Relative degree and nonminimum phase difficulties limit the applicability of input-output feedback linearization; hence the need for approximations. Recent work on predictive control of bilinear systems overcame these problems by means of interpolation between feedback linearization and state feedback, the former providing optimality and the latter guaranteeing feasibility and stability through the use of invariant/feasible polytopes. The current work also makes use of polytopes in preference to ellipsoids but achieves distinctly different objectives. First, it is shown that feedback linearization can be used over particular polytopes without needing to resort to either approximation or interpolation. Then, it is shown that invariant polytopes based on bilinear controllers can be much larger. These two approaches are combined in an algorithm that guarantees stability over much larger initial condition sets and gives much improved closed-loop performance.