I just learned in my programming languages class that "contravariant argument types would actually be safe, but they have not been found useful and are hence not supported in practical languages." Even though they are not supported, I am confused as to why something like this example we were given would still be, in theory, "safe":
class Animal {
...
public bool compare(Panda) { ... }
}
class Panda extends Animal {
...
public bool compare(Animal) { ... }
}
From what I understand, problems with subtyping come up when something is done that could cause a loss of specificity. So what if I did this? :
Panda p = new Panda();
Animal a = new Animal
...
p.compare(a);
When I look at this, it seems like panda could (and probably does) have some extra fields in it that a plain animal wouldn't know about. Thus, even if all of their animal-specific data members are the same, a panda can have other stuff that differs. How would that make it okay to compare it to a plain animal? Would it just consider the animal-only stuff and ignore the rest?
In your example you don't use any generic types. You have Panda extending Animal, and it's an example of inheritance and leads to polymorphism which is more or less what you describe. Check the links.
To get contravariance, you need to consider some generic type. I'll use .NET type IComparer`1[T] as an example. With C# syntax (which I'll use rather than Java), we indicate that IComparer is contravariant in T by writing in in the definition:
public interface IComparer<in T>
{
...
}
Suppose I have a method which returns an IComparer`1[Animal] (or IComaparer<Animal>), like:
static IComparer<Animal> CreateAnimalComparer()
{
// code that returns something here
}
Now in C#, it's legal to say:
IComparer<Panda> myPandaComparer = CreateAnimalComparer();
Now, this is because of contravariance. Note that the type IComparer<Animal> does not derive from (or "extend") the type IComparer<Panda>. Instead, Panda derives from Animal, and this leads to the IComparer<Xxxx> being assignable to each other (in the opposite order, hence "contravariance" (not "covariance")).
The reason why it's meaningful to declare a Comparer<> contravariant, is if you have a comparer that can compare two arbitrary animals, and return a signed number indicating which is greater, then that same comparer can also take in two pandas and compare those. For pandas are animals.
So the relation
any Panda is an Animal
(from inheritance) leads to the relation
any IComparer<Animal> is an IComparer<Panda>
(by contravariance).
For an example with covariance, the same relation
any Panda is an Animal
leads to
any IEnumerable<Panda> is an IEnumerable<Animal>
by covariance (IEnumerable<out T>).
Related
A common question, specifically since Dart 2, is if it is possible to require some or all generic type arguments on some or all types - for example List<int> instead of List or MyType<Foo> instead of MyType.
It's not always clear what the intention is though - i.e. is this a matter of style (you/your team likes to see the types), to prevent bugs (omitting type arguments seems to cause more bugs for you/your team), or as a matter of contract (your library expects a type argument).
For example, on dart-misc, a user writes:
Basically, if I have this:
abstract class Mixin<T> {}
I don't have to specify the type:
// Works class Cls extends Object with Mixin<int> {} // ...also works
class Cls extends Object with Mixin {}
Is there some way to make the second one not allowed?
Strictly speaking, yes, and no.
If you want to enforce that type arguments are always used in your own projects (instead of relying on type inference or defaults), you can use optional linter rules such as always_specify_types. Do note this rule violates the official Dart style guide's recommendation of AVOID redundant type arguments on generic invocations in many cases.
If you want to enforce that generic type arguments are always used when the default would be confusing - such as List implicitly meaning List<dynamic>, no such lint exists yet - though we plan on adding this as a mode of the analyzer: https://github.com/dart-lang/sdk/issues/33119.
Both of the above recommendations will help yourself, but if you are creating a library for others to use, you might be asking if you can require a type argument to use your class. For example, from above:
abstract class Mixin<T> {}
abstract class Class extends Object with Mixin {}
The first thing you could do is add a default bounds to T:
// If T is omitted/not inferred, it defaults to num, not dynamic.
abstract class Mixin<T extends num> {}
If you want to allow anything but want to make it difficult to use your class/mixin when T is dynamic you could choose a different default bound, for example Object, or even better I recommend void:
In practice, I use void to mean “anything and I don’t care about the elements”
abstract class Mixin<T extends void> {
T value;
}
class Class extends Mixin {}
void main() {
var c = Class();
// Compile-time error: 'oops' isn't defined for the class 'void'.
c.value.oops();
}
(You could also use Object for this purpose)
If this is a class under your control, you could add an assertion that prevents the class from being used in a way you don't support or expect. For example:
class AlwaysSpecifyType<T> {
AlwaysSpecifyType() {
assert(T != dynamic);
}
}
Finally, you could write a custom lint or tool to disallow certain generic type arguments from being omitted, but that is likely the most amount of work, and if any of the previous approaches work for you, I'd strongly recommend those!
For example, I'd like to just be able to write:
class Dog {
final String name;
Dog(this.name);
bark() => 'Woof woof said $name';
}
But have #Dog.bark's type definition be () => String.
This previously wasn't possible in Dart 1.x, but I'm hoping type inference can save the day and avoid having to type trivial functions where the return type is inferable (the same as it does for closures today?)
The language team doesn't currently have any plans to do inference on member return types based on their bodies. There are definitely cases like this where it would be nice, but there are other cases (like recursive methods) where it doesn't work.
With inference, we have to balance a few opposing forces:
Having smart inference that handles lots of different cases to alleviate as much typing pain as we can.
Having some explicit type annotations so that things like API boundaries are well-defined. If you change a method body and that changes the inferred return type, now you've made a potentially breaking change to your API.
Having a simple boundary between code that is inferred and code that is not so that users can easily reason about which parts of their code are type safe and which need more attention.
The case you bring up is right at the intersection of those. Personally, I lean towards not inferring. I like my class APIs to be pretty explicitly typed anyway, since I find it makes them easier to read and maintain.
Keep in mind that there are similar cases where inference does come into play:
Dart will infer the return type of an anonymous function based on its body. That makes things like lambdas passed to map() do what you want.
It will infer the return type of a method override from the method it is overriding. You don't need to annotate the return type in Beagle.bark() here:
class Dog {
String bark() => "Bark!";
}
class Beagle extends Dog {
final String name;
Dog(this.name);
bark() => 'Woof woof said $name';
}
I have a class with a large number of properties that map to some JSON data I've parsed into a Map object elsewhere. I'd like to be able to instantiate a class by passing in this map:
class Card {
String name, layout, mana_cost, cmc, type, rarity, text, flavor, artist,
number, power, toughness, loyalty, watermark, border,
timeshifted, hand, life, release_date, starter, original_text, original_type,
source, image_url, set, set_name, id;
int multiverse_id;
List<String> colors, names, supertypes, subtypes, types, printings, variations, legalities;
List<Map> foreign_names, rulings;
// This doesn't work
Card.fromMap(Map card) {
for (var key in card.keys) {
this[key] = card[key];
}
}
}
I'd prefer to not have to assign everything manually. Is there a way to do what I'm trying to do?
I don't think there is a good way to do it in the language itself.
Reflection would be one approach but it's good practice to avoid it in the browser because it can cause code bloat.
There is the reflectable package that limits the negative size impact of reflection and provides almost the same capabilities.
I'd use the code generation approach, where you use tools like build, source_gen to generate the code that assigns the values.
built_value is a package that uses that approach. This might even work directly for your use case.
"What happens if you don't inherit from Object? Nothing terrible. These classes will be slightly more lightweight, however, they will lack some features such as property change notifications, and your objects won't have a common base class. Usually inheriting from Object is what you want." Vala team said.
So I wanted to know how light the classes are with or without inheriting form Object.
So, Here are my test files
test1.vala:
class Aaaa : Object {
public Aaaa () { print ("hello\n"); }
}
void main () { new Aaaa (); }
test2.vala:
class Aaaa {
public Aaaa () { print ("hello\n"); }
}
void main () { new Aaaa (); }
The results after the compilation was totally unexpected, the size of test1 is 9.3 kb and the size of test2 is 14.9 kb and that contradicts what they said. Can someone explain this please?
You are comparing the produced object code / executable size, but that's not what the statement from the tutorial was referring to.
It is refering to the features that your class will support. It's just clarifying that you don't get all the functionality that GLib.Object / GObject provides.
In C# (and in Java, too?) the type system is "rooted" which means all classes always derive implicitly from System.Object. That is not the case for Vala. Vala classes can be "stand alone" classes which means that these stand alone classes don't have any parent class (not even GLib.Object / GObject).
The code size is bigger, because the stand alone class doesn't reuse any functionality from GLib.Object / GObject (which is implemented in glib), so the compiler has to output more boiler plate code (writing classes in C is always involving a lot of boiler plate code).
You can compare yourself with "valac -C yourfile.vala" which will produce a "yourfile.c" file.
That's a very interesting question. The answer will get you deep into how GObjects work. With these kinds of questions a useful feature of valac is to use the --ccode switch. This will produce the C code, instead of the binary. If you look at the C code of the second code sample, which doesn't inherit from Object, it includes a lot more functions, such as aaaa_ref and aaaa_unref. These are basic functions used to handle objects in GLib's object system. When you inherit from Object these functions are already defined in the parent class so the C code and resulting binary are smaller.
By just using class without inheriting from Object you are creating your own GType, but not inheriting all the features of Object so in that sense your classes are lighter weight. This makes them quicker to instantiate. If you time how long it takes to create a huge number of GType objects compared to the same number of GObject inheriting objects you should see the GType object being created more quickly. As you have pointed out GType objects lose some additional features. So the choice depends on your application.
Here's the situation: I have an abstract class with a constructor that takes a boolean (which controls some caching behavior):
abstract class BaseFoo { protected BaseFoo(boolean cache) {...} }
The implementations are all generated source code (many dozens of them). I want to create bindings for all of them automatically, i.e. without explicit hand-coding for each type being bound. I want the injection sites to be able to specify either caching or non-caching (true/false ctor param). For example I might have two injections like:
DependsOnSomeFoos(#Inject #NonCaching AFoo aFoo, #Inject #Caching BFoo bFoo) {...}
(Arguably that's a bad thing to do, since the decision to cache or not might better be in a module. But it seems useful given what I'm working with.)
The question then is: what's the best way to configure bindings to produce a set of generated types in a uniform way, that supports a binding annotation as well as constructor param on the concrete class?
Previously I just had a default constructor on the implementation classes, and simply put an #ImplementedBy on each of the generated interfaces. E.g.:
// This is all generated source...
#ImplementedBy(AFooImpl.class)
interface AFoo { ... }
class AFooImpl extends BaseFoo implements AFoo { AFooImpl() { super(true); } }
But, now I want to allow individual injection points to decide if true or false is passed to BaseFoo, instead of it always defaulting to true. I tried to set up an injection listener to (sneakily) change the true/false value post-construction, but I couldn't see how to "listen" for a range of types injected with a certain annotation.
The problem I keep coming back to is that bindings need to be for a specific type, but I don't want to enumerate all my types centrally.
I also considered:
Writing some kind of scanner to discover all the generated classes and add a pair of bindings for each of them, perhaps using Google Reflections.
Creating additional, trivial "non caching" types (e.g. AFoo.NoCache extends AFoo), which would allow me to go back to #ImplementedBy.
Hard wiring each specific type as either caching/non-caching when it's generated.
I'm not feeling great about any of those ideas. Is there a better way?
UPDATE: Thanks for the comment and answer. I think generating a small module alongside each type and writing out a list of the modules to pull in at runtime via getResources is the winner.
That said, after talking w/ a coworker, we might just dodge the question as I posed it and instead inject a strategy object with a method like boolean shouldCache(Class<? extends BaseFoo> c) into each generated class. The strategy can be implemented on top of the application config and would provide coarse and fine grained control. This gives up on the requirement to vary the behavior by injection site. On the plus side, we don't need the extra modules.
There are two additional approaches to look at (in addition to what you mentioned):
Inject Factory classes instead of your real class; that is, your hand-coded stuff would end up saying:
#Inject
DependsOnSomeFoos(AFoo.Factory aFooFactory, BFoo.Factory bFooFactory) {
AFoo aFoo = aFooFactory.caching();
BFoo bFoo = bFooFactory.nonCaching();
...
}
and your generated code would say:
// In AFoo.java
interface AFoo {
#ImplementedBy(AFooImpl.Factory.class)
interface Factory extends FooFactory<AFoo> {}
// ...
}
// In AFooImpl.java
class AFooImpl extends BaseFoo implements AFoo {
AFooImpl(boolean caching, StuffNeededByAFIConstructor otherStuff) {
super(caching);
// use otherStuff
}
// ...
class Factory implements AFoo.Factory {
#Inject Provider<StuffNeededByAFIConstructor> provider;
public AFoo caching() {
return new AFooImpl(true, provider.get());
}
// ...
}
}
Of course this depends on an interface FooFactory:
interface FooFactory<T> {
T caching();
T nonCaching();
}
Modify the process that does your code generation to generate also a Guice module that you then use in your application setup. I don't know how your code generation is currently structured, but if you have some way of knowing the full set of classes at code generation time you can either do this directly or append to some file that can then be loaded with ClassLoader.getResources as part of a Guice module that autodiscovers what classes to bind.