Bunch up all locally used allocas by the block they are allocated in, and

process them all as a group.  This speeds up SRoA/mem2reg from 28.46s to
0.62s on the testcase from PR209.

llvm-svn: 11100