On 1/11/19 4:44 PM, Alan Modra wrote: > On Fri, Jan 11, 2019 at 11:42:31AM -0700, Jeff Law wrote: >> On 1/10/19 12:19 AM, Alan Modra wrote: >>> bb-reorder is quite seriously broken if get_attr_min_length should >>> return INT_MAX, which it does for hppa on branches with r267666. >> Presumably you're referring to the overflows and such? > > Yes. Even get_uncond_jump_length would have been INT_MAX. All of > the predicates deciding on whether to copy or reorder blocks were > therefore broken. > > The following is fairly obvious and would stop some of the silliness, > but I guess now is not the time to propose this sort of patch. > > * bb-reorder.c (copy_bb_p): Don't overflow size calculation. > (get_uncond_jump_length): Assert length less than INT_MAX and > non-negative. Now seems like a good time to revisit. I ran this through the usual bootstrap and regression text on x86_64. It's also built and tested on a good variety of the embedded targets.
Can't really test the PA right now due to what I believe is a qemu bug. David is still doing bootstraps on real hardware through, so if it causes a problem on the PA I'm sure he'll chime in at some point. I'm going to go ahead and install this on the trunk. jeff