That is odd, because my benchmark indicates the opposite. Here is some benchmark code calling the two version with longer strings and more cuts each time. The results (shown at the end, slightly reformatted: perl 5.6.1) indicate a fairly predictable increase in efficiency of the unpack() method versus the substr() method:

#!/usr/bin/perl -w
use strict;
use Benchmark;
my @args = ('ACGTACGTACGT', 2,4,5);

timethese(-2,{
    japhy  => sub{japhy(@args)},
    danger => sub{danger(@args)},
});

@args = ('ACGTACGTACGT' x 4, (2,4,5) x 4);

timethese(-2,{
    japhy  => sub{japhy(@args)},
    danger => sub{danger(@args)},
});

@args = ('ACGTACGTACGT' x 100, (2,4,5) x 100);

timethese(-2,{
    japhy  => sub{japhy(@args)},
    danger => sub{danger(@args)},
});

sub danger {
    my $s = shift;
    join '-', unpack join('A','',@_,'*') ,$s;
}

sub japhy {
    my $s = shift;
    join '-', map(substr($s, 0, $_, ''), @_),$s;
}

__END__

Benchmark: running danger, japhy, each for at least 2 CPU seconds...
    danger:  1 wallclock secs ( 2.06 usr +  0.00 sys =  2.06 CPU) 
             @ 10438.35/s (n=21503)
     japhy:  1 wallclock secs ( 2.06 usr +  0.00 sys =  2.06 CPU) 
             @ 8694.17/s (n=17910)

Benchmark: running danger, japhy, each for at least 2 CPU seconds...
    danger:  2 wallclock secs ( 2.03 usr +  0.00 sys =  2.03 CPU) 
             @ 4800.49/s (n=9745)
     japhy:  3 wallclock secs ( 2.10 usr +  0.00 sys =  2.10 CPU) 
             @ 3409.05/s (n=7159)

Benchmark: running danger, japhy, each for at least 2 CPU seconds...
    danger:  2 wallclock secs ( 2.13 usr +  0.00 sys =  2.13 CPU) 
             @ 258.22/s (n=550)
     japhy:  3 wallclock secs ( 2.01 usr +  0.00 sys =  2.01 CPU) 
             @ 155.72/s (n=313)
[download]

Of course, it is entirely possible that I've completely messed up the benchmark.