diff --git a/classpath/java-lang.cpp b/classpath/java-lang.cpp
index 508d0fbbf9..ebb272a3e0 100644
--- a/classpath/java-lang.cpp
+++ b/classpath/java-lang.cpp
@@ -476,3 +476,16 @@ Java_java_lang_Double_fillBufferWithDouble(JNIEnv* e, jclass, jdouble val,
   e->ReleaseByteArrayElements(buffer, buf, 0);
   return count;
 }
+
+extern "C" JNIEXPORT void JNICALL
+Java_java_lang_Long_printLong(JNIEnv*, jclass, jlong v)
+{
+  fprintf(stderr, "DEBUG: %lld\n", v);
+}
+
+
+extern "C" JNIEXPORT void JNICALL
+Java_java_lang_Long_printInt(JNIEnv*, jclass, jint v)
+{
+  fprintf(stderr, "DEBUG: %d\n", v);
+}
diff --git a/classpath/java/lang/Long.java b/classpath/java/lang/Long.java
index 013ac883bb..4c006ea549 100644
--- a/classpath/java/lang/Long.java
+++ b/classpath/java/lang/Long.java
@@ -38,11 +38,16 @@ public final class Long extends Number implements Comparable<Long> {
     return String.valueOf(value);
   }
 
+  private static native void printLong(long v);
+  private static native void printInt(long v);
+
   public static String toString(long v, int radix) {
     if (radix < 1 || radix > 36) {
       throw new IllegalArgumentException("radix " + radix + " not in [1,36]");
     }
 
+    printLong(v);
+
     if (v == 0) {
       return "0";
     }
@@ -53,10 +58,13 @@ public final class Long extends Number implements Comparable<Long> {
     int size = (negative ? 1 : 0);
     for (long n = v; n > 0; n /= radix) ++size;
 
+    printInt(size);
+
     char[] array = new char[size];
 
     int i = array.length - 1;
     for (long n = v; n > 0; n /= radix) {
+      printLong(n);
       long digit = n % radix;
       if (digit >= 0 && digit <= 9) {
         array[i] = (char) ('0' + digit);
diff --git a/src/compile.cpp b/src/compile.cpp
index a48398874e..ee81de2aa6 100644
--- a/src/compile.cpp
+++ b/src/compile.cpp
@@ -18,7 +18,7 @@ vmJump(void* address, void* base, void* stack, void* thread);
 
 namespace {
 
-const bool Verbose = false;
+const bool Verbose = true;
 const bool DebugTraces = false;
 
 class MyThread: public Thread {
@@ -617,7 +617,7 @@ class Frame {
   }
 
   void pushLong(Operand* o) {
-    stack = c->push(stack, c->select8(o));
+    stack = c->push(stack, o);
     if (BytesPerWord == 8) {
       stack = c->push(stack, 1);
     }
@@ -654,7 +654,7 @@ class Frame {
   }
 
   Operand* popInt() {
-    Operand* tmp = c->temporary();
+    Operand* tmp = c->select4(c->temporary());
     popInt(tmp);
     return tmp;
   }
@@ -695,7 +695,8 @@ class Frame {
     assert(t, index < codeMaxLocals(t, methodCode(t, method)));
     assert(t, index < parameterFootprint(t, method)
            or getBit(map, index - parameterFootprint(t, method)) == 0);
-    pushInt(c->memory(c->base(), localOffset(t, index, method)));
+    pushInt
+      (c->signExtend4(c->memory(c->base(), localOffset(t, index, method))));
   }
 
   void loadLong(unsigned index) {
@@ -1180,8 +1181,10 @@ compileThrowNew(MyThread* t, Frame* frame, Machine::Type type)
 }
 
 void
-pushReturnValue(MyThread* t, Frame* frame, unsigned code, Operand* result)
+pushReturnValue(MyThread* t, Frame* frame, unsigned code)
 {
+  Object* result = c->result();
+
   switch (code) {
   case ByteField:
   case BooleanField:
@@ -1189,7 +1192,7 @@ pushReturnValue(MyThread* t, Frame* frame, unsigned code, Operand* result)
   case ShortField:
   case FloatField:
   case IntField:
-    frame->pushInt(result);
+    frame->pushInt(c->signExtend4(result));
     break;
 
   case ObjectField:
@@ -1198,7 +1201,7 @@ pushReturnValue(MyThread* t, Frame* frame, unsigned code, Operand* result)
 
   case LongField:
   case DoubleField:
-    frame->pushLong(result);
+    frame->pushLong(c->select8(result));
     break;
 
   case VoidField:
@@ -1207,6 +1210,8 @@ pushReturnValue(MyThread* t, Frame* frame, unsigned code, Operand* result)
   default:
     abort(t);
   }
+
+  c->release(c);
 }
 
 void
@@ -1275,15 +1280,15 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
 
       case faload:
       case iaload:
-        frame->pushInt(c->select4(c->memory(array, ArrayBody, index, 4)));
+        frame->pushInt(c->signExtend4(c->memory(array, ArrayBody, index, 4)));
         break;
 
       case baload:
-        frame->pushInt(c->select1(c->memory(array, ArrayBody, index, 1)));
+        frame->pushInt(c->signExtend1(c->memory(array, ArrayBody, index, 1)));
         break;
 
       case caload:
-        frame->pushInt(c->select2z(c->memory(array, ArrayBody, index, 2)));
+        frame->pushInt(c->zeroExtend2(c->memory(array, ArrayBody, index, 2)));
         break;
 
       case daload:
@@ -1292,7 +1297,7 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
         break;
 
       case saload:
-        frame->pushInt(c->select2(c->memory(array, ArrayBody, index, 2)));
+        frame->pushInt(c->signExtend2(c->memory(array, ArrayBody, index, 2)));
         break;
       }
 
@@ -1350,21 +1355,21 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
 
       case fastore:
       case iastore:
-        c->mov(value, c->select4(c->memory(array, ArrayBody, index, 4)));
+        c->mov(c->select4(value), c->memory(array, ArrayBody, index, 4));
         break;
 
       case bastore:
-        c->mov(value, c->select1(c->memory(array, ArrayBody, index, 1)));
+        c->mov(c->select1(value), c->memory(array, ArrayBody, index, 1));
         break;
 
       case castore:
       case sastore:
-        c->mov(value, c->select2(c->memory(array, ArrayBody, index, 2)));
+        c->mov(c->select2(value), c->memory(array, ArrayBody, index, 2));
         break;
 
       case dastore:
       case lastore:
-        c->mov(value, c->select8(c->memory(array, ArrayBody, index, 8)));
+        c->mov(c->select8(value), c->memory(array, ArrayBody, index, 8));
         break;
       }
 
@@ -1787,25 +1792,30 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
       switch (fieldCode(t, field)) {
       case ByteField:
       case BooleanField:
-        frame->pushInt(c->select1(c->memory(table, fieldOffset(t, field))));
+        frame->pushInt
+          (c->signExtend1(c->memory(table, fieldOffset(t, field))));
         break;
 
       case CharField:
-        frame->pushInt(c->select2z(c->memory(table, fieldOffset(t, field))));
+        frame->pushInt
+          (c->zeroExtend2(c->memory(table, fieldOffset(t, field))));
         break;
 
       case ShortField:
-        frame->pushInt(c->select2(c->memory(table, fieldOffset(t, field))));
+        frame->pushInt
+          (c->signExtend2(c->memory(table, fieldOffset(t, field))));
         break;
 
       case FloatField:
       case IntField:
-        frame->pushInt(c->select4(c->memory(table, fieldOffset(t, field))));
+        frame->pushInt
+          (c->signExtend4(c->memory(table, fieldOffset(t, field))));
         break;
 
       case DoubleField:
       case LongField:
-        frame->pushLong(c->select8(c->memory(table, fieldOffset(t, field))));
+        frame->pushLong
+          (c->select8(c->memory(table, fieldOffset(t, field))));
         break;
 
       case ObjectField:
@@ -1839,12 +1849,12 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
 
     case i2b: {
       Operand* top = frame->topInt();
-      c->mov(c->select1(top), top);
+      c->mov(c->signExtend1(top), top);
     } break;
 
     case i2c: {
       Operand* top = frame->topInt();
-      c->mov(c->select2z(top), top);
+      c->mov(c->zeroExtend2(top), top);
     } break;
 
     case i2d: {
@@ -1865,13 +1875,13 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
 
     case i2l: {
       Operand* a = frame->popInt();
-      frame->pushLong(a);
+      frame->pushLong(c->signExtend4(a));
       c->release(a);
     } break;
 
     case i2s: {
       Operand* top = frame->topInt();
-      c->mov(c->select2(top), top);
+      c->mov(c->signExtend2(top), top);
     } break;
       
     case iadd: {
@@ -2739,22 +2749,22 @@ compile(MyThread* t, Frame* initialFrame, unsigned ip)
       switch (fieldCode(t, field)) {
       case ByteField:
       case BooleanField:
-        c->mov(value, c->select1(c->memory(table, fieldOffset(t, field))));
+        c->mov(c->select1(value), c->memory(table, fieldOffset(t, field)));
         break;
 
       case CharField:
       case ShortField:
-        c->mov(value, c->select2(c->memory(table, fieldOffset(t, field))));
+        c->mov(c->select2(value), c->memory(table, fieldOffset(t, field)));
         break;
             
       case FloatField:
       case IntField:
-        c->mov(value, c->select4(c->memory(table, fieldOffset(t, field))));
+        c->mov(c->select4(value), c->memory(table, fieldOffset(t, field)));
         break;
 
       case DoubleField:
       case LongField:
-        c->mov(value, c->select8(c->memory(table, fieldOffset(t, field))));
+        c->mov(c->select8(value), c->memory(table, fieldOffset(t, field)));
         break;
 
       case ObjectField:
@@ -2997,13 +3007,13 @@ finish(MyThread* t, Compiler* c, object method, Vector* objectPool,
     }
 
     // for debugging:
-    if (false and
+    if (//false and
         strcmp(reinterpret_cast<const char*>
                (&byteArrayBody(t, className(t, methodClass(t, method)), 0)),
-               "java/lang/Throwable") == 0 and
+               "java/lang/Long") == 0 and
         strcmp(reinterpret_cast<const char*>
                (&byteArrayBody(t, methodName(t, method), 0)),
-               "init") == 0)
+               "toString") == 0)
     {
       asm("int3");
     }
diff --git a/src/compiler.cpp b/src/compiler.cpp
index e892c7279f..e1c54e19f2 100644
--- a/src/compiler.cpp
+++ b/src/compiler.cpp
@@ -27,11 +27,14 @@ enum Register {
 };
 
 enum SelectionType {
-  S1Selection,
-  S2Selection,
-  Z2Selection,
-  S4Selection,
-  S8Selection
+  Select1,
+  Select2,
+  Select4,
+  Select8,
+  SignExtend1,
+  SignExtend2,
+  ZeroExtend2,
+  SignExtend4
 };
 
 const bool Verbose = false;
@@ -1112,8 +1115,8 @@ RegisterOperand::apply(Context* c, Operation operation)
       case S8Selection:
         assert(c, selection == S8Selection);
 
-        register_(c, high(c))->apply(c, pop);
         register_(c, value(c))->apply(c, pop);
+        register_(c, high(c))->apply(c, pop);
         break;
 
       default: abort(c);
@@ -1125,16 +1128,10 @@ RegisterOperand::apply(Context* c, Operation operation)
     if (selection == DefaultSelection) {
       c->code.append(0x50 | value(c));
     } else {
-      switch (selection) {
-      case S8Selection:
-        assert(c, selection == S8Selection);
+      assert(c, selection == S8Selection);
 
-        register_(c, value(c))->apply(c, push);
-        register_(c, high(c))->apply(c, push);
-        break;
-
-      default: abort(c);
-      }
+      register_(c, high(c))->apply(c, push);
+      register_(c, value(c))->apply(c, push);
     }
     break;
 
@@ -1579,6 +1576,8 @@ ImmediateOperand::apply(Context* c, Operation operation)
         tmp->release(c);
       }
     } else {
+      assert(c, selection == S8Selection);
+
       immediate(c, (value >> 32) & 0xFFFFFFFF)->apply(c, push);
       immediate(c, (value      ) & 0xFFFFFFFF)->apply(c, push);
     }
@@ -1678,11 +1677,24 @@ MemoryOperand::apply(Context* c, Operation operation)
     if (selection == DefaultSelection) {
       encode(c, 0xff, 6, this, false);
     } else {
-      RegisterOperand* tmp = temporary
-        (c, selection == S8Selection ? S8Selection : DefaultSelection);
-      tmp->accept(c, mov, this);
-      tmp->apply(c, operation);
-      tmp->release(c);
+      switch (selection) {
+      case S8Selection: {
+        MemoryOperand* low = memory(c, base, displacement, index, scale);
+        MemoryOperand* high = memory
+          (c, base, displacement + BytesPerWord, index, scale);
+        
+        high->apply(c, push);
+        low->apply(c, push);
+      } break;
+        
+      default: {
+        RegisterOperand* tmp = temporary
+          (c, selection == S8Selection ? S8Selection : DefaultSelection);
+        tmp->accept(c, mov, this);
+        tmp->apply(c, operation);
+        tmp->release(c);
+      } break;
+      }
     }
     break;
 
@@ -2169,7 +2181,7 @@ class MyCompiler: public Compiler {
       (&c, static_cast<MyPromise*>(machineIp()));
   }
 
-  virtual Operand* indirectCall
+  virtual void indirectCall
   (Operand* address, unsigned argumentCount, ...)
   {
     va_list a; va_start(a, argumentCount);
@@ -2180,8 +2192,6 @@ class MyCompiler: public Compiler {
     call(immediate(&c, c.indirectCaller));
 
     add(immediate(&c, footprint), register_(&c, rsp));
-
-    return register_(&c, rax);
   }
 
   virtual void indirectCallNoReturn
@@ -2196,7 +2206,7 @@ class MyCompiler: public Compiler {
     call(immediate(&c, c.indirectCaller));
   }
 
-  virtual Operand* directCall
+  virtual void directCall
   (Operand* address, unsigned argumentCount, ...)
   {
     va_list a; va_start(a, argumentCount);
@@ -2206,8 +2216,10 @@ class MyCompiler: public Compiler {
     call(address);
 
     add(immediate(&c, footprint), register_(&c, rsp));
+  }
 
-    return register_(&c, rax);
+  virtual Operand* result() {
+    return ::temporary(&c, rax, rdx);
   }
 
   virtual void return_(Operand* v) {
@@ -2216,14 +2228,12 @@ class MyCompiler: public Compiler {
     ret();
   }
 
-  virtual Operand* call(Operand* v) {
+  virtual void call(Operand* v) {
     appendOperation(&c, MyOperand::call, v);
-    return register_(&c, rax);
   }
 
-  virtual Operand* alignedCall(Operand* v) {
+  virtual void alignedCall(Operand* v) {
     appendOperation(&c, MyOperand::alignedCall, v);
-    return register_(&c, rax);
   }
 
   virtual void ret() {
@@ -2322,23 +2332,35 @@ class MyCompiler: public Compiler {
   }
 
   virtual Operand* select1(Operand* v) {
-    return static_cast<MyOperand*>(v)->select(&c, S1Selection);
+    return static_cast<MyOperand*>(v)->select(&c, Select1);
   }
 
   virtual Operand* select2(Operand* v) {
-    return static_cast<MyOperand*>(v)->select(&c, S2Selection);
-  }
-
-  virtual Operand* select2z(Operand* v) {
-    return static_cast<MyOperand*>(v)->select(&c, Z2Selection);
+    return static_cast<MyOperand*>(v)->select(&c, Select2);
   }
 
   virtual Operand* select4(Operand* v) {
-    return static_cast<MyOperand*>(v)->select(&c, S4Selection);
+    return static_cast<MyOperand*>(v)->select(&c, Select4);
   }
 
   virtual Operand* select8(Operand* v) {
-    return static_cast<MyOperand*>(v)->select(&c, S8Selection);
+    return static_cast<MyOperand*>(v)->select(&c, Select8);
+  }
+
+  virtual Operand* signExtend1(Operand* v) {
+    return static_cast<MyOperand*>(v)->select(&c, SignExtend1);
+  }
+
+  virtual Operand* signExtend2(Operand* v) {
+    return static_cast<MyOperand*>(v)->select(&c, SignExtend2);
+  }
+
+  virtual Operand* zeroExtend2(Operand* v) {
+    return static_cast<MyOperand*>(v)->select(&c, ZeroExtend2);
+  }
+
+  virtual Operand* signExtend4(Operand* v) {
+    return static_cast<MyOperand*>(v)->select(&c, SignExtend4);
   }
 
   virtual void prologue() {
diff --git a/src/compiler.h b/src/compiler.h
index c22ba9034e..28b585f11c 100644
--- a/src/compiler.h
+++ b/src/compiler.h
@@ -36,29 +36,34 @@ class Compiler {
 
   virtual Operand* select1(Operand*) = 0;
   virtual Operand* select2(Operand*) = 0;
-  virtual Operand* select2z(Operand*) = 0;
   virtual Operand* select4(Operand*) = 0;
   virtual Operand* select8(Operand*) = 0;
 
+  virtual Operand* signExtend1(Operand*) = 0;
+  virtual Operand* signExtend2(Operand*) = 0;
+  virtual Operand* zeroExtend2(Operand*) = 0;
+  virtual Operand* signExtend4(Operand*) = 0;
+
   virtual Operand* stack() = 0;
   virtual Operand* base() = 0;
   virtual Operand* thread() = 0;
   virtual Operand* indirectTarget() = 0;
   virtual Operand* temporary() = 0;
+  virtual Operand* result() = 0;
   virtual void release(Operand*) = 0;
 
   virtual Operand* label() = 0;
   virtual void mark(Operand*) = 0;
 
-  virtual Operand* indirectCall
+  virtual void indirectCall
   (Operand* address, unsigned argumentCount, ...) = 0;
   virtual void indirectCallNoReturn
   (Operand* address, unsigned argumentCount, ...) = 0;
-  virtual Operand* directCall
+  virtual void directCall
   (Operand* address, unsigned argumentCount, ...) = 0;
 
-  virtual Operand* call(Operand*) = 0;
-  virtual Operand* alignedCall(Operand*) = 0;
+  virtual void call(Operand*) = 0;
+  virtual void alignedCall(Operand*) = 0;
   virtual void return_(Operand*) = 0;
   virtual void ret() = 0;