[squeak-dev] [Vm-dev] Re: [Pharo-dev] Nuking VM ImageSegment support (was Re: Daily Commit Log; System-bf.916)
eliot.miranda at gmail.com
Wed Jul 5 16:07:37 UTC 2017
On Tue, Jul 4, 2017 at 5:01 AM, Bert Freudenberg <bert at freudenbergs.de>
> On Mon, Jul 3, 2017 at 11:43 PM, Eliot Miranda <eliot.miranda at gmail.com>
>> Hi Bert, Hi Tim,
>> On Thu, Sep 15, 2016 at 2:55 PM, <commits at source.squeak.org> wrote:
>>>> Name: System-bf.916
>>>> Ancestors: System-bf.915
>>>> Replace VM-level ImageSegment loading with a Smalltalk implementation
>>>> for old (interpreter-era) projects.
>>>> Also removes support for writing segments.
>>>> This overrides the Spur support introduced in System-eem.758.
>> I am currently putting back the ImageSegment support for saving projects
>> and loading native segments produced in this way.
> Why? I thought we agreed that ImageSegments are not a great idea as a
> file format, for their lack of compatibility. Serialization speed isn't an
> issue anymore thanks to your fast jit.
Quite, but there's legacy...
1. Tert is being ported to Squeak 6.x. Terf (nee Teleplace, nee Qwaq
Forums, nee Croquet) makes heavy use of image segments; they're used to
transfer room contents from server to entering client.
2. Max Leske is using segments for some application in Pharo and has been
waiting patiently for a working version in Spur
3. I myself wanted to be able to transfer projects from working image to
fresh VMMaker image build
4. the code is mature and works, so if it ain't broke why break it ? :-)
What ImageSegments *are* good at is for splitting up an image into segments
> that do not have to be in memory at the same time - but in that case they
> would not be an independent entity but be considered part of one particular
> image. And for that we need only very few methods (that I removed too,
> see storeSegment).
Well, we should indeed clean up, but I am for keeping the ability to save
projects and for applications like Terf to continue.
I note that image segments can only be loaded into a VM using an identical
heap representation, hence, even though 64 & 32 bit Spur have identical
object header formats, loading a 32-bit segment into 64-bit Spur or vice
verse needs to be done by parsing, adding to the suite of parsers written
So while the store-side primitive may be useful, the load-time primitive is
less useful. At some stage we could discard it in favor of a pure
Smalltalk loader, provided performance is good enough, e.g. for Terf.
I''ve added subclasses of ImageSegment called NativeImageSegment and
>> LegacyImageSegment. I have a few questions.
>> 1. I'm curious about the ImageSegment>>scanFrom: method:
>> !ImageSegment methodsFor: 'fileIn/Out' stamp: 'RAA 6/22/2000 17:49'!
>> scanFrom: aStream
>> "Move source code from a fileIn to the changes file for classes in an
>> ImageSegment. Do not compile the methods. They already came in via the
>> image segment. After the ImageSegment in the file, !!ImageSegment new!!
>> captures control, and scanFrom: is called."
>> | val chunk |
>> [aStream atEnd] whileFalse:
>> [aStream skipSeparators.
>> val := (aStream peekFor: $!!)
>> ifTrue: ["Move (aStream nextChunk), find the method or class
>> comment, and install the file location bytes"
>> (Compiler evaluate: aStream nextChunk logged: false)
>> scanFromNoCompile: aStream forSegment: self]
>> ifFalse: [chunk := aStream nextChunk.
>> aStream checkForPreamble: chunk.
>> Compiler evaluate: chunk logged: true].
>> aStream skipStyleChunk].
>> "regular fileIn will close the file"
>> ^ val! !
>> I don't see this in the new (legacy, System-bf.916) image segment loading
>> code. Do Etoys projects not include source code and hence its irrelevant?
> They do, but the code path to load them does not go through this method.
What path does it go through?
The question boils down to whether this should be a method on ImageSegment,
>> and hence inherited by both NativeImageSegment and LegacyImageSegment, or
>> just NativeImageSegment. Also, I don't see how this method attaches
>> sources to the methods brought in. Is it in fact useless and hence that's
>> why you've left it out? If so, would it not be useful to add code that
>> does update the source pointers in loaded methods correctly?
> The ImageSegment nowadays only stores uniclasses as objects. Other changes
> to the system are simply filed in, and not included in the segment. This
> method appears to be unused. I tried to trim everything not strictly
> needed for project loading
OK, I'll nuke it.
> 2. Should I pull in the Etoys extensions deleted in EToys-bf.234?
>> I'm guessing they'll be needed for native loading.
> : is still there, not sure if we need the others.
>> 3. Same for the SMBase extension ImageSegment>>writeForExportOn:. I
>> guess this should come back in too.
> SqueakMap uses ImageSegments, yes. No idea if it's a good idea to write a
> new native segment.
> 4. Where should I look for tests that may have been removed?
> There are no tests, unfortunately ...
> The way I was testing is taking a project written by Etoys (from
> squeakland.org) and dropping it into a trunk image. Before your changes
> this did work, more or less: If you revert to System-pre.956 and
> make allocateMethodContext: and allocateBlockContext: simply return nil,
> you can load old projects again. These contexts aren't needed anyways for
> Etoys projects (if we needed them I'd recreate by recompiling).
> I'm attaching a simple project that I just tested with this.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Squeak-dev