Hmm. I originally used a debug session & timethese(), did a little mental arithmetic, that jelled with what Anonymonk said, so I went with it.

Now I've put it into a proper script, used cmpthese() to do the arithmetic, and I can't reproduce my original results. In fact, I seem to consistantly get the index version coming out quicker, sometimes markedly so??

Maybe I did the arithmetic wrong, or there is a flaw in my benchmark? I used random strings with (or without) randomly positioned ':' to try cover the failing conditions as well as a good spread of the passing ones. The spread of variation in timings seem to indicate that some combinations cause the regex engine to take much longer than others. I've not managed to get the difference less than 9% in favour of index, and on one occasion, it went as high as 28%. Further investigation is called for:)

#! perl -slw
use strict;
use Benchmark qw[ cmpthese ];

sub rndStr{ join'', map{ $_[ rand scalar @_ ] } 0 .. shift }

our @strings = map{ rndStr 8, ':', 'a' .. 'z' } 1 .. 1000;

our( @a, @b );

cmpthese( -3, {
     regex => q[
         our @a = sort{
             ( $a =~ /:/ <=> $b =~ /:/ )
             ||
             $a cmp $b
         } @strings
     ],
     index => q[
         our @b = sort{
             ( ( index($a,':') >= 0 ) <=> ( index($b,':') >= 0) )
             ||
             $a cmp $b;
         } @strings
     ],
 });

print 'Okay' if "@a" eq "@b";

__END__
P:\test>test3
        Rate regex index
regex 14.6/s    --   -9%
index 16.1/s   10%    --
Okay

P:\test>test3
        Rate regex index
regex 13.7/s    --  -15%
index 16.2/s   18%    --
Okay

P:\test>test3
        Rate regex index
regex 14.5/s    --  -11%
index 16.3/s   13%    --
Okay

P:\test>test3
        Rate regex index
regex 13.1/s    --  -19%
index 16.2/s   23%    --
Okay
[download]

Interesting. Running that code exactly gives me a win for 'regex' on both 5.6.1/Solaris-7 running on a Sun and 5.8.0/Linux-2.4.18 running on a PC. On the Sun, the win is usually large: between 15% and 21%. On the Linux box, it's usually only very small: 0% to 2%.

-sauoq
"My two cents aren't worth a dime.";

Intriguing. Maybe it comes down to the underlying C-runtime code? I just tried it with AS 5.8.0, AS 5.6.1 and 5.8.1 built with Borland, and the results always favour index... They also highlight the penalty of unicode support:(.

P:\test>perl5.8.0 test3.pl8
            Rate     regex     index Abi_regex Abi_index
regex     14.6/s        --      -10%      -14%      -21%
index     16.3/s       11%        --       -4%      -13%
Abi_regex 17.0/s       16%        4%        --       -9%
Abi_index 18.6/s       27%       15%       10%        --
Okay

P:\test>e:\perl5.8.1\bin\perl5.8.1 test3.pl8
            Rate     regex     index Abi_regex Abi_index
regex     13.1/s        --      -11%      -27%      -31%
index     14.7/s       12%        --      -18%      -23%
Abi_regex 17.9/s       37%       22%        --       -6%
Abi_index 19.0/s       45%       30%        6%        --
Okay

P:\test>perl5.6.1 test3.pl8
[snip]
            Rate     regex     index Abi_regex Abi_index
regex     13.8/s        --      -11%      -50%      -51%
index     15.5/s       13%        --      -44%      -45%
Abi_regex 27.6/s      100%       77%        --       -3%
Abi_index 28.4/s      106%       83%        3%        --
Okay
[download]

---

Examine what is said, not who speaks.
"Efficiency is intelligent laziness." -David Dunham
"Think for yourself!" - Abigail
Hooray!
Wanted!