diff --git a/src/db/db/dbNetlist.cc b/src/db/db/dbNetlist.cc
index bf5de8610..4114a7e9f 100644
--- a/src/db/db/dbNetlist.cc
+++ b/src/db/db/dbNetlist.cc
@@ -46,7 +46,7 @@ Pin::Pin (const std::string &name)
 //  Device class implementation
 
 Device::Device ()
-  : mp_device_class (0), m_id (0)
+  : mp_device_class (0), m_id (0), mp_circuit (0)
 {
   //  .. nothing yet ..
 }
@@ -61,13 +61,13 @@ Device::~Device ()
 }
 
 Device::Device (DeviceClass *device_class, const std::string &name)
-  : mp_device_class (device_class), m_name (name), m_id (0)
+  : mp_device_class (device_class), m_name (name), m_id (0), mp_circuit (0)
 {
   //  .. nothing yet ..
 }
 
 Device::Device (const Device &other)
-  : mp_device_class (0), m_id (0)
+  : mp_device_class (0), m_id (0), mp_circuit (0)
 {
   operator= (other);
 }
@@ -81,6 +81,11 @@ Device &Device::operator= (const Device &other)
   return *this;
 }
 
+void Device::set_circuit (Circuit *circuit)
+{
+  mp_circuit = circuit;
+}
+
 void Device::set_name (const std::string &n)
 {
   m_name = n;
@@ -176,7 +181,7 @@ void Device::set_parameter_value (const std::string &name, double v)
 //  SubCircuit class implementation
 
 SubCircuit::SubCircuit ()
-  : m_id (0)
+  : m_id (0), mp_circuit (0)
 {
   //  .. nothing yet ..
 }
@@ -191,13 +196,13 @@ SubCircuit::~SubCircuit()
 }
 
 SubCircuit::SubCircuit (Circuit *circuit, const std::string &name)
-  : m_circuit (circuit), m_name (name), m_id (0)
+  : m_circuit_ref (0), m_name (name), m_id (0), mp_circuit (0)
 {
-  //  .. nothing yet ..
+  set_circuit_ref (circuit);
 }
 
 SubCircuit::SubCircuit (const SubCircuit &other)
-  : m_id (0)
+  : m_id (0), mp_circuit (0)
 {
   operator= (other);
 }
@@ -206,8 +211,8 @@ SubCircuit &SubCircuit::operator= (const SubCircuit &other)
 {
   if (this != &other) {
     m_name = other.m_name;
-    m_circuit = other.m_circuit;
     m_trans = other.m_trans;
+    set_circuit_ref (const_cast<Circuit *> (other.circuit_ref ()));
   }
   return *this;
 }
@@ -230,6 +235,17 @@ void SubCircuit::set_pin_ref_for_pin (size_t pin_id, Net::pin_iterator iter)
   m_pin_refs [pin_id] = iter;
 }
 
+void SubCircuit::set_circuit_ref (Circuit *c)
+{
+  if (m_circuit_ref.get ()) {
+    m_circuit_ref->unregister_ref (this);
+  }
+  m_circuit_ref.reset (c);
+  if (m_circuit_ref.get ()) {
+    m_circuit_ref->register_ref (this);
+  }
+}
+
 const Net *SubCircuit::net_for_pin (size_t pin_id) const
 {
   if (pin_id < m_pin_refs.size ()) {
@@ -349,8 +365,8 @@ const Pin *NetPinRef::pin () const
     if (mp_net && mp_net->circuit ()) {
       return mp_net->circuit ()->pin_by_id (m_pin_id);
     }
-  } else if (mp_subcircuit->circuit ()) {
-    return mp_subcircuit->circuit ()->pin_by_id (m_pin_id);
+  } else if (mp_subcircuit->circuit_ref ()) {
+    return mp_subcircuit->circuit_ref ()->pin_by_id (m_pin_id);
   }
   return 0;
 }
@@ -613,6 +629,8 @@ void Circuit::remove_net (Net *net)
 
 void Circuit::add_device (Device *device)
 {
+  device->set_circuit (this);
+
   size_t id = 0;
   if (! m_devices.empty ()) {
     tl_assert (m_devices.back () != 0);
@@ -658,6 +676,8 @@ Device *Circuit::device_by_id (size_t id)
 
 void Circuit::add_subcircuit (SubCircuit *subcircuit)
 {
+  subcircuit->set_circuit (this);
+
   size_t id = 0;
   if (! m_subcircuits.empty ()) {
     tl_assert (m_subcircuits.back () != 0);
@@ -701,12 +721,22 @@ SubCircuit *Circuit::subcircuit_by_id (size_t id)
   return d != m_subcircuit_id_table.end () ? d->second : 0;
 }
 
+void Circuit::register_ref (SubCircuit *r)
+{
+  m_refs.push_back (r);
+}
+
+void Circuit::unregister_ref (SubCircuit *r)
+{
+  m_refs.erase (r);
+}
+
 void Circuit::translate_circuits (const std::map<const Circuit *, Circuit *> &map)
 {
   for (subcircuit_iterator i = m_subcircuits.begin (); i != m_subcircuits.end (); ++i) {
-    std::map<const Circuit *, Circuit *>::const_iterator m = map.find (i->circuit ());
+    std::map<const Circuit *, Circuit *>::const_iterator m = map.find (i->circuit_ref ());
     tl_assert (m != map.end ());
-    i->set_circuit (m->second);
+    i->set_circuit_ref (m->second);
   }
 }
 
diff --git a/src/db/db/dbNetlist.h b/src/db/db/dbNetlist.h
index cf44b1c84..2530ccc02 100644
--- a/src/db/db/dbNetlist.h
+++ b/src/db/db/dbNetlist.h
@@ -608,6 +608,24 @@ public:
     return m_id;
   }
 
+  /**
+   *  @brief Gets the circuit the device lives in (const version)
+   *  This pointer is 0 if the device isn't added to a circuit
+   */
+  const Circuit *circuit () const
+  {
+    return mp_circuit;
+  }
+
+  /**
+   *  @brief Gets the circuit the device lives in (non-const version)
+   *  This pointer is 0 if the device isn't added to a circuit
+   */
+  Circuit *circuit ()
+  {
+    return mp_circuit;
+  }
+
   /**
    *  @brief Sets the name
    */
@@ -675,6 +693,7 @@ private:
   std::vector<Net::terminal_iterator> m_terminal_refs;
   std::vector<double> m_parameters;
   size_t m_id;
+  Circuit *mp_circuit;
 
   /**
    *  @brief Sets the terminal reference for a specific terminal
@@ -696,6 +715,11 @@ private:
   {
     m_id = id;
   }
+
+  /**
+   *  @brief Sets the circuit
+   */
+  void set_circuit (Circuit *circuit);
 };
 
 /**
@@ -727,7 +751,7 @@ public:
   /**
    *  @brief Creates a subcircuit reference to the given circuit
    */
-  SubCircuit (Circuit *circuit, const std::string &name = std::string ());
+  SubCircuit (Circuit *circuit_ref, const std::string &name = std::string ());
 
   /**
    *  @brief Destructor
@@ -746,19 +770,37 @@ public:
   }
 
   /**
-   *  @brief Gets the circuit the reference points to (const version)
+   *  @brief Gets the circuit the subcircuit lives in (const version)
+   *  This pointer is 0 if the subcircuit isn't added to a circuit
    */
   const Circuit *circuit () const
   {
-    return m_circuit.get ();
+    return mp_circuit;
+  }
+
+  /**
+   *  @brief Gets the circuit the subcircuit lives in (non-const version)
+   *  This pointer is 0 if the subcircuit isn't added to a circuit
+   */
+  Circuit *circuit ()
+  {
+    return mp_circuit;
+  }
+
+  /**
+   *  @brief Gets the circuit the reference points to (const version)
+   */
+  const Circuit *circuit_ref () const
+  {
+    return m_circuit_ref.get ();
   }
 
   /**
    *  @brief Gets the circuit the reference points to (non-const version)
    */
-  Circuit *circuit ()
+  Circuit *circuit_ref ()
   {
-    return m_circuit.get ();
+    return m_circuit_ref.get ();
   }
 
   /**
@@ -820,11 +862,12 @@ private:
   friend class Circuit;
   friend class Net;
 
-  tl::weak_ptr<Circuit> m_circuit;
+  tl::weak_ptr<Circuit> m_circuit_ref;
   std::string m_name;
   db::DCplxTrans m_trans;
   std::vector<Net::pin_iterator> m_pin_refs;
   size_t m_id;
+  Circuit *mp_circuit;
 
   /**
    *  @brief Sets the pin reference for a specific pin
@@ -834,9 +877,14 @@ private:
   /**
    *  @brief Sets the circuit reference
    */
+  void set_circuit_ref (Circuit *c);
+
+  /**
+   *  @brief Sets the circuit the subcircuit belongs to
+   */
   void set_circuit (Circuit *c)
   {
-    m_circuit.reset (c);
+    mp_circuit = c;
   }
 
   /**
@@ -870,6 +918,8 @@ public:
   typedef tl::shared_collection<SubCircuit> subcircuit_list;
   typedef subcircuit_list::const_iterator const_subcircuit_iterator;
   typedef subcircuit_list::iterator subcircuit_iterator;
+  typedef tl::weak_collection<SubCircuit>::const_iterator const_refs_iterator;
+  typedef tl::weak_collection<SubCircuit>::iterator refs_iterator;
 
   /**
    *  @brief Constructor
@@ -939,9 +989,44 @@ public:
     return m_cell_index;
   }
 
+  /**
+   *  @brief Gets the references to this circuit (begin, non-const version)
+   *  This iterator will deliver all subcircuits referencing this circuit
+   */
+  refs_iterator begin_refs ()
+  {
+    return m_refs.begin ();
+  }
+
+  /**
+   *  @brief Gets the references to this circuit (end, non-const version)
+   *  This iterator will deliver all subcircuits referencing this circuit
+   */
+  refs_iterator end_refs ()
+  {
+    return m_refs.end ();
+  }
+
+  /**
+   *  @brief Gets the references to this circuit (begin, const version)
+   *  This iterator will deliver all subcircuits referencing this circuit
+   */
+  const_refs_iterator begin_refs () const
+  {
+    return m_refs.begin ();
+  }
+
+  /**
+   *  @brief Gets the references to this circuit (end, const version)
+   *  This iterator will deliver all subcircuits referencing this circuit
+   */
+  const_refs_iterator end_refs () const
+  {
+    return m_refs.end ();
+  }
+
   /**
    *  @brief Adds a pin to this circuit
-   *
    *  The circuit takes over ownership of the object.
    */
   const Pin &add_pin(const Pin &pin);
@@ -1200,6 +1285,7 @@ public:
 private:
   friend class Netlist;
   friend class Net;
+  friend class SubCircuit;
 
   std::string m_name;
   db::cell_index_type m_cell_index;
@@ -1213,12 +1299,16 @@ private:
   std::map<size_t, Device *> m_device_id_table;
   bool m_valid_subcircuit_id_table;
   std::map<size_t, SubCircuit *> m_subcircuit_id_table;
+  tl::weak_collection<SubCircuit> m_refs;
 
   /**
    *  @brief Sets the pin reference for a specific pin
    */
   void set_pin_ref_for_pin (size_t ppin_id, Net::pin_iterator iter);
 
+  void register_ref (SubCircuit *sc);
+  void unregister_ref (SubCircuit *sc);
+
   void translate_circuits (const std::map<const Circuit *, Circuit *> &map);
   void translate_device_classes (const std::map<const DeviceClass *, DeviceClass *> &map);
   void set_netlist (Netlist *netlist);
diff --git a/src/db/db/dbNetlistDeviceExtractor.cc b/src/db/db/dbNetlistDeviceExtractor.cc
index 5664a0884..f0ce1a7f3 100644
--- a/src/db/db/dbNetlistDeviceExtractor.cc
+++ b/src/db/db/dbNetlistDeviceExtractor.cc
@@ -46,10 +46,10 @@ NetlistDeviceExtractorError::NetlistDeviceExtractorError (const std::string &cel
 // ----------------------------------------------------------------------------------------
 //  NetlistDeviceExtractor implementation
 
-NetlistDeviceExtractor::NetlistDeviceExtractor ()
+NetlistDeviceExtractor::NetlistDeviceExtractor (const std::string &name)
   : mp_layout (0), m_cell_index (0), mp_circuit (0)
 {
-  m_device_name_index = 0;
+  m_name = name;
   m_propname_id = 0;
 }
 
@@ -67,8 +67,7 @@ const tl::Variant &NetlistDeviceExtractor::terminal_property_name ()
 void NetlistDeviceExtractor::initialize (db::Netlist *nl)
 {
   m_layer_definitions.clear ();
-  m_device_classes.clear ();
-  m_device_name_index = 0;
+  mp_device_class = 0;
   m_propname_id = 0;
   m_netlist.reset (nl);
 
@@ -108,16 +107,16 @@ void NetlistDeviceExtractor::extract (db::DeepShapeStore &dss, const NetlistDevi
 
   }
 
-  extract_without_initialize (dss.layout (), dss.initial_cell (), layers, nl);
+  extract_without_initialize (dss.layout (), dss.initial_cell (), layers);
 }
 
 void NetlistDeviceExtractor::extract (db::Layout &layout, db::Cell &cell, const std::vector<unsigned int> &layers, db::Netlist *nl)
 {
   initialize (nl);
-  extract_without_initialize (layout, cell, layers, nl);
+  extract_without_initialize (layout, cell, layers);
 }
 
-void NetlistDeviceExtractor::extract_without_initialize (db::Layout &layout, db::Cell &cell, const std::vector<unsigned int> &layers, db::Netlist *nl)
+void NetlistDeviceExtractor::extract_without_initialize (db::Layout &layout, db::Cell &cell, const std::vector<unsigned int> &layers)
 {
   tl_assert (layers.size () == m_layer_definitions.size ());
 
@@ -216,9 +215,16 @@ void NetlistDeviceExtractor::extract_devices (const std::vector<db::Region> & /*
 
 void NetlistDeviceExtractor::register_device_class (DeviceClass *device_class)
 {
+  if (mp_device_class != 0) {
+    throw tl::Exception (tl::to_string (tr ("Device class already set")));
+  }
+
+  tl_assert (device_class != 0);
+  mp_device_class = device_class;
+  mp_device_class->set_name (m_name);
+
   tl_assert (m_netlist.get () != 0);
   m_netlist->add_device_class (device_class);
-  m_device_classes.push_back (device_class);
 }
 
 void NetlistDeviceExtractor::define_layer (const std::string &name, const std::string &description)
@@ -226,11 +232,14 @@ void NetlistDeviceExtractor::define_layer (const std::string &name, const std::s
   m_layer_definitions.push_back (db::NetlistDeviceExtractorLayerDefinition (name, description, m_layer_definitions.size ()));
 }
 
-Device *NetlistDeviceExtractor::create_device (unsigned int device_class_index)
+Device *NetlistDeviceExtractor::create_device ()
 {
+  if (mp_device_class == 0) {
+    throw tl::Exception (tl::to_string (tr ("No device class registered")));
+  }
+
   tl_assert (mp_circuit != 0);
-  tl_assert (device_class_index < m_device_classes.size ());
-  Device *device = new Device (m_device_classes[device_class_index], tl::to_string (++m_device_name_index));
+  Device *device = new Device (mp_device_class);
   mp_circuit->add_device (device);
   return device;
 }
diff --git a/src/db/db/dbNetlistDeviceExtractor.h b/src/db/db/dbNetlistDeviceExtractor.h
index 5f4d458cf..871fc27a1 100644
--- a/src/db/db/dbNetlistDeviceExtractor.h
+++ b/src/db/db/dbNetlistDeviceExtractor.h
@@ -196,16 +196,24 @@ public:
   typedef std::map<std::string, db::Region *> input_layers;
 
   /**
-   *  @brief Default constructor
+   *  @brief Constructor
+   *
+   *  The name is the name of the device class of the devices generated.
    */
-  NetlistDeviceExtractor ();
+  NetlistDeviceExtractor (const std::string &name);
 
   /**
    *  @brief Destructor
    */
   ~NetlistDeviceExtractor ();
 
-  //  TODO: Do we need to declare input layers?
+  /**
+   *  @brief Gets the name of the extractor and the device class
+   */
+  const std::string &name ()
+  {
+    return m_name;
+  }
 
   /**
    *  @brief Gets the property name for the device terminal annotation
@@ -292,8 +300,7 @@ protected:
    *  defining the device classes and setting up the device layers.
    *
    *  At least one device class needs to be defined. Use "register_device_class" to register
-   *  the device classes you need. The first device class registered has device class index 0,
-   *  the further ones 1, 2, etc.
+   *  the device class you need.
    *
    *  The device layers need to be defined by calling "define_layer" once or several times.
    */
@@ -323,7 +330,8 @@ protected:
   /**
    *  @brief Registers a device class
    *  The device class object will become owned by the netlist and must not be deleted by
-   *  the caller.
+   *  the caller. The name of the device class will be changed to the name given to
+   *  the device extractor.
    *  This method shall be used inside the implementation of "setup" to register
    *  the device classes.
    */
@@ -343,7 +351,7 @@ protected:
    *  The device object returned can be configured by the caller, e.g. set parameters.
    *  It will be owned by the netlist and must not be deleted by the caller.
    */
-  Device *create_device (unsigned int device_class_index = 0);
+  Device *create_device ();
 
   /**
    *  @brief Defines a device terminal in the layout (a polygon)
@@ -436,19 +444,23 @@ private:
   db::properties_id_type m_propname_id;
   db::cell_index_type m_cell_index;
   db::Circuit *mp_circuit;
-  std::vector<db::DeviceClass *> m_device_classes;
+  db::DeviceClass *mp_device_class;
+  std::string m_name;
   layer_definitions m_layer_definitions;
   std::vector<unsigned int> m_layers;
-  unsigned int m_device_name_index;
   error_list m_errors;
 
+  //  no copying
+  NetlistDeviceExtractor (const NetlistDeviceExtractor &);
+  NetlistDeviceExtractor &operator= (const NetlistDeviceExtractor &);
+
   /**
    *  @brief Initializes the extractor
    *  This method will produce the device classes required for the device extraction.
    */
   void initialize (db::Netlist *nl);
 
-  void extract_without_initialize (db::Layout &layout, db::Cell &cell, const std::vector<unsigned int> &layers, db::Netlist *nl);
+  void extract_without_initialize (db::Layout &layout, db::Cell &cell, const std::vector<unsigned int> &layers);
 };
 
 }
diff --git a/src/db/db/gsiDeclDbNetlist.cc b/src/db/db/gsiDeclDbNetlist.cc
index 32abccf50..126bff37c 100644
--- a/src/db/db/gsiDeclDbNetlist.cc
+++ b/src/db/db/gsiDeclDbNetlist.cc
@@ -51,6 +51,9 @@ Class<db::Device> decl_dbDevice ("db", "Device",
   gsi::method ("device_class", &db::Device::device_class,
     "@brief Gets the device class the device belongs to.\n"
   ) +
+  gsi::method ("circuit", (db::Circuit *(db::Device::*) ()) &db::Device::circuit,
+    "@brief Gets the circuit the device lives in."
+  ) +
   gsi::method ("id", &db::Device::id,
     "@brief Gets the device ID.\n"
     "The ID is a unique integer which identifies the device.\n"
@@ -125,9 +128,13 @@ static void subcircuit_disconnect_pin1 (db::SubCircuit *subcircuit, const db::Pi
 }
 
 Class<db::SubCircuit> decl_dbSubCircuit ("db", "SubCircuit",
-  gsi::method ("circuit", (db::Circuit *(db::SubCircuit::*) ()) &db::SubCircuit::circuit,
+  gsi::method ("circuit_ref", (db::Circuit *(db::SubCircuit::*) ()) &db::SubCircuit::circuit_ref,
     "@brief Gets the circuit referenced by the subcircuit.\n"
   ) +
+  gsi::method ("circuit", (db::Circuit *(db::SubCircuit::*) ()) &db::SubCircuit::circuit,
+    "@brief Gets the circuit the subcircuit lives in.\n"
+    "This is NOT the circuit which is referenced. For getting the circuit that the subcircuit references, use \\circuit_ref."
+  ) +
   gsi::method ("id", &db::SubCircuit::id,
     "@brief Gets the subcircuit ID.\n"
     "The ID is a unique integer which identifies the subcircuit.\n"
@@ -620,6 +627,9 @@ Class<db::Circuit> decl_dbCircuit ("db", "Circuit",
     "to the outside through such a pin. The pin is added after all existing "
     "pins. For more details see the \\Pin class."
   ) +
+  gsi::iterator ("each_ref", (db::Circuit::refs_iterator (db::Circuit::*) ()) &db::Circuit::begin_refs, (db::Circuit::refs_iterator (db::Circuit::*) ()) &db::Circuit::end_refs,
+    "@brief Iterates over the subcircuit objects referencing this circuit\n"
+  ) +
   gsi::iterator ("each_pin", (db::Circuit::pin_iterator (db::Circuit::*) ()) &db::Circuit::begin_pins, (db::Circuit::pin_iterator (db::Circuit::*) ()) &db::Circuit::end_pins,
     "@brief Iterates over the pins of the circuit"
   ) +
diff --git a/src/db/unit_tests/dbNetlistDeviceExtractorTests.cc b/src/db/unit_tests/dbNetlistDeviceExtractorTests.cc
index a9eaa5ec8..f8450667b 100644
--- a/src/db/unit_tests/dbNetlistDeviceExtractorTests.cc
+++ b/src/db/unit_tests/dbNetlistDeviceExtractorTests.cc
@@ -54,6 +54,7 @@ namespace {
   {
   public:
     DummyDeviceExtractor ()
+      : db::NetlistDeviceExtractor (std::string ("DUMMY"))
     {
       error ("msg1");
       error ("msg2", db::Box (0, 1, 2, 3));
diff --git a/src/db/unit_tests/dbNetlistExtractorTests.cc b/src/db/unit_tests/dbNetlistExtractorTests.cc
index bbf7c72ab..14f6517c4 100644
--- a/src/db/unit_tests/dbNetlistExtractorTests.cc
+++ b/src/db/unit_tests/dbNetlistExtractorTests.cc
@@ -44,172 +44,6 @@
 #include <memory>
 #include <limits>
 
-class MOSFETExtractor
-  : public db::NetlistDeviceExtractor
-{
-public:
-  MOSFETExtractor (db::Layout *debug_out)
-    : db::NetlistDeviceExtractor (), mp_debug_out (debug_out), m_ldiff (0), m_lgate (0)
-  {
-    if (mp_debug_out) {
-      m_ldiff = mp_debug_out->insert_layer (db::LayerProperties (100, 0));
-      m_lgate = mp_debug_out->insert_layer (db::LayerProperties (101, 0));
-    }
-  }
-
-  virtual void setup ()
-  {
-    define_layer ("PD", "P diffusion");
-    define_layer ("ND", "N diffusion");
-    define_layer ("G", "Gate");
-    define_layer ("P", "Poly");
-
-    db::DeviceClassMOS3Transistor *pmos_class = new db::DeviceClassMOS3Transistor ();
-    pmos_class->set_name ("PMOS");
-    register_device_class (pmos_class);
-
-    db::DeviceClassMOS3Transistor *nmos_class = new db::DeviceClassMOS3Transistor ();
-    nmos_class->set_name ("NMOS");
-    register_device_class (nmos_class);
-  }
-
-  virtual db::Connectivity get_connectivity (const db::Layout & /*layout*/, const std::vector<unsigned int> &layers) const
-  {
-    tl_assert (layers.size () == 4);
-
-    unsigned int lpdiff = layers [0];
-    unsigned int lndiff = layers [1];
-    unsigned int gate = layers [2];
-    //  not used for device recognition: poly (3), but used for producing the gate terminals
-
-    //  The layer definition is pdiff, ndiff, gate
-    db::Connectivity conn;
-    //  collect all connected pdiff
-    conn.connect (lpdiff, lpdiff);
-    //  collect all connected ndiff
-    conn.connect (lndiff, lndiff);
-    //  collect all connected gate shapes
-    conn.connect (gate, gate);
-    //  connect gate with pdiff
-    conn.connect (lpdiff, gate);
-    //  connect gate with ndiff
-    conn.connect (lndiff, gate);
-    return conn;
-  }
-
-  virtual void extract_devices (const std::vector<db::Region> &layer_geometry)
-  {
-    const db::Region &rpdiff = layer_geometry [0];
-    const db::Region &rndiff = layer_geometry [1];
-    const db::Region &rgates = layer_geometry [2];
-
-    for (db::Region::const_iterator p = rgates.begin_merged (); !p.at_end (); ++p) {
-
-      db::Region rgate (*p);
-      db::Region rpdiff_on_gate = rpdiff.selected_interacting (rgate);
-      db::Region rndiff_on_gate = rndiff.selected_interacting (rgate);
-
-      if (! rpdiff_on_gate.empty () && ! rndiff_on_gate.empty ()) {
-        error (tl::to_string (tr ("Gate shape touches both ndiff and pdiff - ignored")), *p);
-      } else if (rpdiff_on_gate.empty () && rndiff_on_gate.empty ()) {
-        error (tl::to_string (tr ("Gate shape touches neither ndiff and pdiff - ignored")), *p);
-      } else {
-
-        bool is_pmos = ! rpdiff_on_gate.empty ();
-
-        db::Region &diff = (is_pmos ? rpdiff_on_gate : rndiff_on_gate);
-        unsigned int terminal_geometry_index = (is_pmos ? 0 : 1);
-        unsigned int gate_geometry_index = 3;
-        unsigned int device_class_index = (is_pmos ? 0 /*PMOS*/ : 1 /*NMOS*/);
-
-        if (diff.size () != 2) {
-          error (tl::sprintf (tl::to_string (tr ("Expected two polygons on diff interacting one gate shape (found %d) - gate shape ignored")), int (diff.size ())), *p);
-          continue;
-        }
-
-        db::Edges edges (rgate.edges () & diff.edges ());
-        if (edges.size () != 2) {
-          error (tl::sprintf (tl::to_string (tr ("Expected two edges interacting gate/diff (found %d) - width and length may be incorrect")), int (edges.size ())), *p);
-          continue;
-        }
-
-        if (! p->is_box ()) {
-          error (tl::to_string (tr ("Gate shape is not a box - width and length may be incorrect")), *p);
-        }
-
-        db::Device *device = create_device (device_class_index);
-
-        device->set_parameter_value ("W", dbu () * edges.length () * 0.5);
-        device->set_parameter_value ("L", dbu () * (p->perimeter () - edges.length ()) * 0.5);
-
-        int diff_index = 0;
-        for (db::Region::const_iterator d = diff.begin (); !d.at_end () && diff_index < 2; ++d, ++diff_index) {
-
-          //  count the number of gate shapes attached to this shape and distribute the area of the
-          //  diffusion region to the number of gates
-          int n = rgates.selected_interacting (db::Region (*d)).size ();
-          tl_assert (n > 0);
-
-          device->set_parameter_value (diff_index == 0 ? "AS" : "AD", dbu () * dbu () * d->area () / double (n));
-
-          define_terminal (device, device->device_class ()->terminal_id_for_name (diff_index == 0 ? "S" : "D"), terminal_geometry_index, *d);
-
-        }
-
-        define_terminal (device, device->device_class ()->terminal_id_for_name ("G"), gate_geometry_index, *p);
-
-        //  output the device for debugging
-        device_out (device, diff, rgate);
-
-      }
-
-    }
-  }
-
-private:
-  db::Layout *mp_debug_out;
-  unsigned int m_ldiff, m_lgate;
-
-  void device_out (const db::Device *device, const db::Region &diff, const db::Region &gate)
-  {
-    if (! mp_debug_out) {
-      return;
-    }
-
-    std::string cn = layout ()->cell_name (cell_index ());
-    std::pair<bool, db::cell_index_type> target_cp = mp_debug_out->cell_by_name (cn.c_str ());
-    tl_assert (target_cp.first);
-
-    db::cell_index_type dci = mp_debug_out->add_cell ((device->device_class ()->name () + "_" + device->name ()).c_str ());
-    mp_debug_out->cell (target_cp.second).insert (db::CellInstArray (db::CellInst (dci), db::Trans ()));
-
-    db::Cell &device_cell = mp_debug_out->cell (dci);
-    for (db::Region::const_iterator p = diff.begin (); ! p.at_end (); ++p) {
-      device_cell.shapes (m_ldiff).insert (*p);
-    }
-    for (db::Region::const_iterator p = gate.begin (); ! p.at_end (); ++p) {
-      device_cell.shapes (m_lgate).insert (*p);
-    }
-
-    std::string ps;
-    const std::vector<db::DeviceParameterDefinition> &pd = device->device_class ()->parameter_definitions ();
-    for (std::vector<db::DeviceParameterDefinition>::const_iterator i = pd.begin (); i != pd.end (); ++i) {
-      if (! ps.empty ()) {
-        ps += ",";
-      }
-      ps += i->name () + "=" + tl::to_string (device->parameter_value (i->id ()));
-    }
-    device_cell.shapes (m_ldiff).insert (db::Text (ps, db::Trans (diff.bbox ().center () - db::Point ())));
-  }
-};
-
-static unsigned int define_layer (db::Layout &ly, db::LayerMap &lmap, int gds_layer, int gds_datatype = 0)
-{
-  unsigned int lid = ly.insert_layer (db::LayerProperties (gds_layer, gds_datatype));
-  lmap.map (ly.get_properties (lid), lid);
-  return lid;
-}
-
 static unsigned int layer_of (const db::Region &region)
 {
   return db::DeepLayer (region).layer ();
@@ -248,6 +82,153 @@ static std::string pin_name (const db::Pin &pin)
   }
 }
 
+
+class MOSFETExtractor
+  : public db::NetlistDeviceExtractor
+{
+public:
+  MOSFETExtractor (const std::string &name, db::Layout *debug_out)
+    : db::NetlistDeviceExtractor (name), mp_debug_out (debug_out), m_ldiff (0), m_lgate (0)
+  {
+    if (mp_debug_out) {
+      m_ldiff = mp_debug_out->insert_layer (db::LayerProperties (100, 0));
+      m_lgate = mp_debug_out->insert_layer (db::LayerProperties (101, 0));
+    }
+  }
+
+  virtual void setup ()
+  {
+    define_layer ("SD", "Source/drain diffusion");
+    define_layer ("G", "Gate");
+    define_layer ("P", "Poly");
+
+    register_device_class (new db::DeviceClassMOS3Transistor ());
+  }
+
+  virtual db::Connectivity get_connectivity (const db::Layout & /*layout*/, const std::vector<unsigned int> &layers) const
+  {
+    tl_assert (layers.size () == 3);
+
+    unsigned int diff = layers [0];
+    unsigned int gate = layers [1];
+    //  not used for device recognition: poly (2), but used for producing the gate terminals
+
+    //  The layer definition is diff, gate
+    db::Connectivity conn;
+    //  collect all connected diffusion shapes
+    conn.connect (diff, diff);
+    //  collect all connected gate shapes
+    conn.connect (gate, gate);
+    //  connect gate with diff to detect gate/diffusion boundary
+    conn.connect (diff, gate);
+    return conn;
+  }
+
+  virtual void extract_devices (const std::vector<db::Region> &layer_geometry)
+  {
+    const db::Region &rdiff = layer_geometry [0];
+    const db::Region &rgates = layer_geometry [1];
+
+    for (db::Region::const_iterator p = rgates.begin_merged (); !p.at_end (); ++p) {
+
+      db::Region rgate (*p);
+      db::Region rdiff2gate = rdiff.selected_interacting (rgate);
+
+      if (rdiff2gate.empty ()) {
+        error (tl::to_string (tr ("Gate shape touches no diffusion - ignored")), *p);
+      } else {
+
+        unsigned int terminal_geometry_index = 0;
+        unsigned int gate_geometry_index = 2;
+
+        if (rdiff2gate.size () != 2) {
+          error (tl::sprintf (tl::to_string (tr ("Expected two polygons on diff interacting one gate shape (found %d) - gate shape ignored")), int (rdiff2gate.size ())), *p);
+          continue;
+        }
+
+        db::Edges edges (rgate.edges () & rdiff2gate.edges ());
+        if (edges.size () != 2) {
+          error (tl::sprintf (tl::to_string (tr ("Expected two edges interacting gate/diff (found %d) - width and length may be incorrect")), int (edges.size ())), *p);
+          continue;
+        }
+
+        if (! p->is_box ()) {
+          error (tl::to_string (tr ("Gate shape is not a box - width and length may be incorrect")), *p);
+        }
+
+        db::Device *device = create_device ();
+
+        device->set_parameter_value ("W", dbu () * edges.length () * 0.5);
+        device->set_parameter_value ("L", dbu () * (p->perimeter () - edges.length ()) * 0.5);
+
+        int diff_index = 0;
+        for (db::Region::const_iterator d = rdiff2gate.begin (); !d.at_end () && diff_index < 2; ++d, ++diff_index) {
+
+          //  count the number of gate shapes attached to this shape and distribute the area of the
+          //  diffusion region to the number of gates
+          int n = rgates.selected_interacting (db::Region (*d)).size ();
+          tl_assert (n > 0);
+
+          device->set_parameter_value (diff_index == 0 ? "AS" : "AD", dbu () * dbu () * d->area () / double (n));
+
+          define_terminal (device, device->device_class ()->terminal_id_for_name (diff_index == 0 ? "S" : "D"), terminal_geometry_index, *d);
+
+        }
+
+        define_terminal (device, device->device_class ()->terminal_id_for_name ("G"), gate_geometry_index, *p);
+
+        //  output the device for debugging
+        device_out (device, rdiff2gate, rgate);
+
+      }
+
+    }
+  }
+
+private:
+  db::Layout *mp_debug_out;
+  unsigned int m_ldiff, m_lgate;
+
+  void device_out (const db::Device *device, const db::Region &diff, const db::Region &gate)
+  {
+    if (! mp_debug_out) {
+      return;
+    }
+
+    std::string cn = layout ()->cell_name (cell_index ());
+    std::pair<bool, db::cell_index_type> target_cp = mp_debug_out->cell_by_name (cn.c_str ());
+    tl_assert (target_cp.first);
+
+    db::cell_index_type dci = mp_debug_out->add_cell ((device->device_class ()->name () + "_" + device->circuit ()->name () + "_" + device_name (*device)).c_str ());
+    mp_debug_out->cell (target_cp.second).insert (db::CellInstArray (db::CellInst (dci), db::Trans ()));
+
+    db::Cell &device_cell = mp_debug_out->cell (dci);
+    for (db::Region::const_iterator p = diff.begin (); ! p.at_end (); ++p) {
+      device_cell.shapes (m_ldiff).insert (*p);
+    }
+    for (db::Region::const_iterator p = gate.begin (); ! p.at_end (); ++p) {
+      device_cell.shapes (m_lgate).insert (*p);
+    }
+
+    std::string ps;
+    const std::vector<db::DeviceParameterDefinition> &pd = device->device_class ()->parameter_definitions ();
+    for (std::vector<db::DeviceParameterDefinition>::const_iterator i = pd.begin (); i != pd.end (); ++i) {
+      if (! ps.empty ()) {
+        ps += ",";
+      }
+      ps += i->name () + "=" + tl::to_string (device->parameter_value (i->id ()));
+    }
+    device_cell.shapes (m_ldiff).insert (db::Text (ps, db::Trans (diff.bbox ().center () - db::Point ())));
+  }
+};
+
+static unsigned int define_layer (db::Layout &ly, db::LayerMap &lmap, int gds_layer, int gds_datatype = 0)
+{
+  unsigned int lid = ly.insert_layer (db::LayerProperties (gds_layer, gds_datatype));
+  lmap.map (ly.get_properties (lid), lid);
+  return lid;
+}
+
 //  @@@ TODO: move this somewhere else
 
 static void dump_nets (const db::Netlist &nl, const db::hier_clusters<db::PolygonRef> &clusters, db::Layout &ly, const std::map<unsigned int, unsigned int> &lmap, const db::CellMapping &cmap)
@@ -329,14 +310,15 @@ static std::string netlist2string (const db::Netlist &nl)
     for (db::Circuit::const_subcircuit_iterator sc = c->begin_subcircuits (); sc != c->end_subcircuits (); ++sc) {
       std::string ps;
       const db::SubCircuit &subcircuit = *sc;
-      for (db::Circuit::const_pin_iterator p = sc->circuit ()->begin_pins (); p != sc->circuit ()->end_pins (); ++p) {
-        if (p != sc->circuit ()->begin_pins ()) {
+      const db::Circuit *circuit = sc->circuit_ref ();
+      for (db::Circuit::const_pin_iterator p = circuit->begin_pins (); p != circuit->end_pins (); ++p) {
+        if (p != circuit->begin_pins ()) {
           ps += ",";
         }
         const db::Pin &pin = *p;
         ps += pin_name (pin) + "=" + net_name (subcircuit.net_for_pin (pin.id ()));
       }
-      res += std::string ("  X") + sc->circuit ()->name () + " " + subcircuit_name (*sc) + " (" + ps + ")\n";
+      res += std::string ("  X") + circuit->name () + " " + subcircuit_name (*sc) + " (" + ps + ")\n";
     }
 
   }
@@ -344,7 +326,7 @@ static std::string netlist2string (const db::Netlist &nl)
   return res;
 }
 
-TEST(2_DeviceAndNetExtraction)
+TEST(1_DeviceAndNetExtraction)
 {
   db::Layout ly;
   db::LayerMap lmap;
@@ -396,10 +378,14 @@ TEST(2_DeviceAndNetExtraction)
   db::Region rmetal2_lbl (db::RecursiveShapeIterator (ly, tc, metal2_lbl), dss);
 
   //  derived regions
-  db::Region rgate  = ractive & rpoly;
-  db::Region rsd    = ractive - rgate;
-  db::Region rpdiff = rsd & rnwell;
-  db::Region rndiff = rsd - rnwell;
+
+  db::Region rpactive = ractive & rnwell;
+  db::Region rpgate   = rpactive & rpoly;
+  db::Region rpsd     = rpactive - rpgate;
+
+  db::Region rnactive = ractive - rnwell;
+  db::Region rngate   = rnactive & rpoly;
+  db::Region rnsd     = rnactive - rngate;
 
   //  return the computed layers into the original layout and write it for debugging purposes
 
@@ -408,10 +394,12 @@ TEST(2_DeviceAndNetExtraction)
   unsigned int lpdiff = ly.insert_layer (db::LayerProperties (12, 0));      // 12/0 -> P Diffusion
   unsigned int lndiff = ly.insert_layer (db::LayerProperties (13, 0));      // 13/0 -> N Diffusion
 
-  rgate.insert_into (&ly, tc.cell_index (), lgate);
-  rsd.insert_into (&ly, tc.cell_index (), lsd);
-  rpdiff.insert_into (&ly, tc.cell_index (), lpdiff);
-  rndiff.insert_into (&ly, tc.cell_index (), lndiff);
+  rpgate.insert_into (&ly, tc.cell_index (), lgate);
+  rngate.insert_into (&ly, tc.cell_index (), lgate);
+  rpsd.insert_into (&ly, tc.cell_index (), lsd);
+  rnsd.insert_into (&ly, tc.cell_index (), lsd);
+  rpsd.insert_into (&ly, tc.cell_index (), lpdiff);
+  rnsd.insert_into (&ly, tc.cell_index (), lndiff);
 
   //  perform the extraction
 
@@ -420,15 +408,20 @@ TEST(2_DeviceAndNetExtraction)
   //  NOTE: the device extractor will add more debug layers for the transistors:
   //    20/0 -> Diffusion
   //    21/0 -> Gate
-  MOSFETExtractor ex (&ly);
+  MOSFETExtractor pmos_ex ("PMOS", &ly);
+  MOSFETExtractor nmos_ex ("NMOS", &ly);
 
   db::NetlistDeviceExtractor::input_layers dl;
-  dl["PD"] = &rpdiff;
-  dl["ND"] = &rndiff;
-  dl["G"] = &rgate;
-  dl["P"] = &rpoly;
 
-  ex.extract (dss, dl, &nl);
+  dl["SD"] = &rpsd;
+  dl["G"] = &rpgate;
+  dl["P"] = &rpoly;  //  not needed for extraction but to return terminal shapes
+  pmos_ex.extract (dss, dl, &nl);
+
+  dl["SD"] = &rnsd;
+  dl["G"] = &rngate;
+  dl["P"] = &rpoly;  //  not needed for extraction but to return terminal shapes
+  nmos_ex.extract (dss, dl, &nl);
 
   //  perform the net extraction
 
@@ -436,8 +429,8 @@ TEST(2_DeviceAndNetExtraction)
 
   db::Connectivity conn;
   //  Intra-layer
-  conn.connect (rpdiff);
-  conn.connect (rndiff);
+  conn.connect (rpsd);
+  conn.connect (rnsd);
   conn.connect (rpoly);
   conn.connect (rdiff_cont);
   conn.connect (rpoly_cont);
@@ -445,8 +438,8 @@ TEST(2_DeviceAndNetExtraction)
   conn.connect (rvia1);
   conn.connect (rmetal2);
   //  Inter-layer
-  conn.connect (rpdiff,     rdiff_cont);
-  conn.connect (rndiff,     rdiff_cont);
+  conn.connect (rpsd,       rdiff_cont);
+  conn.connect (rnsd,       rdiff_cont);
   conn.connect (rpoly,      rpoly_cont);
   conn.connect (rpoly_cont, rmetal1);
   conn.connect (rdiff_cont, rmetal1);
@@ -468,9 +461,11 @@ TEST(2_DeviceAndNetExtraction)
   //    206/0 -> Metal1
   //    207/0 -> Via1
   //    208/0 -> Metal2
+  //    210/0 -> N source/drain
+  //    211/0 -> P source/drain
   std::map<unsigned int, unsigned int> dump_map;
-  dump_map [layer_of (rpdiff)    ] = ly.insert_layer (db::LayerProperties (210, 0));
-  dump_map [layer_of (rndiff)    ] = ly.insert_layer (db::LayerProperties (211, 0));
+  dump_map [layer_of (rpsd)      ] = ly.insert_layer (db::LayerProperties (210, 0));
+  dump_map [layer_of (rnsd)      ] = ly.insert_layer (db::LayerProperties (211, 0));
   dump_map [layer_of (rpoly)     ] = ly.insert_layer (db::LayerProperties (203, 0));
   dump_map [layer_of (rdiff_cont)] = ly.insert_layer (db::LayerProperties (204, 0));
   dump_map [layer_of (rpoly_cont)] = ly.insert_layer (db::LayerProperties (205, 0));
@@ -496,10 +491,10 @@ TEST(2_DeviceAndNetExtraction)
     "  XINV2 $9 ($1=$I6,$2=$I45,$3=$I7,$4=VSS,$5=VDD)\n"
     "  XINV2 $10 ($1=$I7,$2=$I46,$3=$I8,$4=VSS,$5=VDD)\n"
     "Circuit INV2 ($1=IN,$2=$2,$3=OUT,$4=$4,$5=$5):\n"
-    "  DPMOS 1 (S=$2,G=IN,D=$5) [L=0.25,W=0.95,AS=0.49875,AD=0.26125]\n"
-    "  DPMOS 2 (S=$5,G=$2,D=OUT) [L=0.25,W=0.95,AS=0.26125,AD=0.49875]\n"
-    "  DNMOS 3 (S=$2,G=IN,D=$4) [L=0.25,W=0.95,AS=0.49875,AD=0.26125]\n"
-    "  DNMOS 4 (S=$4,G=$2,D=OUT) [L=0.25,W=0.95,AS=0.26125,AD=0.49875]\n"
+    "  DPMOS $1 (S=$2,G=IN,D=$5) [L=0.25,W=0.95,AS=0.49875,AD=0.26125]\n"
+    "  DPMOS $2 (S=$5,G=$2,D=OUT) [L=0.25,W=0.95,AS=0.26125,AD=0.49875]\n"
+    "  DNMOS $3 (S=$2,G=IN,D=$4) [L=0.25,W=0.95,AS=0.49875,AD=0.26125]\n"
+    "  DNMOS $4 (S=$4,G=$2,D=OUT) [L=0.25,W=0.95,AS=0.26125,AD=0.49875]\n"
     "  XTRANS $1 ($1=$2,$2=$4,$3=IN)\n"
     "  XTRANS $2 ($1=$2,$2=$5,$3=IN)\n"
     "  XTRANS $3 ($1=$5,$2=OUT,$3=$2)\n"
@@ -523,19 +518,9 @@ TEST(2_DeviceAndNetExtraction)
 //  An attempt to simplify things.
 
 /*
-- layers: use db::Region, or wrapper?
--> use regions, but test whether they are deep regions (?)
-
 TODO:
 - netlist query functions such as net_by_name, device_by_name, circuit_by_name
 - terminal geometry (Polygon) for device, combined device geometry (all terminals)
-- error interface for device extraction
-  //  gets the device extraction errors
-  //  device_extraction_error_iterator begin_device_extraction_errors () const;
-  //  device_extraction_error_iterator end_device_extraction_errors () const;
-  //  bool has_device_extraction_errors () const;
-
-- device extractor needs to declare the layers to allow passing them by name
 - netlist manipulation methods (i.e. flatten certain cells, purging etc.)
 */
 
@@ -551,23 +536,25 @@ public:
   //  the iterator provides the hierarchical selection (enabling/disabling cells etc.)
   LayoutToNetlist (const db::RecursiveShapeIterator &iter);
 
-  //  --- preparation
+  //  --- Step 0: configuration
 
-  //  returns a new'd region
+  void set_threads (unsigned int n);
+  void set_area_ratio (double ar);
+  void set_max_vertex_count (size_t n);
+
+  //  --- Step 1: preparation
+
+  //  returns new'd regions
   db::Region *make_layer (unsigned int layer_index);
   db::Region *make_text_layer (unsigned int layer_index);
   db::Region *make_polygon_layer (unsigned int layer_index);
 
-  //  gets the internal layout and cell
-  const db::Layout &internal_layout () const;
-  const db::Cell &internal_top_cell () const;
-
-  //  --- device extraction
+  //  --- Step 2: device extraction
 
   //  after this, the device extractor will have errors if some occured.
-  void extract_devices (db::NetlistDeviceExtractor *extractor, const std::map<std::string, const db::Region *> &layers);
+  void extract_devices (db::NetlistDeviceExtractor &extractor, const std::map<std::string, db::Region *> &layers);
 
-  //  --- net extraction
+  //  --- Step 3: net extraction
 
   //  define connectivity for the netlist extraction
   void connect (const db::Region &l);
@@ -576,7 +563,11 @@ public:
   //  runs the netlist extraction
   void extract_netlist ();
 
-  //  --- retrieval
+  //  --- Step 4: retrieval
+
+  //  gets the internal layout and cell (0 if not available)
+  const db::Layout *internal_layout () const;
+  const db::Cell *internal_top_cell () const;
 
   //  gets the internal layer index of the given region
   unsigned int layer_of (const db::Region &region) const;
@@ -600,6 +591,11 @@ public:
   //  copies the shapes of the given net from a given layer
   //  (recursive true: include nets from subcircuits)
   db::Region shapes_of_net (const db::Net &net, const db::Region &of_layer, bool recursive);
+
+  //  finds the net by probing a specific location on the given layer looking through the
+  //  hierarchy. Returns 0 if no net was found.
+  db::Net *probe_net (const db::Region &of_region, const db::DPoint &point);
+  db::Net *probe_net (const db::Region &of_region, const db::Point &point);
 };
 
 }
diff --git a/src/db/unit_tests/dbNetlistTests.cc b/src/db/unit_tests/dbNetlistTests.cc
index e1df6d1df..9a2a172d4 100644
--- a/src/db/unit_tests/dbNetlistTests.cc
+++ b/src/db/unit_tests/dbNetlistTests.cc
@@ -165,7 +165,7 @@ static std::string net2string (const db::Net &n)
       res += ",";
     }
     if (i->subcircuit ()) {
-      res += i->subcircuit ()->circuit () ? i->subcircuit ()->circuit ()->name () : "(null)";
+      res += i->subcircuit ()->circuit_ref () ? i->subcircuit ()->circuit_ref ()->name () : "(null)";
       res += ":";
     } else {
       res += "+";
@@ -221,15 +221,15 @@ static std::string netlist2 (const db::Circuit &c)
   }
 
   for (db::Circuit::const_subcircuit_iterator s = c.begin_subcircuits (); s != c.end_subcircuits (); ++s) {
-    if (! s->circuit ()) {
+    if (! s->circuit_ref ()) {
       continue;
     }
     pins.clear ();
-    for (size_t i = 0; i < s->circuit ()->pin_count (); ++i) {
+    for (size_t i = 0; i < s->circuit_ref ()->pin_count (); ++i) {
       if (! pins.empty ()) {
         pins += ",";
       }
-      pins += s->circuit ()->pin_by_id (i)->name ();
+      pins += s->circuit_ref ()->pin_by_id (i)->name ();
       pins += "=";
       const db::Net *net = s->net_for_pin (i);
       pins += net ? net->name () : std::string ("(null)");
@@ -268,7 +268,11 @@ TEST(4_CircuitDevices)
   db::Device *d1 = new db::Device (&dc1, "d1");
   db::Device *d2a = new db::Device (&dc2, "d2a");
   db::Device *d2b = new db::Device (&dc2, "d2b");
+
+  EXPECT_EQ (d1->circuit () == 0, true);
+
   c->add_device (d1);
+  EXPECT_EQ (d1->circuit () == c.get (), true);
   EXPECT_EQ (d1->id (), size_t (1));
   EXPECT_EQ (c->device_by_id (d1->id ()) == d1, true);
   c->add_device (dd);
@@ -386,6 +390,18 @@ static std::string netlist2 (const db::Netlist &nl)
   return res;
 }
 
+static std::string refs2string (const db::Circuit *c)
+{
+  std::string res;
+  for (db::Circuit::const_refs_iterator r = c->begin_refs (); r != c->end_refs (); ++r) {
+    if (!res.empty ()) {
+      res += ",";
+    }
+    res += r->name ();
+  }
+  return res;
+}
+
 TEST(4_NetlistSubcircuits)
 {
   std::auto_ptr<db::Netlist> nl (new db::Netlist ());
@@ -413,14 +429,19 @@ TEST(4_NetlistSubcircuits)
   db::Device *d = new db::Device (dc, "D");
   c2->add_device (d);
 
+  EXPECT_EQ (refs2string (c2), "");
   db::SubCircuit *sc1 = new db::SubCircuit (c2);
   sc1->set_name ("sc1");
+  EXPECT_EQ (refs2string (c2), "sc1");
+  EXPECT_EQ (sc1->circuit () == 0, true);
   c1->add_subcircuit (sc1);
+  EXPECT_EQ (sc1->circuit () == c1, true);
   EXPECT_EQ (sc1->id (), size_t (1));
   EXPECT_EQ (c1->subcircuit_by_id (sc1->id ()) == sc1, true);
 
   db::SubCircuit *sc2 = new db::SubCircuit (c2);
   sc2->set_name ("sc2");
+  EXPECT_EQ (refs2string (c2), "sc1,sc2");
   c1->add_subcircuit (sc2);
   EXPECT_EQ (sc2->id (), size_t (2));
   EXPECT_EQ (c1->subcircuit_by_id (sc2->id ()) == sc2, true);
@@ -778,3 +799,38 @@ TEST(10_NetPinRefBasics)
   EXPECT_EQ (db::NetPinRef (&d1, 1) < db::NetPinRef (&d1, 0), false);
   EXPECT_NE ((db::NetPinRef (&d1, 0) < db::NetPinRef (&d2, 0)), (db::NetPinRef (&d2, 0) < db::NetPinRef (&d1, 0)));
 }
+
+TEST(11_NetlistCircuitRefs)
+{
+  std::auto_ptr<db::Netlist> nl (new db::Netlist ());
+
+  db::Circuit *c1 = new db::Circuit ();
+  c1->set_name ("c1");
+  nl->add_circuit (c1);
+
+  db::Circuit *c2 = new db::Circuit ();
+  c2->set_name ("c2");
+  nl->add_circuit (c2);
+
+  db::SubCircuit *sc1 = new db::SubCircuit (c2);
+  sc1->set_name ("sc1");
+  EXPECT_EQ (refs2string (c2), "sc1");
+  c1->add_subcircuit (sc1);
+
+  db::SubCircuit *sc2 = new db::SubCircuit (c2);
+  sc2->set_name ("sc2");
+  EXPECT_EQ (refs2string (c2), "sc1,sc2");
+  c1->add_subcircuit (sc2);
+
+  *sc2 = db::SubCircuit ();
+  EXPECT_EQ (refs2string (c2), "sc1");
+
+  db::SubCircuit sc3 (c2);
+  sc3.set_name ("sc3");
+  *sc2 = sc3;
+  sc2->set_name ("sc2");
+  EXPECT_EQ (refs2string (c2), "sc1,sc3,sc2");
+
+  sc3 = db::SubCircuit ();
+  EXPECT_EQ (refs2string (c2), "sc1,sc2");
+}
diff --git a/testdata/algo/device_extract_au1.gds b/testdata/algo/device_extract_au1.gds
index 05c3b013f..a970a662c 100644
Binary files a/testdata/algo/device_extract_au1.gds and b/testdata/algo/device_extract_au1.gds differ
diff --git a/testdata/ruby/dbNetlist.rb b/testdata/ruby/dbNetlist.rb
index eeead78df..04249daab 100644
--- a/testdata/ruby/dbNetlist.rb
+++ b/testdata/ruby/dbNetlist.rb
@@ -258,22 +258,35 @@ class DBNetlist_TestClass < TestBase
 
     sc1 = c.create_subcircuit(cc)
     sc1.name = "SC1"
+    assert_equal(sc1.circuit.object_id, c.object_id)
     assert_equal(sc1.name, "SC1")
-    assert_equal(sc1.circuit.name, "CC")
+    assert_equal(sc1.circuit_ref.name, "CC")
     assert_equal(c.subcircuit_by_id(sc1.id).id, 1)
     assert_equal(c.subcircuit_by_id(2).inspect, "nil")
 
+    refs = []
+    cc.each_ref { |r| refs << r.name }
+    assert_equal(refs.join(","), "SC1")
+
     sc2 = c.create_subcircuit(cc)
     sc2.name = "SC2"
 
+    refs = []
+    cc.each_ref { |r| refs << r.name }
+    assert_equal(refs.join(","), "SC1,SC2")
+
     names = []
     ccn = []
-    c.each_subcircuit { |sc| names << sc.name; ccn << sc.circuit.name }
+    c.each_subcircuit { |sc| names << sc.name; ccn << sc.circuit_ref.name }
     assert_equal(names, [ "SC1", "SC2" ])
     assert_equal(ccn, [ "CC", "CC" ])
 
     c.remove_subcircuit(sc2)
     
+    refs = []
+    cc.each_ref { |r| refs << r.name }
+    assert_equal(refs.join(","), "SC1")
+
     names = []
     c.each_subcircuit { |sc| names << sc.name}
     assert_equal(names, [ "SC1" ])
@@ -282,7 +295,7 @@ class DBNetlist_TestClass < TestBase
 
     names = []
     ccn = []
-    c.each_subcircuit { |sc| names << sc.name; ccn << sc.circuit.name }
+    c.each_subcircuit { |sc| names << sc.name; ccn << sc.circuit_ref.name }
     assert_equal(names, [ "SC1", "SC2" ])
     assert_equal(ccn, [ "CC", "CC" ])
 
@@ -531,6 +544,8 @@ class DBNetlist_TestClass < TestBase
     nl.add(c)
     d1 = c.create_device(dc)
 
+    assert_equal(d1.circuit.object_id, c.object_id)
+
     assert_equal(d1.parameter(0), 1.0)
     assert_equal(d1.parameter("U"), 1.0)
     assert_equal(d1.parameter(1), 2.0)