[PATCH] D42717: [JumpThreading] sync DT for LVI analysis (PR 36133)

[Brian Rzycki via Phabricator via llvm-commits]

brzycki added a comment.

@kuhar I had a bug in my test-suite runner script and just fixed it. The new results are require more analysis, particularly sqlite.

  ## COMPARING TIP OF LLVM-PROJECT VS THE SAME SHA WITH THE PROPOSED PATCH FOR D42717
  brzycki at cc01 /work/brzycki/test-suite $ ~/git/me/llvm/litdiff build.upstream/results.json build.D42717/results.json  compile_time
  #  left column: build.upstream/results.json
  # right column: build.D42717/results.json
  #  metric name: compile_time
      336.3600 <- 377.3240     [12.18%]    CTMark/sqlite3/sqlite3.test
      493.8600 -> 486.7000     [1.47%]    CTMark/tramp3d-v4/tramp3d-v4.test
      497.1560 <- 500.4880     [0.67%]    CTMark/ClamAV/clamscan.test
     1265.6800 <- 1272.0480    [0.50%]    CTMark/7zip/7zip-benchmark.test
      589.8000 <- 591.0520     [0.21%]    CTMark/lencod/lencod.test
      387.2960 -> 386.5120     [0.20%]    CTMark/kimwitu++/kc.test
      459.5960 <- 460.3640     [0.17%]    CTMark/SPASS/SPASS.test
      256.1360 <- 256.4000     [0.10%]    CTMark/mafft/pairlocalalign.test
      373.0960 -> 372.7520     [0.09%]    CTMark/consumer-typeset/consumer-typeset.test
      924.1160 -> 923.4120     [0.08%]    CTMark/Bullet/bullet.test
  
  ## COMPARING TIP OF LLVM-PROJECT VS THE SAME SHA WITH THE PROPOSED PATCH FOR D34135
  brzycki at cc01 /work/brzycki/test-suite $ ~/git/me/llvm/litdiff build.upstream/results.json build.D34135/results.json  compile_time
  #  left column: build.upstream/results.json
  # right column: build.D34135/results.json
  #  metric name: compile_time
      336.3600 <- 387.9720     [15.34%]    CTMark/sqlite3/sqlite3.test
      493.8600 -> 485.5160     [1.72%]    CTMark/tramp3d-v4/tramp3d-v4.test
      256.1360 -> 252.7320     [1.35%]    CTMark/mafft/pairlocalalign.test
      387.2960 <- 392.0080     [1.22%]    CTMark/kimwitu++/kc.test
      497.1560 <- 499.8040     [0.53%]    CTMark/ClamAV/clamscan.test
      373.0960 -> 371.5320     [0.42%]    CTMark/consumer-typeset/consumer-typeset.test
      459.5960 <- 461.2160     [0.35%]    CTMark/SPASS/SPASS.test
      589.8000 -> 588.1520     [0.28%]    CTMark/lencod/lencod.test
     1265.6800 <- 1268.2080    [0.20%]    CTMark/7zip/7zip-benchmark.test
      924.1160 -> 923.6120     [0.05%]    CTMark/Bullet/bullet.test
  
  ## COMPARING THE PROPOSED PATCH FOR D34135 VS THE SAME SHA WITH THE PROPOSED PATCH FOR D42717
  brzycki at cc01 /work/brzycki/test-suite $ ~/git/me/llvm/litdiff build.D34135/results.json build.D42717/results.json  compile_time
  #  left column: build.D34135/results.json
  # right column: build.D42717/results.json
  #  metric name: compile_time
      387.9720 -> 377.3240     [2.82%]    CTMark/sqlite3/sqlite3.test
      252.7320 <- 256.4000     [1.45%]    CTMark/mafft/pairlocalalign.test
      392.0080 -> 386.5120     [1.42%]    CTMark/kimwitu++/kc.test
      588.1520 <- 591.0520     [0.49%]    CTMark/lencod/lencod.test
      371.5320 <- 372.7520     [0.33%]    CTMark/consumer-typeset/consumer-typeset.test
     1268.2080 <- 1272.0480    [0.30%]    CTMark/7zip/7zip-benchmark.test
      485.5160 <- 486.7000     [0.24%]    CTMark/tramp3d-v4/tramp3d-v4.test
      461.2160 -> 460.3640     [0.19%]    CTMark/SPASS/SPASS.test
      499.8040 <- 500.4880     [0.14%]    CTMark/ClamAV/clamscan.test
      923.6120 -> 923.4120     [0.02%]    CTMark/Bullet/bullet.test

All the other benchmarks are ~1% deltas and I'm not very concerned with them. The sqlite benchmark loses 12% with this proposed patch and an additional 3% with the patch proposed for https://reviews.llvm.org/D34135. I'm going to do another run of the upstream compiler to "check the cable" but I'm fairly certain these results are accurate.

My conjecture is the DT edge calls on a large tree are very expensive and I'm not sure how to mitigate this...


Repository:
  rL LLVM

https://reviews.llvm.org/D42717