Three memory models will eventually be supported:
- SequentialConsistency, which is the current behaviour.
- TotalStoreOrder, where each thread has a write buffer.
- PartialStoreOrder, where each CRef has a write buffer.
The parameter is currently ignored.
A lot of the places where lists were used, there was an assumed invariant that
there would be no duplicate entries, for some criteria of equality. In some
cases this was enforced by `nub`, in others not. Sets are a much better choice
here, as they actually enforce the invariant, with better complexity for some
operations.
Adds more information to traces with a new Trace' type, which includes the
*sequence* of actions a thread will perform next, and use that for blocking
lookahead. This allows skipping over things like `_concKnowsAbout`, and so
brings the analysis behaviour of `spawn` in-line with `fork`.
For the test cases, this further reduces the average number of runs to 23.
If a decision will immediately block without changing the global state, then
there is no point in making it: no state will become reachable from it which
isn't reachable through some other option we have available.
This has three parts:
- When the prefix runs out and the scheduler is making decisions, it filters
out decisions which will immediately block.
- When a subtree is being added, it records which decisions will immediately
block.
- When backtracking points are being added, it filters out ones in this block
list.
This optimisation is likely to only be useful when threads are communicating a
lot. For instance, a `parMap id` is totally unaffected by this, but the test
cases drop from an average of 64 runs to 42.
This further improves performance in all cases, although it's still far worse
on some of the included tests, and I have yet to figure out why.
See also: "Partial-Order Methods for the Verication of Concurrent Systems"
[Godefroid 1996]
This performs better with "real" code (the Par monad) but surprisingly does far
worse with the included tests! The next thing to do is implement the orthogonal
sleep sets algorithm to cut down on available choices even further and
hopefully correct this issue.
See also: "Bounded Partial-Order Reduction" [Coons, Musuvathi, McKinley 2013]
