An alternative optimising solution to the Schwartzian transform suggested by grandfather is the "Orcish Manoeuvre", which also ensures that the operation evaluating the sort key is only applied once per entry, while using the same hash overhead, but often slightly less processing:

my @sorted;
{
    local %_ = ();
    @sorted =
      sort { $_{ $a } ||= substr( $a, 3, 4 )
               cmp
             $_{ $b } ||= substr( $b, 3, 4 )
           } @array;
}
[download]

BUT, in most cases, I tend to simply pass through the data once creating the inverse hash to that:

my %hash = map { substr( $_, 3, 4 ), $_ } @array;
[download]

Such a hash, using the key info as a hash key, tends to be more useful than a sorted array of keys in the subsequent processing.

Note: if the middle field should vary in length, then the expression ( split /\-/ )[1] should replace the substr function.