The diffs suggest that we are missing some more basic
analysis/transforms, but this keeps the vector path in
sync with the scalar (rL374397). This is again a
preliminary step for introducing the reverse transform
in IR as proposed in D63382.
llvm-svn: 374555