merge

1 files changed, 57 insertions, 40 deletions

diff --git a/indra/llrender/llrender.cpp b/indra/llrender/llrender.cpp
index 2472339ec4..3f70ccacb1 100644
--- a/indra/llrender/llrender.cpp
+++ b/indra/llrender/llrender.cpp
@@ -39,6 +39,7 @@
 #include "llimagegl.h"
 #include "llrendertarget.h"
 #include "lltexture.h"
+#include "llvector4a.h"
 
 LLRender gGL;
 
@@ -53,6 +54,7 @@ U32 LLRender::sUICalls = 0;
 U32 LLRender::sUIVerts = 0;
 
 static const U32 LL_NUM_TEXTURE_LAYERS = 16; 
+static const U32 LL_MAX_UI_STACK_DEPTH = 32;
 
 static GLenum sGLTextureType[] =
 {
@@ -757,14 +759,27 @@ LLRender::LLRender()
     mCount(0),
     mMode(LLRender::TRIANGLES),
     mCurrTextureUnitIndex(0),
-    mMaxAnisotropy(0.f) 
+    mMaxAnisotropy(0.f),
+	mUIStackDepth(0)
 {
 	mBuffer = new LLVertexBuffer(immediate_mask, 0);
 	mBuffer->allocateBuffer(4096, 0, TRUE);
-	mBuffer->getVertexStrider(mVerticesp);
-	mBuffer->getTexCoord0Strider(mTexcoordsp);
-	mBuffer->getColorStrider(mColorsp);
-	
+
+	LLStrider<LLVector3> vert;
+	LLStrider<LLVector2> tc;
+	LLStrider<LLColor4U> color;
+
+	mBuffer->getVertexStrider(vert);
+	mBuffer->getTexCoord0Strider(tc);
+	mBuffer->getColorStrider(color);
+
+	mVerticesp = (LLVector4a*) vert.get();
+	mTexcoordsp = tc.get();
+	mColorsp = color.get();
+
+	mUIOffset = (LLVector4a*) _mm_malloc(LL_MAX_UI_STACK_DEPTH*sizeof(LLVector4a), 16);
+	mUIScale = (LLVector4a*) _mm_malloc(LL_MAX_UI_STACK_DEPTH*sizeof(LLVector4a), 16);
+		
 	mTexUnits.reserve(LL_NUM_TEXTURE_LAYERS);
 	for (U32 i = 0; i < LL_NUM_TEXTURE_LAYERS; i++)
 	{
@@ -800,6 +815,11 @@ void LLRender::shutdown()
 	mTexUnits.clear();
 	delete mDummyTexUnit;
 	mDummyTexUnit = NULL;
+
+	_mm_free(mUIOffset);
+	mUIOffset = NULL;
+	_mm_free(mUIScale);
+	mUIScale = NULL;
 }
 
 void LLRender::refreshState(void)
@@ -848,84 +868,83 @@ void LLRender::popMatrix()
 
 void LLRender::translateUI(F32 x, F32 y, F32 z)
 {
-	if (mUIOffset.empty())
+	if (mUIStackDepth == 0)
 	{
 		llerrs << "Need to push a UI translation frame before offsetting" << llendl;
 	}
 
-	mUIOffset.front().mV[0] += x;
-	mUIOffset.front().mV[1] += y;
-	mUIOffset.front().mV[2] += z;
+	LLVector4a trans(x,y,z);
+	mUIOffset[mUIStackDepth-1].add(trans);
 }
 
 void LLRender::scaleUI(F32 x, F32 y, F32 z)
 {
-	if (mUIScale.empty())
+	if (mUIStackDepth == 0)
 	{
 		llerrs << "Need to push a UI transformation frame before scaling." << llendl;
 	}
 
-	mUIScale.front().scaleVec(LLVector3(x,y,z));
+	LLVector4a scale(x,y,z);
+	mUIScale[mUIStackDepth-1].mul(scale);
 }
 
 void LLRender::pushUIMatrix()
 {
-	if (mUIOffset.empty())
+	if (mUIStackDepth == 0)
 	{
-		mUIOffset.push_front(LLVector3(0,0,0));
+		mUIOffset[0].clear();
+		mUIScale[0].splat(1.f);
 	}
-	else
+	else if (mUIStackDepth < LL_MAX_UI_STACK_DEPTH)
 	{
-		mUIOffset.push_front(mUIOffset.front());
-	}
-	
-	if (mUIScale.empty())
-	{
-		mUIScale.push_front(LLVector3(1,1,1));
+		mUIOffset[mUIStackDepth] = mUIOffset[mUIStackDepth-1];
+		mUIScale[mUIStackDepth] = mUIScale[mUIStackDepth-1];
 	}
 	else
 	{
-		mUIScale.push_front(mUIScale.front());
+		llerrs << "Blown UI matrix stack." << llendl;
 	}
+	
+	++mUIStackDepth;
+	
 }
 
 void LLRender::popUIMatrix()
 {
-	if (mUIOffset.empty())
+	if (mUIStackDepth == 0)
 	{
 		llerrs << "UI offset stack blown." << llendl;
 	}
-	mUIOffset.pop_front();
-	mUIScale.pop_front();
+	--mUIStackDepth;
 }
 
 LLVector3 LLRender::getUITranslation()
 {
-	if (mUIOffset.empty())
+	if (mUIStackDepth == 0)
 	{
 		llerrs << "UI offset stack empty." << llendl;
 	}
-	return mUIOffset.front();
+	return LLVector3(mUIOffset[mUIStackDepth-1].getF32());
 }
 
 LLVector3 LLRender::getUIScale()
 {
-	if (mUIScale.empty())
+	if (mUIStackDepth == 0)
 	{
 		llerrs << "UI scale stack empty." << llendl;
 	}
-	return mUIScale.front();
+	return LLVector3(mUIScale[mUIStackDepth-1].getF32());
 }
 
 
 void LLRender::loadUIIdentity()
 {
-	if (mUIOffset.empty())
+	if (mUIStackDepth == 0)
 	{
 		llerrs << "Need to push UI translation frame before clearing offset." << llendl;
 	}
-	mUIOffset.front().setVec(0,0,0);
-	mUIScale.front().setVec(1,1,1);
+	mUIOffset[mUIStackDepth-1].clear();
+	mUIScale[mUIStackDepth-1].splat(1.f);
 }
 
 void LLRender::setColorMask(bool writeColor, bool writeAlpha)
@@ -1154,7 +1173,7 @@ void LLRender::flush()
 		}
 #endif
 				
-		if (!mUIOffset.empty())
+		if (mUIStackDepth > 0)
 		{
 			sUICalls++;
 			sUIVerts += mCount;
@@ -1206,24 +1225,22 @@ void LLRender::vertex3f(const GLfloat& x, const GLfloat& y, const GLfloat& z)
 		return;
 	}
 
-	if (mUIOffset.empty())
+	LLVector3& v = reinterpret_cast<LLVector3&>(mVerticesp[mCount]);
+	v.set(x,y,z);
+	if (mUIStackDepth != 0)
 	{
-		mVerticesp[mCount] = LLVector3(x,y,z);
-	}
-	else
-	{
-		LLVector3 vert = (LLVector3(x,y,z)+mUIOffset.front()).scaledVec(mUIScale.front());
-		mVerticesp[mCount] = vert;
+		v += reinterpret_cast<LLVector3&>(mUIOffset[mUIStackDepth-1]);
+		v.scaleVec(reinterpret_cast<LLVector3&>(mUIScale[mUIStackDepth-1]));
 	}
 
 	mCount++;
 	if (mCount < 4096)
 	{
-		mVerticesp[mCount] = mVerticesp[mCount-1];
 		mColorsp[mCount] = mColorsp[mCount-1];
 		mTexcoordsp[mCount] = mTexcoordsp[mCount-1];
 	}
 }
+
 void LLRender::vertex2i(const GLint& x, const GLint& y)
 {
 	vertex3f((GLfloat) x, (GLfloat) y, 0);