Latest revision as of 07:30, 31 July 2009

More About Objects

Binding

As you know, there are two principal states that the Mops system can be in. When you first start Mops, the system is waiting for your input, ready to interpret whatever you enter at the keyboard or from a load file on the disk. This is known as run state or interpret state, and when it is active, Mops immediately executes whatever word names you enter into the input stream. On the other hand, you may wish to create colon definitions or methods that compile code to be executed later. After the Mops interpreter sees a colon (or :m) and before it sees the next semicolon (or ;m), Mops will be in compile state, and rather than immediately execute the words that it sees, Mops will compile code to call those words into the dictionary. A colon definition is thus a list of calls to other words that will be executed at a later time, when the name of the word is seen in the input stream and the system is in interpret state.

When Mops sees the name of a word and is in interpret state, it attempts to find a string matching the name string of the desired word in the Mops dictionary. A very fast search word, called FIND, does this in the Mops system. When you send a message to an object, such as:

get: myint

the first thing that happens is that the selector, Get:, is converted into a unique 32-bit code known as a hash value. Then the object name, myint, must be looked up in the Mops dictionary and executed to determine the address of its data. The class of the object myInt is determined, and from that is derived the address in the Mops dictionary of the method that was defined last for that class. This process is known as the binding of a compiled method, and is necessary to determine what code will be processed for that particular message. Note that two different searches must occur before the method can be resolved: the object must be found as a word in the Mops dictionary, and then the compiled method must be found as an entry in the methods dictionary for the object's class.

Early Binding

If we were to enter the above message when Mops was in compile state, the search of both the object and the compiled method would occur at compile time — during the compilation of whatever word or method was being defined. This information would already have been determined by the time that the new definition was actually executed. This is the default manner in which Mops compiles messages, and is known as early binding. Because most of the work of searching is done at compile time, the execution of the message can be very efficient, because it was bound to the actual address of the compiled method in the dictionary.

Late Binding

There are occasions, however, when it is very desirable not to bind the compiled method address when the message is being compiled. Consider, for instance, a situation in which you have a collection of objects that you need to print on the screen. You might have rectangles, strings, bitmapped images, and other objects, all in the same collection. Each of the objects already knows how to print itself by means of a compiled method for Print: that exists somewhere in its inheritance chain. Your program could be much simpler, however, if it didn't have to explicitly concern itself with the class of a particular object at compile time, but could just send the Print: message to the object and have normal method resolution occur at runtime. This would allow you could store the addresses of the various objects that need to be printed in an array or list without concern for the class of each one.

The technique just described is known as late binding, and is used by Smalltalk and some other object-oriented languages as the only style of method resolution. While very powerful and elegant, late binding traditionally has serious performance disadvantages that make most of these languages poor candidates for production of commercial applications. Because Mops provides both late and early binding, you can tailor your application for speed or generality where needed. Even late-bound Mops messages are quite fast, thanks to a highly optimized search primitive. Late binding makes the various objects in your application highly independent of each other, leading to much easier program maintenance. And late binding can greatly simplify the control structure of your code, because much conditional processing can be handled via class differences.

Late Binding and Toolbox Calls

Mops itself uses late binding within many of its Toolbox class methods. For example, when the fEvent object (a default Event object that is predefined in Mops) receives a MouseDown event from the Toolbox, meaning that the user has clicked the mouse button, fEvent hands the processing of the Click over to the window (or menu bar) that was involved. If the Toolbox tells fEvent that a click was received in the Content region of a window, fEvent sends a Content: message to the window involved. This event must be processed differently according to whether the window has controls, editable text, graphics, and so on. In a conventional C or Pascal program, a large switch/case statement would be required that would handle clicks for different types of windows. In Mops, the differential processing is handled automatically by late binding of the Content: method, because the correct processing will occur for the class of the actual window involved. The programmer is then free to define new subclasses of Window with their own Content: methods, and fEvent can still do exactly the same thing.

Carbon Note: Toolbox events in Carbon (called Carbon Events) should be handled by event handler callbacks. So carbon events are processed by callback words installed for each toolbox object for each kind of events in Carbon PowerMops. As the result, fEvent object no longer performs substantial tasks in PowerMops. However, as for the use of late binding, there is essentially no change between 68k Mops and Carbon PowerMops.

Early vs. Late Binding

You can cause late binding to occur in a particular message with a very simple modification of your source:

get: myint         \ early binding
get: [ myint ]     \ late binding

In the first example, Mops would determine at compile time the class of the object myInt, and in the second example this resolution would happen at run time. If myInt is truly an object, using late binding would be a useless waste of time, because the class of myInt could not possibly change. However, the brackets can enclose any code sequence that generates an address of a valid object at runtime. This can be a single Mops word, a sequence of words, messages to other objects, or anything else. Some examples:

(A) get: [ dup ]                      \ message receiver is the object whose address 
                               \ was duplicated on the data stack

(B) get: [ i at: myArray ]            \ receiver is the object whose address is
                               \ at element i in myArray

(C) 0 value theObject         \ create a Value to hold an object address
    myInt -> theObject              \ place the address of myInt in theObject 
    get: [ theObject ]         \ receiver is myInt via theObject

(D) get: [ ]                  \ receiver is object whose addr is top
                               \ of stack

Since the normal use of brackets in Forth is to turn compilation off and on, this particular interpretation of brackets only applies immediately after a selector, and the regular Forth interpretation applies otherwise.
To help avoid (or maybe to add to) confusion, we have added two more ways to specify a late bind

method: **
method: []

to bind to whatever is on the top of the stack at run time. method: [] is really the same as method: [ ] with a space between the brackets.
You will frequently find that it is useful to late bind to Self. This is particularly so with multiple inheritance. For an example, see the (Col) class in the file 'Struct' or 'pStruct'. (Col) knows it will be implementing subclasses multiply inherited with some kind of array, but it doesn't need to worry about what kind of array. It can simply send messages such as AT: to Self, late-bound, and the right kind of array access will be done. We have even provided an extra syntax to make this operation look neater, e.g

at: [self]

Thus the following are all equivalent:

aMethod: [self]
aMethod: [ self ]
self aMethod: **
^base aMethod: **

You can take your pick. But in the case of late binding to Self, I think the first one looks the best. Note that, if you are defining a new class and overriding (replacing) a method (e.g. PRINT:) from a superclass, other methods in the superclass will continue to use the first definition of PRINT: unless all calls to it are late bound.

When to Use Late Binding

You should be able to see that this is a very general and powerful technique. As you become more skilled in building object-based applications, you will find yourself using the power of late binding more and more. The following are some situations in which late binding is particularly useful:

1. Forward referencing
   You may find it convenient to create code that sends messages to an object that won't be defined until later in the source code. For instance, two classes may need to send messages to each other, meaning that one of them will have to be referenced before it is defined. Cases like this can be easily solved by defining a Value that will hold the actual address of an object at runtime, and compiling late-bound messages using the Value rather than an object, as in example C above.
2. Passing objects as arguments
   Frequently, you will find it useful to pass an object as an argument to a Mops word or method. For instance, the following word computes the difference in the areas enclosed by two rectangles:
   

   : ?netArea  { rect1 rect2 -- netArea }   size: [ rect1 ] * size: [ rect2 ] * - ;

   In this example, two named input parameters, rect1 and rect2 are the addresses of objects rect1 and rect2, and are used as receivers of size: messages. This definition compiles exactly the same kind of late-bound reference as if a Value were used. The size: method is looked up and executed at runtime, yielding the dimensions of the rectangle. The area calculation proceeds easily with that information.
3. Algorithmic determination of message receivers
   Because you can use any code sequence that results in an object address between the brackets of a late-bound method call, you can algorithmically determine which object will be the receiver of a given message. This allows you to traverse a list of objects, sending the same message to each one; it also permits sending a message to an object whose address came from another source, such as a Toolbox call. It might be that the routine itself that generates the object address must be dynamically changed at runtime, in which case you could use a vector as message receiver. A vector is a special kind of Value that holds the execution token (xt, see Lesson 20 of the tutorial) of an executable Mops word (not an object); for brevity they are known as Vects, and have to be initiated with the xt of a Mops word:

   ' null vect myvector

   Just as giving the name of a Value puts its contents on the stack, giving the name of a vector causes the word it holds to be executed. A late-bound message to a vector in your code will compile a late-bound reference in which the vector is executed first, which in turn executes the Mops word whose xt it holds; this places an object address on the stack that will be the actual receiver of the late-bound message. By changing the contents of the Vect, you can substitute a new algorithm to generate the object address.
4. Dynamic (heap-based) Objects
   A very important use of late binding is for communicating with dynamic objects, but they are a subject that needs a section all of their own.

References

References are a new feature in PowerMops 4.0. A reference may be thought of as a pointer to an object, but references have many other capabilities as well, since with PowerMops 4.0 they provide the primary means of implementing dynamic objects.

Because references represent a fundamental new Mops feature, there are many places in the manual which will eventually need to mention them. This however has not happened yet, so most of the information you will need to use references effectively is here in this section.
Anyplace you declare an object, you can now put the word ref in front. So, for example, let's say you have a View object

view myView

you can instead make it a reference:

ref view myView

Before you can send messages to it, you need to set it to point to an actual view object. You can do this in two different ways, using one of two prefix operators, -> and new>

Prefix Operator ->

If you want to set a reference to point to an existing object, just get the object's address, and use the -> prefix. The object can be anywhere. So for example if you already have a view someView, you can set the reference myView to point to it thus:

someView -> myView

or, equivalently

addr: someView -> myView

of course you can determine the source object's address in any way you like. Any arbitrary computation can precede the ->. The actual assignment is done by an internal Mops word which does a check that the address you use is the address of an object of the appropriate class, and will give you an error if it's not.

There is only one restriction on using ->. The object must not be an ivar within a record{…}. This is because such objects don't have a header, which prevents Mops performing the above class determination, along with other housekeeping operations associated with references. If you attempt to assign the address of such an object to a reference, you will get the “not an object” error at run time. (Although it might be an object, Mops can't verify that it is.)

Prefix Operator new>

If you want to create a new view object in the heap and set myView to point to it, use the new> prefix:

new> myView

That does the whole job. The heap block can't move (for the technical, it's a pointer-based block), so you don't have to worry about locking or unlocking handles.

Either way you set the myView reference, you can now send messages to it exactly as if it were a normal view.

Note carefully that we clearly distinguish in the syntax if we are sending a message to the object, or performing some operation on the reference itself (as opposed to the referenced object). Any message goes to the object, while the prefix operators -> and new> are operations on the reference, and don't affect the object in any way.

Prefix Operator release>

This is the prefix operator that you use when you are finished with a reference. Now it is important to realize that release> does not cause a heap object to be deleted. What it does is simply reset a reference to point nowhere (by setting it to the nil pointer value nilP). Think of release> as releasing the link between a reference and an object.

So when does a heap object get deleted? Simply when it doesn't have any more references to it. Mops now incorporates a “garbage collection” routine to find these objects and delete them. This should make the programming of complex data structures much simpler than before. Any class can have any number of references in its ivars, and objects of that class can themselves be pointed to by any number of other references, and can be in the dictionary or the heap.

All you have to take care of is doing release> on a reference when you're finished with it. If you accidentally try to send a message through a reference that has had a release>, you'll get an error, not an access to some strange place in memory. But note that multiple release> operations on a reference are harmless. Only the first one will do anything; the following ones will be ignored. This scheme also explains why you can merrily do -> operations on references without having to worry about where they might have been pointing before. Mops simply does an internal release> before assigning the new address to the reference. If the reference was already nil, the internal release> is ignored.

When Mops detects that a heap object has no more references pointing to it, it first sends a release: message to the object (so that the object gets a chance to free any heap blocks that it might have allocated for itself), then it performs a release> on any references in the object, then finally the heap block is deleted

If you are already a Mops user, you’ll be able to see that using references should make certain classes of application — those that involve complex dynamic data structures — much simpler to handle than they were before. But we're not finished with references yet. Lurking beneath that simple interface is another binding method. This allows you to set a reference to point to an object of any subclass of the class of the reference, and when you send a message, you will invoke the method of the subclass rather than the original class.

This mechanism uses a table of addresses associated with each class, which points to the methods of that class in an order which doesn't change when you go to a subclass. This sort of table is used by C++ and called a vtable (for virtual table). It does take a small amount of extra space, but only 4 bytes per public method, which isn't a lot.

For a speed comparison, an early bind only takes a couple of machine instructions, while a vtable bind through a reference takes about 30. A late bind can take as many as 500, but we use a cache which is very effective and reduces the overhead to about 100 instructions most of the time.

You are not forced to use a vtable bind through a reference, however – this is just the default. If you want to have a reference which can point to an object of any class whatsoever, and invoke late binding, declare

ref any <name>

Naturally enough you can't use the new> prefix on a ref any, since there's no class we can use to create the object!

You can also specify early binding through a reference, if you declare

ref <class> <name> no_subclasses

For this kind of reference, you will get an error if you try to assign an object whose class doesn't match exactly. You can’t have an object whose class is a subclass of the reference’s class. But the advantage is that since Mops knows the class at compile time, it can compile early binding code for messages. This can give you some extra speed if you know that the reference will always point to an object of exactly a specific class.

So far references are only implemented for PowerMops. For 68k Mops, you can continue to use object pointers (see below), although they lack a lot of the functionality of references.

But for PowerMops, we now recommend you now use references instead of object pointers, since they can do everything object pointers can do and a whole lot more. (We will however retain object pointers so as not to break old code.)

Dynamic (heap-based) Objects

A very important use of references and late binding is for communicating with dynamic objects. Many applications need to create objects dynamically rather than build them into the dictionary at compile time. For instance, an application that handles multiple windows cannot know in advance how many windows will be open at any one time. It would be clumsy to have to predefine a number of windows in the dictionary called Window1, Window2, and so on. The best approach in this situation is to create a list of window objects that can expand and contract dynamically. To avoid wasting storage, it is most appropriate to create the window objects on the heap when they are needed, and return the heap to the system when a window is closed.

In PowerMops, as we saw in the last section, this is best done with references. In 68k Mops, you will have to use the older method involving ObjHandles, which we describe here.

ObjHandle is a subclass of Handle. In this class we provide methods for creating and accessing heap objects. A heap object can be created thus:

ObjHandle anObjHdl
 ' someClass newObj: anObjHdl

Then, to access the object, the method obj: anObjHdl returns a pointer to the object, and also locks the handle so that the object won't be unceremoniously moved while we are doing things with it. Remember to unlock: anObjHdl when finished. So, using the above example, you can access the object thus:

mssg1: [ get: anObjHdl ]
       ...
       unlock: anObjHdl

When you are completely finished with the object, send release: anObjHdl. This will automatically cause a late-bound Release: to be sent to the object itself, before its storage is released, in case it has some heap storage of its own. Be careful not to send release: anObjHdl if the heap object has already been disposed of, as will happen with a dynamically created Window that has been closed by a click on the close box. The object no longer exists, although the space it occupied is still allocated. You dispose of that with anObjHdl release: class_as> handle (see below) which uses the release method of the superclass Handle.

If you know that a dynamic object has one particular class, you can avoid the time penalty of late binding to it, as we'll now see.

More Ways of Early Binding

Declared References

As we saw above, in PowerMops you may declare a reference to an object in which you specify that messages to the object will use early binding, by using the word no_subclasses.

ref <class> <name> no_subclasses

In 68k Mops, to do the same kind of thing, you will need to use the older feature, Object pointers.

Object Pointers

The idea of an object pointer is to provide a convenient way of early binding to an object whose identity is determined at run time (for example, a heap-based object), but whose class we know at compile time. In cases like this we would like the efficiency of early binding to the object.

With an objPtr, the class is associated with the pointer at compile time, then whenever an object address is stored in the pointer there is a check that the class of the object matches. After that, sending a method to the pointer actually sends it to the object the pointer points to, with early binding (since we know the class at compile time). An object pointer is a 'low-level' entity, rather like a Value. The syntax for object pointers is:

objPtr anObjPtr  class_is theClass
               ...

( get obj addr to the stack ) -> anObjPtr
               ...

aMethod: anObjPtr

Occasionally, the desired class for an object pointer may not be defined at the time the object pointer needs to be defined. In this case, use the syntax

objPtr anObjPtr

then after the class is defined:

' anObjPtr set_to_class theClass

This, of course, must precede any code which sends a message to anObjPtr. See the file Dialog+ for some examples — there I had to implement methods manipulating a chain of pointers to objects of the same class as was being defined. For this purpose I put the set_to_class line straight after the :class line, but before the ivars and methods. This is quite allowable.

Note that the address you store in an objPtr must be an object address. If you use ' (tick) or ['] on an object in the dictionary, you will get the cfa of the object, which isn't the same. As we saw earlier, an object has a pointer to its class at the start (it has some other information there as well). To get the object address, which is the address of its first ivar, you just use the name of the object without any selector. Alternatively, if you already have the cfa of the object, use the word >OBJ to convert it to the object address. So, either of the following will work:

anObj -> anObjPtr

' anObj >obj -> anObjPtr

class_as>

There's a way to force an early bind to an object, without having to set up a reference or an objPtr. The disadvantage is that it's less secure. With earlier versions of Mops you could say

( obj addr on stack ) aMethod: theClass

with the object's class being used as the 'object' to which the method is sent. This syntax was available in Neon, but was undocumented (!). It was, and still is, available in Mops. It can be dangerous if you don't know what you're doing, since there can't be any check on the real class of the object whose address is on the stack (since it's not known at compile time), and there isn't even a check that what's on the stack is a legal address. If it isn't the address of an object of that particular class, an immediate crash is probably the best you can hope for. But if you know what you're doing this syntax can be very handy.

The only difficulty I have had with it is that in reading code it isn't glaringly obvious that you're not sending a normal message to an object. If your class names aren't well chosen, it might appear that the thing following the selector is an object name, not a class name, which would give that code quite a different meaning. (Of course I'm not talking to you, since you always name your classes in such an unambiguous manner that they just couldn't be anything but classes.)

Anyway for those who do sometimes give their classes less than ideal names, with version 2.6 we have a new syntax for the above operation:

<obj addr on stack> msg: class_as> someClass

The old way will still work — I don't plan to delete it and maybe break existing code — but the new way reads less ambiguously (and compiles exactly the same code). As mentioned above (in the section on heap-based objects) this can be a useful way of accessing a method of an objects superclass, if that is more appropriate.

Public and Static ivars

From version 2.5.1, we have provided some extra features relating to ivars. Most simple programs won't need these features, but more experienced Mops users may well find them very useful. Although you may not use these features youself at first, you will find them in use in a number of the fundamental predefined classes, e.g. class Window, so it is as well to be aware of them.

Static (or class) ivars belong to the class rather than to an object. They're rather like globals except that they don't clutter the global namespace. The syntax for accessing them is just the same as for normal ivars.

They're declared like this:

class myClass super{ mySuper }
       var     oneVar
static
{      var     someVar
       int     someInt
}
       var anotherVar

In this example, someVar and someInt are static, oneVar and anotherVar are normal ivars. In the methods of myClass, whenever you access someVar, you are accessing the SAME ivar, no matter what object you are in, and similarly for someInt.

Public ivars can be accessed from outside the class. They're declared this way:

class myClass super{ mySuper }
public
       var     aVar
       int     anInt
end_public
       var anotherVar
       int     anotherInt

They're accessed from outside the class via this syntax:

msg: ivar> anIvar IN someObject

(where someObject is an object of myClass, of course.)

I've considered adding this feature to Mops for some time, doing this for some time, with mixed feelings, but eventually decided it was worth it in some situations where otherwise I'd need to define dozens of pass-through methods. It may not be brilliantly elegant, but it's very practical.

This is really an extension of the public/private distinction which we already have for methods. The default is for ivars to be private, and methods to be public. You can now use PUBLIC or PRIVATE anywhere in the ivar list or method declarations of a class to change this default. You can use END_PUBLIC or END_PRIVATE to restore the default.

If you combine these two new features, you can get a 'public static' ivar. To access this from outside the class, you can't use the above syntax since there's no object to refer to. So the syntax is:

msg: ivar> aStaticIvar IN_CLASS myClass