Why not OpenMP?

Some people wonder if we should spend any effort on dependence analysis, suggesting we might instead rely on pragmas alone. I think such an approach is equivalent to using OpenMP. It's obviously possible, but the results are a bit limited. There's no support for parallelizing recurrences and reductions, and a number of transformations (the ones enabled by dependence analysis) aren't supported. We want to support systems with thousands of threads; OpenMP seems more suitable for systems with dozens of threads.

Expand on this a lot. Focus on benefits of dependence analysis, more than just parallelization.