This test case runs a few odd control flow graphs through the optimizer.

This test compares emitted code against a known-good (via eyeballing)
version, so it is possible for this test to spuriously fail if other
aspects of conversion end up altering the output in innocuous ways.