2014-04-06 22:29 GMT+02:00 Nicolas Cellier <nicolas.cellier.aka.nice@gmail.com>:

Good news,

I've managed to make your issue apparently vanish in Pharo3.0 VM.

For this, I:
- cleaned up unecessary differences with Eliot's VMMaker.oscog branch,
- removed unecessary changes ahead of VMMaker.oscog-eem.333 version (those stamped LucFabresse)

- carefully applied changes from VMMaker.oscog-eem.333 that were not applied

  Note that this version was marked as merged, which it was obviously not

  Please, if you do not fully merge, but just cherry pick some changes, it's better to not merge.

- upgraded to (merged) VMMaker.oscog-eem.335 because it fixes a snafu

- integrated the issues for which I emitted a pull request

All this work can be found publicly at https://github.com/nicolas-cellier-aka-nice/pharo-vm/compare/fixMergeWithEliotVersion333

I cannot tell which missing change exactly was the root cause, and I cannot dissect either, but I saw several + LiteralStart missing... Or it could be related to cogit method/block native code generation...
The changes are most probably here
https://github.com/nicolas-cellier-aka-nice/pharo-vm/commit/af718618eee516d15c0b362123e3a77c8b6fd2e8

Bad news,

the structures at beginning of src/cm/cogit.c are generated out of order, and I didn't find the cause yet.

Once solved, I think my branch should be carefully reviewed and integrated.

Haha! this is Class class>>superclassOrder: which break things...

The order of structures passed on input was mostly good but ruined on output...

The Squeak ChangeSet class>>superclassOrder: does not convert the list of classes asSet, thus somehow preserves provided order.

The Pharo version does not take so much care it seems...

Apparently the method was changed again in Pharo3.0, but neither does it preserve order.

After fixing it, compilation went well...

2014-04-03 4:13 GMT+02:00 Nicolas Cellier <nicolas.cellier.aka.nice@gmail.com>:

2014-04-02 23:59 GMT+02:00 Stephan Eggermont <stephan@stack.nl>:

Nicolas wrote:
>Hmm, don't waste too much time.
>All these methods are unchanged in github pharo-vm head revision (VMMaker-oscog-StefanMarr.313)
>The cause must be completely different in latest pharo vms.

Uhm, why? They haven�t changed.

Sorry, haven't changed means that there's no diff with the corrected version.
IOW, the fixes from Eliot have been integrated for a long time already.

You are seeing another symptom, probably another bug impacting the same code.

There are many diffs between pharo's and Eliot's branch, but not where you're looking at.

Pharo vms 105 and later no longer crash, but all reliably show the bug.
And 236 is the last Squeak one to show the problem.

I might take a look at Esteban�s versions from the same time,
228 and 229

Stephan