Mutable references in a concurrency monad.

Deviations: There is no Eq instance for MonadConc the IORef type. Furthermore, the mkWeakIORef function is not provided.

In a concurrent program, IORef operations may appear out-of-order to another thread, depending on the memory model of the underlying processor architecture. For example, on x86 (which uses total store order), loads can move ahead of stores. Consider this example:

iorefs :: MonadConc m => m (Bool, Bool)
iorefs = do
  r1 <- newIORef False
  r2 <- newIORef False

  x <- spawn $ writeIORef r1 True >> readIORef r2
  y <- spawn $ writeIORef r2 True >> readIORef r1

  (,) <$> readMVar x <*> readMVar y

Under a sequentially consistent memory model the possible results are (True, True), (True, False), and (False, True). Under total or partial store order, (False, False) is also a possible result, even though there is no interleaving of the threads which can lead to this.

We can see this by testing with different memory models:

> autocheckWay defaultWay SequentialConsistency relaxed
[pass] Never Deadlocks
[pass] No Exceptions
[fail] Consistent Result
       (False,True) S0---------S1----S0--S2----S0--

       (True,True) S0---------S1-P2----S1---S0---

       (True,False) S0---------S2----S1----S0---
False

> autocheckWay defaultWay TotalStoreOrder  relaxed
[pass] Never Deadlocks
[pass] No Exceptions
[fail] Consistent Result
        (False,True) S0---------S1----S0--S2----S0--

        (False,False) S0---------S1--P2----S1--S0---

        (True,False) S0---------S2----S1----S0---

        (True,True) S0---------S1-C-S2----S1---S0---
False

Traces for non-sequentially-consistent memory models show where writes to IORefs are committed, which makes a write visible to all threads rather than just the one which performed the write. Only writeIORef is broken up into separate write and commit steps, atomicModifyIORef is still atomic and imposes a memory barrier.