I think your original suggestion, of using Net::CIDR::cidr2range and merging the ranges end-to-end, was better. Your new program seems to be slower than the one I wrote following your Net::CIDR suggestion. Here's my program for comparison:

#!/usr/bin/perl

use Net::CIDR 'cidr2range', 'range2cidr';

my @ranges;
my ($cur_start, $cur_end, $cs, $ce);
while (<>) {
  chomp;
  my @r = cidr2range($_);
  my ($r_start, $r_end) = split /-/, $r[0];
  my ($rs, $re) = map {inet_to_n($_)} $r_start, $r_end;

  if (! defined $cur_start) {
    ($cur_start, $cur_end) = ($r_start, $r_end);
    ($cs, $ce) = ($rs, $re);
  } else {

    if ($rs == $ce + 1) {
      $cur_end = $r_end;
      $ce = $re;
    } else {
      print join "\n", range2cidr("$cur_start-$cur_end"), "";
      ($cur_start, $cur_end) = ($r_start, $r_end);
      ($cs, $ce) = ($rs, $re);
    }
  }
} continue {
  print STDERR  "$. records processed\n" if $. % 1000 == 0;
}

print join "\n", range2cidr("$cur_start-$cur_end"), ""
  if defined $cur_start;


sub inet_to_n {
  unpack "N", pack "C4", split /\./, shift();
}
[download]

I ran your program on nm5.in and waited three minutes. Then I started my program. My program finished first.

Of course, I might have made a mistake in the benchmarking somewhere.

--
Mark Dominus
Perl Paraphernalia