Re: Fix to long-standing crashes in GC

[Stefan Monnier <monnier@iro.umontreal.ca>]

> Now, when we on next gc calls mark_stack, and thus mark_maybe_object /
> mark_maybe_pointer, we may accidentally have a pointer address that
> leads to a (valid) cons, which is checked with mark_maybe_object...
> that's ok, you see that it points to a valid cons block, and it is not
> marked Vdead.

Oh, yes, I understand now, thank you.

> BTW, I think that clearing the cons cells before strings is still a
> proper sequence in gc_sweep, as the old sequence (that your patch
> reinstalls)  implies that after cleaning the strings, there are
> cons cells (unmarked though) that points into freed memory.

I think this is a red-herring: doing it your way leads to temporarily having
non-marked strings whose text-properties might point to deallocated
cons-cells.

The sweep step simply has to be atomic, just like most of the GC itself
(afterall, that's why the mark&sweep algorithm is in the category called
"stop-the-world").


        Stefan