#! /usr/bin/perl
use warnings;
use strict;

my $start = tell DATA;

sub orig {
    seek DATA, $start, 0;
    my $count = 0;
    while (<DATA>) {
        while (/(M+)/g) {
            $count++;
        }
    }
    return $count
}

sub ch {
    my $count = 0;
    seek DATA, $start, 0;
    while (<DATA>) {
        tr/M/M/s;     # tr/MIO/MIO/s seems a bit slower for longer inp
+uts.
        $count += tr/M//;
    }
    return $count
}


use Test::More;
is(ch(), orig(), 'ch4');
done_testing();


use Benchmark qw{ cmpthese };
cmpthese(-3,
         { orig => \&orig,
           ch   => \&ch,
         });

__DATA__
IIIIIIIIIIIMMMMMMMMMMMMOOOOOOOOOOOOMMMMMMMMMIIIIIIIIIMM
IIIIIIMMMMOOOOOMMMMIIIIIIIIIIIIIMMIIII
MIM
IMI
M
[download]

On my machine:

ok 1 - ch4
1..1
        Rate orig   ch
orig 59645/s   -- -34%
ch   89860/s  51%   --
[download]

I tested various other approaches ($count += () = /M+/g, split), but your original's speed was always the same or better.

Update: It seems to scale better with longer inputs, for 15K string, it shows 280%.