AuroraRuntime/Source/Threading/AuSleep.cpp

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

    File: AuSleep.cpp
    Date: 2021-6-12
    Author: Reece
***/
#include <Source/RuntimeInternal.hpp>
#include "AuSleep.hpp"
#include "AuWaitFor.hpp"

#if defined(AURORA_IS_POSIX_DERIVED)
    #include <unistd.h>
#endif

#if defined(AURORA_IS_MODERNNT_DERIVED)
    #include <Source/Time/Time.hpp>
    #include <Source/Threading/Primitives/SMTYield.hpp>
#endif

namespace Aurora::Threading
{
    AUKN_SYM void Sleep(AuUInt64 qwTimeout)
    {
    #if defined(AURORA_IS_MODERNNT_DERIVED)
        ::Sleep(qwTimeout);
    #else
        SleepNs(AuMSToNS<AuUInt64>(qwTimeout));
    #endif 
    }

    AUKN_SYM void SleepNs(AuUInt64 qwTimeout)
    {
    #if defined(AURORA_IS_LINUX_DERIVED) || defined(AURORA_IS_BSD_DERIVED)

        auto qwEndTime = AuTime::SteadyClockNS() + qwTimeout;
        if (auto uSleepMnSec = qwTimeout / 1000)
        {
            if ((usleep(uSleepMnSec) == -1) &&
                (errno == EINTR))
            {
                AuUInt64 qwNow;

                while ((qwNow = AuTime::SteadyClockNS()) < qwEndTime)
                {
                    if (auto uSleepMnSec = (qwEndTime - qwNow) / 1000)
                    {
                        usleep(uSleepMnSec);
                    }
                    else
                    {
                        break;
                    }
                }
            }
        }

    #else

    #if defined(AURORA_IS_MODERNNT_DERIVED)

        if (pNtDelayExecution)
        {
            auto uEndTimeSteadyNS = AuTime::SteadyClockNS() + qwTimeout;
            auto uEndTimeSteadyNS2 = uEndTimeSteadyNS;

        #if defined(_AURORA_WANT_STRICT_NT_WAKEUP_TIME_NO_WALL_CLOCK_SHIFT_ALLOWED)
            auto uEndTimeWall   = AuTime::CurrentClockNS() + qwTimeout;
            auto uTargetTimeNt  = AuTime::ConvertTimestampNs(uEndTimeWall);

            Win32DropSchedulerResolution();
        #else
            if (!AuSwInfo::IsWindows8Point1OrGreater())
            {
                if (qwTimeout < 500000)
                {
                    uEndTimeSteadyNS -= 10'000ull;

                    (void)YieldPollNs(true, uEndTimeSteadyNS, [=]()
                    {
                        return false;
                    });

                    return;
                }

                if (qwTimeout > AuMSToNS<AuUInt64>(60))
                {
                    uEndTimeSteadyNS -= 20'000'000ull;
                }
                else if (qwTimeout > AuMSToNS<AuUInt64>(40))
                {
                    uEndTimeSteadyNS -= 10'000'000ull;
                }
                else if (qwTimeout > AuMSToNS<AuUInt64>(10))
                {
                    uEndTimeSteadyNS -= 4'000'000ull;
                }
                else if (qwTimeout < AuMSToNS<AuUInt64>(1'300))
                {
                    uEndTimeSteadyNS -= 500'000ull;
                }

                Win32DropSchedulerResolution();
            }
            else
            {
                if (qwTimeout < AuMSToNS<AuUInt64>(1))
                {
                    uEndTimeSteadyNS -= 250'000ull;
                }
            }
        #endif

            AuUInt64 uNowNS {};
            while ((uNowNS = AuTime::SteadyClockNS()) < uEndTimeSteadyNS2)
            {
                LARGE_INTEGER word;
            #if defined(_AURORA_WANT_STRICT_NT_WAKEUP_TIME_NO_WALL_CLOCK_SHIFT_ALLOWED)
                word.QuadPart = uTargetTimeNt;
            #else
                if (AuThreadPrimitives::ThrdCfg::gPlatformIsSMPProcessorOptimized)
                {
                    if (uEndTimeSteadyNS2 - uNowNS <= 230000ull)
                    {
                        auto uNow = AuAtomicAdd(&AuThreadPrimitives::gSpinAdaptiveCurrentCount, 1u);
                        if (!AuThreadPrimitives::gSpinAdaptiveThreshold || uNow <= AuThreadPrimitives::gSpinAdaptiveThreshold)
                        {
                        #if (defined(AURORA_ARCH_X64) || defined(AURORA_ARCH_X86))
                            if (AuThreadPrimitives::ThrdCfg::gIsIntelAlderLakeOrGreater)
                            {
                                _tpause(0, __rdtsc() + 10000);
                            }
                            else
                        #endif
                            {
                            #if 0
                                for (AU_ITERATE_N(i, 32))
                                {
                                    AuThreadPrimitives::SMPPause();
                                }
                            #else
                                // shit compiler wont unwrap
                                AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause();
                                AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause();
                                AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause();
                                AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause();
                                // 32 or 16?
                                AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause();
                                AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause();
                                AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause();
                                AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause();
                            #endif
                            }

                            AuAtomicSub(&AuThreadPrimitives::gSpinAdaptiveCurrentCount, 1u);
                            continue;
                        }
                        else
                        {
                            AuAtomicSub(&AuThreadPrimitives::gSpinAdaptiveCurrentCount, 1u);
                        }
                    }
                }

                AuSInt sDelta1 = uEndTimeSteadyNS - uNowNS;
                word.QuadPart = 0;
                if (sDelta1 > 0)
                {
                    word.QuadPart = -(sDelta1 / 100ull);
                }

                if (!word.QuadPart)
                {
                    if (AuThreadPrimitives::ThrdCfg::gPlatformIsSMPProcessorOptimized)
                    {
                        auto uNow = AuAtomicAdd(&AuThreadPrimitives::gSpinAdaptiveCurrentCount, 1u);
                        if (!AuThreadPrimitives::gSpinAdaptiveThreshold || uNow <= AuThreadPrimitives::gSpinAdaptiveThreshold)
                        {
                        #if (defined(AURORA_ARCH_X64) || defined(AURORA_ARCH_X86))
                            if (AuThreadPrimitives::ThrdCfg::gIsIntelAlderLakeOrGreater)
                            {
                                _tpause(0, __rdtsc() + 1000);
                            }
                            else
                        #endif
                            {
                            #if 0
                                for (AU_ITERATE_N(i, 32))
                                {
                                    AuThreadPrimitives::SMPPause();
                                }
                            #else
                                // shit compiler wont unwrap
                                AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause();
                                AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause();
                                AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause();
                                AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause();
                                // 32 or 16?
                                AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause();
                                AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause();
                                AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause();
                                AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause(); AuThreadPrimitives::SMPPause();
                            #endif
                            }

                            AuAtomicSub(&AuThreadPrimitives::gSpinAdaptiveCurrentCount, 1u);
                            continue;
                        }
                        else
                        {
                            AuAtomicSub(&AuThreadPrimitives::gSpinAdaptiveCurrentCount, 1u);
                            break;
                        }
                    }
                    else
                    {
                        if (!uEndTimeSteadyNS)
                        {
                            ContextYield();
                        }
                        else
                        {
                            break;
                        }
                    }
                }
            #endif
                pNtDelayExecution(FALSE, &word);
            }

            return;
        }

    #endif

        auto status = YieldPollNs(true, qwTimeout + Time::SteadyClockNS(), [=]()
        {
            return false;
        });
    #endif
    }

    void InitSleep()
    {
        // ...
    }
}