diff --git a/src/assembler.h b/src/assembler.h
index e97ee597db..1a8a1f6fba 100644
--- a/src/assembler.h
+++ b/src/assembler.h
@@ -136,6 +136,9 @@ class Assembler {
 
     virtual int acquireTemporary(int r = NoRegister) = 0;
     virtual void releaseTemporary(int r) = 0;
+
+    virtual void save(int r) = 0;
+    virtual void restore(int r) = 0;
   };
 
   virtual ~Assembler() { }
diff --git a/src/compile.cpp b/src/compile.cpp
index 9bd513b945..0712548195 100644
--- a/src/compile.cpp
+++ b/src/compile.cpp
@@ -3828,11 +3828,11 @@ finish(MyThread* t, Context* context)
       strcmp
       (reinterpret_cast<const char*>
        (&byteArrayBody(t, className(t, methodClass(t, context->method)), 0)),
-       "java/lang/StringBuilder") == 0 and
+       "java/lang/Class") == 0 and
       strcmp
       (reinterpret_cast<const char*>
        (&byteArrayBody(t, methodName(t, context->method), 0)),
-       "<init>") == 0)
+       "replace") == 0)
   {
     asm("int3");
   }
diff --git a/src/compiler.cpp b/src/compiler.cpp
index 9370f9d565..d24d98aca6 100644
--- a/src/compiler.cpp
+++ b/src/compiler.cpp
@@ -15,10 +15,10 @@ using namespace vm;
 
 namespace {
 
-const bool DebugAppend = true;
-const bool DebugCompile = true;
-const bool DebugStack = true;
-const bool DebugRegisters = true;
+const bool DebugAppend = false;
+const bool DebugCompile = false;
+const bool DebugStack = false;
+const bool DebugRegisters = false;
 
 class Context;
 class Value;
@@ -101,6 +101,7 @@ class Register {
   unsigned size;
   unsigned refCount;
   bool reserved;
+  bool pushed;
 };
 
 class ConstantPoolNode {
@@ -826,6 +827,42 @@ pushNow(Context* c, Stack* start)
   pushNow(c, start, count);
 }
 
+bool
+trySteal(Context* c, int r, Stack* stack)
+{
+  Value* v = c->registers[r].value;
+
+  assert(c, c->registers[r].refCount == 0);
+
+  if (DebugRegisters) {
+    fprintf(stderr, "try steal %d from %p: next: %p\n",
+            r, v, v->sites->next);
+  }
+
+  if (v->sites->next == 0) {
+    unsigned count = 0;
+    Stack* start = 0;
+    for (Stack* s = stack; s and (not s->pushed); s = s->next) {
+      if (s->value == v) {
+        start = s;
+      }
+      if (start) {
+        ++ count;
+      }
+    }
+
+    if (start) {
+      pushNow(c, start, count);
+    } else {
+      return false;
+    }
+  }
+
+  removeSite(c, v, c->registers[r].site);
+
+  return true;
+}
+
 bool
 tryAcquire(Context* c, int r, Stack* stack, unsigned newSize, Value* newValue,
            Site* newSite)
@@ -842,33 +879,9 @@ tryAcquire(Context* c, int r, Stack* stack, unsigned newSize, Value* newValue,
       and oldValue != newValue
       and findSite(c, oldValue, c->registers[r].site))
   {
-    assert(c, c->registers[r].refCount == 0);
-
-    if (DebugRegisters) {
-      fprintf(stderr, "try steal %d from %p: next: %p\n",
-              r, oldValue, oldValue->sites->next);
+    if (not trySteal(c, r, stack)) {
+      return false;
     }
-
-    if (oldValue->sites->next == 0) {
-      unsigned count = 0;
-      Stack* start = 0;
-      for (Stack* s = stack; s and (not s->pushed); s = s->next) {
-        if (s->value == oldValue) {
-          start = s;
-        }
-        if (start) {
-          ++ count;
-        }
-      }
-
-      if (start) {
-        pushNow(c, start, count);
-      } else {
-        return false;
-      }
-    }
-
-    removeSite(c, oldValue, c->registers[r].site);
   }
 
   c->registers[r].size = newSize;
@@ -1182,15 +1195,15 @@ class MoveEvent: public Event {
 
     nextRead(c, src);
 
+    if (cost) {
+      apply(c, type, size, src->source, target);
+    }
+
     if (dst->reads) {
       addSite(c, stack, size, dst, target);
     } else {
       removeSite(c, dst, target);
     }
-
-    if (cost) {
-      apply(c, type, size, src->source, target);
-    }
   }
 
   BinaryOperation type;
@@ -1730,7 +1743,7 @@ compile(Context* c)
 
       for (Event* e = li->firstEvent; e; e = e->next) {
         if (e->stackReset) {
-          fprintf(stderr, "stack reset\n");
+//           fprintf(stderr, "stack reset\n");
           for (Stack* s = e->stack; s; s = s->next) {
             if (s->value->sites) {
               assert(c, s->value->sites->next == 0);
@@ -1863,16 +1876,16 @@ freeRegisterExcept(Context* c, int except, bool allowAcquired)
     }
   }
 
-  for (int i = c->assembler->registerCount() - 1; i >= 0; --i) {
-    if (i != except
-        and c->registers[i].refCount == 0
-        and (not usedExclusively(c, i)))
-    {
-      return i;
-    }
-  }
-
   if (allowAcquired) {
+    for (int i = c->assembler->registerCount() - 1; i >= 0; --i) {
+      if (i != except
+          and c->registers[i].refCount == 0
+          and (not usedExclusively(c, i)))
+      {
+        return i;
+      }
+    }
+
     for (int i = c->assembler->registerCount() - 1; i >= 0; --i) {
       if (i != except
           and c->registers[i].refCount == 0)
@@ -1924,8 +1937,9 @@ class Client: public Assembler::Client {
     if (r == NoRegister) {
       r = freeRegisterExcept(c, NoRegister, false);
     } else {
-      expect(c, c->registers[r].refCount == 0);
-      expect(c, c->registers[r].value == 0);
+      expect(c, (c->registers[r].refCount == 0
+                 and c->registers[r].value == 0)
+             or c->registers[r].pushed);
     }
     increment(c, r);
     return r;
@@ -1935,6 +1949,22 @@ class Client: public Assembler::Client {
     decrement(c, r);
   }
 
+  virtual void save(int r) {
+    if (c->registers[r].refCount or c->registers[r].value) {
+      Assembler::Register operand(r);
+      c->assembler->apply(Push, BytesPerWord, RegisterOperand, &operand);
+      c->registers[r].pushed = true;
+    }
+  }
+
+  virtual void restore(int r) {
+    if (c->registers[r].pushed) {
+      Assembler::Register operand(r);
+      c->assembler->apply(Pop, BytesPerWord, RegisterOperand, &operand);
+      c->registers[r].pushed = false;
+    }
+  }
+
   Context* c;
 };
 
@@ -2092,7 +2122,10 @@ class MyCompiler: public Compiler {
   }
 
   virtual void push(unsigned size) {
-    ::push(&c, size, value(&c));
+    assert(&c, ceiling(size, BytesPerWord));
+
+    c.state->stack = ::stack
+      (&c, value(&c), ceiling(size, BytesPerWord), c.state->stack);
   }
 
   virtual void push(unsigned size, Operand* value) {
diff --git a/src/x86.cpp b/src/x86.cpp
index 25827bde41..9e1f796a10 100644
--- a/src/x86.cpp
+++ b/src/x86.cpp
@@ -598,6 +598,7 @@ moveRR(Context* c, unsigned size, Assembler::Register* a,
   } else {
     switch (size) {
     case 1:
+      rex(c);
       if (BytesPerWord == 8) {
         c->code.append(0x0f);
       }
@@ -606,6 +607,7 @@ moveRR(Context* c, unsigned size, Assembler::Register* a,
       break;
 
     case 2:
+      rex(c);
       if (BytesPerWord == 8) {
         c->code.append(0x0f);
       }
@@ -675,7 +677,7 @@ move4To8RR(Context* c, unsigned size UNUSED, Assembler::Register* a,
   if (BytesPerWord == 8) {
     rex(c);
     c->code.append(0x63);
-    c->code.append(0xc0 | (a->low << 3) | b->low);
+    c->code.append(0xc0 | (b->low << 3) | a->low);
   } else {
     if (a->low == rax and b->low == rax and b->high == rdx) {
       c->code.append(0x99); // cdq
@@ -897,6 +899,8 @@ divideRR(Context* c, unsigned size, Assembler::Register* a,
          Assembler::Register* b)
 {
   if (BytesPerWord == 4 and size == 8) {
+    abort(c); // todo: sync stack first
+
     Assembler::Register axdx(c->client->acquireTemporary(rax),
                              c->client->acquireTemporary(rdx));
 
@@ -918,10 +922,18 @@ divideRR(Context* c, unsigned size, Assembler::Register* a,
     c->client->releaseTemporary(axdx.high);
   } else {
     Assembler::Register ax(rax);
-    Assembler::Register dx(c->client->acquireTemporary(rdx));
+    Assembler::Register divisor(a->low);
+
+    if (a->low == rdx) {
+      divisor.low = c->client->acquireTemporary();
+      moveRR(c, BytesPerWord, a, &divisor);      
+    } else if (b->low != rdx) {
+      c->client->save(rdx);
+    }
 
     if (b->low != rax) {
-      c->client->acquireTemporary(ax.low);
+      c->client->save(rax);
+      c->client->acquireTemporary(rax);
       moveRR(c, BytesPerWord, b, &ax);
     }
     
@@ -929,14 +941,85 @@ divideRR(Context* c, unsigned size, Assembler::Register* a,
     c->code.append(0x99);
     rex(c);
     c->code.append(0xf7);
-    c->code.append(0xf8 | a->low);
+    c->code.append(0xf8 | divisor.low);
 
     if (b->low != rax) {
       moveRR(c, BytesPerWord, &ax, b);
-      c->client->releaseTemporary(ax.low);
+      c->client->releaseTemporary(rax);
+      c->client->restore(rax);
     }
 
-    c->client->releaseTemporary(dx.low);
+    if (a->low == rdx) {
+      moveRR(c, BytesPerWord, &divisor, a);
+      c->client->releaseTemporary(divisor.low);
+    } else if (b->low != rdx) {
+      c->client->restore(rdx);
+    }
+  }
+}
+
+void
+remainderRR(Context* c, unsigned size, Assembler::Register* a,
+            Assembler::Register* b)
+{
+  if (BytesPerWord == 4 and size == 8) {
+    abort(c); // todo: sync stack first
+
+    Assembler::Register axdx(c->client->acquireTemporary(rax),
+                             c->client->acquireTemporary(rdx));
+
+    pushR(c, size, a);
+    pushR(c, size, b);
+    
+    ResolvedPromise addressPromise(reinterpret_cast<intptr_t>(moduloLong));
+    Assembler::Constant address(&addressPromise);
+    callC(c, BytesPerWord, &address);
+
+    ResolvedPromise offsetPromise(16);
+    Assembler::Constant offset(&offsetPromise);
+    Assembler::Register stack(rsp);
+    addCR(c, BytesPerWord, &offset, &stack);
+
+    moveRR(c, size, &axdx, b);
+
+    c->client->releaseTemporary(axdx.low);
+    c->client->releaseTemporary(axdx.high);
+  } else {
+    Assembler::Register ax(rax);
+    Assembler::Register dx(rdx);
+    Assembler::Register divisor(a->low);
+
+    if (a->low == rdx) {
+      divisor.low = c->client->acquireTemporary();
+      moveRR(c, BytesPerWord, a, &divisor);      
+    } else if (b->low != rdx) {
+      c->client->save(rdx);
+    }
+
+    if (b->low != rax) {
+      c->client->save(rax);
+      c->client->acquireTemporary(rax);
+      moveRR(c, BytesPerWord, b, &ax);
+    }
+    
+    rex(c);
+    c->code.append(0x99);
+    rex(c);
+    c->code.append(0xf7);
+    c->code.append(0xf8 | divisor.low);
+
+    if (b->low != rdx) {
+      moveRR(c, BytesPerWord, &dx, b);
+      c->client->releaseTemporary(rax);
+      c->client->restore(rax);
+    }
+
+    if (a->low == rdx) {
+      moveRR(c, BytesPerWord, &divisor, a);
+      c->client->releaseTemporary(divisor.low);
+    } else if (b->low != rdx) {
+      c->client->restore(rdx);
+    }
   }
 }
 
@@ -1167,6 +1250,8 @@ populateTables()
 
   BinaryOperations[INDEX2(Divide, Register, Register)] = CAST2(divideRR);
 
+  BinaryOperations[INDEX2(Remainder, Register, Register)] = CAST2(remainderRR);
+
   BinaryOperations[INDEX2(And, Constant, Register)] = CAST2(andCR);
   BinaryOperations[INDEX2(And, Constant, Memory)] = CAST2(andCM);