[llvm-dev] Speculation and control dependent no wrap flags

[Hal Finkel via llvm-dev]

On 02/03/2017 07:28 AM, Artur Pilipenko wrote:
> I'm looking at the bug (https://llvm.org/bugs/show_bug.cgi?id=31181) 
> which was triggered by my change to make CVP mark adds as no wrap 
> (https://reviews.llvm.org/rL278220) and I'd like to have some broader 
> discussion of the problem. In this bug CVP correctly marks an add as 
> nuw basing on the loop latch check, but later loop rotation pass moves 
> the add to a point before the check. In the new context nuw is no 
> longer valid and leads to an incorrect transformation of the loop. See 
> https://llvm.org/bugs/show_bug.cgi?id=31181#c5 comment in the bug for 
> more details.
>
> Since nsw, nuw flags can be control dependent, it seems like we should 
> be treating them as metadata, i.e. we should be stripping them when we 
> speculate the instruction. I don’t think that we are doing this now 
> anywhere. The problem was noticed on loop rotation, but I expect any 
> other pass which speculates overflowing operations is suffering from 
> the same problem.
>
> Thoughts?

We generally don't strip these because violating the wrapping constraint 
does not immediately cause UB. Instead, it generates a poison value. So 
long as that poison value is not used in way which causes UB, then 
everything is fine. In this case, I suspect that we want to fix IndVars 
to strip the flag, or not do the transformation, when it might introduce 
this kind of issue (i.e. a situation where we might branch on a poison 
value).

  -Hal

>
> Artur

-- 
Hal Finkel
Lead, Compiler Technology and Programming Languages
Leadership Computing Facility
Argonne National Laboratory