[squeak-dev] unloadReloadablePackages

Frank Shearar frank.shearar at gmail.com
Mon Dec 30 13:55:42 UTC 2013


On 30 December 2013 13:46, Tobias Pape <Das.Linux at gmx.de> wrote:
>
> On 30.12.2013, at 14:41, Frank Shearar <frank.shearar at gmail.com> wrote:
>
>> On 30 December 2013 13:21, Tobias Pape <Das.Linux at gmx.de> wrote:
>>>
>>> On 30.12.2013, at 14:18, Frank Shearar <frank.shearar at gmail.com> wrote:
>>>
>>>> On 30 December 2013 12:47, Tobias Pape <Das.Linux at gmx.de> wrote:
>>>>>
>>>>> On 30.12.2013, at 12:05, Frank Shearar <frank.shearar at gmail.com> wrote:
>>>>>
>>>>>> On 29 December 2013 23:35, David T. Lewis <lewis at mail.msen.com> wrote:
>>>>>>> On Sun, Dec 29, 2013 at 08:02:51PM +0000, Frank Shearar wrote:
>>>>>>>> On 29 December 2013 19:34, Tobias Pape <Das.Linux at gmx.de> wrote:
>>>>>>>>> Dear Squeakers
>>>>>>>>>
>>>>>>>>> I give up.
>>>>>>>>> For roughly 6 hours I try to shrink my image using
>>>>>>>>>      Smalltalk unloadReloadablePackages
>>>>>>>>>
>>>>>>>>> It simply does not work currently.
>>>>>>>>> I have the said trunk image (Squeak4.5-13148#712) (NameVersion-Update#CIJob)
>>>>>>>>> but with the obsoletes removed as I explained a few emails ago.
>>>>>>>>>
>>>>>>>>> But to no avail.
>>>>>>>>> * Sometimes (!) ReleaseBuilder retains some obsoletes.
>>>>>>>>> (removing the Obsoletes some time later with fixObsoleteReferences works, but
>>>>>>>>> mostly not during unload)
>>>>>>>>> * same for VersionNumber-bla
>>>>>>>>> * SMLoader always retains obsoletes
>>>>>>>>> * Services-Base itches itself:
>>>>>>>>> When its ServiceRegistry's #isInteractive was unloaded,
>>>>>>>>> ServiceRegistry gets called again and calls #isInteractive
>>>>>>>>> on its current, resulting in an DNU.
>>>>>>>>> Issuing
>>>>>>>>>      Smalltalk at: #SystemChangeNotifier ifPresent: [:scn | scn
>>>>>>>>>              uniqueInstance noMoreNotificationsFor: ServiceRegistry].
>>>>>>>>> manually works, but not as a #preambleOfRemoval.
>>>>>>>>
>>>>>>>> So it sounds like some packages' unload/reloads aren't being tested,
>>>>>>>> which is why they've now broken. I realise you've now given up :), but
>>>>>>>> did your explorations lead you through any #unload implementations?
>>>>>>>
>>>>>>> No it is not a matter of testing. It never worked in the first place, so
>>>>>>> it is an imcomplete implementation, not a bug.
>>>>>>
>>>>>> I don't think it's far off from being right though. I think Tobias has
>>>>>> simply found a bug: when an MCPackage unloads, it should run the
>>>>>> class-side #unload methods for any classes it contains, just like
>>>>>> loading runs the #initialize methods after loading the definitions.
>>>>>
>>>>> It does, when unloading the class definition
>>>>> (which subsequently calls Smalltalk>>#removeFromSystem: (or so)
>>>>> which in turn calls #unload on the class before removing) but too
>>>>> late, IMHO.
>>>>> The problem is, simply executing it at the beginning does not help
>>>>> because #unload would be called twice then…
>>>>
>>>> Yes, but it's useless to call it then, because the things that it
>>>> calls are unloaded already. Unless I completely misunderstood what
>>>> you'd said?
>>>>
>>>
>>> Yep, thats why unload couldn’t be called, because its gone already :(
>>>
>>>> Ah, and the calling-twice problem isn't strictly Monticello's fault,
>>>> is it? #unload gets called by Class >> #removeFromSystem:. And
>>>> #removeFromSystemUnlogged doesn't help, because the unload still gets
>>>> called.
>>>
>>> Exactly.
>>>
>>>>
>>>> I suppose if Monticello was smart enough to figure out that every
>>>> method definition for class Foo was being unloaded, and then Foo too,
>>>> so it need just #removeFromSystem: the class, then we wouldn't have
>>>> this problem.
>>>
>>>> Something like, in MCPackageLoader >> #basicLoad, in "Pass #4: Remove
>>>> the obsolete methods"
>>>>
>>>> removals do: [:ea |
>>>>   ea isClassDefinition ifTrue: [ea unload]
>>>>  "Check the remaining removals to see if they still need to be done:
>>>> no point removing a method from a class that's just been removed"
>>>>   (ea isMethodDefinition and: [ea actualClass notNil]) ifTrue: [ea unload]].
>>>>
>>>> Something like that, at any rate. The idea being that, like with
>>>> loading, we need to do something with the classes before we do things
>>>> with the methods.
>>>
>>> Problem here: are all system notifications sent for each method to be
>>> removed?
>>
>> I figured there'd be loads of things I didn't think of, which is why I
>> hedged with "something like that" :) You're right. The above won't
>> send the notifications for these removed methods, because Environment
>>>> #forgetClass:logged: only notifies of the class removal. But then,
>> this is what happens when you remove a class normally. If we want to
>> start notifying of method removal on class removal, we should change
>> Behavior >> #obsolete to do the notifying.
>
> Bad idea.
> Obsoletion also happens on other occasions, eg, giving a class a new
> instVar. The old class w/o the inst var is obsoloeted and all methods
> are “moved” to the new.

I don't think so. ClassBuilder builds a new class with the new shape,
and copies over the old class' method dictionary. (See ClassBuilder >>
#newSubclassOf:type:instanceVariables:from:. I don't see the old class
being asked to #obsolete itself there.

>   Notifying about method removal would be misleading if not plain wrong.

Yes, if changing a class's shape does indeed cause the old version to
become obsolete.

At any rate, I'm wondering whether removing a class _normally_ causes
notification of the removal of all its methods. If not (do we have
tests for this?), then #basicLoad above doesn't need to do anything
special.

frank

> Best
>         -Tobias
>
>
>


More information about the Squeak-dev mailing list