Delphi class references... aka metaclasses... when to use them

I've read the official documentation and I understand what class references are but I fail to see when and why they are best solution compared to alternatives.

The example cited in the documentation is TCollection which can be instantiated with any descendant of TCollectionItem. The justification for using a class reference is that it allows you to invoke class methods whose type is unknown at compile time (I assume this is the compile time of TCollection). I'm just not seeing how using TCollectionItemClass as an argument is superior to using TCollectionItem. TCollection would still be able to hold any descendant of TCollectionItem and still be able to invoke any method declared in TCollectionItem. Wouldn't it?

Compare this with a generic collection. TObjectList appears to offer much the same functionality as TCollection with the added benefit of type safety. You are freed from the requirement to inherit from TCollectionItem in order to store your object type and you can make a collection as type specific as you want. And if you need to access item's members from within the collection you can use generic constraints. Other than the fact that class references are available to programmers prior to Delphi 2009 is there any other compelling reason to use them over generic containers?

The other example given in the documentation is passing a class refere开发者_如何学编程nce to a function that acts as an object factory. In this case a factory for creating objects of type TControl. Its not really apparent but I'm assuming the TControl factory is invoking the constructor of the descendant type passed to it rather than the constructor of TControl. If this is the case then I'm starting to see at least some reason for using class references.

So I guess what I'm really trying to understand is when and where class references are most appropriate and what do they buy a developer?

MetaClasses and "class procedures"

MetaClasses are all about "class procedures". Starting with a basic class:

type
  TAlgorithm = class
  public
    class procedure DoSomething;virtual;
  end;

Because DoSomething is a class procedure we can call it without an instance of TAlgorithm (it behaves like any other global procedure). We can do this:

TAlgorithm.DoSomething; // this is perfectly valid and doesn't require an instance of TAlgorithm

Once we've got this setup we might create some alternative algorithms, all sharing some bits and pieces of the base algorithm. Like this:

type
  TAlgorithm = class
  protected
    class procedure DoSomethingThatAllDescendentsNeedToDo;
  public
    class procedure DoSomething;virtual;
  end;

  TAlgorithmA = class(TAlgorithm)
  public
    class procedure DoSomething;override;
  end;

  TAlgorithmB = class(TAlgorithm)
  public
    class procedure DoSomething;override;
  end;

We've now got one base class and two descendent classes. The following code is perfectly valid because we declared the methods as "class" methods:

TAlgorithm.DoSomething;
TAlgorithmA.DoSomething;
TAlgorithmB.DoSomething;

Let's introduce the class of type:

type
  TAlgorithmClass = class of TAlgorithm;

procedure Test;
var X:TAlgorithmClass; // This holds a reference to the CLASS, not a instance of the CLASS!
begin
  X := TAlgorithm; // Valid because TAlgorithmClass is supposed to be a "class of TAlgorithm"
  X := TAlgorithmA; // Also valid because TAlgorithmA is an TAlgorithm!
  X := TAlgorithmB;
end;

TAlgorithmClass is a data type that can be used like any other data type, it can be stored in a variable, passed as a parameter to a function. In other words we can do this:

procedure CrunchSomeData(Algo:TAlgorithmClass);
begin
  Algo.DoSomething;
end;

CrunchSomeData(TAlgorithmA);

In this example the procedure CrunchSomeData can use any variation of the algorithm as long as it's an descendant of TAlgorithm.

Here's an example of how this behavior may be used in a real-world application: Imagine a payroll-type application, where some numbers need to be calculated according to an algorithm that's defined by Law. It's conceivable this algorithm will change in time, because the Law is some times changed. Our application needs to calculate salaries for both the current year (using the up-to-date calculator) and for other years, using older versions of the algorithm. Here's how things might look like:

// Algorithm definition
TTaxDeductionCalculator = class
public
  class function ComputeTaxDeduction(Something, SomeOtherThing, ThisOtherThing):Currency;virtual;
end;

// Algorithm "factory"
function GetTaxDeductionCalculator(Year:Integer):TTaxDeductionCalculator;
begin
  case Year of
    2001: Result := TTaxDeductionCalculator_2001;
    2006: Result := TTaxDeductionCalculator_2006;
    else Result := TTaxDeductionCalculator_2010;
  end;
end;

// And we'd use it from some other complex algorithm
procedure Compute;
begin
  Taxes := (NetSalary - GetTaxDeductionCalculator(Year).ComputeTaxDeduction(...)) * 0.16;
end;

Virtual Constructors

A Delphi Constructor works just like a "class function"; If we have a metaclass, and the metaclass knows about an virtual constructor, we're able to create instances of descendant types. This is used by TCollection's IDE Editor to create new items when you hit the "Add" button. All TCollection needs to get this working is a MetaClass for a TCollectionItem.

Yes a Collection would still be able to hold any descendant of TCollectionItem and to invoke methods on it. BUT, it wouldn't be able to instantiate a new instance of any descendant of a TCollectionItem. Calling TCollectionItem.Create constructs an instance of TCollectionItem, whereas

private
  FItemClass: TCollectionItemClass;
...

function AddItem: TCollectionItem;
begin
  Result := FItemClass.Create;
end;

would construct an instance of whatever class of TCollectionItem descendant is held in FItemClass.

I haven't done much with generic containers, but I think that given a choice, I would opt for the generic container. But in either case I'd still have to use a metaclass if I wanted to have the list instantiate and do whatever else needs to be done when an item is added in the container and I do not know the exact class beforehand.

For example an observable TObjectList descendant (or generic container) could have something like:

function AddItem(aClass: TItemClass): TItem;
begin
  Result := Add(aClass.Create);
  FObservers.Notify(Result, cnNew);
  ...
end;

I guess in short the advantage/benefit of metaclasses is any method/class having only knowledge of

type
  TMyThing = class(TObject)
  end;
  TMyThingClass = class of TMyThing;

is able to construct instances of any descendant of TMyThing whereever they may have been declared.

Generics are very useful, and I agree that TObjectList<T> is (usually) more useful than TCollection. But class references are more useful for different scenarios. They're really part of a different paradigm. For example, class references can be useful when you have a virtual method that needs to be overridden. Virtual method overrides have to have the same signature as the original, so the Generics paradigm doesn't apply here.

One place where class references are used heavily is in form streaming. View a DFM as text sometime, and you'll see that every object is referred to by name and class. (And the name is optional, actually.) When the form reader reads the first line of an object definition, it gets the name of the class. It looks it up in a lookup table and retrieves a class reference, and uses that class reference to call that class's override of TComponent.Create(AOwner: TComponent) so it can instantiate the right kind of object, and then it starts applying the properties described in the DFM. This is the sort of thing that class references buy you, and it can't be done with generics.

I also would use a metaclass whenever I need to be able to make a factory that can construct not only one hard-coded class, but any class that inherits from my base class.

Metaclasses are not the term I am familiar with in Delphi circles however. I believe we call them class references, which has a less "magic" sounding name, so it's great that you put both common monikers in your question.

A concrete example of a place I have seen this used well is in the JVCL JvDocking components where a "docking style" component provides metaclass information to the base docking component set, so that when a user drags their mouse and docks a client form to the docking host form, the "tab host" and "conjoin host" forms that show grabber bars (similar in appearance to the title bar of a regular undocked window) can be of a user-defined plug-in class, that provides a customized appearance and customized runtime functionality, on a plug-in basis.

In some of my applications I have a mechanism that connects classes to forms capable of editing instances of one or more of that classes. I have a central list where those pairs are stored: a class refernce and a form class reference. Thus when I have an instance of a class I can lookup the corresponding form class, create a form from it and let it edit the instance.

Of course this could also be implemented by having a class method returning the appropriate form class, but that would require the form class to be known by the class. My approach makes a more modular system. The form must be aware of the class, but not the other way round. This can be a key point when you cannot change the classes.