SLP/Loop vectorizer pass ordering

Thu Oct 9 12:44:07 PDT 2014

I have a patch I've been testing to clean up a lot of the passes around the
vectorizers. I'll add this in and finish testing it, then send it out with
numbers.
On Oct 9, 2014 12:40 PM, "Andrew Trick" <atrick at apple.com> wrote:

>
> On Oct 9, 2014, at 8:48 AM, Hal Finkel <hfinkel at anl.gov> wrote:
>
> ----- Original Message -----
>
> From: "Arnold Schwaighofer" <aschwaighofer at apple.com>
> To: "Zinovy Nis" <zinovy.nis at gmail.com>
> Cc: "Hal Finkel" <hfinkel at anl.gov>, "LLVM Commits" <
> llvm-commits at cs.uiuc.edu>, "Tobias Grosser" <tobias at grosser.es>,
> "Chandler Carruth" <chandlerc at google.com>, "Nadav Rotem" <nrotem at apple.com
> >
> Sent: Thursday, October 9, 2014 10:07:42 AM
> Subject: Re: SLP/Loop vectorizer pass ordering
>
>
> The loop vectorizer now sees this loop:
>
> define void
> @_Z21ambient_occlusion_vecP6_IsectR5vrandILm8EE(%struct._Isect*
> nocapture %isect, %class.vrand* nocapture readonly
> dereferenceable(32) %rng) #0 {
> entry:
>  br label %for.body
>
> for.body:                                         ; preds =
> %for.inc.for.body_crit_edge, %entry
>  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next,
>  %for.inc.for.body_crit_edge ]
>  %occlusion.017 = phi float [ 1.000000e+00, %entry ], [ %phitmp,
>  %for.inc.for.body_crit_edge ]
>  %exitcond = icmp eq i64 %indvars.iv, 63
>  br i1 %exitcond, label %for.end, label %for.inc.for.body_crit_edge
>
> for.inc.for.body_crit_edge:                       ; preds = %for.body
>  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
>  %phitmp = fadd fast float %occlusion.017, 1.000000e+00
>  br label %for.body
>
> for.end:                                          ; preds = %for.body
>  %occlusion.017.lcssa = phi float [ %occlusion.017, %for.body ]
>  %t5 = getelementptr inbounds %struct._Isect* %isect, i64 0, i32 0
>  store float %occlusion.017.lcssa, float* %t5, align 4, !tbaa !1
>  ret void
> }
>
> Notice that the loop exit block is the loop header and the latch is
> not guaranteed to be executed. The loop vectorizer assumes such
> loops have been rotated.
>
>
> If we send this IR through loop-rotate it will vectorize.
>
> The farther away we move the loop vectorizer from loop rotate the
> likelier some optimization will destroy the rotated from. We might
> just want to run loop rotate before the loop vectorizer ...
>
>
> I think that makes sense -- and I don't recall loop rotation being
> expensive, plus is preserves just about everything (and I think does a
> reasonable job cleaning up after itself) ;)
>
> I'd say we run some benchmarks, and barring any issues, we just do it.
>
>
> Well, that is a classic candidate for rotate. So assuming whatever GVN is
> doing is sane, then I’d say it makes sense to rerun rotation.
>
> -Andy
>
>
>
>
> On Oct 9, 2014, at 1:15 AM, Zinovy Nis <zinovy.nis at gmail.com>
> wrote:
>
> Hi.
>
> Did you have a chance to look at my reproducer?
>
> 2014-10-07 21:34 GMT+04:00 Zinovy Nis <zinovy.nis at gmail.com>:
>
> Hi.
>
> I attached a reduced sample, based on
> https://code.google.com/p/aobench/.
>
> Run it first with an old SLP order:
>
> 1) clang -c -Ofast -static -march=core-avx2 aobench.cpp -Rpass=.
> -mllvm -debug-only=loop-vectorize -mllvm
> -run-slp-after-loop-vectorization=0
>
> and then with a new order:
>
> 2) clang -c -Ofast -static -march=core-avx2 aobench.cpp -Rpass=.
> -debug-only=loop-vectorize -mllvm
> -run-slp-after-loop-vectorization=1
>
> and see the logs:
>
> 1) aobench.cpp:59:9: remark: vectorized loop (vectorization
> factor: 8,
> unrolling interleave factor: 1) [-Rpass=loop-vectorize]
> 2) aobench.cpp:59:9: remark: loop ***not*** vectorized: use
> -Rpass-analysis=loop-vectorize for more info
> [-Rpass-missed=loop-vectorize]
>
> LV: Found an unidentified PHI.  %occlusion.017 = phi float [
> 1.000000e+00, %entry ], [ %phitmp, %for.inc.for.body_crit_edge ]
> LV: Can't vectorize the instructions or CFG
> LV: Not vectorizing: Cannot prove legality.
>
> 2014-10-06 17:46 GMT+04:00 Hal Finkel <hfinkel at anl.gov>:
>
> ----- Original Message -----
>
> From: "Zinovy Nis" <zinovy.nis at gmail.com>
> To: "Hal Finkel" <hfinkel at anl.gov>
> Cc: "LLVM Commits" <llvm-commits at cs.uiuc.edu>, "Tobias Grosser"
> <tobias at grosser.es>, "Chandler Carruth"
> <chandlerc at google.com>, "Nadav Rotem" <nrotem at apple.com>,
> "Arnold Schwaighofer" <aschwaighofer at apple.com>
> Sent: Monday, October 6, 2014 8:44:28 AM
> Subject: Re: SLP/Loop vectorizer pass ordering
>
> A bit later. At least GVN creates critical edges which are not
> handled
> by loop vectorizer then.
>
>
> Okay, please do (this is fairly important) -- if you can extract
> some relevant IR, filing a bug report would be great. Are you
> saying that running SLP early inhibits GVN from creating
> critical edges that the loop vectorizer does not understand?
>
> Thanks again,
> Hal
>
>
> 2014-10-06 17:33 GMT+04:00 Hal Finkel <hfinkel at anl.gov>:
>
> ----- Original Message -----
>
> From: "Zinovy Nis" <zinovy.nis at gmail.com>
> To: "Chandler Carruth" <chandlerc at google.com>
> Cc: "LLVM Commits" <llvm-commits at cs.uiuc.edu>, "Tobias
> Grosser"
> <tobias at grosser.es>
> Sent: Monday, October 6, 2014 8:19:24 AM
> Subject: Re: SLP/Loop vectorizer pass ordering
>
> Please wait a while, I'm using it to revert the new order as
> it
> introduces regression in our internal benchmark: SLP was
> creating
> loop
> vectorization opportunities when was called before LV. Now no
> such
> opportunities are available, so we've got a regression.
>
>
> Interesting. Can you provide any further details?
>
> -Hal
>
>
> 2014-10-06 3:28 GMT+04:00 Chandler Carruth
> <chandlerc at google.com>:
>
>
> On Thu, Sep 4, 2014 at 6:32 AM, James Molloy
> <james at jamesmolloy.co.uk>
> wrote:
>
>
> Hi Hal, Chandler,
>
> r217144.
>
>
>
> Is anyone still using the option to disable this? If I don't
> hear
> anything,
> I'll remove this option entirely in the next week.
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>
>
> --
> Hal Finkel
> Assistant Computational Scientist
> Leadership Computing Facility
> Argonne National Laboratory
>
>
>
> --
> Hal Finkel
> Assistant Computational Scientist
> Leadership Computing Facility
> Argonne National Laboratory
>
>
>
>
> --
> Hal Finkel
> Assistant Computational Scientist
> Leadership Computing Facility
> Argonne National Laboratory
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20141009/b82d2ac7/attachment.html>