The basic problem is, two racing async messages can cause parent/child state to diverge for >1 steps if the two initial diverging states allow parent/child to send messages which might allow state to change again, before the other side's racing message has been received. Currently, the type checker ensures that if sending a racing message could cause the parent to transition to state P, then the parent can only receive messages in state P (and similarly for the child). (|discard| is a partial workaround for this problem.) However, there's an important case in which synchronous messages prevent state divergence; rather than describe it, I'll post a test protocol. The type checker should account for this.

State "S3" is what makes the current IPDL compiler complain: it allows the child to send _B or call _R from a state that is reached if A_ and _R race from state S1. The bug is that if A_ and _R *did* race, then because state S2 forces the parent to wait on the child, and because _R is rpc * parent must receive _R before sending any other message (see state S2) * child must receive A_ while waiting on the reply to _R * child must receive A_ before the next message from the parent This means that the error IPDL reports is too conservative, because child/parent state can't diverge on an A_/_R race.

Going to ride this patch along with this bug, r=trivial.