[ARROW-15215] [C++] Consolidate kernel data-copy utilities between replace_with_mask, case_when, coalesce, choose, fill_null_forward, fill_null_backward

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 8.0.0
• Component/s: C++
• Labels: good-second-issue pull-request-available
• External issue URL: https://github.com/apache/arrow/issues/30715
• Assignee: Jabari Booker
• Reporter: David Li
• Created: 29/Dec/21 13:49
• Updated: 11/Jan/23 08:45
• Resolved: 12/Feb/22 17:57

Description

All six kernels use two sets of otherwise very similar kernel utilities for copying slices of an array into an output array. However, there's no reason they can't use the same utilities.

The first set are here: "CopyFixedWidth" https://github.com/apache/arrow/blob/bd356295f6beaba744a2c6b498455701f53a64f8/cpp/src/arrow/compute/kernels/scalar_if_else.cc#L1282-L1284

The second set are here: "ReplaceWithMask::CopyData" https://github.com/apache/arrow/blob/bd356295f6beaba744a2c6b498455701f53a64f8/cpp/src/arrow/compute/kernels/vector_replace.cc#L208-L209 (This is a little confusing because the utilities are intertwined into the kernel implementation)

They would need to be moved into a new header to share them between the codegen units. Also, their interfaces would need to be consolidated.

Additionally, the utilities may be excessively verbose, or generate too much code for what they do. For instance, some of the utilities are templated out for every Arrow type. Instead, we could replace all instantiations for numbers, decimals, temporal types, and so on with a single one for FixedWidthType (an abstract base class). Care should be taken to evaluate the benchmarks for these kernels to ensure there is not a regression.

Issue Links

• relates to: ARROW-14964 [C++] Restructure internal compute/kernel utilities
• links to: GitHub Pull Request #12379