I prefer not to compare apples and oranges:

use Benchmark qw[ cmpthese ];
use List::Util qw[ first ];

our @array = 1 .. 100_000;

cmpthese( -1, {
    func_10th => sub {
        first { length > 3 } @array;
    },
    iter_10th => sub {
        for ( @array ) {
            return $_ if length > 3;
        }
    },
});

cmpthese( -1, {
    func_last => sub {
        first { length > 5 } @array;
    },
    iter_last => sub {
        for ( @array ) {
            return $_ if length > 5;
        }
    },
});
__END__
            Rate func_10th iter_10th
func_10th  718/s        --      -77%
iter_10th 3139/s      337%        --
            Rate iter_last func_last
iter_last 29.4/s        --      -29%
func_last 41.1/s       40%        --
[download]

The results are similar if I run your code though.

Over here, first always wins given a sufficiently large number of iterations.

Strange. Taking your exact code from above and runnning here I get:

P:\test>type 381208-2.pl
#! perl -slw
#use strict;

use Benchmark qw[ cmpthese ];
use List::Util qw[ first ];

our @array = 1 .. 100_000;

cmpthese( -1, {
    func_10th => sub {
        first { length > 3 } @array;
    },
    iter_10th => sub {
        for ( @array ) {
            return $_ if length > 3;
        }
    },
});

cmpthese( -1, {
    func_last => sub {
        $_ = first { length > 5 } @array;
    },
    iter_last => sub {
        for ( @array ) {
            return $_ if length > 5;
        }
    },
});

P:\test>381208-2
            Rate func_10th iter_10th
func_10th 1196/s        --      -61%
iter_10th 3076/s      157%        --
            Rate func_last iter_last
func_last 22.6/s        --      -23%
iter_last 29.3/s       30%        --
[download]

Which indicates something about the difference in OS, but I'm not sure what.

I also wonder what context Benchmark calls the subs in? And is first optimised for the void context case?

---

Examine what is said, not who speaks.

"Efficiency is intelligent laziness." -David Dunham
"Think for yourself!" - Abigail
"Memory, processor, disk in that order on the hardware side. Algorithm, algorithm, algorithm on the code side." - tachyon

Or differences in compilers and flags used to build the versions.

$ perl -v | grep ^This
This is perl, v5.8.3 built for i486-linux
$ perl -V:gccversion
gccversion='3.3.2';
$ perl -V:config_args
config_args='-de -Dprefix=/usr -Dcccdlflags=-fPIC -Dinstallprefix=/usr
+ -Doptimize=-O3 -march=i486 -mcpu=i686 -Darchname=i486-linux';
[download]

The relevant bits for call context are in the runloop() function:

    # ...
    if (ref $c eq 'CODE') {
    $subcode = "sub { for (1 .. $n) { local \$_; package $pack; &\$c; 
+} }";
        $subref  = eval $subcode;
    }
    else {
    $subcode = "sub { for (1 .. $n) { local \$_; package $pack; $c;} }
+";
        $subref  = _doeval($subcode);
    }
    # ...
    &$subref;
[download]

(_doeval is basically a eval.)

So void context, but I don't think that matters here — the compiler doesn't know the context at the time of eval, so shouldn't be optimizing things away. One would have to instrument the code with a few B::Deparse calls to know for sure, of course. But going by the numbers from the benchmark, it doesn't seem to be making the difference.

There just seems to be a high startup overhead for first that amortizes over time, given enough time, for at least some Perls.

Makeshifts last the longest.