I have verified that the optimized assembly for Vec<3> and Vec<4> are consistent with what we generated before.