I'm afraid I don't have any easy way to recreate it, other than throwing a lot of clients at it. It's perl 5.8.4 running under Linux 2.6.8. If I use the list form of accept, I get back the peer address correctly. Here's some debugging code:

        foreach my $socket (@ready)
        {
                print STDERR "Start accept\n";
                my ($connect, $paddr) = $socket->accept();
                my ($port, $iaddr) = sockaddr_in($paddr);
                print STDERR "Peer: ", inet_ntoa($iaddr), ":$port\n";
                if ($connect)
                {
                        if (!$connect->peerhost())
                        {
                                cluck("NULL PEERHOST: " . join(" : ",
                                        $connect->sockhost(),
                                        $connect->sockport(),
                                        $connect->peerhost(),
                                        $connect->peerport(),
                                        $!
                                ));
                                close($connect);
                                print STDERR "End accept (no peerhost)
+\n";
                                next;
                        }
                        $CLIENTS->add($connect);
                        print STDERR "End accept (success)\n";
                        next;
                }
                print STDERR "End accept (no connect)\n";
        }
[download]

Output looks like this:

Start accept
Peer: X.X.X.X:15627
NULL PEERHOST: Y.Y.Y.Y : 1235 :  :  : Transport endpoint is not connected at ./foo.pl line 115
End accept (no peerhost)

Where X.X.X.X is the client ip and Y.Y.Y.Y is my server ip. Perhaps this is an IO::Socket::INET bug?

On a related note, I'm also getting segfaults from perl under heavy load (~2500+ clients). The segfaults don't seem to be related to the null peerhosts. I believe they're occuring on syswrite() calls when I'm sending text back to the clients.