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

[Daniel Berlin via llvm-commits]

If you can get profiles, it may help.
You may discover this is caused by some non-determinism causing something
to do or not do an expensive thing.


On Mon, Feb 12, 2018 at 11:56 AM, Brian Rzycki via Phabricator <
reviews at reviews.llvm.org> wrote:

> brzycki added a comment.
>
> I'm seeing considerable variance on the tramp3d-v4 compile time test
> before and after the preservation of DT on JumpThreading patch:
>
>   brzycki at cc01 /work/brzycki/test-suite $ ~/git/me/llvm/litdiff
> build.pre_dt_1/results.json build.pre_dt_2/results.json compile_time
>   #  left column: build.pre_dt_1/results.json
>   # right column: build.pre_dt_2/results.json
>   #  metric name: compile_time
>       524.8640 -> 474.9400     [ 10.51%]  CTMark/tramp3d-v4/tramp3d-v4.
> test
>       343.1280 <- 369.4960     [  7.68%]  CTMark/sqlite3/sqlite3.test
>       379.4560 <- 386.0880     [  1.75%]  CTMark/kimwitu++/kc.test
>       250.3400 -> 248.0680     [  0.92%]  CTMark/mafft/pairlocalalign.test
>       583.2080 <- 587.5400     [  0.74%]  CTMark/lencod/lencod.test
>       918.8360 -> 912.4920     [  0.70%]  CTMark/Bullet/bullet.test
>       456.7720 -> 454.7480     [  0.45%]  CTMark/SPASS/SPASS.test
>       367.2440 -> 366.3920     [  0.23%]  CTMark/consumer-typeset/
> consumer-typeset.test
>      1254.5240 -> 1252.3720    [  0.17%]  CTMark/7zip/7zip-benchmark.test
>       490.1600 -> 489.4360     [  0.15%]  CTMark/ClamAV/clamscan.test
>   brzycki at cc01 /work/brzycki/test-suite $ ~/git/me/llvm/litdiff
> build.pre_dt_1/results.json build.pre_dt_2/results.json compile_time
>   #  left column: build.pre_dt_1/results.json
>   # right column: build.pre_dt_2/results.json
>   #  metric name: compile_time
>       524.8640 -> 474.9400     [ 10.51%]  CTMark/tramp3d-v4/tramp3d-v4.
> test
>       343.1280 <- 369.4960     [  7.68%]  CTMark/sqlite3/sqlite3.test
>       379.4560 <- 386.0880     [  1.75%]  CTMark/kimwitu++/kc.test
>       250.3400 -> 248.0680     [  0.92%]  CTMark/mafft/pairlocalalign.test
>       583.2080 <- 587.5400     [  0.74%]  CTMark/lencod/lencod.test
>       918.8360 -> 912.4920     [  0.70%]  CTMark/Bullet/bullet.test
>       456.7720 -> 454.7480     [  0.45%]  CTMark/SPASS/SPASS.test
>       367.2440 -> 366.3920     [  0.23%]  CTMark/consumer-typeset/
> consumer-typeset.test
>      1254.5240 -> 1252.3720    [  0.17%]  CTMark/7zip/7zip-benchmark.test
>       490.1600 -> 489.4360     [  0.15%]  CTMark/ClamAV/clamscan.test
>   brzycki at cc01 /work/brzycki/test-suite $ ~/git/me/llvm/litdiff
> build.0cb78d41fca7bac_1/results.json build.0cb78d41fca7bac_2/results.json
> compile_time
>   #  left column: build.0cb78d41fca7bac_1/results.json
>   # right column: build.0cb78d41fca7bac_2/results.json
>   #  metric name: compile_time
>       535.8520 -> 483.3720     [ 10.86%]  CTMark/tramp3d-v4/tramp3d-v4.
> test
>       372.9000 <- 378.4200     [  1.48%]  CTMark/sqlite3/sqlite3.test
>       256.3040 -> 253.7600     [  1.00%]  CTMark/mafft/pairlocalalign.test
>       387.8480 -> 385.0480     [  0.73%]  CTMark/kimwitu++/kc.test
>       369.8760 <- 372.3440     [  0.67%]  CTMark/consumer-typeset/
> consumer-typeset.test
>       458.8480 <- 461.5720     [  0.59%]  CTMark/SPASS/SPASS.test
>       590.1440 -> 588.8520     [  0.22%]  CTMark/lencod/lencod.test
>       499.0920 <- 499.9840     [  0.18%]  CTMark/ClamAV/clamscan.test
>      1270.1200 -> 1267.9480    [  0.17%]  CTMark/7zip/7zip-benchmark.test
>       928.7720 <- 928.8000     [  0.00%]  CTMark/Bullet/bullet.test
>   brzycki at cc01 /work/brzycki/test-suite $ ~/git/me/llvm/litdiff
> build.D42717_1/results.json build.D42717_2/results.json compile_time
>   #  left column: build.D42717_1/results.json
>   # right column: build.D42717_2/results.json
>   #  metric name: compile_time
>       494.1120 -> 487.4120     [  1.37%]  CTMark/tramp3d-v4/tramp3d-v4.
> test
>       369.8320 <- 373.5960     [  1.02%]  CTMark/consumer-typeset/
> consumer-typeset.test
>       386.5280 <- 390.0800     [  0.92%]  CTMark/kimwitu++/kc.test
>       501.9440 -> 497.7200     [  0.85%]  CTMark/ClamAV/clamscan.test
>       255.9640 -> 254.5120     [  0.57%]  CTMark/mafft/pairlocalalign.test
>       347.9000 -> 346.5880     [  0.38%]  CTMark/sqlite3/sqlite3.test
>      1267.7040 <- 1269.8640    [  0.17%]  CTMark/7zip/7zip-benchmark.test
>       462.5080 <- 462.8000     [  0.06%]  CTMark/SPASS/SPASS.test
>       923.5880 -> 923.0320     [  0.06%]  CTMark/Bullet/bullet.test
>       589.6800 <- 589.8840     [  0.03%]  CTMark/lencod/lencod.test
>
> All of these are the same compiler compared against itself on as close to
> a quiescent x86_64 machine as I could find. The https://reviews.llvm.org/
> D42717 version has been rebased on top of llvm-project Git sha
> `0cb78d41fca7bac`.  I'm going to run more iterations of the `pre_dt`
> compiler (llvm-project sha `788d5233ad1f2` from January 12) to see if the
> variance persists.
>
> If it does then this current patch does not meaningfully alter the compile
> time of `tramp3d-v4`.
>
>
> https://reviews.llvm.org/D42717
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180212/39b23cf1/attachment.html>