Re: GC + Java finalization

[Jonas Hahnfeld]

Am Freitag, dem 19.11.2021 um 18:48 +0000 schrieb Maxime Devos:
> Jonas Hahnfeld schreef op vr 19-11-2021 om 15:52 [+0100]:
> > Am Freitag, dem 19.11.2021 um 14:14 +0000 schrieb Maxime Devos:
> > > [...]
> > > 
> > > From your other responses, I now know it is actually related to
> > > (non-
> > > )Java style finalisation, but my comment about ‘separate patch’
> > > still
> > > seems to apply:
> > > 
> > > > 
> > > > Again, as replied in July to the same comment, it *is* a separate
> > > > patch for exactly this reason.
> > > 
> > > More concretely, it is in the same patch as that modified
> > > libguile/random.c.  The patch to libguile/random.c doesn't seem to
> > > be for non-Java finalization reasons. Going by the commit message,
> > > the only possible reason I could find is:
> > > 
> > > ‘There is no point in registering memory with the garbage collector
> > > if it doesn't need to do its job’
> > > 
> > > But I don't see any ‘registering memory’, only replacing
> > > scm_gc_calloc+scm_gc_free by scm_calloc+free, and without any
> > > finalisation in sight. Unless you mean with ‘registering memory’
> > > the "random bignum chunks" argument. But that still seems unrelated
> > > to non-Java finalization.
> > 
> > Any memory allocation through gc implicitly registers the memory.
> 
> I don't mean what you mean with ‘registering memory’. I don't
> see that phrase anywhere at <https://www.hboehm.info/gc/#details>
> or <https://www.hboehm.info/gc/faq.html>.  I only know about
> registering finalisers, but not about registering memory.

Maybe it's not an official term; I call it "registration" because the
garbage collector has to keep track of all memory segments it handed
out, so it "registers" the allocation in a data structure somewhere.

> Also, I'm not sure what you are trying to say here and in the following
> paragraph.  Is this some kind of argument for why the change to
> libguile/random.c should be in the same patch, or general explanation,
> ...?

Both changes address the same issue of removing matching calls to
scm_gc_alloc+scm_gc_free. That's why I think it's one logical change,
even though one is actually a problem when disabling Java-style
finalization - it makes sense even without the following patch.

> 
> > Both changes are unrelated to finalization, they are there to avoid this
> > unnecessary registration.
> Thanks for the clarification, though I have no idea what ‘registration’
> is ...
> >  My previous replies only tried to clarify why
> > any other solution is worse.
> 
> ... but what problem and what replies are you referring to here?
> I haven't seen any e-mails explaining GC problems in libguile/random.c.
> I only have seen replies about non-Java style finalisation, which
> do not apply to libguile/random.c (no objects but the stack have a
> reference to random_chunks anywhere and libguile/random.c is not
> playing with finalizers).

Yes, the one in libguile/random.c is just unnecessary, probably not
actually a problem.

Jonas

signature.asc
Description: This is a digitally signed message part