Re: GNU AUCTeX branch, master, updated. 34befcc17649fd6615fd4ef8756fb78705f7fe55

[Tassilo Horn]

Arash Esbati <arash@gnu.org> writes:

>> Yes, the performance aspect.  The annotation function might be called
>> hundredth or thousands of times, and appending lists is expensive.
>>
>> ;; The current version
>> (benchmark 10000
>>            '(LaTeX--completion-annotation-from-math-menu "aleph"))
>> "Elapsed time: 0.136648s"
>>
>> ;; Your proposed version
>> (benchmark 10000
>>            '(LaTeX--completion-annotation-from-math-menu-with-append 
>> "aleph"))
>> "Elapsed time: 2.190776s (1.388426s in 18 GCs)"
>>
>> So here it would almost be a factor of 20 slower.
>
> Thanks for your responses.  I was aware of `append' being expensive,
> but I didn't expect the hit being so large for this case.

Me neither but I do get those numbers with this definition (and the
current emacs master):

--8<---------------cut here---------------start------------->8---
(defun LaTeX--completion-annotation-from-math-menu-with-append (sym)
  (catch 'found
    (dolist (e (append LaTeX-math-list LaTeX-math-default))
      (let ((val (cadr e)))
        (when (and (stringp val)
                   (string= val sym))
          (let ((char (nth 3 e)))
            (when char
              (throw 'found
                     (concat " " (char-to-string char))))))))))
--8<---------------cut here---------------end--------------->8---

What makes me very curious is that LaTeX-math-list is `nil' so that
`append' needs to copy the first list should be a non-issue here.  And
in fact, `benchmark' suggest appending those two concrete lists is
faster than building a 2-element list containing the values of both
variables which makes sense:

--8<---------------cut here---------------start------------->8---
(benchmark 1000000
           '(list LaTeX-math-list LaTeX-math-default))
"Elapsed time: 0.607044s (0.490079s in 6 GCs)"

(benchmark 1000000
           '(append LaTeX-math-list LaTeX-math-default))
"Elapsed time: 0.170943s"
--8<---------------cut here---------------end--------------->8---

So how can the original benchmarking result be true?

Well, the answer is that I've benchmarked a compiled vs. an interpreted
function.  After evaluating

  (byte-compile 'LaTeX--completion-annotation-from-math-menu-with-append)

the runtimes of the benchmarks are about the same with slight advantages
for the version using `append'.  I guess as soon as `LaTeX-math-list'
has two or more items, it'll tip over to the double-dolist variant.

Bye,
Tassilo