On Tue, 2016-11-15 at 09:06 +0100, Gabriel Paubert wrote: > > I'm really wondering whether this is the kind of thing that would > benefit from transactions. But transactional memory was only implemented > on Power7 and later IIRC, and we still need to support machines without > transactional memory.
Power8 and later. Also this is about *local* atomics not general atomics, so the overhead of starting a transaction in that case is way way higher than the proposed implementation. Cheers, Ben.
