Automatic Reference Counting

About this document

Purpose

The first and primary purpose of this document is to serve as a complete technical specification of Automatic Reference Counting. Given a core Objective-C compiler and runtime, it should be possible to write a compiler and runtime which implements these new semantics.

The secondary purpose is to act as a rationale for why ARC was designed in this way. This should remain tightly focused on the technical design and should not stray into marketing speculation.

Background

This document assumes a basic familiarity with C.

Blocks are a C language extension for creating anonymous functions. Users interact with and transfer block objects using block pointers, which are represented like a normal pointer. A block may capture values from local variables; when this occurs, memory must be dynamically allocated. The initial allocation is done on the stack, but the runtime provides a Block_copy function which, given a block pointer, either copies the underlying block object to the heap, setting its reference count to 1 and returning the new block pointer, or (if the block object is already on the heap) increases its reference count by 1. The paired function is Block_release, which decreases the reference count by 1 and destroys the object if the count reaches zero and is on the heap.

Objective-C is a set of language extensions, significant enough to be considered a different language. It is a strict superset of C. The extensions can also be imposed on C++, producing a language called Objective-C++. The primary feature is a single-inheritance object system; we briefly describe the modern dialect.

Objective-C defines a new type kind, collectively called the object pointer types. This kind has two notable builtin members, id and Class; id is the final supertype of all object pointers. The validity of conversions between object pointer types is not checked at runtime. Users may define classes; each class is a type, and the pointer to that type is an object pointer type. A class may have a superclass; its pointer type is a subtype of its superclass's pointer type. A class has a set of ivars, fields which appear on all instances of that class. For every class T there's an associated metaclass; it has no fields, its superclass is the metaclass of T's superclass, and its metaclass is a global class. Every class has a global object whose class is the class's metaclass; metaclasses have no associated type, so pointers to this object have type Class.

A class declaration (@interface) declares a set of methods. A method has a return type, a list of argument types, and a selector: a name like foo:bar:baz:, where the number of colons corresponds to the number of formal arguments. A method may be an instance method, in which case it can be invoked on objects of the class, or a class method, in which case it can be invoked on objects of the metaclass. A method may be invoked by providing an object (called the receiver) and a list of formal arguments interspersed with the selector, like so:

[receiver foo: fooArg bar: barArg baz: bazArg]

This looks in the dynamic class of the receiver for a method with this name, then in that class's superclass, etc., until it finds something it can execute. The receiver expression may also be the name of a class, in which case the actual receiver is the class object for that class, or (within method definitions) it may be super, in which case the lookup algorithm starts with the static superclass instead of the dynamic class. The actual methods dynamically found in a class are not those declared in the @interface, but those defined in a separate @implementation declaration; however, when compiling a call, typechecking is done based on the methods declared in the @interface.

Method declarations may also be grouped into protocols, which are not inherently associated with any class, but which classes may claim to follow. Object pointer types may be qualified with additional protocols that the object is known to support.

Class extensions are collections of ivars and methods, designed to allow a class's @interface to be split across multiple files; however, there is still a primary implementation file which must see the @interfaces of all class extensions. Categories allow methods (but not ivars) to be declared post hoc on an arbitrary class; the methods in the category's @implementation will be dynamically added to that class's method tables which the category is loaded at runtime, replacing those methods in case of a collision.

In the standard environment, objects are allocated on the heap, and their lifetime is manually managed using a reference count. This is done using two instance methods which all classes are expected to implement: retain increases the object's reference count by 1, whereas release decreases it by 1 and calls the instance method dealloc if the count reaches 0. To simplify certain operations, there is also an autorelease pool, a thread-local list of objects to call release on later; an object can be added to this pool by calling autorelease on it.

Block pointers may be converted to type id; block objects are laid out in a way that makes them compatible with Objective-C objects. There is a builtin class that all block objects are considered to be objects of; this class implements retain by adjusting the reference count, not by calling Block_copy.

General

Automatic Reference Counting implements automatic memory management for Objective-C objects and blocks, freeing the programmer from the need explicitly insert retains and releases. It does not provide a cycle collector; users must explicitly manage lifetime instead.

ARC may be explicitly enabled with the compiler flag -fobjc-arc. It may also be explicitly disabled with the compiler flag -fno-objc-arc. The last of these two flags appearing on the compile line wins.

If ARC is enabled, __has_feature(objc_arc) will expand to 1 in the preprocessor. For more information about __has_feature, see the language extensions document.

Retainable object pointers

This section describes retainable object pointers, their basic operations, and the restrictions imposed on their use under ARC. Note in particular that it covers the rules for pointer values (patterns of bits indicating the location of a pointed-to object), not pointer objects (locations in memory which store pointer values). The rules for objects are covered in the next section.

A retainable object pointer (or retainable pointer) is a value of a retainable object pointer type (retainable type). There are three kinds of retainable object pointer types:

Other pointer types, such as int* and CFStringRef, are not subject to ARC's semantics and restrictions.

Rationale: We are not at liberty to require all code to be recompiled with ARC; therefore, ARC must interoperate with Objective-C code which manages retains and releases manually. In general, there are three requirements in order for a compiler-supported reference-count system to provide reliable interoperation:

  • The type system must reliably identify which objects are to be managed. An int* might be a pointer to a malloc'ed array, or it might be a interior pointer to such an array, or it might point to some field or local variable. In contrast, values of the retainable object pointer types are never interior.
  • The type system must reliably indicate how to manage objects of a type. This usually means that the type must imply a procedure for incrementing and decrementing retain counts. Supporting single-ownership objects requires a lot more explicit mediation in the language.
  • There must be reliable conventions for whether and when ownership is passed between caller and callee, for both arguments and return values. Objective-C methods follow such a convention very reliably, at least for system libraries on Mac OS X, and functions always pass objects at +0. The C-based APIs for Core Foundation objects, on the other hand, have much more varied transfer semantics.

The use of __attribute__((NSObject)) typedefs is not recommended. If it's absolutely necessary to use this attribute, be very explicit about using the typedef, and do not assume that it will be preserved by language features like __typeof and C++ template argument substitution.

Rationale: any compiler operation which incidentally strips type sugar from a type will yield a type without the attribute, which may result in unexpected behavior.

Retain count semantics

A retainable object pointer is either a null pointer or a pointer to a valid object. Furthermore, if it has block pointer type and is not null then it must actually be a pointer to a block object, and if it has Class type (possibly protocol-qualified) then it must actually be a pointer to a class object. Otherwise ARC does not enforce the Objective-C type system as long as the implementing methods follow the signature of the static type. It is undefined behavior if ARC is exposed to an invalid pointer.

For ARC's purposes, a valid object is one with well-behaved retaining operations. Specifically, the object must be laid out such that the Objective-C message send machinery can successfully send it the following messages:

  • retain, taking no arguments and returning a pointer to the object.
  • release, taking no arguments and returning void.
  • autorelease, taking no arguments and returning a pointer to the object.

The behavior of these methods is constrained in the following ways. The term high-level semantics is an intentionally vague term; the intent is that programmers must implement these methods in a way such that the compiler, modifying code in ways it deems safe according to these constraints, will not violate their requirements. For example, if the user puts logging statements in retain, they should not be surprised if those statements are executed more or less often depending on optimization settings. These constraints are not exhaustive of the optimization opportunities: values held in local variables are subject to additional restrictions, described later in this document.

It is undefined behavior if a computation history featuring a send of retain followed by a send of release to the same object, with no intervening release on that object, is not equivalent under the high-level semantics to a computation history in which these sends are removed. Note that this implies that these methods may not raise exceptions.

It is undefined behavior if a computation history features any use whatsoever of an object following the completion of a send of release that is not preceded by a send of retain to the same object.

The behavior of autorelease must be equivalent to sending release when one of the autorelease pools currently in scope is popped. It may not throw an exception.

When the semantics call for performing one of these operations on a retainable object pointer, if that pointer is null then the effect is a no-op.

All of the semantics described in this document are subject to additional optimization rules which permit the removal or optimization of operations based on local knowledge of data flow. The semantics describe the high-level behaviors that the compiler implements, not an exact sequence of operations that a program will be compiled into.

Retainable object pointers as operands and arguments

In general, ARC does not perform retain or release operations when simply using a retainable object pointer as an operand within an expression. This includes:

  • loading a retainable pointer from an object with non-weak ownership,
  • passing a retainable pointer as an argument to a function or method, and
  • receiving a retainable pointer as the result of a function or method call.

Rationale: while this might seem uncontroversial, it is actually unsafe when multiple expressions are evaluated in parallel, as with binary operators and calls, because (for example) one expression might load from an object while another writes to it. However, C and C++ already call this undefined behavior because the evaluations are unsequenced, and ARC simply exploits that here to avoid needing to retain arguments across a large number of calls.

The remainder of this section describes exceptions to these rules, how those exceptions are detected, and what those exceptions imply semantically.

Consumed parameters

A function or method parameter of retainable object pointer type may be marked as consumed, signifying that the callee expects to take ownership of a +1 retain count. This is done by adding the ns_consumed attribute to the parameter declaration, like so:

void foo(__attribute((ns_consumed)) id x);
- (void) foo: (id) __attribute((ns_consumed)) x;

This attribute is part of the type of the function or method, not the type of the parameter. It controls only how the argument is passed and received.

When passing such an argument, ARC retains the argument prior to making the call.

When receiving such an argument, ARC releases the argument at the end of the function, subject to the usual optimizations for local values.

Rationale: this formalizes direct transfers of ownership from a caller to a callee. The most common scenario here is passing the self parameter to init, but it is useful to generalize. Typically, local optimization will remove any extra retains and releases: on the caller side the retain will be merged with a +1 source, and on the callee side the release will be rolled into the initialization of the parameter.

The implicit self parameter of a method may be marked as consumed by adding __attribute__((ns_consumes_self)) to the method declaration. Methods in the init family are treated as if they were implicitly marked with this attribute.

It is undefined behavior if an Objective-C message send to a method with ns_consumed parameters (other than self) is made with a null receiver. It is undefined behavior if the method to which an Objective-C message send statically resolves to has a different set of ns_consumed parameters than the method it dynamically resolves to. It is undefined behavior if a block or function call is made through a static type with a different set of ns_consumed parameters than the implementation of the called block or function.

Rationale: consumed parameters with null receiver are a guaranteed leak. Mismatches with consumed parameters will cause over-retains or over-releases, depending on the direction. The rule about function calls is really just an application of the existing C/C++ rule about calling functions through an incompatible function type, but it's useful to state it explicitly.

Retained return values

A function or method which returns a retainable object pointer type may be marked as returning a retained value, signifying that the caller expects to take ownership of a +1 retain count. This is done by adding the ns_returns_retained attribute to the function or method declaration, like so:

id foo(void) __attribute((ns_returns_retained));
- (id) foo __attribute((ns_returns_retained));

This attribute is part of the type of the function or method.

When returning from such a function or method, ARC retains the value at the point of evaluation of the return statement, before leaving all local scopes.

When receiving a return result from such a function or method, ARC releases the value at the end of the full-expression it is contained within, subject to the usual optimizations for local values.

Rationale: this formalizes direct transfers of ownership from a callee to a caller. The most common scenario this models is the retained return from init, alloc, new, and copy methods, but there are other cases in the frameworks. After optimization there are typically no extra retains and releases required.

Methods in the alloc, copy, init, mutableCopy, and new families are implicitly marked __attribute__((ns_returns_retained)). This may be suppressed by explicitly marking the method __attribute__((ns_returns_not_retained)).

It is undefined behavior if the method to which an Objective-C message send statically resolves has different retain semantics on its result from the method it dynamically resolves to. It is undefined behavior if a block or function call is made through a static type with different retain semantics on its result from the implementation of the called block or function.

Rationale: Mismatches with returned results will cause over-retains or over-releases, depending on the direction. Again, the rule about function calls is really just an application of the existing C/C++ rule about calling functions through an incompatible function type.

Unretained return values

A method or function which returns a retainable object type but does not return a retained value must ensure that the object is still valid across the return boundary.

When returning from such a function or method, ARC retains the value at the point of evaluation of the return statement, then leaves all local scopes, and then balances out the retain while ensuring that the value lives across the call boundary. In the worst case, this may involve an autorelease, but callers must not assume that the value is actually in the autorelease pool.

ARC performs no extra mandatory work on the caller side, although it may elect to do something to shorten the lifetime of the returned value.

Rationale: it is common in non-ARC code to not return an autoreleased value; therefore the convention does not force either path. It is convenient to not be required to do unnecessary retains and autoreleases; this permits optimizations such as eliding retain/autoreleases when it can be shown that the original pointer will still be valid at the point of return.

A method or function may be marked with __attribute__((ns_returns_autoreleased)) to indicate that it returns a pointer which is guaranteed to be valid at least as long as the innermost autorelease pool. There are no additional semantics enforced in the definition of such a method; it merely enables optimizations in callers.

Bridged casts

A bridged cast is a C-style cast annotated with one of three keywords:

  • (__bridge T) op casts the operand to the destination type T. If T is a retainable object pointer type, then op must have a non-retainable pointer type. If T is a non-retainable pointer type, then op must have a retainable object pointer type. Otherwise the cast is ill-formed. There is no transfer of ownership, and ARC inserts no retain operations.
  • (__bridge_retained T) op casts the operand, which must have retainable object pointer type, to the destination type, which must be a non-retainable pointer type. ARC retains the value, subject to the usual optimizations on local values, and the recipient is responsible for balancing that +1.
  • (__bridge_transfer T) op casts the operand, which must have non-retainable pointer type, to the destination type, which must be a retainable object pointer type. ARC will release the value at the end of the enclosing full-expression, subject to the usual optimizations on local values.

These casts are required in order to transfer objects in and out of ARC control; see the rationale in the section on conversion of retainable object pointers.

Using a __bridge_retained or __bridge_transfer cast purely to convince ARC to emit an unbalanced retain or release, respectively, is poor form.

Restrictions

Conversion of retainable object pointers

In general, a program which attempts to implicitly or explicitly convert a value of retainable object pointer type to any non-retainable type, or vice-versa, is ill-formed. For example, an Objective-C object pointer shall not be converted to void*. As an exception, cast to intptr_t is allowed becuase such casts are not transferring ownership. The bridged casts may be used to perform these conversions where necessary.

Rationale: we cannot ensure the correct management of the lifetime of objects if they may be freely passed around as unmanaged types. The bridged casts are provided so that the programmer may explicitly describe whether the cast transfers control into or out of ARC.

An unbridged cast to a retainable object pointer type of the return value of a Objective-C message send which yields a non-retainable pointer is treated as a __bridge_transfer cast if:

  • the method has the cf_returns_retained attribute, or if not that,
  • the method does not have the cf_returns_not_retained attribute and
  • the method's selector family would imply the ns_returns_retained attribute on a method which returned a retainable object pointer type.

Otherwise the cast is treated as a __bridge cast.

Ownership qualification

This section describes the behavior of objects of retainable object pointer type; that is, locations in memory which store retainable object pointers.

A type is a retainable object owner type if it is a retainable object pointer type or an array type whose element type is a retainable object owner type.

An ownership qualifier is a type qualifier which applies only to retainable object owner types. An array type is ownership-qualified according to its element type, and adding an ownership qualifier to an array type so qualifies its element type.

A program is ill-formed if it attempts to apply an ownership qualifier to a type which is already ownership-qualified, even if it is the same qualifier. There is a single exception to this rule: an ownership qualifier may be applied to a substituted template type parameter, which overrides the ownership qualifier provided by the template argument.

Except as described under the inference rules, a program is ill-formed if it attempts to form a pointer or reference type to a retainable object owner type which lacks an ownership qualifier.

Rationale: these rules, together with the inference rules, ensure that all objects and lvalues of retainable object pointer type have an ownership qualifier. The ability to override an ownership qualifier during template substitution is required to counteract the inference of __strong for template type arguments.

There are four ownership qualifiers:

A type is nontrivially ownership-qualified if it is qualified with __autoreleasing, __strong, or __weak.

Spelling

The names of the ownership qualifiers are reserved for the implementation. A program may not assume that they are or are not implemented with macros, or what those macros expand to.

An ownership qualifier may be written anywhere that any other type qualifier may be written.

If an ownership qualifier appears in the declaration-specifiers, the following rules apply:

  • if the type specifier is a retainable object owner type, the qualifier applies to that type;
  • if the outermost non-array part of the declarator is a pointer or block pointer, the qualifier applies to that type;
  • otherwise the program is ill-formed.

If an ownership qualifier appears on the declarator name, or on the declared object, it is applied to outermost pointer or block-pointer type.

If an ownership qualifier appears anywhere else in a declarator, it applies to the type there.

Semantics

There are five managed operations which may be performed on an object of retainable object pointer type. Each qualifier specifies different semantics for each of these operations. It is still undefined behavior to access an object outside of its lifetime.

A load or store with primitive semantics has the same semantics as the respective operation would have on an void* lvalue with the same alignment and non-ownership qualification.

Reading occurs when performing a lvalue-to-rvalue conversion on an object lvalue.

  • For __weak objects, the current pointee is retained and then released at the end of the current full-expression. This must execute atomically with respect to assignments and to the final release of the pointee.
  • For all other objects, the lvalue is loaded with primitive semantics.

Assignment occurs when evaluating an assignment operator. The semantics vary based on the qualification:

  • For __strong objects, the new pointee is first retained; second, the lvalue is loaded with primitive semantics; third, the new pointee is stored into the lvalue with primitive semantics; and finally, the old pointee is released. This is not performed atomically; external synchronization must be used to make this safe in the face of concurrent loads and stores.
  • For __weak objects, the lvalue is updated to point to the new pointee, unless that object is currently undergoing deallocation, in which case it the lvalue is updated to a null pointer. This must execute atomically with respect to other assignments to the object, to reads from the object, and to the final release of the new pointed-to value.
  • For __unsafe_unretained objects, the new pointee is stored into the lvalue using primitive semantics.
  • For __autoreleasing objects, the new pointee is retained, autoreleased, and stored into the lvalue using primitive semantics.

Initialization occurs when an object's lifetime begins, which depends on its storage duration. Initialization proceeds in two stages:

  1. First, a null pointer is stored into the lvalue using primitive semantics. This step is skipped if the object is __unsafe_unretained.
  2. Second, if the object has an initializer, that expression is evaluated and then assigned into the object using the usual assignment semantics.

Destruction occurs when an object's lifetime ends. In all cases it is semantically equivalent to assigning a null pointer to the object, with the proviso that of course the object cannot be legally read after the object's lifetime ends.

Moving occurs in specific situations where an lvalue is moved from, meaning that its current pointee will be used but the object may be left in a different (but still valid) state. This arises with __block variables and rvalue references in C++. For __strong lvalues, moving is equivalent to loading the lvalue with primitive semantics, writing a null pointer to it with primitive semantics, and then releasing the result of the load at the end of the current full-expression. For all other lvalues, moving is equivalent to reading the object.

Restrictions

Storage duration of __autoreleasing objects

A program is ill-formed if it declares an __autoreleasing object of non-automatic storage duration.

Rationale: autorelease pools are tied to the current thread and scope by their nature. While it is possible to have temporary objects whose instance variables are filled with autoreleased objects, there is no way that ARC can provide any sort of safety guarantee there.

It is undefined behavior if a non-null pointer is assigned to an __autoreleasing object while an autorelease pool is in scope and then that object is read after the autorelease pool's scope is left.

Conversion of pointers to ownership-qualified types

A program is ill-formed if an expression of type T* is converted, explicitly or implicitly, to the type U*, where T and U have different ownership qualification, unless:

  • T is qualified with __strong, __autoreleasing, or __unsafe_unretained, and U is qualified with both const and __unsafe_unretained; or
  • either T or U is cv void, where cv is an optional sequence of non-ownership qualifiers; or
  • the conversion is requested with a reinterpret_cast in Objective-C++; or
  • the conversion is a well-formed pass-by-writeback.

The analogous rule applies to T& and U& in Objective-C++.

Rationale: these rules provide a reasonable level of type-safety for indirect pointers, as long as the underlying memory is not deallocated. The conversion to const __unsafe_unretained is permitted because the semantics of reads are equivalent across all these ownership semantics, and that's a very useful and common pattern. The interconversion with void* is useful for allocating memory or otherwise escaping the type system, but use it carefully. reinterpret_cast is considered to be an obvious enough sign of taking responsibility for any problems.

It is undefined behavior to access an ownership-qualified object through an lvalue of a differently-qualified type, except that any non-__weak object may be read through an __unsafe_unretained lvalue.

It is undefined behavior if a managed operation is performed on a __strong or __weak object without a guarantee that it contains a primitive zero bit-pattern, or if the storage for such an object is freed or reused without the object being first assigned a null pointer.

Rationale: ARC cannot differentiate between an assignment operator which is intended to initialize dynamic memory and one which is intended to potentially replace a value. Therefore the object's pointer must be valid before letting ARC at it. Similarly, C and Objective-C do not provide any language hooks for destroying objects held in dynamic memory, so it is the programmer's responsibility to avoid leaks (__strong objects) and consistency errors (__weak objects).

These requirements are followed automatically in Objective-C++ when creating objects of retainable object owner type with new or new[] and destroying them with delete, delete[], or a pseudo-destructor expression. Note that arrays of nontrivially-ownership-qualified type are not ABI compatible with non-ARC code because the element type is non-POD: such arrays that are new[]'d in ARC translation units cannot be delete[]'d in non-ARC translation units and vice-versa.

Passing to an out parameter by writeback

If the argument passed to a parameter of type T __autoreleasing * has type U oq *, where oq is an ownership qualifier, then the argument is a candidate for pass-by-writeback if:

  • oq is __strong or __weak, and
  • it would be legal to initialize a T __strong * with a U __strong *.

For purposes of overload resolution, an implicit conversion sequence requiring a pass-by-writeback is always worse than an implicit conversion sequence not requiring a pass-by-writeback.

The pass-by-writeback is ill-formed if the argument expression does not have a legal form:

  • &var, where var is a scalar variable of automatic storage duration with retainable object pointer type
  • a conditional expression where the second and third operands are both legal forms
  • a cast whose operand is a legal form
  • a null pointer constant

Rationale: the restriction in the form of the argument serves two purposes. First, it makes it impossible to pass the address of an array to the argument, which serves to protect against an otherwise serious risk of mis-inferring an array argument as an out-parameter. Second, it makes it much less likely that the user will see confusing aliasing problems due to the implementation, below, where their store to the writeback temporary is not immediately seen in the original argument variable.

A pass-by-writeback is evaluated as follows:

  1. The argument is evaluated to yield a pointer p of type U oq *.
  2. If p is a null pointer, then a null pointer is passed as the argument, and no further work is required for the pass-by-writeback.
  3. Otherwise, a temporary of type T __autoreleasing is created and initialized to a null pointer.
  4. If the argument is not an Objective-C method parameter marked out, then *p is read, and the result is written into the temporary with primitive semantics.
  5. The address of the temporary is passed as the argument to the actual call.
  6. After the call completes, the temporary is loaded with primitive semantics, and that value is assigned into *p.

Rationale: this is all admittedly convoluted. In an ideal world, we would see that a local variable is being passed to an out-parameter and retroactively modify its type to be __autoreleasing rather than __strong. This would be remarkably difficult and not always well-founded under the C type system. However, it was judged unacceptably invasive to require programmers to write __autoreleasing on all the variables they intend to use for out-parameters. This was the least bad solution.

Ownership-qualified fields of structs and unions

A program is ill-formed if it declares a member of a C struct or union to have a nontrivially ownership-qualified type.

Rationale: the resulting type would be non-POD in the C++ sense, but C does not give us very good language tools for managing the lifetime of aggregates, so it is more convenient to simply forbid them. It is still possible to manage this with a void* or an __unsafe_unretained object.

This restriction does not apply in Objective-C++. However, nontrivally ownership-qualified types are considered non-POD: in C++0x terms, they are not trivially default constructible, copy constructible, move constructible, copy assignable, move assignable, or destructible. It is a violation of C++ One Definition Rule to use a class outside of ARC that, under ARC, would have an ownership-qualified member.

Rationale: unlike in C, we can express all the necessary ARC semantics for ownership-qualified subobjects as suboperations of the (default) special member functions for the class. These functions then become non-trivial. This has the non-obvious repercussion that the class will have a non-trivial copy constructor and non-trivial destructor; if it wouldn't outside of ARC, this means that objects of the type will be passed and returned in an ABI-incompatible manner.

Ownership inference

Objects

If an object is declared with retainable object owner type, but without an explicit ownership qualifier, its type is implicitly adjusted to have __strong qualification.

As a special case, if the object's base type is Class (possibly protocol-qualified), the type is adjusted to have __unsafe_unretained qualification instead.

Indirect parameters

If a function or method parameter has type T*, where T is an ownership-unqualified retainable object pointer type, then:

  • if T is const-qualified or Class, then it is implicitly qualified with __unsafe_unretained;
  • otherwise, it is implicitly qualified with __autoreleasing.

Rationale: __autoreleasing exists mostly for this case, the Cocoa convention for out-parameters. Since a pointer to const is obviously not an out-parameter, we instead use a type more useful for passing arrays. If the user instead intends to pass in a mutable array, inferring __autoreleasing is the wrong thing to do; this directs some of the caution in the following rules about writeback.

Such a type written anywhere else would be ill-formed by the general rule requiring ownership qualifiers.

This rule does not apply in Objective-C++ if a parameter's type is dependent in a template pattern and is only instantiated to a type which would be a pointer to an unqualified retainable object pointer type. Such code is still ill-formed.

Rationale: the convention is very unlikely to be intentional in template code.

Template arguments

If a template argument for a template type parameter is an retainable object owner type that does not have an explicit ownership qualifier, it is adjusted to have __strong qualification. This adjustment occurs regardless of whether the template argument was deduced or explicitly specified.

Rationale: __strong is a useful default for containers (e.g., std::vector<id>), which would otherwise require explicit qualification. Moreover, unqualified retainable object pointer types are unlikely to be useful within templates, since they generally need to have a qualifier applied to the before being used.

Method families

An Objective-C method may fall into a method family, which is a conventional set of behaviors ascribed to it by the Cocoa conventions.

A method is in a certain method family if:

A selector is in a certain selector family if, ignoring any leading underscores, the first component of the selector either consists entirely of the name of the method family or it begins with that name followed by a character other than a lowercase letter. For example, _perform:with: and performWith: would fall into the perform family (if we recognized one), but performing:with would not.

The families and their added restrictions are:

A program is ill-formed if a method's declarations, implementations, and overrides do not all have the same method family.

Explicit method family control

A method may be annotated with the objc_method_family attribute to precisely control which method family it belongs to. If a method in an @implementation does not have this attribute, but there is a method declared in the corresponding @interface that does, then the attribute is copied to the declaration in the @implementation. The attribute is available outside of ARC, and may be tested for with the preprocessor query __has_attribute(objc_method_family).

The attribute is spelled __attribute__((objc_method_family(family))). If family is none, the method has no family, even if it would otherwise be considered to have one based on its selector and type. Otherwise, family must be one of alloc, copy, init, mutableCopy, or new, in which case the method is considered to belong to the corresponding family regardless of its selector. It is an error if a method that is explicitly added to a family in this way does not meet the requirements of the family other than the selector naming convention.

Rationale: the rules codified in this document describe the standard conventions of Objective-C. However, as these conventions have not heretofore been enforced by an unforgiving mechanical system, they are only imperfectly kept, especially as they haven't always even been precisely defined. While it is possible to define low-level ownership semantics with attributes like ns_returns_retained, this attribute allows the user to communicate semantic intent, which of use both to ARC (which, e.g., treats calls to init specially) and the static analyzer.

Semantics of method families

A method's membership in a method family may imply non-standard semantics for its parameters and return type.

Methods in the alloc, copy, mutableCopy, and new families — that is, methods in all the currently-defined families except init — implicitly return a retained object as if they were annotated with the ns_returns_retained attribute. This can be overridden by annotating the method with either of the ns_returns_autoreleased or ns_returns_not_retained attributes.

Semantics of init

Methods in the init family implicitly consume their self parameter and return a retained object. Neither of these properties can be altered through attributes.

A call to an init method with a receiver that is either self (possibly parenthesized or casted) or super is called a delegate init call. It is an error for a delegate init call to be made except from an init method, and excluding blocks within such methods.

As an exception to the usual rule, the variable self is mutable in an init method and has the usual semantics for a __strong variable. However, it is undefined behavior and the program is ill-formed, no diagnostic required, if an init method attempts to use the previous value of self after the completion of a delegate init call. It is conventional, but not required, for an init method to return self.

It is undefined behavior for a program to cause two or more calls to init methods on the same object, except that each init method invocation may perform at most one delegate init call.

Related result types

Certain methods are candidates to have related result types:

  • class methods in the alloc and new method families
  • instance methods in the init family
  • the instance method self
  • outside of ARC, the instance methods retain and autorelease

If the formal result type of such a method is id or protocol-qualified id, or a type equal to the declaring class or a superclass, then it is said to have a related result type. In this case, when invoked in an explicit message send, it is assumed to return a type related to the type of the receiver:

  • if it is a class method, and the receiver is a class name T, the message send expression has type T*; otherwise
  • if it is an instance method, and the receiver has type T, the message send expression has type T; otherwise
  • the message send expression has the normal result type of the method.

This is a new rule of the Objective-C language and applies outside of ARC.

Rationale: ARC's automatic code emission is more prone than most code to signature errors, i.e. errors where a call was emitted against one method signature, but the implementing method has an incompatible signature. Having more precise type information helps drastically lower this risks, as well as catching a number of latent bugs.

Optimization

ARC applies aggressive rules for the optimization of local behavior. These rules are based around a core assumption of local balancing: that other code will perform retains and releases as necessary (and only as necessary) for its own safety, and so the optimizer does not need to consider global properties of the retain and release sequence. For example, if a retain and release immediately bracket a call, the optimizer can delete the retain and release on the assumption that the called function will not do a constant number of unmotivated releases followed by a constant number of balancing retains, such that the local retain/release pair is the only thing preventing the called function from ending up with a dangling reference.

The optimizer assumes that when a new value enters local control, e.g. from a load of a non-local object or as the result of a function call, it is instaneously valid. Subsequently, a retain and release of a value are necessary on a computation path only if there is a use of that value before the release and after any operation which might cause a release of the value (including indirectly or non-locally), and only if the value is not demonstrably already retained.

The complete optimization rules are quite complicated, but it would still be useful to document them here.

Miscellaneous

Special methods

Memory management methods

A program is ill-formed if it contains a method definition, message send, or @selector expression for any of the following selectors:

  • autorelease
  • release
  • retain
  • retainCount

Rationale: retainCount is banned because ARC robs it of consistent semantics. The others were banned after weighing three options for how to deal with message sends:

Honoring them would work out very poorly if a programmer naively or accidentally tried to incorporate code written for manual retain/release code into an ARC program. At best, such code would do twice as much work as necessary; quite frequently, however, ARC and the explicit code would both try to balance the same retain, leading to crashes. The cost is losing the ability to perform unrooted retains, i.e. retains not logically corresponding to a strong reference in the object graph.

Ignoring them would badly violate user expectations about their code. While it would make it easier to develop code simultaneously for ARC and non-ARC, there is very little reason to do so except for certain library developers. ARC and non-ARC translation units share an execution model and can seamlessly interoperate. Within a translation unit, a developer who faithfully maintains their code in non-ARC mode is suffering all the restrictions of ARC for zero benefit, while a developer who isn't testing the non-ARC mode is likely to be unpleasantly surprised if they try to go back to it.

Banning them has the disadvantage of making it very awkward to migrate existing code to ARC. The best answer to that, given a number of other changes and restrictions in ARC, is to provide a specialized tool to assist users in that migration.

Implementing these methods was banned because they are too integral to the semantics of ARC; many tricks which worked tolerably under manual reference counting will misbehave if ARC performs an ephemeral extra retain or two. If absolutely required, it is still possible to implement them in non-ARC code, for example in a category; the implementations must obey the semantics laid out elsewhere in this document.

dealloc

A program is ill-formed if it contains a message send or @selector expression for the selector dealloc.

Rationale: there are no legitimate reasons to call dealloc directly.

A class may provide a method definition for an instance method named dealloc. This method will be called after the final release of the object but before it is deallocated or any of its instance variables are destroyed. The superclass's implementation of dealloc will be called automatically when the method returns.

Rationale: even though ARC destroys instance variables automatically, there are still legitimate reasons to write a dealloc method, such as freeing non-retainable resources. Failing to call [super dealloc] in such a method is nearly always a bug. Sometimes, the object is simply trying to prevent itself from being destroyed, but dealloc is really far too late for the object to be raising such objections. Somewhat more legitimately, an object may have been pool-allocated and should not be deallocated with free; for now, this can only be supported with a dealloc implementation outside of ARC. Such an implementation must be very careful to do all the other work that NSObject's dealloc would, which is outside the scope of this document to describe.

@autoreleasepool

To simplify the use of autorelease pools, and to bring them under the control of the compiler, a new kind of statement is available in Objective-C. It is written @autoreleasepool followed by a compound-statement, i.e. by a new scope delimited by curly braces. Upon entry to this block, the current state of the autorelease pool is captured. When the block is exited normally, whether by fallthrough or directed control flow (such as return or break), the autorelease pool is restored to the saved state, releasing all the objects in it. When the block is exited with an exception, the pool is not drained.

@autoreleasepool may be used in non-ARC translation units, with equivalent semantics.

A program is ill-formed if it refers to the NSAutoreleasePool class.

Rationale: autorelease pools are clearly important for the compiler to reason about, but it is far too much to expect the compiler to accurately reason about control dependencies between two calls. It is also very easy to accidentally forget to drain an autorelease pool when using the manual API, and this can significantly inflate the process's high-water-mark. The introduction of a new scope is unfortunate but basically required for sane interaction with the rest of the language. Not draining the pool during an unwind is apparently required by the Objective-C exceptions implementation.

self

The self parameter variable of an Objective-C method is never actually retained by the implementation. It is undefined behavior, or at least dangerous, to cause an object to be deallocated during a message send to that object. To make this safe, self is implicitly const unless the method is in the init family.

Rationale: the cost of retaining self in all methods was found to be prohibitive, as it tends to be live across calls, preventing the optimizer from proving that the retain and release are unnecessary — for good reason, as it's quite possible in theory to cause an object to be deallocated during its execution without this retain and release. Since it's extremely uncommon to actually do so, even unintentionally, and since there's no natural way for the programmer to remove this retain/release pair otherwise (as there is for other parameters by, say, making the variable __unsafe_unretained), we chose to make this optimizing assumption and shift some amount of risk to the user.

Fast enumeration iteration variables

If a variable is declared in the condition of an Objective-C fast enumeration loop, and the variable has no explicit ownership qualifier, then it is qualified with const __strong and objects encountered during the enumeration are not actually retained.

Rationale: this is an optimization made possible because fast enumeration loops promise to keep the objects retained during enumeration, and the collection itself cannot be synchronously modified. It can be overridden by explicitly qualifying the variable with __strong, which will make the variable mutable again and cause the loop to retain the objects it encounters.

Blocks

The implicit const capture variables created when evaluating a block literal expression have the same ownership semantics as the local variables they capture. The capture is performed by reading from the captured variable and initializing the capture variable with that value; the capture variable is destroyed when the block literal is, i.e. at the end of the enclosing scope.

The inference rules apply equally to __block variables, which is a shift in semantics from non-ARC, where __block variables did not implicitly retain during capture.

__block variables of retainable object owner type are moved off the stack by initializing the heap copy with the result of moving from the stack copy.

With the exception of retains done as part of initializing a __strong parameter variable or reading a __weak variable, whenever these semantics call for retaining a value of block-pointer type, it has the effect of a Block_copy. The optimizer may remove such copies when it sees that the result is used only as an argument to a call.

Exceptions

By default in Objective C, ARC is not exception-safe for normal releases:

  • It does not end the lifetime of __strong variables when their scopes are abnormally terminated by an exception.
  • It does not perform releases which would occur at the end of a full-expression if that full-expression throws an exception.

A program may be compiled with the option -fobjc-arc-exceptions in order to enable these, or with the option -fno-objc-arc-exceptions to explicitly disable them, with the last such argument winning.

Rationale: the standard Cocoa convention is that exceptions signal programmer error and are not intended to be recovered from. Making code exceptions-safe by default would impose severe runtime and code size penalties on code that typically does not actually care about exceptions safety. Therefore, ARC-generated code leaks by default on exceptions, which is just fine if the process is going to be immediately terminated anyway. Programs which do care about recovering from exceptions should enable the option.

In Objective-C++, -fobjc-arc-exceptions is enabled by default.

Rationale: C++ already introduces pervasive exceptions-cleanup code of the sort that ARC introduces. C++ programmers who have not already disabled exceptions are much more likely to actual require exception-safety.

ARC does end the lifetimes of __weak objects when an exception terminates their scope unless exceptions are disabled in the compiler.

Rationale: the consequence of a local __weak object not being destroyed is very likely to be corruption of the Objective-C runtime, so we want to be safer here. Of course, potentially massive leaks are about as likely to take down the process as this corruption is if the program does try to recover from exceptions.

Runtime support

This section describes the interaction between the ARC runtime and the code generated by the ARC compiler. This is not part of the ARC language specification; instead, it is effectively a language-specific ABI supplement, akin to the Itanium generic ABI for C++.

Ownership qualification does not alter the storage requirements for objects, except that it is undefined behavior if a __weak object is inadequately aligned for an object of type id. The other qualifiers may be used on explicitly under-aligned memory.

The runtime tracks __weak objects which holds non-null values. It is undefined behavior to direct modify a __weak object which is being tracked by the runtime except through an objc_storeWeak, objc_destroyWeak, or objc_moveWeak call.

The runtime must provide a number of new entrypoints which the compiler may emit, which are described in the remainder of this section.

Rationale: Several of these functions are semantically equivalent to a message send; we emit calls to C functions instead because:

  • the machine code to do so is significantly smaller,
  • it is much easier to recognize the C functions in the ARC optimizer, and
  • a sufficient sophisticated runtime may be able to avoid the message send in common cases.

Several other of these functions are fused operations which can be described entirely in terms of other operations. We use the fused operations primarily as a code-size optimization, although in some cases there is also a real potential for avoiding redundant operations in the runtime.

id objc_autorelease(id value);

Precondition: value is null or a pointer to a valid object.

If value is null, this call has no effect. Otherwise, it adds the object to the innermost autorelease pool exactly as if the object had been sent the autorelease message.

Always returns value.

void objc_autoreleasePoolPop(void *pool);

Precondition: pool is the result of a previous call to objc_autoreleasePoolPush on the current thread, where neither pool nor any enclosing pool have previously been popped.

Releases all the objects added to the given autorelease pool and any autorelease pools it encloses, then sets the current autorelease pool to the pool directly enclosing pool.

void *objc_autoreleasePoolPush(void);

Creates a new autorelease pool that is enclosed by the current pool, makes that the current pool, and returns an opaque handle to it.

Rationale: while the interface is described as an explicit hierarchy of pools, the rules allow the implementation to just keep a stack of objects, using the stack depth as the opaque pool handle.

id objc_autoreleaseReturnValue(id value);

Precondition: value is null or a pointer to a valid object.

If value is null, this call has no effect. Otherwise, it makes a best effort to hand off ownership of a retain count on the object to a call to objc_retainAutoreleasedReturnValue for the same object in an enclosing call frame. If this is not possible, the object is autoreleased as above.

Always returns value.

void objc_copyWeak(id *dest, id *src);

Precondition: src is a valid pointer which either contains a null pointer or has been registered as a __weak object. dest is a valid pointer which has not been registered as a __weak object.

dest is initialized to be equivalent to src, potentially registering it with the runtime. Equivalent to the following code:

void objc_copyWeak(id *dest, id *src) {
  objc_release(objc_initWeak(dest, objc_loadWeakRetained(src)));
}

Must be atomic with respect to calls to objc_storeWeak on src.

void objc_destroyWeak(id *object);

Precondition: object is a valid pointer which either contains a null pointer or has been registered as a __weak object.

object is unregistered as a weak object, if it ever was. The current value of object is left unspecified; otherwise, equivalent to the following code:

void objc_destroyWeak(id *object) {
  objc_storeWeak(object, nil);
}

Does not need to be atomic with respect to calls to objc_storeWeak on object.

id objc_initWeak(id *object, id value);

Precondition: object is a valid pointer which has not been registered as a __weak object. value is null or a pointer to a valid object.

If value is a null pointer or the object to which it points has begun deallocation, object is zero-initialized. Otherwise, object is registered as a __weak object pointing to value. Equivalent to the following code:

id objc_initWeak(id *object, id value) {
  *object = nil;
  return objc_storeWeak(object, value);
}

Returns the value of object after the call.

Does not need to be atomic with respect to calls to objc_storeWeak on object.

id objc_loadWeak(id *object);

Precondition: object is a valid pointer which either contains a null pointer or has been registered as a __weak object.

If object is registered as a __weak object, and the last value stored into object has not yet been deallocated or begun deallocation, retains and autoreleases that value and returns it. Otherwise returns null. Equivalent to the following code:

id objc_loadWeak(id *object) {
  return objc_autorelease(objc_loadWeakRetained(object));
}

Must be atomic with respect to calls to objc_storeWeak on object.

Rationale: loading weak references would be inherently prone to race conditions without the retain.

id objc_loadWeakRetained(id *object);

Precondition: object is a valid pointer which either contains a null pointer or has been registered as a __weak object.

If object is registered as a __weak object, and the last value stored into object has not yet been deallocated or begun deallocation, retains that value and returns it. Otherwise returns null.

Must be atomic with respect to calls to objc_storeWeak on object.

void objc_moveWeak(id *dest, id *src);

Precondition: src is a valid pointer which either contains a null pointer or has been registered as a __weak object. dest is a valid pointer which has not been registered as a __weak object.

dest is initialized to be equivalent to src, potentially registering it with the runtime. src may then be left in its original state, in which case this call is equivalent to objc_copyWeak, or it may be left as null.

Must be atomic with respect to calls to objc_storeWeak on src.

void objc_release(id value);

Precondition: value is null or a pointer to a valid object.

If value is null, this call has no effect. Otherwise, it performs a release operation exactly as if the object had been sent the release message.

id objc_retain(id value);

Precondition: value is null or a pointer to a valid object.

If value is null, this call has no effect. Otherwise, it performs a retain operation exactly as if the object had been sent the retain message.

Always returns value.

id objc_retainAutorelease(id value);

Precondition: value is null or a pointer to a valid object.

If value is null, this call has no effect. Otherwise, it performs a retain operation followed by an autorelease operation. Equivalent to the following code:

id objc_retainAutorelease(id value) {
  return objc_autorelease(objc_retain(value));
}

Always returns value.

id objc_retainAutoreleaseReturnValue(id value);

Precondition: value is null or a pointer to a valid object.

If value is null, this call has no effect. Otherwise, it performs a retain operation followed by the operation described in objc_autoreleaseReturnValue. Equivalent to the following code:

id objc_retainAutoreleaseReturnValue(id value) {
  return objc_autoreleaseReturnValue(objc_retain(value));
}

Always returns value.

id objc_retainAutoreleasedReturnValue(id value);

Precondition: value is null or a pointer to a valid object.

If value is null, this call has no effect. Otherwise, it attempts to accept a hand off of a retain count from a call to objc_autoreleaseReturnValue on value in a recently-called function or something it calls. If that fails, it performs a retain operation exactly like objc_retain.

Always returns value.

id objc_retainBlock(id value);

Precondition: value is null or a pointer to a valid block object.

If value is null, this call has no effect. Otherwise, if the block pointed to by value is still on the stack, it is copied to the heap and the address of the copy is returned. Otherwise a retain operation is performed on the block exactly as if it had been sent the retain message.

id objc_storeStrong(id *object, id value);

Precondition: object is a valid pointer to a __strong object which is adequately aligned for a pointer. value is null or a pointer to a valid object.

Performs the complete sequence for assigning to a __strong object of non-block type. Equivalent to the following code:

id objc_storeStrong(id *object, id value) {
  value = [value retain];
  id oldValue = *object;
  *object = value;
  [oldValue release];
  return value;
}

Always returns value.

id objc_storeWeak(id *object, id value);

Precondition: object is a valid pointer which either contains a null pointer or has been registered as a __weak object. value is null or a pointer to a valid object.

If value is a null pointer or the object to which it points has begun deallocation, object is assigned null and unregistered as a __weak object. Otherwise, object is registered as a __weak object or has its registration updated to point to value.

Returns the value of object after the call.