Oops, thank you. I missed that one because I downloaded the solutions by crawling the displaycode links on the thread page and he didn't put his entry in CODE tags.

It works correctly and therefor qualifies, but I only get about 340 iterations/s for it. That puts it next to Limbic~Region's entry in my chart. I'm not surprised, as his code is quite complex and involves a true for(;;) loop (relatively slow in Perl). Fast code in Perl means as few opcodes as possible and letting builtins do work implicitly as much as possible. (See GRT, f.ex.)

That is how I arrived at my second version. I copied the first version, ripped everything except the XOR out of the loop, and started benchmarking them against each other as I tried to accelerate extraction of null runs from the bitmask and rotation of the reverse string. Every single operation I added to the (non-functional, skeletal) second version had a dramatic impact on speed. Whatever I did, I found nothing with which to improve upon while( /\0{3,}/g ) { } and substr+chop. I only managed to get a speedup when I constrained the match further so that the regex engine isn't exited to drop into the loop body for null runs that are too short to be candidates. Apparently, avoiding that penalty by skipping matches implicitly more than makes up for the additional cost of having to compile the regex multiple times, which skipping them explicitly didn't require.

The version and compile-time flags of the Perl in use probably matter to some degree, as well.