Playing more than understanding. Using the same benchmarks I got about a 3-5% speed up with a modified version of blazar's which skips lexical assignment by using $_. Tried a few arrangements with length and local and this was the fastest I stumbled on. OS X, v5.8.2.

sub compare_blazar_mod {
    ( $_ = $_[0] ^ $_[1] ) =~ tr/\0//d;
    2 < length;
}
[download]