P0901.bs

<pre class='metadata'>
Title: Size feedback in operator new
Status: D
Shortname: D0901
Group: WG21
Revision: 11
Editor: Andrew Hunter, andrewhhunter@gmail.com
Editor: Chris Kennelly, Google, ckennelly@google.com
Editor: Thomas K&ouml;ppe, Google DeepMind, tkoeppe@google.com
Date: 2023-06-16
Audience: CWG, EWG, LWG
Abstract: Provide access to actual malloc buffer sizes for users.
URL: http://wg21.link/D0901R11
Markup Shorthands: markdown yes
Default Highlight: C++
</pre>

# Motivation # {#mot}

*Throughout this document, "malloc" refers to the* **implementation**
*of* `::operator new` *both as fairly standard practice for implementers, and to
make clear the distinction between the interface and the implementation.*

Everyone's favorite dynamic data structure, `std::vector`, allocates memory with
code that looks something like this (with many details, like `Allocator`,
templating for non `char`, and exception safety, elided):

<xmp>
void vector::reserve(size_t new_cap) {
  if (capacity_ >= new_cap) return;
  const size_t bytes = new_cap;
  void *newp = ::operator new(new_cap);
  memcpy(newp, ptr_, capacity_);
  ptr_ = newp;
  capacity_ = bytes;
}
</xmp>

Consider the sequence of calls:

<xmp>
std::vector<char> v;
v.reserve(37);
// ...
v.reserve(38);
</xmp>

All reasonable implementations of malloc round sizes, both for alignment
requirements and improved performance.  It is extremely unlikely that malloc
provided us exactly 37 bytes.  We do not need to invoke the allocator
here...except that we don't know that for sure, and to use the 38th byte would
be undefined behavior.  We would like that 38th byte to be usable without a
roundtrip through the allocator.

This paper proposes an API making it safe to use that
byte, and explores many of the design choices (not all of which are obvious
without implementation experience.)

## nallocx: not as awesome as it looks## {#nallocx}

The simplest way to help here is to provide an informative API answering the
question "If I ask for N bytes, how many do I actually get?"  [[jemalloc]] and
[[TCMalloc]] call this `nallocx`.  We can then use that hint as a smarter
parameter for operator new:

<xmp>
void vector::reserve(size_t new_cap) {
  if (capacity_ >= new_cap) return;
  const size_t bytes = nallocx(new_cap, 0);
  void *newp = ::operator new(bytes);
  memcpy(newp, ptr_, capacity_);
  ptr_ = newp;
  capacity_ = bytes;
}
</xmp>

This is a good start, and does in fact work to allow vector and friends to use
the true extent of returned objects.  But there are three significant problems
with this approach.

### nallocx must give a conservative answer ### {#whatval}

While many allocators have a deterministic map from requested size to allocated
size, it is by no means guaranteed that all do.  Presumably they can make a
reasonably good guess, but if two calls to `::operator new(37)` might return 64
and 128 bytes, we'd definitely rather know the right answer, not a conservative
approximation.

### nallocx duplicates work ### {#speed}

Allocation is often a crucial limit on performance.  Most allocators compute
the returned size of an object as part of fulfilling that allocation...but if
we make a second call to `nallocx`, we duplicate all that communication, and
also the overhead of the function call.

### nallocx hides information from malloc ### {#feedback}

The biggest problem (for the authors) is that `nallocx` discards information
malloc finds valuable (the user's intended allocation size.)  That is: in our
running example, malloc normally knows that the user wants 37 bytes (then 38),
but with `nallocx`, we will only ever be told that they want 40 (or 48, or
whatever `nallocx(37)` returns.)

Google's malloc implementation ([[TCMalloc]]) rounds requests to one of a small
(<100) number of *sizeclasses*: we maintain local caches of appropriately sized
objects, and cannot do this for every possible size of object.  Originally,
these sizeclasses were just reasonably evenly spaced among the range they
cover.  Since then, we have used extensive telemetry on allocator use in the
wild to tune these choices. In particular, as we know (approximately) how many
objects of any given size are requested, we can solve a fairly simple
optimization problem to minimize the total internal fragmentation for any
choice of N sizeclasses.

Widespread use of `nallocx` breaks this. By the time TCMalloc's telemetry sees
a request that was hinted by nallocx, to the best of our knowledge the user
*wants* exactly as many bytes as we currently provide them.  If a huge number
of callers wanted 40 bytes but were currently getting 48, we'd lose the ability
to know that and optimize for it.

Note that we can't take the same telemetry from `nallocx` calls: we have no
idea how many times the resulting hint will be used (we might not allocate at
all, or we might cache the result and make a million allocations guided by it.)
We would also lose important information in the stack traces we collect from
allocation sites.

Optimization guided by malloc telemetry has been one of our most effective
tools in improving allocator performance.  It is important that we fix this
issue *without* losing the ground truth of what a caller of `::operator new`
wants.

These three issues explain why we don't believe `nallocx` is a sufficient
solution here.

## after allocation is too late ## {#afteralloc}

Another obvious suggestion is to add a way to inspect the size of an object
returned by `::operator new`. Most mallocs provide a way to do this; [[jemalloc]]
calls it `sallocx`.  Vector would look like:

<xmp>
void vector::reserve(size_t new_cap) {
  if (capacity_ >= new_cap) return;
  void *newp = ::operator new(new_cap);
  const size_t bytes = sallocx(newp);
  memcpy(newp, ptr_, capacity_);
  ptr_ = newp;
  capacity_ = bytes;
}
</xmp>

This is worse than nallocx. It fixes the non-constant size problem, and avoids
a feedback loop, but the performance issue is worse (this is the major issue
*fixed* by [[SizedDelete]]!), and what's worse, the above code invokes UB as
soon as we touch byte `new_cap+1`.  We could in principle change the standard,
but this would be an implementation nightmare.

## realloc's day has passed ## {#realloc}

We should also quickly examine why the classic C API `realloc` is insufficient.

<xmp>
void vector::reserve(size_t new_cap) {
  if (capacity_ >= new_cap) return;
  ptr_ = realloc(ptr_, new_cap);
  capacity_ = new_cap;
}
</xmp>

In principle a realloc from 37 to 38 bytes wouldn't carry the full cost of
allocation. But it's dramatically more expensive than making no call at all.
What's more, there are a number of more complicated dynamic data structures that
store variable-sized chunks of data but are never actually resized. These data
structures still deserve the right to use all the memory they're paying for.

Furthermore, `realloc`'s original purpose was not to allow the use of more bytes
the caller already had, but to (hopefully) extend an allocation in place to
adjacent free space.  In a classic malloc implementation this would actually be
possible...but most modern allocators use variants of slab allocation. Even if
the 65th byte in a 64-byte allocation isn't in use, they cannot be combined into
a single object; it's almost certainly required to be used for the next 64-byte
allocation.  In the modern world, `realloc` serves little purpose.

# Proposal # {#proposal}

## Allocation functions

We propose adding new overloads of `::operator new` (so-called "size-returning allocation
functions") that directly inform the user of the size available to them.  C++
makes `::operator new` replaceable (15.5.4.6), allowing a program to provide
its own version different from the standard library's implementation.

We note immediately that we are deliberately not providing this facility as an unrelated
free function, nor leave it as an implementation detail of `std::allocator`; see [[#alt-allocfn]]
for a detailed discussion.

The new overloads are selected by a tag argument of type `std::return_size_t`,
usually provided as the value `std::return_size`. (This is analogous to
`std::nothrow_t`/`std::nothrow` and to `std::align_val_t`.)

The new overloads return a new type, `std::sized_allocation_t`, which
stores a void pointer that corresponds to the pointer returned by the
existing allocation functions, as well as the new size feedback
information.

## *New-expression*s ## {#new}

In the previous revision, R10, we had proposed to make *new-expression*s be able to use the new
allocation functions, so that a placement *new-expression* of the form `new (std;:return_size)
T` would have a novel return type that includes the size feedback. However, during EWG discussion
it became clear that that is not a useful facility; see [[#alt-newexpr]] for details.

Consequently, we are not proposing to expose size-returning allocation functions to
*new-expression*s. We propose that a hypothetical (placement) *new-expression* that would find a
size-returning allocation function during lookup is ill-formed. (No such expressions can
currently exist.)

# Proposed Wording # {#prop}

We propose wording, relative to [[N4917]]:

* Amend [basic.stc.dynamic.general], paragraph 2:

<blockquote>
The library provides default definitions for the global allocation and deallocation functions.
Some global allocation and deallocation functions are replaceable ([new.delete]).
A C++ program shall provide at most one definition of a replaceable allocation or deallocation function.
Any such function definition replaces the default version provided in the library ([replacement.functions]).
The following allocation and deallocation functions ([support.dynamic]) are implicitly declared
in global scope in each translation unit of a program.

<pre>
  [[nodiscard]] void* operator new(std::size_t);
  [[nodiscard]] void* operator new(std::size_t, std::align_val_t);
  <ins>[[nodiscard]] std::sized_allocation_t operator new(std::size_t, std::return_size_t);</ins>
  <ins>[[nodiscard]] std::sized_allocation_t operator new(std::size_t, std::align_val_t,
                                                     std::return_size_t);</ins>
  
  void operator delete(void*) noexcept;
  void operator delete(void*, std::size_t) noexcept;
  void operator delete(void*, std::align_val_t) noexcept;
  void operator delete(void*, std::size_t, std::align_val_t) noexcept;
  
  [[nodiscard]] void* operator new[](std::size_t);
  [[nodiscard]] void* operator new[](std::size_t, std::align_val_t);
  <ins>[[nodiscard]] std::sized_allocation_t operator new[](std::size_t, std::return_size_t);</ins>
  <ins>[[nodiscard]] std::sized_allocation_t operator new[](std::size_t, std::align_val_t,
                                                       std::return_size_t);</ins>

  void operator delete[](void*) noexcept;
  void operator delete[](void*, std::size_t) noexcept;
  void operator delete[](void*, std::align_val_t) noexcept;
  void operator delete[](void*, std::size_t, std::align_val_t) noexcept;
</pre>

These implicit declarations introduce only the function names `operator new`, `operator new[]`, `operator delete`, and `operator delete[]`.

[Note 2: The implicit declarations do not introduce the names `std`, `std::size_t`, `std::align_val_t`, <ins>`std::return_size_t`,</ins> or any other names that the library uses to declare these names.
Thus, a *new-expression*, *delete-expression*, or function call that refers to one of these functions without importing or including the header `<new>` ([new.syn]) or importing a C++ library module ([std.modules]) is well-formed.
However, referring to `std` or `std::size_t` or `std::align_val_t`<ins> or `std::return_size_t`</ins> is ill-formed unless a standard library declaration ([...]) of that name precedes ([basic.lookup.general]) the use of that name.
— end note]

Allocation and/or deallocation functions may also be declared and defined for any class ([class.free]).

</blockquote>

* Amend [basic.stc.dynamic.allocation], paragraph 1:

<blockquote>
An allocation function that is not a class member function shall belong to the global scope and not have a name with internal linkage.
The return type shall be <ins>`std::sized_allocation_t` ([new.syn]) if the allocation function is a size-returning allocation function and </ins>`void*`<ins> otherwise</ins>.
The first parameter shall have type `std::size_t` ([support.types]).
The first parameter shall not have an associated default argument ([dcl.fct.default]).
The value of the first parameter is interpreted as the requested size of the allocation.
<ins>An allocation function is a *size-returning* allocation function if it has a second parameter of type `std::return_size_t`, or it has a second parameter of type `std::align_val_t` and a third parameter of type `std::return_size_t`.</ins>
An allocation function can be a function template.
Such a template shall declare its return type and first parameter as specified above (that is, template parameter types shall not be used in the return type and first parameter type).
Allocation function templates shall have two or more parameters<ins> as specified above</ins>.
</blockquote>

* Insert [basic.stc.dynamic.allocation], after paragraph 1:

<blockquote>
<ins>A size-returning allocation function returns a value with a member subobject that is a pointer suitable to hold the address of potentially allocated storage, and therefore all allocation functions are said to return a pointer value.</ins>
</blockquote>

* Amend [replacement.functions], paragraph 2:

<pre>
operator new(std::size_t)
operator new(std::size_t, std::align_val_t)
operator new(std::size_t, const std::nothrow_t&)
operator new(std::size_t, std::align_val_t, const std::nothrow_t&)
<ins>
operator new(std::size_t size, std::return_size_t)
operator new(std::size_t size, std::align_val_t al, std::return_size_t)
operator new(std::size_t size, std::return_size_t, std::nothrow_t)
operator new(std::size_t size, std::align_val_t al, std::return_size_t, std::nothrow_t)
</ins>
[...]

operator new[](std::size_t)
operator new[](std::size_t, std::align_val_t)
operator new[](std::size_t, const std::nothrow_t&)
operator new[](std::size_t, std::align_val_t, const std::nothrow_t&)
<ins>
operator new[](std::size_t size, std::return_size_t)
operator new[](std::size_t size, std::align_val_t al, std::return_size_t)
operator new[](std::size_t size, std::return_size_t, std::nothrow_t)
operator new[](std::size_t size, std::align_val_t al, std::return_size_t, std::nothrow_t)
</ins>
[...]

</pre>

* Amend header `<new>` synopsis in [new.syn]:

<pre>
namespace std {
  class bad_alloc;
  class bad_array_new_length;

  struct destroying_delete_t {
    explicit destroying_delete_t() = default;
  };
  inline constexpr destroying_delete_t destroying_delete{};
<ins>
  // global operator new control
  struct return_size_t {
    explicit return_size_t() = default;
  };
  inline constexpr return_size_t return_size{};

  struct sized_allocation_t {
    void *ptr;
    size_t bytes;
  };
</ins>
  enum class align_val_t : size_t {};

  [...]
}

[[nodiscard]] void* operator new(std::size_t size);
[[nodiscard]] void* operator new(std::size_t size, std::align_val_t alignment);
[[nodiscard]] void* operator new(std::size_t size, const std::nothrow_t&) noexcept;
[[nodiscard]] void* operator new(std::size_t size, std::align_val_t alignment,
                                 const std::nothrow_t&) noexcept;
<ins>
[[nodiscard]] std::sized_allocation_t operator new(
  std::size_t size, std::return_size_t);
[[nodiscard]] std::sized_allocation_t operator new(
  std::size_t size, std::align_val_t alignment, std::return_size_t);
[[nodiscard]] std::sized_allocation_t operator new(
  std::size_t size, std::return_size_t, std::nothrow_t) noexcept;
[[nodiscard]] std::sized_allocation_t operator new(
  std::size_t size, std::align_val_t alignment, std::return_size_t,
  std::nothrow_t) noexcept;
</ins>
[...]

[[nodiscard]] void* operator new[](std::size_t size);
[[nodiscard]] void* operator new[](std::size_t size, std::align_val_t alignment);
[[nodiscard]] void* operator new[](std::size_t size, const std::nothrow_t&) noexcept;
[[nodiscard]] void* operator new[](std::size_t size, std::align_val_t alignment,
                                   const std::nothrow_t&) noexcept;
<ins>
[[nodiscard]] std::sized_allocation_t operator new[](
  std::size_t size, std::return_size_t);
[[nodiscard]] std::sized_allocation_t operator new[](
  std::size_t size, std::align_val_t alignment, std::return_size_t);
[[nodiscard]] std::sized_allocation_t operator new[](
  std::size_t size, std::return_size_t, std::nothrow_t) noexcept;
[[nodiscard]] std::sized_allocation_t operator new[](
  std::size_t size, std::align_val_t alignment, std::return_size_t,
  std::nothrow_t) noexcept;
</ins>
[...]

[[nodiscard]] void* operator new  (std::size_t size, void* ptr) noexcept;
[[nodiscard]] void* operator new[](std::size_t size, void* ptr) noexcept;
void operator delete  (void* ptr, void*) noexcept;
void operator delete[](void* ptr, void*) noexcept;
</pre>

* Insert [new.delete.single], after paragraph 4:

<blockquote>
<pre>
<ins>
[[nodiscard]] std::sized_allocation_t ::operator new(
  std::size_t size, std::return_size_t);
[[nodiscard]] std::sized_allocation_t ::operator new(
  std::size_t size, std::align_val_t alignment, std::return_size_t);
</ins>
</pre>

*   <ins>*Effects*: Same as above, except these are called by a placement version of
    a *new-expression* when a C++ program prefers the size-returning
    allocation function.</ins>
*   <ins>*Replaceable*: A C++ program may define functions with either of these
    function signatures, and thereby displace the default versions defined by
    the C++ standard library.</ins>
*   <ins>*Required behavior*: Return a `sized_allocation_t` whose `ptr` member
    represents the address of a region of *N* bytes of suitably aligned storage
    ([basic.stc.dynamic]) for some <math>*N* >= *size*</math>, and whose `bytes` member is
    *N*, or else throw a `bad_alloc` exception.  This requirement is binding on
    any replacement versions of these functions.</ins>
*   <ins>*Default behavior*: Returns `std::sized_allocation_t{operator new(size), N}` and
    `std::sized_allocation_t{operator new(size, alignment), N}` respectively. If a
    user-provided operator new is invoked directly or indirectly, N is `size`.</ins>

</blockquote>

* Insert [new.delete.single], after paragraph 9:

<blockquote>
<pre>
<ins>
[[nodiscard]] std::sized_allocation_t ::operator new(
  std::size_t size, std::return_size_t, std::nothrow_t) noexcept;
[[nodiscard]] std::sized_allocation_t ::operator new(
  std::size_t size, std::align_val_t alignment, std::return_size_t, std::nothrow_t) noexcept;
</ins>
</pre>

*   <ins>*Effects*: Same as above, except these are called by a placement version of
    a *new-expression* when a C++ program prefers the size-returning
    allocation function and a null pointer result as an error indication.</ins>
*   <ins>*Replaceable*: A C++ program may define functions with either of these
    function signatures, and thereby displace the default versions defined by
    the C++ standard library.</ins>
*   <ins>*Required behavior*: Return a `sized_allocation_t` whose `ptr` member represents the
    address of a region of *N* bytes of suitably aligned storage
    ([basic.stc.dynamic]) for some <math>*N* >= *size*</math>, and whose `bytes` member is *N*,
    or else return `std::sized_allocation_t{nullptr, 0}`.  Each of these nothrow
    versions of `operator new` returns a pointer obtained as if acquired from
    the (possibly replaced) corresponding non-placement function.  This
    requirement is binding on any replacement versions of these functions.</ins>
*   <ins>*Default behavior*: Returns `std::sized_allocation_t{operator new(size), N}` and
    `std::sized_allocation_t{operator new(size, alignment), N}` respectively. If a
    user-provided operator new is invoked directly or indirectly, N is `size`.
    If the call to `operator new` throws, returns
    `std::sized_allocation_t{nullptr, 0}`.</ins>

</blockquote>

* Amend [new.delete.single], paragraph 11:

<pre>
void operator delete(void* ptr) noexcept;
void operator delete(void* ptr, std::size_t size) noexcept;
void operator delete(void* ptr, std::align_val_t alignment) noexcept;
void operator delete(void* ptr, std::size_t size, std::align_val_t alignment) noexcept;
</pre>
[...]
*   If the `alignment` parameter is not present, `ptr` shall have been
    returned by an allocation function without an alignment parameter. If
    present, the `alignment` argument shall equal the alignment argument passed
    to the allocation function that returned `ptr`. If present, the `size` argument
    shall equal the `size` argument passed to the allocation function that
    returned `ptr`<ins>, if `ptr` was not allocated by a size-returning
    allocation function.  If present, the `size` argument shall satisfy <math>*bytes* >=
    *size* >= *requested*</math> if `ptr` was allocated by a size-returning allocation
    function, where `bytes` is the size returned in `std::sized_allocation_t`
    and `requested` is the size argument passed to the allocation
    function</ins>.

[...]
</blockquote>

* Insert [new.delete.array], after paragraph 4:

<blockquote>
<pre>
<ins>
[[nodiscard]] std::sized_allocation_t ::operator new[](
  std::size_t size, std::return_size_t);
[[nodiscard]] std::sized_allocation_t ::operator new[](
  std::size_t size, std::align_val_t alignment, std::return_size_t);
</ins>
</pre>

*   <ins>*Effects*: Same as above, except these are called by a placement version of
    a *new-expression* when a C++ program prefers the size-returning
    allocation function.</ins>
*   <ins>*Replaceable*: A C++ program may define functions with either of these
    function signatures, and thereby displace the default versions defined by
    the C++ standard library.</ins>
*   <ins>*Required behavior*: Return a `sized_allocation_t` whose `ptr` member
    represents the address of a region of *N* bytes of suitably aligned storage
    ([basic.stc.dynamic]) for some <math>*N* >= *size*</math>, and whose `bytes` member is
    *N*, or else throw a `bad_alloc` exception.  This requirement is binding on
    any replacement versions of these functions.</ins>
*   <ins>*Default behavior*: Returns `std::sized_allocation_t{operator new[](size), N}` and
    `std::sized_allocation_t{operator new[](size, alignment), N}` respectively. If a
    user-provided operator new is invoked directly or indirectly, N is `size`.</ins>

</blockquote>

* Insert [new.delete.array], after paragraph 8:

<blockquote>
<pre>
<ins>
[[nodiscard]] std::sized_allocation_t ::operator new[](
  std::size_t size, std::return_size_t, std::nothrow_t) noexcept;
[[nodiscard]] std::sized_allocation_t ::operator new[](
  std::size_t size, std::align_val_t alignment, std::return_size_t, std::nothrow_t) noexcept;
</ins>
</pre>

*   <ins>*Effects*: Same as above, except these are called by a placement
    version of a *new-expression* when a C++ program prefers the
    size-returning allocation function and a null pointer result as an error
    indication.</ins>
*   <ins>*Replaceable*: A C++ program may define functions with either of these
    function signatures, and thereby displace the default versions defined by
    the C++ standard library.</ins>
*   <ins>*Required behavior*: Return a `sized_allocation_t` whose `ptr`
    member represents the address of a region of *N* bytes of suitably aligned
    storage ([basic.stc.dynamic]) for some <math>*N* >= *size*</math>, and whose `bytes`
    member is *N*, or else return `std::sized_allocation_t{nullptr, 0}`.
    This requirement is binding on any replacement versions of these
    functions.</ins>
*   <ins>*Default behavior*: Returns `std::sized_allocation_t{operator
    new[](size), N}` and `std::sized_allocation_t{operator new[](size,
    alignment), N}` respectively. If a user-provided operator new is invoked
    directly or indirectly, N is `size`.  If the call to `operator new[]`
    throws, returns `std::sized_allocation_t{nullptr, 0}`.</ins>

</blockquote>

* Amend [new.delete.array], paragraph 10:

<pre>
void operator delete[](void* ptr) noexcept;
void operator delete[](void* ptr, std::size_t size) noexcept;
void operator delete[](void* ptr, std::align_val_t alignment) noexcept;
void operator delete[](void* ptr, std::size_t size, std::align_val_t alignment) noexcept;
</pre>
[...]

*   If the `alignment` parameter is not present, `ptr` shall have been
    returned by an allocation function without an `alignment` parameter. If
    present, the `alignment` argument shall equal the `alignment` argument
    passed to the allocation function that returned `ptr`. If present, the
    `size` argument shall equal the `size` argument passed to the allocation
    function that returned `ptr`<ins>, if `ptr` was not allocated by a
    size-returning allocation function.  If present, the `size` argument
    shall satisfy <math>*bytes* >= *size* >= *requested*</math> if `ptr` was allocated by a
    size-returning allocation function, where `bytes` is the size returned in
    `std::sized_allocation_t` and `requested` is the size argument passed
    to the allocation function</ins>.

[...]
</blockquote>

* Amend Table 19 ("Feature-test macros") [cpp.predefined]

<blockquote>
<table>
<tr><th>Name</th><th>Value</th></tr>
<tr><td><ins>`__cpp_size_returning_new`</ins></td><td><ins>PLACEHOLDER DATE</ins></td></tr>
</table>
</blockquote>

# Implementation, Deployment and Performance Experience # {#impldepl}

## Implementability

The fundamental size feedback logic is a shipping part of TCMalloc ([[MallocExtension]]),
as a set of functions such as:
<pre>
tcmalloc::sized_ptr_t tcmalloc_size_returning_operator_new(std::size_t size);
</pre>

We implemented the new `::operator new` overload in Clang. This
presented no noteworthy complication or obstacle. An `::operator new`
whose return type is not `void*` is novel, but entirely implementable.

Building a few large applications from Google's codebase with the
modified version of Clang showed nothing unexpected and continued
working.

## Deployment

We modified the libc++ standard library implementation with minimal effort to
use that function. (The implementation has a convenient
<a href="https://github.com/llvm/llvm-project/blob/ed8dae9a438dd8b76a9760506fb7333ce7057d96/libcxx/include/__memory/allocate_at_least.h#L55">central point</a>
at which the underlying allocation function is called.)
We were able to build a number of significant production binaries of the Google
codebase without problems.

As above, building a few large applications from Google's codebase
succeeded without incident.

## Performance

At Google, we modified libc++'s `basic_string` to use TCMalloc's size-returning allocation
function. This resulted in savings on heap allocations that we will describe below. In summary,
the proposed feature allows vectors and strings to grow more efficiently. While it is hard to
spot these effects in real-world benchmarks, where they are within the noise, we performed
micro-benchmarks and traced allocations to show that string allocations are reduced by about
4%. For the application deployment in Google's production systems, we extrapolate fleet-wide
memory savings from strings of about 0.2%.

One source of inefficiencies is the interplay between malloc and the implementation of
`basic_string`. The libc++ implementation of `basic_string` always allocates memory in multiples
of 16 bytes. The glibc malloc (ptmalloc2) allocates blocks of memory of size
24&nbsp;+&nbsp;<em>n</em>&nbsp;&times;&nbsp;16. For example, a call `reserve(35)` would make
`basic_string` allocate a rounded-up value of 48, but malloc would allocate 56 bytes. With size
feedback, `basic_string` would ask for "at least 35 bytes" and receive 40 bytes from malloc.
With TCMalloc the situation is different, but savings are still possible: its size classes are
8, 16, 32, 48, 64, followed by 8-byte increments up to 128. In this case, size feedback would
allow `basic_string` to hit every size bucket instead of only every even one for strings of
length 64&#x2012;128.

TCMalloc's size feedback is used in Google's implementation of
<a href="https://github.com/abseil/abseil-cpp/blob/1feab4fff90f904518e66cf80971063486fbc984/absl/strings/cord.h#L19-L23">`absl::Cord`</a>
and of <a href="https://protobuf.dev/reference/cpp/api-docs/google.protobuf.repeated_field/">`google::protobuf::RepeatedField`</a>.

# Alternative Designs Considered # {#alternatives}

## Parameters and return types

Another signature we could use would be:

<xmp>
enum class return_size_t : std::size_t {};
void* ::operator new(std::size_t size, std::return_size_t&);
</xmp>

(and so on.)  This is slightly simpler to read as a signature, but arguably
worse in usage:

<xmp>
std::tie(obj.ptr, obj.size) = ::operator new(37, std::return_size_t{});

// ...vs...

// Presumably the object implementation wants to contain a size_t,
// not a return_size_t.
std::return_size_t rs;
obj.ptr = ::operator new(37, rs);
obj.size = rs;
</xmp>

More importantly, this form is less efficient. In practice, underlying malloc
implementations provide actual definitions of `::operator new` symbols which
are called like any other function.  Passing a reference parameter requires us
to actually return the size via memory.

*   Linux ABIs support returning at least two scalar values in registers (even if
    they're members of a trivially copyable struct) which can be dramatically
    more efficient.
*   The [[MicrosoftABI]] returns large types by pointer, but this is no worse
    than making the reference parameter an inherent part of the API.

Whether we use a reference parameter or a second returned value, the
interpretation is the same.

## Why `operator new`? ## {#alt-allocfn}

Given the added complexity of the `operator new` overload set, one might ask why we want to
expose this facility as such an overload, rather than something else. Concretely, we could
provide a completely unrelated free function, or we could ask that the new behavior be exposed
via `std::allocator<T>::allocate_at_least`.

Those alternatives fall short, however, since the `operator new` allocation functions have a
special place in C++ that the alternatives do not have: They are *replaceable* and can be chosen
by the application owner:

*   Compared to an unrelated free function: an unrelated free function would not be used by
    existing code, in particular, code that may be outside the user's control. The replaceable
    allocation function, by contrast, can affect and benefit existing code. Moreover, it would
    require additional integration to make the free function interoperate with `operator delete`,
    whereas we have an existing framework for matching `operator new` and `operator delete`.
    Finally, `operator new` enjoys various special properties regarding implicit lifetimes
    which would need to be added explicitly to an unrelated free function.

*   Compared to `std::allocator`: The implementation of `std::allocator` is generally under the
    control of the standard library vendor, not the user. The vendor is not going to select any
    one particular malloc implementation in their `std::allocator`. In fact, the user may want
    to use different malloc implementations in different projects. By contrast, a replaceable
    allocation function opens the choice of malloc implementation to the user on a
    program-by-program basis. (Note that `std::allocator` is specified to use `::operator new`,
    and while no further details (such as choice of overload) are specified, it is reasonable to
    assume that a size-returning `operator new` would be picked up by
    `std::allocator::allocate_at_least` as a matter of quality-of-implementation.)

Finally, a suggestion that came up during the review of R10 was that we could aim at a *lower*
level and propose a new `malloc`-like facility to both C and C++. While this would undoubtedly
be useful, we believe that it does not replace `operator new`. That is because `operator new` is
generally how one "obtains memory in C++"; it enjoys special properties in the language,
e.g. regarding implicit lifetimes, and also with regards to elision or merging of allocations.

As a data point, the codebase at Google uses TCMalloc and its `operator new` interface, not its
`malloc` interface, which are entirely different. A significant performance gain comes from the
use of `operator new` and sized-`operator delete`, compared to the `malloc`/`free` interface.
(The vast majority of memory allocations in that codebase come from `operator new` and not from
`malloc`.) Exposing the size feedback as part of this established framework makes the performance
benefits available in ways that a less integrated facility would not easily be able to do.

## New expressions ## {#alt-newexpr}

We considered previously to have "size-returning *new-expression*s that expose the excess
allocation space resulting from the allocation that's done as part of the
*new-expression*. However, it is exceedingly awkward to make use of the excess allocation space,
since the lifetime of the object(s) created by the `new-expression` would be distinct from
whatever objects one later creates manually in the excess space.

The following example was given in the previous revision:

    <xmp>
    auto [p, sz] = new (std::return_size) T[5];
    for (int i = 5; i < sz / sizeof(T); i++) {
      new (p[i]) T;
    }
    for (int i = 0; i < sz / sizeof(T); i++) {
      p[i].DoStuff();
    }
    for (int i = 5; i < sz / sizeof(T); i++) {
      p[i].~T();
    }
    delete[] p;
    </xmp>

The example shows how the lifetime of the "requested" objects is managed by the delete
expression, but the lifetime of the manually constructed excess objects has to be handled
manually. This is awkward, easy to misuse, and we do not see any need for such a facility.

Non-array *new-expression*s would be similarly useless.

One conceivable use case, which we ultimately will not pursue, is perhaps worth describing: for
trivial types, such as `char`, one could imagine a chain of expressions `new (std::return_size)
char[N]`. These could successively return excess storage, but in addition, the compiler would be
allowed to merge the allocations, because of the special allowances made to how
*new-expression*s (don't) call allocation functions. This behavior is not achievable with
`operator new` function calls alone.

We also note in passing that in R10 `std::sized_allocation_t` was a template, whose template
parameter would take the type used in the *new-expression*. There were implementer objections to
having a core language facility having to instantiate templates.

### Which kind of "size" ### {#alt-newsize}

We considered alternatives for returning the size.

*   We could return two pointers, the initial object and one past the end of
    the array (minus the array allocation overhead).

    <xmp>
    auto [start, end] = new (std::return_size) T[5];
    for (T* p = start + 5; p != end; p++) {
        new (p) T;
    }
    for (T* p = start; p != end; p++) {
        p->DoStuff();
    }
    for (T* p = start + 5; p != end; p++) {
        p->~T();
    }
    delete[] start;
    </xmp>

    The pair of pointers provides convenience for use with iterator-oriented
    algorithms.  The problem we foresee is that a size-returning allocation
    function may not provide a size that is an appropriate multiple of
    `sizeof(T)` (or fail to be a multiple of `alignof(T)`, a possible, but
    unlikely scenario).  This imposes a need for more extensive logic around
    the *new-expression* in handling the returned allocation size.

*   We could return the size in units of `T`, this leads to an inconsistency
    between the expected usage for `new` and `new[]`:

    *   For `new`, we may only end up fitting a single `T` into an allocator size
        quanta, so the extra space remains unusable.  If we can fit multiple `T`
        into a single allocator size quanta, we now have an array from what was a
        scalar allocation site.  This cannot be foreseen by the compiler as
        `::operator new` is a replaceable function.

    *   For `new[]`, the size in units of `T` can easily be derived from the
        returned size in bytes.

*   We could pass the size in units of `T` or bytes to the constructor of `T`:

    *   For `new`, this is especially useful for tail-padded arrays, but
        neglects default-initialized `T`.

    *   For `new[]`, a common use case is expected to be the allocation of
        arrays of `char`, `int`, etc.  The size of the overall array is
        irrelevant for the individual elements.

*   We could return the size via a reference parameter:

    <xmp>
    std::return_end<T> end;
    T* p = new (end) T[5];
    for (T* p = start + 5; p != end; p++) {
        new (p) T;
    }
    for (T* p = start; p != end; p++) {
        p->DoStuff();
    }
    for (T* p = start + 5; p != end; p++) {
        p->~T();
    }
    </xmp>

    or, demonstrated with bytes:

    <xmp>
    std::return_size_t size;
    T* p = new (size) T[5];
    for (int i = 5; i < size / sizeof(T); i++) {
      new (p[i]) T;
    }
    for (int i = 0; i < size / sizeof(T); i++) {
      p[i].DoStuff();
    }
    for (int i = 5; i < size / sizeof(T); i++) {
      p[i].~T();
    }
    delete[] p;
    </xmp>
    (Casts omitted for clarity.)

    As discussed for `::operator new` in [[#prop]], a reference parameter poses
    difficulties for optimizers and involves returning the size via
    memory (depending on ABI).

For `new[]` expressions, we considered alternatively initializing the returned
(`sz / sizeof(T)`) number of elements.

*   This would avoid the need to explicitly construct / destruct the elements
    with the additional returned space (if any).

    The *new-initializer* is invoked for the returned number of elements,
    rather than the requested number of elements.  This allows `delete[]` to
    destroy the correct number of elements (by storing `sz / sizeof(T)` in the
    array allocation overhead).
*   The presented proposal (leaving this space uninitialized) was chosen for
    consistency with `new`.

# Discussion # {#discuss}

## *How* many `::operator new`'s? ## {#splode}

It is unfortunate that we have so many permutations of `::operator new`--eight
seems like far more than we should really need!  But there really isn't any
significant runtime cost for having them.  Use of raw calls to `::operator new`
is relatively rare:  It's a *building block* for low-level libraries, allocators
([[P0401R4]]), and so on, so the cognitive burden on C++ users is low.

The authors have considered other alternatives to the additional overloads.  At
the Jacksonville meeting, EWG suggested looking at parameter packs.

*   Parameter packs do not reduce the number of symbols introduced.
    Implementers still need to provide implementations each of the n overloads.
*   Retrofitting parameter packs leaves us with *more* mangled variants.
    Implementers need to provide both the legacy symbols as well as the
    parameter pack-mangled symbols.

The authors have also considered APIs where all parameters are passed, thereby
requiring a single new overload.  This adds further overhead for
implementations, as it moves compile-time decisions (is the alignment at or
below the minimum guaranteed by `operator new`) into runtime ones.

The alternative to modifying the handling of *new-expressions* invoking
deallocation functions (when an exception is thrown) would require additional
overloads for `operator delete` / `operator delete[]` whose sole purpose would
be to accept and discard the `std::return_size_t`.

## Implementation difficulty ## {#trivial}

It's worth reiterating that there's a perfectly good trivial implementation of these
functions:

<xmp>
std::sized_allocation_t ::operator new(std::size_t n, std::return_size_t) {
  return {::operator new(n), n};
}
</xmp>

Malloc implementations are free to properly override this with a more impactful
definition, but this paper poses no significant difficulty for toolchain
implementers.

Implementation Experience:
*   TCMalloc has developed an implementation opensourced on GitHub
    ([[MallocExtension]]).  While this requires mapping from an integer size
    class to the true number of bytes, combining this lookup with the
    allocation is more efficient as we avoid recomputing the sizeclass itself
    (given a request) or deriving it from the object's address.
*   jemalloc is prototyping a `smallocx` function providing a C API for this
    functionality [[smallocx]].

## Interaction with Sized Delete ## {#sizeddelete}

For allocations made with `sized_allocation_t`-returning `::operator new`, we need to
relax `::operator delete`'s size argument (16.6.2.1 and 16.6.2.2).  For
allocations of `T`, the size quanta used by the allocator may not be a multiple
of `sizeof(T)`, leading to both the original and returned sizes being
unrecoverable at the time of deletion.

Consider the memory allocated by:
<xmp>
using T = std::aligned_storage<16, 8>::type;

std::vector<T> v(4);
</xmp>
The underlying heap allocation is made with `::operator new(64,
std::return_size_t)`.

*   The memory allocator may return a 72 byte object:  Since there is no `k`
    such that `sizeof(T) * k = 72`, we can't provide that value to `::operator
    delete(void*, size_t)`.  The only option would be storing 72 explicitly,
    which would be wasteful.
*   The memory allocator may instead return an 80 byte object (5 `T`'s):  We
    now cannot represent the original request when deallocating without
    additional storage.

For allocations made with
<xmp>
std::tie(p, m) = ::operator new(n, std::return_size_t{});
</xmp>

we permit `::operator delete(p, s)` where <math>*n* <= *s* <= *m*</math>.

This behavior is consistent with [[jemalloc]]'s `sdallocx`, where the
deallocation size must fall between the request (`n`) and the actual allocated
size (`m`) inclusive. It is also consistent with the standard library's
allocator interface (as amended by [[P0401R4]]).

## Advantages ## {#advantages}

It's easy to see that this approach nicely solves the problems posed by other methods:

*    We pay almost nothing in speed to return an actual-size parameter.  For
     TCMalloc and jemalloc, this is typically a load from to map from
     *sizeclass* to size.  This cost is strictly smaller than with `nallocx` or
     the like, as that same translation must be performed in addition to the
     duplicative work previously discussed.
*    We are told exactly the size we have, without risk of UB.  We can avoid
     subsequent reallocations when growing to a buffer to an already-allocated
     size.
*    Allocator telemetry knows actual request sizes exactly.

## Naming ## {#naming}

The library support type is named  `sized_allocation_t`, based on
LEWG's naming suggestions and for consistency (in choice of `_t`) with the
other allocation library support types (`size_t`, `align_val_t`, `nothrow_t`,
etc.).  We expect this to be spelled rarely.

Its members are `ptr` and `bytes`.  `ptr` for the memory returned is an
intuitive name.  `bytes` conveys units in its name (`bytes`).  Since new
expressions can return additional storage under this proposal, this
distinguishes it from returning the amount of storage available for *whole*
objects.  This contrasts with [[P0401R4]], which does return storage for whole
objects and whose support type's field is named `count`.

# Related work # {#rel}

[[P0401R4]] considers this problem at the level of the `Allocator` concept.
Ironically, the lack of the above API was one significant problem in its
original revision: how could an implementation of `std::allocator` provide the
requested feedback in a way that would work with any underlying malloc
implementation? See also [[#alt-allocfn]].

# History # {#history}

## R10 &rarr; D11 ## {#R11}

Support for *new-expression*s is removed, and the class template
`sized_return_t` is changed to a non-template class with a `void*`
member.

Rationale added why the facility needs to be provided as an allocation
function, rather than some unrelated free function or as a detail of
`std::allocator`.

Reports on implementation and deployment experience and impact on
runtime allocations have been added.

## R9 &rarr; R10 ## {#R10}

A detailed design of *new-expression*s has been added. Previously,
changes to *new-expression*s had only been outlined at a high level.
As a result, previously suggested changes to overload resolution rules
for finding the allocation function (which was allowed to fall back
to a non-size-returning allocation functon, with unclear consequences)
have been removed. The relation between *new-expression*s and allocation
functions is now described explicitly.

Sections are reorganized: alternative designs are now separate from
a discussion of the chosen design, and an expanded section on *new-expressions*
has been moved up before the proposed wording.

## R8 &rarr; R9 ## {#R9}

CWG reviewed [[P0901R8]] via [[CWG2021Telecon]].

Wording changes have been made to the proposal based on CWG feedback and
assistance from Ryan McDougall.

## R7 &rarr; R8 ## {#R8}

LEWG reviewed [[P0901R6]] via telecon.

<blockquote>
Poll: Send P0901R6, after changing `p` to `ptr` and `n` to `bytes` in
`sized_allocation_t`, to electronic ballot to be forwarded to CWG.

<table>
<tr><th>SF</th><th>F</th><th>N</th><th>A</th><th>SA</th></tr>
<tr><td>4</td><td>11</td><td>0</td><td>0</td><td>0</td></tr>
</table>
</blockquote>

*   Applied member name feedback from LEWG to rename `sized_allocation_t`'s
    members from `p` to `ptr` and `n` to `bytes`.  Added naming rationale section.
*   Update reference to revised [[P0401R4]].
*   Refined wording in proposal.

## R6 &rarr; R7 ## {#R7}

*   Removed extraneous `std::` typo

## R5 &rarr; R6 ## {#R6}

LEWG reviewed [[P0901R5]] library support types at [[Prague]].

*   LEWG discussed the similarity to `std::span`.  Participants in the
    discussion were concerned about the potential library layering issue
    (`<new>` layers below practically everything) and that ownership is
    conveyed in the return value.  In contrast, `std::span` is non-owning.
*   LEWG discussed whether the type needed a name.  It needs to be named to
    allow users to replace `operator new` and spell the type.

LEWG took an approval poll of different names

<blockquote>
<table>
<tr><td>5</td><td>`sized_ptr_t`</td></tr>
<tr><td>2</td><td>`sized_ptr`</td></tr>
<tr><td>3</td><td>`memblock_t`</td></tr>
<tr><td>7</td><td>`memblock`</td></tr>
<tr><td>0</td><td>`memspan_t`</td></tr>
<tr><td>2</td><td>`memspan`</td></tr>
<tr><td>8</td><td>`alloc_size_t`</td></tr>
<tr><td>4</td><td>`alloc_size`</td></tr>
<tr><td>5</td><td>`alloc_t`</td></tr>
<tr><td>0</td><td>`alloc`</td></tr>
<tr><td>13</td><td>`sized_allocation_t`</td></tr>
<tr><td>17</td><td>`sized_allocation`</td></tr>
<tr><td>0</td><td>`alloc_span_t`</td></tr>
<tr><td>2</td><td>`alloc_span`</td></tr>
<tr><td>9</td><td>`allocation_t`</td></tr>
<tr><td>9</td><td>`allocation`</td></tr>
<tr><td>16</td><td>`alloc_result_t`</td></tr>
<tr><td>14</td><td>`alloc_result`</td></tr>
<tr><td>12</td><td>`allocation_result_t`</td></tr>
<tr><td>16</td><td>`allocation_result`</td></tr>
</table>
</blockquote>

Based on this poll, LEWG directed the paper authors to consider
`sized_allocation`, `alloc_result_t`, and `allocation_result`.

LEWG additionally polled on whether the type should have a `_t` suffix.

<blockquote>
<table>
<tr><th>SF</th><th>F</th><th>N</th><th>A</th><th>SA</th></tr>
<tr><td>0</td><td>6</td><td>7</td><td>8</td><td>1</td></tr>
</table>
</blockquote>

Based on this feedback, the authors have chosen `sized_allocation_t`, based on
LEWG's naming suggestions and for consistency (in choice of `_t`) with the
other allocation library support types (`size_t`, `align_val_t`, `nothrow_t`,
etc.).  We expect this to be spelled rarely.

## R4 &rarr; R5 ## {#R5}

EWG reviewed P0901R4 at [[Cologne]].

<blockquote>
Poll: P0901R4 as presented, forward to LEWG for C++23, *not* C++20.

<table>
<tr><th>SF</th><th>F</th><th>N</th><th>A</th><th>SA</th></tr>
<tr><td>2</td><td>11</td><td>14</td><td>2</td><td>0</td></tr>
</table>
</blockquote>

*   Fixed typos in examples.
*   Added proposed feature test macro.
*   Added LEWG audience for library support type names.

## R3 &rarr; R4 ## {#R4}

*   Update reference to revised [[P0401R1]].

## R2 &rarr; R3 ## {#R3}

*   Added proposed wording.
*   For newly added allocation functions, `std::nothrow_t` is passed by value,
    rather than reference.

## R1 &rarr; R2 ## {#R2}

Applied feedback from San Diego Mailing

*   Moved from passing `std::return_size_t` parameter by reference to by value.
    For many ABIs, this is more optimizable and to the authors' knowledge, no
    worse on any other.
*   Added rationale for not using parameter packs for this functionality.

## R0 &rarr; R1 ## {#R1}

Applied feedback from [[JacksonvilleMinutes]].

*   Clarified in [[#prop]] the desire to leverage the existing "replacement
    functions" wording of the IS, particularly given the close interoperation
    with the existing `::operator new`/`::operator delete` implementations.
*   Added a discussion of the Microsoft ABI in [[#prop]].
*   Noted in [[#splode]] the possibility of using a parameter pack.
*   Added a proposal for [[#new]], as requested by EWG.

Additionally, a discussion of [[#sizeddelete]] has been added.

<pre class=biblio>
{
    "P0401R1": {
        "href": "http://wg21.link/P0401R1",
        "authors": ["Jonathan Wakely", "Chris Kennelly"],
        "date": "2019-06-11",
        "title": "Providing size feedback in the Allocator interface"
    },

    "P0401R4": {
        "href": "http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2020/p0401r4.html",
        "authors": ["Jonathan Wakely", "Chris Kennelly"],
        "date": "2020-11-14",
        "title": "Providing size feedback in the Allocator interface"
    },

    "Cologne": {
        "date": "2019-09-17",
        "title": "Size feedback in operator new"
    },

    "CWG2021Telecon": {
      "date": "2021-03-22",
      "title": "Paper Review Teleconference 2021-03-22"
    },

    "JacksonvilleMinutes": {
        "date": "2018-03-15",
        "title": "Jacksonville 2018 minutes"
    },

    "jemalloc": {
        "title": "jemalloc(3) - Linux man page",
        "href": "http://jemalloc.net/jemalloc.3.html"
    },

    "MallocExtension": {
        "href": "https://github.com/google/tcmalloc/blob/master/tcmalloc/malloc_extension.h",
        "title": "TCMalloc Malloc Extensions"
    },

    "MicrosoftABI": {
        "title": "Return Values",
        "href": "https://docs.microsoft.com/en-us/cpp/build/return-values-cpp",
        "date": "2016-11-03"
    },

    "N4917": {
        "title": "Working Draft, Standard for Programming Language C++",
        "href": "https://open-std.org/jtc1/sc22/wg21/docs/papers/2022/n4917.pdf",
        "date": "2022-09-05"
    },

    "P0901R5": {
        "title": "Size feedback in operator new",
        "date": "2019-10-06",
        "href": "http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p0901r5.html"
    },

    "P0901R6": {
        "title": "Size feedback in operator new",
        "date": "2020-03-01",
        "href": "http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2020/p0901r6.html"
    },

    "Prague": {
        "title": "LEWG Prague Minutes for P901R5",
        "date": "2020-02-10"
    },

    "SizedDelete": {
        "title": "C++ Sized Deallocation",
        "authors": ["Lawrence Crowl"],
        "href": "http://wg21.link/n3536"
    },

    "smallocx": {
        "href": "https://github.com/jemalloc/jemalloc/pull/1270",
        "title": "Add experimental API to support P0901r0"
    },

    "TCMalloc": {
        "href": "https://github.com/google/tcmalloc",
        "title": "TCMalloc"
    },

    "abslCord": {
        "href": "https://github.com/abseil/abseil-cpp/blob/1feab4fff90f904518e66cf80971063486fbc984/absl/strings/cord.h#L19-L23",
        "title": "`absl::Cord`"
    }
}
</pre>