One possible solution is to check @- and @+:

use warnings;
use strict;

sub calcit {
    my $in = shift;
    my %rules = (A => 1,B => 1,C => 0,D => 1);    
    my ($s,$s1,$s2) = (0)x3;
    while ($in=~/(\d+)([ABCD])/g) {    
        my ($n,$op) = ($1,$2);
        $s += $n * $rules{$op};
        $s1 += $n * $rules{$op}
            unless $op eq 'D' && $-[0]>0;
        $s2 += $n * $rules{$op}
            unless $op eq 'D' && $+[0]<length($in);
    }
    return ($s,$s1,$s2);
}

use Test::More;
is_deeply [calcit("2D40A2C1B4D")], [47,43,45], "2D40A2C1B4D";
is_deeply [calcit("2D40A2C1B")],   [43,43,41], "2D40A2C1B";
done_testing;
[download]

But TIMTOWTDI, for example you could ignore all "D" codes in the initial processing of the string and then use a separate regex (/^(\d+)D/ resp. /(\d+)D$/) to account for them.