The trouble with Generics

This document tries to capture the benefits and doubts we have regarding generics in Wicket 1.4. Ultimately we need to find a way where the Framework and Wicket Community can move onward to Wicket 1.5/2.0 with Java 5 as a basis.

Generics Rock!

Java generics are good for type safety throughout your code. They document the code directly with intent: List<Person> leaves little to the imagination as to what goes into the list.

These benefits are a great asset to Wicket. Wicket uses the concept of models to provide components with the data that needs to be rendered. Until recently Wicket was Java 1.4 based, and a lot of components suffered from the lack of generified models. Take for instance the ListView:

Generics Suck!

TBD

Proposed Approaches

No Generics!

IModel<T> Only

IModel<T> and Component<T>

Hybrid: IModel<T> and Select Parts of the Component Hierarchy (ListView, etc.)

This approach will not work because covariant parameter types are not allowed. The "setters" can't be overridden to take a different parameter type.

Decoupld the "default" Model from Component Completely and Let Subclasses Use Their Own Typed Models

Generics are warranted for IModel but too much verbosity for Component

If there is any place in Wicket which clearly holds typed data, it is the models (the IModel interface). In practice, one ends up creating a lot of custom components without models, such as different subclasses of MarkupContainer and possibly Page, and here the Component generification just adds more verbosity to the code.

Typically when starting to use a custom component, you start by typing its constructor call, so the constructor of a custom component is a crucial place for making clean and easy APIs. The possibility of giving type arguments to the IModel instance being passed in there is great, DropDownChoice is a prime example as it also has constructors with a List that needs to contain objects of the same type as the model. However,

DropDownChoice fruit = new DropDownChoice("fruit", new Model<Fruit>(), Arrays.asList(apple, orange, banana));

is better than

DropDownChoice<Fruit> fruit = new DropDownChoice<Fruit>("fruit", new Model<Fruit>(), Arrays.asList(apple, orange, banana));

Also when using 2.0 the experience was that IModel type was excellent, but Component type was more like noise. When using 1.3, the lack of types has hardly caused trouble; mostly types would be helpful when first creating a new component. Once you have the component constructed and are calling getModelObject() on it, you're typically sure of what to expect. There have not been difficult problems concerning objects of wrong type inside models.

In this approach, using getModelObject() you cast the same as in 1.3, and for getModel() you can do the necessary casting cleanly, which leaves you in a situation better than in 1.3.
(Timo Rantalaiho)

I mostly agree with Timo. The only thing we lose from Component is generification of IConverter. Since IConverter is mostly only important in FormComponent perhaps we can fully type FormComponent which would allow us to have typesafe converters AND validators.
(Igor Vaynberg)

If we dont do Component but we do specific subtypes like FormComponent that the above DropDownChoice example still has that signature..
and then i also vote for Repeaters to be generified, because especially in the onPopulate it is very descriptive what you get..
But i think people will then start complaining because they are using Links a lot and there model object.. So then we do get the question constantly on the list Why is this and Why isnt that... (i think)
I still dont find this that bad.. It really describes it:

DropDownChoice<Fruit> fruit = new DropDownChoice<Fruit>("fruit", new Model<Fruit>(), Arrays.asList(apple, orange, banana));

because what you also could do is this:

DropDownChoice<Fruit> fruit = new DropDownChoice<Fruit>("fruit", Arrays.asList(apple, orange, banana));

and yes i know.. this is somewhere not really type safe..because now a CompoundModel is used and that is based on strings..
But in the code is more descriptive.. but i guess somehow we should have a warning
And this is also the reason why i am also working on type safe property models.
(Johan Compagner)

Suggestion 1 ---- {}setResponsePage(){} ---signature

The initial issue that started this discussion was a problem with this method signature:

public final void setResponsePage(final Class<? extends Page<?>> cls, PageParameters parameters)
{
    getRequestCycle().setResponsePage(cls, parameters);
}

We could change this to:

@SuppressWarnings({"unchecked"})
public final void setResponsePage(final Class<?> cls, PageParameters parameters)
{
    final Class<? extends Page<?>> castCls = (Class<? extends Page<?>>) cls;
    getRequestCycle().setResponsePage(castCls, parameters);
}

This way (a) users migrating to 1.4 don't get tons of warnings/errors and (b) if they pass in the wrong thing, a class cast exception will be thrown quickly.

NOTE: This is just an example - we may wish to move the cast into the getRequestCycle() call ...

My idea is to dial back the generics on some of the API calls like this one and others that cause pain, but leave the generics for useful things like getModelObject().

-Doug Donohoe

This is already resolved with the following signature:

public final <C extends Page<?>> void setResponsePage(final Class<C> cls, PageParameters parameters)
{
    getRequestCycle().setResponsePage(cls, parameters);
}

Gerolf Seitz

Suggestion 2 - RestartResponseAtInterceptPageException constructor

I agree with Doug, and with suggestion 1 above on setResponsePage changing to take Class<?>

I think another example is the constructor of RestartResponseAtInterceptPageException (and similar classes) that take a class as an argument.

public RestartResponseAtInterceptPageException(final Class< ? extends Page< ? >> interceptPageClass)
	{
		if (interceptPageClass == null)
		{
			throw new IllegalStateException("Argument pageClass cannot be null");
		}
		redirectToInterceptPage(interceptPageClass);
	}

We could change this to:

public RestartResponseAtInterceptPageException(final Class<?> interceptPageClass)
{
	if (interceptPageClass == null || !Page.class.isAssignableFrom(interceptPageClass))
	{
		throw new IllegalStateException("Argument pageClass cannot be null and must inherit from Page.class");
	}
	redirectToInterceptPage(interceptPageClass);
}

This, for the same reason as suggestion 1 - it's not very likely that someone is going to pass a non-page-child class into these methods or constructors, so the benefit is minimal, but the strictly generified versions penalize you if your page class is not also generified, which would be confusing to new users.

Same as above. -Gerolf Seitz

Suggestion 3 Static factory methods

new Label<String>("id", new PropertyModel<String>(object, "property"))

is really quite verbose. I got bored with that and wrote

public class LabelFactory {
	public static <T> Label<T> fromProperty(final String labelId, final Object object, final String propertyName) {
		return new Label<T>(labelId, new PropertyModel<T>(object, propertyName));
	}
}

which made constructing Labels just

LabelFactory.fromProperty("id", object, "property")

It would be really nice if wicket contained such factories or had the objects on the methods themselves. It would be really cool to say Label.fromString or Label.fromProperty
-Nik Everett

Generics "Gotchas"

Palette

The current (1.4-m2) Palette class' constructor signature looks like this:

public Palette(String id,
               IModel<Collection<T>> model,
	       IModel<Collection<? extends T>> choicesModel,
               IChoiceRenderer<T> choiceRenderer,
               int rows,
	       boolean allowOrder)

This doesn't allow me to use an IModel<List<T>>. How about if we change it to:

public Palette(String id,
               IModel<? extends Collection<T>> model,
	       IModel<? extends Collection<? extends T>> choicesModel,
               IChoiceRenderer<T> choiceRenderer,
               int rows,
	       boolean allowOrder)

This appears to make things nicer in my local if I change it.

-James Carman

Application.getHomePage()

Currently, the Application.getHomePage() signature looks like:

public abstract Class<? extends Page<?>> getHomePage();

If you choose not to generify your pages, you can't return your home page's class from this method. What if we change the method signature to:

public abstract Class<? extends Page> getHomePage();

-James Carman

The above certainly compiles, but does it keep eventual warnings purely on the framework side of things or are generic users possibly confronted with them? Considering that full backwards compatibility is not the goal of Wicket 1.4 (there are some small API breaks as well I believe), maybe it's OK to break getHomePage()? In any case the solution is generally very simple for users (add one <Void>).

So is backwards compatibility very important? Is allowing raw types in all cases very important? I think questions like these need to be answered to get an overall idea of the goal of generics for Wicket.

-Sebastiaan van Erk

Certainly it won't work if you override the method like so:

@Override
public Class<? extends Page<?>> getHomePage() {
    // Does not compile!
    return Page.class;
}

But this works just fine if you choose not to generify the Page:

@Override
@SuppressWarnings("unchecked")
public Class getHomePage() {
    return Page.class;
}

-David Arnold

Variations of 'public void foo(Class<? extends Component<?>> clazz)'

There seems to be uncertainty about the meaning of the signature 'public void foo(Class<? extends Component<?>> clazz)'

Following code has variations of that signature. It does not compile, it is just for demonstration purposes or as a reference in the Mailing list.

Feel free to change or to add explanations.

public class Generix
{
	public static void main(String[] args)
	{
		Foo foo = new Foo();

		// explanation needed
		foo.bar(Component.class);
		// explanation needed
		foo.bar(IntegerComponent.class);
		// explanation needed
		foo.bar(GenericComponent.class);
		// explanation needed
		foo.bar(GenericIntegerComponent.class);

		foo.cly(Component.class);
		foo.cly(IntegerComponent.class);
		foo.cly(GenericComponent.class);
		foo.cly(GenericIntegerComponent.class);

		foo.dol(Component.class);
		foo.dol(IntegerComponent.class);
		foo.dol(GenericComponent.class);
		foo.dol(GenericIntegerComponent.class);

		foo.elm(Component.class);
		foo.elm(IntegerComponent.class);
		foo.elm(GenericComponent.class);
		foo.elm(GenericIntegerComponent.class);

		foo.far(Component.class);
		foo.far(IntegerComponent.class);
		foo.far(GenericComponent.class);
		foo.far(GenericIntegerComponent.class);

		Class<? extends Component<?>> bar = Component.class;
		bar = IntegerComponent.class;
		bar = GenericComponent.class;
		bar = GnericIntegerComponent.class;

		Class<? extends Component> cly = Component.class;
		cly = IntegerComponent.class;
		cly = GenericComponent.class;
		cly = GenericIntegerComponent.class;

		// not a variation of 'Class<? extends Component<?>>'
		// but nice for clearing the picture.

		// Little surprise Component.class is the super bound of Component<?>
		Class<? super Component<?>> carl = Component.class;
		// No surprises here:
		carl = IntegerComponent.class;
		carl = GenericComponent.class;
		carl = GenericIntegerComponent.class;
	}

	public static class Component<T>
	{
	}
	public static class IntegerComponent extends Component<Integer>
	{
	}
	public static class GenericComponent<T> extends Component<T>
	{
	}

	public static class GenericIntegerComponent<T> extends Component<Integer>
	{
	}

	public static class Foo
	{
		// non generic method with parameter of a generic type
		public void bar(Class<? extends Component<?>> clazz)
		{
		}

		// non generic method with parameter of a generic type
		// Generix.Component is a raw type. References to generic type
		// Generix.Component<T> should be parameterized
		public void cly(Class<? extends Component> clazz)
		{
		}

		// generic method
		public <T extends Component<?>> void dol(Class<T> clazz)
		{
		}

		// generic method
		public <T extends Component<?>> void elm(Class<? extends T> clazz)
		{
		}

		// generic method
		public <M, T extends Component<M>> void far(Class<T> clazz)
		{
		}
	}
}