cleans up all the chains allocated during the processing of each
function so that for very large inputs we don't just grow memory usage
without bound.

llvm-svn: 144533