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.