http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58721
--- Comment #10 from Jan Hubicka <hubicka at gcc dot gnu.org> --- OK, seems that the problem is with Fortran generating internally __builtin_expect to control various construct. I would make a lot more sense to use GIMPLE_PREDICT for those cases. With GIMPLE_PREDICT one can have more fine grained control over the hint and handle it as separate predicter. I wonder if some Fortran maintainer can look into this and update the lowering? c/c-typeck.c: add_stmt (build_predict_expr (PRED_CONTINUE, NOT_TAKEN)); cp/cp-gimplify.c: tree pred = build_predict_expr (PRED_CONTINUE, NOT_TAKEN); are examples how these are used. Unlike builtin_expect you do not add it into the expression, but you drop it on the unlikely code path. Honza
