lua-users home
lua-l archive

[Date Prev][Date Next][Thread Prev][Thread Next] [Date Index] [Thread Index]

On Sat, May 30, 2009 at 18:43, David Manura <> wrote:
> On Sat, May 30, 2009 at 4:19 AM, Cosmin Apreutesei wrote:
>>> That being said, I'd dare making a proposal that I've been giving some
>>> thought, maybe I can hear your opinions about it (well, aside from the
>>> trouble it would mean to implement it): how about an __immediate_gc
>>> metamethod that would ensure finalization of a table or userdata when
>>> the last reference goes out of scope.
> On Sat, May 30, 2009 at 4:51 AM, Bulat Ziganshin wrote:
>> you can't work this way with external resources. imagine for example
>> that your program opens files and process it some way. if file will be
>> not closed until gc/exit, user trying to remove this file from other
>> program will got unexpected fail
> Cosmin's proposal though is to extend Lua to support both tracing
> garbage collection and deterministic destruction, with the latter done
> through reference counting.  The choice between the two would be
> selectable per object, depending on the metamethod defined on the
> object.
> Reference counting is one way to achieve deterministic release of
> resources.  However, it is not the only solution [1].  Perl supports
> the reference counting approach.  Some implementations of Python
> support it too (C not Java [2-3]), but it is an implementation
> decision, and the more portable method is to use the with statement
> [4], which is similar to the proposals made for Lua, where resources
> are released simply upon scope exit, irregardless of reference counts
> (which in a correct program normally should be zero upon scope exit).
>>> But that breaks garbage-collecting semantics, which assures me never
>>> to have invalid pointers no matter what... IMHO this brings me back
>>> down to the C-level of responsibility (watching the stack).
> Even reference counts have their limitations (e.g. cycles [3]), and
> you are not prevented from doing dumb things like setting the variable
> to nil and then trying to access the variable or accidentally storing
> the resource in a global variable, which prevents the resource from
> being released upon scope exit as might be assumed happens (thereby
> only hiding the error).

I think the problem with ref-counting is the performance hit of the
cycle detection algorithm (although it's usually small potatoes
compared to the finalization code itself -- as it was noted before,
even triggering collectgarbage() usually takes less) -- cycles may not
be such a big problem in practice, but if you go down that path, you
gotta have full correctness. There might be better ways to implement
determnism in gc, just don't know of any myself.

I also agree that hanging references could be a problem too, though I
assume it's easy to check with a small checking script, and you could
always impose yourself more discipline :D

I'd like to express just 2 more points and then I'm kicking myself out:

1) I'd rather not have the distinction between a
closed/invalid/useless handler and a nil binding (that is, no
binding), the same way lua assures me I'll never have invalid pointers
thus no memory trashing or access violations ever.

2) the with/scoped approach has IMHO too narrow application. It is
only good when the flow control of the program goes in sync with
resource allocation/finalization (acquire -> use -> dispose). But
think about a Win32 binding, with tons of handlers of many types
floating around in tables, being passed around from function to
function... Who should take care of their destruction? At this point
you've got to implement a resource manager so you can have authority
over finalization, since you're doing it all by yourself -- gc sittin'
there doing nothing for you -- just like java :)

Instead, I don't want to manage resources and create policies about
their use. It's not that I'm lazy, but the bookkeeping really adds no
value to my programs. All I want is to say t = nil, where t is a
spaghetti tree of object handlers intermixed with memory objects
(colors, canvases, panels, ...strings, ...) and be sure they're gone:
no delay, no collectgarbage(), no invalid pointers on scope exit. Out
and clear.

> [1]
> [2]
> [3]
> [4]