I'm looking for shared concurrent vectors in Haskell, so I would accept an answer that points to a performant solution other than what I'm discussing here, which is to combine the vector library with stm or atomic-primops.
I have two threads sharing a mutable vector (Data.Vector.Mutable.IOVector) and some flag that can be switched once atomically (e.g., TVar with stm or IORef with atomic-primops). Thread 1 may write to the vector (nonatomically) before or after switching the flag atomically, and Thread 2 may read from the vector, before or after checking whether the flag has been switched.
For concreteness, here are some minimal examples. Is there a data race that causes them to evaluate undefined?
- Example using STM: http://lpaste.net/362596
- Example using atomic-primops: http://lpaste.net/362595
More generally, do the relative ordering of the atomic operations on the flag imply some ordering between operations on the vector? In particular, I'm wondering whether the following is true:
- If the check happens before the switch, then reads preceding the check cannot see writes following the switch.
- If the check happens after the switch, then reads following the check must see writes preceding the switch.
In the case of atomic-primops I particularly have doubts about using readIORef/readForCAS, since the docs say it does not imply any barriers.