http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49749
--- Comment #6 from William J. Schmidt <wschmidt at gcc dot gnu.org> 2011-07-15 20:15:09 UTC --- We ran some experiments attempting to restore the r161839 behavior, either by lowering the rank of memory references or raising the rank of phi references. Although both experiments restored bwaves code generation to respectability, several other benchmarks suffered, so this is not a general answer. Any thoughts on more limited heuristics that might help the bwaves scenario? It seems that copyrename needs help from reassoc to find the opportunity; perhaps reassoc can be tweaked to recognize situations that will help copyrename downstream?
