[llvm] r273585 - Preserve DebugInfo when replacing values in DAGCombiner

Ismail Donmez via llvm-commits llvm-commits at lists.llvm.org
Tue Jul 26 20:51:38 PDT 2016


Yes! That fixes the problem. Thanks a lot.

On Tuesday, July 26, 2016, Nirav Davé <niravd at google.com> wrote:

> Apparently we're transferring debug values later in the ReplaceAllUsesWith
> than the other N versions. I still don't see how we're getting bad Debug
> Values, but this may fix it. Can you try this patch?
>
> -Nirav
>
> On Tue, Jul 26, 2016 at 9:51 AM, Nirav Davé <niravd at google.com
> <javascript:_e(%7B%7D,'cvml','niravd at google.com');>> wrote:
>
>> Not yet. I've run it for multiple days in multiple contexts but it never
>> crashes for me.
>>
>> I've been working on it from first principles but aside from that minor
>> patch I sent you that slightly cleans up usage, I haven't found anything
>> that could be labelled as wrong. Most upsettingly, none of the Debug Values
>> in question have ever been freed at the time of the crash; They're
>> allocated in a bumpptr allocator which defers freeing until after we're
>> done with SelectionDAG.
>>
>> -Nirav
>>
>>
>> On Jul 26, 2016 07:38, "Ismail Donmez" <ismail at i10z.com
>> <javascript:_e(%7B%7D,'cvml','ismail at i10z.com');>> wrote:
>> >
>> > Any luck with reproducing? This bug affects us really badly, we are
>> > unable to produce a snapshot more than a month now :(
>> >
>> > ismail
>> >
>> > On Fri, Jul 22, 2016 at 9:57 AM, Ismail Donmez <ismail at i10z.com
>> <javascript:_e(%7B%7D,'cvml','ismail at i10z.com');>> wrote:
>> > > Hi Nirav,
>> > >
>> > > On Thu, Jul 21, 2016 at 10:37 PM, Nirav Davé <niravd at google.com
>> <javascript:_e(%7B%7D,'cvml','niravd at google.com');>> wrote:
>> > >> Ismall:
>> > >>
>> > >> I'm still having problems finding reproducing this corruption
>> locally, but I
>> > >> may have a patch. Can you check if this works?
>> > >> If it doesn't, maybe you can spot what I'm doing wrong in my
>> reproduction
>> > >> steps.
>> > >
>> > > Sadly the patch didn't help. What I did to reproduce the crash was
>> > > running the reproducer script from
>> > > https://llvm.org/bugs/show_bug.cgi?id=28613 in a tight loop like:
>> > >
>> > > while true; do ./Loop-blabla.sh && sleep 1; done
>> > >
>> > > After it crashes I used the core file with gdb to debug. Of course
>> > > you'll still need to export MALLOC_CHECK_ and MALLOC_PERTURB_
>> > > beforehand.
>> > >
>> > > Regards,
>> > > ismail
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160727/8d9997c8/attachment.html>


More information about the llvm-commits mailing list