I'm recently been doing work to tune the way that LLVM uses (and abuses) memory. In the past, we typically choose to use a std::vector, std::set, or std::map by default, with little thought of the consequences. Because the options are many and the trade-offs are complex, I've written up a guide on how to pick the right vector-like, set-like, or map-like container for a job.
I'd appreciate it if everyone making data structure choices in LLVM would at least skim through it -- I suspect most people aren't aware of all the options we have:
http://llvm.org/docs/ProgrammersManual.html#datastructure
Comments and feedback are welcome, of course :).
-Chris