This moves the definition of getContainer into a single root-def, avoiding some spuriously materialized negation predicates otherwise needed by the dispatch predicate, and ensuring that the whole relation is cached.

I noticed this predicate getting a bit flaky and sometimes very expensive in response to unrelated changes. In particular, magic started pushing "the set of control-flow nodes that belong to a container" as the context, which is a pretty bad context to use anywhere.

I believe the complicated overriding pattern may have been at fault here:

1. ControlFlowNode.getContainer explicitly referred to this.(Expr).getContainer() even though getContainer() is overridden in Expr, so you'd think the code was dead. Likewise for Stmt. (Ignore, this was just me not paying attention to the class hierarchy.)

2. Due to multiple inheritance, there were several overrides that exist only to choose arbitrarily between two inherited definitions of getContainer, which were equivalent but not obviously so.

3. There wasn't a single root-def for getContainer. For example ClassOrInterface.getContainer did not override ControlFlowNode.getContainer. Not all these root-defs were cached.

There are also a few follow-up performance fixes which are independently useful on their own, but needed to make this safe to land.

Evaluations:

• XSS/smoke-test show a faster start-up time
• Security/nightly with re-run of slowest is leaning on positive