[llvm-dev] SLPVectorizer and full restrict

[Erik Eckstein via llvm-dev]

Oh wow, that's a long time ago since I worked on this. Unfortunately I'm not at the conference next week.
I'm happy to help (if I can remember :-)).

> On 16 Oct 2019, at 22:45, Jeroen Dobbelaere <Jeroen.Dobbelaere at synopsys.com> wrote:
> 
> Hi all,
> 
> Is there anybody who understands how SLPVectorizer.cpp:  BoUpSLP::scheduleBlock(...) works ?
> 
> There is some interference with the noalias intrinsics (https://reviews.llvm.org/D68484),
> and I am not sure on how to fix it (*). 
> I'll be at the conference next week. That would be a good moment to discuss.
> 
> Thanks,
> 
> Jeroen Dobbelaere
> --
> (*) The test-suite build fails for 'SingleSource/Benchmarks/SmallPT':
> on one hand, it looks as if the vectorizer sees through the noalias intrinsics (as expected), but then,
> the scheduler gets stuck and does not schedule all the instructions, probably because some dependencies
> are not correctly resolved.
>