Re: [Sbcl-devel] Re: Better %sxhash-substring?

[Christophe R. <cs...@ca...>]

Juho Snellman <js...@ik...> writes:

> Here are some cl-bench results (read into them what you will; I did
> three runs on both versions, and the results between runs varied at
> most 1-2%):

> COMPILER                 [     53.76]   0.91
> FPRINT/UGLY              [     10.62]   0.87
> FPRINT/PRETTY            [     22.62]   0.93
> HASH-STRINGS             [     11.06]   0.85
> CLOS/defclass            [     10.19]   0.94
> CLOS/defmethod           [     20.93]   0.93
> CLOS/method+after        [     16.01]   0.93
> CLOS/complex-methods     [      5.05]   0.93

Cool!  The FPRINT results are particularly intriguing... I actually
have a performance fix pending for FPRINT/PRETTY, but some extra is
not to be sneezed at :-)  The CLOS tests are, I suspect, making heavy
use of EVAL, and consequently the compiler; that might be why their
performance is improved.

> @@ -103,7 +121,12 @@
>  (defun %sxhash-simple-string (x)
>    (declare (optimize speed))
>    (declare (type simple-string x))
> -  (%sxhash-substring x))
> +  ;; This flet is a kludgy workaround for lack of let conversion in
> +  ;; the cross compiler, as suggested by APD.
> +  (flet ((trick (x)
> +           (%sxhash-substring x)))
> +    (declare (notinline trick))
> +    (trick x)))
>  
>  (defun %sxhash-simple-substring (x count)
>    (declare (optimize speed))

If you're set up already to test the effects of this, could you see
the effect of performing the same rewrite (with FLET TRICK) on
%SXHASH-SIMPLE-SUBSTRING?  It's only called in one place, but that one
place is FIND-SYMBOL*, so it may be relevant in performance terms.

I am currently working on making undefined variable errors be full
warnings (rather than the current style-warnings) in code such as
(defun foo () x) where there has been no previous declaration for X.
This will probably take me a few more days (some interesting problems
have shown up: for kicks, do (apropos "current-catch-block") in a
fresh sbcl), so if someone else would like to merge this patch, please
do :-)

Cheers,

Christophe
-- 
http://www-jcsu.jesus.cam.ac.uk/~csr21/       +44 1223 510 299/+44 7729 383 757
(set-pprint-dispatch 'number (lambda (s o) (declare (special b)) (format s b)))
(defvar b "~&Just another Lisp hacker~%")    (pprint #36rJesusCollegeCambridge)