DO NOT MERGE: EXP: [C++] VisitTwoBitRuns changes

[pitrou]

Based on #50281

[pitrou]

@ursabot please benchmark lang=C++

[rok]

Benchmark runs are scheduled for commit 710cedd. Watch https://buildkite.com/apache-arrow and https://conbench.arrow-dev.org for updates. A comment will be posted here when the runs are complete.

[conbench-apache-arrow]

Thanks for your patience. Conbench analyzed the 2 benchmarking runs that have been run so far on PR commit 710cedd.

There were 48 benchmark results indicating a performance regression:

• Pull Request Run on amd64-m5-4xlarge-linux at 2026-07-06 16:18:36Z
  • ArrayArrayKernel (C++) with params=<AddChecked, Int64Type>/size:1048576/inverse_null_proportion:0, source=cpp-micro, suite=arrow-compute-scalar-arithmetic-benchmark
  • ArrayArrayKernel (C++) with params=<SubtractChecked, UInt8Type>/size:1048576/inverse_null_proportion:0, source=cpp-micro, suite=arrow-compute-scalar-arithmetic-benchmark
• and 46 more (see the report linked below)

The full Conbench report has more details.

[HuaHuaY]

The results were somewhat unexpected. My standard local development environment is a MacBook, but I was unable to reproduce the test-mac-arm benchmark results.

[pitrou]

Yes, compiler code generation can be quite fickle. But overall it seems to me that those benchmark numbers point to a significant improvement.

[HuaHuaY]

But overall it seems to me that those benchmark numbers point to a significant improvement.

Could you please explain how to interpret this report? It seems to me that there are many performance regressions when there are no null values.

Incidentally, adding ARROW_FORCE_INLINE to VisitSetBitRuns provides some optimization for ArrayArrayKernel in my local environment. However, I'm unsure if this optimization always applies to other scenarios.

[pitrou]

Could you please explain how to interpret this report? It seems to me that there are many performance regressions when there are no null values.

The ArrayArrayKernel benchmarks but also the ScalarTemporalBinary benchmarks exercise VisitTwoArrayValuesInline. The fact that some show regressions and some improvements hints that the compiler is not always able to do a good job here.

However, there are also a number of large improvements on the Filter benchmarks.

Incidentally, adding ARROW_FORCE_INLINE to VisitSetBitRuns provides some optimization for ArrayArrayKernel in my local environment.

Hmm, I could try it in this PR.

[pitrou]

@ursabot please benchmark lang=C++

[rok]

Benchmark runs are scheduled for commit b0f622f. Watch https://buildkite.com/apache-arrow and https://conbench.arrow-dev.org for updates. A comment will be posted here when the runs are complete.

[pitrou]

Could you please explain how to interpret this report?

Oh, I forgot: the z-score measures statistical significance. I personally only look at the numbers where abs(z) > 3.
(ideally there should be a way to hide non-significant results @rok )

[rok]

@pitrou yes, having a filter like that sounds like a good idea :)

[conbench-apache-arrow]

Thanks for your patience. Conbench analyzed the 4 benchmarking runs that have been run so far on PR commit b0f622f.

There were 41 benchmark results indicating a performance regression:

• Pull Request Run on amd64-m5-4xlarge-linux at 2026-07-07 09:56:06Z
  • ArrayArrayKernel (C++) with params=<AddChecked, Int16Type>/size:1048576/inverse_null_proportion:0, source=cpp-micro, suite=arrow-compute-scalar-arithmetic-benchmark
  • ArrayArrayKernel (C++) with params=<MultiplyChecked, UInt16Type>/size:1048576/inverse_null_proportion:0, source=cpp-micro, suite=arrow-compute-scalar-arithmetic-benchmark
• and 39 more (see the report linked below)

The full Conbench report has more details.