Sounds like a challenge. I took the dj_string routine on rosettacode and simplified it some.

sub sieve_s2 {
    my ($n, $i, $s, $d, @primes) = (shift, 7);

    my $sieve = '110010101110101110101110111110' .
                '101111101110101110101110111110' x ($n/30);

    for ($sieve) {
        until (($s = $i*$i) > $n) {
           $d = 2*$i;
           do { substr($_, $s, 1, '1') } until ($s += $d) > $n;
           1 while substr($_, $i += 2, 1);
        }
        $_ = substr($_, 1, $n);
        push @primes, pos while m/0/gogo;
        return @primes;
    }
}
[download]

Very nice. About 1.5x faster, better looking code, though 2x memory use.

Addendum: Can you put it on RosettaCode or let me know it's ok to put there? Everything there gets covered by the GNU FDL so I don't want to put there without some sort of permission. More code improvements are also welcome :)

Regarding memory use: the result array footprint probably outweighs the sieve for limits up to 10¹⁴.

Re: code. m//gogo is the sole bit of artistic liberty (or wishful thinking) and can be shortened to just //g. And sure, you're welcome to it, to work on it or repost it. An anonymonk couldn't say no, anyhow.