Wait so it is apparently possible to construct a (lock-free) multiword CAS out of singleword ones?
Is this why GNU libatomic has a hard dependency on CAS primitives? 
@koakuma without memory linear in the number of threads that could be involved?
@robryk I have no idea about memory requirements
@robryk But seems like the schemes I've read so far does involve a per-thread structure yes
@koakuma Actually, I think one can do with just one shared "buffer", at least in the world where you can fit a counter of operations in half of the CAS word -- while filling the buffer, which is composed of <prev version, word> pairs, you CAS its elements so that you start failing if there was an intervening update.
@robryk I wonder if a 2x32b CAS has enough space for the counter...
@koakuma in all of these cases: how fine are you with lock-free unless scheduler has been extremely unfair (in which case it may block)?
@robryk Ideally I'd want something close to a native DWCAS (which I don't think have any fairness guarantee either?)
@robryk Though this is more of a curiosity thing don't take it too seriously lole
@robryk Another thing that I'd like to know is how easy it is to conjure up a lockfree DWCAS out of singleword ones
With the constraint that allocating one time during thread startup is an acceptable thing but having to allocate in the middle of a thread run is a no-no...