Em 2010.08.13. 13:09, Matthias Andree escreveu:
Gabor Kovesdan wrote on 2010-08-13:
Em 2010.08.13. 10:43, Doug Barton escreveu:
My reason is simple, performance. While doing some portmaster work
recently I was regression testing some changes I made to the --index*
options and noticed that things were dramatically slower than the last
time I tested those features. Thinking that I had made a programming
mistake I dug into my code, and while the regexps that I was using
could
be tuned for slightly better performance the problem was not in my
code.
I then installed textproc/gnugrep to compare, and the differences were
very dramatic using a highly pessimized test case (finding a match on
the last line of INDEX). The script I used to test is at
http://people.freebsd.org/~dougb/grep-time-trial.sh.txt and a typical
result was:
GNU grep
Elapsed time: 2 seconds
BSD grep
Elapsed time: 47 seconds
Ok, I'll take care of this soon, and make GNU grep default, again
with a knob to build BSD grep. I agree with you that we cannot allow
such a big performance drawback but I my measures only showed
significant differences for very big searches and I didn't imagine
that it could add up to such a big diference. I'm sorry for the bad
decision I took making it default.
Without knowing any of the details (I am not using 9-CURRENT), Gabor,
I suggest that you check the documentation around Google's RE2 library
(which is in C++); there are quite a few bits of information relating
to (including worst-case) performance of regexp matchers, both
directly in the re2 documentation, as well as indirect through links
and references. Might be worth a read, together with profiling Doug's
test case if he could tell you how to reproduce those.
Thanks, Matthias. I haven't looked deeply at this but iirc it uses
Perl-syntax. We need an efficient, wchar-aware, POSIX(ish) regex library
with a good license and atm only TRE conforms to these criteria.
Besides, we need GNU-style regex support, which will have to be added to
TRE before we can replace our libc-regex.
Gabor
_______________________________________________
freebsd-current@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-current
To unsubscribe, send any mail to "freebsd-current-unsubscr...@freebsd.org"
