Ryan Ingram wrote:
Heinrich Apfelmus wrote:
However, even in a demand-driven implementation, there is one optimization that I would like make: when there are multiple external events, say e1 and e2, the network splits into subnetworks that react only to one of the inputs. For instance, your example would split into two graphs
e1 e2 | \ | \ e3 e4 and e3 e4 | | | | e5 e5 e5 e5
that are independent of each other. Unlike successful filters, these subnetworks are known *statically* and it's worth splitting them out.
Yeah, I realize that as well, although you can get the same problem with a single source, it just makes the network a bit more complicated:
e0 = source e1 = fromLeft <$> filter isLeft e1 e2 = fromRight <$> filter isRight e1 -- rest of network the same
Anyways, the problem I was getting at is that lets say that e1 and e2 are both Event Bool, and e1 has a True event at the same time that e2 has a False event.
Then a behavior derived from e3 is False for that time (assuming behaviors take the 'last' event in the list?), and a behavior from e4 is True for that time.
Yep, that's precisely what will happen. Internally, the four "pillars" in the graph e1 e2 | \-A | \ A-\ e3 e3 e4 e4 | | | | e5 e5 e5 e5 1 2 3 4 -- pillar will simply be executed from left to right (in particular not in depth-first search order). The edge connecting e1 and e4 signifies that the value of e1 will be cached when the first pillar is executed, so that it is available again a few pillars later when it's time for the fourth pillar. Concerning the behaviors, also note that a behaviors change "slightly" later than the events from which they are derived. Semantically, we have stepper x ex = \time -> last $ x : [y | (time', y) <- ex, time' < time] In particular, the strict comparison < means that the behavior still has the previous value when the event happens. So, indeed, a behavior derived from e3 will pick up the last False while a behavior derived from e4 will pick up the last True. Best regards, Heinrich Apfelmus -- http://apfelmus.nfshost.com