abstract static methods (again)

T

Tomas Mikula

I have searched this group for "abstract static methods" and found a
couple of threads, but I think none of them was discussing the kind of
semantics I am going to describe. As you might have guessed, I believe it
would be useful :). I further believe it is fully complatible with the
current language, but there might be caveats I have overlooked. I'm
wonder if you would find it as useful as I do and if you see any problems
with it. I know it is a long post and some parts may be difficult to
understand. Therefore I will be thankful if you can read it all and think
about it.

By "abstract static method" I refer to either a static method in an
interface or an abstract static method of an abstract class.

Put shortly, a declaration of an abstract static method in interface J
(resp. in abstract class A) would mean that any class implementing J
(resp. extending A) must either provide its _own_ implementation of that
static method, or itself be abstract.

Note 1: There would still be no inheritance of static methods.
Note 2: Semantics of calling static methods on instances would remain
unchanged, i.e. static methods are still not virtual.
Note 3: The following would be a compile-time error:

interface J {
public static void f();
}
class X implements J {
public static void f(){...}
}
class Y extends X {
}
// ERROR: class Y does not provide
// its _own_ implementation of f()

Note 4: Abstract constructors could be allowed, too.

interface Serializable {
public abstract Serializable();
}

This would require each implementation of Serializable to provide a
public no-arg constructor.
(Maybe the following would be a better syntax:
interface Serializable<S extends Serializable> {
public abstract S();
}
)


Now I present two examples where it would be useful.

(1) Eliminate or reduce the use of reflection in serialization frameworks.
One example was given in Note 4 --- the presence of no-arg constructor in
a serializable class would be checked at compile-time rather than at run-
time.

For a more sophisticated usage some new API and additional support from
compiler is required. (The following may not be the best way to extend
the API, but I hope it will serve well for illustration.)

Imagine a new magic class Implementation<T>. This class will have no
methods on its own, but on its instances we will be able to call the same
methods as on the class T. (For this, the compiler magic would be
necessary.)

Example:

interface J {
J(int x);
static void f();
void g();
}

class A implements J {
A(int x){...}
static void f(){...}
void g(){...}
}

Implementation<J> I = A.class.asImplementationOf(J.class);
I.new(5); // OK, calling the constructor A(int x)
I.f(); // OK, calling static method A.f()
I.g(); // ERROR, calling an instance method without an instance of J

Notice extending the Class API by adding new method
<T> Implementation<T> asImplementationOf(Class<T> clazz);
The restriction would apply that the type T is known at compile time.

Now back to usage in serialization frameworks. The above API would
automate the verification that all required constructors and static
methods are present in a class:

interface MySerializable<S extends MySerializable<S>> {
public static S readObject(ObjectInputStream in);
}

Class<?> cls = Class.forName("com.example.MySerializableClass");
Implementation<MySerializable> M =
cls.asImplementationOf(MySerializable.class);
MySerializable obj = M.readObject(in);

Note that the verification that MySerializableClass really implements
MySerializable interface would be automatically done in the
Class.asImplementationOf() method, thus saving much of the reflection
code.


(2) The second use case is with generics, but would require reified
generics (which I hope will appear in some future version of Java).

Suppose you have an abstract class Vector which represents a vector in a
vector space (don't confuse with java.util.Vector) and a few
implementations, like Vector2D and Vector3D.

abstract class Vector<V extends Vector<V>> {
public abstract V add(V v); // returns the sum of v and this
...
}

class Vector2D extends Vector<Vector2D> {
public static Vector2D zero(); // returns (0,0)
...
}

class Vector3D extends Vector<Vector3D> {
public static Vector3D zero(); // returns (0,0,0)
...
}

Now let's have a generic class that will use vectors and do operations on
them, but doesn't really care about their actual dimension. So it will
work with abstract type Vector. But for some operations it may be
necessary to obtain the zero vector, without explicitely knowing the
actual type of vector. We may want to write something like this:

class MyClass<V extends Vector<V>> {
public void someMethod(){
V v = V.zero();
...
}
}

This is of course not possible, but could be made possible if Vector
specified abstract static method zero():

abstract class Vector<V extends Vector<V>> {
public static abstract V zero();
public abstract V add(V v);
...
}

We would further change the declaration of MyClass to

class MyClass<V implements Vector<V>> {
...
}

The keywork extends was exchanged for implements (another syntax
extension). The compiler would know that a type parameter V that fully
implements all abstract static methods of Vector is required (so, for
example, Vector itself would not be a valid type parameter of MyClass).
 
T

Tomas Mikula

Tomas said:
I have searched this group for "abstract static methods" and found a
couple of threads, but I think none of them was discussing the kind of
semantics I am going to describe. As you might have guessed, I believe
it would be useful :). [...]

You aren't the first. However, in C++, C#, Java, and languages like
them, you simply are never going to have methods that are both virtual
and static. And since abstract implies virtual, that rules out abstract
static methods too.

I'm not talking about virtual static methods. In this case, abstract
would not imply virtual. My second use case could most probably be
accomplished in C++ by Concepts (though Concepts didn't make it to C+
+0x). In fact, it is achievable with current C++ templates, only without
compile time checking --- it is possible to call T::staticMethod() where
T is a template parameter.
 
M

Marcin Rzeźnicki

I have searched this group for "abstract static methods" and found a
couple of threads, but I think none of them was discussing the kind of
semantics I am going to describe. As you might have guessed, I believe it
would be useful :). I further believe it is fully complatible with the
current language, but there might be caveats I have overlooked. I'm
wonder if you would find it as useful as I do and if you see any problems
with it. I know it is a long post and some parts may be difficult to
understand. Therefore I will be thankful if you can read it all and think
about it.

Hi,
Interesting but I doubt it is going to be useful. First of all,
because statics should remain non-inheritable, static abstract
actually forces each subclass to implement its own definition -
extreme nuisance in my opinion. Example with generics can easily be
substituted by some kind of "trait" parameter or suitable simple
design pattern (for example Factory), or even with classic sub-typing
(zero vector needs not know its dimension, it can simply 'answer' with
neutral element of the ring on which it is constructed for each and
every component query), no big win here either (eliminating type
erasure is extremely welcome but for other reasons). One big advantage
of inheritance is, in my opinion, that it enables you to compose more
specialized classes from generic ones, it is easy to imagine algebraic
ordering relation between types based on inheritance. Your version of
statics breaks this assumption without promise of any reward in
exchange.
However, I like the idea of "interface constructor". It might be handy
to force every client to provide necessary piece of data. To be more
useful for this purpose I'd postulate to create the rule which says
that every derived constructor has to call this 'interface
constructor'. But still, no big deal. It is possible to impose
semantics I am talking about with or without it.
 
J

Joshua Cranmer

I have searched this group for "abstract static methods" and found a
couple of threads, but I think none of them was discussing the kind of
semantics I am going to describe. As you might have guessed, I believe it
would be useful :). I further believe it is fully complatible with the
current language, but there might be caveats I have overlooked. I'm
wonder if you would find it as useful as I do and if you see any problems
with it. I know it is a long post and some parts may be difficult to
understand. Therefore I will be thankful if you can read it all and think
about it.

I saw a more detailed proposal at <http://kijaro.dev.java.net>, which
included a prototype implementation. The specification for said stuff is
linked here: <http://www.jroller.com/jadda/entry/meta_interfaces_revisited>.
 
A

Andreas Leitgeb

Tomas Mikula said:
I have searched this group for "abstract static methods" ...
... presence of no-arg constructor in a serializable class would be
checked at compile-time rather than at run-time.

I think this is easily misunderstood. The newly possible compiletime
check would be for compiling the *concrete class* whose name you later
intend to specify dynamically at runtime. This does have some merit.

Still no compiletime check would of course be possible at the place
where you'd *use* that class dynamically, so nothing at all can be
helped about the reflection-part of this story.

There's still a slight "problem" with abstract static methods:
With respect to static members, *every* class is concrete.
Maybe you thought about static "create"-methods, as alternative
to enforced constructors?
 
T

Tomas Mikula

I think this is easily misunderstood. The newly possible compiletime
check would be for compiling the *concrete class* whose name you later
intend to specify dynamically at runtime. This does have some merit.

Still no compiletime check would of course be possible at the place
where you'd *use* that class dynamically, so nothing at all can be
helped about the reflection-part of this story.

The idea here was to automate the reflection. Suppose a class is required
to have many static methods. Instead of checking if they are actually
present and have the correct signature manually, the method
Class.asImplementationOf(MyInterface.class) would do all the checking for
me.
There's still a slight "problem" with abstract static methods: With
respect to static members, *every* class is concrete. Maybe you thought
about static "create"-methods, as alternative to enforced constructors?

If an abstract class MyClass does not have an implementation of a static
abstract method staticMethod() specified in its superclass/interface, it
can be detected in both the direct use and through reflection:

MyClass.staticMethod(); // this would be a compile-time error
Class.forName("MyClass").asImplementationOf(MyInterface.class);
// this would be a run-time error

My view is that with respect to abstract static members, not every class
would be concrete.
 
T

Tomas Mikula

Hi,
Interesting but I doubt it is going to be useful. First of all, because
statics should remain non-inheritable, static abstract actually forces
each subclass to implement its own definition - extreme nuisance in my
opinion.

Yes (unless the subclass is abstract). I think in some cases it is
reasonable (as with the enforced no-arg constructor in Serializable, or
some other serialization static methods. For example, if I want to
deserialize an immutable object, I need to do it by a static method or a
special constructor, because the non-static readObject(...) method in
Java's serialization API is a mutator method. In my opinion it is
reasonable to enforce own implementation of a (de)serialization method).
Example with generics can easily be substituted by some kind of
"trait" parameter

Sorry, I don't know what you mean by "trait" parameter? Do you mean that
I would call the zero() method on some instance?
Like myDummyVector.zero()?
or suitable simple design pattern (for example
Factory), or even with classic sub-typing (zero vector needs not know
its dimension, it can simply 'answer' with neutral element of the ring
on which it is constructed for each and every component query),

Allowing operations between a concrete vector and this general zero
vector would require to also allow operations between 2D and 3D vetors -
the original type safety would disappear.
no big
win here either (eliminating type erasure is extremely welcome but for
other reasons). One big advantage of inheritance is, in my opinion, that
it enables you to compose more specialized classes from generic ones, it
is easy to imagine algebraic ordering relation between types based on
inheritance. Your version of statics breaks this assumption without
promise of any reward in exchange.

I don't see how it breaks this relation between classes. Also now it is
possible to hide supertype's static methods by own implementation. I
would only add that in some cases this hiding would be required.
 
M

Marcin Rzeźnicki

Yes (unless the subclass is abstract). I think in some cases it is
reasonable (as with the enforced no-arg constructor in Serializable, or
some other serialization static methods. For example, if I want to
deserialize an immutable object, I need to do it by a static method or a
special constructor, because the non-static readObject(...) method in
Java's serialization API is a mutator method. In my opinion it is
reasonable to enforce own implementation of a (de)serialization method).

Hi, possibly it is reasonable, but what is wrong with how it is done
today (readObject/writeObject) which you are not required to implement
if default behavior suffices?
Sorry, I don't know what you mean by "trait" parameter? Do you mean that
I would call the zero() method on some instance?
Like myDummyVector.zero()?

I borrowed the terminology from C++. More or less, you add type
parameter (let's say <Zero extends ZeroVector>) which has a method
like getZeroVector() (strictly speaking ZeroVector has this method).
Actual type parameter provides concrete implementation.
Allowing operations between a concrete vector and this general zero
vector would require to also allow operations between 2D and 3D vetors -
the original type safety would disappear.

I don't get it, could you provide an example?
I don't see how it breaks this relation between classes. Also now it is
possible to hide supertype's static methods by own implementation. I
would only add that in some cases this hiding would be required.

I was not very clear, it was late when I was writing :) I guess what
I was trying to say was that you can impose ordering based on
specialization (as opposed to parent-child relationship). Each class
in an inheritance chain either extends or redefines partially its
ancestor (I am using 'or' as logical or). Therefore each class is
either more specialized (if redefinition occurs and it accepts
stronger contract, as in Rectangle->Square) or equally specialized (if
extension occurs and all redefinitions do not change contract - I
treat extension as an interface extension so that class can be used
_additionally_ in different context). Your proposal forces implementor
to provide implementation for non-inheritable method, so it really
can't take any benefit from redefinitions up the chain. Therefore all
concrete classes are at most equally specialized as their context of
usage is determined by a static method. So it does not play well with
most "inheritance patterns". That's how I see it.
 
M

Marcin Rze¼nicki

I borrowed the terminology from C++. More or less, you add type
parameter (let's say <Zero extends ZeroVector>) which has a method
like getZeroVector() (strictly speaking ZeroVector has this method).
Actual type parameter provides concrete implementation.

I forgot to clarify - you are right, all in all you call zero() method
on some instance :)
 
T

Tomas Mikula

Hi, possibly it is reasonable, but what is wrong with how it is done
today (readObject/writeObject) which you are not required to implement
if default behavior suffices?

I'm saying it is wrong, but just don't like that the implementation
requires a lot of reflection. (I don't mind that implementation of
statndard Java API requires reflection, because someone has already
implemented it for me. But if I want to create my own serialization
framework (e.g. for xml serialization), I need to do a lot of
reflection which could be automated.) Probably one thing I find wrong
with readObject - as I already mentioned, it prevents the object to be
immutable. Though this could also be solved by declaring it static and
use reflection.
I borrowed the terminology from C++. More or less, you add type
parameter (let's say <Zero extends ZeroVector>) which has a method
like getZeroVector() (strictly speaking ZeroVector has this method).
Actual type parameter provides concrete implementation.

I don't see how this would help. Would I call Zero.getZeroVector()?
Probably you meant something else because this leads to the same
problem with calling static method getZeroVector() on a type
parameter. Could you provide an example?
I don't get it, could you provide an example?

If I understood well, you meant something like this:

Class Vector {
public static Vector getZeroVector(){
return someSpecialZeroVectorInstance;
}
public abstract Vector add(Vector v);
}

Class Vector2D {
public Vector add(Vector v){...}
...
}

class MyClass<V extends Vector> {
...
V v; // V is some concrete class, such as Vector2D
...
Vector zero = Vector.getZeroVector();
v.add(zero); // adding a general Vector zero to concrete v
// if this is allowed, then also the following is
v.add(new Vector3D(1,2,3)); // summing 2D and 3D vector
...
}
I was not very clear, it was late when I was writing :) I guess what
I was trying to say was that you can impose ordering based on
specialization (as opposed to parent-child relationship). Each class
in an inheritance chain either extends or redefines partially its
ancestor (I am using 'or' as logical or). Therefore each class is
either more specialized (if redefinition occurs and it accepts
stronger contract, as in Rectangle->Square) or equally specialized (if
extension occurs and all redefinitions do not change contract - I
treat extension as an interface extension so that class can be used
_additionally_ in different context). Your proposal forces implementor
to provide implementation for non-inheritable method, so it really
can't take any benefit from redefinitions up the chain. Therefore all
concrete classes are at most equally specialized as their context of
usage is determined by a static method. So it does not play well with
most "inheritance patterns". That's how I see it.

Now I don't get it. Can you provide an example where you have a class
and its specialized subclass and adding an abstract static method to
their interface removes/prohibits this specialization?
 
M

Marcin Rze¼nicki

I'm saying it is wrong, but just don't like that the implementation
requires a lot of reflection. (I don't mind that implementation of
statndard Java API requires reflection, because someone has already
implemented it for me. But if I want to create my own serialization
framework (e.g. for xml serialization), I need to do a lot of
reflection which could be automated.) Probably one thing I find wrong
with readObject - as I already mentioned, it prevents the object to be
immutable. Though this could also be solved by declaring it static and
use reflection.

Yes, but someone did it for you either - JAXB, xStreams etc. This is
not a type of work you do day-in day-out, so benefits are rarely to be
seen
I don't see how this would help. Would I call Zero.getZeroVector()?
Probably you meant something else because this leads to the same
problem with calling static method getZeroVector() on a type
parameter. Could you provide an example?

OK
public class MyVector<T, Zero extends ZeroVector> extends Vector2D<T>
{
....
public MyVector(Zero zero) { this.zero = zero; }
....
}

If I understood well, you meant something like this:

Class Vector {
public static Vector getZeroVector(){
return someSpecialZeroVectorInstance;
}
public abstract Vector add(Vector v);

}

Class Vector2D {
public Vector add(Vector v){...}
...

}

class MyClass<V extends Vector> {
...
V v; // V is some concrete class, such as Vector2D
...
Vector zero = Vector.getZeroVector();
v.add(zero); // adding a general Vector zero to concrete v
// if this is allowed, then also the following is
v.add(new Vector3D(1,2,3)); // summing 2D and 3D vector
...



}

Right but implementation of addition surely checks for this case,
doesn't it?
public void add(Vector<? extends T> v) { if (v.getDimension() !=
this.getDimension() ) throw new IllegalArgumentException(); }
So it suffices to have a factory method for appropriate zero vectors
Vector zero = Vector.getZeroVector(2);

Now I don't get it. Can you provide an example where you have a class
and its specialized subclass and adding an abstract static method to
their interface removes/prohibits this specialization?

Yes, consider
public abstract class IOStream //for reading disk streams {
public abstract static boolean isReadable(File f) //returns true
for files which a concrete class can hopefully process.
....
}

public class LocalIOStream extends IOstream {
public static boolean isreadable(File f) { return f.isLocalFile(); }
....
}


public class AudioVideoStream extends LocalIOStream {
???
}

in AVStream you have, if I understood you correctly, two choices -
either to redo all work of super-classes which is not really an
option, let's say,
public static boolean isReadable(File f) { return f.isLocalFile() &&
(f instanceof AudioFile && ((AudioFile)f).getAudioCodecID().equals
(...);}
or omit it so then you impose different context. Namely, pretend to be
able to read remote files while you are not.
And one more question:
//client code
Stream s = new AudioVideStream(..);
read10Bytes(s);

public byte[] read10Bytes(Stream s) {
if (!Stream.isReadable(file)) //how would you dispatch it? There is no
way I suppose
}
 
A

Andreas Leitgeb

Tomas Mikula said:
The idea here was to automate the reflection.
Sorry, that is a non-starter.

If you know the class at compiletime, you can just use it with
the "new"-operator and have all kinds of compile time checks.

If you don't know the class at compiletime, then neither does the
compiler, so there's nothing the compiler could possibly do for you
beyond what it already does, namely write bytecode to have the JVM
check it all at runtime.

Even at runtime, there's no saving: both, interface and existence
of relevant methods and constructors, each have to be checked
separately by the JVM.

I still see some merit in being able to enforce that any concrete
class implementing some thusly declared interface had to offer some
particular c'tor, as a means to help developers of such classes to
not forget about it.


About the static methods: if you need that kind of enforcement for
dynamically used classes, then just use instances and non-static
methods as helpers:

public interface Foo { // known to the user at compiletime.
public void pseudoStatic();
}
public class FooBar { // known to the user only at runtime
public void pseudoStatic() { realStatic(); }
public static void realStatic() { /* do something ... */ }
}
// snippet of user's code:
Foo x = (Foo)use_reflection_to_get_instance(implName); // implName=="FooBar"
x.pseudoStatic();

Up to minor syntactical differences this FooBar object does what your
".asImplementationOf()" result was intended to do, if I understood it
correctly. I don't think, that calling static methods on dynamically
named classes is worth such deep changes as you seem to have in mind
for this task.
 
M

Marcin Rze¼nicki

I still see some merit in being able to enforce that any concrete
class implementing some thusly declared interface had to offer some
particular c'tor, as a means to help developers of such classes to
not forget about it.

Yep, this is not bad.
 
A

Arved Sandstrom

Marcin said:
Yep, this is not bad.

I prefer the annotations-based method such as described here:
http://www.javaspecialists.eu/archive/Issue167.html

It works very cleanly - my annotations processors are in a separate JAR
that I include on the javac classpath. Strictly speaking there's no need
to specify the processor path if doing this; it will default to the user
classpath if no processor path is specified. The only change I need to
make to my "real" source is the actual annotations, like
@NoArgsConstructor in the example, and quite frankly on the
implementation classes is where I personally want to enforce a condition
like this.

Because of the @Inherited annotation on the @NoArgsConstructor
annotation it becomes particularly handy. I have found use of this
approach when a large number of JPA @Entity classes inherit from a
@MappedSuperclass - it's not uncommon to want to supply some entities
with useful ctors (and if doing so carelessly the no-args ctor goes
away); using this kind of annotation on the @MappedSuperclass catches
all these problems at compile time.

AHS
 
A

Andreas Leitgeb

Here's my objection: Suppose there's an Entertainer interface
(or abstract class) and ...
...
Okay, it might make sense for the class of Comedians to have a
default stale Joke (a faithful model of reality, perhaps), ... :)
...
The author of Entertainer, who knew nothing about the wants and needs
of those who would come later, ...

Thanks for the entertaining example, but I think it's beside the point.
This type of argument "it's bad for this exemplary usecase, so it must
be bad for all usecases" is obviously flawed. (or was a joke, itself)

On second thought: If the Entertainers were designed to be dynamically
loaded by name, then Comedians just wouldn't have any chance of a individual
default joke. They could offer their Joke- constructor, but unless they
also offered a no-args one, they just wouldn't ever be successfully engaged.

This whole topic is inspired by dynamic loading of classes. Otherwise, there
wouldn't really be any use for dictating constructors at all. Dynamic loading
of classes seems to me of increasing importance with all those AppServers,
J2EE, ... Demanding the default-constructor (or even with a specific set
of arguments) for those classes imposes no new restriction, just formalizes
the restrictions that were already imposed by documentation and use.

PS: In recent threads I spoke up against restrictions, and now I promote them?
It's different types of restrictions, of course: an extra method or c'tor
is easily added as a dummy, but an idly added "final" is much harder to come
by, if deemed improper, later.
 
A

Andreas Leitgeb

Marcin Rzeźnicki said:
Nice, it wins :)

Indeed nice, but what would be the extra effort to create e.g. a
@StringArgConstructor annotation and its processing? And then also
a @StringStringArgConstructor and a @StringMyFooIntArgConstructor, ...

As long as this annotation, its processor, and a mechanism to create
annotations for any particular constructor-signature aren't yet in the
standard, I wouldn't deem them a full substitute...

But then again, they do not depend on my deeming it anything... :)
 
T

Tomas Mikula

Yes, but someone did it for you either - JAXB, xStreams etc. This is
not a type of work you do day-in day-out, so benefits are rarely to be
seen

Occasionally new frameworks appear. Not an everyday work, but for me
it justifies the introduction of a new feature, if there is no hidden
danger we haven't noticed so far. I accept that for you it's not a
sufficient reason.
OK
public class MyVector<T, Zero extends ZeroVector> extends Vector2D<T>
{
...
public MyVector(Zero zero) { this.zero = zero; }
...

}

MyVector<Integer, Zero2D> vec = new MyVector(Vector2D.zero());

OK, but when we already resort to obtaining a zero vector from another
instance, we don't need a reference to zero stored in each instance of
vector. We can just have a nonstatic zero() method:

abstract class Vector<T, V extends Vector<V>> {
public abstract V zero();
}

class Vector2D extends Vector<Integer, Vector2D> {
public static final Vector2D ZERO = new Vector2D(0, 0);
public Vector2D zero(){ return ZERO; }
}

I wanted to obtain zero without a reference to an instance. Using an
instance is unnatural and sometimes an instance is just not at hand.
Right but implementation of addition surely checks for this case,
doesn't it?

Not necessarily:

abstract class Vector<V extends Vector<V>> {
public V add(V v);
}

class Vector2D extends Vector<Vector2D> {
private final int x, y;
public Vector2D(int x, int y){ this.x = x; this.y = y; }
public Vector2D add(Vector2D v){
return new Vector2D(this.x + v.x, this.y + v.y);
}
}

No checking that the argument of addition has the correct type,
because this is enforced by the compiler.
public void add(Vector<? extends T> v) { if (v.getDimension() !=
this.getDimension() ) throw new IllegalArgumentException(); }
So it suffices to have a factory method for appropriate zero vectors
Vector zero = Vector.getZeroVector(2);





Yes, consider
public abstract class IOStream //for reading disk streams {
public abstract static boolean isReadable(File f) //returns true
for files which a concrete class can hopefully process.
...

}

public class LocalIOStream extends IOstream {
public static boolean isreadable(File f) { return f.isLocalFile(); }
...

}

public class AudioVideoStream extends LocalIOStream {
???

}

in AVStream you have, if I understood you correctly, two choices -
either to redo all work of super-classes which is not really an
option, let's say,
public static boolean isReadable(File f) { return f.isLocalFile() &&
(f instanceof AudioFile && ((AudioFile)f).getAudioCodecID().equals
(...);}

You don't have to redo the work, you can call the superclass's static
method as usual:

public static boolean isReadable(File f){
return LocalIOStream.isReadable(f) &&
f instanceof AudioFile &&
((AudioFile)f).getAudioCodecID().equals(...);
}

Furthermore, if we expect that specialized IOStreams will only be able
to process instances specialized instances of File, the IOStreams
could be parametrized by the type of the File.

abstract class IOStream<F extends File> {
public abstract static boolean isReadable(F f);
}

class LocalIOStream<F extends File> extends IOStream<F> {
public static boolean isReadable(F f){
return f.isLocalFile();
}
}

class AudioVideoStream extends LocalIOStream<AudioFile> {
public static boolean isReadable(AudioFile f){
return LocalIOStream.isReadable(f)
&& f.getAudioCodecID().equals(...);
}
}
or omit it so then you impose different context. Namely, pretend to be
able to read remote files while you are not.
And one more question:
//client code
Stream s = new AudioVideStream(..);
read10Bytes(s);

public byte[] read10Bytes(Stream s) {
if (!Stream.isReadable(file)) //how would you dispatch it? There is no
way I suppose

}

This would be a compile-time error, since isReadable() is abstract in
Stream.
 
T

Tomas Mikula

Sorry, that is a non-starter.

If you know the class at compiletime, you can just use it with
the "new"-operator and have all kinds of compile time checks.

If you don't know the class at compiletime, then neither does the
compiler, so there's nothing the compiler could possibly do for you
beyond what it already does, namely write bytecode to have the JVM
check it all at runtime.

In my original post I noted that the use of Class's newly introduced
method
<T> Implementation<T> asImplementationOf(Class<T> clazz);
would have the restriction that the type T is known at compile time.
In this case, the compiler can generate the bytecode to check if the
'this' class implements T.
Even at runtime, there's no saving: both, interface and existence
of relevant methods and constructors, each have to be checked
separately by the JVM.

Although my major intention was to reduce writing reflective code,
there could also be a run-time saving: as soon as the JVM loads a
class A, it will know if it 'statically implements' interface J. (By
the same mechanism as it knows if A implements interface I.)
'Statically implements' would just be a new kind of relationship
between classes, in addition to 'extends' and 'implements'.
I still see some merit in being able to enforce that any concrete
class implementing some thusly declared interface had to offer some
particular c'tor, as a means to help developers of such classes to
not forget about it.

About the static methods: if you need that kind of enforcement for
dynamically used classes, then just use instances and non-static
methods as helpers:

public interface Foo {  // known to the user at compiletime.
   public void pseudoStatic();}

public class FooBar {   // known to the user only at runtime
   public void pseudoStatic() { realStatic(); }
   public static void realStatic() { /* do something ... */ }}

// snippet of user's code:
Foo x = (Foo)use_reflection_to_get_instance(implName); // implName=="FooBar"
x.pseudoStatic();

Up to minor syntactical differences this FooBar object does what your
".asImplementationOf()" result was intended to do, if I understood it
correctly.

Yes, but:
- it requires to get an unnecessary instance (not so bad yet);
- getting this instance requires reflection
- using reflection for getting an instance requires conventions
which cannot be checked at runtime (such as the presence of
some particular (e.g. no-arg) constructor)

If I'm interested in just one static method, it turns out I could just
use reflection to get this Method instead of a dummy instance.
Furthermore, if I forget to override pseudoStatic() or realStatic() in
a subclass, I will get the realStatic() from superclass, which is not
what I want. The compiler will not enforce me in any way to override
them.
 I don't think, that calling static methods on dynamically
named classes is worth such deep changes as you seem to have in mind
for this task.

The good thing about it is that the changes are not real changes, just
extensions. So far I think they are all backward compatible with
current specification. No old code would be broken if these extensions
are introduced.
 
M

Marcin Rze¼nicki

Occasionally new frameworks appear. Not an everyday work, but for me
it justifies the introduction of a new feature, if there is no hidden
danger we haven't noticed so far. I accept that for you it's not a
sufficient reason.

It is, but I am trying to bring up some dangers of your method
throughout this thread.
OK, but when we already resort to obtaining a zero vector from another
instance, we don't need a reference to zero stored in each instance of
vector. We can just have a nonstatic zero() method:

abstract class Vector<T, V extends Vector<V>> {
public abstract V zero();

}

class Vector2D extends Vector<Integer, Vector2D> {
public static final Vector2D ZERO = new Vector2D(0, 0);
public Vector2D zero(){ return ZERO; }

}

I wanted to obtain zero without a reference to an instance. Using an
instance is unnatural and sometimes an instance is just not at hand.

Right, but that was just an example of what is 'trait'. I am not
saying this is necessarily the best design decision in this case.
Not necessarily:

abstract class Vector<V extends Vector<V>> {
public V add(V v);

}

class Vector2D extends Vector<Vector2D> {
private final int x, y;
public Vector2D(int x, int y){ this.x = x; this.y = y; }
public Vector2D add(Vector2D v){
return new Vector2D(this.x + v.x, this.y + v.y);
}

}

No checking that the argument of addition has the correct type,
because this is enforced by the compiler.

Formal arguments have to be invariant with respect to overriding in
Java, you simply created method overload which will be used when
compiler is sure that runtime type of argument will be Vector2D. You
will still have to provide 'generic' add method.
Your example does not help either (or I cannot see how it would)
because you will not be able to dispatch on v's actual type unless you
change how invokestatic works.

You don't have to redo the work, you can call the superclass's static
method as usual:

public static boolean isReadable(File f){
return LocalIOStream.isReadable(f) &&
f instanceof AudioFile &&
((AudioFile)f).getAudioCodecID().equals(...);

}

Yeah, right, but consider what happens when someone implements
multiple interfaces, or when inheritance tree changes, or when someone
inherits multiple interfaces with conflicting statics and so on. This
example is basically hand-crafted implementation of virtual
dispatch :)
Furthermore, if we expect that specialized IOStreams will only be able
to process instances specialized instances of File, the IOStreams
could be parametrized by the type of the File.

abstract class IOStream<F extends File> {
public abstract static boolean isReadable(F f);

}

class LocalIOStream<F extends File> extends IOStream<F> {
public static boolean isReadable(F f){
return f.isLocalFile();
}

}

class AudioVideoStream extends LocalIOStream<AudioFile> {
public static boolean isReadable(AudioFile f){
return LocalIOStream.isReadable(f)
&& f.getAudioCodecID().equals(...);
}

}

Well, ok, but it does not change anything. Still you have to re-
implement invokevirtual by hand all the time :)
or omit it so then you impose different context. Namely, pretend to be
able to read remote files while you are not.
And one more question:
//client code
Stream s = new AudioVideStream(..);
read10Bytes(s);
public byte[] read10Bytes(Stream s) {
if (!Stream.isReadable(file)) //how would you dispatch it? There is no
way I suppose

This would be a compile-time error, since isReadable() is abstract in
Stream.

This is really bad :) Then actually your statics will be usable only
when you know exact type you want to work with.
 

Ask a Question

Want to reply to this thread or ask your own question?

You'll need to choose a username for the site, which only take a couple of moments. After that, you can post your question and our members will help you out.

Ask a Question

Members online

No members online now.

Forum statistics

Threads
473,828
Messages
2,569,736
Members
45,520
Latest member
JudithMorf

Latest Threads

Top