I'm using Perl 5.005_03 by Active State and my results clearly match what eduardo is getting. In fact, to make the test more of a fair match, I realized that I should probably declare $c outside of the "traditional" sub (though I acknowledge that in the real world, a temp variable should not be declared like that.) When I did that to $c, the traditional swap was even faster than the Perl swap.

I can't help but wonder if japhy's results stem from a 5.6 optimization that's not found in older versions of $^X.