With the rise of multi-core processors, researchers are taking a new look at extending the applicability auto-parallelization techniques. In this paper, we identify a dependence pattern on which autoparallelization currently fails. This dependence pattern occurs for ordered sections, i.e. code fragments in a loop that must be executed atomically and in original program order. We discuss why these ordered sections prohibit current auto-parallelizers from working and we present a technique to deal with them. We experimentally demonstrate the efficacy of the technique, yielding significant overall program speedups.