The new 'concurrency' package is starting at version 1.0.0.0 because the
API is already very mature (copied from base).
This breaks the dejafu-0.2 compatibility of async-dejafu.
Closes#51.
The TH boilerplate was really fragile, and by dropping lineNum as well
the entire template-haskell dependency can be dropped. This is a small
sacrifice, as the name needed to be passed in explicitly anyway, and
there are probably more meaningful names than just the line number.
Removing the stuff broke some of the litmus tests, which is bad. It
probably means that those actions were being put into sleep sets, and so
hiding actually interesting interleavings from the POR implementation. I
need to improve the lookahead behaviour to ignore these invisible
actions.
Closes#46.
The final write-to-a-reference is now in the Stop action. This is a
fairly large change as it was easiest to do this by also removing the
Ref and Fixed types, and going for a typeclass for monadic references.
Fixes#43.
Only consider STM transactions dependent if there is overlap in the
TVars they reference. Lookahead is the same worst-case behaviour as
before, as that information isn't available then.
Keep track of thread masking states and only consider a 'throwTo' as
dependent if the action being thrown at can be interrupted. This halves
the execution time of the testsuite on my linode (~190s to ~90s).
Requires dpor-0.2, as thread IDs must be known when stepping the state.
Closes#32, although I feel like there might be yet more gains possible.
-- unconfirmed hunch follows --
I feel like there might be a bug lurking here, but I haven't been able
to exhibit one.
My intuition is that because throwing an exception which can't interrupt
WILL block and so enable other threads, this might have issues in
combination with preemption bounding, although the testsuite still
passes.
