It turns out that qr// is actually the worst way to build regexes, but adding /o mostly fixes the problem.

use Benchmark qw( cmpthese );
print $], "\n"; 
open my $W, '<', '/usr/share/dict/words' or die;
my @words = <$W>;
close $W;
my $s = '(?<![cC])[eE][iI]';
my $re = qr/(?<![cC])[eE][iI]/;
cmpthese(-5, {
   re  => sub { grep /(?<![cC])[eE][iI]/, @words },
   qr  => sub { grep /$re/, @words },
   qro => sub { grep /$re/o, @words },
   s   => sub { grep /$s/, @words },
   so  => sub { grep /$s/o, @words },
});
__END__
5.026001
      Rate   qr    s  qro   so   re
qr  7.14/s   -- -57% -67% -68% -69%
s   16.8/s 135%   -- -22% -26% -26%
qro 21.4/s 200%  28%   --  -5%  -6%
so  22.5/s 215%  34%   5%   --  -1%
re  22.7/s 218%  36%   6%   1%   --
[download]