Details

Reviewers

RKSimon
spatel
hfinkel
craig.topper
dtemirbulatov
anton-afanasyev

Summary

Possibly vectorized operations are extended to the power-of-2 number with UndefValues to allow to use regular vector operations.

For SPEC CPU2017 it gives ~7% perf gain for 526.blender_r (AVX512,
O3+LTOi, -march=native), ~2% gain for 538.imagick_r and 638.imagick_s,
~2% gain for 525.x264_r and 625.x264_s, ~2% gain for 526.blender_r (AVX2
, O3+LTO, -march=native), ~11% gain 526.blender_r, ~3% gain for
544.nab_r and 644.nab_s (AVX512, O3+LTO), ~3% gain
for 526.blender_r, ~2% gain for 544.nab_r and 644.nab_s (AVX2, O3+LTO).

Compile and link time are pretty the same:

AVX512, O3+LTO, -march=native
Metric: compile_time
Geomean difference -0.1% (-1.85 sec)

Metric: link_time
Geomean difference +1.2% (+14.46 sec)

AVX512, O3+LTO
Metric: compile_time
Geomean difference -0.2% (-4.71 sec)

Metric: link_time
Geomean difference -3.6% (-54.53 sec)

AVX2, O3+LTO, -march=native
Metric: compile_time
Geomean difference +0.3% (+10.56 sec)

Metric: link_time
Geomean difference -0.1% (-2.18 sec)

AVX2, O3+LTO
Metric: compile_time
Geomean difference +0.2% (+5.73 sec)

Metric: link_time
Geomean difference -3.4% (-67.45 sec)

Diff Detail

Repository: rG LLVM Github Monorepo

Unit TestsFailed
View All

	Time	Test
	2,530 ms	x64 debian > libarcher.critical::critical.c
	2,560 ms	x64 debian > libarcher.parallel::parallel-simple2.c
	2,630 ms	x64 debian > libarcher.races::critical-unrelated.c
	2,850 ms	x64 debian > libarcher.races::lock-nested-unrelated.c
	2,480 ms	x64 debian > libarcher.races::lock-unrelated.c
		View Full Test Results (18 Failed)

Event Timeline

There are a very large number of changes, so older changes are hidden. Show Older Changes

Rebase

Harbormaster completed remote builds in B81501: Diff 310293.Dec 8 2020, 11:41 AM

AFAICT the only outstanding question is whether the compile time increase is acceptable?

In D57059#2442286, @RKSimon wrote:

AFAICT the only outstanding question is whether the compile time increase is acceptable?

I'd agree that geomean = 0.2% is acceptable for the change with such awesome performance impact, just noted that changed time compilation is significant in comparision with other changes. Generally it looks good to me apart from one minor unaddressed comment.

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
3764	Well, "unsortable" or "unprocessable" term would be more precise. But why did we change `if (sortPtrAccesses)...` to opposite condition? This change just duplicate debug output, since we didn't differentiate it. Also I'd prefer to see the same `if-else` structure as for the load case.

In D57059#2442527, @anton-afanasyev wrote:

In D57059#2442286, @RKSimon wrote:

AFAICT the only outstanding question is whether the compile time increase is acceptable?

I'd agree that geomean = 0.2% is acceptable for the change with such awesome performance impact, just noted that changed time compilation is significant in comparision with other changes. Generally it looks good to me apart from one minor unaddressed comment.

Could the summary of the revision be updated with the performance data? The thread is very long and I didn't spot where the measurements are, so it's hard to say what we're trading off here...

Generally though this level of geomean regression is fine, the issue is usually in the outliers. For example, I spot this this 4-5% regression:

CMakeFiles/lencod.dir/transform8x8.c.o 	3053M 	3188M (+4.41%)

It may be worthwhile to briefly check it in case something can be improved.

In D57059#2442542, @nikic wrote:

In D57059#2442527, @anton-afanasyev wrote:

In D57059#2442286, @RKSimon wrote:

AFAICT the only outstanding question is whether the compile time increase is acceptable?

I'd agree that geomean = 0.2% is acceptable for the change with such awesome performance impact, just noted that changed time compilation is significant in comparision with other changes. Generally it looks good to me apart from one minor unaddressed comment.

Could the summary of the revision be updated with the performance data? The thread is very long and I didn't spot where the measurements are, so it's hard to say what we're trading off here...

Will try to run the benchmarks and get fresh data.

Generally though this level of geomean regression is fine, the issue is usually in the outliers. For example, I spot this this 4-5% regression:
CMakeFiles/lencod.dir/transform8x8.c.o 	3053M 	3188M (+4.41%)
It may be worthwhile to briefly check it in case something can be improved.

I'll check what can be improved in terms of compile time. Not sure that will be able to improve it significantly since the patch itself does not add extensive analysis/transformations, just adds an extra 1 iteration for wider vector analysis. But I'll check it anyway and will try to improve things where possible.

While reviewing the latest update, I think I spotted SLP compile-time failure in SingleSource/Benchmarks/Misc/oourafft.c, here is the reduced testcase to reporduce:
source_filename = "/home/dtemirbulatov/llvm/test-suite/SingleSource/Benchmarks/Misc/oourafft.c"
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"

define dso_local fastcc void @cft1st(double* %a) unnamed_addr #0 {
entry:

%0 = or i64 16, 2
%arrayidx107 = getelementptr inbounds double, double* %a, i64 %0
%1 = or i64 16, 3
%arrayidx114 = getelementptr inbounds double, double* %a, i64 %1
%2 = or i64 16, 4
%arrayidx131 = getelementptr inbounds double, double* %a, i64 %2
%3 = or i64 16, 6
%arrayidx134 = getelementptr inbounds double, double* %a, i64 %3
%4 = load double, double* %arrayidx134, align 8
%5 = or i64 16, 5
%arrayidx138 = getelementptr inbounds double, double* %a, i64 %5
%6 = or i64 16, 7
%arrayidx141 = getelementptr inbounds double, double* %a, i64 %6
%7 = load double, double* %arrayidx141, align 8
%sub149 = fsub double undef, %4
%sub156 = fsub double undef, %7
store double undef, double* %arrayidx131, align 8
store double undef, double* %arrayidx138, align 8
%sub178 = fsub double undef, %sub156
%add179 = fadd double undef, %sub149
%mul180 = fmul double undef, %sub178
%sub182 = fsub double %mul180, undef
store double %sub182, double* %arrayidx107, align 8
%mul186 = fmul double undef, %add179
%add188 = fadd double %mul186, undef
store double %add188, double* %arrayidx114, align 8
unreachable

}

attributes #0 = { "target-features"="+avx,+avx2,+bmi,+bmi2,+cx16,+cx8,+f16c,+fma,+fsgsbase,+fxsr,+invpcid,+lzcnt,+mmx,+movbe,+pclmul,+popcnt,+rdrnd,+sahf,+sse,+sse2,+sse3,+sse4.1,+sse4.2,+ssse3,+x87,+xsave,+xsaveopt" }

!llvm.ident = !{!0}

!0 = !{!"clang version 12.0.0 (https://github.com/llvm/llvm-project.git aaa925795f93c389a96ee01bab73bc2b6b771cbb)"}

In D57059#2443350, @dtemirbulatov wrote:
While reviewing the latest update, I think I spotted SLP compile-time failure in SingleSource/Benchmarks/Misc/oourafft.c, here is the reduced testcase to reporduce:
source_filename = "/home/dtemirbulatov/llvm/test-suite/SingleSource/Benchmarks/Misc/oourafft.c"
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"

define dso_local fastcc void @cft1st(double* %a) unnamed_addr #0 {
entry:
%0 = or i64 16, 2
%arrayidx107 = getelementptr inbounds double, double* %a, i64 %0
%1 = or i64 16, 3
%arrayidx114 = getelementptr inbounds double, double* %a, i64 %1
%2 = or i64 16, 4
%arrayidx131 = getelementptr inbounds double, double* %a, i64 %2
%3 = or i64 16, 6
%arrayidx134 = getelementptr inbounds double, double* %a, i64 %3
%4 = load double, double* %arrayidx134, align 8
%5 = or i64 16, 5
%arrayidx138 = getelementptr inbounds double, double* %a, i64 %5
%6 = or i64 16, 7
%arrayidx141 = getelementptr inbounds double, double* %a, i64 %6
%7 = load double, double* %arrayidx141, align 8
%sub149 = fsub double undef, %4
%sub156 = fsub double undef, %7
store double undef, double* %arrayidx131, align 8
store double undef, double* %arrayidx138, align 8
%sub178 = fsub double undef, %sub156
%add179 = fadd double undef, %sub149
%mul180 = fmul double undef, %sub178
%sub182 = fsub double %mul180, undef
store double %sub182, double* %arrayidx107, align 8
%mul186 = fmul double undef, %add179
%add188 = fadd double %mul186, undef
store double %add188, double* %arrayidx114, align 8
unreachable
}

attributes #0 = { "target-features"="+avx,+avx2,+bmi,+bmi2,+cx16,+cx8,+f16c,+fma,+fsgsbase,+fxsr,+invpcid,+lzcnt,+mmx,+movbe,+pclmul,+popcnt,+rdrnd,+sahf,+sse,+sse2,+sse3,+sse4.1,+sse4.2,+ssse3,+x87,+xsave,+xsaveopt" }

!llvm.ident = !{!0}

!0 = !{!"clang version 12.0.0 (https://github.com/llvm/llvm-project.git aaa925795f93c389a96ee01bab73bc2b6b771cbb)"}

Do you mean compile time increasing? With this patch?

Do you mean compile time increasing? With this patch?

no, just compile-time error.

In D57059#2443492, @dtemirbulatov wrote:

Do you mean compile time increasing? With this patch?

no, just compile-time error.

Crash or incorrect code?

In D57059#2443496, @ABataev wrote:

In D57059#2443492, @dtemirbulatov wrote:

Do you mean compile time increasing? With this patch?

no, just compile-time error.

Crash or incorrect code?

Crash.

wxiao3 added a subscriber: wxiao3.Dec 14 2020, 7:09 AM

ABataev edited the summary of this revision. (Show Details)Feb 10 2021, 6:01 AM

Extra numbers:

AVX512, O3+LTO, -march=native
Metric: SLP.NumVectorInstructions

Program                                                                         lhs      rhs      diff
  test-suite :: External/SPEC/CINT2017rate/500.perlbench_r/500.perlbench_r.test   146.00   148.00   1.4%
 test-suite :: External/SPEC/CINT2017speed/600.perlbench_s/600.perlbench_s.test   146.00   148.00   1.4%
 test-suite :: External/SPEC/CINT2017speed/631.deepsjeng_s/631.deepsjeng_s.test    34.00    34.00   0.0%
             test-suite :: External/SPEC/CINT2017speed/605.mcf_s/605.mcf_s.test    11.00    11.00   0.0%
  test-suite :: External/SPEC/CINT2017rate/531.deepsjeng_r/531.deepsjeng_r.test    34.00    34.00   0.0%
              test-suite :: External/SPEC/CINT2017rate/505.mcf_r/505.mcf_r.test    11.00    11.00   0.0%
  test-suite :: External/SPEC/CINT2017rate/523.xalancbmk_r/523.xalancbmk_r.test  5587.00  5560.00  -0.5%
 test-suite :: External/SPEC/CINT2017speed/623.xalancbmk_s/623.xalancbmk_s.test  5587.00  5560.00  -0.5%
             test-suite :: External/SPEC/CFP2017rate/508.namd_r/508.namd_r.test  7384.00  7341.00  -0.6%
         test-suite :: External/SPEC/CFP2017rate/510.parest_r/510.parest_r.test  9607.00  9359.00  -2.6%
         test-suite :: External/SPEC/CFP2017rate/511.povray_r/511.povray_r.test  5340.00  5178.00  -3.0%
            test-suite :: External/SPEC/CINT2017rate/525.x264_r/525.x264_r.test  1053.00  1006.00  -4.5%
           test-suite :: External/SPEC/CINT2017speed/625.x264_s/625.x264_s.test  1053.00  1006.00  -4.5%
          test-suite :: External/SPEC/CINT2017rate/541.leela_r/541.leela_r.test   141.00   134.00  -5.0%
         test-suite :: External/SPEC/CINT2017speed/641.leela_s/641.leela_s.test   141.00   134.00  -5.0%
      test-suite :: External/SPEC/CFP2017speed/638.imagick_s/638.imagick_s.test  3996.00  3563.00 -10.8%
       test-suite :: External/SPEC/CFP2017rate/538.imagick_r/538.imagick_r.test  3996.00  3563.00 -10.8%
              test-suite :: External/SPEC/CINT2017rate/502.gcc_r/502.gcc_r.test   862.00   767.00 -11.0%
             test-suite :: External/SPEC/CINT2017speed/602.gcc_s/602.gcc_s.test   862.00   767.00 -11.0%
      test-suite :: External/SPEC/CINT2017rate/520.omnetpp_r/520.omnetpp_r.test   524.00   463.00 -11.6%
     test-suite :: External/SPEC/CINT2017speed/620.omnetpp_s/620.omnetpp_s.test   524.00   463.00 -11.6%
              test-suite :: External/SPEC/CFP2017speed/644.nab_s/644.nab_s.test   426.00   370.00 -13.1%
               test-suite :: External/SPEC/CFP2017rate/544.nab_r/544.nab_r.test   426.00   370.00 -13.1%
       test-suite :: External/SPEC/CFP2017rate/526.blender_r/526.blender_r.test 15945.00 12573.00 -21.1%
               test-suite :: External/SPEC/CFP2017rate/519.lbm_r/519.lbm_r.test      NaN    16.00   nan%
              test-suite :: External/SPEC/CFP2017speed/619.lbm_s/619.lbm_s.test      NaN    16.00   nan%
                                                             Geomean difference                     nan%

AVX512, O3+LTO
Metric: SLP.NumVectorInstructions

Program                                                                         lhs      rhs      diff
         test-suite :: External/SPEC/CINT2017speed/641.leela_s/641.leela_s.test    22.00    60.00 172.7%
          test-suite :: External/SPEC/CINT2017rate/541.leela_r/541.leela_r.test    22.00    60.00 172.7%
 test-suite :: External/SPEC/CINT2017speed/600.perlbench_s/600.perlbench_s.test    68.00    72.00   5.9%
  test-suite :: External/SPEC/CINT2017rate/500.perlbench_r/500.perlbench_r.test    68.00    72.00   5.9%
              test-suite :: External/SPEC/CINT2017rate/505.mcf_r/505.mcf_r.test    11.00    11.00   0.0%
 test-suite :: External/SPEC/CINT2017speed/631.deepsjeng_s/631.deepsjeng_s.test    10.00    10.00   0.0%
 test-suite :: External/SPEC/CINT2017speed/623.xalancbmk_s/623.xalancbmk_s.test  3396.00  3396.00   0.0%
             test-suite :: External/SPEC/CINT2017speed/605.mcf_s/605.mcf_s.test    11.00    11.00   0.0%
  test-suite :: External/SPEC/CINT2017rate/531.deepsjeng_r/531.deepsjeng_r.test    10.00    10.00   0.0%
  test-suite :: External/SPEC/CINT2017rate/523.xalancbmk_r/523.xalancbmk_r.test  3396.00  3396.00   0.0%
     test-suite :: External/SPEC/CINT2017speed/620.omnetpp_s/620.omnetpp_s.test   499.00   497.00  -0.4%
      test-suite :: External/SPEC/CINT2017rate/520.omnetpp_r/520.omnetpp_r.test   499.00   497.00  -0.4%
            test-suite :: External/SPEC/CINT2017rate/525.x264_r/525.x264_r.test   838.00   826.00  -1.4%
           test-suite :: External/SPEC/CINT2017speed/625.x264_s/625.x264_s.test   838.00   826.00  -1.4%
         test-suite :: External/SPEC/CFP2017rate/511.povray_r/511.povray_r.test  6090.00  5906.00  -3.0%
              test-suite :: External/SPEC/CFP2017speed/619.lbm_s/619.lbm_s.test   131.00   127.00  -3.1%
               test-suite :: External/SPEC/CFP2017rate/519.lbm_r/519.lbm_r.test   131.00   127.00  -3.1%
             test-suite :: External/SPEC/CFP2017rate/508.namd_r/508.namd_r.test  8815.00  8452.00  -4.1%
      test-suite :: External/SPEC/CFP2017speed/638.imagick_s/638.imagick_s.test  2864.00  2712.00  -5.3%
       test-suite :: External/SPEC/CFP2017rate/538.imagick_r/538.imagick_r.test  2864.00  2712.00  -5.3%
       test-suite :: External/SPEC/CFP2017rate/526.blender_r/526.blender_r.test 16049.00 14753.00  -8.1%
              test-suite :: External/SPEC/CINT2017rate/502.gcc_r/502.gcc_r.test   686.00   621.00  -9.5%
             test-suite :: External/SPEC/CINT2017speed/602.gcc_s/602.gcc_s.test   686.00   621.00  -9.5%
              test-suite :: External/SPEC/CFP2017speed/644.nab_s/644.nab_s.test   551.00   473.00 -14.2%
               test-suite :: External/SPEC/CFP2017rate/544.nab_r/544.nab_r.test   551.00   473.00 -14.2%
         test-suite :: External/SPEC/CFP2017rate/510.parest_r/510.parest_r.test 16240.00 13941.00 -14.2%
                                                             Geomean difference                     4.3%

AVX2, O3+LTO, -march=native
Metric: SLP.NumVectorInstructions

Program                                                                         lhs      rhs      diff
             test-suite :: External/SPEC/CFP2017rate/508.namd_r/508.namd_r.test  7309.00  7341.00   0.4%
              test-suite :: External/SPEC/CINT2017rate/505.mcf_r/505.mcf_r.test    11.00    11.00   0.0%
 test-suite :: External/SPEC/CINT2017speed/631.deepsjeng_s/631.deepsjeng_s.test    34.00    34.00   0.0%
 test-suite :: External/SPEC/CINT2017speed/623.xalancbmk_s/623.xalancbmk_s.test  5490.00  5490.00   0.0%
             test-suite :: External/SPEC/CINT2017speed/605.mcf_s/605.mcf_s.test    11.00    11.00   0.0%
  test-suite :: External/SPEC/CINT2017rate/531.deepsjeng_r/531.deepsjeng_r.test    34.00    34.00   0.0%
  test-suite :: External/SPEC/CINT2017rate/523.xalancbmk_r/523.xalancbmk_r.test  5490.00  5490.00   0.0%
     test-suite :: External/SPEC/CINT2017speed/620.omnetpp_s/620.omnetpp_s.test   462.00   455.00  -1.5%
      test-suite :: External/SPEC/CINT2017rate/520.omnetpp_r/520.omnetpp_r.test   462.00   455.00  -1.5%
         test-suite :: External/SPEC/CFP2017rate/510.parest_r/510.parest_r.test  9508.00  9347.00  -1.7%
         test-suite :: External/SPEC/CFP2017rate/511.povray_r/511.povray_r.test  5393.00  5190.00  -3.8%
            test-suite :: External/SPEC/CINT2017rate/525.x264_r/525.x264_r.test  1066.00   968.00  -9.2%
           test-suite :: External/SPEC/CINT2017speed/625.x264_s/625.x264_s.test  1066.00   968.00  -9.2%
          test-suite :: External/SPEC/CINT2017rate/541.leela_r/541.leela_r.test   151.00   134.00 -11.3%
         test-suite :: External/SPEC/CINT2017speed/641.leela_s/641.leela_s.test   151.00   134.00 -11.3%
  test-suite :: External/SPEC/CINT2017rate/500.perlbench_r/500.perlbench_r.test   160.00   141.00 -11.9%
 test-suite :: External/SPEC/CINT2017speed/600.perlbench_s/600.perlbench_s.test   160.00   141.00 -11.9%
             test-suite :: External/SPEC/CINT2017speed/602.gcc_s/602.gcc_s.test   820.00   722.00 -12.0%
              test-suite :: External/SPEC/CINT2017rate/502.gcc_r/502.gcc_r.test   820.00   722.00 -12.0%
       test-suite :: External/SPEC/CFP2017rate/538.imagick_r/538.imagick_r.test  3605.00  3173.00 -12.0%
      test-suite :: External/SPEC/CFP2017speed/638.imagick_s/638.imagick_s.test  3605.00  3173.00 -12.0%
               test-suite :: External/SPEC/CFP2017rate/544.nab_r/544.nab_r.test   438.00   370.00 -15.5%
              test-suite :: External/SPEC/CFP2017speed/644.nab_s/644.nab_s.test   438.00   370.00 -15.5%
       test-suite :: External/SPEC/CFP2017rate/526.blender_r/526.blender_r.test 14842.00 12463.00 -16.0%
              test-suite :: External/SPEC/CFP2017speed/619.lbm_s/619.lbm_s.test   106.00    79.00 -25.5%
               test-suite :: External/SPEC/CFP2017rate/519.lbm_r/519.lbm_r.test   106.00    79.00 -25.5%
                                                             Geomean difference                    -8.7%

AVX2, O3+LTO
Metric: SLP.NumVectorInstructions

Program                                                                         lhs      rhs      diff
         test-suite :: External/SPEC/CINT2017speed/641.leela_s/641.leela_s.test    22.00    60.00 172.7%
          test-suite :: External/SPEC/CINT2017rate/541.leela_r/541.leela_r.test    22.00    60.00 172.7%
 test-suite :: External/SPEC/CINT2017speed/600.perlbench_s/600.perlbench_s.test    68.00    72.00   5.9%
  test-suite :: External/SPEC/CINT2017rate/500.perlbench_r/500.perlbench_r.test    68.00    72.00   5.9%
              test-suite :: External/SPEC/CINT2017rate/505.mcf_r/505.mcf_r.test    11.00    11.00   0.0%
 test-suite :: External/SPEC/CINT2017speed/631.deepsjeng_s/631.deepsjeng_s.test    10.00    10.00   0.0%
 test-suite :: External/SPEC/CINT2017speed/623.xalancbmk_s/623.xalancbmk_s.test  3396.00  3396.00   0.0%
             test-suite :: External/SPEC/CINT2017speed/605.mcf_s/605.mcf_s.test    11.00    11.00   0.0%
  test-suite :: External/SPEC/CINT2017rate/531.deepsjeng_r/531.deepsjeng_r.test    10.00    10.00   0.0%
  test-suite :: External/SPEC/CINT2017rate/523.xalancbmk_r/523.xalancbmk_r.test  3396.00  3396.00   0.0%
     test-suite :: External/SPEC/CINT2017speed/620.omnetpp_s/620.omnetpp_s.test   499.00   497.00  -0.4%
      test-suite :: External/SPEC/CINT2017rate/520.omnetpp_r/520.omnetpp_r.test   499.00   497.00  -0.4%
            test-suite :: External/SPEC/CINT2017rate/525.x264_r/525.x264_r.test   838.00   826.00  -1.4%
           test-suite :: External/SPEC/CINT2017speed/625.x264_s/625.x264_s.test   838.00   826.00  -1.4%
              test-suite :: External/SPEC/CFP2017speed/619.lbm_s/619.lbm_s.test   131.00   127.00  -3.1%
               test-suite :: External/SPEC/CFP2017rate/519.lbm_r/519.lbm_r.test   131.00   127.00  -3.1%
         test-suite :: External/SPEC/CFP2017rate/511.povray_r/511.povray_r.test  6094.00  5906.00  -3.1%
             test-suite :: External/SPEC/CFP2017rate/508.namd_r/508.namd_r.test  8734.00  8452.00  -3.2%
      test-suite :: External/SPEC/CFP2017speed/638.imagick_s/638.imagick_s.test  2867.00  2712.00  -5.4%
       test-suite :: External/SPEC/CFP2017rate/538.imagick_r/538.imagick_r.test  2867.00  2712.00  -5.4%
       test-suite :: External/SPEC/CFP2017rate/526.blender_r/526.blender_r.test 16026.00 14753.00  -7.9%
              test-suite :: External/SPEC/CINT2017rate/502.gcc_r/502.gcc_r.test   686.00   621.00  -9.5%
             test-suite :: External/SPEC/CINT2017speed/602.gcc_s/602.gcc_s.test   686.00   621.00  -9.5%
         test-suite :: External/SPEC/CFP2017rate/510.parest_r/510.parest_r.test 16241.00 13941.00 -14.2%
              test-suite :: External/SPEC/CFP2017speed/644.nab_s/644.nab_s.test   559.00   473.00 -15.4%
               test-suite :: External/SPEC/CFP2017rate/544.nab_r/544.nab_r.test   559.00   473.00 -15.4%
                                                             Geomean difference                     4.2%

Will update the patch soon.

Rework, bug fixes, rebase

This is an integral patch, going to split it into several smaller patches.

Harbormaster completed remote builds in B88714: Diff 322816.Feb 10 2021, 8:03 PM

Btw, how could it be explained NumVectorInstructions stat reducing after this patch?

In D57059#2553914, @ABataev wrote:

Extra numbers:

AVX512, O3+LTO, -march=native
Metric: SLP.NumVectorInstructions

...
      test-suite :: External/SPEC/CFP2017speed/638.imagick_s/638.imagick_s.test  3996.00  3563.00 -10.8%
       test-suite :: External/SPEC/CFP2017rate/538.imagick_r/538.imagick_r.test  3996.00  3563.00 -10.8%
              test-suite :: External/SPEC/CINT2017rate/502.gcc_r/502.gcc_r.test   862.00   767.00 -11.0%
             test-suite :: External/SPEC/CINT2017speed/602.gcc_s/602.gcc_s.test   862.00   767.00 -11.0%
      test-suite :: External/SPEC/CINT2017rate/520.omnetpp_r/520.omnetpp_r.test   524.00   463.00 -11.6%
...

In D57059#2555233, @ABataev wrote:

This is an integral patch, going to split it into several smaller patches.

Are you planning to send for review all of this patches or just to commit them after this integral review? I.e. should we go on with review here?

In D57059#2578657, @anton-afanasyev wrote:

Btw, how could it be explained NumVectorInstructions stat reducing after this patch?

In D57059#2553914, @ABataev wrote:

Extra numbers:

AVX512, O3+LTO, -march=native
Metric: SLP.NumVectorInstructions

...
      test-suite :: External/SPEC/CFP2017speed/638.imagick_s/638.imagick_s.test  3996.00  3563.00 -10.8%
       test-suite :: External/SPEC/CFP2017rate/538.imagick_r/538.imagick_r.test  3996.00  3563.00 -10.8%
              test-suite :: External/SPEC/CINT2017rate/502.gcc_r/502.gcc_r.test   862.00   767.00 -11.0%
             test-suite :: External/SPEC/CINT2017speed/602.gcc_s/602.gcc_s.test   862.00   767.00 -11.0%
      test-suite :: External/SPEC/CINT2017rate/520.omnetpp_r/520.omnetpp_r.test   524.00   463.00 -11.6%
...

Actually, it is not reducing. This is how test-suite python script works. So, here lhs - number of instructions after this patch, rhs - before. And the less relative number, the more vector instructions we actually generate.

In D57059#2578661, @anton-afanasyev wrote:

In D57059#2555233, @ABataev wrote:

This is an integral patch, going to split it into several smaller patches.

Are you planning to send for review all of this patches or just to commit them after this integral review? I.e. should we go on with review here?

You can publish your comments here, no need to wait for small patches.

Actually, it is not reducing. This is how test-suite python script works. So, here lhs - number of instructions after this patch, rhs - before. And the less relative number, the more vector instructions we actually generate.

Oh, I see. There are still several reducing cases though.

In D57059#2578770, @anton-afanasyev wrote:

Actually, it is not reducing. This is how test-suite python script works. So, here lhs - number of instructions after this patch, rhs - before. And the less relative number, the more vector instructions we actually generate.

Oh, I see. There are still several reducing cases though.

Actually, no. I compared the resulting IR files for these cases - they are absolutely the same as before. It is just we generating fewer ExtractElement/InsertElement instructions and directly emit Shuffle instructions in some cases. That's why it seems it generates fewer vector instructions though it is not.

ABataev added a child revision: D97406: [Instcombiner]Improve emission of logical or/and reductions..Feb 24 2021, 1:06 PM

ABataev mentioned this in rG04ba80ca4dee: [Instcombiner]Improve emission of logical or/and reductions..Mar 4 2021, 8:02 AM

Rebase

Harbormaster completed remote builds in B94525: Diff 331650.Mar 18 2021, 1:08 PM

Removed logical reductions conversion code.

Harbormaster completed remote builds in B94702: Diff 331873.Mar 19 2021, 8:30 AM

ABataev mentioned this in D98967: [Analysis]Add getPointersDiff function to improve compile time..Mar 19 2021, 10:33 AM

ABataev mentioned this in rG065a14a12d26: [Analysis]Add getPointersDiff function to improve compile time..Mar 23 2021, 12:59 PM

ABataev mentioned this in rG99203f2004d0: [Analysis]Add getPointersDiff function to improve compile time..Mar 23 2021, 2:26 PM

Rebase

Harbormaster completed remote builds in B95477: Diff 332970.Mar 24 2021, 11:00 AM

Rebase

Harbormaster completed remote builds in B96712: Diff 334682.Apr 1 2021, 7:42 AM

What is the status of this patch? any blockers? or just lack of reviewers?

In D57059#2666065, @xbolva00 wrote:

What is the status of this patch? any blockers? or just lack of reviewers?

The patch is just too big, I'm splitting it into smaller chunks and commit them step by step. The final chunk would be largest one. But before need to commit other improvements that can be separated from the patch.

ABataev mentioned this in D99796: [SLP]Improve vectorization of the CmpInst instructions..Apr 2 2021, 8:25 AM

ABataev mentioned this in rG00a84f9a7f89: [SLP]Improve vectorization of the CmpInst instructions..Apr 5 2021, 6:50 AM

Rebase

Harbormaster completed remote builds in B97157: Diff 335310.Apr 5 2021, 12:59 PM

ABataev mentioned this in D99980: [SLP]Improve cost model for the vectorized extractelements..Apr 6 2021, 11:09 AM

ABataev mentioned this in rGe99b98cb1bca: [SLP]Improve cost model for the vectorized extractelements..Apr 22 2021, 7:41 AM

ABataev mentioned this in D101109: [SLP]Improve multinode analysis..Apr 22 2021, 2:01 PM

ABataev mentioned this in D101297: [SLP]Allow masked gathers only if allowed by target..Apr 26 2021, 7:41 AM

ABataev mentioned this in rGb5f64768cfee: [SLP]Allow masked gathers only if allowed by target..May 3 2021, 7:05 AM

ABataev mentioned this in rGfd18547e0721: [SLP]Allow masked gathers only if allowed by target..May 3 2021, 8:07 AM

Hi Alexey! Is this patch ready for reviewing or will other patches be splitted from this one?

In D57059#2766508, @anton-afanasyev wrote:

Hi Alexey! Is this patch ready for reviewing or will other patches be splitted from this one?

Just like I said before, it must be split into several smaller patches. I'm doing this step by step, there is D101109, which is part of this big patch. I want to commit it, then rebase this patch and split it again, there are some other parts that can be committed independently.

ABataev mentioned this in D28907: [SLP] Fix for PR30787: Failure to beneficially vectorize 'copyable' elements in integer binary ops..May 20 2021, 6:31 AM

SjoerdMeijer added a subscriber: SjoerdMeijer.May 20 2021, 6:35 AM

ABataev mentioned this in D103247: [SLP]Allow to reorder nodes with >2 scalar values..May 27 2021, 6:48 AM

Is it worth rebasing this to show the remaining diffs that still need to go in?

In D57059#2785073, @RKSimon wrote:

Is it worth rebasing this to show the remaining diffs that still need to go in?

There were not many commits for it, need to commit some extra patches to fix some regressions, but any way I've started rebasing. We'll try to rebase it next week

ABataev mentioned this in D103458: [SLP]Improve gathering of scalar elements..Jun 1 2021, 6:53 AM

ABataev mentioned this in rG89f3bc7698c5: [SLP]Allow to reorder nodes with >2 scalar values..Jun 3 2021, 10:03 AM

ABataev mentioned this in D103638: [SLP]Improve vectorization of PHI instructions..Jun 3 2021, 11:18 AM

ABataev mentioned this in rGa0086add2e52: [SLP]Improve gathering of scalar elements..Jun 9 2021, 5:24 AM

ABataev mentioned this in D104122: [SLP]Improve vectorization of stores..Jun 11 2021, 8:07 AM

ABataev mentioned this in rG908b7536615e: [SLP]Improve vectorization of PHI instructions..Jun 21 2021, 12:27 PM

anton-afanasyev mentioned this in D105042: [SLP][COST][X86]Improve cost model for masked gather..Jul 6 2021, 3:33 PM

ABataev mentioned this in rGc574d2fbaca4: [SLP]Improve vectorization of stores..Jul 8 2021, 12:49 PM

ABataev mentioned this in D105986: [SLP]Improve vectorization of gathered loads..Jul 14 2021, 7:41 AM

In D57059#2785076, @ABataev wrote:

In D57059#2785073, @RKSimon wrote:

Is it worth rebasing this to show the remaining diffs that still need to go in?

There were not many commits for it, need to commit some extra patches to fix some regressions, but any way I've started rebasing. We'll try to rebase it next week

Any chance that you could refresh this patch with your rebase please? I'm investigating a lot of 'float3' style performance issues at the moment (PR50920, PR51075, PR51091) and I'd like to get a better idea of how close the non-pow2 slp support will get us. Thanks.

In D57059#2877539, @RKSimon wrote:

In D57059#2785076, @ABataev wrote:

In D57059#2785073, @RKSimon wrote:

Is it worth rebasing this to show the remaining diffs that still need to go in?

There were not many commits for it, need to commit some extra patches to fix some regressions, but any way I've started rebasing. We'll try to rebase it next week

Any chance that you could refresh this patch with your rebase please? I'm investigating a lot of 'float3' style performance issues at the moment (PR50920, PR51075, PR51091) and I'd like to get a better idea of how close the non-pow2 slp support will get us. Thanks.

I will try to do it ASAP.

That's awesome thanks, it'll definitely help improve the feedback I can give on the patch series.

RKSimon mentioned this in D106399: [VectorCombine] Widening of partial vector loads.Jul 21 2021, 4:59 AM

Rebase. Did not test it thoroughly, just rebased and fixed test cases.

Harbormaster completed remote builds in B115286: Diff 360420.Jul 21 2021, 6:59 AM

Thank you!

Thank you, checked this patch after rebase, trying to fix PR49933. It works well for it, reported to https://bugs.llvm.org/show_bug.cgi?id=49933.

anton-afanasyev mentioned this in rGdd028c359e09: [SLP][Test] Add tests for PR47624 and PR49933.Sep 4 2021, 3:18 PM

nick added a subscriber: nick.Sep 4 2021, 4:30 PM

vporpo added a subscriber: vporpo.Nov 11 2021, 8:01 PM

ABataev mentioned this in rGbd053769867f: [SLP]Improve multinode analysis..Dec 14 2021, 6:18 AM

ABataev mentioned this in D123516: Fix SLP score for out of order contiguous loads.Apr 12 2022, 11:38 AM

liaolucy added a subscriber: liaolucy.May 17 2022, 7:18 PM

Herald added a project: Restricted Project. · View Herald TranscriptMay 17 2022, 7:18 PM

Herald added subscribers: kosarev, StephenFan. · View Herald Transcript

Current status?

Herald added subscribers: • pcwang-thead, nlopes. · View Herald TranscriptSep 7 2022, 12:53 PM

In D57059#3775369, @xbolva00 wrote:

Current status?

Still requires several patches to commit.

In D57059#3775511, @ABataev wrote:

In D57059#3775369, @xbolva00 wrote:

Current status?

Still requires several patches to commit.

Are those patches linked somewhere?

Herald added a subscriber: wangpc. · View Herald TranscriptAug 8 2023, 6:40 AM

In D57059#4569337, @danilaml wrote:

In D57059#3775511, @ABataev wrote:

In D57059#3775369, @xbolva00 wrote:

Current status?

Still requires several patches to commit.

Are those patches linked somewhere?

Almost all my current SLP patches are related to non-power-of-2 support. But even these patches are not the full list. Need to add several others after.

sunshaoce added a subscriber: sunshaoce.Aug 17 2023, 2:19 AM

Please use GitHub pull requests for new patches. Phabricator shutdown timeline

[SLP] Initial support for the vectorization of the non-power-of-2 vectors.Needs ReviewPublicActions

Details

Diff Detail

Unit TestsFailedView All

Event Timeline

Revision ContentsChangeset List

Diff 360420

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

llvm/test/Transforms/SLPVectorizer/AArch64/gather-root.ll

llvm/test/Transforms/SLPVectorizer/AArch64/horizontal.ll

llvm/test/Transforms/SLPVectorizer/AArch64/transpose-inseltpoison.ll

llvm/test/Transforms/SLPVectorizer/AArch64/transpose.ll

llvm/test/Transforms/SLPVectorizer/AArch64/vectorize-free-extracts-inserts.ll

llvm/test/Transforms/SLPVectorizer/AMDGPU/add_sub_sat-inseltpoison.ll

llvm/test/Transforms/SLPVectorizer/AMDGPU/add_sub_sat.ll

llvm/test/Transforms/SLPVectorizer/SystemZ/pr34619.ll

llvm/test/Transforms/SLPVectorizer/X86/PR35865-inseltpoison.ll

llvm/test/Transforms/SLPVectorizer/X86/PR35865.ll

llvm/test/Transforms/SLPVectorizer/X86/PR39774.ll

llvm/test/Transforms/SLPVectorizer/X86/alternate-calls-inseltpoison.ll

llvm/test/Transforms/SLPVectorizer/X86/alternate-calls.ll

llvm/test/Transforms/SLPVectorizer/X86/alternate-cast-inseltpoison.ll

llvm/test/Transforms/SLPVectorizer/X86/alternate-cast.ll

llvm/test/Transforms/SLPVectorizer/X86/alternate-fp-inseltpoison.ll

llvm/test/Transforms/SLPVectorizer/X86/alternate-fp.ll

llvm/test/Transforms/SLPVectorizer/X86/alternate-int-inseltpoison.ll

llvm/test/Transforms/SLPVectorizer/X86/alternate-int.ll

llvm/test/Transforms/SLPVectorizer/X86/arith-fp-inseltpoison.ll

llvm/test/Transforms/SLPVectorizer/X86/arith-fp.ll

llvm/test/Transforms/SLPVectorizer/X86/blending-shuffle-inseltpoison.ll

llvm/test/Transforms/SLPVectorizer/X86/blending-shuffle.ll

llvm/test/Transforms/SLPVectorizer/X86/cmp_commute-inseltpoison.ll

llvm/test/Transforms/SLPVectorizer/X86/cmp_commute.ll

llvm/test/Transforms/SLPVectorizer/X86/commutativity.ll

llvm/test/Transforms/SLPVectorizer/X86/compare-reduce.ll

llvm/test/Transforms/SLPVectorizer/X86/crash_exceed_scheduling.ll

llvm/test/Transforms/SLPVectorizer/X86/crash_reordering_undefs.ll

llvm/test/Transforms/SLPVectorizer/X86/crash_vectorizeTree.ll

llvm/test/Transforms/SLPVectorizer/X86/cse.ll

llvm/test/Transforms/SLPVectorizer/X86/dot-product.ll

llvm/test/Transforms/SLPVectorizer/X86/extract.ll

llvm/test/Transforms/SLPVectorizer/X86/extractelement.ll

llvm/test/Transforms/SLPVectorizer/X86/fptosi-inseltpoison.ll

llvm/test/Transforms/SLPVectorizer/X86/fptosi.ll

llvm/test/Transforms/SLPVectorizer/X86/fptoui.ll

llvm/test/Transforms/SLPVectorizer/X86/geps-non-pow-2.ll

llvm/test/Transforms/SLPVectorizer/X86/horizontal-minmax.ll

llvm/test/Transforms/SLPVectorizer/X86/insert-element-build-vector-inseltpoison.ll

llvm/test/Transforms/SLPVectorizer/X86/insert-element-build-vector.ll

llvm/test/Transforms/SLPVectorizer/X86/load-merge-inseltpoison.ll

llvm/test/Transforms/SLPVectorizer/X86/load-merge.ll

llvm/test/Transforms/SLPVectorizer/X86/lookahead.ll

llvm/test/Transforms/SLPVectorizer/X86/minimum-sizes.ll

llvm/test/Transforms/SLPVectorizer/X86/no_alternate_divrem.ll

llvm/test/Transforms/SLPVectorizer/X86/operandorder.ll

llvm/test/Transforms/SLPVectorizer/X86/ordering-bug.ll

llvm/test/Transforms/SLPVectorizer/X86/phi.ll

llvm/test/Transforms/SLPVectorizer/X86/pr35497.ll

llvm/test/Transforms/SLPVectorizer/X86/pr42022.ll

llvm/test/Transforms/SLPVectorizer/X86/pr47623.ll

llvm/test/Transforms/SLPVectorizer/X86/pr47629-inseltpoison.ll

llvm/test/Transforms/SLPVectorizer/X86/pr47629.ll

llvm/test/Transforms/SLPVectorizer/X86/pr49081.ll

llvm/test/Transforms/SLPVectorizer/X86/reduction-logical.ll

llvm/test/Transforms/SLPVectorizer/X86/reorder_repeated_ops.ll

llvm/test/Transforms/SLPVectorizer/X86/resched.ll

llvm/test/Transforms/SLPVectorizer/X86/revectorized_rdx_crash.ll

llvm/test/Transforms/SLPVectorizer/X86/rgb_phi.ll

llvm/test/Transforms/SLPVectorizer/X86/schedule-bundle.ll

llvm/test/Transforms/SLPVectorizer/X86/shrink_after_reorder.ll

llvm/test/Transforms/SLPVectorizer/X86/supernode.ll

llvm/test/Transforms/SLPVectorizer/X86/value-bug-inseltpoison.ll

llvm/test/Transforms/SLPVectorizer/X86/value-bug.ll

llvm/test/Transforms/SLPVectorizer/X86/vec_list_bias-inseltpoison.ll

llvm/test/Transforms/SLPVectorizer/X86/vec_list_bias.ll

llvm/test/Transforms/SLPVectorizer/X86/vectorize-reorder-reuse.ll

[SLP] Initial support for the vectorization of the non-power-of-2 vectors.
Needs ReviewPublic
Actions

Unit TestsFailed
View All

Revision Contents
Changeset List