Yes, it's interesting and i've made a few more tests to find some optimization.
See below :

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

my $line = ('a' x 500) . 'b!' . ('a' x 20);


cmpthese -10, {
    literal     => sub {$line =~    /a.{1,10}b!/ },
    literal_exp => sub {$line =~    /a(?>.{1,10})b!/ },
    literal_look => sub {$line =~   /a(?>.{1,10})(?<=b!)/ },
    class       => sub {$line =~  / [a]    .{1,10}  [b] [!] /x},
    class_exp   => sub {$line =~   / a     .{1,10}   b  [!] /x},
    class_back  => sub {$line =~   / a (?> .{1,10} ) b  [!] /x},
    class_back_look  => sub {$line =~  /  a (?> .{1,10} ) (?<=b[!]) /x
+},
    class_b  => sub {$line =~          /  a     .{1,10}   (?<=b[!]) /x
+},
};
[download]

                    Rate class_b  class class_back_look literal_look l
+iteral class_back literal_exp class_exp
class_b           2172/s      --   -53%            -72%         -75%  
+ -100%      -100%       -100%     -100%
class             4651/s    114%     --            -40%         -46%  
+  -99%       -99%        -99%      -99%
class_back_look   7809/s    259%    68%              --         -10%  
+  -99%       -99%        -99%      -99%
literal_look      8658/s    299%    86%             11%           --  
+  -99%       -99%        -99%      -99%
literal         640333/s  29376% 13666%           8100%        7296%  
+    --        -9%        -11%      -13%
class_back      704687/s  32338% 15050%           8924%        8039%  
+   10%         --         -2%       -4%
literal_exp     722209/s  33145% 15426%           9149%        8242%  
+   13%         2%          --       -2%
class_exp       733546/s  33667% 15670%           9294%        8373%  
+   15%         4%          2%        --
[download]

Thanks for your interest and your enlightment.

PooLpi