[PATCH] D105996: [AArch64] Enable Upper bound unrolling universally

[JinGu Kang via Phabricator via llvm-commits]

jaykang10 added a comment.

I have checked the performance number from SPEC benchmarks on neoverse-n1 core.

  spec2006 O3 lto	
  Benchmark	score improvement(%)
  400.perlbench	0.35011018
  401.bzip2	-3.003732177
  403.gcc	0.171077131
  429.mcf	2.115844504
  445.gobmk	0.201857868
  456.hmmer	-0.143194012
  458.sjeng	0.207366997
  462.libquantum	3.935575739
  464.h264ref	0.167041326
  471.omnetpp	1.355350874
  473.astar	0.733528789
  483.xalancbmk	1.064585199
  	
  spec2006 Ofast	
  Benchmark	score improvement(%)
  400.perlbench	-0.213126131
  401.bzip2	3.244142727
  403.gcc	1.345302812
  429.mcf	0.194343375
  445.gobmk	0.766024658
  456.hmmer	-0.096760298
  458.sjeng	0.078527561
  462.libquantum	0.214315254
  464.h264ref	0.957874944
  471.omnetpp	0.057501946
  473.astar	0.296876709
  483.xalancbmk	0.261506157
  	
  spec2017 O3 lto	
  Benchmark	score improvement(%)
  500.perlbench_r	1.094911941
  502.gcc_r	0.559111539
  505.mcf_r	0
  520.omnetpp_r	-0.140907845
  523.xalancbmk_r	0.564247335
  525.x264_r	0.198455239
  531.deepsjeng_r	-0.081793385
  541.leela_r	0.038333017
  548.exchange2_r	0.009064456
  557.xz_r	0.167682904
  	
  spec2017 Ofast	
  Benchmark	score improvement(%)
  500.perlbench_r	0.708957995
  502.gcc_r	0.871004714
  505.mcf_r	0
  520.omnetpp_r	1.416668547
  523.xalancbmk_r	-0.30611
  525.x264_r	0.20238427
  531.deepsjeng_r	0.550866022
  541.leela_r	0.095203276
  548.exchange2_r	0.002682641
  557.xz_r	0.896193368

There could be noise... I had problem to get the performance number on other core. It looks it could be better to turn this option for only neoverse-n1 rather than turning it on universally.
Let me update this patch.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D105996/new/

https://reviews.llvm.org/D105996