The LLL demand-matcher has two pieces: a `default-task-supervisor`
which does the work of figuring out what to do for a given change in
an instance of demand, and a driver which computes demand instances
and calls `default-task-supervisor` for each.
An `actions` accumulator is threaded through the
`default-task-supervisor` calls. However, the driver code mistakenly
believed that the result of `default-task-supervisor` calls was a
collection of fresh actions only, and therefore consed together the
previous collection of actions with the "new" ones. Because the
`default-task-supervisor` was returning an unmodified accumulator from
time to time, any actions in the accumulator across a call to
`default-task-supervisor` would end up duplicated.
This change favours the thread-the-accumulator-through perspective,
and changes the way the driver uses `new-actions` to match.
This is a major change to the previous design, and also a change with
respect to the semantics in the ESOP 2016 paper. All the complexity of
echo-cancellation is stripped out of the core dataspace semantics, and
the relaying protocol is changed from one constructor, `at-meta`, to
two, `inbound` and `outbound`. The relay connecting a dataspace to its
container is now completely symmetric with the contained actors: it
initially asserts interest in what it is to relay, just like any other
actor would. Dataspaces no longer treat relaying specially.
This commit has updated all (I think) of the non-graphical examples. The
graphical code remains to be done in a following commit.
Tony Garnock-Jones
Sam Caldwell