I so much stand corrected! When I wrote it, I meant

my $string  = pack "C*" => map { 32 + int rand 95 } 0..1280;

cmpthese (-1, {
    subsingle   => sub { (my $s = $string) =~ s/[^a-zA-Z0-9 _-]//g  },
    subplus     => sub { (my $s = $string) =~ s/[^a-zA-Z0-9 _-]+//g },
    tran        => sub { (my $s = $string) =~ tr/a-zA-Z0-9 _-//cd   },
    });
[download]

Which gives completely different results indeed:

$ perl5.8.8 test.pl
             Rate subsingle   subplus      tran
subsingle  9514/s        --       -2%      -90%
subplus    9660/s        2%        --      -90%
tran      92839/s      876%      861%        --
$ perl5.10.1 test.pl
             Rate subsingle   subplus      tran
subsingle  7802/s        --       -9%      -92%
subplus    8574/s       10%        --      -91%
tran      99211/s     1172%     1057%        --
$ perl5.12.3 test.pl
             Rate subsingle   subplus      tran
subsingle  5191/s        --      -17%      -95%
subplus    6279/s       21%        --      -94%
tran      99211/s     1811%     1480%        --
$ perl5.14.1 test.pl
             Rate subsingle   subplus      tran
subsingle  6081/s        --      -18%      -94%
subplus    7450/s       23%        --      -92%
tran      97303/s     1500%     1206%        --
$ bleadperl test.pl
             Rate subsingle   subplus      tran
subsingle  6221/s        --      -12%      -94%
subplus    7096/s       14%        --      -93%
tran      97303/s     1464%     1271%        --
$
[download]

So the plus indeed speeds up the substitute and tr, even in /c mode, is fastest