[C++] ComputeLogicalNullCount crashes for dictionary arrays with null value type

[goel-skd]

Describe the bug, including details regarding any error messages, version, and platform.

Describe the bug, including details regarding any error messages, version, and platform

Array::ComputeLogicalNullCount segfaults for dictionary arrays whose value type is null(), even though the array passes full validation:

auto type = dictionary(int32(), null());
auto array = DictArrayFromJSON(type, R"([0, 0, null])", "[null]");
ASSERT_OK(array->ValidateFull());
array->ComputeLogicalNullCount();  // segfault

The values array of such a dictionary is a NullArray, which reports null_count == length without having a validity bitmap (buffers[0] is null). The early-out in dict_util::LogicalNullCount only triggers when the dictionary's null count is 0, so the per-index loop unconditionally reads the null validity-bitmap pointer:

https://github.com/apache/arrow/blob/main/cpp/src/arrow/util/dict_util.cc#L56-L59

Note that element-wise Array::IsValid handles the same data correctly (it falls back to null_count == length when there is no bitmap), so the aggregate and per-element views of logical nullness disagree.

Found while working on GH-50338 (#50347).

The crash is also reachable from the bindings through the C++ count kernel, which uses ComputeLogicalNullCount when counting nulls (e.g. pyarrow.compute.count(arr, mode="only_null") on such an array).

Component(s)

C++

Component(s)

C++

[goel-skd]

take