Perhaps you have examples where the gain is large - I've yet to encounter them.
jryan's benchmark code mentioned above can be tweaked with qr for a significant speed optimization. Change the line
my @patterns = ('B.B', 'CB')x10;
to
my @patterns = map qr/$_/, ('B.B', 'CB')x10;
in the &without subroutine.
Your benchmark doesn't show any support for compiled patterns, and that's due to an optimization I've already described here.