AuroraRuntime/Source/Threading/Primitives/AuSemaphore.NT.cpp

/***
    Copyright (C) 2021 J Reece Wilson (a/k/a "Reece"). All rights reserved.

    File: AuSemaphore.NT.cpp
    Date: 2021-6-12
    Author: Reece
***/
#include <Source/RuntimeInternal.hpp>
#include "AuSemaphore.Generic.hpp"
#include "AuSemaphore.NT.hpp"
#include "SMTYield.hpp"
#include "../AuWakeInternal.hpp"

#if !defined(_AURUNTIME_GENERIC_SEMAPHORE)

namespace Aurora::Threading::Primitives
{
    SemaphoreImpl::SemaphoreImpl(AuUInt16 uIntialValue)
    {
        this->dwState_ = uIntialValue;
    }
    
    SemaphoreImpl::~SemaphoreImpl()
    {
    }

    bool SemaphoreImpl::HasOSHandle(AuMach &mach)
    {
        return false;
    }
    
    bool SemaphoreImpl::HasLockImplementation()
    {
        return true;
    }

    bool SemaphoreImpl::TryLockNoSpin()
    {
        auto old = this->dwState_;
        return (old != 0 && AuAtomicCompareExchange(&this->dwState_, old - 1, old) == old);
    }

    bool SemaphoreImpl::TryLockHeavy()
    {
        return DoTryIf([=]()
        {
            return TryLockNoSpin();
        });
    }

    bool SemaphoreImpl::TryLock()
    {
        if (gRuntimeConfig.threadingConfig.bPreferNtSemaphoreSpinTryLock)
        {
            return TryLockHeavy();
        }
        else
        {
            return TryLockNoSpin();
        }
    }
    
    AuUInt32 *SemaphoreImpl::GetSleepCounter()
    {
        return (AuUInt32 *)&this->var;
    }

    bool SemaphoreImpl::LockMS(AuUInt64 uTimeout)
    {
        return LockNS(AuMSToNS<AuUInt64>(uTimeout));
    }

    bool SemaphoreImpl::LockNS(AuUInt64 uTimeout)
    {
        if (this->TryLockNoSpin())
        {
            return true;
        }

        AuUInt64 uStart = AuTime::SteadyClockNS();
        AuUInt64 uEnd   = uTimeout ? uStart + uTimeout : 0;

        if (this->TryLockHeavy())
        {
            return true;
        }

        if (gUseNativeWaitSemapahore)
        {
            AuUInt32 uYieldCounter {};
            auto old = this->dwState_;
            while (!((old != 0) &&
                     (AuAtomicCompareExchange(&this->dwState_, old - 1, old) == old)))
            {
                static const AuUInt32 kExpect { 0 };

                auto pCounter = GetSleepCounter();
                AuAtomicAdd(pCounter, 1u);
                bool bStatus = InternalLTSWaitOnAddressHighRes(&this->dwState_, (void *)&kExpect, sizeof(kExpect), uEnd);
                AuAtomicSub(pCounter, 1u);

                if (!bStatus)
                {
                    return false;
                }

                old = this->dwState_;
            }

            return true;
        }
        else
        {
            this->mutex.Lock();

            while (!TryLockNoSpin())
            {
                if (uTimeout != 0)
                {
                    uStart = Time::SteadyClockNS();
                    if (uStart >= uEnd)
                    {
                        this->mutex.Unlock();
                        return false;
                    }

                    var.WaitForSignalNsEx(&this->mutex, uEnd - uStart);
                }
                else
                {
                    var.WaitForSignalNsEx(&this->mutex, 0);
                }
            }

            this->mutex.Unlock();

            return true;
        }
    }

    bool SemaphoreImpl::LockAbsNS(AuUInt64 qwTimeoutAbs)
    {
        if (this->TryLockHeavy())
        {
            return true;
        }

        if (gUseNativeWaitSemapahore)
        {
            AuUInt32 uYieldCounter {};
            auto old = this->dwState_;
            while (!((old != 0) &&
                     (AuAtomicCompareExchange(&this->dwState_, old - 1, old) == old)))
            {
                static const AuUInt32 kExpect { 0 };

                auto pCounter = GetSleepCounter();
                AuAtomicAdd(pCounter, 1u);
                bool bStatus = InternalLTSWaitOnAddressHighRes(&this->dwState_, (void *)&kExpect, sizeof(kExpect), qwTimeoutAbs);
                AuAtomicSub(pCounter, 1u);

                if (!bStatus)
                {
                    return false;
                }

                old = this->dwState_;
            }

            return true;
        }
        else
        {
            this->mutex.Lock();

            while (!TryLockNoSpin())
            {
                if (qwTimeoutAbs != 0)
                {
                    auto uStart = Time::SteadyClockNS();
                    if (uStart >= qwTimeoutAbs)
                    {
                        this->mutex.Unlock();
                        return false;
                    }

                    var.WaitForSignalNsEx(&this->mutex, qwTimeoutAbs - uStart);
                }
                else
                {
                    var.WaitForSignalNsEx(&this->mutex, 0);
                }
            }

            this->mutex.Unlock();

            return true;
        }
    }
    
    void SemaphoreImpl::Lock()
    {
        auto status = LockNS(0);
        SysAssert(status, "Couldn't lock semaphore");
    }
    
    void SemaphoreImpl::Unlock(AuUInt16 count)
    {
        if (gUseNativeWaitSemapahore)
        {
            AuAtomicAdd<AuUInt32>(&this->dwState_, count);

            if (AuAtomicLoad(GetSleepCounter()))
            {
                if (count == 1)
                {
                    pWakeByAddressSingle(&this->dwState_);
                }
                else
                {
                    pWakeByAddressAll(&this->dwState_);
                }
            }
        }
        else
        {
            // realistically, we cant use the sleep counter optimization trick under windows 7
            // we would have to expand our already oversized by 8, 24-byte x86_64 semaphore for a trivial perf boost
            // we cant efficiently access the conditions state or atomic guarantees...

            this->mutex.Lock(); // do not [re]move this lock fence
            AuAtomicAdd<AuUInt32>(&this->dwState_, count); // this could be moved anywhere above the unlock, including above the lock.
            this->mutex.Unlock();

            if (count == 1)
            {
                this->var.Signal();
            }
            else
            {
                if (count >= 3) // ...this is the only optimization we can hope to achieve
                {
                    // we can always save a few cycles by doing an atomic broadcast on a contended semaphore
                    // waking up the wrong amount of threads probably doesn't matter at this point, on these target platforms
                    this->var.Broadcast();
                }
                else // ...otherwise, do the handshake just a few times
                {
                    // doing the condvar handshake for the exact amount of threads you need, once the cond is contended, can pay off
                    for (AU_ITERATE_N(i, count))
                    {
                        (void)i;
                        this->var.Signal();
                    }
                }
            }
        }
    }
    
    void SemaphoreImpl::Unlock()
    {
        return Unlock(1);
    }
    
    AUKN_SYM ISemaphore *SemaphoreNew(AuUInt16 uIntialValue)
    {
        return _new SemaphoreImpl(uIntialValue);
    }

    AUKN_SYM void SemaphoreRelease(ISemaphore *pSemaphore)
    {
        AuSafeDelete<SemaphoreImpl *>(pSemaphore);
    }

    AUROXTL_INTERFACE_SOO_SRC_EX(AURORA_SYMBOL_EXPORT, Semaphore, SemaphoreImpl)
}
#endif