When introducing reductions into Polly we thought it would be a good idea to model reduction dependences using a "convex_hull" computation. However, this is both computationally a lot more complex than expected and also conflicts made it hard to compute reduction dependences on schedule trees. Fortunately, Sven Verdoolaege suggested (at least parts) of what looks like a good solution. Instead of first computing precise dependences and _then_ reintroducing reduction dependences, we could just model reduction statements as may-writes which will make the dependence analysis automatically introduce the additional dependences needed when removing the inter-reduction-operation dependences. This blocks http://reviews.llvm.org/D5190 This is just a brief reminder. Please ping back in case you need more information.