Vectorised processing is a win according to that paper: it is faster than scalar code. The key point of the paper is that compiled queries---doing loop fusion---is sometimes more of a win (in a database context, where "vectorisation" doesn't always mean SIMD as in this discussion).

Doing fused SIMD-vectorised operations will likely bring the advantages of both, with relatively small downsides. This is a relatively common technique for libraries like Eigen (in C++), that batch a series of operations via "Expression Templates" and then execute them all in as a single sequence, using SIMD when appropriate. (Other examples are C++ ranges and Rust iterators, although these are focused on ensuring loop fusion and any vectorisation is compiler autovectorisation... but they are written with that in mind: effort is put into helping the building blocks vectorise.)