P0401R6
Providing size feedback in the Allocator interface

1. Introduction

This is a library paper to describe how size feedback can be used with allocators:

• In the case of std::allocator, the language feature proposed in [P0901R5] could be used.

• In the case of custom allocators, the availability of the traits interface allows standard library containers to leverage this functionality.

2. Motivation

Consider code adding elements to vector:

std::vector<int> v = {1, 2, 3};
// Expected: v.capacity() == 3

// Add an additional element, triggering a reallocation.
v.push_back(4);

Many allocators only allocate in fixed-sized chunks of memory, rounding up requests. Our underlying heap allocator received a request for 12 bytes (3 * sizeof(int)) while constructing v. For several implementations, this request is turned into a 16 byte region.

2.1. Why not realloc

realloc poses several problems problems:

• We have unclear strategies for growing a container. Consider the two options (a and b) outlined in the comments of the example below. Assume, as with the § 2 Motivation example, that there is a 16 byte size class for our underlying allocator.

  std::vector<int> v = {1, 2, 3};
  // Expected: v.capacity() == 3
  
  // Add an additional element.  Does the vector:
  // a) realloc(sizeof(int) * 4), adding a single element
  // b) realloc(sizeof(int) * 6), attempt to double, needing reallocation
  v.push_back(4);
  
  // Add another element.
  // a) realloc(sizeof(int) * 5), fails in-place, requires reallocation
  // b) size + 1 <= capacity, no reallocation required
  v.push_back(5);
  

  Option a requires every append (after the initial allocation) try to realloc, requiring an interaction with the allocator each time.

  Option b requires guesswork to probe the allocator’s true size. By doubling (the typical growth pattern for std::vector in libc++ and libstdc++), we skip past the existing allocation’s size boundary. While we can use a reduced growth rate, doing so requires more interactions with the allocator (calls to realloc), as well as copies (when we cannot resize in-place).

  Both of these options add extra branching on the control flow for growing the buffer.

• When the allocator cannot resize the buffer in-place, it allocates new memory and memcpy's the contents. For non-trivially copyable types, this is disastrous. The language lacks support for reallocating in-place without moving. While this function could be implemented as magic library function, it could not interact with many underlying malloc implementations (libc, in the case of libc++ and libstdc++) used to build operator new, nor could it interact with user-provided replacement functions.

  In-place reallocation has been considered in [N3495] (throwing std::bad_alloc when unable) and [P0894R1] (returning false when unable). This has sometimes been referred to as an "extend" or an "expand" interface. Extending an existing allocation requires a round-trip to the allocator (and its requisite branch/call overheads) to determine whether the allocation could be expanded. As discussed previously, we need to probe for the true size of the backing object (until we find extend fails) until we give up and apply a multiplicative expansion factor.

For fixed-size allocators it makes more sense, and is much simpler for containers to adapt to, if the allocator is able to over-allocate on the initial request and inform the caller how much memory was made available.

2.2. Why not ask the allocator for the size

We could also explore APIs that answer the question: "If I ask for N bytes, how many do I actually get?" [jemalloc] and [TCMalloc] call this nallocx.

• nallocx requires an extra call / size calculation. This requires we duplicate work, as the size calculation is performed as part of allocation itself.

• nallocx impairs telemetry for monitoring allocation request sizes. If a nallocx return value is cached, the user appears to be asking for exactly as many bytes as nallocx said they would get.

See also [P0901R5]'s discussion "nallocx: not as awesome as it looks."

3. Proposal

Wording relative to [N4868].

We propose a free function to return the allocation size simultaneously with an allocation.

Amend [memory.syn]:

namespace std {
  // ... elided ...

  // [allocator.traits], allocator traits
  template<class Alloc> struct allocator_traits;

  template<class Pointer>
  struct allocation_result {
    Pointer ptr;
    size_t count;
  };

  template<class Allocator>
  [[nodiscard] constexpr allocation_result<typename allocator_traits<Allocator>::pointer> allocate_at_least(
    Allocator& a, size_t n);

  // [default.allocator], the default allocator
  template<class T> class allocator;
  template<class T, class U>
    constexpr bool operator==(const allocator<T>&, const allocator<U>&) noexcept;

  // ... elided ...
}

Amend [allocator.traits.types] to ensure structured binding compatibility:

The class template allocation_result has the template parameters, data members, and special members specified above. It has no base classes or members other than those specified.

Insert [allocator.traits.other]:

template<class Allocator>
[[nodiscard]] constexpr allocation_result<typename allocator_traits<Allocator>::pointer>
  allocate_at_least(Allocator& a, size_t n);

• Returns: a.allocate_at_least(n) if that expression is well-formed; otherwise, {a.allocate(n), n}.

Amend [default.allocator.general]:

namespace std {
  template<class T> class allocator {
   public:
    using value_type                             = T;
    using size_type                              = size_t;
    using difference_type                        = ptrdiff_t;
    using propagate_on_container_move_assignment = true_type;
    using is_always_equal                        = true_type;

    constexpr allocator() noexcept;
    constexpr allocator(const allocator&) noexcept;
    template<class U> constexpr allocator(const allocator<U>&) noexcept;
    constexpr ~allocator();
    constexpr allocator& operator=(const allocator&) = default;

    [[nodiscard]] constexpr T* allocate(size_t n);
    [[nodiscard]] constexpr allocation_result<T*> allocate_at_least(size_t n);
    constexpr void deallocate(T* p, size_t n);
  };
}

Amend [allocator.members]:

Except for the destructor, member functions of the default allocator shall not introduce data races as a result of concurrent calls to those member functions from different threads. Calls to these functions that allocate or deallocate a particular unit of storage shall occur in a single total order, and each such deallocation call shall happen before the next allocation (if any) in this order.

[[nodiscard]] constexpr T* allocate(size_t n);

• Mandates: T is not an incomplete type ([basic.types]).
• Returns: A pointer to the initial element of an array of n T.
• Throws: bad_array_new_length if numeric_limits<size_t>::max() / sizeof(T) < n, or bad_alloc if the storage cannot be obtained.
• Remarks: The storage for the array is obtained by calling ::operator new, but it is unspecified when or how often this function is called. This function starts the lifetime of the array object, but not that of any of the array elements.

[[nodiscard]] constexpr allocation_result<T*> allocate_at_least(size_t n);

• Mandates: T is not an incomplete type ([basic.types]).
• Returns: allocation_result<T*>{ptr, count}, where ptr is a pointer to the initial element of an array of count T and count >= n.
• Throws: bad_array_new_length if numeric_limits<size_t>::max() / sizeof(T) < n, or bad_alloc if the storage cannot be obtained.
• Remarks: The storage for the array is obtained by calling ::operator new, but it is unspecified when or how often this function is called. This function starts the lifetime of the array object, but not that of any of the array elements.

constexpr void deallocate(T* p, size_t n);

• Preconditions: p is a pointer value obtained from allocate. n equals the value passed as the first argument to the invocation of allocate which returned p.
• Preconditions:
  • If p is memory that was obtained by a call to allocate_at_least, let ret be the value returned and req be the value passed as the first argument to that call. p is equal to ret.ptr and n is a value such that $$req ≤ n ≤ ret.count .
  • Otherwise, p is a pointer value obtained from allocate. n equals the value passed as the first argument to the invocation of allocate which returned p.
• Effects: Deallocates the storage referenced by p.
• Remarks: Uses ::operator delete, but it is unspecified when this function is called.

Amend [allocator.requirements.general], [tab:cpp17.allocator]. The return type X::pointer is used in the table for consistency with a.allocate but should be resolved according to the LWG issue.

a.allocate(n)	X::pointer	Memory is allocated for an array of n T and such an object is created but array elements are not constructed. [Example 1: When reusing storage denoted by some pointer value p, launder(reinterpret_­cast<T*>(new (p) byte[n * sizeof(T)])) can be used to implicitly create a suitable array object and obtain a pointer to it. — end example] allocate may throw an appropriate exception. [Note 1: If n == 0, the return value is unspecified. — end note] See Note C, below.
a.allocate(n, y)	X::pointer	Same as a.allocate(n). The use of y is unspecified, but it is intended as an aid to locality.	a.allocate(n)
a.allocate_at_least(n)	allocation_result<X::pointer>	Returns: allocation_result<X::pointer>{ptr, count} where ptr is memory allocated for an array of count T and such an object is created but array elements are not constructed, such that count >= n. allocate_at_least may throw an appropriate exception. See Note C, below.	See Note D, below.
a.deallocate(p,n)	(not used)	Preconditions: p is a value returned by an earlier call to allocate that has not been invalidated by an intervening call to deallocate. n matches the value passed to allocate to obtain this memory. Preconditions: If p is memory that was obtained by a call to a.allocate_at_least, let ret be the value returned and req be the value passed as the first argument of that call. p is equal to ret.ptr and n is a value such that $$req ≤ n ≤ ret.count . Otherwise, p is a pointer value obtained from allocate. n equals the value passed as the first argument to the invocation of allocate which returned p. p has not been invalidated by an intervening call to deallocate. Throws: Nothing.

... elided ...

Note C: If n == 0, the return value is unspecified.

Note D: An allocator need not support allocate_at_least, but no default is provided in allocator_traits. If an allocator has an allocate_at_least member, it shall satisify the requirements.

Amend [version.syn]

#define __cpp_lib_allocate_at_least PLACEHOLDER // also in <memory>

4. Design Considerations

4.1. allocate selection

There are multiple approaches here:

• Return a pointer-size pair, as presented.

• Overload allocate and return via a reference parameter. This potentially hampers optimization depending on ABI, as the value is returned via memory rather than a register.

4.2. Size Return Value

In [Prague], LEWG discussed the return value. For compatibility with the existing allocator APIs, which work in units of objects rather than bytes, this proposal chooses to continue to return an integer number of objects.

Additionally, for types with non-trivial alignment requirements, we must allocate storage for objects, rather than bytes, as raw bytes do not convey the appropriate alignment needs to the allocator.

For example: In the std::vector<T> case, many implementations use 3 pointers for representing the state of the container (begin, end, and capacity). If we preserved the precise value returned by the underlying allocator, we may not be able to legally form the capacity pointer. For these implementations, replacing the capacity pointer with a capacity in bytes would be an ABI break.

4.3. deallocate changes

We now require flexibility on the size we pass to deallocate. For container types using this allocator interface, they are faced with the choice of storing both the original size request as well as the provided size (requiring additional auxillary storage), or being unable to reproduce one during the call to deallocate.

As the true size is expected to be useful for the capacity of a string or vector instance, the returned size is available without additional storage requirements. The original (requested) size is unavailable, so we relax the deallocate size parameter requirements for these allocations.

In [P0901R5], we anticipate use cases where the user does not store enough information about the size to return a precise, byte-level count. We may be allocating T's and receive enough storage to allocate 3.5 objects. Requiring callers return the precise size may be impractical (std::vector frequently stores its capacity as a T*) without additionally auxillary storage.

In this paper, we may be similarly unable to return the true size back. In the [Prague] review of this paper, we discussed cases where we rounded down to the multiple of the size being allocated. Alternatively, custom data structures may use allocator to obtain storage, but quantize that. In Abseil’s [Cord], data is stored in chunks with a length-encoding tag at the start of each chunk. These tags quantize the capacity in AllocatedSizeToTag that can be represented to multiples of 16- and 32-bytes. This makes the encoded length more compact, but means we cannot reproduce the size the allocator provided us precisely.

4.4. Interaction with polymorphic_allocator

std::pmr::memory_resource is implemented using virtual functions. Adding new methods, such as the proposed allocate API would require taking an ABI break.

4.5. Zero-Sized Requests

In Prague, LEWG discussed the behavior of allocate_at_least(allocator, 0). This maximizes implementation freedom.

• The return value may not be dereferenceable (even if not nullptr).

• Other implementations, such as a stack allocator, could just return the current offset.

5. Revision History

5.1. R5 → R6

Applied wording feedback from the reflector.

• In [memory.syn] added missing template parameter and [[nodiscard]].

• In [memory.traits.other] added missing template parameter.

• In [allocator.members]:

  • Removed a..

  • Clarified req is the value of the argument to the call.

  • Used math operators, not C++ operators.

• Removed unneeded parentheses around allocate.

• Fixed wording in [tab:cpp17.allocator].

• Inserted line break between "Note C" and "Note D".

• Changed "note D" from must to shall.

5.2. R4 → R5

Applied wording feedback from [LWGNovTelecon].

• Generally, replaced typename T with class T

• In [allocator.traits], removed change as redundant with [memory.syn]

• In [allocator.traits.type]:

  • Fixed insert editing on [allocator.traits.type]

  • Added note (in paper) for structured binding intent of [allocator.traits.type] wording

• In [allocator.traits.other]

  • Moved allocate_at_least to new stable heading, as it is not a member of allocator_traits

  • Fixed otherwise clause of allocate_at_least

  • Added [[nodiscard]] to allocate_at_least

• In [allocator.members], applied math font in deallocate edits

• In [allocator.members], removed shall in deallocate

• In [allocator.members], applied code font to allocate calls

• In [allocator.members], removed structured binding in deallocate text

• In [allocator.members], bulleted cases for clarity.

• In [allocator.requirements.general]

  • Moved Note 1 from allocate to be "Note C" shared with allocate_at_least

  • Moved "Note C" to "Note D" in the defaults column.

  • Removed value from return type for allocate_at_least

  • Added what allocate_at_least actually returns in third column

  • Removed std:: qualification from (now) "Note D"

  • Removed definition of allocate_at_least (the helper) in (now) "Note D" rather than redefining it.

  • Applied modifications from [allocator.members] deallocate to table.

  • Required that if a allocate_at_least member is provided, it must meet the requirements.

5.3. R3 → R4

• Applied wording feedback from Pablo Halpern and David Stone

• Added discussion about changes to deallocation

• Proposed feature test macro

5.4. R2 → R3

Applied LEWG feedback from [Prague].

• Renamed sized_ptr_t to allocation_result, based on discussion in Prague that this library proposal (which returns a count of storage for objects) have a distinct type from the underlying language proposal ([P0901R5], which returns bytes). This was one of the names bikeshedded during the review of P0901R5 ([P0901R5-Review]).

• Renamed n (the returned size) to count, based on discussion in Prague to clarify that is measured in objects, rather than bytes. Additionally, augmented discussion of the deallocate changes for why implementations may not be able to preserve the precise size returned by the underlying allocator.

Poll: We prefer to return number of bytes.

SF	F	N	A	SA
0	0	6	6	5

• Changed the template parameter to pointer type to allow for fancy pointers.

POLL: Change the template parameter to sized_ptr_t to pointer type (to support fancy ptr).

Unanimous consent

• Added design intent and wording that allocation_result be compatible with structured bindings. This wording was adapted from the proposed resolution of [LWG3373].

• Added constexpr

• Removed default template parameter from allocation_result

• Based on LEWG feedback, changed allocate_at_least to deduce return type from the allocator. Since the pointer type is chosen by the allocator, the underlying object type is less relevant. This also removed the default template parameter.

POLL: We want to deduce the object type from the allocator (alloc_least_n [sic]).

SF	F	N	A	SA
4	8	1	0	0

POLL: We like the design, please revise and do initial wording review (with Pablo, and David Stone) send to LWG (tentatively Ready).

SF	F	N	A	SA
6	12	0	0	0

5.5. R1 → R2

Applied LEWG feedback from [Cologne].

Poll: We want a feature like this.

SF	F	N	A	SA
2	9	2	0	0

• As users may specialize std::allocator_traits, this revision moves to a free function std::allocate_at_least as suggested via a LEWG poll.

Poll: Name choice

5	allocate_with_size
4	overallocate
2	allocate_with_oversubscribe
6	allocate_oversize
14	allocate_at_least

Poll: Prefer a struct rather than a requirement of structured bindings.

SF	F	N	A	SA
2	8	0	2	1