Initial commit

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/algorithm.h

@@ -0,0 +1,33 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include "arrow/result.h"
+
+namespace arrow {
+
+template <typename InputIterator, typename OutputIterator, typename UnaryOperation>
+Status MaybeTransform(InputIterator first, InputIterator last, OutputIterator out,
+                      UnaryOperation unary_op) {
+  for (; first != last; ++first, (void)++out) {
+    ARROW_ASSIGN_OR_RAISE(*out, unary_op(*first));
+  }
+  return Status::OK();
+}
+
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/align_util.h

@@ -0,0 +1,221 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <algorithm>
+
+#include "arrow/memory_pool.h"
+#include "arrow/type_fwd.h"
+#include "arrow/util/bit_util.h"
+
+namespace arrow {
+namespace internal {
+
+struct BitmapWordAlignParams {
+  int64_t leading_bits;
+  int64_t trailing_bits;
+  int64_t trailing_bit_offset;
+  const uint8_t* aligned_start;
+  int64_t aligned_bits;
+  int64_t aligned_words;
+};
+
+// Compute parameters for accessing a bitmap using aligned word instructions.
+// The returned parameters describe:
+// - a leading area of size `leading_bits` before the aligned words
+// - a word-aligned area of size `aligned_bits`
+// - a trailing area of size `trailing_bits` after the aligned words
+template <uint64_t ALIGN_IN_BYTES>
+inline BitmapWordAlignParams BitmapWordAlign(const uint8_t* data, int64_t bit_offset,
+                                             int64_t length) {
+  static_assert(bit_util::IsPowerOf2(ALIGN_IN_BYTES),
+                "ALIGN_IN_BYTES should be a positive power of two");
+  constexpr uint64_t ALIGN_IN_BITS = ALIGN_IN_BYTES * 8;
+
+  BitmapWordAlignParams p;
+
+  // Compute a "bit address" that we can align up to ALIGN_IN_BITS.
+  // We don't care about losing the upper bits since we are only interested in the
+  // difference between both addresses.
+  const uint64_t bit_addr =
+      reinterpret_cast<size_t>(data) * 8 + static_cast<uint64_t>(bit_offset);
+  const uint64_t aligned_bit_addr = bit_util::RoundUpToPowerOf2(bit_addr, ALIGN_IN_BITS);
+
+  p.leading_bits = std::min<int64_t>(length, aligned_bit_addr - bit_addr);
+  p.aligned_words = (length - p.leading_bits) / ALIGN_IN_BITS;
+  p.aligned_bits = p.aligned_words * ALIGN_IN_BITS;
+  p.trailing_bits = length - p.leading_bits - p.aligned_bits;
+  p.trailing_bit_offset = bit_offset + p.leading_bits + p.aligned_bits;
+
+  p.aligned_start = data + (bit_offset + p.leading_bits) / 8;
+  return p;
+}
+}  // namespace internal
+
+namespace util {
+
+// Functions to check if the provided Arrow object is aligned by the specified alignment
+
+/// \brief Special alignment value to use data type-specific alignment
+///
+/// If this is passed as the `alignment` in one of the CheckAlignment or EnsureAlignment
+/// functions, then the function will ensure each buffer is suitably aligned
+/// for the data type of the array.  For example, given an int32 buffer the values
+/// buffer's address must be a multiple of 4.  Given a large_string buffer the offsets
+/// buffer's address must be a multiple of 8.
+constexpr int64_t kValueAlignment = -3;
+
+/// \brief Calculate if the buffer's address is a multiple of `alignment`
+///
+/// If `alignment` is less than or equal to 0 then this method will always return true
+/// \param buffer the buffer to check
+/// \param alignment the alignment (in bytes) to check for
+ARROW_EXPORT bool CheckAlignment(const Buffer& buffer, int64_t alignment);
+/// \brief Calculate if all buffers in the array data are aligned
+///
+/// This will also check the buffers in the dictionary and any children
+/// \param array the array data to check
+/// \param alignment the alignment (in bytes) to check for
+ARROW_EXPORT bool CheckAlignment(const ArrayData& array, int64_t alignment);
+/// \brief Calculate if all buffers in the array are aligned
+///
+/// This will also check the buffers in the dictionary and any children
+/// \param array the array to check
+/// \param alignment the alignment (in bytes) to check for
+ARROW_EXPORT bool CheckAlignment(const Array& array, int64_t alignment);
+
+// Following functions require an additional boolean vector which stores the
+// alignment check bits of the constituent objects.
+// For example, needs_alignment vector for a ChunkedArray will contain the
+// check bits of the constituent Arrays.
+// The boolean vector check was introduced to minimize the repetitive checks
+// of the constituent objects during the EnsureAlignment function where certain
+// objects can be ignored for further checking if we already know that they are
+// completely aligned.
+
+/// \brief Calculate which (if any) chunks in a chunked array are unaligned
+/// \param array the array to check
+/// \param alignment the alignment (in bytes) to check for
+/// \param needs_alignment an output vector that will store the results of the check
+///        it must be set to a valid vector.  Extra elements will be added to the end
+///        of the vector for each chunk that is checked.  `true` will be stored if
+///        the chunk is unaligned.
+/// \param offset the index of the chunk to start checking
+/// \return true if all chunks (starting at `offset`) are aligned, false otherwise
+ARROW_EXPORT bool CheckAlignment(const ChunkedArray& array, int64_t alignment,
+                                 std::vector<bool>* needs_alignment, int offset = 0);
+
+/// \brief calculate which (if any) columns in a record batch are unaligned
+/// \param batch the batch to check
+/// \param alignment the alignment (in bytes) to check for
+/// \param needs_alignment an output vector that will store the results of the
+///        check.  It must be set to a valid vector.  Extra elements will be added
+///        to the end of the vector for each column that is checked.  `true` will be
+///        stored if the column is unaligned.
+ARROW_EXPORT bool CheckAlignment(const RecordBatch& batch, int64_t alignment,
+                                 std::vector<bool>* needs_alignment);
+
+/// \brief calculate which (if any) columns in a table are unaligned
+/// \param table the table to check
+/// \param alignment the alignment (in bytes) to check for
+/// \param needs_alignment an output vector that will store the results of the
+///        check.  It must be set to a valid vector.  Extra elements will be added
+///        to the end of the vector for each column that is checked.  `true` will be
+///        stored if the column is unaligned.
+ARROW_EXPORT bool CheckAlignment(const Table& table, int64_t alignment,
+                                 std::vector<bool>* needs_alignment);
+
+/// \brief return a buffer that has the given alignment and the same data as the input
+/// buffer
+///
+/// If the input buffer is already aligned then this method will return the input buffer
+/// If the input buffer is not already aligned then this method will allocate a new
+/// buffer.  The alignment of the new buffer will have at least
+/// max(kDefaultBufferAlignment, alignment) bytes of alignment.
+///
+/// \param buffer the buffer to check
+/// \param alignment the alignment (in bytes) to check for
+/// \param memory_pool a memory pool that will be used to allocate a new buffer if the
+///        input buffer is not sufficiently aligned
+ARROW_EXPORT Result<std::shared_ptr<Buffer>> EnsureAlignment(
+    std::shared_ptr<Buffer> buffer, int64_t alignment, MemoryPool* memory_pool);
+
+/// \brief return an array data where all buffers are aligned by the given alignment
+///
+/// If any input buffer is already aligned then this method will reuse that same input
+/// buffer.
+///
+/// \param array_data the array data to check
+/// \param alignment the alignment (in bytes) to check for
+/// \param memory_pool a memory pool that will be used to allocate new buffers if any
+///        input buffer is not sufficiently aligned
+ARROW_EXPORT Result<std::shared_ptr<ArrayData>> EnsureAlignment(
+    std::shared_ptr<ArrayData> array_data, int64_t alignment, MemoryPool* memory_pool);
+
+/// \brief return an array where all buffers are aligned by the given alignment
+///
+/// If any input buffer is already aligned then this method will reuse that same input
+/// buffer.
+///
+/// \param array the array to check
+/// \param alignment the alignment (in bytes) to check for
+/// \param memory_pool a memory pool that will be used to allocate new buffers if any
+///        input buffer is not sufficiently aligned
+ARROW_EXPORT Result<std::shared_ptr<Array>> EnsureAlignment(std::shared_ptr<Array> array,
+                                                            int64_t alignment,
+                                                            MemoryPool* memory_pool);
+
+/// \brief return a chunked array where all buffers are aligned by the given alignment
+///
+/// If any input buffer is already aligned then this method will reuse that same input
+/// buffer.
+///
+/// \param array the chunked array to check
+/// \param alignment the alignment (in bytes) to check for
+/// \param memory_pool a memory pool that will be used to allocate new buffers if any
+///        input buffer is not sufficiently aligned
+ARROW_EXPORT Result<std::shared_ptr<ChunkedArray>> EnsureAlignment(
+    std::shared_ptr<ChunkedArray> array, int64_t alignment, MemoryPool* memory_pool);
+
+/// \brief return a record batch where all buffers are aligned by the given alignment
+///
+/// If any input buffer is already aligned then this method will reuse that same input
+/// buffer.
+///
+/// \param batch the batch to check
+/// \param alignment the alignment (in bytes) to check for
+/// \param memory_pool a memory pool that will be used to allocate new buffers if any
+///        input buffer is not sufficiently aligned
+ARROW_EXPORT Result<std::shared_ptr<RecordBatch>> EnsureAlignment(
+    std::shared_ptr<RecordBatch> batch, int64_t alignment, MemoryPool* memory_pool);
+
+/// \brief return a table where all buffers are aligned by the given alignment
+///
+/// If any input buffer is already aligned then this method will reuse that same input
+/// buffer.
+///
+/// \param table the table to check
+/// \param alignment the alignment (in bytes) to check for
+/// \param memory_pool a memory pool that will be used to allocate new buffers if any
+///        input buffer is not sufficiently aligned
+ARROW_EXPORT Result<std::shared_ptr<Table>> EnsureAlignment(std::shared_ptr<Table> table,
+                                                            int64_t alignment,
+                                                            MemoryPool* memory_pool);
+
+}  // namespace util
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/aligned_storage.h

@@ -0,0 +1,126 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cstring>
+#include <type_traits>
+#include <utility>
+
+#include "arrow/util/launder.h"
+#include "arrow/util/macros.h"
+
+namespace arrow {
+namespace internal {
+
+template <typename T>
+class AlignedStorage {
+ public:
+  static constexpr bool can_memcpy = std::is_trivial<T>::value;
+
+  constexpr T* get() noexcept {
+    return arrow::internal::launder(reinterpret_cast<T*>(&data_));
+  }
+
+  constexpr const T* get() const noexcept {
+    // Use fully qualified name to avoid ambiguities with MSVC (ARROW-14800)
+    return arrow::internal::launder(reinterpret_cast<const T*>(&data_));
+  }
+
+  void destroy() noexcept {
+    if (!std::is_trivially_destructible<T>::value) {
+      get()->~T();
+    }
+  }
+
+  template <typename... A>
+  void construct(A&&... args) noexcept {
+    new (&data_) T(std::forward<A>(args)...);
+  }
+
+  template <typename V>
+  void assign(V&& v) noexcept {
+    *get() = std::forward<V>(v);
+  }
+
+  void move_construct(AlignedStorage* other) noexcept {
+    new (&data_) T(std::move(*other->get()));
+  }
+
+  void move_assign(AlignedStorage* other) noexcept { *get() = std::move(*other->get()); }
+
+  template <bool CanMemcpy = can_memcpy>
+  static typename std::enable_if<CanMemcpy>::type move_construct_several(
+      AlignedStorage* ARROW_RESTRICT src, AlignedStorage* ARROW_RESTRICT dest, size_t n,
+      size_t memcpy_length) noexcept {
+    memcpy(dest->get(), src->get(), memcpy_length * sizeof(T));
+  }
+
+  template <bool CanMemcpy = can_memcpy>
+  static typename std::enable_if<CanMemcpy>::type
+  move_construct_several_and_destroy_source(AlignedStorage* ARROW_RESTRICT src,
+                                            AlignedStorage* ARROW_RESTRICT dest, size_t n,
+                                            size_t memcpy_length) noexcept {
+    memcpy(dest->get(), src->get(), memcpy_length * sizeof(T));
+  }
+
+  template <bool CanMemcpy = can_memcpy>
+  static typename std::enable_if<!CanMemcpy>::type move_construct_several(
+      AlignedStorage* ARROW_RESTRICT src, AlignedStorage* ARROW_RESTRICT dest, size_t n,
+      size_t memcpy_length) noexcept {
+    for (size_t i = 0; i < n; ++i) {
+      new (dest[i].get()) T(std::move(*src[i].get()));
+    }
+  }
+
+  template <bool CanMemcpy = can_memcpy>
+  static typename std::enable_if<!CanMemcpy>::type
+  move_construct_several_and_destroy_source(AlignedStorage* ARROW_RESTRICT src,
+                                            AlignedStorage* ARROW_RESTRICT dest, size_t n,
+                                            size_t memcpy_length) noexcept {
+    for (size_t i = 0; i < n; ++i) {
+      new (dest[i].get()) T(std::move(*src[i].get()));
+      src[i].destroy();
+    }
+  }
+
+  static void move_construct_several(AlignedStorage* ARROW_RESTRICT src,
+                                     AlignedStorage* ARROW_RESTRICT dest,
+                                     size_t n) noexcept {
+    move_construct_several(src, dest, n, n);
+  }
+
+  static void move_construct_several_and_destroy_source(
+      AlignedStorage* ARROW_RESTRICT src, AlignedStorage* ARROW_RESTRICT dest,
+      size_t n) noexcept {
+    move_construct_several_and_destroy_source(src, dest, n, n);
+  }
+
+  static void destroy_several(AlignedStorage* p, size_t n) noexcept {
+    if (!std::is_trivially_destructible<T>::value) {
+      for (size_t i = 0; i < n; ++i) {
+        p[i].destroy();
+      }
+    }
+  }
+
+ private:
+  alignas(T) std::byte data_[sizeof(T)];
+};
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/async_generator_fwd.h

@@ -0,0 +1,69 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <functional>
+
+#include "arrow/type_fwd.h"
+#include "arrow/util/type_fwd.h"
+
+namespace arrow {
+
+template <typename T>
+using AsyncGenerator = std::function<Future<T>()>;
+
+template <typename T, typename V>
+class MappingGenerator;
+
+template <typename T, typename ComesAfter, typename IsNext>
+class SequencingGenerator;
+
+template <typename T, typename V>
+class TransformingGenerator;
+
+template <typename T>
+class SerialReadaheadGenerator;
+
+template <typename T>
+class ReadaheadGenerator;
+
+template <typename T>
+class PushGenerator;
+
+template <typename T>
+class MergedGenerator;
+
+template <typename T>
+class EnumeratingGenerator;
+
+template <typename T>
+class TransferringGenerator;
+
+template <typename T>
+class BackgroundGenerator;
+
+template <typename T>
+class GeneratorIterator;
+
+template <typename T>
+struct CancellableGenerator;
+
+template <typename T>
+class DefaultIfEmptyGenerator;
+
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/async_util.h

@@ -0,0 +1,460 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <atomic>
+#include <functional>
+#include <list>
+#include <memory>
+
+#include "arrow/result.h"
+#include "arrow/status.h"
+#include "arrow/util/cancel.h"
+#include "arrow/util/functional.h"
+#include "arrow/util/future.h"
+#include "arrow/util/iterator.h"
+#include "arrow/util/mutex.h"
+#include "arrow/util/thread_pool.h"
+#include "arrow/util/tracing.h"
+
+namespace arrow {
+
+using internal::FnOnce;
+
+namespace util {
+
+/// A utility which keeps tracks of, and schedules, asynchronous tasks
+///
+/// An asynchronous task has a synchronous component and an asynchronous component.
+/// The synchronous component typically schedules some kind of work on an external
+/// resource (e.g. the I/O thread pool or some kind of kernel-based asynchronous
+/// resource like io_uring).  The asynchronous part represents the work
+/// done on that external resource.  Executing the synchronous part will be referred
+/// to as "submitting the task" since this usually includes submitting the asynchronous
+/// portion to the external thread pool.
+///
+/// By default the scheduler will submit the task (execute the synchronous part) as
+/// soon as it is added, assuming the underlying thread pool hasn't terminated or the
+/// scheduler hasn't aborted.  In this mode, the scheduler is simply acting as
+/// a simple task group.
+///
+/// A task scheduler starts with an initial task.  That task, and all subsequent tasks
+/// are free to add subtasks.  Once all submitted tasks finish the scheduler will
+/// finish.  Note, it is not an error to add additional tasks after a scheduler has
+/// aborted. These tasks will be ignored and never submitted.  The scheduler returns a
+/// future which will complete when all submitted tasks have finished executing.  Once all
+/// tasks have been finished the scheduler is invalid and should no longer be used.
+///
+/// Task failure (either the synchronous portion or the asynchronous portion) will cause
+/// the scheduler to enter an aborted state.  The first such failure will be reported in
+/// the final task future.
+class ARROW_EXPORT AsyncTaskScheduler {
+ public:
+  /// Destructor for AsyncTaskScheduler
+  ///
+  /// The lifetime of the task scheduled is managed automatically.  The scheduler
+  /// will remain valid while any tasks are running (and can always be safely accessed)
+  /// within tasks) and will be destroyed as soon as all tasks have finished.
+  virtual ~AsyncTaskScheduler() = default;
+  /// An interface for a task
+  ///
+  /// Users may want to override this, for example, to add priority
+  /// information for use by a queue.
+  class Task {
+   public:
+    virtual ~Task() = default;
+    /// Submit the task
+    ///
+    /// This will be called by the scheduler at most once when there
+    /// is space to run the task.  This is expected to be a fairly quick
+    /// function that simply submits the actual task work to an external
+    /// resource (e.g. I/O thread pool).
+    ///
+    /// If this call fails then the scheduler will enter an aborted state.
+    virtual Result<Future<>> operator()() = 0;
+    /// The cost of the task
+    ///
+    /// A ThrottledAsyncTaskScheduler can be used to limit the number of concurrent tasks.
+    /// A custom cost may be used, for example, if you would like to limit the number of
+    /// tasks based on the total expected RAM usage of the tasks (this is done in the
+    /// scanner)
+    virtual int cost() const { return 1; }
+    /// The name of the task
+    ///
+    /// This is used for debugging and traceability.  The returned view must remain
+    /// valid for the lifetime of the task.
+    virtual std::string_view name() const = 0;
+
+    /// a span tied to the lifetime of the task, for internal use only
+    tracing::Span span;
+  };
+
+  /// Add a task to the scheduler
+  ///
+  /// If the scheduler is in an aborted state this call will return false and the task
+  /// will never be run.  This is harmless and does not need to be guarded against.
+  ///
+  /// The return value for this call can usually be ignored.  There is little harm in
+  /// attempting to add tasks to an aborted scheduler.  It is only included for callers
+  /// that want to avoid future task generation to save effort.
+  ///
+  /// \param task the task to submit
+  ///
+  /// A task's name must remain valid for the duration of the task.  It is used for
+  /// debugging (e.g. when debugging a deadlock to see which tasks still remain) and for
+  /// traceability (the name will be used for spans assigned to the task)
+  ///
+  /// \return true if the task was submitted or queued, false if the task was ignored
+  virtual bool AddTask(std::unique_ptr<Task> task) = 0;
+
+  /// Adds an async generator to the scheduler
+  ///
+  /// The async generator will be visited, one item at a time.  Submitting a task
+  /// will consist of polling the generator for the next future.  The generator's future
+  /// will then represent the task itself.
+  ///
+  /// This visits the task serially without readahead.  If readahead or parallelism
+  /// is desired then it should be added in the generator itself.
+  ///
+  /// The generator itself will be kept alive until all tasks have been completed.
+  /// However, if the scheduler is aborted, the generator will be destroyed as soon as the
+  /// next item would be requested.
+  ///
+  /// \param generator the generator to submit to the scheduler
+  /// \param visitor a function which visits each generator future as it completes
+  /// \param name a name which will be used for each submitted task
+  template <typename T>
+  bool AddAsyncGenerator(std::function<Future<T>()> generator,
+                         std::function<Status(const T&)> visitor, std::string_view name);
+
+  template <typename Callable>
+  struct SimpleTask : public Task {
+    SimpleTask(Callable callable, std::string_view name)
+        : callable(std::move(callable)), name_(name) {}
+    SimpleTask(Callable callable, std::string name)
+        : callable(std::move(callable)), owned_name_(std::move(name)) {
+      name_ = *owned_name_;
+    }
+    Result<Future<>> operator()() override { return callable(); }
+    std::string_view name() const override { return name_; }
+    Callable callable;
+    std::string_view name_;
+    std::optional<std::string> owned_name_;
+  };
+
+  /// Add a task with cost 1 to the scheduler
+  ///
+  /// \param callable a "submit" function that should return a future
+  /// \param name a name for the task
+  ///
+  /// `name` must remain valid until the task has been submitted AND the returned
+  /// future completes.  It is used for debugging and tracing.
+  ///
+  /// \see AddTask for more details
+  template <typename Callable>
+  bool AddSimpleTask(Callable callable, std::string_view name) {
+    return AddTask(std::make_unique<SimpleTask<Callable>>(std::move(callable), name));
+  }
+
+  /// Add a task with cost 1 to the scheduler
+  ///
+  /// This is an overload of \see AddSimpleTask that keeps `name` alive
+  /// in the task.
+  template <typename Callable>
+  bool AddSimpleTask(Callable callable, std::string name) {
+    return AddTask(
+        std::make_unique<SimpleTask<Callable>>(std::move(callable), std::move(name)));
+  }
+
+  /// Construct a scheduler
+  ///
+  /// \param initial_task The initial task which is responsible for adding
+  ///        the first subtasks to the scheduler.
+  /// \param abort_callback A callback that will be triggered immediately after a task
+  ///        fails while other tasks may still be running.  Nothing needs to be done here,
+  ///        when a task fails the scheduler will stop accepting new tasks and eventually
+  ///        return the error.  However, this callback can be used to more quickly end
+  ///        long running tasks that have already been submitted.  Defaults to doing
+  ///        nothing.
+  /// \param stop_token An optional stop token that will allow cancellation of the
+  ///        scheduler.  This will be checked before each task is submitted and, in the
+  ///        event of a cancellation, the scheduler will enter an aborted state. This is
+  ///        a graceful cancellation and submitted tasks will still complete.
+  /// \return A future that will be completed when the initial task and all subtasks have
+  ///         finished.
+  static Future<> Make(
+      FnOnce<Status(AsyncTaskScheduler*)> initial_task,
+      FnOnce<void(const Status&)> abort_callback = [](const Status&) {},
+      StopToken stop_token = StopToken::Unstoppable());
+
+  /// A span tracking execution of the scheduler's tasks, for internal use only
+  virtual const tracing::Span& span() const = 0;
+};
+
+class ARROW_EXPORT ThrottledAsyncTaskScheduler : public AsyncTaskScheduler {
+ public:
+  /// An interface for a task queue
+  ///
+  /// A queue's methods will not be called concurrently
+  class Queue {
+   public:
+    virtual ~Queue() = default;
+    /// Push a task to the queue
+    ///
+    /// \param task the task to enqueue
+    virtual void Push(std::unique_ptr<Task> task) = 0;
+    /// Pop the next task from the queue
+    virtual std::unique_ptr<Task> Pop() = 0;
+    /// Peek the next task in the queue
+    virtual const Task& Peek() = 0;
+    /// Check if the queue is empty
+    virtual bool Empty() = 0;
+    /// Purge the queue of all items
+    virtual void Purge() = 0;
+    virtual std::size_t Size() const = 0;
+  };
+
+  class Throttle {
+   public:
+    virtual ~Throttle() = default;
+    /// Acquire amt permits
+    ///
+    /// If nullopt is returned then the permits were immediately
+    /// acquired and the caller can proceed.  If a future is returned then the caller
+    /// should wait for the future to complete first.  When the returned future completes
+    /// the permits have NOT been acquired and the caller must call Acquire again
+    ///
+    /// \param amt the number of permits to acquire
+    virtual std::optional<Future<>> TryAcquire(int amt) = 0;
+    /// Release amt permits
+    ///
+    /// This will possibly complete waiting futures and should probably not be
+    /// called while holding locks.
+    ///
+    /// \param amt the number of permits to release
+    virtual void Release(int amt) = 0;
+
+    /// The size of the largest task that can run
+    ///
+    /// Incoming tasks will have their cost latched to this value to ensure
+    /// they can still run (although they will be the only thing allowed to
+    /// run at that time).
+    virtual int Capacity() = 0;
+
+    /// Pause the throttle
+    ///
+    /// Any tasks that have been submitted already will continue.  However, no new tasks
+    /// will be run until the throttle is resumed.
+    virtual void Pause() = 0;
+    /// Resume the throttle
+    ///
+    /// Allows task to be submitted again.  If there is a max_concurrent_cost limit then
+    /// it will still apply.
+    virtual void Resume() = 0;
+  };
+
+  /// Pause the throttle
+  ///
+  /// Any tasks that have been submitted already will continue.  However, no new tasks
+  /// will be run until the throttle is resumed.
+  virtual void Pause() = 0;
+  /// Resume the throttle
+  ///
+  /// Allows task to be submitted again.  If there is a max_concurrent_cost limit then
+  /// it will still apply.
+  virtual void Resume() = 0;
+  /// Return the number of tasks queued but not yet submitted
+  virtual std::size_t QueueSize() = 0;
+
+  /// Create a throttled view of a scheduler
+  ///
+  /// Tasks added via this view will be subjected to the throttle and, if the tasks cannot
+  /// run immediately, will be placed into a queue.
+  ///
+  /// Although a shared_ptr is returned it should generally be assumed that the caller
+  /// is being given exclusive ownership.  The shared_ptr is used to share the view with
+  /// queued and submitted tasks and the lifetime of those is unpredictable.  It is
+  /// important the caller keep the returned pointer alive for as long as they plan to add
+  /// tasks to the view.
+  ///
+  /// \param scheduler a scheduler to submit tasks to after throttling
+  ///
+  /// This can be the root scheduler, another throttled scheduler, or a task group.  These
+  /// are all composable.
+  ///
+  /// \param max_concurrent_cost the maximum amount of cost allowed to run at any one time
+  ///
+  /// If a task is added that has a cost greater than max_concurrent_cost then its cost
+  /// will be reduced to max_concurrent_cost so that it is still possible for the task to
+  /// run.
+  ///
+  /// \param queue the queue to use when tasks cannot be submitted
+  ///
+  /// By default a FIFO queue will be used.  However, a custom queue can be provided if
+  /// some tasks have higher priority than other tasks.
+  static std::shared_ptr<ThrottledAsyncTaskScheduler> Make(
+      AsyncTaskScheduler* scheduler, int max_concurrent_cost,
+      std::unique_ptr<Queue> queue = NULLPTR);
+
+  /// @brief Create a ThrottledAsyncTaskScheduler using a custom throttle
+  ///
+  /// \see Make
+  static std::shared_ptr<ThrottledAsyncTaskScheduler> MakeWithCustomThrottle(
+      AsyncTaskScheduler* scheduler, std::unique_ptr<Throttle> throttle,
+      std::unique_ptr<Queue> queue = NULLPTR);
+};
+
+/// A utility to keep track of a collection of tasks
+///
+/// Often it is useful to keep track of some state that only needs to stay alive
+/// for some small collection of tasks, or to perform some kind of final cleanup
+/// when a collection of tasks is finished.
+///
+/// For example, when scanning, we need to keep the file reader alive while all scan
+/// tasks run for a given file, and then we can gracefully close it when we finish the
+/// file.
+class ARROW_EXPORT AsyncTaskGroup : public AsyncTaskScheduler {
+ public:
+  /// Destructor for the task group
+  ///
+  /// The destructor might trigger the finish callback.  If the finish callback fails
+  /// then the error will be reported as a task on the scheduler.
+  ///
+  /// Failure to destroy the async task group will not prevent the scheduler from
+  /// finishing.  If the scheduler finishes before the async task group is done then
+  /// the finish callback will be run immediately when the async task group finishes.
+  ///
+  /// If the scheduler has aborted then the finish callback will not run.
+  ~AsyncTaskGroup() = default;
+  /// Create an async task group
+  ///
+  /// The finish callback will not run until the task group is destroyed and all
+  /// tasks are finished so you will generally want to reset / destroy the returned
+  /// unique_ptr at some point.
+  ///
+  /// \param scheduler The underlying scheduler to submit tasks to
+  /// \param finish_callback A callback that will be run only after the task group has
+  ///                        been destroyed and all tasks added by the group have
+  ///                        finished.
+  ///
+  /// Note: in error scenarios the finish callback may not run.  However, it will still,
+  /// of course, be destroyed.
+  static std::unique_ptr<AsyncTaskGroup> Make(AsyncTaskScheduler* scheduler,
+                                              FnOnce<Status()> finish_callback);
+};
+
+/// Create a task group that is also throttled
+///
+/// This is a utility factory that creates a throttled view of a scheduler and then
+/// wraps that throttled view with a task group that destroys the throttle when finished.
+///
+/// \see ThrottledAsyncTaskScheduler
+/// \see AsyncTaskGroup
+/// \param target the underlying scheduler to submit tasks to
+/// \param max_concurrent_cost the maximum amount of cost allowed to run at any one time
+/// \param queue the queue to use when tasks cannot be submitted
+/// \param finish_callback A callback that will be run only after the task group has
+///                  been destroyed and all tasks added by the group have finished
+ARROW_EXPORT std::unique_ptr<ThrottledAsyncTaskScheduler> MakeThrottledAsyncTaskGroup(
+    AsyncTaskScheduler* target, int max_concurrent_cost,
+    std::unique_ptr<ThrottledAsyncTaskScheduler::Queue> queue,
+    FnOnce<Status()> finish_callback);
+
+// Defined down here to avoid circular dependency between AsyncTaskScheduler and
+// AsyncTaskGroup
+template <typename T>
+bool AsyncTaskScheduler::AddAsyncGenerator(std::function<Future<T>()> generator,
+                                           std::function<Status(const T&)> visitor,
+                                           std::string_view name) {
+  struct State {
+    State(std::function<Future<T>()> generator, std::function<Status(const T&)> visitor,
+          std::unique_ptr<AsyncTaskGroup> task_group, std::string_view name)
+        : generator(std::move(generator)),
+          visitor(std::move(visitor)),
+          task_group(std::move(task_group)),
+          name(name) {}
+    std::function<Future<T>()> generator;
+    std::function<Status(const T&)> visitor;
+    std::unique_ptr<AsyncTaskGroup> task_group;
+    std::string_view name;
+  };
+  struct SubmitTask : public Task {
+    explicit SubmitTask(std::unique_ptr<State> state_holder)
+        : state_holder(std::move(state_holder)) {}
+
+    struct SubmitTaskCallback {
+      SubmitTaskCallback(std::unique_ptr<State> state_holder, Future<> task_completion)
+          : state_holder(std::move(state_holder)),
+            task_completion(std::move(task_completion)) {}
+      void operator()(const Result<T>& maybe_item) {
+        if (!maybe_item.ok()) {
+          task_completion.MarkFinished(maybe_item.status());
+          return;
+        }
+        const auto& item = *maybe_item;
+        if (IsIterationEnd(item)) {
+          task_completion.MarkFinished();
+          return;
+        }
+        Status visit_st = state_holder->visitor(item);
+        if (!visit_st.ok()) {
+          task_completion.MarkFinished(std::move(visit_st));
+          return;
+        }
+        state_holder->task_group->AddTask(
+            std::make_unique<SubmitTask>(std::move(state_holder)));
+        task_completion.MarkFinished();
+      }
+      std::unique_ptr<State> state_holder;
+      Future<> task_completion;
+    };
+
+    Result<Future<>> operator()() {
+      Future<> task = Future<>::Make();
+      // Consume as many items as we can (those that are already finished)
+      // synchronously to avoid recursion / stack overflow.
+      while (true) {
+        Future<T> next = state_holder->generator();
+        if (next.TryAddCallback(
+                [&] { return SubmitTaskCallback(std::move(state_holder), task); })) {
+          return task;
+        }
+        ARROW_ASSIGN_OR_RAISE(T item, next.result());
+        if (IsIterationEnd(item)) {
+          task.MarkFinished();
+          return task;
+        }
+        ARROW_RETURN_NOT_OK(state_holder->visitor(item));
+      }
+    }
+
+    std::string_view name() const { return state_holder->name; }
+
+    std::unique_ptr<State> state_holder;
+  };
+  std::unique_ptr<AsyncTaskGroup> task_group =
+      AsyncTaskGroup::Make(this, [] { return Status::OK(); });
+  AsyncTaskGroup* task_group_view = task_group.get();
+  std::unique_ptr<State> state_holder = std::make_unique<State>(
+      std::move(generator), std::move(visitor), std::move(task_group), name);
+  task_group_view->AddTask(std::make_unique<SubmitTask>(std::move(state_holder)));
+  return true;
+}
+
+}  // namespace util
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/base64.h

@@ -0,0 +1,35 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <string>
+#include <string_view>
+
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+namespace util {
+
+ARROW_EXPORT
+std::string base64_encode(std::string_view s);
+
+ARROW_EXPORT
+std::string base64_decode(std::string_view s);
+
+}  // namespace util
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/basic_decimal.h

@@ -0,0 +1,887 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <array>
+#include <climits>
+#include <cstdint>
+#include <cstring>
+#include <limits>
+#include <string>
+#include <type_traits>
+
+#include "arrow/util/endian.h"
+#include "arrow/util/macros.h"
+#include "arrow/util/type_traits.h"
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+
+enum class DecimalStatus {
+  kSuccess,
+  kDivideByZero,
+  kOverflow,
+  kRescaleDataLoss,
+};
+
+template <typename Derived, int BIT_WIDTH, int NWORDS = BIT_WIDTH / 64>
+class GenericBasicDecimal {
+ protected:
+  struct LittleEndianArrayTag {};
+
+#if ARROW_LITTLE_ENDIAN
+  static constexpr int kHighWordIndex = NWORDS - 1;
+  static constexpr int kLowWordIndex = 0;
+#else
+  static constexpr int kHighWordIndex = 0;
+  static constexpr int kLowWordIndex = NWORDS - 1;
+#endif
+
+ public:
+  static constexpr int kBitWidth = BIT_WIDTH;
+  static constexpr int kByteWidth = kBitWidth / 8;
+  static constexpr int kNumWords = NWORDS;
+
+  // A constructor tag to introduce a little-endian encoded array
+  static constexpr LittleEndianArrayTag LittleEndianArray{};
+
+  using WordArray = std::array<uint64_t, NWORDS>;
+
+  /// \brief Empty constructor creates a decimal with a value of 0.
+  constexpr GenericBasicDecimal() noexcept : array_({0}) {}
+
+  /// \brief Create a decimal from the two's complement representation.
+  ///
+  /// Input array is assumed to be in native endianness.
+  explicit constexpr GenericBasicDecimal(const WordArray& array) noexcept
+      : array_(array) {}
+
+  /// \brief Create a decimal from the two's complement representation.
+  ///
+  /// Input array is assumed to be in little endianness, with native endian elements.
+  GenericBasicDecimal(LittleEndianArrayTag, const WordArray& array) noexcept
+      : GenericBasicDecimal(bit_util::little_endian::ToNative(array)) {}
+
+  /// \brief Create a decimal from any integer not wider than 64 bits.
+  template <typename T,
+            typename = typename std::enable_if<
+                std::is_integral<T>::value && (sizeof(T) <= sizeof(uint64_t)), T>::type>
+  constexpr GenericBasicDecimal(T value) noexcept  // NOLINT(runtime/explicit)
+      : array_(WordsFromLowBits(value)) {}
+
+  /// \brief Create a decimal from an array of bytes.
+  ///
+  /// Bytes are assumed to be in native-endian byte order.
+  explicit GenericBasicDecimal(const uint8_t* bytes) {
+    memcpy(array_.data(), bytes, sizeof(array_));
+  }
+
+  /// \brief Get the bits of the two's complement representation of the number.
+  ///
+  /// The elements are in native endian order. The bits within each uint64_t element
+  /// are in native endian order. For example, on a little endian machine,
+  /// BasicDecimal128(123).native_endian_array() = {123, 0};
+  /// but on a big endian machine,
+  /// BasicDecimal128(123).native_endian_array() = {0, 123};
+  constexpr const WordArray& native_endian_array() const { return array_; }
+
+  /// \brief Get the bits of the two's complement representation of the number.
+  ///
+  /// The elements are in little endian order. However, the bits within each
+  /// uint64_t element are in native endian order.
+  /// For example, BasicDecimal128(123).little_endian_array() = {123, 0};
+  WordArray little_endian_array() const {
+    return bit_util::little_endian::FromNative(array_);
+  }
+
+  const uint8_t* native_endian_bytes() const {
+    return reinterpret_cast<const uint8_t*>(array_.data());
+  }
+
+  uint8_t* mutable_native_endian_bytes() {
+    return reinterpret_cast<uint8_t*>(array_.data());
+  }
+
+  /// \brief Return the raw bytes of the value in native-endian byte order.
+  std::array<uint8_t, kByteWidth> ToBytes() const {
+    std::array<uint8_t, kByteWidth> out{{0}};
+    memcpy(out.data(), array_.data(), kByteWidth);
+    return out;
+  }
+
+  /// \brief Copy the raw bytes of the value in native-endian byte order.
+  void ToBytes(uint8_t* out) const { memcpy(out, array_.data(), kByteWidth); }
+
+  /// Return 1 if positive or zero, -1 if strictly negative.
+  int64_t Sign() const {
+    return 1 | (static_cast<int64_t>(array_[kHighWordIndex]) >> 63);
+  }
+
+  bool IsNegative() const { return static_cast<int64_t>(array_[kHighWordIndex]) < 0; }
+
+  explicit operator bool() const { return array_ != WordArray{}; }
+
+  friend bool operator==(const GenericBasicDecimal& left,
+                         const GenericBasicDecimal& right) {
+    return left.array_ == right.array_;
+  }
+
+  friend bool operator!=(const GenericBasicDecimal& left,
+                         const GenericBasicDecimal& right) {
+    return left.array_ != right.array_;
+  }
+
+ protected:
+  WordArray array_;
+
+  template <typename T>
+  static constexpr uint64_t SignExtend(T low_bits) noexcept {
+    return low_bits >= T{} ? uint64_t{0} : ~uint64_t{0};
+  }
+
+  template <typename T>
+  static constexpr WordArray WordsFromLowBits(T low_bits) {
+    WordArray words{};
+    if (low_bits < T{}) {
+      for (auto& word : words) {
+        word = ~uint64_t{0};
+      }
+    }
+    words[kLowWordIndex] = static_cast<uint64_t>(low_bits);
+    return words;
+  }
+};
+
+template <typename DigitType>
+class ARROW_EXPORT SmallBasicDecimal {
+ public:
+  static_assert(
+      std::is_same_v<DigitType, int32_t> || std::is_same_v<DigitType, int64_t>,
+      "for bitwidths larger than 64 bits use BasicDecimal128 and BasicDecimal256");
+
+  static constexpr int kMaxPrecision = std::numeric_limits<DigitType>::digits10;
+  static constexpr int kMaxScale = kMaxPrecision;
+  static constexpr int kBitWidth = sizeof(DigitType) * CHAR_BIT;
+  static constexpr int kByteWidth = sizeof(DigitType);
+
+  using WordArray = std::array<std::make_unsigned_t<DigitType>, 1>;
+
+  /// \brief Empty constructor creates a decimal with a value of 0.
+  constexpr SmallBasicDecimal() noexcept : value_(0) {}
+
+  /// \brief Create a decimal from any integer not wider than 64 bits.
+  template <typename T,
+            typename = typename std::enable_if<
+                std::is_integral<T>::value && (sizeof(T) <= sizeof(int64_t)), T>::type>
+  constexpr SmallBasicDecimal(T value) noexcept  // NOLINT(runtime/explicit)
+      : value_(static_cast<DigitType>(value)) {}
+
+  /// \brief Create a decimal from an array of bytes.
+  ///
+  /// Bytes are assumed to be in native-endian byte order.
+  explicit SmallBasicDecimal(const uint8_t* bytes) {
+    memcpy(&value_, bytes, sizeof(value_));
+  }
+
+  constexpr const WordArray native_endian_array() const {
+    return WordArray{static_cast<typename WordArray::value_type>(value_)};
+  }
+
+  constexpr const WordArray little_endian_array() const {
+    return bit_util::little_endian::FromNative(
+        WordArray{static_cast<typename WordArray::value_type>(value_)});
+  }
+
+  const uint8_t* native_endian_bytes() const {
+    return reinterpret_cast<const uint8_t*>(&value_);
+  }
+
+  uint8_t* mutable_native_endian_bytes() { return reinterpret_cast<uint8_t*>(&value_); }
+
+  /// \brief Return the raw bytes of the value in native-endian byte order.
+  std::array<uint8_t, kByteWidth> ToBytes() const {
+    std::array<uint8_t, kByteWidth> out{{0}};
+    memcpy(out.data(), &value_, kByteWidth);
+    return out;
+  }
+
+  /// \brief Copy the raw bytes of the value in native-endian byte order
+  void ToBytes(uint8_t* out) const { memcpy(out, &value_, kByteWidth); }
+
+  /// \brief Return 1 if positive or 0, -1 if strictly negative
+  int64_t Sign() const { return 1 | (value_ >> (kBitWidth - 1)); }
+
+  bool IsNegative() const { return value_ < 0; }
+
+  explicit operator bool() const { return value_ != 0; }
+
+  friend bool operator==(const SmallBasicDecimal& left, const SmallBasicDecimal& right) {
+    return left.value_ == right.value_;
+  }
+
+  friend bool operator!=(const SmallBasicDecimal& left, const SmallBasicDecimal& right) {
+    return left.value_ != right.value_;
+  }
+
+  DigitType value() const { return value_; }
+
+  /// \brief count the number of leading binary zeroes.
+  int32_t CountLeadingBinaryZeros() const;
+
+  constexpr uint64_t low_bits() const { return static_cast<uint64_t>(value_); }
+
+ protected:
+  DigitType value_;
+};
+
+class BasicDecimal32;
+class BasicDecimal64;
+
+ARROW_EXPORT bool operator<(const BasicDecimal32& left, const BasicDecimal32& right);
+ARROW_EXPORT bool operator<=(const BasicDecimal32& left, const BasicDecimal32& right);
+ARROW_EXPORT bool operator>(const BasicDecimal32& left, const BasicDecimal32& right);
+ARROW_EXPORT bool operator>=(const BasicDecimal32& left, const BasicDecimal32& right);
+
+ARROW_EXPORT BasicDecimal32 operator-(const BasicDecimal32& self);
+ARROW_EXPORT BasicDecimal32 operator~(const BasicDecimal32& self);
+ARROW_EXPORT BasicDecimal32 operator+(const BasicDecimal32& left,
+                                      const BasicDecimal32& right);
+ARROW_EXPORT BasicDecimal32 operator-(const BasicDecimal32& left,
+                                      const BasicDecimal32& right);
+ARROW_EXPORT BasicDecimal32 operator*(const BasicDecimal32& left,
+                                      const BasicDecimal32& right);
+ARROW_EXPORT BasicDecimal32 operator/(const BasicDecimal32& left,
+                                      const BasicDecimal32& right);
+ARROW_EXPORT BasicDecimal32 operator%(const BasicDecimal32& left,
+                                      const BasicDecimal32& right);
+
+class ARROW_EXPORT BasicDecimal32 : public SmallBasicDecimal<int32_t> {
+ public:
+  using SmallBasicDecimal<int32_t>::SmallBasicDecimal;
+  using ValueType = int32_t;
+
+  /// \brief Negate the current value (in-place)
+  BasicDecimal32& Negate();
+
+  /// \brief Absolute value (in-place)
+  BasicDecimal32& Abs() { return *this < 0 ? Negate() : *this; }
+
+  /// \brief Absolute value
+  static BasicDecimal32 Abs(const BasicDecimal32& in) {
+    BasicDecimal32 result(in);
+    return result.Abs();
+  }
+
+  /// \brief Add a number to this one. The result is truncated to 32 bits.
+  BasicDecimal32& operator+=(const BasicDecimal32& right) {
+    value_ += right.value_;
+    return *this;
+  }
+
+  /// \brief Subtract a number from this one. The result is truncated to 32 bits.
+  BasicDecimal32& operator-=(const BasicDecimal32& right) {
+    value_ -= right.value_;
+    return *this;
+  }
+
+  /// \brief Multiply this number by another. The result is truncated to 32 bits.
+  BasicDecimal32& operator*=(const BasicDecimal32& right) {
+    value_ *= static_cast<uint64_t>(right.value_);
+    return *this;
+  }
+
+  /// \brief Divide this number by the divisor and return the result.
+  ///
+  /// This operation is not destructive.
+  /// The answer rounds to zero. Signs work like:
+  ///   21 /  5 ->  4,  1
+  ///  -21 /  5 -> -4, -1
+  ///   21 / -5 -> -4,  1
+  ///  -21 / -5 ->  4, -1
+  /// \param[in] divisor the number to divide by
+  /// \param[out] result the quotient
+  /// \param[out] remainder the remainder after the division
+  DecimalStatus Divide(const BasicDecimal32& divisor, BasicDecimal32* result,
+                       BasicDecimal32* remainder) const;
+
+  /// \brief In-place division
+  BasicDecimal32& operator/=(const BasicDecimal32& right) {
+    value_ /= right.value_;
+    return *this;
+  }
+
+  /// \brief Bitwise "or" between two BasicDecimal32s
+  BasicDecimal32& operator|=(const BasicDecimal32& right) {
+    value_ |= right.value_;
+    return *this;
+  }
+
+  /// \brief Bitwise "and" between two BasicDecimal32s
+  BasicDecimal32& operator&=(const BasicDecimal32& right) {
+    value_ &= right.value_;
+    return *this;
+  }
+  /// \brief Shift left by the given number of bits.
+  BasicDecimal32& operator<<=(uint32_t bits);
+
+  BasicDecimal32 operator<<(uint32_t bits) const {
+    auto res = *this;
+    res <<= bits;
+    return res;
+  }
+
+  /// \brief Shift right by the given number of bits.
+  ///
+  /// Negative values will sign-extend
+  BasicDecimal32& operator>>=(uint32_t bits);
+
+  BasicDecimal32 operator>>(uint32_t bits) const {
+    auto res = *this;
+    res >>= bits;
+    return res;
+  }
+
+  /// \brief Convert BasicDecimal32 from one scale to another
+  DecimalStatus Rescale(int32_t original_scale, int32_t new_scale,
+                        BasicDecimal32* out) const;
+
+  void GetWholeAndFraction(int scale, BasicDecimal32* whole,
+                           BasicDecimal32* fraction) const;
+
+  /// \brief Scale up.
+  BasicDecimal32 IncreaseScaleBy(int32_t increase_by) const;
+
+  /// \brief Scale down.
+  ///
+  /// - If 'round' is true, the right-most digits are dropped and the result value is
+  ///   rounded up (+1 for +ve, -1 for -ve) based on the value of the dropped digits
+  ///   (>= 10^reduce_by / 2).
+  /// - If 'round' is false, the right-most digits are simply dropped.
+  BasicDecimal32 ReduceScaleBy(int32_t reduce_by, bool round = true) const;
+
+  /// \brief Whether this number fits in the given precision
+  ///
+  /// Return true if the number of significant digits is less or equal to 'precision'.
+  bool FitsInPrecision(int32_t precision) const;
+
+  /// \brief Get the maximum valid unscaled decimal value.
+  static const BasicDecimal32& GetMaxValue();
+  /// \brief Get the maximum valid unscaled decimal value for the given precision.
+  static BasicDecimal32 GetMaxValue(int32_t precision);
+
+  /// \brief Get the maximum decimal value (is not a valid value).
+  static constexpr BasicDecimal32 GetMaxSentinel() {
+    return BasicDecimal32(std::numeric_limits<int32_t>::max());
+  }
+
+  /// \brief Get the minimum decimal value (is not a valid value).
+  static constexpr BasicDecimal32 GetMinSentinel() {
+    return BasicDecimal32(std::numeric_limits<int32_t>::min());
+  }
+
+  /// \brief Scale multiplier for a given scale value.
+  static const BasicDecimal32& GetScaleMultiplier(int32_t scale);
+  /// \brief Half-scale multiplier for a given scale value.
+  static const BasicDecimal32& GetHalfScaleMultiplier(int32_t scale);
+
+  explicit operator BasicDecimal64() const;
+};
+
+ARROW_EXPORT bool operator<(const BasicDecimal64& left, const BasicDecimal64& right);
+ARROW_EXPORT bool operator<=(const BasicDecimal64& left, const BasicDecimal64& right);
+ARROW_EXPORT bool operator>(const BasicDecimal64& left, const BasicDecimal64& right);
+ARROW_EXPORT bool operator>=(const BasicDecimal64& left, const BasicDecimal64& right);
+
+ARROW_EXPORT BasicDecimal64 operator-(const BasicDecimal64& self);
+ARROW_EXPORT BasicDecimal64 operator~(const BasicDecimal64& self);
+ARROW_EXPORT BasicDecimal64 operator+(const BasicDecimal64& left,
+                                      const BasicDecimal64& right);
+ARROW_EXPORT BasicDecimal64 operator-(const BasicDecimal64& left,
+                                      const BasicDecimal64& right);
+ARROW_EXPORT BasicDecimal64 operator*(const BasicDecimal64& left,
+                                      const BasicDecimal64& right);
+ARROW_EXPORT BasicDecimal64 operator/(const BasicDecimal64& left,
+                                      const BasicDecimal64& right);
+ARROW_EXPORT BasicDecimal64 operator%(const BasicDecimal64& left,
+                                      const BasicDecimal64& right);
+
+class ARROW_EXPORT BasicDecimal64 : public SmallBasicDecimal<int64_t> {
+ public:
+  using SmallBasicDecimal<int64_t>::SmallBasicDecimal;
+  using ValueType = int64_t;
+
+  /// \brief Negate the current value (in-place)
+  BasicDecimal64& Negate();
+
+  /// \brief Absolute value (in-place)
+  BasicDecimal64& Abs() { return *this < 0 ? Negate() : *this; }
+
+  /// \brief Absolute value
+  static BasicDecimal64 Abs(const BasicDecimal64& in) {
+    BasicDecimal64 result(in);
+    return result.Abs();
+  }
+
+  /// \brief Add a number to this one. The result is truncated to 32 bits.
+  BasicDecimal64& operator+=(const BasicDecimal64& right) {
+    value_ += right.value_;
+    return *this;
+  }
+
+  /// \brief Subtract a number from this one. The result is truncated to 32 bits.
+  BasicDecimal64& operator-=(const BasicDecimal64& right) {
+    value_ -= right.value_;
+    return *this;
+  }
+
+  /// \brief Multiply this number by another. The result is truncated to 32 bits.
+  BasicDecimal64& operator*=(const BasicDecimal64& right) {
+    value_ *= static_cast<uint64_t>(right.value_);
+    return *this;
+  }
+
+  /// \brief Divide this number by the divisor and return the result.
+  ///
+  /// This operation is not destructive.
+  /// The answer rounds to zero. Signs work like:
+  ///   21 /  5 ->  4,  1
+  ///  -21 /  5 -> -4, -1
+  ///   21 / -5 -> -4,  1
+  ///  -21 / -5 ->  4, -1
+  /// \param[in] divisor the number to divide by
+  /// \param[out] result the quotient
+  /// \param[out] remainder the remainder after the division
+  DecimalStatus Divide(const BasicDecimal64& divisor, BasicDecimal64* result,
+                       BasicDecimal64* remainder) const;
+
+  /// \brief In-place division
+  BasicDecimal64& operator/=(const BasicDecimal64& right) {
+    value_ /= right.value_;
+    return *this;
+  }
+
+  /// \brief Bitwise "or" between two BasicDecimal64s
+  BasicDecimal64& operator|=(const BasicDecimal64& right) {
+    value_ |= right.value_;
+    return *this;
+  }
+
+  /// \brief Bitwise "and" between two BasicDecimal64s
+  BasicDecimal64& operator&=(const BasicDecimal64& right) {
+    value_ &= right.value_;
+    return *this;
+  }
+
+  /// \brief Shift left by the given number of bits.
+  BasicDecimal64& operator<<=(uint32_t bits);
+
+  BasicDecimal64 operator<<(uint32_t bits) const {
+    auto res = *this;
+    res <<= bits;
+    return res;
+  }
+
+  /// \brief Shift right by the given number of bits.
+  ///
+  /// Negative values will sign-extend
+  BasicDecimal64& operator>>=(uint32_t bits);
+
+  BasicDecimal64 operator>>(uint32_t bits) const {
+    auto res = *this;
+    res >>= bits;
+    return res;
+  }
+
+  /// \brief Convert BasicDecimal32 from one scale to another
+  DecimalStatus Rescale(int32_t original_scale, int32_t new_scale,
+                        BasicDecimal64* out) const;
+
+  void GetWholeAndFraction(int scale, BasicDecimal64* whole,
+                           BasicDecimal64* fraction) const;
+
+  /// \brief Scale up.
+  BasicDecimal64 IncreaseScaleBy(int32_t increase_by) const;
+
+  /// \brief Scale down.
+  ///
+  /// - If 'round' is true, the right-most digits are dropped and the result value is
+  ///   rounded up (+1 for +ve, -1 for -ve) based on the value of the dropped digits
+  ///   (>= 10^reduce_by / 2).
+  /// - If 'round' is false, the right-most digits are simply dropped.
+  BasicDecimal64 ReduceScaleBy(int32_t reduce_by, bool round = true) const;
+
+  /// \brief Whether this number fits in the given precision
+  ///
+  /// Return true if the number of significant digits is less or equal to 'precision'.
+  bool FitsInPrecision(int32_t precision) const;
+
+  /// \brief Get the maximum valid unscaled decimal value.
+  static const BasicDecimal64& GetMaxValue();
+  /// \brief Get the maximum valid unscaled decimal value for the given precision.
+  static BasicDecimal64 GetMaxValue(int32_t precision);
+
+  /// \brief Get the maximum decimal value (is not a valid value).
+  static constexpr BasicDecimal64 GetMaxSentinel() {
+    return BasicDecimal64(std::numeric_limits<int32_t>::max());
+  }
+
+  /// \brief Get the minimum decimal value (is not a valid value).
+  static constexpr BasicDecimal64 GetMinSentinel() {
+    return BasicDecimal64(std::numeric_limits<int32_t>::min());
+  }
+
+  /// \brief Scale multiplier for a given scale value.
+  static const BasicDecimal64& GetScaleMultiplier(int32_t scale);
+  /// \brief Half-scale multiplier for a given scale value.
+  static const BasicDecimal64& GetHalfScaleMultiplier(int32_t scale);
+};
+
+/// Represents a signed 128-bit integer in two's complement.
+///
+/// This class is also compiled into LLVM IR - so, it should not have cpp references like
+/// streams and boost.
+class ARROW_EXPORT BasicDecimal128 : public GenericBasicDecimal<BasicDecimal128, 128> {
+ public:
+  static constexpr int kMaxPrecision = 38;
+  static constexpr int kMaxScale = 38;
+
+  using GenericBasicDecimal::GenericBasicDecimal;
+
+  constexpr BasicDecimal128() noexcept : GenericBasicDecimal() {}
+
+  /// \brief Create a BasicDecimal128 from the two's complement representation.
+#if ARROW_LITTLE_ENDIAN
+  constexpr BasicDecimal128(int64_t high, uint64_t low) noexcept
+      : BasicDecimal128(WordArray{low, static_cast<uint64_t>(high)}) {}
+#else
+  constexpr BasicDecimal128(int64_t high, uint64_t low) noexcept
+      : BasicDecimal128(WordArray{static_cast<uint64_t>(high), low}) {}
+#endif
+
+  /// \brief Negate the current value (in-place)
+  BasicDecimal128& Negate();
+
+  /// \brief Absolute value (in-place)
+  BasicDecimal128& Abs();
+
+  /// \brief Absolute value
+  static BasicDecimal128 Abs(const BasicDecimal128& left);
+
+  /// \brief Add a number to this one. The result is truncated to 128 bits.
+  BasicDecimal128& operator+=(const BasicDecimal128& right);
+
+  /// \brief Subtract a number from this one. The result is truncated to 128 bits.
+  BasicDecimal128& operator-=(const BasicDecimal128& right);
+
+  /// \brief Multiply this number by another number. The result is truncated to 128 bits.
+  BasicDecimal128& operator*=(const BasicDecimal128& right);
+
+  /// Divide this number by right and return the result.
+  ///
+  /// This operation is not destructive.
+  /// The answer rounds to zero. Signs work like:
+  ///   21 /  5 ->  4,  1
+  ///  -21 /  5 -> -4, -1
+  ///   21 / -5 -> -4,  1
+  ///  -21 / -5 ->  4, -1
+  /// \param[in] divisor the number to divide by
+  /// \param[out] result the quotient
+  /// \param[out] remainder the remainder after the division
+  DecimalStatus Divide(const BasicDecimal128& divisor, BasicDecimal128* result,
+                       BasicDecimal128* remainder) const;
+
+  /// \brief In-place division.
+  BasicDecimal128& operator/=(const BasicDecimal128& right);
+
+  /// \brief Bitwise "or" between two BasicDecimal128.
+  BasicDecimal128& operator|=(const BasicDecimal128& right);
+
+  /// \brief Bitwise "and" between two BasicDecimal128.
+  BasicDecimal128& operator&=(const BasicDecimal128& right);
+
+  /// \brief Shift left by the given number of bits.
+  BasicDecimal128& operator<<=(uint32_t bits);
+
+  BasicDecimal128 operator<<(uint32_t bits) const {
+    auto res = *this;
+    res <<= bits;
+    return res;
+  }
+
+  /// \brief Shift right by the given number of bits.
+  ///
+  /// Negative values will sign-extend.
+  BasicDecimal128& operator>>=(uint32_t bits);
+
+  BasicDecimal128 operator>>(uint32_t bits) const {
+    auto res = *this;
+    res >>= bits;
+    return res;
+  }
+
+  /// \brief Get the high bits of the two's complement representation of the number.
+  constexpr int64_t high_bits() const {
+#if ARROW_LITTLE_ENDIAN
+    return static_cast<int64_t>(array_[1]);
+#else
+    return static_cast<int64_t>(array_[0]);
+#endif
+  }
+
+  /// \brief Get the low bits of the two's complement representation of the number.
+  constexpr uint64_t low_bits() const {
+#if ARROW_LITTLE_ENDIAN
+    return array_[0];
+#else
+    return array_[1];
+#endif
+  }
+
+  /// \brief separate the integer and fractional parts for the given scale.
+  void GetWholeAndFraction(int32_t scale, BasicDecimal128* whole,
+                           BasicDecimal128* fraction) const;
+
+  /// \brief Scale multiplier for given scale value.
+  static const BasicDecimal128& GetScaleMultiplier(int32_t scale);
+  /// \brief Half-scale multiplier for given scale value.
+  static const BasicDecimal128& GetHalfScaleMultiplier(int32_t scale);
+
+  /// \brief Convert BasicDecimal128 from one scale to another
+  DecimalStatus Rescale(int32_t original_scale, int32_t new_scale,
+                        BasicDecimal128* out) const;
+
+  /// \brief Scale up.
+  BasicDecimal128 IncreaseScaleBy(int32_t increase_by) const;
+
+  /// \brief Scale down.
+  /// - If 'round' is true, the right-most digits are dropped and the result value is
+  ///   rounded up (+1 for +ve, -1 for -ve) based on the value of the dropped digits
+  ///   (>= 10^reduce_by / 2).
+  /// - If 'round' is false, the right-most digits are simply dropped.
+  BasicDecimal128 ReduceScaleBy(int32_t reduce_by, bool round = true) const;
+
+  /// \brief Whether this number fits in the given precision
+  ///
+  /// Return true if the number of significant digits is less or equal to `precision`.
+  bool FitsInPrecision(int32_t precision) const;
+
+  /// \brief count the number of leading binary zeroes.
+  int32_t CountLeadingBinaryZeros() const;
+
+  /// \brief Get the maximum valid unscaled decimal value.
+  static const BasicDecimal128& GetMaxValue();
+
+  /// \brief Get the maximum valid unscaled decimal value for the given precision.
+  static BasicDecimal128 GetMaxValue(int32_t precision);
+
+  /// \brief Get the maximum decimal value (is not a valid value).
+  static constexpr BasicDecimal128 GetMaxSentinel() {
+    return BasicDecimal128(/*high=*/std::numeric_limits<int64_t>::max(),
+                           /*low=*/std::numeric_limits<uint64_t>::max());
+  }
+  /// \brief Get the minimum decimal value (is not a valid value).
+  static constexpr BasicDecimal128 GetMinSentinel() {
+    return BasicDecimal128(/*high=*/std::numeric_limits<int64_t>::min(),
+                           /*low=*/std::numeric_limits<uint64_t>::min());
+  }
+};
+
+ARROW_EXPORT bool operator<(const BasicDecimal128& left, const BasicDecimal128& right);
+ARROW_EXPORT bool operator<=(const BasicDecimal128& left, const BasicDecimal128& right);
+ARROW_EXPORT bool operator>(const BasicDecimal128& left, const BasicDecimal128& right);
+ARROW_EXPORT bool operator>=(const BasicDecimal128& left, const BasicDecimal128& right);
+
+ARROW_EXPORT BasicDecimal128 operator-(const BasicDecimal128& operand);
+ARROW_EXPORT BasicDecimal128 operator~(const BasicDecimal128& operand);
+ARROW_EXPORT BasicDecimal128 operator+(const BasicDecimal128& left,
+                                       const BasicDecimal128& right);
+ARROW_EXPORT BasicDecimal128 operator-(const BasicDecimal128& left,
+                                       const BasicDecimal128& right);
+ARROW_EXPORT BasicDecimal128 operator*(const BasicDecimal128& left,
+                                       const BasicDecimal128& right);
+ARROW_EXPORT BasicDecimal128 operator/(const BasicDecimal128& left,
+                                       const BasicDecimal128& right);
+ARROW_EXPORT BasicDecimal128 operator%(const BasicDecimal128& left,
+                                       const BasicDecimal128& right);
+
+class ARROW_EXPORT BasicDecimal256 : public GenericBasicDecimal<BasicDecimal256, 256> {
+ public:
+  using GenericBasicDecimal::GenericBasicDecimal;
+
+  static constexpr int kMaxPrecision = 76;
+  static constexpr int kMaxScale = 76;
+
+  constexpr BasicDecimal256() noexcept : GenericBasicDecimal() {}
+
+  explicit BasicDecimal256(const BasicDecimal128& value) noexcept
+      : BasicDecimal256(bit_util::little_endian::ToNative<uint64_t, 4>(
+            {value.low_bits(), static_cast<uint64_t>(value.high_bits()),
+             SignExtend(value.high_bits()), SignExtend(value.high_bits())})) {}
+
+  explicit BasicDecimal256(const BasicDecimal64& value) noexcept
+      : BasicDecimal256(bit_util::little_endian::ToNative<uint64_t, 4>(
+            {value.low_bits(), SignExtend(value.value()), SignExtend(value.value()),
+             SignExtend(value.value())})) {}
+
+  explicit BasicDecimal256(const BasicDecimal32& value) noexcept
+      : BasicDecimal256(bit_util::little_endian::ToNative<uint64_t, 4>(
+            {value.low_bits(), SignExtend(value.value()), SignExtend(value.value()),
+             SignExtend(value.value())})) {}
+
+  /// \brief Negate the current value (in-place)
+  BasicDecimal256& Negate();
+
+  /// \brief Absolute value (in-place)
+  BasicDecimal256& Abs();
+
+  /// \brief Absolute value
+  static BasicDecimal256 Abs(const BasicDecimal256& left);
+
+  /// \brief Add a number to this one. The result is truncated to 256 bits.
+  BasicDecimal256& operator+=(const BasicDecimal256& right);
+
+  /// \brief Subtract a number from this one. The result is truncated to 256 bits.
+  BasicDecimal256& operator-=(const BasicDecimal256& right);
+
+  /// \brief Get the lowest bits of the two's complement representation of the number.
+  uint64_t low_bits() const { return bit_util::little_endian::Make(array_)[0]; }
+
+  /// \brief separate the integer and fractional parts for the given scale.
+  void GetWholeAndFraction(int32_t scale, BasicDecimal256* whole,
+                           BasicDecimal256* fraction) const;
+
+  /// \brief Scale multiplier for given scale value.
+  static const BasicDecimal256& GetScaleMultiplier(int32_t scale);
+  /// \brief Half-scale multiplier for given scale value.
+  static const BasicDecimal256& GetHalfScaleMultiplier(int32_t scale);
+
+  /// \brief Convert BasicDecimal256 from one scale to another
+  DecimalStatus Rescale(int32_t original_scale, int32_t new_scale,
+                        BasicDecimal256* out) const;
+
+  /// \brief Scale up.
+  BasicDecimal256 IncreaseScaleBy(int32_t increase_by) const;
+
+  /// \brief Scale down.
+  /// - If 'round' is true, the right-most digits are dropped and the result value is
+  ///   rounded up (+1 for positive, -1 for negative) based on the value of the
+  ///   dropped digits (>= 10^reduce_by / 2).
+  /// - If 'round' is false, the right-most digits are simply dropped.
+  BasicDecimal256 ReduceScaleBy(int32_t reduce_by, bool round = true) const;
+
+  /// \brief Whether this number fits in the given precision
+  ///
+  /// Return true if the number of significant digits is less or equal to `precision`.
+  bool FitsInPrecision(int32_t precision) const;
+
+  /// \brief Multiply this number by another number. The result is truncated to 256 bits.
+  BasicDecimal256& operator*=(const BasicDecimal256& right);
+
+  /// Divide this number by right and return the result.
+  ///
+  /// This operation is not destructive.
+  /// The answer rounds to zero. Signs work like:
+  ///   21 /  5 ->  4,  1
+  ///  -21 /  5 -> -4, -1
+  ///   21 / -5 -> -4,  1
+  ///  -21 / -5 ->  4, -1
+  /// \param[in] divisor the number to divide by
+  /// \param[out] result the quotient
+  /// \param[out] remainder the remainder after the division
+  DecimalStatus Divide(const BasicDecimal256& divisor, BasicDecimal256* result,
+                       BasicDecimal256* remainder) const;
+
+  /// \brief Shift left by the given number of bits.
+  BasicDecimal256& operator<<=(uint32_t bits);
+
+  BasicDecimal256 operator<<(uint32_t bits) const {
+    auto res = *this;
+    res <<= bits;
+    return res;
+  }
+
+  /// \brief Shift right by the given number of bits.
+  ///
+  /// Negative values will sign-extend.
+  BasicDecimal256& operator>>=(uint32_t bits);
+
+  BasicDecimal256 operator>>(uint32_t bits) const {
+    auto res = *this;
+    res >>= bits;
+    return res;
+  }
+
+  /// \brief In-place division.
+  BasicDecimal256& operator/=(const BasicDecimal256& right);
+
+  /// \brief Get the maximum valid unscaled decimal value for the given precision.
+  static BasicDecimal256 GetMaxValue(int32_t precision);
+
+  /// \brief Get the maximum decimal value (is not a valid value).
+  static constexpr BasicDecimal256 GetMaxSentinel() {
+#if ARROW_LITTLE_ENDIAN
+    return BasicDecimal256({std::numeric_limits<uint64_t>::max(),
+                            std::numeric_limits<uint64_t>::max(),
+                            std::numeric_limits<uint64_t>::max(),
+                            static_cast<uint64_t>(std::numeric_limits<int64_t>::max())});
+#else
+    return BasicDecimal256({static_cast<uint64_t>(std::numeric_limits<int64_t>::max()),
+                            std::numeric_limits<uint64_t>::max(),
+                            std::numeric_limits<uint64_t>::max(),
+                            std::numeric_limits<uint64_t>::max()});
+#endif
+  }
+  /// \brief Get the minimum decimal value (is not a valid value).
+  static constexpr BasicDecimal256 GetMinSentinel() {
+#if ARROW_LITTLE_ENDIAN
+    return BasicDecimal256(
+        {0, 0, 0, static_cast<uint64_t>(std::numeric_limits<int64_t>::min())});
+#else
+    return BasicDecimal256(
+        {static_cast<uint64_t>(std::numeric_limits<int64_t>::min()), 0, 0, 0});
+#endif
+  }
+};
+
+ARROW_EXPORT bool operator<(const BasicDecimal256& left, const BasicDecimal256& right);
+
+ARROW_EXPORT inline bool operator<=(const BasicDecimal256& left,
+                                    const BasicDecimal256& right) {
+  return !operator<(right, left);
+}
+
+ARROW_EXPORT inline bool operator>(const BasicDecimal256& left,
+                                   const BasicDecimal256& right) {
+  return operator<(right, left);
+}
+
+ARROW_EXPORT inline bool operator>=(const BasicDecimal256& left,
+                                    const BasicDecimal256& right) {
+  return !operator<(left, right);
+}
+
+ARROW_EXPORT BasicDecimal256 operator-(const BasicDecimal256& operand);
+ARROW_EXPORT BasicDecimal256 operator~(const BasicDecimal256& operand);
+ARROW_EXPORT BasicDecimal256 operator+(const BasicDecimal256& left,
+                                       const BasicDecimal256& right);
+ARROW_EXPORT BasicDecimal256 operator*(const BasicDecimal256& left,
+                                       const BasicDecimal256& right);
+ARROW_EXPORT BasicDecimal256 operator/(const BasicDecimal256& left,
+                                       const BasicDecimal256& right);
+
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/benchmark_util.h

@@ -0,0 +1,211 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#include <algorithm>
+#include <cstdint>
+#include <string>
+
+#include "benchmark/benchmark.h"
+
+#include "arrow/memory_pool.h"
+#include "arrow/type_fwd.h"
+#include "arrow/util/cpu_info.h"
+#include "arrow/util/logging.h"  // IWYU pragma: keep
+
+namespace arrow {
+
+// Benchmark changed its parameter type between releases from
+// int to int64_t. As it doesn't have version macros, we need
+// to apply C++ template magic.
+
+template <typename Func>
+struct BenchmarkArgsType;
+
+// Pattern matching that extracts the vector element type of Benchmark::Args()
+template <typename Values>
+struct BenchmarkArgsType<benchmark::internal::Benchmark* (
+    benchmark::internal::Benchmark::*)(const std::vector<Values>&)> {
+  using type = Values;
+};
+
+using ArgsType =
+    typename BenchmarkArgsType<decltype(&benchmark::internal::Benchmark::Args)>::type;
+
+using internal::CpuInfo;
+
+static const CpuInfo* cpu_info = CpuInfo::GetInstance();
+
+static const int64_t kL1Size = cpu_info->CacheSize(CpuInfo::CacheLevel::L1);
+static const int64_t kL2Size = cpu_info->CacheSize(CpuInfo::CacheLevel::L2);
+static const int64_t kL3Size = cpu_info->CacheSize(CpuInfo::CacheLevel::L3);
+static const int64_t kCantFitInL3Size = kL3Size * 4;
+static const std::vector<int64_t> kMemorySizes = {kL1Size, kL2Size, kL3Size,
+                                                  kCantFitInL3Size};
+// 0 is treated as "no nulls"
+static const std::vector<ArgsType> kInverseNullProportions = {10000, 100, 10, 2, 1, 0};
+
+struct GenericItemsArgs {
+  // number of items processed per iteration
+  const int64_t size;
+
+  // proportion of nulls in generated arrays
+  double null_proportion;
+
+  explicit GenericItemsArgs(benchmark::State& state)
+      : size(state.range(0)), state_(state) {
+    if (state.range(1) == 0) {
+      this->null_proportion = 0.0;
+    } else {
+      this->null_proportion = std::min(1., 1. / static_cast<double>(state.range(1)));
+    }
+  }
+
+  ~GenericItemsArgs() {
+    state_.counters["size"] = static_cast<double>(size);
+    state_.counters["null_percent"] = null_proportion * 100;
+    state_.SetItemsProcessed(state_.iterations() * size);
+  }
+
+ private:
+  benchmark::State& state_;
+};
+
+void BenchmarkSetArgsWithSizes(benchmark::internal::Benchmark* bench,
+                               const std::vector<int64_t>& sizes = kMemorySizes) {
+  bench->Unit(benchmark::kMicrosecond);
+
+  for (const auto size : sizes) {
+    for (const auto inverse_null_proportion : kInverseNullProportions) {
+      bench->Args({static_cast<ArgsType>(size), inverse_null_proportion});
+    }
+  }
+}
+
+void BenchmarkSetArgs(benchmark::internal::Benchmark* bench) {
+  BenchmarkSetArgsWithSizes(bench, kMemorySizes);
+}
+
+void RegressionSetArgs(benchmark::internal::Benchmark* bench) {
+  // Regression do not need to account for cache hierarchy, thus optimize for
+  // the best case.
+  BenchmarkSetArgsWithSizes(bench, {kL1Size});
+}
+
+// RAII struct to handle some of the boilerplate in regression benchmarks
+struct RegressionArgs {
+  // size of memory tested (per iteration) in bytes
+  int64_t size;
+
+  // proportion of nulls in generated arrays
+  double null_proportion;
+
+  // If size_is_bytes is true, then it's a number of bytes, otherwise it's the
+  // number of items processed (for reporting)
+  explicit RegressionArgs(benchmark::State& state, bool size_is_bytes = true)
+      : size(state.range(0)), state_(state), size_is_bytes_(size_is_bytes) {
+    if (state.range(1) == 0) {
+      this->null_proportion = 0.0;
+    } else {
+      this->null_proportion = std::min(1., 1. / static_cast<double>(state.range(1)));
+    }
+  }
+
+  ~RegressionArgs() {
+    state_.counters["size"] = static_cast<double>(size);
+    state_.counters["null_percent"] = null_proportion * 100;
+    if (size_is_bytes_) {
+      state_.SetBytesProcessed(state_.iterations() * size);
+    } else {
+      state_.SetItemsProcessed(state_.iterations() * size);
+    }
+  }
+
+ private:
+  benchmark::State& state_;
+  bool size_is_bytes_;
+};
+
+class MemoryPoolMemoryManager : public benchmark::MemoryManager {
+  void Start() override {
+    memory_pool = std::make_shared<ProxyMemoryPool>(default_memory_pool());
+
+    MemoryPool* default_pool = default_memory_pool();
+    global_allocations_start = default_pool->num_allocations();
+  }
+
+// BENCHMARK_DONT_OPTIMIZE is used here to detect Google Benchmark
+// 1.8.0. We can remove this Stop(Result*) when we require Google
+// Benchmark 1.8.0 or later.
+#ifndef BENCHMARK_DONT_OPTIMIZE
+  void Stop(Result* result) override { Stop(*result); }
+#endif
+
+  void Stop(benchmark::MemoryManager::Result& result) override {
+    // If num_allocations is still zero, we assume that the memory pool wasn't passed down
+    // so we should record them.
+    MemoryPool* default_pool = default_memory_pool();
+    int64_t new_default_allocations =
+        default_pool->num_allocations() - global_allocations_start;
+
+    // Only record metrics if (1) there were allocations and (2) we
+    // recorded at least one.
+    if (new_default_allocations > 0 && memory_pool->num_allocations() > 0) {
+      if (new_default_allocations > memory_pool->num_allocations()) {
+        // If we missed some, let's report that.
+        int64_t missed_allocations =
+            new_default_allocations - memory_pool->num_allocations();
+        ARROW_LOG(WARNING) << "BenchmarkMemoryTracker recorded some allocations "
+                           << "for a benchmark, but missed " << missed_allocations
+                           << " allocations.\n";
+      }
+
+      result.max_bytes_used = memory_pool->max_memory();
+      result.total_allocated_bytes = memory_pool->total_bytes_allocated();
+      result.num_allocs = memory_pool->num_allocations();
+    }
+  }
+
+ public:
+  std::shared_ptr<::arrow::ProxyMemoryPool> memory_pool;
+
+ protected:
+  int64_t global_allocations_start;
+};
+
+/// \brief Track memory pool allocations in benchmarks.
+///
+/// Instantiate as a global variable to register the hooks into Google Benchmark
+/// to collect memory metrics. Before each benchmark, a new ProxyMemoryPool is
+/// created. It can then be accessed with memory_pool(). Once the benchmark is
+/// complete, the hook will record the maximum memory used, the total bytes
+/// allocated, and the total number of allocations. If no allocations were seen,
+/// (for example, if you forgot to pass down the memory pool), then these metrics
+/// will not be saved.
+///
+/// Since this is used as one global variable, this will not work if multiple
+/// benchmarks are run concurrently or for multi-threaded benchmarks (ones
+/// that use `->ThreadRange(...)`).
+class BenchmarkMemoryTracker {
+ public:
+  BenchmarkMemoryTracker() : manager_() { ::benchmark::RegisterMemoryManager(&manager_); }
+  ::arrow::MemoryPool* memory_pool() const { return manager_.memory_pool.get(); }
+
+ protected:
+  ::arrow::MemoryPoolMemoryManager manager_;
+};
+
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/binary_view_util.h

@@ -0,0 +1,115 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <string_view>
+#include <utility>
+
+#include "arrow/type.h"
+#include "arrow/util/span.h"
+
+namespace arrow::util {
+
+inline BinaryViewType::c_type ToInlineBinaryView(const void* data, int32_t size) {
+  assert(size <= BinaryViewType::kInlineSize);
+  // Small string: inlined. Bytes beyond size are zeroed
+  BinaryViewType::c_type out;
+  out.inlined = {size, {}};
+  memcpy(&out.inlined.data, data, size);
+  return out;
+}
+
+inline BinaryViewType::c_type ToInlineBinaryView(std::string_view v) {
+  assert(v.size() <= BinaryViewType::kInlineSize);
+  return ToInlineBinaryView(v.data(), static_cast<int32_t>(v.size()));
+}
+
+inline BinaryViewType::c_type ToNonInlineBinaryView(const void* data, int32_t size,
+                                                    int32_t buffer_index,
+                                                    int32_t offset) {
+  // Large string: store index/offset.
+  BinaryViewType::c_type out;
+  out.ref = {size, {}, buffer_index, offset};
+  memcpy(&out.ref.prefix, data, sizeof(out.ref.prefix));
+  return out;
+}
+
+inline BinaryViewType::c_type ToBinaryView(const void* data, int32_t size,
+                                           int32_t buffer_index, int32_t offset) {
+  if (size <= BinaryViewType::kInlineSize) {
+    return ToInlineBinaryView(data, size);
+  }
+  return ToNonInlineBinaryView(data, size, buffer_index, offset);
+}
+
+inline BinaryViewType::c_type ToBinaryView(std::string_view v, int32_t buffer_index,
+                                           int32_t offset) {
+  return ToBinaryView(v.data(), static_cast<int32_t>(v.size()), buffer_index, offset);
+}
+
+template <typename BufferPtr>
+std::string_view FromBinaryView(const BinaryViewType::c_type& v,
+                                const BufferPtr* data_buffers) {
+  auto* data = v.is_inline() ? v.inlined.data.data()
+                             : data_buffers[v.ref.buffer_index]->data() + v.ref.offset;
+  return {reinterpret_cast<const char*>(data), static_cast<size_t>(v.size())};
+}
+template <typename BufferPtr>
+std::string_view FromBinaryView(BinaryViewType::c_type&&, const BufferPtr*) = delete;
+
+template <typename BufferPtr>
+bool EqualBinaryView(BinaryViewType::c_type l, BinaryViewType::c_type r,
+                     const BufferPtr* l_buffers, const BufferPtr* r_buffers) {
+  int64_t l_size_and_prefix, r_size_and_prefix;
+  memcpy(&l_size_and_prefix, &l, sizeof(l_size_and_prefix));
+  memcpy(&r_size_and_prefix, &r, sizeof(r_size_and_prefix));
+
+  if (l_size_and_prefix != r_size_and_prefix) return false;
+
+  if (l.is_inline()) {
+    // The columnar spec mandates that the inlined part be zero-padded, so we can compare
+    // a word at a time regardless of the exact size.
+    int64_t l_inlined, r_inlined;
+    memcpy(&l_inlined, l.inline_data() + BinaryViewType::kPrefixSize, sizeof(l_inlined));
+    memcpy(&r_inlined, r.inline_data() + BinaryViewType::kPrefixSize, sizeof(r_inlined));
+    return l_inlined == r_inlined;
+  }
+
+  // Sizes are equal and this is not inline, therefore both are out
+  // of line and have kPrefixSize first in common.
+  const uint8_t* l_data = l_buffers[l.ref.buffer_index]->data() + l.ref.offset;
+  const uint8_t* r_data = r_buffers[r.ref.buffer_index]->data() + r.ref.offset;
+  return memcmp(l_data + BinaryViewType::kPrefixSize,
+                r_data + BinaryViewType::kPrefixSize,
+                l.size() - BinaryViewType::kPrefixSize) == 0;
+}
+
+/// \brief Compute the total size of a list of binary views including null
+/// views.
+///
+/// This is useful when calculating the necessary memory to store all the string
+/// data from the views.
+inline int64_t SumOfBinaryViewSizes(const BinaryViewType::c_type* views, int64_t length) {
+  int64_t total = 0;
+  for (int64_t i = 0; i < length; ++i) {
+    total += views[i].size();
+  }
+  return total;
+}
+
+}  // namespace arrow::util

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/bit_block_counter.h

@@ -0,0 +1,570 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <algorithm>
+#include <cstdint>
+#include <limits>
+#include <memory>
+
+#include "arrow/buffer.h"
+#include "arrow/status.h"
+#include "arrow/util/bit_util.h"
+#include "arrow/util/endian.h"
+#include "arrow/util/macros.h"
+#include "arrow/util/ubsan.h"
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+namespace internal {
+namespace detail {
+
+inline uint64_t LoadWord(const uint8_t* bytes) {
+  return bit_util::ToLittleEndian(util::SafeLoadAs<uint64_t>(bytes));
+}
+
+inline uint64_t ShiftWord(uint64_t current, uint64_t next, int64_t shift) {
+  if (shift == 0) {
+    return current;
+  }
+  return (current >> shift) | (next << (64 - shift));
+}
+
+// These templates are here to help with unit tests
+
+template <typename T>
+constexpr T BitNot(T x) {
+  return ~x;
+}
+
+template <>
+constexpr bool BitNot(bool x) {
+  return !x;
+}
+
+struct BitBlockAnd {
+  template <typename T>
+  static constexpr T Call(T left, T right) {
+    return left & right;
+  }
+};
+
+struct BitBlockAndNot {
+  template <typename T>
+  static constexpr T Call(T left, T right) {
+    return left & BitNot(right);
+  }
+};
+
+struct BitBlockOr {
+  template <typename T>
+  static constexpr T Call(T left, T right) {
+    return left | right;
+  }
+};
+
+struct BitBlockOrNot {
+  template <typename T>
+  static constexpr T Call(T left, T right) {
+    return left | BitNot(right);
+  }
+};
+
+}  // namespace detail
+
+/// \brief Return value from bit block counters: the total number of bits and
+/// the number of set bits.
+struct BitBlockCount {
+  int16_t length;
+  int16_t popcount;
+
+  bool NoneSet() const { return this->popcount == 0; }
+  bool AllSet() const { return this->length == this->popcount; }
+};
+
+/// \brief A class that scans through a true/false bitmap to compute popcounts
+/// 64 or 256 bits at a time. This is used to accelerate processing of
+/// mostly-not-null array data.
+class ARROW_EXPORT BitBlockCounter {
+ public:
+  BitBlockCounter(const uint8_t* bitmap, int64_t start_offset, int64_t length)
+      : bitmap_(util::MakeNonNull(bitmap) + start_offset / 8),
+        bits_remaining_(length),
+        offset_(start_offset % 8) {}
+
+  /// \brief The bit size of each word run
+  static constexpr int64_t kWordBits = 64;
+
+  /// \brief The bit size of four words run
+  static constexpr int64_t kFourWordsBits = kWordBits * 4;
+
+  /// \brief Return the next run of available bits, usually 256. The returned
+  /// pair contains the size of run and the number of true values. The last
+  /// block will have a length less than 256 if the bitmap length is not a
+  /// multiple of 256, and will return 0-length blocks in subsequent
+  /// invocations.
+  BitBlockCount NextFourWords() {
+    using detail::LoadWord;
+    using detail::ShiftWord;
+
+    if (!bits_remaining_) {
+      return {0, 0};
+    }
+    int64_t total_popcount = 0;
+    if (offset_ == 0) {
+      if (bits_remaining_ < kFourWordsBits) {
+        return GetBlockSlow(kFourWordsBits);
+      }
+      total_popcount += bit_util::PopCount(LoadWord(bitmap_));
+      total_popcount += bit_util::PopCount(LoadWord(bitmap_ + 8));
+      total_popcount += bit_util::PopCount(LoadWord(bitmap_ + 16));
+      total_popcount += bit_util::PopCount(LoadWord(bitmap_ + 24));
+    } else {
+      // When the offset is > 0, we need there to be a word beyond the last
+      // aligned word in the bitmap for the bit shifting logic.
+      if (bits_remaining_ < 5 * kFourWordsBits - offset_) {
+        return GetBlockSlow(kFourWordsBits);
+      }
+      auto current = LoadWord(bitmap_);
+      auto next = LoadWord(bitmap_ + 8);
+      total_popcount += bit_util::PopCount(ShiftWord(current, next, offset_));
+      current = next;
+      next = LoadWord(bitmap_ + 16);
+      total_popcount += bit_util::PopCount(ShiftWord(current, next, offset_));
+      current = next;
+      next = LoadWord(bitmap_ + 24);
+      total_popcount += bit_util::PopCount(ShiftWord(current, next, offset_));
+      current = next;
+      next = LoadWord(bitmap_ + 32);
+      total_popcount += bit_util::PopCount(ShiftWord(current, next, offset_));
+    }
+    bitmap_ += bit_util::BytesForBits(kFourWordsBits);
+    bits_remaining_ -= kFourWordsBits;
+    return {256, static_cast<int16_t>(total_popcount)};
+  }
+
+  /// \brief Return the next run of available bits, usually 64. The returned
+  /// pair contains the size of run and the number of true values. The last
+  /// block will have a length less than 64 if the bitmap length is not a
+  /// multiple of 64, and will return 0-length blocks in subsequent
+  /// invocations.
+  BitBlockCount NextWord() {
+    using detail::LoadWord;
+    using detail::ShiftWord;
+
+    if (!bits_remaining_) {
+      return {0, 0};
+    }
+    int64_t popcount = 0;
+    if (offset_ == 0) {
+      if (bits_remaining_ < kWordBits) {
+        return GetBlockSlow(kWordBits);
+      }
+      popcount = bit_util::PopCount(LoadWord(bitmap_));
+    } else {
+      // When the offset is > 0, we need there to be a word beyond the last
+      // aligned word in the bitmap for the bit shifting logic.
+      if (bits_remaining_ < 2 * kWordBits - offset_) {
+        return GetBlockSlow(kWordBits);
+      }
+      popcount = bit_util::PopCount(
+          ShiftWord(LoadWord(bitmap_), LoadWord(bitmap_ + 8), offset_));
+    }
+    bitmap_ += kWordBits / 8;
+    bits_remaining_ -= kWordBits;
+    return {64, static_cast<int16_t>(popcount)};
+  }
+
+ private:
+  /// \brief Return block with the requested size when doing word-wise
+  /// computation is not possible due to inadequate bits remaining.
+  BitBlockCount GetBlockSlow(int64_t block_size) noexcept;
+
+  const uint8_t* bitmap_;
+  int64_t bits_remaining_;
+  int64_t offset_;
+};
+
+/// \brief A tool to iterate through a possibly nonexistent validity bitmap,
+/// to allow us to write one code path for both the with-nulls and no-nulls
+/// cases without giving up a lot of performance.
+class ARROW_EXPORT OptionalBitBlockCounter {
+ public:
+  // validity_bitmap may be NULLPTR
+  OptionalBitBlockCounter(const uint8_t* validity_bitmap, int64_t offset, int64_t length);
+
+  // validity_bitmap may be null
+  OptionalBitBlockCounter(const std::shared_ptr<Buffer>& validity_bitmap, int64_t offset,
+                          int64_t length);
+
+  /// Return block count for next word when the bitmap is available otherwise
+  /// return a block with length up to INT16_MAX when there is no validity
+  /// bitmap (so all the referenced values are not null).
+  BitBlockCount NextBlock() {
+    static constexpr int64_t kMaxBlockSize = std::numeric_limits<int16_t>::max();
+    if (has_bitmap_) {
+      BitBlockCount block = counter_.NextWord();
+      position_ += block.length;
+      return block;
+    } else {
+      int16_t block_size =
+          static_cast<int16_t>(std::min(kMaxBlockSize, length_ - position_));
+      position_ += block_size;
+      // All values are non-null
+      return {block_size, block_size};
+    }
+  }
+
+  // Like NextBlock, but returns a word-sized block even when there is no
+  // validity bitmap
+  BitBlockCount NextWord() {
+    static constexpr int64_t kWordSize = 64;
+    if (has_bitmap_) {
+      BitBlockCount block = counter_.NextWord();
+      position_ += block.length;
+      return block;
+    } else {
+      int16_t block_size = static_cast<int16_t>(std::min(kWordSize, length_ - position_));
+      position_ += block_size;
+      // All values are non-null
+      return {block_size, block_size};
+    }
+  }
+
+ private:
+  const bool has_bitmap_;
+  int64_t position_;
+  int64_t length_;
+  BitBlockCounter counter_;
+};
+
+/// \brief A class that computes popcounts on the result of bitwise operations
+/// between two bitmaps, 64 bits at a time. A 64-bit word is loaded from each
+/// bitmap, then the popcount is computed on e.g. the bitwise-and of the two
+/// words.
+class ARROW_EXPORT BinaryBitBlockCounter {
+ public:
+  BinaryBitBlockCounter(const uint8_t* left_bitmap, int64_t left_offset,
+                        const uint8_t* right_bitmap, int64_t right_offset, int64_t length)
+      : left_bitmap_(util::MakeNonNull(left_bitmap) + left_offset / 8),
+        left_offset_(left_offset % 8),
+        right_bitmap_(util::MakeNonNull(right_bitmap) + right_offset / 8),
+        right_offset_(right_offset % 8),
+        bits_remaining_(length) {}
+
+  /// \brief Return the popcount of the bitwise-and of the next run of
+  /// available bits, up to 64. The returned pair contains the size of run and
+  /// the number of true values. The last block will have a length less than 64
+  /// if the bitmap length is not a multiple of 64, and will return 0-length
+  /// blocks in subsequent invocations.
+  BitBlockCount NextAndWord() { return NextWord<detail::BitBlockAnd>(); }
+
+  /// \brief Computes "x & ~y" block for each available run of bits.
+  BitBlockCount NextAndNotWord() { return NextWord<detail::BitBlockAndNot>(); }
+
+  /// \brief Computes "x | y" block for each available run of bits.
+  BitBlockCount NextOrWord() { return NextWord<detail::BitBlockOr>(); }
+
+  /// \brief Computes "x | ~y" block for each available run of bits.
+  BitBlockCount NextOrNotWord() { return NextWord<detail::BitBlockOrNot>(); }
+
+ private:
+  template <class Op>
+  BitBlockCount NextWord() {
+    using detail::LoadWord;
+    using detail::ShiftWord;
+
+    if (!bits_remaining_) {
+      return {0, 0};
+    }
+    // When the offset is > 0, we need there to be a word beyond the last aligned
+    // word in the bitmap for the bit shifting logic.
+    constexpr int64_t kWordBits = BitBlockCounter::kWordBits;
+    const int64_t bits_required_to_use_words =
+        std::max(left_offset_ == 0 ? 64 : 64 + (64 - left_offset_),
+                 right_offset_ == 0 ? 64 : 64 + (64 - right_offset_));
+    if (bits_remaining_ < bits_required_to_use_words) {
+      const int16_t run_length =
+          static_cast<int16_t>(std::min(bits_remaining_, kWordBits));
+      int16_t popcount = 0;
+      for (int64_t i = 0; i < run_length; ++i) {
+        if (Op::Call(bit_util::GetBit(left_bitmap_, left_offset_ + i),
+                     bit_util::GetBit(right_bitmap_, right_offset_ + i))) {
+          ++popcount;
+        }
+      }
+      // This code path should trigger _at most_ 2 times. In the "two times"
+      // case, the first time the run length will be a multiple of 8.
+      left_bitmap_ += run_length / 8;
+      right_bitmap_ += run_length / 8;
+      bits_remaining_ -= run_length;
+      return {run_length, popcount};
+    }
+
+    int64_t popcount = 0;
+    if (left_offset_ == 0 && right_offset_ == 0) {
+      popcount =
+          bit_util::PopCount(Op::Call(LoadWord(left_bitmap_), LoadWord(right_bitmap_)));
+    } else {
+      auto left_word =
+          ShiftWord(LoadWord(left_bitmap_), LoadWord(left_bitmap_ + 8), left_offset_);
+      auto right_word =
+          ShiftWord(LoadWord(right_bitmap_), LoadWord(right_bitmap_ + 8), right_offset_);
+      popcount = bit_util::PopCount(Op::Call(left_word, right_word));
+    }
+    left_bitmap_ += kWordBits / 8;
+    right_bitmap_ += kWordBits / 8;
+    bits_remaining_ -= kWordBits;
+    return {64, static_cast<int16_t>(popcount)};
+  }
+
+  const uint8_t* left_bitmap_;
+  int64_t left_offset_;
+  const uint8_t* right_bitmap_;
+  int64_t right_offset_;
+  int64_t bits_remaining_;
+};
+
+class ARROW_EXPORT OptionalBinaryBitBlockCounter {
+ public:
+  // Any bitmap may be NULLPTR
+  OptionalBinaryBitBlockCounter(const uint8_t* left_bitmap, int64_t left_offset,
+                                const uint8_t* right_bitmap, int64_t right_offset,
+                                int64_t length);
+
+  // Any bitmap may be null
+  OptionalBinaryBitBlockCounter(const std::shared_ptr<Buffer>& left_bitmap,
+                                int64_t left_offset,
+                                const std::shared_ptr<Buffer>& right_bitmap,
+                                int64_t right_offset, int64_t length);
+
+  BitBlockCount NextAndBlock() {
+    static constexpr int64_t kMaxBlockSize = std::numeric_limits<int16_t>::max();
+    switch (has_bitmap_) {
+      case HasBitmap::BOTH: {
+        BitBlockCount block = binary_counter_.NextAndWord();
+        position_ += block.length;
+        return block;
+      }
+      case HasBitmap::ONE: {
+        BitBlockCount block = unary_counter_.NextWord();
+        position_ += block.length;
+        return block;
+      }
+      case HasBitmap::NONE:
+      default: {
+        const int16_t block_size =
+            static_cast<int16_t>(std::min(kMaxBlockSize, length_ - position_));
+        position_ += block_size;
+        // All values are non-null
+        return {block_size, block_size};
+      }
+    }
+  }
+
+  BitBlockCount NextOrNotBlock() {
+    static constexpr int64_t kMaxBlockSize = std::numeric_limits<int16_t>::max();
+    switch (has_bitmap_) {
+      case HasBitmap::BOTH: {
+        BitBlockCount block = binary_counter_.NextOrNotWord();
+        position_ += block.length;
+        return block;
+      }
+      case HasBitmap::ONE: {
+        BitBlockCount block = unary_counter_.NextWord();
+        position_ += block.length;
+        return block;
+      }
+      case HasBitmap::NONE:
+      default: {
+        const int16_t block_size =
+            static_cast<int16_t>(std::min(kMaxBlockSize, length_ - position_));
+        position_ += block_size;
+        // All values are non-null
+        return {block_size, block_size};
+      }
+    }
+  }
+
+ private:
+  enum class HasBitmap : int { BOTH, ONE, NONE };
+
+  const HasBitmap has_bitmap_;
+  int64_t position_;
+  int64_t length_;
+  BitBlockCounter unary_counter_;
+  BinaryBitBlockCounter binary_counter_;
+
+  static HasBitmap HasBitmapFromBitmaps(bool has_left, bool has_right) {
+    switch (static_cast<int>(has_left) + static_cast<int>(has_right)) {
+      case 0:
+        return HasBitmap::NONE;
+      case 1:
+        return HasBitmap::ONE;
+      default:  // 2
+        return HasBitmap::BOTH;
+    }
+  }
+};
+
+// Functional-style bit block visitors.
+
+template <typename VisitNotNull, typename VisitNull>
+static Status VisitBitBlocks(const uint8_t* bitmap, int64_t offset, int64_t length,
+                             VisitNotNull&& visit_not_null, VisitNull&& visit_null) {
+  internal::OptionalBitBlockCounter bit_counter(bitmap, offset, length);
+  int64_t position = 0;
+  while (position < length) {
+    internal::BitBlockCount block = bit_counter.NextBlock();
+    if (block.AllSet()) {
+      for (int64_t i = 0; i < block.length; ++i, ++position) {
+        ARROW_RETURN_NOT_OK(visit_not_null(position));
+      }
+    } else if (block.NoneSet()) {
+      for (int64_t i = 0; i < block.length; ++i, ++position) {
+        ARROW_RETURN_NOT_OK(visit_null());
+      }
+    } else {
+      for (int64_t i = 0; i < block.length; ++i, ++position) {
+        if (bit_util::GetBit(bitmap, offset + position)) {
+          ARROW_RETURN_NOT_OK(visit_not_null(position));
+        } else {
+          ARROW_RETURN_NOT_OK(visit_null());
+        }
+      }
+    }
+  }
+  return Status::OK();
+}
+
+template <typename VisitNotNull, typename VisitNull>
+static void VisitBitBlocksVoid(const uint8_t* bitmap, int64_t offset, int64_t length,
+                               VisitNotNull&& visit_not_null, VisitNull&& visit_null) {
+  internal::OptionalBitBlockCounter bit_counter(bitmap, offset, length);
+  int64_t position = 0;
+  while (position < length) {
+    internal::BitBlockCount block = bit_counter.NextBlock();
+    if (block.AllSet()) {
+      for (int64_t i = 0; i < block.length; ++i, ++position) {
+        visit_not_null(position);
+      }
+    } else if (block.NoneSet()) {
+      for (int64_t i = 0; i < block.length; ++i, ++position) {
+        visit_null();
+      }
+    } else {
+      for (int64_t i = 0; i < block.length; ++i, ++position) {
+        if (bit_util::GetBit(bitmap, offset + position)) {
+          visit_not_null(position);
+        } else {
+          visit_null();
+        }
+      }
+    }
+  }
+}
+
+template <typename VisitNotNull, typename VisitNull>
+static Status VisitTwoBitBlocks(const uint8_t* left_bitmap, int64_t left_offset,
+                                const uint8_t* right_bitmap, int64_t right_offset,
+                                int64_t length, VisitNotNull&& visit_not_null,
+                                VisitNull&& visit_null) {
+  if (left_bitmap == NULLPTR || right_bitmap == NULLPTR) {
+    // At most one bitmap is present
+    if (left_bitmap == NULLPTR) {
+      return VisitBitBlocks(right_bitmap, right_offset, length,
+                            std::forward<VisitNotNull>(visit_not_null),
+                            std::forward<VisitNull>(visit_null));
+    } else {
+      return VisitBitBlocks(left_bitmap, left_offset, length,
+                            std::forward<VisitNotNull>(visit_not_null),
+                            std::forward<VisitNull>(visit_null));
+    }
+  }
+  BinaryBitBlockCounter bit_counter(left_bitmap, left_offset, right_bitmap, right_offset,
+                                    length);
+  int64_t position = 0;
+  while (position < length) {
+    BitBlockCount block = bit_counter.NextAndWord();
+    if (block.AllSet()) {
+      for (int64_t i = 0; i < block.length; ++i, ++position) {
+        ARROW_RETURN_NOT_OK(visit_not_null(position));
+      }
+    } else if (block.NoneSet()) {
+      for (int64_t i = 0; i < block.length; ++i, ++position) {
+        ARROW_RETURN_NOT_OK(visit_null());
+      }
+    } else {
+      for (int64_t i = 0; i < block.length; ++i, ++position) {
+        if (bit_util::GetBit(left_bitmap, left_offset + position) &&
+            bit_util::GetBit(right_bitmap, right_offset + position)) {
+          ARROW_RETURN_NOT_OK(visit_not_null(position));
+        } else {
+          ARROW_RETURN_NOT_OK(visit_null());
+        }
+      }
+    }
+  }
+  return Status::OK();
+}
+
+template <typename VisitNotNull, typename VisitNull>
+static void VisitTwoBitBlocksVoid(const uint8_t* left_bitmap, int64_t left_offset,
+                                  const uint8_t* right_bitmap, int64_t right_offset,
+                                  int64_t length, VisitNotNull&& visit_not_null,
+                                  VisitNull&& visit_null) {
+  if (left_bitmap == NULLPTR || right_bitmap == NULLPTR) {
+    // At most one bitmap is present
+    if (left_bitmap == NULLPTR) {
+      return VisitBitBlocksVoid(right_bitmap, right_offset, length,
+                                std::forward<VisitNotNull>(visit_not_null),
+                                std::forward<VisitNull>(visit_null));
+    } else {
+      return VisitBitBlocksVoid(left_bitmap, left_offset, length,
+                                std::forward<VisitNotNull>(visit_not_null),
+                                std::forward<VisitNull>(visit_null));
+    }
+  }
+  BinaryBitBlockCounter bit_counter(left_bitmap, left_offset, right_bitmap, right_offset,
+                                    length);
+  int64_t position = 0;
+  while (position < length) {
+    BitBlockCount block = bit_counter.NextAndWord();
+    if (block.AllSet()) {
+      for (int64_t i = 0; i < block.length; ++i, ++position) {
+        visit_not_null(position);
+      }
+    } else if (block.NoneSet()) {
+      for (int64_t i = 0; i < block.length; ++i, ++position) {
+        visit_null();
+      }
+    } else {
+      for (int64_t i = 0; i < block.length; ++i, ++position) {
+        if (bit_util::GetBit(left_bitmap, left_offset + position) &&
+            bit_util::GetBit(right_bitmap, right_offset + position)) {
+          visit_not_null(position);
+        } else {
+          visit_null();
+        }
+      }
+    }
+  }
+}
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/bit_run_reader.h

@@ -0,0 +1,539 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cassert>
+#include <cstdint>
+#include <cstring>
+#include <string>
+
+#include "arrow/util/bit_util.h"
+#include "arrow/util/bitmap_reader.h"
+#include "arrow/util/endian.h"
+#include "arrow/util/macros.h"
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+namespace internal {
+
+struct BitRun {
+  int64_t length;
+  // Whether bits are set at this point.
+  bool set;
+
+  std::string ToString() const {
+    return std::string("{Length: ") + std::to_string(length) +
+           ", set=" + std::to_string(set) + "}";
+  }
+};
+
+inline bool operator==(const BitRun& lhs, const BitRun& rhs) {
+  return lhs.length == rhs.length && lhs.set == rhs.set;
+}
+
+inline bool operator!=(const BitRun& lhs, const BitRun& rhs) {
+  return lhs.length != rhs.length || lhs.set != rhs.set;
+}
+
+class BitRunReaderLinear {
+ public:
+  BitRunReaderLinear() = default;
+
+  BitRunReaderLinear(const uint8_t* bitmap, int64_t start_offset, int64_t length)
+      : reader_(bitmap, start_offset, length) {}
+
+  BitRun NextRun() {
+    BitRun rl = {/*length=*/0, reader_.IsSet()};
+    // Advance while the values are equal and not at the end of list.
+    while (reader_.position() < reader_.length() && reader_.IsSet() == rl.set) {
+      rl.length++;
+      reader_.Next();
+    }
+    return rl;
+  }
+
+ private:
+  BitmapReader reader_;
+};
+
+#if ARROW_LITTLE_ENDIAN
+/// A convenience class for counting the number of contiguous set/unset bits
+/// in a bitmap.
+class ARROW_EXPORT BitRunReader {
+ public:
+  BitRunReader() = default;
+
+  /// \brief Constructs new BitRunReader.
+  ///
+  /// \param[in] bitmap source data
+  /// \param[in] start_offset bit offset into the source data
+  /// \param[in] length number of bits to copy
+  BitRunReader(const uint8_t* bitmap, int64_t start_offset, int64_t length);
+
+  /// Returns a new BitRun containing the number of contiguous
+  /// bits with the same value.  length == 0 indicates the
+  /// end of the bitmap.
+  BitRun NextRun() {
+    if (ARROW_PREDICT_FALSE(position_ >= length_)) {
+      return {/*length=*/0, false};
+    }
+    // This implementation relies on a efficient implementations of
+    // CountTrailingZeros and assumes that runs are more often then
+    // not.  The logic is to incrementally find the next bit change
+    // from the current position.  This is done by zeroing all
+    // bits in word_ up to position_ and using the TrailingZeroCount
+    // to find the index of the next set bit.
+
+    // The runs alternate on each call, so flip the bit.
+    current_run_bit_set_ = !current_run_bit_set_;
+
+    int64_t start_position = position_;
+    int64_t start_bit_offset = start_position & 63;
+    // Invert the word for proper use of CountTrailingZeros and
+    // clear bits so CountTrailingZeros can do it magic.
+    word_ = ~word_ & ~bit_util::LeastSignificantBitMask(start_bit_offset);
+
+    // Go  forward until the next change from unset to set.
+    int64_t new_bits = bit_util::CountTrailingZeros(word_) - start_bit_offset;
+    position_ += new_bits;
+
+    if (ARROW_PREDICT_FALSE(bit_util::IsMultipleOf64(position_)) &&
+        ARROW_PREDICT_TRUE(position_ < length_)) {
+      // Continue extending position while we can advance an entire word.
+      // (updates position_ accordingly).
+      AdvanceUntilChange();
+    }
+
+    return {/*length=*/position_ - start_position, current_run_bit_set_};
+  }
+
+ private:
+  void AdvanceUntilChange() {
+    int64_t new_bits = 0;
+    do {
+      // Advance the position of the bitmap for loading.
+      bitmap_ += sizeof(uint64_t);
+      LoadNextWord();
+      new_bits = bit_util::CountTrailingZeros(word_);
+      // Continue calculating run length.
+      position_ += new_bits;
+    } while (ARROW_PREDICT_FALSE(bit_util::IsMultipleOf64(position_)) &&
+             ARROW_PREDICT_TRUE(position_ < length_) && new_bits > 0);
+  }
+
+  void LoadNextWord() { return LoadWord(length_ - position_); }
+
+  // Helper method for Loading the next word.
+  void LoadWord(int64_t bits_remaining) {
+    word_ = 0;
+    // we need at least an extra byte in this case.
+    if (ARROW_PREDICT_TRUE(bits_remaining >= 64)) {
+      std::memcpy(&word_, bitmap_, 8);
+    } else {
+      int64_t bytes_to_load = bit_util::BytesForBits(bits_remaining);
+      auto word_ptr = reinterpret_cast<uint8_t*>(&word_);
+      std::memcpy(word_ptr, bitmap_, bytes_to_load);
+      // Ensure stoppage at last bit in bitmap by reversing the next higher
+      // order bit.
+      bit_util::SetBitTo(word_ptr, bits_remaining,
+                         !bit_util::GetBit(word_ptr, bits_remaining - 1));
+    }
+
+    // Two cases:
+    //   1. For unset, CountTrailingZeros works naturally so we don't
+    //   invert the word.
+    //   2. Otherwise invert so we can use CountTrailingZeros.
+    if (current_run_bit_set_) {
+      word_ = ~word_;
+    }
+  }
+  const uint8_t* bitmap_;
+  int64_t position_;
+  int64_t length_;
+  uint64_t word_;
+  bool current_run_bit_set_;
+};
+#else
+using BitRunReader = BitRunReaderLinear;
+#endif
+
+struct SetBitRun {
+  int64_t position;
+  int64_t length;
+
+  bool AtEnd() const { return length == 0; }
+
+  std::string ToString() const {
+    return std::string("{pos=") + std::to_string(position) +
+           ", len=" + std::to_string(length) + "}";
+  }
+
+  bool operator==(const SetBitRun& other) const {
+    return position == other.position && length == other.length;
+  }
+  bool operator!=(const SetBitRun& other) const {
+    return position != other.position || length != other.length;
+  }
+};
+
+template <bool Reverse>
+class BaseSetBitRunReader {
+ public:
+  /// \brief Constructs new SetBitRunReader.
+  ///
+  /// \param[in] bitmap source data
+  /// \param[in] start_offset bit offset into the source data
+  /// \param[in] length number of bits to copy
+  ARROW_NOINLINE
+  BaseSetBitRunReader(const uint8_t* bitmap, int64_t start_offset, int64_t length)
+      : bitmap_(util::MakeNonNull(bitmap)),
+        length_(length),
+        remaining_(length_),
+        current_word_(0),
+        current_num_bits_(0) {
+    if (Reverse) {
+      bitmap_ += (start_offset + length) / 8;
+      const int8_t end_bit_offset = static_cast<int8_t>((start_offset + length) % 8);
+      if (length > 0 && end_bit_offset) {
+        // Get LSBs from last byte
+        ++bitmap_;
+        current_num_bits_ =
+            std::min(static_cast<int32_t>(length), static_cast<int32_t>(end_bit_offset));
+        current_word_ = LoadPartialWord(8 - end_bit_offset, current_num_bits_);
+      }
+    } else {
+      bitmap_ += start_offset / 8;
+      const int8_t bit_offset = static_cast<int8_t>(start_offset % 8);
+      if (length > 0 && bit_offset) {
+        // Get MSBs from first byte
+        current_num_bits_ =
+            std::min(static_cast<int32_t>(length), static_cast<int32_t>(8 - bit_offset));
+        current_word_ = LoadPartialWord(bit_offset, current_num_bits_);
+      }
+    }
+  }
+
+  ARROW_NOINLINE
+  SetBitRun NextRun() {
+    int64_t pos = 0;
+    int64_t len = 0;
+    if (current_num_bits_) {
+      const auto run = FindCurrentRun();
+      assert(remaining_ >= 0);
+      if (run.length && current_num_bits_) {
+        // The run ends in current_word_
+        return AdjustRun(run);
+      }
+      pos = run.position;
+      len = run.length;
+    }
+    if (!len) {
+      // We didn't get any ones in current_word_, so we can skip any zeros
+      // in the following words
+      SkipNextZeros();
+      if (remaining_ == 0) {
+        return {0, 0};
+      }
+      assert(current_num_bits_);
+      pos = position();
+    } else if (!current_num_bits_) {
+      if (ARROW_PREDICT_TRUE(remaining_ >= 64)) {
+        current_word_ = LoadFullWord();
+        current_num_bits_ = 64;
+      } else if (remaining_ > 0) {
+        current_word_ = LoadPartialWord(/*bit_offset=*/0, remaining_);
+        current_num_bits_ = static_cast<int32_t>(remaining_);
+      } else {
+        // No bits remaining, perhaps we found a run?
+        return AdjustRun({pos, len});
+      }
+      // If current word starts with a zero, we got a full run
+      if (!(current_word_ & kFirstBit)) {
+        return AdjustRun({pos, len});
+      }
+    }
+    // Current word should now start with a set bit
+    len += CountNextOnes();
+    return AdjustRun({pos, len});
+  }
+
+ protected:
+  int64_t position() const {
+    if (Reverse) {
+      return remaining_;
+    } else {
+      return length_ - remaining_;
+    }
+  }
+
+  SetBitRun AdjustRun(SetBitRun run) {
+    if (Reverse) {
+      assert(run.position >= run.length);
+      run.position -= run.length;
+    }
+    return run;
+  }
+
+  uint64_t LoadFullWord() {
+    uint64_t word;
+    if (Reverse) {
+      bitmap_ -= 8;
+    }
+    memcpy(&word, bitmap_, 8);
+    if (!Reverse) {
+      bitmap_ += 8;
+    }
+    return bit_util::ToLittleEndian(word);
+  }
+
+  uint64_t LoadPartialWord(int8_t bit_offset, int64_t num_bits) {
+    assert(num_bits > 0);
+    uint64_t word = 0;
+    const int64_t num_bytes = bit_util::BytesForBits(num_bits);
+    if (Reverse) {
+      // Read in the most significant bytes of the word
+      bitmap_ -= num_bytes;
+      memcpy(reinterpret_cast<char*>(&word) + 8 - num_bytes, bitmap_, num_bytes);
+      // XXX MostSignificantBitmask
+      return (bit_util::ToLittleEndian(word) << bit_offset) &
+             ~bit_util::LeastSignificantBitMask(64 - num_bits);
+    } else {
+      memcpy(&word, bitmap_, num_bytes);
+      bitmap_ += num_bytes;
+      return (bit_util::ToLittleEndian(word) >> bit_offset) &
+             bit_util::LeastSignificantBitMask(num_bits);
+    }
+  }
+
+  void SkipNextZeros() {
+    assert(current_num_bits_ == 0);
+    while (ARROW_PREDICT_TRUE(remaining_ >= 64)) {
+      current_word_ = LoadFullWord();
+      const auto num_zeros = CountFirstZeros(current_word_);
+      if (num_zeros < 64) {
+        // Run of zeros ends here
+        current_word_ = ConsumeBits(current_word_, num_zeros);
+        current_num_bits_ = 64 - num_zeros;
+        remaining_ -= num_zeros;
+        assert(remaining_ >= 0);
+        assert(current_num_bits_ >= 0);
+        return;
+      }
+      remaining_ -= 64;
+    }
+    // Run of zeros continues in last bitmap word
+    if (remaining_ > 0) {
+      current_word_ = LoadPartialWord(/*bit_offset=*/0, remaining_);
+      current_num_bits_ = static_cast<int32_t>(remaining_);
+      const auto num_zeros =
+          std::min<int32_t>(current_num_bits_, CountFirstZeros(current_word_));
+      current_word_ = ConsumeBits(current_word_, num_zeros);
+      current_num_bits_ -= num_zeros;
+      remaining_ -= num_zeros;
+      assert(remaining_ >= 0);
+      assert(current_num_bits_ >= 0);
+    }
+  }
+
+  int64_t CountNextOnes() {
+    assert(current_word_ & kFirstBit);
+
+    int64_t len;
+    if (~current_word_) {
+      const auto num_ones = CountFirstZeros(~current_word_);
+      assert(num_ones <= current_num_bits_);
+      assert(num_ones <= remaining_);
+      remaining_ -= num_ones;
+      current_word_ = ConsumeBits(current_word_, num_ones);
+      current_num_bits_ -= num_ones;
+      if (current_num_bits_) {
+        // Run of ones ends here
+        return num_ones;
+      }
+      len = num_ones;
+    } else {
+      // current_word_ is all ones
+      remaining_ -= 64;
+      current_num_bits_ = 0;
+      len = 64;
+    }
+
+    while (ARROW_PREDICT_TRUE(remaining_ >= 64)) {
+      current_word_ = LoadFullWord();
+      const auto num_ones = CountFirstZeros(~current_word_);
+      len += num_ones;
+      remaining_ -= num_ones;
+      if (num_ones < 64) {
+        // Run of ones ends here
+        current_word_ = ConsumeBits(current_word_, num_ones);
+        current_num_bits_ = 64 - num_ones;
+        return len;
+      }
+    }
+    // Run of ones continues in last bitmap word
+    if (remaining_ > 0) {
+      current_word_ = LoadPartialWord(/*bit_offset=*/0, remaining_);
+      current_num_bits_ = static_cast<int32_t>(remaining_);
+      const auto num_ones = CountFirstZeros(~current_word_);
+      assert(num_ones <= current_num_bits_);
+      assert(num_ones <= remaining_);
+      current_word_ = ConsumeBits(current_word_, num_ones);
+      current_num_bits_ -= num_ones;
+      remaining_ -= num_ones;
+      len += num_ones;
+    }
+    return len;
+  }
+
+  SetBitRun FindCurrentRun() {
+    // Skip any pending zeros
+    const auto num_zeros = CountFirstZeros(current_word_);
+    if (num_zeros >= current_num_bits_) {
+      remaining_ -= current_num_bits_;
+      current_word_ = 0;
+      current_num_bits_ = 0;
+      return {0, 0};
+    }
+    assert(num_zeros <= remaining_);
+    current_word_ = ConsumeBits(current_word_, num_zeros);
+    current_num_bits_ -= num_zeros;
+    remaining_ -= num_zeros;
+    const int64_t pos = position();
+    // Count any ones
+    const auto num_ones = CountFirstZeros(~current_word_);
+    assert(num_ones <= current_num_bits_);
+    assert(num_ones <= remaining_);
+    current_word_ = ConsumeBits(current_word_, num_ones);
+    current_num_bits_ -= num_ones;
+    remaining_ -= num_ones;
+    return {pos, num_ones};
+  }
+
+  inline int CountFirstZeros(uint64_t word);
+  inline uint64_t ConsumeBits(uint64_t word, int32_t num_bits);
+
+  const uint8_t* bitmap_;
+  const int64_t length_;
+  int64_t remaining_;
+  uint64_t current_word_;
+  int32_t current_num_bits_;
+
+  static constexpr uint64_t kFirstBit = Reverse ? 0x8000000000000000ULL : 1;
+};
+
+template <>
+inline int BaseSetBitRunReader<false>::CountFirstZeros(uint64_t word) {
+  return bit_util::CountTrailingZeros(word);
+}
+
+template <>
+inline int BaseSetBitRunReader<true>::CountFirstZeros(uint64_t word) {
+  return bit_util::CountLeadingZeros(word);
+}
+
+template <>
+inline uint64_t BaseSetBitRunReader<false>::ConsumeBits(uint64_t word, int32_t num_bits) {
+  return word >> num_bits;
+}
+
+template <>
+inline uint64_t BaseSetBitRunReader<true>::ConsumeBits(uint64_t word, int32_t num_bits) {
+  return word << num_bits;
+}
+
+using SetBitRunReader = BaseSetBitRunReader</*Reverse=*/false>;
+using ReverseSetBitRunReader = BaseSetBitRunReader</*Reverse=*/true>;
+
+// Functional-style bit run visitors.
+
+template <typename Visit>
+inline Status VisitBitRuns(const uint8_t* bitmap, int64_t offset, int64_t length,
+                           Visit&& visit) {
+  if (bitmap == NULLPTR) {
+    // Assuming all set (as in a null bitmap)
+    return visit(static_cast<int64_t>(0), length, true);
+  }
+  BitRunReader reader(bitmap, offset, length);
+  int64_t position = 0;
+  while (true) {
+    const auto run = reader.NextRun();
+    if (run.length == 0) {
+      break;
+    }
+    ARROW_RETURN_NOT_OK(visit(position, run.length, run.set));
+    position += run.length;
+  }
+  return Status::OK();
+}
+
+// XXX: Try to make this function small so the compiler can inline and optimize
+// the `visit` function, which is normally a hot loop with vectorizable code.
+// - don't inline SetBitRunReader constructor, it doesn't hurt performance
+// - un-inline NextRun hurts 'many null' cases a bit, but improves normal cases
+template <typename Visit>
+inline Status VisitSetBitRuns(const uint8_t* bitmap, int64_t offset, int64_t length,
+                              Visit&& visit) {
+  if (bitmap == NULLPTR) {
+    // Assuming all set (as in a null bitmap)
+    return visit(static_cast<int64_t>(0), static_cast<int64_t>(length));
+  }
+  SetBitRunReader reader(bitmap, offset, length);
+  while (true) {
+    const auto run = reader.NextRun();
+    if (run.length == 0) {
+      break;
+    }
+    ARROW_RETURN_NOT_OK(visit(run.position, run.length));
+  }
+  return Status::OK();
+}
+
+template <typename Visit>
+inline void VisitSetBitRunsVoid(const uint8_t* bitmap, int64_t offset, int64_t length,
+                                Visit&& visit) {
+  if (bitmap == NULLPTR) {
+    // Assuming all set (as in a null bitmap)
+    visit(static_cast<int64_t>(0), static_cast<int64_t>(length));
+    return;
+  }
+  SetBitRunReader reader(bitmap, offset, length);
+  while (true) {
+    const auto run = reader.NextRun();
+    if (run.length == 0) {
+      break;
+    }
+    visit(run.position, run.length);
+  }
+}
+
+template <typename Visit>
+inline Status VisitSetBitRuns(const std::shared_ptr<Buffer>& bitmap, int64_t offset,
+                              int64_t length, Visit&& visit) {
+  return VisitSetBitRuns(bitmap ? bitmap->data() : NULLPTR, offset, length,
+                         std::forward<Visit>(visit));
+}
+
+template <typename Visit>
+inline void VisitSetBitRunsVoid(const std::shared_ptr<Buffer>& bitmap, int64_t offset,
+                                int64_t length, Visit&& visit) {
+  VisitSetBitRunsVoid(bitmap ? bitmap->data() : NULLPTR, offset, length,
+                      std::forward<Visit>(visit));
+}
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/bit_util.h

@@ -0,0 +1,490 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#if defined(_MSC_VER)
+#  if defined(_M_AMD64) || defined(_M_X64)
+#    include <intrin.h>  // IWYU pragma: keep
+#  endif
+
+#  pragma intrinsic(_BitScanReverse)
+#  pragma intrinsic(_BitScanForward)
+#  define ARROW_POPCOUNT64 __popcnt64
+#  define ARROW_POPCOUNT32 __popcnt
+#else
+#  define ARROW_POPCOUNT64 __builtin_popcountll
+#  define ARROW_POPCOUNT32 __builtin_popcount
+#endif
+
+#include <cstdint>
+#include <type_traits>
+
+#include "arrow/util/macros.h"
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+namespace detail {
+
+template <typename Integer>
+typename std::make_unsigned<Integer>::type as_unsigned(Integer x) {
+  return static_cast<typename std::make_unsigned<Integer>::type>(x);
+}
+
+}  // namespace detail
+
+namespace bit_util {
+
+// The number of set bits in a given unsigned byte value, pre-computed
+//
+// Generated with the following Python code
+// output = 'static constexpr uint8_t kBytePopcount[] = {{{0}}};'
+// popcounts = [str(bin(i).count('1')) for i in range(0, 256)]
+// print(output.format(', '.join(popcounts)))
+static constexpr uint8_t kBytePopcount[] = {
+    0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4, 1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3,
+    4, 4, 5, 1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5, 2, 3, 3, 4, 3, 4, 4, 5, 3, 4,
+    4, 5, 4, 5, 5, 6, 1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5, 2, 3, 3, 4, 3, 4, 4,
+    5, 3, 4, 4, 5, 4, 5, 5, 6, 2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6, 3, 4, 4, 5,
+    4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7, 1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5, 2,
+    3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6, 2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5,
+    5, 6, 3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7, 2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4,
+    5, 4, 5, 5, 6, 3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7, 3, 4, 4, 5, 4, 5, 5, 6,
+    4, 5, 5, 6, 5, 6, 6, 7, 4, 5, 5, 6, 5, 6, 6, 7, 5, 6, 6, 7, 6, 7, 7, 8};
+
+static inline uint64_t PopCount(uint64_t bitmap) { return ARROW_POPCOUNT64(bitmap); }
+static inline uint32_t PopCount(uint32_t bitmap) { return ARROW_POPCOUNT32(bitmap); }
+
+//
+// Bit-related computations on integer values
+//
+
+// Returns the ceil of value/divisor
+constexpr int64_t CeilDiv(int64_t value, int64_t divisor) {
+  return (value == 0) ? 0 : 1 + (value - 1) / divisor;
+}
+
+// Return the number of bytes needed to fit the given number of bits
+constexpr int64_t BytesForBits(int64_t bits) {
+  // This formula avoids integer overflow on very large `bits`
+  return (bits >> 3) + ((bits & 7) != 0);
+}
+
+constexpr bool IsPowerOf2(int64_t value) {
+  return value > 0 && (value & (value - 1)) == 0;
+}
+
+constexpr bool IsPowerOf2(uint64_t value) {
+  return value > 0 && (value & (value - 1)) == 0;
+}
+
+// Returns the smallest power of two that contains v.  If v is already a
+// power of two, it is returned as is.
+static inline int64_t NextPower2(int64_t n) {
+  // Taken from
+  // http://graphics.stanford.edu/~seander/bithacks.html#RoundUpPowerOf2
+  n--;
+  n |= n >> 1;
+  n |= n >> 2;
+  n |= n >> 4;
+  n |= n >> 8;
+  n |= n >> 16;
+  n |= n >> 32;
+  n++;
+  return n;
+}
+
+constexpr bool IsMultipleOf64(int64_t n) { return (n & 63) == 0; }
+
+constexpr bool IsMultipleOf8(int64_t n) { return (n & 7) == 0; }
+
+// Returns a mask for the bit_index lower order bits.
+// Only valid for bit_index in the range [0, 64).
+constexpr uint64_t LeastSignificantBitMask(int64_t bit_index) {
+  return (static_cast<uint64_t>(1) << bit_index) - 1;
+}
+
+// Returns 'value' rounded up to the nearest multiple of 'factor'
+constexpr int64_t RoundUp(int64_t value, int64_t factor) {
+  return CeilDiv(value, factor) * factor;
+}
+
+// Returns 'value' rounded down to the nearest multiple of 'factor'
+constexpr int64_t RoundDown(int64_t value, int64_t factor) {
+  return (value / factor) * factor;
+}
+
+// Returns 'value' rounded up to the nearest multiple of 'factor' when factor
+// is a power of two.
+// The result is undefined on overflow, i.e. if `value > 2**64 - factor`,
+// since we cannot return the correct result which would be 2**64.
+constexpr int64_t RoundUpToPowerOf2(int64_t value, int64_t factor) {
+  // DCHECK(value >= 0);
+  // DCHECK(IsPowerOf2(factor));
+  return (value + (factor - 1)) & ~(factor - 1);
+}
+
+constexpr uint64_t RoundUpToPowerOf2(uint64_t value, uint64_t factor) {
+  // DCHECK(IsPowerOf2(factor));
+  return (value + (factor - 1)) & ~(factor - 1);
+}
+
+constexpr int64_t RoundUpToMultipleOf8(int64_t num) { return RoundUpToPowerOf2(num, 8); }
+
+constexpr int64_t RoundUpToMultipleOf64(int64_t num) {
+  return RoundUpToPowerOf2(num, 64);
+}
+
+// Returns the number of bytes covering a sliced bitmap. Find the length
+// rounded to cover full bytes on both extremities.
+//
+// The following example represents a slice (offset=10, length=9)
+//
+// 0       8       16     24
+// |-------|-------|------|
+//           [       ]          (slice)
+//         [             ]      (same slice aligned to bytes bounds, length=16)
+//
+// The covering bytes is the length (in bytes) of this new aligned slice.
+constexpr int64_t CoveringBytes(int64_t offset, int64_t length) {
+  return (bit_util::RoundUp(length + offset, 8) - bit_util::RoundDown(offset, 8)) / 8;
+}
+
+// Returns the 'num_bits' least-significant bits of 'v'.
+static inline uint64_t TrailingBits(uint64_t v, int num_bits) {
+  if (ARROW_PREDICT_FALSE(num_bits == 0)) return 0;
+  if (ARROW_PREDICT_FALSE(num_bits >= 64)) return v;
+  int n = 64 - num_bits;
+  return (v << n) >> n;
+}
+
+/// \brief Count the number of leading zeros in an unsigned integer.
+static inline int CountLeadingZeros(uint32_t value) {
+#if defined(__clang__) || defined(__GNUC__)
+  if (value == 0) return 32;
+  return static_cast<int>(__builtin_clz(value));
+#elif defined(_MSC_VER)
+  unsigned long index;                                               // NOLINT
+  if (_BitScanReverse(&index, static_cast<unsigned long>(value))) {  // NOLINT
+    return 31 - static_cast<int>(index);
+  } else {
+    return 32;
+  }
+#else
+  int bitpos = 0;
+  while (value != 0) {
+    value >>= 1;
+    ++bitpos;
+  }
+  return 32 - bitpos;
+#endif
+}
+
+static inline int CountLeadingZeros(uint64_t value) {
+#if defined(__clang__) || defined(__GNUC__)
+  if (value == 0) return 64;
+  return static_cast<int>(__builtin_clzll(value));
+#elif defined(_MSC_VER)
+  unsigned long index;                     // NOLINT
+  if (_BitScanReverse64(&index, value)) {  // NOLINT
+    return 63 - static_cast<int>(index);
+  } else {
+    return 64;
+  }
+#else
+  int bitpos = 0;
+  while (value != 0) {
+    value >>= 1;
+    ++bitpos;
+  }
+  return 64 - bitpos;
+#endif
+}
+
+static inline int CountTrailingZeros(uint32_t value) {
+#if defined(__clang__) || defined(__GNUC__)
+  if (value == 0) return 32;
+  return static_cast<int>(__builtin_ctzl(value));
+#elif defined(_MSC_VER)
+  unsigned long index;  // NOLINT
+  if (_BitScanForward(&index, value)) {
+    return static_cast<int>(index);
+  } else {
+    return 32;
+  }
+#else
+  int bitpos = 0;
+  if (value) {
+    while (value & 1 == 0) {
+      value >>= 1;
+      ++bitpos;
+    }
+  } else {
+    bitpos = 32;
+  }
+  return bitpos;
+#endif
+}
+
+static inline int CountTrailingZeros(uint64_t value) {
+#if defined(__clang__) || defined(__GNUC__)
+  if (value == 0) return 64;
+  return static_cast<int>(__builtin_ctzll(value));
+#elif defined(_MSC_VER)
+  unsigned long index;  // NOLINT
+  if (_BitScanForward64(&index, value)) {
+    return static_cast<int>(index);
+  } else {
+    return 64;
+  }
+#else
+  int bitpos = 0;
+  if (value) {
+    while (value & 1 == 0) {
+      value >>= 1;
+      ++bitpos;
+    }
+  } else {
+    bitpos = 64;
+  }
+  return bitpos;
+#endif
+}
+
+// Returns the minimum number of bits needed to represent an unsigned value
+static inline int NumRequiredBits(uint64_t x) { return 64 - CountLeadingZeros(x); }
+
+// Returns ceil(log2(x)).
+static inline int Log2(uint64_t x) {
+  // DCHECK_GT(x, 0);
+  return NumRequiredBits(x - 1);
+}
+
+//
+// Utilities for reading and writing individual bits by their index
+// in a memory area.
+//
+
+// Bitmask selecting the k-th bit in a byte
+static constexpr uint8_t kBitmask[] = {1, 2, 4, 8, 16, 32, 64, 128};
+
+// the bitwise complement version of kBitmask
+static constexpr uint8_t kFlippedBitmask[] = {254, 253, 251, 247, 239, 223, 191, 127};
+
+// Bitmask selecting the (k - 1) preceding bits in a byte
+static constexpr uint8_t kPrecedingBitmask[] = {0, 1, 3, 7, 15, 31, 63, 127};
+static constexpr uint8_t kPrecedingWrappingBitmask[] = {255, 1, 3, 7, 15, 31, 63, 127};
+
+// the bitwise complement version of kPrecedingBitmask
+static constexpr uint8_t kTrailingBitmask[] = {255, 254, 252, 248, 240, 224, 192, 128};
+
+static constexpr bool GetBit(const uint8_t* bits, uint64_t i) {
+  return (bits[i >> 3] >> (i & 0x07)) & 1;
+}
+
+// Gets the i-th bit from a byte. Should only be used with i <= 7.
+static constexpr bool GetBitFromByte(uint8_t byte, uint8_t i) {
+  return byte & kBitmask[i];
+}
+
+static inline void ClearBit(uint8_t* bits, int64_t i) {
+  bits[i / 8] &= kFlippedBitmask[i % 8];
+}
+
+static inline void SetBit(uint8_t* bits, int64_t i) { bits[i / 8] |= kBitmask[i % 8]; }
+
+static inline void SetBitTo(uint8_t* bits, int64_t i, bool bit_is_set) {
+  // https://graphics.stanford.edu/~seander/bithacks.html
+  // "Conditionally set or clear bits without branching"
+  // NOTE: this seems to confuse Valgrind as it reads from potentially
+  // uninitialized memory
+  bits[i / 8] ^= static_cast<uint8_t>(-static_cast<uint8_t>(bit_is_set) ^ bits[i / 8]) &
+                 kBitmask[i % 8];
+}
+
+/// \brief set or clear a range of bits quickly
+ARROW_EXPORT
+void SetBitsTo(uint8_t* bits, int64_t start_offset, int64_t length, bool bits_are_set);
+
+/// \brief Sets all bits in the bitmap to true
+ARROW_EXPORT
+void SetBitmap(uint8_t* data, int64_t offset, int64_t length);
+
+/// \brief Clears all bits in the bitmap (set to false)
+ARROW_EXPORT
+void ClearBitmap(uint8_t* data, int64_t offset, int64_t length);
+
+/// Returns a mask with lower i bits set to 1. If i >= sizeof(Word)*8, all-ones will be
+/// returned
+/// ex:
+/// ref: https://stackoverflow.com/a/59523400
+template <typename Word>
+constexpr Word PrecedingWordBitmask(const unsigned int i) {
+  return static_cast<Word>(static_cast<Word>(i < sizeof(Word) * 8)
+                           << (i & (sizeof(Word) * 8 - 1))) -
+         1;
+}
+static_assert(PrecedingWordBitmask<uint8_t>(0) == 0x00, "");
+static_assert(PrecedingWordBitmask<uint8_t>(4) == 0x0f, "");
+static_assert(PrecedingWordBitmask<uint8_t>(8) == 0xff, "");
+static_assert(PrecedingWordBitmask<uint16_t>(8) == 0x00ff, "");
+
+/// \brief Create a word with low `n` bits from `low` and high `sizeof(Word)-n` bits
+/// from `high`.
+/// Word ret
+/// for (i = 0; i < sizeof(Word)*8; i++){
+///     ret[i]= i < n ? low[i]: high[i];
+/// }
+template <typename Word>
+constexpr Word SpliceWord(int n, Word low, Word high) {
+  return (high & ~PrecedingWordBitmask<Word>(n)) | (low & PrecedingWordBitmask<Word>(n));
+}
+
+/// \brief Pack integers into a bitmap in batches of 8
+template <int batch_size>
+void PackBits(const uint32_t* values, uint8_t* out) {
+  for (int i = 0; i < batch_size / 8; ++i) {
+    *out++ = static_cast<uint8_t>(values[0] | values[1] << 1 | values[2] << 2 |
+                                  values[3] << 3 | values[4] << 4 | values[5] << 5 |
+                                  values[6] << 6 | values[7] << 7);
+    values += 8;
+  }
+}
+
+constexpr int64_t MaxLEB128ByteLen(int64_t n_bits) { return CeilDiv(n_bits, 7); }
+
+template <typename Int>
+constexpr int64_t kMaxLEB128ByteLenFor = MaxLEB128ByteLen(sizeof(Int) * 8);
+
+/// Write a integer as LEB128
+///
+/// Write the input value as LEB128 into the outptut buffer and return the number of bytes
+/// written.
+/// If the output buffer size is insufficient, return 0 but the output may have been
+/// written to.
+/// The input value can be a signed integer, but must be non negative.
+///
+/// \see https://en.wikipedia.org/wiki/LEB128
+/// \see MaxLEB128ByteLenFor
+template <typename Int>
+constexpr int32_t WriteLEB128(Int value, uint8_t* out, int32_t max_out_size) {
+  constexpr Int kLow7Mask = Int(0x7F);
+  constexpr Int kHigh7Mask = ~kLow7Mask;
+  constexpr uint8_t kContinuationBit = 0x80;
+
+  // This encoding does not work for negative values
+  if constexpr (std::is_signed_v<Int>) {
+    if (ARROW_PREDICT_FALSE(value < 0)) {
+      return 0;
+    }
+  }
+
+  const auto out_first = out;
+
+  // Write as many bytes as we could be for the given input
+  while ((value & kHigh7Mask) != Int(0)) {
+    // We do not have enough room to write the LEB128
+    if (ARROW_PREDICT_FALSE(out - out_first >= max_out_size)) {
+      return 0;
+    }
+
+    // Write the encoded byte with continuation bit
+    *out = static_cast<uint8_t>(value & kLow7Mask) | kContinuationBit;
+    ++out;
+    // Shift remaining data
+    value >>= 7;
+  }
+
+  // We do not have enough room to write the LEB128
+  if (ARROW_PREDICT_FALSE(out - out_first >= max_out_size)) {
+    return 0;
+  }
+
+  // Write last non-continuing byte
+  *out = static_cast<uint8_t>(value & kLow7Mask);
+  ++out;
+
+  return static_cast<int32_t>(out - out_first);
+}
+
+/// Parse a leading LEB128
+///
+/// Take as input a data pointer and the maximum number of bytes that can be read from it
+/// (typically the array size).
+/// When a valid LEB128 is found at the start of the data, the function writes it to the
+/// out pointer and return the number of bytes read.
+/// Otherwise, the out pointer is unmodified and zero is returned.
+///
+/// \see https://en.wikipedia.org/wiki/LEB128
+/// \see MaxLEB128ByteLenFor
+template <typename Int>
+constexpr int32_t ParseLeadingLEB128(const uint8_t* data, int32_t max_data_size,
+                                     Int* out) {
+  constexpr auto kMaxBytes = static_cast<int32_t>(kMaxLEB128ByteLenFor<Int>);
+  static_assert(kMaxBytes >= 1);
+  constexpr uint8_t kLow7Mask = 0x7F;
+  constexpr uint8_t kContinuationBit = 0x80;
+  constexpr int32_t kSignBitCount = std::is_signed_v<Int> ? 1 : 0;
+  // Number of bits allowed for encoding data on the last byte to avoid overflow
+  constexpr uint8_t kHighBitCount = (8 * sizeof(Int) - kSignBitCount) % 7;
+  // kHighBitCount least significant `0` bits and the rest with `1`
+  constexpr uint8_t kHighForbiddenMask = ~((1 << kHighBitCount) - 1);
+
+  // Iteratively building the value
+  std::make_unsigned_t<Int> value = 0;
+
+  // Read as many bytes as we could be for the given output.
+  for (int32_t i = 0; i < kMaxBytes - 1; i++) {
+    // We have not finished reading a valid LEB128, yet we run out of data
+    if (ARROW_PREDICT_FALSE(i >= max_data_size)) {
+      return 0;
+    }
+
+    // Read the byte and set its 7 LSB to in the final value
+    const uint8_t byte = data[i];
+    value |= static_cast<Int>(byte & kLow7Mask) << (7 * i);
+
+    // Check for lack of continuation flag in MSB
+    if ((byte & kContinuationBit) == 0) {
+      *out = value;
+      return i + 1;
+    }
+  }
+
+  // Process the last index avoiding overflowing
+  constexpr int32_t last = kMaxBytes - 1;
+
+  // We have not finished reading a valid LEB128, yet we run out of data
+  if (ARROW_PREDICT_FALSE(last >= max_data_size)) {
+    return 0;
+  }
+
+  const uint8_t byte = data[last];
+
+  // Need to check if there are bits that would overflow the output.
+  // Also checks that there is no continuation.
+  if (ARROW_PREDICT_FALSE((byte & kHighForbiddenMask) != 0)) {
+    return 0;
+  }
+
+  // No longer need to mask since we ensured
+  value |= static_cast<Int>(byte) << (7 * last);
+  *out = value;
+  return last + 1;
+}
+}  // namespace bit_util
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/bitmap.h

@@ -0,0 +1,466 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <algorithm>
+#include <array>
+#include <bitset>
+#include <cassert>
+#include <cstdint>
+#include <cstring>
+#include <memory>
+#include <string>
+#include <string_view>
+#include <utility>
+
+#include "arrow/buffer.h"
+#include "arrow/util/bit_util.h"
+#include "arrow/util/bitmap_ops.h"
+#include "arrow/util/bitmap_reader.h"
+#include "arrow/util/bitmap_writer.h"
+#include "arrow/util/compare.h"
+#include "arrow/util/endian.h"
+#include "arrow/util/functional.h"
+#include "arrow/util/span.h"
+#include "arrow/util/string_util.h"
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+
+class BooleanArray;
+
+namespace internal {
+
+class ARROW_EXPORT Bitmap : public util::ToStringOstreamable<Bitmap>,
+                            public util::EqualityComparable<Bitmap> {
+ public:
+  Bitmap() = default;
+
+  Bitmap(const std::shared_ptr<Buffer>& buffer, int64_t offset, int64_t length)
+      : data_(buffer->data()), offset_(offset), length_(length) {
+    if (buffer->is_mutable()) {
+      mutable_data_ = buffer->mutable_data();
+    }
+  }
+
+  Bitmap(const void* data, int64_t offset, int64_t length)
+      : data_(reinterpret_cast<const uint8_t*>(data)), offset_(offset), length_(length) {}
+
+  Bitmap(void* data, int64_t offset, int64_t length)
+      : data_(reinterpret_cast<const uint8_t*>(data)),
+        mutable_data_(reinterpret_cast<uint8_t*>(data)),
+        offset_(offset),
+        length_(length) {}
+
+  Bitmap Slice(int64_t offset) const {
+    if (mutable_data_ != NULLPTR) {
+      return {mutable_data_, offset_ + offset, length_ - offset};
+    } else {
+      return {data_, offset_ + offset, length_ - offset};
+    }
+  }
+
+  Bitmap Slice(int64_t offset, int64_t length) const {
+    if (mutable_data_ != NULLPTR) {
+      return {mutable_data_, offset_ + offset, length};
+    } else {
+      return {data_, offset_ + offset, length};
+    }
+  }
+
+  std::string ToString() const;
+
+  bool Equals(const Bitmap& other) const;
+
+  std::string Diff(const Bitmap& other) const;
+
+  bool GetBit(int64_t i) const { return bit_util::GetBit(data_, i + offset_); }
+
+  bool operator[](int64_t i) const { return GetBit(i); }
+
+  void SetBitTo(int64_t i, bool v) const {
+    bit_util::SetBitTo(mutable_data_, i + offset_, v);
+  }
+
+  void SetBitsTo(bool v) { bit_util::SetBitsTo(mutable_data_, offset_, length_, v); }
+
+  void CopyFrom(const Bitmap& other);
+  void CopyFromInverted(const Bitmap& other);
+
+  /// \brief Visit bits from each bitmap as bitset<N>
+  ///
+  /// All bitmaps must have identical length.
+  template <size_t N, typename Visitor>
+  static void VisitBits(const Bitmap (&bitmaps)[N], Visitor&& visitor) {
+    int64_t bit_length = BitLength(bitmaps, N);
+    std::bitset<N> bits;
+    for (int64_t bit_i = 0; bit_i < bit_length; ++bit_i) {
+      for (size_t i = 0; i < N; ++i) {
+        bits[i] = bitmaps[i].GetBit(bit_i);
+      }
+      visitor(bits);
+    }
+  }
+
+  /// \brief Visit bits from each bitmap as bitset<N>
+  ///
+  /// All bitmaps must have identical length.
+  template <size_t N, typename Visitor>
+  static void VisitBits(const std::array<Bitmap, N>& bitmaps, Visitor&& visitor) {
+    int64_t bit_length = BitLength(bitmaps);
+    std::bitset<N> bits;
+    for (int64_t bit_i = 0; bit_i < bit_length; ++bit_i) {
+      for (size_t i = 0; i < N; ++i) {
+        bits[i] = bitmaps[i].GetBit(bit_i);
+      }
+      visitor(bits);
+    }
+  }
+
+  /// \brief Visit words of bits from each bitmap as array<Word, N>
+  ///
+  /// All bitmaps must have identical length. The first bit in a visited bitmap
+  /// may be offset within the first visited word, but words will otherwise contain
+  /// densely packed bits loaded from the bitmap. That offset within the first word is
+  /// returned.
+  ///
+  /// TODO(bkietz) allow for early termination
+  // NOTE: this function is efficient on 3+ sufficiently large bitmaps.
+  // It also has a large prolog / epilog overhead and should be used
+  // carefully in other cases.
+  // For 2 bitmaps or less, and/or smaller bitmaps, see also VisitTwoBitBlocksVoid
+  // and BitmapUInt64Reader.
+  template <size_t N, typename Visitor,
+            typename Word = typename std::decay<
+                internal::call_traits::argument_type<0, Visitor&&>>::type::value_type>
+  static int64_t VisitWords(const Bitmap (&bitmaps_arg)[N], Visitor&& visitor) {
+    constexpr int64_t kBitWidth = sizeof(Word) * 8;
+
+    // local, mutable variables which will be sliced/decremented to represent consumption:
+    Bitmap bitmaps[N];
+    int64_t offsets[N];
+    int64_t bit_length = BitLength(bitmaps_arg, N);
+    util::span<const Word> words[N];
+    for (size_t i = 0; i < N; ++i) {
+      bitmaps[i] = bitmaps_arg[i];
+      offsets[i] = bitmaps[i].template word_offset<Word>();
+      assert(offsets[i] >= 0 && offsets[i] < kBitWidth);
+      words[i] = bitmaps[i].template words<Word>();
+    }
+
+    auto consume = [&](int64_t consumed_bits) {
+      for (size_t i = 0; i < N; ++i) {
+        bitmaps[i] = bitmaps[i].Slice(consumed_bits, bit_length - consumed_bits);
+        offsets[i] = bitmaps[i].template word_offset<Word>();
+        assert(offsets[i] >= 0 && offsets[i] < kBitWidth);
+        words[i] = bitmaps[i].template words<Word>();
+      }
+      bit_length -= consumed_bits;
+    };
+
+    std::array<Word, N> visited_words;
+    visited_words.fill(0);
+
+    if (bit_length <= kBitWidth * 2) {
+      // bitmaps fit into one or two words so don't bother with optimization
+      while (bit_length > 0) {
+        auto leading_bits = std::min(bit_length, kBitWidth);
+        SafeLoadWords(bitmaps, 0, leading_bits, false, &visited_words);
+        visitor(visited_words);
+        consume(leading_bits);
+      }
+      return 0;
+    }
+
+    int64_t max_offset = *std::max_element(offsets, offsets + N);
+    int64_t min_offset = *std::min_element(offsets, offsets + N);
+    if (max_offset > 0) {
+      // consume leading bits
+      auto leading_bits = kBitWidth - min_offset;
+      SafeLoadWords(bitmaps, 0, leading_bits, true, &visited_words);
+      visitor(visited_words);
+      consume(leading_bits);
+    }
+    assert(*std::min_element(offsets, offsets + N) == 0);
+
+    int64_t whole_word_count = bit_length / kBitWidth;
+    assert(whole_word_count >= 1);
+
+    if (min_offset == max_offset) {
+      // all offsets were identical, all leading bits have been consumed
+      assert(
+          std::all_of(offsets, offsets + N, [](int64_t offset) { return offset == 0; }));
+
+      for (int64_t word_i = 0; word_i < whole_word_count; ++word_i) {
+        for (size_t i = 0; i < N; ++i) {
+          visited_words[i] = words[i][word_i];
+        }
+        visitor(visited_words);
+      }
+      consume(whole_word_count * kBitWidth);
+    } else {
+      // leading bits from potentially incomplete words have been consumed
+
+      // word_i such that words[i][word_i] and words[i][word_i + 1] are lie entirely
+      // within the bitmap for all i
+      for (int64_t word_i = 0; word_i < whole_word_count - 1; ++word_i) {
+        for (size_t i = 0; i < N; ++i) {
+          if (offsets[i] == 0) {
+            visited_words[i] = words[i][word_i];
+          } else {
+            auto words0 = bit_util::ToLittleEndian(words[i][word_i]);
+            auto words1 = bit_util::ToLittleEndian(words[i][word_i + 1]);
+            visited_words[i] = bit_util::FromLittleEndian(
+                (words0 >> offsets[i]) | (words1 << (kBitWidth - offsets[i])));
+          }
+        }
+        visitor(visited_words);
+      }
+      consume((whole_word_count - 1) * kBitWidth);
+
+      SafeLoadWords(bitmaps, 0, kBitWidth, false, &visited_words);
+
+      visitor(visited_words);
+      consume(kBitWidth);
+    }
+
+    // load remaining bits
+    if (bit_length > 0) {
+      SafeLoadWords(bitmaps, 0, bit_length, false, &visited_words);
+      visitor(visited_words);
+    }
+
+    return min_offset;
+  }
+
+  template <size_t N, size_t M, typename ReaderT, typename WriterT, typename Visitor,
+            typename Word = typename std::decay<
+                internal::call_traits::argument_type<0, Visitor&&>>::type::value_type>
+  static void RunVisitWordsAndWriteLoop(int64_t bit_length,
+                                        std::array<ReaderT, N>& readers,
+                                        std::array<WriterT, M>& writers,
+                                        Visitor&& visitor) {
+    constexpr int64_t kBitWidth = sizeof(Word) * 8;
+
+    std::array<Word, N> visited_words;
+    std::array<Word, M> output_words;
+
+    // every reader will have same number of words, since they are same length'ed
+    // TODO($JIRA) this will be inefficient in some cases. When there are offsets beyond
+    //  Word boundary, every Word would have to be created from 2 adjoining Words
+    auto n_words = readers[0].words();
+    bit_length -= n_words * kBitWidth;
+    while (n_words--) {
+      // first collect all words to visited_words array
+      for (size_t i = 0; i < N; i++) {
+        visited_words[i] = readers[i].NextWord();
+      }
+      visitor(visited_words, &output_words);
+      for (size_t i = 0; i < M; i++) {
+        writers[i].PutNextWord(output_words[i]);
+      }
+    }
+
+    // every reader will have same number of trailing bytes, because of the above reason
+    // tailing portion could be more than one word! (ref: BitmapWordReader constructor)
+    // remaining full/ partial words to write
+
+    if (bit_length) {
+      // convert the word visitor lambda to a byte_visitor
+      auto byte_visitor = [&](const std::array<uint8_t, N>& in,
+                              std::array<uint8_t, M>* out) {
+        std::array<Word, N> in_words;
+        std::array<Word, M> out_words;
+        std::copy(in.begin(), in.end(), in_words.begin());
+        visitor(in_words, &out_words);
+        for (size_t i = 0; i < M; i++) {
+          out->at(i) = static_cast<uint8_t>(out_words[i]);
+        }
+      };
+
+      std::array<uint8_t, N> visited_bytes;
+      std::array<uint8_t, M> output_bytes;
+      int n_bytes = readers[0].trailing_bytes();
+      while (n_bytes--) {
+        visited_bytes.fill(0);
+        output_bytes.fill(0);
+        int valid_bits;
+        for (size_t i = 0; i < N; i++) {
+          visited_bytes[i] = readers[i].NextTrailingByte(valid_bits);
+        }
+        byte_visitor(visited_bytes, &output_bytes);
+        for (size_t i = 0; i < M; i++) {
+          writers[i].PutNextTrailingByte(output_bytes[i], valid_bits);
+        }
+      }
+    }
+  }
+
+  /// \brief Visit words of bits from each input bitmap as array<Word, N> and collects
+  /// outputs to an array<Word, M>, to be written into the output bitmaps accordingly.
+  ///
+  /// All bitmaps must have identical length. The first bit in a visited bitmap
+  /// may be offset within the first visited word, but words will otherwise contain
+  /// densely packed bits loaded from the bitmap. That offset within the first word is
+  /// returned.
+  /// Visitor is expected to have the following signature
+  ///     [](const std::array<Word, N>& in_words, std::array<Word, M>* out_words){...}
+  ///
+  // NOTE: this function is efficient on 3+ sufficiently large bitmaps.
+  // It also has a large prolog / epilog overhead and should be used
+  // carefully in other cases.
+  // For 2 bitmaps or less, and/or smaller bitmaps, see also VisitTwoBitBlocksVoid
+  // and BitmapUInt64Reader.
+  template <size_t N, size_t M, typename Visitor,
+            typename Word = typename std::decay<
+                internal::call_traits::argument_type<0, Visitor&&>>::type::value_type>
+  static void VisitWordsAndWrite(const std::array<Bitmap, N>& bitmaps_arg,
+                                 std::array<Bitmap, M>* out_bitmaps_arg,
+                                 Visitor&& visitor) {
+    int64_t bit_length = BitLength(bitmaps_arg);
+    assert(bit_length == BitLength(*out_bitmaps_arg));
+
+    // if both input and output bitmaps have no byte offset, then use special template
+    if (std::all_of(bitmaps_arg.begin(), bitmaps_arg.end(),
+                    [](const Bitmap& b) { return b.offset_ % 8 == 0; }) &&
+        std::all_of(out_bitmaps_arg->begin(), out_bitmaps_arg->end(),
+                    [](const Bitmap& b) { return b.offset_ % 8 == 0; })) {
+      std::array<BitmapWordReader<Word, /*may_have_byte_offset=*/false>, N> readers;
+      for (size_t i = 0; i < N; ++i) {
+        const Bitmap& in_bitmap = bitmaps_arg[i];
+        readers[i] = BitmapWordReader<Word, /*may_have_byte_offset=*/false>(
+            in_bitmap.data_, in_bitmap.offset_, in_bitmap.length_);
+      }
+
+      std::array<BitmapWordWriter<Word, /*may_have_byte_offset=*/false>, M> writers;
+      for (size_t i = 0; i < M; ++i) {
+        const Bitmap& out_bitmap = out_bitmaps_arg->at(i);
+        writers[i] = BitmapWordWriter<Word, /*may_have_byte_offset=*/false>(
+            out_bitmap.mutable_data_, out_bitmap.offset_, out_bitmap.length_);
+      }
+
+      RunVisitWordsAndWriteLoop(bit_length, readers, writers, visitor);
+    } else {
+      std::array<BitmapWordReader<Word>, N> readers;
+      for (size_t i = 0; i < N; ++i) {
+        const Bitmap& in_bitmap = bitmaps_arg[i];
+        readers[i] =
+            BitmapWordReader<Word>(in_bitmap.data_, in_bitmap.offset_, in_bitmap.length_);
+      }
+
+      std::array<BitmapWordWriter<Word>, M> writers;
+      for (size_t i = 0; i < M; ++i) {
+        const Bitmap& out_bitmap = out_bitmaps_arg->at(i);
+        writers[i] = BitmapWordWriter<Word>(out_bitmap.mutable_data_, out_bitmap.offset_,
+                                            out_bitmap.length_);
+      }
+
+      RunVisitWordsAndWriteLoop(bit_length, readers, writers, visitor);
+    }
+  }
+
+  const uint8_t* data() const { return data_; }
+  uint8_t* mutable_data() { return mutable_data_; }
+
+  /// offset of first bit relative to buffer().data()
+  int64_t offset() const { return offset_; }
+
+  /// number of bits in this Bitmap
+  int64_t length() const { return length_; }
+
+  /// span of all bytes which contain any bit in this Bitmap
+  util::span<const uint8_t> bytes() const {
+    auto byte_offset = offset_ / 8;
+    auto byte_count = bit_util::CeilDiv(offset_ + length_, 8) - byte_offset;
+    return {data_ + byte_offset, static_cast<size_t>(byte_count)};
+  }
+
+ private:
+  /// span of all Words which contain any bit in this Bitmap
+  ///
+  /// For example, given Word=uint16_t and a bitmap spanning bits [20, 36)
+  /// words() would span bits [16, 48).
+  ///
+  /// 0       16      32     48     64
+  /// |-------|-------|------|------| (buffer)
+  ///           [       ]             (bitmap)
+  ///         |-------|------|        (returned words)
+  ///
+  /// \warning The words may contain bytes which lie outside the buffer or are
+  /// uninitialized.
+  template <typename Word>
+  util::span<const Word> words() const {
+    auto bytes_addr = reinterpret_cast<intptr_t>(bytes().data());
+    auto words_addr = bytes_addr - bytes_addr % sizeof(Word);
+    auto word_byte_count =
+        bit_util::RoundUpToPowerOf2(static_cast<int64_t>(bytes_addr + bytes().size()),
+                                    static_cast<int64_t>(sizeof(Word))) -
+        words_addr;
+    return {reinterpret_cast<const Word*>(words_addr),
+            static_cast<size_t>(word_byte_count / sizeof(Word))};
+  }
+
+  /// offset of first bit relative to words<Word>().data()
+  template <typename Word>
+  int64_t word_offset() const {
+    return offset_ + 8 * (reinterpret_cast<intptr_t>(data_) -
+                          reinterpret_cast<intptr_t>(words<Word>().data()));
+  }
+
+  /// load words from bitmaps bitwise
+  template <size_t N, typename Word>
+  static void SafeLoadWords(const Bitmap (&bitmaps)[N], int64_t offset,
+                            int64_t out_length, bool set_trailing_bits,
+                            std::array<Word, N>* out) {
+    out->fill(0);
+
+    int64_t out_offset = set_trailing_bits ? sizeof(Word) * 8 - out_length : 0;
+
+    Bitmap slices[N], out_bitmaps[N];
+    for (size_t i = 0; i < N; ++i) {
+      slices[i] = bitmaps[i].Slice(offset, out_length);
+      out_bitmaps[i] = Bitmap(&out->at(i), out_offset, out_length);
+    }
+
+    int64_t bit_i = 0;
+    Bitmap::VisitBits(slices, [&](std::bitset<N> bits) {
+      for (size_t i = 0; i < N; ++i) {
+        out_bitmaps[i].SetBitTo(bit_i, bits[i]);
+      }
+      ++bit_i;
+    });
+  }
+
+  /// assert bitmaps have identical length and return that length
+  static int64_t BitLength(const Bitmap* bitmaps, size_t N);
+
+  template <size_t N>
+  static int64_t BitLength(const std::array<Bitmap, N>& bitmaps) {
+    for (size_t i = 1; i < N; ++i) {
+      assert(bitmaps[i].length() == bitmaps[0].length());
+    }
+    return bitmaps[0].length();
+  }
+
+  const uint8_t* data_ = NULLPTR;
+  uint8_t* mutable_data_ = NULLPTR;
+  int64_t offset_ = 0, length_ = 0;
+};
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/bitmap_builders.h

@@ -0,0 +1,44 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cstdint>
+#include <memory>
+#include <vector>
+
+#include "arrow/result.h"
+#include "arrow/type_fwd.h"
+#include "arrow/util/span.h"
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+namespace internal {
+
+/// \brief Generate Bitmap with all position to `value` except for one found
+/// at `straggler_pos`.
+ARROW_EXPORT
+Result<std::shared_ptr<Buffer>> BitmapAllButOne(MemoryPool* pool, int64_t length,
+                                                int64_t straggler_pos, bool value = true);
+
+/// \brief Convert vector of bytes to bitmap buffer
+ARROW_EXPORT
+Result<std::shared_ptr<Buffer>> BytesToBits(util::span<const uint8_t> bytes,
+                                            MemoryPool* pool = default_memory_pool());
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/bitmap_generate.h

@@ -0,0 +1,112 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cstdint>
+#include <memory>
+
+#include "arrow/buffer.h"
+#include "arrow/memory_pool.h"
+#include "arrow/result.h"
+#include "arrow/util/bit_util.h"
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+namespace internal {
+
+// A std::generate() like function to write sequential bits into a bitmap area.
+// Bits preceding the bitmap area are preserved, bits following the bitmap
+// area may be clobbered.
+
+template <class Generator>
+void GenerateBits(uint8_t* bitmap, int64_t start_offset, int64_t length, Generator&& g) {
+  if (length == 0) {
+    return;
+  }
+  uint8_t* cur = bitmap + start_offset / 8;
+  uint8_t bit_mask = bit_util::kBitmask[start_offset % 8];
+  uint8_t current_byte = *cur & bit_util::kPrecedingBitmask[start_offset % 8];
+
+  for (int64_t index = 0; index < length; ++index) {
+    const bool bit = g();
+    current_byte = bit ? (current_byte | bit_mask) : current_byte;
+    bit_mask = static_cast<uint8_t>(bit_mask << 1);
+    if (bit_mask == 0) {
+      bit_mask = 1;
+      *cur++ = current_byte;
+      current_byte = 0;
+    }
+  }
+  if (bit_mask != 1) {
+    *cur++ = current_byte;
+  }
+}
+
+// Like GenerateBits(), but unrolls its main loop for higher performance.
+
+template <class Generator>
+void GenerateBitsUnrolled(uint8_t* bitmap, int64_t start_offset, int64_t length,
+                          Generator&& g) {
+  static_assert(std::is_same<decltype(std::declval<Generator>()()), bool>::value,
+                "Functor passed to GenerateBitsUnrolled must return bool");
+
+  if (length == 0) {
+    return;
+  }
+  uint8_t current_byte;
+  uint8_t* cur = bitmap + start_offset / 8;
+  const uint64_t start_bit_offset = start_offset % 8;
+  uint8_t bit_mask = bit_util::kBitmask[start_bit_offset];
+  int64_t remaining = length;
+
+  if (bit_mask != 0x01) {
+    current_byte = *cur & bit_util::kPrecedingBitmask[start_bit_offset];
+    while (bit_mask != 0 && remaining > 0) {
+      current_byte |= g() * bit_mask;
+      bit_mask = static_cast<uint8_t>(bit_mask << 1);
+      --remaining;
+    }
+    *cur++ = current_byte;
+  }
+
+  int64_t remaining_bytes = remaining / 8;
+  uint8_t out_results[8];
+  while (remaining_bytes-- > 0) {
+    for (int i = 0; i < 8; ++i) {
+      out_results[i] = g();
+    }
+    *cur++ = static_cast<uint8_t>(out_results[0] | out_results[1] << 1 |
+                                  out_results[2] << 2 | out_results[3] << 3 |
+                                  out_results[4] << 4 | out_results[5] << 5 |
+                                  out_results[6] << 6 | out_results[7] << 7);
+  }
+
+  int64_t remaining_bits = remaining % 8;
+  if (remaining_bits) {
+    current_byte = 0;
+    bit_mask = 0x01;
+    while (remaining_bits-- > 0) {
+      current_byte |= g() * bit_mask;
+      bit_mask = static_cast<uint8_t>(bit_mask << 1);
+    }
+    *cur++ = current_byte;
+  }
+}
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/bitmap_ops.h

@@ -0,0 +1,246 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cstdint>
+#include <memory>
+
+#include "arrow/result.h"
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+
+class Buffer;
+class MemoryPool;
+
+namespace internal {
+
+// ----------------------------------------------------------------------
+// Bitmap utilities
+
+/// Copy a bit range of an existing bitmap
+///
+/// \param[in] pool memory pool to allocate memory from
+/// \param[in] bitmap source data
+/// \param[in] offset bit offset into the source data
+/// \param[in] length number of bits to copy
+/// \param[in] out_offset bit offset into the output buffer
+///
+/// \return Status message
+ARROW_EXPORT
+Result<std::shared_ptr<Buffer>> CopyBitmap(MemoryPool* pool, const uint8_t* bitmap,
+                                           int64_t offset, int64_t length,
+                                           int64_t out_offset = 0);
+
+/// Copy a bit range of an existing bitmap into an existing bitmap
+///
+/// \param[in] bitmap source data
+/// \param[in] offset bit offset into the source data
+/// \param[in] length number of bits to copy
+/// \param[in] dest_offset bit offset into the destination
+/// \param[out] dest the destination buffer, must have at least space for
+/// (offset + length) bits
+ARROW_EXPORT
+void CopyBitmap(const uint8_t* bitmap, int64_t offset, int64_t length, uint8_t* dest,
+                int64_t dest_offset);
+
+/// Invert a bit range of an existing bitmap into an existing bitmap
+///
+/// \param[in] bitmap source data
+/// \param[in] offset bit offset into the source data
+/// \param[in] length number of bits to copy
+/// \param[in] dest_offset bit offset into the destination
+/// \param[out] dest the destination buffer, must have at least space for
+/// (offset + length) bits
+ARROW_EXPORT
+void InvertBitmap(const uint8_t* bitmap, int64_t offset, int64_t length, uint8_t* dest,
+                  int64_t dest_offset);
+
+/// Invert a bit range of an existing bitmap
+///
+/// \param[in] pool memory pool to allocate memory from
+/// \param[in] bitmap source data
+/// \param[in] offset bit offset into the source data
+/// \param[in] length number of bits to copy
+///
+/// \return Status message
+ARROW_EXPORT
+Result<std::shared_ptr<Buffer>> InvertBitmap(MemoryPool* pool, const uint8_t* bitmap,
+                                             int64_t offset, int64_t length);
+
+/// Reverse a bit range of an existing bitmap into an existing bitmap
+///
+/// \param[in] bitmap source data
+/// \param[in] offset bit offset into the source data
+/// \param[in] length number of bits to reverse
+/// \param[in] dest_offset bit offset into the destination
+/// \param[out] dest the destination buffer, must have at least space for
+/// (offset + length) bits
+ARROW_EXPORT
+void ReverseBitmap(const uint8_t* bitmap, int64_t offset, int64_t length, uint8_t* dest,
+                   int64_t dest_offset);
+
+/// Reverse a bit range of an existing bitmap
+///
+/// \param[in] pool memory pool to allocate memory from
+/// \param[in] bitmap source data
+/// \param[in] offset bit offset into the source data
+/// \param[in] length number of bits to reverse
+///
+/// \return Status message
+ARROW_EXPORT
+Result<std::shared_ptr<Buffer>> ReverseBitmap(MemoryPool* pool, const uint8_t* bitmap,
+                                              int64_t offset, int64_t length);
+
+/// Compute the number of 1's in the given data array
+///
+/// \param[in] data a packed LSB-ordered bitmap as a byte array
+/// \param[in] bit_offset a bitwise offset into the bitmap
+/// \param[in] length the number of bits to inspect in the bitmap relative to
+/// the offset
+///
+/// \return The number of set (1) bits in the range
+ARROW_EXPORT
+int64_t CountSetBits(const uint8_t* data, int64_t bit_offset, int64_t length);
+
+/// Compute the number of 1's in the result of an "and" (&) of two bitmaps
+///
+/// \param[in] left_bitmap a packed LSB-ordered bitmap as a byte array
+/// \param[in] left_offset a bitwise offset into the left bitmap
+/// \param[in] right_bitmap a packed LSB-ordered bitmap as a byte array
+/// \param[in] right_offset a bitwise offset into the right bitmap
+/// \param[in] length the length of the bitmaps (must be the same)
+///
+/// \return The number of set (1) bits in the "and" of the two bitmaps
+ARROW_EXPORT
+int64_t CountAndSetBits(const uint8_t* left_bitmap, int64_t left_offset,
+                        const uint8_t* right_bitmap, int64_t right_offset,
+                        int64_t length);
+
+ARROW_EXPORT
+bool BitmapEquals(const uint8_t* left, int64_t left_offset, const uint8_t* right,
+                  int64_t right_offset, int64_t length);
+
+// Same as BitmapEquals, but considers a NULL bitmap pointer the same as an
+// all-ones bitmap.
+ARROW_EXPORT
+bool OptionalBitmapEquals(const uint8_t* left, int64_t left_offset, const uint8_t* right,
+                          int64_t right_offset, int64_t length);
+
+ARROW_EXPORT
+bool OptionalBitmapEquals(const std::shared_ptr<Buffer>& left, int64_t left_offset,
+                          const std::shared_ptr<Buffer>& right, int64_t right_offset,
+                          int64_t length);
+
+/// \brief Do a "bitmap and" on right and left buffers starting at
+/// their respective bit-offsets for the given bit-length and put
+/// the results in out_buffer starting at the given bit-offset.
+///
+/// out_buffer will be allocated and initialized to zeros using pool before
+/// the operation.
+ARROW_EXPORT
+Result<std::shared_ptr<Buffer>> BitmapAnd(MemoryPool* pool, const uint8_t* left,
+                                          int64_t left_offset, const uint8_t* right,
+                                          int64_t right_offset, int64_t length,
+                                          int64_t out_offset);
+
+/// \brief Do a "bitmap and" on right and left buffers starting at
+/// their respective bit-offsets for the given bit-length and put
+/// the results in out starting at the given bit-offset.
+ARROW_EXPORT
+void BitmapAnd(const uint8_t* left, int64_t left_offset, const uint8_t* right,
+               int64_t right_offset, int64_t length, int64_t out_offset, uint8_t* out);
+
+/// \brief Do a "bitmap or" for the given bit length on right and left buffers
+/// starting at their respective bit-offsets and put the results in out_buffer
+/// starting at the given bit-offset.
+///
+/// out_buffer will be allocated and initialized to zeros using pool before
+/// the operation.
+ARROW_EXPORT
+Result<std::shared_ptr<Buffer>> BitmapOr(MemoryPool* pool, const uint8_t* left,
+                                         int64_t left_offset, const uint8_t* right,
+                                         int64_t right_offset, int64_t length,
+                                         int64_t out_offset);
+
+/// \brief Do a "bitmap or" for the given bit length on right and left buffers
+/// starting at their respective bit-offsets and put the results in out
+/// starting at the given bit-offset.
+ARROW_EXPORT
+void BitmapOr(const uint8_t* left, int64_t left_offset, const uint8_t* right,
+              int64_t right_offset, int64_t length, int64_t out_offset, uint8_t* out);
+
+/// \brief Do a "bitmap xor" for the given bit-length on right and left
+/// buffers starting at their respective bit-offsets and put the results in
+/// out_buffer starting at the given bit offset.
+///
+/// out_buffer will be allocated and initialized to zeros using pool before
+/// the operation.
+ARROW_EXPORT
+Result<std::shared_ptr<Buffer>> BitmapXor(MemoryPool* pool, const uint8_t* left,
+                                          int64_t left_offset, const uint8_t* right,
+                                          int64_t right_offset, int64_t length,
+                                          int64_t out_offset);
+
+/// \brief Do a "bitmap xor" for the given bit-length on right and left
+/// buffers starting at their respective bit-offsets and put the results in
+/// out starting at the given bit offset.
+ARROW_EXPORT
+void BitmapXor(const uint8_t* left, int64_t left_offset, const uint8_t* right,
+               int64_t right_offset, int64_t length, int64_t out_offset, uint8_t* out);
+
+/// \brief Do a "bitmap and not" on right and left buffers starting at
+/// their respective bit-offsets for the given bit-length and put
+/// the results in out_buffer starting at the given bit-offset.
+///
+/// out_buffer will be allocated and initialized to zeros using pool before
+/// the operation.
+ARROW_EXPORT
+Result<std::shared_ptr<Buffer>> BitmapAndNot(MemoryPool* pool, const uint8_t* left,
+                                             int64_t left_offset, const uint8_t* right,
+                                             int64_t right_offset, int64_t length,
+                                             int64_t out_offset);
+
+/// \brief Do a "bitmap and not" on right and left buffers starting at
+/// their respective bit-offsets for the given bit-length and put
+/// the results in out starting at the given bit-offset.
+ARROW_EXPORT
+void BitmapAndNot(const uint8_t* left, int64_t left_offset, const uint8_t* right,
+                  int64_t right_offset, int64_t length, int64_t out_offset, uint8_t* out);
+
+/// \brief Do a "bitmap or not" on right and left buffers starting at
+/// their respective bit-offsets for the given bit-length and put
+/// the results in out_buffer starting at the given bit-offset.
+///
+/// out_buffer will be allocated and initialized to zeros using pool before
+/// the operation.
+ARROW_EXPORT
+Result<std::shared_ptr<Buffer>> BitmapOrNot(MemoryPool* pool, const uint8_t* left,
+                                            int64_t left_offset, const uint8_t* right,
+                                            int64_t right_offset, int64_t length,
+                                            int64_t out_offset);
+
+/// \brief Do a "bitmap or not" on right and left buffers starting at
+/// their respective bit-offsets for the given bit-length and put
+/// the results in out starting at the given bit-offset.
+ARROW_EXPORT
+void BitmapOrNot(const uint8_t* left, int64_t left_offset, const uint8_t* right,
+                 int64_t right_offset, int64_t length, int64_t out_offset, uint8_t* out);
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/bitmap_reader.h

@@ -0,0 +1,275 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cassert>
+#include <cstdint>
+#include <cstring>
+
+#include "arrow/buffer.h"
+#include "arrow/util/bit_util.h"
+#include "arrow/util/endian.h"
+#include "arrow/util/macros.h"
+
+namespace arrow {
+namespace internal {
+
+class BitmapReader {
+ public:
+  BitmapReader() = default;
+
+  BitmapReader(const uint8_t* bitmap, int64_t start_offset, int64_t length)
+      : bitmap_(bitmap), position_(0), length_(length) {
+    current_byte_ = 0;
+    byte_offset_ = start_offset / 8;
+    bit_offset_ = start_offset % 8;
+    if (length > 0) {
+      current_byte_ = bitmap[byte_offset_];
+    }
+  }
+
+  bool IsSet() const { return (current_byte_ & (1 << bit_offset_)) != 0; }
+
+  bool IsNotSet() const { return (current_byte_ & (1 << bit_offset_)) == 0; }
+
+  void Next() {
+    ++bit_offset_;
+    ++position_;
+    if (ARROW_PREDICT_FALSE(bit_offset_ == 8)) {
+      bit_offset_ = 0;
+      ++byte_offset_;
+      if (ARROW_PREDICT_TRUE(position_ < length_)) {
+        current_byte_ = bitmap_[byte_offset_];
+      }
+    }
+  }
+
+  int64_t position() const { return position_; }
+
+  int64_t length() const { return length_; }
+
+ private:
+  const uint8_t* bitmap_;
+  int64_t position_;
+  int64_t length_;
+
+  uint8_t current_byte_;
+  int64_t byte_offset_;
+  int64_t bit_offset_;
+};
+
+// XXX Cannot name it BitmapWordReader because the name is already used
+// in bitmap_ops.cc
+
+class BitmapUInt64Reader {
+ public:
+  BitmapUInt64Reader(const uint8_t* bitmap, int64_t start_offset, int64_t length)
+      : bitmap_(util::MakeNonNull(bitmap) + start_offset / 8),
+        num_carry_bits_(8 - start_offset % 8),
+        length_(length),
+        remaining_length_(length_),
+        carry_bits_(0) {
+    if (length_ > 0) {
+      // Load carry bits from the first byte's MSBs
+      if (length_ >= num_carry_bits_) {
+        carry_bits_ =
+            LoadPartialWord(static_cast<int8_t>(8 - num_carry_bits_), num_carry_bits_);
+      } else {
+        carry_bits_ = LoadPartialWord(static_cast<int8_t>(8 - num_carry_bits_), length_);
+      }
+    }
+  }
+
+  uint64_t NextWord() {
+    if (ARROW_PREDICT_TRUE(remaining_length_ >= 64 + num_carry_bits_)) {
+      // We can load a full word
+      uint64_t next_word = LoadFullWord();
+      // Carry bits come first, then the (64 - num_carry_bits_) LSBs from next_word
+      uint64_t word = carry_bits_ | (next_word << num_carry_bits_);
+      carry_bits_ = next_word >> (64 - num_carry_bits_);
+      remaining_length_ -= 64;
+      return word;
+    } else if (remaining_length_ > num_carry_bits_) {
+      // We can load a partial word
+      uint64_t next_word =
+          LoadPartialWord(/*bit_offset=*/0, remaining_length_ - num_carry_bits_);
+      uint64_t word = carry_bits_ | (next_word << num_carry_bits_);
+      carry_bits_ = next_word >> (64 - num_carry_bits_);
+      remaining_length_ = std::max<int64_t>(remaining_length_ - 64, 0);
+      return word;
+    } else {
+      remaining_length_ = 0;
+      return carry_bits_;
+    }
+  }
+
+  int64_t position() const { return length_ - remaining_length_; }
+
+  int64_t length() const { return length_; }
+
+ private:
+  uint64_t LoadFullWord() {
+    uint64_t word;
+    memcpy(&word, bitmap_, 8);
+    bitmap_ += 8;
+    return bit_util::ToLittleEndian(word);
+  }
+
+  uint64_t LoadPartialWord(int8_t bit_offset, int64_t num_bits) {
+    uint64_t word = 0;
+    const int64_t num_bytes = bit_util::BytesForBits(num_bits);
+    memcpy(&word, bitmap_, num_bytes);
+    bitmap_ += num_bytes;
+    return (bit_util::ToLittleEndian(word) >> bit_offset) &
+           bit_util::LeastSignificantBitMask(num_bits);
+  }
+
+  const uint8_t* bitmap_;
+  const int64_t num_carry_bits_;  // in [1, 8]
+  const int64_t length_;
+  int64_t remaining_length_;
+  uint64_t carry_bits_;
+};
+
+// BitmapWordReader here is faster than BitmapUInt64Reader (in bitmap_reader.h)
+// on sufficiently large inputs.  However, it has a larger prolog / epilog overhead
+// and should probably not be used for small bitmaps.
+
+template <typename Word, bool may_have_byte_offset = true>
+class BitmapWordReader {
+ public:
+  BitmapWordReader() = default;
+  BitmapWordReader(const uint8_t* bitmap, int64_t offset, int64_t length)
+      : offset_(static_cast<int64_t>(may_have_byte_offset) * (offset % 8)),
+        bitmap_(bitmap + offset / 8),
+        bitmap_end_(bitmap_ + bit_util::BytesForBits(offset_ + length)) {
+    // decrement word count by one as we may touch two adjacent words in one iteration
+    nwords_ = length / (sizeof(Word) * 8) - 1;
+    if (nwords_ < 0) {
+      nwords_ = 0;
+    }
+    trailing_bits_ = static_cast<int>(length - nwords_ * sizeof(Word) * 8);
+    trailing_bytes_ = static_cast<int>(bit_util::BytesForBits(trailing_bits_));
+
+    if (nwords_ > 0) {
+      current_data.word_ = load<Word>(bitmap_);
+    } else if (length > 0) {
+      current_data.epi.byte_ = load<uint8_t>(bitmap_);
+    }
+  }
+
+  Word NextWord() {
+    bitmap_ += sizeof(Word);
+    const Word next_word = load<Word>(bitmap_);
+    Word word = current_data.word_;
+    if (may_have_byte_offset && offset_) {
+      // combine two adjacent words into one word
+      // |<------ next ----->|<---- current ---->|
+      // +-------------+-----+-------------+-----+
+      // |     ---     |  A  |      B      | --- |
+      // +-------------+-----+-------------+-----+
+      //                  |         |       offset
+      //                  v         v
+      //               +-----+-------------+
+      //               |  A  |      B      |
+      //               +-----+-------------+
+      //               |<------ word ----->|
+      word >>= offset_;
+      word |= next_word << (sizeof(Word) * 8 - offset_);
+    }
+    current_data.word_ = next_word;
+    return word;
+  }
+
+  uint8_t NextTrailingByte(int& valid_bits) {
+    uint8_t byte;
+    assert(trailing_bits_ > 0);
+
+    if (trailing_bits_ <= 8) {
+      // last byte
+      valid_bits = trailing_bits_;
+      trailing_bits_ = 0;
+      byte = 0;
+      internal::BitmapReader reader(bitmap_, offset_, valid_bits);
+      for (int i = 0; i < valid_bits; ++i) {
+        byte >>= 1;
+        if (reader.IsSet()) {
+          byte |= 0x80;
+        }
+        reader.Next();
+      }
+      byte >>= (8 - valid_bits);
+    } else {
+      ++bitmap_;
+      const uint8_t next_byte = load<uint8_t>(bitmap_);
+      byte = current_data.epi.byte_;
+      if (may_have_byte_offset && offset_) {
+        byte >>= offset_;
+        byte |= next_byte << (8 - offset_);
+      }
+      current_data.epi.byte_ = next_byte;
+      trailing_bits_ -= 8;
+      trailing_bytes_--;
+      valid_bits = 8;
+    }
+    return byte;
+  }
+
+  int64_t words() const { return nwords_; }
+  int trailing_bytes() const { return trailing_bytes_; }
+
+ private:
+  int64_t offset_;
+  const uint8_t* bitmap_;
+
+  const uint8_t* bitmap_end_;
+  int64_t nwords_;
+  int trailing_bits_;
+  int trailing_bytes_;
+  union {
+    Word word_;
+    struct {
+#if ARROW_LITTLE_ENDIAN == 0
+      uint8_t padding_bytes_[sizeof(Word) - 1];
+#endif
+      uint8_t byte_;
+    } epi;
+  } current_data;
+
+  template <typename DType>
+  DType load(const uint8_t* bitmap) {
+    assert(bitmap + sizeof(DType) <= bitmap_end_);
+    return bit_util::ToLittleEndian(util::SafeLoadAs<DType>(bitmap));
+  }
+};
+
+/// \brief Index into a possibly nonexistent bitmap
+struct OptionalBitIndexer {
+  const uint8_t* bitmap;
+  const int64_t offset;
+
+  explicit OptionalBitIndexer(const uint8_t* buffer = NULLPTR, int64_t offset = 0)
+      : bitmap(buffer), offset(offset) {}
+
+  bool operator[](int64_t i) const {
+    return bitmap == NULLPTR || bit_util::GetBit(bitmap, offset + i);
+  }
+};
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/bitmap_visit.h

@@ -0,0 +1,88 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cstdint>
+
+#include "arrow/util/bit_util.h"
+#include "arrow/util/bitmap_reader.h"
+
+namespace arrow {
+namespace internal {
+
+// A function that visits each bit in a bitmap and calls a visitor function with a
+// boolean representation of that bit. This is intended to be analogous to
+// GenerateBits.
+template <class Visitor>
+void VisitBits(const uint8_t* bitmap, int64_t start_offset, int64_t length,
+               Visitor&& visit) {
+  BitmapReader reader(bitmap, start_offset, length);
+  for (int64_t index = 0; index < length; ++index) {
+    visit(reader.IsSet());
+    reader.Next();
+  }
+}
+
+// Like VisitBits(), but unrolls its main loop for better performance.
+template <class Visitor>
+void VisitBitsUnrolled(const uint8_t* bitmap, int64_t start_offset, int64_t length,
+                       Visitor&& visit) {
+  if (length == 0) {
+    return;
+  }
+
+  // Start by visiting any bits preceding the first full byte.
+  int64_t num_bits_before_full_bytes =
+      bit_util::RoundUpToMultipleOf8(start_offset) - start_offset;
+  // Truncate num_bits_before_full_bytes if it is greater than length.
+  if (num_bits_before_full_bytes > length) {
+    num_bits_before_full_bytes = length;
+  }
+  // Use the non loop-unrolled VisitBits since we don't want to add branches
+  VisitBits<Visitor>(bitmap, start_offset, num_bits_before_full_bytes, visit);
+
+  // Shift the start pointer to the first full byte and compute the
+  // number of full bytes to be read.
+  const uint8_t* first_full_byte = bitmap + bit_util::CeilDiv(start_offset, 8);
+  const int64_t num_full_bytes = (length - num_bits_before_full_bytes) / 8;
+
+  // Iterate over each full byte of the input bitmap and call the visitor in
+  // a loop-unrolled manner.
+  for (int64_t byte_index = 0; byte_index < num_full_bytes; ++byte_index) {
+    // Get the current bit-packed byte value from the bitmap.
+    const uint8_t byte = *(first_full_byte + byte_index);
+
+    // Execute the visitor function on each bit of the current byte.
+    visit(bit_util::GetBitFromByte(byte, 0));
+    visit(bit_util::GetBitFromByte(byte, 1));
+    visit(bit_util::GetBitFromByte(byte, 2));
+    visit(bit_util::GetBitFromByte(byte, 3));
+    visit(bit_util::GetBitFromByte(byte, 4));
+    visit(bit_util::GetBitFromByte(byte, 5));
+    visit(bit_util::GetBitFromByte(byte, 6));
+    visit(bit_util::GetBitFromByte(byte, 7));
+  }
+
+  // Write any leftover bits in the last byte.
+  const int64_t num_bits_after_full_bytes = (length - num_bits_before_full_bytes) % 8;
+  VisitBits<Visitor>(first_full_byte + num_full_bytes, 0, num_bits_after_full_bytes,
+                     visit);
+}
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/bitmap_writer.h

@@ -0,0 +1,286 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cstdint>
+#include <cstring>
+
+#include "arrow/util/bit_util.h"
+#include "arrow/util/endian.h"
+#include "arrow/util/macros.h"
+
+namespace arrow {
+namespace internal {
+
+class BitmapWriter {
+  // A sequential bitwise writer that preserves surrounding bit values.
+
+ public:
+  BitmapWriter(uint8_t* bitmap, int64_t start_offset, int64_t length)
+      : bitmap_(bitmap), position_(0), length_(length) {
+    byte_offset_ = start_offset / 8;
+    bit_mask_ = bit_util::kBitmask[start_offset % 8];
+    if (length > 0) {
+      current_byte_ = bitmap[byte_offset_];
+    } else {
+      current_byte_ = 0;
+    }
+  }
+
+  void Set() { current_byte_ |= bit_mask_; }
+
+  void Clear() { current_byte_ &= bit_mask_ ^ 0xFF; }
+
+  void Next() {
+    bit_mask_ = static_cast<uint8_t>(bit_mask_ << 1);
+    ++position_;
+    if (bit_mask_ == 0) {
+      // Finished this byte, need advancing
+      bit_mask_ = 0x01;
+      bitmap_[byte_offset_++] = current_byte_;
+      if (ARROW_PREDICT_TRUE(position_ < length_)) {
+        current_byte_ = bitmap_[byte_offset_];
+      }
+    }
+  }
+
+  void Finish() {
+    // Store current byte if we didn't went past bitmap storage
+    if (length_ > 0 && (bit_mask_ != 0x01 || position_ < length_)) {
+      bitmap_[byte_offset_] = current_byte_;
+    }
+  }
+
+  int64_t position() const { return position_; }
+
+ private:
+  uint8_t* bitmap_;
+  int64_t position_;
+  int64_t length_;
+
+  uint8_t current_byte_;
+  uint8_t bit_mask_;
+  int64_t byte_offset_;
+};
+
+class FirstTimeBitmapWriter {
+  // Like BitmapWriter, but any bit values *following* the bits written
+  // might be clobbered.  It is hence faster than BitmapWriter, and can
+  // also avoid false positives with Valgrind.
+
+ public:
+  FirstTimeBitmapWriter(uint8_t* bitmap, int64_t start_offset, int64_t length)
+      : bitmap_(bitmap), position_(0), length_(length) {
+    current_byte_ = 0;
+    byte_offset_ = start_offset / 8;
+    bit_mask_ = bit_util::kBitmask[start_offset % 8];
+    if (length > 0) {
+      current_byte_ =
+          bitmap[byte_offset_] & bit_util::kPrecedingBitmask[start_offset % 8];
+    } else {
+      current_byte_ = 0;
+    }
+  }
+
+  /// Appends number_of_bits from word to valid_bits and valid_bits_offset.
+  ///
+  /// \param[in] word The LSB bitmap to append. Any bits past number_of_bits are assumed
+  ///            to be unset (i.e. 0).
+  /// \param[in] number_of_bits The number of bits to append from word.
+  void AppendWord(uint64_t word, int64_t number_of_bits) {
+    if (ARROW_PREDICT_FALSE(number_of_bits == 0)) {
+      return;
+    }
+
+    // Location that the first byte needs to be written to.
+    uint8_t* append_position = bitmap_ + byte_offset_;
+
+    // Update state variables except for current_byte_ here.
+    position_ += number_of_bits;
+    int64_t bit_offset = bit_util::CountTrailingZeros(static_cast<uint32_t>(bit_mask_));
+    bit_mask_ = bit_util::kBitmask[(bit_offset + number_of_bits) % 8];
+    byte_offset_ += (bit_offset + number_of_bits) / 8;
+
+    if (bit_offset != 0) {
+      // We are in the middle of the byte. This code updates the byte and shifts
+      // bits appropriately within word so it can be memcpy'd below.
+      int64_t bits_to_carry = 8 - bit_offset;
+      // Carry over bits from word to current_byte_. We assume any extra bits in word
+      // unset so no additional accounting is needed for when number_of_bits <
+      // bits_to_carry.
+      current_byte_ |= (word & bit_util::kPrecedingBitmask[bits_to_carry]) << bit_offset;
+      // Check if everything is transferred into current_byte_.
+      if (ARROW_PREDICT_FALSE(number_of_bits < bits_to_carry)) {
+        return;
+      }
+      *append_position = current_byte_;
+      append_position++;
+      // Move the carry bits off of word.
+      word = word >> bits_to_carry;
+      number_of_bits -= bits_to_carry;
+    }
+    word = bit_util::ToLittleEndian(word);
+    int64_t bytes_for_word = ::arrow::bit_util::BytesForBits(number_of_bits);
+    std::memcpy(append_position, &word, bytes_for_word);
+    // At this point, the previous current_byte_ has been written to bitmap_.
+    // The new current_byte_ is either the last relevant byte in 'word'
+    // or cleared if the new position is byte aligned (i.e. a fresh byte).
+    if (bit_mask_ == 0x1) {
+      current_byte_ = 0;
+    } else {
+      current_byte_ = *(append_position + bytes_for_word - 1);
+    }
+  }
+
+  void Set() { current_byte_ |= bit_mask_; }
+
+  void Clear() {}
+
+  void Next() {
+    bit_mask_ = static_cast<uint8_t>(bit_mask_ << 1);
+    ++position_;
+    if (bit_mask_ == 0) {
+      // Finished this byte, need advancing
+      bit_mask_ = 0x01;
+      bitmap_[byte_offset_++] = current_byte_;
+      current_byte_ = 0;
+    }
+  }
+
+  void Finish() {
+    // Store current byte if we didn't went go bitmap storage
+    if (length_ > 0 && (bit_mask_ != 0x01 || position_ < length_)) {
+      bitmap_[byte_offset_] = current_byte_;
+    }
+  }
+
+  int64_t position() const { return position_; }
+
+ private:
+  uint8_t* bitmap_;
+  int64_t position_;
+  int64_t length_;
+
+  uint8_t current_byte_;
+  uint8_t bit_mask_;
+  int64_t byte_offset_;
+};
+
+template <typename Word, bool may_have_byte_offset = true>
+class BitmapWordWriter {
+ public:
+  BitmapWordWriter() = default;
+  BitmapWordWriter(uint8_t* bitmap, int64_t offset, int64_t length)
+      : offset_(static_cast<int64_t>(may_have_byte_offset) * (offset % 8)),
+        bitmap_(bitmap + offset / 8),
+        bitmap_end_(bitmap_ + bit_util::BytesForBits(offset_ + length)),
+        mask_((1U << offset_) - 1) {
+    if (offset_) {
+      if (length >= static_cast<int>(sizeof(Word) * 8)) {
+        current_data.word_ = load<Word>(bitmap_);
+      } else if (length > 0) {
+        current_data.epi.byte_ = load<uint8_t>(bitmap_);
+      }
+    }
+  }
+
+  void PutNextWord(Word word) {
+    if (may_have_byte_offset && offset_) {
+      // split one word into two adjacent words, don't touch unused bits
+      //               |<------ word ----->|
+      //               +-----+-------------+
+      //               |  A  |      B      |
+      //               +-----+-------------+
+      //                  |         |
+      //                  v         v       offset
+      // +-------------+-----+-------------+-----+
+      // |     ---     |  A  |      B      | --- |
+      // +-------------+-----+-------------+-----+
+      // |<------ next ----->|<---- current ---->|
+      word = (word << offset_) | (word >> (sizeof(Word) * 8 - offset_));
+      Word next_word = load<Word>(bitmap_ + sizeof(Word));
+      current_data.word_ = (current_data.word_ & mask_) | (word & ~mask_);
+      next_word = (next_word & ~mask_) | (word & mask_);
+      store<Word>(bitmap_, current_data.word_);
+      store<Word>(bitmap_ + sizeof(Word), next_word);
+      current_data.word_ = next_word;
+    } else {
+      store<Word>(bitmap_, word);
+    }
+    bitmap_ += sizeof(Word);
+  }
+
+  void PutNextTrailingByte(uint8_t byte, int valid_bits) {
+    if (valid_bits == 8) {
+      if (may_have_byte_offset && offset_) {
+        byte = (byte << offset_) | (byte >> (8 - offset_));
+        uint8_t next_byte = load<uint8_t>(bitmap_ + 1);
+        current_data.epi.byte_ = (current_data.epi.byte_ & mask_) | (byte & ~mask_);
+        next_byte = (next_byte & ~mask_) | (byte & mask_);
+        store<uint8_t>(bitmap_, current_data.epi.byte_);
+        store<uint8_t>(bitmap_ + 1, next_byte);
+        current_data.epi.byte_ = next_byte;
+      } else {
+        store<uint8_t>(bitmap_, byte);
+      }
+      ++bitmap_;
+    } else {
+      assert(valid_bits > 0);
+      assert(valid_bits < 8);
+      assert(bitmap_ + bit_util::BytesForBits(offset_ + valid_bits) <= bitmap_end_);
+      internal::BitmapWriter writer(bitmap_, offset_, valid_bits);
+      for (int i = 0; i < valid_bits; ++i) {
+        (byte & 0x01) ? writer.Set() : writer.Clear();
+        writer.Next();
+        byte >>= 1;
+      }
+      writer.Finish();
+    }
+  }
+
+ private:
+  int64_t offset_;
+  uint8_t* bitmap_;
+
+  const uint8_t* bitmap_end_;
+  uint64_t mask_;
+  union {
+    Word word_;
+    struct {
+#if ARROW_LITTLE_ENDIAN == 0
+      uint8_t padding_bytes_[sizeof(Word) - 1];
+#endif
+      uint8_t byte_;
+    } epi;
+  } current_data;
+
+  template <typename DType>
+  DType load(const uint8_t* bitmap) {
+    assert(bitmap + sizeof(DType) <= bitmap_end_);
+    return bit_util::ToLittleEndian(util::SafeLoadAs<DType>(bitmap));
+  }
+
+  template <typename DType>
+  void store(uint8_t* bitmap, DType data) {
+    assert(bitmap + sizeof(DType) <= bitmap_end_);
+    util::SafeStore(bitmap, bit_util::FromLittleEndian(data));
+  }
+};
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/byte_size.h

@@ -0,0 +1,88 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cstdint>
+
+#include "arrow/type_fwd.h"
+
+namespace arrow {
+
+namespace util {
+
+/// \brief The sum of bytes in each buffer referenced by the array
+///
+/// Note: An array may only reference a portion of a buffer.
+///       This method will overestimate in this case and return the
+///       byte size of the entire buffer.
+/// Note: If a buffer is referenced multiple times then it will
+///       only be counted once.
+ARROW_EXPORT int64_t TotalBufferSize(const ArrayData& array_data);
+/// \brief The sum of bytes in each buffer referenced by the array
+/// \see TotalBufferSize(const ArrayData& array_data) for details
+ARROW_EXPORT int64_t TotalBufferSize(const Array& array);
+/// \brief The sum of bytes in each buffer referenced by the array
+/// \see TotalBufferSize(const ArrayData& array_data) for details
+ARROW_EXPORT int64_t TotalBufferSize(const ChunkedArray& chunked_array);
+/// \brief The sum of bytes in each buffer referenced by the batch
+/// \see TotalBufferSize(const ArrayData& array_data) for details
+ARROW_EXPORT int64_t TotalBufferSize(const RecordBatch& record_batch);
+/// \brief The sum of bytes in each buffer referenced by the table
+/// \see TotalBufferSize(const ArrayData& array_data) for details
+ARROW_EXPORT int64_t TotalBufferSize(const Table& table);
+
+/// \brief Calculate the buffer ranges referenced by the array
+///
+/// These ranges will take into account array offsets
+///
+/// The ranges may contain duplicates
+///
+/// Dictionary arrays will ignore the offset of their containing array
+///
+/// The return value will be a struct array corresponding to the schema:
+/// schema({field("start", uint64()), field("offset", uint64()), field("length",
+/// uint64()))
+ARROW_EXPORT Result<std::shared_ptr<Array>> ReferencedRanges(const ArrayData& array_data);
+
+/// \brief Returns the sum of bytes from all buffer ranges referenced
+///
+/// Unlike TotalBufferSize this method will account for array
+/// offsets.
+///
+/// If buffers are shared between arrays then the shared
+/// portion will be counted multiple times.
+///
+/// Dictionary arrays will always be counted in their entirety
+/// even if the array only references a portion of the dictionary.
+ARROW_EXPORT Result<int64_t> ReferencedBufferSize(const ArrayData& array_data);
+/// \brief Returns the sum of bytes from all buffer ranges referenced
+/// \see ReferencedBufferSize(const ArrayData& array_data) for details
+ARROW_EXPORT Result<int64_t> ReferencedBufferSize(const Array& array_data);
+/// \brief Returns the sum of bytes from all buffer ranges referenced
+/// \see ReferencedBufferSize(const ArrayData& array_data) for details
+ARROW_EXPORT Result<int64_t> ReferencedBufferSize(const ChunkedArray& array_data);
+/// \brief Returns the sum of bytes from all buffer ranges referenced
+/// \see ReferencedBufferSize(const ArrayData& array_data) for details
+ARROW_EXPORT Result<int64_t> ReferencedBufferSize(const RecordBatch& array_data);
+/// \brief Returns the sum of bytes from all buffer ranges referenced
+/// \see ReferencedBufferSize(const ArrayData& array_data) for details
+ARROW_EXPORT Result<int64_t> ReferencedBufferSize(const Table& array_data);
+
+}  // namespace util
+
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/cancel.h

@@ -0,0 +1,118 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <functional>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "arrow/status.h"
+#include "arrow/type_fwd.h"
+#include "arrow/util/macros.h"
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+
+class StopToken;
+
+struct StopSourceImpl;
+
+/// EXPERIMENTAL
+class ARROW_EXPORT StopSource {
+ public:
+  StopSource();
+  ~StopSource();
+
+  // Consumer API (the side that stops)
+  void RequestStop();
+  void RequestStop(Status error);
+  // Async-signal-safe. TODO Deprecate this?
+  void RequestStopFromSignal(int signum);
+
+  StopToken token();
+
+  // For internal use only
+  void Reset();
+
+ protected:
+  std::shared_ptr<StopSourceImpl> impl_;
+};
+
+/// EXPERIMENTAL
+class ARROW_EXPORT StopToken {
+ public:
+  // Public for Cython
+  StopToken() = default;
+
+  explicit StopToken(std::shared_ptr<StopSourceImpl> impl) : impl_(std::move(impl)) {}
+
+  // A trivial token that never propagates any stop request
+  static StopToken Unstoppable() { return StopToken(); }
+
+  /// \brief Check if the stop source has been cancelled.
+  ///
+  /// Producers should call this method, whenever convenient, to check and
+  /// see if they should stop producing early (i.e. have been cancelled).
+  /// Failure to call this method often enough will lead to an unresponsive
+  /// cancellation.
+  ///
+  /// This is part of the producer API (the side that gets asked to stop)
+  /// This method is thread-safe
+  ///
+  /// \return An OK status if the stop source has not been cancelled or a
+  ///         cancel error if the source has been cancelled.
+  Status Poll() const;
+  bool IsStopRequested() const;
+
+ protected:
+  std::shared_ptr<StopSourceImpl> impl_;
+};
+
+/// EXPERIMENTAL: Set a global StopSource that can receive signals
+///
+/// The only allowed order of calls is the following:
+/// - SetSignalStopSource()
+/// - any number of pairs of (RegisterCancellingSignalHandler,
+///   UnregisterCancellingSignalHandler) calls
+/// - ResetSignalStopSource()
+///
+/// Beware that these settings are process-wide.  Typically, only one
+/// thread should call these APIs, even in a multithreaded setting.
+ARROW_EXPORT
+Result<StopSource*> SetSignalStopSource();
+
+/// EXPERIMENTAL: Reset the global signal-receiving StopSource
+///
+/// This will invalidate the pointer returned by SetSignalStopSource.
+ARROW_EXPORT
+void ResetSignalStopSource();
+
+/// EXPERIMENTAL: Register signal handler triggering the signal-receiving StopSource
+///
+/// Note that those handlers are automatically un-registered in a fork()ed process,
+/// therefore the child process will need to call RegisterCancellingSignalHandler()
+/// if desired.
+ARROW_EXPORT
+Status RegisterCancellingSignalHandler(const std::vector<int>& signals);
+
+/// EXPERIMENTAL: Unregister signal handler set up by RegisterCancellingSignalHandler
+ARROW_EXPORT
+void UnregisterCancellingSignalHandler();
+
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/checked_cast.h

@@ -0,0 +1,61 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <memory>
+#include <type_traits>
+#include <utility>
+
+namespace arrow {
+namespace internal {
+
+template <typename OutputType, typename InputType>
+inline OutputType checked_cast(InputType&& value) {
+  static_assert(std::is_class<typename std::remove_pointer<
+                    typename std::remove_reference<InputType>::type>::type>::value,
+                "checked_cast input type must be a class");
+  static_assert(std::is_class<typename std::remove_pointer<
+                    typename std::remove_reference<OutputType>::type>::type>::value,
+                "checked_cast output type must be a class");
+#ifdef NDEBUG
+  return static_cast<OutputType>(value);
+#else
+  return dynamic_cast<OutputType>(value);
+#endif
+}
+
+template <class T, class U>
+std::shared_ptr<T> checked_pointer_cast(std::shared_ptr<U> r) noexcept {
+#ifdef NDEBUG
+  return std::static_pointer_cast<T>(std::move(r));
+#else
+  return std::dynamic_pointer_cast<T>(std::move(r));
+#endif
+}
+
+template <class T, class U>
+std::unique_ptr<T> checked_pointer_cast(std::unique_ptr<U> r) noexcept {
+#ifdef NDEBUG
+  return std::unique_ptr<T>(static_cast<T*>(r.release()));
+#else
+  return std::unique_ptr<T>(dynamic_cast<T*>(r.release()));
+#endif
+}
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/compare.h

@@ -0,0 +1,62 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <memory>
+#include <type_traits>
+#include <utility>
+
+#include "arrow/util/macros.h"
+
+namespace arrow {
+namespace util {
+
+/// CRTP helper for declaring equality comparison. Defines operator== and operator!=
+template <typename T>
+class EqualityComparable {
+ public:
+  ~EqualityComparable() {
+    static_assert(
+        std::is_same<decltype(std::declval<const T>().Equals(std::declval<const T>())),
+                     bool>::value,
+        "EqualityComparable depends on the method T::Equals(const T&) const");
+  }
+
+  template <typename... Extra>
+  bool Equals(const std::shared_ptr<T>& other, Extra&&... extra) const {
+    if (other == NULLPTR) {
+      return false;
+    }
+    return cast().Equals(*other, std::forward<Extra>(extra)...);
+  }
+
+  struct PtrsEqual {
+    bool operator()(const std::shared_ptr<T>& l, const std::shared_ptr<T>& r) const {
+      return l->Equals(*r);
+    }
+  };
+
+  friend bool operator==(T const& a, T const& b) { return a.Equals(b); }
+  friend bool operator!=(T const& a, T const& b) { return !(a == b); }
+
+ private:
+  const T& cast() const { return static_cast<const T&>(*this); }
+};
+
+}  // namespace util
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/compression.h

@@ -0,0 +1,241 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cstdint>
+#include <limits>
+#include <memory>
+#include <optional>
+#include <string>
+
+#include "arrow/result.h"
+#include "arrow/status.h"
+#include "arrow/util/type_fwd.h"
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+namespace util {
+
+constexpr int kUseDefaultCompressionLevel = std::numeric_limits<int>::min();
+
+/// \brief Streaming compressor interface
+///
+class ARROW_EXPORT Compressor {
+ public:
+  virtual ~Compressor() = default;
+
+  struct CompressResult {
+    int64_t bytes_read;
+    int64_t bytes_written;
+  };
+  struct FlushResult {
+    int64_t bytes_written;
+    bool should_retry;
+  };
+  struct EndResult {
+    int64_t bytes_written;
+    bool should_retry;
+  };
+
+  /// \brief Compress some input.
+  ///
+  /// If bytes_read is 0 on return, then a larger output buffer should be supplied.
+  virtual Result<CompressResult> Compress(int64_t input_len, const uint8_t* input,
+                                          int64_t output_len, uint8_t* output) = 0;
+
+  /// \brief Flush part of the compressed output.
+  ///
+  /// If should_retry is true on return, Flush() should be called again
+  /// with a larger buffer.
+  virtual Result<FlushResult> Flush(int64_t output_len, uint8_t* output) = 0;
+
+  /// \brief End compressing, doing whatever is necessary to end the stream.
+  ///
+  /// If should_retry is true on return, End() should be called again
+  /// with a larger buffer.  Otherwise, the Compressor should not be used anymore.
+  ///
+  /// End() implies Flush().
+  virtual Result<EndResult> End(int64_t output_len, uint8_t* output) = 0;
+
+  // XXX add methods for buffer size heuristics?
+};
+
+/// \brief Streaming decompressor interface
+///
+class ARROW_EXPORT Decompressor {
+ public:
+  virtual ~Decompressor() = default;
+
+  struct DecompressResult {
+    // XXX is need_more_output necessary? (Brotli?)
+    int64_t bytes_read;
+    int64_t bytes_written;
+    bool need_more_output;
+  };
+
+  /// \brief Decompress some input.
+  ///
+  /// If need_more_output is true on return, a larger output buffer needs
+  /// to be supplied.
+  virtual Result<DecompressResult> Decompress(int64_t input_len, const uint8_t* input,
+                                              int64_t output_len, uint8_t* output) = 0;
+
+  /// \brief Return whether the compressed stream is finished.
+  ///
+  /// This is a heuristic.  If true is returned, then it is guaranteed
+  /// that the stream is finished.  If false is returned, however, it may
+  /// simply be that the underlying library isn't able to provide the information.
+  virtual bool IsFinished() = 0;
+
+  /// \brief Reinitialize decompressor, making it ready for a new compressed stream.
+  virtual Status Reset() = 0;
+
+  // XXX add methods for buffer size heuristics?
+};
+
+/// \brief Compression codec options
+class ARROW_EXPORT CodecOptions {
+ public:
+  explicit CodecOptions(int compression_level = kUseDefaultCompressionLevel)
+      : compression_level(compression_level) {}
+
+  virtual ~CodecOptions() = default;
+
+  int compression_level;
+};
+
+// ----------------------------------------------------------------------
+// GZip codec options implementation
+
+enum class GZipFormat {
+  ZLIB,
+  DEFLATE,
+  GZIP,
+};
+
+class ARROW_EXPORT GZipCodecOptions : public CodecOptions {
+ public:
+  GZipFormat gzip_format = GZipFormat::GZIP;
+  std::optional<int> window_bits;
+};
+
+// ----------------------------------------------------------------------
+// brotli codec options implementation
+
+class ARROW_EXPORT BrotliCodecOptions : public CodecOptions {
+ public:
+  std::optional<int> window_bits;
+};
+
+/// \brief Compression codec
+class ARROW_EXPORT Codec {
+ public:
+  virtual ~Codec() = default;
+
+  /// \brief Return special value to indicate that a codec implementation
+  /// should use its default compression level
+  static int UseDefaultCompressionLevel();
+
+  /// \brief Return a string name for compression type
+  static const std::string& GetCodecAsString(Compression::type t);
+
+  /// \brief Return compression type for name (all lower case)
+  static Result<Compression::type> GetCompressionType(const std::string& name);
+
+  /// \brief Create a codec for the given compression algorithm with CodecOptions
+  static Result<std::unique_ptr<Codec>> Create(
+      Compression::type codec, const CodecOptions& codec_options = CodecOptions{});
+
+  /// \brief Create a codec for the given compression algorithm
+  static Result<std::unique_ptr<Codec>> Create(Compression::type codec,
+                                               int compression_level);
+
+  /// \brief Return true if support for indicated codec has been enabled
+  static bool IsAvailable(Compression::type codec);
+
+  /// \brief Return true if indicated codec supports setting a compression level
+  static bool SupportsCompressionLevel(Compression::type codec);
+
+  /// \brief Return the smallest supported compression level for the codec
+  /// Note: This function creates a temporary Codec instance
+  static Result<int> MinimumCompressionLevel(Compression::type codec);
+
+  /// \brief Return the largest supported compression level for the codec
+  /// Note: This function creates a temporary Codec instance
+  static Result<int> MaximumCompressionLevel(Compression::type codec);
+
+  /// \brief Return the default compression level
+  /// Note: This function creates a temporary Codec instance
+  static Result<int> DefaultCompressionLevel(Compression::type codec);
+
+  /// \brief Return the smallest supported compression level
+  virtual int minimum_compression_level() const = 0;
+
+  /// \brief Return the largest supported compression level
+  virtual int maximum_compression_level() const = 0;
+
+  /// \brief Return the default compression level
+  virtual int default_compression_level() const = 0;
+
+  /// \brief One-shot decompression function
+  ///
+  /// output_buffer_len must be correct and therefore be obtained in advance.
+  /// The actual decompressed length is returned.
+  ///
+  /// \note One-shot decompression is not always compatible with streaming
+  /// compression.  Depending on the codec (e.g. LZ4), different formats may
+  /// be used.
+  virtual Result<int64_t> Decompress(int64_t input_len, const uint8_t* input,
+                                     int64_t output_buffer_len,
+                                     uint8_t* output_buffer) = 0;
+
+  /// \brief One-shot compression function
+  ///
+  /// output_buffer_len must first have been computed using MaxCompressedLen().
+  /// The actual compressed length is returned.
+  ///
+  /// \note One-shot compression is not always compatible with streaming
+  /// decompression.  Depending on the codec (e.g. LZ4), different formats may
+  /// be used.
+  virtual Result<int64_t> Compress(int64_t input_len, const uint8_t* input,
+                                   int64_t output_buffer_len, uint8_t* output_buffer) = 0;
+
+  virtual int64_t MaxCompressedLen(int64_t input_len, const uint8_t* input) = 0;
+
+  /// \brief Create a streaming compressor instance
+  virtual Result<std::shared_ptr<Compressor>> MakeCompressor() = 0;
+
+  /// \brief Create a streaming compressor instance
+  virtual Result<std::shared_ptr<Decompressor>> MakeDecompressor() = 0;
+
+  /// \brief This Codec's compression type
+  virtual Compression::type compression_type() const = 0;
+
+  /// \brief The name of this Codec's compression type
+  const std::string& name() const { return GetCodecAsString(compression_type()); }
+
+  /// \brief This Codec's compression level, if applicable
+  virtual int compression_level() const { return UseDefaultCompressionLevel(); }
+
+ private:
+  /// \brief Initializes the codec's resources.
+  virtual Status Init();
+};
+
+}  // namespace util
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/concurrent_map.h

@@ -0,0 +1,68 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <unordered_map>
+#include <utility>
+
+#include "arrow/util/mutex.h"
+
+namespace arrow {
+namespace util {
+
+template <typename K, typename V>
+class ConcurrentMap {
+ public:
+  void Insert(const K& key, const V& value) {
+    auto lock = mutex_.Lock();
+    map_.insert({key, value});
+  }
+
+  template <typename ValueFunc>
+  V GetOrInsert(const K& key, ValueFunc&& compute_value_func) {
+    auto lock = mutex_.Lock();
+    auto it = map_.find(key);
+    if (it == map_.end()) {
+      auto pair = map_.emplace(key, compute_value_func());
+      it = pair.first;
+    }
+    return it->second;
+  }
+
+  void Erase(const K& key) {
+    auto lock = mutex_.Lock();
+    map_.erase(key);
+  }
+
+  void Clear() {
+    auto lock = mutex_.Lock();
+    map_.clear();
+  }
+
+  size_t size() const {
+    auto lock = mutex_.Lock();
+    return map_.size();
+  }
+
+ private:
+  std::unordered_map<K, V> map_;
+  mutable arrow::util::Mutex mutex_;
+};
+
+}  // namespace util
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/config.h

@@ -0,0 +1,69 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#define ARROW_VERSION_MAJOR 22
+#define ARROW_VERSION_MINOR 0
+#define ARROW_VERSION_PATCH 0
+#define ARROW_VERSION ((ARROW_VERSION_MAJOR * 1000) + ARROW_VERSION_MINOR) * 1000 + ARROW_VERSION_PATCH
+
+#define ARROW_VERSION_STRING "22.0.0"
+
+#define ARROW_SO_VERSION "2200"
+#define ARROW_FULL_SO_VERSION "2200.0.0"
+
+#define ARROW_CXX_COMPILER_ID "GNU"
+#define ARROW_CXX_COMPILER_VERSION "14.2.1"
+#define ARROW_CXX_COMPILER_FLAGS " -Wno-noexcept-type -Wno-self-move -Wno-subobject-linkage  -fdiagnostics-color=always  -Wall -fno-semantic-interposition -msse4.2 "
+
+#define ARROW_BUILD_TYPE "RELEASE"
+
+#define ARROW_PACKAGE_KIND "python-wheel-manylinux228"
+
+#define ARROW_COMPUTE
+#define ARROW_CSV
+/* #undef ARROW_CUDA */
+#define ARROW_DATASET
+#define ARROW_FILESYSTEM
+#define ARROW_FLIGHT
+/* #undef ARROW_FLIGHT_SQL */
+#define ARROW_IPC
+#define ARROW_JEMALLOC
+#define ARROW_JEMALLOC_VENDORED
+#define ARROW_JSON
+#define ARROW_MIMALLOC
+#define ARROW_ORC
+#define ARROW_PARQUET
+#define ARROW_SUBSTRAIT
+
+#define ARROW_AZURE
+#define ARROW_ENABLE_THREADING
+#define ARROW_GCS
+#define ARROW_HDFS
+#define ARROW_S3
+/* #undef ARROW_USE_GLOG */
+#define ARROW_USE_NATIVE_INT128
+#define ARROW_WITH_BROTLI
+#define ARROW_WITH_BZ2
+#define ARROW_WITH_LZ4
+/* #undef ARROW_WITH_MUSL */
+/* #undef ARROW_WITH_OPENTELEMETRY */
+#define ARROW_WITH_RE2
+#define ARROW_WITH_SNAPPY
+#define ARROW_WITH_UTF8PROC
+#define ARROW_WITH_ZLIB
+#define ARROW_WITH_ZSTD
+#define PARQUET_REQUIRE_ENCRYPTION

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/converter.h

@@ -0,0 +1,411 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#include <memory>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "arrow/array.h"
+#include "arrow/chunked_array.h"
+#include "arrow/status.h"
+#include "arrow/type.h"
+#include "arrow/type_traits.h"
+#include "arrow/util/checked_cast.h"
+#include "arrow/visit_type_inline.h"
+
+namespace arrow {
+namespace internal {
+
+template <typename BaseConverter, template <typename...> class ConverterTrait>
+static Result<std::unique_ptr<BaseConverter>> MakeConverter(
+    std::shared_ptr<DataType> type, typename BaseConverter::OptionsType options,
+    MemoryPool* pool);
+
+template <typename Input, typename Options>
+class Converter {
+ public:
+  using Self = Converter<Input, Options>;
+  using InputType = Input;
+  using OptionsType = Options;
+
+  virtual ~Converter() = default;
+
+  Status Construct(std::shared_ptr<DataType> type, OptionsType options,
+                   MemoryPool* pool) {
+    type_ = std::move(type);
+    options_ = std::move(options);
+    return Init(pool);
+  }
+
+  virtual Status Append(InputType value) { return Status::NotImplemented("Append"); }
+
+  virtual Status Extend(InputType values, int64_t size, int64_t offset = 0) {
+    return Status::NotImplemented("Extend");
+  }
+
+  virtual Status ExtendMasked(InputType values, InputType mask, int64_t size,
+                              int64_t offset = 0) {
+    return Status::NotImplemented("ExtendMasked");
+  }
+
+  const std::shared_ptr<ArrayBuilder>& builder() const { return builder_; }
+
+  const std::shared_ptr<DataType>& type() const { return type_; }
+
+  OptionsType options() const { return options_; }
+
+  bool may_overflow() const { return may_overflow_; }
+
+  bool rewind_on_overflow() const { return rewind_on_overflow_; }
+
+  virtual Status Reserve(int64_t additional_capacity) {
+    return builder_->Reserve(additional_capacity);
+  }
+
+  Status AppendNull() { return builder_->AppendNull(); }
+
+  virtual Result<std::shared_ptr<Array>> ToArray() { return builder_->Finish(); }
+
+  virtual Result<std::shared_ptr<Array>> ToArray(int64_t length) {
+    ARROW_ASSIGN_OR_RAISE(auto arr, this->ToArray());
+    return arr->Slice(0, length);
+  }
+
+  virtual Result<std::shared_ptr<ChunkedArray>> ToChunkedArray() {
+    ARROW_ASSIGN_OR_RAISE(auto array, ToArray());
+    std::vector<std::shared_ptr<Array>> chunks = {std::move(array)};
+    return std::make_shared<ChunkedArray>(chunks);
+  }
+
+ protected:
+  virtual Status Init(MemoryPool* pool) { return Status::OK(); }
+
+  std::shared_ptr<DataType> type_;
+  std::shared_ptr<ArrayBuilder> builder_;
+  OptionsType options_;
+  bool may_overflow_ = false;
+  bool rewind_on_overflow_ = false;
+};
+
+template <typename ArrowType, typename BaseConverter>
+class PrimitiveConverter : public BaseConverter {
+ public:
+  using BuilderType = typename TypeTraits<ArrowType>::BuilderType;
+
+ protected:
+  Status Init(MemoryPool* pool) override {
+    this->builder_ = std::make_shared<BuilderType>(this->type_, pool);
+    // Narrow variable-sized binary types may overflow
+    this->may_overflow_ = is_binary_like(this->type_->id());
+    primitive_type_ = checked_cast<const ArrowType*>(this->type_.get());
+    primitive_builder_ = checked_cast<BuilderType*>(this->builder_.get());
+    return Status::OK();
+  }
+
+  const ArrowType* primitive_type_;
+  BuilderType* primitive_builder_;
+};
+
+template <typename ArrowType, typename BaseConverter,
+          template <typename...> class ConverterTrait>
+class ListConverter : public BaseConverter {
+ public:
+  using BuilderType = typename TypeTraits<ArrowType>::BuilderType;
+  using ConverterType = typename ConverterTrait<ArrowType>::type;
+
+ protected:
+  Status Init(MemoryPool* pool) override {
+    list_type_ = checked_cast<const ArrowType*>(this->type_.get());
+    ARROW_ASSIGN_OR_RAISE(value_converter_,
+                          (MakeConverter<BaseConverter, ConverterTrait>(
+                              list_type_->value_type(), this->options_, pool)));
+    this->builder_ =
+        std::make_shared<BuilderType>(pool, value_converter_->builder(), this->type_);
+    list_builder_ = checked_cast<BuilderType*>(this->builder_.get());
+    // Narrow list types may overflow
+    this->may_overflow_ = this->rewind_on_overflow_ =
+        sizeof(typename ArrowType::offset_type) < sizeof(int64_t);
+    return Status::OK();
+  }
+
+  const ArrowType* list_type_;
+  BuilderType* list_builder_;
+  std::unique_ptr<BaseConverter> value_converter_;
+};
+
+template <typename BaseConverter, template <typename...> class ConverterTrait>
+class StructConverter : public BaseConverter {
+ public:
+  using ConverterType = typename ConverterTrait<StructType>::type;
+
+  Status Reserve(int64_t additional_capacity) override {
+    ARROW_RETURN_NOT_OK(this->builder_->Reserve(additional_capacity));
+    for (const auto& child : children_) {
+      ARROW_RETURN_NOT_OK(child->Reserve(additional_capacity));
+    }
+    return Status::OK();
+  }
+
+ protected:
+  Status Init(MemoryPool* pool) override {
+    std::unique_ptr<BaseConverter> child_converter;
+    std::vector<std::shared_ptr<ArrayBuilder>> child_builders;
+
+    struct_type_ = checked_cast<const StructType*>(this->type_.get());
+    for (const auto& field : struct_type_->fields()) {
+      ARROW_ASSIGN_OR_RAISE(child_converter,
+                            (MakeConverter<BaseConverter, ConverterTrait>(
+                                field->type(), this->options_, pool)));
+      this->may_overflow_ |= child_converter->may_overflow();
+      this->rewind_on_overflow_ = this->may_overflow_;
+      child_builders.push_back(child_converter->builder());
+      children_.push_back(std::move(child_converter));
+    }
+
+    this->builder_ =
+        std::make_shared<StructBuilder>(this->type_, pool, std::move(child_builders));
+    struct_builder_ = checked_cast<StructBuilder*>(this->builder_.get());
+
+    return Status::OK();
+  }
+
+  const StructType* struct_type_;
+  StructBuilder* struct_builder_;
+  std::vector<std::unique_ptr<BaseConverter>> children_;
+};
+
+template <typename ValueType, typename BaseConverter>
+class DictionaryConverter : public BaseConverter {
+ public:
+  using BuilderType = DictionaryBuilder<ValueType>;
+
+ protected:
+  Status Init(MemoryPool* pool) override {
+    std::unique_ptr<ArrayBuilder> builder;
+    ARROW_RETURN_NOT_OK(MakeDictionaryBuilder(pool, this->type_, NULLPTR, &builder));
+    this->builder_ = std::move(builder);
+    this->may_overflow_ = false;
+    dict_type_ = checked_cast<const DictionaryType*>(this->type_.get());
+    value_type_ = checked_cast<const ValueType*>(dict_type_->value_type().get());
+    value_builder_ = checked_cast<BuilderType*>(this->builder_.get());
+    return Status::OK();
+  }
+
+  const DictionaryType* dict_type_;
+  const ValueType* value_type_;
+  BuilderType* value_builder_;
+};
+
+template <typename BaseConverter, template <typename...> class ConverterTrait>
+struct MakeConverterImpl {
+  template <typename T, typename ConverterType = typename ConverterTrait<T>::type>
+  Status Visit(const T&) {
+    out.reset(new ConverterType());
+    return out->Construct(std::move(type), std::move(options), pool);
+  }
+
+  Status Visit(const DictionaryType& t) {
+    switch (t.value_type()->id()) {
+#define DICTIONARY_CASE(TYPE)                                                       \
+  case TYPE::type_id:                                                               \
+    out = std::make_unique<                                                         \
+        typename ConverterTrait<DictionaryType>::template dictionary_type<TYPE>>(); \
+    break;
+      DICTIONARY_CASE(BooleanType);
+      DICTIONARY_CASE(Int8Type);
+      DICTIONARY_CASE(Int16Type);
+      DICTIONARY_CASE(Int32Type);
+      DICTIONARY_CASE(Int64Type);
+      DICTIONARY_CASE(UInt8Type);
+      DICTIONARY_CASE(UInt16Type);
+      DICTIONARY_CASE(UInt32Type);
+      DICTIONARY_CASE(UInt64Type);
+      DICTIONARY_CASE(FloatType);
+      DICTIONARY_CASE(DoubleType);
+      DICTIONARY_CASE(BinaryType);
+      DICTIONARY_CASE(StringType);
+      DICTIONARY_CASE(FixedSizeBinaryType);
+#undef DICTIONARY_CASE
+      default:
+        return Status::NotImplemented("DictionaryArray converter for type ", t.ToString(),
+                                      " not implemented");
+    }
+    return out->Construct(std::move(type), std::move(options), pool);
+  }
+
+  Status Visit(const DataType& t) { return Status::NotImplemented(t.name()); }
+
+  std::shared_ptr<DataType> type;
+  typename BaseConverter::OptionsType options;
+  MemoryPool* pool;
+  std::unique_ptr<BaseConverter> out;
+};
+
+template <typename BaseConverter, template <typename...> class ConverterTrait>
+static Result<std::unique_ptr<BaseConverter>> MakeConverter(
+    std::shared_ptr<DataType> type, typename BaseConverter::OptionsType options,
+    MemoryPool* pool) {
+  MakeConverterImpl<BaseConverter, ConverterTrait> visitor{
+      std::move(type), std::move(options), pool, NULLPTR};
+  ARROW_RETURN_NOT_OK(VisitTypeInline(*visitor.type, &visitor));
+  return std::move(visitor.out);
+}
+
+template <typename Converter>
+class Chunker {
+ public:
+  using InputType = typename Converter::InputType;
+
+  explicit Chunker(std::unique_ptr<Converter> converter)
+      : converter_(std::move(converter)) {}
+
+  Status Reserve(int64_t additional_capacity) {
+    ARROW_RETURN_NOT_OK(converter_->Reserve(additional_capacity));
+    reserved_ += additional_capacity;
+    return Status::OK();
+  }
+
+  Status AppendNull() {
+    auto status = converter_->AppendNull();
+    if (ARROW_PREDICT_FALSE(status.IsCapacityError())) {
+      if (converter_->builder()->length() == 0) {
+        // Builder length == 0 means the individual element is too large to append.
+        // In this case, no need to try again.
+        return status;
+      }
+      ARROW_RETURN_NOT_OK(FinishChunk());
+      return converter_->AppendNull();
+    }
+    ++length_;
+    return status;
+  }
+
+  Status Append(InputType value) {
+    auto status = converter_->Append(value);
+    if (ARROW_PREDICT_FALSE(status.IsCapacityError())) {
+      if (converter_->builder()->length() == 0) {
+        return status;
+      }
+      ARROW_RETURN_NOT_OK(FinishChunk());
+      return Append(value);
+    }
+    ++length_;
+    return status;
+  }
+
+  Status Extend(InputType values, int64_t size, int64_t offset = 0) {
+    while (offset < size) {
+      auto length_before = converter_->builder()->length();
+      auto status = converter_->Extend(values, size, offset);
+      auto length_after = converter_->builder()->length();
+      auto num_converted = length_after - length_before;
+
+      offset += num_converted;
+      length_ += num_converted;
+
+      if (status.IsCapacityError()) {
+        if (converter_->builder()->length() == 0) {
+          // Builder length == 0 means the individual element is too large to append.
+          // In this case, no need to try again.
+          return status;
+        } else if (converter_->rewind_on_overflow()) {
+          // The list-like and binary-like conversion paths may raise  a capacity error,
+          // we need to handle them differently. While the binary-like converters check
+          // the capacity before append/extend the list-like converters just check after
+          // append/extend. Thus depending on the implementation semantics we may need
+          // to rewind (slice) the output chunk by one.
+          length_ -= 1;
+          offset -= 1;
+        }
+        ARROW_RETURN_NOT_OK(FinishChunk());
+      } else if (!status.ok()) {
+        return status;
+      }
+    }
+    return Status::OK();
+  }
+
+  Status ExtendMasked(InputType values, InputType mask, int64_t size,
+                      int64_t offset = 0) {
+    while (offset < size) {
+      auto length_before = converter_->builder()->length();
+      auto status = converter_->ExtendMasked(values, mask, size, offset);
+      auto length_after = converter_->builder()->length();
+      auto num_converted = length_after - length_before;
+
+      offset += num_converted;
+      length_ += num_converted;
+
+      if (status.IsCapacityError()) {
+        if (converter_->builder()->length() == 0) {
+          // Builder length == 0 means the individual element is too large to append.
+          // In this case, no need to try again.
+          return status;
+        } else if (converter_->rewind_on_overflow()) {
+          // The list-like and binary-like conversion paths may raise  a capacity error,
+          // we need to handle them differently. While the binary-like converters check
+          // the capacity before append/extend the list-like converters just check after
+          // append/extend. Thus depending on the implementation semantics we may need
+          // to rewind (slice) the output chunk by one.
+          length_ -= 1;
+          offset -= 1;
+        }
+        ARROW_RETURN_NOT_OK(FinishChunk());
+      } else if (!status.ok()) {
+        return status;
+      }
+    }
+    return Status::OK();
+  }
+
+  Status FinishChunk() {
+    ARROW_ASSIGN_OR_RAISE(auto chunk, converter_->ToArray(length_));
+    chunks_.push_back(chunk);
+    // Reserve space for the remaining items.
+    // Besides being an optimization, it is also required if the converter's
+    // implementation relies on unsafe builder methods in converter->Append().
+    auto remaining = reserved_ - length_;
+    Reset();
+    return Reserve(remaining);
+  }
+
+  Result<std::shared_ptr<ChunkedArray>> ToChunkedArray() {
+    ARROW_RETURN_NOT_OK(FinishChunk());
+    return std::make_shared<ChunkedArray>(chunks_);
+  }
+
+ protected:
+  void Reset() {
+    converter_->builder()->Reset();
+    length_ = 0;
+    reserved_ = 0;
+  }
+
+  int64_t length_ = 0;
+  int64_t reserved_ = 0;
+  std::unique_ptr<Converter> converter_;
+  std::vector<std::shared_ptr<Array>> chunks_;
+};
+
+template <typename T>
+static Result<std::unique_ptr<Chunker<T>>> MakeChunker(std::unique_ptr<T> converter) {
+  return std::make_unique<Chunker<T>>(std::move(converter));
+}
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/cpu_info.h

@@ -0,0 +1,114 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+// From Apache Impala (incubating) as of 2016-01-29. Pared down to a minimal
+// set of functions needed for Apache Arrow / Apache parquet-cpp
+
+#pragma once
+
+#include <cstdint>
+#include <memory>
+#include <string>
+
+#include "arrow/util/macros.h"
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+namespace internal {
+
+/// CpuInfo is an interface to query for cpu information at runtime.  The caller can
+/// ask for the sizes of the caches and what hardware features are supported.
+/// On Linux, this information is pulled from a couple of sys files (/proc/cpuinfo and
+/// /sys/devices)
+class ARROW_EXPORT CpuInfo {
+ public:
+  ~CpuInfo();
+
+  /// x86 features
+  static constexpr int64_t SSSE3 = (1LL << 0);
+  static constexpr int64_t SSE4_1 = (1LL << 1);
+  static constexpr int64_t SSE4_2 = (1LL << 2);
+  static constexpr int64_t POPCNT = (1LL << 3);
+  static constexpr int64_t AVX = (1LL << 4);
+  static constexpr int64_t AVX2 = (1LL << 5);
+  static constexpr int64_t AVX512F = (1LL << 6);
+  static constexpr int64_t AVX512CD = (1LL << 7);
+  static constexpr int64_t AVX512VL = (1LL << 8);
+  static constexpr int64_t AVX512DQ = (1LL << 9);
+  static constexpr int64_t AVX512BW = (1LL << 10);
+  static constexpr int64_t AVX512 = AVX512F | AVX512CD | AVX512VL | AVX512DQ | AVX512BW;
+  static constexpr int64_t BMI1 = (1LL << 11);
+  static constexpr int64_t BMI2 = (1LL << 12);
+
+  /// Arm features
+  static constexpr int64_t ASIMD = (1LL << 32);
+
+  /// Cache enums for L1 (data), L2 and L3
+  enum class CacheLevel { L1 = 0, L2, L3, Last = L3 };
+
+  /// CPU vendors
+  enum class Vendor { Unknown, Intel, AMD };
+
+  static const CpuInfo* GetInstance();
+
+  /// Returns all the flags for this cpu
+  int64_t hardware_flags() const;
+
+  /// Returns the number of cores (including hyper-threaded) on this machine.
+  int num_cores() const;
+
+  /// Returns the vendor of the cpu.
+  Vendor vendor() const;
+
+  /// Returns the model name of the cpu (e.g. Intel i7-2600)
+  const std::string& model_name() const;
+
+  /// Returns the size of the cache in KB at this cache level
+  int64_t CacheSize(CacheLevel level) const;
+
+  /// \brief Returns whether or not the given feature is enabled.
+  ///
+  /// IsSupported() is true iff IsDetected() is also true and the feature
+  /// wasn't disabled by the user (for example by setting the ARROW_USER_SIMD_LEVEL
+  /// environment variable).
+  bool IsSupported(int64_t flags) const;
+
+  /// Returns whether or not the given feature is available on the CPU.
+  bool IsDetected(int64_t flags) const;
+
+  /// Determine if the CPU meets the minimum CPU requirements and if not, issue an error
+  /// and terminate.
+  void VerifyCpuRequirements() const;
+
+  /// Toggle a hardware feature on and off.  It is not valid to turn on a feature
+  /// that the underlying hardware cannot support. This is useful for testing.
+  void EnableFeature(int64_t flag, bool enable);
+
+  bool HasEfficientBmi2() const {
+    // BMI2 (pext, pdep) is only efficient on Intel X86 processors.
+    return vendor() == Vendor::Intel && IsSupported(BMI2);
+  }
+
+ private:
+  CpuInfo();
+
+  struct Impl;
+  std::unique_ptr<Impl> impl_;
+};
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/crc32.h

@@ -0,0 +1,36 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#include <cstddef>
+#include <cstdint>
+
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+namespace internal {
+
+/// \brief Compute the CRC32 checksum of the given data
+///
+/// This function computes CRC32 with the polynomial 0x04C11DB7,
+/// as used in zlib and others (note this is different from CRC32C).
+/// To compute a running CRC32, pass the previous value in `prev`,
+/// otherwise `prev` should be 0.
+ARROW_EXPORT
+uint32_t crc32(uint32_t prev, const void* data, size_t length);
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/debug.h

@@ -0,0 +1,29 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+namespace internal {
+
+ARROW_EXPORT
+void DebugTrap();
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/decimal.h

@@ -0,0 +1,523 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cstdint>
+#include <iosfwd>
+#include <limits>
+#include <string>
+#include <string_view>
+#include <utility>
+
+#include "arrow/result.h"
+#include "arrow/status.h"
+#include "arrow/type_fwd.h"
+#include "arrow/util/basic_decimal.h"
+
+namespace arrow {
+
+class Decimal64;
+
+namespace internal {
+
+ARROW_EXPORT
+Status ToArrowStatus(DecimalStatus);
+
+}  // namespace internal
+
+template <>
+struct IntoStatus<DecimalStatus> {
+  static inline Status ToStatus(DecimalStatus st) { return internal::ToArrowStatus(st); }
+};
+
+/// Represents a signed 32-bit decimal value in two's complement.
+/// Calulations wrap around and overflow is ignored.
+/// The max decimal precision that can be safely represented is
+/// 9 significant digits.
+///
+/// The implementation is split into two parts :
+///
+/// 1. BasicDecimal32
+///    - can be safely compiled to IR without references to libstdc++
+/// 2. Decimal32
+///    - has additional functionality on top of BasicDecimal32 to deal with
+///      strings and streams
+class ARROW_EXPORT Decimal32 : public BasicDecimal32 {
+ public:
+  /// \cond FALSE
+  // (need to avoid a duplicate definition in sphinx)
+  using BasicDecimal32::BasicDecimal32;
+  /// \endcond
+
+  /// \brief constructor creates a Decimal32 from a BasicDecimal32
+  constexpr Decimal32(const BasicDecimal32& value) noexcept  // NOLINT runtime/explicit
+      : BasicDecimal32(value) {}
+
+  /// \brief Parse the number from a base 10 string representation
+  explicit Decimal32(const std::string& value);
+
+  /// \brief Empty constructor creates a Decimal32 with a value of 0
+  /// this is required for some older compilers
+  constexpr Decimal32() noexcept : BasicDecimal32() {}
+
+  /// \brief Divide this number by right and return the result.
+  ///
+  /// This operation is not destructive.
+  /// The answer rounds to zero. Signs work like:
+  ///   21 /  5 ->  4,  1
+  ///  -21 /  5 -> -4, -1
+  ///   21 / -5 -> -4,  1
+  ///  -21 / -5 ->  4, -1
+  /// \param[in] divisor the number to divide by
+  /// \return the pair of the quotient and the remainder
+  Result<std::pair<Decimal32, Decimal32>> Divide(const Decimal32& divisor) const {
+    std::pair<Decimal32, Decimal32> result;
+    ARROW_RETURN_NOT_OK(BasicDecimal32::Divide(divisor, &result.first, &result.second));
+    return result;
+  }
+
+  /// \brief Convert the Decimal32 value to a base 10 decimal string with the given scale
+  std::string ToString(int32_t scale) const;
+
+  /// \brief Convert the value to an integer string
+  std::string ToIntegerString() const;
+
+  /// \brief Cast this value to an int64_t
+  explicit operator int64_t() const;
+
+  explicit operator Decimal64() const;
+
+  /// \brief Convert a decimal string to a Decimal value, optionally including
+  /// precision and scale if they're passed in and not null.
+  static Status FromString(std::string_view s, Decimal32* out, int32_t* precision,
+                           int32_t* scale = NULLPTR);
+  static Status FromString(const std::string& s, Decimal32* out, int32_t* precision,
+                           int32_t* scale = NULLPTR);
+  static Status FromString(const char* s, Decimal32* out, int32_t* precision,
+                           int32_t* scale = NULLPTR);
+  static Result<Decimal32> FromString(std::string_view s);
+  static Result<Decimal32> FromString(const std::string& s);
+  static Result<Decimal32> FromString(const char* s);
+
+  static Result<Decimal32> FromReal(double real, int32_t precision, int32_t scale);
+  static Result<Decimal32> FromReal(float real, int32_t precision, int32_t scale);
+
+  /// \brief Convert from a big-endian byte representation. The length must be
+  ///        between 1 and 4
+  /// \return error status if the length is an invalid value
+  static Result<Decimal32> FromBigEndian(const uint8_t* data, int32_t length);
+
+  /// \brief Convert Decimal32 from one scale to another
+  Result<Decimal32> Rescale(int32_t original_scale, int32_t new_scale) const {
+    Decimal32 out;
+    ARROW_RETURN_NOT_OK(BasicDecimal32::Rescale(original_scale, new_scale, &out));
+    return out;
+  }
+
+  /// \brief Convert to a signed integer
+  template <typename T, typename = internal::EnableIfIsOneOf<T, int32_t, int64_t>>
+  Result<T> ToInteger() const {
+    return static_cast<T>(value_);
+  }
+
+  /// \brief Convert to a signed integer
+  template <typename T, typename = internal::EnableIfIsOneOf<T, int32_t, int64_t>>
+  Status ToInteger(T* out) const {
+    return ToInteger<T>().Value(out);
+  }
+
+  /// \brief Convert to a floating-point number (scaled)
+  float ToFloat(int32_t scale) const;
+  /// \brief Convert to a floating-point number (scaled)
+  double ToDouble(int32_t scale) const;
+
+  /// \brief Convert to a floating-point number (scaled)
+  template <typename T, typename = std::enable_if_t<std::is_floating_point_v<T>>>
+  T ToReal(int32_t scale) const {
+    static_assert(std::is_same_v<T, float> || std::is_same_v<T, double>,
+                  "Unexpected floating-point type");
+    if constexpr (std::is_same_v<T, float>) {
+      return ToFloat(scale);
+    } else {
+      return ToDouble(scale);
+    }
+  }
+
+  ARROW_FRIEND_EXPORT friend std::ostream& operator<<(std::ostream& os,
+                                                      const Decimal32& decimal);
+};
+
+class ARROW_EXPORT Decimal64 : public BasicDecimal64 {
+ public:
+  /// \cond FALSE
+  // (need to avoid a duplicate definition in sphinx)
+  using BasicDecimal64::BasicDecimal64;
+  /// \endcond
+
+  /// \brief constructor creates a Decimal64 from a BasicDecimal64
+  constexpr Decimal64(const BasicDecimal64& value) noexcept  // NOLINT runtime/explicit
+      : BasicDecimal64(value) {}
+
+  explicit Decimal64(const BasicDecimal32& value) noexcept
+      : BasicDecimal64(static_cast<int64_t>(value.value())) {}
+
+  /// \brief Parse the number from a base 10 string representation
+  explicit Decimal64(const std::string& value);
+
+  /// \brief Empty constructor creates a Decimal64 with a value of 0
+  /// this is required for some older compilers
+  constexpr Decimal64() noexcept : BasicDecimal64() {}
+
+  /// \brief Divide this number by right and return the result.
+  ///
+  /// This operation is not destructive.
+  /// The answer rounds to zero. Signs work like:
+  ///   21 /  5 ->  4,  1
+  ///  -21 /  5 -> -4, -1
+  ///   21 / -5 -> -4,  1
+  ///  -21 / -5 ->  4, -1
+  /// \param[in] divisor the number to divide by
+  /// \return the pair of the quotient and the remainder
+  Result<std::pair<Decimal64, Decimal64>> Divide(const Decimal64& divisor) const {
+    std::pair<Decimal64, Decimal64> result;
+    ARROW_RETURN_NOT_OK(BasicDecimal64::Divide(divisor, &result.first, &result.second));
+    return result;
+  }
+
+  /// \brief Convert the Decimal64 value to a base 10 decimal string with the given scale
+  std::string ToString(int32_t scale) const;
+
+  /// \brief Convert the value to an integer string
+  std::string ToIntegerString() const;
+
+  /// \brief Cast this value to an int64_t
+  explicit operator int64_t() const;
+
+  /// \brief Convert a decimal string to a Decimal value, optionally including
+  /// precision and scale if they're passed in and not null.
+  static Status FromString(std::string_view s, Decimal64* out, int32_t* precision,
+                           int32_t* scale = NULLPTR);
+  static Status FromString(const std::string& s, Decimal64* out, int32_t* precision,
+                           int32_t* scale = NULLPTR);
+  static Status FromString(const char* s, Decimal64* out, int32_t* precision,
+                           int32_t* scale = NULLPTR);
+  static Result<Decimal64> FromString(std::string_view s);
+  static Result<Decimal64> FromString(const std::string& s);
+  static Result<Decimal64> FromString(const char* s);
+
+  static Result<Decimal64> FromReal(double real, int32_t precision, int32_t scale);
+  static Result<Decimal64> FromReal(float real, int32_t precision, int32_t scale);
+
+  /// \brief Convert from a big-endian byte representation. The length must be
+  ///        between 1 and 4
+  /// \return error status if the length is an invalid value
+  static Result<Decimal64> FromBigEndian(const uint8_t* data, int32_t length);
+
+  /// \brief Convert Decimal64 from one scale to another
+  Result<Decimal64> Rescale(int32_t original_scale, int32_t new_scale) const {
+    Decimal64 out;
+    ARROW_RETURN_NOT_OK(BasicDecimal64::Rescale(original_scale, new_scale, &out));
+    return out;
+  }
+
+  /// \brief Convert to a signed integer
+  template <typename T, typename = internal::EnableIfIsOneOf<T, int32_t, int64_t>>
+  Result<T> ToInteger() const {
+    return static_cast<T>(value_);
+  }
+
+  /// \brief Convert to a signed integer
+  template <typename T, typename = internal::EnableIfIsOneOf<T, int32_t, int64_t>>
+  Status ToInteger(T* out) const {
+    return ToInteger<T>().Value(out);
+  }
+
+  /// \brief Convert to a floating-point number (scaled)
+  float ToFloat(int32_t scale) const;
+  /// \brief Convert to a floating-point number (scaled)
+  double ToDouble(int32_t scale) const;
+
+  /// \brief Convert to a floating-point number (scaled)
+  template <typename T, typename = std::enable_if_t<std::is_floating_point_v<T>>>
+  T ToReal(int32_t scale) const {
+    static_assert(std::is_same_v<T, float> || std::is_same_v<T, double>,
+                  "Unexpected floating-point type");
+    if constexpr (std::is_same_v<T, float>) {
+      return ToFloat(scale);
+    } else {
+      return ToDouble(scale);
+    }
+  }
+
+  ARROW_FRIEND_EXPORT friend std::ostream& operator<<(std::ostream& os,
+                                                      const Decimal64& decimal);
+};
+
+/// Represents a signed 128-bit integer in two's complement.
+/// Calculations wrap around and overflow is ignored.
+/// The max decimal precision that can be safely represented is
+/// 38 significant digits.
+///
+/// For a discussion of the algorithms, look at Knuth's volume 2,
+/// Semi-numerical Algorithms section 4.3.1.
+///
+/// Adapted from the Apache ORC C++ implementation
+///
+/// The implementation is split into two parts :
+///
+/// 1. BasicDecimal128
+///    - can be safely compiled to IR without references to libstdc++.
+/// 2. Decimal128
+///    - has additional functionality on top of BasicDecimal128 to deal with
+///      strings and streams.
+class ARROW_EXPORT Decimal128 : public BasicDecimal128 {
+ public:
+  /// \cond FALSE
+  // (need to avoid a duplicate definition in Sphinx)
+  using BasicDecimal128::BasicDecimal128;
+  /// \endcond
+
+  /// \brief constructor creates a Decimal128 from a BasicDecimal128.
+  constexpr Decimal128(const BasicDecimal128& value) noexcept  // NOLINT runtime/explicit
+      : BasicDecimal128(value) {}
+
+  /// \brief Parse the number from a base 10 string representation.
+  explicit Decimal128(const std::string& value);
+
+  /// \brief Empty constructor creates a Decimal128 with a value of 0.
+  // This is required on some older compilers.
+  constexpr Decimal128() noexcept : BasicDecimal128() {}
+
+  /// Divide this number by right and return the result.
+  ///
+  /// This operation is not destructive.
+  /// The answer rounds to zero. Signs work like:
+  ///   21 /  5 ->  4,  1
+  ///  -21 /  5 -> -4, -1
+  ///   21 / -5 -> -4,  1
+  ///  -21 / -5 ->  4, -1
+  /// \param[in] divisor the number to divide by
+  /// \return the pair of the quotient and the remainder
+  Result<std::pair<Decimal128, Decimal128>> Divide(const Decimal128& divisor) const {
+    std::pair<Decimal128, Decimal128> result;
+    ARROW_RETURN_NOT_OK(BasicDecimal128::Divide(divisor, &result.first, &result.second));
+    return result;
+  }
+
+  /// \brief Convert the Decimal128 value to a base 10 decimal string with the given
+  /// scale.
+  std::string ToString(int32_t scale) const;
+
+  /// \brief Convert the value to an integer string
+  std::string ToIntegerString() const;
+
+  /// \brief Cast this value to an int64_t.
+  explicit operator int64_t() const;
+
+  /// \brief Convert a decimal string to a Decimal128 value, optionally including
+  /// precision and scale if they're passed in and not null.
+  static Status FromString(std::string_view s, Decimal128* out, int32_t* precision,
+                           int32_t* scale = NULLPTR);
+  static Status FromString(const std::string& s, Decimal128* out, int32_t* precision,
+                           int32_t* scale = NULLPTR);
+  static Status FromString(const char* s, Decimal128* out, int32_t* precision,
+                           int32_t* scale = NULLPTR);
+  static Result<Decimal128> FromString(std::string_view s);
+  static Result<Decimal128> FromString(const std::string& s);
+  static Result<Decimal128> FromString(const char* s);
+
+  static Result<Decimal128> FromReal(double real, int32_t precision, int32_t scale);
+  static Result<Decimal128> FromReal(float real, int32_t precision, int32_t scale);
+
+  /// \brief Convert from a big-endian byte representation. The length must be
+  ///        between 1 and 16.
+  /// \return error status if the length is an invalid value
+  static Result<Decimal128> FromBigEndian(const uint8_t* data, int32_t length);
+
+  /// \brief Convert Decimal128 from one scale to another
+  Result<Decimal128> Rescale(int32_t original_scale, int32_t new_scale) const {
+    Decimal128 out;
+    ARROW_RETURN_NOT_OK(BasicDecimal128::Rescale(original_scale, new_scale, &out));
+    return out;
+  }
+
+  /// \brief Convert to a signed integer
+  template <typename T, typename = internal::EnableIfIsOneOf<T, int32_t, int64_t>>
+  Result<T> ToInteger() const {
+    constexpr auto min_value = std::numeric_limits<T>::min();
+    constexpr auto max_value = std::numeric_limits<T>::max();
+    const auto& self = *this;
+    if (self < min_value || self > max_value) {
+      return Status::Invalid("Invalid cast from Decimal128 to ", sizeof(T),
+                             " byte integer");
+    }
+    return static_cast<T>(low_bits());
+  }
+
+  /// \brief Convert to a signed integer
+  template <typename T, typename = internal::EnableIfIsOneOf<T, int32_t, int64_t>>
+  Status ToInteger(T* out) const {
+    return ToInteger<T>().Value(out);
+  }
+
+  /// \brief Convert to a floating-point number (scaled)
+  float ToFloat(int32_t scale) const;
+  /// \brief Convert to a floating-point number (scaled)
+  double ToDouble(int32_t scale) const;
+
+  /// \brief Convert to a floating-point number (scaled)
+  template <typename T, typename = std::enable_if_t<std::is_floating_point_v<T>>>
+  T ToReal(int32_t scale) const {
+    static_assert(std::is_same_v<T, float> || std::is_same_v<T, double>,
+                  "Unexpected floating-point type");
+    if constexpr (std::is_same_v<T, float>) {
+      return ToFloat(scale);
+    } else {
+      return ToDouble(scale);
+    }
+  }
+
+  ARROW_FRIEND_EXPORT friend std::ostream& operator<<(std::ostream& os,
+                                                      const Decimal128& decimal);
+};
+
+/// Represents a signed 256-bit integer in two's complement.
+/// The max decimal precision that can be safely represented is
+/// 76 significant digits.
+///
+/// The implementation is split into two parts :
+///
+/// 1. BasicDecimal256
+///    - can be safely compiled to IR without references to libstdc++.
+/// 2. Decimal256
+///    - (TODO) has additional functionality on top of BasicDecimal256 to deal with
+///      strings and streams.
+class ARROW_EXPORT Decimal256 : public BasicDecimal256 {
+ public:
+  /// \cond FALSE
+  // (need to avoid a duplicate definition in Sphinx)
+  using BasicDecimal256::BasicDecimal256;
+  /// \endcond
+
+  /// \brief constructor creates a Decimal256 from a BasicDecimal256.
+  constexpr Decimal256(const BasicDecimal256& value) noexcept  // NOLINT(runtime/explicit)
+      : BasicDecimal256(value) {}
+
+  /// \brief Parse the number from a base 10 string representation.
+  explicit Decimal256(const std::string& value);
+
+  /// \brief Empty constructor creates a Decimal256 with a value of 0.
+  // This is required on some older compilers.
+  constexpr Decimal256() noexcept : BasicDecimal256() {}
+
+  /// \brief Convert the Decimal256 value to a base 10 decimal string with the given
+  /// scale.
+  std::string ToString(int32_t scale) const;
+
+  /// \brief Convert the value to an integer string
+  std::string ToIntegerString() const;
+
+  /// \brief Convert a decimal string to a Decimal256 value, optionally including
+  /// precision and scale if they're passed in and not null.
+  static Status FromString(std::string_view s, Decimal256* out, int32_t* precision,
+                           int32_t* scale = NULLPTR);
+  static Status FromString(const std::string& s, Decimal256* out, int32_t* precision,
+                           int32_t* scale = NULLPTR);
+  static Status FromString(const char* s, Decimal256* out, int32_t* precision,
+                           int32_t* scale = NULLPTR);
+  static Result<Decimal256> FromString(std::string_view s);
+  static Result<Decimal256> FromString(const std::string& s);
+  static Result<Decimal256> FromString(const char* s);
+
+  /// \brief Convert Decimal256 from one scale to another
+  Result<Decimal256> Rescale(int32_t original_scale, int32_t new_scale) const {
+    Decimal256 out;
+    ARROW_RETURN_NOT_OK(BasicDecimal256::Rescale(original_scale, new_scale, &out));
+    return out;
+  }
+
+  /// Divide this number by right and return the result.
+  ///
+  /// This operation is not destructive.
+  /// The answer rounds to zero. Signs work like:
+  ///   21 /  5 ->  4,  1
+  ///  -21 /  5 -> -4, -1
+  ///   21 / -5 -> -4,  1
+  ///  -21 / -5 ->  4, -1
+  /// \param[in] divisor the number to divide by
+  /// \return the pair of the quotient and the remainder
+  Result<std::pair<Decimal256, Decimal256>> Divide(const Decimal256& divisor) const {
+    std::pair<Decimal256, Decimal256> result;
+    ARROW_RETURN_NOT_OK(BasicDecimal256::Divide(divisor, &result.first, &result.second));
+    return result;
+  }
+
+  /// \brief Convert from a big-endian byte representation. The length must be
+  ///        between 1 and 32.
+  /// \return error status if the length is an invalid value
+  static Result<Decimal256> FromBigEndian(const uint8_t* data, int32_t length);
+
+  static Result<Decimal256> FromReal(double real, int32_t precision, int32_t scale);
+  static Result<Decimal256> FromReal(float real, int32_t precision, int32_t scale);
+
+  /// \brief Convert to a floating-point number (scaled).
+  /// May return infinity in case of overflow.
+  float ToFloat(int32_t scale) const;
+  /// \brief Convert to a floating-point number (scaled)
+  double ToDouble(int32_t scale) const;
+
+  /// \brief Convert to a floating-point number (scaled)
+  template <typename T, typename = std::enable_if_t<std::is_floating_point_v<T>>>
+  T ToReal(int32_t scale) const {
+    static_assert(std::is_same_v<T, float> || std::is_same_v<T, double>,
+                  "Unexpected floating-point type");
+    if constexpr (std::is_same_v<T, float>) {
+      return ToFloat(scale);
+    } else {
+      return ToDouble(scale);
+    }
+  }
+
+  ARROW_FRIEND_EXPORT friend std::ostream& operator<<(std::ostream& os,
+                                                      const Decimal256& decimal);
+};
+
+/// For an integer type, return the max number of decimal digits
+/// (=minimal decimal precision) it can represent.
+inline Result<int32_t> MaxDecimalDigitsForInteger(Type::type type_id) {
+  switch (type_id) {
+    case Type::INT8:
+    case Type::UINT8:
+      return 3;
+    case Type::INT16:
+    case Type::UINT16:
+      return 5;
+    case Type::INT32:
+    case Type::UINT32:
+      return 10;
+    case Type::INT64:
+      return 19;
+    case Type::UINT64:
+      return 20;
+    default:
+      break;
+  }
+  return Status::Invalid("Not an integer type: ", type_id);
+}
+
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/delimiting.h

@@ -0,0 +1,181 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cstdint>
+#include <memory>
+#include <string_view>
+
+#include "arrow/status.h"
+#include "arrow/util/macros.h"
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+
+class Buffer;
+
+class ARROW_EXPORT BoundaryFinder {
+ public:
+  BoundaryFinder() = default;
+
+  virtual ~BoundaryFinder();
+
+  /// \brief Find the position of the first delimiter inside block
+  ///
+  /// `partial` is taken to be the beginning of the block, and `block`
+  /// its continuation.  Also, `partial` doesn't contain a delimiter.
+  ///
+  /// The returned `out_pos` is relative to `block`'s start and should point
+  /// to the first character after the first delimiter.
+  /// `out_pos` will be -1 if no delimiter is found.
+  virtual Status FindFirst(std::string_view partial, std::string_view block,
+                           int64_t* out_pos) = 0;
+
+  /// \brief Find the position of the last delimiter inside block
+  ///
+  /// The returned `out_pos` is relative to `block`'s start and should point
+  /// to the first character after the last delimiter.
+  /// `out_pos` will be -1 if no delimiter is found.
+  virtual Status FindLast(std::string_view block, int64_t* out_pos) = 0;
+
+  /// \brief Find the position of the Nth delimiter inside the block
+  ///
+  /// `partial` is taken to be the beginning of the block, and `block`
+  /// its continuation.  Also, `partial` doesn't contain a delimiter.
+  ///
+  /// The returned `out_pos` is relative to `block`'s start and should point
+  /// to the first character after the first delimiter.
+  /// `out_pos` will be -1 if no delimiter is found.
+  ///
+  /// The returned `num_found` is the number of delimiters actually found
+  virtual Status FindNth(std::string_view partial, std::string_view block, int64_t count,
+                         int64_t* out_pos, int64_t* num_found) = 0;
+
+  static constexpr int64_t kNoDelimiterFound = -1;
+
+ protected:
+  ARROW_DISALLOW_COPY_AND_ASSIGN(BoundaryFinder);
+};
+
+ARROW_EXPORT
+std::shared_ptr<BoundaryFinder> MakeNewlineBoundaryFinder();
+
+/// \brief A reusable block-based chunker for delimited data
+///
+/// The chunker takes a block of delimited data and helps carve a sub-block
+/// which begins and ends on delimiters (suitable for consumption by parsers
+/// which can only parse whole objects).
+class ARROW_EXPORT Chunker {
+ public:
+  explicit Chunker(std::shared_ptr<BoundaryFinder> delimiter);
+  ~Chunker();
+
+  /// \brief Carve up a chunk in a block of data to contain only whole objects
+  ///
+  /// Pre-conditions:
+  /// - `block` is the start of a valid block of delimited data
+  ///   (i.e. starts just after a delimiter)
+  ///
+  /// Post-conditions:
+  /// - block == whole + partial
+  /// - `whole` is a valid block of delimited data
+  ///   (i.e. starts just after a delimiter and ends with a delimiter)
+  /// - `partial` doesn't contain an entire delimited object
+  ///   (IOW: `partial` is generally small)
+  ///
+  /// This method will look for the last delimiter in `block` and may
+  /// therefore be costly.
+  ///
+  /// \param[in] block data to be chunked
+  /// \param[out] whole subrange of block containing whole delimited objects
+  /// \param[out] partial subrange of block starting with a partial delimited object
+  Status Process(std::shared_ptr<Buffer> block, std::shared_ptr<Buffer>* whole,
+                 std::shared_ptr<Buffer>* partial);
+
+  /// \brief Carve the completion of a partial object out of a block
+  ///
+  /// Pre-conditions:
+  /// - `partial` is the start of a valid block of delimited data
+  ///   (i.e. starts just after a delimiter)
+  /// - `block` follows `partial` in file order
+  ///
+  /// Post-conditions:
+  /// - block == completion + rest
+  /// - `partial + completion` is a valid block of delimited data
+  ///   (i.e. starts just after a delimiter and ends with a delimiter)
+  /// - `completion` doesn't contain an entire delimited object
+  ///   (IOW: `completion` is generally small)
+  ///
+  /// This method will look for the first delimiter in `block` and should
+  /// therefore be reasonably cheap.
+  ///
+  /// \param[in] partial incomplete delimited data
+  /// \param[in] block delimited data following partial
+  /// \param[out] completion subrange of block containing the completion of partial
+  /// \param[out] rest subrange of block containing what completion does not cover
+  Status ProcessWithPartial(std::shared_ptr<Buffer> partial,
+                            std::shared_ptr<Buffer> block,
+                            std::shared_ptr<Buffer>* completion,
+                            std::shared_ptr<Buffer>* rest);
+
+  /// \brief Like ProcessWithPartial, but for the last block of a file
+  ///
+  /// This method allows for a final delimited object without a trailing delimiter
+  /// (ProcessWithPartial would return an error in that case).
+  ///
+  /// Pre-conditions:
+  /// - `partial` is the start of a valid block of delimited data
+  /// - `block` follows `partial` in file order and is the last data block
+  ///
+  /// Post-conditions:
+  /// - block == completion + rest
+  /// - `partial + completion` is a valid block of delimited data
+  /// - `completion` doesn't contain an entire delimited object
+  ///   (IOW: `completion` is generally small)
+  ///
+  Status ProcessFinal(std::shared_ptr<Buffer> partial, std::shared_ptr<Buffer> block,
+                      std::shared_ptr<Buffer>* completion, std::shared_ptr<Buffer>* rest);
+
+  /// \brief Skip count number of rows
+  /// Pre-conditions:
+  /// - `partial` is the start of a valid block of delimited data
+  ///   (i.e. starts just after a delimiter)
+  /// - `block` follows `partial` in file order
+  ///
+  /// Post-conditions:
+  /// - `count` is updated to indicate the number of rows that still need to be skipped
+  /// - If `count` is > 0 then `rest` is an incomplete block that should be a future
+  /// `partial`
+  /// - Else `rest` could be one or more valid blocks of delimited data which need to be
+  /// parsed
+  ///
+  /// \param[in] partial incomplete delimited data
+  /// \param[in] block delimited data following partial
+  /// \param[in] final whether this is the final chunk
+  /// \param[in,out] count number of rows that need to be skipped
+  /// \param[out] rest subrange of block containing what was not skipped
+  Status ProcessSkip(std::shared_ptr<Buffer> partial, std::shared_ptr<Buffer> block,
+                     bool final, int64_t* count, std::shared_ptr<Buffer>* rest);
+
+ protected:
+  ARROW_DISALLOW_COPY_AND_ASSIGN(Chunker);
+
+  std::shared_ptr<BoundaryFinder> boundary_finder_;
+};
+
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/endian.h

@@ -0,0 +1,245 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#ifdef _WIN32
+#  define ARROW_LITTLE_ENDIAN 1
+#else
+#  if defined(__APPLE__) || defined(__FreeBSD__)
+#    include <machine/endian.h>  // IWYU pragma: keep
+#  elif defined(sun) || defined(__sun)
+#    include <sys/byteorder.h>  // IWYU pragma: keep
+#  elif !defined(_AIX)
+#    include <endian.h>  // IWYU pragma: keep
+#  endif
+#
+#  ifndef __BYTE_ORDER__
+#    error "__BYTE_ORDER__ not defined"
+#  endif
+#
+#  ifndef __ORDER_LITTLE_ENDIAN__
+#    error "__ORDER_LITTLE_ENDIAN__ not defined"
+#  endif
+#
+#  if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+#    define ARROW_LITTLE_ENDIAN 1
+#  else
+#    define ARROW_LITTLE_ENDIAN 0
+#  endif
+#endif
+
+#if defined(_MSC_VER)
+#  include <intrin.h>  // IWYU pragma: keep
+#  define ARROW_BYTE_SWAP64 _byteswap_uint64
+#  define ARROW_BYTE_SWAP32 _byteswap_ulong
+#else
+#  define ARROW_BYTE_SWAP64 __builtin_bswap64
+#  define ARROW_BYTE_SWAP32 __builtin_bswap32
+#endif
+
+#include <algorithm>
+#include <array>
+
+#include "arrow/util/type_traits.h"
+#include "arrow/util/ubsan.h"
+
+namespace arrow {
+namespace bit_util {
+
+//
+// Byte-swap 16-bit, 32-bit and 64-bit values
+//
+
+// Swap the byte order (i.e. endianness)
+static inline int64_t ByteSwap(int64_t value) { return ARROW_BYTE_SWAP64(value); }
+static inline uint64_t ByteSwap(uint64_t value) {
+  return static_cast<uint64_t>(ARROW_BYTE_SWAP64(value));
+}
+static inline int32_t ByteSwap(int32_t value) { return ARROW_BYTE_SWAP32(value); }
+static inline uint32_t ByteSwap(uint32_t value) {
+  return static_cast<uint32_t>(ARROW_BYTE_SWAP32(value));
+}
+static inline int16_t ByteSwap(int16_t value) {
+  constexpr auto m = static_cast<int16_t>(0xff);
+  return static_cast<int16_t>(((value >> 8) & m) | ((value & m) << 8));
+}
+static inline uint16_t ByteSwap(uint16_t value) {
+  return static_cast<uint16_t>(ByteSwap(static_cast<int16_t>(value)));
+}
+static inline uint8_t ByteSwap(uint8_t value) { return value; }
+static inline int8_t ByteSwap(int8_t value) { return value; }
+static inline double ByteSwap(double value) {
+  const uint64_t swapped = ARROW_BYTE_SWAP64(util::SafeCopy<uint64_t>(value));
+  return util::SafeCopy<double>(swapped);
+}
+static inline float ByteSwap(float value) {
+  const uint32_t swapped = ARROW_BYTE_SWAP32(util::SafeCopy<uint32_t>(value));
+  return util::SafeCopy<float>(swapped);
+}
+
+// Write the swapped bytes into dst. Src and dst cannot overlap.
+static inline void ByteSwap(void* dst, const void* src, int len) {
+  switch (len) {
+    case 1:
+      *reinterpret_cast<int8_t*>(dst) = *reinterpret_cast<const int8_t*>(src);
+      return;
+    case 2:
+      *reinterpret_cast<int16_t*>(dst) = ByteSwap(*reinterpret_cast<const int16_t*>(src));
+      return;
+    case 4:
+      *reinterpret_cast<int32_t*>(dst) = ByteSwap(*reinterpret_cast<const int32_t*>(src));
+      return;
+    case 8:
+      *reinterpret_cast<int64_t*>(dst) = ByteSwap(*reinterpret_cast<const int64_t*>(src));
+      return;
+    default:
+      break;
+  }
+
+  auto d = reinterpret_cast<uint8_t*>(dst);
+  auto s = reinterpret_cast<const uint8_t*>(src);
+  for (int i = 0; i < len; ++i) {
+    d[i] = s[len - i - 1];
+  }
+}
+
+// Convert to little/big endian format from the machine's native endian format.
+#if ARROW_LITTLE_ENDIAN
+template <typename T, typename = internal::EnableIfIsOneOf<
+                          T, int64_t, uint64_t, int32_t, uint32_t, int16_t, uint16_t,
+                          uint8_t, int8_t, float, double, bool>>
+static inline T ToBigEndian(T value) {
+  return ByteSwap(value);
+}
+
+template <typename T, typename = internal::EnableIfIsOneOf<
+                          T, int64_t, uint64_t, int32_t, uint32_t, int16_t, uint16_t,
+                          uint8_t, int8_t, float, double, bool>>
+static inline T ToLittleEndian(T value) {
+  return value;
+}
+#else
+template <typename T, typename = internal::EnableIfIsOneOf<
+                          T, int64_t, uint64_t, int32_t, uint32_t, int16_t, uint16_t,
+                          uint8_t, int8_t, float, double, bool>>
+static inline T ToBigEndian(T value) {
+  return value;
+}
+
+template <typename T, typename = internal::EnableIfIsOneOf<
+                          T, int64_t, uint64_t, int32_t, uint32_t, int16_t, uint16_t,
+                          uint8_t, int8_t, float, double, bool>>
+static inline T ToLittleEndian(T value) {
+  return ByteSwap(value);
+}
+#endif
+
+// Convert from big/little endian format to the machine's native endian format.
+#if ARROW_LITTLE_ENDIAN
+template <typename T, typename = internal::EnableIfIsOneOf<
+                          T, int64_t, uint64_t, int32_t, uint32_t, int16_t, uint16_t,
+                          uint8_t, int8_t, float, double, bool>>
+static inline T FromBigEndian(T value) {
+  return ByteSwap(value);
+}
+
+template <typename T, typename = internal::EnableIfIsOneOf<
+                          T, int64_t, uint64_t, int32_t, uint32_t, int16_t, uint16_t,
+                          uint8_t, int8_t, float, double, bool>>
+static inline T FromLittleEndian(T value) {
+  return value;
+}
+#else
+template <typename T, typename = internal::EnableIfIsOneOf<
+                          T, int64_t, uint64_t, int32_t, uint32_t, int16_t, uint16_t,
+                          uint8_t, int8_t, float, double, bool>>
+static inline T FromBigEndian(T value) {
+  return value;
+}
+
+template <typename T, typename = internal::EnableIfIsOneOf<
+                          T, int64_t, uint64_t, int32_t, uint32_t, int16_t, uint16_t,
+                          uint8_t, int8_t, float, double, bool>>
+static inline T FromLittleEndian(T value) {
+  return ByteSwap(value);
+}
+#endif
+
+// Handle endianness in *word* granularity (keep individual array element untouched)
+namespace little_endian {
+
+namespace detail {
+
+// Read a native endian array as little endian
+template <typename T, size_t N>
+struct Reader {
+  const std::array<T, N>& native_array;
+
+  explicit Reader(const std::array<T, N>& native_array) : native_array(native_array) {}
+
+  const T& operator[](size_t i) const {
+    return native_array[ARROW_LITTLE_ENDIAN ? i : N - 1 - i];
+  }
+};
+
+// Read/write a native endian array as little endian
+template <typename T, size_t N>
+struct Writer {
+  std::array<T, N>* native_array;
+
+  explicit Writer(std::array<T, N>* native_array) : native_array(native_array) {}
+
+  const T& operator[](size_t i) const {
+    return (*native_array)[ARROW_LITTLE_ENDIAN ? i : N - 1 - i];
+  }
+  T& operator[](size_t i) { return (*native_array)[ARROW_LITTLE_ENDIAN ? i : N - 1 - i]; }
+};
+
+}  // namespace detail
+
+// Construct array reader and try to deduce template augments
+template <typename T, size_t N>
+static inline detail::Reader<T, N> Make(const std::array<T, N>& native_array) {
+  return detail::Reader<T, N>(native_array);
+}
+
+// Construct array writer and try to deduce template augments
+template <typename T, size_t N>
+static inline detail::Writer<T, N> Make(std::array<T, N>* native_array) {
+  return detail::Writer<T, N>(native_array);
+}
+
+// Convert little endian array to native endian
+template <typename T, size_t N>
+static inline std::array<T, N> ToNative(std::array<T, N> array) {
+  if (!ARROW_LITTLE_ENDIAN) {
+    std::reverse(array.begin(), array.end());
+  }
+  return array;
+}
+
+// Convert native endian array to little endian
+template <typename T, size_t N>
+static inline std::array<T, N> FromNative(std::array<T, N> array) {
+  return ToNative(array);
+}
+
+}  // namespace little_endian
+
+}  // namespace bit_util
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/float16.h

@@ -0,0 +1,206 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <array>
+#include <cstdint>
+#include <cstring>
+#include <iosfwd>
+#include <limits>
+#include <type_traits>
+
+#include "arrow/util/endian.h"
+#include "arrow/util/macros.h"
+#include "arrow/util/ubsan.h"
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+namespace util {
+
+/// \brief Class representing an IEEE half-precision float, encoded as a `uint16_t`
+///
+/// The exact format is as follows (from LSB to MSB):
+/// - bits 0-10:  mantissa
+/// - bits 10-15: exponent
+/// - bit 15:     sign
+///
+class ARROW_EXPORT Float16 {
+ public:
+  Float16() = default;
+  explicit Float16(float f) : Float16(FromFloat(f)) {}
+  explicit Float16(double d) : Float16(FromDouble(d)) {}
+  template <typename T,
+            typename std::enable_if_t<std::is_convertible_v<T, double>>* = NULLPTR>
+  explicit Float16(T v) : Float16(static_cast<double>(v)) {}
+
+  /// \brief Create a `Float16` from its exact binary representation
+  constexpr static Float16 FromBits(uint16_t bits) { return Float16{bits, bool{}}; }
+  /// \brief Create a `Float16` from a 32-bit float (may lose precision)
+  static Float16 FromFloat(float f);
+  /// \brief Create a `Float16` from a 64-bit float (may lose precision)
+  static Float16 FromDouble(double d);
+
+  /// \brief Read a `Float16` from memory in native-endian byte order
+  static Float16 FromBytes(const uint8_t* src) {
+    return FromBits(SafeLoadAs<uint16_t>(src));
+  }
+
+  /// \brief Read a `Float16` from memory in little-endian byte order
+  static Float16 FromLittleEndian(const uint8_t* src) {
+    return FromBits(::arrow::bit_util::FromLittleEndian(SafeLoadAs<uint16_t>(src)));
+  }
+
+  /// \brief Read a `Float16` from memory in big-endian byte order
+  static Float16 FromBigEndian(const uint8_t* src) {
+    return FromBits(::arrow::bit_util::FromBigEndian(SafeLoadAs<uint16_t>(src)));
+  }
+
+  /// \brief Return the value's binary representation as a `uint16_t`
+  constexpr uint16_t bits() const { return bits_; }
+
+  /// \brief Return true if the value is negative (sign bit is set)
+  constexpr bool signbit() const { return (bits_ & 0x8000) != 0; }
+
+  /// \brief Return true if the value is NaN
+  constexpr bool is_nan() const { return (bits_ & 0x7fff) > 0x7c00; }
+  /// \brief Return true if the value is positive/negative infinity
+  constexpr bool is_infinity() const { return (bits_ & 0x7fff) == 0x7c00; }
+  /// \brief Return true if the value is finite and not NaN
+  constexpr bool is_finite() const { return (bits_ & 0x7c00) != 0x7c00; }
+  /// \brief Return true if the value is positive/negative zero
+  constexpr bool is_zero() const { return (bits_ & 0x7fff) == 0; }
+
+  /// \brief Convert to a 32-bit float
+  float ToFloat() const;
+  /// \brief Convert to a 64-bit float
+  double ToDouble() const;
+
+  explicit operator float() const { return ToFloat(); }
+  explicit operator double() const { return ToDouble(); }
+
+  /// \brief Copy the value's bytes in native-endian byte order
+  void ToBytes(uint8_t* dest) const { std::memcpy(dest, &bits_, sizeof(bits_)); }
+  /// \brief Return the value's bytes in native-endian byte order
+  constexpr std::array<uint8_t, 2> ToBytes() const {
+#if ARROW_LITTLE_ENDIAN
+    return ToLittleEndian();
+#else
+    return ToBigEndian();
+#endif
+  }
+
+  /// \brief Copy the value's bytes in little-endian byte order
+  void ToLittleEndian(uint8_t* dest) const {
+    const auto bytes = ToLittleEndian();
+    std::memcpy(dest, bytes.data(), bytes.size());
+  }
+  /// \brief Return the value's bytes in little-endian byte order
+  constexpr std::array<uint8_t, 2> ToLittleEndian() const {
+    return {uint8_t(bits_ & 0xff), uint8_t(bits_ >> 8)};
+  }
+
+  /// \brief Copy the value's bytes in big-endian byte order
+  void ToBigEndian(uint8_t* dest) const {
+    const auto bytes = ToBigEndian();
+    std::memcpy(dest, bytes.data(), bytes.size());
+  }
+  /// \brief Return the value's bytes in big-endian byte order
+  constexpr std::array<uint8_t, 2> ToBigEndian() const {
+    return {uint8_t(bits_ >> 8), uint8_t(bits_ & 0xff)};
+  }
+
+  constexpr Float16 operator-() const { return FromBits(bits_ ^ 0x8000); }
+  constexpr Float16 operator+() const { return FromBits(bits_); }
+
+  friend constexpr bool operator==(Float16 lhs, Float16 rhs) {
+    if (lhs.is_nan() || rhs.is_nan()) return false;
+    return Float16::CompareEq(lhs, rhs);
+  }
+  friend constexpr bool operator!=(Float16 lhs, Float16 rhs) { return !(lhs == rhs); }
+
+  friend constexpr bool operator<(Float16 lhs, Float16 rhs) {
+    if (lhs.is_nan() || rhs.is_nan()) return false;
+    return Float16::CompareLt(lhs, rhs);
+  }
+  friend constexpr bool operator>(Float16 lhs, Float16 rhs) { return rhs < lhs; }
+
+  friend constexpr bool operator<=(Float16 lhs, Float16 rhs) {
+    if (lhs.is_nan() || rhs.is_nan()) return false;
+    return !Float16::CompareLt(rhs, lhs);
+  }
+  friend constexpr bool operator>=(Float16 lhs, Float16 rhs) { return rhs <= lhs; }
+
+  ARROW_FRIEND_EXPORT friend std::ostream& operator<<(std::ostream& os, Float16 arg);
+
+  static constexpr Float16 zero() { return FromBits(0); }
+  static constexpr Float16 one() { return FromBits(0x3c00); }
+
+ protected:
+  uint16_t bits_;
+
+ private:
+  constexpr Float16(uint16_t bits, bool) : bits_(bits) {}
+
+  // Comparison helpers that assume neither operand is NaN
+  static constexpr bool CompareEq(Float16 lhs, Float16 rhs) {
+    return (lhs.bits() == rhs.bits()) || (lhs.is_zero() && rhs.is_zero());
+  }
+  static constexpr bool CompareLt(Float16 lhs, Float16 rhs) {
+    if (lhs.signbit()) {
+      if (rhs.signbit()) {
+        // Both are negative
+        return lhs.bits() > rhs.bits();
+      } else {
+        // Handle +/-0
+        return !lhs.is_zero() || rhs.bits() != 0;
+      }
+    } else if (rhs.signbit()) {
+      return false;
+    } else {
+      // Both are positive
+      return lhs.bits() < rhs.bits();
+    }
+  }
+};
+
+static_assert(std::is_standard_layout_v<Float16>);
+static_assert(std::is_trivial_v<Float16>);
+static_assert(sizeof(Float16) == sizeof(uint16_t));
+
+}  // namespace util
+}  // namespace arrow
+
+// TODO: Not complete
+template <>
+class std::numeric_limits<arrow::util::Float16> {
+  using T = arrow::util::Float16;
+
+ public:
+  static constexpr bool is_specialized = true;
+  static constexpr bool is_signed = true;
+  static constexpr bool has_infinity = true;
+  static constexpr bool has_quiet_NaN = true;
+
+  static constexpr T min() { return T::FromBits(0b0000010000000000); }
+  static constexpr T max() { return T::FromBits(0b0111101111111111); }
+  static constexpr T lowest() { return -max(); }
+
+  static constexpr T infinity() { return T::FromBits(0b0111110000000000); }
+
+  static constexpr T quiet_NaN() { return T::FromBits(0b0111111111111111); }
+};

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/formatting.h

@@ -0,0 +1,667 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+// This is a private header for number-to-string formatting utilities
+
+#pragma once
+
+#include <array>
+#include <cassert>
+#include <chrono>
+#include <limits>
+#include <memory>
+#include <string>
+#include <string_view>
+#include <type_traits>
+#include <utility>
+
+#include "arrow/status.h"
+#include "arrow/type_fwd.h"
+#include "arrow/type_traits.h"
+#include "arrow/util/macros.h"
+#include "arrow/util/string.h"
+#include "arrow/util/time.h"
+#include "arrow/util/visibility.h"
+#include "arrow/vendored/datetime.h"
+
+namespace arrow {
+namespace internal {
+
+/// \brief The entry point for conversion to strings.
+template <typename ARROW_TYPE, typename Enable = void>
+class StringFormatter;
+
+template <typename T>
+struct is_formattable {
+  template <typename U, typename = typename StringFormatter<U>::value_type>
+  static std::true_type Test(U*);
+
+  template <typename U>
+  static std::false_type Test(...);
+
+  static constexpr bool value = decltype(Test<T>(NULLPTR))::value;
+};
+
+template <typename T, typename R = void>
+using enable_if_formattable = enable_if_t<is_formattable<T>::value, R>;
+
+template <typename Appender>
+using Return = decltype(std::declval<Appender>()(std::string_view{}));
+
+/////////////////////////////////////////////////////////////////////////
+// Boolean formatting
+
+template <>
+class StringFormatter<BooleanType> {
+ public:
+  explicit StringFormatter(const DataType* = NULLPTR) {}
+
+  using value_type = bool;
+
+  template <typename Appender>
+  Return<Appender> operator()(bool value, Appender&& append) {
+    if (value) {
+      const char string[] = "true";
+      return append(std::string_view(string));
+    } else {
+      const char string[] = "false";
+      return append(std::string_view(string));
+    }
+  }
+};
+
+/////////////////////////////////////////////////////////////////////////
+// Decimals formatting
+
+template <typename ARROW_TYPE>
+class DecimalToStringFormatterMixin {
+ public:
+  explicit DecimalToStringFormatterMixin(const DataType* type)
+      : scale_(static_cast<const ARROW_TYPE*>(type)->scale()) {}
+
+  using value_type = typename TypeTraits<ARROW_TYPE>::CType;
+
+  template <typename Appender>
+  Return<Appender> operator()(const value_type& value, Appender&& append) {
+    return append(value.ToString(scale_));
+  }
+
+ private:
+  int32_t scale_;
+};
+
+template <>
+class StringFormatter<Decimal32Type>
+    : public DecimalToStringFormatterMixin<Decimal32Type> {
+  using DecimalToStringFormatterMixin::DecimalToStringFormatterMixin;
+};
+
+template <>
+class StringFormatter<Decimal64Type>
+    : public DecimalToStringFormatterMixin<Decimal64Type> {
+  using DecimalToStringFormatterMixin::DecimalToStringFormatterMixin;
+};
+
+template <>
+class StringFormatter<Decimal128Type>
+    : public DecimalToStringFormatterMixin<Decimal128Type> {
+  using DecimalToStringFormatterMixin::DecimalToStringFormatterMixin;
+};
+
+template <>
+class StringFormatter<Decimal256Type>
+    : public DecimalToStringFormatterMixin<Decimal256Type> {
+  using DecimalToStringFormatterMixin::DecimalToStringFormatterMixin;
+};
+
+/////////////////////////////////////////////////////////////////////////
+// Integer formatting
+
+namespace detail {
+
+// A 2x100 direct table mapping integers in [0..99] to their decimal representations.
+ARROW_EXPORT extern const char digit_pairs[];
+
+// Based on fmtlib's format_int class:
+// Write digits from right to left into a stack allocated buffer.
+// \pre *cursor points to the byte after the one that will be written.
+// \post *cursor points to the byte that was written.
+inline void FormatOneChar(char c, char** cursor) { *(--(*cursor)) = c; }
+
+template <typename Int>
+void FormatOneDigit(Int value, char** cursor) {
+  assert(value >= 0 && value <= 9);
+  FormatOneChar(static_cast<char>('0' + value), cursor);
+}
+
+// GH-35662: I don't know why but the following combination causes SEGV:
+// * template implementation without inline
+// * MinGW
+// * Release build
+template <typename Int>
+inline void FormatTwoDigits(Int value, char** cursor) {
+  assert(value >= 0 && value <= 99);
+  auto digit_pair = &digit_pairs[value * 2];
+  FormatOneChar(digit_pair[1], cursor);
+  FormatOneChar(digit_pair[0], cursor);
+}
+
+template <typename Int>
+void FormatAllDigits(Int value, char** cursor) {
+  assert(value >= 0);
+  while (value >= 100) {
+    FormatTwoDigits(value % 100, cursor);
+    value /= 100;
+  }
+
+  if (value >= 10) {
+    FormatTwoDigits(value, cursor);
+  } else {
+    FormatOneDigit(value, cursor);
+  }
+}
+
+template <typename Int>
+void FormatAllDigitsLeftPadded(Int value, size_t pad, char pad_char, char** cursor) {
+  auto end = *cursor - pad;
+  FormatAllDigits(value, cursor);
+  while (*cursor > end) {
+    FormatOneChar(pad_char, cursor);
+  }
+}
+
+template <size_t BUFFER_SIZE>
+std::string_view ViewDigitBuffer(const std::array<char, BUFFER_SIZE>& buffer,
+                                 char* cursor) {
+  auto buffer_end = buffer.data() + BUFFER_SIZE;
+  return {cursor, static_cast<size_t>(buffer_end - cursor)};
+}
+
+template <typename Int, typename UInt = typename std::make_unsigned<Int>::type>
+constexpr UInt Abs(Int value) {
+  return value < 0 ? ~static_cast<UInt>(value) + 1 : static_cast<UInt>(value);
+}
+
+template <typename Int>
+constexpr size_t Digits10(Int value) {
+  return value <= 9 ? 1 : Digits10(value / 10) + 1;
+}
+
+}  // namespace detail
+
+template <typename ARROW_TYPE>
+class IntToStringFormatterMixin {
+ public:
+  explicit IntToStringFormatterMixin(const DataType* = NULLPTR) {}
+
+  using value_type = typename ARROW_TYPE::c_type;
+
+  template <typename Appender>
+  Return<Appender> operator()(value_type value, Appender&& append) {
+    constexpr size_t buffer_size =
+        detail::Digits10(std::numeric_limits<value_type>::max()) + 1;
+
+    std::array<char, buffer_size> buffer;
+    char* cursor = buffer.data() + buffer_size;
+    detail::FormatAllDigits(detail::Abs(value), &cursor);
+    if (value < 0) {
+      detail::FormatOneChar('-', &cursor);
+    }
+    return append(detail::ViewDigitBuffer(buffer, cursor));
+  }
+};
+
+template <>
+class StringFormatter<Int8Type> : public IntToStringFormatterMixin<Int8Type> {
+  using IntToStringFormatterMixin::IntToStringFormatterMixin;
+};
+
+template <>
+class StringFormatter<Int16Type> : public IntToStringFormatterMixin<Int16Type> {
+  using IntToStringFormatterMixin::IntToStringFormatterMixin;
+};
+
+template <>
+class StringFormatter<Int32Type> : public IntToStringFormatterMixin<Int32Type> {
+  using IntToStringFormatterMixin::IntToStringFormatterMixin;
+};
+
+template <>
+class StringFormatter<Int64Type> : public IntToStringFormatterMixin<Int64Type> {
+  using IntToStringFormatterMixin::IntToStringFormatterMixin;
+};
+
+template <>
+class StringFormatter<UInt8Type> : public IntToStringFormatterMixin<UInt8Type> {
+  using IntToStringFormatterMixin::IntToStringFormatterMixin;
+};
+
+template <>
+class StringFormatter<UInt16Type> : public IntToStringFormatterMixin<UInt16Type> {
+  using IntToStringFormatterMixin::IntToStringFormatterMixin;
+};
+
+template <>
+class StringFormatter<UInt32Type> : public IntToStringFormatterMixin<UInt32Type> {
+  using IntToStringFormatterMixin::IntToStringFormatterMixin;
+};
+
+template <>
+class StringFormatter<UInt64Type> : public IntToStringFormatterMixin<UInt64Type> {
+  using IntToStringFormatterMixin::IntToStringFormatterMixin;
+};
+
+/////////////////////////////////////////////////////////////////////////
+// Floating-point formatting
+
+class ARROW_EXPORT FloatToStringFormatter {
+ public:
+  FloatToStringFormatter();
+  FloatToStringFormatter(int flags, const char* inf_symbol, const char* nan_symbol,
+                         char exp_character, int decimal_in_shortest_low,
+                         int decimal_in_shortest_high,
+                         int max_leading_padding_zeroes_in_precision_mode,
+                         int max_trailing_padding_zeroes_in_precision_mode);
+  ~FloatToStringFormatter();
+
+  // Returns the number of characters written
+  int FormatFloat(float v, char* out_buffer, int out_size);
+  int FormatFloat(double v, char* out_buffer, int out_size);
+  int FormatFloat(uint16_t v, char* out_buffer, int out_size);
+
+ protected:
+  struct Impl;
+  std::unique_ptr<Impl> impl_;
+};
+
+template <typename ARROW_TYPE>
+class FloatToStringFormatterMixin : public FloatToStringFormatter {
+ public:
+  using value_type = typename ARROW_TYPE::c_type;
+
+  static constexpr int buffer_size = 50;
+
+  explicit FloatToStringFormatterMixin(const DataType* = NULLPTR) {}
+
+  FloatToStringFormatterMixin(int flags, const char* inf_symbol, const char* nan_symbol,
+                              char exp_character, int decimal_in_shortest_low,
+                              int decimal_in_shortest_high,
+                              int max_leading_padding_zeroes_in_precision_mode,
+                              int max_trailing_padding_zeroes_in_precision_mode)
+      : FloatToStringFormatter(flags, inf_symbol, nan_symbol, exp_character,
+                               decimal_in_shortest_low, decimal_in_shortest_high,
+                               max_leading_padding_zeroes_in_precision_mode,
+                               max_trailing_padding_zeroes_in_precision_mode) {}
+
+  template <typename Appender>
+  Return<Appender> operator()(value_type value, Appender&& append) {
+    char buffer[buffer_size];
+    int size = FormatFloat(value, buffer, buffer_size);
+    return append(std::string_view(buffer, size));
+  }
+};
+
+template <>
+class StringFormatter<HalfFloatType> : public FloatToStringFormatterMixin<HalfFloatType> {
+ public:
+  using FloatToStringFormatterMixin::FloatToStringFormatterMixin;
+};
+
+template <>
+class StringFormatter<FloatType> : public FloatToStringFormatterMixin<FloatType> {
+ public:
+  using FloatToStringFormatterMixin::FloatToStringFormatterMixin;
+};
+
+template <>
+class StringFormatter<DoubleType> : public FloatToStringFormatterMixin<DoubleType> {
+ public:
+  using FloatToStringFormatterMixin::FloatToStringFormatterMixin;
+};
+
+/////////////////////////////////////////////////////////////////////////
+// Temporal formatting
+
+namespace detail {
+
+constexpr size_t BufferSizeYYYY_MM_DD() {
+  // "-"? "99999-12-31"
+  return 1 + detail::Digits10(99999) + 1 + detail::Digits10(12) + 1 +
+         detail::Digits10(31);
+}
+
+inline void FormatYYYY_MM_DD(arrow_vendored::date::year_month_day ymd, char** cursor) {
+  FormatTwoDigits(static_cast<unsigned>(ymd.day()), cursor);
+  FormatOneChar('-', cursor);
+  FormatTwoDigits(static_cast<unsigned>(ymd.month()), cursor);
+  FormatOneChar('-', cursor);
+  auto year = static_cast<int>(ymd.year());
+  const auto is_neg_year = year < 0;
+  year = std::abs(year);
+  assert(year <= 99999);
+  FormatTwoDigits(year % 100, cursor);
+  year /= 100;
+  FormatTwoDigits(year % 100, cursor);
+  if (year >= 100) {
+    FormatOneDigit(year / 100, cursor);
+  }
+  if (is_neg_year) {
+    FormatOneChar('-', cursor);
+  }
+}
+
+template <typename Duration>
+constexpr size_t BufferSizeHH_MM_SS() {
+  // "23:59:59" ("." "9"+)?
+  return detail::Digits10(23) + 1 + detail::Digits10(59) + 1 + detail::Digits10(59) + 1 +
+         detail::Digits10(Duration::period::den) - 1;
+}
+
+template <typename Duration>
+void FormatHH_MM_SS(arrow_vendored::date::hh_mm_ss<Duration> hms, char** cursor) {
+  constexpr size_t subsecond_digits = Digits10(Duration::period::den) - 1;
+  if (subsecond_digits != 0) {
+    FormatAllDigitsLeftPadded(hms.subseconds().count(), subsecond_digits, '0', cursor);
+    FormatOneChar('.', cursor);
+  }
+  FormatTwoDigits(hms.seconds().count(), cursor);
+  FormatOneChar(':', cursor);
+  FormatTwoDigits(hms.minutes().count(), cursor);
+  FormatOneChar(':', cursor);
+  FormatTwoDigits(hms.hours().count(), cursor);
+}
+
+// Some out-of-bound datetime values would result in erroneous printing
+// because of silent integer wraparound in the `arrow_vendored::date` library.
+//
+// To avoid such misprinting, we must therefore check the bounds explicitly.
+// The bounds correspond to start of year -32767 and end of year 32767,
+// respectively (-32768 is an invalid year value in `arrow_vendored::date`).
+//
+// Note these values are the same as documented for C++20:
+// https://en.cppreference.com/w/cpp/chrono/year_month_day/operator_days
+template <typename Unit>
+bool IsDateTimeInRange(Unit duration) {
+  constexpr Unit kMinIncl =
+      std::chrono::duration_cast<Unit>(arrow_vendored::date::days{-12687428});
+  constexpr Unit kMaxExcl =
+      std::chrono::duration_cast<Unit>(arrow_vendored::date::days{11248738});
+  return duration >= kMinIncl && duration < kMaxExcl;
+}
+
+// IsDateTimeInRange() specialization for nanoseconds: a 64-bit number of
+// nanoseconds cannot represent years outside of the [-32767, 32767]
+// range, and the {kMinIncl, kMaxExcl} constants above would overflow.
+constexpr bool IsDateTimeInRange(std::chrono::nanoseconds duration) { return true; }
+
+template <typename Unit>
+bool IsTimeInRange(Unit duration) {
+  constexpr Unit kMinIncl = std::chrono::duration_cast<Unit>(std::chrono::seconds{0});
+  constexpr Unit kMaxExcl = std::chrono::duration_cast<Unit>(std::chrono::seconds{86400});
+  return duration >= kMinIncl && duration < kMaxExcl;
+}
+
+template <typename RawValue, typename Appender>
+Return<Appender> FormatOutOfRange(RawValue&& raw_value, Appender&& append) {
+  // XXX locale-sensitive but good enough for now
+  std::string formatted = "<value out of range: " + ToChars(raw_value) + ">";
+  return append(std::move(formatted));
+}
+
+const auto kEpoch = arrow_vendored::date::sys_days{arrow_vendored::date::jan / 1 / 1970};
+
+}  // namespace detail
+
+template <>
+class StringFormatter<DurationType> : public IntToStringFormatterMixin<DurationType> {
+  using IntToStringFormatterMixin::IntToStringFormatterMixin;
+};
+
+class DateToStringFormatterMixin {
+ public:
+  explicit DateToStringFormatterMixin(const DataType* = NULLPTR) {}
+
+ protected:
+  template <typename Appender>
+  Return<Appender> FormatDays(arrow_vendored::date::days since_epoch, Appender&& append) {
+    arrow_vendored::date::sys_days timepoint_days{since_epoch};
+
+    constexpr size_t buffer_size = detail::BufferSizeYYYY_MM_DD();
+
+    std::array<char, buffer_size> buffer;
+    char* cursor = buffer.data() + buffer_size;
+
+    detail::FormatYYYY_MM_DD(arrow_vendored::date::year_month_day{timepoint_days},
+                             &cursor);
+    return append(detail::ViewDigitBuffer(buffer, cursor));
+  }
+};
+
+template <>
+class StringFormatter<Date32Type> : public DateToStringFormatterMixin {
+ public:
+  using value_type = typename Date32Type::c_type;
+
+  using DateToStringFormatterMixin::DateToStringFormatterMixin;
+
+  template <typename Appender>
+  Return<Appender> operator()(value_type value, Appender&& append) {
+    const auto since_epoch = arrow_vendored::date::days{value};
+    if (!ARROW_PREDICT_TRUE(detail::IsDateTimeInRange(since_epoch))) {
+      return detail::FormatOutOfRange(value, append);
+    }
+    return FormatDays(since_epoch, std::forward<Appender>(append));
+  }
+};
+
+template <>
+class StringFormatter<Date64Type> : public DateToStringFormatterMixin {
+ public:
+  using value_type = typename Date64Type::c_type;
+
+  using DateToStringFormatterMixin::DateToStringFormatterMixin;
+
+  template <typename Appender>
+  Return<Appender> operator()(value_type value, Appender&& append) {
+    const auto since_epoch = std::chrono::milliseconds{value};
+    if (!ARROW_PREDICT_TRUE(detail::IsDateTimeInRange(since_epoch))) {
+      return detail::FormatOutOfRange(value, append);
+    }
+    return FormatDays(std::chrono::duration_cast<arrow_vendored::date::days>(since_epoch),
+                      std::forward<Appender>(append));
+  }
+};
+
+template <>
+class StringFormatter<TimestampType> {
+ public:
+  using value_type = int64_t;
+
+  explicit StringFormatter(const DataType* type)
+      : unit_(checked_cast<const TimestampType&>(*type).unit()),
+        timezone_(checked_cast<const TimestampType&>(*type).timezone()) {}
+
+  template <typename Duration, typename Appender>
+  Return<Appender> operator()(Duration, value_type value, Appender&& append) {
+    using arrow_vendored::date::days;
+
+    const Duration since_epoch{value};
+    if (!ARROW_PREDICT_TRUE(detail::IsDateTimeInRange(since_epoch))) {
+      return detail::FormatOutOfRange(value, append);
+    }
+
+    const auto timepoint = detail::kEpoch + since_epoch;
+    // Round days towards zero
+    // (the naive approach of using arrow_vendored::date::floor() would
+    //  result in UB for very large negative timestamps, similarly as
+    //  https://github.com/HowardHinnant/date/issues/696)
+    auto timepoint_days = std::chrono::time_point_cast<days>(timepoint);
+    Duration since_midnight;
+    if (timepoint_days <= timepoint) {
+      // Year >= 1970
+      since_midnight = timepoint - timepoint_days;
+    } else {
+      // Year < 1970
+      since_midnight = days(1) - (timepoint_days - timepoint);
+      timepoint_days -= days(1);
+    }
+
+    // YYYY_MM_DD " " HH_MM_SS "Z"?
+    constexpr size_t buffer_size =
+        detail::BufferSizeYYYY_MM_DD() + 1 + detail::BufferSizeHH_MM_SS<Duration>() + 1;
+
+    std::array<char, buffer_size> buffer;
+    char* cursor = buffer.data() + buffer_size;
+
+    if (timezone_.size() > 0) {
+      detail::FormatOneChar('Z', &cursor);
+    }
+    detail::FormatHH_MM_SS(arrow_vendored::date::make_time(since_midnight), &cursor);
+    detail::FormatOneChar(' ', &cursor);
+    detail::FormatYYYY_MM_DD(timepoint_days, &cursor);
+    return append(detail::ViewDigitBuffer(buffer, cursor));
+  }
+
+  template <typename Appender>
+  Return<Appender> operator()(value_type value, Appender&& append) {
+    return util::VisitDuration(unit_, *this, value, std::forward<Appender>(append));
+  }
+
+ private:
+  TimeUnit::type unit_;
+  std::string timezone_;
+};
+
+template <typename T>
+class StringFormatter<T, enable_if_time<T>> {
+ public:
+  using value_type = typename T::c_type;
+
+  explicit StringFormatter(const DataType* type)
+      : unit_(checked_cast<const T&>(*type).unit()) {}
+
+  template <typename Duration, typename Appender>
+  Return<Appender> operator()(Duration, value_type count, Appender&& append) {
+    const Duration since_midnight{count};
+    if (!ARROW_PREDICT_TRUE(detail::IsTimeInRange(since_midnight))) {
+      return detail::FormatOutOfRange(count, append);
+    }
+
+    constexpr size_t buffer_size = detail::BufferSizeHH_MM_SS<Duration>();
+
+    std::array<char, buffer_size> buffer;
+    char* cursor = buffer.data() + buffer_size;
+
+    detail::FormatHH_MM_SS(arrow_vendored::date::make_time(since_midnight), &cursor);
+    return append(detail::ViewDigitBuffer(buffer, cursor));
+  }
+
+  template <typename Appender>
+  Return<Appender> operator()(value_type value, Appender&& append) {
+    return util::VisitDuration(unit_, *this, value, std::forward<Appender>(append));
+  }
+
+ private:
+  TimeUnit::type unit_;
+};
+
+template <>
+class StringFormatter<MonthIntervalType> {
+ public:
+  using value_type = MonthIntervalType::c_type;
+
+  explicit StringFormatter(const DataType*) {}
+
+  template <typename Appender>
+  Return<Appender> operator()(value_type interval, Appender&& append) {
+    constexpr size_t buffer_size =
+        /*'m'*/ 3 + /*negative signs*/ 1 +
+        /*months*/ detail::Digits10(std::numeric_limits<value_type>::max());
+    std::array<char, buffer_size> buffer;
+    char* cursor = buffer.data() + buffer_size;
+
+    detail::FormatOneChar('M', &cursor);
+    detail::FormatAllDigits(detail::Abs(interval), &cursor);
+    if (interval < 0) detail::FormatOneChar('-', &cursor);
+
+    return append(detail::ViewDigitBuffer(buffer, cursor));
+  }
+};
+
+template <>
+class StringFormatter<DayTimeIntervalType> {
+ public:
+  using value_type = DayTimeIntervalType::DayMilliseconds;
+
+  explicit StringFormatter(const DataType*) {}
+
+  template <typename Appender>
+  Return<Appender> operator()(value_type interval, Appender&& append) {
+    constexpr size_t buffer_size =
+        /*d, ms*/ 3 + /*negative signs*/ 2 +
+        /*days/milliseconds*/ 2 * detail::Digits10(std::numeric_limits<int32_t>::max());
+    std::array<char, buffer_size> buffer;
+    char* cursor = buffer.data() + buffer_size;
+
+    detail::FormatOneChar('s', &cursor);
+    detail::FormatOneChar('m', &cursor);
+    detail::FormatAllDigits(detail::Abs(interval.milliseconds), &cursor);
+    if (interval.milliseconds < 0) detail::FormatOneChar('-', &cursor);
+
+    detail::FormatOneChar('d', &cursor);
+    detail::FormatAllDigits(detail::Abs(interval.days), &cursor);
+    if (interval.days < 0) detail::FormatOneChar('-', &cursor);
+
+    return append(detail::ViewDigitBuffer(buffer, cursor));
+  }
+};
+
+template <>
+class StringFormatter<MonthDayNanoIntervalType> {
+ public:
+  using value_type = MonthDayNanoIntervalType::MonthDayNanos;
+
+  explicit StringFormatter(const DataType*) {}
+
+  template <typename Appender>
+  Return<Appender> operator()(value_type interval, Appender&& append) {
+    constexpr size_t buffer_size =
+        /*m, d, ns*/ 4 + /*negative signs*/ 3 +
+        /*months/days*/ 2 * detail::Digits10(std::numeric_limits<int32_t>::max()) +
+        /*nanoseconds*/ detail::Digits10(std::numeric_limits<int64_t>::max());
+    std::array<char, buffer_size> buffer;
+    char* cursor = buffer.data() + buffer_size;
+
+    detail::FormatOneChar('s', &cursor);
+    detail::FormatOneChar('n', &cursor);
+    detail::FormatAllDigits(detail::Abs(interval.nanoseconds), &cursor);
+    if (interval.nanoseconds < 0) detail::FormatOneChar('-', &cursor);
+
+    detail::FormatOneChar('d', &cursor);
+    detail::FormatAllDigits(detail::Abs(interval.days), &cursor);
+    if (interval.days < 0) detail::FormatOneChar('-', &cursor);
+
+    detail::FormatOneChar('M', &cursor);
+    detail::FormatAllDigits(detail::Abs(interval.months), &cursor);
+    if (interval.months < 0) detail::FormatOneChar('-', &cursor);
+
+    return append(detail::ViewDigitBuffer(buffer, cursor));
+  }
+};
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/functional.h

@@ -0,0 +1,160 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <memory>
+#include <tuple>
+#include <type_traits>
+
+#include "arrow/result.h"
+#include "arrow/util/macros.h"
+
+namespace arrow {
+namespace internal {
+
+struct Empty {
+  static Result<Empty> ToResult(Status s) {
+    if (ARROW_PREDICT_TRUE(s.ok())) {
+      return Empty{};
+    }
+    return s;
+  }
+};
+
+/// Helper struct for examining lambdas and other callables.
+/// TODO(ARROW-12655) support function pointers
+struct call_traits {
+ public:
+  template <typename R, typename... A>
+  static std::false_type is_overloaded_impl(R(A...));
+
+  template <typename F>
+  static std::false_type is_overloaded_impl(decltype(&F::operator())*);
+
+  template <typename F>
+  static std::true_type is_overloaded_impl(...);
+
+  template <typename F, typename R, typename... A>
+  static R return_type_impl(R (F::*)(A...));
+
+  template <typename F, typename R, typename... A>
+  static R return_type_impl(R (F::*)(A...) const);
+
+  template <std::size_t I, typename F, typename R, typename... A>
+  static typename std::tuple_element<I, std::tuple<A...>>::type argument_type_impl(
+      R (F::*)(A...));
+
+  template <std::size_t I, typename F, typename R, typename... A>
+  static typename std::tuple_element<I, std::tuple<A...>>::type argument_type_impl(
+      R (F::*)(A...) const);
+
+  template <std::size_t I, typename F, typename R, typename... A>
+  static typename std::tuple_element<I, std::tuple<A...>>::type argument_type_impl(
+      R (F::*)(A...) &&);
+
+  template <typename F, typename R, typename... A>
+  static std::integral_constant<int, sizeof...(A)> argument_count_impl(R (F::*)(A...));
+
+  template <typename F, typename R, typename... A>
+  static std::integral_constant<int, sizeof...(A)> argument_count_impl(R (F::*)(A...)
+                                                                           const);
+
+  template <typename F, typename R, typename... A>
+  static std::integral_constant<int, sizeof...(A)> argument_count_impl(R (F::*)(A...) &&);
+
+  /// bool constant indicating whether F is a callable with more than one possible
+  /// signature. Will be true_type for objects which define multiple operator() or which
+  /// define a template operator()
+  template <typename F>
+  using is_overloaded =
+      decltype(is_overloaded_impl<typename std::decay<F>::type>(NULLPTR));
+
+  template <typename F, typename T = void>
+  using enable_if_overloaded = typename std::enable_if<is_overloaded<F>::value, T>::type;
+
+  template <typename F, typename T = void>
+  using disable_if_overloaded =
+      typename std::enable_if<!is_overloaded<F>::value, T>::type;
+
+  /// If F is not overloaded, the argument types of its call operator can be
+  /// extracted via call_traits::argument_type<Index, F>
+  template <std::size_t I, typename F>
+  using argument_type = decltype(argument_type_impl<I>(&std::decay<F>::type::operator()));
+
+  template <typename F>
+  using argument_count = decltype(argument_count_impl(&std::decay<F>::type::operator()));
+
+  template <typename F>
+  using return_type = decltype(return_type_impl(&std::decay<F>::type::operator()));
+
+  template <typename F, typename T, typename RT = T>
+  using enable_if_return =
+      typename std::enable_if<std::is_same<return_type<F>, T>::value, RT>;
+
+  template <typename T, typename R = void>
+  using enable_if_empty = typename std::enable_if<std::is_same<T, Empty>::value, R>::type;
+
+  template <typename T, typename R = void>
+  using enable_if_not_empty =
+      typename std::enable_if<!std::is_same<T, Empty>::value, R>::type;
+};
+
+/// A type erased callable object which may only be invoked once.
+/// It can be constructed from any lambda which matches the provided call signature.
+/// Invoking it results in destruction of the lambda, freeing any state/references
+/// immediately. Invoking a default constructed FnOnce or one which has already been
+/// invoked will segfault.
+template <typename Signature>
+class FnOnce;
+
+template <typename R, typename... A>
+class FnOnce<R(A...)> {
+ public:
+  FnOnce() = default;
+
+  template <typename Fn,
+            typename = typename std::enable_if<std::is_convertible<
+                decltype(std::declval<Fn&&>()(std::declval<A>()...)), R>::value>::type>
+  FnOnce(Fn fn) : impl_(new FnImpl<Fn>(std::move(fn))) {  // NOLINT runtime/explicit
+  }
+
+  explicit operator bool() const { return impl_ != NULLPTR; }
+
+  R operator()(A... a) && {
+    auto bye = std::move(impl_);
+    return bye->invoke(std::forward<A&&>(a)...);
+  }
+
+ private:
+  struct Impl {
+    virtual ~Impl() = default;
+    virtual R invoke(A&&... a) = 0;
+  };
+
+  template <typename Fn>
+  struct FnImpl : Impl {
+    explicit FnImpl(Fn fn) : fn_(std::move(fn)) {}
+    R invoke(A&&... a) override { return std::move(fn_)(std::forward<A&&>(a)...); }
+    Fn fn_;
+  };
+
+  std::unique_ptr<Impl> impl_;
+};
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/future.h

@@ -0,0 +1,882 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <atomic>
+#include <cmath>
+#include <functional>
+#include <memory>
+#include <optional>
+#include <type_traits>
+#include <utility>
+#include <vector>
+
+#include "arrow/result.h"
+#include "arrow/status.h"
+#include "arrow/type_fwd.h"
+#include "arrow/type_traits.h"
+#include "arrow/util/config.h"
+#include "arrow/util/functional.h"
+#include "arrow/util/macros.h"
+#include "arrow/util/tracing.h"
+#include "arrow/util/type_fwd.h"
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+
+template <typename>
+struct EnsureFuture;
+
+namespace detail {
+
+template <typename>
+struct is_future : std::false_type {};
+
+template <typename T>
+struct is_future<Future<T>> : std::true_type {};
+
+template <typename Signature, typename Enable = void>
+struct result_of;
+
+template <typename Fn, typename... A>
+struct result_of<Fn(A...),
+                 internal::void_t<decltype(std::declval<Fn>()(std::declval<A>()...))>> {
+  using type = decltype(std::declval<Fn>()(std::declval<A>()...));
+};
+
+template <typename Signature>
+using result_of_t = typename result_of<Signature>::type;
+
+// Helper to find the synchronous counterpart for a Future
+template <typename T>
+struct SyncType {
+  using type = Result<T>;
+};
+
+template <>
+struct SyncType<internal::Empty> {
+  using type = Status;
+};
+
+template <typename Fn>
+using first_arg_is_status =
+    std::is_same<typename std::decay<internal::call_traits::argument_type<0, Fn>>::type,
+                 Status>;
+
+template <typename Fn, typename Then, typename Else,
+          typename Count = internal::call_traits::argument_count<Fn>>
+using if_has_no_args = typename std::conditional<Count::value == 0, Then, Else>::type;
+
+/// Creates a callback that can be added to a future to mark a `dest` future finished
+template <typename Source, typename Dest, bool SourceEmpty = Source::is_empty,
+          bool DestEmpty = Dest::is_empty>
+struct MarkNextFinished {};
+
+/// If the source and dest are both empty we can pass on the status
+template <typename Source, typename Dest>
+struct MarkNextFinished<Source, Dest, true, true> {
+  void operator()(const Status& status) && { next.MarkFinished(status); }
+  Dest next;
+};
+
+/// If the source is not empty but the dest is then we can take the
+/// status out of the result
+template <typename Source, typename Dest>
+struct MarkNextFinished<Source, Dest, false, true> {
+  void operator()(const Result<typename Source::ValueType>& res) && {
+    next.MarkFinished(internal::Empty::ToResult(res.status()));
+  }
+  Dest next;
+};
+
+/// If neither are empty we pass on the result
+template <typename Source, typename Dest>
+struct MarkNextFinished<Source, Dest, false, false> {
+  void operator()(const Result<typename Source::ValueType>& res) && {
+    next.MarkFinished(res);
+  }
+  Dest next;
+};
+
+/// Helper that contains information about how to apply a continuation
+struct ContinueFuture {
+  template <typename Return>
+  struct ForReturnImpl;
+
+  template <typename Return>
+  using ForReturn = typename ForReturnImpl<Return>::type;
+
+  template <typename Signature>
+  using ForSignature = ForReturn<result_of_t<Signature>>;
+
+  // If the callback returns void then we return Future<> that always finishes OK.
+  template <typename ContinueFunc, typename... Args,
+            typename ContinueResult = result_of_t<ContinueFunc && (Args && ...)>,
+            typename NextFuture = ForReturn<ContinueResult>>
+  typename std::enable_if<std::is_void<ContinueResult>::value>::type operator()(
+      NextFuture next, ContinueFunc&& f, Args&&... a) const {
+    std::forward<ContinueFunc>(f)(std::forward<Args>(a)...);
+    next.MarkFinished();
+  }
+
+  /// If the callback returns a non-future then we return Future<T>
+  /// and mark the future finished with the callback result.  It will get promoted
+  /// to Result<T> as part of MarkFinished if it isn't already.
+  ///
+  /// If the callback returns Status and we return Future<> then also send the callback
+  /// result as-is to the destination future.
+  template <typename ContinueFunc, typename... Args,
+            typename ContinueResult = result_of_t<ContinueFunc && (Args && ...)>,
+            typename NextFuture = ForReturn<ContinueResult>>
+  typename std::enable_if<
+      !std::is_void<ContinueResult>::value && !is_future<ContinueResult>::value &&
+      (!NextFuture::is_empty || std::is_same<ContinueResult, Status>::value)>::type
+  operator()(NextFuture next, ContinueFunc&& f, Args&&... a) const {
+    next.MarkFinished(std::forward<ContinueFunc>(f)(std::forward<Args>(a)...));
+  }
+
+  /// If the callback returns a Result and the next future is Future<> then we mark
+  /// the future finished with the callback result.
+  ///
+  /// It may seem odd that the next future is Future<> when the callback returns a
+  /// result but this can occur if the OnFailure callback returns a result while the
+  /// OnSuccess callback is void/Status (e.g. you would get this calling the one-arg
+  /// version of Then with an OnSuccess callback that returns void)
+  template <typename ContinueFunc, typename... Args,
+            typename ContinueResult = result_of_t<ContinueFunc && (Args && ...)>,
+            typename NextFuture = ForReturn<ContinueResult>>
+  typename std::enable_if<!std::is_void<ContinueResult>::value &&
+                          !is_future<ContinueResult>::value && NextFuture::is_empty &&
+                          !std::is_same<ContinueResult, Status>::value>::type
+  operator()(NextFuture next, ContinueFunc&& f, Args&&... a) const {
+    next.MarkFinished(std::forward<ContinueFunc>(f)(std::forward<Args>(a)...).status());
+  }
+
+  /// If the callback returns a Future<T> then we return Future<T>.  We create a new
+  /// future and add a callback to the future given to us by the user that forwards the
+  /// result to the future we just created
+  template <typename ContinueFunc, typename... Args,
+            typename ContinueResult = result_of_t<ContinueFunc && (Args && ...)>,
+            typename NextFuture = ForReturn<ContinueResult>>
+  typename std::enable_if<is_future<ContinueResult>::value>::type operator()(
+      NextFuture next, ContinueFunc&& f, Args&&... a) const {
+    ContinueResult signal_to_complete_next =
+        std::forward<ContinueFunc>(f)(std::forward<Args>(a)...);
+    MarkNextFinished<ContinueResult, NextFuture> callback{std::move(next)};
+    signal_to_complete_next.AddCallback(std::move(callback));
+  }
+
+  /// Helpers to conditionally ignore arguments to ContinueFunc
+  template <typename ContinueFunc, typename NextFuture, typename... Args>
+  void IgnoringArgsIf(std::true_type, NextFuture&& next, ContinueFunc&& f,
+                      Args&&...) const {
+    operator()(std::forward<NextFuture>(next), std::forward<ContinueFunc>(f));
+  }
+  template <typename ContinueFunc, typename NextFuture, typename... Args>
+  void IgnoringArgsIf(std::false_type, NextFuture&& next, ContinueFunc&& f,
+                      Args&&... a) const {
+    operator()(std::forward<NextFuture>(next), std::forward<ContinueFunc>(f),
+               std::forward<Args>(a)...);
+  }
+};
+
+/// Helper struct which tells us what kind of Future gets returned from `Then` based on
+/// the return type of the OnSuccess callback
+template <>
+struct ContinueFuture::ForReturnImpl<void> {
+  using type = Future<>;
+};
+
+template <>
+struct ContinueFuture::ForReturnImpl<Status> {
+  using type = Future<>;
+};
+
+template <typename R>
+struct ContinueFuture::ForReturnImpl {
+  using type = Future<R>;
+};
+
+template <typename T>
+struct ContinueFuture::ForReturnImpl<Result<T>> {
+  using type = Future<T>;
+};
+
+template <typename T>
+struct ContinueFuture::ForReturnImpl<Future<T>> {
+  using type = Future<T>;
+};
+
+}  // namespace detail
+
+/// A Future's execution or completion status
+enum class FutureState : int8_t { PENDING, SUCCESS, FAILURE };
+
+inline bool IsFutureFinished(FutureState state) { return state != FutureState::PENDING; }
+
+/// \brief Describe whether the callback should be scheduled or run synchronously
+enum class ShouldSchedule {
+  /// Always run the callback synchronously (the default)
+  Never = 0,
+  /// Schedule a new task only if the future is not finished when the
+  /// callback is added
+  IfUnfinished = 1,
+  /// Always schedule the callback as a new task
+  Always = 2,
+  /// Schedule a new task only if it would run on an executor other than
+  /// the specified executor.
+  IfDifferentExecutor = 3,
+};
+
+/// \brief Options that control how a continuation is run
+struct CallbackOptions {
+  /// Describe whether the callback should be run synchronously or scheduled
+  ShouldSchedule should_schedule = ShouldSchedule::Never;
+  /// If the callback is scheduled then this is the executor it should be scheduled
+  /// on.  If this is NULL then should_schedule must be Never
+  internal::Executor* executor = NULLPTR;
+
+  static CallbackOptions Defaults() { return {}; }
+};
+
+// Untyped private implementation
+class ARROW_EXPORT FutureImpl : public std::enable_shared_from_this<FutureImpl> {
+ public:
+  FutureImpl();
+  virtual ~FutureImpl() = default;
+
+  FutureState state() { return state_.load(); }
+
+  static std::unique_ptr<FutureImpl> Make();
+  static std::unique_ptr<FutureImpl> MakeFinished(FutureState state);
+
+#ifdef ARROW_WITH_OPENTELEMETRY
+  void SetSpan(util::tracing::Span* span) { span_ = span; }
+#endif
+
+  // Future API
+  void MarkFinished();
+  void MarkFailed();
+  void Wait();
+  bool Wait(double seconds);
+  template <typename ValueType>
+  Result<ValueType>* CastResult() const {
+    return static_cast<Result<ValueType>*>(result_.get());
+  }
+
+  using Callback = internal::FnOnce<void(const FutureImpl& impl)>;
+  void AddCallback(Callback callback, CallbackOptions opts);
+  bool TryAddCallback(const std::function<Callback()>& callback_factory,
+                      CallbackOptions opts);
+
+  std::atomic<FutureState> state_{FutureState::PENDING};
+
+  // Type erased storage for arbitrary results
+  // XXX small objects could be stored inline instead of boxed in a pointer
+  using Storage = std::unique_ptr<void, void (*)(void*)>;
+  Storage result_{NULLPTR, NULLPTR};
+
+  struct CallbackRecord {
+    Callback callback;
+    CallbackOptions options;
+  };
+  std::vector<CallbackRecord> callbacks_;
+#ifdef ARROW_WITH_OPENTELEMETRY
+  util::tracing::Span* span_ = NULLPTR;
+#endif
+};
+
+// ---------------------------------------------------------------------
+// Public API
+
+/// \brief EXPERIMENTAL A std::future-like class with more functionality.
+///
+/// A Future represents the results of a past or future computation.
+/// The Future API has two sides: a producer side and a consumer side.
+///
+/// The producer API allows creating a Future and setting its result or
+/// status, possibly after running a computation function.
+///
+/// The consumer API allows querying a Future's current state, wait for it
+/// to complete, and composing futures with callbacks.
+template <typename T>
+class [[nodiscard]] Future {
+ public:
+  using ValueType = T;
+  using SyncType = typename detail::SyncType<T>::type;
+  static constexpr bool is_empty = std::is_same<T, internal::Empty>::value;
+  // The default constructor creates an invalid Future.  Use Future::Make()
+  // for a valid Future.  This constructor is mostly for the convenience
+  // of being able to presize a vector of Futures.
+  Future() = default;
+
+#ifdef ARROW_WITH_OPENTELEMETRY
+  void SetSpan(util::tracing::Span* span) { impl_->SetSpan(span); }
+#endif
+
+  // Consumer API
+
+  bool is_valid() const { return impl_ != NULLPTR; }
+
+  /// \brief Return the Future's current state
+  ///
+  /// A return value of PENDING is only indicative, as the Future can complete
+  /// concurrently.  A return value of FAILURE or SUCCESS is definitive, though.
+  FutureState state() const {
+    CheckValid();
+    return impl_->state();
+  }
+
+  /// \brief Whether the Future is finished
+  ///
+  /// A false return value is only indicative, as the Future can complete
+  /// concurrently.  A true return value is definitive, though.
+  bool is_finished() const {
+    CheckValid();
+    return IsFutureFinished(impl_->state());
+  }
+
+  /// \brief Wait for the Future to complete and return its Result
+  const Result<ValueType>& result() const& {
+    Wait();
+    return *GetResult();
+  }
+
+  /// \brief Returns an rvalue to the result.  This method is potentially unsafe
+  ///
+  /// The future is not the unique owner of the result, copies of a future will
+  /// also point to the same result.  You must make sure that no other copies
+  /// of the future exist.  Attempts to add callbacks after you move the result
+  /// will result in undefined behavior.
+  Result<ValueType>&& MoveResult() {
+    Wait();
+    return std::move(*GetResult());
+  }
+
+  /// \brief Wait for the Future to complete and return its Status
+  const Status& status() const { return result().status(); }
+
+  /// \brief Future<T> is convertible to Future<>, which views only the
+  /// Status of the original. Marking the returned Future Finished is not supported.
+  explicit operator Future<>() const {
+    Future<> status_future;
+    status_future.impl_ = impl_;
+    return status_future;
+  }
+
+  /// \brief Wait for the Future to complete
+  void Wait() const {
+    CheckValid();
+    impl_->Wait();
+  }
+
+  /// \brief Wait for the Future to complete, or for the timeout to expire
+  ///
+  /// `true` is returned if the Future completed, `false` if the timeout expired.
+  /// Note a `false` value is only indicative, as the Future can complete
+  /// concurrently.
+  bool Wait(double seconds) const {
+    CheckValid();
+    return impl_->Wait(seconds);
+  }
+
+  // Producer API
+
+  /// \brief Producer API: mark Future finished
+  ///
+  /// The Future's result is set to `res`.
+  void MarkFinished(Result<ValueType> res) { DoMarkFinished(std::move(res)); }
+
+  /// \brief Mark a Future<> completed with the provided Status.
+  template <typename E = ValueType, typename = typename std::enable_if<
+                                        std::is_same<E, internal::Empty>::value>::type>
+  void MarkFinished(Status s = Status::OK()) {
+    return DoMarkFinished(E::ToResult(std::move(s)));
+  }
+
+  /// \brief Producer API: instantiate a valid Future
+  ///
+  /// The Future's state is initialized with PENDING.  If you are creating a future with
+  /// this method you must ensure that future is eventually completed (with success or
+  /// failure).  Creating a future, returning it, and never completing the future can lead
+  /// to memory leaks (for example, see Loop).
+  static Future Make() {
+    Future fut;
+    fut.impl_ = FutureImpl::Make();
+    return fut;
+  }
+
+  /// \brief Producer API: instantiate a finished Future
+  static Future<ValueType> MakeFinished(Result<ValueType> res) {
+    Future<ValueType> fut;
+    fut.InitializeFromResult(std::move(res));
+    return fut;
+  }
+
+  /// \brief Make a finished Future<> with the provided Status.
+  template <typename E = ValueType, typename = typename std::enable_if<
+                                        std::is_same<E, internal::Empty>::value>::type>
+  static Future<> MakeFinished(Status s = Status::OK()) {
+    return MakeFinished(E::ToResult(std::move(s)));
+  }
+
+  struct WrapResultOnComplete {
+    template <typename OnComplete>
+    struct Callback {
+      void operator()(const FutureImpl& impl) && {
+        std::move(on_complete)(*impl.CastResult<ValueType>());
+      }
+      OnComplete on_complete;
+    };
+  };
+
+  struct WrapStatusyOnComplete {
+    template <typename OnComplete>
+    struct Callback {
+      static_assert(std::is_same<internal::Empty, ValueType>::value,
+                    "Only callbacks for Future<> should accept Status and not Result");
+
+      void operator()(const FutureImpl& impl) && {
+        std::move(on_complete)(impl.CastResult<ValueType>()->status());
+      }
+      OnComplete on_complete;
+    };
+  };
+
+  template <typename OnComplete>
+  using WrapOnComplete = typename std::conditional<
+      detail::first_arg_is_status<OnComplete>::value, WrapStatusyOnComplete,
+      WrapResultOnComplete>::type::template Callback<OnComplete>;
+
+  /// \brief Consumer API: Register a callback to run when this future completes
+  ///
+  /// The callback should receive the result of the future (const Result<T>&)
+  /// For a void or statusy future this should be (const Status&)
+  ///
+  /// There is no guarantee to the order in which callbacks will run.  In
+  /// particular, callbacks added while the future is being marked complete
+  /// may be executed immediately, ahead of, or even the same time as, other
+  /// callbacks that have been previously added.
+  ///
+  /// WARNING: callbacks may hold arbitrary references, including cyclic references.
+  /// Since callbacks will only be destroyed after they are invoked, this can lead to
+  /// memory leaks if a Future is never marked finished (abandoned):
+  ///
+  /// {
+  ///     auto fut = Future<>::Make();
+  ///     fut.AddCallback([fut]() {});
+  /// }
+  ///
+  /// In this example `fut` falls out of scope but is not destroyed because it holds a
+  /// cyclic reference to itself through the callback.
+  template <typename OnComplete, typename Callback = WrapOnComplete<OnComplete>>
+  void AddCallback(OnComplete on_complete,
+                   CallbackOptions opts = CallbackOptions::Defaults()) const {
+    // We know impl_ will not be dangling when invoking callbacks because at least one
+    // thread will be waiting for MarkFinished to return. Thus it's safe to keep a
+    // weak reference to impl_ here
+    impl_->AddCallback(Callback{std::move(on_complete)}, opts);
+  }
+
+  /// \brief Overload of AddCallback that will return false instead of running
+  /// synchronously
+  ///
+  /// This overload will guarantee the callback is never run synchronously.  If the future
+  /// is already finished then it will simply return false.  This can be useful to avoid
+  /// stack overflow in a situation where you have recursive Futures.  For an example
+  /// see the Loop function
+  ///
+  /// Takes in a callback factory function to allow moving callbacks (the factory function
+  /// will only be called if the callback can successfully be added)
+  ///
+  /// Returns true if a callback was actually added and false if the callback failed
+  /// to add because the future was marked complete.
+  template <typename CallbackFactory,
+            typename OnComplete = detail::result_of_t<CallbackFactory()>,
+            typename Callback = WrapOnComplete<OnComplete>>
+  bool TryAddCallback(CallbackFactory callback_factory,
+                      CallbackOptions opts = CallbackOptions::Defaults()) const {
+    return impl_->TryAddCallback([&]() { return Callback{callback_factory()}; }, opts);
+  }
+
+  template <typename OnSuccess, typename OnFailure>
+  struct ThenOnComplete {
+    static constexpr bool has_no_args =
+        internal::call_traits::argument_count<OnSuccess>::value == 0;
+
+    using ContinuedFuture = detail::ContinueFuture::ForSignature<
+        detail::if_has_no_args<OnSuccess, OnSuccess && (), OnSuccess && (const T&)>>;
+
+    static_assert(
+        std::is_same<detail::ContinueFuture::ForSignature<OnFailure && (const Status&)>,
+                     ContinuedFuture>::value,
+        "OnSuccess and OnFailure must continue with the same future type");
+
+    struct DummyOnSuccess {
+      void operator()(const T&);
+    };
+    using OnSuccessArg = typename std::decay<internal::call_traits::argument_type<
+        0, detail::if_has_no_args<OnSuccess, DummyOnSuccess, OnSuccess>>>::type;
+
+    static_assert(
+        !std::is_same<OnSuccessArg, typename EnsureResult<OnSuccessArg>::type>::value,
+        "OnSuccess' argument should not be a Result");
+
+    void operator()(const Result<T>& result) && {
+      detail::ContinueFuture continue_future;
+      if (ARROW_PREDICT_TRUE(result.ok())) {
+        // move on_failure to a(n immediately destroyed) temporary to free its resources
+        ARROW_UNUSED(OnFailure(std::move(on_failure)));
+        continue_future.IgnoringArgsIf(
+            detail::if_has_no_args<OnSuccess, std::true_type, std::false_type>{},
+            std::move(next), std::move(on_success), result.ValueOrDie());
+      } else {
+        ARROW_UNUSED(OnSuccess(std::move(on_success)));
+        continue_future(std::move(next), std::move(on_failure), result.status());
+      }
+    }
+
+    OnSuccess on_success;
+    OnFailure on_failure;
+    ContinuedFuture next;
+  };
+
+  template <typename OnSuccess>
+  struct PassthruOnFailure {
+    using ContinuedFuture = detail::ContinueFuture::ForSignature<
+        detail::if_has_no_args<OnSuccess, OnSuccess && (), OnSuccess && (const T&)>>;
+
+    Result<typename ContinuedFuture::ValueType> operator()(const Status& s) { return s; }
+  };
+
+  /// \brief Consumer API: Register a continuation to run when this future completes
+  ///
+  /// The continuation will run in the same thread that called MarkFinished (whatever
+  /// callback is registered with this function will run before MarkFinished returns).
+  /// Avoid long-running callbacks in favor of submitting a task to an Executor and
+  /// returning the future.
+  ///
+  /// Two callbacks are supported:
+  /// - OnSuccess, called with the result (const ValueType&) on successful completion.
+  ///              for an empty future this will be called with nothing ()
+  /// - OnFailure, called with the error (const Status&) on failed completion.
+  ///              This callback is optional and defaults to a passthru of any errors.
+  ///
+  /// Then() returns a Future whose ValueType is derived from the return type of the
+  /// callbacks. If a callback returns:
+  /// - void, a Future<> will be returned which will completes successfully as soon
+  ///   as the callback runs.
+  /// - Status, a Future<> will be returned which will complete with the returned Status
+  ///   as soon as the callback runs.
+  /// - V or Result<V>, a Future<V> will be returned which will complete with the result
+  ///   of invoking the callback as soon as the callback runs.
+  /// - Future<V>, a Future<V> will be returned which will be marked complete when the
+  ///   future returned by the callback completes (and will complete with the same
+  ///   result).
+  ///
+  /// The continued Future type must be the same for both callbacks.
+  ///
+  /// Note that OnFailure can swallow errors, allowing continued Futures to successfully
+  /// complete even if this Future fails.
+  ///
+  /// If this future is already completed then the callback will be run immediately
+  /// and the returned future may already be marked complete.
+  ///
+  /// See AddCallback for general considerations when writing callbacks.
+  template <typename OnSuccess, typename OnFailure = PassthruOnFailure<OnSuccess>,
+            typename OnComplete = ThenOnComplete<OnSuccess, OnFailure>,
+            typename ContinuedFuture = typename OnComplete::ContinuedFuture>
+  ContinuedFuture Then(OnSuccess on_success, OnFailure on_failure = {},
+                       CallbackOptions options = CallbackOptions::Defaults()) const {
+    auto next = ContinuedFuture::Make();
+    AddCallback(OnComplete{std::forward<OnSuccess>(on_success),
+                           std::forward<OnFailure>(on_failure), next},
+                options);
+    return next;
+  }
+
+  /// \brief Implicit constructor to create a finished future from a value
+  Future(ValueType val) : Future() {  // NOLINT runtime/explicit
+    impl_ = FutureImpl::MakeFinished(FutureState::SUCCESS);
+    SetResult(std::move(val));
+  }
+
+  /// \brief Implicit constructor to create a future from a Result, enabling use
+  ///     of macros like ARROW_ASSIGN_OR_RAISE.
+  Future(Result<ValueType> res) : Future() {  // NOLINT runtime/explicit
+    if (ARROW_PREDICT_TRUE(res.ok())) {
+      impl_ = FutureImpl::MakeFinished(FutureState::SUCCESS);
+    } else {
+      impl_ = FutureImpl::MakeFinished(FutureState::FAILURE);
+    }
+    SetResult(std::move(res));
+  }
+
+  /// \brief Implicit constructor to create a future from a Status, enabling use
+  ///     of macros like ARROW_RETURN_NOT_OK.
+  Future(Status s)  // NOLINT runtime/explicit
+      : Future(Result<ValueType>(std::move(s))) {}
+
+ protected:
+  void InitializeFromResult(Result<ValueType> res) {
+    if (ARROW_PREDICT_TRUE(res.ok())) {
+      impl_ = FutureImpl::MakeFinished(FutureState::SUCCESS);
+    } else {
+      impl_ = FutureImpl::MakeFinished(FutureState::FAILURE);
+    }
+    SetResult(std::move(res));
+  }
+
+  void Initialize() { impl_ = FutureImpl::Make(); }
+
+  Result<ValueType>* GetResult() const { return impl_->CastResult<ValueType>(); }
+
+  void SetResult(Result<ValueType> res) {
+    impl_->result_ = {new Result<ValueType>(std::move(res)),
+                      [](void* p) { delete static_cast<Result<ValueType>*>(p); }};
+  }
+
+  void DoMarkFinished(Result<ValueType> res) {
+    SetResult(std::move(res));
+
+    if (ARROW_PREDICT_TRUE(GetResult()->ok())) {
+      impl_->MarkFinished();
+    } else {
+      impl_->MarkFailed();
+    }
+  }
+
+  void CheckValid() const {
+#ifndef NDEBUG
+    if (!is_valid()) {
+      Status::Invalid("Invalid Future (default-initialized?)").Abort();
+    }
+#endif
+  }
+
+  explicit Future(std::shared_ptr<FutureImpl> impl) : impl_(std::move(impl)) {}
+
+  std::shared_ptr<FutureImpl> impl_;
+
+  friend struct detail::ContinueFuture;
+
+  template <typename U>
+  friend class Future;
+  friend class WeakFuture<T>;
+
+  FRIEND_TEST(FutureRefTest, ChainRemoved);
+  FRIEND_TEST(FutureRefTest, TailRemoved);
+  FRIEND_TEST(FutureRefTest, HeadRemoved);
+};
+
+template <typename T>
+typename Future<T>::SyncType FutureToSync(const Future<T>& fut) {
+  return fut.result();
+}
+
+template <>
+inline typename Future<internal::Empty>::SyncType FutureToSync<internal::Empty>(
+    const Future<internal::Empty>& fut) {
+  return fut.status();
+}
+
+template <>
+inline Future<>::Future(Status s) : Future(internal::Empty::ToResult(std::move(s))) {}
+
+template <typename T>
+class WeakFuture {
+ public:
+  explicit WeakFuture(const Future<T>& future) : impl_(future.impl_) {}
+
+  Future<T> get() { return Future<T>{impl_.lock()}; }
+
+ private:
+  std::weak_ptr<FutureImpl> impl_;
+};
+
+/// \defgroup future-utilities Functions for working with Futures
+/// @{
+
+/// If a Result<Future> holds an error instead of a Future, construct a finished Future
+/// holding that error.
+template <typename T>
+static Future<T> DeferNotOk(Result<Future<T>> maybe_future) {
+  if (ARROW_PREDICT_FALSE(!maybe_future.ok())) {
+    return Future<T>::MakeFinished(std::move(maybe_future).status());
+  }
+  return std::move(maybe_future).MoveValueUnsafe();
+}
+
+/// \brief Create a Future which completes when all of `futures` complete.
+///
+/// The future's result is a vector of the results of `futures`.
+/// Note that this future will never be marked "failed"; failed results
+/// will be stored in the result vector alongside successful results.
+template <typename T>
+Future<std::vector<Result<T>>> All(std::vector<Future<T>> futures) {
+  struct State {
+    explicit State(std::vector<Future<T>> f)
+        : futures(std::move(f)), n_remaining(futures.size()) {}
+
+    std::vector<Future<T>> futures;
+    std::atomic<size_t> n_remaining;
+  };
+
+  if (futures.size() == 0) {
+    return {std::vector<Result<T>>{}};
+  }
+
+  auto state = std::make_shared<State>(std::move(futures));
+
+  auto out = Future<std::vector<Result<T>>>::Make();
+  for (const Future<T>& future : state->futures) {
+    future.AddCallback([state, out](const Result<T>&) mutable {
+      if (state->n_remaining.fetch_sub(1) != 1) return;
+
+      std::vector<Result<T>> results(state->futures.size());
+      for (size_t i = 0; i < results.size(); ++i) {
+        results[i] = state->futures[i].result();
+      }
+      out.MarkFinished(std::move(results));
+    });
+  }
+  return out;
+}
+
+/// \brief Create a Future which completes when all of `futures` complete.
+///
+/// The future will be marked complete if all `futures` complete
+/// successfully. Otherwise, it will be marked failed with the status of
+/// the first failing future.
+ARROW_EXPORT
+Future<> AllComplete(const std::vector<Future<>>& futures);
+
+/// \brief Create a Future which completes when all of `futures` complete.
+///
+/// The future will finish with an ok status if all `futures` finish with
+/// an ok status. Otherwise, it will be marked failed with the status of
+/// one of the failing futures.
+///
+/// Unlike AllComplete this Future will not complete immediately when a
+/// failure occurs.  It will wait until all futures have finished.
+ARROW_EXPORT
+Future<> AllFinished(const std::vector<Future<>>& futures);
+
+/// @}
+
+struct Continue {
+  template <typename T>
+  operator std::optional<T>() && {  // NOLINT explicit
+    return {};
+  }
+};
+
+template <typename T = internal::Empty>
+std::optional<T> Break(T break_value = {}) {
+  return std::optional<T>{std::move(break_value)};
+}
+
+template <typename T = internal::Empty>
+using ControlFlow = std::optional<T>;
+
+/// \brief Loop through an asynchronous sequence
+///
+/// \param[in] iterate A generator of Future<ControlFlow<BreakValue>>. On completion
+/// of each yielded future the resulting ControlFlow will be examined. A Break will
+/// terminate the loop, while a Continue will re-invoke `iterate`.
+///
+/// \return A future which will complete when a Future returned by iterate completes with
+/// a Break
+template <typename Iterate,
+          typename Control = typename detail::result_of_t<Iterate()>::ValueType,
+          typename BreakValueType = typename Control::value_type>
+Future<BreakValueType> Loop(Iterate iterate) {
+  struct Callback {
+    bool CheckForTermination(const Result<Control>& control_res) {
+      if (!control_res.ok()) {
+        break_fut.MarkFinished(control_res.status());
+        return true;
+      }
+      if (control_res->has_value()) {
+        break_fut.MarkFinished(**control_res);
+        return true;
+      }
+      return false;
+    }
+
+    void operator()(const Result<Control>& maybe_control) && {
+      if (CheckForTermination(maybe_control)) return;
+
+      auto control_fut = iterate();
+      while (true) {
+        if (control_fut.TryAddCallback([this]() { return *this; })) {
+          // Adding a callback succeeded; control_fut was not finished
+          // and we must wait to CheckForTermination.
+          return;
+        }
+        // Adding a callback failed; control_fut was finished and we
+        // can CheckForTermination immediately. This also avoids recursion and potential
+        // stack overflow.
+        if (CheckForTermination(control_fut.result())) return;
+
+        control_fut = iterate();
+      }
+    }
+
+    Iterate iterate;
+
+    // If the future returned by control_fut is never completed then we will be hanging on
+    // to break_fut forever even if the listener has given up listening on it.  Instead we
+    // rely on the fact that a producer (the caller of Future<>::Make) is always
+    // responsible for completing the futures they create.
+    // TODO: Could avoid this kind of situation with "future abandonment" similar to mesos
+    Future<BreakValueType> break_fut;
+  };
+
+  auto break_fut = Future<BreakValueType>::Make();
+  auto control_fut = iterate();
+  control_fut.AddCallback(Callback{std::move(iterate), break_fut});
+
+  return break_fut;
+}
+
+inline Future<> ToFuture(Status status) {
+  return Future<>::MakeFinished(std::move(status));
+}
+
+template <typename T>
+Future<T> ToFuture(T value) {
+  return Future<T>::MakeFinished(std::move(value));
+}
+
+template <typename T>
+Future<T> ToFuture(Result<T> maybe_value) {
+  return Future<T>::MakeFinished(std::move(maybe_value));
+}
+
+template <typename T>
+Future<T> ToFuture(Future<T> fut) {
+  return fut;
+}
+
+template <typename T>
+struct EnsureFuture {
+  using type = decltype(ToFuture(std::declval<T>()));
+};
+
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/hash_util.h

@@ -0,0 +1,66 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+namespace arrow {
+namespace internal {
+
+// ----------------------------------------------------------------------
+// BEGIN Hash utilities from Boost
+
+namespace detail {
+
+#if defined(_MSC_VER)
+#  define ARROW_HASH_ROTL32(x, r) _rotl(x, r)
+#else
+#  define ARROW_HASH_ROTL32(x, r) (x << r) | (x >> (32 - r))
+#endif
+
+template <typename SizeT>
+inline void hash_combine_impl(SizeT& seed, SizeT value) {
+  seed ^= value + 0x9e3779b9 + (seed << 6) + (seed >> 2);
+}
+
+inline void hash_combine_impl(uint32_t& h1, uint32_t k1) {
+  const uint32_t c1 = 0xcc9e2d51;
+  const uint32_t c2 = 0x1b873593;
+
+  k1 *= c1;
+  k1 = ARROW_HASH_ROTL32(k1, 15);
+  k1 *= c2;
+
+  h1 ^= k1;
+  h1 = ARROW_HASH_ROTL32(h1, 13);
+  h1 = h1 * 5 + 0xe6546b64;
+}
+
+#undef ARROW_HASH_ROTL32
+
+}  // namespace detail
+
+template <class T>
+inline void hash_combine(std::size_t& seed, T const& v) {
+  std::hash<T> hasher;
+  return ::arrow::internal::detail::hash_combine_impl(seed, hasher(v));
+}
+
+// END Hash utilities from Boost
+// ----------------------------------------------------------------------
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/hashing.h

@@ -0,0 +1,984 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+// Private header, not to be exported
+
+#pragma once
+
+#include <algorithm>
+#include <cassert>
+#include <cmath>
+#include <cstdint>
+#include <cstring>
+#include <limits>
+#include <memory>
+#include <string>
+#include <type_traits>
+#include <utility>
+#include <vector>
+
+#include "arrow/array/builder_binary.h"
+#include "arrow/buffer_builder.h"
+#include "arrow/result.h"
+#include "arrow/status.h"
+#include "arrow/type_fwd.h"
+#include "arrow/type_traits.h"
+#include "arrow/util/bit_util.h"
+#include "arrow/util/bitmap_builders.h"
+#include "arrow/util/endian.h"
+#include "arrow/util/float16.h"
+#include "arrow/util/logging.h"
+#include "arrow/util/macros.h"
+#include "arrow/util/ubsan.h"
+
+#define XXH_INLINE_ALL
+
+#include "arrow/vendored/xxhash.h"  // IWYU pragma: keep
+
+namespace arrow {
+namespace internal {
+
+// XXX would it help to have a 32-bit hash value on large datasets?
+typedef uint64_t hash_t;
+
+// Notes about the choice of a hash function.
+// - XXH3 is extremely fast on most data sizes, from small to huge;
+//   faster even than HW CRC-based hashing schemes
+// - our custom hash function for tiny values (< 16 bytes) is still
+//   significantly faster (~30%), at least on this machine and compiler
+
+template <uint64_t AlgNum>
+inline hash_t ComputeStringHash(const void* data, int64_t length);
+
+/// \brief A hash function for bitmaps that can handle offsets and lengths in
+/// terms of number of bits. The hash only depends on the bits actually hashed.
+///
+/// It's the caller's responsibility to ensure that bits_offset + num_bits are
+/// readable from the bitmap.
+///
+/// \pre bits_offset >= 0
+/// \pre num_bits >= 0
+/// \pre (bits_offset + num_bits + 7) / 8 <= readable length in bytes from bitmap
+///
+/// \param bitmap The pointer to the bitmap.
+/// \param seed The seed for the hash function (useful when chaining hash functions).
+/// \param bits_offset The offset in bits relative to the start of the bitmap.
+/// \param num_bits The number of bits after the offset to be hashed.
+ARROW_EXPORT hash_t ComputeBitmapHash(const uint8_t* bitmap, hash_t seed,
+                                      int64_t bits_offset, int64_t num_bits);
+
+template <typename Scalar, uint64_t AlgNum>
+struct ScalarHelperBase {
+  static bool CompareScalars(Scalar u, Scalar v) { return u == v; }
+
+  static hash_t ComputeHash(const Scalar& value) {
+    // Generic hash computation for scalars.  Simply apply the string hash
+    // to the bit representation of the value.
+
+    // XXX in the case of FP values, we'd like equal values to have the same hash,
+    // even if they have different bit representations...
+    return ComputeStringHash<AlgNum>(&value, sizeof(value));
+  }
+};
+
+template <typename Scalar, uint64_t AlgNum = 0, typename Enable = void>
+struct ScalarHelper : public ScalarHelperBase<Scalar, AlgNum> {};
+
+template <typename Scalar, uint64_t AlgNum>
+struct ScalarHelper<Scalar, AlgNum, enable_if_t<std::is_integral<Scalar>::value>>
+    : public ScalarHelperBase<Scalar, AlgNum> {
+  // ScalarHelper specialization for integers
+
+  static hash_t ComputeHash(const Scalar& value) {
+    // Faster hash computation for integers.
+
+    // Two of xxhash's prime multipliers (which are chosen for their
+    // bit dispersion properties)
+    static constexpr uint64_t multipliers[] = {11400714785074694791ULL,
+                                               14029467366897019727ULL};
+
+    // Multiplying by the prime number mixes the low bits into the high bits,
+    // then byte-swapping (which is a single CPU instruction) allows the
+    // combined high and low bits to participate in the initial hash table index.
+    auto h = static_cast<hash_t>(value);
+    return bit_util::ByteSwap(multipliers[AlgNum] * h);
+  }
+};
+
+template <typename Scalar, uint64_t AlgNum>
+struct ScalarHelper<Scalar, AlgNum,
+                    enable_if_t<std::is_same<std::string_view, Scalar>::value>>
+    : public ScalarHelperBase<Scalar, AlgNum> {
+  // ScalarHelper specialization for std::string_view
+
+  static hash_t ComputeHash(std::string_view value) {
+    return ComputeStringHash<AlgNum>(value.data(), static_cast<int64_t>(value.size()));
+  }
+};
+
+template <typename Scalar, uint64_t AlgNum>
+struct ScalarHelper<Scalar, AlgNum, enable_if_t<std::is_floating_point<Scalar>::value>>
+    : public ScalarHelperBase<Scalar, AlgNum> {
+  // ScalarHelper specialization for reals
+
+  static bool CompareScalars(Scalar u, Scalar v) {
+    if (std::isnan(u)) {
+      // XXX should we do a bit-precise comparison?
+      return std::isnan(v);
+    }
+    return u == v;
+  }
+};
+
+template <typename Scalar, uint64_t AlgNum>
+struct ScalarHelper<Scalar, AlgNum,
+                    enable_if_t<std::is_same_v<Scalar, ::arrow::util::Float16>>>
+    : public ScalarHelperBase<Scalar, AlgNum> {
+  // ScalarHelper specialization for Float16
+
+  static bool CompareScalars(Scalar u, Scalar v) {
+    if (u.is_nan()) {
+      // XXX should we do a bit-precise comparison?
+      return v.is_nan();
+    }
+    return u == v;
+  }
+};
+
+template <uint64_t AlgNum = 0>
+hash_t ComputeStringHash(const void* data, int64_t length) {
+  if (ARROW_PREDICT_TRUE(length <= 16)) {
+    // Specialize for small hash strings, as they are quite common as
+    // hash table keys.  Even XXH3 isn't quite as fast.
+    auto p = reinterpret_cast<const uint8_t*>(data);
+    auto n = static_cast<uint32_t>(length);
+    if (n <= 8) {
+      if (n <= 3) {
+        if (n == 0) {
+          return 1U;
+        }
+        uint32_t x = (n << 24) ^ (p[0] << 16) ^ (p[n / 2] << 8) ^ p[n - 1];
+        return ScalarHelper<uint32_t, AlgNum>::ComputeHash(x);
+      }
+      // 4 <= length <= 8
+      // We can read the string as two overlapping 32-bit ints, apply
+      // different hash functions to each of them in parallel, then XOR
+      // the results
+      uint32_t x, y;
+      hash_t hx, hy;
+      x = util::SafeLoadAs<uint32_t>(p + n - 4);
+      y = util::SafeLoadAs<uint32_t>(p);
+      hx = ScalarHelper<uint32_t, AlgNum>::ComputeHash(x);
+      hy = ScalarHelper<uint32_t, AlgNum ^ 1>::ComputeHash(y);
+      return n ^ hx ^ hy;
+    }
+    // 8 <= length <= 16
+    // Apply the same principle as above
+    uint64_t x, y;
+    hash_t hx, hy;
+    x = util::SafeLoadAs<uint64_t>(p + n - 8);
+    y = util::SafeLoadAs<uint64_t>(p);
+    hx = ScalarHelper<uint64_t, AlgNum>::ComputeHash(x);
+    hy = ScalarHelper<uint64_t, AlgNum ^ 1>::ComputeHash(y);
+    return n ^ hx ^ hy;
+  }
+
+#if XXH3_SECRET_SIZE_MIN != 136
+#  error XXH3_SECRET_SIZE_MIN changed, please fix kXxh3Secrets
+#endif
+
+  // XXH3_64bits_withSeed generates a secret based on the seed, which is too slow.
+  // Instead, we use hard-coded random secrets.  To maximize cache efficiency,
+  // they reuse the same memory area.
+  static constexpr unsigned char kXxh3Secrets[XXH3_SECRET_SIZE_MIN + 1] = {
+      0xe7, 0x8b, 0x13, 0xf9, 0xfc, 0xb5, 0x8e, 0xef, 0x81, 0x48, 0x2c, 0xbf, 0xf9, 0x9f,
+      0xc1, 0x1e, 0x43, 0x6d, 0xbf, 0xa6, 0x6d, 0xb5, 0x72, 0xbc, 0x97, 0xd8, 0x61, 0x24,
+      0x0f, 0x12, 0xe3, 0x05, 0x21, 0xf7, 0x5c, 0x66, 0x67, 0xa5, 0x65, 0x03, 0x96, 0x26,
+      0x69, 0xd8, 0x29, 0x20, 0xf8, 0xc7, 0xb0, 0x3d, 0xdd, 0x7d, 0x18, 0xa0, 0x60, 0x75,
+      0x92, 0xa4, 0xce, 0xba, 0xc0, 0x77, 0xf4, 0xac, 0xb7, 0x03, 0x53, 0xf0, 0x98, 0xce,
+      0xe6, 0x2b, 0x20, 0xc7, 0x82, 0x91, 0xab, 0xbf, 0x68, 0x5c, 0x62, 0x4d, 0x33, 0xa3,
+      0xe1, 0xb3, 0xff, 0x97, 0x54, 0x4c, 0x44, 0x34, 0xb5, 0xb9, 0x32, 0x4c, 0x75, 0x42,
+      0x89, 0x53, 0x94, 0xd4, 0x9f, 0x2b, 0x76, 0x4d, 0x4e, 0xe6, 0xfa, 0x15, 0x3e, 0xc1,
+      0xdb, 0x71, 0x4b, 0x2c, 0x94, 0xf5, 0xfc, 0x8c, 0x89, 0x4b, 0xfb, 0xc1, 0x82, 0xa5,
+      0x6a, 0x53, 0xf9, 0x4a, 0xba, 0xce, 0x1f, 0xc0, 0x97, 0x1a, 0x87};
+
+  static_assert(AlgNum < 2, "AlgNum too large");
+  static constexpr auto secret = kXxh3Secrets + AlgNum;
+  return XXH3_64bits_withSecret(data, static_cast<size_t>(length), secret,
+                                XXH3_SECRET_SIZE_MIN);
+}
+
+// XXX add a HashEq<ArrowType> struct with both hash and compare functions?
+
+// ----------------------------------------------------------------------
+// An open-addressing insert-only hash table (no deletes)
+
+template <typename Payload>
+class HashTable {
+ public:
+  static constexpr hash_t kSentinel = 0ULL;
+  static constexpr int64_t kLoadFactor = 2UL;
+
+  struct Entry {
+    hash_t h;
+    Payload payload;
+
+    // An entry is valid if the hash is different from the sentinel value
+    operator bool() const { return h != kSentinel; }
+  };
+
+  HashTable(MemoryPool* pool, uint64_t capacity) : entries_builder_(pool) {
+    ARROW_DCHECK_NE(pool, nullptr);
+    // Minimum of 32 elements
+    capacity = std::max<uint64_t>(capacity, 32UL);
+    capacity_ = bit_util::NextPower2(capacity);
+    capacity_mask_ = capacity_ - 1;
+    size_ = 0;
+
+    ARROW_DCHECK_OK(UpsizeBuffer(capacity_));
+  }
+
+  // Lookup with non-linear probing
+  // cmp_func should have signature bool(const Payload*).
+  // Return a (Entry*, found) pair.
+  template <typename CmpFunc>
+  std::pair<Entry*, bool> Lookup(hash_t h, CmpFunc&& cmp_func) {
+    auto p = Lookup<DoCompare, CmpFunc>(h, entries_, capacity_mask_,
+                                        std::forward<CmpFunc>(cmp_func));
+    return {&entries_[p.first], p.second};
+  }
+
+  template <typename CmpFunc>
+  std::pair<const Entry*, bool> Lookup(hash_t h, CmpFunc&& cmp_func) const {
+    auto p = Lookup<DoCompare, CmpFunc>(h, entries_, capacity_mask_,
+                                        std::forward<CmpFunc>(cmp_func));
+    return {&entries_[p.first], p.second};
+  }
+
+  Status Insert(Entry* entry, hash_t h, const Payload& payload) {
+    // Ensure entry is empty before inserting
+    assert(!*entry);
+    entry->h = FixHash(h);
+    entry->payload = payload;
+    ++size_;
+
+    if (ARROW_PREDICT_FALSE(NeedUpsizing())) {
+      // Resize less frequently since it is expensive
+      return Upsize(capacity_ * kLoadFactor * 2);
+    }
+    return Status::OK();
+  }
+
+  uint64_t size() const { return size_; }
+
+  // Visit all non-empty entries in the table
+  // The visit_func should have signature void(const Entry*)
+  template <typename VisitFunc>
+  void VisitEntries(VisitFunc&& visit_func) const {
+    for (uint64_t i = 0; i < capacity_; i++) {
+      const auto& entry = entries_[i];
+      if (entry) {
+        visit_func(&entry);
+      }
+    }
+  }
+
+ protected:
+  // NoCompare is for when the value is known not to exist in the table
+  enum CompareKind { DoCompare, NoCompare };
+
+  // The workhorse lookup function
+  template <CompareKind CKind, typename CmpFunc>
+  std::pair<uint64_t, bool> Lookup(hash_t h, const Entry* entries, uint64_t size_mask,
+                                   CmpFunc&& cmp_func) const {
+    static constexpr uint8_t perturb_shift = 5;
+
+    uint64_t index, perturb;
+    const Entry* entry;
+
+    h = FixHash(h);
+    index = h & size_mask;
+    perturb = (h >> perturb_shift) + 1U;
+
+    while (true) {
+      entry = &entries[index];
+      if (CompareEntry<CKind, CmpFunc>(h, entry, std::forward<CmpFunc>(cmp_func))) {
+        // Found
+        return {index, true};
+      }
+      if (entry->h == kSentinel) {
+        // Empty slot
+        return {index, false};
+      }
+
+      // Perturbation logic inspired from CPython's set / dict object.
+      // The goal is that all 64 bits of the unmasked hash value eventually
+      // participate in the probing sequence, to minimize clustering.
+      index = (index + perturb) & size_mask;
+      perturb = (perturb >> perturb_shift) + 1U;
+    }
+  }
+
+  template <CompareKind CKind, typename CmpFunc>
+  bool CompareEntry(hash_t h, const Entry* entry, CmpFunc&& cmp_func) const {
+    if (CKind == NoCompare) {
+      return false;
+    } else {
+      return entry->h == h && cmp_func(&entry->payload);
+    }
+  }
+
+  bool NeedUpsizing() const {
+    // Keep the load factor <= 1/2
+    return size_ * kLoadFactor >= capacity_;
+  }
+
+  Status UpsizeBuffer(uint64_t capacity) {
+    RETURN_NOT_OK(entries_builder_.Resize(capacity));
+    entries_ = entries_builder_.mutable_data();
+    memset(static_cast<void*>(entries_), 0, capacity * sizeof(Entry));
+
+    return Status::OK();
+  }
+
+  Status Upsize(uint64_t new_capacity) {
+    assert(new_capacity > capacity_);
+    uint64_t new_mask = new_capacity - 1;
+    assert((new_capacity & new_mask) == 0);  // it's a power of two
+
+    // Stash old entries and seal builder, effectively resetting the Buffer
+    const Entry* old_entries = entries_;
+    ARROW_ASSIGN_OR_RAISE(auto previous, entries_builder_.FinishWithLength(capacity_));
+    // Allocate new buffer
+    RETURN_NOT_OK(UpsizeBuffer(new_capacity));
+
+    for (uint64_t i = 0; i < capacity_; i++) {
+      const auto& entry = old_entries[i];
+      if (entry) {
+        // Dummy compare function will not be called
+        auto p = Lookup<NoCompare>(entry.h, entries_, new_mask,
+                                   [](const Payload*) { return false; });
+        // Lookup<NoCompare> (and CompareEntry<NoCompare>) ensure that an
+        // empty slots is always returned
+        assert(!p.second);
+        entries_[p.first] = entry;
+      }
+    }
+    capacity_ = new_capacity;
+    capacity_mask_ = new_mask;
+
+    return Status::OK();
+  }
+
+  hash_t FixHash(hash_t h) const { return (h == kSentinel) ? 42U : h; }
+
+  // The number of slots available in the hash table array.
+  uint64_t capacity_;
+  uint64_t capacity_mask_;
+  // The number of used slots in the hash table array.
+  uint64_t size_;
+
+  Entry* entries_;
+  TypedBufferBuilder<Entry> entries_builder_;
+};
+
+// XXX typedef memo_index_t int32_t ?
+
+constexpr int32_t kKeyNotFound = -1;
+
+// ----------------------------------------------------------------------
+// A base class for memoization table.
+
+class MemoTable {
+ public:
+  virtual ~MemoTable() = default;
+
+  virtual int32_t size() const = 0;
+};
+
+// ----------------------------------------------------------------------
+// A memoization table for memory-cheap scalar values.
+
+// The memoization table remembers and allows to look up the insertion
+// index for each key.
+
+template <typename Scalar, template <class> class HashTableTemplateType = HashTable>
+class ScalarMemoTable : public MemoTable {
+ public:
+  explicit ScalarMemoTable(MemoryPool* pool, int64_t entries = 0)
+      : hash_table_(pool, static_cast<uint64_t>(entries)) {}
+
+  template <typename Value>
+  int32_t Get(Value&& v) const {
+    const Scalar value(std::forward<Value>(v));
+    auto cmp_func = [value](const Payload* payload) -> bool {
+      return ScalarHelper<Scalar, 0>::CompareScalars(payload->value, value);
+    };
+    hash_t h = ComputeHash(value);
+    auto p = hash_table_.Lookup(h, cmp_func);
+    if (p.second) {
+      return p.first->payload.memo_index;
+    } else {
+      return kKeyNotFound;
+    }
+  }
+
+  template <typename Value, typename Func1, typename Func2>
+  Status GetOrInsert(Value&& v, Func1&& on_found, Func2&& on_not_found,
+                     int32_t* out_memo_index) {
+    const Scalar value(std::forward<Value>(v));
+    auto cmp_func = [value](const Payload* payload) -> bool {
+      return ScalarHelper<Scalar, 0>::CompareScalars(value, payload->value);
+    };
+    hash_t h = ComputeHash(value);
+    auto p = hash_table_.Lookup(h, cmp_func);
+    int32_t memo_index;
+    if (p.second) {
+      memo_index = p.first->payload.memo_index;
+      on_found(memo_index);
+    } else {
+      memo_index = size();
+      RETURN_NOT_OK(hash_table_.Insert(p.first, h, {value, memo_index}));
+      on_not_found(memo_index);
+    }
+    *out_memo_index = memo_index;
+    return Status::OK();
+  }
+
+  template <typename Value>
+  Status GetOrInsert(Value&& value, int32_t* out_memo_index) {
+    return GetOrInsert(
+        value, [](int32_t i) {}, [](int32_t i) {}, out_memo_index);
+  }
+
+  int32_t GetNull() const { return null_index_; }
+
+  template <typename Func1, typename Func2>
+  int32_t GetOrInsertNull(Func1&& on_found, Func2&& on_not_found) {
+    int32_t memo_index = GetNull();
+    if (memo_index != kKeyNotFound) {
+      on_found(memo_index);
+    } else {
+      null_index_ = memo_index = size();
+      on_not_found(memo_index);
+    }
+    return memo_index;
+  }
+
+  int32_t GetOrInsertNull() {
+    return GetOrInsertNull([](int32_t i) {}, [](int32_t i) {});
+  }
+
+  // The number of entries in the memo table +1 if null was added.
+  // (which is also 1 + the largest memo index)
+  int32_t size() const override {
+    return static_cast<int32_t>(hash_table_.size()) + (GetNull() != kKeyNotFound);
+  }
+
+  // Copy values starting from index `start` into `out_data`
+  template <typename Value>
+  void CopyValues(int32_t start, Value* out_data) const {
+    // So that both uint16_t and Float16 are allowed
+    static_assert(sizeof(Value) == sizeof(Scalar));
+    Scalar* out = reinterpret_cast<Scalar*>(out_data);
+    hash_table_.VisitEntries([=](const HashTableEntry* entry) {
+      int32_t index = entry->payload.memo_index - start;
+      if (index >= 0) {
+        out[index] = entry->payload.value;
+      }
+    });
+    // Zero-initialize the null entry
+    if (null_index_ != kKeyNotFound) {
+      int32_t index = null_index_ - start;
+      if (index >= 0) {
+        out[index] = Scalar{};
+      }
+    }
+  }
+
+  template <typename Value>
+  void CopyValues(Value* out_data) const {
+    CopyValues(0, out_data);
+  }
+
+ protected:
+  struct Payload {
+    Scalar value;
+    int32_t memo_index;
+  };
+
+  using HashTableType = HashTableTemplateType<Payload>;
+  using HashTableEntry = typename HashTableType::Entry;
+  HashTableType hash_table_;
+  int32_t null_index_ = kKeyNotFound;
+
+  hash_t ComputeHash(const Scalar& value) const {
+    return ScalarHelper<Scalar, 0>::ComputeHash(value);
+  }
+
+ public:
+  // defined here so that `HashTableType` is visible
+  // Merge entries from `other_table` into `this->hash_table_`.
+  Status MergeTable(const ScalarMemoTable& other_table) {
+    const HashTableType& other_hashtable = other_table.hash_table_;
+
+    other_hashtable.VisitEntries([this](const HashTableEntry* other_entry) {
+      int32_t unused;
+      ARROW_DCHECK_OK(this->GetOrInsert(other_entry->payload.value, &unused));
+    });
+    // TODO: ARROW-17074 - implement proper error handling
+    return Status::OK();
+  }
+};
+
+// ----------------------------------------------------------------------
+// A memoization table for small scalar values, using direct indexing
+
+template <typename Scalar, typename Enable = void>
+struct SmallScalarTraits {};
+
+template <>
+struct SmallScalarTraits<bool> {
+  static constexpr int32_t cardinality = 2;
+
+  static uint32_t AsIndex(bool value) { return value ? 1 : 0; }
+};
+
+template <typename Scalar>
+struct SmallScalarTraits<Scalar, enable_if_t<std::is_integral<Scalar>::value>> {
+  using Unsigned = typename std::make_unsigned<Scalar>::type;
+
+  static constexpr int32_t cardinality = 1U + std::numeric_limits<Unsigned>::max();
+
+  static uint32_t AsIndex(Scalar value) { return static_cast<Unsigned>(value); }
+};
+
+template <typename Scalar, template <class> class HashTableTemplateType = HashTable>
+class SmallScalarMemoTable : public MemoTable {
+ public:
+  explicit SmallScalarMemoTable(MemoryPool* pool, int64_t entries = 0) {
+    std::fill(value_to_index_, value_to_index_ + cardinality + 1, kKeyNotFound);
+    index_to_value_.reserve(cardinality);
+  }
+
+  int32_t Get(const Scalar value) const {
+    auto value_index = AsIndex(value);
+    return value_to_index_[value_index];
+  }
+
+  template <typename Func1, typename Func2>
+  Status GetOrInsert(const Scalar value, Func1&& on_found, Func2&& on_not_found,
+                     int32_t* out_memo_index) {
+    auto value_index = AsIndex(value);
+    auto memo_index = value_to_index_[value_index];
+    if (memo_index == kKeyNotFound) {
+      memo_index = static_cast<int32_t>(index_to_value_.size());
+      index_to_value_.push_back(value);
+      value_to_index_[value_index] = memo_index;
+      ARROW_DCHECK_LT(memo_index, cardinality + 1);
+      on_not_found(memo_index);
+    } else {
+      on_found(memo_index);
+    }
+    *out_memo_index = memo_index;
+    return Status::OK();
+  }
+
+  Status GetOrInsert(const Scalar value, int32_t* out_memo_index) {
+    return GetOrInsert(
+        value, [](int32_t i) {}, [](int32_t i) {}, out_memo_index);
+  }
+
+  int32_t GetNull() const { return value_to_index_[cardinality]; }
+
+  template <typename Func1, typename Func2>
+  int32_t GetOrInsertNull(Func1&& on_found, Func2&& on_not_found) {
+    auto memo_index = GetNull();
+    if (memo_index == kKeyNotFound) {
+      memo_index = value_to_index_[cardinality] = size();
+      index_to_value_.push_back(0);
+      on_not_found(memo_index);
+    } else {
+      on_found(memo_index);
+    }
+    return memo_index;
+  }
+
+  int32_t GetOrInsertNull() {
+    return GetOrInsertNull([](int32_t i) {}, [](int32_t i) {});
+  }
+
+  // The number of entries in the memo table
+  // (which is also 1 + the largest memo index)
+  int32_t size() const override { return static_cast<int32_t>(index_to_value_.size()); }
+
+  // Merge entries from `other_table` into `this`.
+  Status MergeTable(const SmallScalarMemoTable& other_table) {
+    for (const Scalar& other_val : other_table.index_to_value_) {
+      int32_t unused;
+      RETURN_NOT_OK(this->GetOrInsert(other_val, &unused));
+    }
+    return Status::OK();
+  }
+
+  // Copy values starting from index `start` into `out_data`
+  void CopyValues(int32_t start, Scalar* out_data) const {
+    ARROW_DCHECK_GE(start, 0);
+    ARROW_DCHECK_LE(static_cast<size_t>(start), index_to_value_.size());
+    int64_t offset = start * static_cast<int32_t>(sizeof(Scalar));
+    memcpy(out_data, index_to_value_.data() + offset, (size() - start) * sizeof(Scalar));
+  }
+
+  void CopyValues(Scalar* out_data) const { CopyValues(0, out_data); }
+
+  const std::vector<Scalar>& values() const { return index_to_value_; }
+
+ protected:
+  static constexpr auto cardinality = SmallScalarTraits<Scalar>::cardinality;
+  static_assert(cardinality <= 256, "cardinality too large for direct-addressed table");
+
+  uint32_t AsIndex(Scalar value) const {
+    return SmallScalarTraits<Scalar>::AsIndex(value);
+  }
+
+  // The last index is reserved for the null element.
+  int32_t value_to_index_[cardinality + 1];
+  std::vector<Scalar> index_to_value_;
+};
+
+// ----------------------------------------------------------------------
+// A memoization table for variable-sized binary data.
+
+template <typename BinaryBuilderT>
+class BinaryMemoTable : public MemoTable {
+ public:
+  using builder_offset_type = typename BinaryBuilderT::offset_type;
+  explicit BinaryMemoTable(MemoryPool* pool, int64_t entries = 0,
+                           int64_t values_size = -1)
+      : hash_table_(pool, static_cast<uint64_t>(entries)), binary_builder_(pool) {
+    const int64_t data_size = (values_size < 0) ? entries * 4 : values_size;
+    ARROW_DCHECK_OK(binary_builder_.Resize(entries));
+    ARROW_DCHECK_OK(binary_builder_.ReserveData(data_size));
+  }
+
+  int32_t Get(const void* data, builder_offset_type length) const {
+    hash_t h = ComputeStringHash<0>(data, length);
+    auto p = Lookup(h, data, length);
+    if (p.second) {
+      return p.first->payload.memo_index;
+    } else {
+      return kKeyNotFound;
+    }
+  }
+
+  int32_t Get(std::string_view value) const {
+    return Get(value.data(), static_cast<builder_offset_type>(value.length()));
+  }
+
+  template <typename Func1, typename Func2>
+  Status GetOrInsert(const void* data, builder_offset_type length, Func1&& on_found,
+                     Func2&& on_not_found, int32_t* out_memo_index) {
+    hash_t h = ComputeStringHash<0>(data, length);
+    auto p = Lookup(h, data, length);
+    int32_t memo_index;
+    if (p.second) {
+      memo_index = p.first->payload.memo_index;
+      on_found(memo_index);
+    } else {
+      memo_index = size();
+      // Insert string value
+      RETURN_NOT_OK(binary_builder_.Append(static_cast<const char*>(data), length));
+      // Insert hash entry
+      RETURN_NOT_OK(
+          hash_table_.Insert(const_cast<HashTableEntry*>(p.first), h, {memo_index}));
+
+      on_not_found(memo_index);
+    }
+    *out_memo_index = memo_index;
+    return Status::OK();
+  }
+
+  template <typename Func1, typename Func2>
+  Status GetOrInsert(std::string_view value, Func1&& on_found, Func2&& on_not_found,
+                     int32_t* out_memo_index) {
+    return GetOrInsert(value.data(), static_cast<builder_offset_type>(value.length()),
+                       std::forward<Func1>(on_found), std::forward<Func2>(on_not_found),
+                       out_memo_index);
+  }
+
+  Status GetOrInsert(const void* data, builder_offset_type length,
+                     int32_t* out_memo_index) {
+    return GetOrInsert(
+        data, length, [](int32_t i) {}, [](int32_t i) {}, out_memo_index);
+  }
+
+  Status GetOrInsert(std::string_view value, int32_t* out_memo_index) {
+    return GetOrInsert(value.data(), static_cast<builder_offset_type>(value.length()),
+                       out_memo_index);
+  }
+
+  int32_t GetNull() const { return null_index_; }
+
+  template <typename Func1, typename Func2>
+  int32_t GetOrInsertNull(Func1&& on_found, Func2&& on_not_found) {
+    int32_t memo_index = GetNull();
+    if (memo_index == kKeyNotFound) {
+      memo_index = null_index_ = size();
+      ARROW_DCHECK_OK(binary_builder_.AppendNull());
+      on_not_found(memo_index);
+    } else {
+      on_found(memo_index);
+    }
+    return memo_index;
+  }
+
+  int32_t GetOrInsertNull() {
+    return GetOrInsertNull([](int32_t i) {}, [](int32_t i) {});
+  }
+
+  // The number of entries in the memo table
+  // (which is also 1 + the largest memo index)
+  int32_t size() const override {
+    return static_cast<int32_t>(hash_table_.size() + (GetNull() != kKeyNotFound));
+  }
+
+  int64_t values_size() const { return binary_builder_.value_data_length(); }
+
+  // Copy (n + 1) offsets starting from index `start` into `out_data`
+  template <class Offset>
+  void CopyOffsets(int32_t start, Offset* out_data) const {
+    ARROW_DCHECK_LE(start, size());
+
+    const builder_offset_type* offsets = binary_builder_.offsets_data();
+    const builder_offset_type delta =
+        start < binary_builder_.length() ? offsets[start] : 0;
+    for (int32_t i = start; i < size(); ++i) {
+      const builder_offset_type adjusted_offset = offsets[i] - delta;
+      Offset cast_offset = static_cast<Offset>(adjusted_offset);
+      assert(static_cast<builder_offset_type>(cast_offset) ==
+             adjusted_offset);  // avoid truncation
+      *out_data++ = cast_offset;
+    }
+
+    // Copy last value since BinaryBuilder only materializes it on in Finish()
+    *out_data = static_cast<Offset>(binary_builder_.value_data_length() - delta);
+  }
+
+  template <class Offset>
+  void CopyOffsets(Offset* out_data) const {
+    CopyOffsets(0, out_data);
+  }
+
+  // Copy values starting from index `start` into `out_data`
+  void CopyValues(int32_t start, uint8_t* out_data) const {
+    CopyValues(start, -1, out_data);
+  }
+
+  // Same as above, but check output size in debug mode
+  void CopyValues(int32_t start, int64_t out_size, uint8_t* out_data) const {
+    ARROW_DCHECK_LE(start, size());
+
+    // The absolute byte offset of `start` value in the binary buffer.
+    const builder_offset_type offset = binary_builder_.offset(start);
+    const auto length = binary_builder_.value_data_length() - static_cast<size_t>(offset);
+
+    if (out_size != -1) {
+      assert(static_cast<int64_t>(length) <= out_size);
+    }
+
+    auto view = binary_builder_.GetView(start);
+    memcpy(out_data, view.data(), length);
+  }
+
+  void CopyValues(uint8_t* out_data) const { CopyValues(0, -1, out_data); }
+
+  void CopyValues(int64_t out_size, uint8_t* out_data) const {
+    CopyValues(0, out_size, out_data);
+  }
+
+  void CopyFixedWidthValues(int32_t start, int32_t width_size, int64_t out_size,
+                            uint8_t* out_data) const {
+    // This method exists to cope with the fact that the BinaryMemoTable does
+    // not know the fixed width when inserting the null value. The data
+    // buffer hold a zero length string for the null value (if found).
+    //
+    // Thus, the method will properly inject an empty value of the proper width
+    // in the output buffer.
+    //
+    if (start >= size()) {
+      return;
+    }
+
+    int32_t null_index = GetNull();
+    if (null_index < start) {
+      // Nothing to skip, proceed as usual.
+      CopyValues(start, out_size, out_data);
+      return;
+    }
+
+    builder_offset_type left_offset = binary_builder_.offset(start);
+
+    // Ensure that the data length is exactly missing width_size bytes to fit
+    // in the expected output (n_values * width_size).
+#ifndef NDEBUG
+    int64_t data_length = values_size() - static_cast<size_t>(left_offset);
+    assert(data_length + width_size == out_size);
+    ARROW_UNUSED(data_length);
+#endif
+
+    auto in_data = binary_builder_.value_data() + left_offset;
+    // The null use 0-length in the data, slice the data in 2 and skip by
+    // width_size in out_data. [part_1][width_size][part_2]
+    auto null_data_offset = binary_builder_.offset(null_index);
+    auto left_size = null_data_offset - left_offset;
+    if (left_size > 0) {
+      memcpy(out_data, in_data + left_offset, left_size);
+    }
+    // Zero-initialize the null entry
+    memset(out_data + left_size, 0, width_size);
+
+    auto right_size = values_size() - static_cast<size_t>(null_data_offset);
+    if (right_size > 0) {
+      // skip the null fixed size value.
+      auto out_offset = left_size + width_size;
+      assert(out_data + out_offset + right_size == out_data + out_size);
+      memcpy(out_data + out_offset, in_data + null_data_offset, right_size);
+    }
+  }
+
+  // Visit the stored values in insertion order.
+  // The visitor function should have the signature `void(std::string_view)`
+  // or `void(const std::string_view&)`.
+  template <typename VisitFunc>
+  void VisitValues(int32_t start, VisitFunc&& visit) const {
+    for (int32_t i = start; i < size(); ++i) {
+      visit(binary_builder_.GetView(i));
+    }
+  }
+
+  // Visit the stored value at a specific index in insertion order.
+  // The visitor function should have the signature `void(std::string_view)`
+  // or `void(const std::string_view&)`.
+  template <typename VisitFunc>
+  void VisitValue(int32_t idx, VisitFunc&& visit) const {
+    visit(binary_builder_.GetView(idx));
+  }
+
+ protected:
+  struct Payload {
+    int32_t memo_index;
+  };
+
+  using HashTableType = HashTable<Payload>;
+  using HashTableEntry = typename HashTable<Payload>::Entry;
+  HashTableType hash_table_;
+  BinaryBuilderT binary_builder_;
+
+  int32_t null_index_ = kKeyNotFound;
+
+  std::pair<const HashTableEntry*, bool> Lookup(hash_t h, const void* data,
+                                                builder_offset_type length) const {
+    auto cmp_func = [&](const Payload* payload) {
+      std::string_view lhs = binary_builder_.GetView(payload->memo_index);
+      std::string_view rhs(static_cast<const char*>(data), length);
+      return lhs == rhs;
+    };
+    return hash_table_.Lookup(h, cmp_func);
+  }
+
+ public:
+  Status MergeTable(const BinaryMemoTable& other_table) {
+    other_table.VisitValues(0, [this](std::string_view other_value) {
+      int32_t unused;
+      ARROW_DCHECK_OK(this->GetOrInsert(other_value, &unused));
+    });
+    return Status::OK();
+  }
+};
+
+template <typename T, typename Enable = void>
+struct HashTraits {};
+
+template <>
+struct HashTraits<BooleanType> {
+  using MemoTableType = SmallScalarMemoTable<bool>;
+};
+
+template <typename T>
+struct HashTraits<T, enable_if_8bit_int<T>> {
+  using c_type = typename T::c_type;
+  using MemoTableType = SmallScalarMemoTable<typename T::c_type>;
+};
+
+template <typename T>
+struct HashTraits<T, enable_if_t<has_c_type<T>::value && !is_8bit_int<T>::value>> {
+  using c_type = typename T::c_type;
+  using MemoTableType = ScalarMemoTable<c_type, HashTable>;
+};
+
+template <>
+struct HashTraits<HalfFloatType> {
+  using MemoTableType = ScalarMemoTable<::arrow::util::Float16>;
+};
+
+template <typename T>
+struct HashTraits<T, enable_if_t<has_string_view<T>::value &&
+                                 !std::is_base_of<LargeBinaryType, T>::value>> {
+  using MemoTableType = BinaryMemoTable<BinaryBuilder>;
+};
+
+template <typename T>
+struct HashTraits<T, enable_if_decimal<T>> {
+  using MemoTableType = BinaryMemoTable<BinaryBuilder>;
+};
+
+template <typename T>
+struct HashTraits<T, enable_if_t<std::is_base_of<LargeBinaryType, T>::value>> {
+  using MemoTableType = BinaryMemoTable<LargeBinaryBuilder>;
+};
+
+template <typename MemoTableType>
+static inline Status ComputeNullBitmap(MemoryPool* pool, const MemoTableType& memo_table,
+                                       int64_t start_offset, int64_t* null_count,
+                                       std::shared_ptr<Buffer>* null_bitmap) {
+  int64_t dict_length = static_cast<int64_t>(memo_table.size()) - start_offset;
+  int64_t null_index = memo_table.GetNull();
+
+  *null_count = 0;
+  *null_bitmap = nullptr;
+
+  if (null_index != kKeyNotFound && null_index >= start_offset) {
+    null_index -= start_offset;
+    *null_count = 1;
+    ARROW_ASSIGN_OR_RAISE(*null_bitmap,
+                          internal::BitmapAllButOne(pool, dict_length, null_index));
+  }
+
+  return Status::OK();
+}
+
+struct StringViewHash {
+  // std::hash compatible hasher for use with std::unordered_*
+  // (the std::hash specialization provided by nonstd constructs std::string
+  // temporaries then invokes std::hash<std::string> against those)
+  hash_t operator()(std::string_view value) const {
+    return ComputeStringHash<0>(value.data(), static_cast<int64_t>(value.size()));
+  }
+};
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/int_util.h

@@ -0,0 +1,137 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cstdint>
+#include <type_traits>
+
+#include "arrow/status.h"
+
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+
+class DataType;
+struct ArraySpan;
+struct Scalar;
+
+namespace internal {
+
+ARROW_EXPORT
+uint8_t DetectUIntWidth(const uint64_t* values, int64_t length, uint8_t min_width = 1);
+
+ARROW_EXPORT
+uint8_t DetectUIntWidth(const uint64_t* values, const uint8_t* valid_bytes,
+                        int64_t length, uint8_t min_width = 1);
+
+ARROW_EXPORT
+uint8_t DetectIntWidth(const int64_t* values, int64_t length, uint8_t min_width = 1);
+
+ARROW_EXPORT
+uint8_t DetectIntWidth(const int64_t* values, const uint8_t* valid_bytes, int64_t length,
+                       uint8_t min_width = 1);
+
+ARROW_EXPORT
+void DowncastInts(const int64_t* source, int8_t* dest, int64_t length);
+
+ARROW_EXPORT
+void DowncastInts(const int64_t* source, int16_t* dest, int64_t length);
+
+ARROW_EXPORT
+void DowncastInts(const int64_t* source, int32_t* dest, int64_t length);
+
+ARROW_EXPORT
+void DowncastInts(const int64_t* source, int64_t* dest, int64_t length);
+
+ARROW_EXPORT
+void DowncastUInts(const uint64_t* source, uint8_t* dest, int64_t length);
+
+ARROW_EXPORT
+void DowncastUInts(const uint64_t* source, uint16_t* dest, int64_t length);
+
+ARROW_EXPORT
+void DowncastUInts(const uint64_t* source, uint32_t* dest, int64_t length);
+
+ARROW_EXPORT
+void DowncastUInts(const uint64_t* source, uint64_t* dest, int64_t length);
+
+ARROW_EXPORT
+void UpcastInts(const int32_t* source, int64_t* dest, int64_t length);
+
+template <typename InputInt, typename OutputInt>
+inline typename std::enable_if<(sizeof(InputInt) >= sizeof(OutputInt))>::type CastInts(
+    const InputInt* source, OutputInt* dest, int64_t length) {
+  DowncastInts(source, dest, length);
+}
+
+template <typename InputInt, typename OutputInt>
+inline typename std::enable_if<(sizeof(InputInt) < sizeof(OutputInt))>::type CastInts(
+    const InputInt* source, OutputInt* dest, int64_t length) {
+  UpcastInts(source, dest, length);
+}
+
+template <typename InputInt, typename OutputInt>
+ARROW_EXPORT void TransposeInts(const InputInt* source, OutputInt* dest, int64_t length,
+                                const int32_t* transpose_map);
+
+ARROW_EXPORT
+Status TransposeInts(const DataType& src_type, const DataType& dest_type,
+                     const uint8_t* src, uint8_t* dest, int64_t src_offset,
+                     int64_t dest_offset, int64_t length, const int32_t* transpose_map);
+
+/// \brief Do vectorized boundschecking of integer-type array indices. The
+/// indices must be nonnegative and strictly less than the passed upper
+/// limit (which is usually the length of an array that is being indexed-into).
+ARROW_EXPORT
+Status CheckIndexBounds(const ArraySpan& values, uint64_t upper_limit);
+
+/// \brief Boundscheck integer values to determine if they are all between the
+/// passed upper and lower limits (inclusive). Upper and lower bounds must be
+/// the same type as the data and are not currently casted.
+ARROW_EXPORT
+Status CheckIntegersInRange(const ArraySpan& values, const Scalar& bound_lower,
+                            const Scalar& bound_upper);
+
+/// \brief Use CheckIntegersInRange to determine whether the passed integers
+/// can fit safely in the passed integer type. This helps quickly determine if
+/// integer narrowing (e.g. int64->int32) is safe to do.
+ARROW_EXPORT
+Status IntegersCanFit(const ArraySpan& values, const DataType& target_type);
+
+/// \brief Convenience for boundschecking a single Scalar value
+ARROW_EXPORT
+Status IntegersCanFit(const Scalar& value, const DataType& target_type);
+
+/// Upcast an integer to the largest possible width (currently 64 bits)
+
+template <typename Integer>
+typename std::enable_if<
+    std::is_integral<Integer>::value && std::is_signed<Integer>::value, int64_t>::type
+UpcastInt(Integer v) {
+  return v;
+}
+
+template <typename Integer>
+typename std::enable_if<
+    std::is_integral<Integer>::value && std::is_unsigned<Integer>::value, uint64_t>::type
+UpcastInt(Integer v) {
+  return v;
+}
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/int_util_overflow.h

@@ -0,0 +1,118 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cstdint>
+#include <limits>
+#include <type_traits>
+
+#include "arrow/status.h"
+#include "arrow/util/macros.h"
+#include "arrow/util/visibility.h"
+
+// "safe-math.h" includes <intsafe.h> from the Windows headers.
+#include "arrow/util/windows_compatibility.h"
+#include "arrow/vendored/portable-snippets/safe-math.h"
+// clang-format off (avoid include reordering)
+#include "arrow/util/windows_fixup.h"
+// clang-format on
+
+namespace arrow {
+namespace internal {
+
+// Define functions AddWithOverflow, SubtractWithOverflow, MultiplyWithOverflow
+// with the signature `bool(T u, T v, T* out)` where T is an integer type.
+// On overflow, these functions return true.  Otherwise, false is returned
+// and `out` is updated with the result of the operation.
+
+#define OP_WITH_OVERFLOW(_func_name, _psnip_op, _type, _psnip_type)           \
+  [[nodiscard]] static inline bool _func_name(_type u, _type v, _type* out) { \
+    return !psnip_safe_##_psnip_type##_##_psnip_op(out, u, v);                \
+  }
+
+#define OPS_WITH_OVERFLOW(_func_name, _psnip_op)            \
+  OP_WITH_OVERFLOW(_func_name, _psnip_op, int8_t, int8)     \
+  OP_WITH_OVERFLOW(_func_name, _psnip_op, int16_t, int16)   \
+  OP_WITH_OVERFLOW(_func_name, _psnip_op, int32_t, int32)   \
+  OP_WITH_OVERFLOW(_func_name, _psnip_op, int64_t, int64)   \
+  OP_WITH_OVERFLOW(_func_name, _psnip_op, uint8_t, uint8)   \
+  OP_WITH_OVERFLOW(_func_name, _psnip_op, uint16_t, uint16) \
+  OP_WITH_OVERFLOW(_func_name, _psnip_op, uint32_t, uint32) \
+  OP_WITH_OVERFLOW(_func_name, _psnip_op, uint64_t, uint64)
+
+OPS_WITH_OVERFLOW(AddWithOverflow, add)
+OPS_WITH_OVERFLOW(SubtractWithOverflow, sub)
+OPS_WITH_OVERFLOW(MultiplyWithOverflow, mul)
+OPS_WITH_OVERFLOW(DivideWithOverflow, div)
+
+#undef OP_WITH_OVERFLOW
+#undef OPS_WITH_OVERFLOW
+
+// Define function NegateWithOverflow with the signature `bool(T u, T* out)`
+// where T is a signed integer type.  On overflow, these functions return true.
+// Otherwise, false is returned and `out` is updated with the result of the
+// operation.
+
+#define UNARY_OP_WITH_OVERFLOW(_func_name, _psnip_op, _type, _psnip_type) \
+  [[nodiscard]] static inline bool _func_name(_type u, _type* out) {      \
+    return !psnip_safe_##_psnip_type##_##_psnip_op(out, u);               \
+  }
+
+#define SIGNED_UNARY_OPS_WITH_OVERFLOW(_func_name, _psnip_op)   \
+  UNARY_OP_WITH_OVERFLOW(_func_name, _psnip_op, int8_t, int8)   \
+  UNARY_OP_WITH_OVERFLOW(_func_name, _psnip_op, int16_t, int16) \
+  UNARY_OP_WITH_OVERFLOW(_func_name, _psnip_op, int32_t, int32) \
+  UNARY_OP_WITH_OVERFLOW(_func_name, _psnip_op, int64_t, int64)
+
+SIGNED_UNARY_OPS_WITH_OVERFLOW(NegateWithOverflow, neg)
+
+#undef UNARY_OP_WITH_OVERFLOW
+#undef SIGNED_UNARY_OPS_WITH_OVERFLOW
+
+/// Signed addition with well-defined behaviour on overflow (as unsigned)
+template <typename SignedInt>
+SignedInt SafeSignedAdd(SignedInt u, SignedInt v) {
+  using UnsignedInt = typename std::make_unsigned<SignedInt>::type;
+  return static_cast<SignedInt>(static_cast<UnsignedInt>(u) +
+                                static_cast<UnsignedInt>(v));
+}
+
+/// Signed subtraction with well-defined behaviour on overflow (as unsigned)
+template <typename SignedInt>
+SignedInt SafeSignedSubtract(SignedInt u, SignedInt v) {
+  using UnsignedInt = typename std::make_unsigned<SignedInt>::type;
+  return static_cast<SignedInt>(static_cast<UnsignedInt>(u) -
+                                static_cast<UnsignedInt>(v));
+}
+
+/// Signed negation with well-defined behaviour on overflow (as unsigned)
+template <typename SignedInt>
+SignedInt SafeSignedNegate(SignedInt u) {
+  using UnsignedInt = typename std::make_unsigned<SignedInt>::type;
+  return static_cast<SignedInt>(~static_cast<UnsignedInt>(u) + 1);
+}
+
+/// Signed left shift with well-defined behaviour on negative numbers or overflow
+template <typename SignedInt, typename Shift>
+SignedInt SafeLeftShift(SignedInt u, Shift shift) {
+  using UnsignedInt = typename std::make_unsigned<SignedInt>::type;
+  return static_cast<SignedInt>(static_cast<UnsignedInt>(u) << shift);
+}
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/io_util.h

@@ -0,0 +1,458 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#ifndef _WIN32
+#  define ARROW_HAVE_SIGACTION 1
+#endif
+
+#include <atomic>
+#include <memory>
+#include <optional>
+#include <string>
+#include <utility>
+#include <vector>
+
+#if ARROW_HAVE_SIGACTION
+#  include <csignal>  // Needed for struct sigaction
+#endif
+
+#include "arrow/result.h"
+#include "arrow/status.h"
+#include "arrow/type_fwd.h"
+#include "arrow/util/macros.h"
+#include "arrow/util/windows_fixup.h"
+
+namespace arrow::internal {
+
+// NOTE: 8-bit path strings on Windows are encoded using UTF-8.
+// Using MBCS would fail encoding some paths.
+
+#if defined(_WIN32)
+using NativePathString = std::wstring;
+#else
+using NativePathString = std::string;
+#endif
+
+class ARROW_EXPORT PlatformFilename {
+ public:
+  struct Impl;
+
+  ~PlatformFilename();
+  PlatformFilename();
+  PlatformFilename(const PlatformFilename&);
+  PlatformFilename(PlatformFilename&&);
+  PlatformFilename& operator=(const PlatformFilename&);
+  PlatformFilename& operator=(PlatformFilename&&);
+  explicit PlatformFilename(NativePathString path);
+  explicit PlatformFilename(const NativePathString::value_type* path);
+
+  const NativePathString& ToNative() const;
+  std::string ToString() const;
+
+  PlatformFilename Parent() const;
+  Result<PlatformFilename> Real() const;
+
+  // These functions can fail for character encoding reasons.
+  static Result<PlatformFilename> FromString(std::string_view file_name);
+  Result<PlatformFilename> Join(std::string_view child_name) const;
+
+  PlatformFilename Join(const PlatformFilename& child_name) const;
+
+  bool operator==(const PlatformFilename& other) const;
+  bool operator!=(const PlatformFilename& other) const;
+
+  // Made public to avoid the proliferation of friend declarations.
+  const Impl* impl() const { return impl_.get(); }
+
+ private:
+  std::unique_ptr<Impl> impl_;
+
+  explicit PlatformFilename(Impl impl);
+};
+
+/// Create a directory if it doesn't exist.
+///
+/// Return whether the directory was created.
+ARROW_EXPORT
+Result<bool> CreateDir(const PlatformFilename& dir_path);
+
+/// Create a directory and its parents if it doesn't exist.
+///
+/// Return whether the directory was created.
+ARROW_EXPORT
+Result<bool> CreateDirTree(const PlatformFilename& dir_path);
+
+/// Delete a directory's contents (but not the directory itself) if it exists.
+///
+/// Return whether the directory existed.
+ARROW_EXPORT
+Result<bool> DeleteDirContents(const PlatformFilename& dir_path,
+                               bool allow_not_found = true);
+
+/// Delete a directory tree if it exists.
+///
+/// Return whether the directory existed.
+ARROW_EXPORT
+Result<bool> DeleteDirTree(const PlatformFilename& dir_path, bool allow_not_found = true);
+
+// Non-recursively list the contents of the given directory.
+// The returned names are the children's base names, not including dir_path.
+ARROW_EXPORT
+Result<std::vector<PlatformFilename>> ListDir(const PlatformFilename& dir_path);
+
+/// Delete a file if it exists.
+///
+/// Return whether the file existed.
+ARROW_EXPORT
+Result<bool> DeleteFile(const PlatformFilename& file_path, bool allow_not_found = true);
+
+/// Return whether a file exists.
+ARROW_EXPORT
+Result<bool> FileExists(const PlatformFilename& path);
+
+// TODO expose this more publicly to make it available from io/file.h?
+/// A RAII wrapper for a file descriptor.
+///
+/// The underlying file descriptor is automatically closed on destruction.
+/// Moving is supported with well-defined semantics.
+/// Furthermore, closing is idempotent.
+class ARROW_EXPORT FileDescriptor {
+ public:
+  FileDescriptor() = default;
+  explicit FileDescriptor(int fd) : fd_(fd) {}
+  FileDescriptor(FileDescriptor&&);
+  FileDescriptor& operator=(FileDescriptor&&);
+
+  ~FileDescriptor();
+
+  Status Close();
+
+  /// May return -1 if closed or default-initialized
+  int fd() const { return fd_.load(); }
+
+  /// Detach and return the underlying file descriptor
+  int Detach();
+
+  bool closed() const { return fd_.load() == -1; }
+
+ protected:
+  static void CloseFromDestructor(int fd);
+
+  std::atomic<int> fd_{-1};
+};
+
+/// Open a file for reading and return a file descriptor.
+ARROW_EXPORT
+Result<FileDescriptor> FileOpenReadable(const PlatformFilename& file_name);
+
+/// Open a file for writing and return a file descriptor.
+ARROW_EXPORT
+Result<FileDescriptor> FileOpenWritable(const PlatformFilename& file_name,
+                                        bool write_only = true, bool truncate = true,
+                                        bool append = false);
+
+/// Read from current file position.  Return number of bytes read.
+ARROW_EXPORT
+Result<int64_t> FileRead(int fd, uint8_t* buffer, int64_t nbytes);
+/// Read from given file position.  Return number of bytes read.
+ARROW_EXPORT
+Result<int64_t> FileReadAt(int fd, uint8_t* buffer, int64_t position, int64_t nbytes);
+
+ARROW_EXPORT
+Status FileWrite(int fd, const uint8_t* buffer, const int64_t nbytes);
+ARROW_EXPORT
+Status FileTruncate(int fd, const int64_t size);
+
+ARROW_EXPORT
+Status FileSeek(int fd, int64_t pos);
+ARROW_EXPORT
+Status FileSeek(int fd, int64_t pos, int whence);
+ARROW_EXPORT
+Result<int64_t> FileTell(int fd);
+ARROW_EXPORT
+Result<int64_t> FileGetSize(int fd);
+
+ARROW_EXPORT
+Status FileClose(int fd);
+
+struct Pipe {
+  FileDescriptor rfd;
+  FileDescriptor wfd;
+
+  Status Close() { return rfd.Close() & wfd.Close(); }
+};
+
+ARROW_EXPORT
+Result<Pipe> CreatePipe();
+
+ARROW_EXPORT
+Status SetPipeFileDescriptorNonBlocking(int fd);
+
+class ARROW_EXPORT SelfPipe {
+ public:
+  static Result<std::shared_ptr<SelfPipe>> Make(bool signal_safe);
+  virtual ~SelfPipe();
+
+  /// \brief Wait for a wakeup.
+  ///
+  /// Status::Invalid is returned if the pipe has been shutdown.
+  /// Otherwise the next sent payload is returned.
+  virtual Result<uint64_t> Wait() = 0;
+
+  /// \brief Wake up the pipe by sending a payload.
+  ///
+  /// This method is async-signal-safe if `signal_safe` was set to true.
+  virtual void Send(uint64_t payload) = 0;
+
+  /// \brief Wake up the pipe and shut it down.
+  virtual Status Shutdown() = 0;
+};
+
+ARROW_EXPORT
+int64_t GetPageSize();
+
+struct MemoryRegion {
+  void* addr;
+  size_t size;
+};
+
+ARROW_EXPORT
+Status MemoryMapRemap(void* addr, size_t old_size, size_t new_size, int fildes,
+                      void** new_addr);
+ARROW_EXPORT
+Status MemoryAdviseWillNeed(const std::vector<MemoryRegion>& regions);
+
+ARROW_EXPORT
+Result<std::string> GetEnvVar(const char* name);
+ARROW_EXPORT
+Result<std::string> GetEnvVar(const std::string& name);
+ARROW_EXPORT
+Result<NativePathString> GetEnvVarNative(const char* name);
+ARROW_EXPORT
+Result<NativePathString> GetEnvVarNative(const std::string& name);
+
+ARROW_EXPORT
+Status SetEnvVar(const char* name, const char* value);
+ARROW_EXPORT
+Status SetEnvVar(const std::string& name, const std::string& value);
+ARROW_EXPORT
+Status DelEnvVar(const char* name);
+ARROW_EXPORT
+Status DelEnvVar(const std::string& name);
+
+ARROW_EXPORT
+std::string ErrnoMessage(int errnum);
+#if _WIN32
+ARROW_EXPORT
+std::string WinErrorMessage(int errnum);
+#endif
+
+ARROW_EXPORT
+std::shared_ptr<StatusDetail> StatusDetailFromErrno(int errnum);
+ARROW_EXPORT
+std::optional<int> ErrnoFromStatusDetail(const StatusDetail& detail);
+#if _WIN32
+ARROW_EXPORT
+std::shared_ptr<StatusDetail> StatusDetailFromWinError(int errnum);
+#endif
+ARROW_EXPORT
+std::shared_ptr<StatusDetail> StatusDetailFromSignal(int signum);
+
+template <typename... Args>
+Status StatusFromErrno(int errnum, StatusCode code, Args&&... args) {
+  return Status::FromDetailAndArgs(code, StatusDetailFromErrno(errnum),
+                                   std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+Status IOErrorFromErrno(int errnum, Args&&... args) {
+  return StatusFromErrno(errnum, StatusCode::IOError, std::forward<Args>(args)...);
+}
+
+#if _WIN32
+template <typename... Args>
+Status StatusFromWinError(int errnum, StatusCode code, Args&&... args) {
+  return Status::FromDetailAndArgs(code, StatusDetailFromWinError(errnum),
+                                   std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+Status IOErrorFromWinError(int errnum, Args&&... args) {
+  return StatusFromWinError(errnum, StatusCode::IOError, std::forward<Args>(args)...);
+}
+#endif
+
+template <typename... Args>
+Status StatusFromSignal(int signum, StatusCode code, Args&&... args) {
+  return Status::FromDetailAndArgs(code, StatusDetailFromSignal(signum),
+                                   std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+Status CancelledFromSignal(int signum, Args&&... args) {
+  return StatusFromSignal(signum, StatusCode::Cancelled, std::forward<Args>(args)...);
+}
+
+ARROW_EXPORT
+int ErrnoFromStatus(const Status&);
+
+// Always returns 0 on non-Windows platforms (for Python).
+ARROW_EXPORT
+int WinErrorFromStatus(const Status&);
+
+ARROW_EXPORT
+int SignalFromStatus(const Status&);
+
+class ARROW_EXPORT TemporaryDir {
+ public:
+  ~TemporaryDir();
+
+  /// '/'-terminated path to the temporary dir
+  const PlatformFilename& path() { return path_; }
+
+  /// Create a temporary subdirectory in the system temporary dir,
+  /// named starting with `prefix`.
+  static Result<std::unique_ptr<TemporaryDir>> Make(const std::string& prefix);
+
+ private:
+  PlatformFilename path_;
+
+  explicit TemporaryDir(PlatformFilename&&);
+};
+
+class ARROW_EXPORT SignalHandler {
+ public:
+  using Callback = void (*)(int);
+
+  SignalHandler();
+  explicit SignalHandler(Callback cb);
+#if ARROW_HAVE_SIGACTION
+  explicit SignalHandler(const struct sigaction& sa);
+#endif
+
+  Callback callback() const;
+#if ARROW_HAVE_SIGACTION
+  const struct sigaction& action() const;
+#endif
+
+ protected:
+#if ARROW_HAVE_SIGACTION
+  // Storing the full sigaction allows to restore the entire signal handling
+  // configuration.
+  struct sigaction sa_;
+#else
+  Callback cb_;
+#endif
+};
+
+/// \brief Return the current handler for the given signal number.
+ARROW_EXPORT
+Result<SignalHandler> GetSignalHandler(int signum);
+
+/// \brief Set a new handler for the given signal number.
+///
+/// The old signal handler is returned.
+ARROW_EXPORT
+Result<SignalHandler> SetSignalHandler(int signum, const SignalHandler& handler);
+
+/// \brief Reinstate the signal handler
+///
+/// For use in signal handlers.  This is needed on platforms without sigaction()
+/// such as Windows, as the default signal handler is restored there as
+/// soon as a signal is raised.
+ARROW_EXPORT
+void ReinstateSignalHandler(int signum, SignalHandler::Callback handler);
+
+/// \brief Send a signal to the current process
+///
+/// The thread which will receive the signal is unspecified.
+ARROW_EXPORT
+Status SendSignal(int signum);
+
+/// \brief Send a signal to the given thread
+///
+/// This function isn't supported on Windows.
+ARROW_EXPORT
+Status SendSignalToThread(int signum, uint64_t thread_id);
+
+/// \brief Get an unpredictable random seed
+///
+/// This function may be slightly costly, so should only be used to initialize
+/// a PRNG, not to generate a large amount of random numbers.
+/// It is better to use this function rather than std::random_device, unless
+/// absolutely necessary (e.g. to generate a cryptographic secret).
+ARROW_EXPORT
+int64_t GetRandomSeed();
+
+/// \brief Get the current thread id
+///
+/// In addition to having the same properties as std::thread, the returned value
+/// is a regular integer value, which is more convenient than an opaque type.
+ARROW_EXPORT
+uint64_t GetThreadId();
+
+/// \brief Get the current memory used by the current process in bytes
+///
+/// This function supports Windows, Linux, and Mac and will return 0 otherwise
+ARROW_EXPORT
+int64_t GetCurrentRSS();
+
+/// \brief Get the total memory available to the system in bytes
+///
+/// This function supports Windows, Linux, and Mac and will return 0 otherwise
+ARROW_EXPORT
+int64_t GetTotalMemoryBytes();
+
+/// \brief Get the number of affinity core on the system.
+///
+/// This is only implemented on Linux.
+/// If a value is returned, it is guaranteed to be greater or equal to one.
+ARROW_EXPORT Result<int32_t> GetNumAffinityCores();
+
+/// \brief Load a dynamic library
+///
+/// This wraps dlopen() except on Windows, where LoadLibrary() is called.
+/// These two platforms handle absolute paths consistently; relative paths
+/// or the library's bare name may be handled but inconsistently.
+///
+/// \return An opaque handle for the dynamic library, which can be used for
+///         subsequent symbol lookup. Nullptr will never be returned; instead
+///         an error will be raised.
+ARROW_EXPORT Result<void*> LoadDynamicLibrary(const PlatformFilename& path);
+
+/// \brief Load a dynamic library
+///
+/// An overload taking null terminated string.
+ARROW_EXPORT Result<void*> LoadDynamicLibrary(const char* path);
+
+/// \brief Retrieve a symbol by name from a library handle.
+///
+/// This wraps dlsym() except on Windows, where GetProcAddress() is called.
+///
+/// \return The address associated with the named symbol. Nullptr will never be
+///         returned; instead an error will be raised.
+ARROW_EXPORT Result<void*> GetSymbol(void* handle, const char* name);
+
+template <typename T>
+Result<T*> GetSymbolAs(void* handle, const char* name) {
+  ARROW_ASSIGN_OR_RAISE(void* sym, GetSymbol(handle, name));
+  return reinterpret_cast<T*>(sym);
+}
+
+}  // namespace arrow::internal

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/iterator.h

@@ -0,0 +1,582 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cassert>
+#include <functional>
+#include <memory>
+#include <optional>
+#include <tuple>
+#include <type_traits>
+#include <utility>
+#include <vector>
+
+#include "arrow/result.h"
+#include "arrow/status.h"
+#include "arrow/util/compare.h"
+#include "arrow/util/functional.h"
+#include "arrow/util/macros.h"
+#include "arrow/util/type_fwd.h"
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+
+template <typename T>
+class Iterator;
+
+template <typename T>
+struct IterationTraits {
+  /// \brief a reserved value which indicates the end of iteration. By
+  /// default this is NULLPTR since most iterators yield pointer types.
+  /// Specialize IterationTraits if different end semantics are required.
+  ///
+  /// Note: This should not be used to determine if a given value is a
+  /// terminal value.  Use IsIterationEnd (which uses IsEnd) instead.  This
+  /// is only for returning terminal values.
+  static T End() { return T(NULLPTR); }
+
+  /// \brief Checks to see if the value is a terminal value.
+  /// A method is used here since T is not necessarily comparable in many
+  /// cases even though it has a distinct final value
+  static bool IsEnd(const T& val) { return val == End(); }
+};
+
+template <typename T>
+T IterationEnd() {
+  return IterationTraits<T>::End();
+}
+
+template <typename T>
+bool IsIterationEnd(const T& val) {
+  return IterationTraits<T>::IsEnd(val);
+}
+
+template <typename T>
+struct IterationTraits<std::optional<T>> {
+  /// \brief by default when iterating through a sequence of optional,
+  /// nullopt indicates the end of iteration.
+  /// Specialize IterationTraits if different end semantics are required.
+  static std::optional<T> End() { return std::nullopt; }
+
+  /// \brief by default when iterating through a sequence of optional,
+  /// nullopt (!has_value()) indicates the end of iteration.
+  /// Specialize IterationTraits if different end semantics are required.
+  static bool IsEnd(const std::optional<T>& val) { return !val.has_value(); }
+
+  // TODO(bkietz) The range-for loop over Iterator<optional<T>> yields
+  // Result<optional<T>> which is unnecessary (since only the unyielded end optional
+  // is nullopt. Add IterationTraits::GetRangeElement() to handle this case
+};
+
+template <typename T>
+struct IterationTraits<Enumerated<T>> {
+  static Enumerated<T> End() { return Enumerated<T>{IterationEnd<T>(), -1, false}; }
+  static bool IsEnd(const Enumerated<T>& val) { return val.index < 0; }
+};
+
+/// \brief A generic Iterator that can return errors
+template <typename T>
+class Iterator : public util::EqualityComparable<Iterator<T>> {
+ public:
+  /// \brief Iterator may be constructed from any type which has a member function
+  /// with signature Result<T> Next();
+  /// End of iterator is signalled by returning IteratorTraits<T>::End();
+  ///
+  /// The argument is moved or copied to the heap and kept in a unique_ptr<void>. Only
+  /// its destructor and its Next method (which are stored in function pointers) are
+  /// referenced after construction.
+  ///
+  /// This approach is used to dodge MSVC linkage hell (ARROW-6244, ARROW-6558) when using
+  /// an abstract template base class: instead of being inlined as usual for a template
+  /// function the base's virtual destructor will be exported, leading to multiple
+  /// definition errors when linking to any other TU where the base is instantiated.
+  template <typename Wrapped>
+  explicit Iterator(Wrapped has_next)
+      : ptr_(new Wrapped(std::move(has_next)), Delete<Wrapped>), next_(Next<Wrapped>) {}
+
+  Iterator() : ptr_(NULLPTR, [](void*) {}) {}
+
+  /// \brief Return the next element of the sequence, IterationTraits<T>::End() when the
+  /// iteration is completed.
+  Result<T> Next() {
+    if (ptr_) {
+      auto next_result = next_(ptr_.get());
+      if (next_result.ok() && IsIterationEnd(next_result.ValueUnsafe())) {
+        ptr_.reset(NULLPTR);
+      }
+      return next_result;
+    } else {
+      return IterationTraits<T>::End();
+    }
+  }
+
+  /// Pass each element of the sequence to a visitor. Will return any error status
+  /// returned by the visitor, terminating iteration.
+  template <typename Visitor>
+  Status Visit(Visitor&& visitor) {
+    for (;;) {
+      ARROW_ASSIGN_OR_RAISE(auto value, Next());
+
+      if (IsIterationEnd(value)) break;
+
+      ARROW_RETURN_NOT_OK(visitor(std::move(value)));
+    }
+
+    return Status::OK();
+  }
+
+  /// Iterators will only compare equal if they are both null.
+  /// Equality comparability is required to make an Iterator of Iterators
+  /// (to check for the end condition).
+  bool Equals(const Iterator& other) const { return ptr_ == other.ptr_; }
+
+  explicit operator bool() const { return ptr_ != NULLPTR; }
+
+  class RangeIterator {
+   public:
+    RangeIterator() : value_(IterationTraits<T>::End()) {}
+
+    explicit RangeIterator(Iterator i)
+        : value_(IterationTraits<T>::End()),
+          iterator_(std::make_shared<Iterator>(std::move(i))) {
+      Next();
+    }
+
+    bool operator!=(const RangeIterator& other) const { return value_ != other.value_; }
+
+    RangeIterator& operator++() {
+      Next();
+      return *this;
+    }
+
+    Result<T> operator*() {
+      ARROW_RETURN_NOT_OK(value_);
+
+      auto value = std::move(value_);
+      value_ = IterationTraits<T>::End();
+      return value;
+    }
+
+   private:
+    void Next() {
+      if (!value_.ok()) {
+        value_ = IterationTraits<T>::End();
+        return;
+      }
+      value_ = iterator_->Next();
+    }
+
+    Result<T> value_;
+    std::shared_ptr<Iterator> iterator_;
+  };
+
+  RangeIterator begin() { return RangeIterator(std::move(*this)); }
+
+  RangeIterator end() { return RangeIterator(); }
+
+  /// \brief Move every element of this iterator into a vector.
+  Result<std::vector<T>> ToVector() {
+    std::vector<T> out;
+    for (auto maybe_element : *this) {
+      ARROW_ASSIGN_OR_RAISE(auto element, maybe_element);
+      out.push_back(std::move(element));
+    }
+    return out;
+  }
+
+ private:
+  /// Implementation of deleter for ptr_: Casts from void* to the wrapped type and
+  /// deletes that.
+  template <typename HasNext>
+  static void Delete(void* ptr) {
+    delete static_cast<HasNext*>(ptr);
+  }
+
+  /// Implementation of Next: Casts from void* to the wrapped type and invokes that
+  /// type's Next member function.
+  template <typename HasNext>
+  static Result<T> Next(void* ptr) {
+    return static_cast<HasNext*>(ptr)->Next();
+  }
+
+  /// ptr_ is a unique_ptr to void with a custom deleter: a function pointer which first
+  /// casts from void* to a pointer to the wrapped type then deletes that.
+  std::unique_ptr<void, void (*)(void*)> ptr_;
+
+  /// next_ is a function pointer which first casts from void* to a pointer to the wrapped
+  /// type then invokes its Next member function.
+  Result<T> (*next_)(void*) = NULLPTR;
+};
+
+template <typename T>
+struct TransformFlow {
+  using YieldValueType = T;
+
+  TransformFlow(YieldValueType value, bool ready_for_next)
+      : finished_(false),
+        ready_for_next_(ready_for_next),
+        yield_value_(std::move(value)) {}
+  TransformFlow(bool finished, bool ready_for_next)
+      : finished_(finished), ready_for_next_(ready_for_next), yield_value_() {}
+
+  bool HasValue() const { return yield_value_.has_value(); }
+  bool Finished() const { return finished_; }
+  bool ReadyForNext() const { return ready_for_next_; }
+  T Value() const { return *yield_value_; }
+
+  bool finished_ = false;
+  bool ready_for_next_ = false;
+  std::optional<YieldValueType> yield_value_;
+};
+
+struct TransformFinish {
+  template <typename T>
+  operator TransformFlow<T>() && {  // NOLINT explicit
+    return TransformFlow<T>(true, true);
+  }
+};
+
+struct TransformSkip {
+  template <typename T>
+  operator TransformFlow<T>() && {  // NOLINT explicit
+    return TransformFlow<T>(false, true);
+  }
+};
+
+template <typename T>
+TransformFlow<T> TransformYield(T value = {}, bool ready_for_next = true) {
+  return TransformFlow<T>(std::move(value), ready_for_next);
+}
+
+template <typename T, typename V>
+using Transformer = std::function<Result<TransformFlow<V>>(T)>;
+
+template <typename T, typename V>
+class TransformIterator {
+ public:
+  explicit TransformIterator(Iterator<T> it, Transformer<T, V> transformer)
+      : it_(std::move(it)),
+        transformer_(std::move(transformer)),
+        last_value_(),
+        finished_() {}
+
+  Result<V> Next() {
+    while (!finished_) {
+      ARROW_ASSIGN_OR_RAISE(std::optional<V> next, Pump());
+      if (next.has_value()) {
+        return std::move(*next);
+      }
+      ARROW_ASSIGN_OR_RAISE(last_value_, it_.Next());
+    }
+    return IterationTraits<V>::End();
+  }
+
+ private:
+  // Calls the transform function on the current value.  Can return in several ways
+  // * If the next value is requested (e.g. skip) it will return an empty optional
+  // * If an invalid status is encountered that will be returned
+  // * If finished it will return IterationTraits<V>::End()
+  // * If a value is returned by the transformer that will be returned
+  Result<std::optional<V>> Pump() {
+    if (!finished_ && last_value_.has_value()) {
+      auto next_res = transformer_(*last_value_);
+      if (!next_res.ok()) {
+        finished_ = true;
+        return next_res.status();
+      }
+      auto next = std::move(*next_res);
+      if (next.ReadyForNext()) {
+        if (IsIterationEnd(*last_value_)) {
+          finished_ = true;
+        }
+        last_value_.reset();
+      }
+      if (next.Finished()) {
+        finished_ = true;
+      }
+      if (next.HasValue()) {
+        return next.Value();
+      }
+    }
+    if (finished_) {
+      return IterationTraits<V>::End();
+    }
+    return std::nullopt;
+  }
+
+  Iterator<T> it_;
+  Transformer<T, V> transformer_;
+  std::optional<T> last_value_;
+  bool finished_ = false;
+};
+
+/// \brief Transforms an iterator according to a transformer, returning a new Iterator.
+///
+/// The transformer will be called on each element of the source iterator and for each
+/// call it can yield a value, skip, or finish the iteration.  When yielding a value the
+/// transformer can choose to consume the source item (the default, ready_for_next = true)
+/// or to keep it and it will be called again on the same value.
+///
+/// This is essentially a more generic form of the map operation that can return 0, 1, or
+/// many values for each of the source items.
+///
+/// The transformer will be exposed to the end of the source sequence
+/// (IterationTraits::End) in case it needs to return some penultimate item(s).
+///
+/// Any invalid status returned by the transformer will be returned immediately.
+template <typename T, typename V>
+Iterator<V> MakeTransformedIterator(Iterator<T> it, Transformer<T, V> op) {
+  return Iterator<V>(TransformIterator<T, V>(std::move(it), std::move(op)));
+}
+
+template <typename T>
+struct IterationTraits<Iterator<T>> {
+  // The end condition for an Iterator of Iterators is a default constructed (null)
+  // Iterator.
+  static Iterator<T> End() { return Iterator<T>(); }
+  static bool IsEnd(const Iterator<T>& val) { return !val; }
+};
+
+template <typename Fn, typename T>
+class FunctionIterator {
+ public:
+  explicit FunctionIterator(Fn fn) : fn_(std::move(fn)) {}
+
+  Result<T> Next() { return fn_(); }
+
+ private:
+  Fn fn_;
+};
+
+/// \brief Construct an Iterator which invokes a callable on Next()
+template <typename Fn,
+          typename Ret = typename internal::call_traits::return_type<Fn>::ValueType>
+Iterator<Ret> MakeFunctionIterator(Fn fn) {
+  return Iterator<Ret>(FunctionIterator<Fn, Ret>(std::move(fn)));
+}
+
+template <typename T>
+Iterator<T> MakeEmptyIterator() {
+  return MakeFunctionIterator([]() -> Result<T> { return IterationTraits<T>::End(); });
+}
+
+template <typename T>
+Iterator<T> MakeErrorIterator(Status s) {
+  return MakeFunctionIterator([s]() -> Result<T> {
+    ARROW_RETURN_NOT_OK(s);
+    return IterationTraits<T>::End();
+  });
+}
+
+/// \brief Simple iterator which yields the elements of a std::vector
+template <typename T>
+class VectorIterator {
+ public:
+  explicit VectorIterator(std::vector<T> v) : elements_(std::move(v)) {}
+
+  Result<T> Next() {
+    if (i_ == elements_.size()) {
+      return IterationTraits<T>::End();
+    }
+    return std::move(elements_[i_++]);
+  }
+
+ private:
+  std::vector<T> elements_;
+  size_t i_ = 0;
+};
+
+template <typename T>
+Iterator<T> MakeVectorIterator(std::vector<T> v) {
+  return Iterator<T>(VectorIterator<T>(std::move(v)));
+}
+
+/// \brief Simple iterator which yields *pointers* to the elements of a std::vector<T>.
+/// This is provided to support T where IterationTraits<T>::End is not specialized
+template <typename T>
+class VectorPointingIterator {
+ public:
+  explicit VectorPointingIterator(std::vector<T> v) : elements_(std::move(v)) {}
+
+  Result<T*> Next() {
+    if (i_ == elements_.size()) {
+      return NULLPTR;
+    }
+    return &elements_[i_++];
+  }
+
+ private:
+  std::vector<T> elements_;
+  size_t i_ = 0;
+};
+
+template <typename T>
+Iterator<T*> MakeVectorPointingIterator(std::vector<T> v) {
+  return Iterator<T*>(VectorPointingIterator<T>(std::move(v)));
+}
+
+/// \brief MapIterator takes ownership of an iterator and a function to apply
+/// on every element. The mapped function is not allowed to fail.
+template <typename Fn, typename I, typename O>
+class MapIterator {
+ public:
+  explicit MapIterator(Fn map, Iterator<I> it)
+      : map_(std::move(map)), it_(std::move(it)) {}
+
+  Result<O> Next() {
+    ARROW_ASSIGN_OR_RAISE(I i, it_.Next());
+
+    if (IsIterationEnd(i)) {
+      return IterationTraits<O>::End();
+    }
+
+    return map_(std::move(i));
+  }
+
+ private:
+  Fn map_;
+  Iterator<I> it_;
+};
+
+/// \brief MapIterator takes ownership of an iterator and a function to apply
+/// on every element. The mapped function is not allowed to fail.
+template <typename Fn, typename From = internal::call_traits::argument_type<0, Fn>,
+          typename To = internal::call_traits::return_type<Fn>>
+Iterator<To> MakeMapIterator(Fn map, Iterator<From> it) {
+  return Iterator<To>(MapIterator<Fn, From, To>(std::move(map), std::move(it)));
+}
+
+/// \brief Like MapIterator, but where the function can fail.
+template <typename Fn, typename From = internal::call_traits::argument_type<0, Fn>,
+          typename To = typename internal::call_traits::return_type<Fn>::ValueType>
+Iterator<To> MakeMaybeMapIterator(Fn map, Iterator<From> it) {
+  return Iterator<To>(MapIterator<Fn, From, To>(std::move(map), std::move(it)));
+}
+
+struct FilterIterator {
+  enum Action { ACCEPT, REJECT };
+
+  template <typename To>
+  static Result<std::pair<To, Action>> Reject() {
+    return std::make_pair(IterationTraits<To>::End(), REJECT);
+  }
+
+  template <typename To>
+  static Result<std::pair<To, Action>> Accept(To out) {
+    return std::make_pair(std::move(out), ACCEPT);
+  }
+
+  template <typename To>
+  static Result<std::pair<To, Action>> MaybeAccept(Result<To> maybe_out) {
+    return std::move(maybe_out).Map(Accept<To>);
+  }
+
+  template <typename To>
+  static Result<std::pair<To, Action>> Error(Status s) {
+    return s;
+  }
+
+  template <typename Fn, typename From, typename To>
+  class Impl {
+   public:
+    explicit Impl(Fn filter, Iterator<From> it) : filter_(filter), it_(std::move(it)) {}
+
+    Result<To> Next() {
+      To out = IterationTraits<To>::End();
+      Action action;
+
+      for (;;) {
+        ARROW_ASSIGN_OR_RAISE(From i, it_.Next());
+
+        if (IsIterationEnd(i)) {
+          return IterationTraits<To>::End();
+        }
+
+        ARROW_ASSIGN_OR_RAISE(std::tie(out, action), filter_(std::move(i)));
+
+        if (action == ACCEPT) return out;
+      }
+    }
+
+   private:
+    Fn filter_;
+    Iterator<From> it_;
+  };
+};
+
+/// \brief Like MapIterator, but where the function can fail or reject elements.
+template <
+    typename Fn, typename From = typename internal::call_traits::argument_type<0, Fn>,
+    typename Ret = typename internal::call_traits::return_type<Fn>::ValueType,
+    typename To = typename std::tuple_element<0, Ret>::type,
+    typename Enable = typename std::enable_if<std::is_same<
+        typename std::tuple_element<1, Ret>::type, FilterIterator::Action>::value>::type>
+Iterator<To> MakeFilterIterator(Fn filter, Iterator<From> it) {
+  return Iterator<To>(
+      FilterIterator::Impl<Fn, From, To>(std::move(filter), std::move(it)));
+}
+
+/// \brief FlattenIterator takes an iterator generating iterators and yields a
+/// unified iterator that flattens/concatenates in a single stream.
+template <typename T>
+class FlattenIterator {
+ public:
+  explicit FlattenIterator(Iterator<Iterator<T>> it) : parent_(std::move(it)) {}
+
+  Result<T> Next() {
+    if (IsIterationEnd(child_)) {
+      // Pop from parent's iterator.
+      ARROW_ASSIGN_OR_RAISE(child_, parent_.Next());
+
+      // Check if final iteration reached.
+      if (IsIterationEnd(child_)) {
+        return IterationTraits<T>::End();
+      }
+
+      return Next();
+    }
+
+    // Pop from child_ and check for depletion.
+    ARROW_ASSIGN_OR_RAISE(T out, child_.Next());
+    if (IsIterationEnd(out)) {
+      // Reset state such that we pop from parent on the recursive call
+      child_ = IterationTraits<Iterator<T>>::End();
+
+      return Next();
+    }
+
+    return out;
+  }
+
+ private:
+  Iterator<Iterator<T>> parent_;
+  Iterator<T> child_ = IterationTraits<Iterator<T>>::End();
+};
+
+template <typename T>
+Iterator<T> MakeFlattenIterator(Iterator<Iterator<T>> it) {
+  return Iterator<T>(FlattenIterator<T>(std::move(it)));
+}
+
+template <typename Reader>
+Iterator<typename Reader::ValueType> MakeIteratorFromReader(
+    const std::shared_ptr<Reader>& reader) {
+  return MakeFunctionIterator([reader] { return reader->Next(); });
+}
+
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/key_value_metadata.h

@@ -0,0 +1,99 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cstdint>
+#include <memory>
+#include <string>
+#include <string_view>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#include "arrow/result.h"
+#include "arrow/status.h"
+#include "arrow/util/macros.h"
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+
+/// \brief A container for key-value pair type metadata. Not thread-safe
+class ARROW_EXPORT KeyValueMetadata {
+ public:
+  KeyValueMetadata();
+  KeyValueMetadata(std::vector<std::string> keys, std::vector<std::string> values);
+  explicit KeyValueMetadata(const std::unordered_map<std::string, std::string>& map);
+
+  static std::shared_ptr<KeyValueMetadata> Make(std::vector<std::string> keys,
+                                                std::vector<std::string> values);
+
+  void ToUnorderedMap(std::unordered_map<std::string, std::string>* out) const;
+  void Append(std::string key, std::string value);
+
+  Result<std::string> Get(std::string_view key) const;
+  bool Contains(std::string_view key) const;
+  // Note that deleting may invalidate known indices
+  Status Delete(std::string_view key);
+  Status Delete(int64_t index);
+  Status DeleteMany(std::vector<int64_t> indices);
+  Status Set(std::string key, std::string value);
+
+  void reserve(int64_t n);
+
+  int64_t size() const;
+  const std::string& key(int64_t i) const;
+  const std::string& value(int64_t i) const;
+  const std::vector<std::string>& keys() const { return keys_; }
+  const std::vector<std::string>& values() const { return values_; }
+
+  std::vector<std::pair<std::string, std::string>> sorted_pairs() const;
+
+  /// \brief Perform linear search for key, returning -1 if not found
+  int FindKey(std::string_view key) const;
+
+  std::shared_ptr<KeyValueMetadata> Copy() const;
+
+  /// \brief Return a new KeyValueMetadata by combining the passed metadata
+  /// with this KeyValueMetadata. Colliding keys will be overridden by the
+  /// passed metadata. Assumes keys in both containers are unique
+  std::shared_ptr<KeyValueMetadata> Merge(const KeyValueMetadata& other) const;
+
+  bool Equals(const KeyValueMetadata& other) const;
+  std::string ToString() const;
+
+ private:
+  std::vector<std::string> keys_;
+  std::vector<std::string> values_;
+
+  ARROW_DISALLOW_COPY_AND_ASSIGN(KeyValueMetadata);
+};
+
+/// \brief Create a KeyValueMetadata instance
+///
+/// \param pairs key-value mapping
+ARROW_EXPORT std::shared_ptr<KeyValueMetadata> key_value_metadata(
+    const std::unordered_map<std::string, std::string>& pairs);
+
+/// \brief Create a KeyValueMetadata instance
+///
+/// \param keys sequence of metadata keys
+/// \param values sequence of corresponding metadata values
+ARROW_EXPORT std::shared_ptr<KeyValueMetadata> key_value_metadata(
+    std::vector<std::string> keys, std::vector<std::string> values);
+
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/launder.h

@@ -0,0 +1,35 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <new>
+
+namespace arrow {
+namespace internal {
+
+#if __cpp_lib_launder
+using std::launder;
+#else
+template <class T>
+constexpr T* launder(T* p) noexcept {
+  return p;
+}
+#endif
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/list_util.h

@@ -0,0 +1,55 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cstdint>
+#include <utility>
+
+#include "arrow/array/data.h"
+#include "arrow/result.h"
+
+namespace arrow {
+namespace list_util {
+namespace internal {
+
+/// \brief Calculate the smallest continuous range of values used by the
+/// var-length list-like input (list, map and list-view types).
+///
+/// \param input The input array such that is_var_length_list_like(input.type)
+/// is true
+/// \return A pair of (offset, length) describing the range
+ARROW_EXPORT Result<std::pair<int64_t, int64_t>> RangeOfValuesUsed(
+    const ArraySpan& input);
+
+/// \brief Calculate the sum of the sizes of all valid lists or list-views
+///
+/// This is usually the same as the length of the RangeOfValuesUsed() range, but
+/// it can be:
+/// - Smaller: when the child array contains many values that are not
+/// referenced by the lists or list-views in the parent array
+/// - Greater: when the list-views share child array ranges
+///
+/// \param input The input array such that is_var_length_list_like(input.type)
+/// is true
+/// \return The sum of all list or list-view sizes
+ARROW_EXPORT Result<int64_t> SumOfLogicalListSizes(const ArraySpan& input);
+
+}  // namespace internal
+
+}  // namespace list_util
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/logger.h

@@ -0,0 +1,186 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <chrono>
+#include <iosfwd>
+#include <memory>
+#include <string_view>
+
+#include "arrow/result.h"
+#include "arrow/status.h"
+#include "arrow/util/logging.h"
+#include "arrow/util/macros.h"
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+namespace util {
+
+struct SourceLocation {
+  const char* file = "";
+  int line = 0;
+};
+
+struct LogDetails {
+  ArrowLogLevel severity = ArrowLogLevel::ARROW_INFO;
+  std::chrono::system_clock::time_point timestamp = std::chrono::system_clock::now();
+  SourceLocation source_location{};
+  std::string_view message = "";
+};
+
+/// \brief A base interface for custom loggers.
+///
+/// Loggers can be added to the LoggerRegistry for global access or directly provided to
+/// certain logging utilities.
+class Logger {
+ public:
+  virtual ~Logger() = default;
+
+  virtual void Log(const LogDetails& details) = 0;
+
+  virtual bool Flush(std::chrono::microseconds timeout) { return true; }
+  bool Flush() { return this->Flush(std::chrono::microseconds::max()); }
+
+  virtual bool is_enabled() const { return true; }
+
+  virtual ArrowLogLevel severity_threshold() const { return ArrowLogLevel::ARROW_TRACE; }
+};
+
+/// \brief Creates a simple logger that redirects output to std::cerr
+ARROW_EXPORT std::shared_ptr<Logger> MakeOStreamLogger(ArrowLogLevel severity_threshold);
+/// \brief Creates a simple logger that redirects output to the provided ostream
+ARROW_EXPORT std::shared_ptr<Logger> MakeOStreamLogger(ArrowLogLevel severity_threshold,
+                                                       std::ostream& sink);
+
+class ARROW_EXPORT LoggerRegistry {
+ public:
+  /// \brief Add a logger to the registry with the associated name
+  ///
+  /// Returns Invalid if a logger with the provided name already exists. Users should call
+  /// `UnregisterLogger` first if they wish to overwrite it.
+  static Status RegisterLogger(std::string_view name, std::shared_ptr<Logger> logger);
+
+  /// \brief Remove a logger from the registry
+  static void UnregisterLogger(std::string_view name);
+
+  /// \brief Return the logger associated with the provided name
+  ///
+  /// If `name` is empty, the default logger is returned. If `name` doesn't match any of
+  /// the registered loggers then a non-null noop logger is returned
+  static std::shared_ptr<Logger> GetLogger(std::string_view name = "");
+
+  /// \brief Return the default logger
+  static std::shared_ptr<Logger> GetDefaultLogger();
+  /// \brief Set the default logger
+  static void SetDefaultLogger(std::shared_ptr<Logger> logger);
+};
+
+/// \brief Represents a single log record to be emitted by an underlying logger
+class ARROW_EXPORT LogMessage {
+ public:
+  /// \brief Construct a LogMessage with the provided underlying logger
+  LogMessage(ArrowLogLevel severity, std::shared_ptr<Logger> logger,
+             SourceLocation source_location = {});
+  /// \brief Construct a LogMessage with the provided logger name, which will be used to
+  /// find an underlying logger in the registry
+  LogMessage(ArrowLogLevel severity, std::string_view logger_name,
+             SourceLocation source_location = {});
+
+  std::ostream& Stream();
+
+  // Convenience method - mainly for use in ARROW_LOG_* macros. This prevents unnecessary
+  // argument evaluation when log statements are stripped in certain builds
+  template <typename... Args>
+  LogMessage& Append(Args&&... args) {
+    if constexpr (sizeof...(Args) > 0) {
+      if (CheckIsEnabled()) {
+        (Stream() << ... << args);
+      }
+    }
+    return *this;
+  }
+
+ private:
+  bool CheckIsEnabled();
+
+  class Impl;
+  std::shared_ptr<Impl> impl_;
+};
+
+}  // namespace util
+}  // namespace arrow
+
+// For the following macros, log statements with a lower severity than
+// `ARROW_MINIMUM_LOG_LEVEL` will be stripped from the build
+#ifndef ARROW_MINIMUM_LOG_LEVEL
+#  define ARROW_MINIMUM_LOG_LEVEL -1000
+#endif
+
+#define ARROW_LOGGER_INTERNAL(LOGGER, LEVEL)                                      \
+  (::arrow::util::LogMessage(::arrow::util::ArrowLogLevel::ARROW_##LEVEL, LOGGER, \
+                             ::arrow::util::SourceLocation{__FILE__, __LINE__}))
+
+static_assert(static_cast<int>(::arrow::util::ArrowLogLevel::ARROW_TRACE) == -2);
+#if ARROW_MINIMUM_LOG_LEVEL <= -2
+#  define ARROW_LOGGER_TRACE(LOGGER, ...) \
+    (ARROW_LOGGER_INTERNAL(LOGGER, TRACE).Append(__VA_ARGS__))
+#else
+#  define ARROW_LOGGER_TRACE(...) ARROW_UNUSED(0)
+#endif
+
+static_assert(static_cast<int>(::arrow::util::ArrowLogLevel::ARROW_DEBUG) == -1);
+#if ARROW_MINIMUM_LOG_LEVEL <= -1
+#  define ARROW_LOGGER_DEBUG(LOGGER, ...) \
+    (ARROW_LOGGER_INTERNAL(LOGGER, DEBUG).Append(__VA_ARGS__))
+#else
+#  define ARROW_LOGGER_DEBUG(...) ARROW_UNUSED(0)
+#endif
+
+static_assert(static_cast<int>(::arrow::util::ArrowLogLevel::ARROW_INFO) == 0);
+#if ARROW_MINIMUM_LOG_LEVEL <= 0
+#  define ARROW_LOGGER_INFO(LOGGER, ...) \
+    (ARROW_LOGGER_INTERNAL(LOGGER, INFO).Append(__VA_ARGS__))
+#else
+#  define ARROW_LOGGER_INFO(...) ARROW_UNUSED(0)
+#endif
+
+static_assert(static_cast<int>(::arrow::util::ArrowLogLevel::ARROW_WARNING) == 1);
+#if ARROW_MINIMUM_LOG_LEVEL <= 1
+#  define ARROW_LOGGER_WARNING(LOGGER, ...) \
+    (ARROW_LOGGER_INTERNAL(LOGGER, WARNING).Append(__VA_ARGS__))
+#else
+#  define ARROW_LOGGER_WARNING(...) ARROW_UNUSED(0)
+#endif
+
+static_assert(static_cast<int>(::arrow::util::ArrowLogLevel::ARROW_ERROR) == 2);
+#if ARROW_MINIMUM_LOG_LEVEL <= 2
+#  define ARROW_LOGGER_ERROR(LOGGER, ...) \
+    (ARROW_LOGGER_INTERNAL(LOGGER, ERROR).Append(__VA_ARGS__))
+#else
+#  define ARROW_LOGGER_ERROR(...) ARROW_UNUSED(0)
+#endif
+
+static_assert(static_cast<int>(::arrow::util::ArrowLogLevel::ARROW_FATAL) == 3);
+#if ARROW_MINIMUM_LOG_LEVEL <= 3
+#  define ARROW_LOGGER_FATAL(LOGGER, ...) \
+    (ARROW_LOGGER_INTERNAL(LOGGER, FATAL).Append(__VA_ARGS__))
+#else
+#  define ARROW_LOGGER_FATAL(...) ARROW_UNUSED(0)
+#endif
+
+#define ARROW_LOGGER_CALL(LOGGER, LEVEL, ...) ARROW_LOGGER_##LEVEL(LOGGER, __VA_ARGS__)

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/logging.h

@@ -0,0 +1,251 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#ifdef GANDIVA_IR
+
+// The LLVM IR code doesn't have an NDEBUG mode. And, it shouldn't include references to
+// streams or stdc++. So, making the DCHECK calls void in that case.
+
+#  define ARROW_IGNORE_EXPR(expr) ((void)(expr))
+
+#  define ARROW_DCHECK(condition) ARROW_IGNORE_EXPR(condition)
+#  define ARROW_DCHECK_OK(status) ARROW_IGNORE_EXPR(status)
+#  define ARROW_DCHECK_EQ(val1, val2) ARROW_IGNORE_EXPR(val1)
+#  define ARROW_DCHECK_NE(val1, val2) ARROW_IGNORE_EXPR(val1)
+#  define ARROW_DCHECK_LE(val1, val2) ARROW_IGNORE_EXPR(val1)
+#  define ARROW_DCHECK_LT(val1, val2) ARROW_IGNORE_EXPR(val1)
+#  define ARROW_DCHECK_GE(val1, val2) ARROW_IGNORE_EXPR(val1)
+#  define ARROW_DCHECK_GT(val1, val2) ARROW_IGNORE_EXPR(val1)
+
+#else  // !GANDIVA_IR
+
+#  include <memory>
+#  include <ostream>
+#  include <string>
+
+#  include "arrow/util/macros.h"
+#  include "arrow/util/visibility.h"
+
+namespace arrow {
+namespace util {
+
+enum class ArrowLogLevel : int {
+  ARROW_TRACE = -2,
+  ARROW_DEBUG = -1,
+  ARROW_INFO = 0,
+  ARROW_WARNING = 1,
+  ARROW_ERROR = 2,
+  ARROW_FATAL = 3
+};
+
+#  define ARROW_LOG_INTERNAL(level) ::arrow::util::ArrowLog(__FILE__, __LINE__, level)
+#  define ARROW_LOG(level) ARROW_LOG_INTERNAL(::arrow::util::ArrowLogLevel::ARROW_##level)
+
+#  define ARROW_IGNORE_EXPR(expr) ((void)(expr))
+
+#  define ARROW_CHECK_OR_LOG(condition, level) \
+    ARROW_PREDICT_TRUE(condition)              \
+    ? ARROW_IGNORE_EXPR(0)                     \
+    : ::arrow::util::Voidify() & ARROW_LOG(level) << " Check failed: " #condition " "
+
+#  define ARROW_CHECK(condition) ARROW_CHECK_OR_LOG(condition, FATAL)
+
+// If 'to_call' returns a bad status, CHECK immediately with a logged message
+// of 'msg' followed by the status.
+#  define ARROW_CHECK_OK_PREPEND(to_call, msg, level)                 \
+    do {                                                              \
+      ::arrow::Status _s = ::arrow::ToStatus(to_call);                \
+      ARROW_CHECK_OR_LOG(_s.ok(), level)                              \
+          << "Operation failed: " << ARROW_STRINGIFY(to_call) << "\n" \
+          << (msg) << ": " << _s.ToString();                          \
+    } while (false)
+
+// If the status is bad, CHECK immediately, appending the status to the
+// logged message.
+#  define ARROW_CHECK_OK(s) ARROW_CHECK_OK_PREPEND(s, "Bad status", FATAL)
+
+#  define ARROW_CHECK_EQ(val1, val2) ARROW_CHECK((val1) == (val2))
+#  define ARROW_CHECK_NE(val1, val2) ARROW_CHECK((val1) != (val2))
+#  define ARROW_CHECK_LE(val1, val2) ARROW_CHECK((val1) <= (val2))
+#  define ARROW_CHECK_LT(val1, val2) ARROW_CHECK((val1) < (val2))
+#  define ARROW_CHECK_GE(val1, val2) ARROW_CHECK((val1) >= (val2))
+#  define ARROW_CHECK_GT(val1, val2) ARROW_CHECK((val1) > (val2))
+
+#  ifdef NDEBUG
+#    define ARROW_DFATAL ::arrow::util::ArrowLogLevel::ARROW_WARNING
+
+// CAUTION: DCHECK_OK() always evaluates its argument, but other DCHECK*() macros
+// only do so in debug mode.
+
+#    define ARROW_DCHECK(condition)               \
+      while (false) ARROW_IGNORE_EXPR(condition); \
+      while (false) ::arrow::util::detail::NullLog()
+#    define ARROW_DCHECK_OK(s) \
+      ARROW_IGNORE_EXPR(s);    \
+      while (false) ::arrow::util::detail::NullLog()
+#    define ARROW_DCHECK_EQ(val1, val2)      \
+      while (false) ARROW_IGNORE_EXPR(val1); \
+      while (false) ARROW_IGNORE_EXPR(val2); \
+      while (false) ::arrow::util::detail::NullLog()
+#    define ARROW_DCHECK_NE(val1, val2)      \
+      while (false) ARROW_IGNORE_EXPR(val1); \
+      while (false) ARROW_IGNORE_EXPR(val2); \
+      while (false) ::arrow::util::detail::NullLog()
+#    define ARROW_DCHECK_LE(val1, val2)      \
+      while (false) ARROW_IGNORE_EXPR(val1); \
+      while (false) ARROW_IGNORE_EXPR(val2); \
+      while (false) ::arrow::util::detail::NullLog()
+#    define ARROW_DCHECK_LT(val1, val2)      \
+      while (false) ARROW_IGNORE_EXPR(val1); \
+      while (false) ARROW_IGNORE_EXPR(val2); \
+      while (false) ::arrow::util::detail::NullLog()
+#    define ARROW_DCHECK_GE(val1, val2)      \
+      while (false) ARROW_IGNORE_EXPR(val1); \
+      while (false) ARROW_IGNORE_EXPR(val2); \
+      while (false) ::arrow::util::detail::NullLog()
+#    define ARROW_DCHECK_GT(val1, val2)      \
+      while (false) ARROW_IGNORE_EXPR(val1); \
+      while (false) ARROW_IGNORE_EXPR(val2); \
+      while (false) ::arrow::util::detail::NullLog()
+
+#  else
+#    define ARROW_DFATAL ::arrow::util::ArrowLogLevel::ARROW_FATAL
+
+#    define ARROW_DCHECK ARROW_CHECK
+#    define ARROW_DCHECK_OK ARROW_CHECK_OK
+#    define ARROW_DCHECK_EQ ARROW_CHECK_EQ
+#    define ARROW_DCHECK_NE ARROW_CHECK_NE
+#    define ARROW_DCHECK_LE ARROW_CHECK_LE
+#    define ARROW_DCHECK_LT ARROW_CHECK_LT
+#    define ARROW_DCHECK_GE ARROW_CHECK_GE
+#    define ARROW_DCHECK_GT ARROW_CHECK_GT
+
+#  endif  // NDEBUG
+
+// This code is adapted from
+// https://github.com/ray-project/ray/blob/master/src/ray/util/logging.h.
+
+// To make the logging lib pluggable with other logging libs and make
+// the implementation unawared by the user, ArrowLog is only a declaration
+// which hide the implementation into logging.cc file.
+// In logging.cc, we can choose different log libs using different macros.
+
+// This is also a null log which does not output anything.
+class ARROW_EXPORT ArrowLogBase {
+ public:
+  virtual ~ArrowLogBase() {}
+
+  virtual bool IsEnabled() const { return false; }
+
+  template <typename T>
+  ArrowLogBase& operator<<(const T& t) {
+    if (IsEnabled()) {
+      Stream() << t;
+    }
+    return *this;
+  }
+
+ protected:
+  virtual std::ostream& Stream() = 0;
+};
+
+class ARROW_EXPORT ArrowLog : public ArrowLogBase {
+ public:
+  ArrowLog(const char* file_name, int line_number, ArrowLogLevel severity);
+  ~ArrowLog() override;
+
+  /// Return whether or not current logging instance is enabled.
+  ///
+  /// \return True if logging is enabled and false otherwise.
+  bool IsEnabled() const override;
+
+  /// The init function of arrow log for a program which should be called only once.
+  ///
+  /// \param appName The app name which starts the log.
+  /// \param severity_threshold Logging threshold for the program.
+  /// \param logDir Logging output file name. If empty, the log won't output to file.
+  static void StartArrowLog(const std::string& appName,
+                            ArrowLogLevel severity_threshold = ArrowLogLevel::ARROW_INFO,
+                            const std::string& logDir = "");
+
+  /// The shutdown function of arrow log, it should be used with StartArrowLog as a pair.
+  static void ShutDownArrowLog();
+
+  /// Install the failure signal handler to output call stack when crash.
+  /// If glog is not installed, this function won't do anything.
+  static void InstallFailureSignalHandler();
+
+  /// Uninstall the signal actions installed by InstallFailureSignalHandler.
+  static void UninstallSignalAction();
+
+  /// Return whether or not the log level is enabled in current setting.
+  ///
+  /// \param log_level The input log level to test.
+  /// \return True if input log level is not lower than the threshold.
+  static bool IsLevelEnabled(ArrowLogLevel log_level);
+
+ private:
+  ARROW_DISALLOW_COPY_AND_ASSIGN(ArrowLog);
+
+  // Hide the implementation of log provider by void *.
+  // Otherwise, lib user may define the same macro to use the correct header file.
+  void* logging_provider_;
+  /// True if log messages should be logged and false if they should be ignored.
+  bool is_enabled_;
+
+  static ArrowLogLevel severity_threshold_;
+
+ protected:
+  std::ostream& Stream() override;
+};
+
+// This class make ARROW_CHECK compilation pass to change the << operator to void.
+// This class is copied from glog.
+class ARROW_EXPORT Voidify {
+ public:
+  Voidify() {}
+  // This has to be an operator with a precedence lower than << but
+  // higher than ?:
+  void operator&(ArrowLogBase&) {}
+};
+
+namespace detail {
+
+/// @brief A helper for the nil log sink.
+///
+/// Using this helper is analogous to sending log messages to /dev/null:
+/// nothing gets logged.
+class NullLog {
+ public:
+  /// The no-op output operator.
+  ///
+  /// @param [in] t
+  ///   The object to send into the nil sink.
+  /// @return Reference to the updated object.
+  template <class T>
+  NullLog& operator<<(const T& t) {
+    return *this;
+  }
+};
+
+}  // namespace detail
+}  // namespace util
+}  // namespace arrow
+
+#endif  // GANDIVA_IR

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/macros.h

@@ -0,0 +1,252 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cstdint>
+
+#define ARROW_EXPAND(x) x
+#define ARROW_STRINGIFY(x) #x
+#define ARROW_CONCAT(x, y) x##y
+
+// From Google gutil
+#ifndef ARROW_DISALLOW_COPY_AND_ASSIGN
+#  define ARROW_DISALLOW_COPY_AND_ASSIGN(TypeName) \
+    TypeName(const TypeName&) = delete;            \
+    void operator=(const TypeName&) = delete
+#endif
+
+#ifndef ARROW_DEFAULT_MOVE_AND_ASSIGN
+#  define ARROW_DEFAULT_MOVE_AND_ASSIGN(TypeName) \
+    TypeName(TypeName&&) = default;               \
+    TypeName& operator=(TypeName&&) = default
+#endif
+
+// With ARROW_PREDICT_FALSE, GCC and clang can be told that a certain branch is
+// not likely to be taken (for instance, a CHECK failure), and use that information in
+// static analysis. Giving the compiler this information can affect the generated code
+// layout in the absence of better information (i.e. -fprofile-arcs). [1] explains how
+// this feature can be used to improve code generation. It was written as a positive
+// comment to a negative article about the use of these annotations.
+//
+// ARROW_COMPILER_ASSUME allows the compiler to assume that a given expression is
+// true, without evaluating it, and to optimise based on this assumption [2]. If this
+// condition is violated at runtime, the behavior is undefined. This can be useful to
+// generate both faster and smaller code in compute kernels.
+//
+// IMPORTANT: Different optimisers are likely to react differently to this annotation!
+// It should be used with care when we can prove by some means that the assumption
+// is (1) guaranteed to always hold and (2) is useful for optimization [3]. If the
+// assumption is pessimistic, it might even block the compiler from decisions that
+// could lead to better code [4]. If you have a good intuition for what the compiler
+// can do with assumptions [5], you can use this macro to guide it and end up with
+// results you would only get with more complex code transformations.
+// `clang -S -emit-llvm` can be used to check how the generated code changes with
+// your specific use of this macro.
+//
+// [1] https://lobste.rs/s/uwgtkt/don_t_use_likely_unlikely_attributes#c_xi3wmc
+// [2] "Portable assumptions"
+//     https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2021/p1774r4.pdf
+// [3] "Assertions Are Pessimistic, Assumptions Are Optimistic"
+//     https://blog.regehr.org/archives/1096
+// [4] https://discourse.llvm.org/t/llvm-assume-blocks-optimization/71609
+// [5] J. Doerfert et al. 2019. "Performance Exploration Through Optimistic Static
+//     Program Annotations". https://github.com/jdoerfert/PETOSPA/blob/master/ISC19.pdf
+#define ARROW_UNUSED(x) (void)(x)
+#ifdef ARROW_WARN_DOCUMENTATION
+#  define ARROW_ARG_UNUSED(x) x
+#else
+#  define ARROW_ARG_UNUSED(x)
+#endif
+#if defined(__GNUC__)  // GCC and compatible compilers (clang, Intel ICC)
+#  define ARROW_NORETURN __attribute__((noreturn))
+#  define ARROW_NOINLINE __attribute__((noinline))
+#  define ARROW_FORCE_INLINE __attribute__((always_inline))
+#  define ARROW_PREDICT_FALSE(x) (__builtin_expect(!!(x), 0))
+#  define ARROW_PREDICT_TRUE(x) (__builtin_expect(!!(x), 1))
+#  define ARROW_RESTRICT __restrict
+#  if defined(__clang__)  // clang-specific
+#    define ARROW_COMPILER_ASSUME(expr) __builtin_assume(expr)
+#  else  // GCC-specific
+#    if __GNUC__ >= 13
+#      define ARROW_COMPILER_ASSUME(expr) __attribute__((assume(expr)))
+#    else
+// GCC does not have a built-in assume intrinsic before GCC 13, so we use an
+// if statement and __builtin_unreachable() to achieve the same effect [2].
+// Unlike clang's __builtin_assume and C++23's [[assume(expr)]], using this
+// on GCC won't warn about side-effects in the expression, so make sure expr
+// is side-effect free when working with GCC versions before 13 (Jan-2024),
+// otherwise clang/MSVC builds will fail in CI.
+#      define ARROW_COMPILER_ASSUME(expr) \
+        if (expr) {                       \
+        } else {                          \
+          __builtin_unreachable();        \
+        }
+#    endif  // __GNUC__ >= 13
+#  endif
+#elif defined(_MSC_VER)  // MSVC
+#  define ARROW_NORETURN __declspec(noreturn)
+#  define ARROW_NOINLINE __declspec(noinline)
+#  define ARROW_FORCE_INLINE __forceinline
+#  define ARROW_PREDICT_FALSE(x) (x)
+#  define ARROW_PREDICT_TRUE(x) (x)
+#  define ARROW_RESTRICT __restrict
+#  define ARROW_COMPILER_ASSUME(expr) __assume(expr)
+#else
+#  define ARROW_NORETURN
+#  define ARROW_NOINLINE
+#  define ARROW_FORCE_INLINE
+#  define ARROW_PREDICT_FALSE(x) (x)
+#  define ARROW_PREDICT_TRUE(x) (x)
+#  define ARROW_RESTRICT
+#  define ARROW_COMPILER_ASSUME(expr)
+#endif
+
+// ----------------------------------------------------------------------
+// C++/CLI support macros (see ARROW-1134)
+
+#ifndef NULLPTR
+
+#  ifdef __cplusplus_cli
+#    define NULLPTR __nullptr
+#  else
+#    define NULLPTR nullptr
+#  endif
+
+#endif  // ifndef NULLPTR
+
+// ----------------------------------------------------------------------
+
+// clang-format off
+// [[deprecated]] is only available in C++14, use this for the time being
+// This macro takes an optional deprecation message
+#ifdef __COVERITY__
+#  define ARROW_DEPRECATED(...)
+#else
+#  define ARROW_DEPRECATED(...) [[deprecated(__VA_ARGS__)]]
+#endif
+
+#ifdef __COVERITY__
+#  define ARROW_DEPRECATED_ENUM_VALUE(...)
+#else
+#  define ARROW_DEPRECATED_ENUM_VALUE(...) [[deprecated(__VA_ARGS__)]]
+#endif
+
+// clang-format on
+
+// Macros to disable deprecation warnings
+
+#ifdef __clang__
+#  define ARROW_SUPPRESS_DEPRECATION_WARNING \
+    _Pragma("clang diagnostic push");        \
+    _Pragma("clang diagnostic ignored \"-Wdeprecated-declarations\"")
+#  define ARROW_UNSUPPRESS_DEPRECATION_WARNING _Pragma("clang diagnostic pop")
+#elif defined(__GNUC__)
+#  define ARROW_SUPPRESS_DEPRECATION_WARNING \
+    _Pragma("GCC diagnostic push");          \
+    _Pragma("GCC diagnostic ignored \"-Wdeprecated-declarations\"")
+#  define ARROW_UNSUPPRESS_DEPRECATION_WARNING _Pragma("GCC diagnostic pop")
+#elif defined(_MSC_VER)
+#  define ARROW_SUPPRESS_DEPRECATION_WARNING \
+    __pragma(warning(push)) __pragma(warning(disable : 4996))
+#  define ARROW_UNSUPPRESS_DEPRECATION_WARNING __pragma(warning(pop))
+#else
+#  define ARROW_SUPPRESS_DEPRECATION_WARNING
+#  define ARROW_UNSUPPRESS_DEPRECATION_WARNING
+#endif
+
+// ----------------------------------------------------------------------
+
+// Macros to disable warnings about undeclared global functions
+#if defined(__GNUC__)
+#  define ARROW_SUPPRESS_MISSING_DECLARATIONS_WARNING \
+    _Pragma("GCC diagnostic push");                   \
+    _Pragma("GCC diagnostic ignored \"-Wmissing-declarations\"")
+#  define ARROW_UNSUPPRESS_MISSING_DECLARATIONS_WARNING _Pragma("GCC diagnostic pop")
+#else
+#  define ARROW_SUPPRESS_MISSING_DECLARATIONS_WARNING
+#  define ARROW_UNSUPPRESS_MISSING_DECLARATIONS_WARNING
+#endif
+
+// ----------------------------------------------------------------------
+
+// macros to disable padding
+// these macros are portable across different compilers and platforms
+//[https://github.com/google/flatbuffers/blob/master/include/flatbuffers/flatbuffers.h#L1355]
+#if !defined(MANUALLY_ALIGNED_STRUCT)
+#  if defined(_MSC_VER)
+#    define MANUALLY_ALIGNED_STRUCT(alignment) \
+      __pragma(pack(1));                       \
+      struct __declspec(align(alignment))
+#    define STRUCT_END(name, size) \
+      __pragma(pack());            \
+      static_assert(sizeof(name) == size, "compiler breaks packing rules")
+#  elif defined(__GNUC__) || defined(__clang__)
+#    define MANUALLY_ALIGNED_STRUCT(alignment) \
+      _Pragma("pack(1)") struct __attribute__((aligned(alignment)))
+#    define STRUCT_END(name, size)                          \
+      _Pragma("pack()") static_assert(sizeof(name) == size, \
+                                      "compiler breaks packing rules")
+#  else
+#    error Unknown compiler, please define structure alignment macros
+#  endif
+#endif  // !defined(MANUALLY_ALIGNED_STRUCT)
+
+// ----------------------------------------------------------------------
+// Convenience macro disabling a particular UBSan check in a function
+
+#if defined(__clang__)
+#  define ARROW_DISABLE_UBSAN(feature) __attribute__((no_sanitize(feature)))
+#else
+#  define ARROW_DISABLE_UBSAN(feature)
+#endif
+
+// ----------------------------------------------------------------------
+// Machine information
+
+#if INTPTR_MAX == INT64_MAX
+#  define ARROW_BITNESS 64
+#elif INTPTR_MAX == INT32_MAX
+#  define ARROW_BITNESS 32
+#else
+#  error Unexpected INTPTR_MAX
+#endif
+
+// ----------------------------------------------------------------------
+// From googletest
+// (also in parquet-cpp)
+
+// When you need to test the private or protected members of a class,
+// use the FRIEND_TEST macro to declare your tests as friends of the
+// class.  For example:
+//
+// class MyClass {
+//  private:
+//   void MyMethod();
+//   FRIEND_TEST(MyClassTest, MyMethod);
+// };
+//
+// class MyClassTest : public testing::Test {
+//   // ...
+// };
+//
+// TEST_F(MyClassTest, MyMethod) {
+//   // Can call MyClass::MyMethod() here.
+// }
+
+#define FRIEND_TEST(test_case_name, test_name) \
+  friend class test_case_name##_##test_name##_Test

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/math_constants.h

@@ -0,0 +1,32 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cmath>
+
+// Not provided by default in MSVC,
+// and _USE_MATH_DEFINES is not reliable with unity builds
+#ifndef M_PI
+#  define M_PI 3.14159265358979323846
+#endif
+#ifndef M_PI_2
+#  define M_PI_2 1.57079632679489661923
+#endif
+#ifndef M_PI_4
+#  define M_PI_4 0.785398163397448309616
+#endif

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/mutex.h

@@ -0,0 +1,85 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <memory>
+
+#include "arrow/util/macros.h"
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+namespace util {
+
+/// A wrapper around std::mutex since we can't use it directly in
+/// public headers due to C++/CLI.
+/// https://docs.microsoft.com/en-us/cpp/standard-library/mutex#remarks
+class ARROW_EXPORT Mutex {
+ public:
+  Mutex();
+  Mutex(Mutex&&) = default;
+  Mutex& operator=(Mutex&&) = default;
+
+  /// A Guard is falsy if a lock could not be acquired.
+  class ARROW_EXPORT Guard {
+   public:
+    Guard() : locked_(NULLPTR, [](Mutex* mutex) {}) {}
+    Guard(Guard&&) = default;
+    Guard& operator=(Guard&&) = default;
+
+    explicit operator bool() const { return bool(locked_); }
+
+    void Unlock() { locked_.reset(); }
+
+   private:
+    explicit Guard(Mutex* locked);
+
+    std::unique_ptr<Mutex, void (*)(Mutex*)> locked_;
+    friend Mutex;
+  };
+
+  Guard TryLock();
+  Guard Lock();
+
+ private:
+  struct Impl;
+  std::unique_ptr<Impl, void (*)(Impl*)> impl_;
+};
+
+#ifndef _WIN32
+/// Return a pointer to a process-wide, process-specific Mutex that can be used
+/// at any point in a child process.  NULL is returned when called in the parent.
+///
+/// The rule is to first check that getpid() corresponds to the parent process pid
+/// and, if not, call this function to lock any after-fork reinitialization code.
+/// Like this:
+///
+///   std::atomic<pid_t> pid{getpid()};
+///   ...
+///   if (pid.load() != getpid()) {
+///     // In child process
+///     auto lock = GlobalForkSafeMutex()->Lock();
+///     if (pid.load() != getpid()) {
+///       // Reinitialize internal structures after fork
+///       ...
+///       pid.store(getpid());
+ARROW_EXPORT
+Mutex* GlobalForkSafeMutex();
+#endif
+
+}  // namespace util
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/parallel.h

@@ -0,0 +1,104 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <utility>
+#include <vector>
+
+#include "arrow/status.h"
+#include "arrow/util/functional.h"
+#include "arrow/util/thread_pool.h"
+#include "arrow/util/vector.h"
+
+namespace arrow {
+namespace internal {
+
+// A parallelizer that takes a `Status(int)` function and calls it with
+// arguments between 0 and `num_tasks - 1`, on an arbitrary number of threads.
+
+template <class FUNCTION>
+Status ParallelFor(int num_tasks, FUNCTION&& func,
+                   Executor* executor = internal::GetCpuThreadPool()) {
+  std::vector<Future<>> futures(num_tasks);
+
+  for (int i = 0; i < num_tasks; ++i) {
+    ARROW_ASSIGN_OR_RAISE(futures[i], executor->Submit(func, i));
+  }
+  auto st = Status::OK();
+  for (auto& fut : futures) {
+    st &= fut.status();
+  }
+  return st;
+}
+
+template <class FUNCTION, typename T,
+          typename R = typename internal::call_traits::return_type<FUNCTION>::ValueType>
+Future<std::vector<R>> ParallelForAsync(std::vector<T> inputs, FUNCTION&& func,
+                                        Executor* executor = internal::GetCpuThreadPool(),
+                                        TaskHints hints = TaskHints{}) {
+  std::vector<Future<R>> futures(inputs.size());
+  for (size_t i = 0; i < inputs.size(); ++i) {
+    ARROW_ASSIGN_OR_RAISE(futures[i],
+                          executor->Submit(hints, func, i, std::move(inputs[i])));
+  }
+  return All(std::move(futures))
+      .Then([](const std::vector<Result<R>>& results) -> Result<std::vector<R>> {
+        return UnwrapOrRaise(results);
+      });
+}
+
+// A parallelizer that takes a `Status(int)` function and calls it with
+// arguments between 0 and `num_tasks - 1`, in sequence or in parallel,
+// depending on the input boolean.
+
+template <class FUNCTION>
+Status OptionalParallelFor(bool use_threads, int num_tasks, FUNCTION&& func,
+                           Executor* executor = internal::GetCpuThreadPool()) {
+  if (use_threads) {
+    return ParallelFor(num_tasks, std::forward<FUNCTION>(func), executor);
+  } else {
+    for (int i = 0; i < num_tasks; ++i) {
+      RETURN_NOT_OK(func(i));
+    }
+    return Status::OK();
+  }
+}
+
+// A parallelizer that takes a `Result<R>(int index, T item)` function and
+// calls it with each item from the input array, in sequence or in parallel,
+// depending on the input boolean.
+
+template <class FUNCTION, typename T,
+          typename R = typename internal::call_traits::return_type<FUNCTION>::ValueType>
+Future<std::vector<R>> OptionalParallelForAsync(
+    bool use_threads, std::vector<T> inputs, FUNCTION&& func,
+    Executor* executor = internal::GetCpuThreadPool(), TaskHints hints = TaskHints{}) {
+  if (use_threads) {
+    return ParallelForAsync(std::move(inputs), std::forward<FUNCTION>(func), executor,
+                            hints);
+  } else {
+    std::vector<R> result(inputs.size());
+    for (size_t i = 0; i < inputs.size(); ++i) {
+      ARROW_ASSIGN_OR_RAISE(result[i], func(i, inputs[i]));
+    }
+    return result;
+  }
+}
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/pcg_random.h

@@ -0,0 +1,33 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include "arrow/vendored/pcg/pcg_random.hpp"  // IWYU pragma: export
+
+namespace arrow {
+namespace random {
+
+using pcg32 = ::arrow_vendored::pcg32;
+using pcg64 = ::arrow_vendored::pcg64;
+using pcg32_fast = ::arrow_vendored::pcg32_fast;
+using pcg64_fast = ::arrow_vendored::pcg64_fast;
+using pcg32_oneseq = ::arrow_vendored::pcg32_oneseq;
+using pcg64_oneseq = ::arrow_vendored::pcg64_oneseq;
+
+}  // namespace random
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/prefetch.h

@@ -0,0 +1,31 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#if defined(__GNUC__)  // GCC and compatible compilers (clang, Intel ICC)
+#  define ARROW_PREFETCH(addr) __builtin_prefetch(addr)
+#elif defined(_MSC_VER)  // MSVC
+#  if defined(ARROW_HAVE_SSE4_2) || defined(ARROW_HAVE_RUNTIME_SSE4_2)
+#    include <nmmintrin.h>
+#    define ARROW_PREFETCH(addr) _mm_prefetch((const char*)(addr), _MM_HINT_T0)
+#  else
+#    define ARROW_PREFETCH(addr)
+#  endif
+#else
+#  define ARROW_PREFETCH(addr)
+#endif

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/queue.h

@@ -0,0 +1,29 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include "arrow/vendored/ProducerConsumerQueue.h"
+
+namespace arrow {
+namespace util {
+
+template <typename T>
+using SpscQueue = arrow_vendored::folly::ProducerConsumerQueue<T>;
+
+}
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/range.h

@@ -0,0 +1,258 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
+#include <iterator>
+#include <numeric>
+#include <tuple>
+#include <utility>
+#include <vector>
+
+namespace arrow::internal {
+
+/// Create a vector containing the values from start up to stop
+template <typename T>
+std::vector<T> Iota(T start, T stop) {
+  if (start > stop) {
+    return {};
+  }
+  std::vector<T> result(static_cast<size_t>(stop - start));
+  std::iota(result.begin(), result.end(), start);
+  return result;
+}
+
+/// Create a vector containing the values from 0 up to length
+template <typename T>
+std::vector<T> Iota(T length) {
+  return Iota(static_cast<T>(0), length);
+}
+
+/// Create a range from a callable which takes a single index parameter
+/// and returns the value of iterator on each call and a length.
+/// Only iterators obtained from the same range should be compared, the
+/// behaviour generally similar to other STL containers.
+template <typename Generator>
+class LazyRange {
+ private:
+  // callable which generates the values
+  // has to be defined at the beginning of the class for type deduction
+  const Generator gen_;
+  // the length of the range
+  int64_t length_;
+#ifdef _MSC_VER
+  // workaround to VS2010 not supporting decltype properly
+  // see https://stackoverflow.com/questions/21782846/decltype-for-class-member-function
+  static Generator gen_static_;
+#endif
+
+ public:
+#ifdef _MSC_VER
+  using return_type = decltype(gen_static_(0));
+#else
+  using return_type = decltype(gen_(0));
+#endif
+
+  /// Construct a new range from a callable and length
+  LazyRange(Generator gen, int64_t length) : gen_(gen), length_(length) {}
+
+  // Class of the dependent iterator, created implicitly by begin and end
+  class RangeIter {
+   public:
+    using difference_type = int64_t;
+    using value_type = return_type;
+    using reference = const value_type&;
+    using pointer = const value_type*;
+    using iterator_category = std::forward_iterator_tag;
+
+#ifdef _MSC_VER
+    // msvc complains about unchecked iterators,
+    // see https://stackoverflow.com/questions/21655496/error-c4996-checked-iterators
+    using _Unchecked_type = typename LazyRange<Generator>::RangeIter;
+#endif
+
+    RangeIter() = delete;
+    RangeIter(const RangeIter& other) = default;
+    RangeIter& operator=(const RangeIter& other) = default;
+
+    RangeIter(const LazyRange<Generator>& range, int64_t index)
+        : range_(&range), index_(index) {}
+
+    const return_type operator*() const { return range_->gen_(index_); }
+
+    RangeIter operator+(difference_type length) const {
+      return RangeIter(*range_, index_ + length);
+    }
+
+    // pre-increment
+    RangeIter& operator++() {
+      ++index_;
+      return *this;
+    }
+
+    // post-increment
+    RangeIter operator++(int) {
+      auto copy = RangeIter(*this);
+      ++index_;
+      return copy;
+    }
+
+    bool operator==(const typename LazyRange<Generator>::RangeIter& other) const {
+      return this->index_ == other.index_ && this->range_ == other.range_;
+    }
+
+    bool operator!=(const typename LazyRange<Generator>::RangeIter& other) const {
+      return this->index_ != other.index_ || this->range_ != other.range_;
+    }
+
+    int64_t operator-(const typename LazyRange<Generator>::RangeIter& other) const {
+      return this->index_ - other.index_;
+    }
+
+    bool operator<(const typename LazyRange<Generator>::RangeIter& other) const {
+      return this->index_ < other.index_;
+    }
+
+   private:
+    // parent range reference
+    const LazyRange* range_;
+    // current index
+    int64_t index_;
+  };
+
+  friend class RangeIter;
+
+  // Create a new begin const iterator
+  RangeIter begin() { return RangeIter(*this, 0); }
+
+  // Create a new end const iterator
+  RangeIter end() { return RangeIter(*this, length_); }
+};
+
+/// Helper function to create a lazy range from a callable (e.g. lambda) and length
+template <typename Generator>
+LazyRange<Generator> MakeLazyRange(Generator&& gen, int64_t length) {
+  return LazyRange<Generator>(std::forward<Generator>(gen), length);
+}
+
+/// \brief A helper for iterating multiple ranges simultaneously, similar to C++23's
+/// zip() view adapter modelled after python's built-in zip() function.
+///
+/// \code {.cpp}
+/// const std::vector<SomeTable>& tables = ...
+/// std::function<std::vector<std::string>()> GetNames = ...
+/// for (auto [table, name] : Zip(tables, GetNames())) {
+///   static_assert(std::is_same_v<decltype(table), const SomeTable&>);
+///   static_assert(std::is_same_v<decltype(name), std::string&>);
+///   // temporaries (like this vector of strings) are kept alive for the
+///   // duration of a loop and are safely movable).
+///   RegisterTableWithName(std::move(name), &table);
+/// }
+/// \endcode
+///
+/// The zipped sequence ends as soon as any of its member ranges ends.
+///
+/// Always use `auto` for the loop's declaration; it will always be a tuple
+/// of references so for example using `const auto&` will compile but will
+/// *look* like forcing const-ness even though the members of the tuple are
+/// still mutable references.
+///
+/// NOTE: we *could* make Zip a more full fledged range and enable things like
+/// - gtest recognizing it as a container; it currently doesn't since Zip is
+///   always mutable so this breaks:
+///       EXPECT_THAT(Zip(std::vector{0}, std::vector{1}),
+///                   ElementsAre(std::tuple{0, 1}));
+/// - letting it be random access when possible so we can do things like *sort*
+///   parallel ranges
+/// - ...
+///
+/// However doing this will increase the compile time overhead of using Zip as
+/// long as we're still using headers. Therefore until we can use c++20 modules:
+/// *don't* extend Zip.
+template <typename Ranges, typename Indices>
+struct Zip;
+
+template <typename... Ranges>
+Zip(Ranges&&...) -> Zip<std::tuple<Ranges...>, std::index_sequence_for<Ranges...>>;
+
+template <typename... Ranges, size_t... I>
+struct Zip<std::tuple<Ranges...>, std::index_sequence<I...>> {
+  explicit Zip(Ranges... ranges) : ranges_(std::forward<Ranges>(ranges)...) {}
+
+  std::tuple<Ranges...> ranges_;
+
+  using sentinel = std::tuple<decltype(std::end(std::get<I>(ranges_)))...>;
+  constexpr sentinel end() { return {std::end(std::get<I>(ranges_))...}; }
+
+  struct iterator : std::tuple<decltype(std::begin(std::get<I>(ranges_)))...> {
+    using std::tuple<decltype(std::begin(std::get<I>(ranges_)))...>::tuple;
+
+    constexpr auto operator*() {
+      return std::tuple<decltype(*std::get<I>(*this))...>{*std::get<I>(*this)...};
+    }
+
+    constexpr iterator& operator++() {
+      (++std::get<I>(*this), ...);
+      return *this;
+    }
+
+    constexpr bool operator!=(const sentinel& s) const {
+      bool all_iterators_valid = (... && (std::get<I>(*this) != std::get<I>(s)));
+      return all_iterators_valid;
+    }
+  };
+  constexpr iterator begin() { return {std::begin(std::get<I>(ranges_))...}; }
+};
+
+/// \brief A lazy sequence of integers which starts from 0 and never stops.
+///
+/// This can be used in conjunction with Zip() to emulate python's built-in
+/// enumerate() function:
+///
+/// \code {.cpp}
+/// const std::vector<SomeTable>& tables = ...
+/// for (auto [i, table] : Zip(Enumerate<>, tables)) {
+///   std::cout << "#" << i << ": " << table.name() << std::endl;
+/// }
+/// \endcode
+template <typename I = size_t>
+constexpr auto Enumerate = [] {
+  struct {
+    struct sentinel {};
+    constexpr sentinel end() const { return {}; }
+
+    struct iterator {
+      I value{0};
+
+      constexpr I operator*() { return value; }
+
+      constexpr iterator& operator++() {
+        ++value;
+        return *this;
+      }
+
+      constexpr std::true_type operator!=(sentinel) const { return {}; }
+    };
+    constexpr iterator begin() const { return {}; }
+  } out;
+
+  return out;
+}();
+
+}  // namespace arrow::internal

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/ree_util.h

@@ -0,0 +1,584 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <algorithm>
+#include <cassert>
+#include <cstdint>
+
+#include "arrow/array/data.h"
+#include "arrow/type_traits.h"
+#include "arrow/util/checked_cast.h"
+#include "arrow/util/macros.h"
+
+namespace arrow {
+namespace ree_util {
+
+/// \brief Get the child array holding the run ends from an REE array
+inline const ArraySpan& RunEndsArray(const ArraySpan& span) { return span.child_data[0]; }
+
+/// \brief Get the child array holding the data values from an REE array
+inline const ArraySpan& ValuesArray(const ArraySpan& span) { return span.child_data[1]; }
+
+/// \brief Get a pointer to run ends values of an REE array
+template <typename RunEndCType>
+const RunEndCType* RunEnds(const ArraySpan& span) {
+  assert(RunEndsArray(span).type->id() == CTypeTraits<RunEndCType>::ArrowType::type_id);
+  return RunEndsArray(span).GetValues<RunEndCType>(1);
+}
+
+/// \brief Perform basic validations on the parameters of an REE array
+/// and its two children arrays
+///
+/// All the checks complete in O(1) time. Consequently, this function:
+/// - DOES NOT check that run_ends is sorted and all-positive
+/// - DOES NOT check the actual contents of the run_ends and values arrays
+Status ValidateRunEndEncodedChildren(const RunEndEncodedType& type,
+                                     int64_t logical_length,
+                                     const std::shared_ptr<ArrayData>& run_ends_data,
+                                     const std::shared_ptr<ArrayData>& values_data,
+                                     int64_t null_count, int64_t logical_offset);
+
+/// \brief Compute the logical null count of an REE array
+int64_t LogicalNullCount(const ArraySpan& span);
+
+namespace internal {
+
+/// \brief Uses binary-search to find the physical offset given a logical offset
+/// and run-end values
+///
+/// \return the physical offset or run_ends_size if the physical offset is not
+/// found in run_ends
+template <typename RunEndCType>
+int64_t FindPhysicalIndex(const RunEndCType* run_ends, int64_t run_ends_size, int64_t i,
+                          int64_t absolute_offset) {
+  assert(absolute_offset + i >= 0);
+  auto it = std::upper_bound(run_ends, run_ends + run_ends_size, absolute_offset + i);
+  int64_t result = std::distance(run_ends, it);
+  assert(result <= run_ends_size);
+  return result;
+}
+
+/// \brief Uses binary-search to calculate the range of physical values (and
+/// run-ends) necessary to represent the logical range of values from
+/// offset to length
+///
+/// \return a pair of physical offset and physical length
+template <typename RunEndCType>
+std::pair<int64_t, int64_t> FindPhysicalRange(const RunEndCType* run_ends,
+                                              int64_t run_ends_size, int64_t length,
+                                              int64_t offset) {
+  const int64_t physical_offset =
+      FindPhysicalIndex<RunEndCType>(run_ends, run_ends_size, 0, offset);
+  // The physical length is calculated by finding the offset of the last element
+  // and adding 1 to it, so first we ensure there is at least one element.
+  if (length == 0) {
+    return {physical_offset, 0};
+  }
+  const int64_t physical_index_of_last = FindPhysicalIndex<RunEndCType>(
+      run_ends + physical_offset, run_ends_size - physical_offset, length - 1, offset);
+
+  assert(physical_index_of_last < run_ends_size - physical_offset);
+  return {physical_offset, physical_index_of_last + 1};
+}
+
+/// \brief Uses binary-search to calculate the number of physical values (and
+/// run-ends) necessary to represent the logical range of values from
+/// offset to length
+template <typename RunEndCType>
+int64_t FindPhysicalLength(const RunEndCType* run_ends, int64_t run_ends_size,
+                           int64_t length, int64_t offset) {
+  auto [_, physical_length] =
+      FindPhysicalRange<RunEndCType>(run_ends, run_ends_size, length, offset);
+  // GH-37107: This is a workaround for GCC 7. GCC 7 doesn't ignore
+  // variables in structured binding automatically from unused
+  // variables when one of these variables are used.
+  // See also: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81767
+  ARROW_UNUSED(_);
+  return physical_length;
+}
+
+/// \brief Find the physical index into the values array of the REE ArraySpan
+///
+/// This function uses binary-search, so it has a O(log N) cost.
+template <typename RunEndCType>
+int64_t FindPhysicalIndex(const ArraySpan& span, int64_t i, int64_t absolute_offset) {
+  const int64_t run_ends_size = RunEndsArray(span).length;
+  return FindPhysicalIndex(RunEnds<RunEndCType>(span), run_ends_size, i, absolute_offset);
+}
+
+/// \brief Find the physical length of an REE ArraySpan
+///
+/// The physical length of an REE is the number of physical values (and
+/// run-ends) necessary to represent the logical range of values from
+/// offset to length.
+///
+/// Avoid calling this function if the physical length can be established in
+/// some other way (e.g. when iterating over the runs sequentially until the
+/// end). This function uses binary-search, so it has a O(log N) cost.
+template <typename RunEndCType>
+int64_t FindPhysicalLength(const ArraySpan& span) {
+  return FindPhysicalLength(
+      /*run_ends=*/RunEnds<RunEndCType>(span),
+      /*run_ends_size=*/RunEndsArray(span).length,
+      /*length=*/span.length,
+      /*offset=*/span.offset);
+}
+
+template <typename RunEndCType>
+struct PhysicalIndexFinder;
+
+// non-inline implementations for each run-end type
+ARROW_EXPORT int64_t FindPhysicalIndexImpl16(PhysicalIndexFinder<int16_t>& self,
+                                             int64_t i);
+ARROW_EXPORT int64_t FindPhysicalIndexImpl32(PhysicalIndexFinder<int32_t>& self,
+                                             int64_t i);
+ARROW_EXPORT int64_t FindPhysicalIndexImpl64(PhysicalIndexFinder<int64_t>& self,
+                                             int64_t i);
+
+/// \brief Stateful version of FindPhysicalIndex() that caches the result of
+/// the previous search and uses it to optimize the next search.
+///
+/// When new queries for the physical index of a logical index come in,
+/// binary search is performed again but the first candidate checked is the
+/// result of the previous search (cached physical index) instead of the
+/// midpoint of the run-ends array.
+///
+/// If that test fails, internal::FindPhysicalIndex() is called with one of the
+/// partitions defined by the cached index. If the queried logical indices
+/// follow an increasing or decreasing pattern, this first test is much more
+/// effective in (1) finding the answer right away (close logical indices belong
+/// to the same runs) or (2) discarding many more candidates than probing
+/// the midpoint would.
+///
+/// The most adversarial case (i.e. alternating between 0 and length-1 queries)
+/// only adds one extra binary search probe when compared to always starting
+/// binary search from the midpoint without any of these optimizations.
+///
+/// \tparam RunEndCType The numeric type of the run-ends array.
+template <typename RunEndCType>
+struct PhysicalIndexFinder {
+  const ArraySpan array_span;
+  const RunEndCType* run_ends;
+  int64_t last_physical_index = 0;
+
+  explicit PhysicalIndexFinder(const ArrayData& data)
+      : array_span(data),
+        run_ends(RunEndsArray(array_span).template GetValues<RunEndCType>(1)) {
+    assert(CTypeTraits<RunEndCType>::ArrowType::type_id ==
+           ::arrow::internal::checked_cast<const RunEndEncodedType&>(*data.type)
+               .run_end_type()
+               ->id());
+  }
+
+  /// \brief Find the physical index into the values array of the REE array.
+  ///
+  /// \pre 0 <= i < array_span.length()
+  /// \param i the logical index into the REE array
+  /// \return the physical index into the values array
+  int64_t FindPhysicalIndex(int64_t i) {
+    if constexpr (std::is_same_v<RunEndCType, int16_t>) {
+      return FindPhysicalIndexImpl16(*this, i);
+    } else if constexpr (std::is_same_v<RunEndCType, int32_t>) {
+      return FindPhysicalIndexImpl32(*this, i);
+    } else {
+      static_assert(std::is_same_v<RunEndCType, int64_t>, "Unsupported RunEndCType.");
+      return FindPhysicalIndexImpl64(*this, i);
+    }
+  }
+};
+
+}  // namespace internal
+
+/// \brief Find the physical index into the values array of the REE ArraySpan
+///
+/// This function uses binary-search, so it has a O(log N) cost.
+ARROW_EXPORT int64_t FindPhysicalIndex(const ArraySpan& span, int64_t i,
+                                       int64_t absolute_offset);
+
+/// \brief Find the physical length of an REE ArraySpan
+///
+/// The physical length of an REE is the number of physical values (and
+/// run-ends) necessary to represent the logical range of values from
+/// offset to length.
+///
+/// Avoid calling this function if the physical length can be established in
+/// some other way (e.g. when iterating over the runs sequentially until the
+/// end). This function uses binary-search, so it has a O(log N) cost.
+ARROW_EXPORT int64_t FindPhysicalLength(const ArraySpan& span);
+
+/// \brief Find the physical range of physical values referenced by the REE in
+/// the logical range from offset to offset + length
+///
+/// \return a pair of physical offset and physical length
+ARROW_EXPORT std::pair<int64_t, int64_t> FindPhysicalRange(const ArraySpan& span,
+                                                           int64_t offset,
+                                                           int64_t length);
+
+// Publish PhysicalIndexFinder outside of the internal namespace.
+template <typename RunEndCType>
+using PhysicalIndexFinder = internal::PhysicalIndexFinder<RunEndCType>;
+
+template <typename RunEndCType>
+class RunEndEncodedArraySpan {
+ private:
+  struct PrivateTag {};
+
+ public:
+  /// \brief Iterator representing the current run during iteration over a
+  /// run-end encoded array
+  class Iterator {
+   public:
+    Iterator(PrivateTag, const RunEndEncodedArraySpan& span, int64_t logical_pos,
+             int64_t physical_pos)
+        : span(span), logical_pos_(logical_pos), physical_pos_(physical_pos) {}
+
+    /// \brief Return the physical index of the run
+    ///
+    /// The values array can be addressed with this index to get the value
+    /// that makes up the run.
+    ///
+    /// NOTE: if this Iterator is equal to RunEndEncodedArraySpan::end(),
+    /// the value returned is undefined.
+    int64_t index_into_array() const { return physical_pos_; }
+
+    /// \brief Return the initial logical position of the run
+    ///
+    /// If this Iterator is equal to RunEndEncodedArraySpan::end(), this is
+    /// the same as RunEndEncodedArraySpan::length().
+    int64_t logical_position() const { return logical_pos_; }
+
+    /// \brief Return the logical position immediately after the run.
+    ///
+    /// Pre-condition: *this != RunEndEncodedArraySpan::end()
+    int64_t run_end() const { return span.run_end(physical_pos_); }
+
+    /// \brief Returns the logical length of the run.
+    ///
+    /// Pre-condition: *this != RunEndEncodedArraySpan::end()
+    int64_t run_length() const { return run_end() - logical_pos_; }
+
+    /// \brief Check if the iterator is at the end of the array.
+    ///
+    /// This can be used to avoid paying the cost of a call to
+    /// RunEndEncodedArraySpan::end().
+    ///
+    /// \return true if the iterator is at the end of the array
+    bool is_end(const RunEndEncodedArraySpan& span) const {
+      return logical_pos_ >= span.length();
+    }
+
+    Iterator& operator++() {
+      logical_pos_ = span.run_end(physical_pos_);
+      physical_pos_ += 1;
+      return *this;
+    }
+
+    Iterator operator++(int) {
+      const Iterator prev = *this;
+      ++(*this);
+      return prev;
+    }
+
+    Iterator& operator--() {
+      physical_pos_ -= 1;
+      logical_pos_ = (physical_pos_ > 0) ? span.run_end(physical_pos_ - 1) : 0;
+      return *this;
+    }
+
+    Iterator operator--(int) {
+      const Iterator prev = *this;
+      --(*this);
+      return prev;
+    }
+
+    bool operator==(const Iterator& other) const {
+      return logical_pos_ == other.logical_pos_;
+    }
+
+    bool operator!=(const Iterator& other) const {
+      return logical_pos_ != other.logical_pos_;
+    }
+
+   public:
+    const RunEndEncodedArraySpan& span;
+
+   private:
+    int64_t logical_pos_;
+    int64_t physical_pos_;
+  };
+
+  // Prevent implicit ArrayData -> ArraySpan conversion in
+  // RunEndEncodedArraySpan instantiation.
+  explicit RunEndEncodedArraySpan(const ArrayData& data) = delete;
+
+  /// \brief Construct a RunEndEncodedArraySpan from an ArraySpan and new
+  /// absolute offset and length.
+  ///
+  /// RunEndEncodedArraySpan{span, off, len} is equivalent to:
+  ///
+  ///   span.SetSlice(off, len);
+  ///   RunEndEncodedArraySpan{span}
+  ///
+  /// ArraySpan::SetSlice() updates the null_count to kUnknownNullCount, but
+  /// we don't need that here as REE arrays have null_count set to 0 by
+  /// convention.
+  explicit RunEndEncodedArraySpan(const ArraySpan& array_span, int64_t offset,
+                                  int64_t length)
+      : array_span_{array_span},
+        run_ends_(RunEnds<RunEndCType>(array_span_)),
+        length_(length),
+        offset_(offset) {
+    assert(array_span_.type->id() == Type::RUN_END_ENCODED);
+  }
+
+  explicit RunEndEncodedArraySpan(const ArraySpan& array_span)
+      : RunEndEncodedArraySpan(array_span, array_span.offset, array_span.length) {}
+
+  int64_t offset() const { return offset_; }
+  int64_t length() const { return length_; }
+
+  int64_t PhysicalIndex(int64_t logical_pos) const {
+    return internal::FindPhysicalIndex(run_ends_, RunEndsArray(array_span_).length,
+                                       logical_pos, offset_);
+  }
+
+  /// \brief Create an iterator from a logical position and its
+  /// pre-computed physical offset into the run ends array
+  ///
+  /// \param logical_pos is an index in the [0, length()] range
+  /// \param physical_offset the pre-calculated PhysicalIndex(logical_pos)
+  Iterator iterator(int64_t logical_pos, int64_t physical_offset) const {
+    return Iterator{PrivateTag{}, *this, logical_pos, physical_offset};
+  }
+
+  /// \brief Create an iterator from a logical position
+  ///
+  /// \param logical_pos is an index in the [0, length()] range
+  Iterator iterator(int64_t logical_pos) const {
+    if (logical_pos < length()) {
+      return iterator(logical_pos, PhysicalIndex(logical_pos));
+    }
+    // If logical_pos is above the valid range, use length() as the logical
+    // position and calculate the physical address right after the last valid
+    // physical position. Which is the physical index of the last logical
+    // position, plus 1.
+    return (length() == 0) ? iterator(0, PhysicalIndex(0))
+                           : iterator(length(), PhysicalIndex(length() - 1) + 1);
+  }
+
+  /// \brief Create an iterator representing the logical begin of the run-end
+  /// encoded array
+  Iterator begin() const { return iterator(0, PhysicalIndex(0)); }
+
+  /// \brief Create an iterator representing the first invalid logical position
+  /// of the run-end encoded array
+  ///
+  /// \warning Avoid calling end() in a loop, as it will recompute the physical
+  /// length of the array on each call (O(log N) cost per call).
+  ///
+  /// You can write your loops like this instead:
+  ///
+  /// \code
+  /// for (auto it = array.begin(), end = array.end(); it != end; ++it) {
+  ///   // ...
+  /// }
+  /// \endcode
+  ///
+  /// Or this version that does not look like idiomatic C++, but removes
+  /// the need for calling end() completely:
+  ///
+  /// \code
+  /// for (auto it = array.begin(); !it.is_end(array); ++it) {
+  ///   // ...
+  /// }
+  /// \endcode
+  Iterator end() const {
+    return iterator(length(),
+                    (length() == 0) ? PhysicalIndex(0) : PhysicalIndex(length() - 1) + 1);
+  }
+
+  // Pre-condition: physical_pos < RunEndsArray(array_span_).length);
+  inline int64_t run_end(int64_t physical_pos) const {
+    assert(physical_pos < RunEndsArray(array_span_).length);
+    // Logical index of the end of the run at physical_pos with offset applied
+    const int64_t logical_run_end =
+        std::max<int64_t>(static_cast<int64_t>(run_ends_[physical_pos]) - offset(), 0);
+    // The current run may go further than the logical length, cap it
+    return std::min(logical_run_end, length());
+  }
+
+ private:
+  const ArraySpan& array_span_;
+  const RunEndCType* run_ends_;
+  const int64_t length_;
+  const int64_t offset_;
+};
+
+/// \brief Iterate over two run-end encoded arrays in runs or sub-runs that are
+/// inside run boundaries on both inputs
+///
+/// Both RunEndEncodedArraySpan should have the same logical length. Instances
+/// of this iterator only hold references to the RunEndEncodedArraySpan inputs.
+template <typename Left, typename Right>
+class MergedRunsIterator {
+ private:
+  using LeftIterator = typename Left::Iterator;
+  using RightIterator = typename Right::Iterator;
+
+  MergedRunsIterator(LeftIterator left_it, RightIterator right_it,
+                     int64_t common_logical_length, int64_t common_logical_pos)
+      : ree_iterators_{std::move(left_it), std::move(right_it)},
+        logical_length_(common_logical_length),
+        logical_pos_(common_logical_pos) {}
+
+ public:
+  /// \brief Construct a MergedRunsIterator positioned at logical position 0.
+  ///
+  /// Pre-condition: left.length() == right.length()
+  MergedRunsIterator(const Left& left, const Right& right)
+      : MergedRunsIterator(left.begin(), right.begin(), left.length(), 0) {
+    assert(left.length() == right.length());
+  }
+
+  static Result<MergedRunsIterator> MakeBegin(const Left& left, const Right& right) {
+    if (left.length() != right.length()) {
+      return Status::Invalid(
+          "MergedRunsIterator expects RunEndEncodedArraySpans of the same length");
+    }
+    return MergedRunsIterator(left, right);
+  }
+
+  static Result<MergedRunsIterator> MakeEnd(const Left& left, const Right& right) {
+    if (left.length() != right.length()) {
+      return Status::Invalid(
+          "MergedRunsIterator expects RunEndEncodedArraySpans of the same length");
+    }
+    return MergedRunsIterator(left.end(), right.end(), left.length(), left.length());
+  }
+
+  /// \brief Return the left RunEndEncodedArraySpan child
+  const Left& left() const { return std::get<0>(ree_iterators_).span; }
+
+  /// \brief Return the right RunEndEncodedArraySpan child
+  const Right& right() const { return std::get<1>(ree_iterators_).span; }
+
+  /// \brief Return the initial logical position of the run
+  ///
+  /// If is_end(), this is the same as length().
+  int64_t logical_position() const { return logical_pos_; }
+
+  /// \brief Whether the iterator is at logical position 0.
+  bool is_begin() const { return logical_pos_ == 0; }
+
+  /// \brief Whether the iterator has reached the end of both arrays
+  bool is_end() const { return logical_pos_ == logical_length_; }
+
+  /// \brief Return the logical position immediately after the run.
+  ///
+  /// Pre-condition: !is_end()
+  int64_t run_end() const {
+    const auto& left_it = std::get<0>(ree_iterators_);
+    const auto& right_it = std::get<1>(ree_iterators_);
+    return std::min(left_it.run_end(), right_it.run_end());
+  }
+
+  /// \brief returns the logical length of the current run
+  ///
+  /// Pre-condition: !is_end()
+  int64_t run_length() const { return run_end() - logical_pos_; }
+
+  /// \brief Return a physical index into the values array of a given input,
+  /// pointing to the value of the current run
+  template <size_t input_id>
+  int64_t index_into_array() const {
+    return std::get<input_id>(ree_iterators_).index_into_array();
+  }
+
+  int64_t index_into_left_array() const { return index_into_array<0>(); }
+  int64_t index_into_right_array() const { return index_into_array<1>(); }
+
+  MergedRunsIterator& operator++() {
+    auto& left_it = std::get<0>(ree_iterators_);
+    auto& right_it = std::get<1>(ree_iterators_);
+
+    const int64_t left_run_end = left_it.run_end();
+    const int64_t right_run_end = right_it.run_end();
+
+    if (left_run_end < right_run_end) {
+      logical_pos_ = left_run_end;
+      ++left_it;
+    } else if (left_run_end > right_run_end) {
+      logical_pos_ = right_run_end;
+      ++right_it;
+    } else {
+      logical_pos_ = left_run_end;
+      ++left_it;
+      ++right_it;
+    }
+    return *this;
+  }
+
+  MergedRunsIterator operator++(int) {
+    MergedRunsIterator prev = *this;
+    ++(*this);
+    return prev;
+  }
+
+  MergedRunsIterator& operator--() {
+    auto& left_it = std::get<0>(ree_iterators_);
+    auto& right_it = std::get<1>(ree_iterators_);
+
+    // The logical position of each iterator is the run_end() of the previous run.
+    const int64_t left_logical_pos = left_it.logical_position();
+    const int64_t right_logical_pos = right_it.logical_position();
+
+    if (left_logical_pos < right_logical_pos) {
+      --right_it;
+      logical_pos_ = std::max(left_logical_pos, right_it.logical_position());
+    } else if (left_logical_pos > right_logical_pos) {
+      --left_it;
+      logical_pos_ = std::max(left_it.logical_position(), right_logical_pos);
+    } else {
+      --left_it;
+      --right_it;
+      logical_pos_ = std::max(left_it.logical_position(), right_it.logical_position());
+    }
+    return *this;
+  }
+
+  MergedRunsIterator operator--(int) {
+    MergedRunsIterator prev = *this;
+    --(*this);
+    return prev;
+  }
+
+  bool operator==(const MergedRunsIterator& other) const {
+    return logical_pos_ == other.logical_position();
+  }
+
+  bool operator!=(const MergedRunsIterator& other) const { return !(*this == other); }
+
+ private:
+  std::tuple<LeftIterator, RightIterator> ree_iterators_;
+  const int64_t logical_length_;
+  int64_t logical_pos_;
+};
+
+}  // namespace ree_util
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/regex.h

@@ -0,0 +1,51 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cassert>
+#include <initializer_list>
+#include <regex>
+#include <string_view>
+#include <type_traits>
+
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+namespace internal {
+
+/// Match regex against target and produce string_views out of matches.
+inline bool RegexMatch(const std::regex& regex, std::string_view target,
+                       std::initializer_list<std::string_view*> out_matches) {
+  assert(regex.mark_count() == out_matches.size());
+
+  std::match_results<decltype(target.begin())> match;
+  if (!std::regex_match(target.begin(), target.end(), match, regex)) {
+    return false;
+  }
+
+  // Match #0 is the whole matched sequence
+  assert(regex.mark_count() + 1 == match.size());
+  auto out_it = out_matches.begin();
+  for (size_t i = 1; i < match.size(); ++i) {
+    **out_it++ = target.substr(match.position(i), match.length(i));
+  }
+  return true;
+}
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/rows_to_batches.h

@@ -0,0 +1,163 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include "arrow/record_batch.h"
+#include "arrow/result.h"
+#include "arrow/status.h"
+#include "arrow/table_builder.h"
+#include "arrow/util/iterator.h"
+
+#include <type_traits>
+
+namespace arrow::util {
+
+namespace detail {
+
+// Default identity function row accessor. Used to for the common case where the value
+// of each row iterated over is it's self also directly iterable.
+[[nodiscard]] constexpr inline auto MakeDefaultRowAccessor() {
+  return [](auto& x) -> Result<decltype(std::ref(x))> { return std::ref(x); };
+}
+
+// Meta-function to check if a type `T` is a range (iterable using `std::begin()` /
+// `std::end()`). `is_range<T>::value` will be false if `T` is not a valid range.
+template <typename T, typename = void>
+struct is_range : std::false_type {};
+
+template <typename T>
+struct is_range<T, std::void_t<decltype(std::begin(std::declval<T>())),
+                               decltype(std::end(std::declval<T>()))>> : std::true_type {
+};
+
+}  // namespace detail
+
+/// Delete overload for `const Range&& rows` because the data's lifetime must exceed
+/// the lifetime of the function call. `data` will be read when client uses the
+/// `RecordBatchReader`
+template <class Range, class DataPointConvertor,
+          class RowAccessor = decltype(detail::MakeDefaultRowAccessor())>
+[[nodiscard]] typename std::enable_if_t<detail::is_range<Range>::value,
+                                        Result<std::shared_ptr<RecordBatchReader>>>
+/* Result<std::shared_ptr<RecordBatchReader>>> */ RowsToBatches(
+    const std::shared_ptr<Schema>& schema, const Range&& rows,
+    DataPointConvertor&& data_point_convertor,
+    RowAccessor&& row_accessor = detail::MakeDefaultRowAccessor(),
+    MemoryPool* pool = default_memory_pool(),
+    const std::size_t batch_size = 1024) = delete;
+
+/// \brief Utility function for converting any row-based structure into an
+/// `arrow::RecordBatchReader` (this can be easily converted to an `arrow::Table` using
+/// `arrow::RecordBatchReader::ToTable()`).
+///
+/// Examples of supported types:
+/// - `std::vector<std::vector<std::variant<int, bsl::string>>>`
+/// - `std::vector<MyRowStruct>`
+
+/// If `rows` (client’s row-based structure) is not a valid C++ range, the client will
+/// need to either make it iterable, or make an adapter/wrapper that is a valid C++
+/// range.
+
+/// The client must provide a `DataPointConvertor` callable type that will convert the
+/// structure’s data points into the corresponding arrow types.
+
+/// Complex nested rows can be supported by providing a custom `row_accessor` instead
+/// of the default.
+
+/// Example usage:
+/// \code{.cpp}
+/// auto IntConvertor = [](ArrayBuilder& array_builder, int value) {
+///  return static_cast<Int64Builder&>(array_builder).Append(value);
+/// };
+/// std::vector<std::vector<int>> data = {{1, 2, 4}, {5, 6, 7}};
+/// auto batches = RowsToBatches(kTestSchema, data, IntConvertor);
+/// \endcode
+
+/// \param[in] schema - The schema to be used in the `RecordBatchReader`
+
+/// \param[in] rows - Iterable row-based structure that will be converted to arrow
+/// batches
+
+/// \param[in] data_point_convertor - Client provided callable type that will convert
+/// the structure’s data points into the corresponding arrow types. The convertor must
+/// return an error `Status` if an error happens during conversion.
+
+/// \param[in] row_accessor - In the common case where the value of each row iterated
+/// over is it's self also directly iterable, the client can just use the default.
+/// The provided callable must take the values of the `rows` range and return a
+/// `std::reference_wrapper<Range>` to the data points in a given row. The data points
+/// must be in order of their corresponding fields in the schema.
+/// see: /ref `MakeDefaultRowAccessor`
+
+/// \param[in] pool - The MemoryPool to use for allocations.
+
+/// \param[in] batch_size - Number of rows to insert into each RecordBatch.
+
+/// \return `Result<std::shared_ptr<RecordBatchReader>>>` result will be a
+/// `std::shared_ptr<RecordBatchReader>>` if not errors occurred, else an error status.
+template <class Range, class DataPointConvertor,
+          class RowAccessor = decltype(detail::MakeDefaultRowAccessor())>
+[[nodiscard]] typename std::enable_if_t<detail::is_range<Range>::value,
+                                        Result<std::shared_ptr<RecordBatchReader>>>
+/* Result<std::shared_ptr<RecordBatchReader>>> */ RowsToBatches(
+    const std::shared_ptr<Schema>& schema, const Range& rows,
+    DataPointConvertor&& data_point_convertor,
+    RowAccessor&& row_accessor = detail::MakeDefaultRowAccessor(),
+    MemoryPool* pool = default_memory_pool(), const std::size_t batch_size = 1024) {
+  auto make_next_batch =
+      [pool = pool, batch_size = batch_size, rows_ittr = std::begin(rows),
+       rows_ittr_end = std::end(rows), schema = schema,
+       row_accessor = std::forward<RowAccessor>(row_accessor),
+       data_point_convertor = std::forward<DataPointConvertor>(
+           data_point_convertor)]() mutable -> Result<std::shared_ptr<RecordBatch>> {
+    if (rows_ittr == rows_ittr_end) return NULLPTR;
+
+    ARROW_ASSIGN_OR_RAISE(auto record_batch_builder,
+                          RecordBatchBuilder::Make(schema, pool, batch_size));
+
+    for (size_t i = 0; i < batch_size && (rows_ittr != rows_ittr_end);
+         i++, std::advance(rows_ittr, 1)) {
+      int col_index = 0;
+      ARROW_ASSIGN_OR_RAISE(const auto row, row_accessor(*rows_ittr));
+
+      // If the accessor returns a `std::reference_wrapper` unwrap if
+      const auto& row_unwrapped = [&]() {
+        if constexpr (detail::is_range<decltype(row)>::value)
+          return row;
+        else
+          return row.get();
+      }();
+
+      for (auto& data_point : row_unwrapped) {
+        ArrayBuilder* array_builder = record_batch_builder->GetField(col_index);
+        ARROW_RETURN_IF(array_builder == NULLPTR,
+                        Status::Invalid("array_builder == NULLPTR"));
+
+        ARROW_RETURN_NOT_OK(data_point_convertor(*array_builder, data_point));
+        col_index++;
+      }
+    }
+
+    ARROW_ASSIGN_OR_RAISE(auto result, record_batch_builder->Flush());
+    return result;
+  };
+  return RecordBatchReader::MakeFromIterator(MakeFunctionIterator(make_next_batch),
+                                             schema);
+}
+
+}  // namespace arrow::util

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/secure_string.h

@@ -0,0 +1,72 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cstdint>
+#include <string>
+
+#include "arrow/util/span.h"
+#include "arrow/util/visibility.h"
+
+namespace arrow::util {
+/**
+ * A secure string that ensures the wrapped string is cleared from memory on
+ * deconstruction. This class can only be created from std::string that are securely
+ * erased after creation.
+ *
+ * Note: This class does not provide a constructor / assignment operator that copies a
+ * std::string because that would allow code to create a SecureString while accidentally
+ * not noticing the need to securely erasing the argument after invoking the constructor /
+ * calling the assignment operator.
+ */
+class ARROW_EXPORT SecureString {
+ public:
+  SecureString() = default;
+  SecureString(SecureString&&) noexcept;
+  SecureString(const SecureString&) = default;
+  explicit SecureString(std::string&&) noexcept;
+  explicit SecureString(size_t, char) noexcept;
+
+  SecureString& operator=(SecureString&&) noexcept;
+  SecureString& operator=(const SecureString&);
+  SecureString& operator=(std::string&&) noexcept;
+
+  bool operator==(const SecureString&) const;
+  bool operator!=(const SecureString&) const;
+
+  ~SecureString() { Dispose(); }
+
+  [[nodiscard]] bool empty() const;
+  [[nodiscard]] std::size_t size() const;
+  [[nodiscard]] std::size_t length() const;
+  [[nodiscard]] std::size_t capacity() const;
+
+  [[nodiscard]] span<uint8_t> as_span();
+  [[nodiscard]] span<const uint8_t> as_span() const;
+  [[nodiscard]] std::string_view as_view() const;
+
+  void Dispose();
+
+  static void SecureClear(std::string*);
+  static void SecureClear(uint8_t* data, size_t size);
+
+ private:
+  std::string secret_;
+};
+
+}  // namespace arrow::util

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/simd.h

@@ -0,0 +1,51 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#ifdef _MSC_VER
+// MSVC x86_64/arm64
+
+#  if defined(_M_AMD64) || defined(_M_X64)
+#    include <intrin.h>
+#  endif
+
+#else
+// gcc/clang (possibly others)
+
+#  if defined(ARROW_HAVE_BMI2) || defined(ARROW_HAVE_RUNTIME_BMI2)
+#    include <x86intrin.h>
+#  endif
+
+#  if defined(ARROW_HAVE_AVX2) || defined(ARROW_HAVE_AVX512) || \
+      defined(ARROW_HAVE_RUNTIME_AVX2) || defined(ARROW_HAVE_RUNTIME_AVX512)
+#    include <immintrin.h>
+#  elif defined(ARROW_HAVE_SSE4_2) || defined(ARROW_HAVE_RUNTIME_SSE4_2)
+#    include <nmmintrin.h>
+#  endif
+
+#  ifdef ARROW_HAVE_NEON
+#    include <arm_neon.h>
+#  endif
+
+// GH-44098: Workaround for missing _mm256_set_m128i in older versions of GCC.
+#  if defined(__GNUC__) && !defined(__clang__) && __GNUC__ < 8
+#    define _mm256_set_m128i(hi, lo) \
+      _mm256_inserti128_si256(_mm256_castsi128_si256(lo), (hi), 1)
+#  endif
+
+#endif

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/small_vector.h

@@ -0,0 +1,512 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <algorithm>
+#include <cassert>
+#include <cstddef>
+#include <initializer_list>
+#include <iterator>
+#include <limits>
+#include <new>
+#include <type_traits>
+#include <utility>
+#include <vector>
+
+#include "arrow/util/aligned_storage.h"
+#include "arrow/util/macros.h"
+
+namespace arrow {
+namespace internal {
+
+template <typename T, size_t N, bool NonTrivialDestructor>
+struct StaticVectorStorageBase {
+  using storage_type = AlignedStorage<T>;
+
+  storage_type static_data_[N];
+  size_t size_ = 0;
+
+  void destroy() noexcept {}
+};
+
+template <typename T, size_t N>
+struct StaticVectorStorageBase<T, N, true> {
+  using storage_type = AlignedStorage<T>;
+
+  storage_type static_data_[N];
+  size_t size_ = 0;
+
+  ~StaticVectorStorageBase() noexcept { destroy(); }
+
+  void destroy() noexcept { storage_type::destroy_several(static_data_, size_); }
+};
+
+template <typename T, size_t N, bool D = !std::is_trivially_destructible<T>::value>
+struct StaticVectorStorage : public StaticVectorStorageBase<T, N, D> {
+  using Base = StaticVectorStorageBase<T, N, D>;
+  using typename Base::storage_type;
+
+  using Base::size_;
+  using Base::static_data_;
+
+  StaticVectorStorage() noexcept = default;
+
+  constexpr storage_type* storage_ptr() { return static_data_; }
+
+  constexpr const storage_type* const_storage_ptr() const { return static_data_; }
+
+  // Adjust storage size, but don't initialize any objects
+  void bump_size(size_t addend) {
+    assert(size_ + addend <= N);
+    size_ += addend;
+  }
+
+  void ensure_capacity(size_t min_capacity) { assert(min_capacity <= N); }
+
+  // Adjust storage size, but don't destroy any objects
+  void reduce_size(size_t reduce_by) {
+    assert(reduce_by <= size_);
+    size_ -= reduce_by;
+  }
+
+  // Move objects from another storage, but don't destroy any objects currently
+  // stored in *this.
+  // You need to call destroy() first if necessary (e.g. in a
+  // move assignment operator).
+  void move_construct(StaticVectorStorage&& other) noexcept {
+    size_ = other.size_;
+    if (size_ != 0) {
+      // Use a compile-time memcpy size (N) for trivial types
+      storage_type::move_construct_several(other.static_data_, static_data_, size_, N);
+    }
+  }
+
+  constexpr size_t capacity() const { return N; }
+
+  constexpr size_t max_size() const { return N; }
+
+  void reserve(size_t n) {}
+
+  void clear() {
+    storage_type::destroy_several(static_data_, size_);
+    size_ = 0;
+  }
+};
+
+template <typename T, size_t N>
+struct SmallVectorStorage {
+  using storage_type = AlignedStorage<T>;
+
+  storage_type static_data_[N];
+  size_t size_ = 0;
+  storage_type* data_ = static_data_;
+  size_t dynamic_capacity_ = 0;
+
+  SmallVectorStorage() noexcept = default;
+
+  ~SmallVectorStorage() { destroy(); }
+
+  constexpr storage_type* storage_ptr() { return data_; }
+
+  constexpr const storage_type* const_storage_ptr() const { return data_; }
+
+  void bump_size(size_t addend) {
+    const size_t new_size = size_ + addend;
+    ensure_capacity(new_size);
+    size_ = new_size;
+  }
+
+  void ensure_capacity(size_t min_capacity) {
+    if (dynamic_capacity_) {
+      // Grow dynamic storage if necessary
+      if (min_capacity > dynamic_capacity_) {
+        size_t new_capacity = std::max(dynamic_capacity_ * 2, min_capacity);
+        reallocate_dynamic(new_capacity);
+      }
+    } else if (min_capacity > N) {
+      switch_to_dynamic(min_capacity);
+    }
+  }
+
+  void reduce_size(size_t reduce_by) {
+    assert(reduce_by <= size_);
+    size_ -= reduce_by;
+  }
+
+  void destroy() noexcept {
+    storage_type::destroy_several(data_, size_);
+    if (dynamic_capacity_) {
+      delete[] data_;
+    }
+  }
+
+  void move_construct(SmallVectorStorage&& other) noexcept {
+    size_ = other.size_;
+    dynamic_capacity_ = other.dynamic_capacity_;
+    if (dynamic_capacity_) {
+      data_ = other.data_;
+      other.data_ = other.static_data_;
+      other.dynamic_capacity_ = 0;
+      other.size_ = 0;
+    } else if (size_ != 0) {
+      // Use a compile-time memcpy size (N) for trivial types
+      storage_type::move_construct_several(other.static_data_, static_data_, size_, N);
+    }
+  }
+
+  constexpr size_t capacity() const { return dynamic_capacity_ ? dynamic_capacity_ : N; }
+
+  constexpr size_t max_size() const { return std::numeric_limits<size_t>::max(); }
+
+  void reserve(size_t n) {
+    if (dynamic_capacity_) {
+      if (n > dynamic_capacity_) {
+        reallocate_dynamic(n);
+      }
+    } else if (n > N) {
+      switch_to_dynamic(n);
+    }
+  }
+
+  void clear() {
+    storage_type::destroy_several(data_, size_);
+    size_ = 0;
+  }
+
+ private:
+  void switch_to_dynamic(size_t new_capacity) {
+    dynamic_capacity_ = new_capacity;
+    data_ = new storage_type[new_capacity];
+    storage_type::move_construct_several_and_destroy_source(static_data_, data_, size_);
+  }
+
+  void reallocate_dynamic(size_t new_capacity) {
+    assert(new_capacity >= size_);
+    auto new_data = new storage_type[new_capacity];
+    storage_type::move_construct_several_and_destroy_source(data_, new_data, size_);
+    delete[] data_;
+    dynamic_capacity_ = new_capacity;
+    data_ = new_data;
+  }
+};
+
+template <typename T, size_t N, typename Storage>
+class StaticVectorImpl {
+ private:
+  Storage storage_;
+
+  T* data_ptr() { return storage_.storage_ptr()->get(); }
+
+  constexpr const T* const_data_ptr() const {
+    return storage_.const_storage_ptr()->get();
+  }
+
+ public:
+  using size_type = size_t;
+  using difference_type = ptrdiff_t;
+  using value_type = T;
+  using pointer = T*;
+  using const_pointer = const T*;
+  using reference = T&;
+  using const_reference = const T&;
+  using iterator = T*;
+  using const_iterator = const T*;
+  using reverse_iterator = std::reverse_iterator<iterator>;
+  using const_reverse_iterator = std::reverse_iterator<const_iterator>;
+
+  constexpr StaticVectorImpl() noexcept = default;
+
+  // Move and copy constructors
+  StaticVectorImpl(StaticVectorImpl&& other) noexcept {
+    storage_.move_construct(std::move(other.storage_));
+  }
+
+  StaticVectorImpl& operator=(StaticVectorImpl&& other) noexcept {
+    if (ARROW_PREDICT_TRUE(&other != this)) {
+      // TODO move_assign?
+      storage_.destroy();
+      storage_.move_construct(std::move(other.storage_));
+    }
+    return *this;
+  }
+
+  StaticVectorImpl(const StaticVectorImpl& other) {
+    init_by_copying(other.storage_.size_, other.const_data_ptr());
+  }
+
+  StaticVectorImpl& operator=(const StaticVectorImpl& other) noexcept {
+    if (ARROW_PREDICT_TRUE(&other != this)) {
+      assign_by_copying(other.storage_.size_, other.data());
+    }
+    return *this;
+  }
+
+  // Automatic conversion from std::vector<T>, for convenience
+  StaticVectorImpl(const std::vector<T>& other) {  // NOLINT: explicit
+    init_by_copying(other.size(), other.data());
+  }
+
+  StaticVectorImpl(std::vector<T>&& other) noexcept {  // NOLINT: explicit
+    init_by_moving(other.size(), other.data());
+  }
+
+  StaticVectorImpl& operator=(const std::vector<T>& other) {
+    assign_by_copying(other.size(), other.data());
+    return *this;
+  }
+
+  StaticVectorImpl& operator=(std::vector<T>&& other) noexcept {
+    assign_by_moving(other.size(), other.data());
+    return *this;
+  }
+
+  // Constructing from count and optional initialization value
+  explicit StaticVectorImpl(size_t count) {
+    storage_.bump_size(count);
+    auto* p = storage_.storage_ptr();
+    for (size_t i = 0; i < count; ++i) {
+      p[i].construct();
+    }
+  }
+
+  StaticVectorImpl(size_t count, const T& value) {
+    storage_.bump_size(count);
+    auto* p = storage_.storage_ptr();
+    for (size_t i = 0; i < count; ++i) {
+      p[i].construct(value);
+    }
+  }
+
+  StaticVectorImpl(std::initializer_list<T> values) {
+    storage_.bump_size(values.size());
+    auto* p = storage_.storage_ptr();
+    for (auto&& v : values) {
+      // Unfortunately, cannot move initializer values
+      p++->construct(v);
+    }
+  }
+
+  // Size inspection
+
+  constexpr bool empty() const { return storage_.size_ == 0; }
+
+  constexpr size_t size() const { return storage_.size_; }
+
+  constexpr size_t capacity() const { return storage_.capacity(); }
+
+  constexpr size_t max_size() const { return storage_.max_size(); }
+
+  // Data access
+
+  T& operator[](size_t i) { return data_ptr()[i]; }
+
+  constexpr const T& operator[](size_t i) const { return const_data_ptr()[i]; }
+
+  T& front() { return data_ptr()[0]; }
+
+  constexpr const T& front() const { return const_data_ptr()[0]; }
+
+  T& back() { return data_ptr()[storage_.size_ - 1]; }
+
+  constexpr const T& back() const { return const_data_ptr()[storage_.size_ - 1]; }
+
+  T* data() { return data_ptr(); }
+
+  constexpr const T* data() const { return const_data_ptr(); }
+
+  // Iterators
+
+  iterator begin() { return iterator(data_ptr()); }
+
+  constexpr const_iterator begin() const { return const_iterator(const_data_ptr()); }
+
+  constexpr const_iterator cbegin() const { return const_iterator(const_data_ptr()); }
+
+  iterator end() { return iterator(data_ptr() + storage_.size_); }
+
+  constexpr const_iterator end() const {
+    return const_iterator(const_data_ptr() + storage_.size_);
+  }
+
+  constexpr const_iterator cend() const {
+    return const_iterator(const_data_ptr() + storage_.size_);
+  }
+
+  reverse_iterator rbegin() { return reverse_iterator(end()); }
+
+  constexpr const_reverse_iterator rbegin() const {
+    return const_reverse_iterator(end());
+  }
+
+  constexpr const_reverse_iterator crbegin() const {
+    return const_reverse_iterator(end());
+  }
+
+  reverse_iterator rend() { return reverse_iterator(begin()); }
+
+  constexpr const_reverse_iterator rend() const {
+    return const_reverse_iterator(begin());
+  }
+
+  constexpr const_reverse_iterator crend() const {
+    return const_reverse_iterator(begin());
+  }
+
+  // Mutations
+
+  void reserve(size_t n) { storage_.reserve(n); }
+
+  void clear() { storage_.clear(); }
+
+  void push_back(const T& value) {
+    storage_.bump_size(1);
+    storage_.storage_ptr()[storage_.size_ - 1].construct(value);
+  }
+
+  void push_back(T&& value) {
+    storage_.bump_size(1);
+    storage_.storage_ptr()[storage_.size_ - 1].construct(std::move(value));
+  }
+
+  template <typename... Args>
+  void emplace_back(Args&&... args) {
+    storage_.bump_size(1);
+    storage_.storage_ptr()[storage_.size_ - 1].construct(std::forward<Args>(args)...);
+  }
+
+  template <typename InputIt>
+  iterator insert(const_iterator insert_at, InputIt first, InputIt last) {
+    const size_t n = storage_.size_;
+    const size_t it_size = static_cast<size_t>(last - first);  // XXX might be O(n)?
+    const size_t pos = static_cast<size_t>(insert_at - const_data_ptr());
+    storage_.bump_size(it_size);
+    auto* p = storage_.storage_ptr();
+    if (it_size == 0) {
+      return p[pos].get();
+    }
+    const size_t end_pos = pos + it_size;
+
+    // Move [pos; n) to [end_pos; end_pos + n - pos)
+    size_t i = n;
+    size_t j = end_pos + n - pos;
+    while (j > std::max(n, end_pos)) {
+      p[--j].move_construct(&p[--i]);
+    }
+    while (j > end_pos) {
+      p[--j].move_assign(&p[--i]);
+    }
+    assert(j == end_pos);
+    // Copy [first; last) to [pos; end_pos)
+    j = pos;
+    while (j < std::min(n, end_pos)) {
+      p[j++].assign(*first++);
+    }
+    while (j < end_pos) {
+      p[j++].construct(*first++);
+    }
+    assert(first == last);
+    return p[pos].get();
+  }
+
+  void resize(size_t n) {
+    const size_t old_size = storage_.size_;
+    if (n > storage_.size_) {
+      storage_.bump_size(n - old_size);
+      auto* p = storage_.storage_ptr();
+      for (size_t i = old_size; i < n; ++i) {
+        p[i].construct(T{});
+      }
+    } else {
+      auto* p = storage_.storage_ptr();
+      for (size_t i = n; i < old_size; ++i) {
+        p[i].destroy();
+      }
+      storage_.reduce_size(old_size - n);
+    }
+  }
+
+  void resize(size_t n, const T& value) {
+    const size_t old_size = storage_.size_;
+    if (n > storage_.size_) {
+      storage_.bump_size(n - old_size);
+      auto* p = storage_.storage_ptr();
+      for (size_t i = old_size; i < n; ++i) {
+        p[i].construct(value);
+      }
+    } else {
+      auto* p = storage_.storage_ptr();
+      for (size_t i = n; i < old_size; ++i) {
+        p[i].destroy();
+      }
+      storage_.reduce_size(old_size - n);
+    }
+  }
+
+ private:
+  template <typename InputIt>
+  void init_by_copying(size_t n, InputIt src) {
+    storage_.bump_size(n);
+    auto* dest = storage_.storage_ptr();
+    for (size_t i = 0; i < n; ++i, ++src) {
+      dest[i].construct(*src);
+    }
+  }
+
+  template <typename InputIt>
+  void init_by_moving(size_t n, InputIt src) {
+    init_by_copying(n, std::make_move_iterator(src));
+  }
+
+  template <typename InputIt>
+  void assign_by_copying(size_t n, InputIt src) {
+    const size_t old_size = storage_.size_;
+    if (n > old_size) {
+      storage_.bump_size(n - old_size);
+      auto* dest = storage_.storage_ptr();
+      for (size_t i = 0; i < old_size; ++i, ++src) {
+        dest[i].assign(*src);
+      }
+      for (size_t i = old_size; i < n; ++i, ++src) {
+        dest[i].construct(*src);
+      }
+    } else {
+      auto* dest = storage_.storage_ptr();
+      for (size_t i = 0; i < n; ++i, ++src) {
+        dest[i].assign(*src);
+      }
+      for (size_t i = n; i < old_size; ++i) {
+        dest[i].destroy();
+      }
+      storage_.reduce_size(old_size - n);
+    }
+  }
+
+  template <typename InputIt>
+  void assign_by_moving(size_t n, InputIt src) {
+    assign_by_copying(n, std::make_move_iterator(src));
+  }
+};
+
+template <typename T, size_t N>
+using StaticVector = StaticVectorImpl<T, N, StaticVectorStorage<T, N>>;
+
+template <typename T, size_t N>
+using SmallVector = StaticVectorImpl<T, N, SmallVectorStorage<T, N>>;
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/span.h

@@ -0,0 +1,132 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
+#include <cstring>
+#include <iterator>
+#include <type_traits>
+
+namespace arrow::util {
+
+template <class T>
+class span;
+
+/// std::span polyfill.
+///
+/// Does not support static extents.
+template <typename T>
+class span {
+  static_assert(sizeof(T),
+                R"(
+std::span allows contiguous_iterators instead of just pointers, the enforcement
+of which requires T to be a complete type. arrow::util::span does not support
+contiguous_iterators, but T is still required to be a complete type to prevent
+writing code which would break when it is replaced by std::span.)");
+
+ public:
+  using element_type = T;
+  using value_type = std::remove_cv_t<T>;
+  using iterator = T*;
+  using const_iterator = T const*;
+
+  span() = default;
+  span(const span&) = default;
+  span& operator=(const span&) = default;
+
+  template <typename M, typename = std::enable_if_t<std::is_same_v<T, M const>>>
+  // NOLINTNEXTLINE runtime/explicit
+  constexpr span(span<M> mut) : span{mut.data(), mut.size()} {}
+
+  constexpr span(T* data, size_t count) : data_{data}, size_{count} {}
+
+  constexpr span(T* begin, T* end)
+      : data_{begin}, size_{static_cast<size_t>(end - begin)} {}
+
+  template <typename R, typename RD = decltype(std::data(std::declval<R>())),
+            typename RS = decltype(std::size(std::declval<R>())),
+            typename E = std::enable_if_t<std::is_constructible_v<T*, RD> &&
+                                          std::is_constructible_v<size_t, RS>>>
+  // NOLINTNEXTLINE runtime/explicit, non-const reference
+  constexpr span(R&& range) : data_{std::data(range)}, size_{std::size(range)} {}
+
+  constexpr T* begin() const { return data_; }
+  constexpr T* end() const { return data_ + size_; }
+  constexpr T* data() const { return data_; }
+
+  constexpr size_t size() const { return size_; }
+  constexpr size_t size_bytes() const { return size_ * sizeof(T); }
+  constexpr bool empty() const { return size_ == 0; }
+
+  constexpr T& operator[](size_t i) { return data_[i]; }
+  constexpr const T& operator[](size_t i) const { return data_[i]; }
+
+  constexpr span subspan(size_t offset) const {
+    if (offset > size_) return {data_, data_};
+    return {data_ + offset, size_ - offset};
+  }
+
+  constexpr span subspan(size_t offset, size_t count) const {
+    auto out = subspan(offset);
+    if (count < out.size_) {
+      out.size_ = count;
+    }
+    return out;
+  }
+
+  constexpr bool operator==(const span& other) const {
+    if (size_ != other.size_) return false;
+
+    if constexpr (std::is_integral_v<T>) {
+      if (size_ == 0) {
+        return true;  // memcmp does not handle null pointers, even if size_ == 0
+      }
+      return std::memcmp(data_, other.data_, size_bytes()) == 0;
+    } else {
+      T* ptr = data_;
+      for (T const& e : other) {
+        if (*ptr++ != e) return false;
+      }
+      return true;
+    }
+  }
+  constexpr bool operator!=(const span& other) const { return !(*this == other); }
+
+ private:
+  T* data_{};
+  size_t size_{};
+};
+
+template <typename R>
+span(R& range) -> span<std::remove_pointer_t<decltype(std::data(range))>>;
+
+template <typename T>
+span(T*, size_t) -> span<T>;
+
+template <typename T>
+constexpr span<std::byte const> as_bytes(span<T> s) {
+  return {reinterpret_cast<const std::byte*>(s.data()), s.size_bytes()};
+}
+
+template <typename T>
+constexpr span<std::byte> as_writable_bytes(span<T> s) {
+  return {reinterpret_cast<std::byte*>(s.data()), s.size_bytes()};
+}
+
+}  // namespace arrow::util

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/string.h

@@ -0,0 +1,173 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cassert>
+#include <optional>
+#include <string>
+#include <string_view>
+#include <type_traits>
+#include <utility>
+#include <vector>
+
+#if __has_include(<charconv>)
+#  include <charconv>
+#endif
+
+#include "arrow/result.h"
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+
+class Status;
+
+ARROW_EXPORT std::string HexEncode(const uint8_t* data, size_t length);
+
+ARROW_EXPORT std::string Escape(const char* data, size_t length);
+
+ARROW_EXPORT std::string HexEncode(const char* data, size_t length);
+
+ARROW_EXPORT std::string HexEncode(std::string_view str);
+
+ARROW_EXPORT std::string Escape(std::string_view str);
+
+ARROW_EXPORT Status ParseHexValue(const char* hex_pair, uint8_t* out);
+
+ARROW_EXPORT Status ParseHexValues(std::string_view hex_string, uint8_t* out);
+
+namespace internal {
+
+/// Like std::string_view::starts_with in C++20
+inline bool StartsWith(std::string_view s, std::string_view prefix) {
+  return s.length() >= prefix.length() &&
+         (s.empty() || s.substr(0, prefix.length()) == prefix);
+}
+
+/// Like std::string_view::ends_with in C++20
+inline bool EndsWith(std::string_view s, std::string_view suffix) {
+  return s.length() >= suffix.length() &&
+         (s.empty() || s.substr(s.length() - suffix.length()) == suffix);
+}
+
+/// \brief Split a string with a delimiter
+ARROW_EXPORT
+std::vector<std::string_view> SplitString(std::string_view v, char delim,
+                                          int64_t limit = 0);
+
+/// \brief Join strings with a delimiter
+ARROW_EXPORT
+std::string JoinStrings(const std::vector<std::string_view>& strings,
+                        std::string_view delimiter);
+
+/// \brief Join strings with a delimiter
+ARROW_EXPORT
+std::string JoinStrings(const std::vector<std::string>& strings,
+                        std::string_view delimiter);
+
+/// \brief Trim whitespace from left and right sides of string
+ARROW_EXPORT
+std::string TrimString(std::string value);
+
+ARROW_EXPORT
+bool AsciiEqualsCaseInsensitive(std::string_view left, std::string_view right);
+
+ARROW_EXPORT
+std::string AsciiToLower(std::string_view value);
+
+ARROW_EXPORT
+std::string AsciiToUpper(std::string_view value);
+
+/// \brief Search for the first instance of a token and replace it or return nullopt if
+/// the token is not found.
+ARROW_EXPORT
+std::optional<std::string> Replace(std::string_view s, std::string_view token,
+                                   std::string_view replacement);
+
+/// \brief Get boolean value from string
+///
+/// If "1", "true" (case-insensitive), returns true
+/// If "0", "false" (case-insensitive), returns false
+/// Otherwise, returns Status::Invalid
+ARROW_EXPORT
+arrow::Result<bool> ParseBoolean(std::string_view value);
+
+#if __has_include(<charconv>)
+
+namespace detail {
+template <typename T, typename = void>
+struct can_to_chars : public std::false_type {};
+
+template <typename T>
+struct can_to_chars<
+    T, std::void_t<decltype(std::to_chars(std::declval<char*>(), std::declval<char*>(),
+                                          std::declval<std::remove_reference_t<T>>()))>>
+    : public std::true_type {};
+}  // namespace detail
+
+/// \brief Whether std::to_chars exists for the current value type.
+///
+/// This is useful as some C++ libraries do not implement all specified overloads
+/// for std::to_chars.
+template <typename T>
+inline constexpr bool have_to_chars = detail::can_to_chars<T>::value;
+
+/// \brief An ergonomic wrapper around std::to_chars, returning a std::string
+///
+/// For most inputs, the std::string result will not incur any heap allocation
+/// thanks to small string optimization.
+///
+/// Compared to std::to_string, this function gives locale-agnostic results
+/// and might also be faster.
+template <typename T, typename... Args>
+std::string ToChars(T value, Args&&... args) {
+  if constexpr (!have_to_chars<T>) {
+    // Some C++ standard libraries do not yet implement std::to_chars for all types,
+    // in which case we have to fallback to std::string.
+    return std::to_string(value);
+  } else {
+    // According to various sources, the GNU libstdc++ and Microsoft's C++ STL
+    // allow up to 15 bytes of small string optimization, while clang's libc++
+    // goes up to 22 bytes. Choose the pessimistic value.
+    std::string out(15, 0);
+    auto res = std::to_chars(&out.front(), &out.back(), value, args...);
+    while (res.ec != std::errc{}) {
+      assert(res.ec == std::errc::value_too_large);
+      out.resize(out.capacity() * 2);
+      res = std::to_chars(&out.front(), &out.back(), value, args...);
+    }
+    const auto length = res.ptr - out.data();
+    assert(length <= static_cast<int64_t>(out.length()));
+    out.resize(length);
+    return out;
+  }
+}
+
+#else  // !__has_include(<charconv>)
+
+template <typename T>
+inline constexpr bool have_to_chars = false;
+
+template <typename T, typename... Args>
+std::string ToChars(T value, Args&&... args) {
+  return std::to_string(value);
+}
+
+#endif
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/string_util.h

@@ -0,0 +1,82 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License. template <typename T>
+
+#pragma once
+
+#include <memory>
+#include <ostream>
+#include <string>
+#include <type_traits>
+#include <utility>
+
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+
+namespace internal {
+
+class ARROW_EXPORT StringStreamWrapper {
+ public:
+  StringStreamWrapper();
+  ~StringStreamWrapper();
+
+  std::ostream& stream() { return ostream_; }
+  std::string str();
+
+ protected:
+  std::unique_ptr<std::ostringstream> sstream_;
+  std::ostream& ostream_;
+};
+
+template <typename... Args>
+std::string JoinToString(Args&&... args) {
+  StringStreamWrapper ss;
+  (
+      [&ss](auto&& arg) {
+        // Avoid losing precision when printing floating point numbers
+        if constexpr (std::is_floating_point_v<std::decay_t<decltype(arg)>>) {
+          ss.stream() << std::to_string(arg);
+        } else {
+          ss.stream() << arg;
+        }
+      }(std::forward<Args>(args)),
+      ...);
+  return ss.str();
+}
+}  // namespace internal
+
+namespace util {
+/// CRTP helper for declaring string representation. Defines operator<<
+template <typename T>
+class ToStringOstreamable {
+ public:
+  ~ToStringOstreamable() {
+    static_assert(
+        std::is_same<decltype(std::declval<const T>().ToString()), std::string>::value,
+        "ToStringOstreamable depends on the method T::ToString() const");
+  }
+
+ private:
+  const T& cast() const { return static_cast<const T&>(*this); }
+
+  friend inline std::ostream& operator<<(std::ostream& os, const ToStringOstreamable& t) {
+    return os << t.cast().ToString();
+  }
+};
+
+}  // namespace util
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/task_group.h

@@ -0,0 +1,106 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <memory>
+#include <utility>
+
+#include "arrow/status.h"
+#include "arrow/type_fwd.h"
+#include "arrow/util/cancel.h"
+#include "arrow/util/functional.h"
+#include "arrow/util/macros.h"
+#include "arrow/util/type_fwd.h"
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+namespace internal {
+
+/// \brief A group of related tasks
+///
+/// A TaskGroup executes tasks with the signature `Status()`.
+/// Execution can be serial or parallel, depending on the TaskGroup
+/// implementation.  When Finish() returns, it is guaranteed that all
+/// tasks have finished, or at least one has errored.
+///
+/// Once an error has occurred any tasks that are submitted to the task group
+/// will not run.  The call to Append will simply return without scheduling the
+/// task.
+///
+/// If the task group is parallel it is possible that multiple tasks could be
+/// running at the same time and one of those tasks fails.  This will put the
+/// task group in a failure state (so additional tasks cannot be run) however
+/// it will not interrupt running tasks.  Finish will not complete
+/// until all running tasks have finished, even if one task fails.
+///
+/// Once a task group has finished new tasks may not be added to it.  If you need to start
+/// a new batch of work then you should create a new task group.
+class ARROW_EXPORT TaskGroup : public std::enable_shared_from_this<TaskGroup> {
+ public:
+  /// Add a Status-returning function to execute.  Execution order is
+  /// undefined.  The function may be executed immediately or later.
+  template <typename Function>
+  void Append(Function&& func) {
+    return AppendReal(std::forward<Function>(func));
+  }
+
+  /// Wait for execution of all tasks (and subgroups) to be finished,
+  /// or for at least one task (or subgroup) to error out.
+  /// The returned Status propagates the error status of the first failing
+  /// task (or subgroup).
+  virtual Status Finish() = 0;
+
+  /// Returns a future that will complete the first time all tasks are finished.
+  /// This should be called only after all top level tasks
+  /// have been added to the task group.
+  ///
+  /// If you are using a TaskGroup asynchronously there are a few considerations to keep
+  /// in mind.  The tasks should not block on I/O, etc (defeats the purpose of using
+  /// futures) and should not be doing any nested locking or you run the risk of the tasks
+  /// getting stuck in the thread pool waiting for tasks which cannot get scheduled.
+  ///
+  /// Primarily this call is intended to help migrate existing work written with TaskGroup
+  /// in mind to using futures without having to do a complete conversion on the first
+  /// pass.
+  virtual Future<> FinishAsync() = 0;
+
+  /// The current aggregate error Status.  Non-blocking, useful for stopping early.
+  virtual Status current_status() = 0;
+
+  /// Whether some tasks have already failed.  Non-blocking, useful for stopping early.
+  virtual bool ok() const = 0;
+
+  /// How many tasks can typically be executed in parallel.
+  /// This is only a hint, useful for testing or debugging.
+  virtual int parallelism() = 0;
+
+  static std::shared_ptr<TaskGroup> MakeSerial(StopToken = StopToken::Unstoppable());
+  static std::shared_ptr<TaskGroup> MakeThreaded(internal::Executor*,
+                                                 StopToken = StopToken::Unstoppable());
+
+  virtual ~TaskGroup() = default;
+
+ protected:
+  TaskGroup() = default;
+  ARROW_DISALLOW_COPY_AND_ASSIGN(TaskGroup);
+
+  virtual void AppendReal(FnOnce<Status()> task) = 0;
+};
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/test_common.h

@@ -0,0 +1,90 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <iosfwd>
+
+#include "arrow/testing/gtest_util.h"
+#include "arrow/util/iterator.h"
+
+namespace arrow {
+
+struct TestInt {
+  TestInt();
+  TestInt(int i);  // NOLINT runtime/explicit
+  int value;
+
+  bool operator==(const TestInt& other) const;
+
+  friend std::ostream& operator<<(std::ostream& os, const TestInt& v);
+};
+
+template <>
+struct IterationTraits<TestInt> {
+  static TestInt End() { return TestInt(); }
+  static bool IsEnd(const TestInt& val) { return val == IterationTraits<TestInt>::End(); }
+};
+
+struct TestStr {
+  TestStr();
+  TestStr(const std::string& s);  // NOLINT runtime/explicit
+  TestStr(const char* s);         // NOLINT runtime/explicit
+  explicit TestStr(const TestInt& test_int);
+  std::string value;
+
+  bool operator==(const TestStr& other) const;
+
+  friend std::ostream& operator<<(std::ostream& os, const TestStr& v);
+};
+
+template <>
+struct IterationTraits<TestStr> {
+  static TestStr End() { return TestStr(); }
+  static bool IsEnd(const TestStr& val) { return val == IterationTraits<TestStr>::End(); }
+};
+
+std::vector<TestInt> RangeVector(unsigned int max, unsigned int step = 1);
+
+template <typename T>
+inline Iterator<T> VectorIt(std::vector<T> v) {
+  return MakeVectorIterator<T>(std::move(v));
+}
+
+template <typename T>
+inline Iterator<T> PossiblySlowVectorIt(std::vector<T> v, bool slow = false) {
+  auto iterator = MakeVectorIterator<T>(std::move(v));
+  if (slow) {
+    return MakeTransformedIterator<T, T>(std::move(iterator),
+                                         [](T item) -> Result<TransformFlow<T>> {
+                                           SleepABit();
+                                           return TransformYield(item);
+                                         });
+  } else {
+    return iterator;
+  }
+}
+
+template <typename T>
+inline void AssertIteratorExhausted(Iterator<T>& it) {
+  ASSERT_OK_AND_ASSIGN(T next, it.Next());
+  ASSERT_TRUE(IsIterationEnd(next));
+}
+
+Transformer<TestInt, TestStr> MakeFilter(std::function<bool(TestInt&)> filter);
+
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/thread_pool.h

@@ -0,0 +1,643 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cstdint>
+#include <memory>
+#include <queue>
+#include <type_traits>
+#include <unordered_set>
+#include <utility>
+
+#include "arrow/result.h"
+#include "arrow/status.h"
+#include "arrow/util/cancel.h"
+#include "arrow/util/config.h"
+#include "arrow/util/functional.h"
+#include "arrow/util/future.h"
+#include "arrow/util/iterator.h"
+#include "arrow/util/macros.h"
+#include "arrow/util/visibility.h"
+
+#if defined(_MSC_VER)
+// Disable harmless warning for decorated name length limit
+#  pragma warning(disable : 4503)
+#endif
+
+namespace arrow {
+
+/// \brief Get the capacity of the global thread pool
+///
+/// Return the number of worker threads in the thread pool to which
+/// Arrow dispatches various CPU-bound tasks.  This is an ideal number,
+/// not necessarily the exact number of threads at a given point in time.
+///
+/// You can change this number using SetCpuThreadPoolCapacity().
+ARROW_EXPORT int GetCpuThreadPoolCapacity();
+
+/// \brief Set the capacity of the global thread pool
+///
+/// Set the number of worker threads int the thread pool to which
+/// Arrow dispatches various CPU-bound tasks.
+///
+/// The current number is returned by GetCpuThreadPoolCapacity().
+ARROW_EXPORT Status SetCpuThreadPoolCapacity(int threads);
+
+namespace internal {
+
+// Hints about a task that may be used by an Executor.
+// They are ignored by the provided ThreadPool implementation.
+struct TaskHints {
+  // The lower, the more urgent
+  int32_t priority = 0;
+  // The IO transfer size in bytes
+  int64_t io_size = -1;
+  // The approximate CPU cost in number of instructions
+  int64_t cpu_cost = -1;
+  // An application-specific ID
+  int64_t external_id = -1;
+};
+
+class ARROW_EXPORT Executor {
+ public:
+  using StopCallback = internal::FnOnce<void(const Status&)>;
+
+  virtual ~Executor();
+
+  // Spawn a fire-and-forget task.
+  template <typename Function>
+  Status Spawn(Function&& func) {
+    return SpawnReal(TaskHints{}, std::forward<Function>(func), StopToken::Unstoppable(),
+                     StopCallback{});
+  }
+  template <typename Function>
+  Status Spawn(Function&& func, StopToken stop_token) {
+    return SpawnReal(TaskHints{}, std::forward<Function>(func), std::move(stop_token),
+                     StopCallback{});
+  }
+  template <typename Function>
+  Status Spawn(TaskHints hints, Function&& func) {
+    return SpawnReal(hints, std::forward<Function>(func), StopToken::Unstoppable(),
+                     StopCallback{});
+  }
+  template <typename Function>
+  Status Spawn(TaskHints hints, Function&& func, StopToken stop_token) {
+    return SpawnReal(hints, std::forward<Function>(func), std::move(stop_token),
+                     StopCallback{});
+  }
+  template <typename Function>
+  Status Spawn(TaskHints hints, Function&& func, StopToken stop_token,
+               StopCallback stop_callback) {
+    return SpawnReal(hints, std::forward<Function>(func), std::move(stop_token),
+                     std::move(stop_callback));
+  }
+
+  // Transfers a future to this executor.  Any continuations added to the
+  // returned future will run in this executor.  Otherwise they would run
+  // on the same thread that called MarkFinished.
+  //
+  // This is necessary when (for example) an I/O task is completing a future.
+  // The continuations of that future should run on the CPU thread pool keeping
+  // CPU heavy work off the I/O thread pool.  So the I/O task should transfer
+  // the future to the CPU executor before returning.
+  //
+  // By default this method will only transfer if the future is not already completed.  If
+  // the future is already completed then any callback would be run synchronously and so
+  // no transfer is typically necessary.  However, in cases where you want to force a
+  // transfer (e.g. to help the scheduler break up units of work across multiple cores)
+  // then you can override this behavior with `always_transfer`.
+  template <typename T>
+  Future<T> Transfer(Future<T> future) {
+    return DoTransfer(std::move(future), false);
+  }
+
+  // Overload of Transfer which will always schedule callbacks on new threads even if the
+  // future is finished when the callback is added.
+  //
+  // This can be useful in cases where you want to ensure parallelism
+  template <typename T>
+  Future<T> TransferAlways(Future<T> future) {
+    return DoTransfer(std::move(future), true);
+  }
+
+  // Submit a callable and arguments for execution.  Return a future that
+  // will return the callable's result value once.
+  // The callable's arguments are copied before execution.
+  template <typename Function, typename... Args,
+            typename FutureType = typename ::arrow::detail::ContinueFuture::ForSignature<
+                Function && (Args && ...)>>
+  Result<FutureType> Submit(TaskHints hints, StopToken stop_token, Function&& func,
+                            Args&&... args) {
+    using ValueType = typename FutureType::ValueType;
+
+    auto future = FutureType::Make();
+    auto task = std::bind(::arrow::detail::ContinueFuture{}, future,
+                          std::forward<Function>(func), std::forward<Args>(args)...);
+    struct {
+      WeakFuture<ValueType> weak_fut;
+
+      void operator()(const Status& st) {
+        auto fut = weak_fut.get();
+        if (fut.is_valid()) {
+          fut.MarkFinished(st);
+        }
+      }
+    } stop_callback{WeakFuture<ValueType>(future)};
+    ARROW_RETURN_NOT_OK(SpawnReal(hints, std::move(task), std::move(stop_token),
+                                  std::move(stop_callback)));
+
+    return future;
+  }
+
+  template <typename Function, typename... Args,
+            typename FutureType = typename ::arrow::detail::ContinueFuture::ForSignature<
+                Function && (Args && ...)>>
+  Result<FutureType> Submit(StopToken stop_token, Function&& func, Args&&... args) {
+    return Submit(TaskHints{}, stop_token, std::forward<Function>(func),
+                  std::forward<Args>(args)...);
+  }
+
+  template <typename Function, typename... Args,
+            typename FutureType = typename ::arrow::detail::ContinueFuture::ForSignature<
+                Function && (Args && ...)>>
+  Result<FutureType> Submit(TaskHints hints, Function&& func, Args&&... args) {
+    return Submit(std::move(hints), StopToken::Unstoppable(),
+                  std::forward<Function>(func), std::forward<Args>(args)...);
+  }
+
+  template <typename Function, typename... Args,
+            typename FutureType = typename ::arrow::detail::ContinueFuture::ForSignature<
+                Function && (Args && ...)>>
+  Result<FutureType> Submit(Function&& func, Args&&... args) {
+    return Submit(TaskHints{}, StopToken::Unstoppable(), std::forward<Function>(func),
+                  std::forward<Args>(args)...);
+  }
+
+  // Return the level of parallelism (the number of tasks that may be executed
+  // concurrently).  This may be an approximate number.
+  virtual int GetCapacity() = 0;
+
+  // Return true if the thread from which this function is called is owned by this
+  // Executor. Returns false if this Executor does not support this property.
+  virtual bool OwnsThisThread() { return false; }
+
+  // Return true if this is the current executor being called
+  // n.b. this defaults to just calling OwnsThisThread
+  // unless the threadpool is disabled
+  virtual bool IsCurrentExecutor() { return OwnsThisThread(); }
+
+  /// \brief An interface to represent something with a custom destructor
+  ///
+  /// \see KeepAlive
+  class ARROW_EXPORT Resource {
+   public:
+    virtual ~Resource() = default;
+  };
+
+  /// \brief Keep a resource alive until all executor threads have terminated
+  ///
+  /// Executors may have static storage duration.  In particular, the CPU and I/O
+  /// executors are currently implemented this way.  These threads may access other
+  /// objects with static storage duration such as the OpenTelemetry runtime context
+  /// the default memory pool, or other static executors.
+  ///
+  /// The order in which these objects are destroyed is difficult to control.  In order
+  /// to ensure those objects remain alive until all threads have finished those objects
+  /// should be wrapped in a Resource object and passed into this method.  The given
+  /// shared_ptr will be kept alive until all threads have finished their worker loops.
+  virtual void KeepAlive(std::shared_ptr<Resource> resource);
+
+ protected:
+  ARROW_DISALLOW_COPY_AND_ASSIGN(Executor);
+
+  Executor() = default;
+
+  template <typename T, typename FT = Future<T>, typename FTSync = typename FT::SyncType>
+  Future<T> DoTransfer(Future<T> future, bool always_transfer = false) {
+    auto transferred = Future<T>::Make();
+    if (always_transfer) {
+      CallbackOptions callback_options = CallbackOptions::Defaults();
+      callback_options.should_schedule = ShouldSchedule::Always;
+      callback_options.executor = this;
+      auto sync_callback = [transferred](const FTSync& result) mutable {
+        transferred.MarkFinished(result);
+      };
+      future.AddCallback(sync_callback, callback_options);
+      return transferred;
+    }
+
+    // We could use AddCallback's ShouldSchedule::IfUnfinished but we can save a bit of
+    // work by doing the test here.
+    auto callback = [this, transferred](const FTSync& result) mutable {
+      auto spawn_status =
+          Spawn([transferred, result]() mutable { transferred.MarkFinished(result); });
+      if (!spawn_status.ok()) {
+        transferred.MarkFinished(spawn_status);
+      }
+    };
+    auto callback_factory = [&callback]() { return callback; };
+    if (future.TryAddCallback(callback_factory)) {
+      return transferred;
+    }
+    // If the future is already finished and we aren't going to force spawn a thread
+    // then we don't need to add another layer of callback and can return the original
+    // future
+    return future;
+  }
+
+  // Subclassing API
+  virtual Status SpawnReal(TaskHints hints, FnOnce<void()> task, StopToken,
+                           StopCallback&&) = 0;
+};
+
+/// \brief An executor implementation that runs all tasks on a single thread using an
+/// event loop.
+///
+/// Note: Any sort of nested parallelism will deadlock this executor.  Blocking waits are
+/// fine but if one task needs to wait for another task it must be expressed as an
+/// asynchronous continuation.
+class ARROW_EXPORT SerialExecutor : public Executor {
+ public:
+  template <typename T = ::arrow::internal::Empty>
+  using TopLevelTask = internal::FnOnce<Future<T>(Executor*)>;
+
+  ~SerialExecutor() override;
+
+  int GetCapacity() override { return 1; };
+  bool OwnsThisThread() override;
+  Status SpawnReal(TaskHints hints, FnOnce<void()> task, StopToken,
+                   StopCallback&&) override;
+
+  // Return the number of tasks either running or in the queue.
+  int GetNumTasks();
+
+  /// \brief Runs the TopLevelTask and any scheduled tasks
+  ///
+  /// The TopLevelTask (or one of the tasks it schedules) must either return an invalid
+  /// status or call the finish signal. Failure to do this will result in a deadlock.  For
+  /// this reason it is preferable (if possible) to use the helper methods (below)
+  /// RunSynchronously/RunSerially which delegates the responsibility onto a Future
+  /// producer's existing responsibility to always mark a future finished (which can
+  /// someday be aided by ARROW-12207).
+  template <typename T = internal::Empty, typename FT = Future<T>,
+            typename FTSync = typename FT::SyncType>
+  static FTSync RunInSerialExecutor(TopLevelTask<T> initial_task) {
+    Future<T> fut = SerialExecutor().Run<T>(std::move(initial_task));
+    return FutureToSync(fut);
+  }
+
+  /// \brief Transform an AsyncGenerator into an Iterator
+  ///
+  /// An event loop will be created and each call to Next will power the event loop with
+  /// the calling thread until the next item is ready to be delivered.
+  ///
+  /// Note: The iterator's destructor will run until the given generator is fully
+  /// exhausted. If you wish to abandon iteration before completion then the correct
+  /// approach is to use a stop token to cause the generator to exhaust early.
+  template <typename T>
+  static Iterator<T> IterateGenerator(
+      internal::FnOnce<Result<std::function<Future<T>()>>(Executor*)> initial_task) {
+    auto serial_executor = std::unique_ptr<SerialExecutor>(new SerialExecutor());
+    auto maybe_generator = std::move(initial_task)(serial_executor.get());
+    if (!maybe_generator.ok()) {
+      return MakeErrorIterator<T>(maybe_generator.status());
+    }
+    auto generator = maybe_generator.MoveValueUnsafe();
+    struct SerialIterator {
+      SerialIterator(std::unique_ptr<SerialExecutor> executor,
+                     std::function<Future<T>()> generator)
+          : executor(std::move(executor)), generator(std::move(generator)) {}
+      ARROW_DISALLOW_COPY_AND_ASSIGN(SerialIterator);
+      ARROW_DEFAULT_MOVE_AND_ASSIGN(SerialIterator);
+      ~SerialIterator() {
+        // A serial iterator must be consumed before it can be destroyed.  Allowing it to
+        // do otherwise would lead to resource leakage.  There will likely be deadlocks at
+        // this spot in the future but these will be the result of other bugs and not the
+        // fact that we are forcing consumption here.
+
+        // If a streaming API needs to support early abandonment then it should be done so
+        // with a cancellation token and not simply discarding the iterator and expecting
+        // the underlying work to clean up correctly.
+        if (executor && !executor->IsFinished()) {
+          while (true) {
+            Result<T> maybe_next = Next();
+            if (!maybe_next.ok() || IsIterationEnd(*maybe_next)) {
+              break;
+            }
+          }
+        }
+      }
+
+      Result<T> Next() {
+        executor->Unpause();
+        // This call may lead to tasks being scheduled in the serial executor
+        Future<T> next_fut = generator();
+        next_fut.AddCallback([this](const Result<T>& res) {
+          // If we're done iterating we should drain the rest of the tasks in the executor
+          if (!res.ok() || IsIterationEnd(*res)) {
+            executor->Finish();
+            return;
+          }
+          // Otherwise we will break out immediately, leaving the remaining tasks for
+          // the next call.
+          executor->Pause();
+        });
+#ifdef ARROW_ENABLE_THREADING
+        // future must run on this thread
+        // Borrow this thread and run tasks until the future is finished
+        executor->RunLoop();
+#else
+        next_fut.Wait();
+#endif
+        if (!next_fut.is_finished()) {
+          // Not clear this is possible since RunLoop wouldn't generally exit
+          // unless we paused/finished which would imply next_fut has been
+          // finished.
+          return Status::Invalid(
+              "Serial executor terminated before next result computed");
+        }
+        // At this point we may still have tasks in the executor, that is ok.
+        // We will run those tasks the next time through.
+        return next_fut.result();
+      }
+
+      std::unique_ptr<SerialExecutor> executor;
+      std::function<Future<T>()> generator;
+    };
+    return Iterator<T>(SerialIterator{std::move(serial_executor), std::move(generator)});
+  }
+
+#ifndef ARROW_ENABLE_THREADING
+  // run a pending task from loop
+  // returns true if any tasks were run in the last go round the loop (i.e. if it
+  // returns false, all executors are waiting)
+  static bool RunTasksOnAllExecutors();
+  static SerialExecutor* GetCurrentExecutor();
+
+  bool IsCurrentExecutor() override;
+
+#endif
+
+ protected:
+  virtual void RunLoop();
+
+  // State uses mutex
+  struct State;
+  std::shared_ptr<State> state_;
+
+  SerialExecutor();
+
+  // We mark the serial executor "finished" when there should be
+  // no more tasks scheduled on it.  It's not strictly needed but
+  // can help catch bugs where we are trying to use the executor
+  // after we are done with it.
+  void Finish();
+  bool IsFinished();
+  // We pause the executor when we are running an async generator
+  // and we have received an item that we can deliver.
+  void Pause();
+  void Unpause();
+
+  template <typename T, typename FTSync = typename Future<T>::SyncType>
+  Future<T> Run(TopLevelTask<T> initial_task) {
+    auto final_fut = std::move(initial_task)(this);
+    final_fut.AddCallback([this](const FTSync&) { Finish(); });
+    RunLoop();
+    return final_fut;
+  }
+
+#ifndef ARROW_ENABLE_THREADING
+  // we have to run tasks from all live executors
+  // during RunLoop if we don't have threading
+  static std::unordered_set<SerialExecutor*> all_executors;
+  // a pointer to the last one called by the loop
+  // so all tasks get spawned equally
+  // on multiple calls to RunTasksOnAllExecutors
+  static SerialExecutor* last_called_executor;
+  // without threading we can't tell which executor called the
+  // current process - so we set it in spawning the task
+  static SerialExecutor* current_executor;
+#endif  // ARROW_ENABLE_THREADING
+};
+
+#ifdef ARROW_ENABLE_THREADING
+
+/// An Executor implementation spawning tasks in FIFO manner on a fixed-size
+/// pool of worker threads.
+///
+/// Note: Any sort of nested parallelism will deadlock this executor.  Blocking waits are
+/// fine but if one task needs to wait for another task it must be expressed as an
+/// asynchronous continuation.
+class ARROW_EXPORT ThreadPool : public Executor {
+ public:
+  // Construct a thread pool with the given number of worker threads
+  static Result<std::shared_ptr<ThreadPool>> Make(int threads);
+
+  // Like Make(), but takes care that the returned ThreadPool is compatible
+  // with destruction late at process exit.
+  static Result<std::shared_ptr<ThreadPool>> MakeEternal(int threads);
+
+  // Destroy thread pool; the pool will first be shut down
+  ~ThreadPool() override;
+
+  // Return the desired number of worker threads.
+  // The actual number of workers may lag a bit before being adjusted to
+  // match this value.
+  int GetCapacity() override;
+
+  // Return the number of tasks either running or in the queue.
+  int GetNumTasks();
+
+  bool OwnsThisThread() override;
+  // Dynamically change the number of worker threads.
+  //
+  // This function always returns immediately.
+  // If fewer threads are running than this number, new threads are spawned
+  // on-demand when needed for task execution.
+  // If more threads are running than this number, excess threads are reaped
+  // as soon as possible.
+  Status SetCapacity(int threads);
+
+  // Heuristic for the default capacity of a thread pool for CPU-bound tasks.
+  // This is exposed as a static method to help with testing.
+  // The number returned is guaranteed to be greater or equal to one.
+  static int DefaultCapacity();
+
+  // Shutdown the pool.  Once the pool starts shutting down, new tasks
+  // cannot be submitted anymore.
+  // If "wait" is true, shutdown waits for all pending tasks to be finished.
+  // If "wait" is false, workers are stopped as soon as currently executing
+  // tasks are finished.
+  Status Shutdown(bool wait = true);
+
+  // Wait for the thread pool to become idle
+  //
+  // This is useful for sequencing tests
+  void WaitForIdle();
+
+  void KeepAlive(std::shared_ptr<Executor::Resource> resource) override;
+
+  struct State;
+
+ protected:
+  FRIEND_TEST(TestThreadPool, SetCapacity);
+  FRIEND_TEST(TestGlobalThreadPool, Capacity);
+  ARROW_FRIEND_EXPORT friend ThreadPool* GetCpuThreadPool();
+
+  ThreadPool();
+
+  Status SpawnReal(TaskHints hints, FnOnce<void()> task, StopToken,
+                   StopCallback&&) override;
+
+  // Collect finished worker threads, making sure the OS threads have exited
+  void CollectFinishedWorkersUnlocked();
+  // Launch a given number of additional workers
+  void LaunchWorkersUnlocked(int threads);
+  // Get the current actual capacity
+  int GetActualCapacity();
+
+  static std::shared_ptr<ThreadPool> MakeCpuThreadPool();
+
+  std::shared_ptr<State> sp_state_;
+  State* state_;
+  bool shutdown_on_destroy_;
+};
+#else  // ARROW_ENABLE_THREADING
+// an executor implementation which pretends to be a thread pool but runs everything
+// on the main thread using a static queue (shared between all thread pools, otherwise
+// cross-threadpool dependencies will break everything)
+class ARROW_EXPORT ThreadPool : public SerialExecutor {
+ public:
+  ARROW_FRIEND_EXPORT friend ThreadPool* GetCpuThreadPool();
+
+  static Result<std::shared_ptr<ThreadPool>> Make(int threads);
+
+  // Like Make(), but takes care that the returned ThreadPool is compatible
+  // with destruction late at process exit.
+  static Result<std::shared_ptr<ThreadPool>> MakeEternal(int threads);
+
+  // Destroy thread pool; the pool will first be shut down
+  ~ThreadPool() override;
+
+  // Return the desired number of worker threads.
+  // The actual number of workers may lag a bit before being adjusted to
+  // match this value.
+  int GetCapacity() override;
+
+  virtual int GetActualCapacity();
+
+  bool OwnsThisThread() override { return true; }
+
+  // Dynamically change the number of worker threads.
+  // without threading this is equal to the
+  // number of tasks that can be running at once
+  // (inside each other)
+  Status SetCapacity(int threads);
+
+  static int DefaultCapacity() { return 8; }
+
+  // Shutdown the pool.  Once the pool starts shutting down, new tasks
+  // cannot be submitted anymore.
+  // If "wait" is true, shutdown waits for all pending tasks to be finished.
+  // If "wait" is false, workers are stopped as soon as currently executing
+  // tasks are finished.
+  Status Shutdown(bool wait = true);
+
+  // Wait for the thread pool to become idle
+  //
+  // This is useful for sequencing tests
+  void WaitForIdle();
+
+ protected:
+  static std::shared_ptr<ThreadPool> MakeCpuThreadPool();
+  ThreadPool();
+};
+
+#endif  // ARROW_ENABLE_THREADING
+
+// Return the process-global thread pool for CPU-bound tasks.
+ARROW_EXPORT ThreadPool* GetCpuThreadPool();
+
+/// \brief Potentially run an async operation serially (if use_threads is false)
+/// \see RunSerially
+///
+/// If `use_threads` is true, the global CPU executor is used.
+/// If `use_threads` is false, a temporary SerialExecutor is used.
+/// `get_future` is called (from this thread) with the chosen executor and must
+/// return a future that will eventually finish. This function returns once the
+/// future has finished.
+template <typename Fut, typename ValueType = typename Fut::ValueType>
+typename Fut::SyncType RunSynchronously(FnOnce<Fut(Executor*)> get_future,
+                                        bool use_threads) {
+  if (use_threads) {
+    auto fut = std::move(get_future)(GetCpuThreadPool());
+    return FutureToSync(fut);
+  } else {
+    return SerialExecutor::RunInSerialExecutor<ValueType>(std::move(get_future));
+  }
+}
+
+/// \brief Potentially iterate an async generator serially (if use_threads is false)
+///   using a potentially custom Executor
+/// \see IterateGenerator
+///
+/// If `use_threads` is true, the custom executor or, if null,
+///   the global CPU executor will be used.  Each call to
+///   the iterator will simply wait until the next item is available.  Tasks may run in
+///   the background between calls.
+///
+/// If `use_threads` is false, the calling thread only will be used.  Each call to
+///   the iterator will use the calling thread to do enough work to generate one item.
+///   Tasks will be left in a queue until the next call and no work will be done between
+///   calls.
+template <typename T>
+Iterator<T> IterateSynchronously(
+    FnOnce<Result<std::function<Future<T>()>>(Executor*)> get_gen, bool use_threads,
+    Executor* executor) {
+  if (use_threads) {
+    auto used_executor = executor != NULLPTR ? executor : GetCpuThreadPool();
+    auto maybe_gen = std::move(get_gen)(used_executor);
+    if (!maybe_gen.ok()) {
+      return MakeErrorIterator<T>(maybe_gen.status());
+    }
+    return MakeGeneratorIterator(*maybe_gen);
+  } else {
+    return SerialExecutor::IterateGenerator(std::move(get_gen));
+  }
+}
+
+/// \brief Potentially iterate an async generator serially (if use_threads is false)
+///   using the default CPU thread pool
+/// \see IterateGenerator
+///
+/// If `use_threads` is true, the global CPU executor will be used.  Each call to
+///   the iterator will simply wait until the next item is available.  Tasks may run in
+///   the background between calls.
+///
+/// If `use_threads` is false, the calling thread only will be used.  Each call to
+///   the iterator will use the calling thread to do enough work to generate one item.
+///   Tasks will be left in a queue until the next call and no work will be done between
+///   calls.
+template <typename T>
+Iterator<T> IterateSynchronously(
+    FnOnce<Result<std::function<Future<T>()>>(Executor*)> get_gen, bool use_threads) {
+  return IterateSynchronously(std::move(get_gen), use_threads, NULLPTR);
+}
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/time.h

@@ -0,0 +1,83 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <chrono>
+#include <memory>
+#include <utility>
+
+#include "arrow/type_fwd.h"
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+namespace util {
+
+enum DivideOrMultiply {
+  MULTIPLY,
+  DIVIDE,
+};
+
+ARROW_EXPORT
+std::pair<DivideOrMultiply, int64_t> GetTimestampConversion(TimeUnit::type in_unit,
+                                                            TimeUnit::type out_unit);
+
+// Converts a Timestamp value into another Timestamp value.
+//
+// This function takes care of properly transforming from one unit to another.
+//
+// \param[in] in the input type. Must be TimestampType.
+// \param[in] out the output type. Must be TimestampType.
+// \param[in] value the input value.
+//
+// \return The converted value, or an error.
+ARROW_EXPORT Result<int64_t> ConvertTimestampValue(const std::shared_ptr<DataType>& in,
+                                                   const std::shared_ptr<DataType>& out,
+                                                   int64_t value);
+
+template <typename Visitor, typename... Args>
+decltype(std::declval<Visitor>()(std::chrono::seconds{}, std::declval<Args&&>()...))
+VisitDuration(TimeUnit::type unit, Visitor&& visitor, Args&&... args) {
+  switch (unit) {
+    default:
+    case TimeUnit::SECOND:
+      break;
+    case TimeUnit::MILLI:
+      return visitor(std::chrono::milliseconds{}, std::forward<Args>(args)...);
+    case TimeUnit::MICRO:
+      return visitor(std::chrono::microseconds{}, std::forward<Args>(args)...);
+    case TimeUnit::NANO:
+      return visitor(std::chrono::nanoseconds{}, std::forward<Args>(args)...);
+  }
+  return visitor(std::chrono::seconds{}, std::forward<Args>(args)...);
+}
+
+/// Convert a count of seconds to the corresponding count in a different TimeUnit
+struct CastSecondsToUnitImpl {
+  template <typename Duration>
+  int64_t operator()(Duration, int64_t seconds) {
+    auto duration = std::chrono::duration_cast<Duration>(std::chrono::seconds{seconds});
+    return static_cast<int64_t>(duration.count());
+  }
+};
+
+inline int64_t CastSecondsToUnit(TimeUnit::type unit, int64_t seconds) {
+  return VisitDuration(unit, CastSecondsToUnitImpl{}, seconds);
+}
+
+}  // namespace util
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/tracing.h

@@ -0,0 +1,45 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <memory>
+
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+namespace util {
+namespace tracing {
+
+class ARROW_EXPORT SpanDetails {
+ public:
+  virtual ~SpanDetails() {}
+};
+
+class ARROW_EXPORT Span {
+ public:
+  Span() noexcept;
+  /// True if this span has been started with START_SPAN
+  bool valid() const;
+  /// End the span early
+  void reset();
+  std::unique_ptr<SpanDetails> details;
+};
+
+}  // namespace tracing
+}  // namespace util
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/type_fwd.h

@@ -0,0 +1,82 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+namespace arrow {
+
+namespace internal {
+struct Empty;
+}  // namespace internal
+
+template <typename T = internal::Empty>
+class WeakFuture;
+class FutureWaiter;
+
+class TimestampParser;
+
+namespace internal {
+
+class Executor;
+class TaskGroup;
+class ThreadPool;
+class CpuInfo;
+
+namespace tracing {
+
+struct Scope;
+
+}  // namespace tracing
+}  // namespace internal
+
+struct Compression {
+  /// \brief Compression algorithm
+  enum type {
+    UNCOMPRESSED,
+    SNAPPY,
+    GZIP,
+    BROTLI,
+    ZSTD,
+    LZ4,
+    LZ4_FRAME,
+    LZO,
+    BZ2,
+    LZ4_HADOOP
+  };
+};
+
+namespace util {
+class AsyncTaskScheduler;
+class Compressor;
+class Decompressor;
+class Codec;
+class Uri;
+}  // namespace util
+
+template <typename T>
+struct Enumerated {
+  T value;
+  int index;
+  bool last;
+
+  friend inline bool operator==(const Enumerated<T>& left, const Enumerated<T>& right) {
+    return left.index == right.index && left.last == right.last &&
+           left.value == right.value;
+  }
+};
+
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/type_traits.h

@@ -0,0 +1,73 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cstdint>
+#include <type_traits>
+
+namespace arrow {
+namespace internal {
+
+/// \brief Metafunction to allow checking if a type matches any of another set of types
+template <typename...>
+struct IsOneOf : std::false_type {};  /// Base case: nothing has matched
+
+template <typename T, typename U, typename... Args>
+struct IsOneOf<T, U, Args...> {
+  /// Recursive case: T == U or T matches any other types provided (not including U).
+  static constexpr bool value = std::is_same<T, U>::value || IsOneOf<T, Args...>::value;
+};
+
+/// \brief Shorthand for using IsOneOf + std::enable_if
+template <typename T, typename... Args>
+using EnableIfIsOneOf = typename std::enable_if<IsOneOf<T, Args...>::value, T>::type;
+
+/// \brief is_null_pointer from C++17
+template <typename T>
+struct is_null_pointer : std::is_same<std::nullptr_t, typename std::remove_cv<T>::type> {
+};
+
+template <int kNumBytes>
+struct SizedIntImpl;
+
+template <>
+struct SizedIntImpl<1> {
+  using type = int8_t;
+};
+
+template <>
+struct SizedIntImpl<2> {
+  using type = int16_t;
+};
+
+template <>
+struct SizedIntImpl<4> {
+  using type = int32_t;
+};
+
+template <>
+struct SizedIntImpl<8> {
+  using type = int64_t;
+};
+
+// Map a number of bytes to a type
+template <int kNumBytes>
+using SizedInt = typename SizedIntImpl<kNumBytes>::type;
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/ubsan.h

@@ -0,0 +1,87 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+// Contains utilities for making UBSan happy.
+
+#pragma once
+
+#include <cstring>
+#include <memory>
+#include <type_traits>
+
+#include "arrow/util/macros.h"
+
+namespace arrow {
+namespace util {
+
+namespace internal {
+
+constexpr uint8_t kNonNullFiller = 0;
+
+}  // namespace internal
+
+/// \brief Returns maybe_null if not null or a non-null pointer to an arbitrary memory
+/// that shouldn't be dereferenced.
+///
+/// Memset/Memcpy are undefined when a nullptr is passed as an argument use this utility
+/// method to wrap locations where this could happen.
+///
+/// Note: Flatbuffers has UBSan warnings if a zero length vector is passed.
+/// https://github.com/google/flatbuffers/pull/5355 is trying to resolve
+/// them.
+template <typename T>
+inline T* MakeNonNull(T* maybe_null = NULLPTR) {
+  if (ARROW_PREDICT_TRUE(maybe_null != NULLPTR)) {
+    return maybe_null;
+  }
+
+  return const_cast<T*>(reinterpret_cast<const T*>(&internal::kNonNullFiller));
+}
+
+template <typename T>
+inline std::enable_if_t<std::is_trivially_copyable_v<T>, T> SafeLoadAs(
+    const uint8_t* unaligned) {
+  std::remove_const_t<T> ret;
+  std::memcpy(&ret, unaligned, sizeof(T));
+  return ret;
+}
+
+template <typename T>
+inline std::enable_if_t<std::is_trivially_copyable_v<T>, T> SafeLoad(const T* unaligned) {
+  std::remove_const_t<T> ret;
+  std::memcpy(&ret, static_cast<const void*>(unaligned), sizeof(T));
+  return ret;
+}
+
+template <typename U, typename T>
+inline std::enable_if_t<std::is_trivially_copyable_v<T> &&
+                            std::is_trivially_copyable_v<U> && sizeof(T) == sizeof(U),
+                        U>
+SafeCopy(T value) {
+  std::remove_const_t<U> ret;
+  std::memcpy(&ret, static_cast<const void*>(&value), sizeof(T));
+  return ret;
+}
+
+template <typename T>
+inline std::enable_if_t<std::is_trivially_copyable_v<T>, void> SafeStore(void* unaligned,
+                                                                         T value) {
+  std::memcpy(unaligned, &value, sizeof(T));
+}
+
+}  // namespace util
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/union_util.h

@@ -0,0 +1,31 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#include <cstdint>
+#include "arrow/array/data.h"
+
+namespace arrow {
+namespace union_util {
+
+/// \brief Compute the number of of logical nulls in a sparse union array
+int64_t LogicalSparseUnionNullCount(const ArraySpan& span);
+
+/// \brief Compute the number of of logical nulls in a dense union array
+int64_t LogicalDenseUnionNullCount(const ArraySpan& span);
+
+}  // namespace union_util
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/unreachable.h

@@ -0,0 +1,30 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include "arrow/util/visibility.h"
+
+#include <string_view>
+
+namespace arrow {
+
+[[noreturn]] ARROW_EXPORT void Unreachable(const char* message = "Unreachable");
+
+[[noreturn]] ARROW_EXPORT void Unreachable(std::string_view message);
+
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/uri.h

@@ -0,0 +1,119 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cstdint>
+#include <memory>
+#include <string>
+#include <string_view>
+#include <utility>
+#include <vector>
+
+#include "arrow/type_fwd.h"
+#include "arrow/util/visibility.h"
+
+namespace arrow::util {
+
+/// \brief A parsed URI
+class ARROW_EXPORT Uri {
+ public:
+  Uri();
+  ~Uri();
+  Uri(Uri&&);
+  Uri& operator=(Uri&&);
+
+  // XXX Should we use std::string_view instead?  These functions are
+  // not performance-critical.
+
+  /// The URI scheme, such as "http", or the empty string if the URI has no
+  /// explicit scheme.
+  std::string scheme() const;
+
+  /// Convenience function that returns true if the scheme() is "file"
+  bool is_file_scheme() const;
+
+  /// Whether the URI has an explicit host name.  This may return true if
+  /// the URI has an empty host (e.g. "file:///tmp/foo"), while it returns
+  /// false is the URI has not host component at all (e.g. "file:/tmp/foo").
+  bool has_host() const;
+  /// The URI host name, such as "localhost", "127.0.0.1" or "::1", or the empty
+  /// string is the URI does not have a host component.
+  std::string host() const;
+
+  /// The URI port number, as a string such as "80", or the empty string is the URI
+  /// does not have a port number component.
+  std::string port_text() const;
+  /// The URI port parsed as an integer, or -1 if the URI does not have a port
+  /// number component.
+  int32_t port() const;
+
+  /// The username specified in the URI.
+  std::string username() const;
+  /// The password specified in the URI.
+  std::string password() const;
+
+  /// The URI path component.
+  std::string path() const;
+
+  /// The URI query string
+  std::string query_string() const;
+
+  /// The URI query items
+  ///
+  /// Note this API doesn't allow differentiating between an empty value
+  /// and a missing value, such in "a&b=1" vs. "a=&b=1".
+  Result<std::vector<std::pair<std::string, std::string>>> query_items() const;
+
+  /// Get the string representation of this URI.
+  const std::string& ToString() const;
+
+  /// Factory function to parse a URI from its string representation.
+  Status Parse(const std::string& uri_string);
+
+  /// Factory function to parse a URI from its string representation.
+  static Result<Uri> FromString(const std::string& uri_string);
+
+ private:
+  struct Impl;
+  std::unique_ptr<Impl> impl_;
+};
+
+/// Percent-encode the input string, for use e.g. as a URI query parameter.
+///
+/// This will escape directory separators, making this function unsuitable
+/// for encoding URI paths directly. See UriFromAbsolutePath() instead.
+ARROW_EXPORT
+std::string UriEscape(std::string_view s);
+
+ARROW_EXPORT
+std::string UriUnescape(std::string_view s);
+
+/// Encode a host for use within a URI, such as "localhost",
+/// "127.0.0.1", or "[::1]".
+ARROW_EXPORT
+std::string UriEncodeHost(std::string_view host);
+
+/// Whether the string is a syntactically valid URI scheme according to RFC 3986.
+ARROW_EXPORT
+bool IsValidUriScheme(std::string_view s);
+
+/// Create a file uri from a given absolute path
+ARROW_EXPORT
+Result<std::string> UriFromAbsolutePath(std::string_view path);
+
+}  // namespace arrow::util

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/utf8.h

@@ -0,0 +1,59 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cstdint>
+#include <cstring>
+#include <string>
+#include <string_view>
+
+#include "arrow/type_fwd.h"
+#include "arrow/util/macros.h"
+#include "arrow/util/visibility.h"
+
+namespace arrow {
+namespace util {
+
+// Convert a UTF8 string to a wstring (either UTF16 or UTF32, depending
+// on the wchar_t width).
+ARROW_EXPORT Result<std::wstring> UTF8ToWideString(std::string_view source);
+
+// Similarly, convert a wstring to a UTF8 string.
+ARROW_EXPORT Result<std::string> WideStringToUTF8(const std::wstring& source);
+
+// Convert UTF8 string to a UTF16 string.
+ARROW_EXPORT Result<std::u16string> UTF8StringToUTF16(std::string_view source);
+
+// Convert UTF16 string to a UTF8 string.
+ARROW_EXPORT Result<std::string> UTF16StringToUTF8(std::u16string_view source);
+
+// This function needs to be called before doing UTF8 validation.
+ARROW_EXPORT void InitializeUTF8();
+
+ARROW_EXPORT bool ValidateUTF8(const uint8_t* data, int64_t size);
+
+ARROW_EXPORT bool ValidateUTF8(std::string_view str);
+
+// Skip UTF8 byte order mark, if any.
+ARROW_EXPORT
+Result<const uint8_t*> SkipUTF8BOM(const uint8_t* data, int64_t size);
+
+static constexpr uint32_t kMaxUnicodeCodepoint = 0x110000;
+
+}  // namespace util
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/value_parsing.h

@@ -0,0 +1,947 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+// This is a private header for string-to-number parsing utilities
+
+#pragma once
+
+#include <cassert>
+#include <chrono>
+#include <cstddef>
+#include <cstdint>
+#include <limits>
+#include <memory>
+#include <string>
+#include <type_traits>
+
+#include "arrow/type.h"
+#include "arrow/type_traits.h"
+#include "arrow/util/checked_cast.h"
+#include "arrow/util/config.h"
+#include "arrow/util/float16.h"
+#include "arrow/util/macros.h"
+#include "arrow/util/time.h"
+#include "arrow/util/visibility.h"
+#include "arrow/vendored/datetime.h"
+#include "arrow/vendored/strptime.h"
+
+namespace arrow {
+
+/// \brief A virtual string to timestamp parser
+class ARROW_EXPORT TimestampParser {
+ public:
+  virtual ~TimestampParser() = default;
+
+  virtual bool operator()(const char* s, size_t length, TimeUnit::type out_unit,
+                          int64_t* out,
+                          bool* out_zone_offset_present = NULLPTR) const = 0;
+
+  virtual const char* kind() const = 0;
+
+  virtual const char* format() const;
+
+  /// \brief Create a TimestampParser that recognizes strptime-like format strings
+  static std::shared_ptr<TimestampParser> MakeStrptime(std::string format);
+
+  /// \brief Create a TimestampParser that recognizes (locale-agnostic) ISO8601
+  /// timestamps
+  static std::shared_ptr<TimestampParser> MakeISO8601();
+};
+
+namespace internal {
+
+/// \brief The entry point for conversion from strings.
+///
+/// Specializations of StringConverter for `ARROW_TYPE` must define:
+/// - A default constructible member type `value_type` which will be yielded on a
+///   successful parse.
+/// - The static member function `Convert`, callable with signature
+///   `(const ARROW_TYPE& t, const char* s, size_t length, value_type* out)`.
+///   `Convert` returns truthy for successful parses and assigns the parsed values to
+///   `*out`. Parameters required for parsing (for example a timestamp's TimeUnit)
+///   are acquired from the type parameter `t`.
+template <typename ARROW_TYPE, typename Enable = void>
+struct StringConverter;
+
+template <typename T>
+struct is_parseable {
+  template <typename U, typename = typename StringConverter<U>::value_type>
+  static std::true_type Test(U*);
+
+  template <typename U>
+  static std::false_type Test(...);
+
+  static constexpr bool value = decltype(Test<T>(NULLPTR))::value;
+};
+
+template <typename T, typename R = void>
+using enable_if_parseable = enable_if_t<is_parseable<T>::value, R>;
+
+template <>
+struct StringConverter<BooleanType> {
+  using value_type = bool;
+
+  bool Convert(const BooleanType&, const char* s, size_t length, value_type* out) {
+    if (length == 1) {
+      // "0" or "1"?
+      if (s[0] == '0') {
+        *out = false;
+        return true;
+      }
+      if (s[0] == '1') {
+        *out = true;
+        return true;
+      }
+      return false;
+    }
+    if (length == 4) {
+      // "true"?
+      *out = true;
+      return ((s[0] == 't' || s[0] == 'T') && (s[1] == 'r' || s[1] == 'R') &&
+              (s[2] == 'u' || s[2] == 'U') && (s[3] == 'e' || s[3] == 'E'));
+    }
+    if (length == 5) {
+      // "false"?
+      *out = false;
+      return ((s[0] == 'f' || s[0] == 'F') && (s[1] == 'a' || s[1] == 'A') &&
+              (s[2] == 'l' || s[2] == 'L') && (s[3] == 's' || s[3] == 'S') &&
+              (s[4] == 'e' || s[4] == 'E'));
+    }
+    return false;
+  }
+};
+
+// Ideas for faster float parsing:
+// - http://rapidjson.org/md_doc_internals.html#ParsingDouble
+// - https://github.com/google/double-conversion [used here]
+// - https://github.com/achan001/dtoa-fast
+
+ARROW_EXPORT
+bool StringToFloat(const char* s, size_t length, char decimal_point, float* out);
+
+ARROW_EXPORT
+bool StringToFloat(const char* s, size_t length, char decimal_point, double* out);
+
+ARROW_EXPORT
+bool StringToFloat(const char* s, size_t length, char decimal_point,
+                   ::arrow::util::Float16* out);
+
+template <>
+struct StringConverter<FloatType> {
+  using value_type = float;
+
+  explicit StringConverter(char decimal_point = '.') : decimal_point(decimal_point) {}
+
+  bool Convert(const FloatType&, const char* s, size_t length, value_type* out) {
+    return ARROW_PREDICT_TRUE(StringToFloat(s, length, decimal_point, out));
+  }
+
+ private:
+  const char decimal_point;
+};
+
+template <>
+struct StringConverter<DoubleType> {
+  using value_type = double;
+
+  explicit StringConverter(char decimal_point = '.') : decimal_point(decimal_point) {}
+
+  bool Convert(const DoubleType&, const char* s, size_t length, value_type* out) {
+    return ARROW_PREDICT_TRUE(StringToFloat(s, length, decimal_point, out));
+  }
+
+ private:
+  const char decimal_point;
+};
+
+template <>
+struct StringConverter<HalfFloatType> {
+  using value_type = ::arrow::util::Float16;
+
+  explicit StringConverter(char decimal_point = '.') : decimal_point(decimal_point) {}
+
+  bool Convert(const HalfFloatType&, const char* s, size_t length, value_type* out) {
+    return ARROW_PREDICT_TRUE(StringToFloat(s, length, decimal_point, out));
+  }
+
+ private:
+  const char decimal_point;
+};
+
+// NOTE: HalfFloatType would require a half<->float conversion library
+
+inline uint8_t ParseDecimalDigit(char c) { return static_cast<uint8_t>(c - '0'); }
+
+#define PARSE_UNSIGNED_ITERATION(C_TYPE)          \
+  if (length > 0) {                               \
+    uint8_t digit = ParseDecimalDigit(*s++);      \
+    result = static_cast<C_TYPE>(result * 10U);   \
+    length--;                                     \
+    if (ARROW_PREDICT_FALSE(digit > 9U)) {        \
+      /* Non-digit */                             \
+      return false;                               \
+    }                                             \
+    result = static_cast<C_TYPE>(result + digit); \
+  } else {                                        \
+    break;                                        \
+  }
+
+#define PARSE_UNSIGNED_ITERATION_LAST(C_TYPE)                                     \
+  if (length > 0) {                                                               \
+    if (ARROW_PREDICT_FALSE(result > std::numeric_limits<C_TYPE>::max() / 10U)) { \
+      /* Overflow */                                                              \
+      return false;                                                               \
+    }                                                                             \
+    uint8_t digit = ParseDecimalDigit(*s++);                                      \
+    result = static_cast<C_TYPE>(result * 10U);                                   \
+    C_TYPE new_result = static_cast<C_TYPE>(result + digit);                      \
+    if (ARROW_PREDICT_FALSE(--length > 0)) {                                      \
+      /* Too many digits */                                                       \
+      return false;                                                               \
+    }                                                                             \
+    if (ARROW_PREDICT_FALSE(digit > 9U)) {                                        \
+      /* Non-digit */                                                             \
+      return false;                                                               \
+    }                                                                             \
+    if (ARROW_PREDICT_FALSE(new_result < result)) {                               \
+      /* Overflow */                                                              \
+      return false;                                                               \
+    }                                                                             \
+    result = new_result;                                                          \
+  }
+
+inline bool ParseUnsigned(const char* s, size_t length, uint8_t* out) {
+  uint8_t result = 0;
+
+  do {
+    PARSE_UNSIGNED_ITERATION(uint8_t);
+    PARSE_UNSIGNED_ITERATION(uint8_t);
+    PARSE_UNSIGNED_ITERATION_LAST(uint8_t);
+  } while (false);
+  *out = result;
+  return true;
+}
+
+inline bool ParseUnsigned(const char* s, size_t length, uint16_t* out) {
+  uint16_t result = 0;
+  do {
+    PARSE_UNSIGNED_ITERATION(uint16_t);
+    PARSE_UNSIGNED_ITERATION(uint16_t);
+    PARSE_UNSIGNED_ITERATION(uint16_t);
+    PARSE_UNSIGNED_ITERATION(uint16_t);
+    PARSE_UNSIGNED_ITERATION_LAST(uint16_t);
+  } while (false);
+  *out = result;
+  return true;
+}
+
+inline bool ParseUnsigned(const char* s, size_t length, uint32_t* out) {
+  uint32_t result = 0;
+  do {
+    PARSE_UNSIGNED_ITERATION(uint32_t);
+    PARSE_UNSIGNED_ITERATION(uint32_t);
+    PARSE_UNSIGNED_ITERATION(uint32_t);
+    PARSE_UNSIGNED_ITERATION(uint32_t);
+    PARSE_UNSIGNED_ITERATION(uint32_t);
+
+    PARSE_UNSIGNED_ITERATION(uint32_t);
+    PARSE_UNSIGNED_ITERATION(uint32_t);
+    PARSE_UNSIGNED_ITERATION(uint32_t);
+    PARSE_UNSIGNED_ITERATION(uint32_t);
+
+    PARSE_UNSIGNED_ITERATION_LAST(uint32_t);
+  } while (false);
+  *out = result;
+  return true;
+}
+
+inline bool ParseUnsigned(const char* s, size_t length, uint64_t* out) {
+  uint64_t result = 0;
+  do {
+    PARSE_UNSIGNED_ITERATION(uint64_t);
+    PARSE_UNSIGNED_ITERATION(uint64_t);
+    PARSE_UNSIGNED_ITERATION(uint64_t);
+    PARSE_UNSIGNED_ITERATION(uint64_t);
+    PARSE_UNSIGNED_ITERATION(uint64_t);
+
+    PARSE_UNSIGNED_ITERATION(uint64_t);
+    PARSE_UNSIGNED_ITERATION(uint64_t);
+    PARSE_UNSIGNED_ITERATION(uint64_t);
+    PARSE_UNSIGNED_ITERATION(uint64_t);
+    PARSE_UNSIGNED_ITERATION(uint64_t);
+
+    PARSE_UNSIGNED_ITERATION(uint64_t);
+    PARSE_UNSIGNED_ITERATION(uint64_t);
+    PARSE_UNSIGNED_ITERATION(uint64_t);
+    PARSE_UNSIGNED_ITERATION(uint64_t);
+    PARSE_UNSIGNED_ITERATION(uint64_t);
+
+    PARSE_UNSIGNED_ITERATION(uint64_t);
+    PARSE_UNSIGNED_ITERATION(uint64_t);
+    PARSE_UNSIGNED_ITERATION(uint64_t);
+    PARSE_UNSIGNED_ITERATION(uint64_t);
+
+    PARSE_UNSIGNED_ITERATION_LAST(uint64_t);
+  } while (false);
+  *out = result;
+  return true;
+}
+
+#undef PARSE_UNSIGNED_ITERATION
+#undef PARSE_UNSIGNED_ITERATION_LAST
+
+template <typename T>
+bool ParseHex(const char* s, size_t length, T* out) {
+  // lets make sure that the length of the string is not too big
+  if (!ARROW_PREDICT_TRUE(sizeof(T) * 2 >= length && length > 0)) {
+    return false;
+  }
+  T result = 0;
+  for (size_t i = 0; i < length; i++) {
+    result = static_cast<T>(result << 4);
+    if (s[i] >= '0' && s[i] <= '9') {
+      result = static_cast<T>(result | (s[i] - '0'));
+    } else if (s[i] >= 'A' && s[i] <= 'F') {
+      result = static_cast<T>(result | (s[i] - 'A' + 10));
+    } else if (s[i] >= 'a' && s[i] <= 'f') {
+      result = static_cast<T>(result | (s[i] - 'a' + 10));
+    } else {
+      /* Non-digit */
+      return false;
+    }
+  }
+  *out = result;
+  return true;
+}
+
+template <class ARROW_TYPE>
+struct StringToUnsignedIntConverterMixin {
+  using value_type = typename ARROW_TYPE::c_type;
+
+  bool Convert(const ARROW_TYPE&, const char* s, size_t length, value_type* out) {
+    if (ARROW_PREDICT_FALSE(length == 0)) {
+      return false;
+    }
+    // If it starts with 0x then its hex
+    if (length > 2 && s[0] == '0' && ((s[1] == 'x') || (s[1] == 'X'))) {
+      length -= 2;
+      s += 2;
+
+      return ARROW_PREDICT_TRUE(ParseHex(s, length, out));
+    }
+    // Skip leading zeros
+    while (length > 0 && *s == '0') {
+      length--;
+      s++;
+    }
+    return ParseUnsigned(s, length, out);
+  }
+};
+
+template <>
+struct StringConverter<UInt8Type> : public StringToUnsignedIntConverterMixin<UInt8Type> {
+  using StringToUnsignedIntConverterMixin<UInt8Type>::StringToUnsignedIntConverterMixin;
+};
+
+template <>
+struct StringConverter<UInt16Type>
+    : public StringToUnsignedIntConverterMixin<UInt16Type> {
+  using StringToUnsignedIntConverterMixin<UInt16Type>::StringToUnsignedIntConverterMixin;
+};
+
+template <>
+struct StringConverter<UInt32Type>
+    : public StringToUnsignedIntConverterMixin<UInt32Type> {
+  using StringToUnsignedIntConverterMixin<UInt32Type>::StringToUnsignedIntConverterMixin;
+};
+
+template <>
+struct StringConverter<UInt64Type>
+    : public StringToUnsignedIntConverterMixin<UInt64Type> {
+  using StringToUnsignedIntConverterMixin<UInt64Type>::StringToUnsignedIntConverterMixin;
+};
+
+template <class ARROW_TYPE>
+struct StringToSignedIntConverterMixin {
+  using value_type = typename ARROW_TYPE::c_type;
+  using unsigned_type = typename std::make_unsigned<value_type>::type;
+
+  bool Convert(const ARROW_TYPE&, const char* s, size_t length, value_type* out) {
+    static constexpr auto max_positive =
+        static_cast<unsigned_type>(std::numeric_limits<value_type>::max());
+    // Assuming two's complement
+    static constexpr unsigned_type max_negative = max_positive + 1;
+    bool negative = false;
+    unsigned_type unsigned_value = 0;
+
+    if (ARROW_PREDICT_FALSE(length == 0)) {
+      return false;
+    }
+    // If it starts with 0x then its hex
+    if (length > 2 && s[0] == '0' && ((s[1] == 'x') || (s[1] == 'X'))) {
+      length -= 2;
+      s += 2;
+
+      if (!ARROW_PREDICT_TRUE(ParseHex(s, length, &unsigned_value))) {
+        return false;
+      }
+      *out = static_cast<value_type>(unsigned_value);
+      return true;
+    }
+
+    if (*s == '-') {
+      negative = true;
+      s++;
+      if (--length == 0) {
+        return false;
+      }
+    }
+    // Skip leading zeros
+    while (length > 0 && *s == '0') {
+      length--;
+      s++;
+    }
+    if (!ARROW_PREDICT_TRUE(ParseUnsigned(s, length, &unsigned_value))) {
+      return false;
+    }
+    if (negative) {
+      if (ARROW_PREDICT_FALSE(unsigned_value > max_negative)) {
+        return false;
+      }
+      // To avoid both compiler warnings (with unsigned negation)
+      // and undefined behaviour (with signed negation overflow),
+      // use the expanded formula for 2's complement negation.
+      *out = static_cast<value_type>(~unsigned_value + 1);
+    } else {
+      if (ARROW_PREDICT_FALSE(unsigned_value > max_positive)) {
+        return false;
+      }
+      *out = static_cast<value_type>(unsigned_value);
+    }
+    return true;
+  }
+};
+
+template <>
+struct StringConverter<Int8Type> : public StringToSignedIntConverterMixin<Int8Type> {
+  using StringToSignedIntConverterMixin<Int8Type>::StringToSignedIntConverterMixin;
+};
+
+template <>
+struct StringConverter<Int16Type> : public StringToSignedIntConverterMixin<Int16Type> {
+  using StringToSignedIntConverterMixin<Int16Type>::StringToSignedIntConverterMixin;
+};
+
+template <>
+struct StringConverter<Int32Type> : public StringToSignedIntConverterMixin<Int32Type> {
+  using StringToSignedIntConverterMixin<Int32Type>::StringToSignedIntConverterMixin;
+};
+
+template <>
+struct StringConverter<Int64Type> : public StringToSignedIntConverterMixin<Int64Type> {
+  using StringToSignedIntConverterMixin<Int64Type>::StringToSignedIntConverterMixin;
+};
+
+namespace detail {
+
+// Inline-able ISO-8601 parser
+
+using ts_type = TimestampType::c_type;
+
+template <typename Duration>
+static inline bool ParseHH(const char* s, Duration* out) {
+  uint8_t hours = 0;
+  if (ARROW_PREDICT_FALSE(!ParseUnsigned(s + 0, 2, &hours))) {
+    return false;
+  }
+  if (ARROW_PREDICT_FALSE(hours >= 24)) {
+    return false;
+  }
+  *out = std::chrono::duration_cast<Duration>(std::chrono::hours(hours));
+  return true;
+}
+
+template <typename Duration>
+static inline bool ParseHH_MM(const char* s, Duration* out) {
+  uint8_t hours = 0;
+  uint8_t minutes = 0;
+  if (ARROW_PREDICT_FALSE(s[2] != ':')) {
+    return false;
+  }
+  if (ARROW_PREDICT_FALSE(!ParseUnsigned(s + 0, 2, &hours))) {
+    return false;
+  }
+  if (ARROW_PREDICT_FALSE(!ParseUnsigned(s + 3, 2, &minutes))) {
+    return false;
+  }
+  if (ARROW_PREDICT_FALSE(hours >= 24)) {
+    return false;
+  }
+  if (ARROW_PREDICT_FALSE(minutes >= 60)) {
+    return false;
+  }
+  *out = std::chrono::duration_cast<Duration>(std::chrono::hours(hours) +
+                                              std::chrono::minutes(minutes));
+  return true;
+}
+
+template <typename Duration>
+static inline bool ParseHHMM(const char* s, Duration* out) {
+  uint8_t hours = 0;
+  uint8_t minutes = 0;
+  if (ARROW_PREDICT_FALSE(!ParseUnsigned(s + 0, 2, &hours))) {
+    return false;
+  }
+  if (ARROW_PREDICT_FALSE(!ParseUnsigned(s + 2, 2, &minutes))) {
+    return false;
+  }
+  if (ARROW_PREDICT_FALSE(hours >= 24)) {
+    return false;
+  }
+  if (ARROW_PREDICT_FALSE(minutes >= 60)) {
+    return false;
+  }
+  *out = std::chrono::duration_cast<Duration>(std::chrono::hours(hours) +
+                                              std::chrono::minutes(minutes));
+  return true;
+}
+
+template <typename Duration>
+static inline bool ParseHH_MM_SS(const char* s, Duration* out) {
+  uint8_t hours = 0;
+  uint8_t minutes = 0;
+  uint8_t seconds = 0;
+  if (ARROW_PREDICT_FALSE(s[2] != ':') || ARROW_PREDICT_FALSE(s[5] != ':')) {
+    return false;
+  }
+  if (ARROW_PREDICT_FALSE(!ParseUnsigned(s + 0, 2, &hours))) {
+    return false;
+  }
+  if (ARROW_PREDICT_FALSE(!ParseUnsigned(s + 3, 2, &minutes))) {
+    return false;
+  }
+  if (ARROW_PREDICT_FALSE(!ParseUnsigned(s + 6, 2, &seconds))) {
+    return false;
+  }
+  if (ARROW_PREDICT_FALSE(hours >= 24)) {
+    return false;
+  }
+  if (ARROW_PREDICT_FALSE(minutes >= 60)) {
+    return false;
+  }
+  if (ARROW_PREDICT_FALSE(seconds >= 60)) {
+    return false;
+  }
+  *out = std::chrono::duration_cast<Duration>(std::chrono::hours(hours) +
+                                              std::chrono::minutes(minutes) +
+                                              std::chrono::seconds(seconds));
+  return true;
+}
+
+static inline bool ParseSubSeconds(const char* s, size_t length, TimeUnit::type unit,
+                                   uint32_t* out) {
+  // The decimal point has been peeled off at this point
+
+  // Fail if number of decimal places provided exceeds what the unit can hold.
+  // Calculate how many trailing decimal places are omitted for the unit
+  // e.g. if 4 decimal places are provided and unit is MICRO, 2 are missing
+  size_t omitted = 0;
+  switch (unit) {
+    case TimeUnit::MILLI:
+      if (ARROW_PREDICT_FALSE(length > 3)) {
+        return false;
+      }
+      if (length < 3) {
+        omitted = 3 - length;
+      }
+      break;
+    case TimeUnit::MICRO:
+      if (ARROW_PREDICT_FALSE(length > 6)) {
+        return false;
+      }
+      if (length < 6) {
+        omitted = 6 - length;
+      }
+      break;
+    case TimeUnit::NANO:
+      if (ARROW_PREDICT_FALSE(length > 9)) {
+        return false;
+      }
+      if (length < 9) {
+        omitted = 9 - length;
+      }
+      break;
+    default:
+      return false;
+  }
+
+  if (ARROW_PREDICT_TRUE(omitted == 0)) {
+    return ParseUnsigned(s, length, out);
+  } else {
+    uint32_t subseconds = 0;
+    bool success = ParseUnsigned(s, length, &subseconds);
+    if (ARROW_PREDICT_TRUE(success)) {
+      switch (omitted) {
+        case 1:
+          *out = subseconds * 10;
+          break;
+        case 2:
+          *out = subseconds * 100;
+          break;
+        case 3:
+          *out = subseconds * 1000;
+          break;
+        case 4:
+          *out = subseconds * 10000;
+          break;
+        case 5:
+          *out = subseconds * 100000;
+          break;
+        case 6:
+          *out = subseconds * 1000000;
+          break;
+        case 7:
+          *out = subseconds * 10000000;
+          break;
+        case 8:
+          *out = subseconds * 100000000;
+          break;
+        default:
+          // Impossible case
+          break;
+      }
+      return true;
+    } else {
+      return false;
+    }
+  }
+}
+
+}  // namespace detail
+
+template <typename Duration>
+static inline bool ParseYYYY_MM_DD(const char* s, Duration* since_epoch) {
+  uint16_t year = 0;
+  uint8_t month = 0;
+  uint8_t day = 0;
+  if (ARROW_PREDICT_FALSE(s[4] != '-') || ARROW_PREDICT_FALSE(s[7] != '-')) {
+    return false;
+  }
+  if (ARROW_PREDICT_FALSE(!ParseUnsigned(s + 0, 4, &year))) {
+    return false;
+  }
+  if (ARROW_PREDICT_FALSE(!ParseUnsigned(s + 5, 2, &month))) {
+    return false;
+  }
+  if (ARROW_PREDICT_FALSE(!ParseUnsigned(s + 8, 2, &day))) {
+    return false;
+  }
+  arrow_vendored::date::year_month_day ymd{arrow_vendored::date::year{year},
+                                           arrow_vendored::date::month{month},
+                                           arrow_vendored::date::day{day}};
+  if (ARROW_PREDICT_FALSE(!ymd.ok())) return false;
+
+  *since_epoch = std::chrono::duration_cast<Duration>(
+      arrow_vendored::date::sys_days{ymd}.time_since_epoch());
+  return true;
+}
+
+static inline bool ParseTimestampISO8601(const char* s, size_t length,
+                                         TimeUnit::type unit, TimestampType::c_type* out,
+                                         bool* out_zone_offset_present = NULLPTR) {
+  using seconds_type = std::chrono::duration<TimestampType::c_type>;
+
+  // We allow the following zone offset formats:
+  // - (none)
+  // - Z
+  // - [+-]HH(:?MM)?
+  //
+  // We allow the following formats for all units:
+  // - "YYYY-MM-DD"
+  // - "YYYY-MM-DD[ T]hhZ?"
+  // - "YYYY-MM-DD[ T]hh:mmZ?"
+  // - "YYYY-MM-DD[ T]hh:mm:ssZ?"
+  //
+  // We allow the following formats for unit == MILLI, MICRO, or NANO:
+  // - "YYYY-MM-DD[ T]hh:mm:ss.s{1,3}Z?"
+  //
+  // We allow the following formats for unit == MICRO, or NANO:
+  // - "YYYY-MM-DD[ T]hh:mm:ss.s{4,6}Z?"
+  //
+  // We allow the following formats for unit == NANO:
+  // - "YYYY-MM-DD[ T]hh:mm:ss.s{7,9}Z?"
+  //
+  // UTC is always assumed, and the DataType's timezone is ignored.
+  //
+
+  if (ARROW_PREDICT_FALSE(length < 10)) return false;
+
+  seconds_type seconds_since_epoch;
+  if (ARROW_PREDICT_FALSE(!ParseYYYY_MM_DD(s, &seconds_since_epoch))) {
+    return false;
+  }
+
+  if (length == 10) {
+    *out = util::CastSecondsToUnit(unit, seconds_since_epoch.count());
+    return true;
+  }
+
+  if (ARROW_PREDICT_FALSE(s[10] != ' ') && ARROW_PREDICT_FALSE(s[10] != 'T')) {
+    return false;
+  }
+
+  if (out_zone_offset_present) {
+    *out_zone_offset_present = false;
+  }
+
+  seconds_type zone_offset(0);
+  if (s[length - 1] == 'Z') {
+    --length;
+    if (out_zone_offset_present) *out_zone_offset_present = true;
+  } else if (s[length - 3] == '+' || s[length - 3] == '-') {
+    // [+-]HH
+    length -= 3;
+    if (ARROW_PREDICT_FALSE(!detail::ParseHH(s + length + 1, &zone_offset))) {
+      return false;
+    }
+    if (s[length] == '+') zone_offset *= -1;
+    if (out_zone_offset_present) *out_zone_offset_present = true;
+  } else if (s[length - 5] == '+' || s[length - 5] == '-') {
+    // [+-]HHMM
+    length -= 5;
+    if (ARROW_PREDICT_FALSE(!detail::ParseHHMM(s + length + 1, &zone_offset))) {
+      return false;
+    }
+    if (s[length] == '+') zone_offset *= -1;
+    if (out_zone_offset_present) *out_zone_offset_present = true;
+  } else if ((s[length - 6] == '+' || s[length - 6] == '-') && (s[length - 3] == ':')) {
+    // [+-]HH:MM
+    length -= 6;
+    if (ARROW_PREDICT_FALSE(!detail::ParseHH_MM(s + length + 1, &zone_offset))) {
+      return false;
+    }
+    if (s[length] == '+') zone_offset *= -1;
+    if (out_zone_offset_present) *out_zone_offset_present = true;
+  }
+
+  seconds_type seconds_since_midnight;
+  switch (length) {
+    case 13:  // YYYY-MM-DD[ T]hh
+      if (ARROW_PREDICT_FALSE(!detail::ParseHH(s + 11, &seconds_since_midnight))) {
+        return false;
+      }
+      break;
+    case 16:  // YYYY-MM-DD[ T]hh:mm
+      if (ARROW_PREDICT_FALSE(!detail::ParseHH_MM(s + 11, &seconds_since_midnight))) {
+        return false;
+      }
+      break;
+    case 19:  // YYYY-MM-DD[ T]hh:mm:ss
+    case 21:  // YYYY-MM-DD[ T]hh:mm:ss.s
+    case 22:  // YYYY-MM-DD[ T]hh:mm:ss.ss
+    case 23:  // YYYY-MM-DD[ T]hh:mm:ss.sss
+    case 24:  // YYYY-MM-DD[ T]hh:mm:ss.ssss
+    case 25:  // YYYY-MM-DD[ T]hh:mm:ss.sssss
+    case 26:  // YYYY-MM-DD[ T]hh:mm:ss.ssssss
+    case 27:  // YYYY-MM-DD[ T]hh:mm:ss.sssssss
+    case 28:  // YYYY-MM-DD[ T]hh:mm:ss.ssssssss
+    case 29:  // YYYY-MM-DD[ T]hh:mm:ss.sssssssss
+      if (ARROW_PREDICT_FALSE(!detail::ParseHH_MM_SS(s + 11, &seconds_since_midnight))) {
+        return false;
+      }
+      break;
+    default:
+      return false;
+  }
+
+  seconds_since_epoch += seconds_since_midnight;
+  seconds_since_epoch += zone_offset;
+
+  if (length <= 19) {
+    *out = util::CastSecondsToUnit(unit, seconds_since_epoch.count());
+    return true;
+  }
+
+  if (ARROW_PREDICT_FALSE(s[19] != '.')) {
+    return false;
+  }
+
+  uint32_t subseconds = 0;
+  if (ARROW_PREDICT_FALSE(
+          !detail::ParseSubSeconds(s + 20, length - 20, unit, &subseconds))) {
+    return false;
+  }
+
+  *out = util::CastSecondsToUnit(unit, seconds_since_epoch.count()) + subseconds;
+  return true;
+}
+
+#if defined(_WIN32) || defined(ARROW_WITH_MUSL)
+static constexpr bool kStrptimeSupportsZone = false;
+#else
+static constexpr bool kStrptimeSupportsZone = true;
+#endif
+
+/// \brief Returns time since the UNIX epoch in the requested unit
+static inline bool ParseTimestampStrptime(const char* buf, size_t length,
+                                          const char* format, bool ignore_time_in_day,
+                                          bool allow_trailing_chars, TimeUnit::type unit,
+                                          int64_t* out) {
+  // NOTE: strptime() is more than 10x faster than arrow_vendored::date::parse().
+  // The buffer may not be nul-terminated
+  std::string clean_copy(buf, length);
+  struct tm result;
+  memset(&result, 0, sizeof(struct tm));
+#ifdef _WIN32
+  char* ret = arrow_strptime(clean_copy.c_str(), format, &result);
+#else
+  char* ret = strptime(clean_copy.c_str(), format, &result);
+#endif
+  if (ret == NULLPTR) {
+    return false;
+  }
+  if (!allow_trailing_chars && static_cast<size_t>(ret - clean_copy.c_str()) != length) {
+    return false;
+  }
+  // ignore the time part
+  arrow_vendored::date::sys_seconds secs =
+      arrow_vendored::date::sys_days(arrow_vendored::date::year(result.tm_year + 1900) /
+                                     (result.tm_mon + 1) / std::max(result.tm_mday, 1));
+  if (!ignore_time_in_day) {
+    secs += (std::chrono::hours(result.tm_hour) + std::chrono::minutes(result.tm_min) +
+             std::chrono::seconds(result.tm_sec));
+#if !defined(_WIN32) && !defined(_AIX)
+    secs -= std::chrono::seconds(result.tm_gmtoff);
+#endif
+  }
+  *out = util::CastSecondsToUnit(unit, secs.time_since_epoch().count());
+  return true;
+}
+
+template <>
+struct StringConverter<TimestampType> {
+  using value_type = int64_t;
+
+  bool Convert(const TimestampType& type, const char* s, size_t length, value_type* out) {
+    return ParseTimestampISO8601(s, length, type.unit(), out);
+  }
+};
+
+template <>
+struct StringConverter<DurationType>
+    : public StringToSignedIntConverterMixin<DurationType> {
+  using StringToSignedIntConverterMixin<DurationType>::StringToSignedIntConverterMixin;
+};
+
+template <typename DATE_TYPE>
+struct StringConverter<DATE_TYPE, enable_if_date<DATE_TYPE>> {
+  using value_type = typename DATE_TYPE::c_type;
+
+  using duration_type =
+      typename std::conditional<std::is_same<DATE_TYPE, Date32Type>::value,
+                                arrow_vendored::date::days,
+                                std::chrono::milliseconds>::type;
+
+  bool Convert(const DATE_TYPE& type, const char* s, size_t length, value_type* out) {
+    if (ARROW_PREDICT_FALSE(length != 10)) {
+      return false;
+    }
+
+    duration_type since_epoch;
+    if (ARROW_PREDICT_FALSE(!ParseYYYY_MM_DD(s, &since_epoch))) {
+      return false;
+    }
+
+    *out = static_cast<value_type>(since_epoch.count());
+    return true;
+  }
+};
+
+template <typename TIME_TYPE>
+struct StringConverter<TIME_TYPE, enable_if_time<TIME_TYPE>> {
+  using value_type = typename TIME_TYPE::c_type;
+
+  // We allow the following formats for all units:
+  // - "hh:mm"
+  // - "hh:mm:ss"
+  //
+  // We allow the following formats for unit == MILLI, MICRO, or NANO:
+  // - "hh:mm:ss.s{1,3}"
+  //
+  // We allow the following formats for unit == MICRO, or NANO:
+  // - "hh:mm:ss.s{4,6}"
+  //
+  // We allow the following formats for unit == NANO:
+  // - "hh:mm:ss.s{7,9}"
+
+  bool Convert(const TIME_TYPE& type, const char* s, size_t length, value_type* out) {
+    const auto unit = type.unit();
+    std::chrono::seconds since_midnight;
+
+    if (length == 5) {
+      if (ARROW_PREDICT_FALSE(!detail::ParseHH_MM(s, &since_midnight))) {
+        return false;
+      }
+      *out =
+          static_cast<value_type>(util::CastSecondsToUnit(unit, since_midnight.count()));
+      return true;
+    }
+
+    if (ARROW_PREDICT_FALSE(length < 8)) {
+      return false;
+    }
+    if (ARROW_PREDICT_FALSE(!detail::ParseHH_MM_SS(s, &since_midnight))) {
+      return false;
+    }
+
+    *out = static_cast<value_type>(util::CastSecondsToUnit(unit, since_midnight.count()));
+
+    if (length == 8) {
+      return true;
+    }
+
+    if (ARROW_PREDICT_FALSE(s[8] != '.')) {
+      return false;
+    }
+
+    uint32_t subseconds_count = 0;
+    if (ARROW_PREDICT_FALSE(
+            !detail::ParseSubSeconds(s + 9, length - 9, unit, &subseconds_count))) {
+      return false;
+    }
+
+    *out += subseconds_count;
+    return true;
+  }
+};
+
+/// \brief Convenience wrappers around internal::StringConverter.
+template <typename T>
+bool ParseValue(const T& type, const char* s, size_t length,
+                typename StringConverter<T>::value_type* out) {
+  return StringConverter<T>{}.Convert(type, s, length, out);
+}
+
+template <typename T>
+enable_if_parameter_free<T, bool> ParseValue(
+    const char* s, size_t length, typename StringConverter<T>::value_type* out) {
+  static T type;
+  return StringConverter<T>{}.Convert(type, s, length, out);
+}
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/vector.h

@@ -0,0 +1,172 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <algorithm>
+#include <utility>
+#include <vector>
+
+#include "arrow/result.h"
+#include "arrow/util/algorithm.h"
+#include "arrow/util/functional.h"
+#include "arrow/util/logging.h"
+
+namespace arrow {
+namespace internal {
+
+template <typename T>
+std::vector<T> DeleteVectorElement(const std::vector<T>& values, size_t index) {
+  ARROW_DCHECK(!values.empty());
+  ARROW_DCHECK_LT(index, values.size());
+  std::vector<T> out;
+  out.reserve(values.size() - 1);
+  for (size_t i = 0; i < index; ++i) {
+    out.push_back(values[i]);
+  }
+  for (size_t i = index + 1; i < values.size(); ++i) {
+    out.push_back(values[i]);
+  }
+  return out;
+}
+
+template <typename T>
+std::vector<T> AddVectorElement(const std::vector<T>& values, size_t index,
+                                T new_element) {
+  ARROW_DCHECK_LE(index, values.size());
+  std::vector<T> out;
+  out.reserve(values.size() + 1);
+  for (size_t i = 0; i < index; ++i) {
+    out.push_back(values[i]);
+  }
+  out.emplace_back(std::move(new_element));
+  for (size_t i = index; i < values.size(); ++i) {
+    out.push_back(values[i]);
+  }
+  return out;
+}
+
+template <typename T>
+std::vector<T> ReplaceVectorElement(const std::vector<T>& values, size_t index,
+                                    T new_element) {
+  ARROW_DCHECK_LE(index, values.size());
+  std::vector<T> out;
+  out.reserve(values.size());
+  for (size_t i = 0; i < index; ++i) {
+    out.push_back(values[i]);
+  }
+  out.emplace_back(std::move(new_element));
+  for (size_t i = index + 1; i < values.size(); ++i) {
+    out.push_back(values[i]);
+  }
+  return out;
+}
+
+template <typename T, typename Predicate>
+std::vector<T> FilterVector(std::vector<T> values, Predicate&& predicate) {
+  auto new_end = std::remove_if(values.begin(), values.end(),
+                                [&](const T& value) { return !predicate(value); });
+  values.erase(new_end, values.end());
+  return values;
+}
+
+template <typename Fn, typename From,
+          typename To = decltype(std::declval<Fn>()(std::declval<From>()))>
+std::vector<To> MapVector(Fn&& map, const std::vector<From>& source) {
+  std::vector<To> out;
+  out.reserve(source.size());
+  std::transform(source.begin(), source.end(), std::back_inserter(out),
+                 std::forward<Fn>(map));
+  return out;
+}
+
+template <typename Fn, typename From,
+          typename To = decltype(std::declval<Fn>()(std::declval<From>()))>
+std::vector<To> MapVector(Fn&& map, std::vector<From>&& source) {
+  std::vector<To> out;
+  out.reserve(source.size());
+  std::transform(std::make_move_iterator(source.begin()),
+                 std::make_move_iterator(source.end()), std::back_inserter(out),
+                 std::forward<Fn>(map));
+  return out;
+}
+
+/// \brief Like MapVector, but where the function can fail.
+template <typename Fn, typename From = internal::call_traits::argument_type<0, Fn>,
+          typename To = typename internal::call_traits::return_type<Fn>::ValueType>
+Result<std::vector<To>> MaybeMapVector(Fn&& map, const std::vector<From>& source) {
+  std::vector<To> out;
+  out.reserve(source.size());
+  ARROW_RETURN_NOT_OK(MaybeTransform(source.begin(), source.end(),
+                                     std::back_inserter(out), std::forward<Fn>(map)));
+  return out;
+}
+
+template <typename Fn, typename From = internal::call_traits::argument_type<0, Fn>,
+          typename To = typename internal::call_traits::return_type<Fn>::ValueType>
+Result<std::vector<To>> MaybeMapVector(Fn&& map, std::vector<From>&& source) {
+  std::vector<To> out;
+  out.reserve(source.size());
+  ARROW_RETURN_NOT_OK(MaybeTransform(std::make_move_iterator(source.begin()),
+                                     std::make_move_iterator(source.end()),
+                                     std::back_inserter(out), std::forward<Fn>(map)));
+  return std::move(out);
+}
+
+template <typename T>
+std::vector<T> FlattenVectors(const std::vector<std::vector<T>>& vecs) {
+  std::size_t sum = 0;
+  for (const auto& vec : vecs) {
+    sum += vec.size();
+  }
+  std::vector<T> out;
+  out.reserve(sum);
+  for (const auto& vec : vecs) {
+    out.insert(out.end(), vec.begin(), vec.end());
+  }
+  return out;
+}
+
+template <typename T>
+Result<std::vector<T>> UnwrapOrRaise(std::vector<Result<T>>&& results) {
+  std::vector<T> out;
+  out.reserve(results.size());
+  auto end = std::make_move_iterator(results.end());
+  for (auto it = std::make_move_iterator(results.begin()); it != end; it++) {
+    if (!it->ok()) {
+      return it->status();
+    }
+    out.push_back(it->MoveValueUnsafe());
+  }
+  return out;
+}
+
+template <typename T>
+Result<std::vector<T>> UnwrapOrRaise(const std::vector<Result<T>>& results) {
+  std::vector<T> out;
+  out.reserve(results.size());
+  for (const auto& result : results) {
+    if (!result.ok()) {
+      return result.status();
+    }
+    out.push_back(result.ValueUnsafe());
+  }
+  return out;
+}
+
+}  // namespace internal
+}  // namespace arrow

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/visibility.h

@@ -0,0 +1,95 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#if defined(_WIN32) || defined(__CYGWIN__)
+// Windows
+
+#  if defined(_MSC_VER)
+#    pragma warning(disable : 4251)
+#  else
+#    pragma GCC diagnostic ignored "-Wattributes"
+#  endif
+
+#  if defined(__cplusplus) && defined(__GNUC__) && !defined(__clang__)
+// Use C++ attribute syntax where possible to avoid GCC parser bug
+// (https://stackoverflow.com/questions/57993818/gcc-how-to-combine-attribute-dllexport-and-nodiscard-in-a-struct-de)
+#    define ARROW_DLLEXPORT [[gnu::dllexport]]
+#    define ARROW_DLLIMPORT [[gnu::dllimport]]
+#  else
+#    define ARROW_DLLEXPORT __declspec(dllexport)
+#    define ARROW_DLLIMPORT __declspec(dllimport)
+#  endif
+
+// _declspec(dllexport) even when #included by a non-arrow source
+#  define ARROW_FORCE_EXPORT ARROW_DLLEXPORT
+
+#  ifdef ARROW_STATIC
+#    define ARROW_EXPORT
+#    define ARROW_FRIEND_EXPORT
+#    define ARROW_TEMPLATE_EXPORT
+#  elif defined(ARROW_EXPORTING)
+#    define ARROW_EXPORT ARROW_DLLEXPORT
+// For some reason [[gnu::dllexport]] doesn't work well with friend declarations
+#    define ARROW_FRIEND_EXPORT __declspec(dllexport)
+#    define ARROW_TEMPLATE_EXPORT ARROW_DLLEXPORT
+#  else
+#    define ARROW_EXPORT ARROW_DLLIMPORT
+#    define ARROW_FRIEND_EXPORT __declspec(dllimport)
+#    define ARROW_TEMPLATE_EXPORT ARROW_DLLIMPORT
+#  endif
+
+#  define ARROW_NO_EXPORT
+
+#else
+
+// Non-Windows
+
+#  if defined(__cplusplus) && (defined(__GNUC__) || defined(__clang__))
+#    ifndef ARROW_EXPORT
+#      define ARROW_EXPORT [[gnu::visibility("default")]]
+#    endif
+#    ifndef ARROW_NO_EXPORT
+#      define ARROW_NO_EXPORT [[gnu::visibility("hidden")]]
+#    endif
+// The C++ language does not have clear rules for how to export explicit template
+// instantiations, and clang/gcc have differing syntax. See
+// https://github.com/llvm/llvm-project/issues/29464 and
+// https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0537r0.html
+#    if defined(__clang__)
+#      define ARROW_TEMPLATE_EXPORT
+#    else
+#      define ARROW_TEMPLATE_EXPORT ARROW_EXPORT
+#    endif
+#  else
+// Not C++, or not gcc/clang
+#    ifndef ARROW_EXPORT
+#      define ARROW_EXPORT
+#    endif
+#    ifndef ARROW_NO_EXPORT
+#      define ARROW_NO_EXPORT
+#    endif
+#    define ARROW_TEMPLATE_EXPORT
+#  endif
+
+#  define ARROW_FRIEND_EXPORT
+
+// [[gnu::visibility("default")]] even when #included by a non-arrow source
+#  define ARROW_FORCE_EXPORT [[gnu::visibility("default")]]
+
+#endif  // Non-Windows

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/windows_compatibility.h

@@ -0,0 +1,39 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#ifdef _WIN32
+
+// Windows defines min and max macros that mess up std::min/max
+#  ifndef NOMINMAX
+#    define NOMINMAX
+#  endif
+
+#  define WIN32_LEAN_AND_MEAN
+
+// Set Windows 7 as a conservative minimum for Apache Arrow
+#  if defined(_WIN32_WINNT) && _WIN32_WINNT < 0x601
+#    undef _WIN32_WINNT
+#  endif
+#  ifndef _WIN32_WINNT
+#    define _WIN32_WINNT 0x601
+#  endif
+
+#  include <winsock2.h>
+
+#  include "arrow/util/windows_fixup.h"
+
+#endif  // _WIN32

venv/lib/python3.10/site-packages/pyarrow/include/arrow/util/windows_fixup.h

@@ -0,0 +1,52 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+// This header needs to be included multiple times.
+
+#ifdef _WIN32
+
+#  ifdef max
+#    undef max
+#  endif
+#  ifdef min
+#    undef min
+#  endif
+
+// The Windows API defines macros from *File resolving to either
+// *FileA or *FileW.  Need to undo them.
+#  ifdef CopyFile
+#    undef CopyFile
+#  endif
+#  ifdef CreateFile
+#    undef CreateFile
+#  endif
+#  ifdef DeleteFile
+#    undef DeleteFile
+#  endif
+
+// Other annoying Windows macro definitions...
+#  ifdef IN
+#    undef IN
+#  endif
+#  ifdef OUT
+#    undef OUT
+#  endif
+
+// Note that we can't undefine OPTIONAL, because it can be used in other
+// Windows headers...
+
+#endif  // _WIN32