in a successor block for every block...resulting in some O(N^k) algorithm
which wasn't very good for performance. Calculating this information up
front and keeping it in a map made it much faster.

llvm-svn: 30697