Path: utzoo!utgpu!news-server.csri.toronto.edu!rpi!zaphod.mps.ohio-state.edu!cis.ohio-state.edu!ucbvax!elmer-fudd.berkeley.edu!konstan
From: konstan@elmer-fudd.berkeley.edu (Joe Konstan)
Newsgroups: comp.lang.clos
Subject: Re: CLOS Private Methods
Message-ID: <42474@ucbvax.BERKELEY.EDU>
Date: 26 Jun 91 05:51:53 GMT
References: <LGM.91Jun25201133@cbnewsc.ATT.COM> <9106260502.AA11848@kuwait>
Sender: nobody@ucbvax.BERKELEY.EDU
Organization: Picasso Research Group, UCB Computer Science Division
Lines: 83

Re:  Encapsulation and fear of packages

I think that the lack of encapsulation problem is more than merely a fear of
packages.  With CLOS, there are many cases where a useful functionality is
unavailable unless you have (and can modify) the source code to a class and/or
the methods which specialize on that class.  For one example (from the paper
_Developing a GUIDE Using Object-Oriented Programming_ by me and Lawrence A.
Rowe to appear at this year's OOPSLA) take the classes HOUSE and BOAT.
Logically, as a user of multiple inheritance, I would like to create a
child class HOUSEBOAT that inherits from both.  Methods defined for one or the
other of the superclasses work fine as do methods completely overridden in the
subclass.  Let's take the example of the generic function CLEAN, though.  

(defmethod CLEAN ((self HOUSE))
  (wash-windows self)
  (scrub-floors self))

(defmethod CLEAN ((self BOAT))
  (scrape-barnacles self))		;; I show my ignorance about boats

The default behavior I want for HOUSEBOAT is to do both.  I can't do this by
not defining a method, since neither HOUSE not BOAT does a call-next-method
and therefore only the first superclass will have its method called.  I can't
even write a method such as:

(defmethod CLEAN ((self HOUSEBOAT))
  (CLEAN (HOUSE self))
  (CLEAN (BOAT self)))

In fact, I not only have to add call-next-method into the CLEAN method for at
least one superclass, but I may have to make other changes as well (perhaps
HOUSE is a subclass of DWELLING and DWELLING has a CLEAN method HOUSE doesn't
call.....)

This type of problem is at the heart of what I consider a predisposition
against encapsulation.  I realize that metaclass hacking can get around almost
any individual problem like this, and that clever coding can sometimes do so
as well, but for the average CLOS user, these problems constitute a failure in
encapsulation.  

As for the issue of packages, I think that the reason most Common Lisp
programmers fear or avoid them is related to the nasty interaction they have
with the reader and with compiled-versus-interpreted code.  One recent problem
I'd hoped to use packages to solve illustrates some of these problems.  I
wanted to have two packages, VIDEO and FAKE-VIDEO, which would export the same
symbols (functions such as PLAY and state variables such as PLAYER-STATE) and
of which only one package would be in use at a time (based on whether the
hardware could support video). 

I found this to be a difficult enough problem so as to make it not worth
pursuing far.  When interpreted, I could load one set of code or the other,
use that package, and then load the code that called the video successfully.
But, there was no way to compile the code that called the video or even load
it in advance since to compile or load that code it needed to resolve the
package of the symbols and would create them locally if they weren't already
imported.  

Similarly, it is frustrating that even an export call creates symbols in the
current package (so that I end up using 

	(defpackage foo
         (:exports "FOO-SYM-1"
                   "FOO-SYM-2"))

since putting symbols in there would create them in the current package).  

One solution would be to have some sort of external declaration such as

	(defpackage video-user
         (:external play player-status))

which would indicate that the symbols play and player-status should be
resolved at use-time rather than compile or load-time.

This is getting a bit away from CLOS, but the point I'm trying to make is that
CLOS really doesn't have viable encapsulation and that packages, while they
could be used for that, are awkward enough that most CL programmers don't use
them much.  So long as this is the case, application writers desiring
encapsulation will likely avoid CLOS.

Joe Konstan
konstan@cs.berkeley.edu