On modern architectures, a missed optimization can translate into performance degradations reaching orders of magnitude. More than ever, translating Moore's law into actual performance improvements depends on the effectiveness of the compiler. Moreover, missing an optimization and putting the blame on the programmer is not a viable strategy: we must strive for portability of performance or the majority of the software industry will see no benefit in future many-core processors. As a consequence, an optimizing compiler must also be a parallelizing one; it must take care of the memory hierarchy and of (re)partitioning computation to best suit the target architecture Polyhedral compilation is a program optimization and parallelization framework capable of expressing extremely complex transformation sequences. The ability to build and traverse a tractable search space of such transformations remains challenging, and existing model-based heuristics can easily be beaten in identifying profitable parallelism/locality trade-offs. We propose a hybrid iterative and model-driven algorithm for automatic tiling, fusion, distribution and parallelization of programs in the polyhedral model. Our experiments demonstrate the effectiveness of this approach, both in obtaining solid performance improvements over existing auto-parallelizing compilers, and in achieving portability of performance on various modern multi-core architectures.