Circular references (was: Defining VCL-like framework for Py

[William Tanksley]

On Thu, 20 May 1999 15:00:03 -0500, Gordon McMillan wrote:

>Despite these drawbacks, I generally prefer ref counting to garbage
>collection. GC is generally nondeterminant - you know it _will_
>happen, but you have no idea _when_. So if your objects have some
>kind of outside resource open, you need to have a special free()
>method anyway instead of relying on freeing it in the destructor. 

I prefer refcounting as well -- it feels like a simpler, more controlled
solution.  It's not particularly "more deterministic", though; any
assignment to a pointer can result in a massive flurry of deallocations
:).  Ken Thompson recently mentioned the way they do GC in Inferno.  They
keep reference counts, and also do a lazy generational scan every once in
a while.  The reference counting controls the vast majority of the
garbage; the lazy scan eventually finds and cleans up the occasional
cycle, without spending too much time or causing too many other problems.
It looks like a nice solution to the problem of cycles.

The only problem is, of course, that an object which is stuck in a cycle
can't be trusted to shut down properly, just like in full GC.

>I realize many people feel strongly the other way, and I'm not going 
>to fight about it. Neither solution is perfect. I'm just stating my 
>preference.

Well, I agree -- but I would like to remove the problem of cycles.  It
would just be one less thing to complicate my Python code.

I've tried to find other methods, but nothing simple seems to work well --
everything requires some pretty nasty complications and overhead.

>- Gordon

-- 
-William "Billy" Tanksley
"But you shall not escape my iambics."
           -- Gaius Valerius Catullus