The goal of this is to improve compile and load times.
It also would be nice for libraries wanting to take on smaller dependencies.
Additionally, LoopVectorization itself might no longer depend on llvmcall at all in the future, but may still want other parts.
The pieces: