GH-36176: [C++] Fix regression for single-key Table sorting

[benibus]

Rationale for this change

Fixes a regression introduced in #35727.

What changes are included in this PR?

Re-implements a branch in the Table sorter that defers to the ChunkedArray sorter for single sort keys.

Are these changes tested?

Covered by existing tests.

Are there any user-facing changes?

No.

• Closes: [C++] Large regression in single-key table sort performance #36176

[benibus]

Local benchmark results (approx. best-case)

Before

-------------------------------------------------------------------------------------------------------
Benchmark                                             Time             CPU   Iterations UserCounters...
-------------------------------------------------------------------------------------------------------
TableSortIndicesInt64Narrow/1048576/100/1/32  367381664 ns    367357156 ns            2 chunks=32 columns=1 items_per_second=2.85438M/s null_percent=1
TableSortIndicesInt64Narrow/1048576/4/1/32    281608431 ns    281605789 ns            2 chunks=32 columns=1 items_per_second=3.72356M/s null_percent=25
TableSortIndicesInt64Narrow/1048576/0/1/32    354414823 ns    354413272 ns            2 chunks=32 columns=1 items_per_second=2.95863M/s null_percent=0
TableSortIndicesInt64Narrow/1048576/100/1/4   315257284 ns    315253535 ns            2 chunks=4 columns=1 items_per_second=3.32614M/s null_percent=1
TableSortIndicesInt64Narrow/1048576/4/1/4     240298213 ns    240291188 ns            3 chunks=4 columns=1 items_per_second=4.36377M/s null_percent=25
TableSortIndicesInt64Narrow/1048576/0/1/4     311795614 ns    311790481 ns            2 chunks=4 columns=1 items_per_second=3.36308M/s null_percent=0
TableSortIndicesInt64Narrow/1048576/100/1/1   282068938 ns    282067950 ns            2 chunks=1 columns=1 items_per_second=3.71746M/s null_percent=1
TableSortIndicesInt64Narrow/1048576/4/1/1     210222675 ns    210216608 ns            3 chunks=1 columns=1 items_per_second=4.98807M/s null_percent=25
TableSortIndicesInt64Narrow/1048576/0/1/1     281105329 ns    281101193 ns            2 chunks=1 columns=1 items_per_second=3.73024M/s null_percent=0

After

-------------------------------------------------------------------------------------------------------
Benchmark                                             Time             CPU   Iterations UserCounters...
-------------------------------------------------------------------------------------------------------
TableSortIndicesInt64Narrow/1048576/100/1/32  148802109 ns    148797635 ns            4 chunks=32 columns=1 items_per_second=7.04699M/s null_percent=1
TableSortIndicesInt64Narrow/1048576/4/1/32    125102993 ns    125100621 ns            6 chunks=32 columns=1 items_per_second=8.38186M/s null_percent=25
TableSortIndicesInt64Narrow/1048576/0/1/32    147291350 ns    147287710 ns            5 chunks=32 columns=1 items_per_second=7.11924M/s null_percent=0
TableSortIndicesInt64Narrow/1048576/100/1/4    72243393 ns     72241902 ns           10 chunks=4 columns=1 items_per_second=14.5148M/s null_percent=1
TableSortIndicesInt64Narrow/1048576/4/1/4      66780395 ns     66779462 ns           11 chunks=4 columns=1 items_per_second=15.7021M/s null_percent=25
TableSortIndicesInt64Narrow/1048576/0/1/4      70065128 ns     70063740 ns           10 chunks=4 columns=1 items_per_second=14.966M/s null_percent=0
TableSortIndicesInt64Narrow/1048576/100/1/1    19336900 ns     19335762 ns           36 chunks=1 columns=1 items_per_second=54.2299M/s null_percent=1
TableSortIndicesInt64Narrow/1048576/4/1/1      25458573 ns     25457697 ns           27 chunks=1 columns=1 items_per_second=41.189M/s null_percent=25
TableSortIndicesInt64Narrow/1048576/0/1/1      17580368 ns     17577897 ns           40 chunks=1 columns=1 items_per_second=59.6531M/s null_percent=0

[benibus]

cc @pitrou

[pitrou]

This fixes the regression here.

[pitrou]

Thanks for the quick fix @benibus !

[conbench-apache-arrow]

Conbench analyzed the 6 benchmark runs on commit 8a4c19bd.

There were 7 benchmark results indicating a performance regression:

• Commit Run on ursa-thinkcentre-m75q at 2023-06-21 07:09:25Z
  • params=1048576/0, source=cpp-micro, suite=arrow-acero-aggregate-benchmark
  • source=java-micro, suite=arrow.memory.util.ByteFunctionHelpersBenchmarks
• and 5 more (see the report linked below)

The full Conbench report has more details.