From c306c12763782700e659760ecbb37ea1bdad25dd Mon Sep 17 00:00:00 2001
From: Jamie Reece Wilson <me@reece.sx>
Date: Sun, 30 Jul 2023 09:34:39 +0100
Subject: [PATCH] [*] Improve WakeOnAddress by hash binning by
 kDefaultWaitPerProcess instead the previous iteration before BST or HashTree
 lookup

---
 Source/Threading/AuWakeOnAddress.cpp | 39 +++++++++++++++++++++++-----
 Source/Threading/AuWakeOnAddress.hpp | 14 ++++++++--
 2 files changed, 44 insertions(+), 9 deletions(-)

diff --git a/Source/Threading/AuWakeOnAddress.cpp b/Source/Threading/AuWakeOnAddress.cpp
index 064308cf..88e8ae0d 100644
--- a/Source/Threading/AuWakeOnAddress.cpp
+++ b/Source/Threading/AuWakeOnAddress.cpp
@@ -183,11 +183,18 @@ namespace Aurora::Threading
 
     bool WaitEntry::TryWakeNoLockNoReallyNoLock(const void *pAddress)
     {
+    #if 0
         if (AuReinterpretCast<const char *>(this->pAddress) > AuReinterpretCast<const char *>(pAddress) ||
             AuReinterpretCast<const char *>(this->pAddress) + this->uSize <= AuReinterpretCast<const char *>(pAddress))
         {
             return false;
         }
+    #else
+        if (this->pAddress != pAddress)
+        {
+            return false;
+        }
+    #endif
 
         this->uAtomic = 0;
         this->variable.Signal();
@@ -196,11 +203,18 @@ namespace Aurora::Threading
 
     bool WaitEntry::TryWakeNoLock(const void *pAddress)
     {
+    #if 0
         if (AuReinterpretCast<const char *>(this->pAddress) > AuReinterpretCast<const char *>(pAddress) ||
           AuReinterpretCast<const char *>(this->pAddress) + this->uSize <= AuReinterpretCast<const char *>(pAddress))
         {
             return false;
         }
+    #else
+        if (this->pAddress != pAddress)
+        {
+            return false;
+        }
+    #endif
 
         AU_LOCK_GUARD(this->mutex);
         this->uAtomic = 0;
@@ -238,7 +252,7 @@ namespace Aurora::Threading
         }
     }
 
-    WaitEntry *ProcessWaitContainer::WaitBufferFrom(void *pAddress, AuUInt8 uSize)
+    WaitEntry *ProcessWaitNodeContainer::WaitBufferFrom(void *pAddress, AuUInt8 uSize)
     {
     #if defined(HACK_NO_INVALID_ACCESS_LEAK_SHARED_REF_ON_DESTROYED_THREAD)
         auto pReturn = tlsWaitEntry.get();
@@ -271,7 +285,7 @@ namespace Aurora::Threading
     }
 
     template <typename T>
-    bool ProcessWaitContainer::IterateWake(T callback)
+    bool ProcessWaitNodeContainer::IterateWake(T callback)
     {
         bool bRetStatus { true };
 
@@ -324,16 +338,27 @@ namespace Aurora::Threading
         return bRetStatus;
     }
 
-    void ProcessWaitContainer::Lock()
+    void ProcessWaitNodeContainer::Lock()
     {
         DoSpinLockOnVar(&this->uAtomic);
     }
 
-    void ProcessWaitContainer::Unlock()
+    void ProcessWaitNodeContainer::Unlock()
     {
         this->uAtomic = 0;
     }
-    
+
+    WaitEntry *ProcessWaitContainer::WaitBufferFrom(void *pAddress, AuUInt8 uSize)
+    {
+        return this->list[AuHashCode(pAddress) % AuArraySize(this->list)].WaitBufferFrom(pAddress, uSize);
+    }
+
+    template <typename T>
+    bool ProcessWaitContainer::IterateWake(void *pAddress, T callback)
+    {
+        return this->list[AuHashCode(pAddress) % AuArraySize(this->list)].IterateWake(callback);
+    }
+
     AUKN_SYM bool IsWaitOnRecommended()
     {
     #if defined(AURORA_IS_MODERNNT_DERIVED)
@@ -810,7 +835,7 @@ namespace Aurora::Threading
         }
         else
         {
-            (void)gProcessWaitables.IterateWake([&](WaitEntry &entry) -> AuPair<bool, bool>
+            (void)gProcessWaitables.IterateWake(pTargetAddress, [&](WaitEntry &entry) -> AuPair<bool, bool>
             {
                 if (!uNMaximumThreads)
                 {
@@ -843,7 +868,7 @@ namespace Aurora::Threading
         }
         else
         {
-            (void)gProcessWaitables.IterateWake([&](WaitEntry &entry) -> AuPair<bool, bool>
+            (void)gProcessWaitables.IterateWake(pTargetAddress, [&](WaitEntry &entry) -> AuPair<bool, bool>
             {
                 return AuMakePair(true, entry.TryWakeNoLockNoReallyNoLock(pTargetAddress));
             });
diff --git a/Source/Threading/AuWakeOnAddress.hpp b/Source/Threading/AuWakeOnAddress.hpp
index 108efd84..ec8f8f8c 100644
--- a/Source/Threading/AuWakeOnAddress.hpp
+++ b/Source/Threading/AuWakeOnAddress.hpp
@@ -12,7 +12,7 @@
 
 namespace Aurora::Threading
 {
-    static const auto kDefaultWaitPerProcess = 12;
+    static const auto kDefaultWaitPerProcess = 64;
 
     struct WaitState;
 
@@ -71,7 +71,7 @@ namespace Aurora::Threading
         WaitEntry *pTail {};
     };
 
-    struct ProcessWaitContainer
+    struct ProcessWaitNodeContainer
     {
         AuUInt32 uAtomic {};
     #if defined(WOA_ENABLE_OLD_SHORT_LIST)
@@ -88,4 +88,14 @@ namespace Aurora::Threading
 
         void Unlock();
     };
+
+    struct ProcessWaitContainer
+    {
+        ProcessWaitNodeContainer list[kDefaultWaitPerProcess];
+
+        WaitEntry *WaitBufferFrom(void *pAddress, AuUInt8 uSize);
+
+        template <typename T>
+        bool IterateWake(void *pAddress, T callback);
+    };
 }
\ No newline at end of file