I get

Benchmark: timing 100000 iterations of rollahead, unroll, unroll2...
 rollahead: 22.9776 wallclock secs (22.93 usr +  0.02 sys = 22.95 CPU)
+ @ 4357.30/s (n=100000)
    unroll: 23.0715 wallclock secs (23.03 usr +  0.02 sys = 23.05 CPU)
+ @ 4338.39/s (n=100000)
   unroll2: 22.3888 wallclock secs (22.35 usr +  0.03 sys = 22.38 CPU)
+ @ 4468.28/s (n=100000)
            Rate    unroll rollahead   unroll2
unroll    4338/s        --       -0%       -3%
rollahead 4357/s        0%        --       -2%
unroll2   4468/s        3%        3%        --
[download]

which still rates rollahead as faster than unroll, but barely. Interesting that runs on different machines can produce a 15% change in relative performance. (I am running perl 5.8.6.)

unroll2 is unroll with a more efficient regexp (one fewer alternation). That it is slightly faster suggests that the alternation is indeed slowing things down.

sub unroll2 {
    my @x = $_[0] =~ /(?:^|@)((?:#[#@]|[^#@])*)/g;
    for(@x){
        $_ =~ s/##/#/g;
        $_ =~ s/#@/@/g;
    } @x
}
[download]