On Mon, Jul 11, 2005 at 09:46:30AM -0400, Stevan Little wrote:
: Ingo,
: On Jul 11, 2005, at 9:16 AM, Ingo Blechschmidt wrote:
: >Hi,
: >
: >  class Foo {}
: >  class Bar is Foo {}
: >
: >  Bar.new.isa(Object);    # true
: >  Bar.new.isa(Class);     # false
: >  Bar.new.isa(Foo);       # true
: >  Bar.new.isa(Bar);       # true
: >  # These are clear, I think.
: Yes, these all make sense to me.
: >
: >  Bar.isa(Object);        # true
: >  Bar.isa(Class);         # true
: >  Bar.isa(Foo);           # ? (my guess: false)
: >  Bar.isa(Bar);           # ? (my guess: false)
: I am not sure about this. I think that .isa as a class method should 
: behave much as it does for an instance method.

Right, or you can't easily decide whether Bar isa Foo in the abstract.
You need to able to reason about the relationships of user-defined
classes in the absence of instances.

: If we start supporting 
: things like Bar.isa(Class) then we start exposing the soft underbelly 
: of the meta-model to the outside world. Which IMO might not be a good 
: idea.

Bar.isa(Class) probably false in any event, since I think Class
is probably a role rather than a class.  But "Class" is almost
certainly not the name of the metaclass either.  Or at least,
it's not the name of *both* metaclasses (counting the Class role
as one of them).  Basically every user-defined class has a Platonic
description (known as its "Class") and an Aristotelian description
(known as its "MetaClass").  If "Class" is a valid name for a class,
it's the Platonic one, not the Aristotelian one, and it would be
a class only because "Class" can refer either to the role or to an
anonymous class generated from the role, if we follow the idea that Int
is both a role and a class.  So maybe Bar.isa(Class) in that sense.
Or maybe we should force .isa false even if there is an associated
anonymomus class, just to keep things straight.

So what is certainly true is that Bar.does(Class), where the Class role
describes the Platonic interface of items like Bar.  That is to say,
Bar is the stand-in for all the members of its class when you don't
actually have a member.  Its Platonic role is to know about Barness,
not about classness.  We call call it the dispatcher class because
it knows how to dispatch things of type Bar whether you actually have
an instance of type Bar or not.  That's why method calls on it can do
things like call constructors.  But it doesn't know much about any of
the grubby Aristotelian workings of the metaclass.  It just knows there
is one, and that it can delegate all the messy practicalities to it.
I suspect the metaclass of "Class" actually has the name "class" or
"CLASS", and the metaclass of a Role is actually "role" or "ROLE".
Perhaps the "class" and "role" keywords are just specific examples
of the grammar being smart enough to treat declared metaclasses as
BEGIN analogues.  But maybe it's smarter to keep the "class" and "CLASS"
identifiers at arms length from each other.  (If for no other reason,
to keep our sanity while discussing them.)

So anyway, that gives us something like:

    Bar.isa(Foo);               # true
    Bar.isa(Bar);               # true
    Bar.isa(Class)              # false presuming Class is only a role
    Bar.does(Class)             # true
    Bar.does(CLASS)             # false
    Bar.meta.isa(Foo);          # false
    Bar.meta.isa(Bar);          # false
    Bar.meta.does(Class)        # false
    Bar.meta.isa(CLASS)         # false presuming CLASS is only a role
    Bar.meta.does(CLASS)        # true

You know, this class/role distinction might just be what saves our
collective bacon on the usual infinite regress of metaclasses, if a
metaclass turns out to be an object of anonymous class but named
role.  The bootstrap might reside entirely in the code that constructs
the metaclass instance of anonymous class but filling the CLASS role
in its abstract interface.

I mentioned the idea of autogenerating a class Int from a role Int,
but maybe some roles don't allow themselves to be autogenerated into
classes, and maybe CLASS is one of them.  And it sounds to me like
the absence of that feature is caused simply by ROLE's autogenerate
interface being left with an implementation of {...} and not overridden
in the anonymous metaclass representing the CLASS role.

So I guess the deep answer to the infinite recursion problem is that,
yes, it's metaclasses all the way down, but if you actually try to
pursue it, you'll at some point run into "method not yet implemented"
because nobody has cared about it that deeply yet.  So we basically
get the possibility of mapping to anyone's metamodel without actually
having to commit to the fanciest one.  I think I like that, even if
I don't understand it.  Or maybe *because* I don't understand it.


Reply via email to