From e68d5e248f73180def7c8928b32482347dd91de4 Mon Sep 17 00:00:00 2001
From: "Brad Payne (Vir Linden)" <vir@lindenlab.com>
Date: Wed, 14 Dec 2011 16:18:19 -0500
Subject: SH-2789 WIP - add asserts to check object address alignment where
 needed

---
 indra/llmath/llvector4a.h | 2 ++
 1 file changed, 2 insertions(+)

(limited to 'indra/llmath/llvector4a.h')

diff --git a/indra/llmath/llvector4a.h b/indra/llmath/llvector4a.h
index 596082509d..eada6ed1b7 100644
--- a/indra/llmath/llvector4a.h
+++ b/indra/llmath/llvector4a.h
@@ -32,6 +32,7 @@ class LLRotation;
 
 #include <assert.h>
 #include "llpreprocessor.h"
+#include "llmemory.h"
 
 ///////////////////////////////////
 // FIRST TIME USERS PLEASE READ
@@ -90,6 +91,7 @@ public:
 	
 	LLVector4a()
 	{ //DO NOT INITIALIZE -- The overhead is completely unnecessary
+		ll_assert_aligned(this,16);
 	}
 	
 	LLVector4a(F32 x, F32 y, F32 z, F32 w = 0.f)
-- 
cgit v1.3


From e4a1b12e7a9c4bb8cbc2a1e41fddb8cec5a1b1e0 Mon Sep 17 00:00:00 2001
From: "Brad Payne (Vir Linden)" <vir@lindenlab.com>
Date: Thu, 15 Dec 2011 12:48:07 -0500
Subject: SH-2789 WIP - build without tcmalloc, force alignment in various
 places

---
 indra/cmake/FindGooglePerfTools.cmake |  2 ++
 indra/cmake/GooglePerfTools.cmake     |  2 +-
 indra/llmath/llcamera.h               | 12 ++++++------
 indra/llmath/llmatrix3a.h             |  2 +-
 indra/llmath/llmatrix4a.h             |  2 +-
 indra/llmath/llplane.h                |  4 +++-
 indra/llmath/llvector4a.h             |  3 ++-
 indra/newview/CMakeLists.txt          |  5 +++--
 indra/newview/lldrawable.h            |  7 ++++---
 indra/newview/llspatialpartition.h    | 12 ++++++------
 indra/newview/llviewercamera.h        |  4 +++-
 indra/newview/llvoavatar.cpp          |  2 +-
 indra/newview/llvoavatar.h            |  2 +-
 13 files changed, 34 insertions(+), 25 deletions(-)
 mode change 100644 => 100755 indra/cmake/FindGooglePerfTools.cmake
 mode change 100644 => 100755 indra/cmake/GooglePerfTools.cmake
 mode change 100644 => 100755 indra/llmath/llcamera.h
 mode change 100644 => 100755 indra/llmath/llmatrix4a.h
 mode change 100644 => 100755 indra/llmath/llplane.h
 mode change 100644 => 100755 indra/llmath/llvector4a.h
 mode change 100644 => 100755 indra/newview/CMakeLists.txt
 mode change 100644 => 100755 indra/newview/lldrawable.h
 mode change 100644 => 100755 indra/newview/llspatialpartition.h
 mode change 100644 => 100755 indra/newview/llviewercamera.h
 mode change 100644 => 100755 indra/newview/llvoavatar.cpp
 mode change 100644 => 100755 indra/newview/llvoavatar.h

(limited to 'indra/llmath/llvector4a.h')

diff --git a/indra/cmake/FindGooglePerfTools.cmake b/indra/cmake/FindGooglePerfTools.cmake
old mode 100644
new mode 100755
index bb125d538e..1c785d3461
--- a/indra/cmake/FindGooglePerfTools.cmake
+++ b/indra/cmake/FindGooglePerfTools.cmake
@@ -64,3 +64,5 @@ MARK_AS_ADVANCED(
   PROFILER_LIBRARY
   GOOGLE_PERFTOOLS_INCLUDE_DIR
   )
+
+SET(GOOGLE_PERFTOOLS_FOUND "NO")
diff --git a/indra/cmake/GooglePerfTools.cmake b/indra/cmake/GooglePerfTools.cmake
old mode 100644
new mode 100755
index d9f91193be..2e96912769
--- a/indra/cmake/GooglePerfTools.cmake
+++ b/indra/cmake/GooglePerfTools.cmake
@@ -29,7 +29,7 @@ if (GOOGLE_PERFTOOLS_FOUND)
 endif (GOOGLE_PERFTOOLS_FOUND)
 
 if (WINDOWS)
-    set(USE_GOOGLE_PERFTOOLS ON)
+   set(USE_GOOGLE_PERFTOOLS OFF)
 endif (WINDOWS)
 
 if (USE_GOOGLE_PERFTOOLS)
diff --git a/indra/llmath/llcamera.h b/indra/llmath/llcamera.h
old mode 100644
new mode 100755
index ec67b91d05..0b591be622
--- a/indra/llmath/llcamera.h
+++ b/indra/llmath/llcamera.h
@@ -60,7 +60,7 @@ static const F32 MAX_FIELD_OF_VIEW = 175.f * DEG_TO_RAD;
 // roll(), pitch(), yaw()
 // etc...
 
-
+LL_ALIGN_PREFIX(16)
 class LLCamera
 : 	public LLCoordFrame
 {
@@ -108,7 +108,7 @@ public:
 	};
 
 private:
-	LLPlane mAgentPlanes[7];  //frustum planes in agent space a la gluUnproject (I'm a bastard, I know) - DaveP
+	LL_ALIGN_16(LLPlane mAgentPlanes[7]);  //frustum planes in agent space a la gluUnproject (I'm a bastard, I know) - DaveP
 	U8 mPlaneMask[8];         // 8 for alignment	
 	
 	F32 mView;					// angle between top and bottom frustum planes in radians.
@@ -116,13 +116,13 @@ private:
 	S32 mViewHeightInPixels;	// for ViewHeightInPixels() only
 	F32 mNearPlane;
 	F32 mFarPlane;
-	LLPlane mLocalPlanes[4];
+	LL_ALIGN_16(LLPlane mLocalPlanes[4]);
 	F32 mFixedDistance;			// Always return this distance, unless < 0
 	LLVector3 mFrustCenter;		// center of frustum and radius squared for ultra-quick exclusion test
 	F32 mFrustRadiusSquared;
 	
-	LLPlane mWorldPlanes[PLANE_NUM];
-	LLPlane mHorizPlanes[HORIZ_PLANE_NUM];
+	LL_ALIGN_16(LLPlane mWorldPlanes[PLANE_NUM]);
+	LL_ALIGN_16(LLPlane mHorizPlanes[HORIZ_PLANE_NUM]);
 
 	U32 mPlaneCount;  //defaults to 6, if setUserClipPlane is called, uses user supplied clip plane in
 
@@ -208,7 +208,7 @@ protected:
 	void calculateFrustumPlanes(F32 left, F32 right, F32 top, F32 bottom);
 	void calculateFrustumPlanesFromWindow(F32 x1, F32 y1, F32 x2, F32 y2);
 	void calculateWorldFrustumPlanes();
-};
+} LL_ALIGN_POSTFIX(16);
 
 
 #endif
diff --git a/indra/llmath/llmatrix3a.h b/indra/llmath/llmatrix3a.h
index adb7e3389d..9916cfd2da 100644
--- a/indra/llmath/llmatrix3a.h
+++ b/indra/llmath/llmatrix3a.h
@@ -111,7 +111,7 @@ public:
 
 protected:
 
-	LLVector4a mColumns[3];
+	LL_ALIGN_16(LLVector4a mColumns[3]);
 
 };
 
diff --git a/indra/llmath/llmatrix4a.h b/indra/llmath/llmatrix4a.h
old mode 100644
new mode 100755
index 27cf5b79f6..c4cefdb4fa
--- a/indra/llmath/llmatrix4a.h
+++ b/indra/llmath/llmatrix4a.h
@@ -34,7 +34,7 @@
 class LLMatrix4a
 {
 public:
-	LLVector4a mMatrix[4];
+	LL_ALIGN_16(LLVector4a mMatrix[4]);
 
 	inline void clear()
 	{
diff --git a/indra/llmath/llplane.h b/indra/llmath/llplane.h
old mode 100644
new mode 100755
index a611894721..3c32441b11
--- a/indra/llmath/llplane.h
+++ b/indra/llmath/llplane.h
@@ -36,6 +36,8 @@
 // The plane normal = [A, B, C]
 // The closest approach = D / sqrt(A*A + B*B + C*C)
 
+
+LL_ALIGN_PREFIX(16)
 class LLPlane
 {
 public:
@@ -94,7 +96,7 @@ public:
 		
 private:
 	LLVector4a mV;
-};
+} LL_ALIGN_POSTFIX(16);
 
 
 
diff --git a/indra/llmath/llvector4a.h b/indra/llmath/llvector4a.h
old mode 100644
new mode 100755
index eada6ed1b7..9de0e66774
--- a/indra/llmath/llvector4a.h
+++ b/indra/llmath/llvector4a.h
@@ -47,6 +47,7 @@ class LLRotation;
 // LLVector3/LLVector4. 
 /////////////////////////////////
 
+LL_ALIGN_PREFIX(16)
 class LLVector4a
 {
 public:
@@ -315,7 +316,7 @@ public:
 
 private:
 	LLQuad mQ;
-};
+} LL_ALIGN_POSTFIX(16);
 
 inline void update_min_max(LLVector4a& min, LLVector4a& max, const LLVector4a& p)
 {
diff --git a/indra/newview/CMakeLists.txt b/indra/newview/CMakeLists.txt
old mode 100644
new mode 100755
index 6b2fe1e45a..825cbbef8f
--- a/indra/newview/CMakeLists.txt
+++ b/indra/newview/CMakeLists.txt
@@ -1509,7 +1509,8 @@ if (WINDOWS)
         PROPERTIES
         # *TODO -reenable this once we get server usage sorted out
         #LINK_FLAGS "/debug /NODEFAULTLIB:LIBCMT /SUBSYSTEM:WINDOWS /INCLUDE:\"__tcmalloc\""
-        LINK_FLAGS "/debug /NODEFAULTLIB:LIBCMT /SUBSYSTEM:WINDOWS /INCLUDE:__tcmalloc"
+#        LINK_FLAGS "/debug /NODEFAULTLIB:LIBCMT /SUBSYSTEM:WINDOWS /INCLUDE:__tcmalloc"
+        LINK_FLAGS "/debug /NODEFAULTLIB:LIBCMT /SUBSYSTEM:WINDOWS"
         LINK_FLAGS_DEBUG "/NODEFAULTLIB:\"LIBCMT;LIBCMTD;MSVCRT\" /INCREMENTAL:NO"
         LINK_FLAGS_RELEASE ""
         )
@@ -1764,7 +1765,7 @@ target_link_libraries(${VIEWER_BINARY_NAME}
     ${CRYPTO_LIBRARIES}
     ${LLLOGIN_LIBRARIES}
     ${LLCONVEXDECOMP_LIBRARY}
-    ${TCMALLOC_LIBRARIES}
+#    ${TCMALLOC_LIBRARIES}
     )
 
 if (USE_KDU)
diff --git a/indra/newview/lldrawable.h b/indra/newview/lldrawable.h
old mode 100644
new mode 100755
index e268640a21..244b87b2e7
--- a/indra/newview/lldrawable.h
+++ b/indra/newview/lldrawable.h
@@ -59,6 +59,7 @@ class LLViewerTexture;
 const U32 SILHOUETTE_HIGHLIGHT = 0;
 
 // All data for new renderer goes into this class.
+LL_ALIGN_PREFIX(16)
 class LLDrawable : public LLRefCount
 {
 public:
@@ -280,8 +281,8 @@ public:
 	} EDrawableFlags;
 
 private: //aligned members
-	LLVector4a		mExtents[2];
-	LLVector4a		mPositionGroup;
+	LL_ALIGN_16(LLVector4a		mExtents[2]);
+	LL_ALIGN_16(LLVector4a		mPositionGroup);
 	
 public:
 	LLXformMatrix       mXform;
@@ -322,7 +323,7 @@ private:
 
 	static U32 sNumZombieDrawables;
 	static LLDynamicArrayPtr<LLPointer<LLDrawable> > sDeadList;
-};
+} LL_ALIGN_POSTFIX(16);
 
 
 inline LLFace* LLDrawable::getFace(const S32 i) const
diff --git a/indra/newview/llspatialpartition.h b/indra/newview/llspatialpartition.h
old mode 100644
new mode 100755
index f0c8a372ee..cbd4507f84
--- a/indra/newview/llspatialpartition.h
+++ b/indra/newview/llspatialpartition.h
@@ -372,12 +372,12 @@ public:
 		V4_COUNT = 10
 	} eV4Index;
 
-	LLVector4a mBounds[2]; // bounding box (center, size) of this node and all its children (tight fit to objects)
-	LLVector4a mExtents[2]; // extents (min, max) of this node and all its children
-	LLVector4a mObjectExtents[2]; // extents (min, max) of objects in this node
-	LLVector4a mObjectBounds[2]; // bounding box (center, size) of objects in this node
-	LLVector4a mViewAngle;
-	LLVector4a mLastUpdateViewAngle;
+	LL_ALIGN_16(LLVector4a mBounds[2]); // bounding box (center, size) of this node and all its children (tight fit to objects)
+	LL_ALIGN_16(LLVector4a mExtents[2]); // extents (min, max) of this node and all its children
+	LL_ALIGN_16(LLVector4a mObjectExtents[2]); // extents (min, max) of objects in this node
+	LL_ALIGN_16(LLVector4a mObjectBounds[2]); // bounding box (center, size) of objects in this node
+	LL_ALIGN_16(LLVector4a mViewAngle);
+	LL_ALIGN_16(LLVector4a mLastUpdateViewAngle);
 		
 private:
 	U32                     mCurUpdatingTime ;
diff --git a/indra/newview/llviewercamera.h b/indra/newview/llviewercamera.h
old mode 100644
new mode 100755
index cc3395115b..7257297cfd
--- a/indra/newview/llviewercamera.h
+++ b/indra/newview/llviewercamera.h
@@ -51,6 +51,7 @@ const BOOL NOT_FOR_SELECTION = FALSE;
 extern template class LLViewerCamera* LLSingleton<class LLViewerCamera>::getInstance();
 #endif
 
+LL_ALIGN_PREFIX(16)
 class LLViewerCamera : public LLCamera, public LLSingleton<LLViewerCamera>
 {
 public:
@@ -137,6 +138,7 @@ protected:
 	S16					mZoomSubregion;
 
 public:
-};
+} LL_ALIGN_POSTFIX(16);
+
 
 #endif // LL_LLVIEWERCAMERA_H
diff --git a/indra/newview/llvoavatar.cpp b/indra/newview/llvoavatar.cpp
old mode 100644
new mode 100755
index 4767ba2bed..ad5ea15bbb
--- a/indra/newview/llvoavatar.cpp
+++ b/indra/newview/llvoavatar.cpp
@@ -2595,7 +2595,7 @@ void LLVOAvatar::idleUpdateMisc(bool detailed_update)
 
 	if (isImpostor() && !mNeedsImpostorUpdate)
 	{
-		LLVector4a ext[2];
+		LL_ALIGN_16(LLVector4a ext[2]);
 		F32 distance;
 		LLVector3 angle;
 
diff --git a/indra/newview/llvoavatar.h b/indra/newview/llvoavatar.h
old mode 100644
new mode 100755
index 59796370ae..82157436b9
--- a/indra/newview/llvoavatar.h
+++ b/indra/newview/llvoavatar.h
@@ -212,7 +212,7 @@ public:
 	bool isBuilt() const { return mIsBuilt; }
 
 private: //aligned members
-	LLVector4a	mImpostorExtents[2];
+	LL_ALIGN_16(LLVector4a	mImpostorExtents[2]);
 
 private:
 	BOOL			mSupportsAlphaLayers; // For backwards compatibility, TRUE for 1.23+ clients
-- 
cgit v1.3


From 1435a8b9e6203911d2ebe9e3ba217f8eb20e3140 Mon Sep 17 00:00:00 2001
From: "Brad Payne (Vir Linden)" <vir@lindenlab.com>
Date: Wed, 4 Jan 2012 15:21:23 -0500
Subject: SH-2789 WIP - stricter calling of memcpyNonAliased16

---
 indra/llcommon/llmemory.cpp             |  3 +++
 indra/llcommon/llmemory.h               | 12 ++++++------
 indra/llmath/llvector4a.cpp             |  3 ---
 indra/llmath/llvector4a.h               |  1 +
 indra/llmath/tests/alignment_test.cpp   |  6 ++++--
 indra/newview/llfloatermodelpreview.cpp |  3 ++-
 indra/newview/llviewerjointmesh.cpp     |  6 ++++--
 7 files changed, 20 insertions(+), 14 deletions(-)
 mode change 100644 => 100755 indra/llcommon/llmemory.cpp
 mode change 100644 => 100755 indra/newview/llviewerjointmesh.cpp

(limited to 'indra/llmath/llvector4a.h')

diff --git a/indra/llcommon/llmemory.cpp b/indra/llcommon/llmemory.cpp
old mode 100644
new mode 100755
index 22204e756a..afaf366668
--- a/indra/llcommon/llmemory.cpp
+++ b/indra/llcommon/llmemory.cpp
@@ -63,11 +63,14 @@ LLPrivateMemoryPoolManager::mem_allocation_info_t LLPrivateMemoryPoolManager::sM
 
 void ll_assert_aligned_func(uintptr_t ptr,U32 alignment)
 {
+#ifdef SHOW_ASSERT
+	// Redundant, place to set breakpoints.
 	if (ptr%alignment!=0)
 	{
 		llwarns << "alignment check failed" << llendl;
 	}
 	llassert(ptr%alignment==0);
+#endif
 }
 
 //static
diff --git a/indra/llcommon/llmemory.h b/indra/llcommon/llmemory.h
index 3eaf700bf1..1b54b56107 100755
--- a/indra/llcommon/llmemory.h
+++ b/indra/llcommon/llmemory.h
@@ -83,7 +83,7 @@ inline void ll_aligned_free_16(void *p)
 inline void* ll_aligned_malloc_32(size_t size) // returned hunk MUST be freed with ll_aligned_free_32().
 {
 #if defined(LL_WINDOWS)
-	return _mm_malloc(size, 32);
+	return _aligned_malloc(size, 32);
 #elif defined(LL_DARWIN)
 	return ll_aligned_malloc( size, 32 );
 #else
@@ -98,7 +98,7 @@ inline void* ll_aligned_malloc_32(size_t size) // returned hunk MUST be freed wi
 inline void ll_aligned_free_32(void *p)
 {
 #if defined(LL_WINDOWS)
-	_mm_free(p);
+	_aligned_free(p);
 #elif defined(LL_DARWIN)
 	ll_aligned_free( p );
 #else
@@ -107,10 +107,12 @@ inline void ll_aligned_free_32(void *p)
 }
 
 #else // USE_TCMALLOC
-// ll_aligned_foo are noops now that we use tcmalloc everywhere (tcmalloc aligns automatically at appropriate intervals)
+// ll_aligned_foo are noops now that we use tcmalloc everywhere
+// (tcmalloc aligns automatically at appropriate intervals)
 #define ll_aligned_malloc( size, align ) malloc(size)
 #define ll_aligned_free( ptr ) free(ptr)
 #define ll_aligned_malloc_16 malloc
+#define ll_aligned_realloc_16 realloc
 #define ll_aligned_free_16 free
 #define ll_aligned_malloc_32 malloc
 #define ll_aligned_free_32 free
@@ -524,11 +526,9 @@ void  LLPrivateMemoryPoolTester::operator delete[](void* addr)
 
 // LLSingleton moved to llsingleton.h
 
-#define CHECK_ALIGNMENT
-
 LL_COMMON_API void ll_assert_aligned_func(uintptr_t ptr,U32 alignment);
 
-#ifdef CHECK_ALIGNMENT
+#ifdef SHOW_ASSERT
 #define ll_assert_aligned(ptr,alignment) ll_assert_aligned_func(reinterpret_cast<uintptr_t>(ptr),((U32)alignment))
 #else
 #define ll_assert_aligned(ptr,alignment)
diff --git a/indra/llmath/llvector4a.cpp b/indra/llmath/llvector4a.cpp
index 480ccf4ed9..6edeb0fefe 100755
--- a/indra/llmath/llvector4a.cpp
+++ b/indra/llmath/llvector4a.cpp
@@ -41,8 +41,6 @@ extern const LLVector4a LL_V4A_EPSILON = reinterpret_cast<const LLVector4a&> ( F
 
 /*static */void LLVector4a::memcpyNonAliased16(F32* __restrict dst, const F32* __restrict src, size_t bytes)
 {
-//	memcpy((void*)dst,(const void*)src,bytes);
-#if 1
 	assert(src != NULL);
 	assert(dst != NULL);
 	assert(bytes > 0);
@@ -92,7 +90,6 @@ extern const LLVector4a LL_V4A_EPSILON = reinterpret_cast<const LLVector4a&> ( F
 		dst += 4;
 		src += 4;
 	}
-#endif
 }
 
 void LLVector4a::setRotated( const LLRotation& rot, const LLVector4a& vec )
diff --git a/indra/llmath/llvector4a.h b/indra/llmath/llvector4a.h
index 9de0e66774..0526793d3a 100755
--- a/indra/llmath/llvector4a.h
+++ b/indra/llmath/llvector4a.h
@@ -84,6 +84,7 @@ public:
 	}
 
 	// Copy words 16-byte blocks from src to dst. Source and destination must not overlap. 
+	// Source and dest must be 16-byte aligned and size must be multiple of 16.
 	static void memcpyNonAliased16(F32* __restrict dst, const F32* __restrict src, size_t bytes);
 
 	////////////////////////////////////
diff --git a/indra/llmath/tests/alignment_test.cpp b/indra/llmath/tests/alignment_test.cpp
index 51a7051e69..dc9b41957d 100755
--- a/indra/llmath/tests/alignment_test.cpp
+++ b/indra/llmath/tests/alignment_test.cpp
@@ -72,6 +72,10 @@ void alignment_test_object_t::test<1>()
 	{
 		align_ptr = ll_aligned_malloc_16(sizeof(MyVector4a));
 		ensure("ll_aligned_malloc_16 failed", is_aligned(align_ptr,16));
+
+		align_ptr = ll_aligned_realloc_16(2*sizeof(MyVector4a));
+		ensure("ll_aligned_realloc_16 failed", is_aligned(align_ptr,16));
+
 		ll_aligned_free_16(align_ptr);
 
 		align_ptr = ll_aligned_malloc_32(sizeof(MyVector4a));
@@ -84,8 +88,6 @@ void alignment_test_object_t::test<1>()
 template<> template<>
 void alignment_test_object_t::test<2>()
 {
-	ensure("LLAlignment reality is broken: ", (1==1));
-
 	MyVector4a vec1;
 	ensure("LLAlignment vec1 unaligned", is_aligned(&vec1,16));
 	
diff --git a/indra/newview/llfloatermodelpreview.cpp b/indra/newview/llfloatermodelpreview.cpp
index 64bdcccd9f..8aa3b95578 100755
--- a/indra/newview/llfloatermodelpreview.cpp
+++ b/indra/newview/llfloatermodelpreview.cpp
@@ -4770,7 +4770,8 @@ void LLModelPreview::genBuffers(S32 lod, bool include_skin_weights)
 			if (vf.mTexCoords)
 			{
 				vb->getTexCoord0Strider(tc_strider);
-				LLVector4a::memcpyNonAliased16((F32*) tc_strider.get(), (F32*) vf.mTexCoords, num_vertices*2*sizeof(F32));
+				S32 tex_size = (num_vertices*2*sizeof(F32)+0xF) & ~0xF;
+				LLVector4a::memcpyNonAliased16((F32*) tc_strider.get(), (F32*) vf.mTexCoords, tex_size);
 			}
 			
 			if (vf.mNormals)
diff --git a/indra/newview/llviewerjointmesh.cpp b/indra/newview/llviewerjointmesh.cpp
old mode 100644
new mode 100755
index 76f4e18c27..d604687678
--- a/indra/newview/llviewerjointmesh.cpp
+++ b/indra/newview/llviewerjointmesh.cpp
@@ -731,8 +731,10 @@ void LLViewerJointMesh::updateFaceData(LLFace *face, F32 pixel_area, BOOL damp_w
 				F32* vw = (F32*) vertex_weightsp.get();
 				F32* cw = (F32*) clothing_weightsp.get();	
 
-				LLVector4a::memcpyNonAliased16(tc, (F32*) mMesh->getTexCoords(), num_verts*2*sizeof(F32));
-				LLVector4a::memcpyNonAliased16(vw, (F32*) mMesh->getWeights(), num_verts*sizeof(F32));	
+				S32 tc_size = (num_verts*2*sizeof(F32)+0xF) & ~0xF;
+				LLVector4a::memcpyNonAliased16(tc, (F32*) mMesh->getTexCoords(), tc_size);
+				S32 vw_size = (num_verts*sizeof(F32)+0xF) & ~0xF;	
+				LLVector4a::memcpyNonAliased16(vw, (F32*) mMesh->getWeights(), vw_size);	
 				LLVector4a::memcpyNonAliased16(cw, (F32*) mMesh->getClothingWeights(), num_verts*4*sizeof(F32));	
 			}
 
-- 
cgit v1.3