Languguage OS 2/Languguage OS II Version 10-94 (Knowledge Media)(1994).ISO/gnu/objcissu.lha/multiple-inheritance

Internet Message Format | 1993-02-07 | 11KB

Date: Thu, 28 Jan 93 20:13:49 -0600
From: Paul Burchard <[e-mail]>
To: [e-mail]
Subject: How to do better than Multiple Inheritance
Cc: [e-mail], [e-mail]
Reply-To: [e-mail]

[Since this is an important issue I am forwarding this to a couple of
the GNU mailing lists...]

Dan LaLiberte <[e-mail]> writes:
> What do you think about the possibility of adding
> multiple inheritance to GNU Objective-C?

I understand this temptation, arising from the deficiencies of
delegation, but I don't think MI is the answer you're looking for.  MI
is actually a fairly crude tool for inheriting different pieces of
implementation from different classes:

(1) MI can't handle the situation where several instances want to
share a "parent" object to which substantial implementation should be
delegated.

For example, if you have a document object supports multiple "views"
of itself, you may want each document-view to respond to most of the
messages that the document object itself would respond to.  MI doesn't
address this because it assumes that the "parent" must always be
imbedded in the "child".

(2) In cases of conflicting inheritance, you lose all but the coarsest
control over what comes from where.  While fully explicit inheritance
rules can sometimes help, there will still be plenty of situations
where you'll get boxed in by MI.

For example, in an intertwined MI hierarchy of Controls and Cells,
GlorpButton might want to inherit from Button and GlorpCell.  Let's
suppose that GlorpCell acts much like ButtonCell, but is implemented
completely differently.  Now MI is hosed!  Button would already have
multi-inherited most Cell methods from ButtonCell.  Thus, even if MI
allowed us to choose a "preferred" parent class, it would provide no
way for GlorpButton to inherit just Cell functionality from GlorpCell.
We're back to explicit delegation.

(3) One of Obj-C's hallmarks is that each of its small number of
features is very poweful and flexible for the amount of complexity it
adds to the language.  For example, Protocols are a great addition
because, despite their utter simplicity, they directly support the
correct basis for typing in OOP---functionality, rather than
implementation.  MI, by contrast, is only a partial solution to the
problem it tries to address---reuse of implementation.


I'd really like to see a good solution to the implementation re-use
problem.  Explicit delegation is flexible enough, but as you know
suffers from several problems:

* It generates lots of boilerplate which becomes a maintenance
headache (when you're overriding to inherit from the "other
parent")...or maybe a C preprocessor headache.

* It's up to twice as slow as ordinary messaging due to the extra
message call overhead.  Penalties go even higher if you try to make
things more compact by using general runtime delegation mechanisms
like forward::.


Here's one possible solution (sadly, not compatible with the current
GNU Obj-C run-time).  It provides a concise, but flexible @super
notation to support inheritance of implementation from different parts
of the class hierarchy.  It is transparent to implementation code,
affecting only interface declarations.  In particular, overriding and
modification of inherited implementations looks just like it always
did.

The idea is that we can designate Protocols for which the meaning of
"super" in overridden methods---and the meaning of "self" in
non-overridden methods---is modified.  For example, here is how I
imagine the GlorpButton example above being implemented:

    @interface GlorpButton : Button
    {
        // Inherit <Cells> methods from GlorpCell class.
        // Make inherited implementations act on state in "cell".

        @super<Cells> GlorpCell *cell;
        ...
    }
    ...
    @end

Normally "super" invokes a fixed implementation of the method on a
parent object which lives parasitically inside the child.  But by
designating an associated object to be "@super" for some protocol, an
appropriate fixed implementation can be invoked on a different,
external "parent" object (the designated instance var) instead.

In the same way, if a method in the protocol was not overridden, it
would be dispatched by default to the object designated as @super for
that protocol, using the same implementation that would be obtained if
"self" were equal to that associated object.

So the intended effect of the above example is that:

* All <Cells> methods sent to a GlorpButton would, by default, really
be sent the GlorpButton's "cell" instance variable.  Just as with
ordinary dispatch, the implementation would be chosen at run-time,
depending on the actual sublclass of GlorpCell occupied by "cell".

* If GlorpButton overrode a <Cells> method, then any message to
"super" within such an implementation would invoke GlorpCell's
implementation of that message on "cell" (indepedent of the real class
of "cell").  For example:

    @implementation GlorpButton

    - (int)intValue
    {
        // Compute value using the *inherited* intValue.
        // (This comes from "cell", considered as a GlorpCell.)

        return ([super intValue] + 1);
    }
    ...
    @end

(To get the intValue of cell according to its runtime class, [cell
intValue] would work as always.)

Based on the above example we see that it would have to be permissible
to make stricter redeclarations of existing instance variables, at
least for the purpose of @super designation.  This shouldn't be a big
problem because it doesn't require any retroactive changes to
previously compiled code in superclasses.  Also, note that any @super
variables would have to be statically typed, since the purpose is to
inherit implementation from somewhere.

This proposal is not (portably) compatible with the current GNU
Objective-C run-time because the "self" argument to method
implementation calls is hard-compiled in.  If the current message-call
conventions are kept, it also seems that any implementation of @super
would have to be equivalent to explicit delegation or forwarding, and
would thereby incur the corresponding penalties in execution time.

--------------------------------------------------------------------
Paul Burchard    <[e-mail]>
``I'm still learning how to count backwards from infinity...''
--------------------------------------------------------------------



From: [e-mail] (Bill Burcham)
Date: Fri, 29 Jan 93 09:34:19 -0600
To: [e-mail]
Subject: Re: How to do better than Multiple Inheritance
Cc: [e-mail]

Cool ideas!  I like the idea of using protocols for conflict
resolution in MI.  What I don't see is the need for:

    @interface GlorpButton : Button
    {
        // Inherit <Cells> methods from GlorpCell class.
        // Make inherited implementations act on state in "cell".

        @super<Cells> GlorpCell *cell;
        ...
    }
    ...
    @end

the cell instance variable to be visible to an instance of GlorpButton
(or maybe visible, but the implementer of GlorpCell shouldn't have to
think up a name for that instance).  Also, are you insinuating that I
would have to alloc a GlorpCell myself in - init?  Would the runtime
do it for me?  Let's say that the runtime _will_ alloc objects and
init all @super outlets with instances of the proper class -- then how
is the better than the runtime alloc'ing a single object with the
superset of all the ivars?  In asking that last question I start to
think maybe there _are_ advantages: maybe it would be nice to have an
ivar in GlorpButton with a name the same as one in Button, but with a
different meaning (this would be nice, so I wouldn't have to worry
about ivar name conflicts in my new subclass: GlorpButton).  I can see
arguments for having seperate objects, but I don't see the need for
the cell ivar in the .h file.  Would this do:

    @interface GlorpButton : Button, <Cells> GlorpCell
    {
        ...
    }
    ...
    @end

You could still have the runtime actually create a GlorpButton(as a
subclass of Button -- just like it works now) and a GlorpCell.  The
GlorpButton would not usually need direct access to the instance of
GlorpCell.  When GlorpButton needs to get at the id of one of these
superclass objects, we could have an Object method to do so perhaps:
    - superclassInstance:(id)aClassObject;

So we could write something like:

    [self superclassInstance:GlorpCell];

Just some random thoughts.  I also am very intrigued with
prototype-instance languages like (I think) GemStone, and Self, where
there is no distinction between Class objects and "instance" objects.
Any object can be used as a prototype for other objects.  Methods and
ivars can be added/removed/changed at runtime, and objects that
inherit those things dynamically get the altered behavior.  I realize
this is beyond the scope of the GNU Objective-C project but it is
really cool to think about anyhow.

---
+--------------------------------+----------------------------------+
|          Bill Burcham          | "Make no small plans; they have  |
| First National Bank of Chicago | no magic to stir men's souls"    |
|    [e-mail]  (NeXTmail)  |      Daniel J. Burnham           |
+--------------------------------+----------------------------------+

Date: Fri, 29 Jan 93 12:19:34 -0600
From: Paul Burchard <[e-mail]>
To: [e-mail] (Bill Burcham)
Subject: Re: How to do better than Multiple Inheritance
Cc: [e-mail]
Reply-To: [e-mail]

> Cool ideas!  I like the idea of using protocols for conflict
> resolution in MI.  What I don't see is the need for:
> [...]
> the cell instance variable to be visible to an instance of
> GlorpButton

Yes, this is the difficult part of how to do notation for this scheme.
Because of the examples I gave, I think it's important to be able to
use instance variables to hold state for implementations culled from
other parts of the hierarchy, rather than always using imbedded parent
objects.

The problem is that the appropriate instance variables are often
already inside some other parent object, and therefore syntactically
inaccessible (for good reason, to be sure).  So there needs to be some
notational system which lets the compiler resolve this.

> Would this do:
>     @interface GlorpButton : Button, <Cells> GlorpCell
>     {
>         ...
>     }
>     ...
>     @end

This could work if the compiler can figure out where to put the state
for <Cells> methods.  If Control had already set its "cell" instance
var to be a parent object for <Cells> methods, for example, then
perhaps that could imply an interpretation of the above.  Unless such
resolution is possible, maybe use of parent vars for MI should not be
allowed?

I appreciate your thoughts...

--------------------------------------------------------------------
Paul Burchard    <[e-mail]>
``I'm still learning how to count backwards from infinity...''
--------------------------------------------------------------------