Make use of the exact bit when optimizing '(X >>exact 3) << 1' to eliminate the

'and' that would zero out the trailing bits, and to produce an exact shift
ourselves.

llvm-svn: 147391