11 files changed, 171 insertions, 193 deletions

diff --git a/indra/llrender/llfontfreetype.cpp b/indra/llrender/llfontfreetype.cpp
index 5f04664267..d37b16ce0c 100644
--- a/indra/llrender/llfontfreetype.cpp
+++ b/indra/llrender/llfontfreetype.cpp
@@ -56,9 +56,9 @@
 
 FT_Render_Mode gFontRenderMode = FT_RENDER_MODE_NORMAL;
 
-LLFontManager *gFontManagerp = NULL;
+LLFontManager *gFontManagerp = nullptr;
 
-FT_Library gFTLibrary = NULL;
+FT_Library gFTLibrary = nullptr;
 
 //static
 void LLFontManager::initClass()
@@ -73,7 +73,7 @@ void LLFontManager::initClass()
 void LLFontManager::cleanupClass()
 {
     delete gFontManagerp;
-    gFontManagerp = NULL;
+    gFontManagerp = nullptr;
 }
 
 LLFontManager::LLFontManager()
@@ -101,6 +101,7 @@ LLFontManager::LLFontManager()
 LLFontManager::~LLFontManager()
 {
     FT_Done_FreeType(gFTLibrary);
+    unloadAllFonts();
 }
 
 
@@ -139,12 +140,8 @@ LLFontFreetype::LLFontFreetype()
     mAscender(0.f),
     mDescender(0.f),
     mLineHeight(0.f),
-#ifdef LL_WINDOWS
-    pFileStream(NULL),
-    pFtStream(NULL),
-#endif
     mIsFallback(false),
-    mFTFace(NULL),
+    mFTFace(nullptr),
     mRenderGlyphCount(0),
     mStyle(0),
     mPointSize(0)
@@ -157,35 +154,16 @@ LLFontFreetype::~LLFontFreetype()
     // Clean up freetype libs.
     if (mFTFace)
         FT_Done_Face(mFTFace);
-    mFTFace = NULL;
+    mFTFace = nullptr;
 
     // Delete glyph info
     std::for_each(mCharGlyphInfoMap.begin(), mCharGlyphInfoMap.end(), DeletePairedPointer());
     mCharGlyphInfoMap.clear();
 
-#ifdef LL_WINDOWS
-    delete pFileStream; // closed by FT_Done_Face
-    delete pFtStream;
-#endif
     delete mFontBitmapCachep;
     // mFallbackFonts cleaned up by LLPointer destructor
 }
 
-#ifdef LL_WINDOWS
-unsigned long ft_read_cb(FT_Stream stream, unsigned long offset, unsigned char *buffer, unsigned long count) {
-    if (count <= 0) return count;
-    llifstream *file_stream = static_cast<llifstream *>(stream->descriptor.pointer);
-    file_stream->seekg(offset, std::ios::beg);
-    file_stream->read((char*)buffer, count);
-    return (unsigned long)file_stream->gcount();
-}
-
-void ft_close_cb(FT_Stream stream) {
-    llifstream *file_stream = static_cast<llifstream *>(stream->descriptor.pointer);
-    file_stream->close();
-}
-#endif
-
 bool LLFontFreetype::loadFace(const std::string& filename, F32 point_size, F32 vert_dpi, F32 horz_dpi, bool is_fallback, S32 face_n)
 {
     // Don't leak face objects.  This is also needed to deal with
@@ -193,26 +171,21 @@ bool LLFontFreetype::loadFace(const std::string& filename, F32 point_size, F32 v
     if (mFTFace)
     {
         FT_Done_Face(mFTFace);
-        mFTFace = NULL;
+        mFTFace = nullptr;
     }
 
-    int error;
-#ifdef LL_WINDOWS
-    error = ftOpenFace(filename, face_n);
-#else
-    error = FT_New_Face( gFTLibrary,
-                         filename.c_str(),
-                         0,
-                         &mFTFace);
-#endif
+    FT_Open_Args openArgs;
+    memset( &openArgs, 0, sizeof( openArgs ) );
+    openArgs.memory_base = gFontManagerp->loadFont( filename, openArgs.memory_size );
+
+    if( !openArgs.memory_base )
+        return false;
+
+    openArgs.flags = FT_OPEN_MEMORY;
+    int error = FT_Open_Face( gFTLibrary, &openArgs, 0, &mFTFace );
 
     if (error)
-    {
-#ifdef LL_WINDOWS
-        clearFontStreams();
-#endif
         return false;
-    }
 
     mIsFallback = is_fallback;
     F32 pixels_per_em = (point_size / 72.f)*vert_dpi; // Size in inches * dpi
@@ -227,10 +200,8 @@ bool LLFontFreetype::loadFace(const std::string& filename, F32 point_size, F32 v
     {
         // Clean up freetype libs.
         FT_Done_Face(mFTFace);
-#ifdef LL_WINDOWS
-        clearFontStreams();
-#endif
-        mFTFace = NULL;
+
+        mFTFace = nullptr;
         return false;
     }
 
@@ -286,73 +257,30 @@ S32 LLFontFreetype::getNumFaces(const std::string& filename)
     if (mFTFace)
     {
         FT_Done_Face(mFTFace);
-        mFTFace = NULL;
+        mFTFace = nullptr;
     }
 
     S32 num_faces = 1;
 
-#ifdef LL_WINDOWS
-    int error = ftOpenFace(filename, 0);
+    FT_Open_Args openArgs;
+    memset( &openArgs, 0, sizeof( openArgs ) );
+    openArgs.memory_base = gFontManagerp->loadFont( filename, openArgs.memory_size );
+    if( !openArgs.memory_base )
+        return 0;
+    openArgs.flags = FT_OPEN_MEMORY;
+    int error = FT_Open_Face( gFTLibrary, &openArgs, 0, &mFTFace );
 
     if (error)
-    {
         return 0;
-    }
     else
-    {
         num_faces = mFTFace->num_faces;
-    }
 
     FT_Done_Face(mFTFace);
-    clearFontStreams();
-    mFTFace = NULL;
-#endif
+    mFTFace = nullptr;
 
     return num_faces;
 }
 
-#ifdef LL_WINDOWS
-S32 LLFontFreetype::ftOpenFace(const std::string& filename, S32 face_n)
-{
-    S32 error = -1;
-    pFileStream = new llifstream(filename, std::ios::binary);
-    if (pFileStream->is_open())
-    {
-        std::streampos beg = pFileStream->tellg();
-        pFileStream->seekg(0, std::ios::end);
-        std::streampos end = pFileStream->tellg();
-        std::size_t file_size = end - beg;
-        pFileStream->seekg(0, std::ios::beg);
-
-        pFtStream = new LLFT_Stream();
-        pFtStream->base = 0;
-        pFtStream->pos = 0;
-        pFtStream->size = static_cast<unsigned long>(file_size);
-        pFtStream->descriptor.pointer = pFileStream;
-        pFtStream->read = ft_read_cb;
-        pFtStream->close = ft_close_cb;
-
-        FT_Open_Args args;
-        args.flags = FT_OPEN_STREAM;
-        args.stream = (FT_StreamRec*)pFtStream;
-        error = FT_Open_Face(gFTLibrary, &args, face_n, &mFTFace);
-    }
-    return error;
-}
-
-void LLFontFreetype::clearFontStreams()
-{
-    if (pFileStream)
-    {
-        pFileStream->close();
-    }
-    delete pFileStream;
-    delete pFtStream;
-    pFileStream = NULL;
-    pFtStream = NULL;
-}
-#endif
-
 void LLFontFreetype::addFallbackFont(const LLPointer<LLFontFreetype>& fallback_font,
                                      const char_functor_t& functor)
 {
@@ -376,7 +304,7 @@ F32 LLFontFreetype::getDescenderHeight() const
 
 F32 LLFontFreetype::getXAdvance(llwchar wch) const
 {
-    if (mFTFace == NULL)
+    if (mFTFace == nullptr)
         return 0.0;
 
     // Return existing info only if it is current
@@ -400,7 +328,7 @@ F32 LLFontFreetype::getXAdvance(llwchar wch) const
 
 F32 LLFontFreetype::getXAdvance(const LLFontGlyphInfo* glyph) const
 {
-    if (mFTFace == NULL)
+    if (mFTFace == nullptr)
         return 0.0;
 
     return glyph->mXAdvance;
@@ -408,7 +336,7 @@ F32 LLFontFreetype::getXAdvance(const LLFontGlyphInfo* glyph) const
 
 F32 LLFontFreetype::getXKerning(llwchar char_left, llwchar char_right) const
 {
-    if (mFTFace == NULL)
+    if (mFTFace == nullptr)
         return 0.0;
 
     //llassert(!mIsFallback);
@@ -427,7 +355,7 @@ F32 LLFontFreetype::getXKerning(llwchar char_left, llwchar char_right) const
 
 F32 LLFontFreetype::getXKerning(const LLFontGlyphInfo* left_glyph_info, const LLFontGlyphInfo* right_glyph_info) const
 {
-    if (mFTFace == NULL)
+    if (mFTFace == nullptr)
         return 0.0;
 
     U32 left_glyph = left_glyph_info ? left_glyph_info->mGlyphIndex : 0;
@@ -450,7 +378,7 @@ LLFontGlyphInfo* LLFontFreetype::addGlyph(llwchar wch, EFontGlyphType glyph_type
 {
     if (!mFTFace)
     {
-        return NULL;
+        return nullptr;
     }
 
     llassert(!mIsFallback);
@@ -541,14 +469,14 @@ LLFontGlyphInfo* LLFontFreetype::addGlyph(llwchar wch, EFontGlyphType glyph_type
     {
         return addGlyphFromFont(this, wch, glyph_index, glyph_type);
     }
-    return NULL;
+    return nullptr;
 }
 
 LLFontGlyphInfo* LLFontFreetype::addGlyphFromFont(const LLFontFreetype *fontp, llwchar wch, U32 glyph_index, EFontGlyphType requested_glyph_type) const
 {
     LL_PROFILE_ZONE_SCOPED;
-    if (mFTFace == NULL)
-        return NULL;
+    if (mFTFace == nullptr)
+        return nullptr;
 
     llassert(!mIsFallback);
     fontp->renderGlyph(requested_glyph_type, glyph_index, wch);
@@ -600,7 +528,7 @@ LLFontGlyphInfo* LLFontFreetype::addGlyphFromFont(const LLFontFreetype *fontp, l
     {
         U8 *buffer_data = fontp->mFTFace->glyph->bitmap.buffer;
         S32 buffer_row_stride = fontp->mFTFace->glyph->bitmap.pitch;
-        U8 *tmp_graydata = NULL;
+        U8 *tmp_graydata = nullptr;
 
         if (fontp->mFTFace->glyph->bitmap.pixel_mode
             == FT_PIXEL_MODE_MONO)
@@ -704,7 +632,7 @@ void LLFontFreetype::insertGlyphInfo(llwchar wch, LLFontGlyphInfo* gi) const
 
 void LLFontFreetype::renderGlyph(EFontGlyphType bitmap_type, U32 glyph_index, llwchar wch) const
 {
-    if (mFTFace == NULL)
+    if (mFTFace == nullptr)
         return;
 
     FT_Int32 load_flags = FT_LOAD_FORCE_AUTOHINT;
@@ -918,3 +846,58 @@ void LLFontFreetype::setSubImageLuminanceAlpha(U32 x, U32 y, U32 bitmap_num, U32
     }
 }
 
+
+namespace ll
+{
+    namespace fonts
+    {
+        class LoadedFont
+        {
+            public:
+            LoadedFont( std::string aName , std::string const &aAddress, std::size_t aSize )
+            : mAddress( aAddress )
+            {
+                mName = aName;
+                mSize = aSize;
+                mRefs = 1;
+            }
+            std::string mName;
+            std::string mAddress;
+            std::size_t mSize;
+            U32  mRefs;
+        };
+    }
+}
+
+U8 const* LLFontManager::loadFont( std::string const &aFilename, long &a_Size)
+{
+    a_Size = 0;
+    std::map< std::string, std::shared_ptr<ll::fonts::LoadedFont> >::iterator itr = m_LoadedFonts.find( aFilename );
+    if( itr != m_LoadedFonts.end() )
+    {
+        ++itr->second->mRefs;
+        // A possible overflow cannot happen here, as it is asserted that the size is less than std::numeric_limits<long>::max() a few lines below.
+        a_Size = static_cast<long>(itr->second->mSize);
+        return reinterpret_cast<U8 const*>(itr->second->mAddress.c_str());
+    }
+
+    auto strContent = LLFile::getContents(aFilename);
+
+    if( strContent.empty() )
+        return nullptr;
+
+    // For fontconfig a type of long is required, std::string::size() returns size_t. I think it is safe to limit this to 2GiB and not support fonts that huge (can that even be a thing?)
+    llassert_always( strContent.size() < std::numeric_limits<long>::max() );
+
+    a_Size = static_cast<long>(strContent.size());
+
+    auto pCache = std::make_shared<ll::fonts::LoadedFont>( aFilename,  strContent, a_Size );
+    itr = m_LoadedFonts.insert( std::make_pair( aFilename, pCache ) ).first;
+
+    return reinterpret_cast<U8 const*>(itr->second->mAddress.c_str());
+}
+
+void LLFontManager::unloadAllFonts()
+{
+    m_LoadedFonts.clear();
+}
diff --git a/indra/llrender/llfontfreetype.h b/indra/llrender/llfontfreetype.h
index 783bf4a4b3..a9b3a944ee 100644
--- a/indra/llrender/llfontfreetype.h
+++ b/indra/llrender/llfontfreetype.h
@@ -43,15 +43,28 @@ typedef struct FT_FaceRec_* LLFT_Face;
 struct FT_StreamRec_;
 typedef struct FT_StreamRec_ LLFT_Stream;
 
+namespace ll
+{
+    namespace fonts
+    {
+        class LoadedFont;
+    }
+}
+
 class LLFontManager
 {
 public:
     static void initClass();
     static void cleanupClass();
 
+    U8 const *loadFont( std::string const &aFilename, long &a_Size );
+
 private:
     LLFontManager();
     ~LLFontManager();
+
+    void unloadAllFonts();
+    std::map< std::string, std::shared_ptr<ll::fonts::LoadedFont> > m_LoadedFonts;
 };
 
 struct LLFontGlyphInfo
@@ -90,11 +103,6 @@ public:
 
     S32 getNumFaces(const std::string& filename);
 
-#ifdef LL_WINDOWS
-    S32 ftOpenFace(const std::string& filename, S32 face_n);
-    void clearFontStreams();
-#endif
-
     typedef std::function<bool(llwchar)> char_functor_t;
     void addFallbackFont(const LLPointer<LLFontFreetype>& fallback_font, const char_functor_t& functor = nullptr);
 
@@ -170,11 +178,6 @@ private:
 
     LLFT_Face mFTFace;
 
-#ifdef LL_WINDOWS
-    llifstream *pFileStream;
-    LLFT_Stream *pFtStream;
-#endif
-
     bool mIsFallback;
     typedef std::pair<LLPointer<LLFontFreetype>, char_functor_t> fallback_font_t;
     typedef std::vector<fallback_font_t> fallback_font_vector_t;
diff --git a/indra/llrender/llgl.cpp b/indra/llrender/llgl.cpp
index 0be27b8975..d13b98e274 100644
--- a/indra/llrender/llgl.cpp
+++ b/indra/llrender/llgl.cpp
@@ -1123,17 +1123,6 @@ bool LLGLManager::initGL()
     if (mGLVersion >= 2.f)
     {
         parse_glsl_version(mGLSLVersionMajor, mGLSLVersionMinor);
-
-#if 0 && LL_DARWIN
-        // TODO maybe switch to using a core profile for GL 3.2?
-        // https://stackoverflow.com/a/19868861
-        //never use GLSL greater than 1.20 on OSX
-        if (mGLSLVersionMajor > 1 || mGLSLVersionMinor > 30)
-        {
-            mGLSLVersionMajor = 1;
-            mGLSLVersionMinor = 30;
-        }
-#endif
     }
 
     if (mGLVersion >= 2.1f && LLImageGL::sCompressTextures)
@@ -1247,7 +1236,7 @@ bool LLGLManager::initGL()
     // there's some implementation that reports a crazy value
     mMaxUniformBlockSize = llmin(mMaxUniformBlockSize, 65536);
 
-    if (mGLVersion >= 4.59f)
+    if (mHasAnisotropic)
     {
         glGetFloatv(GL_MAX_TEXTURE_MAX_ANISOTROPY, &mMaxAnisotropy);
     }
@@ -1411,6 +1400,11 @@ void LLGLManager::initExtensions()
     mHasCubeMapArray = mGLVersion >= 3.99f;
     mHasTransformFeedback = mGLVersion >= 3.99f;
     mHasDebugOutput = mGLVersion >= 4.29f;
+    mHasAnisotropic = mGLVersion >= 4.59f;
+    if(!mHasAnisotropic && gGLHExts.mSysExts)
+    {
+        mHasAnisotropic = ExtensionExists("GL_EXT_texture_filter_anisotropic", gGLHExts.mSysExts);
+    }
 
     // Misc
     glGetIntegerv(GL_MAX_ELEMENTS_VERTICES, (GLint*) &mGLMaxVertexRange);
@@ -2728,7 +2722,7 @@ void LLGLUserClipPlane::setPlane(F32 a, F32 b, F32 c, F32 d)
     if(cplane[2] < 0)
         cplane *= -1;
 
-    glm::mat4 suffix;
+    glm::mat4 suffix = glm::identity<glm::mat4>();
     suffix = glm::row(suffix, 2, cplane);
     glm::mat4 newP = suffix * P;
     gGL.matrixMode(LLRender::MM_PROJECTION);
diff --git a/indra/llrender/llglslshader.cpp b/indra/llrender/llglslshader.cpp
index 3735a99109..9cd5dc8145 100644
--- a/indra/llrender/llglslshader.cpp
+++ b/indra/llrender/llglslshader.cpp
@@ -421,8 +421,11 @@ bool LLGLSLShader::createShader()
     llassert_always(!mShaderFiles.empty());
 
 #if LL_DARWIN
-    // work-around missing mix(vec3,vec3,bvec3)
-    mDefines["OLD_SELECT"] = "1";
+    if(!gGLManager.mIsApple)
+    {
+        // work-around missing mix(vec3,vec3,bvec3)
+        mDefines["OLD_SELECT"] = "1";
+    }
 #endif
 
     mShaderHash = hash();
@@ -543,7 +546,7 @@ bool LLGLSLShader::createShader()
         }
     }
 
-#ifdef LL_PROFILER_ENABLE_RENDER_DOC
+#if LL_PROFILER_ENABLE_RENDER_DOC
     setLabel(mName.c_str());
 #endif
 
@@ -2078,7 +2081,7 @@ LLUUID LLGLSLShader::hash()
     return hash_obj.digest();
 }
 
-#ifdef LL_PROFILER_ENABLE_RENDER_DOC
+#if LL_PROFILER_ENABLE_RENDER_DOC
 void LLGLSLShader::setLabel(const char* label) {
     LL_LABEL_OBJECT_GL(GL_PROGRAM, mProgramObject, strlen(label), label);
 }
diff --git a/indra/llrender/llglslshader.h b/indra/llrender/llglslshader.h
index 873ab0cff5..4702a27cc5 100644
--- a/indra/llrender/llglslshader.h
+++ b/indra/llrender/llglslshader.h
@@ -361,7 +361,7 @@ public:
     // hacky flag used for optimization in LLDrawPoolAlpha
     bool mCanBindFast = false;
 
-#ifdef LL_PROFILER_ENABLE_RENDER_DOC
+#if LL_PROFILER_ENABLE_RENDER_DOC
     void setLabel(const char* label);
 #endif
 
@@ -381,7 +381,7 @@ extern LLGLSLShader         gSolidColorProgram;
 //Alpha mask shader (declared here so llappearance can access properly)
 extern LLGLSLShader         gAlphaMaskProgram;
 
-#ifdef LL_PROFILER_ENABLE_RENDER_DOC
+#if LL_PROFILER_ENABLE_RENDER_DOC
 #define LL_SET_SHADER_LABEL(shader) shader.setLabel(#shader)
 #else
 #define LL_SET_SHADER_LABEL(shader, label)
diff --git a/indra/llrender/llimagegl.cpp b/indra/llrender/llimagegl.cpp
index 1db36d91f9..52738ec626 100644
--- a/indra/llrender/llimagegl.cpp
+++ b/indra/llrender/llimagegl.cpp
@@ -338,6 +338,7 @@ S32 LLImageGL::dataFormatBits(S32 dataformat)
     case GL_BGRA:                                   return 32;      // Used for QuickTime media textures on the Mac
     case GL_DEPTH_COMPONENT:                        return 24;
     case GL_DEPTH_COMPONENT24:                      return 24;
+    case GL_RGBA16:                                 return 64;
     case GL_R16F:                                   return 16;
     case GL_RG16F:                                  return 32;
     case GL_RGB16F:                                 return 48;
diff --git a/indra/llrender/llrender.cpp b/indra/llrender/llrender.cpp
index 1dc87a66ce..4f646cdc33 100644
--- a/indra/llrender/llrender.cpp
+++ b/indra/llrender/llrender.cpp
@@ -118,7 +118,7 @@ static const GLenum sGLBlendFactor[] =
 
 LLTexUnit::LLTexUnit(S32 index)
     : mCurrTexType(TT_NONE),
-    mCurrColorScale(1), mCurrAlphaScale(1), mCurrTexture(0),
+    mCurrTexture(0),
     mHasMipMaps(false),
     mIndex(index)
 {
@@ -207,6 +207,12 @@ void LLTexUnit::bindFast(LLTexture* texture)
     }
     glBindTexture(sGLTextureType[gl_tex->getTarget()], mCurrTexture);
     mHasMipMaps = gl_tex->mHasMipMaps;
+    if (gl_tex->mTexOptionsDirty)
+    {
+        gl_tex->mTexOptionsDirty = false;
+        setTextureAddressModeFast(gl_tex->mAddressMode);
+        setTextureFilteringOptionFast(gl_tex->mFilterOption);
+    }
 }
 
 bool LLTexUnit::bind(LLTexture* texture, bool for_rendering, bool forceBind)
@@ -461,11 +467,16 @@ void LLTexUnit::setTextureAddressMode(eTextureAddressMode mode)
 
     activate();
 
-    glTexParameteri (sGLTextureType[mCurrTexType], GL_TEXTURE_WRAP_S, sGLAddressMode[mode]);
-    glTexParameteri (sGLTextureType[mCurrTexType], GL_TEXTURE_WRAP_T, sGLAddressMode[mode]);
+    setTextureAddressModeFast(mode);
+}
+
+void LLTexUnit::setTextureAddressModeFast(eTextureAddressMode mode)
+{
+    glTexParameteri(sGLTextureType[mCurrTexType], GL_TEXTURE_WRAP_S, sGLAddressMode[mode]);
+    glTexParameteri(sGLTextureType[mCurrTexType], GL_TEXTURE_WRAP_T, sGLAddressMode[mode]);
     if (mCurrTexType == TT_CUBE_MAP)
     {
-        glTexParameteri (GL_TEXTURE_CUBE_MAP, GL_TEXTURE_WRAP_R, sGLAddressMode[mode]);
+        glTexParameteri(GL_TEXTURE_CUBE_MAP, GL_TEXTURE_WRAP_R, sGLAddressMode[mode]);
     }
 }
 
@@ -475,6 +486,11 @@ void LLTexUnit::setTextureFilteringOption(LLTexUnit::eTextureFilterOptions optio
 
     gGL.flush();
 
+    setTextureFilteringOptionFast(option);
+}
+
+void LLTexUnit::setTextureFilteringOptionFast(LLTexUnit::eTextureFilterOptions option)
+{
     if (option == TFO_POINT)
     {
         glTexParameteri(sGLTextureType[mCurrTexType], GL_TEXTURE_MAG_FILTER, GL_NEAREST);
@@ -511,7 +527,7 @@ void LLTexUnit::setTextureFilteringOption(LLTexUnit::eTextureFilterOptions optio
         }
     }
 
-    if (gGLManager.mGLVersion >= 4.59f)
+    if (gGLManager.mHasAnisotropic)
     {
         if (LLImageGL::sGlobalUseAnisotropic && option == TFO_ANISOTROPIC)
         {
@@ -600,26 +616,6 @@ GLint LLTexUnit::getTextureSourceType(eTextureBlendSrc src, bool isAlpha)
     }
 }
 
-void LLTexUnit::setColorScale(S32 scale)
-{
-    if (mCurrColorScale != scale || gGL.mDirty)
-    {
-        mCurrColorScale = scale;
-        gGL.flush();
-        glTexEnvi( GL_TEXTURE_ENV, GL_RGB_SCALE, scale );
-    }
-}
-
-void LLTexUnit::setAlphaScale(S32 scale)
-{
-    if (mCurrAlphaScale != scale || gGL.mDirty)
-    {
-        mCurrAlphaScale = scale;
-        gGL.flush();
-        glTexEnvi( GL_TEXTURE_ENV, GL_ALPHA_SCALE, scale );
-    }
-}
-
 // Useful for debugging that you've manually assigned a texture operation to the correct
 // texture unit based on the currently set active texture in opengl.
 void LLTexUnit::debugTextureUnit(void)
@@ -1283,9 +1279,7 @@ void LLRender::translateUI(F32 x, F32 y, F32 z)
         LL_ERRS() << "Need to push a UI translation frame before offsetting" << LL_ENDL;
     }
 
-    mUIOffset.back().mV[0] += x;
-    mUIOffset.back().mV[1] += y;
-    mUIOffset.back().mV[2] += z;
+    mUIOffset.back().add(LLVector4a(x, y, z));
 }
 
 void LLRender::scaleUI(F32 x, F32 y, F32 z)
@@ -1295,14 +1289,14 @@ void LLRender::scaleUI(F32 x, F32 y, F32 z)
         LL_ERRS() << "Need to push a UI transformation frame before scaling." << LL_ENDL;
     }
 
-    mUIScale.back().scaleVec(LLVector3(x,y,z));
+    mUIScale.back().mul(LLVector4a(x, y, z));
 }
 
 void LLRender::pushUIMatrix()
 {
     if (mUIOffset.empty())
     {
-        mUIOffset.emplace_back(0.f,0.f,0.f);
+        mUIOffset.emplace_back(0.f);
     }
     else
     {
@@ -1311,7 +1305,7 @@ void LLRender::pushUIMatrix()
 
     if (mUIScale.empty())
     {
-        mUIScale.emplace_back(1.f,1.f,1.f);
+        mUIScale.emplace_back(1.f);
     }
     else
     {
@@ -1333,18 +1327,20 @@ LLVector3 LLRender::getUITranslation()
 {
     if (mUIOffset.empty())
     {
-        return LLVector3(0,0,0);
+        return LLVector3::zero;
     }
-    return mUIOffset.back();
+
+    return LLVector3(mUIOffset.back().getF32ptr());
 }
 
 LLVector3 LLRender::getUIScale()
 {
     if (mUIScale.empty())
     {
-        return LLVector3(1,1,1);
+        return LLVector3::all_one;
     }
-    return mUIScale.back();
+
+    return LLVector3(mUIScale.back().getF32ptr());
 }
 
 
@@ -1354,8 +1350,9 @@ void LLRender::loadUIIdentity()
     {
         LL_ERRS() << "Need to push UI translation frame before clearing offset." << LL_ENDL;
     }
-    mUIOffset.back().setVec(0,0,0);
-    mUIScale.back().setVec(1,1,1);
+
+    mUIOffset.back().clear();
+    mUIScale.back().splat(1);
 }
 
 void LLRender::setColorMask(bool writeColor, bool writeAlpha)
@@ -1783,8 +1780,10 @@ void LLRender::vertex3f(const GLfloat& x, const GLfloat& y, const GLfloat& z)
     }
     else
     {
-        LLVector3 vert = (LLVector3(x,y,z)+mUIOffset.back()).scaledVec(mUIScale.back());
-        mVerticesp[mCount].set(vert.mV[VX], vert.mV[VY], vert.mV[VZ]);
+        LLVector4a vert(x, y, z);
+        vert.add(mUIOffset.back());
+        vert.mul(mUIScale.back());
+        mVerticesp[mCount] = vert;
     }
 
     mCount++;
diff --git a/indra/llrender/llrender.h b/indra/llrender/llrender.h
index 7b6bd4198b..755aee4bd6 100644
--- a/indra/llrender/llrender.h
+++ b/indra/llrender/llrender.h
@@ -43,6 +43,7 @@
 #include "llglheaders.h"
 #include "llmatrix4a.h"
 #include "glm/mat4x4.hpp"
+#include <boost/align/aligned_allocator.hpp>
 
 #include <array>
 #include <list>
@@ -207,11 +208,15 @@ public:
     // Warning: this stays set for the bound texture forever,
     // make sure you want to permanently change the address mode  for the bound texture.
     void setTextureAddressMode(eTextureAddressMode mode);
+    // MUST already be active and bound
+    void setTextureAddressModeFast(eTextureAddressMode mode);
 
     // Sets the filtering options used to sample the texture
     // Warning: this stays set for the bound texture forever,
     // make sure you want to permanently change the filtering for the bound texture.
     void setTextureFilteringOption(LLTexUnit::eTextureFilterOptions option);
+    // MUST already be active and bound
+    void setTextureFilteringOptionFast(LLTexUnit::eTextureFilterOptions option);
 
     static U32 getInternalType(eTextureType type);
 
@@ -227,13 +232,9 @@ protected:
     S32                 mIndex;
     U32                 mCurrTexture;
     eTextureType        mCurrTexType;
-    S32                 mCurrColorScale;
-    S32                 mCurrAlphaScale;
     bool                mHasMipMaps;
 
     void debugTextureUnit(void);
-    void setColorScale(S32 scale);
-    void setAlphaScale(S32 scale);
     GLint getTextureSource(eTextureBlendSrc src);
     GLint getTextureSourceType(eTextureBlendSrc src, bool isAlpha = false);
 };
@@ -526,8 +527,8 @@ private:
     eBlendFactor mCurrBlendAlphaSFactor;
     eBlendFactor mCurrBlendAlphaDFactor;
 
-    std::vector<LLVector3> mUIOffset;
-    std::vector<LLVector3> mUIScale;
+    std::vector<LLVector4a, boost::alignment::aligned_allocator<LLVector4a, 16> > mUIOffset;
+    std::vector<LLVector4a, boost::alignment::aligned_allocator<LLVector4a, 16> > mUIScale;
 };
 
 extern F32 gGLModelView[16];
diff --git a/indra/llrender/llshadermgr.cpp b/indra/llrender/llshadermgr.cpp
index a4d5282b0c..9cdd02f403 100644
--- a/indra/llrender/llshadermgr.cpp
+++ b/indra/llrender/llshadermgr.cpp
@@ -458,7 +458,7 @@ GLuint LLShaderMgr::loadShaderFile(const std::string& filename, S32 & shader_lev
 
 // endsure work-around for missing GLSL funcs gets propogated to feature shader files (e.g. srgbF.glsl)
 #if LL_DARWIN
-    if (defines)
+    if (!gGLManager.mIsApple && defines)
     {
         (*defines)["OLD_SELECT"] = "1";
     }
diff --git a/indra/llrender/llvertexbuffer.cpp b/indra/llrender/llvertexbuffer.cpp
index 1f0c424188..d59ddd0fec 100644
--- a/indra/llrender/llvertexbuffer.cpp
+++ b/indra/llrender/llvertexbuffer.cpp
@@ -787,12 +787,6 @@ void LLVertexBuffer::drawElements(U32 mode, const LLVector4a* pos, const LLVecto
 
     gGL.syncMatrices();
 
-    U32 mask = LLVertexBuffer::MAP_VERTEX;
-    if (tc)
-    {
-        mask = mask | LLVertexBuffer::MAP_TEXCOORD0;
-    }
-
     unbind();
 
     gGL.begin(mode);
@@ -885,7 +879,7 @@ bool LLVertexBuffer::validateRange(U32 start, U32 end, U32 count, U32 indices_of
     return true;
 }
 
-#ifdef LL_PROFILER_ENABLE_RENDER_DOC
+#if LL_PROFILER_ENABLE_RENDER_DOC
 void LLVertexBuffer::setLabel(const char* label) {
     LL_LABEL_OBJECT_GL(GL_BUFFER, mGLBuffer, strlen(label), label);
 }
diff --git a/indra/llrender/llvertexbuffer.h b/indra/llrender/llvertexbuffer.h
index 375ad76fb8..faaa6ba0f0 100644
--- a/indra/llrender/llvertexbuffer.h
+++ b/indra/llrender/llvertexbuffer.h
@@ -279,7 +279,7 @@ public:
     //for debugging, validate data in given range is valid
     bool validateRange(U32 start, U32 end, U32 count, U32 offset) const;
 
-    #ifdef LL_PROFILER_ENABLE_RENDER_DOC
+    #if LL_PROFILER_ENABLE_RENDER_DOC
     void setLabel(const char* label);
     #endif
 
@@ -340,7 +340,7 @@ public:
     static U32 sVertexCount;
 };
 
-#ifdef LL_PROFILER_ENABLE_RENDER_DOC
+#if LL_PROFILER_ENABLE_RENDER_DOC
 #define LL_LABEL_VERTEX_BUFFER(buf, name) buf->setLabel(name)
 #else
 #define LL_LABEL_VERTEX_BUFFER(buf, name)