https://git.project-hobbit.eu/dj16/ricec/blob/c9d3dceb1c3b1c03a42077e0461e3ce5a2615a51/data/atomicops.h#L248

// ©2013-2016 Cameron Desrochers.
// Distributed under the simplified BSD license (see the license file that
// should have come with this header).
// Uses Jeff Preshing's semaphore implementation (under the terms of its
// separate zlib license, embedded below).

#pragma once

// Provides portable (VC++2010+, Intel ICC 13, GCC 4.7+, and anything C++11 compliant) implementation
// of low-level memory barriers, plus a few semi-portable utility macros (for inlining and alignment).
// Also has a basic atomic type (limited to hardware-supported atomics with no memory ordering guarantees).
// Uses the AE_* prefix for macros (historical reasons), and the "moodycamel" namespace for symbols.

#include <cassert>
#include <type_traits>
#include <cerrno>
#include <cstdint>
#include <ctime>

// Platform detection
#if defined(__INTEL_COMPILER)
#define AE_ICC
#elif defined(_MSC_VER)
#define AE_VCPP
#elif defined(__GNUC__)
#define AE_GCC
#endif

#if defined(_M_IA64) || defined(__ia64__)
#define AE_ARCH_IA64
#elif defined(_WIN64) || defined(__amd64__) || defined(_M_X64) || defined(__x86_64__)
#define AE_ARCH_X64
#elif defined(_M_IX86) || defined(__i386__)
#define AE_ARCH_X86
#elif defined(_M_PPC) || defined(__powerpc__)
#define AE_ARCH_PPC
#else
#define AE_ARCH_UNKNOWN
#endif

// AE_UNUSED
#define AE_UNUSED(x) ((void)x)

// AE_FORCEINLINE
#if defined(AE_VCPP) || defined(AE_ICC)
#define AE_FORCEINLINE __forceinline
#elif defined(AE_GCC)
//#define AE_FORCEINLINE __attribute__((always_inline)) 
#define AE_FORCEINLINE inline
#else
#define AE_FORCEINLINE inline
#endif

// AE_ALIGN
#if defined(AE_VCPP) || defined(AE_ICC)
#define AE_ALIGN(x) __declspec(align(x))
#elif defined(AE_GCC)
#define AE_ALIGN(x) __attribute__((aligned(x)))
#else
// Assume GCC compliant syntax...
#define AE_ALIGN(x) __attribute__((aligned(x)))
#endif

// Portable atomic fences implemented below:

namespace concurent {

enum memory_order {
    memory_order_relaxed,
    memory_order_acquire,
    memory_order_release,
    memory_order_acq_rel,
    memory_order_seq_cst,

    // memory_order_sync: Forces a full sync:
    // #LoadLoad, #LoadStore, #StoreStore, and most significantly, #StoreLoad
    memory_order_sync = memory_order_seq_cst
};

}    // end namespace moodycamel

#if (defined(AE_VCPP) && (_MSC_VER < 1700 || defined(__cplusplus_cli))) || defined(AE_ICC)
// VS2010 and ICC13 don't support std::atomic_*_fence, implement our own fences

#include <intrin.h>

#if defined(AE_ARCH_X64) || defined(AE_ARCH_X86)
#define AeFullSync _mm_mfence
#define AeLiteSync _mm_mfence
#elif defined(AE_ARCH_IA64)
#define AeFullSync __mf
#define AeLiteSync __mf
#elif defined(AE_ARCH_PPC)
#include <ppcintrinsics.h>
#define AeFullSync __sync
#define AeLiteSync __lwsync
#endif

#ifdef AE_VCPP
#pragma warning(push)
#pragma warning(disable: 4365)        // Disable erroneous 'conversion from long to unsigned int, signed/unsigned mismatch' error when using `assert`
#ifdef __cplusplus_cli
#pragma managed(push, off)
#endif
#endif

namespace moodycamel {

AE_FORCEINLINE void compiler_fence(memory_order order)
{
    switch (order) {
        case memory_order_relaxed: break;
        case memory_order_acquire: _ReadBarrier(); break;
        case memory_order_release: _WriteBarrier(); break;
        case memory_order_acq_rel: _ReadWriteBarrier(); break;
        case memory_order_seq_cst: _ReadWriteBarrier(); break;
        default: assert(false);
    }
}

// x86/x64 have a strong memory model -- all loads and stores have
// acquire and release semantics automatically (so only need compiler
// barriers for those).
#if defined(AE_ARCH_X86) || defined(AE_ARCH_X64)
AE_FORCEINLINE void fence(memory_order order)
{
    switch (order) {
        case memory_order_relaxed: break;
        case memory_order_acquire: _ReadBarrier(); break;
        case memory_order_release: _WriteBarrier(); break;
        case memory_order_acq_rel: _ReadWriteBarrier(); break;
        case memory_order_seq_cst:
            _ReadWriteBarrier();
            AeFullSync();
            _ReadWriteBarrier();
            break;
        default: assert(false);
    }
}
#else
AE_FORCEINLINE void fence(memory_order order)
{
    // Non-specialized arch, use heavier memory barriers everywhere just in case :-(
    switch (order) {
        case memory_order_relaxed:
            break;
        case memory_order_acquire:
            _ReadBarrier();
            AeLiteSync();
            _ReadBarrier();
            break;
        case memory_order_release:
            _WriteBarrier();
            AeLiteSync();
            _WriteBarrier();
            break;
        case memory_order_acq_rel:
            _ReadWriteBarrier();
            AeLiteSync();
            _ReadWriteBarrier();
            break;
        case memory_order_seq_cst:
            _ReadWriteBarrier();
            AeFullSync();
            _ReadWriteBarrier();
            break;
        default: assert(false);
    }
}
#endif
}    // end namespace moodycamel
#else
// Use standard library of atomics
#include <atomic>

namespace concurent {

AE_FORCEINLINE void compiler_fence(memory_order order)
{
    switch (order) {
        case memory_order_relaxed: break;
        case memory_order_acquire: std::atomic_signal_fence(std::memory_order_acquire); break;
        case memory_order_release: std::atomic_signal_fence(std::memory_order_release); break;
        case memory_order_acq_rel: std::atomic_signal_fence(std::memory_order_acq_rel); break;
        case memory_order_seq_cst: std::atomic_signal_fence(std::memory_order_seq_cst); break;
        default: assert(false);
    }
}

AE_FORCEINLINE void fence(memory_order order)
{
    switch (order) {
        case memory_order_relaxed: break;
        case memory_order_acquire: std::atomic_thread_fence(std::memory_order_acquire); break;
        case memory_order_release: std::atomic_thread_fence(std::memory_order_release); break;
        case memory_order_acq_rel: std::atomic_thread_fence(std::memory_order_acq_rel); break;
        case memory_order_seq_cst: std::atomic_thread_fence(std::memory_order_seq_cst); break;
        default: assert(false);
    }
}

}    // end namespace moodycamel

#endif

#if !defined(AE_VCPP) || (_MSC_VER >= 1700 && !defined(__cplusplus_cli))
#define AE_USE_STD_ATOMIC_FOR_WEAK_ATOMIC
#endif

#ifdef AE_USE_STD_ATOMIC_FOR_WEAK_ATOMIC
#include <atomic>
#endif
#include <utility>

// WARNING: *NOT* A REPLACEMENT FOR std::atomic. READ CAREFULLY:
// Provides basic support for atomic variables -- no memory ordering guarantees are provided.
// The guarantee of atomicity is only made for types that already have atomic load and store guarantees
// at the hardware level -- on most platforms this generally means aligned pointers and integers (only).
namespace concurent {
template<typename T>
class weak_atomic
{
public:
    weak_atomic() { }
#ifdef AE_VCPP
#pragma warning(disable: 4100)        // Get rid of (erroneous) 'unreferenced formal parameter' warning
#endif
    template<typename U> weak_atomic(U&& x) : value(std::forward<U>(x)) {  }
#ifdef __cplusplus_cli
    // Work around bug with universal reference/nullptr combination that only appears when /clr is on
    weak_atomic(nullptr_t) : value(nullptr) {  }
#endif
    weak_atomic(weak_atomic const& other) : value(other.value) {  }
    weak_atomic(weak_atomic&& other) : value(std::move(other.value)) {  }
#ifdef AE_VCPP
#pragma warning(default: 4100)
#endif

    AE_FORCEINLINE operator T() const { return load(); }

    
#ifndef AE_USE_STD_ATOMIC_FOR_WEAK_ATOMIC
    template<typename U> AE_FORCEINLINE weak_atomic const& operator=(U&& x) { value = std::forward<U>(x); return *this; }
    AE_FORCEINLINE weak_atomic const& operator=(weak_atomic const& other) { value = other.value; return *this; }
    
    AE_FORCEINLINE T load() const { return value; }
    
    AE_FORCEINLINE T fetch_add_acquire(T increment)
    {
#if defined(AE_ARCH_X64) || defined(AE_ARCH_X86)
        if (sizeof(T) == 4) return _InterlockedExchangeAdd((long volatile*)&value, (long)increment);
#if defined(_M_AMD64)
        else if (sizeof(T) == 8) return _InterlockedExchangeAdd64((long long volatile*)&value, (long long)increment);
#endif
#else
#error Unsupported platform
#endif
        assert(false && "T must be either a 32 or 64 bit type");
        return value;
    }
    
    AE_FORCEINLINE T fetch_add_release(T increment)
    {
#if defined(AE_ARCH_X64) || defined(AE_ARCH_X86)
        if (sizeof(T) == 4) return _InterlockedExchangeAdd((long volatile*)&value, (long)increment);
#if defined(_M_AMD64)
        else if (sizeof(T) == 8) return _InterlockedExchangeAdd64((long long volatile*)&value, (long long)increment);
#endif
#else
#error Unsupported platform
#endif
        assert(false && "T must be either a 32 or 64 bit type");
        return value;
    }
#else
    template<typename U>
    AE_FORCEINLINE weak_atomic const& operator=(U&& x)
    {
        value.store(std::forward<U>(x), std::memory_order_relaxed);
        return *this;
    }
    
    AE_FORCEINLINE weak_atomic const& operator=(weak_atomic const& other)
    {
        value.store(other.value.load(std::memory_order_relaxed), std::memory_order_relaxed);
        return *this;
    }

    AE_FORCEINLINE T load() const { return value.load(std::memory_order_relaxed); }
    
    AE_FORCEINLINE T fetch_add_acquire(T increment)
    {
        return value.fetch_add(increment, std::memory_order_acquire);
    }
    
    AE_FORCEINLINE T fetch_add_release(T increment)
    {
        return value.fetch_add(increment, std::memory_order_release);
    }
#endif
    

private:
#ifndef AE_USE_STD_ATOMIC_FOR_WEAK_ATOMIC
    // No std::atomic support, but still need to circumvent compiler optimizations.
    // `volatile` will make memory access slow, but is guaranteed to be reliable.
    volatile T value;
#else
    std::atomic<T> value;
#endif
};

}    // end namespace moodycamel

// Portable single-producer, single-consumer semaphore below:

#if defined(_WIN32)
// Avoid including windows.h in a header; we only need a handful of
// items, so we'll redeclare them here (this is relatively safe since
// the API generally has to remain stable between Windows versions).
// I know this is an ugly hack but it still beats polluting the global
// namespace with thousands of generic names or adding a .cpp for nothing.
extern "C" {
    struct _SECURITY_ATTRIBUTES;
    __declspec(dllimport) void* __stdcall CreateSemaphoreW(_SECURITY_ATTRIBUTES* lpSemaphoreAttributes, long lInitialCount, long lMaximumCount, const wchar_t* lpName);
    __declspec(dllimport) int __stdcall CloseHandle(void* hObject);
    __declspec(dllimport) unsigned long __stdcall WaitForSingleObject(void* hHandle, unsigned long dwMilliseconds);
    __declspec(dllimport) int __stdcall ReleaseSemaphore(void* hSemaphore, long lReleaseCount, long* lpPreviousCount);
}
#elif defined(__MACH__)
#include <mach/mach.h>
#elif defined(__unix__)
#include <semaphore.h>
#endif

namespace concurent
{
    // Code in the spsc_sema namespace below is an adaptation of Jeff Preshing's
    // portable + lightweight semaphore implementations, originally from
    // https://github.com/preshing/cpp11-on-multicore/blob/master/common/sema.h
    // LICENSE:
    // Copyright (c) 2015 Jeff Preshing
    //
    // This software is provided 'as-is', without any express or implied
    // warranty. In no event will the authors be held liable for any damages
    // arising from the use of this software.
    //
    // Permission is granted to anyone to use this software for any purpose,
    // including commercial applications, and to alter it and redistribute it
    // freely, subject to the following restrictions:
    //
    // 1. The origin of this software must not be misrepresented; you must not
    //    claim that you wrote the original software. If you use this software
    //    in a product, an acknowledgement in the product documentation would be
    //    appreciated but is not required.
    // 2. Altered source versions must be plainly marked as such, and must not be
    //    misrepresented as being the original software.
    // 3. This notice may not be removed or altered from any source distribution.
    namespace spsc_sema
    {
#if defined(_WIN32)
        class Semaphore
        {
        private:
            void* m_hSema;
            
            Semaphore(const Semaphore& other);
            Semaphore& operator=(const Semaphore& other);

        public:
            Semaphore(int initialCount = 0)
            {
                assert(initialCount >= 0);
                const long maxLong = 0x7fffffff;
                m_hSema = CreateSemaphoreW(nullptr, initialCount, maxLong, nullptr);
            }

            ~Semaphore()
            {
                CloseHandle(m_hSema);
            }

            void wait()
            {
                const unsigned long infinite = 0xffffffff;
                WaitForSingleObject(m_hSema, infinite);
            }

            bool try_wait()
            {
                const unsigned long RC_WAIT_TIMEOUT = 0x00000102;
                return WaitForSingleObject(m_hSema, 0) != RC_WAIT_TIMEOUT;
            }

            bool timed_wait(std::uint64_t usecs)
            {
                const unsigned long RC_WAIT_TIMEOUT = 0x00000102;
                return WaitForSingleObject(m_hSema, (unsigned long)(usecs / 1000)) != RC_WAIT_TIMEOUT;
            }

            void signal(int count = 1)
            {
                ReleaseSemaphore(m_hSema, count, nullptr);
            }
        };
#elif defined(__MACH__)
        //---------------------------------------------------------
        // Semaphore (Apple iOS and OSX)
        // Can't use POSIX semaphores due to http://lists.apple.com/archives/darwin-kernel/2009/Apr/msg00010.html
        //---------------------------------------------------------
        class Semaphore
        {
        private:
            semaphore_t m_sema;

            Semaphore(const Semaphore& other);
            Semaphore& operator=(const Semaphore& other);

        public:
            Semaphore(int initialCount = 0)
            {
                assert(initialCount >= 0);
                semaphore_create(mach_task_self(), &m_sema, SYNC_POLICY_FIFO, initialCount);
            }

            ~Semaphore()
            {
                semaphore_destroy(mach_task_self(), m_sema);
            }

            void wait()
            {
                semaphore_wait(m_sema);
            }

            bool try_wait()
            {
                return timed_wait(0);
            }

            bool timed_wait(std::int64_t timeout_usecs)
            {
                mach_timespec_t ts;
                ts.tv_sec = timeout_usecs / 1000000;
                ts.tv_nsec = (timeout_usecs % 1000000) * 1000;

                // added in OSX 10.10: https://developer.apple.com/library/prerelease/mac/documentation/General/Reference/APIDiffsMacOSX10_10SeedDiff/modules/Darwin.html
                kern_return_t rc = semaphore_timedwait(m_sema, ts);

                return rc != KERN_OPERATION_TIMED_OUT;
            }

            void signal()
            {
                semaphore_signal(m_sema);
            }

            void signal(int count)
            {
                while (count-- > 0)
                {
                    semaphore_signal(m_sema);
                }
            }
        };
#elif defined(__unix__)
        //---------------------------------------------------------
        // Semaphore (POSIX, Linux)
        //---------------------------------------------------------
        class Semaphore
        {
        private:
            sem_t m_sema;

            Semaphore(const Semaphore& other);
            Semaphore& operator=(const Semaphore& other);

        public:
            Semaphore(int initialCount = 0)
            {
                assert(initialCount >= 0);
                sem_init(&m_sema, 0, initialCount);
            }

            ~Semaphore()
            {
                sem_destroy(&m_sema);
            }

            void wait()
            {
                // http://stackoverflow.com/questions/2013181/gdb-causes-sem-wait-to-fail-with-eintr-error
                int rc;
                do
                {
                    rc = sem_wait(&m_sema);
                }
                while (rc == -1 && errno == EINTR);
            }

            bool try_wait()
            {
                int rc;
                do {
                    rc = sem_trywait(&m_sema);
                } while (rc == -1 && errno == EINTR);
                return !(rc == -1 && errno == EAGAIN);
            }

            bool timed_wait(std::uint64_t usecs)
            {
                struct timespec ts;
                const int usecs_in_1_sec = 1000000;
                const int nsecs_in_1_sec = 1000000000;
                clock_gettime(CLOCK_REALTIME, &ts);
                ts.tv_sec += usecs / usecs_in_1_sec;
                ts.tv_nsec += (usecs % usecs_in_1_sec) * 1000;
                // sem_timedwait bombs if you have more than 1e9 in tv_nsec
                // so we have to clean things up before passing it in
                if (ts.tv_nsec > nsecs_in_1_sec) {
                    ts.tv_nsec -= nsecs_in_1_sec;
                    ++ts.tv_sec;
                }

                int rc;
                do {
                    rc = sem_timedwait(&m_sema, &ts);
                } while (rc == -1 && errno == EINTR);
                return !(rc == -1 && errno == ETIMEDOUT);
            }

            void signal()
            {
                sem_post(&m_sema);
            }

            void signal(int count)
            {
                while (count-- > 0)
                {
                    sem_post(&m_sema);
                }
            }
        };
#else
#error Unsupported platform! (No semaphore wrapper available)
#endif

        //---------------------------------------------------------
        // LightweightSemaphore
        //---------------------------------------------------------
        class LightweightSemaphore
        {
        public:
            typedef std::make_signed<std::size_t>::type ssize_t;
            
        private:
            weak_atomic<ssize_t> m_count;
            Semaphore m_sema;

            bool waitWithPartialSpinning(std::int64_t timeout_usecs = -1)
            {
                ssize_t oldCount;
                // Is there a better way to set the initial spin count?
                // If we lower it to 1000, testBenaphore becomes 15x slower on my Core i7-5930K Windows PC,
                // as threads start hitting the kernel semaphore.
                int spin = 10000;
                while (--spin >= 0)
                {
                    if (m_count.load() > 0)
                    {
                        m_count.fetch_add_acquire(-1);
                        return true;
                    }
                    compiler_fence(memory_order_acquire);     // Prevent the compiler from collapsing the loop.
                }
                oldCount = m_count.fetch_add_acquire(-1);
                if (oldCount > 0)
                    return true;
                if (timeout_usecs < 0)
                {
                    m_sema.wait();
                    return true;
                }
                if (m_sema.timed_wait(timeout_usecs))
                    return true;
                // At this point, we've timed out waiting for the semaphore, but the
                // count is still decremented indicating we may still be waiting on
                // it. So we have to re-adjust the count, but only if the semaphore
                // wasn't signaled enough times for us too since then. If it was, we
                // need to release the semaphore too.
                while (true)
                {
                    oldCount = m_count.fetch_add_release(1);
                    if (oldCount < 0)
                        return false;    // successfully restored things to the way they were
                    // Oh, the producer thread just signaled the semaphore after all. Try again:
                    oldCount = m_count.fetch_add_acquire(-1);
                    if (oldCount > 0 && m_sema.try_wait())
                        return true;
                }
            }

        public:
            LightweightSemaphore(ssize_t initialCount = 0) : m_count(initialCount)
            {
                assert(initialCount >= 0);
            }

            bool tryWait()
            {
                if (m_count.load() > 0)
                {
                    m_count.fetch_add_acquire(-1);
                    return true;
                }
                return false;
            }

            void wait()
            {
                if (!tryWait())
                    waitWithPartialSpinning();
            }

            bool wait(std::int64_t timeout_usecs)
            {
                return tryWait() || waitWithPartialSpinning(timeout_usecs);
            }

            void signal(ssize_t count = 1)
            {
                assert(count >= 0);
                ssize_t oldCount = m_count.fetch_add_release(count);
                assert(oldCount >= -1);
                if (oldCount < 0)
                {
                    m_sema.signal(1);
                }
            }
            
            ssize_t availableApprox() const
            {
                ssize_t count = m_count.load();
                return count > 0 ? count : 0;
            }
        };
    }    // end namespace spsc_sema
}    // end namespace moodycamel

#if defined(AE_VCPP) && (_MSC_VER < 1700 || defined(__cplusplus_cli))
#pragma warning(pop)
#ifdef __cplusplus_cli
#pragma managed(pop)
#endi