Re: GC with finalisation?

[John Skaller <skaller@maxtal.com.au>]

At 09:13 27/08/99 +0200, STARYNKEVITCH Basile wrote:
>>>>>> "John" == John Skaller <skaller@maxtal.com.au> writes:
>
>    John> I could use some advice on this. I'm implementing a Python
>    John> interpreter/compiler in Ocaml. Python class instances
>    John> support __del__ methods -- that is, destructors.  
>
>3 choices are then proposed. I suggest a fourth one:
>
> 4) patch the existing ocaml 2.02 garbage collector to allow finalized
>    objects containing (for example) a single Ocaml value:
>
>this is not very difficult; 

	I had thought of that, but I didn't include it in my list
for two reasons:

	1) I do not have the experience to undertake this task
without losing confidence in the resulting system. I have that
confidence now because I know the existing collector was written
by experts and has been tested by many users.

	2) It will mean my clients cannot build my product,
without installing and compiling a patched version of Ocaml:
bytecode versions won't work either.

>you'll have to create a new tag, i.e. add
>
>    #define Finalwithvalue_tag 249
>
>to mlvalues.h
>
>    then grep for Final_tag in the source, and add the case for
>    Finalwithvalue_tag (which should handle as value its second word
>    after the header, the first being the finalizer).
>
>Notice that ocaml finalizer (including those alreday provided thru
>Final_tag) should *not* allocate any value in the Ocaml heap (in
>contrast with Java finalizers).

	I see. This could be a problem in the first instance,
since the __del__ methods of Python class instances can
execute arbitrary code, invoking my ocaml Python interpreter:
this will certainly cause heap allocations.

	[This does not rule out a solution using indirection]

>On the other hand, I consider that finalized objects supported by the
>language (such as provided by Java & Python) are a mistake. 

	I have a strong tendancy to agree with you; but I have
a problem implementing a language which does support automatic
finalisation, which I myself have used heavily. So I need
some kind of solution anyhow.

	I have a tendancy _not_ to patch ocaml for just this reason:
however, I do wonder how other people handle finalisation issues;
including (but not limited to) releasing system resources.

>In my opinion, finalized objects should not have a user-provided finalizer,
>but should only deal with system resources (file descriptors, X11
>windows, coded only in C)

	That has the disadvantage that it only handles system resources
for which you have a suitable implementation. It makes sense to
create 'user defined' resources, and manage them in ocaml.

	It would seem that a module defining a 'generic' collector
with finalisation could help solve this and other related problems;
without compromising the ocaml memory collector.

	Such a collector would have to be 'layered ontop' of
a system, instead of leveraging the ocaml collector, but that
would probably be an advantage in many ways: collecting
non-memory resources may well require different semantics
than optimised memory collection.

-------------------------------------------------------
John Skaller    email: skaller@maxtal.com.au
		http://www.maxtal.com.au/~skaller
		phone: 61-2-96600850
		snail: 10/1 Toxteth Rd, Glebe NSW 2037, Australia