llvm dev - Sep 2021 - [llvm-reduce] Reduction to undef/poison/null?

I frequently use llvm-reduce just to minimize a crash caused by some change
and present that to the author of a change to look at. I don't think that
having tons of freeze poisons in a repro file is nice to work with. If a
crash repros with a `0` as opposed to a `freeze poison` the `0` seems much
more appealing to present.

We could add a flag to reduce to the various options here if people have
different needs.

On Mon, Aug 30, 2021 at 11:31 AM Johannes Doerfert via llvm-dev <
llvm-dev at lists.llvm.org> wrote:
> On 8/30/21 1:22 PM, Roman Lebedev via llvm-dev wrote:
> > I've been thinking we should be using `freeze poison`,
> > but i don't think this question matters for the patch at hand,
> > it should just stick to the current practice of using undef.
>
> I like freeze poison. It conveys the idea w/o making things UB all the
> time.
> It basically is an oracle w/o the side effects.
>
> FWIW, when I ported tests to the Attributor, e.g., from
> ArgumentPromotion or IPSCCP,
> I had to manually remove all the UB that made the test meaningless
> first. In general,
> tests that contain statically provable UB are less likely to be
> meaningful over time
> and/or be reusable.
>
> ~ Johannes
>
>
> > Roman.
> >
> > On Mon, Aug 30, 2021 at 9:14 PM Florian Hahn via llvm-dev
> > <llvm-dev at lists.llvm.org> wrote:
> >>
> >>
> >>> On 30 Aug 2021, at 19:59, David Blaikie via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
> >>>
> >>> Nicer because it's less likely to introduce new UB? Or
some other
> reason?
> >>>
> >> Using undef/poison is problematic, because there are multiple ways
this
> could cause new UB (e.g. branch on undef, passing poison/undef to a
> function with a noundef argument).
> >>
> >> I’m not sure if using zero will work well in certain cases,
because it
> can introduce UB as well (e.g. load from null, passing as nonnull
argument).
> >>
> >> I think ideally we would have a way to materialise values we know
> nothing about, but are not undef. Perhaps we could add some oracle
> function, but that would come with its own drawbacks.
> >>
> >> Cheers,
> >> Florian
> >>
> >> _______________________________________________
> >> LLVM Developers mailing list
> >> llvm-dev at lists.llvm.org
> >> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
> > _______________________________________________
> > LLVM Developers mailing list
> > llvm-dev at lists.llvm.org
> > https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
> --
> ──────────────────────
> ∽ Johannes Doerfert ∽
> ∽ Pronouns: he/him  ∽
> ∽ Researcher @ ANL  ∽
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20210830/e5b54fe8/attachment.html>

On 8/30/21 1:36 PM, Arthur Eubanks wrote:
> I frequently use llvm-reduce just to minimize a crash caused by some change
> and present that to the author of a change to look at. I don't think
that
> having tons of freeze poisons in a repro file is nice to work with. If a
> crash repros with a `0` as opposed to a `freeze poison` the `0` seems much
> more appealing to present.
>
> We could add a flag to reduce to the various options here if people have
> different needs.
Yeah, maybe. I think it's clear undef is often not the best choice.

> On Mon, Aug 30, 2021 at 11:31 AM Johannes Doerfert via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
>
>> On 8/30/21 1:22 PM, Roman Lebedev via llvm-dev wrote:
>>> I've been thinking we should be using `freeze poison`,
>>> but i don't think this question matters for the patch at hand,
>>> it should just stick to the current practice of using undef.
>> I like freeze poison. It conveys the idea w/o making things UB all the
>> time.
>> It basically is an oracle w/o the side effects.
>>
>> FWIW, when I ported tests to the Attributor, e.g., from
>> ArgumentPromotion or IPSCCP,
>> I had to manually remove all the UB that made the test meaningless
>> first. In general,
>> tests that contain statically provable UB are less likely to be
>> meaningful over time
>> and/or be reusable.
>>
>> ~ Johannes
>>
>>
>>> Roman.
>>>
>>> On Mon, Aug 30, 2021 at 9:14 PM Florian Hahn via llvm-dev
>>> <llvm-dev at lists.llvm.org> wrote:
>>>>
>>>>> On 30 Aug 2021, at 19:59, David Blaikie via llvm-dev <
>> llvm-dev at lists.llvm.org> wrote:
>>>>> Nicer because it's less likely to introduce new UB? Or
some other
>> reason?
>>>> Using undef/poison is problematic, because there are multiple
ways this
>> could cause new UB (e.g. branch on undef, passing poison/undef to a
>> function with a noundef argument).
>>>> I’m not sure if using zero will work well in certain cases,
because it
>> can introduce UB as well (e.g. load from null, passing as nonnull
argument).
>>>> I think ideally we would have a way to materialise values we
know
>> nothing about, but are not undef. Perhaps we could add some oracle
>> function, but that would come with its own drawbacks.
>>>> Cheers,
>>>> Florian
>>>>
>>>> _______________________________________________
>>>> LLVM Developers mailing list
>>>> llvm-dev at lists.llvm.org
>>>> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>> _______________________________________________
>>> LLVM Developers mailing list
>>> llvm-dev at lists.llvm.org
>>> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>> --
>> ──────────────────────
>> ∽ Johannes Doerfert ∽
>> ∽ Pronouns: he/him  ∽
>> ∽ Researcher @ ANL  ∽
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>

> On Aug 31, 2021, at 9:04 AM, Johannes Doerfert via llvm-dev <llvm-dev at
lists.llvm.org> wrote:
> 
> 
> On 8/30/21 1:36 PM, Arthur Eubanks wrote:
>> I frequently use llvm-reduce just to minimize a crash caused by some
change
>> and present that to the author of a change to look at. I don't
think that
>> having tons of freeze poisons in a repro file is nice to work with. If
a
>> crash repros with a `0` as opposed to a `freeze poison` the `0` seems
much
>> more appealing to present.
>> 
>> We could add a flag to reduce to the various options here if people
have
>> different needs.
> 
> Yeah, maybe. I think it's clear undef is often not the best choice.
I agree with the meta point here - use of undef will make the behavior far less
predictable, and therefore make reduction more annoying.  I think it makes a lot
of sense to use a deterministic zero value instead of undef.

Is there any _disadvantage_ to doing that?

-Chris