Flow analysis and optimization with computed gotos

[Joern Rennecke]

> We'd need modifications to the gcse/lcm code to allow it to work
> on subgraphs (which would likely reduce max memory footprint at
> the same time).  We'd also need code to appropriately partition
> the CFG; presumably there are well-known algorithms for doing this.

If we handle the abnormal edges as killing all expressions, this will
automatically partition the graph into as many subgraphs as necessary.
We can still pretend to work on the entire graph, and thus get all
subgraphs.