[Pharo-project] Another fixes to finalization (was Re: [update 1.2] #12161 - #12172)
Henrik Sperre Johansen
henrik.s.johansen at veloxit.no
Mon Oct 4 17:29:37 EDT 2010
On 04.10.2010 22:26, Levente Uzonyi wrote:
> On Mon, 4 Oct 2010, Igor Stasenko wrote:
>> On 4 October 2010 22:51, Henrik Sperre Johansen
> <henrik.s.johansen at veloxit.no> wrote:
>> On 04.10.2010 21:47, Igor Stasenko wrote:
>>> On 4 October 2010 22:09, Stéphane Ducasse<stephane.ducasse at inria.fr>
>>>> so let us know in the bug entry what is the conclusion :)
>>> I think someone should verify my benchmarks i.e.
>>> [ self loadsomething ] timeToRun
>>> before and after patch.
>>> And conclusion is better be written by Henrik, because he's having
>>> concerns about speed,
>>> while i don't. :)
>> I already have:
> emm.. wait.
> WeakKeyDict should not delete associations with nil-ed keys
> automatically, because otherwise
> you won't be able to use it in weak finalization scheme.
It could in 1.0, without anyone noticing.
#rehash truncated duplicate nil keys, thus if it was triggered (f.ex. by
a manual removal, or adding enough to cause growth) in a thread with
priority higher than finalization process (ie ran after GC but before
finalization thread), the nil keys could be truncated before
finalization being run. An unlikely scenario, I admit, which I guess is
why noone encountered it.
#grow suffered the same problem. (well, actually, it didn't add any of
the nil keyed associations)
> There is two distinct use cases of weak-key dicts:
> a) for weak finalization
> b) for attaching some extra info(value) per object(key), which can be
> automatically discarded once object become garbage
> so, while in case (b) you can mercilessly kill/reuse associations with
> nil keys, once they discovered
> in case (a) you should preserve them until there is explicit request
> from outside to finalize values.
Yes, that's the point I've been trying to make these past couple of
> The need in (a), apparently prohibits from using (b) in most efficient
> So, i think, the solution would be to introduce a specialized weak-key
> dicts, which can work much better for (b).
The difference between a and b, and the fact Pharo 1.1's implementation
severly screws with the performance if b) (and honestly, isn't that good
for a) either) is the point I've been trying to make for months now...
In Pharo 1.1 we've moved to the other extreme compared to 1.0, nil keys
are never removed. b) always works, but a) HAS to be registered to work
satisfactory at all.
It was changed to do an extra step of keeping finalized assocs in a
special state where they can be simply replaced without rehashing after
finalization instead of doing rehash, but that's not really the major
result of the change.
> Squeak's implementation works well in both cases.
The way I read it (correct me if I'm wrong), the solution in Squeak was
adding a finalizer inst var to WeakKeyDictionary, that way you can
distinguish the two cases and handle them appropriately. (in addition
there were speed improvements for rehashing, growing, etc.)
I looked for another solution which would keep WeakKeyDictionary
oblivious to whether or not it was being finalized or not and could not
find one, thus I think it is worth porting.
Keeping the (pardon my fren.. err, norwegian) POS that is currectly in
1.1, or reverting to the 1.0 version for 1.2, is not a good alternative.
More information about the Pharo-dev