Breaking it down further, I'd say what you are really benchmarking is the speed difference between creating a new set of scalars and upping the refcount on an existing set of scalars.

Both move a single scalar, both create and return a list, but one creates new scalars and another must run an array down and return all it's guts after changing each one. Running the benchmark below, I get split losing just barely on linux but the difference varied from 1-5% so I'd say that they are basically equivalent Benchmark wise. I'd say the difference you see is either a poorer split implementation or a variation in how the OS's deal with memory allocation and paging. And, I'd stop worrying about it and start coding the way you like.

#!/usr/bin/perl -w
use strict;
use Benchmark qw(cmpthese);
use vars qw($d $e @t);
$d = <DATA>;
my @e = split " ", $d;
$e= \@e;

sub sss {
        my $q=shift;
        return (split m/ /, $q)
}

sub aaa {my $q=shift; return @$q}

print $d,$/,sss($d),$/,aaa($e),$/;

cmpthese ( -10,
  {
    split=> '@t = sss($d)',
    arref=> '@t = aaa($e)',
  });
__DATA__
a b c d e f g h i j k l m n o p q r s t u v w x y z
[download]

--
$you = new YOU;
honk() if $you->love(perl)