Unfortunately, I don't think you'll get any great gains because of how arguments work. @_ will still have to be created for the inlined function, and I imagine that's where most of the time is spent.

Reults of my fairest approximation:

                  Rate          copy        nocopy   inline_copy inlin
+e_nocopy
copy          534208/s            --           -9%          -13%      
+    -23%
nocopy        587704/s           10%            --           -4%      
+    -16%
inline_copy   610942/s           14%            4%            --      
+    -12%
inline_nocopy 697762/s           31%           19%           14%      
+      --
[download]

use strict;
use warnings;

use Benchmark qw( cmpthese );

sub attrib_copy   { my ($self) = @_; $self->{attrib} }
sub attrib_nocopy {                  $_[0]->{attrib} }

our $self = bless({ attrib => 'some_value' });

my %tests = (
    copy          => 'my $x = do { local @_ = $self; &main::attrib_cop
+y };',
    nocopy        => 'my $x = do { local @_ = $self; &main::attrib_noc
+opy };',
    inline_copy   => 'my $x = do { local @_ = $self; my ($self) = @_; 
+$self->{attrib} };',
    inline_nocopy => 'my $x = do { local @_ = $self;                  
+$_[0]->{attrib} };',
);

$_ = 'use strict; use warnings; our $self; ' . $_
    for values %tests;

cmpthese(-3, \%tests);
[download]

See how you can almost get the same benefit from not copying the @_ into local vars?