Here's my fevered imaginings.

With more than two derived classes, I'd use a hash or perhaps an array in B holding references to the constructors of the derived classes. It depends on how much the program may be extended in the future.

package B;

sub new {
    my $class = shift;    # ignore
    my $type = shift;
    if ($type == 1) {
        return D1->new(@_);
    } elsif ($type == 2) {
        return D2->new(@_);
    } else {
        return undef;
    }    
}

# more as appropriate

package D1;

@D1::ISA = qw( B );

sub new {
    # paranoia check
    # you may not want this and I haven't tested it anyway
    return unless caller eq 'B';
    # more as appropriate
}

package D2;

@D2::ISA = qw( B );

sub new {
    # you may not want this and I haven't tested it anyway
    return unless caller eq 'B';
}
[download]

Update: You might have to do something like the following:

package B;

my @models;    # factory model, get it?

sub add {
    my $self = shift;
    my $child_coderef = shift;
    push @models, $child_coderef;
}

sub new {
    my $class = shift;    # ignore
    my $type = shift;
    return unless $type;
    my $child_coderef = @models[$type - 1];
    return unless defined &$child_coderef;
    return $child_coderef->(@_);
}

package D3;

@D3::ISA = qw( B );
use B;    # gotta be done here

B->add(\&new);

sub new {
    # etc
}
[download]

That's where I'd start my experimenting. Good luck!