Modifying jeroenes's modification of davorg's modified benchmark (threw in tye's neat trick for good measure):

           Rate    join    argv  hybrid linesep sysread
join     5459/s      --    -44%    -71%    -81%    -91%
argv     9690/s     78%      --    -49%    -67%    -85%
hybrid  19084/s    250%     97%      --    -34%    -70%
linesep 29070/s    433%    200%     52%      --    -55%
sysread 64103/s   1074%    562%    236%    121%      --
[download]

Nice, dkubb! I think we have a winner. Not only is _sysread() clearly the fastest, but it burns the least CPU*. I like it! Hopefully there are no "catches."

update: _argv() is almost as fast as _linesep() when working with large files. _join() and _hybrid() tend to fall behind.

* Linux 2.4, Celeron-366, Perl 5.6.0

---

#!/usr/bin/perl -w

use strict;
use Benchmark qw(cmpthese);

my $file = $0;
open(IN, $file) or die "$file: $!\n";

cmpthese(50_000, {
    join     => \&_join,
    linesep  => \&_linesep,
    hybrid   => \&_hybrid,
    sysread  => \&_sysread,
    argv     => \&_argv
});

close(IN);

sub _join { 
    seek(IN, 0, 0); my $content = join '', <IN>; 
}

sub _linesep { 
    seek(IN, 0, 0); my $content = do { local $/; <IN> } 
}

sub _hybrid { 
    seek(IN, 0, 0); my $content = do { local $/; join '', <IN> } 
}

sub _sysread { 
    seek(IN, 0, 0); sysread IN, my $content, -s IN 
}

sub _argv { 
    my $content = 
        do { local(*ARGV, $/); @ARGV = ($file); <> } 
}
[download]