Wierd, though the tail version gets relatively more efficient if you increase the recursion level:

cmpthese( -10, {
        "normal" => sub { normal(500000) },
        "tail"   => sub {   tail(500000) },
        "faster" => sub { faster(500000) },
    }
);
[download]

       s/iter normal   tail faster
normal   2.26     --   -22%   -40%
tail     1.77    28%     --   -23%
faster   1.36    66%    30%     --
[download]