1. Motivation
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,
, allocates memory with
code that looks something like this (with many details, like
,
templating for non
, and
exception safety, elided):
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 ; }
Consider the sequence of calls:
std :: vector < char > v ; v . reserve ( 37 ); // ... v . reserve ( 38 );
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.)
1.1. nallocx: not as awesome as it looks
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
. We can then use that hint as a smarter
parameter for operator new:
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 ; }
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.
1.1.1. nallocx must give a conservative answer
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
might return 64
and 128 bytes, we’d definitely rather know the right answer, not a conservative
approximation.
1.1.2. nallocx duplicates work
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
, we duplicate all that communication, and
also the overhead of the function call.
1.1.3. nallocx hides information from malloc
The biggest problem (for the authors) is that
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
, we will only ever be told that they want 40 (or 48, or
whatever
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
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
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
wants.
These three issues explain why we don’t believe
is a sufficient
solution here.
1.2. after allocation is too late
Another obvious suggestion is to add a way to inspect the size of an object
returned by
. Most mallocs provide a way to do this; [jemalloc] calls it
. Vector would look like:
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 ; }
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
. We could in principle change the standard,
but this would be an implementation nightmare.
1.3. realloc’s day has passed
We should also quickly examine why the classic C API
is insufficient.
void vector::reserve ( size_t new_cap ) { if ( capacity_ >= new_cap ) return ; ptr_ = realloc ( ptr_ , new_cap ); capacity_ = new_cap ; }
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,
'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,
serves little purpose.
2. Proposal
We propose adding new overloads of
(size-returning allocation
functions) that directly inform the user of the size available to them. C++
makes
replaceable (15.5.4.6), allowing a program to provide
its own version different from the standard library’s implementation.
We propose wording, relative to [CppDraft]:
-
Amend [basic.stc.dynamic.allocation], paragraph 1:
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.[[ nodiscard ]] void * operator new ( std :: size_t ); [[ nodiscard ]] void * operator new ( std :: size_t , std :: align_val_t ); [[ nodiscard ]] std :: sized_allocation_t operator new ( std :: size_t , std :: return_size_t ); [[ nodiscard ]] std :: sized_allocation_t operator new ( std :: size_t , std :: align_val_t , std :: return_size_t ); 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 ); [[ nodiscard ]] std :: sized_allocation_t operator new []( std :: size_t , std :: return_size_t ); [[ nodiscard ]] std :: sized_allocation_t operator new []( std :: size_t , std :: align_val_t , std :: return_size_t ); 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 ; These implicit declarations introduce only the function names
,
operator new ,
operator new [] , and
operator delete .
operator delete [] [ Note 2: The implicit declarations do not introduce the names
,
std ,
std :: size_t ,
std :: align_val_t 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
std :: return_size_t is well-formed. However, referring to
< new > or
std or
std :: size_t or
std :: align_val_t is ill-formed unless the name has been declared by importing or including the appropriate header. — end note ]
std :: return_size_t Allocation and/or deallocation functions may also be declared and defined for any class ([class.free]).
-
Amend [basic.stc.dynamic.allocation], paragraph 1:
An allocation function that is not a class member function shall belong to the global scope and not have a name with internal linkage. An allocation function is a size-returning allocation function if it has a second parameter of type, or it has a second parameter of type
std :: return_size_t and a third parameter of type
std :: align_val_t . The return type shall be
std :: return_size_t ([new.syn]) if the allocation function is a size-returning allocation function and
std :: sized_allocation_t < void > otherwise . The first parameter shall have type
void * ([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. 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.
std :: size_t
-
Insert [basic.stc.dynamic.allocation], after paragraph 1:
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 can be said to return a pointer value.
-
Insert [expr.new], after paragraph 9:
A new-pointer is pointer that results from evaluating the new-expression:
If the new-expression is a size-returning placement new expression then new-pointer is the ptr member of
([new.syn])
std :: sized_allocation_t < T > Otherwise new-pointer is a prvalue of type "pointer to T"
-
Amend [expr.new], paragraph 10:
When the allocated type is "array of N T" (that is, the noptr-new-declarator syntax is used or the new-type-id or type-id denotes an array type),the new-expression yields a prvalue of type "pointer to T" that points to the initial element (if any) of the array. Otherwise, let T be the allocated type; the new-expression is a prvalue of type "pointer to T" that points to the object createdthe new-pointer points to the initial element (if any) of the array. Otherwise, let T be the allocated type; the new-pointer points to the object created .
-
Amend [expr.new], paragraph 16:
When a new-expression calls an allocation function and that allocation has not been extended, the new-expression passes the amount of space requested to the allocation function as the first argument of type. That argument shall be no less than the size of the object being created; it may be greater than the size of the object being created only if the object is an array and the allocation function is not a non-allocating form ([new.delete.placement]). For arrays of
std :: size_t ,
char , and
unsigned char , the difference between the
std :: byte result of the new-expressionnew-pointer and the address returned by the allocation function shall be an integral multiple of the strictest fundamental alignment requirement of any object type whose size is no greater than the size of the array being created.
-
Insert [expr.new], after paragraph 17:
The result of a size-returning placement new-expression is an object of type([new.syn]) whose
std :: sized_allocation_t < T > member points to the object created and bytes member satisfies
ptr , where
requested <= bytes <= returned
is the number of bytes of valid storage starting from
bytes , and
ptr when the allocated type is "array of N T"
is
requested , and
sizeof ( T ) * N
is the number of bytes of storage returned by the size-returning allocation function less any array allocation overhead,
returned otherwise
is
requested , and
sizeof ( T )
is the number of bytes of storage returned by the size-returning allocation function.
returned
-
Amend [expr.new], paragraph 19:
Overload resolution is performed on a function call created by assembling an argument list. The first argument is the amount of space requested, and has type. If the type of the allocated object has new-extended alignment, the next argument is the type’s alignment, and has type
std :: size_t . The next argument is
std :: align_val_t . If the new-placement syntax is used, the initializer-clauses in its expression-list are the succeeding arguments. If no matching function is found then
std :: return_size_t
if the allocated object type has new-extended alignment, the alignment argument is removed from the argument list;
otherwise, an argument that is the type’s alignment and has type
is added into the argument list immediately after the first argument;
std :: align_val_t and then overload resolution is performed again.
If no matching function is found then the return size argument is removed from the argument list and then overload resolution is performed again.
-
Amend [expr.new], paragraph 19:
[Example 6:
results in one of the following calls:
new T
- operator new(sizeof(T), std::size_return_t())
- operator new(sizeof(T), std::align_val_t(alignof(T)), std::size_return_t())
operator new(sizeof(T))
operator new(sizeof(T), std::align_val_t(alignof(T)))
results in one of the following calls:
new ( 2 , f ) T
- operator new(sizeof(T), std::size_return_t(), 2, f)
- operator new(sizeof(T), std::align_val_t(alignof(T)), std::size_return_t(), 2, f)
operator new(sizeof(T), 2, f)
operator new(sizeof(T), std::align_val_t(alignof(T)), 2, f)
results in one of the following calls:
new T [ 5 ]
results in one of the following calls:
new ( 2 , f ) T [ 5 ] Here, each instance of x is a non-negative unspecified value representing array allocation overhead; the result of the new-expression will be offset by this amount from the value returned by
. This overhead may be applied in all array new-expressions, including those referencing a placement allocation function, except when referencing the library function
operator new [] . The amount of overhead may vary from one invocation of new to another. — end example]
operator new []( std :: size_t , void * )
-
Amend [expr.new], paragraph 21:
[Note 10: Unless an allocation function has a non-throwing exception specification, it indicates failure to allocate storage by throwing aexception ([basic.stc.dynamic.allocation], [except], [bad.alloc]);
std :: bad_alloc it returnsnew-pointer is a non-null pointer otherwise. If the allocation function has a non-throwing exception specification,it returnsnew-pointer is null to indicate failure to allocate storage and new-pointer is a non-null pointer otherwise. — end note]If the allocation function is a non-allocating form ([new.delete.placement])
that returnswhose new-pointer null, the behavior is undefined. Otherwise, if the allocation function 's new-pointer isreturnsnull, initialization shall not be done, the deallocation function shall not be called, and the value of the new-expression shall be null.
-
Amend [expr.new], paragraph 22:
[Note 11: When the allocation function 's new-pointer is
returnsa value other than null, it must be a pointer to a block of storage in which space for the object has been reserved. The block of storage is assumed to be appropriately aligned and of the requested size. The address of the created object will not necessarily be the same as that of the block if the object is an array. — end note]
-
Amend [expr.new], paragraph 29:
If a new-expression calls a deallocation function, it passes thevalue returnednew-pointer from the allocation function call as the first argument of type void*. If a placement deallocation function is called, it is passed the same additional arguments as were passed to the placement allocation function, that is, the same arguments as those specified with the new-placement syntax. If the implementation is allowed to introduce a temporary object or make a copy of any argument as part of the call to the allocation function, it is unspecified whether the same object is used in the call to both the allocation and deallocation functions.
-
Amend [expr.delete], paragraph 2:
If the operand has a class type, the operand is converted to a pointer type by calling the above-mentioned conversion function, and the converted operand is used in place of the original operand for the remainder of this subclause. In a single-object delete expression, the value of the operand ofmay be a null pointer value,
delete a pointer value that resulted fromthe new-pointer of a previous non-array new-expression, or a pointer to a base class subobject of an object created by such a new-expression. If not, the behavior is undefined. In an array delete expression, the value of the operand ofmay be a null pointer value or
delete a pointer value that resulted fromthe new-pointer of a previous array new-expression. If not, the behavior is undefined.[Note 1: This means that the syntax of the delete-expression must match the type of the object allocated by new, not the syntax of the new-expression. — end note]
[Note 2: A pointer to a const type can be the operand of a delete-expression; it is not necessary to cast away the constness ([expr.const.cast]) of the pointer expression before it is used as the operand of the delete-expression. — end note]
-
Amend [expr.delete], paragraph 7:
If the value of the operand of the delete-expression is not a null pointer value, then:If the allocation call for the new-expression for the object to be deleted was not omitted and the allocation was not extended ([expr.new]), the delete-expression shall call a deallocation function. The
value returned from the allocation call of the new-expressionnew-pointer shall be passed as the first argument to the deallocation function.Otherwise, if the allocation was extended or was provided by extending the allocation of another new-expression, and the delete-expression for every other
pointer value produced by a new-expressionnew-pointer that had storage provided by the extended new-expression has been evaluated, the delete-expression shall call a deallocation function. Thevalue returned from the allocation callnew-pointer of the extended new-expression shall be passed as the first argument to the deallocation function.Otherwise, the delete-expression will not call a deallocation function.
[Note 3: The deallocation function is called regardless of whether the destructor for the object or some element of the array throws an exception. — end note]
If the value of the operand of the delete-expression is a null pointer value, it is unspecified whether a deallocation function will be called as described above.
-
Amend [replacement.functions], paragraph 2:
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 & ) 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 ) [...] 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 & ) 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 ) [...]
-
Amend header
synopsis in [new.syn]:< new >
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 {}; // global operator new control struct return_size_t { explicit return_size_t () = default ; }; inline constexpr return_size_t return_size {}; template < typename T = void > struct sized_allocation_t { T * ptr ; size_t bytes ; }; 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 ; [[ nodiscard ]] std :: sized_allocation_t < void > :: operator new ( std :: size_t size , std :: return_size_t ); [[ nodiscard ]] std :: sized_allocation_t < void > :: operator new ( std :: size_t size , std :: align_val_t alignment , std :: return_size_t ); [[ nodiscard ]] std :: sized_allocation_t < void > :: operator new ( std :: size_t size , std :: return_size_t , std :: nothrow_t ) noexcept ; [[ nodiscard ]] std :: sized_allocation_t < void > :: operator new ( std :: size_t size , std :: align_val_t alignment , std :: return_size_t , std :: nothrow_t ) noexcept ; [...] [[ 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 ; [[ nodiscard ]] std :: sized_allocation_t < void > :: operator new []( std :: size_t size , std :: return_size_t ); [[ nodiscard ]] std :: sized_allocation_t < void > :: operator new []( std :: size_t size , std :: align_val_t alignment , std :: return_size_t ); [[ nodiscard ]] std :: sized_allocation_t < void > :: operator new []( std :: size_t size , std :: return_size_t , std :: nothrow_t ) noexcept ; [[ nodiscard ]] std :: sized_allocation_t < void > :: operator new []( std :: size_t size , std :: align_val_t alignment , std :: return_size_t , std :: nothrow_t ) noexcept ; [...] [[ 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 ;
-
Insert [new.delete.single], after paragraph 4:
[[ nodiscard ]] std :: sized_allocation_t < void > :: operator new ( std :: size_t size , std :: return_size_t ); [[ nodiscard ]] std :: sized_allocation_t < void > :: operator new ( std :: size_t size , std :: align_val_t alignment , std :: return_size_t );
- 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.
- 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.
- Required behavior: Return a
whose
sized_allocation_t < void > member represents the address of a region of N bytes of suitably aligned storage ([basic.stc.dynamic]) for some , and whose
ptr member is N, or else throw a
bytes exception. This requirement is binding on any replacement versions of these functions.
bad_alloc - Default behavior: Returns
and
std :: sized_allocation_t < void > { operator new ( size ), N } respectively. If a user-provided operator new is invoked directly or indirectly, N is
std :: sized_allocation_t < void > { operator new ( size , alignment ), N } .
size
-
Insert [new.delete.single], after paragraph 9:
[[ 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 ;
- 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.
- 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.
- Required behavior: Return a
whose
sized_allocation_t < void > member represents the address of a region of N bytes of suitably aligned storage ([basic.stc.dynamic]) for some , and whose
ptr member is N, or else return
bytes . Each of these nothrow versions of
std :: sized_allocation_t < void > { nullptr , 0 } 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.
operator new - Default behavior: Returns
and
std :: sized_allocation_t < void > { operator new ( size ), N } respectively. If a user-provided operator new is invoked directly or indirectly, N is
std :: sized_allocation_t < void > { operator new ( size , alignment ), N } . If the call to
size throws, returns
operator new .
std :: sized_allocation_t < void > { nullptr , 0 }
-
Amend [new.delete.single], paragraph 12:
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 ;
[...]
-
If the
parameter is not present,alignment
shall have been returned by an allocation function without an alignment parameter. If present, theptr
argument shall equal the alignment argument passed to the allocation function that returnedalignment
. If present, theptr
argument shall equal thesize
argument passed to the allocation function that returnedsize
, ifptr
was not allocated by a size-returning allocation function. If present, theptr
argument shall satisfy ifsize
was allocated by a size-returning allocation function, whereptr
is the size returned inbytes
andstd :: sized_allocation_t < void >
is the size argument passed to the allocation function .requested
[...]
-
Insert [new.delete.array], after paragraph 4:
[[ 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 );
- 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.
- 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.
- Required behavior: Return a
whose
sized_allocation_t < void > member represents the address of a region of N bytes of suitably aligned storage ([basic.stc.dynamic]) for some , and whose
ptr member is N, or else throw a
bytes exception. This requirement is binding on any replacement versions of these functions.
bad_alloc - Default behavior: Returns
and
std :: sized_allocation_t < void > { operator new []( size ), N } respectively. If a user-provided operator new is invoked directly or indirectly, N is
std :: sized_allocation_t < void > { operator new []( size , alignment ), N } .
size
-
Insert [new.delete.array], after paragraph 8:
[[ 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 ;
- 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.
- 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.
- Required behavior: Return a
whose
sized_allocation_t < void > member represents the address of a region of N bytes of suitably aligned storage ([basic.stc.dynamic]) for some , and whose
ptr member is N, or else return
bytes . This requirement is binding on any replacement versions of these functions.
std :: sized_allocation_t < void > { nullptr , 0 } - Default behavior: Returns
and
std :: sized_allocation_t < void > { operator new []( size ), N } respectively. If a user-provided operator new is invoked directly or indirectly, N is
std :: sized_allocation_t < void > { operator new []( size , alignment ), N } . If the call to
size throws, returns
operator new [] .
std :: sized_allocation_t < void > { nullptr , 0 }
-
Amend [new.delete.array], paragraph 11:
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 ;
[...]
-
If the
parameter is not present,alignment
shall have been returned by an allocation function without anptr
parameter. If present, thealignment
argument shall equal thealignment
argument passed to the allocation function that returnedalignment
. If present, theptr
argument shall equal thesize
argument passed to the allocation function that returnedsize
, ifptr
was not allocated by a size-returning allocation function. If present, theptr
argument shall satisfy ifsize
was allocated by a size-returning allocation function, whereptr
is the size returned inbytes
andstd :: sized_allocation_t < void >
is the size argument passed to the allocation function .requested
[...]
-
Amend Table 19 ("Feature-test macros") [cpp.predefined]
Name Value
__cpp_size_returning_new PLACEHOLDER DATE
3. Alternative Designs Considered
Another signature we could use would be:
enum class return_size_t : std :: size_t {}; void * :: operator new ( std :: size_t size , std :: return_size_t & );
(and so on.) This is slightly simpler to read as a signature, but arguably worse in usage:
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 ;
More importantly, this form is less efficient. In practice, underlying malloc
implementations provide actual definitions of
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.
3.1. How many :: operator new
's?
It is unfortunate that we have so many permutations of
--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
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
) 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
/
whose sole purpose would
be to accept and discard the
.
3.2. Implementation difficulty
It’s worth reiterating that there’s a perfectly good trivial implementation of these functions:
std :: sized_allocation_t < void > :: operator new ( std :: size_t n , std :: return_size_t ) { return { :: operator new ( n ), n }; }
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
function providing a C API for this functionality [smallocx].smallocx
3.3. Interaction with Sized Delete
For allocations made with
-returning
, we need to
relax
's size argument (16.6.2.1 and 16.6.2.2). For
allocations of
, the size quanta used by the allocator may not be a multiple
of
, leading to both the original and returned sizes being
unrecoverable at the time of deletion.
Consider the memory allocated by:
using T = std :: aligned_storage < 16 , 8 >:: type ; std :: vector < T > v ( 4 );
The underlying heap allocation is made with
.
-
The memory allocator may return a 72 byte object: Since there is no
such thatk
, we can’t provide that value tosizeof ( T ) * k = 72
. The only option would be storing 72 explicitly, which would be wasteful.:: operator delete ( void * , size_t ) -
The memory allocator may instead return an 80 byte object (5
's): We now cannot represent the original request when deallocating without additional storage.T
For allocations made with
std :: tie ( p , m ) = :: operator new ( n , std :: return_size_t {});
we permit
where
.
This behavior is consistent with [jemalloc]'s
, where the
deallocation size must fall between the request (
) and the actual allocated
size (
) inclusive.
3.4. 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
or the like, as that same translation must be performed in addition to the duplicative work previously discussed.nallocx -
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.
4. New Expressions
Additionally, we propose expanding this functionality to
expressions by returning:
-
For
, a pointer to the object created and the size of the allocation in bytes.new -
For
, a pointers to the initial element of the array and the size of the allocation in bytes (less overhead):new [] 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 ;
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).
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 ; The pair of pointers provides convience 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
(or fail to be a multiple ofsizeof ( T )
, a possible, but unlikely scenario). This imposes a need for more extensive logic around the new-expression in handling the returned allocation size.alignof ( T ) -
We could return the size in units of
, this leads to an inconsistency between the expected usage forT
andnew
:new [] -
For
, we may only end up fitting a singlenew
into an allocator size quanta, so the extra space remains unusable. If we can fit multipleT
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 asT
is a replaceable function.:: operator new -
For
, the size in units ofnew []
can easily be derived from the returned size in bytes.T
-
-
We could pass the size in units of
or bytes to the constructor ofT
:T -
For
, this is especially useful for tail-padded arrays, but neglects default-initializednew
.T -
For
, a common use case is expected to be the allocation of arrays ofnew []
,char
, etc. The size of the overall array is irrelevant for the individual elements.int
-
-
We could return the size via a reference parameter:
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 (); } or, demonstrated with bytes:
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 ; (Casts omitted for clarity.)
As discussed for
in § 2 Proposal, a reference parameter poses difficulties for optimizers and involves returning the size via memory (depending on ABI).:: operator new
For
expressions, we considered alternatively initializing the returned
(
) 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
to destroy the correct number of elements (by storingdelete []
in the array allocation overhead).sz / sizeof ( T ) -
The presented proposal (leaving this space uninitialized) was chosen for consistency with
.new
5. Naming
The library support type is named
, based on
LEWG’s naming suggestions and for consistency (in choice of
) with the
other allocation library support types (
,
,
,
etc.). We expect this to be spelled rarely.
Its members are
and
.
for the memory returned is an
intuitive name.
conveys units in its name (
). 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
.
6. Related work
[P0401R4] considers this problem at the level of the
concept.
Ironically, the lack of the above API was one significant problem in its
original revision: how could an implementation of
provide the
requested feedback in a way that would work with any underlying malloc
implementation?
7. History
7.1. R8 → R9
CWG reviewed [P0901R8] via [CWG2021Telecon].
Wording changes have been made to the proposal based on CWG feedback and assistance from Ryan McDougall.
7.2. R7 → R8
LEWG reviewed [P0901R6] via telecon.
Poll: Send P0901R6, after changingto
p and
ptr to
n in
bytes , to electronic ballot to be forwarded to CWG.
sized_allocation_t
SF F N A SA 4 11 0 0 0
-
Applied member name feedback from LEWG to rename
's members fromsized_allocation_t
top
andptr
ton
. Added naming rationale section.bytes -
Update reference to revised [P0401R4].
-
Refined wording in proposal.
7.3. R6 → R7
-
Removed extraneous
typostd ::
7.4. R5 → R6
LEWG reviewed [P0901R5] library support types at [Prague].
-
LEWG discussed the similarity to
. Participants in the discussion were concerned about the potential library layering issue (std :: span
layers below practically everything) and that ownership is conveyed in the return value. In contrast,< new >
is non-owning.std :: span -
LEWG discussed whether the type needed a name. It needs to be named to allow users to replace
and spell the type.operator new
LEWG took an approval poll of different names
5
sized_ptr_t 2
sized_ptr 3
memblock_t 7
memblock 0
memspan_t 2
memspan 8
alloc_size_t 4
alloc_size 5
alloc_t 0
alloc 13
sized_allocation_t 17
sized_allocation 0
alloc_span_t 2
alloc_span 9
allocation_t 9
allocation 16
alloc_result_t 14
alloc_result 12
allocation_result_t 16
allocation_result
Based on this poll, LEWG directed the paper authors to consider
,
, and
.
LEWG additionally polled on whether the type should have a
suffix.
SF F N A SA 0 6 7 8 1
Based on this feedback, the authors have chosen
, based on
LEWG’s naming suggestions and for consistency (in choice of
) with the
other allocation library support types (
,
,
,
etc.). We expect this to be spelled rarely.
7.5. R4 → R5
EWG reviewed P0901R4 at [Cologne].
Poll: P0901R4 as presented, forward to LEWG for C++23, not C++20.
SF F N A SA 2 11 14 2 0
-
Fixed typos in examples.
-
Added proposed feature test macro.
-
Added LEWG audience for library support type names.
7.6. R3 → R4
-
Update reference to revised [P0401R1].
7.7. R2 → R3
-
Added proposed wording.
-
For newly added allocation functions,
is passed by value, rather than reference.std :: nothrow_t
7.8. R1 → R2
Applied feedback from San Diego Mailing
-
Moved from passing
parameter by reference to by value. For many ABIs, this is more optimizable and to the authors' knowledge, no worse on any other.std :: return_size_t -
Added rationale for not using parameter packs for this functionality.
7.9. R0 → R1
Applied feedback from [JacksonvilleMinutes].
-
Clarified in § 2 Proposal the desire to leverage the existing "replacement functions" wording of the IS, particularly given the close interoperation with the existing
/:: operator new
implementations.:: operator delete -
Added a discussion of the Microsoft ABI in § 2 Proposal.
-
Noted in § 3.1 How many ::operator new's? the possibility of using a parameter pack.
-
Added a proposal for § 4 New Expressions, as requested by EWG.
Additionally, a discussion of § 3.3 Interaction with Sized Delete has been added.