From 22a4d3383d555c4368ac7a4924d479890b0861cf Mon Sep 17 00:00:00 2001
From: Jamie Reece Wilson <me@reece.sx>
Date: Mon, 29 Apr 2024 17:08:31 +0100
Subject: [PATCH] [*] NT/Win32-only: improved AuThreading::SleepNs(qwTimeout)

---
 Source/Threading/AuSleep.cpp | 112 +++++++++++++++++++++++++++--------
 1 file changed, 88 insertions(+), 24 deletions(-)

diff --git a/Source/Threading/AuSleep.cpp b/Source/Threading/AuSleep.cpp
index 906ef719..d126dca8 100644
--- a/Source/Threading/AuSleep.cpp
+++ b/Source/Threading/AuSleep.cpp
@@ -14,25 +14,18 @@
 #endif
 
 #if defined(AURORA_IS_MODERNNT_DERIVED)
-    #include <Time/Time.hpp>
+    #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_LINUX_DERIVED) 
-
-        SleepNs(AuMSToNS<AuUInt64>(qwTimeout));
-
-    #elif defined(AURORA_IS_MODERNNT_DERIVED)
-        
+    #if defined(AURORA_IS_MODERNNT_DERIVED)
         ::Sleep(qwTimeout);
-    
     #else
-        std::atomic<bool> value = false;
-        BooleanWaitable waitable(value);
-        WaitFor(&waitable, qwTimeout);
+        SleepNs(AuMSToNS<AuUInt64>(qwTimeout));
     #endif 
     }
 
@@ -41,43 +34,114 @@ namespace Aurora::Threading
     #if defined(AURORA_IS_LINUX_DERIVED) || defined(AURORA_IS_BSD_DERIVED)
 
         auto qwEndTime = AuTime::SteadyClockNS() + qwTimeout;
-
-        if ((usleep(qwTimeout / 1000) == -1) && 
-            (errno == EINTR))
+        if (auto uSleepMnSec = qwTimeout / 1000)
         {
-            AuUInt64 qwNow;
-
-            while ((qwNow = AuTime::SteadyClockNS()) < qwEndTime)
+            if ((usleep(uSleepMnSec) == -1) &&
+                (errno == EINTR))
             {
-                usleep((qwEndTime - qwNow) / 1000);
+                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 uWndTimeSteadyNS = AuTime::SteadyClockNS() + qwTimeout;
+            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);
-        #endif
 
             Win32DropSchedulerResolution();
+        #else
+            if (!AuSwInfo::IsWindows8Point1OrGreater())
+            {
+                if (qwTimeout < 500000)
+                {
+                    uEndTimeSteadyNS -= 10'000ull;
+
+                    (void)YieldPollNs(true, uEndTimeSteadyNS, [=]()
+                    {
+                        return false;
+                    });
+
+                    return;
+                }
+
+                if (qwTimeout < AuMSToNS<AuUInt64>(1))
+                {
+                    uEndTimeSteadyNS -= 250'000ull;
+                }
+
+                Win32DropSchedulerResolution();
+            }
+            else
+            {
+                if (qwTimeout < AuMSToNS<AuUInt64>(1))
+                {
+                    uEndTimeSteadyNS -= 100'000ull;
+                }
+            }
+        #endif
 
             AuUInt64 uNowNS {};
-            while ((uNowNS = AuTime::SteadyClockNS()) < uWndTimeSteadyNS)
+            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
-                word.QuadPart = -((uWndTimeSteadyNS - uNowNS) / 100ull);
+                if (AuThreadPrimitives::ThrdCfg::gPlatformIsSMPProcessorOptimized)
+                {
+                    if (uEndTimeSteadyNS2 - uNowNS <= 100000ull)
+                    {
+                        for (AU_ITERATE_N(i, 32))
+                        {
+                            AuThreadPrimitives::SMPPause();
+                        }
+
+                        continue;
+                    }
+                }
+
+                AuSInt sDelta1 = uEndTimeSteadyNS - uNowNS;
+                word.QuadPart = 0;
+                if (sDelta1 > 0)
+                {
+                    word.QuadPart = -(sDelta1 / 100ull);
+                }
+
                 if (!word.QuadPart)
                 {
-                    break;
+                    if (AuThreadPrimitives::ThrdCfg::gPlatformIsSMPProcessorOptimized)
+                    {
+                        for (AU_ITERATE_N(i, 32))
+                        {
+                            AuThreadPrimitives::SMPPause();
+                        }
+
+                        continue;
+                    }
+                    else
+                    {
+                        break;
+                    }
                 }
             #endif
                 pNtDelayExecution(FALSE, &word);