I wonder what it takes to trigger that optimization.

davido:~/scripts$ perlbrew exec --with perl-5.10.1 perl mytest.pl
Perl version: 5.010001

          Rate retn bare
retn 6849315/s   --  -8%
bare 7462687/s   9%   --

(25000000, 25000000)


davido:~/scripts$ perlbrew exec --with perl-5.28.0 perl mytest.pl
Perl version: 5.028000

           Rate bare retn
bare 10729614/s   --  -3%
retn 11013216/s   3%   --

(25000000, 25000000)
[download]

I mean I do see a difference, but other optimizations between Perl 5.10 and Perl 5.28 seem to be far more significant (probably the integer optimization).

Here is the sample code:

use strict;                      
use warnings;                    
                                 
use Benchmark qw(cmpthese);      
                                 
my $x;                           
my $y;                           
                                 
sub bare { ++$x; $x }            
sub retn { ++$y; return $y }     
                                 
print "Perl version: $]\n\n";    
                                 
cmpthese(25_000_000, {           
    bare => sub {my $t = bare()},
    retn => sub {my $t = retn()},
});                              
                                 
print "\n($x, $y)\n";
[download]

The reason for the double sub call is to assure that the return value is being obtained, and not discarded due to context. So we do have the double-call overhead, but it's the same overhead for both sides of the test.

Anyway, where there had been a 9% difference in favor of bare returns in 5.10, there's now a statistically insignificant 3% difference that seems to go either direction.