I think you are making a mistake. The surprising hypothesis you are testing is whether the access to a literal hash ref

{ a => 'b' }->{$k};
[download]

is slower than saving the hash ref in a var and accessing

my $hash_ref = { a => 'b' };
# and then
$hash_ref->{$k}
[download]

And your benchmark suggested it is so indeed. I guess you're wondering: but there should be an overhead to access the variable, so that using a var should be slower. But the big problem with your benchmark code is that the literal hash ref is being constructed inside the loop over and over, and that causes an overhead greater than the one of acessing the variable.

I think the hypothesis "access to a literal hash ref is slower than access via a hash ref saved in a variable" is not true. The following code adds a more fair alternative with a constant to save the literal (so that is not built over and over inside the loop) and you'll see that this is the fastest option.

Another alternative using a lexical variable (my) instead of package variable (our) was introduced as well. But there is no visible performance difference in this case.

use strict;
use warnings;

use Benchmark qw(cmpthese);

our $global_hash = { 'a' => 'A', 'b' => 'B'};
my $lexical_hash = { 'a' => 'A', 'b' => 'B'};
use constant CONST_HASH => { 'a' => 'A', 'b' => 'B'};

cmpthese(-2, {
    'hash' => sub { for my $x (qw(a b a b a b)) { my $y = { 'a' => 'A'
+, 'b' => 'B'}->{$x}; }},
    'global_ref' => sub { for my $x (qw(a b a b a b)) { my $y = $globa
+l_hash->{$x}; }},
    'lexical_ref' => sub { for my $x (qw(a b a b a b)) { my $y = $lexi
+cal_hash->{$x}; }},
    'const' => sub { for my $x (qw(a b a b a b)) { my $y = CONST_HASH(
+)->{$x}; }},
});
[download]

and a sample result:

                Rate        hash  global_ref lexical_ref       const
hash         27571/s          --        -73%        -74%        -76%
global_ref  102819/s        273%          --         -3%        -10%
lexical_ref 105720/s        283%          3%          --         -8%
const       114361/s        315%         11%          8%          --
[download]