http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48636
--- Comment #27 from Jan Hubicka <hubicka at gcc dot gnu.org> 2012-10-20 10:34:58 UTC --- Thank you for testing. It seems that the patch works well for small benchmarks, I will look into lapack/test_fpu slowdown. There is problem that it really causes inacceptable growth on SPEC2k6 and 2k in non-LTO mode. I will need to analyze some of these testcases and see why we predict so much of speedup when there are no benefits in runtime. Jakub: the plan is to make ipa-cp to handle propagation across aggregates in general (jump functions are already in place), that will handle the array descriptors, too. The fatigue is however different case - the values happens to be loop invariant of the outer loop the function is called from. So inlining enables a lot of invariant code motion. This is similar to cray. Both these cases are now understood by ipa-inline-analysis but the fact is not really used w/o this patch.
