dart2js: remove side effect builder call graph

[sigmundch]

During global inference we are effectively building 2 call-graph representations at once. One in the typegraphnode representation, the other with the SideEfffectBuilder objects.

Our hope is that we can combine the two:

• remove SideEffectBuilders altogether
• store a SideEffect object on each member
• propagate effects through the callgraph at the end of inference.

Doing so might also let us use a canonical instance for each side-effect in the lattice (it's not that big!).

Note: this has to be done carefully if we reintroduce indirection in type-inference. Up until a92a9f1ba21f7fff440f1197aaf120f5f52c9332 we were using indirection for all == methods during inference. That was OK for computing types, but would have been a regression to use the same graph to compute the side-effects.

When using indirection, we want to make sure it is not used globally for all calls, but only for calls that have a large number of possible targets. Then simple calls will likely not be polluted with unnecessary effects.

[fishythefish]

To clarify, we want to separate the type graph and call graph representations. Then we should be able to store a SideEffect object on each member in the call graph and omit it entirely from the type graph, right?

[sigmundch]

... separate type graph from call graph representations.

Correct!

... store a SideEffect object on each member in the call graph ... ?

Yes, but there is a representation question here to be considered: should it be on the call graph itself or on the side? Depending on our representation, we may end up with a callgraph that contains no data, or contains an opaque storage for data. That way it can be agnostic of the analysis it will be used for (whether it is side-effects, type graphs, or something else in the future).