diff --git a/src/bootimage.cpp b/src/bootimage.cpp
index 7179328580..e57cfd0f3f 100644
--- a/src/bootimage.cpp
+++ b/src/bootimage.cpp
@@ -31,12 +31,6 @@ endsWith(const char* suffix, const char* s, unsigned length)
     and memcmp(suffix, s + (length - suffixLength), suffixLength) == 0;
 }
 
-unsigned
-codeMapSize(unsigned codeSize)
-{
-  return ceiling(codeSize, BitsPerWord) * BytesPerWord;
-}
-
 object
 makeCodeImage(Thread* t, Zone* zone, BootImage* image, uint8_t* code,
               unsigned capacity)
@@ -82,12 +76,6 @@ makeCodeImage(Thread* t, Zone* zone, BootImage* image, uint8_t* code,
   return constants;
 }
 
-unsigned
-heapMapSize(unsigned heapSize)
-{
-  return ceiling(heapSize, BitsPerWord * 8) * BytesPerWord;
-}
-
 unsigned
 objectSize(Thread* t, object o)
 {
@@ -238,15 +226,18 @@ writeBootImage(Thread* t, FILE* out)
 
   image.magic = BootImage::Magic;
 
-//   fprintf(stderr, "heap size %d code size %d\n",
-//           image.heapSize, image.codeSize);
-  fwrite(&image, sizeof(BootImage), 1, out);
+  if (true) {
+    fprintf(stderr, "heap size %d code size %d\n",
+            image.heapSize, image.codeSize);
+  } else {
+    fwrite(&image, sizeof(BootImage), 1, out);
 
-  fwrite(heapMap, pad(heapMapSize(image.heapSize)), 1, out);
-  fwrite(heap, pad(image.heapSize), 1, out);
+    fwrite(heapMap, pad(heapMapSize(image.heapSize)), 1, out);
+    fwrite(heap, pad(image.heapSize), 1, out);
 
-  fwrite(codeMap, pad(codeMapSize(image.codeSize)), 1, out);
-  fwrite(code, pad(image.codeSize), 1, out);
+    fwrite(codeMap, pad(codeMapSize(image.codeSize)), 1, out);
+    fwrite(code, pad(image.codeSize), 1, out);
+  }
 }
 
 } // namespace
diff --git a/src/bootimage.h b/src/bootimage.h
index 2e09e44d89..cfd73171d1 100644
--- a/src/bootimage.h
+++ b/src/bootimage.h
@@ -28,21 +28,40 @@ class BootImage {
   unsigned stringMap;
   unsigned types;
 
-  uintptr_t codeBase;
   unsigned callTable;
   unsigned methodTree;
   unsigned methodTreeSentinal;
-  unsigned objectPools;
+
+  uintptr_t codeBase;
 
   unsigned defaultThunk;
   unsigned nativeThunk;
   unsigned aioobThunk;
-  
-#define THUNK(s) unsigned s##Thunk;
+
+  unsigned thunkTable;
+  unsigned thunkSize;
+
+  unsigned compileMethodCall;
+  unsigned invokeNativeCall;
+  unsigned throwArrayIndexOutOfBoundsCall;
+
+#define THUNK(s) unsigned s##Call;
 #include "thunks.cpp"
 #undef THUNK
 };
 
+inline unsigned
+codeMapSize(unsigned codeSize)
+{
+  return ceiling(codeSize, BitsPerWord) * BytesPerWord;
+}
+
+inline unsigned
+heapMapSize(unsigned heapSize)
+{
+  return ceiling(heapSize, BitsPerWord * 8) * BytesPerWord;
+}
+
 } // namespace vm
 
 #endif//BOOTIMAGE_H
diff --git a/src/compile.cpp b/src/compile.cpp
index 3441be0dae..4c68c5bf95 100644
--- a/src/compile.cpp
+++ b/src/compile.cpp
@@ -1738,19 +1738,28 @@ compileDirectInvoke(MyThread* t, Frame* frame, object target)
     } else {
       BootContext* bc = frame->context->bootContext;
       if (bc) {
-        Promise* p = new (bc->zone->allocate(sizeof(ListenPromise)))
-          ListenPromise(t->m->system, bc->zone);
+        if (objectClass(t, target) == objectClass(t, frame->context->method)) {
+          Promise* p = new (bc->zone->allocate(sizeof(ListenPromise)))
+            ListenPromise(t->m->system, bc->zone);
 
-        PROTECT(t, target);
-        object pointer = makePointer(t, p);
-        bc->calls = makeTriple(t, target, pointer, bc->calls);
+          PROTECT(t, target);
+          object pointer = makePointer(t, p);
+          bc->calls = makeTriple(t, target, pointer, bc->calls);
 
-        result = c->call
-          (c->promiseConstant(p),
-           0,
-           frame->trace(0, false),
-           rSize,
-           0);
+          result = c->call
+            (c->promiseConstant(p),
+             0,
+             frame->trace(0, false),
+             rSize,
+             0);
+        } else {
+          result = c->call
+            (c->constant(defaultThunk(t)),
+             Compiler::Aligned,
+             frame->trace(target, false),
+             rSize,
+             0);
+        }
       } else if (methodCompiled(t, target) == defaultThunk(t)) {
         result = c->call
           (c->constant(defaultThunk(t)),
@@ -4654,6 +4663,31 @@ class SegFaultHandler: public System::SignalHandler {
   Machine* m;
 };
 
+object
+fixupCallTable(MyThread* t, object oldTable, uintptr_t oldBase,
+               uintptr_t newBase)
+{
+  PROTECT(t, oldTable);
+
+  object newTable = makeArray(t, arrayLength(t, oldTable), true);
+
+  for (unsigned i = 0; i < arrayLength(t, oldTable); ++i) {
+    object next;
+    for (object p = arrayBody(t, oldTable, i); p; p = next) {
+      next = callNodeNext(t, p);
+
+      intptr_t k = (callNodeAddress(t, p) - oldBase) + newBase;
+
+      unsigned index = k & (arrayLength(t, newTable) - 1);
+
+      set(t, p, CallNodeNext, arrayBody(t, newTable, index));
+      set(t, newTable, ArrayBody + (index * BytesPerWord), p);
+    }
+  }
+
+  return newTable;
+}
+
 class FixedAllocator: public Allocator {
  public:
   FixedAllocator(Thread* t, uint8_t* base, unsigned capacity):
@@ -5073,7 +5107,43 @@ class MyProcessor: public Processor {
     image->callTable = w->visitRoot(callTable);
     image->methodTree = w->visitRoot(methodTree);
     image->methodTreeSentinal = w->visitRoot(methodTreeSentinal);
-    image->objectPools = w->visitRoot(objectPools);
+  }
+
+  virtual void boot(Thread* t, BootImage* image, uintptr_t* heap,
+                    uint8_t* code)
+  {
+    methodTree = reinterpret_cast<object>(heap + image->methodTree);
+    methodTreeSentinal = reinterpret_cast<object>
+      (heap + image->methodTreeSentinal);
+
+    callTable = fixupCallTable
+      (static_cast<MyThread*>(t),
+       reinterpret_cast<object>(heap + image->callTable),
+       image->codeBase,
+       reinterpret_cast<uintptr_t>(code));
+
+    defaultThunk = code + image->defaultThunk;
+    { void* p = voidPointer(::compileMethod);
+      memcpy(code + image->compileMethodCall, &p, BytesPerWord); }
+
+    nativeThunk = code + image->nativeThunk;
+    { void* p = voidPointer(invokeNative);
+      memcpy(code + image->invokeNativeCall, &p, BytesPerWord); }
+
+    aioobThunk = code + image->aioobThunk;
+    { void* p = voidPointer(throwArrayIndexOutOfBounds);
+      memcpy(code + image->throwArrayIndexOutOfBoundsCall, &p, BytesPerWord); }
+
+    thunkTable = code + image->thunkTable;
+    thunkSize = image->thunkSize;
+
+#define THUNK(s)                                                \
+    { void* p = voidPointer(s);                                 \
+      memcpy(code + image->s##Call, &p, BytesPerWord); }
+
+#include "thunks.cpp"
+
+#undef THUNK
   }
   
   System* s;
@@ -5183,38 +5253,47 @@ compileThunks(MyThread* t, Allocator* allocator, MyProcessor* p,
   p->defaultThunk = finish
     (t, allocator, defaultContext.context.assembler, "default");
 
-  { void* p = defaultContext.promise.listener->resolve
+  { void* call = defaultContext.promise.listener->resolve
       (reinterpret_cast<intptr_t>(voidPointer(compileMethod))); 
     if (image) {
-      image->defaultThunk = static_cast<uint8_t*>(p) - imageBase;
+      image->defaultThunk = p->defaultThunk - imageBase;
+      image->compileMethodCall = static_cast<uint8_t*>(call) - imageBase;
     }
   }
 
   p->nativeThunk = finish
     (t, allocator, nativeContext.context.assembler, "native");
 
-  { void* p = nativeContext.promise.listener->resolve
+  { void* call = nativeContext.promise.listener->resolve
       (reinterpret_cast<intptr_t>(voidPointer(invokeNative)));
 
     if (image) {
-      image->nativeThunk = static_cast<uint8_t*>(p) - imageBase;
+      image->nativeThunk = p->nativeThunk - imageBase;
+      image->invokeNativeCall = static_cast<uint8_t*>(call) - imageBase;
     }
   }
 
   p->aioobThunk = finish
     (t, allocator, aioobContext.context.assembler, "aioob");
 
-  { void* p = aioobContext.promise.listener->resolve
+  { void* call = aioobContext.promise.listener->resolve
       (reinterpret_cast<intptr_t>(voidPointer(throwArrayIndexOutOfBounds)));
 
     if (image) {
-      image->aioobThunk = static_cast<uint8_t*>(p) - imageBase;
+      image->aioobThunk = p->aioobThunk - imageBase;
+      image->throwArrayIndexOutOfBoundsCall
+        = static_cast<uint8_t*>(call) - imageBase;
     }
   }
 
   p->thunkTable = static_cast<uint8_t*>
     (allocator->allocate(p->thunkSize * ThunkCount));
 
+  if (image) {
+    image->thunkTable = p->thunkTable - imageBase;
+    image->thunkSize = p->thunkSize;
+  }
+
   logCompile(t, p->thunkTable, p->thunkSize * ThunkCount, 0, "thunkTable", 0);
 
   uint8_t* start = p->thunkTable;
@@ -5222,10 +5301,10 @@ compileThunks(MyThread* t, Allocator* allocator, MyProcessor* p,
 #define THUNK(s)                                                        \
   tableContext.context.assembler->writeTo(start);                       \
   start += p->thunkSize;                                                \
-  { void* p = tableContext.promise.listener->resolve                    \
+  { void* call = tableContext.promise.listener->resolve                 \
       (reinterpret_cast<intptr_t>(voidPointer(s)));                     \
     if (image) {                                                        \
-      image->s##Thunk = static_cast<uint8_t*>(p) - imageBase;           \
+      image->s##Call = static_cast<uint8_t*>(call) - imageBase;         \
     }                                                                   \
   }
 
diff --git a/src/heap.cpp b/src/heap.cpp
index a70f176b3e..93fb925b2d 100644
--- a/src/heap.cpp
+++ b/src/heap.cpp
@@ -121,12 +121,12 @@ class Segment {
         for (; word <= wordLimit and (word < wordLimit or bit < bitLimit);
              ++word)
         {
-          uintptr_t* p = map->data() + word;
-          if (*p) {
+          uintptr_t w = map->data()[word];
+          if (2) {
             for (; bit < BitsPerWord and (word < wordLimit or bit < bitLimit);
                  ++bit)
             {
-              if (map->data()[word] & (static_cast<uintptr_t>(1) << bit)) {
+              if (w & (static_cast<uintptr_t>(1) << bit)) {
                 index = ::indexOf(word, bit);
 //                 printf("hit at index %d\n", index);
                 return true;
diff --git a/src/machine.cpp b/src/machine.cpp
index dd9b4cdd4f..78a0063995 100644
--- a/src/machine.cpp
+++ b/src/machine.cpp
@@ -1502,6 +1502,156 @@ bootJavaClass(Thread* t, Machine::Type type, int superType, const char* name,
   hashMapInsert(t, t->m->bootstrapClassMap, n, class_, byteArrayHash);
 }
 
+void
+boot(Thread* t, BootImage* image)
+{
+  assert(t, image->magic == BootImage::Magic);
+
+  uintptr_t* heapMap = reinterpret_cast<uintptr_t*>(image + 1);
+  unsigned heapMapSizeInWords = ceiling
+    (heapMapSize(image->heapSize), BytesPerWord);
+  uintptr_t* heap = heapMap + heapMapSizeInWords;
+  
+  for (unsigned word = 0; word < heapMapSizeInWords; ++word) {
+    uintptr_t w = heapMap[word];
+    if (w) {
+      for (unsigned bit = 0; bit < BitsPerWord; ++bit) {
+        if (w & (static_cast<uintptr_t>(1) << bit)) {
+          unsigned index = ::indexOf(word, bit);
+          uintptr_t* p = heap + index;
+          if (*p) {
+            *p = reinterpret_cast<uintptr_t>(heap + *p - 1);
+          }
+        }
+      }
+    }
+  }
+
+  t->m->loader = reinterpret_cast<object>(heap + image->loader);
+  t->m->stringMap = reinterpret_cast<object>(heap + image->stringMap);
+  t->m->types = reinterpret_cast<object>(heap + image->types);
+
+  uintptr_t* codeMap = heap + ceiling(image->heapSize, BytesPerWord);
+  unsigned codeMapSizeInWords = ceiling
+    (codeMapSize(image->codeSize), BytesPerWord);
+  uint8_t* code = reinterpret_cast<uint8_t*>(codeMap + codeMapSizeInWords);
+ 
+  for (unsigned word = 0; word < codeMapSizeInWords; ++word) {
+    uintptr_t w = codeMap[word];
+    if (w) {
+      for (unsigned bit = 0; bit < BitsPerWord; ++bit) {
+        if (w & (static_cast<uintptr_t>(1) << bit)) {
+          unsigned index = ::indexOf(word, bit);
+          uintptr_t v; memcpy(&v, code + index, BytesPerWord);
+          if (v) {
+            v = reinterpret_cast<uintptr_t>(heap + v - 1);
+            memcpy(code + index, &v, BytesPerWord);
+          }
+        }
+      }
+    }
+  }
+
+  t->m->processor->boot(t, image, heap, code);
+
+  for (HashMapIterator it(t, systemClassLoaderMap(t, t->m->loader));
+       it.hasMore();)
+  {
+    object c = tripleSecond(t, it.next());
+
+    if (classMethodTable(t, c)) {
+      for (unsigned i = 0; i < arrayLength(t, classMethodTable(t, c)); ++i) {
+        object method = arrayBody(t, classMethodTable(t, c), i);
+        methodCompiled(t, method)
+          = (methodCompiled(t, method) - image->codeBase)
+          + reinterpret_cast<uintptr_t>(code);
+      }
+    }
+
+    t->m->processor->initVtable(t, c);
+  }
+}
+
+void
+boot(Thread* t)
+{
+  Machine* m = t->m;
+
+  m->loader = allocate(t, sizeof(void*) * 3, true);
+  memset(m->loader, 0, sizeof(void*) * 2);
+
+  m->types = allocate(t, pad((TypeCount + 2) * BytesPerWord), true);
+  arrayLength(t, m->types) = TypeCount;
+  memset(&arrayBody(t, m->types, 0), 0, TypeCount * BytesPerWord);
+
+#include "type-initializations.cpp"
+
+  object arrayClass = arrayBody(t, m->types, Machine::ArrayType);
+  set(t, m->types, 0, arrayClass);
+
+  object loaderClass = arrayBody
+    (t, m->types, Machine::SystemClassLoaderType);
+  set(t, m->loader, 0, loaderClass);
+
+  object objectClass = arrayBody(t, m->types, Machine::JobjectType);
+
+  object classClass = arrayBody(t, m->types, Machine::ClassType);
+  set(t, classClass, 0, classClass);
+  set(t, classClass, ClassSuper, objectClass);
+
+  object intArrayClass = arrayBody(t, m->types, Machine::IntArrayType);
+  set(t, intArrayClass, 0, classClass);
+  set(t, intArrayClass, ClassSuper, objectClass);
+
+  m->unsafe = false;
+
+  classVmFlags(t, arrayBody(t, m->types, Machine::SingletonType))
+    |= SingletonFlag;
+
+  classVmFlags(t, arrayBody(t, m->types, Machine::JreferenceType))
+    |= ReferenceFlag;
+  classVmFlags(t, arrayBody(t, m->types, Machine::WeakReferenceType))
+    |= ReferenceFlag | WeakReferenceFlag;
+  classVmFlags(t, arrayBody(t, m->types, Machine::PhantomReferenceType))
+    |= ReferenceFlag | WeakReferenceFlag;
+
+  classVmFlags(t, arrayBody(t, m->types, Machine::JbooleanType))
+    |= PrimitiveFlag;
+  classVmFlags(t, arrayBody(t, m->types, Machine::JbyteType))
+    |= PrimitiveFlag;
+  classVmFlags(t, arrayBody(t, m->types, Machine::JcharType))
+    |= PrimitiveFlag;
+  classVmFlags(t, arrayBody(t, m->types, Machine::JshortType))
+    |= PrimitiveFlag;
+  classVmFlags(t, arrayBody(t, m->types, Machine::JintType))
+    |= PrimitiveFlag;
+  classVmFlags(t, arrayBody(t, m->types, Machine::JlongType))
+    |= PrimitiveFlag;
+  classVmFlags(t, arrayBody(t, m->types, Machine::JfloatType))
+    |= PrimitiveFlag;
+  classVmFlags(t, arrayBody(t, m->types, Machine::JdoubleType))
+    |= PrimitiveFlag;
+  classVmFlags(t, arrayBody(t, m->types, Machine::JvoidType))
+    |= PrimitiveFlag;
+
+  m->bootstrapClassMap = makeHashMap(t, 0, 0);
+
+  { object loaderMap = makeHashMap(t, 0, 0);
+    set(t, m->loader, SystemClassLoaderMap, loaderMap);
+  }
+
+  m->stringMap = makeWeakHashMap(t, 0, 0);
+
+  { object bootCode = makeCode(t, 0, 0, 0, 0, 0, 1, false);
+    codeBody(t, bootCode, 0) = impdep1;
+    object bootMethod = makeMethod
+      (t, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, bootCode, 0);
+    PROTECT(t, bootMethod);
+
+#include "type-java-initializations.cpp"
+  }
+}
+
 class HeapClient: public Heap::Client {
  public:
   HeapClient(Machine* m): m(m) { }
@@ -1716,86 +1866,30 @@ Thread::init()
       abort(this);
     }
 
-    Thread* t = this;
+    BootImage* image = 0;
+    const char* imageFunctionName = findProperty(m, "avian.bootimage");
+    if (imageFunctionName) {
+      void* p = m->libraries->resolve(imageFunctionName);
+      if (p) {
+        BootImage* (*function)(unsigned*);
+        memcpy(&function, &p, BytesPerWord);
 
-    t->m->loader = allocate(t, sizeof(void*) * 3, true);
-    memset(t->m->loader, 0, sizeof(void*) * 2);
+        unsigned size;
+        image = function(&size);
+      }
+    }
 
-    t->m->types = allocate(t, pad((TypeCount + 2) * BytesPerWord), true);
-    arrayLength(t, t->m->types) = TypeCount;
-    memset(&arrayBody(t, t->m->types, 0), 0, TypeCount * BytesPerWord);
-
-#include "type-initializations.cpp"
-
-    object arrayClass = arrayBody(t, t->m->types, Machine::ArrayType);
-    set(t, t->m->types, 0, arrayClass);
-
-    object loaderClass = arrayBody
-      (t, t->m->types, Machine::SystemClassLoaderType);
-    set(t, t->m->loader, 0, loaderClass);
-
-    object objectClass = arrayBody(t, m->types, Machine::JobjectType);
-
-    object classClass = arrayBody(t, m->types, Machine::ClassType);
-    set(t, classClass, 0, classClass);
-    set(t, classClass, ClassSuper, objectClass);
-
-    object intArrayClass = arrayBody(t, m->types, Machine::IntArrayType);
-    set(t, intArrayClass, 0, classClass);
-    set(t, intArrayClass, ClassSuper, objectClass);
-
-    m->unsafe = false;
-
-    classVmFlags(t, arrayBody(t, m->types, Machine::SingletonType))
-      |= SingletonFlag;
-
-    classVmFlags(t, arrayBody(t, m->types, Machine::JreferenceType))
-      |= ReferenceFlag;
-    classVmFlags(t, arrayBody(t, m->types, Machine::WeakReferenceType))
-      |= ReferenceFlag | WeakReferenceFlag;
-    classVmFlags(t, arrayBody(t, m->types, Machine::PhantomReferenceType))
-      |= ReferenceFlag | WeakReferenceFlag;
-
-    classVmFlags(t, arrayBody(t, m->types, Machine::JbooleanType))
-      |= PrimitiveFlag;
-    classVmFlags(t, arrayBody(t, m->types, Machine::JbyteType))
-      |= PrimitiveFlag;
-    classVmFlags(t, arrayBody(t, m->types, Machine::JcharType))
-      |= PrimitiveFlag;
-    classVmFlags(t, arrayBody(t, m->types, Machine::JshortType))
-      |= PrimitiveFlag;
-    classVmFlags(t, arrayBody(t, m->types, Machine::JintType))
-      |= PrimitiveFlag;
-    classVmFlags(t, arrayBody(t, m->types, Machine::JlongType))
-      |= PrimitiveFlag;
-    classVmFlags(t, arrayBody(t, m->types, Machine::JfloatType))
-      |= PrimitiveFlag;
-    classVmFlags(t, arrayBody(t, m->types, Machine::JdoubleType))
-      |= PrimitiveFlag;
-    classVmFlags(t, arrayBody(t, m->types, Machine::JvoidType))
-      |= PrimitiveFlag;
-
-    m->bootstrapClassMap = makeHashMap(this, 0, 0);
-
-    { object loaderMap = makeHashMap(this, 0, 0);
-      set(t, m->loader, SystemClassLoaderMap, loaderMap);
+    if (image) {
+      boot(this, image);
+    } else {
+      boot(this);
     }
 
     m->monitorMap = makeWeakHashMap(this, 0, 0);
-    m->stringMap = makeWeakHashMap(this, 0, 0);
 
     m->jniMethodTable = makeVector(this, 0, 0, false);
 
     m->localThread->set(this);
-
-    { object bootCode = makeCode(t, 0, 0, 0, 0, 0, 1, false);
-      codeBody(t, bootCode, 0) = impdep1;
-      object bootMethod = makeMethod
-        (t, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, bootCode, 0);
-      PROTECT(t, bootMethod);
-
-#include "type-java-initializations.cpp"
-    }
   } else {
     peer = parent->child;
     parent->child = this;
diff --git a/src/processor.h b/src/processor.h
index 621bd5460a..ee5d728de3 100644
--- a/src/processor.h
+++ b/src/processor.h
@@ -127,6 +127,9 @@ class Processor {
   virtual void
   visitRoots(BootImage* image, HeapWalker* w) = 0;
 
+  virtual void
+  boot(Thread* t, BootImage* image, uintptr_t* heap, uint8_t* code) = 0;
+
   object
   invoke(Thread* t, object method, object this_, ...)
   {