llvm dev - Oct 2014 - [LLVMdev] RFC: Should we have (something like) -extra-vectorizer-passes in -O2?

----- Original Message -----
> From: "Chandler Carruth" <chandlerc at google.com>
> To: "Zinovy Nis" <zinovy.nis at gmail.com>
> Cc: "Hal Finkel" <hfinkel at anl.gov>, "James
Molloy" <james at jamesmolloy.co.uk>, "LLVM Developers Mailing
List"
> <llvmdev at cs.uiuc.edu>
> Sent: Thursday, October 16, 2014 1:21:19 PM
> Subject: Re: [LLVMdev] RFC: Should we have (something like)
-extra-vectorizer-passes in -O2?
> 
> 
> 
> 
> 
> On Thu, Oct 16, 2014 at 7:55 AM, Zinovy Nis < zinovy.nis at gmail.com
>
> wrote:
> 
> 
> 
> Seems that adding -extra-vectorizer-passes doesn't help to vectorizer
> in my case. LoopRotation re-run does nothing.
> 
> Can you reduce this to a test case you can share?
He had posted one earlier here:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20141006/238660.html
(Arnold had posted some analysis here:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20141006/239144.html
- which I imagine you saw)

 -Hal

-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory

On Thu, Oct 16, 2014 at 11:38 AM, Hal Finkel <hfinkel at anl.gov> wrote:
> He had posted one earlier here:
>
>
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20141006/238660.html
> (Arnold had posted some analysis here:
>
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20141006/239144.html
> - which I imagine you saw)
>
Doh, sorry. I saw Arnold's analysis but (wrongly) assumed that a complete
working test case wasn't available which is why Arnold expected loop-rotate
to fix this when it didn't. My bad.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20141016/6eb19691/attachment.html>

I quickly took a look again. Instcombine is removing the fast-math flag on the
reduction so the vectorizer does not touch it.

bin/clang -O3 -ffast-math -mllvm -extra-vectorizer-passes aobench.cpp -emit-llvm
-S -o aobench.2.ll -mllvm -debug-only=loop-vectorize -mllvm -print-before-all
-mllvm -print-after-all 2> aobench.2.debug.ll

*** IR Dump Before Combine redundant instructions ***
; Function Attrs: noinline nounwind ssp uwtable
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
  %8 = phi float [ %conv.i, %entry ], [ %.pre, %for.inc.for.body_crit_edge ]
  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next,
%for.inc.for.body_crit_edge ]
  %occlusion.017 = phi float [ 0.000000e+00, %entry ], [ %conv4,
%for.inc.for.body_crit_edge ]
  %arrayidx = getelementptr inbounds [8 x float]* %rand1, i64 0, i64 %indvars.iv
  %9 = load float* %arrayidx, align 4, !tbaa !5
  %mul3 = fmul fast float %8, %9
  %mul.i.i = fmul fast float %mul3, %mul3
  %cmp.i = fcmp olt float %mul.i.i, 0x3C670EF540000000
  %conv4 = fadd fast float %occlusion.017, 1.000000e+00  <============= FAST
  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
  %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
  %arrayidx2.phi.trans.insert = getelementptr inbounds [8 x float]* %rand2, i64
0, i64 %indvars.iv.next
  %.pre = load float* %arrayidx2.phi.trans.insert, align 4, !tbaa !5
  br label %for.body

for.end:                                          ; preds = %for.body
  %t5 = getelementptr inbounds %struct._Isect* %isect, i64 0, i32 0
  store float %conv4, float* %t5, align 4, !tbaa !7
  ret void
}

*** IR Dump After Combine redundant instructions ***
; Function Attrs: noinline nounwind ssp uwtable
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 float %occlusion.017, 1.000000e+00 <=============== NOT FAST
  br label %for.body

for.end:                                          ; preds = %for.body
  %t5 = getelementptr inbounds %struct._Isect* %isect, i64 0, i32 0
  store float %occlusion.017, float* %t5, align 4, !tbaa !1
  ret void
}

> On Oct 16, 2014, at 11:52 AM, Chandler Carruth <chandlerc at
google.com> wrote:
> 
> 
> On Thu, Oct 16, 2014 at 11:38 AM, Hal Finkel <hfinkel at anl.gov>
wrote:
> He had posted one earlier here:
>
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20141006/238660.html
> (Arnold had posted some analysis here:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20141006/239144.html
- which I imagine you saw)
> 
> Doh, sorry. I saw Arnold's analysis but (wrongly) assumed that a
complete working test case wasn't available which is why Arnold expected
loop-rotate to fix this when it didn't. My bad.
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev