SELF extension

Self is a prototype-based OO language that uses objects that only have slots. This page presents a Tcl package by the same name that was inspired by Self.

(For the self TclOO command, see self - TclOO)

See Also

the pure Tcl extension which inspired this extension


MJ: In a recent (oct-2006) OO frenzy on the Tcl chatroom resulting in code gems like neo and eos, I decided to write a Self-like extension for Tcl.

Initially this was a C extension to allow payload with procedure definitions. But I have rewritten it in plain Tcl using [interp alias] available from

After loading the extension, one object (Object) is known and the two commands (self and next) are imported. Every other object in the system will be cloned or copied from Object or its clones and will have slots defined containing values or methods.


The generic prototype Object can be used to create new objects with clone or copy. The generic object contains the following slots:

  • clone new - creates a cloned object with name $new and sets the clone receiver as parent
  • copy new - creates a copy of the receiver with name $new
  • name: value - create a value slot name that will return $value
  • name: args body - will create a method slot name
  • parents* - a value or method slot that returns a list with all the parents of the object. This information is used for method dispatch. Object initially has no parents.
  • _state - Show the internal object state (use for debugging only). Because it doesn't use the dispatcher this can't be overridden or redefined.

NB: The slot assignment syntax has be changed from slot name to name: to be less verbose and better in line with Self.


clone will create a new object with only the parents* slot defined. The parents* slot will contain the name of the receiver of the clone message. The parents* slot will be used in slot dispatch resulting in inheritance and can be updated during runtime, allowing for mixin-like behaviour.

cloned object

  • parents* {receiver of the clone message}


copy will create an exact copy of the receiver (in other words: _state is exactly the same).

Command dispatch

Slots are executed by sending messages to the object. A depth-first search of the object slots and the parents* list will be done to find the implementation of a slot. When a slot is found the slot is executed in the context of the object receiving the message.

If no implementation of a slot is found, the dispatcher tries to call an unknown slot with the slotname and arguments as args. If that also fails, a standard error is returned.

During evaluation of a slot, three additional commands are available:

  • self slotname args: will call slot $slotname on the receiver of the initial message. Without arguments it will return the name of the receiver.
  • next: will call the same slot with the same args as currently executing, but the dispatcher will start looking for the slot only in the parents of the implementer of the currently executing slot, which is not the same as self for inherited slots.
  • next slotname args: will call a different slot $slotname on the parents of the implementor of the currently executing slot.


Currently there are some missing features which would be nice to have when using the extension:

  • Variable slots are not real variables, so it is not possible to add traces to them.
  • Currently method dispatch in a very deep parent child chain is slow. Doing dispatch on a slot that is defined 999 objects higher in the inheritance tree takes approximately 1000 microseconds on my machine, where as in the TIP 257 [L1 ] implementation it takes only 12. This can be resolved using slot/call chain caching.
  • It would be nice to be able to delegate methods based on their signature (an unknown on steroids), which is very useful for building megawidgets. For instance:
 a slot delegate* {}
 a delegate* {{cget .t} {* {self unknown}}}
 # the unknown slot is now a normal slot.
 # delegates will be called with the slotname and args


package require self

# create a Point object.
Object clone Point

# add a to_s slot to display information of the object
Object to_s: {} {
    return "[self]"

# add x and y slots for the point, notice that these slots cannot be called for now.
Point x: {args} {error "abstract slot, override in clone"}
Point y: {args} {error "abstract slot, override in clone"} 

# extend default behavior from parent (Object)
Point to_s: {} {
    return "id: [next] ([self x],[self y])"
    # Here next will search for a slot named to_s in the parents of the implementor of the current method (Point)
    # finding the Object slot to_s and the execute it in the context of the receiver (which will be a clone of Point) 

# define a point factory
Point create: {name x y} {
    self clone $name
    $name x: $x
    $name y: $y

# clone a Point
Point clone p1

# to_s will fail because the x and y slots in Point are called
catch {p1 to_s} err
puts $err

# use the Point factory which will define x and y slots
Point create p1 0 0

# to_s will now work
puts [p1 to_s]

Intercepting slot calls for debugging purposes

Object clone A
A test: args {return}

A clone a
a test

A clone debug
debug test: {args} {puts "called test with $args"; next}

a parents*: {debug}
a test 1 2 3

Demonstrating unknown to create a read-only text widget

 # example demostrating how to override a widget
 package require self
 package require Tk

 proc debugtext name {
        text $name
        rename $name _$name
        Object clone $name
        $name unknown: args {
           puts "[self] $args"
           _[self] {*}$args
        $name destroy: {} {
                destroy _[self]
                rename _[self] {}
        return $name

 debugtext .t 

 pack .t -expand 1 -fill both
 button .b -text "Make readonly" -command make_ro
 pack .b
 proc make_ro {} {
  # allows on the fly redefining of behaviour
  .t insert: args {puts stderr "readonly"}
  .t delete: args {puts stderr "readonly"}

With all of these Self-like extensions, is it possible to make singleton objects by removing the clone function? Does that even make sense for prototype-based object systems? -- escargo 20 Oct 2006

NEM: A singleton would just be an object. Perhaps an example of what you would be using the singleton for would be useful? I tend to avoid singletons. About the only place I use them is when defining the base case of some structure (e.g., if you define a binary search tree as two cases: Branch(left,right,val) and Empty, then the Empty case can be a singleton).

MJ: As NEM already mentions above, a singleton only makes sense in a class based OO system where you want to instantiate a class only once. In a prototype based OO system everything is a singleton (there are no classes). However if you just want to disallow cloning of a specific object you can use the fact that clone is just a slot and redefine it e.g.:

 % Object clone a
 % a clone: {args} {error "cannot clone"}
 % a clone b
 cannot clone

NEM: I've not tried the implementation yet, but I very much like the specification of this extension. If I make a slot contain an object, what is the syntax for sending messages to that object? From your description, it sounds like it would be something like:

 MyObj pos: [Point create $x $y] 
 puts "pos = [[MyObj pos] to_s]"

Is that correct? Would it be possible to make it like the following?

 puts "pos = [MyObj pos to_s]"

MJ: In the point implementation from above that create call should actually be Point create pos $x $y (note that automatic clone naming is trivial to add in the clone slot). Apart from that, you are correct. I guess it would be possible to implement this, but I cannot see a clear way to add this in the current implementation and not break anything else. It will certainly make slot dispatch more complicated; it has to do number of arguments checking for instance. Even figuring out if a slot contains another object is not straightforward in the current implementation. However, one could implement it with the existing functionality something like this:

 Object addChild: {object} {
   self $object: {slot args} "return \[$object \$slot \{*\}\$args \]"   

 Object clone a   
 Object clone pos

 pos to_s: {} {return "I am [self]"}
 a addChild pos
 puts [a pos to_s]

 # even nested 
 Object clone b
 b to_s: {} {return "I am [self]"}
 pos addChild b

 puts [a pos b to_s]

Or more elaborate:

 package require self

 Object children*: {}

 Object addChild: {name object} {
   self $name: {slot args} "return \[$object \$slot \{*\}\$args \]"
   self children* [lappend [self children*] $object]

 Object delete: {} {
    foreach child [self children*] {
        $child delete
    self destroy

 Object clone Point

  Object to_s: {} {
  return "[self]"

 namespace eval self {
  namespace eval objs {
   variable counter

 Point x: {args} {error "abstract slot, override in clone"}
 Point y: {args} {error "abstract slot, override in clone"}

 Point to_s: {} {
  return "id: [next] ([self x],[self y])"

 Point create: {x y} {
  set obj [self clone ::self::objs::obj[incr ::self::objs::counter]]
  $obj x: $x
  $obj y: $y
  return $obj

 Object clone MousePointer
 MousePointer addChild pos [Point create 130 140]
 MousePointer pos x
 MousePointer pos to_s
 MousePointer delete
 # child is gone
 info commands ::self::objs::*

On a side note, implementing something like this will take away some of the simplicity of the design IMO and I have tried to make the extension as simple as possible while still offering enough flexibility.

Zarutian 2006-10-26 15:35 UTC: I find this extension interesting but I haven't tried it out yet but plan to do just that.