With the other cleanup:

sub comb {
    map  permute( '', $_, @_ ), ( 0 .. $#_ );
}

sub permute {
    my ( $str, $depth, @chars ) = @_;
    map $str.$_,
      $depth
    ? (map permute( $chars[$_],
            $depth-1, @chars[ ( $_ + 1 ) .. ($#chars) ] )
       , (0..$#chars))
        : @chars;
}
[download]

That benchmarks about 10% faster than the original for me, but the powerset code below is 15x faster than either of them.

sub powerset {
    my ($car, @cdr) = @_;
    my @cdr_powerset = @cdr ? powerset(@cdr) : ();
    return
#     sort {length $a <=> length $b or $a cmp $b}
      ($car, map("$car$_", @cdr_powerset), @cdr_powerset);
}
[download]

If you include the sort that is currently commented out, the order will be correct, but the performance will degrade, though it is still about 7x faster for A..G. When I increase the range, the advantage for powerset increases. Even with the sort in place, it is 18x faster for A..L (30x without the sort).