private:

        friend class _UnrealizedChore;

        void _Construct()
        {
            _M_pOwningContext = NULL;
            _M_inlineFlags = _S_structured;
        }

        
        _CONCRTIMP void _Abort();

        _CONCRTIMP void _CleanupToken();

        
        bool _TaskCleanup()
        {
            //
            // Users are required to call Wait() before letting the destructor run. Otherwise, throw. Note that before throwing,
            // we must actually wait on the tasks because they contain pointers into stack frames and unwinding without the wait is
            // instant stack corruption.
            //
            if (_M_unpoppedChores > 0)
            {
                _Abort();

                if (!__uncaught_exception())
                {
                    return false;
                }
            }

            return true;
        }

        
        void _Initialize();

        
        void _WaitOnStolenChores(long _StolenChoreCount);

        
        void _CountUp();

        
        static void __cdecl _CancelViaToken(_StructuredTaskCollection *_PCollection);

        //
        // _StructuredTaskCollection::_M_event is used to construct an structured event object only when it is needed to block. The structured event object
        //  has no state to cleanup, therefore no dtor code is required.
        //
        _CONCRT_BUFFER _M_event[(sizeof(void*) + sizeof(_CONCRT_BUFFER) - 1) / sizeof(_CONCRT_BUFFER)];
    };

    
    class _TaskCollection : public _TaskCollectionBase
    {
    public:

        
        _CONCRTIMP _TaskCollection();
        _TaskCollection(const _TaskCollection&) = delete;

        _TaskCollection& operator=(const _TaskCollection&) = delete;
        
        _CONCRTIMP _TaskCollection(_CancellationTokenState *_PTokenState);

        
        _CONCRTIMP ~_TaskCollection();

        
        _CONCRTIMP void _Schedule(_UnrealizedChore * _PChore, location * _PLocation);

        
        _CONCRTIMP void _Schedule(_UnrealizedChore * _PChore);

        
        _CONCRTIMP void _Cancel();

        
        _CONCRTIMP bool _IsCanceling();

        
        _CONCRTIMP _TaskCollectionStatus __stdcall _RunAndWait(_UnrealizedChore * _PChore = NULL);

        
        _TaskCollectionStatus _Wait()
        {
            return _RunAndWait();
        }

        
        bool _IsMarkedForAbnormalExit() const;

        
        _TaskCollection * _OriginalCollection() const;

        
        bool _IsAlias() const;

        
        void _RegisterCompletionHandler(TaskProc _Func, void * _PCompletionContext);

    private:

        friend class _UnrealizedChore;
        friend class ::Concurrency::details::ContextBase;

        
        bool _IsStaleAlias() const;

        
        void _ReleaseAlias();

        
        _TaskCollection(_TaskCollection * _POriginCollection, bool _FDirectAlias);

        
        _TaskCollection * _Alias();

        
        void _Abort(bool _FLeaveCanceled = false);

        
        bool _IsIndirectAlias() const;

        
        bool _IsDirectAlias() const;

        
        bool _HasDirectAlias() const;

        
        void _Cancel(bool _InsideException, _TaskCollection * _PSnapPoint);

        
        void _NotifyNewChore();

        
        void _NotifyCompletedChoreAndFree(_UnrealizedChore * _PChore = NULL);

        
        void _FullAliasWait(_TaskCollection * _PSnapPoint);

        
        void _Reset(_TaskCollection * _PSnapPoint);

        
        void _RaisedException();

        
        void _RaisedCancel();

        
        bool _SetCancelState(long _Status);

        
        void _CancelFromArbitraryThread(bool _InsideException);

        
        void _CancelDirectAliases(bool _InsideException, _TaskCollection * _PSnapPoint);

        
        void _CancelStolenContexts(bool _InsideException, bool _FInlineGated);

        
        void *_GetStealTrackingList() const;

        
        void _Initialize();

        
        void _AbortiveSweep(void *_PCtx);

        
        static bool __cdecl _CollectionMatchPredicate(_UnrealizedChore *_PChore, void *_PData);

        
        static bool __cdecl _SweepAbortedChore(_UnrealizedChore *_PChore, void *_PData);

        
        bool _TaskCleanup(bool _FExceptional);

        
        static void __cdecl _CancelViaToken(_TaskCollection *_PCollection);

        
        _CONCRT_BUFFER _M_stealTracker[_SAFERWLIST_SIZE];

        
        long _M_activeStealersForCancellation;

        
        volatile long _M_exitCode;

        
        volatile long _M_executionStatus;

        
        event _M_event;

        _TaskCollection * _M_pOriginalCollection;
        _TaskCollection * _M_pNextAlias;
        void * _M_pTaskExtension;

        int _M_taskCookies[2];

        volatile long _M_flags;
        volatile long _M_chaining;

        DWORD _M_boundQueueId;
        int _M_stackPos;

        TaskProc _M_completionHandler;
        void * _M_pCompletionContext;
    };

    class _StackGuard
    {
    public:
        _StackGuard() : _Depth(_GetCurrentInlineDepth())
        {
            // _Depth is the reference to the depth slot on context.
            ++_Depth;
        }
        _StackGuard(const _StackGuard&) = delete;

        _StackGuard& operator=(const _StackGuard&) = delete;

        ~_StackGuard()
        {
            // _Depth is the reference to the depth slot on context.
            --_Depth;
        }

        bool _ShouldInline(_TaskInliningMode _InliningMode) const
        {
            // As _TaskInliningMode is defined as inlining threshold, we can directly convert
            // it into size_t, and compare with current context inlining depth.
            return _Depth <= static_cast<size_t>(_InliningMode);
        }
    private:
        size_t & _Depth;

        _CONCRTIMP static size_t & __cdecl _GetCurrentInlineDepth();
    };

    class _AsyncTaskCollection : public _RefCounterBase
    {
    public:
        _AsyncTaskCollection(const _AsyncTaskCollection&) = delete;
        _AsyncTaskCollection& operator=(const _AsyncTaskCollection&) = delete;
        _CONCRTIMP static _AsyncTaskCollection * __cdecl _NewCollection(_CancellationTokenState *_PTokenState);

        _TaskCollectionStatus _ScheduleWithAutoInline(_UnrealizedChore * _PChore, _TaskInliningMode _InliningMode)
        {
            _CONCRT_ASSERT(_PChore != nullptr);
            _Reference();

            if (_InliningMode == _NoInline)
            {
                _M_taskCollection._Schedule(_PChore);
                return _NotComplete;
            }
            else
            {
                _StackGuard _Guard;
                if (_Guard._ShouldInline(_InliningMode))
                {
                    return _M_taskCollection._RunAndWait(_PChore);
                }
                else
                {
                    _M_taskCollection._Schedule(_PChore);
                    return _NotComplete;
                }
            }
        }

        void _Cancel()
        {
            _M_taskCollection._Cancel();
        }

        _TaskCollectionStatus _RunAndWait()
        {
            // Note that _Guard is NOT unused variable, the constructor and destructor will be called to maintain inline depth.
            _StackGuard _Guard;
            return _M_taskCollection._RunAndWait();
        }

    private:

        void _NotificationHandler();

        _CONCRTIMP virtual void _Destroy();

        // Private constructor
        _AsyncTaskCollection(_CancellationTokenState *_PTokenState);

        __declspec(noinline)
        static void __cdecl _CompletionHandler(void * _PCompletionContext);

    private:

        // Underlying task collection where the chore is scheduled to run
        _TaskCollection _M_taskCollection;
    };

    struct _Beacon_reference
    {
        volatile long _M_signals;
    };

    class _Cancellation_beacon
    {
    public:

        _CONCRTIMP _Cancellation_beacon();

        _CONCRTIMP ~_Cancellation_beacon();

        bool _Is_signaled() const
        {
            return (_M_pRef->_M_signals != 0);
        }

        // This method should only be called when the beacon is signaled. It confirms whether a cancellation is indeed happening and that the beacon
        // was not flagged due to a false positive race. If the cancellation is not confirmed, the beacon is lowered.
        _CONCRTIMP bool _Confirm_cancel();

        void _Raise()
        {
            _InterlockedIncrement(&_M_pRef->_M_signals);
        }

        void _Lower()
        {
            _InterlockedDecrement(&_M_pRef->_M_signals);
        }

    private:

        _Beacon_reference *_M_pRef;

    };

    //
    // Internal stub class.
    //
    class _TimerStub;

    //
    // Internal wrapper around timers in order to allow timer messaging blocks to share implementation with internal ConcRT runtime
    // timers.
    //
    class _Timer
    {
    protected:
        // Constructs a new timer.
        //
        // _Ms: The duration and period of the timer in milliseconds.
        // _FRepeating: An indication of whether the timer is repeating (periodic) or not.
        _CONCRTIMP _Timer(unsigned int _Ms, bool _FRepeating);

        // Destroys the timer.
        _CONCRTIMP virtual ~_Timer();

        // Starts the timer.
        _CONCRTIMP void _Start();

        // Stops the timer.
        _CONCRTIMP void _Stop();

    private:
        friend class _TimerStub;

        // Called when the timer fires.
        virtual void _Fire() =0;

        // The actual timer
        HANDLE _M_hTimer;

        // The duration and period of the timer.
        unsigned int _M_ms;

        // Whether the timer is repeating (periodic by _M_ms)
        bool _M_fRepeating;
    };

    //
    // Internal runtime structure that holds the trace flags and level for ETW events
    // provided by the Concurrent runtime.
    //
    struct _CONCRT_TRACE_INFO
    {
        volatile unsigned long EnableFlags;    // Determines which class of events to log
        volatile unsigned char EnableLevel;    // Determines the severity of events to log

        void _EnableTrace(unsigned char _Level, unsigned long _Flags)
        {
            EnableFlags = _Flags;
            EnableLevel = _Level;
        }
#pragma warning ( push )
#pragma warning ( disable : 5393 )  // unreachable code
        void _DisableTrace()
        {
            EnableLevel = 0;
            EnableFlags = 0;
        }
#pragma warning ( pop )

        bool _IsEnabled(unsigned char _Level, unsigned long _Flags) const
        {
            return ((_Level <= EnableLevel) &&  ((EnableFlags & _Flags) == _Flags));
        }
    };

    
    _CONCRTIMP const _CONCRT_TRACE_INFO * _GetConcRTTraceInfo();

    
    void _RegisterConcRTEventTracing();

    
    void _UnregisterConcRTEventTracing();

} // namespace details



__declspec(deprecated("Concurrency::EnableTracing is a deprecated function.")) _CONCRTIMP HRESULT __cdecl EnableTracing();


__declspec(deprecated("Concurrency::DisableTracing is a deprecated function.")) _CONCRTIMP HRESULT __cdecl DisableTracing();


enum ConcRT_EventType
{
    
    CONCRT_EVENT_GENERIC    = 0,
    
    CONCRT_EVENT_START      = 1,
    
    CONCRT_EVENT_END        = 2,
    
    CONCRT_EVENT_BLOCK      = 3,
    
    CONCRT_EVENT_UNBLOCK    = 4,
    
    CONCRT_EVENT_YIELD      = 5,
    
    CONCRT_EVENT_IDLE       = 6,
    
    CONCRT_EVENT_ATTACH     = 7,
    
    CONCRT_EVENT_DETACH     = 8,
};

// Common trace header structure for all ConcRT diagnostic events
//      struct CONCRT_TRACE_EVENT_HEADER_COMMON
//      {
//          EVENT_TRACE_HEADER header;
//          DWORD VirtualProcessorID;
//          DWORD SchedulerID;
//          DWORD ContextID;
//          DWORD ScheduleGroupID;
//      };


_CONCRTIMP extern const GUID ConcRT_ProviderGuid;

//
// GUIDS for events
//


_CONCRTIMP extern const GUID ConcRTEventGuid;


_CONCRTIMP extern const GUID SchedulerEventGuid;


_CONCRTIMP extern const GUID ScheduleGroupEventGuid;


_CONCRTIMP extern const GUID ContextEventGuid;


_CONCRTIMP extern const GUID ChoreEventGuid;


_CONCRTIMP extern const GUID VirtualProcessorEventGuid;


_CONCRTIMP extern const GUID LockEventGuid;


_CONCRTIMP extern const GUID ResourceManagerEventGuid;


_CONCRTIMP extern const GUID PPLParallelInvokeEventGuid;


_CONCRTIMP extern const GUID PPLParallelForEventGuid;


_CONCRTIMP extern const GUID PPLParallelForeachEventGuid;


_CONCRTIMP extern const GUID AgentEventGuid;

// Trace an event signaling a parallel function
_CONCRTIMP void __cdecl _Trace_ppl_function(const GUID& _Guid, unsigned char _Level, ConcRT_EventType _Type);


enum Concrt_TraceFlags
{
    SchedulerEventFlag              = 0x1,
    ContextEventFlag                = 0x2,
    VirtualProcessorEventFlag       = 0x4,
    ResourceManagerEventFlag        = 0x8,
    PPLEventFlag                    = 0x10,
    AgentEventFlag                  = 0x20,

    AllEventsFlag                   = 0xFFFFFFFF
};


enum Agents_EventType
{
    
    AGENTS_EVENT_CREATE   = 0,

    
    AGENTS_EVENT_START    = 1,

    
    AGENTS_EVENT_END      = 2,

    
    AGENTS_EVENT_DESTROY  = 3,

    
    AGENTS_EVENT_SCHEDULE = 4,

    
    AGENTS_EVENT_LINK     = 5,

    
    AGENTS_EVENT_UNLINK   = 6,

    
    AGENTS_EVENT_NAME     = 7

};

//        // Common trace payload for agents
//
//        struct AGENTS_TRACE_PAYLOAD
//        {
//            // Identifier of the agent or message block that is emitting the event
//            __int64  AgentId1;
//            union
//            {
//                // The identifier of a target block for link/unlink event
//                __int64  AgentId2;
//
//                // Count of messages processed for the end event
//                long    Count;
//
//                // Name of this agent for the purposes of the ETW trace
//                wchar_t Name[32];
//            };
//        };

// Emit a trace event specific to the agents library of the given type and payload
_CONCRTIMP void __cdecl _Trace_agents(Agents_EventType _Type, __int64 _AgentId, ...);
}

#ifndef _NO_DEFAULT_CONCRT_LIB

#undef _DEBUG_AFFIX
#undef _IDL_AFFIX
#undef _IDL_DEFAULT
#undef _LIB_STEM

#ifdef _DEBUG
    #define _DEBUG_AFFIX "d"
    #define _IDL_DEFAULT 2
#else
    #define _DEBUG_AFFIX ""
    #define _IDL_DEFAULT 0
#endif /* _DEBUG */

#if defined(_DLL) && !defined(_STATIC_CPPLIB)
    #define _LIB_STEM "concrt"
#else /* defined(_DLL) && !defined(_STATIC_CPPLIB) */
    #define _LIB_STEM "libconcrt"
    #if _ITERATOR_DEBUG_LEVEL != _IDL_DEFAULT
        #define _IDL_AFFIX _CRT_STRINGIZE(_ITERATOR_DEBUG_LEVEL)
    #endif /* _ITERATOR_DEBUG_LEVEL != _IDL_DEFAULT */
#endif /* defined(_DLL) && !defined(_STATIC_CPPLIB) */

#ifndef _IDL_AFFIX
    #define _IDL_AFFIX ""
#endif /* _IDL_AFFIX */

#pragma comment(lib, _LIB_STEM _DEBUG_AFFIX _IDL_AFFIX)

#undef _DEBUG_AFFIX
#undef _IDL_AFFIX
#undef _IDL_DEFAULT
#undef _LIB_STEM

#endif /* _NO_DEFAULT_CONCRT_LIB */

namespace concurrency = ::Concurrency;

#pragma pop_macro("_YieldProcessor")
#pragma pop_macro("new")
#pragma pack(pop)