predecessors before returning.  Otherwise, if multiple predecessor edges need
splitting, we only get one of them per iteration.  This makes a small but
measurable compile time improvement with -enable-full-load-pre.

llvm-svn: 97521