diff --git a/Bullet/BroadphaseCollision/AxisSweep3.cpp b/Bullet/BroadphaseCollision/AxisSweep3.cpp
index 235afa1f9..535369a41 100644
--- a/Bullet/BroadphaseCollision/AxisSweep3.cpp
+++ b/Bullet/BroadphaseCollision/AxisSweep3.cpp
@@ -215,7 +215,7 @@ void AxisSweep3::RemoveHandle(unsigned short handle)
 {
 	Handle* pHandle = GetHandle(handle);
 
-	RemoveOverlappingPairsContainingProxy(pHandle);
+	//RemoveOverlappingPairsContainingProxy(pHandle);
 
 
 	// compute current limit of edge arrays
diff --git a/Bullet/BroadphaseCollision/BroadphaseProxy.h b/Bullet/BroadphaseCollision/BroadphaseProxy.h
index 0e64d0632..b26a2701d 100644
--- a/Bullet/BroadphaseCollision/BroadphaseProxy.h
+++ b/Bullet/BroadphaseCollision/BroadphaseProxy.h
@@ -42,6 +42,9 @@ IMPLICIT_CONVEX_SHAPES_START_HERE,
 CONCAVE_SHAPES_START_HERE,
 	//keep all the convex shapetype below here, for the check IsConvexShape in broadphase proxy!
 	TRIANGLE_MESH_SHAPE_PROXYTYPE,
+	///used for demo integration FAST/Swift collision library and Bullet
+	FAST_CONCAVE_MESH_PROXYTYPE,
+
 	EMPTY_SHAPE_PROXYTYPE,
 	STATIC_PLANE_PROXYTYPE,
 CONCAVE_SHAPES_END_HERE,
@@ -124,17 +127,26 @@ struct BroadphasePair
 		}
 	}
 	BroadphasePair(BroadphaseProxy& proxy0,BroadphaseProxy& proxy1)
-		:
-		m_pProxy0(&proxy0),
-		m_pProxy1(&proxy1)
 	{
+
+		//keep them sorted, so the std::set operations work
+		if (&proxy0 < &proxy1)
+        { 
+            m_pProxy0 = &proxy0; 
+            m_pProxy1 = &proxy1; 
+        }
+        else 
+        { 
+			m_pProxy0 = &proxy1; 
+            m_pProxy1 = &proxy0; 
+        }
+
 		for (int i=0;i<SIMPLE_MAX_ALGORITHMS;i++)
-	{
+		{
 			m_algorithms[i] = 0;
 		}
 
 	}
-
 	
 	BroadphaseProxy* m_pProxy0;
 	BroadphaseProxy* m_pProxy1;
@@ -142,5 +154,14 @@ struct BroadphasePair
 	mutable CollisionAlgorithm* m_algorithms[SIMPLE_MAX_ALGORITHMS];
 };
 
+
+//comparison for set operation, see Solid DT_Encounter
+inline bool operator<(const BroadphasePair& a, const BroadphasePair& b) 
+{ 
+    return a.m_pProxy0 < b.m_pProxy0 || 
+        (a.m_pProxy0 == b.m_pProxy0 && a.m_pProxy1 < b.m_pProxy1); 
+}
+
+
 #endif //BROADPHASE_PROXY_H
 
diff --git a/Bullet/BroadphaseCollision/Dispatcher.h b/Bullet/BroadphaseCollision/Dispatcher.h
index 5edfbbdd3..47a006f69 100644
--- a/Bullet/BroadphaseCollision/Dispatcher.h
+++ b/Bullet/BroadphaseCollision/Dispatcher.h
@@ -87,7 +87,7 @@ public:
 
 	virtual	void	ReleaseManifoldResult(ManifoldResult*)=0;
 
-	virtual void	DispatchAllCollisionPairs(struct BroadphasePair* pairs,int numPairs,DispatcherInfo& dispatchInfo)=0;
+	virtual void	DispatchAllCollisionPairs(OverlappingPairCache* pairCache,DispatcherInfo& dispatchInfo)=0;
 
 	virtual int GetNumManifolds() const = 0;
 
diff --git a/Bullet/BroadphaseCollision/OverlappingPairCache.cpp b/Bullet/BroadphaseCollision/OverlappingPairCache.cpp
index 89ee1ee73..f8d3e7b54 100644
--- a/Bullet/BroadphaseCollision/OverlappingPairCache.cpp
+++ b/Bullet/BroadphaseCollision/OverlappingPairCache.cpp
@@ -24,26 +24,24 @@ subject to the following restrictions:
 
 OverlappingPairCache::OverlappingPairCache(int maxOverlap):
 m_blockedForChanges(false),
-m_NumOverlapBroadphasePair(0),
+//m_NumOverlapBroadphasePair(0),
 m_maxOverlap(maxOverlap)
 {
-	m_OverlappingPairs = new BroadphasePair[maxOverlap];
 }
 
 
 OverlappingPairCache::~OverlappingPairCache()
 {
-	delete [] m_OverlappingPairs;
+	//todo/test: show we erase/delete data, or is it automatic
 }
 
 
 void	OverlappingPairCache::RemoveOverlappingPair(BroadphasePair& pair)
 {
 	CleanOverlappingPair(pair);
-	int	index = &pair - &m_OverlappingPairs[0];
-	//remove efficiently, swap with the last
-	m_OverlappingPairs[index] = m_OverlappingPairs[m_NumOverlapBroadphasePair-1];
-	m_NumOverlapBroadphasePair--;
+	std::set<BroadphasePair>::iterator it = m_overlappingPairSet.find(pair);
+	assert(it != m_overlappingPairSet.end());
+	m_overlappingPairSet.erase(pair);
 }
 
 
@@ -75,25 +73,22 @@ void	OverlappingPairCache::AddOverlappingPair(BroadphaseProxy* proxy0,Broadphase
 
 
 	BroadphasePair pair(*proxy0,*proxy1);
-	m_OverlappingPairs[m_NumOverlapBroadphasePair] = pair;
-
+	
+	m_overlappingPairSet.insert(pair);
+	
+	
+#ifdef _DEBUG
+	/*
+	BroadphasePair& pr = (*newElem);
 	int i;
 	for (i=0;i<SIMPLE_MAX_ALGORITHMS;i++)
 	{
 		assert(!m_OverlappingPairs[m_NumOverlapBroadphasePair].m_algorithms[i]);
-		m_OverlappingPairs[m_NumOverlapBroadphasePair].m_algorithms[i] = 0;
+		//m_OverlappingPairs[m_NumOverlapBroadphasePair].m_algorithms[i] = 0;
 	}
+	*/
 
-	if (m_NumOverlapBroadphasePair >= m_maxOverlap)
-	{
-		//printf("Error: too many overlapping objects: m_NumOverlapBroadphasePair: %d\n",m_NumOverlapBroadphasePair);
-#ifdef DEBUG
-		assert(0);
-#endif 
-	} else
-	{
-		m_NumOverlapBroadphasePair++;
-	}
+#endif _DEBUG
 
 
 }
@@ -104,27 +99,25 @@ void	OverlappingPairCache::AddOverlappingPair(BroadphaseProxy* proxy0,Broadphase
 ///Also we can use a 2D bitmap, which can be useful for a future GPU implementation
 BroadphasePair*	OverlappingPairCache::FindPair(BroadphaseProxy* proxy0,BroadphaseProxy* proxy1)
 {
-	BroadphasePair* foundPair = 0;
+	if (!NeedsCollision(proxy0,proxy1))
+		return 0;
 
-	int i;
-	for (i=m_NumOverlapBroadphasePair-1;i>=0;i--)
-	{
-		BroadphasePair& pair = m_OverlappingPairs[i];
-		if (((pair.m_pProxy0 == proxy0) && (pair.m_pProxy1 == proxy1)) ||
-			((pair.m_pProxy0 == proxy1) && (pair.m_pProxy1 == proxy0)))
-		{
-			foundPair = &pair;
-			return foundPair;
-		}
-	}	
+	BroadphasePair tmpPair(*proxy0,*proxy1);
+	std::set<BroadphasePair>::iterator it = m_overlappingPairSet.find(tmpPair);
+	if ((it == m_overlappingPairSet.end()))
+		return 0;
 
-	return foundPair;
+	//assert(it != m_overlappingPairSet.end());
+	BroadphasePair* pair = &(*it);
+	return pair;
 }
 
 
 
 void	OverlappingPairCache::CleanProxyFromPairs(BroadphaseProxy* proxy)
 {
+	assert(0);
+	/*
 	for (int i=0;i<m_NumOverlapBroadphasePair;i++)
 	{
 		BroadphasePair& pair = m_OverlappingPairs[i];
@@ -134,11 +127,17 @@ void	OverlappingPairCache::CleanProxyFromPairs(BroadphaseProxy* proxy)
 			CleanOverlappingPair(pair);
 		}
 	}
+	*/
+
 }
 
 void	OverlappingPairCache::RemoveOverlappingPairsContainingProxy(BroadphaseProxy* proxy)
 {
+
+	assert(0);
+	/*
 	int i;
+	
 	for ( i=m_NumOverlapBroadphasePair-1;i>=0;i--)
 	{
 		BroadphasePair& pair = m_OverlappingPairs[i];
@@ -148,7 +147,23 @@ void	OverlappingPairCache::RemoveOverlappingPairsContainingProxy(BroadphaseProxy
 			RemoveOverlappingPair(pair);
 		}
 	}
+	*/
+
 }
 
 
 
+void	OverlappingPairCache::ProcessAllOverlappingPairs(OverlapCallback* callback)
+{
+	std::set<BroadphasePair>::iterator it = m_overlappingPairSet.begin();
+	for (; !(it==m_overlappingPairSet.end());it++)
+	{
+		BroadphasePair& pair = (*it);
+		if (callback->ProcessOverlap(pair))
+		{
+			assert(0);
+			m_overlappingPairSet.erase(it);
+		}
+	}
+}
+
diff --git a/Bullet/BroadphaseCollision/OverlappingPairCache.h b/Bullet/BroadphaseCollision/OverlappingPairCache.h
index d19d579f1..22b1ca37a 100644
--- a/Bullet/BroadphaseCollision/OverlappingPairCache.h
+++ b/Bullet/BroadphaseCollision/OverlappingPairCache.h
@@ -21,35 +21,33 @@ subject to the following restrictions:
 #include "BroadphaseInterface.h"
 #include "BroadphaseProxy.h"
 #include "SimdPoint3.h"
+#include <set>
 
 
+struct	OverlapCallback
+{
+	//return true for deletion of the pair
+	virtual bool	ProcessOverlap(BroadphasePair& pair) = 0;
+};
+
 ///OverlappingPairCache maintains the objects with overlapping AABB
 ///Typically managed by the Broadphase, Axis3Sweep or SimpleBroadphase
 class	OverlappingPairCache : public BroadphaseInterface
 {
-
-	BroadphasePair*	m_OverlappingPairs;
-	int	m_NumOverlapBroadphasePair;
+	//avoid brute-force finding all the time
+	std::set<BroadphasePair>	m_overlappingPairSet;
+	
 	int m_maxOverlap;
 	
 	//during the dispatch, check that user doesn't destroy/create proxy
 	bool		m_blockedForChanges;
-
 	
 	public:
 		
 	OverlappingPairCache(int maxOverlap);	
 	virtual ~OverlappingPairCache();
-	
-	int		GetNumOverlappingPairs() const
-	{
-		return m_NumOverlapBroadphasePair;
-	}
 
-	BroadphasePair& GetOverlappingPair(int index)
-	{
-		return m_OverlappingPairs[index];
-	}
+	void	ProcessAllOverlappingPairs(OverlapCallback*);
 
 	void	RemoveOverlappingPair(BroadphasePair& pair);
 
@@ -58,8 +56,7 @@ class	OverlappingPairCache : public BroadphaseInterface
 	void	AddOverlappingPair(BroadphaseProxy* proxy0,BroadphaseProxy* proxy1);
 
 	BroadphasePair*	FindPair(BroadphaseProxy* proxy0,BroadphaseProxy* proxy1);
-	
-	
+		
 	
 	void	CleanProxyFromPairs(BroadphaseProxy* proxy);
 
diff --git a/Bullet/BroadphaseCollision/SimpleBroadphase.cpp b/Bullet/BroadphaseCollision/SimpleBroadphase.cpp
index 5a648e1d3..f00b71776 100644
--- a/Bullet/BroadphaseCollision/SimpleBroadphase.cpp
+++ b/Bullet/BroadphaseCollision/SimpleBroadphase.cpp
@@ -95,7 +95,27 @@ BroadphaseProxy*	SimpleBroadphase::CreateProxy(  const SimdVector3& min,  const
 	return proxy;
 }
 
+class	RemovingOverlapCallback : public OverlapCallback
+{
+	virtual bool	ProcessOverlap(BroadphasePair& pair)
+	{
+		assert(0);
+	}
+};
 
+class RemovePairContainingProxy
+{
+
+	BroadphaseProxy*	m_targetProxy;
+
+	virtual bool ProcessOverlap(BroadphasePair& pair)
+	{
+		SimpleBroadphaseProxy* proxy0 = static_cast<SimpleBroadphaseProxy*>(pair.m_pProxy0);
+		SimpleBroadphaseProxy* proxy1 = static_cast<SimpleBroadphaseProxy*>(pair.m_pProxy1);
+
+		return ((m_targetProxy == proxy0 || m_targetProxy == proxy1));
+	};
+};
 
 void	SimpleBroadphase::DestroyProxy(BroadphaseProxy* proxyOrg)
 {
@@ -109,7 +129,24 @@ void	SimpleBroadphase::DestroyProxy(BroadphaseProxy* proxyOrg)
 		assert (index < m_maxProxies);
 		m_freeProxies[--m_firstFreeProxy] = index;
 
-		RemoveOverlappingPairsContainingProxy(proxyOrg);
+		//RemoveOverlappingPairsContainingProxy(proxyOrg);
+
+		assert(0);
+		//then remove non-overlapping ones
+		/*for (i=0;i<GetNumOverlappingPairs();i++)
+		{
+			BroadphasePair& pair = GetOverlappingPair(i);
+
+			SimpleBroadphaseProxy* proxy0 = GetSimpleProxyFromProxy(pair.m_pProxy0);
+			SimpleBroadphaseProxy* proxy1 = GetSimpleProxyFromProxy(pair.m_pProxy1);
+			if ((proxy0==proxyOrg) || (proxy1==proxyOrg))
+			{
+				RemoveOverlappingPair(pair);
+			}
+		}
+		*/
+
+
 
 		
 		for (i=0;i<m_numProxies;i++)
@@ -172,6 +209,8 @@ void	SimpleBroadphase::RefreshOverlappingPairs()
 		}
 	}
 
+	assert(0);
+	/*
 	//then remove non-overlapping ones
 	for (i=0;i<GetNumOverlappingPairs();i++)
 	{
@@ -185,7 +224,8 @@ void	SimpleBroadphase::RefreshOverlappingPairs()
 		}
 	}
 
-	
+	*/
+
 
 }
 
diff --git a/Bullet/CollisionDispatch/CollisionDispatcher.cpp b/Bullet/CollisionDispatch/CollisionDispatcher.cpp
index 62f555c60..93d590c1b 100644
--- a/Bullet/CollisionDispatch/CollisionDispatcher.cpp
+++ b/Bullet/CollisionDispatch/CollisionDispatcher.cpp
@@ -265,7 +265,72 @@ void	CollisionDispatcher::ReleaseManifoldResult(ManifoldResult*)
 }
 
 
-void	CollisionDispatcher::DispatchAllCollisionPairs(BroadphasePair* pairs,int numPairs,DispatcherInfo& dispatchInfo)
+class CollisionPairCallback : public OverlapCallback
+{
+	DispatcherInfo& m_dispatchInfo;
+	CollisionDispatcher*	m_dispatcher;
+	int		m_dispatcherId;
+public:
+
+	CollisionPairCallback(DispatcherInfo& dispatchInfo,CollisionDispatcher*	dispatcher,int		dispatcherId)
+	:m_dispatchInfo(dispatchInfo),
+	m_dispatcher(dispatcher),
+	m_dispatcherId(dispatcherId)
+	{
+	}
+
+	virtual bool	ProcessOverlap(BroadphasePair& pair)
+	{
+		if (m_dispatcherId>= 0)
+		{
+			//dispatcher will keep algorithms persistent in the collision pair
+			if (!pair.m_algorithms[m_dispatcherId])
+			{
+				pair.m_algorithms[m_dispatcherId] = m_dispatcher->FindAlgorithm(
+					*pair.m_pProxy0,
+					*pair.m_pProxy1);
+			}
+
+			if (pair.m_algorithms[m_dispatcherId])
+			{
+				if (m_dispatchInfo.m_dispatchFunc == 		DispatcherInfo::DISPATCH_DISCRETE)
+				{
+					pair.m_algorithms[m_dispatcherId]->ProcessCollision(pair.m_pProxy0,pair.m_pProxy1,m_dispatchInfo);
+				} else
+				{
+					float toi = pair.m_algorithms[m_dispatcherId]->CalculateTimeOfImpact(pair.m_pProxy0,pair.m_pProxy1,m_dispatchInfo);
+					if (m_dispatchInfo.m_timeOfImpact > toi)
+						m_dispatchInfo.m_timeOfImpact = toi;
+
+				}
+			}
+		} else
+		{
+			//non-persistent algorithm dispatcher
+			CollisionAlgorithm* algo = m_dispatcher->FindAlgorithm(
+				*pair.m_pProxy0,
+				*pair.m_pProxy1);
+
+			if (algo)
+			{
+				if (m_dispatchInfo.m_dispatchFunc == 		DispatcherInfo::DISPATCH_DISCRETE)
+				{
+					algo->ProcessCollision(pair.m_pProxy0,pair.m_pProxy1,m_dispatchInfo);
+				} else
+				{
+					float toi = algo->CalculateTimeOfImpact(pair.m_pProxy0,pair.m_pProxy1,m_dispatchInfo);
+					if (m_dispatchInfo.m_timeOfImpact > toi)
+						m_dispatchInfo.m_timeOfImpact = toi;
+				}
+			}
+		}
+		return false;
+
+	}
+};
+
+
+void	CollisionDispatcher::DispatchAllCollisionPairs(OverlappingPairCache* pairCache,DispatcherInfo& dispatchInfo)
 {
 	//m_blockedForChanges = true;
 
@@ -273,58 +338,9 @@ void	CollisionDispatcher::DispatchAllCollisionPairs(BroadphasePair* pairs,int nu
 
 	int dispatcherId = GetUniqueId();
 
-	
+	CollisionPairCallback	collisionCallback(dispatchInfo,this,dispatcherId);
 
-	for (i=0;i<numPairs;i++)
-	{
-
-		BroadphasePair& pair = pairs[i];
-
-		if (dispatcherId>= 0)
-		{
-			//dispatcher will keep algorithms persistent in the collision pair
-			if (!pair.m_algorithms[dispatcherId])
-			{
-				pair.m_algorithms[dispatcherId] = FindAlgorithm(
-					*pair.m_pProxy0,
-					*pair.m_pProxy1);
-			}
-
-			if (pair.m_algorithms[dispatcherId])
-			{
-				if (dispatchInfo.m_dispatchFunc == 		DispatcherInfo::DISPATCH_DISCRETE)
-				{
-					pair.m_algorithms[dispatcherId]->ProcessCollision(pair.m_pProxy0,pair.m_pProxy1,dispatchInfo);
-				} else
-				{
-					float toi = pair.m_algorithms[dispatcherId]->CalculateTimeOfImpact(pair.m_pProxy0,pair.m_pProxy1,dispatchInfo);
-					if (dispatchInfo.m_timeOfImpact > toi)
-						dispatchInfo.m_timeOfImpact = toi;
-
-				}
-			}
-		} else
-		{
-			//non-persistent algorithm dispatcher
-			CollisionAlgorithm* algo = FindAlgorithm(
-				*pair.m_pProxy0,
-				*pair.m_pProxy1);
-
-			if (algo)
-			{
-				if (dispatchInfo.m_dispatchFunc == 		DispatcherInfo::DISPATCH_DISCRETE)
-				{
-					algo->ProcessCollision(pair.m_pProxy0,pair.m_pProxy1,dispatchInfo);
-				} else
-				{
-					float toi = algo->CalculateTimeOfImpact(pair.m_pProxy0,pair.m_pProxy1,dispatchInfo);
-					if (dispatchInfo.m_timeOfImpact > toi)
-						dispatchInfo.m_timeOfImpact = toi;
-				}
-			}
-		}
-
-	}
+	pairCache->ProcessAllOverlappingPairs(&collisionCallback);
 
 	//m_blockedForChanges = false;
 
diff --git a/Bullet/CollisionDispatch/CollisionDispatcher.h b/Bullet/CollisionDispatch/CollisionDispatcher.h
index 85679b95f..c2a85cf4d 100644
--- a/Bullet/CollisionDispatch/CollisionDispatcher.h
+++ b/Bullet/CollisionDispatch/CollisionDispatcher.h
@@ -105,7 +105,7 @@ public:
 
 	virtual int GetUniqueId() { return RIGIDBODY_DISPATCHER;}
 	
-	virtual void	DispatchAllCollisionPairs(BroadphasePair* pairs,int numPairs,DispatcherInfo& dispatchInfo);
+	virtual void	DispatchAllCollisionPairs(OverlappingPairCache* pairCache,DispatcherInfo& dispatchInfo);
 
 	
 
diff --git a/Bullet/CollisionDispatch/CollisionWorld.cpp b/Bullet/CollisionDispatch/CollisionWorld.cpp
index f1d687777..e750e29b5 100644
--- a/Bullet/CollisionDispatch/CollisionWorld.cpp
+++ b/Bullet/CollisionDispatch/CollisionWorld.cpp
@@ -108,7 +108,7 @@ void	CollisionWorld::PerformDiscreteCollisionDetection()
 
 	Dispatcher* dispatcher = GetDispatcher();
 	if (dispatcher)
-		dispatcher->DispatchAllCollisionPairs(&m_pairCache->GetOverlappingPair(0),m_pairCache->GetNumOverlappingPairs(),dispatchInfo);
+		dispatcher->DispatchAllCollisionPairs(m_pairCache,dispatchInfo);
 
 }
 
diff --git a/Demos/UserCollisionAlgorithm/SphereSphereCollisionAlgorithm.cpp b/Bullet/CollisionDispatch/SphereSphereCollisionAlgorithm.cpp
similarity index 100%
rename from Demos/UserCollisionAlgorithm/SphereSphereCollisionAlgorithm.cpp
rename to Bullet/CollisionDispatch/SphereSphereCollisionAlgorithm.cpp
diff --git a/Demos/UserCollisionAlgorithm/SphereSphereCollisionAlgorithm.h b/Bullet/CollisionDispatch/SphereSphereCollisionAlgorithm.h
similarity index 100%
rename from Demos/UserCollisionAlgorithm/SphereSphereCollisionAlgorithm.h
rename to Bullet/CollisionDispatch/SphereSphereCollisionAlgorithm.h
diff --git a/Demos/CcdPhysicsDemo/CcdPhysicsDemo.cpp b/Demos/CcdPhysicsDemo/CcdPhysicsDemo.cpp
index cb293a48c..dc45b6ba9 100644
--- a/Demos/CcdPhysicsDemo/CcdPhysicsDemo.cpp
+++ b/Demos/CcdPhysicsDemo/CcdPhysicsDemo.cpp
@@ -48,7 +48,7 @@ subject to the following restrictions:
 #include "IDebugDraw.h"
 
 #include "GLDebugDrawer.h"
-
+#include "CollisionDispatch/SphereSphereCollisionAlgorithm.h"
 
 
 
@@ -72,13 +72,13 @@ const int maxProxies = 32766;
 const int maxOverlap = 65535;
 
 bool createConstraint = true;//false;
-bool useCompound = true;//false;
+bool useCompound = false;//true;//false;
 
 
 #ifdef _DEBUG
-const int numObjects = 50;
+const int numObjects = 250;
 #else
-const int numObjects = 120;
+const int numObjects = 3250;
 #endif
 
 const int maxNumObjects = 32760;
@@ -255,6 +255,8 @@ void	CcdPhysicsDemo::initPhysics()
 	SimdVector3 worldAabbMax(30000,30000,30000);
 
 	OverlappingPairCache* broadphase = new AxisSweep3(worldAabbMin,worldAabbMax,maxProxies,maxOverlap);
+	dispatcher->RegisterCollisionCreateFunc(SPHERE_SHAPE_PROXYTYPE,SPHERE_SHAPE_PROXYTYPE,new SphereSphereCollisionAlgorithm::CreateFunc);
+	
 	//OverlappingPairCache* broadphase = new SimpleBroadphase(maxProxies,maxOverlap);
 
 #ifdef USE_PARALLEL_DISPATCHER
diff --git a/Demos/OpenGL/DemoApplication.cpp b/Demos/OpenGL/DemoApplication.cpp
index cb9ec12c0..4cc8fa221 100644
--- a/Demos/OpenGL/DemoApplication.cpp
+++ b/Demos/OpenGL/DemoApplication.cpp
@@ -27,7 +27,7 @@ subject to the following restrictions:
 #include "BMF_Api.h"
 
 int numObjects = 0;
-const int maxNumObjects = 450;
+const int maxNumObjects = 4096;
 DefaultMotionState ms[maxNumObjects];
 CcdPhysicsController* physObjects[maxNumObjects];
 SimdTransform startTransforms[maxNumObjects];
@@ -335,6 +335,9 @@ void DemoApplication::keyboardCallback(unsigned char key, int x, int y)
         break;
     }
 
+	if (m_physicsEnvironmentPtr)
+		m_physicsEnvironmentPtr->setDebugMode(m_debugMode);
+
 	glutPostRedisplay();
 
 }
diff --git a/Demos/UserCollisionAlgorithm/UserCollisionAlgorithm.cpp b/Demos/UserCollisionAlgorithm/UserCollisionAlgorithm.cpp
index 06985eae6..b9460fddc 100644
--- a/Demos/UserCollisionAlgorithm/UserCollisionAlgorithm.cpp
+++ b/Demos/UserCollisionAlgorithm/UserCollisionAlgorithm.cpp
@@ -27,6 +27,8 @@ subject to the following restrictions:
 #include "ConstraintSolver/SequentialImpulseConstraintSolver.h"
 #include "CollisionDispatch/CollisionDispatcher.h"
 #include "BroadphaseCollision/SimpleBroadphase.h"
+#include "BroadphaseCollision/AxisSweep3.h"
+
 #include "CollisionShapes/TriangleMeshShape.h"
 #include "CollisionShapes/TriangleIndexVertexArray.h"
 #include "CollisionShapes/BvhTriangleMeshShape.h"
@@ -39,7 +41,7 @@ subject to the following restrictions:
 #include "GlutStuff.h"
 
 //The user defined collision algorithm
-#include "SphereSphereCollisionAlgorithm.h"
+#include "CollisionDispatch/SphereSphereCollisionAlgorithm.h"
 
 GLDebugDrawer	debugDrawer;
 
@@ -139,7 +141,8 @@ void	UserCollisionAlgorithm::initPhysics()
 	
 	CollisionDispatcher* dispatcher = new	CollisionDispatcher();
 		
-	OverlappingPairCache* broadphase = new SimpleBroadphase();
+	SimdVector3 maxAabb(10000,10000,10000);
+	OverlappingPairCache* broadphase = new AxisSweep3(-maxAabb,maxAabb);//SimpleBroadphase();
 	dispatcher->RegisterCollisionCreateFunc(SPHERE_SHAPE_PROXYTYPE,SPHERE_SHAPE_PROXYTYPE,new SphereSphereCollisionAlgorithm::CreateFunc);
 	
 
diff --git a/Doxyfile b/Doxyfile
index 50bef14e5..2ce47d3cf 100644
--- a/Doxyfile
+++ b/Doxyfile
@@ -267,7 +267,7 @@ WARN_LOGFILE           =
 # directories like "/usr/src/myproject". Separate the files or directories 
 # with spaces. 
 
-INPUT                  = Bullet BulletDynamics LinearMath Extras/PhysicsInterface
+INPUT                  = Bullet BulletDynamics LinearMath Extras/PhysicsInterface Demos
 
 
 # If the value of the INPUT tag contains directories, you can use the 
diff --git a/Extras/PhysicsInterface/CcdPhysics/CcdPhysicsEnvironment.cpp b/Extras/PhysicsInterface/CcdPhysics/CcdPhysicsEnvironment.cpp
index b529b30fe..83c79e7f7 100644
--- a/Extras/PhysicsInterface/CcdPhysics/CcdPhysicsEnvironment.cpp
+++ b/Extras/PhysicsInterface/CcdPhysics/CcdPhysicsEnvironment.cpp
@@ -687,7 +687,8 @@ bool	CcdPhysicsEnvironment::proceedDeltaTimeOneStep(float timeStep)
 	dispatchInfo.m_debugDraw = this->m_debugDrawer;
 
 	scene->RefreshOverlappingPairs();
-	GetCollisionWorld()->GetDispatcher()->DispatchAllCollisionPairs(&scene->GetOverlappingPair(0),scene->GetNumOverlappingPairs(),dispatchInfo);
+	
+	GetCollisionWorld()->GetDispatcher()->DispatchAllCollisionPairs(scene,dispatchInfo);
 
 
 #ifdef USE_QUICKPROF
@@ -856,7 +857,7 @@ bool	CcdPhysicsEnvironment::proceedDeltaTimeOneStep(float timeStep)
 				dispatchInfo.m_dispatchFunc = DispatcherInfo::DISPATCH_CONTINUOUS;
 
 				//pairCache->RefreshOverlappingPairs();//??
-				GetCollisionWorld()->GetDispatcher()->DispatchAllCollisionPairs(&scene->GetOverlappingPair(0),scene->GetNumOverlappingPairs(),dispatchInfo);
+				GetCollisionWorld()->GetDispatcher()->DispatchAllCollisionPairs(scene,dispatchInfo);
 				
 				toi = dispatchInfo.m_timeOfImpact;
 
diff --git a/Extras/PhysicsInterface/CcdPhysics/ParallelIslandDispatcher.cpp b/Extras/PhysicsInterface/CcdPhysics/ParallelIslandDispatcher.cpp
index 89a8af1fd..57d004869 100644
--- a/Extras/PhysicsInterface/CcdPhysics/ParallelIslandDispatcher.cpp
+++ b/Extras/PhysicsInterface/CcdPhysics/ParallelIslandDispatcher.cpp
@@ -282,7 +282,7 @@ void	ParallelIslandDispatcher::ReleaseManifoldResult(ManifoldResult*)
 }
 
 
-void	ParallelIslandDispatcher::DispatchAllCollisionPairs(BroadphasePair* pairs,int numPairs,DispatcherInfo& dispatchInfo)
+void	ParallelIslandDispatcher::DispatchAllCollisionPairs(OverlappingPairCache* pairCache,DispatcherInfo& dispatchInfo)
 {
 	//m_blockedForChanges = true;
 
@@ -290,8 +290,9 @@ void	ParallelIslandDispatcher::DispatchAllCollisionPairs(BroadphasePair* pairs,i
 
 	int dispatcherId = GetUniqueId();
 
+	assert(0);
 	
-
+/*
 	for (i=0;i<numPairs;i++)
 	{
 
@@ -342,7 +343,7 @@ void	ParallelIslandDispatcher::DispatchAllCollisionPairs(BroadphasePair* pairs,i
 		}
 
 	}
-
+*/
 	//m_blockedForChanges = false;
 
 }
diff --git a/Extras/PhysicsInterface/CcdPhysics/ParallelIslandDispatcher.h b/Extras/PhysicsInterface/CcdPhysics/ParallelIslandDispatcher.h
index bdee39cdc..2a47ff619 100644
--- a/Extras/PhysicsInterface/CcdPhysics/ParallelIslandDispatcher.h
+++ b/Extras/PhysicsInterface/CcdPhysics/ParallelIslandDispatcher.h
@@ -118,7 +118,7 @@ public:
 
 	virtual int GetUniqueId() { return RIGIDBODY_DISPATCHER;}
 
-	virtual void	DispatchAllCollisionPairs(BroadphasePair* pairs,int numPairs,DispatcherInfo& dispatchInfo);
+	virtual void	DispatchAllCollisionPairs(OverlappingPairCache* pairCache,DispatcherInfo& dispatchInfo);
 	
 	
 
diff --git a/Extras/PhysicsInterface/CcdPhysics/ParallelPhysicsEnvironment.cpp b/Extras/PhysicsInterface/CcdPhysics/ParallelPhysicsEnvironment.cpp
index 602396c0f..b645ff3be 100644
--- a/Extras/PhysicsInterface/CcdPhysics/ParallelPhysicsEnvironment.cpp
+++ b/Extras/PhysicsInterface/CcdPhysics/ParallelPhysicsEnvironment.cpp
@@ -110,8 +110,13 @@ bool	ParallelPhysicsEnvironment::proceedDeltaTimeOneStep(float timeStep)
 	
 	//this is a brute force approach, will rethink later about more subtle ways
 	int i;
+
+			assert(0);
+/*
 	for (i=0;i<	scene->GetNumOverlappingPairs();i++)
 	{
+
+
 		BroadphasePair* pair = &scene->GetOverlappingPair(i);
 
 		CollisionObject*	col0 = static_cast<CollisionObject*>(pair->m_pProxy0->m_clientObject);
@@ -124,7 +129,9 @@ bool	ParallelPhysicsEnvironment::proceedDeltaTimeOneStep(float timeStep)
 		{
 			simulationIslands[col1->m_islandTag1].m_overlappingPairIndices.push_back(i);
 		}
+		
 	}
+	*/
 	
 	//store constraint indices for each island
 	for (i=0;i<m_constraints.size();i++)
@@ -176,14 +183,19 @@ bool	ParallelPhysicsEnvironment::proceedDeltaTimeOneStep(float timeStep)
 
 
 
+	assert(0);
+	/*
 	//Each simulation island can be processed in parallel (will be put on a job queue)
 	for (k=0;k<simulationIslands.size();k++)
 	{
 		if (simulationIslands[k].m_controllers.size())
 		{
+			assert(0);//seems to be wrong, passing ALL overlapping pairs
 			simulationIslands[k].Simulate(m_debugDrawer,m_numIterations, constraintBase ,&scene->GetOverlappingPair(0),dispatcher,GetBroadphase(),m_solver,timeStep);
 		}
 	}
+	*/
+
 
 #ifdef USE_QUICKPROF
 	Profiler::endBlock("SimulateIsland");
diff --git a/Extras/PhysicsInterface/CcdPhysics/SimulationIsland.cpp b/Extras/PhysicsInterface/CcdPhysics/SimulationIsland.cpp
index 60a35e985..eb3b7f235 100644
--- a/Extras/PhysicsInterface/CcdPhysics/SimulationIsland.cpp
+++ b/Extras/PhysicsInterface/CcdPhysics/SimulationIsland.cpp
@@ -102,7 +102,8 @@ bool	SimulationIsland::Simulate(IDebugDraw* debugDrawer,int numSolverIterations,
 	//pairCache->RefreshOverlappingPairs();
 	if (overlappingPairs.size())
 	{
-		dispatcher->DispatchAllCollisionPairs(&overlappingPairs[0],overlappingPairs.size(),dispatchInfo);///numsubstep,g);
+		assert(0);
+		//dispatcher->DispatchAllCollisionPairs(&overlappingPairs[0],overlappingPairs.size(),dispatchInfo);///numsubstep,g);
 	}
 
 	//scatter overlapping pair info, mainly the created algorithms/contact caches
diff --git a/LinearMath/SimdTransform.h b/LinearMath/SimdTransform.h
index d2d9b53f9..621dc854c 100644
--- a/LinearMath/SimdTransform.h
+++ b/LinearMath/SimdTransform.h
@@ -23,6 +23,10 @@ subject to the following restrictions:
 
 
 class SimdTransform {
+	
+
+public:
+	
 	enum { 
 		TRANSLATION = 0x01,
 		ROTATION    = 0x02,
@@ -32,12 +36,8 @@ class SimdTransform {
 		AFFINE      = TRANSLATION | LINEAR
 	};
 
-public:
 	SimdTransform() {}
 
-	//	template <typename Scalar2>
-	//	explicit Transform(const Scalar2 *m) { setValue(m); }
-
 	explicit SIMD_FORCE_INLINE SimdTransform(const SimdQuaternion& q, 
 		const SimdVector3& c = SimdVector3(SimdScalar(0), SimdScalar(0), SimdScalar(0))) 
 		: m_basis(q),