bug#18013: 24.3.92; looking-back "[^\n]" taking a lot of time in large buffers (was: Infloop in re_search_2)

[Nicolas Richard <theonewiththeevillook@yahoo.fr>]

Andreas Schwab <schwab@suse.de> writes:

> Nicolas Richard <theonewiththeevillook@yahoo.fr> writes:
>
>> I'll admit that (looking-back "[^\n]") is not exactly the canonical way
>> to test for (not (bolp)), but should it make an infloop ? I can't
>> reproduce though.
>
> I don't think it infloops, it just takes a very long time.  Since this
> is running in a process filter interrupts are disabled, so you have to
> be extra careful with what you do here.

It seems you were totally right, thanks ! Here's a repro test case :

# pick up a big file or make one :

$ yes | dd bs=1MB iflag=fullblock count=50 > foobartest 
50+0 enregistrements lus
50+0 enregistrements écrits
50000000 octets (50 MB) copiés, 0,853576 s, 58,6 MB/s

# open it in an emacs buffer and try looking-back :

$ time emacs --batch -Q --eval '(with-temp-buffer (insert-file-contents "foobartest") (goto-char (point-max)) (princ "Looking back...") (looking-back "[^\n]") (princ "done!") (terpri))'

"Looking back..."

"done!"


real	0m8.577s
user	0m8.541s
sys	0m0.047s

As you see it takes more than 8 seconds on my system, most of that time
is spent looking-back (inserting the file is quick).

(looking-back "[^\n]") is bad code, so I totally deserve this. Still, is
it possible to make the search smarter when the match is supposed to be
"anchored" at point ? Otherwise let's just close this bug.

-- 
Nico.