Re: Design questions for a small s-expression library

[Antonio Vieiro <antonio@...>]

El 18/01/14 17:58, lua-l-request@lists.lua.org escribió:
> Date: Sat, 18 Jan 2014 17:38:31 +0100
> From: Philipp Janda <siffiejoe@gmx.net>
> Subject: Re: Design questions for a small s-expression library
> To: lua-l@lists.lua.org
> Message-ID: <lbealr$6nt$1@ger.gmane.org>
> Content-Type: text/plain; charset=UTF-8; format=flowed
>
> Am 18.01.2014 15:04 schröbte Antonio Vieiro:
> [...]
> C) Use closures for cons-cells. See [3].
> D) Similar to A, but use the keys `1` and `2` (i.e. the array part of
> the table) for head and tail.
>
> `cons`, `car`, `cdr`, `setcar`, and `setcdr` are the only functions that
> need to know the difference, so you can even change the implementation
> later ...
>
> Memory per cons-cell in bytes (as reported by `collectgarbage"count"`):
> table-based Lua 5.x:   144
> table-based LuaJIT:     80
> closure-based Lua 5.1: 136
> closure-based Lua 5.2: 128
> closure-based LuaJIT:   76
> array-based Lua 5.x:    96
> array-based LuaJIT:     56

I thought closures would be more expensive than tables, I may try that. 
Thanks, Philipp, for the idea!

> > 4. There's no problem in having circular dependencies using luaL_ref's
> > between different cons cells, because the Lua GC will detect that and
> > cleanse cons-cells appropriately.
> >
> > Am I right in these assumptions? I'm a little worried about point 4.
>
> With good reason. You are basically putting reference counting on top of
> garbage collection. Reference counting cannot handle cycles, so the
> cycles never become garbage. Even if you have chains instead of cycles
> you probably need multiple garbage collection runs before the whole
> chain is collected (the `__gc` function has to run before the garbage
> collector can realize that the next cons-cell may be collected as well).
> So I would advise against B even if you can make sure that there will be
> no cycles.

Well, this seems logical: Lua has no way to know that a userdata keeps a 
reference to another Lua object, so it's reasonable it cannot detect 
cycles between objects storing luaL_ref's among them.

Another option could be doing GC on top of GC, i.e., building a custom 
garbage collector for cons-cells. It would be great if one could run 
this custom garbage-collector immediately before Lua's garbage 
collector. But as far as I can tell there's no "hook" available in Lua 
that runs a custom C function before Lua's GC is run.

There's a LUA_MASKCOUNT hook [1], though, that could be used, but only 
if nobody else is using it. I may explore using it.

Thanks for the advice,
Antonio


[1] http://www.lua.org/manual/5.1/manual.html#lua_sethook