Nice one jsprat++!

I spent the last hour playing with your benchmark trying to redeem something from whatever brainfart it was that caused me to post without benchmarking first. I failed.

I tried varying the size of the array and the size of the elements and the only time I saw any benefit from the non-hash version was when the size of array * size of elements pushed the boundaries of my installed memory. The hash takes a fair amount more memory than the composite string, so the non-hash version survives a while longer at the pathelogical extremes.

I managed a marginal improvement to your improved version by doing away with interpolation all together vis.

sub improved {
    my $c = ord(1);
    my $temp = join $c, @array2;
    @array1 = grep { -1==index($temp, $c . $_ . $c ) } @array1;
}
[download]

I also saw an equally marginal improvement (~2%) in the faq version by using slice assignment rather than a for loop, but I'm pretty sure that this has been noted elsewhere.

sub perlfaq {
  my %hash;
  @hash{@array2} = ((undef) x @array2);
  @array1 = grep { not exists $hash{$_} } @array1;
}
[download]

Either is at best a micro-optomisation. When I thought about the fact that index has to search the whole string every time, it's fairly obvious that the hash will win.