Get the latest tech news
Making Sense of Acquire-Release Semantics
February 2023 Multiprocessor Synchronization was one of my favorite classes during my undergrad — it had a clear progression from theory to practice, starting at the theory of consensus numbers and moving onto atomic operations and then synchronization primitives and lock-free data structures, all from first principles. With careful thought and a little intuition-bending, every problem could be stated clearly and solved in a neat, orderly way … in Java.
So far the CPU has come up with two different ways to break our code, and both were the kinds of subtle, impossible-to-repro problems that can slip through review and testing and then run roughshod through production. Past this point, the consumer can pick up the item off the queue at any time, so we need to be certain this thread isn’t still running code that might be using the shared memory, in this method or in the calling context. This puts us in a kind of funny situation: according to the stdatomic docs, and the C++ spec, our original queue is technically wrong, but on some pretty popular CPUs, it’ll work just fine anyways.
Or read this on Hacker News
