I've been optimizing selector matching, but haven't started in too much on the cascade. The myspace testcase is pretty light on the cascade, whereas the obama testcase exercises it more. I did some profiling today and found a bunch of stuff we can do here.
I did my usual trick of hacking up a patch to make some stuff out of line: https://github.com/bholley/gecko/tree/out_of_line_cascade That gives us this profile here: https://perfht.ml/2oU1fRY , which shows us spending 35ms in the cascade. Gecko spends about 15-18ms in WalkRuleTree for the same testcase.