Fix for PR929. The PHI nodes were being gone through for each instruction

Fix for PR929. The PHI nodes were being gone through for each instruction
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.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30697 91177308-0d34-0410-b5e6-96231b3b80d8