Hans Boehm wrote: [... RMW/store-load barrier free (on fast path) release for lock with a queue ...]
Interestingly, there is a 2011 POPL paper by Attiya et al that proves, IIRC, that store-load ordering or atomic RMW operations are unavoidable in some cases. I haven't looked into this enough to understand how to reconcile that with the above result. The kind of scheduler hack needed in step 6 above is probably outside their model.
http://infoscience.epfl.ch/record/161286/files/popl168gf-attiya.pdf (Laws of Order: Expensive Synchronization in Concurrent Algorithms Cannot be Eliminated) Nice title. ;-) I think they are talking about impossibility of RMW/store-load barrier free lock acquisition. RMW/store-load barrier free (on fast path) release for queued lock is out of scope of their paper, AFAICS. regards, alexander.