[polly] r264118 - [ScopInfo] Fix domains after loops.

[Tobias Grosser via llvm-commits]

On 04/04/2016 01:34 PM, Michael Kruse wrote:
> 2016-04-04 11:55 GMT+02:00 Johannes Doerfert <doerfert at cs.uni-saarland.de>:

>>> Your example above suggests that you have a correctness issue in mind. As
>>> said above, do we already have (or can we draft) such a test case?
>>>
>>>> D18450 cannot help when the loop
>>>> has multiple exits.
>>>
>>> Agreed. And for whatever reason it also did not fix the bug we see on LNT.
>> That statement is not helping. AFAIK, there is little evidence that the
>> additional constraints cause the lnt failure.
>
> Huh?

Even though I think we managed to move to a very productive discussion 
culture (thanks to both of you for this!), I wanted to briefly remind 
that some phrases unintentionally may be perceived in ways we did not 
expect.

Both, 'That statement is not helping' as well as 'Huh' might belong in 
this category. (Not claiming I don't unintentionally reply in such way 
sometimes as well).

Regarding Johannes question:

 >> That statement is not helping. AFAIK, there is little evidence that
 >> the additional constraints cause the lnt failure.

The additional constraints are part of the reason we see the LNT 
failure, as they prevent us from bailing out of the code generation,
such that we cannot fail to generate code for this case.

As we could just fix the code generation to not break, the question that 
remains is if the code we would generate would be correct or not.
I think understanding this question is important.

Best,
Tobias