Re: [Sbcl-devel] Can the source transform for mapfoo make its initial cons dynamic-extent?

[Paul K. <pv...@pv...>]

Douglas Katzman wrote:
> Suppose you have the following
>
> (defun highly-popular-function (x) (mapcar #'something x))
>
> where not only is this function called very often, but almost always
> with X being NIL.
>
> On the whole it conses madly despite NIL input.  Two solutions come to mind:
> 1. the loop over the input should check whether there is anything to do.
> if not, just return nil
> 2. the "extra" cons cell that is pre-allocated for the purpose of being
> able to do RPLACD down the chain can perhaps be stack-allocated
[...]
> Specifically, in this loop in 'mapfoo-transform' and the similar one for
> 'nconc', and also for the LOOP macro (which I once found where to insert
> the dx but have since forgotten), I am proposing that MAP-RESULT be
> declared truly-dynamic-extent.
>
>                (let ((temp (gensym))
>                      (map-result (gensym)))
>                  `(let ((,map-result (list nil))) ; <-- this guy
>                     (do-anonymous ((,temp ,map-result) . ,(do-clauses))
>                       (,endtest (truly-the list (cdr ,map-result)))
>                       (rplacd ,temp (setq ,temp (list ,call)))))))

Sounds right to me. As far as I understand, DX contagion only happens 
when a value is bound or rebound to a variable that is declared DX. I'm 
tempted to implement both suggestions and expand into something like
  (unless [end-tests]
    (tagbody [...] (when [end-tests] (return ...)))
The replication ought to help rotate the loop nicely, and it also 
exposes more information to branch prediction. Downside: more code, but 
that's POLICY-able.

Paul Khuong