Mostly related to making sure that some hot paths get
inlined better, particularly for allocation/sweeping.

These changes make a simple for-loop benchmark (which
is just dependent on allocation speed) about 25% faster.