1. First off, i43s's code fails to compile for me using perl 5.6.0 with -w, with the following:

   Warning: Use of "shift" without parens is ambiguous
     at benchadd.pl line 3.
   Type of arg 1 to shift must be array (not shift)
     at benchadd.pl line 3, near "shift,"
   [download]

   This is fixed by adding explicit parenthesis after every shift().

2. Second, mdillon's code as is returns undef, because the "my $sum" is part of the loop (I think that's the reason, at least). Separating the declaration of the variable from the suming loop fixes the problem.
3. Third, I have benchmarked th three solutions proposed so far (i43s, davorg and mdillon) using the code below. The results show:

   Benchmark: timing 100000 iterations of davorg, i43s, mdillon...
       davorg: 70 wallclock secs (55.52 usr +  0.06 sys = 
           55.58 CPU) @ 1799.21/s (n=100000)
         i43s: 20 wallclock secs (15.47 usr +  0.00 sys = 
           15.47 CPU) @ 6464.12/s (n=100000)
      mdillon:  8 wallclock secs ( 6.77 usr +  0.00 sys =  
           6.77 CPU) @ 14771.05/s (n=100000)
   [download]

   So my comments are:
   • davorg: nice trick with eval, but eval is just too slow.
   • i43s: nice recursive subroutine, but recursion is also expensive, and IMHO it complicates the code in this case.
   • mdillon: that's how I would have done it too :-)

--ZZamboni

The benchmarking code:

use Benchmark;

sub add_i43s { @_ > 1 ? add_i43s( shift() + shift(), @_ ) : shift() }
sub add_davorg { eval join '+', @_ };
sub add_mdillon { my $sum; $sum += $_ for @_; $sum }

timethese(100000, {
        'i43s' => sub { add_i43s(1,2,3,4,5,6,7,8,9,10); },
        'davorg' => sub { add_davorg(1,2,3,4,5,6,7,8,9,10); },
        'mdillon' => sub { add_mdillon(1,2,3,4,5,6,7,8,9,10); } });
[download]