Are you aware of the work I did that enables class fault detection in Squeak?

If a class has a nil methodDictionary, it causes a fault (ie it sends recoverFromMDFault to the class). � That code then restores the MD in addition to any anyalysis you want. I never cleaned it up for posterity, but it enabled lots of interesting work...

� � � �By nilling all but a few MDs, and then running some application, you could
� � � � � � � �tell exactly what classes were needed and which were not.
� � � � � � � �See the method discoverActiveClasses.
� � � �A special feature even recorded the call stack that caused each class to be
� � � � � � � �touched.
� � � �What I was doing was restructuring the entire system as a bunch of
� � � � � � � �image segments that could be swapped out and brought in
� � � � � � � �on demand -- a sort of user-programmable swapping memory.
� � � �This all worked at one point. �Swapping of projects was a bit flakier
� � � � � � � �because of our weak discipline on pointer scope, but it
� � � � � � � �even worked in many cases.

Where I was going with the whole "reorganizeEverything" project was that the SystemDictionary would be partitioned into many separate dictionaries in each project, and that namespaces would be the same as the project swapping unit, and the whole thing would be simple and fast. �It actually worked once (;-) also.

Loom was cool, but it could never come close to the speed of imageSegments. �I played around with swapping out the entire VM Support category with all classes and methods, and it would swap in the entirety in less than 200 milliseconds, and this was a decade ago.

Sorry for the long digression, but maybe some of this thinking or even the actual mechanism can be of use to your experiments.


Hi Dan. First let me say that it is really cool to have you in the list :)

Yes, I know about that MD faults and I did some experiments with it. As part of my work, I remove that from core and I put it in a separate package. I also rewrite some things a little. I invite you, if you want/have time to look at it. It is in

http://www.squeaksource.com/ClaseUseDiscovery

You can see the class comments of CUDClassesUseDiscoverer and also the CUDClassesUseDiscovererTest
This tool is working perfect. I sent an email some weeks ago:
http://n4.nabble.com/ANN-ClaseUseDiscovery-tool-td1296859.html#a1296861

Now, my conclusions were that it is a great tool but it has the problem that when ever you do a Smalltalk allClassesDo or similar, it brings all the classes back. I saw there was the *gentlly*� so that you could do allClassesDoGently�� etc� and the gently checks for isInMemory, where in ImageSegmentRootStub answers false. The problem is that you have to put those gently everywhere and once you forget about that....you are dead.

The other problem is the gradunalirity. I mean, even when sending a simple message to a single object of a class, the whole class is loaded.

Even with all that, it was great to see and have in mind a percentage of the used classes against different scenarios.

Cheers

Mariano
�
� � � �- Dan

>
>> �James Ladd <james_ladd@hotmail.com> wrote...
>>>> Hi Pharo-ites,
>>>>
>>>> I'm working on a port of Pharo to the Java Virtual Machine called Redline Smalltalk.
>>>> (read more here: http://jamesladdcode.com/?p=323)
>>>>
>>>> I'm wanting to port every single class over time, but initially enough to get the
>>>> compiler working and have a minimum runtime.
>>>>
>>>> Of the 1600+/- classes which are those necessary to make a small workable compiler
>>>> and runtime environment?
>>
>> Hi, James -
>>
>> Bert Freudenberg forwarded this along to me, as I'm not on a Pharo mailing list.
>>
>> I think you know of my JSqueak project (a.k.a. Potato) -- a Squeak VM written in Java. �The image that I used with it, is the same that I always use for this kind of experiment -- mini.image built in Squeak 2.2.
>>
>> The reason it is so cool is that it includes browser, editor, compiler, inspector debugger, and files -- everything you need for self-support. �Plus it has the decompiler, and special temp-name hack, so you can actually browse sources with temp names before any file system works. �My favorite thing about this image is that although it is only 600k, if you decompile all its sources, it comprises over 850k of Smalltalk, so it's like a compressed version of the sources with a full-function IDE thrown in for free ;-).
>>
>> As I said, the image is about 600k, of which 240k is code -- 202 classes (x2 if you count the metas) with 4590 methods. �This should be a good guide for you in choosing classes to start with.
>>
>> You can actually run this image live in JSqueak on the web if you have Java installed and the planets are aligned correctly...
>> � � � � http://Weather-dimensions.com/Dan/JSqueak.jnlp
>> Another interesting statistic is that the Java .jar file is only 433k, including not only the VM, but also the entire image! �This is possible because the image is so small that the pointers in it are mostly zeroes, so it compresses nicely. �I just checked that this runs, and was happy to see that it putts along at 21 million bytecodes/sec on my laptop.
> >
>> I haven't looked recently, but I think you'll find Mini2.2.image or something like it in the same place where all the historical releases of Squeak are kept. �Let me know if you can't find it.
>>
>> Have fun with your project! �I hope this helps
>>
>> � � � � - Dan
>> _______________________________________________
>> Pharo-project mailing list
>> Pharo-project@lists.gforge.inria.fr
>> http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/pharo-project
>
>
>_______________________________________________
>Pharo-project mailing list
>Pharo-project@lists.gforge.inria.fr
>http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/pharo-project


_______________________________________________
Pharo-project mailing list
Pharo-project@lists.gforge.inria.fr
http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/pharo-project