#! perl -slw
use strict;
use Data::Dump qw[ pp ];
use threads;
use Time::HiRes qw[ time ];


sub thread {
    my $tid = threads->tid;
    my( $start, $end ) = @_;

    warn "$tid: processing $start to $end\n";
    my $count = 0;
    for my $mon ( $start .. $end ) {
        $count += unpack '%32b*', pack 'N', $mon;
    }
    return $count;
}


our $T //= 1;
our $M //= 1e6;

my( $step, $s, @ranges ) = ( int $M / $T, 0 );
push( @ranges, [ $s, $s+$step -1] ), $s+=$step for 1 ..$T;
$ranges[ -1][1] = $M;
my $start = time;

my @threads = map{
    threads->create( \&thread, @$_ );
} @ranges;

my $total = 0;
$total += $_->join for @threads;

printf "Total = %d (in %f seconds using $T threads)\n", $total, time()-$start;

__END__
C:\test>888714 -M=100e6 -T=1
1: processing 0 to 100e6
Total = 1314447116 (in 46.340000 seconds using 1 threads)

C:\test>888714 -M=100e6 -T=2
1: processing 0 to 49999999
2: processing 50000000 to 100e6
Total = 1314447116 (in 23.269000 seconds using 2 threads)

C:\test>888714 -M=100e6 -T=3
1: processing 0 to 33333332
2: processing 33333333 to 66666665
3: processing 66666666 to 100e6
Total = 1314447116 (in 15.878000 seconds using 3 threads)

C:\test>888714 -M=100e6 -T=4
1: processing 0 to 24999999
2: processing 25000000 to 49999999
3: processing 50000000 to 74999999
4: processing 75000000 to 100e6
Total = 1314447116 (in 12.960000 seconds using 4 threads)