Why is it necessary that the operations must be made on non-overlapping addresses if the processor is going to hold the whole cache line until it is done?
bpr
@ant, this condition comes from Intel's manuals on x86. There is no guarantee that Intel is doing these operations on entire cache lines. They may make an optimization for atomic operations with different starting addresses, such that non-overlapping operations no longer have false sharing.
If you make them on overlapping addresses, the operations may or may not work.
Why is it necessary that the operations must be made on non-overlapping addresses if the processor is going to hold the whole cache line until it is done?
@ant, this condition comes from Intel's manuals on x86. There is no guarantee that Intel is doing these operations on entire cache lines. They may make an optimization for atomic operations with different starting addresses, such that non-overlapping operations no longer have false sharing.
If you make them on overlapping addresses, the operations may or may not work.