Fix ASAN errors from LLVector4a::memcpyNonAliased16

Found by running with -fsanitze=thread
Suggestion to avoid accessing invalid memory:

In both cases memory will be allocated by can be accessed beyond bounds.

In LLPolyMesh it can be off by at least one (+x%2). Though I am not even sure if even in best case it always will be a multiple of 16.

In LLViewerJointMesh::updateFaceData the code tries to account for padding by, but the allocation in LLPolyMeshSharedData::allocateVertexData is done without any padding. Thus the sizes must not match.

Replacing the calls with memcpy as a quick fix to see if the error goes away fixed address sanitzer complaining.

It is up to debate if memcpy is a good replacement. LLVector4a::memcpyNonAliased16 was invented for performance. But on the other hand one could argue that nowadays every stdlib maintainer will very heavily optmize functions like memcpy themselves and could take advantage of CPU features the old LL implementation does not take into account.

AVX comes to mind. In any case did I not measure any of this.

2 files changed, 11 insertions, 5 deletions

diff --git a/indra/llappearance/llpolymesh.cpp b/indra/llappearance/llpolymesh.cpp
index 97f9ca68b6..6bfb7456e8 100644
--- a/indra/llappearance/llpolymesh.cpp
+++ b/indra/llappearance/llpolymesh.cpp
@@ -981,7 +981,7 @@ void LLPolyMesh::initializeForMorph()
     LLVector4a::memcpyNonAliased16((F32*) mScaledNormals, (F32*) mSharedData->mBaseNormals, sizeof(LLVector4a) * mSharedData->mNumVertices);
     LLVector4a::memcpyNonAliased16((F32*) mBinormals, (F32*) mSharedData->mBaseNormals, sizeof(LLVector4a) * mSharedData->mNumVertices);
     LLVector4a::memcpyNonAliased16((F32*) mScaledBinormals, (F32*) mSharedData->mBaseNormals, sizeof(LLVector4a) * mSharedData->mNumVertices);
-    LLVector4a::memcpyNonAliased16((F32*) mTexCoords, (F32*) mSharedData->mTexCoords, sizeof(LLVector2) * (mSharedData->mNumVertices + mSharedData->mNumVertices%2));
+	memcpy((F32*) mTexCoords, (F32*) mSharedData->mTexCoords, sizeof(LLVector2) * (mSharedData->mNumVertices)); // allocated in LLPolyMeshSharedData::allocateVertexData
 
     for (S32 i = 0; i < mSharedData->mNumVertices; ++i)
     {
diff --git a/indra/newview/llviewerjointmesh.cpp b/indra/newview/llviewerjointmesh.cpp
index f0567b18c4..54b6c29e31 100644
--- a/indra/newview/llviewerjointmesh.cpp
+++ b/indra/newview/llviewerjointmesh.cpp
@@ -407,10 +407,16 @@ void LLViewerJointMesh::updateFaceData(LLFace *face, F32 pixel_area, bool damp_w
                 F32* vw = (F32*) vertex_weightsp.get();
                 F32* cw = (F32*) clothing_weightsp.get();
 
-                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);
+				//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);
+
+                // Both allocated in LLPolyMeshSharedData::allocateVertexData(unsigned int)
+
+                memcpy(tc, mMesh->getTexCoords(), num_verts*2*sizeof(F32) );
+                memcpy(vw, mMesh->getWeights(), num_verts*sizeof(F32) );
+
                 LLVector4a::memcpyNonAliased16(cw, (F32*) mMesh->getClothingWeights(), num_verts*4*sizeof(F32));
             }