At one point I used an actual code attribute to register subroutines for inlining; the attribute sub itself did some processing of a B::Deparse'd version of the code (gotten with coderef2text or something) and then saved some metadata to a hash. Later on, I'd dump the entire thing to a file on an INIT block.

I've deprecated that system already for unrelated reasons, but it was pretty nice for a lazy way of automatically gathering stuff from files. I found attributes themselves to be useful most of the time, my gripe is purely syntactical.