memory costs of storing lambdas?

memory costs of storing lambdas?

[Han-Wen Nienhuys]

(please CC replies, I'm not subscribed to guile-user)

Hi there,

I have a small question: I was wondering how much memory it costs to
store an anonymous function.  I have to make choices between storing
various entities in lilypond either as a function, or as an
expression.

The following fictional example illustrates the distinction (#
introduces in-line Scheme in Lilypond syntax):


1. using anon functions:

  (define (style-setter style)
    (lambda (head) (set-note-head-style head style)))
  ...

  \property Voice.NoteHead \set #'style-procedure
    = #(style-setter "cross")

During the lilypond run, the anonymous function is called. Storage
cost: one anonymous function

2. using expressions

  \property Voice.NoteHead \set #'style-expression
    = #(list set-note-head-style "cross")

During the lilypond-run, a note-head pointer is prepended to the
argument list, and the list (set-note-head-style head style) is
evaluated.

Storage cost: a list of length 2.


I have the feeling that solution 1. is more expensive in terms of
memory than solution 2. My question: is this true, and how could I
have found out this answer myself? (i.e. where is the related
interpreter code?)

Thanks,

Han-Wen

-- 

Han-Wen Nienhuys   |   hanwen@cs.uu.nl   |   http://www.cs.uu.nl/~hanwen 


_______________________________________________
Guile-user mailing list
Guile-user@gnu.org
http://mail.gnu.org/mailman/listinfo/guile-user

Re: memory costs of storing lambdas?

[Mikael Djurfeldt]

Han-Wen Nienhuys <hanwen@cs.uu.nl> writes:

> 1. using anon functions:
>
>   (define (style-setter style)
>     (lambda (head) (set-note-head-style head style)))
>   ...
>
>   \property Voice.NoteHead \set #'style-procedure
>     = #(style-setter "cross")
>
> During the lilypond run, the anonymous function is called. Storage
> cost: one anonymous function

According to eval.c:scm_closure and procs.h:

closure = < closcar, env >
closcar = < code, properties >
code = < formals, body >

So, the extra cost (not counting the cost of the list of expressions
in `body') is 3 pairs + pairs in formals list.

In your case, you also surround the closure with an environment
captured by the closure.  This environment has the structure
(explained in the Guile docs):

new-env = < frame, env >
frame = < formals, values >
formals = < 'style, '() >
values = < style-value, '() >

So, you have 4 pairs for the environment, 3 pairs for closure
overhead, 1 pair for formals list, 1 pair for list of expressions in
body, and, 3 pairs for the expression.  So, in total, the object
returned by `style-setter' above will consume 12 pairs.

> During the lilypond-run, a note-head pointer is prepended to the
> argument list, and the list (set-note-head-style head style) is
> evaluated.
>
> Storage cost: a list of length 2.

If you have only two items to store, the most efficient data structure
available is one pair.

For more items, the most efficient structure is a GOOPS object or a
vector.  The cost of ordinary (not applicable) GOOPS objects is one
pair + mallocated memory of size sizeof(ptr) * n where n is number of
items.

Best regards,
Mikael


_______________________________________________
Guile-user mailing list
Guile-user@gnu.org
http://mail.gnu.org/mailman/listinfo/guile-user

Re: memory costs of storing lambdas? (correction)

[Mikael Djurfeldt]

Mikael Djurfeldt <djurfeldt@nada.kth.se> writes:

> In your case, you also surround the closure with an environment
> captured by the closure.  This environment has the structure
> (explained in the Guile docs):
>
> new-env = < frame, env >
> frame = < formals, values >
> formals = < 'style, '() >
> values = < style-value, '() >
>
> So, you have 4 pairs for the environment

Oops.  The `formals' pair is constant, so it doesn't add to the cost
of new environments.

> So, in total, the object returned by `style-setter' above will
> consume 12

Rather 11, then.


_______________________________________________
Guile-user mailing list
Guile-user@gnu.org
http://mail.gnu.org/mailman/listinfo/guile-user