I think this benchmark code confirms my point regarding the expense of continually creating and invoking anonymous subroutines.

use strict;
use warnings;

use Benchmark qw{ cmpthese };

my @arr = ( 1 .. 1e6 );

my $rcOneElement  = sub
   {
       my $aref = shift;
       sub { my $var = shift }->( $_ ) for @$aref;
   };

my $rcAllElements = sub
   {
       my $aref = shift;
       sub { my $var = shift while @_ }->( @$aref );
   };

cmpthese (
   -5,
   {
       OneElement  => sub { $rcOneElement->( \ @arr ) },
       AllElements => sub { $rcAllElements->( \ @arr ) },
   }
   );
[download]

The output.

              Rate  OneElement AllElements
OneElement  2.59/s          --        -58%
AllElements 6.16/s        138%          --
[download]

However, I am good at cocking up benchmarks so I could be deluding myself :-/