Use a Bitvector instead, we didn't need the smaller memory footprint anyway.
This makes the greedy register allocator 10% faster.

llvm-svn: 129390