The benchmark above shows the cost of the "my" in the loop. This compares both lexical approaches with the package variable (aka "global") approach:

my $foo;
our $foo2;

cmpthese(shift, {
        'test_1' => sub { $foo = 1; },
        'test_2' => sub { my $bar = 1; },
        'test_3' => sub { $foo2 = 1; },
    });

__END__
            Rate test_2 test_1 test_3
test_2 3731343/s     --   -30%   -37%
test_1 5347594/s    43%     --   -10%
test_3 5952381/s    60%    11%     --
[download]

But the apparent gaps diminish rapidly as you start doing even a little work, particularly if system interaction is involved:

my $foo;
our $foo2;

cmpthese(shift, {
        'test_1' => sub { $foo = time; $foo += 1 if __PACKAGE__ eq 'ma
+in' },
        'test_2' => sub { my $bar = time; $bar += 1 if __PACKAGE__ eq 
+'main' },
        'test_3' => sub { $foo2 = time; $foo2 += 1 if __PACKAGE__ eq '
+main' },
    });

__END__
           Rate test_2 test_3 test_1
test_2 442478/s     --    -4%   -10%
test_3 462963/s     5%     --    -6%
test_1 492611/s    11%     6%     --
[download]

Interesting that the order reverses, thought it might just margin of error