Client code can't copy the object (except through
object->clone() or the like), because the client has only a
pointer to some abstract base type.
That's one situation, appropriate for some applications, but not
everywhere. What happens with the template method pattern, for
example, where the client has to derive himself in order for the
pattern to work. Most of the time, the "client" is responsible
for creating the object.
(I put "client" in quotes, because I'm not sure the word is
really appropriate. Most such objects in my applications are at
the application level---there is no higher level which can be
considered a "client", for which they are a service provider.)
Within the factory that creates the objects, though, there's
no mystery about the exact derived types involved, so they can
be created, copied, or what have you. The factory has to
allocate the object dynamically so that the object will
continue to exist after the factory function returns. The
alternative, which I posted elsewhere in this thread, is for
the client code to give the factory a "receiver" to which the
object's address can be passed. The code in the receiver
completes before the factory function returns, and therefore
before the object goes out of scope.
But the receiver has to know the exact type, which sort of
defeats the role of the factory.
At any rate, I don't see what this buys you. Instead of just
calling delete (on a pointer to the base, or more often, on
this), the code has to find the correct type, to find the
receiver, and remove the object from the receiver. This sounds
like a lot of extra complication (and obfuscation) for nothing.
Since ever. I don't know what you mean by "identity is identity."
Just what I said. An entity object has identity. You can't
copy it---two different objects are two different objects, and
one can't replace the other.
The typical example I use is that of a monetary amount and a
bank account. A monetary amount is a value. You can (and
usually do) copy and assign it. You practically never allocate
it dynamically. A bank account is an entity. A copy of my bank
account is not my bank account, and if you're crediting a
monetary amount to it, you have to do it on the specific object,
not some copy which happens to be equal.
Normally, entity objects do NOT support copy and assignment.
This isn't always strictly true for copy---transaction
management often involves copies in order to implement roll
back. But I've never seen one which supported assignment---in
practice, assignment and polymorphism don't work well together.
Whether you can copy a particular object almost always depends
on whether you have a pointer to its most-derived type, or to
some abstract base.
Whether you can copy a particular object depends on whether it
supports copy. Most polymorphic objects don't (or shouldn't).
And of course, if the object is to be kept in a container, it
needs to support assignment as well.
The factory that creates objects implementing a particular
interface knows the derived types directly; if it didn't, it
couldn't instantiate them. There's no reason the factory
can't copy those objects at will.
Except that by design, the object doesn't support copy and
assignment.
(I have one or two cases in my code where an object supports
copy only because I need to bind it to a reference in some
cases. In such cases, the copy constructor only works if the
object has just been created; once it has been used for anything
else, calling the copy constructor will cause an assertion
failure. Such a scheme could be used in your case. But again:
why?)
It does not (IMO) make the code more complicated. The factory
no longer needs to call new or delete directly, no longer
needs an explicitly declared destructor to delete the objects,
and has immediate access to all objects of a particular
derived type.
Whoever decides to delete the object has to know the most
derived type, in order to find the appropriate container.
That's an enormous complication, and goes against all of the
principles of OO design. And it doesn't buy you anything that I
can see---what's the difference between calling delete (which
every C++ programmer knows will delete the object), and calling
erase on some container (where the relationship between the
pointer one has and the element in the container that is being
deleted may not be obvious)? I'd call using containers for this
obfuscation and additional complication.
And of course, you're not always using a factory---things like
the template method pattern are quite popular in GUI libraries,
for example.
Suppose a factory stores a single vector of
pointers-to-Abstract, rather than a separate list for each
derived type. Now suppose the factory needs to access all
objects of some particular derived type.
Why would it need to do that? Is it a factory, or is it
something else---a factory has nothing to do with the object
once it has created it.
The popular solution is to walk through the entire vector,
checking each element's dynamic type with dynamic_cast.
The usual solution is to forget the derived type entirely once
the object has been created. There are some cases where
dynamic_cast is appropriate, but not very many. The whole point
of derivation (in this case) is that the derived type isA base
type---you need to know the derived type in order to create the
object, but once that's done, you have a base, and that's all
you want to know. (There are exceptions, of course, but they
aren't that frequent.)
stream& operator<<(std: