[llvm] r315329 - Revert "[SCCP] Propagate integer range info for parameters in IPSCCP."

Florian Hahn via llvm-commits llvm-commits at lists.llvm.org
Thu Oct 12 13:33:59 PDT 2017


On 12/10/2017 20:36, Davide Italiano wrote:
> On Thu, Oct 12, 2017 at 12:30 PM, Florian Hahn <florian.hahn at arm.com> wrote:
> >
> >
> > On 12/10/2017 17:55, Bruno Cardoso Lopes wrote:
> >>
> >> On Wed, Oct 11, 2017 at 1:11 PM, Florian Hahn <florian.hahn at arm.com>
> >> wrote:
> >>>
> >>>
> >>> On 11/10/2017 17:45, Bruno Cardoso Lopes wrote:
> >>>>
> >>>>
> >>>> On Wed, Oct 11, 2017 at 7:04 AM, Florian Hahn <florian.hahn at arm.com>
> >>>> wrote:
> >>>>>
> >>>>>
> >>>>>
> >>>>> On 11/10/2017 05:31, Davide Italiano wrote:
> >>>>>>
> >>>>>>
> >>>>>> Do you have a repro?
> >>>>>>
> >>>>> I tried to reproduce the failure locally, by building building
> >>>>> yaml-bench
> >>>>> with LTO and a Clang build using the commit introducing the SCCP
> >>>>> change.
> >>>>> Unfortunately, I did not manage to reproduce the failure yet :(
> >>>>
> >>>>
> >>>> If you re-introduce the change I can try to get a reproducer. The same
> >>>> bot is red as of this morning, if you can wait until I can get it
> >>>> green again that would be ideal. I'll give you a heads up soon.
> >>>>
> >>>
> >>> That would be extremely helpful! Would it work for you if I land the
> >>> change
> >>> tomorrow around the 7pm UK time?
> >>
> >>
> >> Just landed both commits again in r315593, I'll keep an eye on the bots.
> >>
> >> Thanks!
> >
> >
> > Great thanks! It's failing again, with
> >
> > Assertion failed: (I != ValueState.end() && "V not found in ValueState nor
> > Paramstate map!"), function getLatticeValueFor, file
> > /Users/buildslave/jenkins/workspace/clang-stage1-configure-RA/llvm/lib/Transforms/Scalar/SCCP.cpp,
> > line 310.
> >
>
> FWIW, as it's asserting (and not miscompiling), you should be able to
> dump the IR before SCCP (or IPSCCP), and run opt -sccp/-ipsccp on it
> (and then reduce from there). Bruno, can you please attach the IR?


I am off now. It would be great if you could revert the commit again, once you got a reproducer.

Again, thank you very much for spending time on this!

Cheers,
Florian


More information about the llvm-commits mailing list