Seemed odd to me, so I ran it on 5.8.8 and the current bleadperl 5.11. The absolute numbers for my computer are, of course, different, but the general observation holds for 5.11. The really disturbing part is that 5.8.8 executes the atomic1/2 cases at about the same speed as "normal". Oh, and on a side note: The difference in execution speed of 5.11's "normal" and 5.8.8's "normal" may be due to 5.11 being a debugging build and all, so it's not necessarily significant.

#!/usr/bin/perl
use strict;
use warnings;
use version;
use Benchmark qw(cmpthese);

my $str = "bea" x 100;
my $re = qr/(?:be|ea|a)/;
cmpthese(-2, {
    atomic1   => sub { die if $str =~ m/(?>$re+)\d/ },
    atomic2   => sub { die if $str =~ m/(?>$re)+\d/ },
    normal    => sub { die if $str =~ m/$re+\d/ },
    ( version->new($]) >= version->new("5.10.0")
      ? (posessive => sub { die if $str =~ m/$re++\d/ })
      : ()
    ),
});
__END__
5.8.8:

          Rate  normal atomic2 atomic1
normal  5000/s      --     -4%    -18%
atomic2 5208/s      4%      --    -15%
atomic1 6132/s     23%     18%      --

5.11:
            Rate posessive   atomic1   atomic2    normal
posessive 59.9/s        --       -1%      -97%      -97%
atomic1   60.4/s        1%        --      -97%      -97%
atomic2   1953/s     3159%     3134%        --       -9%
normal    2156/s     3500%     3472%       10%        --
[download]

I'm hoping for demerphq to chime in with an explanation. :)

Cheers,
Steffen