On Sun, 24 Oct 2010, Eliot Miranda wrote:
Hi Levente,
in my experience playing with the Dan and Alan Borning's multiple
inheritance implementation getting things to work was one thing, reasonably
straighforward, but getting tool support right (file-out format so that
things filed back in correctly, etc) was another, an dth ebulk of the work.
Looking at your proposal below I see no problem getting it to work but see
lots of problems integrating it, e.g. with Monticello, etc. Traits is
You're right, integration is hard. It would probably be harder than the
integration of Traits was.
already integrated with Monticello, supported with tests, etc. As such it
is a much less risky or effortful proposition than a new scheme. So with
that in mind are there important functional benefits of your scheme's direct
state access as compared with traits that use accessors to access state that
each user of a trait must implement? i.e. what are the key benefits you see
in direct inst var access?
I see four benefits:
- a bit better performance on current VMs
- no name conflict (accessors)
- accessor methods can also be reused
- the ability to reuse the code of existing classes (this may cause
problems)
Let's see the same example as a Trait:
Trait named: #TBidirectionalLink
uses: #()
category: 'ClassCompositionExample'
TBidirectionalLink >> unlink
self previous next: self next.
self next previous: self previous
TBidirectionalLink >> linkAfter: aBidirectionalLink
self previous: aBidirectionalLink.
self next: aBidirectionalLink next.
aBidirectionalLink next: self.
self next previous: self
And the user:
ECSelectorEntry subclass: #ODatedEntry
uses: TBidirectionalLink
instanceVariableNames: 'date next previous'
classVariableNames: ''
poolDictionaries: ''
category: 'Ocompletion'
ODatedEntry (and all other users of TBidirectionalLink) has to implement
four methods (the accessors):
ODatedEntry >> next
^next
ODatedEntry >> next: anODatedEntry
next := anODatedEntry
ODatedEntry >> previous
^previous
ODatedEntry >> previous: anODatedEntry
previous := anODatedEntry
This is only possible if ODatedEntry (or a superclass of it) doesn't
implement any of these methods, otherwise the Trait can't be used.
Levente
best
Eliot
2010/10/24 Levente Uzonyi <[email protected]>
On Sun, 24 Oct 2010, Stéphane Ducasse wrote:
2) well, this is difficult to get the money for the butter and the
butter - we are trying.
If you can compose classes the way you can add a trait to a class now
with class and optional method level instance+class variable mapping, then
you're done. It would be a lot simpler to use it _and_ it would also be a
lot easier to implement it. Especially the tools part.
I'm interested to hear more about that.
so can you explain what you meant because I did not understand it.
Okay, here is an example of my class composition idea. It works like
Traits, but it supports state and there's no distinction between a trait and
a class. Let's say I have a class named BidirectionalLink which can be used
as a link in a linked list (with a head element). This will be used as a
stateful trait. It's definition is like this:
Object subclass: #BidirectionalLink
instanceVariableNames: 'next previous'
classVariableNames: ''
poolDictionaries: ''
category: 'ClassCompositionExample'
It has a few methods:
BidirectionalLink >> next
^next
BidirectionalLink >> next: aBidirectionalLink
next := aBidirectionalLink
BidirectionalLink >> previous
^previous
BidirectionalLink >> previous: aBidirectionalLink
previous := aBidirectionalLink
BidirectionalLink >> unlink
previous next: next.
next previous: previous
BidirectionalLink >> linkAfter: aBidirectionalLink
previous := aBidirectionalLink.
next := aBidirectionalLink next.
aBidirectionalLink next: self.
next previous: self
I have an existing class, let's call it ODatedEntry :). It has the
following definition:
ECSelectorEntry subclass: #ODatedEntry
instanceVariableNames: 'date'
classVariableNames: ''
poolDictionaries: ''
category: 'Ocompletion'
As you can see, it inherits some state and behavior from ECSelectorEntry.
My goal is to use the instances of this class in a linked list. So these
objects should implement the same protocol with the same behavior as
BidirectionalLink. I can't make it a subclass of BidirectionalLink, because
I also need the behavior and state from ECSelectorEntry and we don't have
multiple inheritance. So I'll compose the two classes. First I add the
necessary instance variables to the class. Let's call them nextEntry and
previousEntry:
ECSelectorEntry subclass: #ODatedEntry
instanceVariableNames: 'date nextEntry previousEntry'
classVariableNames: ''
poolDictionaries: ''
category: 'Ocompletion'
Then do the composition:
ECSelectorEntry subclass: #ODatedEntry
uses: BidirectionalLink
instanceVariableNames: 'date nextEntry previousEntry'
classVariableNames: ''
poolDictionaries: ''
category: 'Ocompletion'
Now this doesn't work, because ODatedEntry doesn't have instance variables
named next and previous, but the methods of BidirectionalLink would like to
use them. Of course I could have used those names in the previous step and
(with a working implementation) this would just work out of the box. But the
example is about the instance variable mapping. Let's say ~ is the
composition operator which defines variable mapping for a class or a method.
Then I can write the following:
ECSelectorEntry subclass: #ODatedEntry
uses: BidirectionalLink ~ { #nextEntry -> #next. #previousEntry ->
#previous }
instanceVariableNames: 'date nextEntry previousEntry'
classVariableNames: ''
poolDictionaries: ''
category: 'Ocompletion'
This means: take all methods from BidirectionalLink, but replace the
variable named next with nextEntry and previous with previousEntry.
So now ODatedEntry understands #next, #next:, #previous, #unlink, etc.
This is the basic concept. There are some open questions, like:
- what happens when a composed method (or a method sent by a composed
method, etc) has a super send?
- will the class also get the methods of the superclasses of the "trait"?
- how does it work on the class side?
- what about class variables?
- what if I don't want to use all methods, just a few?
Tell us more. The problem we faced was
- offset access = you cannot reuse bytecode of a trait because
the order of the offset can be different in each trait users
If you mean that a CompiledMethod of a trait cannot be added to the
class' method dictionary, than that's not an issue. The current Trait
implementation was changed, because shared CompiledMethods caused other
problems.
no this is not what I meant
If you mean that the same bytecodes can't be used, than that's neither a
problem, because you can and should be compile the method again. Sharing
trailer bytes may cause problems.
So adding a method from a trait to a class is simply recompiling it in
the class' context.
this is what we wanted to avoid.
Also because you may have to recompile all the other methods of the class
hierarchy because if a trait add an instance
variable then you should recompile the subclasses when a trait get added
with a state in the superclass.
My idea is that traits don't add instance variables. The user of the trait
maps the trait's variables to their own by name. So if a trait gets
a new variable, then only the trait and subclasses have to be recompiled.
The recompilation is postponed until a method of the trait which uses the
new instance variable is added to a class.
Instance variables should be used by name during compilation. If there's
a name collision then use the instance variable map I mentioned above.
what is that the instance variable map?
take the time to write an example
See above.
Levente
- initialization of instances variables at the trait level and the
composition at the class levele
You can always rename a trait's method in your class. So if the trait has
an #initialize method, then simply rename it to #initializeFooBar and send
it from the class' #initialize method.
Yes this is what the javascript implementation does but this is not that
nice but may be there is no better solution.
So indeed we could think about adding state.
Levente
3) again if nobody does anything and we just all cry on ourselves then
nothing will happen.
Tools are a must. No tools - no users.
Exact.
So for now identifying traits and learning is the way. Then we can
refactor, redesign
Well, Traits are in Squeak since 2006, IIRC they were available a few
years earlier. So in the last X (at least 4) years the only good candidate
to become a Trait was Magnitude.
Come on.
I will not answer to such statement because I'm positive thinking.
Stef
_______________________________________________
Pharo-project mailing list
[email protected]
http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/pharo-project
_______________________________________________
Pharo-project mailing list
[email protected]
http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/pharo-project
_______________________________________________
Pharo-project mailing list
[email protected]
http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/pharo-project
_______________________________________________
Pharo-project mailing list
[email protected]
http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/pharo-project
_______________________________________________
Pharo-project mailing list
[email protected]
http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/pharo-project
