Even after thirty years, the polyhedral model is far from being an unequivocal success, even on the restricted domain where it is applicable. We do not (yet) have compilers for general-purpose processors that can generate code approaching either the machine peak, or the algorithmic peak of the source program, or hand tuned libraries developed by “heroes of high performance computing.” I will try to explain why this is so, and suggest a plan of how we can improve this state of affairs.