You pays your money and takes your choice:^)

#! perl -slw
use strict;
use Benchmark qw[cmpthese];

use constant ARG1 => 0;
use constant ARG2 => 1;
use constant LAST => -1;

sub directly{ return "$_[ARG1] - $_[ARG2]" if $_[LAST]; }
sub shifty{ my ($arg1, $arg2, $arg3) = (shift, shift, shift); return "
+$arg1 - $arg2" if $arg3; }
sub listy{ my ($arg1, $arg2, @others) = @_; return "$arg1 - $arg2" if 
+$others[-1]; }
sub listya{ my (@args) = @_; return "$args[1] - $args[1]" if $args[-1]
+; }

our @a100 = 1..100;

cmpthese( -3, {
    directly3    => 'directly 1, "TWO", 3.0;',
    directly100    => 'directly @a100;',
    shifty3        => 'shifty 1, "TWO", 3.0;',
    shifty100    => 'shifty @a100;',
    listy3        => 'listy 1, "TWO", 3.0;',
    listya        => 'listya 1, "TWO", 3.0',
    listya100    => 'listya @a100;',
});
__END__
Name "main::a100" used only once: possible typo at C:\test\233863.pl l
+ine 14.
Benchmark:
running
 directly100, directly3, listy3, listya, listya100, shifty100, shifty3
, each for at least 3 CPU seconds

directly100:  5 wallclock secs ( 3.19 usr +  0.00 sys =  3.19 CPU) @ 2
+7542.27/s (n=87970)
  directly3:  1 wallclock secs ( 3.03 usr +  0.00 sys =  3.03 CPU) @ 4
+7871.13/s (n=145241)
     listy3:  4 wallclock secs ( 3.10 usr +  0.00 sys =  3.10 CPU) @ 1
+9624.88/s (n=60896)
     listya:  4 wallclock secs ( 3.04 usr +  0.00 sys =  3.04 CPU) @ 1
+9505.42/s (n=59394)
  listya100:  3 wallclock secs ( 3.12 usr +  0.00 sys =  3.12 CPU) @ 3
+179.52/s  (n=9936)
  shifty100:  4 wallclock secs ( 3.04 usr +  0.00 sys =  3.04 CPU) @ 2
+1964.84/s (n=66839)
    shifty3:  5 wallclock secs ( 3.09 usr +  0.00 sys =  3.09 CPU) @ 3
+0639.55/s (n=94523)

               Rate listya100 listya listy3 shifty100 directly100 shif
+ty3 directly3
listya100    3180/s        --   -84%   -84%      -86%        -88%    -
+90%      -93%
listya      19505/s      513%     --    -1%      -11%        -29%    -
+36%      -59%
listy3      19625/s      517%     1%     --      -11%        -29%    -
+36%      -59%
shifty100   21965/s      591%    13%    12%        --        -20%    -
+28%      -54%
directly100 27542/s      766%    41%    40%       25%          --    -
+10%      -42%
shifty3     30640/s      864%    57%    56%       39%         11%     
+ --      -36%
directly3   47871/s     1406%   145%   144%      118%         74%     
+56%        --
[download]

Direct access is hands down fastest, with shift second and list assignment last. Directly accessing 100 args is nearly as fast as shifting 3!;

However, what if any discernable difference it makes will depend very much on how many times your calling the sub, how deeply the call stack grows and how much you are doing in the sub. Calling small subs with many parameters many times benefiting much more than a few calls to subs that do lots of work.