[llvm-dev] Dataflow analysis regression in 3.7

[Johan Engelen via llvm-dev]

Hi Chad and Davide,
  Thanks for working on this.
I don't know enough about LLVM internals (yet?) to contribute here, but
it's nice to "listen-in" on the discussion and watch the progress :)

- Johan

On Fri, Jul 7, 2017 at 10:59 PM, Davide Italiano <davide at freebsd.org> wrote:

> On Fri, Jul 7, 2017 at 1:47 PM, Chad Rosier <mcrosier at codeaurora.org>
> wrote:
> > David/Johan,
> >
> > I would love to claim victory, but I don't think that D34901 catches this
> > case.
> >
>
> Hi Chad, thanks for taking another look at this.
> Maybe I didn't bisect correctly. Apologies. Anyway, more fun for us.
>
> > However, I got interested and threw this together quickly:
> > https://reviews.llvm.org/D35140.
> >
> > This does catch the below case.  If people are interested I can add test
> > cases and submit for formal review.  FWIW, it does hit about 1/3 of all
> of
> > the SPEC benchmarks.  I haven't done any performance analysis to see if
> it
> > matters, however.
> >
>
> As I stated earlier, I think this calls for a slightly better VRP
> [and/or improvements in VRP] & I'm not entirely sure `InstCombine` is
> the best place for this transform.
> OTOH, the fact this triggers a lot is a good thing, if your WIP patch
> actually results in performance improvements, we should consider a
> more general solution to address this kind of issues.
>
> --
> Davide
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170708/b2b8a6fb/attachment.html>