I was just curious about it, so here you go:

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

my @data = split /\n/, <<END;
800-800-1212
(800)800-1212
(800) 800-1212
800 800 1212
END
 
sub dosloop {
        my $re = shift;
        (my $phone = $_) =~ s/$re//g foreach @data;
}
 
sub doyloop {
        my $fake = shift;
        (my $phone = $_) =~ y/0-9//cd foreach @data;
}
 
cmpthese(-10, {
                single => sub { dosloop(qr/\D/) },
                multiple=> sub { dosloop(qr/\D+/) },
                y => sub { doyloop(qr/\D/) },  # Just a fake parameter
        }
);
[download]

I tried to be as fair as possible and have the same overhead for all calls. The results are the following:

             Rate multiple   single        y
multiple  95932/s       --      -6%     -39%
single   101881/s       6%       --     -35%
y        156303/s      63%      53%       --
[download]

The simple translation outperforms the other, as expected. I was particularly curious about putting the '+' or not. I vaguely remember some discussion about it before, but I couldn't tell you where.