In this talk, I will use our experience with a complete scientific application---three-dimensional, nonuniform, iterative Magnetic Resonance image reconstruction---to discuss the challenges in developing a declarative programming language which nevertheless exposes aspects of control flow to the user to enable high levels of optimization. In particular, I will explain how we have had to modify our thinking, to present a more conventional view of the multi-loop first proposed at CASCON 2007 than the original view which was much closer to the hypergraph intermediate language used in Coconut, and how that led to modifications to our scheduling algorithm Explicitly Staged Software Pipelining.