Do really not unroll the vector loop in combination with register tiling

The previous commit lacked a test case for register tiling + pre-vectorization
and we obviously got it immediately wrong.

llvm-svn: 245599