mm_mfence: wait for all me[mory] operations to complete,
some reference here
pagerank
It is interesting to see how abstraction and implementation are related here. The more accurate we express the synchronization needs (fewer constraint while keeping the computation correct), there will be more space for the compiler and hardware to do the optimization. Different grammars have different expression ability, and the best one should be the one can express the needs the computation exactly without any unnecessary constraints.
How are these fences implemented in hardware?
mm_mfence: wait for all me[mory] operations to complete, some reference here
It is interesting to see how abstraction and implementation are related here. The more accurate we express the synchronization needs (fewer constraint while keeping the computation correct), there will be more space for the compiler and hardware to do the optimization. Different grammars have different expression ability, and the best one should be the one can express the needs the computation exactly without any unnecessary constraints.