> This has been broken because since mainline revision r160124, which > (correctly) causes 'y' to be flagged as noreturn -- y never returns via > the normal function return route. If ipa-pure-const.c is hacked to > check that a function does not perform non-local gotos before marking > it noreturn, the original (correct) offset is restored -- but that is > simply papering over the problem.
Agreed. > So, on to the attached fix. As mentioned previously, GCC tracks > elimination offsets throughout each function, which can vary as e.g. > other functions are called. There is a concept of an "initial" offset: > I believe that when nonlocal gotos are executed, and hence at all the > labels a nonlocal goto may reach, the elimination offsets must have > their "initial" values. Agreed. > We can find out if a label is the target of a nonlocal goto (and hence > force the use of initial offsets) in a somewhat roundabout way, by > looking up its containing basic block, seeing if that BB is a nonlocal > goto target, then seeing if the label is the first instruction in the > block. This seems slightly clumsy, but I didn't find another way of > doing it. These labels are on the nonlocal_goto_handler_labels chain. You presumably just need to apply the same treatment to them in set_initial_label_offsets as the one applied to forced labels. > Any comments, or OK to apply? OK for the adjusted patch if it works, mainline and 4.6 branch once it reopens. Please mention PR rtl-optimization/47918 in the ChangeLog. Thanks for fixing this. -- Eric Botcazou
