It might be better to separate the data structure name change (niter_desc to loop_desc) into a different patch. Other than that, the patch looks good to me (for google branches only) Unroller really needs more heuristics like this instead of just looking at size.
David http://codereview.appspot.com/6099055/
