On Thu, Apr 21, 2016 at 3:12 PM, Richard Biener <rguent...@suse.de> wrote: > On Thu, 21 Apr 2016, Jakub Jelinek wrote: > >> On Thu, Apr 21, 2016 at 02:20:34PM +0200, Richard Biener wrote: >> > >> > The following fixes the followup ICEs in the testcase for PR70725 >> > where Markes patch only fixed the first one. >> > >> > Bootstrapped on x86_64-unknown-linux-gnu, testing in progress. >> > >> > Jakub - the patch should be safe in that any testcase running into >> > the changed paths now would have caused SSA verifications without >> > the patch. There's still the question what to do for GCC 6.1 - we >> > can revert Mareks patch or apply this one ontop. I don't have a >> > strong opinion here. >> >> I think reverting Marek's patch for 6.1 and reapplying it for 6.2 together >> with this would be safest, but am not strongly against just applying this to >> 6.1 either. > > Ok, it doesn't apply cleanly and I won't get to it today so please > revert Mareks change on the branch if you want to create RC2 today. Hi, Hmm, I think this fix is not needed on GCC 6 because the ICE is introduced by my patch @GCC7. The default behavior of GCC6 is simply return false, thus loop is not if converted. Sorry I should have kept the old behavior before my next patch fixing PHIs that can be degenerated.
Thanks, bin > > Applied to trunk sofar. > > Richard. > >> > 2016-04-21 Richard Biener <rguent...@suse.de> >> > >> > PR tree-optimization/70725 >> > * tree-if-conv.c (if_convertible_phi_p): Adjust guard >> > for phi_convertible_by_degenerating_args. >> > (predicate_all_scalar_phis): Handle single-argument PHIs.
