AutoTVM and CPU vectorization: should I split?

If the vectorization is impossible due to hardware constraints, at the worst case it may generate ordinary loop (even if you specify as vectorized in TVM IR)