Merge branch 'main' into maint-b

8 files changed, 543 insertions, 186 deletions

diff --git a/indra/llrender/llfontgl.cpp b/indra/llrender/llfontgl.cpp
index 3714bb1883..2ac2615c0a 100644
--- a/indra/llrender/llfontgl.cpp
+++ b/indra/llrender/llfontgl.cpp
@@ -270,10 +270,10 @@ S32 LLFontGL::render(const LLWString &wstr, S32 begin_offset, F32 x, F32 y, cons
 
     const LLFontGlyphInfo* next_glyph = NULL;
 
-    const S32 GLYPH_BATCH_SIZE = 30;
-    LLVector3 vertices[GLYPH_BATCH_SIZE * 4];
-    LLVector2 uvs[GLYPH_BATCH_SIZE * 4];
-    LLColor4U colors[GLYPH_BATCH_SIZE * 4];
+    const S32 GLYPH_BATCH_SIZE = 120;
+    LLVector4a vertices[GLYPH_BATCH_SIZE * 6];
+    LLVector2 uvs[GLYPH_BATCH_SIZE * 6];
+    LLColor4U colors[GLYPH_BATCH_SIZE * 6];
 
     LLColor4U text_color(color);
     // Preserve the transparency to render fading emojis in fading text (e.g.
@@ -305,9 +305,9 @@ S32 LLFontGL::render(const LLWString &wstr, S32 begin_offset, F32 x, F32 y, cons
             // otherwise the queued glyphs will be taken from wrong textures.
             if (glyph_count > 0)
             {
-                gGL.begin(LLRender::QUADS);
+                gGL.begin(LLRender::TRIANGLES);
                 {
-                    gGL.vertexBatchPreTransformed(vertices, uvs, colors, glyph_count * 4);
+                    gGL.vertexBatchPreTransformed(vertices, uvs, colors, glyph_count * 6);
                 }
                 gGL.end();
                 glyph_count = 0;
@@ -338,12 +338,11 @@ S32 LLFontGL::render(const LLWString &wstr, S32 begin_offset, F32 x, F32 y, cons
 
         if (glyph_count >= GLYPH_BATCH_SIZE)
         {
-            gGL.begin(LLRender::QUADS);
+            gGL.begin(LLRender::TRIANGLES);
             {
-                gGL.vertexBatchPreTransformed(vertices, uvs, colors, glyph_count * 4);
+                gGL.vertexBatchPreTransformed(vertices, uvs, colors, glyph_count * 6);
             }
             gGL.end();
-
             glyph_count = 0;
         }
 
@@ -376,9 +375,9 @@ S32 LLFontGL::render(const LLWString &wstr, S32 begin_offset, F32 x, F32 y, cons
         cur_render_y = cur_y;
     }
 
-    gGL.begin(LLRender::QUADS);
+    gGL.begin(LLRender::TRIANGLES);
     {
-        gGL.vertexBatchPreTransformed(vertices, uvs, colors, glyph_count * 4);
+        gGL.vertexBatchPreTransformed(vertices, uvs, colors, glyph_count * 6);
     }
     gGL.end();
 
@@ -1226,31 +1225,42 @@ LLFontGL &LLFontGL::operator=(const LLFontGL &source)
     return *this;
 }
 
-void LLFontGL::renderQuad(LLVector3* vertex_out, LLVector2* uv_out, LLColor4U* colors_out, const LLRectf& screen_rect, const LLRectf& uv_rect, const LLColor4U& color, F32 slant_amt) const
+void LLFontGL::renderTriangle(LLVector4a* vertex_out, LLVector2* uv_out, LLColor4U* colors_out, const LLRectf& screen_rect, const LLRectf& uv_rect, const LLColor4U& color, F32 slant_amt) const
 {
     S32 index = 0;
 
-    vertex_out[index] = LLVector3(screen_rect.mRight, screen_rect.mTop, 0.f);
+    vertex_out[index] = LLVector4a(screen_rect.mRight, screen_rect.mTop, 0.f);
     uv_out[index] = LLVector2(uv_rect.mRight, uv_rect.mTop);
     colors_out[index] = color;
     index++;
 
-    vertex_out[index] = LLVector3(screen_rect.mLeft, screen_rect.mTop, 0.f);
+    vertex_out[index] = LLVector4a(screen_rect.mLeft, screen_rect.mTop, 0.f);
     uv_out[index] = LLVector2(uv_rect.mLeft, uv_rect.mTop);
     colors_out[index] = color;
     index++;
 
-    vertex_out[index] = LLVector3(screen_rect.mLeft, screen_rect.mBottom, 0.f);
+    vertex_out[index] = LLVector4a(screen_rect.mLeft, screen_rect.mBottom, 0.f);
+    uv_out[index] = LLVector2(uv_rect.mLeft, uv_rect.mBottom);
+    colors_out[index] = color;
+    index++;
+
+
+    vertex_out[index] = LLVector4a(screen_rect.mRight, screen_rect.mTop, 0.f);
+    uv_out[index] = LLVector2(uv_rect.mRight, uv_rect.mTop);
+    colors_out[index] = color;
+    index++;
+
+    vertex_out[index] = LLVector4a(screen_rect.mLeft, screen_rect.mBottom, 0.f);
     uv_out[index] = LLVector2(uv_rect.mLeft, uv_rect.mBottom);
     colors_out[index] = color;
     index++;
 
-    vertex_out[index] = LLVector3(screen_rect.mRight, screen_rect.mBottom, 0.f);
+    vertex_out[index] = LLVector4a(screen_rect.mRight, screen_rect.mBottom, 0.f);
     uv_out[index] = LLVector2(uv_rect.mRight, uv_rect.mBottom);
     colors_out[index] = color;
 }
 
-void LLFontGL::drawGlyph(S32& glyph_count, LLVector3* vertex_out, LLVector2* uv_out, LLColor4U* colors_out, const LLRectf& screen_rect, const LLRectf& uv_rect, const LLColor4U& color, U8 style, ShadowType shadow, F32 drop_shadow_strength) const
+void LLFontGL::drawGlyph(S32& glyph_count, LLVector4a* vertex_out, LLVector2* uv_out, LLColor4U* colors_out, const LLRectf& screen_rect, const LLRectf& uv_rect, const LLColor4U& color, U8 style, ShadowType shadow, F32 drop_shadow_strength) const
 {
     F32 slant_offset;
     slant_offset = ((style & ITALIC) ? ( -mFontFreetype->getAscenderHeight() * 0.2f) : 0.f);
@@ -1264,7 +1274,7 @@ void LLFontGL::drawGlyph(S32& glyph_count, LLVector3* vertex_out, LLVector2* uv_
             LLRectf screen_rect_offset = screen_rect;
 
             screen_rect_offset.translate((F32)(pass * BOLD_OFFSET), 0.f);
-            renderQuad(&vertex_out[glyph_count * 4], &uv_out[glyph_count * 4], &colors_out[glyph_count * 4], screen_rect_offset, uv_rect, color, slant_offset);
+            renderTriangle(&vertex_out[glyph_count * 6], &uv_out[glyph_count * 6], &colors_out[glyph_count * 6], screen_rect_offset, uv_rect, color, slant_offset);
             glyph_count++;
         }
     }
@@ -1295,10 +1305,10 @@ void LLFontGL::drawGlyph(S32& glyph_count, LLVector3* vertex_out, LLVector2* uv_
                 break;
             }
 
-            renderQuad(&vertex_out[glyph_count * 4], &uv_out[glyph_count * 4], &colors_out[glyph_count * 4], screen_rect_offset, uv_rect, shadow_color, slant_offset);
+            renderTriangle(&vertex_out[glyph_count * 6], &uv_out[glyph_count * 6], &colors_out[glyph_count * 6], screen_rect_offset, uv_rect, shadow_color, slant_offset);
             glyph_count++;
         }
-        renderQuad(&vertex_out[glyph_count * 4], &uv_out[glyph_count * 4], &colors_out[glyph_count * 4], screen_rect, uv_rect, color, slant_offset);
+        renderTriangle(&vertex_out[glyph_count * 6], &uv_out[glyph_count * 6], &colors_out[glyph_count * 6], screen_rect, uv_rect, color, slant_offset);
         glyph_count++;
     }
     else if (shadow == DROP_SHADOW)
@@ -1307,14 +1317,14 @@ void LLFontGL::drawGlyph(S32& glyph_count, LLVector3* vertex_out, LLVector2* uv_
         shadow_color.mV[VALPHA] = U8(color.mV[VALPHA] * drop_shadow_strength);
         LLRectf screen_rect_shadow = screen_rect;
         screen_rect_shadow.translate(1.f, -1.f);
-        renderQuad(&vertex_out[glyph_count * 4], &uv_out[glyph_count * 4], &colors_out[glyph_count * 4], screen_rect_shadow, uv_rect, shadow_color, slant_offset);
+        renderTriangle(&vertex_out[glyph_count * 6], &uv_out[glyph_count * 6], &colors_out[glyph_count * 6], screen_rect_shadow, uv_rect, shadow_color, slant_offset);
         glyph_count++;
-        renderQuad(&vertex_out[glyph_count * 4], &uv_out[glyph_count * 4], &colors_out[glyph_count * 4], screen_rect, uv_rect, color, slant_offset);
+        renderTriangle(&vertex_out[glyph_count * 6], &uv_out[glyph_count * 6], &colors_out[glyph_count * 6], screen_rect, uv_rect, color, slant_offset);
         glyph_count++;
     }
     else // normal rendering
     {
-        renderQuad(&vertex_out[glyph_count * 4], &uv_out[glyph_count * 4], &colors_out[glyph_count * 4], screen_rect, uv_rect, color, slant_offset);
+        renderTriangle(&vertex_out[glyph_count * 6], &uv_out[glyph_count * 6], &colors_out[glyph_count * 6], screen_rect, uv_rect, color, slant_offset);
         glyph_count++;
     }
 }
diff --git a/indra/llrender/llfontgl.h b/indra/llrender/llfontgl.h
index 65f0a8cbfd..81674f4a64 100644
--- a/indra/llrender/llfontgl.h
+++ b/indra/llrender/llfontgl.h
@@ -238,8 +238,8 @@ private:
     LLFontDescriptor mFontDescriptor;
     LLPointer<LLFontFreetype> mFontFreetype;
 
-    void renderQuad(LLVector3* vertex_out, LLVector2* uv_out, LLColor4U* colors_out, const LLRectf& screen_rect, const LLRectf& uv_rect, const LLColor4U& color, F32 slant_amt) const;
-    void drawGlyph(S32& glyph_count, LLVector3* vertex_out, LLVector2* uv_out, LLColor4U* colors_out, const LLRectf& screen_rect, const LLRectf& uv_rect, const LLColor4U& color, U8 style, ShadowType shadow, F32 drop_shadow_fade) const;
+    void renderTriangle(LLVector4a* vertex_out, LLVector2* uv_out, LLColor4U* colors_out, const LLRectf& screen_rect, const LLRectf& uv_rect, const LLColor4U& color, F32 slant_amt) const;
+    void drawGlyph(S32& glyph_count, LLVector4a* vertex_out, LLVector2* uv_out, LLColor4U* colors_out, const LLRectf& screen_rect, const LLRectf& uv_rect, const LLColor4U& color, U8 style, ShadowType shadow, F32 drop_shadow_fade) const;
 
     // Registry holds all instantiated fonts.
     static LLFontRegistry* sFontRegistry;
diff --git a/indra/llrender/llgl.cpp b/indra/llrender/llgl.cpp
index 99777c5f62..eae395e90c 100644
--- a/indra/llrender/llgl.cpp
+++ b/indra/llrender/llgl.cpp
@@ -997,6 +997,7 @@ LLGLManager::LLGLManager() :
     mIsAMD(FALSE),
     mIsNVIDIA(FALSE),
     mIsIntel(FALSE),
+    mIsApple(FALSE),
 #if LL_DARWIN
     mIsMobileGF(FALSE),
 #endif
@@ -1170,6 +1171,11 @@ bool LLGLManager::initGL()
         mGLVendorShort = "INTEL";
         mIsIntel = TRUE;
     }
+    else if(mGLVendor.find("APPLE") != std::string::npos)
+    {
+        mGLVendorShort = "APPLE";
+        mIsApple = TRUE;
+    }
     else
     {
         mGLVendorShort = "MISC";
@@ -1369,6 +1375,7 @@ void LLGLManager::asLLSD(LLSD& info)
     info["is_ati"] = mIsAMD;  // note, do not rename is_ati to is_amd without coordinating with DW
     info["is_nvidia"] = mIsNVIDIA;
     info["is_intel"] = mIsIntel;
+    info["is_apple"] = mIsApple;
 
     info["gl_renderer"] = mGLRenderer;
 }
diff --git a/indra/llrender/llgl.h b/indra/llrender/llgl.h
index 5a7ad943df..e4b106c999 100644
--- a/indra/llrender/llgl.h
+++ b/indra/llrender/llgl.h
@@ -100,6 +100,7 @@ public:
     BOOL mIsAMD;
     BOOL mIsNVIDIA;
     BOOL mIsIntel;
+    BOOL mIsApple;
 
 #if LL_DARWIN
     // Needed to distinguish problem cards on older Macs that break with Materials
diff --git a/indra/llrender/llrender.cpp b/indra/llrender/llrender.cpp
index f28eefcb63..7ca628bac5 100644
--- a/indra/llrender/llrender.cpp
+++ b/indra/llrender/llrender.cpp
@@ -50,6 +50,8 @@ extern void APIENTRY gl_debug_callback(GLenum source,
 
 thread_local LLRender gGL;
 
+const U32 BATCH_SIZE = 16334;
+
 // Handy copies of last good GL matrices
 F32 gGLModelView[16];
 F32 gGLLastModelView[16];
@@ -66,7 +68,7 @@ S32 gGLViewport[4];
 U32 LLRender::sUICalls = 0;
 U32 LLRender::sUIVerts = 0;
 U32 LLTexUnit::sWhiteTexture = 0;
-bool LLRender::sGLCoreProfile = false;
+bool LLRender::sGLCoreProfile = true;
 bool LLRender::sNsightDebugSupport = false;
 LLVector2 LLRender::sUIGLScaleFactor = LLVector2(1.f, 1.f);
 
@@ -920,7 +922,7 @@ void LLRender::initVertexBuffer()
     llassert_always(mBuffer.isNull());
     stop_glerror();
     mBuffer = new LLVertexBuffer(immediate_mask);
-    mBuffer->allocateBuffer(4096, 0);
+    mBuffer->allocateBuffer(BATCH_SIZE, 0);
     mBuffer->getVertexStrider(mVerticesp);
     mBuffer->getTexCoord0Strider(mTexcoordsp);
     mBuffer->getColorStrider(mColorsp);
@@ -1333,9 +1335,8 @@ 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;
+    LLVector4a add(x,y,z);
+    mUIOffset.back().add(add);
 }
 
 void LLRender::scaleUI(F32 x, F32 y, F32 z)
@@ -1345,14 +1346,15 @@ 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));
+    LLVector4a scale(x,y,z);
+    mUIScale.back().mul(scale);
 }
 
 void LLRender::pushUIMatrix()
 {
     if (mUIOffset.empty())
     {
-        mUIOffset.push_back(LLVector3(0,0,0));
+        mUIOffset.emplace_back(LLVector4a::getZero());
     }
     else
     {
@@ -1361,7 +1363,7 @@ void LLRender::pushUIMatrix()
 
     if (mUIScale.empty())
     {
-        mUIScale.push_back(LLVector3(1,1,1));
+        mUIScale.emplace_back(LLVector4a(1.f));
     }
     else
     {
@@ -1385,7 +1387,7 @@ LLVector3 LLRender::getUITranslation()
     {
         return LLVector3(0,0,0);
     }
-    return mUIOffset.back();
+    return LLVector3(mUIOffset.back().getF32ptr());
 }
 
 LLVector3 LLRender::getUIScale()
@@ -1394,18 +1396,17 @@ LLVector3 LLRender::getUIScale()
     {
         return LLVector3(1,1,1);
     }
-    return mUIScale.back();
+    return LLVector3(mUIScale.back().getF32ptr());
 }
 
-
 void LLRender::loadUIIdentity()
 {
     if (mUIOffset.empty())
     {
         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().splat(0.f);
+    mUIScale.back().splat(1.f);
 }
 
 void LLRender::setColorMask(bool writeColor, bool writeAlpha)
@@ -1594,7 +1595,7 @@ void LLRender::end()
         mMode != LLRender::LINES &&
         mMode != LLRender::TRIANGLES &&
         mMode != LLRender::POINTS) ||
-        mCount > 2048)
+        mCount > (BATCH_SIZE / 2))
     {
         flush();
     }
@@ -1614,28 +1615,28 @@ void LLRender::flush()
         //store mCount in a local variable to avoid re-entrance (drawArrays may call flush)
         U32 count = mCount;
 
-            if (mMode == LLRender::QUADS && !sGLCoreProfile)
+        if (mMode == LLRender::QUADS && !sGLCoreProfile)
+        {
+            if (mCount%4 != 0)
             {
-                if (mCount%4 != 0)
-                {
                 count -= (mCount % 4);
                 LL_WARNS() << "Incomplete quad requested." << LL_ENDL;
-                }
             }
+        }
 
-            if (mMode == LLRender::TRIANGLES)
+        if (mMode == LLRender::TRIANGLES)
+        {
+            if (mCount%3 != 0)
             {
-                if (mCount%3 != 0)
-                {
                 count -= (mCount % 3);
                 LL_WARNS() << "Incomplete triangle requested." << LL_ENDL;
-                }
             }
+        }
 
-            if (mMode == LLRender::LINES)
+        if (mMode == LLRender::LINES)
+        {
+            if (mCount%2 != 0)
             {
-                if (mCount%2 != 0)
-                {
                 count -= (mCount % 2);
                 LL_WARNS() << "Incomplete line requested." << LL_ENDL;
             }
@@ -1697,7 +1698,7 @@ void LLRender::flush()
 
                 vb->setBuffer();
 
-                vb->setPositionData((LLVector4a*) mVerticesp.get());
+                vb->setPositionData(mVerticesp.get());
 
                 if (attribute_mask & LLVertexBuffer::MAP_TEXCOORD0)
                 {
@@ -1764,10 +1765,10 @@ void LLRender::flush()
     }
 }
 
-void LLRender::vertex3f(const GLfloat& x, const GLfloat& y, const GLfloat& z)
+void LLRender::vertex4a(const LLVector4a& vertex)
 {
     //the range of mVerticesp, mColorsp and mTexcoordsp is [0, 4095]
-    if (mCount > 2048)
+    if (mCount > BATCH_SIZE / 2)
     { //break when buffer gets reasonably full to keep GL command buffers happy and avoid overflow below
         switch (mMode)
         {
@@ -1778,20 +1779,20 @@ void LLRender::vertex3f(const GLfloat& x, const GLfloat& y, const GLfloat& z)
         }
     }
 
-    if (mCount > 4094)
+    if (mCount > BATCH_SIZE - 2)
     {
-    //  LL_WARNS() << "GL immediate mode overflow.  Some geometry not drawn." << LL_ENDL;
+        LL_WARNS() << "GL immediate mode overflow.  Some geometry not drawn." << LL_ENDL;
         return;
     }
 
     if (mUIOffset.empty())
     {
-        mVerticesp[mCount] = LLVector3(x,y,z);
+        mVerticesp[mCount] = vertex;
     }
     else
     {
-        LLVector3 vert = (LLVector3(x,y,z)+mUIOffset.back()).scaledVec(mUIScale.back());
-        mVerticesp[mCount] = vert;
+        mVerticesp[mCount].setAdd(vertex, mUIOffset.back());
+        mVerticesp[mCount].mul(mUIScale.back());
     }
 
     if (mMode == LLRender::QUADS && LLRender::sGLCoreProfile)
@@ -1819,9 +1820,9 @@ void LLRender::vertex3f(const GLfloat& x, const GLfloat& y, const GLfloat& z)
     mTexcoordsp[mCount] = mTexcoordsp[mCount-1];
 }
 
-void LLRender::vertexBatchPreTransformed(LLVector3* verts, S32 vert_count)
+void LLRender::vertexBatchPreTransformed(LLVector4a* verts, S32 vert_count)
 {
-    if (mCount + vert_count > 4094)
+    if (mCount + vert_count > BATCH_SIZE - 2)
     {
         //  LL_WARNS() << "GL immediate mode overflow.  Some geometry not drawn." << LL_ENDL;
         return;
@@ -1877,9 +1878,9 @@ void LLRender::vertexBatchPreTransformed(LLVector3* verts, S32 vert_count)
         mVerticesp[mCount] = mVerticesp[mCount-1];
 }
 
-void LLRender::vertexBatchPreTransformed(LLVector3* verts, LLVector2* uvs, S32 vert_count)
+void LLRender::vertexBatchPreTransformed(LLVector4a* verts, LLVector2* uvs, S32 vert_count)
 {
-    if (mCount + vert_count > 4094)
+    if (mCount + vert_count > BATCH_SIZE - 2)
     {
         //  LL_WARNS() << "GL immediate mode overflow.  Some geometry not drawn." << LL_ENDL;
         return;
@@ -1938,9 +1939,9 @@ void LLRender::vertexBatchPreTransformed(LLVector3* verts, LLVector2* uvs, S32 v
     }
 }
 
-void LLRender::vertexBatchPreTransformed(LLVector3* verts, LLVector2* uvs, LLColor4U* colors, S32 vert_count)
+void LLRender::vertexBatchPreTransformed(LLVector4a* verts, LLVector2* uvs, LLColor4U* colors, S32 vert_count)
 {
-    if (mCount + vert_count > 4094)
+    if (mCount + vert_count > BATCH_SIZE - 2)
     {
         //  LL_WARNS() << "GL immediate mode overflow.  Some geometry not drawn." << LL_ENDL;
         return;
@@ -2001,25 +2002,6 @@ void LLRender::vertexBatchPreTransformed(LLVector3* verts, LLVector2* uvs, LLCol
     }
 }
 
-void LLRender::vertex2i(const GLint& x, const GLint& y)
-{
-    vertex3f((GLfloat) x, (GLfloat) y, 0);
-}
-
-void LLRender::vertex2f(const GLfloat& x, const GLfloat& y)
-{
-    vertex3f(x,y,0);
-}
-
-void LLRender::vertex2fv(const GLfloat* v)
-{
-    vertex3f(v[0], v[1], 0);
-}
-
-void LLRender::vertex3fv(const GLfloat* v)
-{
-    vertex3f(v[0], v[1], v[2]);
-}
 
 void LLRender::texCoord2f(const GLfloat& x, const GLfloat& y)
 {
diff --git a/indra/llrender/llrender.h b/indra/llrender/llrender.h
index 020682f526..555234356f 100644
--- a/indra/llrender/llrender.h
+++ b/indra/llrender/llrender.h
@@ -416,11 +416,15 @@ public:
 
     void begin(const GLuint& mode);
     void end();
-    void vertex2i(const GLint& x, const GLint& y);
-    void vertex2f(const GLfloat& x, const GLfloat& y);
-    void vertex3f(const GLfloat& x, const GLfloat& y, const GLfloat& z);
-    void vertex2fv(const GLfloat* v);
-    void vertex3fv(const GLfloat* v);
+
+    LL_FORCE_INLINE void vertex2i(const GLint& x, const GLint& y) { vertex4a(LLVector4a((GLfloat)x,(GLfloat)y,0.f)); }
+    LL_FORCE_INLINE void vertex2f(const GLfloat& x, const GLfloat& y) { vertex4a(LLVector4a(x,y,0.f)); }
+    LL_FORCE_INLINE void vertex3f(const GLfloat& x, const GLfloat& y, const GLfloat& z) { vertex4a(LLVector4a(x,y,z)); }
+    LL_FORCE_INLINE void vertex2fv(const GLfloat* v) { vertex4a(LLVector4a(v[0],v[1],0.f)); }
+    LL_FORCE_INLINE void vertex3fv(const GLfloat* v) { vertex4a(LLVector4a(v[0],v[1],v[2])); }
+
+    void vertex4a(const LLVector4a& v);
+
 
     void texCoord2i(const GLint& x, const GLint& y);
     void texCoord2f(const GLfloat& x, const GLfloat& y);
@@ -440,9 +444,9 @@ public:
     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);
-    void vertexBatchPreTransformed(LLVector3* verts, LLVector2* uvs, LLColor4U*, S32 vert_count);
+    void vertexBatchPreTransformed(LLVector4a* verts, S32 vert_count);
+    void vertexBatchPreTransformed(LLVector4a* verts, LLVector2* uvs, S32 vert_count);
+    void vertexBatchPreTransformed(LLVector4a* verts, LLVector2* uvs, LLColor4U*, S32 vert_count);
 
     void setColorMask(bool writeColor, bool writeAlpha);
     void setColorMask(bool writeColorR, bool writeColorG, bool writeColorB, bool writeAlpha);
@@ -500,7 +504,7 @@ private:
     bool                mCurrColorMask[4];
 
     LLPointer<LLVertexBuffer>   mBuffer;
-    LLStrider<LLVector3>        mVerticesp;
+    LLStrider<LLVector4a>        mVerticesp;
     LLStrider<LLVector2>        mTexcoordsp;
     LLStrider<LLColor4U>        mColorsp;
     std::array<LLTexUnit, LL_NUM_TEXTURE_LAYERS> mTexUnits;
@@ -512,8 +516,8 @@ private:
     eBlendFactor mCurrBlendAlphaSFactor;
     eBlendFactor mCurrBlendAlphaDFactor;
 
-    std::vector<LLVector3> mUIOffset;
-    std::vector<LLVector3> mUIScale;
+    std::vector<LLVector4a> mUIOffset;
+    std::vector<LLVector4a> mUIScale;
 
 };
 
diff --git a/indra/llrender/llrender2dutils.cpp b/indra/llrender/llrender2dutils.cpp
index 5939fe68dc..c04b571798 100644
--- a/indra/llrender/llrender2dutils.cpp
+++ b/indra/llrender/llrender2dutils.cpp
@@ -119,14 +119,18 @@ void gl_rect_2d(S32 left, S32 top, S32 right, S32 bottom, BOOL filled )
 {
     gGL.getTexUnit(0)->unbind(LLTexUnit::TT_TEXTURE);
 
-    // Counterclockwise quad will face the viewer
     if( filled )
     {
-        gGL.begin( LLRender::QUADS );
+        gGL.begin( LLRender::TRIANGLES );
+        {
             gGL.vertex2i(left, top);
             gGL.vertex2i(left, bottom);
             gGL.vertex2i(right, bottom);
+
+            gGL.vertex2i(left, top);
+            gGL.vertex2i(right, bottom);
             gGL.vertex2i(right, top);
+        }
         gGL.end();
     }
     else
@@ -171,51 +175,73 @@ void gl_drop_shadow(S32 left, S32 top, S32 right, S32 bottom, const LLColor4 &st
 
     LLColor4 end_color = start_color;
     end_color.mV[VALPHA] = 0.f;
+    gGL.begin(LLRender::TRIANGLES);
 
-    gGL.begin(LLRender::QUADS);
-
-    // Right edge, CCW faces screen
+    // Right edge
+    gGL.color4fv(start_color.mV);
+    gGL.vertex2i(right, top-lines);
+    gGL.vertex2i(right, bottom);
+    gGL.color4fv(end_color.mV);
+    gGL.vertex2i(right+lines, bottom);
     gGL.color4fv(start_color.mV);
-    gGL.vertex2i(right,     top-lines);
-    gGL.vertex2i(right,     bottom);
+    gGL.vertex2i(right, top-lines);
     gGL.color4fv(end_color.mV);
     gGL.vertex2i(right+lines, bottom);
     gGL.vertex2i(right+lines, top-lines);
 
-    // Bottom edge, CCW faces screen
+    // Bottom edge
     gGL.color4fv(start_color.mV);
-    gGL.vertex2i(right,     bottom);
-    gGL.vertex2i(left+lines,    bottom);
+    gGL.vertex2i(right, bottom);
+    gGL.vertex2i(left+lines, bottom);
+    gGL.color4fv(end_color.mV);
+    gGL.vertex2i(left+lines, bottom-lines);
+    gGL.color4fv(start_color.mV);
+    gGL.vertex2i(right, bottom);
     gGL.color4fv(end_color.mV);
-    gGL.vertex2i(left+lines,    bottom-lines);
-    gGL.vertex2i(right,     bottom-lines);
+    gGL.vertex2i(left+lines, bottom-lines);
+    gGL.vertex2i(right, bottom-lines);
 
     // bottom left Corner
     gGL.color4fv(start_color.mV);
-    gGL.vertex2i(left+lines,    bottom);
+    gGL.vertex2i(left+lines, bottom);
+    gGL.color4fv(end_color.mV);
+    gGL.vertex2i(left, bottom);
+
+    // bottom left corner
+    gGL.vertex2i(left+1, bottom-lines+1);
+    gGL.color4fv(start_color.mV);
+    gGL.vertex2i(left+lines, bottom);
     gGL.color4fv(end_color.mV);
-    gGL.vertex2i(left,      bottom);
-    // make the bottom left corner not sharp
-    gGL.vertex2i(left+1,        bottom-lines+1);
-    gGL.vertex2i(left+lines,    bottom-lines);
+    gGL.vertex2i(left+1, bottom-lines+1);
+    gGL.vertex2i(left+lines, bottom-lines);
 
     // bottom right corner
     gGL.color4fv(start_color.mV);
-    gGL.vertex2i(right,     bottom);
+    gGL.vertex2i(right, bottom);
     gGL.color4fv(end_color.mV);
-    gGL.vertex2i(right,     bottom-lines);
+    gGL.vertex2i(right, bottom-lines);
+
     // make the rightmost corner not sharp
     gGL.vertex2i(right+lines-1, bottom-lines+1);
-    gGL.vertex2i(right+lines,   bottom);
+    gGL.color4fv(start_color.mV);
+    gGL.vertex2i(right, bottom);
+    gGL.color4fv(end_color.mV);
+    gGL.vertex2i(right+lines-1, bottom-lines+1);
+    gGL.vertex2i(right+lines, bottom);
 
     // top right corner
     gGL.color4fv(start_color.mV);
-    gGL.vertex2i( right,            top-lines );
+    gGL.vertex2i( right, top-lines );
     gGL.color4fv(end_color.mV);
-    gGL.vertex2i( right+lines,  top-lines );
+    gGL.vertex2i( right+lines, top-lines );
+
     // make the corner not sharp
-    gGL.vertex2i( right+lines-1,    top-1 );
-    gGL.vertex2i( right,            top );
+    gGL.vertex2i( right+lines-1, top-1 );
+    gGL.color4fv(start_color.mV);
+    gGL.vertex2i( right, top-lines );
+    gGL.color4fv(end_color.mV);
+    gGL.vertex2i( right+lines-1, top-1 );
+    gGL.vertex2i( right, top );
 
     gGL.end();
     stop_glerror();
@@ -429,165 +455,237 @@ void gl_draw_scaled_image_with_border(S32 x, S32 y, S32 width, S32 height, LLTex
 
         gGL.color4fv(color.mV);
 
-        const S32 NUM_VERTICES = 9 * 4; // 9 quads
+        const S32 NUM_VERTICES = 9 * 6;
         LLVector2 uv[NUM_VERTICES];
-        LLVector3 pos[NUM_VERTICES];
+        LLVector4a pos[NUM_VERTICES];
 
         S32 index = 0;
 
-        gGL.begin(LLRender::QUADS);
+        gGL.begin(LLRender::TRIANGLES);
         {
             // draw bottom left
             uv[index] = LLVector2(uv_outer_rect.mLeft, uv_outer_rect.mBottom);
-            pos[index] = LLVector3(draw_outer_rect.mLeft, draw_outer_rect.mBottom, 0.f);
+            pos[index] = LLVector4a(draw_outer_rect.mLeft, draw_outer_rect.mBottom, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_center_rect.mLeft, uv_outer_rect.mBottom);
-            pos[index] = LLVector3(draw_center_rect.mLeft, draw_outer_rect.mBottom, 0.f);
+            pos[index] = LLVector4a(draw_center_rect.mLeft, draw_outer_rect.mBottom, 0.f);
+            index++;
+
+            uv[index] = LLVector2(uv_center_rect.mLeft, uv_center_rect.mBottom);
+            pos[index] = LLVector4a(draw_center_rect.mLeft, draw_center_rect.mBottom, 0.f);
+            index++;
+
+            uv[index] = LLVector2(uv_outer_rect.mLeft, uv_outer_rect.mBottom);
+            pos[index] = LLVector4a(draw_outer_rect.mLeft, draw_outer_rect.mBottom, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_center_rect.mLeft, uv_center_rect.mBottom);
-            pos[index] = LLVector3(draw_center_rect.mLeft, draw_center_rect.mBottom, 0.f);
+            pos[index] = LLVector4a(draw_center_rect.mLeft, draw_center_rect.mBottom, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_outer_rect.mLeft, uv_center_rect.mBottom);
-            pos[index] = LLVector3(draw_outer_rect.mLeft, draw_center_rect.mBottom, 0.f);
+            pos[index] = LLVector4a(draw_outer_rect.mLeft, draw_center_rect.mBottom, 0.f);
             index++;
 
             // draw bottom middle
             uv[index] = LLVector2(uv_center_rect.mLeft, uv_outer_rect.mBottom);
-            pos[index] = LLVector3(draw_center_rect.mLeft, draw_outer_rect.mBottom, 0.f);
+            pos[index] = LLVector4a(draw_center_rect.mLeft, draw_outer_rect.mBottom, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_center_rect.mRight, uv_outer_rect.mBottom);
-            pos[index] = LLVector3(draw_center_rect.mRight, draw_outer_rect.mBottom, 0.f);
+            pos[index] = LLVector4a(draw_center_rect.mRight, draw_outer_rect.mBottom, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_center_rect.mRight, uv_center_rect.mBottom);
-            pos[index] = LLVector3(draw_center_rect.mRight, draw_center_rect.mBottom, 0.f);
+            pos[index] = LLVector4a(draw_center_rect.mRight, draw_center_rect.mBottom, 0.f);
+            index++;
+
+            uv[index] = LLVector2(uv_center_rect.mLeft, uv_outer_rect.mBottom);
+            pos[index] = LLVector4a(draw_center_rect.mLeft, draw_outer_rect.mBottom, 0.f);
+            index++;
+
+            uv[index] = LLVector2(uv_center_rect.mRight, uv_center_rect.mBottom);
+            pos[index] = LLVector4a(draw_center_rect.mRight, draw_center_rect.mBottom, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_center_rect.mLeft, uv_center_rect.mBottom);
-            pos[index] = LLVector3(draw_center_rect.mLeft, draw_center_rect.mBottom, 0.f);
+            pos[index] = LLVector4a(draw_center_rect.mLeft, draw_center_rect.mBottom, 0.f);
             index++;
 
             // draw bottom right
             uv[index] = LLVector2(uv_center_rect.mRight, uv_outer_rect.mBottom);
-            pos[index] = LLVector3(draw_center_rect.mRight, draw_outer_rect.mBottom, 0.f);
+            pos[index] = LLVector4a(draw_center_rect.mRight, draw_outer_rect.mBottom, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_outer_rect.mRight, uv_outer_rect.mBottom);
-            pos[index] = LLVector3(draw_outer_rect.mRight, draw_outer_rect.mBottom, 0.f);
+            pos[index] = LLVector4a(draw_outer_rect.mRight, draw_outer_rect.mBottom, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_outer_rect.mRight, uv_center_rect.mBottom);
-            pos[index] = LLVector3(draw_outer_rect.mRight, draw_center_rect.mBottom, 0.f);
+            pos[index] = LLVector4a(draw_outer_rect.mRight, draw_center_rect.mBottom, 0.f);
+            index++;
+
+            uv[index] = LLVector2(uv_center_rect.mRight, uv_outer_rect.mBottom);
+            pos[index] = LLVector4a(draw_center_rect.mRight, draw_outer_rect.mBottom, 0.f);
+            index++;
+
+            uv[index] = LLVector2(uv_outer_rect.mRight, uv_center_rect.mBottom);
+            pos[index] = LLVector4a(draw_outer_rect.mRight, draw_center_rect.mBottom, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_center_rect.mRight, uv_center_rect.mBottom);
-            pos[index] = LLVector3(draw_center_rect.mRight, draw_center_rect.mBottom, 0.f);
+            pos[index] = LLVector4a(draw_center_rect.mRight, draw_center_rect.mBottom, 0.f);
             index++;
 
             // draw left
             uv[index] = LLVector2(uv_outer_rect.mLeft, uv_center_rect.mBottom);
-            pos[index] = LLVector3(draw_outer_rect.mLeft, draw_center_rect.mBottom, 0.f);
+            pos[index] = LLVector4a(draw_outer_rect.mLeft, draw_center_rect.mBottom, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_center_rect.mLeft, uv_center_rect.mBottom);
-            pos[index] = LLVector3(draw_center_rect.mLeft, draw_center_rect.mBottom, 0.f);
+            pos[index] = LLVector4a(draw_center_rect.mLeft, draw_center_rect.mBottom, 0.f);
+            index++;
+
+            uv[index] = LLVector2(uv_center_rect.mLeft, uv_center_rect.mTop);
+            pos[index] = LLVector4a(draw_center_rect.mLeft, draw_center_rect.mTop, 0.f);
+            index++;
+
+            uv[index] = LLVector2(uv_outer_rect.mLeft, uv_center_rect.mBottom);
+            pos[index] = LLVector4a(draw_outer_rect.mLeft, draw_center_rect.mBottom, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_center_rect.mLeft, uv_center_rect.mTop);
-            pos[index] = LLVector3(draw_center_rect.mLeft, draw_center_rect.mTop, 0.f);
+            pos[index] = LLVector4a(draw_center_rect.mLeft, draw_center_rect.mTop, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_outer_rect.mLeft, uv_center_rect.mTop);
-            pos[index] = LLVector3(draw_outer_rect.mLeft, draw_center_rect.mTop, 0.f);
+            pos[index] = LLVector4a(draw_outer_rect.mLeft, draw_center_rect.mTop, 0.f);
             index++;
 
             // draw middle
             uv[index] = LLVector2(uv_center_rect.mLeft, uv_center_rect.mBottom);
-            pos[index] = LLVector3(draw_center_rect.mLeft, draw_center_rect.mBottom, 0.f);
+            pos[index] = LLVector4a(draw_center_rect.mLeft, draw_center_rect.mBottom, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_center_rect.mRight, uv_center_rect.mBottom);
-            pos[index] = LLVector3(draw_center_rect.mRight, draw_center_rect.mBottom, 0.f);
+            pos[index] = LLVector4a(draw_center_rect.mRight, draw_center_rect.mBottom, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_center_rect.mRight, uv_center_rect.mTop);
-            pos[index] = LLVector3(draw_center_rect.mRight, draw_center_rect.mTop, 0.f);
+            pos[index] = LLVector4a(draw_center_rect.mRight, draw_center_rect.mTop, 0.f);
+            index++;
+
+            uv[index] = LLVector2(uv_center_rect.mLeft, uv_center_rect.mBottom);
+            pos[index] = LLVector4a(draw_center_rect.mLeft, draw_center_rect.mBottom, 0.f);
+            index++;
+
+            uv[index] = LLVector2(uv_center_rect.mRight, uv_center_rect.mTop);
+            pos[index] = LLVector4a(draw_center_rect.mRight, draw_center_rect.mTop, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_center_rect.mLeft, uv_center_rect.mTop);
-            pos[index] = LLVector3(draw_center_rect.mLeft, draw_center_rect.mTop, 0.f);
+            pos[index] = LLVector4a(draw_center_rect.mLeft, draw_center_rect.mTop, 0.f);
             index++;
 
             // draw right
             uv[index] = LLVector2(uv_center_rect.mRight, uv_center_rect.mBottom);
-            pos[index] = LLVector3(draw_center_rect.mRight, draw_center_rect.mBottom, 0.f);
+            pos[index] = LLVector4a(draw_center_rect.mRight, draw_center_rect.mBottom, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_outer_rect.mRight, uv_center_rect.mBottom);
-            pos[index] = LLVector3(draw_outer_rect.mRight, draw_center_rect.mBottom, 0.f);
+            pos[index] = LLVector4a(draw_outer_rect.mRight, draw_center_rect.mBottom, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_outer_rect.mRight, uv_center_rect.mTop);
-            pos[index] = LLVector3(draw_outer_rect.mRight, draw_center_rect.mTop, 0.f);
+            pos[index] = LLVector4a(draw_outer_rect.mRight, draw_center_rect.mTop, 0.f);
+            index++;
+
+            uv[index] = LLVector2(uv_center_rect.mRight, uv_center_rect.mBottom);
+            pos[index] = LLVector4a(draw_center_rect.mRight, draw_center_rect.mBottom, 0.f);
+            index++;
+
+            uv[index] = LLVector2(uv_outer_rect.mRight, uv_center_rect.mTop);
+            pos[index] = LLVector4a(draw_outer_rect.mRight, draw_center_rect.mTop, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_center_rect.mRight, uv_center_rect.mTop);
-            pos[index] = LLVector3(draw_center_rect.mRight, draw_center_rect.mTop, 0.f);
+            pos[index] = LLVector4a(draw_center_rect.mRight, draw_center_rect.mTop, 0.f);
             index++;
 
             // draw top left
             uv[index] = LLVector2(uv_outer_rect.mLeft, uv_center_rect.mTop);
-            pos[index] = LLVector3(draw_outer_rect.mLeft, draw_center_rect.mTop, 0.f);
+            pos[index] = LLVector4a(draw_outer_rect.mLeft, draw_center_rect.mTop, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_center_rect.mLeft, uv_center_rect.mTop);
-            pos[index] = LLVector3(draw_center_rect.mLeft, draw_center_rect.mTop, 0.f);
+            pos[index] = LLVector4a(draw_center_rect.mLeft, draw_center_rect.mTop, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_center_rect.mLeft, uv_outer_rect.mTop);
-            pos[index] = LLVector3(draw_center_rect.mLeft, draw_outer_rect.mTop, 0.f);
+            pos[index] = LLVector4a(draw_center_rect.mLeft, draw_outer_rect.mTop, 0.f);
+            index++;
+
+            uv[index] = LLVector2(uv_outer_rect.mLeft, uv_center_rect.mTop);
+            pos[index] = LLVector4a(draw_outer_rect.mLeft, draw_center_rect.mTop, 0.f);
+            index++;
+
+            uv[index] = LLVector2(uv_center_rect.mLeft, uv_outer_rect.mTop);
+            pos[index] = LLVector4a(draw_center_rect.mLeft, draw_outer_rect.mTop, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_outer_rect.mLeft, uv_outer_rect.mTop);
-            pos[index] = LLVector3(draw_outer_rect.mLeft, draw_outer_rect.mTop, 0.f);
+            pos[index] = LLVector4a(draw_outer_rect.mLeft, draw_outer_rect.mTop, 0.f);
             index++;
 
             // draw top middle
             uv[index] = LLVector2(uv_center_rect.mLeft, uv_center_rect.mTop);
-            pos[index] = LLVector3(draw_center_rect.mLeft, draw_center_rect.mTop, 0.f);
+            pos[index] = LLVector4a(draw_center_rect.mLeft, draw_center_rect.mTop, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_center_rect.mRight, uv_center_rect.mTop);
-            pos[index] = LLVector3(draw_center_rect.mRight, draw_center_rect.mTop, 0.f);
+            pos[index] = LLVector4a(draw_center_rect.mRight, draw_center_rect.mTop, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_center_rect.mRight, uv_outer_rect.mTop);
-            pos[index] = LLVector3(draw_center_rect.mRight, draw_outer_rect.mTop, 0.f);
+            pos[index] = LLVector4a(draw_center_rect.mRight, draw_outer_rect.mTop, 0.f);
+            index++;
+
+            uv[index] = LLVector2(uv_center_rect.mLeft, uv_center_rect.mTop);
+            pos[index] = LLVector4a(draw_center_rect.mLeft, draw_center_rect.mTop, 0.f);
+            index++;
+
+            uv[index] = LLVector2(uv_center_rect.mRight, uv_outer_rect.mTop);
+            pos[index] = LLVector4a(draw_center_rect.mRight, draw_outer_rect.mTop, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_center_rect.mLeft, uv_outer_rect.mTop);
-            pos[index] = LLVector3(draw_center_rect.mLeft, draw_outer_rect.mTop, 0.f);
+            pos[index] = LLVector4a(draw_center_rect.mLeft, draw_outer_rect.mTop, 0.f);
             index++;
 
             // draw top right
             uv[index] = LLVector2(uv_center_rect.mRight, uv_center_rect.mTop);
-            pos[index] = LLVector3(draw_center_rect.mRight, draw_center_rect.mTop, 0.f);
+            pos[index] = LLVector4a(draw_center_rect.mRight, draw_center_rect.mTop, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_outer_rect.mRight, uv_center_rect.mTop);
-            pos[index] = LLVector3(draw_outer_rect.mRight, draw_center_rect.mTop, 0.f);
+            pos[index] = LLVector4a(draw_outer_rect.mRight, draw_center_rect.mTop, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_outer_rect.mRight, uv_outer_rect.mTop);
-            pos[index] = LLVector3(draw_outer_rect.mRight, draw_outer_rect.mTop, 0.f);
+            pos[index] = LLVector4a(draw_outer_rect.mRight, draw_outer_rect.mTop, 0.f);
+            index++;
+
+            uv[index] = LLVector2(uv_center_rect.mRight, uv_center_rect.mTop);
+            pos[index] = LLVector4a(draw_center_rect.mRight, draw_center_rect.mTop, 0.f);
+            index++;
+
+            uv[index] = LLVector2(uv_outer_rect.mRight, uv_outer_rect.mTop);
+            pos[index] = LLVector4a(draw_outer_rect.mRight, draw_outer_rect.mTop, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_center_rect.mRight, uv_outer_rect.mTop);
-            pos[index] = LLVector3(draw_center_rect.mRight, draw_outer_rect.mTop, 0.f);
+            pos[index] = LLVector4a(draw_center_rect.mRight, draw_outer_rect.mTop, 0.f);
             index++;
 
             gGL.vertexBatchPreTransformed(pos, uv, NUM_VERTICES);
@@ -629,11 +727,11 @@ void gl_draw_scaled_rotated_image(S32 x, S32 y, S32 width, S32 height, F32 degre
 
     if (degrees == 0.f)
     {
-        const S32 NUM_VERTICES = 4; // 9 quads
+        const S32 NUM_VERTICES = 6;
         LLVector2 uv[NUM_VERTICES];
-        LLVector3 pos[NUM_VERTICES];
+        LLVector4a pos[NUM_VERTICES];
 
-        gGL.begin(LLRender::QUADS);
+        gGL.begin(LLRender::TRIANGLES);
         {
             LLVector3 ui_scale = gGL.getUIScale();
             LLVector3 ui_translation = gGL.getUITranslation();
@@ -645,19 +743,27 @@ void gl_draw_scaled_rotated_image(S32 x, S32 y, S32 width, S32 height, F32 degre
             S32 scaled_height = ll_round(height * ui_scale.mV[VY]);
 
             uv[index] = LLVector2(uv_rect.mRight, uv_rect.mTop);
-            pos[index] = LLVector3(ui_translation.mV[VX] + scaled_width, ui_translation.mV[VY] + scaled_height, 0.f);
+            pos[index] = LLVector4a(ui_translation.mV[VX] + scaled_width, ui_translation.mV[VY] + scaled_height, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_rect.mLeft, uv_rect.mTop);
-            pos[index] = LLVector3(ui_translation.mV[VX], ui_translation.mV[VY] + scaled_height, 0.f);
+            pos[index] = LLVector4a(ui_translation.mV[VX], ui_translation.mV[VY] + scaled_height, 0.f);
+            index++;
+
+            uv[index] = LLVector2(uv_rect.mLeft, uv_rect.mBottom);
+            pos[index] = LLVector4a(ui_translation.mV[VX], ui_translation.mV[VY], 0.f);
+            index++;
+
+            uv[index] = LLVector2(uv_rect.mRight, uv_rect.mTop);
+            pos[index] = LLVector4a(ui_translation.mV[VX] + scaled_width, ui_translation.mV[VY] + scaled_height, 0.f);
             index++;
 
             uv[index] = LLVector2(uv_rect.mLeft, uv_rect.mBottom);
-            pos[index] = LLVector3(ui_translation.mV[VX], ui_translation.mV[VY], 0.f);
+            pos[index] = LLVector4a(ui_translation.mV[VX], ui_translation.mV[VY], 0.f);
             index++;
 
             uv[index] = LLVector2(uv_rect.mRight, uv_rect.mBottom);
-            pos[index] = LLVector3(ui_translation.mV[VX] + scaled_width, ui_translation.mV[VY], 0.f);
+            pos[index] = LLVector4a(ui_translation.mV[VX] + scaled_width, ui_translation.mV[VY], 0.f);
             index++;
 
             gGL.vertexBatchPreTransformed(pos, uv, NUM_VERTICES);
@@ -687,26 +793,32 @@ void gl_draw_scaled_rotated_image(S32 x, S32 y, S32 width, S32 height, F32 degre
 
         gGL.color4fv(color.mV);
 
-        gGL.begin(LLRender::QUADS);
+        gGL.begin(LLRender::TRIANGLES);
         {
-            LLVector3 v;
+            LLVector3 v1 = LLVector3(offset_x, offset_y, 0.f) * quat;
+            LLVector3 v2 = LLVector3(-offset_x, offset_y, 0.f) * quat;
+            LLVector3 v3 = LLVector3(-offset_x, -offset_y, 0.f) * quat;
+            LLVector3 v4 = LLVector3(offset_x, -offset_y, 0.f) * quat;
 
-            v = LLVector3(offset_x, offset_y, 0.f) * quat;
             gGL.texCoord2f(uv_rect.mRight, uv_rect.mTop);
-            gGL.vertex2f(v.mV[0], v.mV[1] );
+            gGL.vertex2f(v1.mV[0], v1.mV[1] );
 
-            v = LLVector3(-offset_x, offset_y, 0.f) * quat;
             gGL.texCoord2f(uv_rect.mLeft, uv_rect.mTop);
-            gGL.vertex2f(v.mV[0], v.mV[1] );
+            gGL.vertex2f(v2.mV[0], v2.mV[1] );
+
+            gGL.texCoord2f(uv_rect.mLeft, uv_rect.mBottom);
+            gGL.vertex2f(v3.mV[0], v3.mV[1] );
+
+            gGL.texCoord2f(uv_rect.mRight, uv_rect.mTop);
+            gGL.vertex2f(v1.mV[0], v1.mV[1] );
 
-            v = LLVector3(-offset_x, -offset_y, 0.f) * quat;
             gGL.texCoord2f(uv_rect.mLeft, uv_rect.mBottom);
-            gGL.vertex2f(v.mV[0], v.mV[1] );
+            gGL.vertex2f(v3.mV[0], v3.mV[1] );
 
-            v = LLVector3(offset_x, -offset_y, 0.f) * quat;
             gGL.texCoord2f(uv_rect.mRight, uv_rect.mBottom);
-            gGL.vertex2f(v.mV[0], v.mV[1] );
+            gGL.vertex2f(v4.mV[0], v4.mV[1] );
         }
+
         gGL.end();
         gGL.popUIMatrix();
     }
@@ -948,8 +1060,8 @@ void gl_washer_segment_2d(F32 outer_radius, F32 inner_radius, F32 start_radians,
 
 void gl_rect_2d_simple_tex( S32 width, S32 height )
 {
-    gGL.begin( LLRender::QUADS );
-
+    gGL.begin( LLRender::TRIANGLES );
+    {
         gGL.texCoord2f(1.f, 1.f);
         gGL.vertex2i(width, height);
 
@@ -959,19 +1071,30 @@ void gl_rect_2d_simple_tex( S32 width, S32 height )
         gGL.texCoord2f(0.f, 0.f);
         gGL.vertex2i(0, 0);
 
+        gGL.texCoord2f(1.f, 1.f);
+        gGL.vertex2i(width, height);
+
+        gGL.texCoord2f(0.f, 0.f);
+        gGL.vertex2i(0, 0);
+
         gGL.texCoord2f(1.f, 0.f);
         gGL.vertex2i(width, 0);
-
+    }
     gGL.end();
 }
 
 void gl_rect_2d_simple( S32 width, S32 height )
 {
-    gGL.begin( LLRender::QUADS );
+    gGL.begin( LLRender::TRIANGLES );
+    {
         gGL.vertex2i(width, height);
         gGL.vertex2i(0, height);
         gGL.vertex2i(0, 0);
+
+        gGL.vertex2i(width, height);
+        gGL.vertex2i(0, 0);
         gGL.vertex2i(width, 0);
+    }
     gGL.end();
 }
 
@@ -1011,7 +1134,7 @@ void gl_segmented_rect_2d_tex(const S32 left,
     LLVector2 width_vec((F32)width, 0.f);
     LLVector2 height_vec(0.f, (F32)height);
 
-    gGL.begin(LLRender::QUADS);
+    gGL.begin(LLRender::TRIANGLES);
     {
         // draw bottom left
         gGL.texCoord2f(0.f, 0.f);
@@ -1023,6 +1146,12 @@ void gl_segmented_rect_2d_tex(const S32 left,
         gGL.texCoord2f(border_uv_scale.mV[VX], border_uv_scale.mV[VY]);
         gGL.vertex2fv((border_width_left + border_height_bottom).mV);
 
+        gGL.texCoord2f(0.f, 0.f);
+        gGL.vertex2f(0.f, 0.f);
+
+        gGL.texCoord2f(border_uv_scale.mV[VX], border_uv_scale.mV[VY]);
+        gGL.vertex2fv((border_width_left + border_height_bottom).mV);
+
         gGL.texCoord2f(0.f, border_uv_scale.mV[VY]);
         gGL.vertex2fv(border_height_bottom.mV);
 
@@ -1036,6 +1165,12 @@ void gl_segmented_rect_2d_tex(const S32 left,
         gGL.texCoord2f(1.f - border_uv_scale.mV[VX], border_uv_scale.mV[VY]);
         gGL.vertex2fv((width_vec - border_width_right + border_height_bottom).mV);
 
+        gGL.texCoord2f(border_uv_scale.mV[VX], 0.f);
+        gGL.vertex2fv(border_width_left.mV);
+
+        gGL.texCoord2f(1.f - border_uv_scale.mV[VX], border_uv_scale.mV[VY]);
+        gGL.vertex2fv((width_vec - border_width_right + border_height_bottom).mV);
+
         gGL.texCoord2f(border_uv_scale.mV[VX], border_uv_scale.mV[VY]);
         gGL.vertex2fv((border_width_left + border_height_bottom).mV);
 
@@ -1049,6 +1184,12 @@ void gl_segmented_rect_2d_tex(const S32 left,
         gGL.texCoord2f(1.f, border_uv_scale.mV[VY]);
         gGL.vertex2fv((width_vec + border_height_bottom).mV);
 
+        gGL.texCoord2f(1.f - border_uv_scale.mV[VX], 0.f);
+        gGL.vertex2fv((width_vec - border_width_right).mV);
+
+        gGL.texCoord2f(1.f, border_uv_scale.mV[VY]);
+        gGL.vertex2fv((width_vec + border_height_bottom).mV);
+
         gGL.texCoord2f(1.f - border_uv_scale.mV[VX], border_uv_scale.mV[VY]);
         gGL.vertex2fv((width_vec - border_width_right + border_height_bottom).mV);
 
@@ -1062,6 +1203,12 @@ void gl_segmented_rect_2d_tex(const S32 left,
         gGL.texCoord2f(border_uv_scale.mV[VX], 1.f - border_uv_scale.mV[VY]);
         gGL.vertex2fv((border_width_left + height_vec - border_height_top).mV);
 
+        gGL.texCoord2f(0.f, border_uv_scale.mV[VY]);
+        gGL.vertex2fv(border_height_bottom.mV);
+
+        gGL.texCoord2f(border_uv_scale.mV[VX], 1.f - border_uv_scale.mV[VY]);
+        gGL.vertex2fv((border_width_left + height_vec - border_height_top).mV);
+
         gGL.texCoord2f(0.f, 1.f - border_uv_scale.mV[VY]);
         gGL.vertex2fv((height_vec - border_height_top).mV);
 
@@ -1075,6 +1222,12 @@ void gl_segmented_rect_2d_tex(const S32 left,
         gGL.texCoord2f(1.f - border_uv_scale.mV[VX], 1.f - border_uv_scale.mV[VY]);
         gGL.vertex2fv((width_vec - border_width_right + height_vec - border_height_top).mV);
 
+        gGL.texCoord2f(border_uv_scale.mV[VX], border_uv_scale.mV[VY]);
+        gGL.vertex2fv((border_width_left + border_height_bottom).mV);
+
+        gGL.texCoord2f(1.f - border_uv_scale.mV[VX], 1.f - border_uv_scale.mV[VY]);
+        gGL.vertex2fv((width_vec - border_width_right + height_vec - border_height_top).mV);
+
         gGL.texCoord2f(border_uv_scale.mV[VX], 1.f - border_uv_scale.mV[VY]);
         gGL.vertex2fv((border_width_left + height_vec - border_height_top).mV);
 
@@ -1088,6 +1241,12 @@ void gl_segmented_rect_2d_tex(const S32 left,
         gGL.texCoord2f(1.f, 1.f - border_uv_scale.mV[VY]);
         gGL.vertex2fv((width_vec + height_vec - border_height_top).mV);
 
+        gGL.texCoord2f(1.f - border_uv_scale.mV[VX], border_uv_scale.mV[VY]);
+        gGL.vertex2fv((width_vec - border_width_right + border_height_bottom).mV);
+
+        gGL.texCoord2f(1.f, 1.f - border_uv_scale.mV[VY]);
+        gGL.vertex2fv((width_vec + height_vec - border_height_top).mV);
+
         gGL.texCoord2f(1.f - border_uv_scale.mV[VX], 1.f - border_uv_scale.mV[VY]);
         gGL.vertex2fv((width_vec - border_width_right + height_vec - border_height_top).mV);
 
@@ -1101,6 +1260,12 @@ void gl_segmented_rect_2d_tex(const S32 left,
         gGL.texCoord2f(border_uv_scale.mV[VX], 1.f);
         gGL.vertex2fv((border_width_left + height_vec).mV);
 
+        gGL.texCoord2f(0.f, 1.f - border_uv_scale.mV[VY]);
+        gGL.vertex2fv((height_vec - border_height_top).mV);
+
+        gGL.texCoord2f(border_uv_scale.mV[VX], 1.f);
+        gGL.vertex2fv((border_width_left + height_vec).mV);
+
         gGL.texCoord2f(0.f, 1.f);
         gGL.vertex2fv((height_vec).mV);
 
@@ -1114,6 +1279,12 @@ void gl_segmented_rect_2d_tex(const S32 left,
         gGL.texCoord2f(1.f - border_uv_scale.mV[VX], 1.f);
         gGL.vertex2fv((width_vec - border_width_right + height_vec).mV);
 
+        gGL.texCoord2f(border_uv_scale.mV[VX], 1.f - border_uv_scale.mV[VY]);
+        gGL.vertex2fv((border_width_left + height_vec - border_height_top).mV);
+
+        gGL.texCoord2f(1.f - border_uv_scale.mV[VX], 1.f);
+        gGL.vertex2fv((width_vec - border_width_right + height_vec).mV);
+
         gGL.texCoord2f(border_uv_scale.mV[VX], 1.f);
         gGL.vertex2fv((border_width_left + height_vec).mV);
 
@@ -1127,6 +1298,12 @@ void gl_segmented_rect_2d_tex(const S32 left,
         gGL.texCoord2f(1.f, 1.f);
         gGL.vertex2fv((width_vec + height_vec).mV);
 
+        gGL.texCoord2f(1.f - border_uv_scale.mV[VX], 1.f - border_uv_scale.mV[VY]);
+        gGL.vertex2fv((width_vec - border_width_right + height_vec - border_height_top).mV);
+
+        gGL.texCoord2f(1.f, 1.f);
+        gGL.vertex2fv((width_vec + height_vec).mV);
+
         gGL.texCoord2f(1.f - border_uv_scale.mV[VX], 1.f);
         gGL.vertex2fv((width_vec - border_width_right + height_vec).mV);
     }
@@ -1181,7 +1358,7 @@ void gl_segmented_rect_2d_fragment_tex(const LLRect& rect,
     LLVector2 x_min;
     LLVector2 x_max;
 
-    gGL.begin(LLRender::QUADS);
+    gGL.begin(LLRender::TRIANGLES);
     {
         if (start_fragment < middle_start)
         {
@@ -1200,6 +1377,12 @@ void gl_segmented_rect_2d_fragment_tex(const LLRect& rect,
             gGL.texCoord2f(u_max, border_uv_scale.mV[VY]);
             gGL.vertex2fv((x_max + border_height_bottom).mV);
 
+            gGL.texCoord2f(u_min, 0.f);
+            gGL.vertex2fv(x_min.mV);
+
+            gGL.texCoord2f(u_max, border_uv_scale.mV[VY]);
+            gGL.vertex2fv((x_max + border_height_bottom).mV);
+
             gGL.texCoord2f(u_min, border_uv_scale.mV[VY]);
             gGL.vertex2fv((x_min + border_height_bottom).mV);
 
@@ -1213,6 +1396,12 @@ void gl_segmented_rect_2d_fragment_tex(const LLRect& rect,
             gGL.texCoord2f(u_max, 1.f - border_uv_scale.mV[VY]);
             gGL.vertex2fv((x_max + height_vec - border_height_top).mV);
 
+            gGL.texCoord2f(u_min, border_uv_scale.mV[VY]);
+            gGL.vertex2fv((x_min + border_height_bottom).mV);
+
+            gGL.texCoord2f(u_max, 1.f - border_uv_scale.mV[VY]);
+            gGL.vertex2fv((x_max + height_vec - border_height_top).mV);
+
             gGL.texCoord2f(u_min, 1.f - border_uv_scale.mV[VY]);
             gGL.vertex2fv((x_min + height_vec - border_height_top).mV);
 
@@ -1226,6 +1415,12 @@ void gl_segmented_rect_2d_fragment_tex(const LLRect& rect,
             gGL.texCoord2f(u_max, 1.f);
             gGL.vertex2fv((x_max + height_vec).mV);
 
+            gGL.texCoord2f(u_min, 1.f - border_uv_scale.mV[VY]);
+            gGL.vertex2fv((x_min + height_vec - border_height_top).mV);
+
+            gGL.texCoord2f(u_max, 1.f);
+            gGL.vertex2fv((x_max + height_vec).mV);
+
             gGL.texCoord2f(u_min, 1.f);
             gGL.vertex2fv((x_min + height_vec).mV);
         }
@@ -1245,6 +1440,12 @@ void gl_segmented_rect_2d_fragment_tex(const LLRect& rect,
             gGL.texCoord2f(1.f - border_uv_scale.mV[VX], border_uv_scale.mV[VY]);
             gGL.vertex2fv((x_max + border_height_bottom).mV);
 
+            gGL.texCoord2f(border_uv_scale.mV[VX], 0.f);
+            gGL.vertex2fv(x_min.mV);
+
+            gGL.texCoord2f(1.f - border_uv_scale.mV[VX], border_uv_scale.mV[VY]);
+            gGL.vertex2fv((x_max + border_height_bottom).mV);
+
             gGL.texCoord2f(border_uv_scale.mV[VX], border_uv_scale.mV[VY]);
             gGL.vertex2fv((x_min + border_height_bottom).mV);
 
@@ -1258,6 +1459,12 @@ void gl_segmented_rect_2d_fragment_tex(const LLRect& rect,
             gGL.texCoord2f(1.f - border_uv_scale.mV[VX], 1.f - border_uv_scale.mV[VY]);
             gGL.vertex2fv((x_max + height_vec - border_height_top).mV);
 
+            gGL.texCoord2f(border_uv_scale.mV[VX], border_uv_scale.mV[VY]);
+            gGL.vertex2fv((x_min + border_height_bottom).mV);
+
+            gGL.texCoord2f(1.f - border_uv_scale.mV[VX], 1.f - border_uv_scale.mV[VY]);
+            gGL.vertex2fv((x_max + height_vec - border_height_top).mV);
+
             gGL.texCoord2f(border_uv_scale.mV[VX], 1.f - border_uv_scale.mV[VY]);
             gGL.vertex2fv((x_min + height_vec - border_height_top).mV);
 
@@ -1271,6 +1478,12 @@ void gl_segmented_rect_2d_fragment_tex(const LLRect& rect,
             gGL.texCoord2f(1.f - border_uv_scale.mV[VX], 1.f);
             gGL.vertex2fv((x_max + height_vec).mV);
 
+            gGL.texCoord2f(border_uv_scale.mV[VX], 1.f - border_uv_scale.mV[VY]);
+            gGL.vertex2fv((x_min + height_vec - border_height_top).mV);
+
+            gGL.texCoord2f(1.f - border_uv_scale.mV[VX], 1.f);
+            gGL.vertex2fv((x_max + height_vec).mV);
+
             gGL.texCoord2f(border_uv_scale.mV[VX], 1.f);
             gGL.vertex2fv((x_min + height_vec).mV);
         }
@@ -1292,6 +1505,12 @@ void gl_segmented_rect_2d_fragment_tex(const LLRect& rect,
             gGL.texCoord2f(u_max, border_uv_scale.mV[VY]);
             gGL.vertex2fv((x_max + border_height_bottom).mV);
 
+            gGL.texCoord2f(u_min, 0.f);
+            gGL.vertex2fv((x_min).mV);
+
+            gGL.texCoord2f(u_max, border_uv_scale.mV[VY]);
+            gGL.vertex2fv((x_max + border_height_bottom).mV);
+
             gGL.texCoord2f(u_min, border_uv_scale.mV[VY]);
             gGL.vertex2fv((x_min + border_height_bottom).mV);
 
@@ -1305,6 +1524,12 @@ void gl_segmented_rect_2d_fragment_tex(const LLRect& rect,
             gGL.texCoord2f(u_max, 1.f - border_uv_scale.mV[VY]);
             gGL.vertex2fv((x_max + height_vec - border_height_top).mV);
 
+            gGL.texCoord2f(u_min, border_uv_scale.mV[VY]);
+            gGL.vertex2fv((x_min + border_height_bottom).mV);
+
+            gGL.texCoord2f(u_max, 1.f - border_uv_scale.mV[VY]);
+            gGL.vertex2fv((x_max + height_vec - border_height_top).mV);
+
             gGL.texCoord2f(u_min, 1.f - border_uv_scale.mV[VY]);
             gGL.vertex2fv((x_min + height_vec - border_height_top).mV);
 
@@ -1318,6 +1543,12 @@ void gl_segmented_rect_2d_fragment_tex(const LLRect& rect,
             gGL.texCoord2f(u_max, 1.f);
             gGL.vertex2fv((x_max + height_vec).mV);
 
+            gGL.texCoord2f(u_min, 1.f - border_uv_scale.mV[VY]);
+            gGL.vertex2fv((x_min + height_vec - border_height_top).mV);
+
+            gGL.texCoord2f(u_max, 1.f);
+            gGL.vertex2fv((x_max + height_vec).mV);
+
             gGL.texCoord2f(u_min, 1.f);
             gGL.vertex2fv((x_min + height_vec).mV);
         }
@@ -1332,7 +1563,7 @@ void gl_segmented_rect_3d_tex(const LLRectf& clip_rect, const LLRectf& center_uv
 {
     LL_PROFILE_ZONE_SCOPED_CATEGORY_UI;
 
-    gGL.begin(LLRender::QUADS);
+   gGL.begin(LLRender::TRIANGLES);
     {
         // draw bottom left
         gGL.texCoord2f(clip_rect.mLeft, clip_rect.mBottom);
@@ -1344,6 +1575,12 @@ void gl_segmented_rect_3d_tex(const LLRectf& clip_rect, const LLRectf& center_uv
         gGL.texCoord2f(center_uv_rect.mLeft, center_uv_rect.mBottom);
         gGL.vertex3fv((center_draw_rect.mLeft * width_vec + center_draw_rect.mBottom * height_vec).mV);
 
+        gGL.texCoord2f(clip_rect.mLeft, clip_rect.mBottom);
+        gGL.vertex3f(0.f, 0.f, 0.f);
+
+        gGL.texCoord2f(center_uv_rect.mLeft, center_uv_rect.mBottom);
+        gGL.vertex3fv((center_draw_rect.mLeft * width_vec + center_draw_rect.mBottom * height_vec).mV);
+
         gGL.texCoord2f(clip_rect.mLeft, center_uv_rect.mBottom);
         gGL.vertex3fv((center_draw_rect.mBottom * height_vec).mV);
 
@@ -1357,6 +1594,12 @@ void gl_segmented_rect_3d_tex(const LLRectf& clip_rect, const LLRectf& center_uv
         gGL.texCoord2f(center_uv_rect.mRight, center_uv_rect.mBottom);
         gGL.vertex3fv((center_draw_rect.mRight * width_vec + center_draw_rect.mBottom * height_vec).mV);
 
+        gGL.texCoord2f(center_uv_rect.mLeft, clip_rect.mBottom);
+        gGL.vertex3fv((center_draw_rect.mLeft * width_vec).mV);
+
+        gGL.texCoord2f(center_uv_rect.mRight, center_uv_rect.mBottom);
+        gGL.vertex3fv((center_draw_rect.mRight * width_vec + center_draw_rect.mBottom * height_vec).mV);
+
         gGL.texCoord2f(center_uv_rect.mLeft, center_uv_rect.mBottom);
         gGL.vertex3fv((center_draw_rect.mLeft * width_vec + center_draw_rect.mBottom * height_vec).mV);
 
@@ -1370,6 +1613,12 @@ void gl_segmented_rect_3d_tex(const LLRectf& clip_rect, const LLRectf& center_uv
         gGL.texCoord2f(clip_rect.mRight, center_uv_rect.mBottom);
         gGL.vertex3fv((width_vec + center_draw_rect.mBottom * height_vec).mV);
 
+        gGL.texCoord2f(center_uv_rect.mRight, clip_rect.mBottom);
+        gGL.vertex3fv((center_draw_rect.mRight * width_vec).mV);
+
+        gGL.texCoord2f(clip_rect.mRight, center_uv_rect.mBottom);
+        gGL.vertex3fv((width_vec + center_draw_rect.mBottom * height_vec).mV);
+
         gGL.texCoord2f(center_uv_rect.mRight, center_uv_rect.mBottom);
         gGL.vertex3fv((center_draw_rect.mRight * width_vec + center_draw_rect.mBottom * height_vec).mV);
 
@@ -1383,6 +1632,12 @@ void gl_segmented_rect_3d_tex(const LLRectf& clip_rect, const LLRectf& center_uv
         gGL.texCoord2f(center_uv_rect.mLeft, center_uv_rect.mTop);
         gGL.vertex3fv((center_draw_rect.mLeft * width_vec + center_draw_rect.mTop * height_vec).mV);
 
+        gGL.texCoord2f(clip_rect.mLeft, center_uv_rect.mBottom);
+        gGL.vertex3fv((center_draw_rect.mBottom * height_vec).mV);
+
+        gGL.texCoord2f(center_uv_rect.mLeft, center_uv_rect.mTop);
+        gGL.vertex3fv((center_draw_rect.mLeft * width_vec + center_draw_rect.mTop * height_vec).mV);
+
         gGL.texCoord2f(clip_rect.mLeft, center_uv_rect.mTop);
         gGL.vertex3fv((center_draw_rect.mTop * height_vec).mV);
 
@@ -1396,6 +1651,12 @@ void gl_segmented_rect_3d_tex(const LLRectf& clip_rect, const LLRectf& center_uv
         gGL.texCoord2f(center_uv_rect.mRight, center_uv_rect.mTop);
         gGL.vertex3fv((center_draw_rect.mRight * width_vec + center_draw_rect.mTop * height_vec).mV);
 
+        gGL.texCoord2f(center_uv_rect.mLeft, center_uv_rect.mBottom);
+        gGL.vertex3fv((center_draw_rect.mLeft * width_vec + center_draw_rect.mBottom * height_vec).mV);
+
+        gGL.texCoord2f(center_uv_rect.mRight, center_uv_rect.mTop);
+        gGL.vertex3fv((center_draw_rect.mRight * width_vec + center_draw_rect.mTop * height_vec).mV);
+
         gGL.texCoord2f(center_uv_rect.mLeft, center_uv_rect.mTop);
         gGL.vertex3fv((center_draw_rect.mLeft * width_vec + center_draw_rect.mTop * height_vec).mV);
 
@@ -1409,6 +1670,12 @@ void gl_segmented_rect_3d_tex(const LLRectf& clip_rect, const LLRectf& center_uv
         gGL.texCoord2f(clip_rect.mRight, center_uv_rect.mTop);
         gGL.vertex3fv((width_vec + center_draw_rect.mTop * height_vec).mV);
 
+        gGL.texCoord2f(center_uv_rect.mRight, center_uv_rect.mBottom);
+        gGL.vertex3fv((center_draw_rect.mRight * width_vec + center_draw_rect.mBottom * height_vec).mV);
+
+        gGL.texCoord2f(clip_rect.mRight, center_uv_rect.mTop);
+        gGL.vertex3fv((width_vec + center_draw_rect.mTop * height_vec).mV);
+
         gGL.texCoord2f(center_uv_rect.mRight, center_uv_rect.mTop);
         gGL.vertex3fv((center_draw_rect.mRight * width_vec + center_draw_rect.mTop * height_vec).mV);
 
@@ -1422,6 +1689,12 @@ void gl_segmented_rect_3d_tex(const LLRectf& clip_rect, const LLRectf& center_uv
         gGL.texCoord2f(center_uv_rect.mLeft, clip_rect.mTop);
         gGL.vertex3fv((center_draw_rect.mLeft * width_vec + height_vec).mV);
 
+        gGL.texCoord2f(clip_rect.mLeft, center_uv_rect.mTop);
+        gGL.vertex3fv((center_draw_rect.mTop * height_vec).mV);
+
+        gGL.texCoord2f(center_uv_rect.mLeft, clip_rect.mTop);
+        gGL.vertex3fv((center_draw_rect.mLeft * width_vec + height_vec).mV);
+
         gGL.texCoord2f(clip_rect.mLeft, clip_rect.mTop);
         gGL.vertex3fv((height_vec).mV);
 
@@ -1435,6 +1708,12 @@ void gl_segmented_rect_3d_tex(const LLRectf& clip_rect, const LLRectf& center_uv
         gGL.texCoord2f(center_uv_rect.mRight, clip_rect.mTop);
         gGL.vertex3fv((center_draw_rect.mRight * width_vec + height_vec).mV);
 
+        gGL.texCoord2f(center_uv_rect.mLeft, center_uv_rect.mTop);
+        gGL.vertex3fv((center_draw_rect.mLeft * width_vec + center_draw_rect.mTop * height_vec).mV);
+
+        gGL.texCoord2f(center_uv_rect.mRight, clip_rect.mTop);
+        gGL.vertex3fv((center_draw_rect.mRight * width_vec + height_vec).mV);
+
         gGL.texCoord2f(center_uv_rect.mLeft, clip_rect.mTop);
         gGL.vertex3fv((center_draw_rect.mLeft * width_vec + height_vec).mV);
 
@@ -1448,6 +1727,12 @@ void gl_segmented_rect_3d_tex(const LLRectf& clip_rect, const LLRectf& center_uv
         gGL.texCoord2f(clip_rect.mRight, clip_rect.mTop);
         gGL.vertex3fv((width_vec + height_vec).mV);
 
+        gGL.texCoord2f(center_uv_rect.mRight, center_uv_rect.mTop);
+        gGL.vertex3fv((center_draw_rect.mRight * width_vec + center_draw_rect.mTop * height_vec).mV);
+
+        gGL.texCoord2f(clip_rect.mRight, clip_rect.mTop);
+        gGL.vertex3fv((width_vec + height_vec).mV);
+
         gGL.texCoord2f(center_uv_rect.mRight, clip_rect.mTop);
         gGL.vertex3fv((center_draw_rect.mRight * width_vec + height_vec).mV);
     }
diff --git a/indra/llrender/llvertexbuffer.cpp b/indra/llrender/llvertexbuffer.cpp
index 7caf20f40b..52fb58187a 100644
--- a/indra/llrender/llvertexbuffer.cpp
+++ b/indra/llrender/llvertexbuffer.cpp
@@ -359,7 +359,8 @@ public:
             mMisses++;
             name = gen_buffer();
             glBindBuffer(type, name);
-            glBufferData(type, size, nullptr, GL_DYNAMIC_DRAW);
+            //glBufferData(type, size, nullptr, GL_DYNAMIC_DRAW);
+            glBufferData(type, size, nullptr, GL_STREAM_DRAW);
             if (type == GL_ELEMENT_ARRAY_BUFFER)
             {
                 LLVertexBuffer::sGLRenderIndices = name;
@@ -1147,20 +1148,87 @@ static void flush_vbo(GLenum target, U32 start, U32 end, void* data)
 {
     if (end != 0)
     {
+        //Note (observeur): I maintained the profile "glBufferSubData" names because i'm not sure if it would impact any statistics part somewhere in the code.
         LL_PROFILE_ZONE_NAMED_CATEGORY_VERTEX("glBufferSubData");
         LL_PROFILE_ZONE_NUM(start);
         LL_PROFILE_ZONE_NUM(end);
         LL_PROFILE_ZONE_NUM(end-start);
 
-        constexpr U32 block_size = 8192;
+        U32 size = end-start+1;
+        U32 block_size = 65536;
+
+        //Note (observeur): The following code is executed on non Apple gpus. Using glMapBufferRange() didn't show obvious benefit on the other tested platforms (intel igpu, amd igpu and nVidia dgpus).
+        if(!gGLManager.mIsApple)
+        {
+            for (U32 i = start; i <= end; i += block_size)
+            {
+                LL_PROFILE_ZONE_NAMED_CATEGORY_VERTEX("glBufferSubData block");
+                LL_PROFILE_GPU_ZONE("glBufferSubData");
+                U32 tend = llmin(i + block_size, end);
+                U32 size = tend - i + 1;
+                glBufferSubData(target, i, size, (U8*) data + (i-start));
+            }
+
+            return;
+        }
+
+        //Note (observeur): glBufferSubData() was causing synchronization stalls on Apple GPUs resulting to heavy stutters and lower performance in the world and UI rendering. Using glMapBufferRange() benefits Macs with Apple gpus enormously.
+
+        //Note (observeur): Other bits such as GL_MAP_INVALIDATE_RANGE_BIT or GL_MAP_UNSYNCHRONIZED_BIT didn't seem to make much of a difference on Apple gpus, so we stick to the simple way.
+        U32 MapBits = GL_MAP_WRITE_BIT;
+
+        //Note (observeur): Using a block size of 0 will call the following block and map the buffer all in once. It doesn't bother Apple machines, it might actually benefit them a little bit. A larger value is also fine. The largest buffers I observed where around 2mb or 3mb while most of buffers are smaller than 50000 bytes.
+        block_size = 524288;
+
+        //Note (observeur): This is called in case block_size is set to 0 (All in one mapping).
+        if(block_size == 0)
+        {
+            U8 * mptr = NULL;
+            LL_PROFILE_ZONE_NAMED_CATEGORY_VERTEX("glBufferSubData block");
+            LL_PROFILE_GPU_ZONE("glBufferSubData");
+
+            mptr = (U8*) glMapBufferRange( target, start, size, MapBits);
+
+            if(mptr)
+            {
+                std::memcpy(mptr, (U8*) data, size);
+                glUnmapBuffer(target);
+            }
+            else
+            {
+                LL_WARNS() << "glMapBufferRange() returned NULL" << LL_ENDL;
+            }
+            return;
+        }
+
+        //Note (observeur): The following code is executed in case of block_size is superior to 0
+
+        //Note (observeur): This is for analysis purpose only
+        //if(size > block_size)
+        //{
+        //    LL_INFOS() << "Large data range (MB MODE) : " << size << LL_ENDL;
+        //}
+
+        U8 * mptr = NULL;
 
         for (U32 i = start; i <= end; i += block_size)
         {
             LL_PROFILE_ZONE_NAMED_CATEGORY_VERTEX("glBufferSubData block");
-            //LL_PROFILE_GPU_ZONE("glBufferSubData");
+            LL_PROFILE_GPU_ZONE("glBufferSubData");
             U32 tend = llmin(i + block_size, end);
-            U32 size = tend - i + 1;
-            glBufferSubData(target, i, size, (U8*) data + (i-start));
+            size = tend - i + 1;
+
+            mptr = (U8*) glMapBufferRange( target, i, size, MapBits );
+
+            if(mptr)
+            {
+                std::memcpy(mptr, (U8*) data + (i-start), size);
+                glUnmapBuffer(target);
+            }
+            else
+            {
+                LL_WARNS() << "glMapBufferRange() returned NULL" << LL_ENDL;
+            }
         }
     }
 }