You're right to be concerned. These are the results of your orignal on my system:

C:\test>786728-1.pl
          Rate three   one   two
three  95432/s    --  -13%  -35%
one   110191/s   15%    --  -25%
two   146082/s   53%   33%    --
[download]

And these excluding the setup overhead:

C:\test>786728-2.pl
           Rate   one   two three
one   6935150/s    --   -0%   -6%
two   6965018/s    0%    --   -5%
three 7345874/s    6%    5%    --
[download]

As you can see, the setup swamps the code under test and skws the results horribly.

Here's my version of the benchmark:

use strict;
use warnings;
use Benchmark;

our @strings = qw(exception:tex exception:mex asdf tex:exception:mex);

Benchmark::cmpthese( -5, {
    'one'   => q[ my @filtered = grep { /exception:(?!tex)/} @unfilter
+ed; ],
    'two'   => q[ my @filtered = grep { /exception/ && !/tex/ } @unfil
+tered; ],
    'three' => q[ my @filtered = grep { /exception:/g && !/\Gtex/ } @u
+nfiltered; ],
});
[download]