From cf08f101629eecaf0dcb7e00e1266daac96435c2 Mon Sep 17 00:00:00 2001
From: narzoul <narzoul@users.noreply.github.com>
Date: Sat, 21 May 2016 15:21:43 +0200
Subject: [PATCH] Improved GetPixel performance via DDLOCK_READONLY

Reduces Diablo II startup time.
---
 DDrawCompat/CompatGdi.cpp            | 28 ++++++++++++++++++++++------
 DDrawCompat/CompatGdi.h              |  2 +-
 DDrawCompat/CompatGdiDcFunctions.cpp | 18 +++++++++++++++++-
 3 files changed, 40 insertions(+), 8 deletions(-)

diff --git a/DDrawCompat/CompatGdi.cpp b/DDrawCompat/CompatGdi.cpp
index 4361486..a3e78db 100644
--- a/DDrawCompat/CompatGdi.cpp
+++ b/DDrawCompat/CompatGdi.cpp
@@ -17,6 +17,7 @@ namespace
 {
 	std::atomic<int> g_disableEmulationCount = 0;
 	DWORD g_renderingRefCount = 0;
+	DWORD g_ddLockFlags = 0;
 	DWORD g_ddLockThreadRenderingRefCount = 0;
 	DWORD g_ddLockThreadId = 0;
 	HANDLE g_ddUnlockBeginEvent = nullptr;
@@ -53,16 +54,22 @@ namespace
 		return TRUE;
 	}
 
-	bool lockPrimarySurface()
+	bool lockPrimarySurface(DWORD lockFlags)
 	{
 		DDSURFACEDESC2 desc = {};
 		desc.dwSize = sizeof(desc);
 		auto primary(CompatPrimarySurface::getPrimary());
-		if (FAILED(primary->Lock(primary, nullptr, &desc, DDLOCK_WAIT, nullptr)))
+		if (FAILED(primary->Lock(primary, nullptr, &desc, lockFlags | DDLOCK_WAIT, nullptr)))
 		{
 			return false;
 		}
 
+		g_ddLockFlags = lockFlags;
+		if (0 != lockFlags)
+		{
+			EnterCriticalSection(&CompatGdi::g_gdiCriticalSection);
+		}
+
 		g_ddLockThreadId = GetCurrentThreadId();
 		CompatGdiDcCache::setDdLockThreadId(g_ddLockThreadId);
 		CompatGdiDcCache::setSurfaceMemory(desc.lpSurface, desc.lPitch);
@@ -74,8 +81,17 @@ namespace
 		GdiFlush();
 		auto primary(CompatPrimarySurface::getPrimary());
 		primary->Unlock(primary, nullptr);
-		RealPrimarySurface::invalidate(nullptr);
-		RealPrimarySurface::update();
+		if (DDLOCK_READONLY != g_ddLockFlags)
+		{
+			RealPrimarySurface::invalidate(nullptr);
+			RealPrimarySurface::update();
+		}
+
+		if (0 != g_ddLockFlags)
+		{
+			LeaveCriticalSection(&CompatGdi::g_gdiCriticalSection);
+		}
+		g_ddLockFlags = 0;
 
 		Compat::origProcs.ReleaseDDThreadLock();
 	}
@@ -85,7 +101,7 @@ namespace CompatGdi
 {
 	CRITICAL_SECTION g_gdiCriticalSection;
 
-	bool beginGdiRendering()
+	bool beginGdiRendering(DWORD lockFlags)
 	{
 		if (!isEmulationEnabled())
 		{
@@ -99,7 +115,7 @@ namespace CompatGdi
 			LeaveCriticalSection(&g_gdiCriticalSection);
 			Compat::origProcs.AcquireDDThreadLock();
 			EnterCriticalSection(&g_gdiCriticalSection);
-			if (!lockPrimarySurface())
+			if (!lockPrimarySurface(lockFlags))
 			{
 				Compat::origProcs.ReleaseDDThreadLock();
 				return false;
diff --git a/DDrawCompat/CompatGdi.h b/DDrawCompat/CompatGdi.h
index 80f8021..a2668e7 100644
--- a/DDrawCompat/CompatGdi.h
+++ b/DDrawCompat/CompatGdi.h
@@ -6,7 +6,7 @@
 
 namespace CompatGdi
 {
-	bool beginGdiRendering();
+	bool beginGdiRendering(DWORD lockFlags = 0);
 	void endGdiRendering();
 
 	void disableEmulation();
diff --git a/DDrawCompat/CompatGdiDcFunctions.cpp b/DDrawCompat/CompatGdiDcFunctions.cpp
index e363069..90a0b35 100644
--- a/DDrawCompat/CompatGdiDcFunctions.cpp
+++ b/DDrawCompat/CompatGdiDcFunctions.cpp
@@ -10,6 +10,9 @@ namespace
 {
 	std::unordered_map<void*, const char*> g_funcNames;
 
+	template <typename OrigFuncPtr, OrigFuncPtr origFunc, typename... Params>
+	DWORD getDdLockFlags(Params... params);
+
 	BOOL WINAPI GdiDrawStream(HDC, DWORD, DWORD) { return FALSE; }
 	BOOL WINAPI PolyPatBlt(HDC, DWORD, DWORD, DWORD, DWORD) { return FALSE; }
 
@@ -67,7 +70,8 @@ namespace
 		Compat::LogEnter(g_funcNames[origFunc], params...);
 #endif
 
-		if (!hasDisplayDcArg(params...) || !CompatGdi::beginGdiRendering())
+		if (!hasDisplayDcArg(params...) ||
+			!CompatGdi::beginGdiRendering(getDdLockFlags<OrigFuncPtr, origFunc>(params...)))
 		{
 			Result result = Compat::getOrigFuncPtr<OrigFuncPtr, origFunc>()(params...);
 
@@ -107,6 +111,18 @@ namespace
 		return &compatGdiDcFunc<OrigFuncPtr, origFunc, Result, Params...>;
 	}
 
+	template <typename OrigFuncPtr, OrigFuncPtr origFunc, typename... Params>
+	DWORD getDdLockFlags(Params...)
+	{
+		return 0;
+	}
+
+	template <>
+	DWORD getDdLockFlags<decltype(&GetPixel), &GetPixel>(HDC, int, int)
+	{
+		return DDLOCK_READONLY;
+	}
+
 	template <typename OrigFuncPtr, OrigFuncPtr origFunc>
 	void hookGdiDcFunction(const char* moduleName, const char* funcName)
 	{