[MLton] Re: [MLton-commit] r6608

[Matthew Fluet]

On Thu, 24 Apr 2008, Vesa Karvonen wrote:
> Loop over lists (of declarations) rather than vectors (of ...) in the
> translation pass, because this allows the elements to be GC'ed, which
> reduces the memory requirements dramatically (verified by monitoring
> process memory usage).
>
> ----------------------------------------------------------------------
>
> U   mlton/trunk/mlton/defunctorize/defunctorize.fun
>
> ----------------------------------------------------------------------
>
> Modified: mlton/trunk/mlton/defunctorize/defunctorize.fun
> ===================================================================
> --- mlton/trunk/mlton/defunctorize/defunctorize.fun	2008-04-23 00:23:08 UTC (rev 6607)
> +++ mlton/trunk/mlton/defunctorize/defunctorize.fun	2008-04-24 14:27:24 UTC (rev 6608)
> @@ -912,7 +912,10 @@
>                end
>          end
>       and loopDecs (ds: Cdec.t vector, (e: Xexp.t, t: Xtype.t)): Xexp.t =
> -         Vector.foldr (ds, e, fn (d, e) => loopDec (d, e, t))
> +         loopDecsList (Vector.toList ds, (e, t))
> +      (* Convert vector->list to allow processed Cdecs to be GC'ed. *)
> +      and loopDecsList (ds: Cdec.t list, (e: Xexp.t, t: Xtype.t)): Xexp.t =
> +         List.foldr (ds, e, fn (d, e) => loopDec (d, e, t))
>       and loopExp (e: Cexp.t): Xexp.t * Xtype.t =
>          let
>             val (n, ty) = Cexp.dest e

It's an interesting observation, with a relatively simple fix.  We 
generally prefer vectors to lists, because vectors are more space 
efficient.  But, as observed, when mapping or folding over a vector, the 
vector will remain live until the end of the map/fold.

There may be other places where breaking a vector into a list before 
processing would allow elements to be GCed early.  On the other hand, I 
suspect that it is particularly bad with the defunctorize pass because the 
top-level representation of a CoreML program is:
             datatype t = T of {decs: Dec.t vector}
So, we would have retained the entire CoreML program while we built up the 
entire XML program.  (Worse than that, actually, as the XML program is 
build up with the Xml.DirectExp functions, which do things in 
continuation passing style, so we build up a lot of closures (presumable 
larger than the XML program they denote).)