Re: GC puzzler...

[David Jones <drj@...>]

On Feb 15, 2005, at 13:02, Adam D. Moss wrote:

> Hi there!
>
> foo = {}
> wtable[foo] = setmode({1, 2, foo, 4}, "k")
> foo = nil
> [...do lots of other stuff not involving wtable]

You haven't obviously assigned wtable (to be a table value) before 
indexing it with foo, and you haven't explained what setmode does.

But I'll assume that you mean to put an ordinary (strong) reference to 
foo in a table which is only reachable by using foo as a weak key in 
some other (weak) table.  Or something like that.

> Will 'foo' be GC'able (and more to the point will Lua actually
> GC it)?  The only surviving way to 'get at' foo is indirectly
> through the [foo] key of wtable, and wtable is weak-keyed, so
> foo is the only thing keeping itself alive and thus it might
> reasonably be argued that 'foo' is GC'able.

I agree, foo should be GC'able (the GC should be permitted to remove 
it).  In particular all paths of references from the roots to 
the-table-formerly-known-as-foo have at least one weak reference in the 
path.  Or in other words there is no path of only strong (by which I 
mean an ordinary, not weak, reference) references from the roots to 
foo.
> It looks like it'd probably be tough to implement such GC
> semantics though, so I'm guessing that no-one does.

It's not tough, the Memory Pool System, 
http://www.ravenbrook.com/project/mps/, does it (and I'm pretty sure 
most of Sun's JVM GC's do it; certainly their explanation of 
reachability with respect to weak references makes it appear like they 
ought to).  All you have to arrange is that you first identify the set 
of objects that are reachable from the roots via paths that have only 
strong references.

I have no idea what Lua's GC actually does.

David Jones