property list hack, profiling, compiler optimization, looking at code

On 8/11/2012 2:38 PM, Stavros Macrakis wrote:
> On Sat, Aug 11, 2012 at 2:11 PM, Richard Fateman 
> <fateman at eecs.berkeley.edu <mailto:fateman at eecs.berkeley.edu>> wrote:
>
>     An additional observation regarding speed and  zerop1, which is
>     defined this way:
>
>      (defun zerop1 (x)
>       (or (and (integerp x) (= 0 x))
>           (and (floatp x) (= 0.0 x))
>           (and ($bfloatp x) (= 0 (second x)))))
>
>
> This is a strange definition, as for a non-zero integer, it still 
> checks floatp and $bfloatp.
oops!

>  More straightforward would be:
>
> (defun zerop1 (x)
>     (declare (inline $bfloatp))   ; worthwhile for such a common 
> function  (requires declaim/inline etc. @ $bfloatp)
>     (cond ((fixnump x) (= 0 x))   ; a bignum can't be 0
>               ((floatp x) (= 0.0 x))
>               (($bfloatp x) (= 0 (second x)))
>               (t nil)))
>
maybe we could use typecase

  (defun zerop7(x) (declare (optimize (speed 3)(safety 0)))
      (typecase x (number (zerop x))
                           (list (and (eq (caar x) 'bigfloat) (= (second 
x) 0)))))


or even

(defun zerop10(x) (declare (optimize (speed 3)(safety 0)))
   (if (numberp x)(zerop x)
             (and(listp x)(if (eq (caar x) 'bigfloat)(eq (second x) 0))


;; this one is half the length (in assembler , in allegro) of the 
original, and probably  faster.
;; note that (= 0 0.0) is true, so there is no need to discriminate 
between floats and fixnum and integers.
;; also if x is a fixnum I think that (eq x 0)  will work faster than (= 
x 0).


> Of course, the straightforward definition isn't necessarily the 
> fastest, and may depend on the particular compiler and processor.
yes
>
> Though I find micro-optimizing like this peculiarly entertaining,
the hacker mentality...

> I'm pretty sure this will make no noticeable difference to the user 
> experience,
agreed, though it may make things 1% faster :)

> certainly less difference than improving functionality in any one of a 
> number of ways.
>
>             -s
>
RJF