On 14.09.2011 2:44, Andrei Alexandrescu wrote:
On 9/13/11 4:37 PM, Dmitry Olshansky wrote:
OK in this particular benchmark, we are slighlty behind Google inc. :)
Namely RE2 and irregex.
On my machine RE2 is ~5 sec while FReD does the job in ~7sec.
As for irregex, ... uhm-oh I haven't actually tested it (failed to just
compile) but it beats RE2 in this test by some ~30%.
I'm judging by this page, RE2 I mentioned is "C++ GNU g++ #2" but with
OpenMP declarations stripped:
http://shootout.alioth.debian.org/u32/performance.php?test=regexdna
Damn looks like we could be #4... these things change too fast :(
Anyway, the rest of competition is not even close, and I still haven't
spent all of my tricks.
Sounds great! I advise you to invest effort in reducing the 25% handicap
to <1%.
Yes, sir! :)
As regexen have a fairly standardized interface, speed is the
most impactful differentiatior people look at.
Well in this particular benchmark there is little I can do, some small
scale optimizations and such, I expect no more then 10%.
The bigger picture is that our best bet here is static regex, the
competitors use some aggressive JIT so it's a fair play. I actually
received pull request this morning do just that (thx, Martin) though
dmd uses so much memory it's not for mortals yet:
I had 2G of ram + some 600mb of swap - fail on 3rd pattern of 9
+ 2 Gb of swap - 6 out of 9
Come on, it's a sport!
+ 4 Gb of swap and after few restless minutes I got it working
Fresh timings:
FReD C-T 3.4 sec
FReD R-T 6.6 sec
RE2 4.8 sec
V8 JS 3.7 sec
Look ma, we are on top! :)
So, by the end of day, D does kick ass, even though it's a bit of
Pyrrhic victory.
---
Dmitry Olshansky