The patch http://reviews.llvm.org/D8943 is committed as r251592 (sorry for forgetting to include the differential revision in the commit message). This patch added a flag that enables calculating vectorization factor (VF) using the smallest type instead of the widest one. Due to imprecise cost model, the largest VF could not be used with larger cost than smaller VFs. I will compose more patches later to adjust the cost model. Here I am asking people to test this patch by turning on the flag vectorizer-maximize-bandwidth.