We don't optimize but emit straight C-code summing each element in
turn, so that it will work with any arithmetic C type down the road.