diff --git a/src/BulletCollision/BroadphaseCollision/btDbvt.h b/src/BulletCollision/BroadphaseCollision/btDbvt.h
index ae765e32e..9dae440f6 100644
--- a/src/BulletCollision/BroadphaseCollision/btDbvt.h
+++ b/src/BulletCollision/BroadphaseCollision/btDbvt.h
@@ -568,6 +568,8 @@ DBVT_INLINE btScalar	Proximity(	const btDbvtAabbMm& a,
 	return(btFabs(d.x())+btFabs(d.y())+btFabs(d.z()));
 }
 
+
+
 //
 DBVT_INLINE int			Select(	const btDbvtAabbMm& o,
 							   const btDbvtAabbMm& a,
@@ -577,6 +579,14 @@ DBVT_INLINE int			Select(	const btDbvtAabbMm& o,
 	static DBVT_ALIGN const unsigned __int32	mask[]={0x7fffffff,0x7fffffff,0x7fffffff,0x7fffffff};
 	// TODO: the intrinsic version is 11% slower
 #if DBVT_USE_INTRINSIC_SSE
+
+	union btSSEUnion ///NOTE: if we use more intrinsics, move btSSEUnion into the LinearMath directory
+	{
+	   __m128		ssereg;
+	   float		floats[4];
+	   int			ints[4];
+	};
+
 	__m128	omi(_mm_load_ps(o.mi));
 	omi=_mm_add_ps(omi,_mm_load_ps(o.mx));
 	__m128	ami(_mm_load_ps(a.mi));
@@ -590,10 +600,14 @@ DBVT_INLINE int			Select(	const btDbvtAabbMm& o,
 	__m128	t0(_mm_movehl_ps(ami,ami));
 	ami=_mm_add_ps(ami,t0);
 	ami=_mm_add_ss(ami,_mm_shuffle_ps(ami,ami,1));
-	__m128	t1(_mm_movehl_ps(bmi,bmi));
+	__m128 t1(_mm_movehl_ps(bmi,bmi));
 	bmi=_mm_add_ps(bmi,t1);
 	bmi=_mm_add_ss(bmi,_mm_shuffle_ps(bmi,bmi,1));
-	return(_mm_cmple_ss(bmi,ami).m128_u32[0]&1);
+	
+	btSSEUnion tmp;
+	tmp.ssereg = _mm_cmple_ss(bmi,ami);
+	return tmp.ints[0]&1;
+
 #else
 	DBVT_ALIGN __int32	r[1];
 	__asm
diff --git a/src/BulletCollision/CollisionShapes/btBvhTriangleMeshShape.cpp b/src/BulletCollision/CollisionShapes/btBvhTriangleMeshShape.cpp
index 5317bce52..85b4ac461 100644
--- a/src/BulletCollision/CollisionShapes/btBvhTriangleMeshShape.cpp
+++ b/src/BulletCollision/CollisionShapes/btBvhTriangleMeshShape.cpp
@@ -143,10 +143,19 @@ void	btBvhTriangleMeshShape::performRaycast (btTriangleCallback* callback, const
 			for (int j=2;j>=0;j--)
 			{
 				int graphicsindex = indicestype==PHY_SHORT?((unsigned short*)gfxbase)[j]:gfxbase[j];
-
-				btScalar* graphicsbase = (btScalar*)(vertexbase+graphicsindex*stride);
-
-				m_triangle[j] = btVector3(graphicsbase[0]*meshScaling.getX(),graphicsbase[1]*meshScaling.getY(),graphicsbase[2]*meshScaling.getZ());		
+				
+				if (type == PHY_FLOAT)
+				{
+					float* graphicsbase = (float*)(vertexbase+graphicsindex*stride);
+					
+					m_triangle[j] = btVector3(graphicsbase[0]*meshScaling.getX(),graphicsbase[1]*meshScaling.getY(),graphicsbase[2]*meshScaling.getZ());		
+				}
+				else
+				{
+					double* graphicsbase = (double*)(vertexbase+graphicsindex*stride);
+					
+					m_triangle[j] = btVector3(graphicsbase[0]*meshScaling.getX(),graphicsbase[1]*meshScaling.getY(),graphicsbase[2]*meshScaling.getZ());		
+				}
 			}
 
 			/* Perform ray vs. triangle collision here */
@@ -204,9 +213,18 @@ void	btBvhTriangleMeshShape::performConvexcast (btTriangleCallback* callback, co
 			{
 				int graphicsindex = indicestype==PHY_SHORT?((unsigned short*)gfxbase)[j]:gfxbase[j];
 
-				btScalar* graphicsbase = (btScalar*)(vertexbase+graphicsindex*stride);
+				if (type == PHY_FLOAT)
+				{
+					float* graphicsbase = (float*)(vertexbase+graphicsindex*stride);
 
-				m_triangle[j] = btVector3(graphicsbase[0]*meshScaling.getX(),graphicsbase[1]*meshScaling.getY(),graphicsbase[2]*meshScaling.getZ());		
+					m_triangle[j] = btVector3(graphicsbase[0]*meshScaling.getX(),graphicsbase[1]*meshScaling.getY(),graphicsbase[2]*meshScaling.getZ());		
+				}
+				else
+				{
+					double* graphicsbase = (double*)(vertexbase+graphicsindex*stride);
+					
+					m_triangle[j] = btVector3(graphicsbase[0]*meshScaling.getX(),graphicsbase[1]*meshScaling.getY(),graphicsbase[2]*meshScaling.getZ());		
+				}
 			}
 
 			/* Perform ray vs. triangle collision here */
@@ -281,12 +299,24 @@ void	btBvhTriangleMeshShape::processAllTriangles(btTriangleCallback* callback,co
 #ifdef DEBUG_TRIANGLE_MESH
 				printf("%d ,",graphicsindex);
 #endif //DEBUG_TRIANGLE_MESH
-				btScalar* graphicsbase = (btScalar*)(vertexbase+graphicsindex*stride);
+				if (type == PHY_FLOAT)
+				{
+					float* graphicsbase = (float*)(vertexbase+graphicsindex*stride);
+					
+					m_triangle[j] = btVector3(
+																		graphicsbase[0]*meshScaling.getX(),
+																		graphicsbase[1]*meshScaling.getY(),
+																		graphicsbase[2]*meshScaling.getZ());
+				}
+				else
+				{
+					double* graphicsbase = (double*)(vertexbase+graphicsindex*stride);
 
-				m_triangle[j] = btVector3(
-					graphicsbase[0]*meshScaling.getX(),
-					graphicsbase[1]*meshScaling.getY(),
-					graphicsbase[2]*meshScaling.getZ());
+					m_triangle[j] = btVector3(
+						graphicsbase[0]*meshScaling.getX(),
+						graphicsbase[1]*meshScaling.getY(),
+						graphicsbase[2]*meshScaling.getZ());
+				}
 #ifdef DEBUG_TRIANGLE_MESH
 				printf("triangle vertices:%f,%f,%f\n",triangle[j].x(),triangle[j].y(),triangle[j].z());
 #endif //DEBUG_TRIANGLE_MESH
diff --git a/src/BulletCollision/CollisionShapes/btConvexTriangleMeshShape.cpp b/src/BulletCollision/CollisionShapes/btConvexTriangleMeshShape.cpp
index db9e4493c..0a47a6cfe 100644
--- a/src/BulletCollision/CollisionShapes/btConvexTriangleMeshShape.cpp
+++ b/src/BulletCollision/CollisionShapes/btConvexTriangleMeshShape.cpp
@@ -269,15 +269,12 @@ void btConvexTriangleMeshShape::calculatePrincipalAxisTransform(btTransform& pri
          btVector3 a = triangle[0] - center;
          btVector3 b = triangle[1] - center;
          btVector3 c = triangle[2] - center;
-         btVector3 abc = a + b + c;
          btScalar volNeg = -btFabs(a.triple(b, c)) * btScalar(1. / 6);
          for (int j = 0; j < 3; j++)
          {
             for (int k = 0; k <= j; k++)
             {
-               i[j][k] = i[k][j] = volNeg * (center[j] * center[k]
-                 + btScalar(0.25) * (center[j] * abc[k] + center[k] * abc[j])
-                  + btScalar(0.1) * (a[j] * a[k] + b[j] * b[k] + c[j] * c[k])
+               i[j][k] = i[k][j] = volNeg * (btScalar(0.1) * (a[j] * a[k] + b[j] * b[k] + c[j] * c[k])
                   + btScalar(0.05) * (a[j] * b[k] + a[k] * b[j] + a[j] * c[k] + a[k] * c[j] + b[j] * c[k] + b[k] * c[j]));
             }
          }
diff --git a/src/BulletCollision/CollisionShapes/btOptimizedBvh.cpp b/src/BulletCollision/CollisionShapes/btOptimizedBvh.cpp
index a248c55b5..3e7305939 100644
--- a/src/BulletCollision/CollisionShapes/btOptimizedBvh.cpp
+++ b/src/BulletCollision/CollisionShapes/btOptimizedBvh.cpp
@@ -319,19 +319,19 @@ void	btOptimizedBvh::updateBvhNodes(btStridingMeshInterface* meshInterface,int f
 				{
 					
 					int graphicsindex = indicestype==PHY_SHORT?((unsigned short*)gfxbase)[j]:gfxbase[j];
-					btScalar* graphicsbase = (btScalar*)(vertexbase+graphicsindex*stride);
-#ifdef DEBUG_PATCH_COLORS
-					btVector3 mycolor = color[index&3];
-					graphicsbase[8] = mycolor.getX();
-					graphicsbase[9] = mycolor.getY();
-					graphicsbase[10] = mycolor.getZ();
-#endif //DEBUG_PATCH_COLORS
-
-
-					triangleVerts[j] = btVector3(
-						graphicsbase[0]*meshScaling.getX(),
-						graphicsbase[1]*meshScaling.getY(),
-						graphicsbase[2]*meshScaling.getZ());
+					if (type == PHY_FLOAT)
+					{
+						float* graphicsbase = (float*)(vertexbase+graphicsindex*stride);
+						triangleVerts[j] = btVector3(
+							graphicsbase[0]*meshScaling.getX(),
+							graphicsbase[1]*meshScaling.getY(),
+							graphicsbase[2]*meshScaling.getZ());
+					}
+					else
+					{
+						double* graphicsbase = (double*)(vertexbase+graphicsindex*stride);
+						triangleVerts[j] = btVector3( btScalar(graphicsbase[0]*meshScaling.getX()), btScalar(graphicsbase[1]*meshScaling.getY()), btScalar(graphicsbase[2]*meshScaling.getZ()));
+					}
 				}