Merge v3.6.7 in

16 files changed, 456 insertions, 161 deletions

diff --git a/indra/llrender/llgl.cpp b/indra/llrender/llgl.cpp
index 088ba95b75..acfb3c085a 100755
--- a/indra/llrender/llgl.cpp
+++ b/indra/llrender/llgl.cpp
@@ -59,6 +59,7 @@ BOOL gDebugGL = FALSE;
 BOOL gClothRipple = FALSE;
 BOOL gHeadlessClient = FALSE;
 BOOL gGLActive = FALSE;
+BOOL gGLDebugLoggingEnabled = TRUE;
 
 static const std::string HEADLESS_VENDOR_STRING("Linden Lab");
 static const std::string HEADLESS_RENDERER_STRING("Headless");
@@ -80,6 +81,8 @@ void APIENTRY gl_debug_callback(GLenum source,
                                 const GLchar* message,
                                 GLvoid* userParam)
 {
+	if (gGLDebugLoggingEnabled)
+	{
 	if (severity == GL_DEBUG_SEVERITY_HIGH_ARB)
 	{
 		llwarns << "----- GL ERROR --------" << llendl;
@@ -98,6 +101,7 @@ void APIENTRY gl_debug_callback(GLenum source,
 		llerrs << "Halting on GL Error" << llendl;
 	}
 }
+}
 #endif
 
 void parse_glsl_version(S32& major, S32& minor);
@@ -258,6 +262,7 @@ PFNGLBEGINTRANSFORMFEEDBACKPROC glBeginTransformFeedback = NULL;
 PFNGLENDTRANSFORMFEEDBACKPROC glEndTransformFeedback = NULL;
 PFNGLTRANSFORMFEEDBACKVARYINGSPROC glTransformFeedbackVaryings = NULL;
 PFNGLBINDBUFFERRANGEPROC glBindBufferRange = NULL;
+PFNGLBINDBUFFERBASEPROC glBindBufferBase = NULL;
 
 //GL_ARB_debug_output
 PFNGLDEBUGMESSAGECONTROLARBPROC glDebugMessageControlARB = NULL;
@@ -647,7 +652,8 @@ bool LLGLManager::initGL()
 		}
 #if LL_DARWIN
 		else if ((mGLRenderer.find("9400M") != std::string::npos)
-			  || (mGLRenderer.find("9600M") != std::string::npos))
+			  || (mGLRenderer.find("9600M") != std::string::npos)
+			  || (mGLRenderer.find("9800M") != std::string::npos))
 		{
 			mIsMobileGF = TRUE;
 		}
@@ -1154,7 +1160,8 @@ void LLGLManager::initExtensions()
 	// Misc
 	glGetIntegerv(GL_MAX_ELEMENTS_VERTICES, (GLint*) &mGLMaxVertexRange);
 	glGetIntegerv(GL_MAX_ELEMENTS_INDICES, (GLint*) &mGLMaxIndexRange);
-	
+	glGetIntegerv(GL_MAX_TEXTURE_SIZE, (GLint*) &mGLMaxTextureSize);
+
 #if (LL_WINDOWS || LL_LINUX || LL_SOLARIS) && !LL_MESA_HEADLESS
 	LL_DEBUGS("RenderInit") << "GL Probe: Getting symbols" << LL_ENDL;
 	if (mHasVertexBufferObject)
@@ -1245,6 +1252,7 @@ void LLGLManager::initExtensions()
 		glEndTransformFeedback = (PFNGLENDTRANSFORMFEEDBACKPROC) GLH_EXT_GET_PROC_ADDRESS("glEndTransformFeedback");
 		glTransformFeedbackVaryings = (PFNGLTRANSFORMFEEDBACKVARYINGSPROC) GLH_EXT_GET_PROC_ADDRESS("glTransformFeedbackVaryings");
 		glBindBufferRange = (PFNGLBINDBUFFERRANGEPROC) GLH_EXT_GET_PROC_ADDRESS("glBindBufferRange");
+		glBindBufferBase = (PFNGLBINDBUFFERBASEPROC) GLH_EXT_GET_PROC_ADDRESS("glBindBufferBase");
 	}
 	if (mHasDebugOutput)
 	{
@@ -1499,7 +1507,7 @@ void do_assert_glerror()
 
 void assert_glerror()
 {
-	if (!gGLActive)
+/*	if (!gGLActive)
 	{
 		//llwarns << "GL used while not active!" << llendl;
 
@@ -1508,8 +1516,13 @@ void assert_glerror()
 			//ll_fail("GL used while not active");
 		}
 	}
+*/
 
-	if (gDebugGL) 
+	if (!gDebugGL) 
+	{
+		//funny looking if for branch prediction -- gDebugGL is almost always false and assert_glerror is called often
+	}
+	else
 	{
 		do_assert_glerror();
 	}
diff --git a/indra/llrender/llgl.h b/indra/llrender/llgl.h
index 60597fd090..75e5fe86ec 100755
--- a/indra/llrender/llgl.h
+++ b/indra/llrender/llgl.h
@@ -155,6 +155,7 @@ public:
 	S32 mVRAM; // VRAM in MB
 	S32 mGLMaxVertexRange;
 	S32 mGLMaxIndexRange;
+	S32 mGLMaxTextureSize;
 	
 	void getPixelFormat(); // Get the best pixel format
 
diff --git a/indra/llrender/llglheaders.h b/indra/llrender/llglheaders.h
index dace572953..5a80a8faa4 100755
--- a/indra/llrender/llglheaders.h
+++ b/indra/llrender/llglheaders.h
@@ -543,6 +543,7 @@ extern PFNGLBEGINTRANSFORMFEEDBACKPROC glBeginTransformFeedback;
 extern PFNGLENDTRANSFORMFEEDBACKPROC glEndTransformFeedback;
 extern PFNGLTRANSFORMFEEDBACKVARYINGSPROC glTransformFeedbackVaryings;
 extern PFNGLBINDBUFFERRANGEPROC glBindBufferRange;
+extern PFNGLBINDBUFFERBASEPROC glBindBufferBase;
 
 
 #elif LL_WINDOWS
@@ -787,6 +788,7 @@ extern PFNGLBEGINTRANSFORMFEEDBACKPROC glBeginTransformFeedback;
 extern PFNGLENDTRANSFORMFEEDBACKPROC glEndTransformFeedback;
 extern PFNGLTRANSFORMFEEDBACKVARYINGSPROC glTransformFeedbackVaryings;
 extern PFNGLBINDBUFFERRANGEPROC glBindBufferRange;
+extern PFNGLBINDBUFFERBASEPROC glBindBufferBase;
 
 //GL_ARB_debug_output
 extern PFNGLDEBUGMESSAGECONTROLARBPROC glDebugMessageControlARB;
@@ -1009,12 +1011,7 @@ extern void glGetBufferPointervARB (GLenum, GLenum, GLvoid* *);
 }
 #endif
 
-#if __MAC_OS_X_VERSION_MAX_ALLOWED <= 1070
 #include <OpenGL/gl.h>
-#else
-#include <AGL/gl.h>
-#endif
-
 
 #endif // LL_MESA / LL_WINDOWS / LL_DARWIN
 
diff --git a/indra/llrender/llglslshader.cpp b/indra/llrender/llglslshader.cpp
index 62191b4c1a..35620bb656 100755
--- a/indra/llrender/llglslshader.cpp
+++ b/indra/llrender/llglslshader.cpp
@@ -149,6 +149,9 @@ void LLGLSLShader::clearStats()
 	mTimeElapsed = 0;
 	mSamplesDrawn = 0;
 	mDrawCalls = 0;
+	mTextureStateFetched = false;
+	mTextureMagFilter.clear();
+	mTextureMinFilter.clear();
 }
 
 void LLGLSLShader::dumpStats()
@@ -161,6 +164,16 @@ void LLGLSLShader::dumpStats()
 		{
 			llinfos << mShaderFiles[i].first << llendl;
 		}
+		for (U32 i = 0; i < mTexture.size(); ++i)
+		{
+			GLint idx = mTexture[i];
+			
+			if (idx >= 0)
+			{
+				GLint uniform_idx = getUniformLocation(i);
+				llinfos << mUniformNameMap[uniform_idx] << " - " << std::hex << mTextureMagFilter[i] << "/" << mTextureMinFilter[i] << std::dec << llendl;
+			}
+		}
 		llinfos << "=============================================" << llendl;
 
 		F32 ms = mTimeElapsed/1000000.f;
@@ -211,6 +224,39 @@ void LLGLSLShader::placeProfileQuery()
 		glGenQueriesARB(1, &mTimerQuery);
 	}
 
+	if (!mTextureStateFetched)
+	{
+		mTextureStateFetched = true;
+		mTextureMagFilter.resize(mTexture.size());
+		mTextureMinFilter.resize(mTexture.size());
+
+		U32 cur_active = gGL.getCurrentTexUnitIndex();
+
+		for (U32 i = 0; i < mTexture.size(); ++i)
+		{
+			GLint idx = mTexture[i];
+
+			if (idx >= 0)
+			{
+				gGL.getTexUnit(idx)->activate();
+
+				U32 mag = 0xFFFFFFFF;
+				U32 min = 0xFFFFFFFF;
+
+				U32 type = LLTexUnit::getInternalType(gGL.getTexUnit(idx)->getCurrType());
+
+				glGetTexParameteriv(type, GL_TEXTURE_MAG_FILTER, (GLint*) &mag);
+				glGetTexParameteriv(type, GL_TEXTURE_MIN_FILTER, (GLint*) &min);
+
+				mTextureMagFilter[i] = mag;
+				mTextureMinFilter[i] = min;
+			}
+		}
+
+		gGL.getTexUnit(cur_active)->activate();
+	}
+
+
 	glBeginQueryARB(GL_SAMPLES_PASSED, 1);
 	glBeginQueryARB(GL_TIME_ELAPSED, mTimerQuery);
 #endif
@@ -309,8 +355,8 @@ void LLGLSLShader::unload()
 	stop_glerror();
 }
 
-BOOL LLGLSLShader::createShader(vector<string> * attributes,
-								vector<string> * uniforms,
+BOOL LLGLSLShader::createShader(std::vector<LLStaticHashedString> * attributes,
+								std::vector<LLStaticHashedString> * uniforms,
 								U32 varying_count,
 								const char** varyings)
 {
@@ -329,6 +375,11 @@ BOOL LLGLSLShader::createShader(vector<string> * attributes,
 	// Create program
 	mProgramObject = glCreateProgramObjectARB();
 	
+#if LL_DARWIN
+    // work-around missing mix(vec3,vec3,bvec3)
+    mDefines["OLD_SELECT"] = "1";
+#endif
+	
 	//compile new source
 	vector< pair<string,GLenum> >::iterator fileIter = mShaderFiles.begin();
 	for ( ; fileIter != mShaderFiles.end(); fileIter++ )
@@ -392,7 +443,8 @@ BOOL LLGLSLShader::createShader(vector<string> * attributes,
 
 		for (S32 i = 0; i < channel_count; i++)
 		{
-			uniform1i(llformat("tex%d", i), i);
+			LLStaticHashedString uniName(llformat("tex%d", i));
+			uniform1i(uniName, i);
 		}
 
 		S32 cur_tex = channel_count; //adjust any texture channels that might have been overwritten
@@ -449,7 +501,7 @@ void LLGLSLShader::attachObjects(GLhandleARB* objects, S32 count)
 	}
 }
 
-BOOL LLGLSLShader::mapAttributes(const vector<string> * attributes)
+BOOL LLGLSLShader::mapAttributes(const std::vector<LLStaticHashedString> * attributes)
 {
 	//before linking, make sure reserved attributes always have consistent locations
 	for (U32 i = 0; i < LLShaderMgr::instance()->mReservedAttribs.size(); i++)
@@ -486,7 +538,7 @@ BOOL LLGLSLShader::mapAttributes(const vector<string> * attributes)
 		{
 			for (U32 i = 0; i < numAttributes; i++)
 			{
-				const char* name = (*attributes)[i].c_str();
+				const char* name = (*attributes)[i].String().c_str();
 				S32 index = glGetAttribLocationARB(mProgramObject, name);
 				if (index != -1)
 				{
@@ -502,7 +554,7 @@ BOOL LLGLSLShader::mapAttributes(const vector<string> * attributes)
 	return FALSE;
 }
 
-void LLGLSLShader::mapUniform(GLint index, const vector<string> * uniforms)
+void LLGLSLShader::mapUniform(GLint index, const vector<LLStaticHashedString> * uniforms)
 {
 	if (index == -1)
 	{
@@ -572,7 +624,10 @@ void LLGLSLShader::mapUniform(GLint index, const vector<string> * uniforms)
 			is_array[0] = 0;
 		}
 
-		mUniformMap[name] = location;
+		LLStaticHashedString hashedName(name);
+		mUniformNameMap[location] = name;
+		mUniformMap[hashedName] = location;
+
 		LL_DEBUGS("ShaderLoading") << "Uniform " << name << " is at location " << location << LL_ENDL;
 	
 		//find the index of this uniform
@@ -593,7 +648,7 @@ void LLGLSLShader::mapUniform(GLint index, const vector<string> * uniforms)
 			for (U32 i = 0; i < uniforms->size(); i++)
 			{
 				if ( (mUniform[i+LLShaderMgr::instance()->mReservedUniforms.size()] == -1)
-					&& ((*uniforms)[i] == name))
+					&& ((*uniforms)[i].String() == name))
 				{
 					//found it
 					mUniform[i+LLShaderMgr::instance()->mReservedUniforms.size()] = location;
@@ -627,7 +682,7 @@ GLint LLGLSLShader::mapUniformTextureChannel(GLint location, GLenum type)
 	return -1;
 }
 
-BOOL LLGLSLShader::mapUniforms(const vector<string> * uniforms)
+BOOL LLGLSLShader::mapUniforms(const vector<LLStaticHashedString> * uniforms)
 {
 	BOOL res = TRUE;
 	
@@ -635,6 +690,7 @@ BOOL LLGLSLShader::mapUniforms(const vector<string> * uniforms)
 	mActiveTextureChannels = 0;
 	mUniform.clear();
 	mUniformMap.clear();
+	mUniformNameMap.clear();
 	mTexture.clear();
 	mValue.clear();
 	//initialize arrays
@@ -1088,18 +1144,18 @@ void LLGLSLShader::uniformMatrix4fv(U32 index, U32 count, GLboolean transpose, c
 	}
 }
 
-GLint LLGLSLShader::getUniformLocation(const string& uniform)
+GLint LLGLSLShader::getUniformLocation(const LLStaticHashedString& uniform)
 {
 	GLint ret = -1;
 	if (mProgramObject > 0)
 	{
-		std::map<string, GLint>::iterator iter = mUniformMap.find(uniform);
+		LLStaticStringTable<GLint>::iterator iter = mUniformMap.find(uniform);
 		if (iter != mUniformMap.end())
 		{
 			if (gDebugGL)
 			{
 				stop_glerror();
-				if (iter->second != glGetUniformLocationARB(mProgramObject, uniform.c_str()))
+				if (iter->second != glGetUniformLocationARB(mProgramObject, uniform.String().c_str()))
 				{
 					llerrs << "Uniform does not match." << llendl;
 				}
@@ -1136,7 +1192,7 @@ GLint LLGLSLShader::getAttribLocation(U32 attrib)
 	}
 }
 
-void LLGLSLShader::uniform1i(const string& uniform, GLint v)
+void LLGLSLShader::uniform1i(const LLStaticHashedString& uniform, GLint v)
 {
 	GLint location = getUniformLocation(uniform);
 				
@@ -1152,72 +1208,73 @@ void LLGLSLShader::uniform1i(const string& uniform, GLint v)
 	}
 }
 
-void LLGLSLShader::uniform1f(const string& uniform, GLfloat v)
+void LLGLSLShader::uniform2i(const LLStaticHashedString& uniform, GLint i, GLint j)
 {
 	GLint location = getUniformLocation(uniform);
 				
 	if (location >= 0)
 	{
 		std::map<GLint, LLVector4>::iterator iter = mValue.find(location);
-		LLVector4 vec(v,0.f,0.f,0.f);
+		LLVector4 vec(i,j,0.f,0.f);
 		if (iter == mValue.end() || shouldChange(iter->second,vec))
 		{
-			glUniform1fARB(location, v);
+			glUniform2iARB(location, i, j);
 			mValue[location] = vec;
 		}
 	}
 }
 
-void LLGLSLShader::uniform2f(const string& uniform, GLfloat x, GLfloat y)
+
+void LLGLSLShader::uniform1f(const LLStaticHashedString& uniform, GLfloat v)
 {
 	GLint location = getUniformLocation(uniform);
 				
 	if (location >= 0)
 	{
 		std::map<GLint, LLVector4>::iterator iter = mValue.find(location);
-		LLVector4 vec(x,y,0.f,0.f);
+		LLVector4 vec(v,0.f,0.f,0.f);
 		if (iter == mValue.end() || shouldChange(iter->second,vec))
 		{
-			glUniform2fARB(location, x,y);
+			glUniform1fARB(location, v);
 			mValue[location] = vec;
 		}
 	}
-
 }
 
-void LLGLSLShader::uniform3f(const string& uniform, GLfloat x, GLfloat y, GLfloat z)
+void LLGLSLShader::uniform2f(const LLStaticHashedString& uniform, GLfloat x, GLfloat y)
 {
 	GLint location = getUniformLocation(uniform);
 				
 	if (location >= 0)
 	{
 		std::map<GLint, LLVector4>::iterator iter = mValue.find(location);
-		LLVector4 vec(x,y,z,0.f);
+		LLVector4 vec(x,y,0.f,0.f);
 		if (iter == mValue.end() || shouldChange(iter->second,vec))
 		{
-			glUniform3fARB(location, x,y,z);
+			glUniform2fARB(location, x,y);
 			mValue[location] = vec;
 		}
 	}
+
 }
 
-void LLGLSLShader::uniform4f(const string& uniform, GLfloat x, GLfloat y, GLfloat z, GLfloat w)
+void LLGLSLShader::uniform3f(const LLStaticHashedString& uniform, GLfloat x, GLfloat y, GLfloat z)
 {
 	GLint location = getUniformLocation(uniform);
-
+				
 	if (location >= 0)
 	{
 		std::map<GLint, LLVector4>::iterator iter = mValue.find(location);
-		LLVector4 vec(x,y,z,w);
+		LLVector4 vec(x,y,z,0.f);
 		if (iter == mValue.end() || shouldChange(iter->second,vec))
 		{
-			glUniform4fARB(location, x,y,z,w);
+			glUniform3fARB(location, x,y,z);
 			mValue[location] = vec;
 		}
 	}
 }
 
-void LLGLSLShader::uniform1fv(const string& uniform, U32 count, const GLfloat* v)
+void LLGLSLShader::uniform1fv(const LLStaticHashedString& uniform, U32 count, const GLfloat* v)
 {
 	GLint location = getUniformLocation(uniform);
 
@@ -1233,7 +1290,7 @@ void LLGLSLShader::uniform1fv(const string& uniform, U32 count, const GLfloat* v
 	}
 }
 
-void LLGLSLShader::uniform2fv(const string& uniform, U32 count, const GLfloat* v)
+void LLGLSLShader::uniform2fv(const LLStaticHashedString& uniform, U32 count, const GLfloat* v)
 {
 	GLint location = getUniformLocation(uniform);
 				
@@ -1249,7 +1306,7 @@ void LLGLSLShader::uniform2fv(const string& uniform, U32 count, const GLfloat* v
 	}
 }
 
-void LLGLSLShader::uniform3fv(const string& uniform, U32 count, const GLfloat* v)
+void LLGLSLShader::uniform3fv(const LLStaticHashedString& uniform, U32 count, const GLfloat* v)
 {
 	GLint location = getUniformLocation(uniform);
 				
@@ -1265,7 +1322,7 @@ void LLGLSLShader::uniform3fv(const string& uniform, U32 count, const GLfloat* v
 	}
 }
 
-void LLGLSLShader::uniform4fv(const string& uniform, U32 count, const GLfloat* v)
+void LLGLSLShader::uniform4fv(const LLStaticHashedString& uniform, U32 count, const GLfloat* v)
 {
 	GLint location = getUniformLocation(uniform);
 
@@ -1283,27 +1340,7 @@ void LLGLSLShader::uniform4fv(const string& uniform, U32 count, const GLfloat* v
 	}
 }
 
-void LLGLSLShader::uniformMatrix2fv(const string& uniform, U32 count, GLboolean transpose, const GLfloat* v)
-{
-	GLint location = getUniformLocation(uniform);
-				
-	if (location >= 0)
-	{
-		glUniformMatrix2fvARB(location, count, transpose, v);
-	}
-}
-
-void LLGLSLShader::uniformMatrix3fv(const string& uniform, U32 count, GLboolean transpose, const GLfloat* v)
-{
-	GLint location = getUniformLocation(uniform);
-				
-	if (location >= 0)
-	{
-		glUniformMatrix3fvARB(location, count, transpose, v);
-	}
-}
-
-void LLGLSLShader::uniformMatrix4fv(const string& uniform, U32 count, GLboolean transpose, const GLfloat* v)
+void LLGLSLShader::uniformMatrix4fv(const LLStaticHashedString& uniform, U32 count, GLboolean transpose, const GLfloat* v)
 {
 	GLint location = getUniformLocation(uniform);
 				
diff --git a/indra/llrender/llglslshader.h b/indra/llrender/llglslshader.h
index 3c775cde27..7b2f5f04c2 100755
--- a/indra/llrender/llglslshader.h
+++ b/indra/llrender/llglslshader.h
@@ -29,6 +29,7 @@
 
 #include "llgl.h"
 #include "llrender.h"
+#include "llstaticstringtable.h"
 
 class LLShaderFeatures
 {
@@ -90,16 +91,16 @@ public:
 	void placeProfileQuery();
 	void readProfileQuery(U32 count, U32 mode);
 
-	BOOL createShader(std::vector<std::string> * attributes,
-						std::vector<std::string> * uniforms,
+	BOOL createShader(std::vector<LLStaticHashedString> * attributes,
+						std::vector<LLStaticHashedString> * uniforms,
 						U32 varying_count = 0,
 						const char** varyings = NULL);
 	BOOL attachObject(std::string object);
 	void attachObject(GLhandleARB object);
 	void attachObjects(GLhandleARB* objects = NULL, S32 count = 0);
-	BOOL mapAttributes(const std::vector<std::string> * attributes);
-	BOOL mapUniforms(const std::vector<std::string> * uniforms);
-	void mapUniform(GLint index, const std::vector<std::string> * uniforms);
+	BOOL mapAttributes(const std::vector<LLStaticHashedString> * attributes);
+	BOOL mapUniforms(const std::vector<LLStaticHashedString> *);
+	void mapUniform(GLint index, const std::vector<LLStaticHashedString> *);
 	void uniform1i(U32 index, GLint i);
 	void uniform1f(U32 index, GLfloat v);
 	void uniform2f(U32 index, GLfloat x, GLfloat y);
@@ -110,29 +111,27 @@ public:
 	void uniform2fv(U32 index, U32 count, const GLfloat* v);
 	void uniform3fv(U32 index, U32 count, const GLfloat* v);
 	void uniform4fv(U32 index, U32 count, const GLfloat* v);
-	void uniform1i(const std::string& uniform, GLint i);
-	void uniform1f(const std::string& uniform, GLfloat v);
-	void uniform2f(const std::string& uniform, GLfloat x, GLfloat y);
-	void uniform3f(const std::string& uniform, GLfloat x, GLfloat y, GLfloat z);
-	void uniform4f(const std::string& uniform, GLfloat x, GLfloat y, GLfloat z, GLfloat w);
-	void uniform1iv(const std::string& uniform, U32 count, const GLint* i);
-	void uniform1fv(const std::string& uniform, U32 count, const GLfloat* v);
-	void uniform2fv(const std::string& uniform, U32 count, const GLfloat* v);
-	void uniform3fv(const std::string& uniform, U32 count, const GLfloat* v);
-	void uniform4fv(const std::string& uniform, U32 count, const GLfloat* v);
+	void uniform2i(const LLStaticHashedString& uniform, GLint i, GLint j);
 	void uniformMatrix2fv(U32 index, U32 count, GLboolean transpose, const GLfloat *v);
 	void uniformMatrix3fv(U32 index, U32 count, GLboolean transpose, const GLfloat *v);
 	void uniformMatrix4fv(U32 index, U32 count, GLboolean transpose, const GLfloat *v);
-	void uniformMatrix2fv(const std::string& uniform, U32 count, GLboolean transpose, const GLfloat *v);
-	void uniformMatrix3fv(const std::string& uniform, U32 count, GLboolean transpose, const GLfloat *v);
-	void uniformMatrix4fv(const std::string& uniform, U32 count, GLboolean transpose, const GLfloat *v);
+	void uniform1i(const LLStaticHashedString& uniform, GLint i);
+	void uniform1f(const LLStaticHashedString& uniform, GLfloat v);
+	void uniform2f(const LLStaticHashedString& uniform, GLfloat x, GLfloat y);
+	void uniform3f(const LLStaticHashedString& uniform, GLfloat x, GLfloat y, GLfloat z);
+	void uniform1fv(const LLStaticHashedString& uniform, U32 count, const GLfloat* v);
+	void uniform2fv(const LLStaticHashedString& uniform, U32 count, const GLfloat* v);
+	void uniform3fv(const LLStaticHashedString& uniform, U32 count, const GLfloat* v);
+	void uniform4fv(const LLStaticHashedString& uniform, U32 count, const GLfloat* v);
+	void uniformMatrix4fv(const LLStaticHashedString& uniform, U32 count, GLboolean transpose, const GLfloat *v);
 
 	void setMinimumAlpha(F32 minimum);
 
 	void vertexAttrib4f(U32 index, GLfloat x, GLfloat y, GLfloat z, GLfloat w);
 	void vertexAttrib4fv(U32 index, GLfloat* v);
 	
-	GLint getUniformLocation(const std::string& uniform);
+	//GLint getUniformLocation(const std::string& uniform);
+	GLint getUniformLocation(const LLStaticHashedString& uniform);	
 	GLint getUniformLocation(U32 index);
 
 	GLint getAttribLocation(U32 attrib);
@@ -169,7 +168,8 @@ public:
 	std::vector<GLint> mAttribute; //lookup table of attribute enum to attribute channel
 	U32 mAttributeMask;  //mask of which reserved attributes are set (lines up with LLVertexBuffer::getTypeMask())
 	std::vector<GLint> mUniform;   //lookup table of uniform enum to uniform location
-	std::map<std::string, GLint> mUniformMap;  //lookup map of uniform name to uniform location
+	LLStaticStringTable<GLint> mUniformMap; //lookup map of uniform name to uniform location
+	std::map<GLint, std::string> mUniformNameMap; //lookup map of uniform location to uniform name
 	std::map<GLint, LLVector4> mValue; //lookup map of uniform location to last known value
 	std::vector<GLint> mTexture;
 	S32 mTotalUniformSize;
@@ -192,6 +192,11 @@ public:
 	static U64 sTotalSamplesDrawn;
 	U32 mDrawCalls;
 	static U32 sTotalDrawCalls;
+
+	bool mTextureStateFetched;
+	std::vector<U32> mTextureMagFilter;
+	std::vector<U32> mTextureMinFilter;
+	
 };
 
 //UI shader (declared here so llui_libtest will link properly)
diff --git a/indra/llrender/llimagegl.cpp b/indra/llrender/llimagegl.cpp
index 38764eba23..6a37d31415 100644..100755
--- a/indra/llrender/llimagegl.cpp
+++ b/indra/llrender/llimagegl.cpp
@@ -715,9 +715,9 @@ void LLImageGL::setImage(const U8* data_in, BOOL data_hasmips)
 
 					mMipLevels = wpo2(llmax(w, h));
 
-					//use legacy mipmap generation mode
+					//use legacy mipmap generation mode (note: making this condional can cause rendering issues)
 					glTexParameteri(mTarget, GL_GENERATE_MIPMAP, GL_TRUE);
-					
+
 					LLImageGL::setManualImage(mTarget, 0, mFormatInternal,
 								 w, h, 
 								 mFormatPrimary, mFormatType,
@@ -1089,6 +1089,16 @@ void LLImageGL::generateTextures(LLTexUnit::eTextureType type, U32 format, S32 n
 	LLFastTimer t(FTM_GENERATE_TEXTURES);
 	bool empty = true;
 
+	if (LLRender::sGLCoreProfile)
+	{
+		switch (format)
+		{
+			case GL_LUMINANCE8: format = GL_RGB8; break;
+			case GL_LUMINANCE8_ALPHA8:
+			case GL_ALPHA8: format = GL_RGBA8; break;
+		}
+	}
+
 	dead_texturelist_t::iterator iter = sDeadTextureList[type].find(format);
 	
 	if (iter != sDeadTextureList[type].end())
diff --git a/indra/llrender/llpostprocess.cpp b/indra/llrender/llpostprocess.cpp
index c0045c8044..4c36185b08 100755
--- a/indra/llrender/llpostprocess.cpp
+++ b/indra/llrender/llpostprocess.cpp
@@ -31,6 +31,21 @@
 #include "llsdserialize.h"
 #include "llrender.h"
 
+static LLStaticHashedString sRenderTexture("RenderTexture");
+static LLStaticHashedString sBrightness("brightness");
+static LLStaticHashedString sContrast("contrast");
+static LLStaticHashedString sContrastBase("contrastBase");
+static LLStaticHashedString sSaturation("saturation");
+static LLStaticHashedString sLumWeights("lumWeights");
+static LLStaticHashedString sNoiseTexture("NoiseTexture");
+static LLStaticHashedString sBrightMult("brightMult");
+static LLStaticHashedString sNoiseStrength("noiseStrength");
+static LLStaticHashedString sExtractLow("extractLow");
+static LLStaticHashedString sExtractHigh("extractHigh");
+static LLStaticHashedString sBloomStrength("bloomStrength");
+static LLStaticHashedString sTexelSize("texelSize");
+static LLStaticHashedString sBlurDirection("blurDirection");
+static LLStaticHashedString sBlurWidth("blurWidth");
 
 LLPostProcess * gPostProcess = NULL;
 
@@ -258,12 +273,12 @@ void LLPostProcess::applyColorFilterShader(void)
 void LLPostProcess::createColorFilterShader(void)
 {
 	/// Define uniform names
-	colorFilterUniforms["RenderTexture"] = 0;
-	colorFilterUniforms["brightness"] = 0;
-	colorFilterUniforms["contrast"] = 0;
-	colorFilterUniforms["contrastBase"] = 0;
-	colorFilterUniforms["saturation"] = 0;
-	colorFilterUniforms["lumWeights"] = 0;
+	colorFilterUniforms[sRenderTexture] = 0;
+	colorFilterUniforms[sBrightness] = 0;
+	colorFilterUniforms[sContrast] = 0;
+	colorFilterUniforms[sContrastBase] = 0;
+	colorFilterUniforms[sSaturation] = 0;
+	colorFilterUniforms[sLumWeights] = 0;
 }
 
 void LLPostProcess::applyNightVisionShader(void)
@@ -307,11 +322,11 @@ void LLPostProcess::applyNightVisionShader(void)
 void LLPostProcess::createNightVisionShader(void)
 {
 	/// Define uniform names
-	nightVisionUniforms["RenderTexture"] = 0;
-	nightVisionUniforms["NoiseTexture"] = 0;
-	nightVisionUniforms["brightMult"] = 0;	
-	nightVisionUniforms["noiseStrength"] = 0;
-	nightVisionUniforms["lumWeights"] = 0;	
+	nightVisionUniforms[sRenderTexture] = 0;
+	nightVisionUniforms[sNoiseTexture] = 0;
+	nightVisionUniforms[sBrightMult] = 0;	
+	nightVisionUniforms[sNoiseStrength] = 0;
+	nightVisionUniforms[sLumWeights] = 0;	
 
 	createNoiseTexture(mNoiseTexture);
 }
@@ -326,25 +341,25 @@ void LLPostProcess::createBloomShader(void)
 	createTexture(mTempBloomTexture, unsigned(screenW * 0.5), unsigned(screenH * 0.5));
 
 	/// Create Bloom Extract Shader
-	bloomExtractUniforms["RenderTexture"] = 0;
-	bloomExtractUniforms["extractLow"] = 0;
-	bloomExtractUniforms["extractHigh"] = 0;	
-	bloomExtractUniforms["lumWeights"] = 0;	
+	bloomExtractUniforms[sRenderTexture] = 0;
+	bloomExtractUniforms[sExtractLow] = 0;
+	bloomExtractUniforms[sExtractHigh] = 0;	
+	bloomExtractUniforms[sLumWeights] = 0;	
 	
 	/// Create Bloom Blur Shader
-	bloomBlurUniforms["RenderTexture"] = 0;
-	bloomBlurUniforms["bloomStrength"] = 0;	
-	bloomBlurUniforms["texelSize"] = 0;
-	bloomBlurUniforms["blurDirection"] = 0;
-	bloomBlurUniforms["blurWidth"] = 0;
+	bloomBlurUniforms[sRenderTexture] = 0;
+	bloomBlurUniforms[sBloomStrength] = 0;	
+	bloomBlurUniforms[sTexelSize] = 0;
+	bloomBlurUniforms[sBlurDirection] = 0;
+	bloomBlurUniforms[sBlurWidth] = 0;
 }
 
 void LLPostProcess::getShaderUniforms(glslUniforms & uniforms, GLhandleARB & prog)
 {
 	/// Find uniform locations and insert into map	
-	std::map<const char *, GLuint>::iterator i;
+	glslUniforms::iterator i;
 	for (i  = uniforms.begin(); i != uniforms.end(); ++i){
-		i->second = glGetUniformLocationARB(prog, i->first);
+		i->second = glGetUniformLocationARB(prog, i->first.String().c_str());
 	}
 }
 
diff --git a/indra/llrender/llpostprocess.h b/indra/llrender/llpostprocess.h
index e19de44c60..ce17b6693d 100755
--- a/indra/llrender/llpostprocess.h
+++ b/indra/llrender/llpostprocess.h
@@ -31,6 +31,7 @@
 #include <fstream>
 #include "llgl.h"
 #include "llglheaders.h"
+#include "llstaticstringtable.h"
 
 class LLPostProcess 
 {
@@ -44,7 +45,7 @@ public:
 	} QuadType;
 
 	/// GLSL Shader Encapsulation Struct
-	typedef std::map<const char *, GLuint> glslUniforms;
+	typedef LLStaticStringTable<GLuint> glslUniforms;
 
 	struct PostProcessTweaks : public LLSD {
 		inline PostProcessTweaks() : LLSD(LLSD::emptyMap())
diff --git a/indra/llrender/llrender.cpp b/indra/llrender/llrender.cpp
index c60eb8d9d9..0ac30b4d63 100755
--- a/indra/llrender/llrender.cpp
+++ b/indra/llrender/llrender.cpp
@@ -225,26 +225,15 @@ void LLTexUnit::disable(void)
 bool LLTexUnit::bind(LLTexture* texture, bool for_rendering, bool forceBind)
 {
 	stop_glerror();
-	if (mIndex < 0) return false;
-
+	if (mIndex >= 0)
+	{
 	gGL.flush();
 
 	LLImageGL* gl_tex = NULL ;
-	if (texture == NULL || !(gl_tex = texture->getGLTexture()))
+		if (texture != NULL && (gl_tex = texture->getGLTexture()))
 	{
-		llwarns << "NULL LLTexUnit::bind texture" << llendl;
-		return false;
-	}
-
-	if (!gl_tex->getTexName()) //if texture does not exist
+			if (gl_tex->getTexName()) //if texture exists
 	{
-		//if deleted, will re-generate it immediately
-		texture->forceImmediateUpdate() ;
-
-		gl_tex->forceUpdateBindStats() ;
-		return texture->bindDefaultImage(mIndex);
-	}
-
 	//in audit, replace the selected texture by the default one.
 	if ((mCurrTexture != gl_tex->getTexName()) || forceBind)
 	{
@@ -265,6 +254,27 @@ bool LLTexUnit::bind(LLTexture* texture, bool for_rendering, bool forceBind)
 			setTextureFilteringOption(gl_tex->mFilterOption);
 		}
 	}
+			}
+			else
+			{
+				//if deleted, will re-generate it immediately
+				texture->forceImmediateUpdate() ;
+
+				gl_tex->forceUpdateBindStats() ;
+				return texture->bindDefaultImage(mIndex);
+			}
+		}
+		else
+		{
+			llwarns << "NULL LLTexUnit::bind texture" << llendl;
+			return false;
+		}
+	}
+	else
+	{ // mIndex < 0
+		return false;
+	}
+
 	return true;
 }
 
@@ -1058,6 +1068,16 @@ LLRender::~LLRender()
 
 void LLRender::init()
 {
+	if (sGLCoreProfile && !LLVertexBuffer::sUseVAO)
+	{ //bind a dummy vertex array object so we're core profile compliant
+#ifdef GL_ARB_vertex_array_object
+		U32 ret;
+		glGenVertexArrays(1, &ret);
+		glBindVertexArray(ret);
+#endif
+	}
+
+
 	llassert_always(mBuffer.isNull()) ;
 	stop_glerror();
 	mBuffer = new LLVertexBuffer(immediate_mask, 0);
@@ -1849,13 +1869,15 @@ void LLRender::flush()
 			sUIVerts += mCount;
 		}
 		
-		if (gDebugGL)
-		{
+		//store mCount in a local variable to avoid re-entrance (drawArrays may call flush)
+		U32 count = mCount;
+
 			if (mMode == LLRender::QUADS && !sGLCoreProfile)
 			{
 				if (mCount%4 != 0)
 				{
-					llerrs << "Incomplete quad rendered." << llendl;
+				count -= (mCount % 4);
+				llwarns << "Incomplete quad requested." << llendl;
 				}
 			}
 			
@@ -1863,7 +1885,8 @@ void LLRender::flush()
 			{
 				if (mCount%3 != 0)
 				{
-					llerrs << "Incomplete triangle rendered." << llendl;
+				count -= (mCount % 3);
+				llwarns << "Incomplete triangle requested." << llendl;
 				}
 			}
 			
@@ -1871,13 +1894,11 @@ void LLRender::flush()
 			{
 				if (mCount%2 != 0)
 				{
-					llerrs << "Incomplete line rendered." << llendl;
-				}
+				count -= (mCount % 2);
+				llwarns << "Incomplete line requested." << llendl;
 			}
 		}
 
-		//store mCount in a local variable to avoid re-entrance (drawArrays may call flush)
-		U32 count = mCount;
 		mCount = 0;
 
 		if (mBuffer->useVBOs() && !mBuffer->isLocked())
@@ -2280,6 +2301,22 @@ void LLRender::diffuseColor4ubv(const U8* c)
 	}
 }
 
+void LLRender::diffuseColor4ub(U8 r, U8 g, U8 b, U8 a)
+{
+	LLGLSLShader* shader = LLGLSLShader::sCurBoundShaderPtr;
+	llassert(!LLGLSLShader::sNoFixedFunction || shader != NULL);
+
+	if (shader)
+	{
+		shader->uniform4f(LLShaderMgr::DIFFUSE_COLOR, r/255.f, g/255.f, b/255.f, a/255.f);
+	}
+	else
+	{
+		glColor4ub(r,g,b,a);
+	}
+}
+
+
 void LLRender::debugTexUnits(void)
 {
 	LL_INFOS("TextureUnit") << "Active TexUnit: " << mCurrTextureUnitIndex << LL_ENDL;
diff --git a/indra/llrender/llrender.h b/indra/llrender/llrender.h
index 98222939e7..42b02a8159 100755
--- a/indra/llrender/llrender.h
+++ b/indra/llrender/llrender.h
@@ -396,6 +396,7 @@ public:
 	void diffuseColor4f(F32 r, F32 g, F32 b, F32 a);
 	void diffuseColor4fv(const F32* c);
 	void diffuseColor4ubv(const U8* c);
+	void diffuseColor4ub(U8 r, U8 g, U8 b, U8 a);
 
 	void vertexBatchPreTransformed(LLVector3* verts, S32 vert_count);
 	void vertexBatchPreTransformed(LLVector3* verts, LLVector2* uvs, S32 vert_count);
diff --git a/indra/llrender/llrendertarget.cpp b/indra/llrender/llrendertarget.cpp
index 5fb4fc8e52..353e61105a 100755
--- a/indra/llrender/llrendertarget.cpp
+++ b/indra/llrender/llrendertarget.cpp
@@ -53,11 +53,19 @@ void check_framebuffer_status()
 bool LLRenderTarget::sUseFBO = false;
 U32 LLRenderTarget::sCurFBO = 0;
 
+
+extern S32 gGLViewport[4];
+
+U32 LLRenderTarget::sCurResX = 0;
+U32 LLRenderTarget::sCurResY = 0;
+
 LLRenderTarget::LLRenderTarget() :
 	mResX(0),
 	mResY(0),
 	mFBO(0),
 	mPreviousFBO(0),
+	mPreviousResX(0),
+	mPreviousResY(0),
 	mDepth(0),
 	mStencil(0),
 	mUseDepth(false),
@@ -71,7 +79,7 @@ LLRenderTarget::~LLRenderTarget()
 	release();
 }
 
-void LLRenderTarget::resize(U32 resx, U32 resy, U32 color_fmt)
+void LLRenderTarget::resize(U32 resx, U32 resy)
 { 
 	//for accounting, get the number of pixels added/subtracted
 	S32 pix_diff = (resx*resy)-(mResX*mResY);
@@ -79,10 +87,12 @@ void LLRenderTarget::resize(U32 resx, U32 resy, U32 color_fmt)
 	mResX = resx;
 	mResY = resy;
 
+	llassert(mInternalFormat.size() == mTex.size());
+
 	for (U32 i = 0; i < mTex.size(); ++i)
 	{ //resize color attachments
 		gGL.getTexUnit(0)->bindManual(mUsage, mTex[i]);
-		LLImageGL::setManualImage(LLTexUnit::getInternalType(mUsage), 0, color_fmt, mResX, mResY, GL_RGBA, GL_UNSIGNED_BYTE, NULL, false);
+		LLImageGL::setManualImage(LLTexUnit::getInternalType(mUsage), 0, mInternalFormat[i], mResX, mResY, GL_RGBA, GL_UNSIGNED_BYTE, NULL, false);
 		sBytesAllocated += pix_diff*4;
 	}
 
@@ -109,8 +119,8 @@ void LLRenderTarget::resize(U32 resx, U32 resy, U32 color_fmt)
 
 bool LLRenderTarget::allocate(U32 resx, U32 resy, U32 color_fmt, bool depth, bool stencil, LLTexUnit::eTextureType usage, bool use_fbo, S32 samples)
 {
-	resx = llmin(resx, (U32) 4096);
-	resy = llmin(resy, (U32) 4096);
+	resx = llmin(resx, (U32) gGLManager.mGLMaxTextureSize);
+	resy = llmin(resy, (U32) gGLManager.mGLMaxTextureSize);
 
 	stop_glerror();
 	release();
@@ -244,12 +254,15 @@ bool LLRenderTarget::addColorAttachment(U32 color_fmt)
 	mTex.push_back(tex);
 	mInternalFormat.push_back(color_fmt);
 
+#if !LL_DARWIN
 	if (gDebugGL)
 	{ //bind and unbind to validate target
 		bindTarget();
 		flush();
 	}
-
+#endif
+    
+    
 	return true;
 }
 
@@ -390,13 +403,12 @@ void LLRenderTarget::bindTarget()
 {
 	if (mFBO)
 	{
-		mPreviousFBO = sCurFBO;
-
 		stop_glerror();
 		
+		mPreviousFBO = sCurFBO;
 		glBindFramebuffer(GL_FRAMEBUFFER, mFBO);
 		sCurFBO = mFBO;
-
+		
 		stop_glerror();
 		if (gGLManager.mHasDrawBuffers)
 		{ //setup multiple render targets
@@ -418,7 +430,12 @@ void LLRenderTarget::bindTarget()
 		stop_glerror();
 	}
 
+	mPreviousResX = sCurResX;
+	mPreviousResY = sCurResY;
 	glViewport(0, 0, mResX, mResY);
+	sCurResX = mResX;
+	sCurResY = mResY;
+
 	sBoundTarget = this;
 }
 
@@ -458,6 +475,12 @@ U32 LLRenderTarget::getTexture(U32 attachment) const
 	return mTex[attachment];
 }
 
+U32 LLRenderTarget::getNumTextures() const
+{
+	return mTex.size();
+}
+
+
 void LLRenderTarget::bindTexture(U32 index, S32 channel)
 {
 	gGL.getTexUnit(channel)->bindManual(mUsage, getTexture(index));
@@ -489,6 +512,20 @@ void LLRenderTarget::flush(bool fetch_depth)
 		stop_glerror();
 		glBindFramebuffer(GL_FRAMEBUFFER, mPreviousFBO);
 		sCurFBO = mPreviousFBO;
+
+		if (mPreviousFBO)
+		{
+			glViewport(0, 0, mPreviousResX, mPreviousResY);
+			sCurResX = mPreviousResX;
+			sCurResY = mPreviousResY;
+		}
+		else
+		{
+			glViewport(gGLViewport[0],gGLViewport[1],gGLViewport[2],gGLViewport[3]);
+			sCurResX = gGLViewport[2];
+			sCurResY = gGLViewport[3];
+		}
+						
 		stop_glerror();
 	}
 }
@@ -546,8 +583,10 @@ void LLRenderTarget::copyContentsToFramebuffer(LLRenderTarget& source, S32 srcX0
 {
 	if (!source.mFBO)
 	{
-		llerrs << "Cannot copy framebuffer contents for non FBO render targets." << llendl;
+		llwarns << "Cannot copy framebuffer contents for non FBO render targets." << llendl;
+		return;
 	}
+
 	{
 		GLboolean write_depth = mask & GL_DEPTH_BUFFER_BIT ? TRUE : FALSE;
 
diff --git a/indra/llrender/llrendertarget.h b/indra/llrender/llrendertarget.h
index 765a727b5b..6dc84d978d 100755
--- a/indra/llrender/llrendertarget.h
+++ b/indra/llrender/llrendertarget.h
@@ -63,6 +63,9 @@ public:
 	static bool sUseFBO; 
 	static U32 sBytesAllocated;
 	static U32 sCurFBO;
+	static U32 sCurResX;
+	static U32 sCurResY;
+
 
 	LLRenderTarget();
 	~LLRenderTarget();
@@ -76,7 +79,7 @@ public:
 	// CAUTION: if the GL runs out of memory attempting to resize, this render target will be undefined
 	// DO NOT use for screen space buffers or for scratch space for an image that might be uploaded
 	// DO use for render targets that resize often and aren't likely to ruin someone's day if they break
-	void resize(U32 resx, U32 resy, U32 color_fmt);
+	void resize(U32 resx, U32 resy);
 
 	//add color buffer attachment
 	//limit of 4 color attachments per render target
@@ -112,6 +115,7 @@ public:
 	LLTexUnit::eTextureType getUsage(void) const { return mUsage; }
 
 	U32 getTexture(U32 attachment = 0) const;
+	U32 getNumTextures() const;
 
 	U32 getDepth(void) const { return mDepth; }
 	bool hasStencil() const { return mStencil; }
@@ -146,6 +150,9 @@ protected:
 	std::vector<U32> mInternalFormat;
 	U32 mFBO;
 	U32 mPreviousFBO;
+	U32 mPreviousResX;
+	U32 mPreviousResY;
+
 	U32 mDepth;
 	bool mStencil;
 	bool mUseDepth;
diff --git a/indra/llrender/llshadermgr.cpp b/indra/llrender/llshadermgr.cpp
index fea4ee2819..d230574752 100755
--- a/indra/llrender/llshadermgr.cpp
+++ b/indra/llrender/llshadermgr.cpp
@@ -993,7 +993,9 @@ void LLShaderMgr::initAttribsAndUniforms()
 	mReservedUniforms.push_back("texture_matrix1");
 	mReservedUniforms.push_back("texture_matrix2");
 	mReservedUniforms.push_back("texture_matrix3");
-	llassert(mReservedUniforms.size() == LLShaderMgr::TEXTURE_MATRIX3+1);
+	mReservedUniforms.push_back("object_plane_s");
+	mReservedUniforms.push_back("object_plane_t");
+	llassert(mReservedUniforms.size() == LLShaderMgr::OBJECT_PLANE_T+1);
 
 	mReservedUniforms.push_back("viewport");
 
@@ -1135,13 +1137,54 @@ void LLShaderMgr::initAttribsAndUniforms()
 	mReservedUniforms.push_back("lightMap");
 	mReservedUniforms.push_back("bloomMap");
 	mReservedUniforms.push_back("projectionMap");
-	
+	mReservedUniforms.push_back("norm_mat");
+
 	mReservedUniforms.push_back("global_gamma");
 	mReservedUniforms.push_back("texture_gamma");
 	
 	mReservedUniforms.push_back("specular_color");
 	mReservedUniforms.push_back("env_intensity");
 
+	mReservedUniforms.push_back("matrixPalette");
+	
+	mReservedUniforms.push_back("screenTex");
+	mReservedUniforms.push_back("screenDepth");
+	mReservedUniforms.push_back("refTex");
+	mReservedUniforms.push_back("eyeVec");
+	mReservedUniforms.push_back("time");
+	mReservedUniforms.push_back("d1");
+	mReservedUniforms.push_back("d2");
+	mReservedUniforms.push_back("lightDir");
+	mReservedUniforms.push_back("specular");
+	mReservedUniforms.push_back("lightExp");
+	mReservedUniforms.push_back("waterFogColor");
+	mReservedUniforms.push_back("waterFogDensity");
+	mReservedUniforms.push_back("waterFogKS");
+	mReservedUniforms.push_back("refScale");
+	mReservedUniforms.push_back("waterHeight");
+	mReservedUniforms.push_back("waterPlane");
+	mReservedUniforms.push_back("normScale");
+	mReservedUniforms.push_back("fresnelScale");
+	mReservedUniforms.push_back("fresnelOffset");
+	mReservedUniforms.push_back("blurMultiplier");
+	mReservedUniforms.push_back("sunAngle");
+	mReservedUniforms.push_back("scaledAngle");
+	mReservedUniforms.push_back("sunAngle2");
+	
+	mReservedUniforms.push_back("camPosLocal");
+
+	mReservedUniforms.push_back("gWindDir");
+	mReservedUniforms.push_back("gSinWaveParams");
+	mReservedUniforms.push_back("gGravity");
+
+	mReservedUniforms.push_back("detail_0");
+	mReservedUniforms.push_back("detail_1");
+	mReservedUniforms.push_back("detail_2");
+	mReservedUniforms.push_back("detail_3");
+	mReservedUniforms.push_back("alpha_ramp");
+
+	mReservedUniforms.push_back("origin");
+	mReservedUniforms.push_back("display_gamma");
 	llassert(mReservedUniforms.size() == END_RESERVED_UNIFORMS);
 
 	std::set<std::string> dupe_check;
diff --git a/indra/llrender/llshadermgr.h b/indra/llrender/llshadermgr.h
index c049e935b8..51c27fc8b6 100755
--- a/indra/llrender/llshadermgr.h
+++ b/indra/llrender/llshadermgr.h
@@ -47,6 +47,8 @@ public:
 		TEXTURE_MATRIX1,
 		TEXTURE_MATRIX2,
 		TEXTURE_MATRIX3,
+		OBJECT_PLANE_S,
+		OBJECT_PLANE_T,
 		VIEWPORT,
 		LIGHT_POSITION,
 		LIGHT_DIRECTION,
@@ -165,13 +167,54 @@ public:
 		DEFERRED_LIGHT,
 		DEFERRED_BLOOM,
 		DEFERRED_PROJECTION,
-		
+		DEFERRED_NORM_MATRIX,
+
 		GLOBAL_GAMMA,
 		TEXTURE_GAMMA,
 		
 		SPECULAR_COLOR,
 		ENVIRONMENT_INTENSITY,
 		
+		AVATAR_MATRIX,
+
+		WATER_SCREENTEX,
+		WATER_SCREENDEPTH,
+		WATER_REFTEX,
+		WATER_EYEVEC,
+		WATER_TIME,
+		WATER_WAVE_DIR1,
+		WATER_WAVE_DIR2,
+		WATER_LIGHT_DIR,
+		WATER_SPECULAR,
+		WATER_SPECULAR_EXP,
+		WATER_FOGCOLOR,
+		WATER_FOGDENSITY,
+		WATER_FOGKS,
+		WATER_REFSCALE,
+		WATER_WATERHEIGHT,
+		WATER_WATERPLANE,
+		WATER_NORM_SCALE,
+		WATER_FRESNEL_SCALE,
+		WATER_FRESNEL_OFFSET,
+		WATER_BLUR_MULTIPLIER,
+		WATER_SUN_ANGLE,
+		WATER_SCALED_ANGLE,
+		WATER_SUN_ANGLE2,
+		
+		WL_CAMPOSLOCAL,
+
+		AVATAR_WIND,
+		AVATAR_SINWAVE,
+		AVATAR_GRAVITY,
+
+		TERRAIN_DETAIL0,
+		TERRAIN_DETAIL1,
+		TERRAIN_DETAIL2,
+		TERRAIN_DETAIL3,
+		TERRAIN_ALPHARAMP,
+		
+		SHINY_ORIGIN,
+DISPLAY_GAMMA,
 		END_RESERVED_UNIFORMS
 	} eGLSLReservedUniforms;
 
diff --git a/indra/llrender/llvertexbuffer.cpp b/indra/llrender/llvertexbuffer.cpp
index 01541026b1..f168b3af14 100755
--- a/indra/llrender/llvertexbuffer.cpp
+++ b/indra/llrender/llvertexbuffer.cpp
@@ -85,6 +85,7 @@ const U32 LL_VBO_POOL_SEED_COUNT = vbo_block_index(LL_VBO_POOL_MAX_SEED_SIZE);
 //static
 LLVBOPool LLVertexBuffer::sStreamVBOPool(GL_STREAM_DRAW_ARB, GL_ARRAY_BUFFER_ARB);
 LLVBOPool LLVertexBuffer::sDynamicVBOPool(GL_DYNAMIC_DRAW_ARB, GL_ARRAY_BUFFER_ARB);
+LLVBOPool LLVertexBuffer::sDynamicCopyVBOPool(GL_DYNAMIC_COPY_ARB, GL_ARRAY_BUFFER_ARB);
 LLVBOPool LLVertexBuffer::sStreamIBOPool(GL_STREAM_DRAW_ARB, GL_ELEMENT_ARRAY_BUFFER_ARB);
 LLVBOPool LLVertexBuffer::sDynamicIBOPool(GL_DYNAMIC_DRAW_ARB, GL_ELEMENT_ARRAY_BUFFER_ARB);
 
@@ -199,7 +200,10 @@ volatile U8* LLVBOPool::allocate(U32& name, U32 size, bool for_seed)
 		if (LLVertexBuffer::sDisableVBOMapping || mUsage != GL_DYNAMIC_DRAW_ARB)
 		{
 			glBufferDataARB(mType, size, 0, mUsage);
-			ret = (U8*) ll_aligned_malloc_16(size);
+			if (mUsage != GL_DYNAMIC_COPY_ARB)
+			{ //data will be provided by application
+				ret = (U8*) ll_aligned_malloc(size, 64);
+			}
 		}
 		else
 		{ //always use a true hint of static draw when allocating non-client-backed buffers
@@ -252,7 +256,7 @@ void LLVBOPool::release(U32 name, volatile U8* buffer, U32 size)
 	llassert(vbo_block_size(size) == size);
 
 	deleteBuffer(name);
-	ll_aligned_free_16((U8*) buffer);
+	ll_aligned_free((U8*) buffer);
 
 	if (mType == GL_ARRAY_BUFFER_ARB)
 	{
@@ -306,7 +310,7 @@ void LLVBOPool::cleanup()
 			
 			if (r.mClientData)
 			{
-				ll_aligned_free_16((void*) r.mClientData);
+				ll_aligned_free((void*) r.mClientData);
 			}
 
 			l.pop_front();
@@ -412,6 +416,7 @@ void LLVertexBuffer::seedPools()
 {
 	sStreamVBOPool.seedPool();
 	sDynamicVBOPool.seedPool();
+	sDynamicCopyVBOPool.seedPool();
 	sStreamIBOPool.seedPool();
 	sDynamicIBOPool.seedPool();
 }
@@ -910,6 +915,7 @@ void LLVertexBuffer::cleanupClass()
 	sDynamicIBOPool.cleanup();
 	sStreamVBOPool.cleanup();
 	sDynamicVBOPool.cleanup();
+	sDynamicCopyVBOPool.cleanup();
 
 	if(sPrivatePoolp)
 	{
@@ -946,6 +952,8 @@ S32 LLVertexBuffer::determineUsage(S32 usage)
 	
 	if (ret_usage && ret_usage != GL_STREAM_DRAW_ARB)
 	{ //only stream_draw and dynamic_draw are supported when using VBOs, dynamic draw is the default
+		if (ret_usage != GL_DYNAMIC_COPY_ARB)
+		{
 		if (sDisableVBOMapping)
 		{ //always use stream draw if VBO mapping is disabled
 			ret_usage = GL_STREAM_DRAW_ARB;
@@ -955,6 +963,7 @@ S32 LLVertexBuffer::determineUsage(S32 usage)
 			ret_usage = GL_DYNAMIC_DRAW_ARB;
 		}
 	}
+	}
 	
 	return ret_usage;
 }
@@ -1102,10 +1111,15 @@ void LLVertexBuffer::genBuffer(U32 size)
 	{
 		mMappedData = sStreamVBOPool.allocate(mGLBuffer, mSize);
 	}
-	else
+	else if (mUsage == GL_DYNAMIC_DRAW_ARB)
 	{
 		mMappedData = sDynamicVBOPool.allocate(mGLBuffer, mSize);
 	}
+	else
+	{
+		mMappedData = sDynamicCopyVBOPool.allocate(mGLBuffer, mSize);
+	}
+	
 	
 	sGLCount++;
 }
@@ -2144,21 +2158,14 @@ bool LLVertexBuffer::bindGLBuffer(bool force_bind)
 
 	if (useVBOs() && (force_bind || (mGLBuffer && (mGLBuffer != sGLRenderBuffer || !sVBOActive))))
 	{
-		LLFastTimer t(FTM_BIND_GL_BUFFER);
-		/*if (sMapped)
-		{
-			llerrs << "VBO bound while another VBO mapped!" << llendl;
-		}*/
+		//LLFastTimer t(FTM_BIND_GL_BUFFER); <-- this timer is showing up as a hotspot (irony)
+		
 		glBindBufferARB(GL_ARRAY_BUFFER_ARB, mGLBuffer);
 		sGLRenderBuffer = mGLBuffer;
 		sBindCount++;
 		sVBOActive = true;
 
-		if (mGLArray)
-		{
-			llassert(sGLRenderArray == mGLArray);
-			//mCachedRenderBuffer = mGLBuffer;
-		}
+		llassert(!mGLArray || sGLRenderArray == mGLArray);
 
 		ret = true;
 	}
@@ -2239,7 +2246,38 @@ void LLVertexBuffer::setBuffer(U32 data_mask)
 
 			if ((data_mask & required_mask) != required_mask)
 			{
-				llwarns << "Shader consumption mismatches data provision." << llendl;
+				
+				U32 unsatisfied_mask = (required_mask & ~data_mask);
+				U32 i = 0;
+
+				while (i < TYPE_MAX)
+				{
+                    U32 unsatisfied_flag = unsatisfied_mask & (1 << i);
+					switch (unsatisfied_flag)
+					{
+						case MAP_VERTEX: llinfos << "Missing vert pos" << llendl; break;
+						case MAP_NORMAL: llinfos << "Missing normals" << llendl; break;
+						case MAP_TEXCOORD0: llinfos << "Missing TC 0" << llendl; break;
+						case MAP_TEXCOORD1: llinfos << "Missing TC 1" << llendl; break;
+						case MAP_TEXCOORD2: llinfos << "Missing TC 2" << llendl; break;
+						case MAP_TEXCOORD3: llinfos << "Missing TC 3" << llendl; break;
+						case MAP_COLOR: llinfos << "Missing vert color" << llendl; break;
+						case MAP_EMISSIVE: llinfos << "Missing emissive" << llendl; break;
+						case MAP_TANGENT: llinfos << "Missing tangent" << llendl; break;
+						case MAP_WEIGHT: llinfos << "Missing weight" << llendl; break;
+						case MAP_WEIGHT4: llinfos << "Missing weightx4" << llendl; break;
+						case MAP_CLOTHWEIGHT: llinfos << "Missing clothweight" << llendl; break;
+						case MAP_TEXTURE_INDEX: llinfos << "Missing tex index" << llendl; break;
+						default: llinfos << "Missing who effin knows: " << unsatisfied_flag << llendl;
+					}					
+				}
+
+            if (unsatisfied_mask & (1 << TYPE_INDEX))
+            {
+               llinfos << "Missing indices" << llendl;
+            }
+
+				llerrs << "Shader consumption mismatches data provision." << llendl;
 			}
 		}
 	}
@@ -2409,7 +2447,8 @@ void LLVertexBuffer::setupVertexBuffer(U32 data_mask)
 		if (data_mask & MAP_COLOR)
 		{
 			S32 loc = TYPE_COLOR;
-			void* ptr = (void*)(base + mOffsets[TYPE_COLOR]);
+			//bind emissive instead of color pointer if emissive is present
+			void* ptr = (data_mask & MAP_EMISSIVE) ? (void*)(base + mOffsets[TYPE_EMISSIVE]) : (void*)(base + mOffsets[TYPE_COLOR]);
 			glVertexAttribPointerARB(loc, 4, GL_UNSIGNED_BYTE, GL_TRUE, LLVertexBuffer::sTypeSize[TYPE_COLOR], ptr);
 		}
 		if (data_mask & MAP_EMISSIVE)
@@ -2417,6 +2456,12 @@ void LLVertexBuffer::setupVertexBuffer(U32 data_mask)
 			S32 loc = TYPE_EMISSIVE;
 			void* ptr = (void*)(base + mOffsets[TYPE_EMISSIVE]);
 			glVertexAttribPointerARB(loc, 4, GL_UNSIGNED_BYTE, GL_TRUE, LLVertexBuffer::sTypeSize[TYPE_EMISSIVE], ptr);
+
+			if (!(data_mask & MAP_COLOR))
+			{ //map emissive to color channel when color is not also being bound to avoid unnecessary shader swaps
+				loc = TYPE_COLOR;
+				glVertexAttribPointerARB(loc, 4, GL_UNSIGNED_BYTE, GL_TRUE, LLVertexBuffer::sTypeSize[TYPE_EMISSIVE], ptr);
+			}
 		}
 		if (data_mask & MAP_WEIGHT)
 		{
@@ -2502,8 +2547,7 @@ void LLVertexBuffer::setupVertexBuffer(U32 data_mask)
 LLVertexBuffer::MappedRegion::MappedRegion(S32 type, S32 index, S32 count)
 : mType(type), mIndex(index), mCount(count)
 { 
-	llassert(mType == LLVertexBuffer::TYPE_INDEX || 
-			mType < LLVertexBuffer::TYPE_TEXTURE_INDEX);
+	mEnd = mIndex+mCount;	
 }	
 
 
diff --git a/indra/llrender/llvertexbuffer.h b/indra/llrender/llvertexbuffer.h
index 04806c1d8c..de58207c23 100755
--- a/indra/llrender/llvertexbuffer.h
+++ b/indra/llrender/llvertexbuffer.h
@@ -107,6 +107,7 @@ public:
 		S32 mType;
 		S32 mIndex;
 		S32 mCount;
+		S32 mEnd;
 		
 		MappedRegion(S32 type, S32 index, S32 count);
 	};
@@ -125,9 +126,10 @@ public:
 
 	static LLVBOPool sStreamVBOPool;
 	static LLVBOPool sDynamicVBOPool;
+	static LLVBOPool sDynamicCopyVBOPool;
 	static LLVBOPool sStreamIBOPool;
 	static LLVBOPool sDynamicIBOPool;
-
+	
 	static std::list<U32> sAvailableVAOName;
 	static U32 sCurVAOName;