On Tue, 26 Apr 2016, Uros Bizjak wrote: > On Tue, Apr 26, 2016 at 11:17 AM, Richard Biener <rguent...@suse.de> wrote: > > On Mon, 25 Apr 2016, Uros Bizjak wrote: > > > >> On Mon, Apr 25, 2016 at 4:47 PM, H.J. Lu <hjl.to...@gmail.com> wrote: > >> > On Mon, Apr 25, 2016 at 7:18 AM, Uros Bizjak <ubiz...@gmail.com> wrote: > >> >> On Mon, Apr 25, 2016 at 2:51 PM, H.J. Lu <hongjiu...@intel.com> wrote: > >> >>> Tested on Linux/x86-64. OK for trunk? > >> >> > >> >>> + /* FIXME: Since the CSE pass may change dominance info, which isn't > >> >>> + expected by the fwprop pass, call free_dominance_info to > >> >>> + invalidate dominance info. Otherwise, the fwprop pass may crash > >> >>> + when dominance info is changed. */ > >> >>> + if (TARGET_64BIT) > >> >>> + free_dominance_info (CDI_DOMINATORS); > >> >>> + > >> >> > >> >> Please resolve the above problem first, target-dependent sources are > >> >> not the place to apply band-aids for middle-end problems. The thread > >> >> with the proposed fix died in [1]. > >> >> > >> >> [1] https://gcc.gnu.org/ml/gcc/2016-03/msg00143.html > >> > > >> > free_dominance_info (CDI_DOMINATORS) has been called in other > >> > places to avoid this middle-end issue. I don't know when the middle-end > >> > will be fixed. I don't think this target optimization should be > >> > penalized by > >> > the middle-end issue. > >> > >> Let's ask Richard if he is OK with the workaround... > > > > Well, it's ultimately your call (it's a workaround in the target). > > Oh well, ... > > > Of course I'd like to see the underlying issue fixed and the > > workarounds in "other places" be removed. > > ... then at least a reference to a relevant PR should be added to a > FIXME comment.
HJ, can you please open a bug with 1) a testcase, 2) a patch to revert the workaround so it shows the ICE and 3) a pointer to the ml thread with your preliminary analysis? The advantage is that with the patch in we have a reproducer at least. Thanks, Richard.
