p0019R8
Atomic Ref

Published Proposal,

This version:
https://github.com/ORNL/cpp-proposals-pub/blob/master/P0019/P0019r7.html
Authors:
Audience:
LWG
Toggle Diffs:
Project:
ISO JTC1/SC22/WG21: Programming Language C++

Abstract

Extension to the atomic operations library to allow atomic operations to apply to non-atomic objects.

1. Revision History

[P0019r8]

[P0019r7]

[P0019r6]

[P0019r5]

[P0019r4]

[P0019r3]

2. Overview

This paper proposes an extension to the atomic operations library [atomics] to allow atomic operations to apply to non-atomic objects. As required by [atomics.types.generic] the value type T must be trivially copyable.

This paper includes atomic floating point capability defined in [P0020r5].

Note: A reference implementation](https://github.com/ORNL/cpp-proposals-pub/blob/master/P0019/atomic_ref.hpp) is available that works on compilers which support the GNU atomic builtin functions including recent versions of g++, icpc, and clang++. --end note

This paper is currently targeting C++20.

3. Motivation

3.1. Atomic Operations on a Single Non-atomic Object

An atomic reference is used to perform atomic operations on a referenced non-atomic object. The intent is for atomic reference to provide the best-performing implementation of atomic operations for the non-atomic object type. All atomic operations performed through an atomic reference on a referenced non-atomic object are atomic with respect to any other atomic reference that references the same object, as defined by equality of pointers to that object. The intent is for atomic operations to directly update the referenced object. An atomic reference constructor might acquire a resource, such as a lock from a collection of address-sharded locks, to perform atomic operations. Such atomic reference objects are not lock free and not address free. When such a resource is necessary, subsequent copy and move constructors and assignment operators might reduce overhead by copying or moving the previously acquired resource as opposed to re-acquiring that resource.

Introducing concurrency within legacy code might require replacing operations on existing non-atomic objects with atomic operations such that the non-atomic object cannot be replaced with an atomic object.

An object could be heavily used non-atomically in well-defined phases of an application. Forcing such objects to be exclusively atomic would incur an unnecessary performance penalty.

3.2. Atomic Operations on Members of a Very Large Array

High-performance computing (HPC) applications use very large arrays. Computations with these arrays typically have distinct phases that allocate and initialize members of the array, update members of the array, and read members of the array. Parallel algorithms for initialization (e.g., zero fill) have non-conflicting access when assigning member values. Parallel algorithms for updates have conflicting access to members which must be guarded by atomic operations. Parallel algorithms with read-only access require best-performing streaming read access, random read access, vectorization, or other guaranteed non-conflicting HPC pattern.

4. Reference-ability Constraints

An object referenced by an atomic reference must satisfy possibly architecture-specific constraints. For example, the object might need to be properly aligned in memory or might not be allowed to reside in GPU register memory. We do not enumerate all potential constraints or specify behavior when these constraints are violated. It is a quality-of-implementation issue to generate appropriate information when constraints are violated.

Note: Whether an implementation of atomic<T> is lock free, does not necessarily constrain whether the corresponding implementation of atomic_ref<T> is lock free. --end note

5. Concern with atomic<T> and padding bits in T

A concern has been discussed for atomic<T> where T is a class type that contains padding bits and how construction and compare_exchange operations are effected by the value of those padding bits. We require that the resolution of padding bits follow [P0528r2].

6. Naming atomic reference of T

2018-03-15 Suggested that atomic reference of T be named atomic_ref<T> instead of atomic<T&>?

Result of SG1 poll

SF F N A SA
0 7 11 3 1

Those against atomic<T&> raised the concern that it allows dangerous errors to creep into generic code, which requires users to be aware of this edge case to avoid. Also, after an atomic<T> is constructed it does not have data races with other objects, while an atomic reference of T does. Furthermore atomic<T&> does not have volatile member functions. Consequently, atomic<T&> is a specialization of atomic<T> with weaker guarantees.

The arguments for atomic<T&> is that it is more concise and reduces the vocabulary terms that a user needs to know.

We decided to keep the name of an atomic reference of T as atomic_ref<T> for two reasons. First, using the name atomic_ref<T> removes any possibility of impacting existing generic code which uses atomic<T>. Second, when trying to create wording for atomic<T&> the specializations had a distinct vector<bool> feel where each specialization needed to walk back from guarantees made by the primary template. In particular, the atomic<T&> specializations would be unable to use the phrase "Descriptions are provided below only for members that differ from the primary template".

7. Future Work

  1. Rewrite atomic<T> in terms of atomic_ref<T>?

    Result of SG1 poll

    SF F N A SA
    4 3 13 0 0

    [P0019r8] duplicates much of the wording of atomic<T> and it is desirable to reduce this overhead. However, rewriting atomic<T> would require revisiting the contentious issue of atomic<T> having an exposition only member of type T. While there is strong support for this rewrite in SG1 and LWG, we are limiting the scope of this paper to atomic reference to reduce unnecessary conflicts.

  2. The free functions and macros intended for C compatibility are deliberately omitted from this paper, but can be added if/when a need for them arises.

8. Proposal

The proposed changes are relative to the working draft of the standard as of [N4727].

Text in blockquotes is not proposed wording

The � character is used to denote a placeholder section number which the editor shall determine.

[ Note: Most of the wording in this paper is duplicated from the wording for atomic<T>. A future paper is needed to define atomic<T> in terms of atomic_ref<T> and improve the current wording for atomic objects. --end note]
Apply the following changes to 32.2.� [atomics.syn]:
namespace std {

// 3.� atomic ref
template<class T> struct atomic_ref;
// 3.� atomic ref partial specialization for pointers
template<class T> struct atomic_ref<T*>;

}
Add a new subsection [atomics.ref.generic] before [atomics.types.generic]

Class template atomic_ref

template<class T> struct atomic_ref {
private:
  T* ptr; // exposition only
public:
  using value_type = T;
  static constexpr bool is_always_lock_free = implementation-defined;
  static constexpr size_t required_alignment = implementation-defined;

  atomic_ref() = delete;
  atomic_ref& operator=(const atomic_ref&) = delete;

  explicit atomic_ref(T&);
  atomic_ref(const atomic_ref&) noexcept;

  T operator=(T) const noexcept;
  operator T() const noexcept;

  bool is_lock_free() const noexcept;
  void store(T, memory_order = memory_order_seq_cst) const noexcept;
  T load(memory_order = memory_order_seq_cst) const noexcept;
  T exchange(T, memory_order = memory_order_seq_cst) const noexcept;
  bool compare_exchange_weak(T&, T, 
                             memory_order, memory_order) const noexcept;
  bool compare_exchange_strong(T&, T,
                               memory_order, memory_order) const noexcept;
  bool compare_exchange_weak(T&, T,
                             memory_order = memory_order_seq_cst) const noexcept;
  bool compare_exchange_strong(T&, T,
                               memory_order = memory_order_seq_cst) const noexcept;
};

An atomic_ref object applies atomic operations [atomics.general] to the object referenced by *ptr such that, for the lifetime [basic.life] of the atomic_ref object, the object referenced by *ptr is an atomic object [intro.races].

The template argument for T shall be trivially copyable [basic.types].

The lifetime [basic.life] of an object referenced by *ptr shall exceed the lifetime of all atomic_refs that reference the object. While any atomic_ref instances exist which reference the *ptr object all accesses to that object shall exclusively occur through those atomic_ref instances. No subobject of the object referenced by atomic_ref shall be concurrently referenced by any other atomic_ref object.

Atomic operations applied to an object through a referencing atomic_ref are atomic with respect to atomic operations applied through any other atomic_ref referencing the same object.

[Note: Atomic operations or the atomic_ref constructor could acquire a shared resource, such as a lock associated with the referenced object, to enable atomic operations applied to the referenced object. - end note]

Add a new subsubsection [atomics.ref.operations] after [atomics.ref.generic]

Operations on atomic types

static constexpr bool is_always_lockfree;

The static data member is_always_lock_free is true if the atomic_ref type’s operations are always lock-free, and false otherwise.

static constexpr size_t required_alignment;

The required alignment of an object to be referenced by an atomic reference, which is at least alignof(T).

[Note: Hardware could require that an object to be referenced by an atomic_ref have stricter alignment [basic.align] than other objects of type T. Further, whether operations on an atomic_ref are lock-free could depend on the alignment of the referenced object. For example, lock-free operations on std::complex<double> could be supported only if aligned to 2*alignof(double). - end note ]

atomic_ref(T& obj);

Requires: The referenced object shall be aligned to required_alignment.

Effects: Constructs an atomic reference that references the object.

Throws: Nothing.

atomic_ref(const atomic_ref& ref) noexcept;

Effects: Construct an atomic reference that references the object referenced by ref.

T operator=(T desired) const noexcept;

Effects: Equivalent to:

  store(desired);
  return desired;

operator T() const noexcept;

Effects: Equivalent to: return load();

bool is_lock_free() const noexcept;

Returns: true if the object’s operations are lock-free, false otherwise.

void store(T desired, memory_order order = memory_order_seq_cst) const noexcept;

Requires: The order argument shall not be memory_order_consume, memory_order_acquire, nor memory_order_acq_rel.

Effects: Atomically replaces the value referenced by *ptr with the value of desired. Memory is affected according to the value of order.

T load(memory_order order = memory_order_seq_cst) const noexcept;

Requires: The order argument shall not be memory_order_release nor memory_order_acq_rel.

Effects: Memory is affected according to the value of order.

Returns: Atomically returns the value referenced by *ptr.

T exchange(T desired, memory_order order = memory_order_seq_cst) noexcept;

Effects: Atomically replaces the value referenced by *ptr with desired. Memory is affected according to the value of order. These operations are atomic read-modify-write operations [intro.multithread].

Returns: Atomically returns the value referenced by *ptr immediately before the effects.

bool compare_exchange_weak(T& expected, T desired, memory_order success, memory_order failure) const noexcept;

bool compare_exchange_strong(T& expected, T desired, memory_order success, memory_order failure) const noexcept;

bool compare_exchange_weak(T& expected, T desired, memory_order order = memory_order_seq_cst) const noexcept;

bool compare_exchange_strong(T& expected, T desired, memory_order order = memory_order_seq_cst) const noexcept;

Requires: The failure argument shall not be memory_order_release nor memory_order_acq_rel.

Effects: Retrieves the value in expected. It then atomically compares the value referenced by *ptr for equality with that previously retrieved from expected, and if true, replaces the value referenced by *ptr with that in desired. If and only if the comparison is true, memory is affected according to the value of success, and if the comparison is false, memory is affected according to the value of failure. When only one memory_order argument is supplied, the value of success is order, and the value of failure is order except that a value of memory_order_acq_rel shall be replaced by the value memory_order_acquire and a value of memory_order_release shall be replaced by the value memory_order_relaxed. If and only if the comparison is false then, after the atomic operation, the contents of the memory in expected are replaced by the value read from the value referenced by *ptr during the atomic comparison. If the operation returns true, these operations are atomic read-modify-write operations [intro.races] on the value referenced by *ptr. Otherwise, these operations are atomic load operations on that memory.

Returns: The result of the comparison.

Remarks: A weak compare-and-exchange operation may fail spuriously. That is, even when the contents of memory referred to by expected and ptr are equal, it may return false and store back to expected the same memory contents that were originally there. [ Note: This spurious failure enables implementation of compare-and-exchange on a broader class of machines, e.g., load-locked store-conditional machines. A consequence of spurious failure is that nearly all uses of weak compare-and-exchange will be in a loop. When a compare-and-exchange is in a loop, the weak version will yield better performance on some platforms. When a weak compare-and-exchange would require a loop and a strong one would not, the strong one is preferable. — end note ]

Add a new subsubsection [atomics.ref.int] following the [atomics.ref.operations] subsubsection

Specializations for integral types

There are specializations of the atomic template for the integral types char, signed char, unsigned char, short, unsigned short, int, unsigned int, long, unsigned long, long long, unsigned long long, char16_t, char32_t, wchar_t, and any other types needed by the typedefs in the header <cstdint>. For each such integral type integral, the specialization atomic_ref<integral> provides additional atomic operations appropriate to integral types. [ Note: For the specialization atomic_ref<bool>, see [atomics.ref.generic]. — end note ]

template<> struct atomic_ref<integral> {
private:
  integral* ptr; // exposition only
public:
  using value_type = integral;
  using difference_type = value_type;
  static constexpr bool is_always_lock_free = implementation-defined;
  static constexpr size_t required_alignment = implementation-defined;

  atomic_ref() = delete;
  atomic_ref& operator = (const atomic_ref&) = delete;

  explicit atomic_ref(integral&);
  atomic_ref(const atomic_ref&) noexcept;

  integral operator=(integral) const noexcept;
  operator integral () const noexcept;

  bool is_lock_free() const noexcept;
  void store(integral, memory_order = memory_order_seq_cst) const noexcept;
  integral load(memory_order = memory_order_seq_cst) const noexcept;
  integral exchange(integral,
                    memory_order = memory_order_seq_cst) const noexcept;
  bool compare_exchange_weak(integral&, integral,
                             memory_order, memory_order) const noexcept;
  bool compare_exchange_strong(integral&, integral,
                               memory_order, memory_order) const noexcept;
  bool compare_exchange_weak(integral&, integral,
                             memory_order = memory_order_seq_cst) const noexcept;
  bool compare_exchange_strong(integral&, integral,
                               memory_order = memory_order_seq_cst) const noexcept;

  integral fetch_add(integral,
                     memory_order = memory_order_seq_cst) const noexcept;
  integral fetch_sub(integral,
                     memory_order = memory_order_seq_cst) const noexcept;
  integral fetch_and(integral,
                     memory_order = memory_order_seq_cst) const noexcept;
  integral fetch_or(integral,
                    memory_order = memory_order_seq_cst) const noexcept;
  integral fetch_xor(integral,
                     memory_order = memory_order_seq_cst) const noexcept;

  integral operator++(int) const noexcept;
  integral operator--(int) const noexcept;
  integral operator++() const noexcept;
  integral operator--() const noexcept;
  integral operator+=(integral) const noexcept;
  integral operator-=(integral) const noexcept;
  integral operator&=(integral) const noexcept;
  integral operator|=(integral) const noexcept;
  integral operator^=(integral) const noexcept;
};

Descriptions are provided below only for members that differ from the primary template.

The following operations perform arithmetic computations. The key, operator, and computation correspondence are identified in Table 130 [atomics.types.int].

integral fetch_key(integral operand, memory_order order = memory_order_seq_cst) const noexcept;

Effects: Atomically replaces the value referenced by *ptr with the result of the computation applied to the value referenced by *ptr and the given operand. Memory is affected according to the value of order. These operations are atomic read-modify-write operations [intro.races].

Returns: Atomically, the value referenced by *ptr immediately before the effects.

Remarks: For signed integer types, arithmetic is defined to use two’s complement representation. There are no undefined results.

integral operator op=(integral operand) const noexcept;

Effects: Equivalent to: return fetch_key(operand) op operand;

Add a new subsubsection [atomics.ref.float] following the [atomics.ref.int] subsubsection

Specializations for floating-point types

There are specializations of the atomic_ref template for the floating-point types float, double, and long double. For each such floating-point type floating-point, the specialization atomic_ref<floating-point> provides additional atomic operations appropriate to floating-point types.

template<> struct atomic_ref<floating-point> {
private:
  floating-point* ptr; // exposition only
public:
  using value_type = floating-point;
  using difference_type = value_type;
  static constexpr bool is_always_lock_free = implementation-defined;
  static constexpr size_t required_alignment = implementation-defined;

  atomic_ref() = delete;
  atomic_ref& operator = (const atomic_ref&) = delete;

  explicit atomic_ref(floating-point&);
  atomic_ref(const atomic_ref&) noexcept;

  floating-point operator=(floating-point) noexcept;
  operator floating-point () const noexcept;

  bool is_lock_free() const noexcept;
  void store(floating-point, memory_order = memory_order_seq_cst) const noexcept;
  floating-point load(memory_order = memory_order_seq_cst) const noexcept;
  floating-point exchange(floating-point,
                          memory_order = memory_order_seq_cst) const noexcept;
  bool compare_exchange_weak(floating-point&, floating-point,
                             memory_order, memory_order) const noexcept;
  bool compare_exchange_strong(floating-point&, floating-point,
                               memory_order, memory_order) const noexcept;
  bool compare_exchange_weak(floating-point&, floating-point,
                             memory_order = memory_order_seq_cst) const noexcept;
  bool compare_exchange_strong(floating-point&, floating-point,
                               memory_order = memory_order_seq_cst) const noexcept;

  floating-point fetch_add(floating-point,
                           memory_order = memory_order_seq_cst) const noexcept;
  floating-point fetch_sub(floating-point,
                           memory_order = memory_order_seq_cst) const noexcept;

  floating-point operator+=(floating-point) const noexcept;
  floating-point operator-=(floating-point) const noexcept;
};

Descriptions are provided below only for members that differ from the primary template.

The following operations perform arithmetic computations. The key, operator, and computation correspondence are identified in Table 130 [atomics.types.int].

floating-point fetch_key(floating-point operand, memory_order order = memory_order_seq_cst) const noexcept;

Effects: Atomically replaces the value referenced by *ptr with the result of the computation applied to the value referenced by *ptr and the given operand. Memory is affected according to the value of order. These operations are atomic read-modify-write operations [intro.races].

Returns: Atomically, the value referenced by *ptr immediately before the effects.

Remarks: If the result is not a representable value for its type [expr.pre] the result is unspecified, but the operations otherwise have no undefined behavior. Atomic arithmetic operations on floating-point should conform to the std::numeric_limits<floating-point> traits associated with the floating-point type [limits.syn]. The floating-point environment [cfenv] for atomic arithmetic operations on floating-point may be different than the calling thread’s floating-point environment.

floating-point operator op=(floating-point operand) const noexcept;

Effects: Equivalent to: return fetch_key(operand) op operand;

Add a new subsubsection [atomics.ref.pointer] following the [atomics.ref.float] subsubsection

Partial specialization for pointers

template<class T> struct atomic_ref<T*> {
private:
  T** ptr; // exposition only
public:
  using value_type = T*;
  using difference_type = ptrdiff_t;
  static constexpr bool is_always_lock_free = implementation-defined;
  static constexpr size_t required_alignment = implementation-defined;

  atomic_ref() = delete;
  atomic_ref& operator = (const atomic_ref&) = delete;

  explicit atomic_ref(T*&);
  atomic_ref(const atomic_ref&) noexcept;

  T* operator=(T*) const noexcept;
  operator T* () const noexcept;

  bool is_lock_free() const noexcept;
  void store(T*, memory_order = memory_order_seq_cst) const noexcept;
  T* load(memory_order = memory_order_seq_cst) const noexcept;
  T* exchange(T*, memory_order = memory_order_seq_cst) const noexcept;
  bool compare_exchange_weak(T*&, T*,
                             memory_order, memory_order) const noexcept;
  bool compare_exchange_strong(T*&, T*,
                               memory_order, memory_order) const noexcept;
  bool compare_exchange_weak(T*&, T*,
                             memory_order = memory_order_seq_cst) const noexcept;
  bool compare_exchange_strong(T*&, T*,
                               memory_order = memory_order_seq_cst) const noexcept;

  T* fetch_add(difference_type, memory_order = memory_order_seq_cst) const noexcept;
  T* fetch_sub(difference_type, memory_order = memory_order_seq_cst) const noexcept;

  T* operator++(int) const noexcept;
  T* operator--(int) const noexcept;
  T* operator++() const noexcept;
  T* operator--() const noexcept;
  T* operator+=(difference_type) const noexcept;
  T* operator-=(difference_type) const noexcept;
};

Descriptions are provided below only for members that differ from the primary template.

The following operations perform arithmetic computations. The key, operator, and computation correspondence are identified in Table 130 [atomics.types.pointer].

T* fetch_key(difference_type operand, memory_order order = memory_order_seq_cst) const noexcept;

Requires: T shall be an object type, otherwise the program is ill-formed

Effects: Atomically replaces the value referenced by *ptr with the result of the computation applied to the value referenced by *ptr and the given operand. Memory is affected according to the value of order. These operations are atomic read-modify-write operations [intro.races].

Returns: Atomically, the value referenced by *ptr immediately before the effects.

Remarks: The result may be an undefined address, but the operations otherwise have no undefined behavior.

T* operator op=(difference_type operand) const noexcept;

Effects: Equivalent to: return fetch_key(operand) op operand;

Add a new subsubsection [atomics.ref.memberops] following the [atomics.ref.pointer] subsubsection

Member operators of atomic_ref common to integers and pointers to objects

T* operator++(int) const noexcept;

Effects: Equivalent to: return fetch_add(1);

T* operator--(int) const noexcept;

Effects: Equivalent to: return fetch_sub(1);

T* operator++() const noexcept;

Effects: Equivalent to: return fetch_add(1) + 1;

T* operator--(int) const noexcept;

Effects: Equivalent to: return fetch_sub(1) - 1;

9. Feature Testing

The __cpp_lib_atomic_ref feature test macro should be added.

References

Informative References

[N4727]
Richard Smith. Working Draft, Standard for Programming Language C++. URL: https://wg21.link/n4727
[P0019r3]
H. Carter Edwards, Hans Boehm, Olivier Giroux, James Reus. Atomic View. 14 October 2016. URL: https://wg21.link/p0019r3
[P0019r4]
H. Carter Edwards, Hans Boehm, Olivier Giroux, James Reus. Atomic View. 9 November 2016. URL: https://wg21.link/p0019r4
[P0019r5]
H. Carter Edwards, Hans Boehm, Olivier Giroux, James Reus. Atomic View. URL: https://wg21.link/p0019r5
[P0019r6]
H. Carter Edwards, Hans Boehm, Olivier Giroux, James Reus. Atomic View. URL: https://wg21.link/p0019r6
[P0020r5]
H. Carter Edwards, Hans Boehm, Olivier Giroux, JF Bastien, James Reus. Floating Point Atomic. URL: https://wg21.link/p0020r5
[P0558r1]
Resolving atomic named base class inconsistencies. URL: https://wg21.link/p0558r1