diff --git a/makefile b/makefile
index 2a149fda60..5cd3d5d06d 100644
--- a/makefile
+++ b/makefile
@@ -105,13 +105,13 @@ ifeq ($(platform),windows)
 endif
 
 ifeq ($(mode),debug)
-	cflags += -O0 -g3
+	cflags += -O0 -g3 -DNDEBUG
 endif
 ifeq ($(mode),stress)
-	cflags += -O0 -g3 -DVM_STRESS
+	cflags += -O0 -g3 -DNDEBUG -DVM_STRESS
 endif
 ifeq ($(mode),stress-major)
-	cflags += -O0 -g3 -DVM_STRESS -DVM_STRESS_MAJOR
+	cflags += -O0 -g3 -DNDEBUG -DVM_STRESS -DVM_STRESS_MAJOR
 endif
 ifeq ($(mode),fast)
 	cflags += -O3 -g3 -DNDEBUG
diff --git a/src/builtin.cpp b/src/builtin.cpp
index 6ce717ba02..c96218d64c 100644
--- a/src/builtin.cpp
+++ b/src/builtin.cpp
@@ -108,7 +108,7 @@ Java_java_lang_Object_clone(Thread* t, jclass, jobject o)
     // clear any object header flags:
     setObjectClass(t, *o, objectClass(t, *o));
   } else {
-    clone = make(t, objectClass(t, *o));
+    clone = make(t, class_);
     memcpy(reinterpret_cast<void**>(clone) + 1,
            reinterpret_cast<void**>(*o) + 1,
            size - BytesPerWord);
diff --git a/src/compile2.cpp b/src/compile2.cpp
index ffb87c1e34..0892c380ea 100644
--- a/src/compile2.cpp
+++ b/src/compile2.cpp
@@ -2749,8 +2749,8 @@ finish(MyThread* t, Compiler* c, object method, Buffer* objectPool,
       traceLog->get(i + BytesPerWord + 2 + 1 + 4, map, mapSize);
 
       object node = makeTraceNode
-        (t, reinterpret_cast<intptr_t>(start + offset->value()), 0, 0, method,
-         target, virtualCall, mapSize / BytesPerWord, false);
+        (t, reinterpret_cast<intptr_t>(start + offset->value()), 0, 0, 0,
+         method, target, virtualCall, mapSize / BytesPerWord, false);
 
       if (mapSize) {
         memcpy(&traceNodeMap(t, node, 0), map, mapSize);
@@ -3328,6 +3328,7 @@ class MyProcessor: public Processor {
     defaultCompiled(0),
     nativeCompiled(0),
     addressTree(0),
+    addressTreeSentinal(0),
     indirectCaller(0)
   { }
 
@@ -3587,13 +3588,41 @@ class MyProcessor: public Processor {
   object defaultCompiled;
   object nativeCompiled;
   object addressTree;
+  object addressTreeSentinal;
   void* indirectCaller;
 };
 
+MyProcessor*
+processor(MyThread* t)
+{
+  MyProcessor* p = static_cast<MyProcessor*>(t->m->processor);
+  if (p->addressTree == 0) {
+    ACQUIRE(t, t->m->classLock);
+
+    if (p->addressTree == 0) {
+      p->addressTreeSentinal = makeTraceNode(t, 0, 0, 0, 0, 0, 0, 0, 0, false);
+      set(t, p->addressTreeSentinal, TreeNodeLeft, p->addressTreeSentinal);
+      set(t, p->addressTreeSentinal, TreeNodeRight, p->addressTreeSentinal);
+      p->addressTree = p->addressTreeSentinal;
+
+      Compiler* c = makeCompiler(t->m->system, 0);
+
+      c->mov(c->base(), c->memory(c->thread(), difference(&(t->base), t)));
+      c->mov(c->stack(), c->memory(c->thread(), difference(&(t->stack), t)));
+
+      c->jmp(c->indirectTarget());
+
+      p->indirectCaller = t->m->system->allocate(c->size());
+      c->writeTo(p->indirectCaller);
+    }
+  }
+  return p;
+}
+
 void
 compile(MyThread* t, object method)
 {
-  MyProcessor* p = static_cast<MyProcessor*>(t->m->processor);
+  MyProcessor* p = processor(t);
 
   object stub = p->getDefaultCompiled(t);
 
@@ -3603,18 +3632,6 @@ compile(MyThread* t, object method)
     ACQUIRE(t, t->m->classLock);
     
     if (methodCompiled(t, method) == stub) {
-      if (p->indirectCaller == 0) {
-        Compiler* c = makeCompiler(t->m->system, 0);
-
-        c->mov(c->base(), c->memory(c->thread(), difference(&(t->base), t)));
-        c->mov(c->stack(), c->memory(c->thread(), difference(&(t->stack), t)));
-
-        c->jmp(c->indirectTarget());
-
-        p->indirectCaller = t->m->system->allocate(c->size());
-        c->writeTo(p->indirectCaller);
-      }
-
       PROTECT(t, method);
 
       Compiler* c = makeCompiler(t->m->system, p->indirectCaller);
@@ -3630,15 +3647,17 @@ compile(MyThread* t, object method)
 object
 findTraceNode(MyThread* t, void* address)
 {
-  MyProcessor* p = static_cast<MyProcessor*>(t->m->processor);
-  return treeQuery(t, p->addressTree, reinterpret_cast<intptr_t>(address));
+  MyProcessor* p = processor(t);
+  return treeQuery(t, p->addressTree, reinterpret_cast<intptr_t>(address),
+                   p->addressTreeSentinal);
 }
 
 void
 insertTraceNode(MyThread* t, object node)
 {
-  MyProcessor* p = static_cast<MyProcessor*>(t->m->processor);
-  p->addressTree = treeInsert(t, p->addressTree, node);
+  MyProcessor* p = processor(t);
+  p->addressTree = treeInsert
+    (t, p->addressTree, node, p->addressTreeSentinal);
 }
 
 } // namespace
diff --git a/src/types.def b/src/types.def
index 831c5deab8..f982b33044 100644
--- a/src/types.def
+++ b/src/types.def
@@ -93,14 +93,21 @@
 
 (type treeNode
   (intptr_t key)
+  (uintptr_t red)
   (object left)
   (object right))
  
+(type treePath
+  (uintptr_t fresh)
+  (object node)
+  (object root)
+  (object ancestors))
+
 (type traceNode
   (extends treeNode)
   (object method)
   (object target)
-  (uint8_t virtualCall)
+  (uintptr_t virtualCall)
   (array uintptr_t map))
 
 (type array
diff --git a/src/util.cpp b/src/util.cpp
index a1d932344d..dd40d3e7b5 100644
--- a/src/util.cpp
+++ b/src/util.cpp
@@ -1,5 +1,208 @@
 #include "util.h"
 
+using namespace vm;
+
+namespace {
+
+object
+clone(Thread* t, object o)
+{
+  object class_ = objectClass(t, o);
+  unsigned size = baseSize(t, o, class_) * BytesPerWord;
+
+  object clone = make(t, class_);
+  memcpy(reinterpret_cast<void**>(clone) + 1,
+         reinterpret_cast<void**>(o) + 1,
+         size - BytesPerWord);
+
+  return clone;
+}
+
+object
+treeFind(Thread* t, object oldRoot, object node, object sentinal)
+{
+  object newRoot = clone(t, oldRoot);
+  object ancestors = 0;
+
+  object old = oldRoot;
+  object new_ = newRoot;
+  while (old != sentinal) {
+    ancestors = makePair(t, new_, ancestors);
+
+    intptr_t difference = treeNodeKey(t, node) - treeNodeKey(t, old);
+    if (difference < 0) {
+      old = treeNodeLeft(t, old);
+      object n = clone(t, old);
+      set(t, new_, TreeNodeLeft, n);
+      new_ = n;
+    } else if (difference > 0) {
+      old = treeNodeRight(t, old);
+      object n = clone(t, old);
+      set(t, new_, TreeNodeRight, n);
+      new_ = n;
+    } else {
+      return makeTreePath(t, false, new_, newRoot, pairSecond(t, ancestors));
+    }
+  }
+
+  object class_ = objectClass(t, node);
+  unsigned size = baseSize(t, node, class_) * BytesPerWord;
+  unsigned treeNodeSize = classFixedSize
+    (t, arrayBody(t, t->m->types, Machine::TreeNodeType)) * BytesPerWord;
+  unsigned diff = treeNodeSize - size;
+  
+  if (diff) {
+    memcpy(reinterpret_cast<uint8_t*>(new_) + treeNodeSize,
+           reinterpret_cast<uint8_t*>(node) + treeNodeSize,
+           diff);
+  }
+
+  return makeTreePath(t, true, new_, newRoot, ancestors);
+}
+
+object
+leftRotate(Thread* t, object n)
+{
+  object child = clone(t, treeNodeRight(t, n));
+  set(t, n, TreeNodeRight, treeNodeLeft(t, child));
+  set(t, child, TreeNodeLeft, n);
+  return child;
+}
+
+object
+rightRotate(Thread* t, object n)
+{
+  object child = clone(t, treeNodeLeft(t, n));
+  set(t, n, TreeNodeLeft, treeNodeRight(t, child));
+  set(t, child, TreeNodeRight, n);
+  return child;
+}
+
+object
+treeAdd(Thread* t, object path)
+{
+  object new_ = treePathNode(t, path);
+  object newRoot = treePathRoot(t, path);
+  object ancestors = treePathAncestors(t, path);
+
+  // rebalance
+  treeNodeRed(t, new_) = true;
+  while (ancestors != 0 and treeNodeRed(t, pairFirst(t, ancestors))) {
+    if (pairFirst(t, ancestors)
+        == treeNodeLeft(t, pairFirst(t, pairSecond(t, ancestors))))
+    {
+      if (treeNodeRed
+          (t, treeNodeRight(t, pairFirst(t, pairSecond(t, ancestors)))))
+      {
+        treeNodeRed(t, pairFirst(t, ancestors)) = true;
+
+        object n = clone
+          (t, treeNodeRight(t, pairFirst(t, pairSecond(t, ancestors))));
+
+        set(t, pairFirst(t, pairSecond(t, ancestors)), TreeNodeRight, n);
+
+        treeNodeRed
+          (t, treeNodeRight
+           (t, pairFirst(t, pairSecond(t, ancestors)))) = false;
+
+        treeNodeRed(t, pairFirst(t, pairSecond(t, ancestors))) = false;
+
+        new_ = pairFirst(t, pairSecond(t, ancestors));
+        ancestors = pairSecond(t, pairSecond(t, ancestors));
+      } else {
+        if (new_ == treeNodeRight(t, pairFirst(t, ancestors))) {
+          new_ = pairFirst(t, ancestors);
+          ancestors = pairSecond(t, ancestors);
+
+          object n = leftRotate(t, new_);
+
+          if (new_ == treeNodeRight(t, pairFirst(t, ancestors))) {
+            set(t, pairFirst(t, ancestors), TreeNodeRight, n);
+          } else {
+            set(t, pairFirst(t, ancestors), TreeNodeLeft, n);
+          }
+          ancestors = makePair(t, n, ancestors);
+        }
+        treeNodeRed(t, pairFirst(t, ancestors)) = false;
+        treeNodeRed(t, pairFirst(t, pairSecond(t, ancestors))) = true;
+
+        object n = rightRotate(t, pairFirst(t, pairSecond(t, ancestors)));
+        if (pairSecond(t, pairSecond(t, ancestors)) == 0) {
+          newRoot = n;
+        } else if (treeNodeRight
+                   (t, pairFirst(t, pairSecond(t, pairSecond(t, ancestors))))
+                   == pairFirst(t, pairSecond(t, ancestors)))
+        {
+          set(t, pairFirst(t, pairSecond(t, pairSecond(t, ancestors))),
+              TreeNodeRight, n);
+        } else {
+          set(t, pairFirst(t, pairSecond(t, pairSecond(t, ancestors))),
+              TreeNodeLeft, n);
+        }
+        // done
+      }
+    } else { // this is just the reverse of the code above (right and
+             // left swapped):
+      if (treeNodeRed
+          (t, treeNodeLeft(t, pairFirst(t, pairSecond(t, ancestors)))))
+      {
+        treeNodeRed(t, pairFirst(t, ancestors)) = true;
+
+        object n = clone
+          (t, treeNodeLeft(t, pairFirst(t, pairSecond(t, ancestors))));
+
+        set(t, pairFirst(t, pairSecond(t, ancestors)), TreeNodeLeft, n);
+
+        treeNodeRed
+          (t, treeNodeLeft
+           (t, pairFirst(t, pairSecond(t, ancestors)))) = false;
+
+        treeNodeRed(t, pairFirst(t, pairSecond(t, ancestors))) = false;
+
+        new_ = pairFirst(t, pairSecond(t, ancestors));
+        ancestors = pairSecond(t, pairSecond(t, ancestors));
+      } else {
+        if (new_ == treeNodeLeft(t, pairFirst(t, ancestors))) {
+          new_ = pairFirst(t, ancestors);
+          ancestors = pairSecond(t, ancestors);
+
+          object n = rightRotate(t, new_);
+
+          if (new_ == treeNodeLeft(t, pairFirst(t, ancestors))) {
+            set(t, pairFirst(t, ancestors), TreeNodeLeft, n);
+          } else {
+            set(t, pairFirst(t, ancestors), TreeNodeRight, n);
+          }
+          ancestors = makePair(t, n, ancestors);
+        }
+        treeNodeRed(t, pairFirst(t, ancestors)) = false;
+        treeNodeRed(t, pairFirst(t, pairSecond(t, ancestors))) = true;
+
+        object n = leftRotate(t, pairFirst(t, pairSecond(t, ancestors)));
+        if (pairSecond(t, pairSecond(t, ancestors)) == 0) {
+          newRoot = n;
+        } else if (treeNodeLeft
+                   (t, pairFirst(t, pairSecond(t, pairSecond(t, ancestors))))
+                   == pairFirst(t, pairSecond(t, ancestors)))
+        {
+          set(t, pairFirst(t, pairSecond(t, pairSecond(t, ancestors))),
+              TreeNodeLeft, n);
+        } else {
+          set(t, pairFirst(t, pairSecond(t, pairSecond(t, ancestors))),
+              TreeNodeRight, n);
+        }
+        // done
+      }
+    }
+  }
+
+  treeNodeRed(t, newRoot) = false;
+
+  return newRoot;
+}
+
+} // namespace
+
 namespace vm {
 
 object
@@ -249,4 +452,33 @@ vectorAppend(Thread* t, object vector, object value)
   return vector;
 }
 
+object
+treeQuery(Thread* t, object tree, intptr_t key, object sentinal)
+{
+  object node = tree;
+  while (node != sentinal) {
+    intptr_t difference = key - treeNodeKey(t, node);
+    if (difference < 0) {
+      node = treeNodeLeft(t, node);
+    } else if (difference > 0) {
+      node = treeNodeRight(t, node);
+    } else {
+      return node;
+    }
+  }
+
+  return 0;
+}
+
+object
+treeInsert(Thread* t, object tree, object node, object sentinal)
+{
+  object path = treeFind(t, tree, node, sentinal);
+  if (treePathFresh(t, path)) {
+    return treeAdd(t, path);
+  } else {
+    return tree;
+  }  
+}
+
 } // namespace vm
diff --git a/src/util.h b/src/util.h
index c9a37d8911..ef217b70dc 100644
--- a/src/util.h
+++ b/src/util.h
@@ -74,10 +74,10 @@ object
 vectorAppend(Thread* t, object vector, object value);
 
 object
-treeQuery(Thread* t, object tree, intptr_t key);
+treeQuery(Thread* t, object tree, intptr_t key, object sentinal);
 
 object
-treeInsert(Thread* t, object tree, object node);
+treeInsert(Thread* t, object tree, object node, object sentinal);
 
 } // vm