Dear Matz, thanks for your replies

In the discussion appended below (that I have now unearthed) the usefulness of a "message
consuming nil" is presented. I believe that Objective-C uses this approach. I personally
use ST/X by exept.de and I have yet to get this pattern working in this environment
either (due to the fact that the compiler inlines the ifTrue:ifFalse: behaviour and you
cant redefine it. (it works in the interpreter though)

It appears that the only efficient way to do this is to request that it be included as a
language feature. Obviously if it is genuinely useful then I would have thought it would
be snapped up.

The stuff I have implemented gives me 90% of what it needed but perhaps you might be
interested to put this in. I think it would give a lot of utility for relatively little
effort.

-----------------
 Documentation:

Copyright 2000, by Nevin Pratt

In Smalltalk, as we all know, nil is the distinguished object which is typically used as
the initial value of all
variables. It answers true to the message #isNil, and throws a 'Does Not Understand'
exception when almost
any other message is sent to it.

Do all dynamic languages have a similar concept, and a similar nil? No, they do not. In
this article, I am going to
briefly compare Smalltalk's nil behavior to that of another dynamic language-Objective-C.
Even though it has
been about six years since I have done any programming in Objective-C (because I switched
to Smalltalk), I
have found some Objective-C techniques to be helpful and useful to my Smalltalk career.
In particular, there are
certain situations where I actually prefer Objective-C's concept of a nil over
Smalltalk's. This article explores
some of those situations.

If you answer nil in Objective-C, the nil that is returned is, in some respects, a lot
like Smalltalk's nil, except
that instead of the nil generating exceptions when you message it, it just silently eats
the messages. Thus, in
Objective-C, nil acts more like a 'black hole'. It is emptiness. It is nothingness. If
you send anything to nil, nil
is all you get back. No exceptions. No return values. Nothing.

Obviously, if we wanted this same behavior in Smalltalk, we would simply override the
#doesNotUnderstand:
instance message of the UndefinedObject class to just answer self. But would making this
change be a good
idea in Smalltalk? No it wouldn't-but I'm getting ahead of myself in saying that. So,
let's look at it a bit closer,
because in doing so it will eventually lead us to what I really do like!

In Objective-C, nil didn't always have this 'message eating' behavior. Brad Cox, the
inventor of Objective-C,
originally gave nil behavior that more closely modeled Smalltalk's nil. That is,
messaging nil originally
generated a runtime exception, as is evidenced in his release (via StepStone Corporation)
of his 'ICPack 101'
class library and accompanying compiler and runtime. Then, beginning with 'ICPack 201'
(and accompanying
compiler and runtime), this behavior was changed to the current 'message eating'
behavior. And, NeXT
Computers followed suit with their Objective-C implementation, in which they gave their
nil a 'message
eating' behavior, as did the Free Software Foundation with their GNU Objective-C
compiler. But it wasn't
always that way.

So why did it change?

As you might guess, this change created two diverging camps among the programmers. On one
side sat the
programmers that preferred the original 'exception throwing' behavior, and on the other
side sat the
programmers who preferred the new 'message eating' behavior. And they each gave their
best arguments to
try and illustrate why the philosophy of their side was superior to the other. It was a
lively and interesting
debate that had no victors, other than the de-facto victor voiced by the compiler
implementers themselves;
namely NeXT Computers, StepStone Corp, and later the FSF, all of which chose the 'message
eating'
behavior. But, the opinions voiced in the 'pro' and 'con' arguments were interesting, and
especially interesting
in what they both actually agreed upon!

Both sides agreed that the 'message eating' behavior tended to create more elegant code!

But of course, the 'exception throwing' side responded by saying, 'seemingly more elegant
code, yes, but
potentially troublesome and unreliable', and they gave their reasons for asserting this.
We will look at some of
those arguments, but first I will demonstrate how the 'message eating' behavior can tend
to make the code
more elegant looking.

Suppose, for example, that we wanted to find out the last telephone number that you
dialed from your office
telephone, and we wanted to save this last number into a variable called, say,
`lastNumber'. Suppose further
that we wanted to save it as a string so that we could display it in a GUI widget (as
well as so we could use it
later). If you are the `person' in the message sequence below, would the message sequence
to accomplish this
request be as follows?

lastNumber := person office phone lastNumberDialed asString.
widget setStringValue: lastNumber.

Maybe.

But then, what if you don't have an office? Or, what if you have an office, but the
office doesn't have a phone?
Or, what if it is new phone, and the phone has never been dialed yet? In any of these
cases, using the
'exception throwing' nil convention, an exception will be thrown, thus potentially
halting the program if an
exception handler hasn't been created to handle that exception.

But, what if nil has the 'message eating' behavior? In this case, `lastNumber' could
potentially have a final
value of nil, but the code above works just fine for this. Even passing nil as an
argument to the
#setStringValue:1 method doesn't hurt, because the 'message eating' nil convention is
used by the widgetry as
well. It doesn't matter that the argument is nil. Everything still works, and there's no
exception thrown, and no
immediately apparent strange side-effects (but we'll analyze that one some more later).

To contrast this, how then would you have to code it if nil has the 'exception throwing'
behavior? You would do
it similar to this:

  | tmp |
   tmp := person office.
   tmp notNil ifTrue: [tmp := tmp phone].
   tmp notNil ifTrue: [tmp := tmp lastNumberDialed].
   tmp notNil ifTrue: [lastNumber := tmp asString].
   widget setStringValue: lastNumber.

Yuck...all those explicit tests for nil are really ugly! Of course, you could have
instead wrapped the original code
in an exception handler, and thus avoided the nil tests, something like as follows:

   [lastNumber := person office phone lastNumberDialed asString.
   widget setStringValue: lastNumber]
      on: Object messageNotUnderstoodSignal do: [].

This looks a bit simpler than the previous example, but even this example contrasts
poorly to the first example.
The first example of these three is much simpler! You just 'do it', without worrying
about exceptions, exception
handlers, or explicit tests.

But, is this kind of code common? With the 'exception throwing' nil, do we really end up
typically testing for nil
like this, or else setting up exception handlers like this?

Yes, it is common. While the above example was contrived, let's look at a real-life
example, from the
#objectWantingControl method of the VisualPart class of VisualWorks. The VisualPart class
is a superclass of
the View class, and views in VisualWorks are coded to generally expect to have a
collaborating controller
object for processing user input (mouse and keyboard events). Thus, the
#objectWantingControl method is a
method of the view object, and it asks it's controller if it wants the user focus. If it
does, #objectWantingControl
answers self, otherwise it answers nil. If nil is answered, then the view is considered
to be read only, and will
not process user input. The actual implementation of #objectWantingControl is as follows:

objectWantingControl


   | ctrl |
   ctrl := self getController.
   ctrl isNil ifTrue: [^nil].
   ' Trap errors occurring while searching for
   the object wanting control. '
   ^Object errorSignal
   handle: [:ex |
   Controller badControllerSignal
   raiseErrorString:
   'Bad controller in objectWantingControl']
   do: [ctrl isControlWanted ifTrue: [self] ifFalse: [nil]]

Notice that this method has both an explicit #isNil test, as well as an exception
handler. How would this
method instead be written if the system had a 'message eating' nil throughout? While
there are a several
variations of possibilities, including at least one variation that is shorter (but not
necessarily clearer), we would
probably write it as follows:

objectWantingControl
   self getController isControlWanted ifTrue: [^self].
   ^nil

Notice how much simpler it suddenly became. The programmer's intentions are much clearer.
No #isNil checks,
no exceptions, no extra code to confuse the issue.

Furthermore, with the 'exception throwing' nil, even when the code is written to avoid
#isNil tests and/or
exception handlers, the coding style is usually altered in other ways to compensate. And,
invariably, these style
alterations don't produce as elegant of code as if you had a 'message eating' nil.

'Message eating' nil creates simpler, more elegant code. This was almost the unanimous
opinion of both camps
of the Objective-C debate on this. But of course, the 'exception throwing' camp argued
that this 'simpler' code
was also potentially more troublesome, and sometimes even buggy. And, they gave examples
to illustrate. But
their examples also all seemed to fall into one of two arguments.

The first argument boiled down to the observation that, with a 'message eating' nil, if a
message sequence
produces nil as the final result, it is more difficult to determine exactly where the
breakdown occurred. In other
words, what was the message that produced the first nil?

And, the response to this argument was: the programmer typically doesn't care what
message produced the
first nil, and even if he did, he would explicitly test for it.

And of course, the response to this response was: the programmer should care, but
typically won't care,
therefore the 'message eating' nil is a feature which promotes bad programming habits.

And of course, this in turn illicited a response that essentially just disagreed with
their conclusions and
challenged their statements, such as why the programmer should care, etc.

And so the debate raged. But none-the-less, both sides seemed to admit that the 'message
eating' nil tended
to create simpler, more elegant looking code. And, the existing code base tended to
substantiate this
conclusion. And simple code is good code, as long as it is also accurate code.

So, with a 'message eating' nil, is the resulting code accurate? Or does the 'message
eating' nil tend to
introduce subtle bugs? To this question, the 'exception throwing' crowd said it
introduces subtle bugs, and they
gave specific examples. Interestingly enough, all of their examples that I looked at were
with statically declared
variables, and those examples typically illustrated the platform dependent idiosyncracies
that developed when a
nil was coerced into a static type. One specific example was illustrated via the
following code snippet (which I
have modified to conform to Smalltalk syntax instead of Objective-C syntax):

   value := widget floatValue

In this example, if `value' is statically declared to be a variable of type float (floats
are not objects in
Objective-C, but are instead native data types), and the #floatValue2 message returns nil
instead of a valid
float number, then after the assignment has completed, `value' will equal zero on the
Motorola M68K family of
processors, but on the Intel processor family, it ends up being a non-zero value, because
of the peculiarities of
implicit casting of a nil to a native float datatype. This is clearly an undesirable
result, and can lead to subtle
bugs.

But, while those examples might be relevant for Objective-C, they are totally irrelevant
for Smalltalk. There is
no static declaration of variable types in Smalltalk, nor are there native data types
(non-objects). It's a
non-issue in Smalltalk.

So, should the semantics of nil in Smalltalk be changed such that it eats messages? This
is easy to do by
changing #doesNotUnderstand: to just answer self. Should we do it?

No, I don't think so. There has been too much code already written that is now expecting
nil to throw
exceptions. To change the semantics of nil from 'exception throwing' to 'message eating'
at this time would
likely break a large body of that code. It could be a very painful change, indeed.

Furthermore, even in Smalltalk, the first objection to a 'message eating' nil still
stands; to whit, in a given
message sequence whose final value is nil, it is difficult to determine what object first
returned the nil. While it
is purely a subjective opinion as to how important that objection really is, I don't know
how anyone could not
agree that it is indeed a valid objection. Minor perhaps (and perhaps not), but valid.

So, instead of modifying Smalltalk's nil, let's now briefly look at an alternative, that
of sending back a
specialized Null object that has message-eating semantics. The first public document that
I am aware of that
explored this alternative is Bobby Woolf's excellent white paper, 'The Null Object
Pattern'3, although earlier
works likely do exist. In that paper (which is now about five years old-almost an
eternity in computer time), he
also uses the VisualPart example from above to illustrate his pattern. In fact, that is
precisely why I also chose
to illustrate the results of using a 'message eating' null via this same VisualPart
example. That way, I could
keep things simple and consistent, without introducing too much additional code for all
of the illustrations.

The 'Null Object Pattern' essentially recommends the creation of a 'do nothing' null
object which implements
the same protocol as the original object, but does nothing in response to that protocol.
For the VisualPart
example above, this pattern requires the creation of a class called NoController which
implements the protocol
of a controller, but does nothing in response to it.

Doing nothing, however, means something special to a controller. For example, the
NoController is expected to
answer false to the #isControlWanted message. Why is this important? Because clients of
NoController expect
a boolean result to the #isControlWanted message, and they might in turn try sending
#ifTrue: or #ifFalse: to
that result, and only booleans (and perhaps 'message eating' nils) respond to #ifTrue:
and #ifFalse. The
NoController has to return something that will respond to these boolean messages, or else
the NoController is
not going to be plug-compatible with a real controller.

But, suppose we instead had #isControlWanted return a 'message eating' nil? Or better
yet, what if the
#getController method of the VisualPart returned a 'message eating' nil? I believe
everything would still 'just
work', and that this also would be a simple way to generalize the 'Null Object Pattern'
of Woolf's paper.

Interestingly enough, in Woolf's paper, he describes an advantage of the 'Null Object
Pattern' by saying it...

...simplifies client code. Clients can treat real collaborators and null collaborators
uniformly.
Clients normally don't know (and shouldn't care) whether they're dealing with a real or a
null
collaborator. This simplifies client code, because it avoids having to write special
testing code
to handle the null collaborator.

This testimony dovetails nicely with the NeXTSTEP community's assertion that the 'message
eating' nil
behavior of Objective-C appears to simplify code, as I have already demonstrated.

But, to implement the 'Null Object Pattern', do we create a NoController class, and a
NoOffice, and a NoPhone,
and a NoLastNumberDialed class? Where does it end? Indeed, this potential class explosion
of the 'Null Object
Pattern' is also mentioned by Woolf, as follows:

[One of] the disadvantages of the Null Object pattern [is]...class explosion. The pattern
can
necessitate creating a new NullObject class for every new AbstractObject class.

A 'message eating' nil would avoid this class explosion, as it is protocol-general
instead of protocol-specific. I
personally feel that this difference is even a bit reminiscent of the static typing vs.
dynamic typing differences
(and ensuing debates), as the following chart illustrates:

Static Typing vs. Dynamic Typing

Should we allow any type of object to be
handled  (assigned to) this variable,
or only objects of a specific type?

Null Objects vs. Message Eating Nil
Should we allow any type of message to
be handled by (sent to) this object, or
only messages of a specific type
(protocol)?

I make no secret that I prefer dynamic typing over static typing. And, I also believe
that often a general
'message eating' nil is more desirable than the more specific 'Null Object Pattern',
provided of course that the
'message eating' nil is implemented correctly. What follows is my implementation of a
'message eating' nil,
which I call a null, which is an instance of my class Null.

Recall that the first objection against the null was that in a given message sequence
whose final value is null, it
is difficult to determine what object first returned the null. How do we handle that
objection?

Simple. Just ask it.

The Null class should have an originator instance variable that records who originally
invoked the `Null new',
as well as a sentFromMethod instance variable that records from what method of the
originator the `Null new'
was invoked.

But does that mean that the creator of the null must now tell the null so that the null
can tell you? That sounds
like a lot of work! And, what if someone forgets those extra steps?

Simple. Don't require the extra steps. Anybody should be able to send `Null new', and the
Null class itself
should be able to figure this information out.

But that is not possible to do unless the Null class can somehow automatically determine
who is calling one of
it's instance creation methods. In other words, we need to detect who the sender of the
message is. How do we
do that? It is not part of standard Smalltalk!

Well, here is how to do it in VisualWorks:

   Object>>sender
      ^thisContext sender sender receiver

   Object>>sentFromMethod
      ^thisContext sender sender selector


Now, in your other code, anytime you want to return a nil that also has message-eating
semantics (which I call
a null), you use `^Null new' from your code instead of `^nil'. Then, your caller can
easily discover the originator
of the null if it wishes to simply by asking.

If you are concerned about the potential proliferation of nulls with such a scheme,
another trick you might try is
to create a default null using a `Default' class variable:

   Null class>>default
   Default == nil
   ifTrue:
   [Default := super new initialize.
      Default originator: Null.
      Default fromMethod: #default].
 ^Default

The default null can then later be accessed via `Null default' instead of `Null new'. I
actually use this quite often
for automatic instance variable initialization in my abstract DomainModel class, which is
the superclass of all of
my domain objects:

DomainObject>>initialize
   1 to: self class instSize
      do: [:ea |
         (self instVarAt: ea) isNil
            ifTrue:
               [self instVarAt: ea put: Null default]].
   ^self

I have found that such a scheme does indeed simplify the domain logic, just as this
article indicates that it
should. In fact, sometimes it has dramatically simplified things. And, I have never had
any problems with this
scheme, as long as I have limited its use to the domain layer only. I have, however, had
problems trying to
integrate some of these ideas into the GUI layer, and decided long ago that it was a bad
idea in that layer.

My own implementation of the Null class was originally written in VisualAge, and was
originally part of a much
larger domain-specific class library. This class library originally tried a number of
ideas on an experimental
basis, to see if problems resulted from their use. The use of the Null pattern described
in this article was one of
those experimental ideas. Even though it is actually a small idea, it's widespread use in
the domain layer was
encouraged from my previous Objective-C experience, but I still didn't know if I would
run into other subtle
issues while using it in Smalltalk. But I feel comfortable with it now in the domain
layer (but not in the GUI
layer).

Some time after creating the class library I mentioned above, the entire class library
was ported to GemStone,
and then finally the entire class library was moved to VisualWorks. A filein of the
VisualWorks implementation
of the Null class follows. Email me at nevinop / xmission.com if you want either the
VisualAge or GemStone
versions, and I'll try to dig them out.

If you instead decide to create your own Null class in VisualAge, another thing to
realize is that #isNil is inlined
in VA (but not in VW). Thus, something like `Null new isNil' will always answer false in
VA, even though your
Null>>isNil method explicitly answers true. Hence, in your domain code, with VA you
probably want to create
an #isNull method and use that instead of #isNil. That is what I originally did, and that
convention carried to the
GemStone version, but I have since broken that convention in the VisualWorks version.

1.As a sidebar, one could also argue about the appropriateness of a #setStringValue:
method in this
example, and its implied limitation of only setting, or showing, strings, rather than
having perhaps a
more generic #show: method that can show other types as well. To this, I have three
things to say:

first, consider the commonly used #show: method of the Transcript class in Smalltalk, and
the argument
type it expects (strings)
second, #setStringValue: is the actual method name used for TextField widgets in NeXTSTEP

third, who cares, this is just a contrived example anyway.

2.#floatValue also is an actual message implemented by TextFields under NeXTSTEP, just as
the
#setStringValue: of the earlier code snippets is.
3.Published in Pattern Languages of Program Design. Addison-Wesley, James Coplien and
Douglas
Schmidt (editors). Reading, MA, 1995; http://www.awl.com/cseng/titles/0-201-60734-4.