GH-49888: [C++][Compute] Fix count for run-end encoded arrays with nulls

[fenfeng9]

Rationale for this change

The count kernel used GetNullCount(), which reports the physical null count. For run-end encoded arrays, this ignored nulls in the encoded values child.

What changes are included in this PR?

Use ComputeLogicalNullCount() in the count kernel so run-end encoded arrays are counted correctly. Add C++ and Python tests for this case.

Are these changes tested?

Yes.

Are there any user-facing changes?

No.

• GitHub Issue: [C++][Compute] count kernel miscounts when run-end encoded array contains null #49888

[fenfeng9]

The added tests use a simplified version of the reproducer from the issue.

Reproduce

import pyarrow as pa
import pyarrow.compute as pc

array = pa.array([1, None])
encoded = pc.run_end_encode(array)

print(f"plain only_null: ", pc.count(array, mode="only_null"))
print(f"run_end_encode only_null: ", pc.count(encoded, mode="only_null"))

Result

plain only_null:  1
run_end_encode only_null:  0

[tadeja]

Thank you for the fix, @fenfeng9 ! 🌞 A few additional test suggestions from my end as inline comments.
The MinGW64 CI failures look unrelated. There were some overlapping failures on main.

[fenfeng9]

Thanks for the suggestion. I updated the C++ and Python tests.

[fenfeng9]

The original behavior is:

Reproduce

import pyarrow as pa
import pyarrow.compute as pc


array = pa.array([1, 1, None, None, None, 2, 2, 2, None, 3])
encoded = pc.run_end_encode(array)
# Logical slice: [None, None, 2, 2, 2, None].
slice_plain = array.slice(3, 6)
slice_encoded = encoded.slice(3, 6)

print("pyarrow:", pa.__version__)
print()

print(f"{'case':<12} {'only_valid':>10} {'only_null':>10} {'all':>6}")
for name, value in [
    ("plain", array),
    ("ree", encoded),
    ("slice plain", slice_plain),
    ("slice ree", slice_encoded),
]:
    print(
        f"{name:<12} "
        f"{pc.count(value, mode='only_valid').as_py():>10} "
        f"{pc.count(value, mode='only_null').as_py():>10} "
        f"{pc.count(value, mode='all').as_py():>6}"
    )

Result

pyarrow: 24.0.0

case         only_valid  only_null    all
plain                 6          4     10
ree                  10          0     10
slice plain           3          3      6
slice ree             6          0      6

[fenfeng9]

The test failures look unrelated.

[pitrou]

Thank you @fenfeng9 !

[pitrou]

And thanks @tadeja for the useful reviewing!

[conbench-apache-arrow]

After merging your PR, Conbench analyzed the 0 benchmarking runs that have been run so far on merge-commit 1b3f313.

None of the specified runs were found on the Conbench server.

The full Conbench report has more details.