Blame - docs/AutomaticReferenceCounting.html - platform/external/clang

blob: 1150614254f85f5a03967827057ddab7169e6113 [file] [log] [blame]

John McCall	8246702	2011-06-15 21:21:53 +0000	[diff] [blame]	1	<html>
				2	<head>
				3	<title>Objective-C Automatic Reference Counting (ARC)</title>
				4	<link type="text/css" rel="stylesheet" href="../menu.css" />
				5	<link type="text/css" rel="stylesheet" href="../content.css" />
				6	<style type="text/css">
				7	/* Collapse the items in the ToC to the left. */
				8	div#toc ul {
				9	padding-left: 0
				10	}
				11
				12	/* Rationales appear in italic. */
				13	div.rationale {
				14	font-style: italic
				15	}
				16
John McCall	f3d08a6	2011-06-18 07:31:30 +0000	[diff] [blame^]	17	div.rationale em {
				18	font-style: normal
				19	}
				20
John McCall	8246702	2011-06-15 21:21:53 +0000	[diff] [blame]	21	div h1 { font-size: 2em; margin: .67em 0 }
				22	div div h1 { font-size: 1.5em; margin: .75em 0 }
				23	div div div h1 { font-size: 1.17em; margin: .83em 0 }
				24	div div div div h1 { margin: 1.12em 0 }
				25
				26	span.term { font-style: italic; font-weight: bold }
				27	</style>
				28
				29	<script lang="javascript">
				30	/// A little script to recursively build a table of contents.
				31	function buildTOC(div, toc, ancestry) {
				32	var children = div.childNodes;
				33	var len = children.length;
				34
				35	var childNumber = 0;
				36
				37	var list = null;
				38	for (var i = 0; i < len; ++i) {
				39	var child = children[i];
				40	if (child.nodeName != "DIV") continue;
				41	if (child.getAttribute("class") == "rationale") continue;
				42	if (child.id == "toc") continue;
				43
				44	// Okay, we're actually going to build a list node.
				45	if (list === null) list = document.createElement("ul");
				46
				47	var childAncestry = ancestry + ++childNumber + ".";
				48
				49	var headerNode = child.childNodes[1];
				50	var title = headerNode.innerHTML;
				51	headerNode.insertBefore(document.createTextNode(childAncestry + " "),
				52	headerNode.firstChild);
				53
				54	var item = document.createElement("li");
				55	item.appendChild(document.createTextNode(childAncestry + " "));
				56
				57	var anchor = document.createElement("a");
				58	anchor.href = "#" + child.id;
				59	anchor.innerHTML = title;
				60	item.appendChild(anchor);
				61
				62	buildTOC(child, item, childAncestry);
				63
				64	list.appendChild(item);
				65	}
				66	if (list) toc.appendChild(list);
				67	}
				68
				69	function onLoad() {
				70	var toc = document.getElementById("toc");
				71	var content = document.getElementById("content");
				72	buildTOC(content, toc, "");
				73	}
				74	window.onload = onLoad;
				75
				76	</script>
				77	</head>
				78	<body>
				79
				80	<!--#include virtual="../menu.html.incl"-->
				81
				82	<div id="content">
				83	<h1>Automatic Reference Counting</h1>
				84
				85	<div id="toc">
				86	</div>
				87
				88	<div id="meta">
				89	<h1>About this document</h1>
				90
				91	<div id="meta.purpose">
				92	<h1>Purpose</h1>
				93
				94	<p>The first and primary purpose of this document is to serve as a
				95	complete technical specification of Automatic Reference Counting.
				96	Given a core Objective-C compiler and runtime, it should be possible
				97	to write a compiler and runtime which implements these new
				98	semantics.</p>
				99
				100	<p>The secondary purpose is to act as a rationale for why ARC was
				101	designed in this way. This should remain tightly focused on the
				102	technical design and should not stray into marketing speculation.</p>
				103
				104	</div> <!-- meta.purpose -->
				105
				106	<div id="meta.background">
				107	<h1>Background</h1>
				108
				109	<p>This document assumes a basic familiarity with C.</p>
				110
				111	<p><span class="term">Blocks</span> are a C language extension for
				112	creating anonymous functions. Users interact with and transfer block
				113	objects using <span class="term">block pointers</span>, which are
				114	represented like a normal pointer. A block may capture values from
				115	local variables; when this occurs, memory must be dynamically
				116	allocated. The initial allocation is done on the stack, but the
				117	runtime provides a <tt>Block_copy</tt> function which, given a block
				118	pointer, either copies the underlying block object to the heap,
				119	setting its reference count to 1 and returning the new block pointer,
				120	or (if the block object is already on the heap) increases its
				121	reference count by 1. The paired function is <tt>Block_release</tt>,
				122	which decreases the reference count by 1 and destroys the object if
				123	the count reaches zero and is on the heap.</p>
				124
				125	<p>Objective-C is a set of language extensions, significant enough to
				126	be considered a different language. It is a strict superset of C.
				127	The extensions can also be imposed on C++, producing a language called
				128	Objective-C++. The primary feature is a single-inheritance object
				129	system; we briefly describe the modern dialect.</p>
				130
				131	<p>Objective-C defines a new type kind, collectively called
				132	the <span class="term">object pointer types</span>. This kind has two
				133	notable builtin members, <tt>id</tt> and <tt>Class</tt>; <tt>id</tt>
				134	is the final supertype of all object pointers. The validity of
				135	conversions between object pointer types is not checked at runtime.
				136	Users may define <span class="term">classes</span>; each class is a
				137	type, and the pointer to that type is an object pointer type. A class
				138	may have a superclass; its pointer type is a subtype of its
				139	superclass's pointer type. A class has a set
				140	of <span class="term">ivars</span>, fields which appear on all
				141	instances of that class. For every class <i>T</i> there's an
				142	associated metaclass; it has no fields, its superclass is the
				143	metaclass of <i>T</i>'s superclass, and its metaclass is a global
				144	class. Every class has a global object whose class is the
				145	class's metaclass; metaclasses have no associated type, so pointers to
				146	this object have type <tt>Class</tt>.</p>
				147
				148	<p>A class declaration (<tt>@interface</tt>) declares a set
				149	of <span class="term">methods</span>. A method has a return type, a
				150	list of argument types, and a <span class="term">selector</span>: a
				151	name like <tt>foo:bar:baz:</tt>, where the number of colons
				152	corresponds to the number of formal arguments. A method may be an
				153	instance method, in which case it can be invoked on objects of the
				154	class, or a class method, in which case it can be invoked on objects
				155	of the metaclass. A method may be invoked by providing an object
				156	(called the <span class="term">receiver</span>) and a list of formal
				157	arguments interspersed with the selector, like so:</p>
				158
				159	<pre>[receiver foo: fooArg bar: barArg baz: bazArg]</pre>
				160
				161	<p>This looks in the dynamic class of the receiver for a method with
				162	this name, then in that class's superclass, etc., until it finds
				163	something it can execute. The receiver <q>expression</q> may also be
				164	the name of a class, in which case the actual receiver is the class
				165	object for that class, or (within method definitions) it may
				166	be <tt>super</tt>, in which case the lookup algorithm starts with the
				167	static superclass instead of the dynamic class. The actual methods
				168	dynamically found in a class are not those declared in the
				169	<tt>@interface</tt>, but those defined in a separate
				170	<tt>@implementation</tt> declaration; however, when compiling a
				171	call, typechecking is done based on the methods declared in the
				172	<tt>@interface</tt>.</p>
				173
				174	<p>Method declarations may also be grouped into
				175	<span class="term">protocols</span>, which are not inherently
				176	associated with any class, but which classes may claim to follow.
				177	Object pointer types may be qualified with additional protocols that
				178	the object is known to support.</p>
				179
				180	<p><span class="term">Class extensions</span> are collections of ivars
				181	and methods, designed to allow a class's <tt>@interface</tt> to be
				182	split across multiple files; however, there is still a primary
				183	implementation file which must see the <tt>@interface</tt>s of all
				184	class extensions.
				185	<span class="term">Categories</span> allow methods (but not ivars) to
				186	be declared <i>post hoc</i> on an arbitrary class; the methods in the
				187	category's <tt>@implementation</tt> will be dynamically added to that
				188	class's method tables which the category is loaded at runtime,
				189	replacing those methods in case of a collision.</p>
				190
				191	<p>In the standard environment, objects are allocated on the heap, and
				192	their lifetime is manually managed using a reference count. This is
				193	done using two instance methods which all classes are expected to
				194	implement: <tt>retain</tt> increases the object's reference count by
				195	1, whereas <tt>release</tt> decreases it by 1 and calls the instance
				196	method <tt>dealloc</tt> if the count reaches 0. To simplify certain
				197	operations, there is also an <span class="term">autorelease
				198	pool</span>, a thread-local list of objects to call <tt>release</tt>
				199	on later; an object can be added to this pool by
				200	calling <tt>autorelease</tt> on it.</p>
				201
				202	<p>Block pointers may be converted to type <tt>id</tt>; block objects
				203	are laid out in a way that makes them compatible with Objective-C
				204	objects. There is a builtin class that all block objects are
				205	considered to be objects of; this class implements <tt>retain</tt> by
				206	adjusting the reference count, not by calling <tt>Block_copy</tt>.</p>
				207
				208	</div> <!-- meta.background -->
				209
				210	</div> <!-- meta -->
				211
				212	<div id="general">
				213	<h1>General</h1>
				214
				215	<p>Automatic Reference Counting implements automatic memory management
				216	for Objective-C objects and blocks, freeing the programmer from the
				217	need explicitly insert retains and releases. It does not provide a
				218	cycle collector; users must explicitly manage lifetime instead.</p>
				219
				220	<p>ARC may be explicitly enabled with the compiler
				221	flag <tt>-fobjc-arc</tt>. It may also be explicitly disabled with the
				222	compiler flag <tt>-fno-objc-arc</tt>. The last of these two flags
				223	appearing on the compile line <q>wins</q>.</p>
				224
				225	<p>If ARC is enabled, <tt>__has_feature(objc_arc)</tt> will expand to
				226	1 in the preprocessor. For more information about <tt>__has_feature</tt>,
				227	see the <a href="LanguageExtensions.html#__has_feature_extension">language
				228	extensions</a> document.</p>
				229
				230	</div>
				231
				232	<div id="objects">
				233	<h1>Retainable object pointers</h1>
				234
				235	<p>This section describes retainable object pointers, their basic
				236	operations, and the restrictions imposed on their use under ARC. Note
				237	in particular that it covers the rules for pointer <em>values</em>
				238	(patterns of bits indicating the location of a pointed-to object), not
				239	pointer
				240	<em>objects</em> (locations in memory which store pointer values).
				241	The rules for objects are covered in the next section.</p>
				242
				243	<p>A <span class="term">retainable object pointer</span>
				244	(or <q>retainable pointer</q>) is a value of
				245	a <span class="term">retainable object pointer type</span>
				246	(<q>retainable type</q>). There are three kinds of retainable object
				247	pointer types:</p>
				248	<ul>
				249	<li>block pointers (formed by applying the caret (<tt>^</tt>)
				250	declarator sigil to a function type)</li>
				251	<li>Objective-C object pointers (<tt>id</tt>, <tt>Class</tt>, <tt>NSFoo*</tt>, etc.)</li>
				252	<li>typedefs marked with <tt>__attribute__((NSObject))</tt></li>
				253	</ul>
				254
				255	<p>Other pointer types, such as <tt>int*</tt> and <tt>CFStringRef</tt>,
				256	are not subject to ARC's semantics and restrictions.</p>
				257
				258	<div class="rationale">
				259
				260	<p>Rationale: We are not at liberty to require
				261	all code to be recompiled with ARC; therefore, ARC must interoperate
				262	with Objective-C code which manages retains and releases manually. In
				263	general, there are three requirements in order for a
				264	compiler-supported reference-count system to provide reliable
				265	interoperation:</p>
				266
				267	<ul>
				268	<li>The type system must reliably identify which objects are to be
				269	managed. An <tt>int*</tt> might be a pointer to a <tt>malloc</tt>'ed
				270	array, or it might be a interior pointer to such an array, or it might
				271	point to some field or local variable. In contrast, values of the
				272	retainable object pointer types are never interior.</li>
				273	<li>The type system must reliably indicate how to
				274	manage objects of a type. This usually means that the type must imply
				275	a procedure for incrementing and decrementing retain counts.
				276	Supporting single-ownership objects requires a lot more explicit
				277	mediation in the language.</li>
				278	<li>There must be reliable conventions for whether and
				279	when <q>ownership</q> is passed between caller and callee, for both
				280	arguments and return values. Objective-C methods follow such a
				281	convention very reliably, at least for system libraries on Mac OS X,
				282	and functions always pass objects at +0. The C-based APIs for Core
				283	Foundation objects, on the other hand, have much more varied transfer
				284	semantics.</li>
				285	</ul>
				286	</div> <!-- rationale -->
				287
				288	<p>The use of <tt>__attribute__((NSObject))</tt> typedefs is not
				289	recommended. If it's absolutely necessary to use this attribute, be
				290	very explicit about using the typedef, and do not assume that it will
				291	be preserved by language features like <tt>__typeof</tt> and C++
				292	template argument substitution.</p>
				293
				294	<div class="rationale"><p>Rationale: any compiler operation which
				295	incidentally strips type <q>sugar</q> from a type will yield a type
				296	without the attribute, which may result in unexpected
				297	behavior.</p></div>
				298
				299	<div id="objects.retains">
				300	<h1>Retain count semantics</h1>
				301
				302	<p>A retainable object pointer is either a <span class="term">null
				303	pointer</span> or a pointer to a valid object. Furthermore, if it has
				304	block pointer type and is not <tt>null</tt> then it must actually be a
				305	pointer to a block object, and if it has <tt>Class</tt> type (possibly
				306	protocol-qualified) then it must actually be a pointer to a class
				307	object. Otherwise ARC does not enforce the Objective-C type system as
				308	long as the implementing methods follow the signature of the static
				309	type. It is undefined behavior if ARC is exposed to an invalid
				310	pointer.</p>
				311
				312	<p>For ARC's purposes, a valid object is one with <q>well-behaved</q>
				313	retaining operations. Specifically, the object must be laid out such
				314	that the Objective-C message send machinery can successfully send it
				315	the following messages:</p>
				316
				317	<ul>
				318	<li><tt>retain</tt>, taking no arguments and returning a pointer to
				319	the object.</li>
				320	<li><tt>release</tt>, taking no arguments and returning <tt>void</tt>.</li>
				321	<li><tt>autorelease</tt>, taking no arguments and returning a pointer
				322	to the object.</li>
				323	</ul>
				324
				325	<p>The behavior of these methods is constrained in the following ways.
				326	The term <span class="term">high-level semantics</span> is an
				327	intentionally vague term; the intent is that programmers must
				328	implement these methods in a way such that the compiler, modifying
				329	code in ways it deems safe according to these constraints, will not
				330	violate their requirements. For example, if the user puts logging
				331	statements in <tt>retain</tt>, they should not be surprised if those
				332	statements are executed more or less often depending on optimization
				333	settings. These constraints are not exhaustive of the optimization
				334	opportunities: values held in local variables are subject to
				335	additional restrictions, described later in this document.</p>
				336
				337	<p>It is undefined behavior if a computation history featuring a send
				338	of <tt>retain</tt> followed by a send of <tt>release</tt> to the same
				339	object, with no intervening <tt>release</tt> on that object, is not
				340	equivalent under the high-level semantics to a computation
				341	history in which these sends are removed. Note that this implies that
				342	these methods may not raise exceptions.</p>
				343
				344	<p>It is undefined behavior if a computation history features any use
				345	whatsoever of an object following the completion of a send
				346	of <tt>release</tt> that is not preceded by a send of <tt>retain</tt>
				347	to the same object.</p>
				348
				349	<p>The behavior of <tt>autorelease</tt> must be equivalent to sending
				350	<tt>release</tt> when one of the autorelease pools currently in scope
				351	is popped. It may not throw an exception.</p>
				352
				353	<p>When the semantics call for performing one of these operations on a
				354	retainable object pointer, if that pointer is <tt>null</tt> then the
				355	effect is a no-op.</p>
				356
				357	<p>All of the semantics described in this document are subject to
				358	additional <a href="#optimization">optimization rules</a> which permit
				359	the removal or optimization of operations based on local knowledge of
				360	data flow. The semantics describe the high-level behaviors that the
				361	compiler implements, not an exact sequence of operations that a
				362	program will be compiled into.</p>
				363
				364	</div> <!-- objects.retains -->
				365
				366	<div id="objects.operands">
				367	<h1>Retainable object pointers as operands and arguments</h1>
				368
				369	<p>In general, ARC does not perform retain or release operations when
				370	simply using a retainable object pointer as an operand within an
				371	expression. This includes:</p>
				372	<ul>
				373	<li>loading a retainable pointer from an object with non-weak
				374	<a href="#ownership">ownership</a>,</li>
				375	<li>passing a retainable pointer as an argument to a function or
				376	method, and</li>
				377	<li>receiving a retainable pointer as the result of a function or
				378	method call.</li>
				379	</ul>
				380
				381	<div class="rationale"><p>Rationale: while this might seem
				382	uncontroversial, it is actually unsafe when multiple expressions are
				383	evaluated in <q>parallel</q>, as with binary operators and calls,
				384	because (for example) one expression might load from an object while
				385	another writes to it. However, C and C++ already call this undefined
				386	behavior because the evaluations are unsequenced, and ARC simply
				387	exploits that here to avoid needing to retain arguments across a large
				388	number of calls.</p></div>
				389
				390	<p>The remainder of this section describes exceptions to these rules,
				391	how those exceptions are detected, and what those exceptions imply
				392	semantically.</p>
				393
				394	<div id="objects.operands.consumed">
				395	<h1>Consumed parameters</h1>
				396
				397	<p>A function or method parameter of retainable object pointer type
				398	may be marked as <span class="term">consumed</span>, signifying that
				399	the callee expects to take ownership of a +1 retain count. This is
				400	done by adding the <tt>ns_consumed</tt> attribute to the parameter
				401	declaration, like so:</p>
				402
				403	<pre>void foo(__attribute((ns_consumed)) id x);
				404	- (void) foo: (id) __attribute((ns_consumed)) x;</pre>
				405
				406	<p>This attribute is part of the type of the function or method, not
				407	the type of the parameter. It controls only how the argument is
				408	passed and received.</p>
				409
				410	<p>When passing such an argument, ARC retains the argument prior to
				411	making the call.</p>
				412
				413	<p>When receiving such an argument, ARC releases the argument at the
				414	end of the function, subject to the usual optimizations for local
				415	values.</p>
				416
				417	<div class="rationale"><p>Rationale: this formalizes direct transfers
				418	of ownership from a caller to a callee. The most common scenario here
				419	is passing the <tt>self</tt> parameter to <tt>init</tt>, but it is
				420	useful to generalize. Typically, local optimization will remove any
				421	extra retains and releases: on the caller side the retain will be
				422	merged with a +1 source, and on the callee side the release will be
				423	rolled into the initialization of the parameter.</p></div>
				424
				425	<p>The implicit <tt>self</tt> parameter of a method may be marked as
				426	consumed by adding <tt>__attribute__((ns_consumes_self))</tt> to the
				427	method declaration. Methods in
				428	the <tt>init</tt> <a href="#family">family</a> are implicitly
				429	marked <tt>__attribute__((ns_consumes_self))</tt>.</p>
				430
				431	<p>It is undefined behavior if an Objective-C message send of a method
				432	with <tt>ns_consumed</tt> parameters (other than self) is made to a
				433	null pointer.</p>
				434
				435	<div class="rationale"><p>Rationale: in fact, it's probably a
				436	guaranteed leak.</p></div>
				437
				438	</div>
				439
				440	<div id="objects.operands.retained-returns">
				441	<h1>Retained return values</h1>
				442
				443	<p>A function or method which returns a retainable object pointer type
				444	may be marked as returning a retained value, signifying that the
				445	caller expects to take ownership of a +1 retain count. This is done
				446	by adding the <tt>ns_returns_retained</tt> attribute to the function or
				447	method declaration, like so:</p>
				448
				449	<pre>id foo(void) __attribute((ns_returns_retained));
				450	- (id) foo __attribute((ns_returns_retained));</pre>
				451
				452	<p>This attribute is part of the type of the function or method.</p>
				453
				454	<p>When returning from such a function or method, ARC retains the
				455	value at the point of evaluation of the return statement, before
				456	leaving all local scopes.</p>
				457
				458	<p>When receiving a return result from such a function or method, ARC
				459	releases the value at the end of the full-expression it is contained
				460	within, subject to the usual optimizations for local values.</p>
				461
				462	<div class="rationale"><p>Rationale: this formalizes direct transfers of
				463	ownership from a callee to a caller. The most common scenario this
				464	models is the retained return from <tt>init</tt>, <tt>alloc</tt>,
				465	<tt>new</tt>, and <tt>copy</tt> methods, but there are other cases in
				466	the frameworks. After optimization there are typically no extra
				467	retains and releases required.</p></div>
				468
				469	<p>Methods in
				470	the <tt>alloc</tt>, <tt>copy</tt>, <tt>init</tt>, <tt>mutableCopy</tt>,
				471	and <tt>new</tt> <a href="#family">families</a> are implicitly marked
				472	<tt>__attribute__((ns_returns_retained))</tt>. This may be suppressed
				473	by explicitly marking the
				474	method <tt>__attribute__((ns_returns_not_retained))</tt>.</p>
				475	</div>
				476
				477	<div id="objects.operands.other-returns">
				478	<h1>Unretained return values</h1>
				479
				480	<p>A method or function which returns a retainable object type but
				481	does not return a retained value must ensure that the object is
				482	still valid across the return boundary.</p>
				483
				484	<p>When returning from such a function or method, ARC retains the
				485	value at the point of evaluation of the return statement, then leaves
				486	all local scopes, and then balances out the retain while ensuring that
				487	the value lives across the call boundary. In the worst case, this may
				488	involve an <tt>autorelease</tt>, but callers must not assume that the
				489	value is actually in the autorelease pool.</p>
				490
				491	<p>ARC performs no extra mandatory work on the caller side, although
				492	it may elect to do something to shorten the lifetime of the returned
				493	value.</p>
				494
				495	<div class="rationale"><p>Rationale: it is common in non-ARC code to not
				496	return an autoreleased value; therefore the convention does not force
				497	either path. It is convenient to not be required to do unnecessary
				498	retains and autoreleases; this permits optimizations such as eliding
				499	retain/autoreleases when it can be shown that the original pointer
				500	will still be valid at the point of return.</p></div>
				501
				502	<p>A method or function may be marked
				503	with <tt>__attribute__((ns_returns_autoreleased))</tt> to indicate
				504	that it returns a pointer which is guaranteed to be valid at least as
				505	long as the innermost autorelease pool. There are no additional
				506	semantics enforced in the definition of such a method; it merely
				507	enables optimizations in callers.</p>
				508	</div>
				509
				510	<div id="objects.operands.casts">
				511	<h1>Bridged casts</h1>
				512
				513	<p>A <span class="term">bridged cast</span> is a C-style cast
				514	annotated with one of three keywords:</p>
				515
				516	<ul>
				517	<li><tt>(__bridge T) op</tt> casts the operand to the destination
				518	type <tt>T</tt>. If <tt>T</tt> is a retainable object pointer type,
				519	then <tt>op</tt> must have a non-retainable pointer type.
				520	If <tt>T</tt> is a non-retainable pointer type, then <tt>op</tt> must
				521	have a retainable object pointer type. Otherwise the cast is
				522	ill-formed. There is no transfer of ownership, and ARC inserts
				523	no retain operations.</li>
				524
				525	<li><tt>(__bridge_retained T) op</tt> casts the operand, which must
				526	have retainable object pointer type, to the destination type, which
				527	must be a non-retainable pointer type. ARC retains the value, subject
				528	to the usual optimizations on local values, and the recipient is
				529	responsible for balancing that +1.</li>
				530
				531	<li><tt>(__bridge_transfer T) op</tt> casts the operand, which must
				532	have non-retainable pointer type, to the destination type, which must
				533	be a retainable object pointer type. ARC will release the value at
				534	the end of the enclosing full-expression, subject to the usual
				535	optimizations on local values.</li>
				536	</ul>
				537
				538	<p>These casts are required in order to transfer objects in and out of
				539	ARC control; see the rationale in the section
				540	on <a href="#objects.restrictions.conversion">conversion of retainable
				541	object pointers</a>.</p>
				542
				543	<p>Using a <tt>__bridge_retained</tt> or <tt>__bridge_transfer</tt>
				544	cast purely to convince ARC to emit an unbalanced retain or release,
				545	respectively, is poor form.</p>
				546
				547	</div>
				548
				549	</div>
				550
				551	<div id="objects.restrictions">
				552	<h1>Restrictions</h1>
				553
				554	<div id="objects.restrictions.conversion">
				555	<h1>Conversion of retainable object pointers</h1>
				556
				557	<p>In general, a program which attempts to implicitly or explicitly
				558	convert a value of retainable object pointer type to any
				559	non-retainable type, or vice-versa, is ill-formed. For example, an
				560	Objective-C object pointer shall not be converted to <tt>intptr_t</tt>
				561	or <tt>void*</tt>. The <a href="#objects.operands.casts">bridged
				562	casts</a> may be used to perform these conversions where
				563	necessary.</p>
				564
				565	<div class="rationale"><p>Rationale: we cannot ensure the correct
				566	management of the lifetime of objects if they may be freely passed
				567	around as unmanaged types. The bridged casts are provided so that the
				568	programmer may explicitly describe whether the cast transfers control
				569	into or out of ARC.</p></div>
				570	</div>
				571
				572	<p>An unbridged cast to a retainable object pointer type of the return
				573	value of a Objective-C message send which yields a non-retainable
				574	pointer is treated as a <tt>__bridge_transfer</tt> cast
				575	if:</p>
				576
				577	<ul>
				578	<li>the method has the <tt>cf_returns_retained</tt> attribute, or if
				579	not that,</li>
				580	<li>the method does not have the <tt>cf_returns_not_retained</tt>
				581	attribute and</li>
				582	<li>the method's <a href="#family">selector family</a> would imply
				583	the <tt>ns_returns_retained</tt> attribute on a method which returned
				584	a retainable object pointer type.</li>
				585	</ul>
				586
				587	<p>Otherwise the cast is treated as a <tt>__bridge</tt> cast.</p>
				588
				589	</div>
				590
				591	</div>
				592
				593	<div id="ownership">
				594	<h1>Ownership qualification</h1>
				595
				596	<p>This section describes the behavior of <em>objects</em> of
				597	retainable object pointer type; that is, locations in memory which
				598	store retainable object pointers.</p>
				599
				600	<p>A type is a <span class="term">retainable object owner type</span>
				601	if it is a retainable object pointer type or an array type whose
				602	element type is a retainable object owner type.</p>
				603
				604	<p>An <span class="term">ownership qualifier</span> is a type
Douglas Gregor	4020cae	2011-06-17 23:16:24 +0000	[diff] [blame]	605	qualifier which applies only to retainable object owner types. An array type is
				606	ownership-qualified according to its element type, and adding an ownership
				607	qualifier to an array type so qualifies its element type.</p>
				608
				609	<p>A program is ill-formed if it attempts to apply an ownership qualifier
John McCall	8246702	2011-06-15 21:21:53 +0000	[diff] [blame]	610	to a type which is already ownership-qualified, even if it is the same
Douglas Gregor	4020cae	2011-06-17 23:16:24 +0000	[diff] [blame]	611	qualifier. There is a single exception to this rule: an ownership qualifier
				612	may be applied to a substituted template type parameter, which overrides the
				613	ownership qualifier provided by the template argument.</p>
John McCall	8246702	2011-06-15 21:21:53 +0000	[diff] [blame]	614
				615	<p>Except as described under
				616	the <a href="#ownership.inference">inference rules</a>, a program is
				617	ill-formed if it attempts to form a pointer or reference type to a
				618	retainable object owner type which lacks an ownership qualifier.</p>
				619
				620	<div class="rationale"><p>Rationale: these rules, together with the
				621	inference rules, ensure that all objects and lvalues of retainable
Douglas Gregor	4020cae	2011-06-17 23:16:24 +0000	[diff] [blame]	622	object pointer type have an ownership qualifier. The ability to override an ownership qualifier during template substitution is required to counteract the <a href="#ownership.inference.template_arguments">inference of <tt>__strong</tt> for template type arguments</a>. </p></div>
John McCall	8246702	2011-06-15 21:21:53 +0000	[diff] [blame]	623
				624	<p>There are four ownership qualifiers:</p>
				625
				626	<ul>
				627	<li><tt>__autoreleasing</tt></li>
				628	<li><tt>__strong</tt></li>
				629	<li><tt>__unsafe_unretained</tt></li>
				630	<li><tt>__weak</tt></li>
				631	</ul>
				632
				633	<p>A type is <span class="term">nontrivially ownership-qualified</span>
				634	if it is qualified with <tt>__autoreleasing</tt>, <tt>__strong</tt>, or
				635	<tt>__weak</tt>.</p>
				636
				637	<div id="ownership.spelling">
				638	<h1>Spelling</h1>
				639
				640	<p>The names of the ownership qualifiers are reserved for the
				641	implementation. A program may not assume that they are or are not
				642	implemented with macros, or what those macros expand to.</p>
				643
				644	<p>An ownership qualifier may be written anywhere that any other type
				645	qualifier may be written.</p>
				646
				647	<p>If an ownership qualifier appears in
				648	the <i>declaration-specifiers</i>, the following rules apply:</p>
				649
				650	<ul>
				651	<li>if the type specifier is a retainable object owner type, the
				652	qualifier applies to that type;</li>
				653	<li>if the outermost non-array part of the declarator is a pointer or
				654	block pointer, the qualifier applies to that type;</li>
				655	<li>otherwise the program is ill-formed.</li>
				656	</ul>
				657
				658	<p>If an ownership qualifier appears on the declarator name, or on the
				659	declared object, it is applied to outermost pointer or block-pointer
				660	type.</p>
				661
				662	<p>If an ownership qualifier appears anywhere else in a declarator, it
				663	applies to the type there.</p>
				664
				665	</div> <!-- ownership.spelling -->
				666
				667	<div id="ownership.semantics">
				668	<h1>Semantics</h1>
				669
				670	<p>There are five <span class="term">managed operations</span> which
				671	may be performed on an object of retainable object pointer type. Each
				672	qualifier specifies different semantics for each of these operations.
				673	It is still undefined behavior to access an object outside of its
				674	lifetime.</p>
				675
				676	<p>A load or store with <q>primitive semantics</q> has the same
				677	semantics as the respective operation would have on an <tt>void*</tt>
				678	lvalue with the same alignment and non-ownership qualification.</p>
				679
				680	<p><span class="term">Reading</span> occurs when performing a
				681	lvalue-to-rvalue conversion on an object lvalue.
				682
				683	<ul>
				684	<li>For <tt>__weak</tt> objects, the current pointee is retained and
				685	then released at the end of the current full-expression. This must
				686	execute atomically with respect to assignments and to the final
				687	release of the pointee.</li>
				688	<li>For all other objects, the lvalue is loaded with primitive
				689	semantics.</li>
				690	</ul>
				691	</p>
				692
				693	<p><span class="term">Assignment</span> occurs when evaluating
				694	an assignment operator. The semantics vary based on the qualification:
				695	<ul>
				696	<li>For <tt>__strong</tt> objects, the new pointee is first retained;
				697	second, the lvalue is loaded with primitive semantics; third, the new
				698	pointee is stored into the lvalue with primitive semantics; and
				699	finally, the old pointee is released. This is not performed
				700	atomically; external synchronization must be used to make this safe in
				701	the face of concurrent loads and stores.</li>
				702	<li>For <tt>__weak</tt> objects, the lvalue is updated to point to the
				703	new pointee, unless that object is currently undergoing deallocation,
				704	in which case it the lvalue is updated to a null pointer. This must
				705	execute atomically with respect to other assignments to the object, to
				706	reads from the object, and to the final release of the new pointed-to
				707	value.</li>
				708	<li>For <tt>__unsafe_unretained</tt> objects, the new pointee is
				709	stored into the lvalue using primitive semantics.</li>
				710	<li>For <tt>__autoreleasing</tt> objects, the new pointee is retained,
				711	autoreleased, and stored into the lvalue using primitive semantics.</li>
				712	</ul>
				713	</p>
				714
				715	<p><span class="term">Initialization</span> occurs when an object's
				716	lifetime begins, which depends on its storage duration.
				717	Initialization proceeds in two stages:
				718	<ol>
				719	<li>First, a null pointer is stored into the lvalue using primitive
				720	semantics. This step is skipped if the object
				721	is <tt>__unsafe_unretained</tt>.</li>
				722	<li>Second, if the object has an initializer, that expression is
				723	evaluated and then assigned into the object using the usual assignment
				724	semantics.</li>
				725	</ol>
				726	</p>
				727
				728	<p><span class="term">Destruction</span> occurs when an object's
				729	lifetime ends. In all cases it is semantically equivalent to
				730	assigning a null pointer to the object, with the proviso that of
				731	course the object cannot be legally read after the object's lifetime
				732	ends.</p>
				733
				734	<p><span class="term">Moving</span> occurs in specific situations
				735	where an lvalue is <q>moved from</q>, meaning that its current pointee
				736	will be used but the object may be left in a different (but still
				737	valid) state. This arises with <tt>__block</tt> variables and rvalue
				738	references in C++. For <tt>__strong</tt> lvalues, moving is equivalent
				739	to loading the lvalue with primitive semantics, writing a null pointer
				740	to it with primitive semantics, and then releasing the result of the
				741	load at the end of the current full-expression. For all other
				742	lvalues, moving is equivalent to reading the object.</p>
				743
				744	</div> <!-- ownership.semantics -->
				745
				746	<div id="ownership.restrictions">
				747	<h1>Restrictions</h1>
				748
				749	<div id="ownership.restrictions.autoreleasing">
				750	<h1>Storage duration of<tt> __autoreleasing</tt> objects</h1>
				751
				752	<p>A program is ill-formed if it declares an <tt>__autoreleasing</tt>
				753	object of non-automatic storage duration.</p>
				754
				755	<div class="rationale"><p>Rationale: autorelease pools are tied to the
				756	current thread and scope by their nature. While it is possible to
				757	have temporary objects whose instance variables are filled with
				758	autoreleased objects, there is no way that ARC can provide any sort of
				759	safety guarantee there.</p></div>
				760
				761	<p>It is undefined behavior if a non-null pointer is assigned to
				762	an <tt>__autoreleasing</tt> object while an autorelease pool is in
				763	scope and then that object is read after the autorelease pool's scope
				764	is left.</p>
				765
				766	</div>
				767
				768	<div id="ownership.restrictions.conversion.indirect">
				769	<h1>Conversion of pointers to ownership-qualified types</h1>
				770
				771	<p>A program is ill-formed if an expression of type <tt>T*</tt> is
				772	converted, explicitly or implicitly, to the type <tt>U*</tt>,
				773	where <tt>T</tt> and <tt>U</tt> have different ownership
				774	qualification, unless:
				775	<ul>
				776	<li><tt>T</tt> is qualified with <tt>__strong</tt>,
				777	<tt>__autoreleasing</tt>, or <tt>__unsafe_unretained</tt>, and
				778	<tt>U</tt> is qualified with both <tt>const</tt> and
				779	<tt>__unsafe_unretained</tt>; or</li>
				780	<li>either <tt>T</tt> or <tt>U</tt> is <tt>cv void</tt>, where
				781	<tt>cv</tt> is an optional sequence of non-ownership qualifiers; or</li>
				782	<li>the conversion is requested with a <tt>reinterpret_cast</tt> in
				783	Objective-C++; or</li>
				784	<li>the conversion is a
				785	well-formed <a href="#ownership.restrictions.pass_by_writeback">pass-by-writeback</a>.</li>
				786	</ul>
				787	</p>
				788
				789	<p>The analogous rule applies to <tt>T&</tt> and <tt>U&</tt> in
				790	Objective-C++.</p>
				791
				792	<div class="rationale"><p>Rationale: these rules provide a reasonable
				793	level of type-safety for indirect pointers, as long as the underlying
				794	memory is not deallocated. The conversion to <tt>const
				795	__unsafe_unretained</tt> is permitted because the semantics of reads
				796	are equivalent across all these ownership semantics, and that's a very
				797	useful and common pattern. The interconversion with <tt>void*</tt> is
				798	useful for allocating memory or otherwise escaping the type system,
				799	but use it carefully. <tt>reinterpret_cast</tt> is considered to be
				800	an obvious enough sign of taking responsibility for any
				801	problems.</p></div>
				802
				803	<p>It is undefined behavior to access an ownership-qualified object
				804	through an lvalue of a differently-qualified type, except that any
				805	non-<tt>__weak</tt> object may be read through
				806	an <tt>__unsafe_unretained</tt> lvalue.</p>
				807
				808	<p>It is undefined behavior if a managed operation is performed on
				809	a <tt>__strong</tt> or <tt>__weak</tt> object without a guarantee that
				810	it contains a primitive zero bit-pattern, or if the storage for such
				811	an object is freed or reused without the object being first assigned a
				812	null pointer.</p>
				813
				814	<div class="rationale"><p>Rationale: ARC cannot differentiate between
				815	an assignment operator which is intended to <q>initialize</q> dynamic
				816	memory and one which is intended to potentially replace a value.
				817	Therefore the object's pointer must be valid before letting ARC at it.
				818	Similarly, C and Objective-C do not provide any language hooks for
				819	destroying objects held in dynamic memory, so it is the programmer's
				820	responsibility to avoid leaks (<tt>__strong</tt> objects) and
				821	consistency errors (<tt>__weak</tt> objects).</p>
				822
				823	<p>These requirements are followed automatically in Objective-C++ when
				824	creating objects of retainable object owner type with <tt>new</tt>
				825	or <tt>new[]</tt> and destroying them with <tt>delete</tt>,
				826	<tt>delete[]</tt>, or a pseudo-destructor expression. Note that
				827	arrays of nontrivially-ownership-qualified type are not ABI compatible
				828	with non-ARC code because the element type is non-POD: such arrays
				829	that are <tt>new[]</tt>'d in ARC translation units cannot
				830	be <tt>delete[]</tt>'d in non-ARC translation units and
				831	vice-versa.</p></div>
				832
				833	</div>
				834
				835	<div id="ownership.restrictions.pass_by_writeback">
				836	<h1>Passing to an out parameter by writeback</h1>
				837
				838	<p>If the argument passed to a parameter of type
				839	<tt>T __autoreleasing </tt> has type <tt>U oq </tt>,
				840	where <tt>oq</tt> is an ownership qualifier, then the argument is a
				841	candidate for <span class="term">pass-by-writeback</span> if:</p>
				842
				843	<ul>
				844	<li><tt>oq</tt> is <tt>__strong</tt> or <tt>__weak</tt>, and
				845	<li>it would be legal to initialize a <tt>T __strong *</tt> with
				846	a <tt>U __strong *</tt>.</li>
				847	</ul>
				848
				849	<p>For purposes of overload resolution, an implicit conversion
				850	sequence requiring a pass-by-writeback is always worse than an
				851	implicit conversion sequence not requiring a pass-by-writeback.</p>
				852
				853	<p>The pass-by-writeback is ill-formed if the argument expression does
				854	not have a legal form:</p>
				855
				856	<ul>
				857	<li><tt>&var</tt>, where <tt>var</tt> is a scalar variable of
				858	automatic storage duration with retainable object pointer type</li>
				859	<li>a conditional expression where the second and third operands are
				860	both legal forms</li>
				861	<li>a cast whose operand is a legal form</li>
				862	<li>a null pointer constant</li>
				863	</ul>
				864
				865	<div class="rationale"><p>Rationale: the restriction in the form of
				866	the argument serves two purposes. First, it makes it impossible to
				867	pass the address of an array to the argument, which serves to protect
				868	against an otherwise serious risk of mis-inferring an <q>array</q>
				869	argument as an out-parameter. Second, it makes it much less likely
				870	that the user will see confusing aliasing problems due to the
				871	implementation, below, where their store to the writeback temporary is
				872	not immediately seen in the original argument variable.</p></div>
				873
				874	<p>A pass-by-writeback is evaluated as follows:
				875	<ol>
				876	<li>The argument is evaluated to yield a pointer <tt>p</tt> of
				877	type <tt>U oq *</tt>.</li>
				878	<li>If <tt>p</tt> is a null pointer, then a null pointer is passed as
				879	the argument, and no further work is required for the pass-by-writeback.</li>
				880	<li>Otherwise, a temporary of type <tt>T __autoreleasing</tt> is
				881	created and initialized to a null pointer.</li>
				882	<li>If the argument is not an Objective-C method parameter marked
				883	<tt>out</tt>, then <tt>*p</tt> is read, and the result is written
				884	into the temporary with primitive semantics.</li>
				885	<li>The address of the temporary is passed as the argument to the
				886	actual call.</li>
				887	<li>After the call completes, the temporary is loaded with primitive
				888	semantics, and that value is assigned into <tt>*p</tt>.</li>
				889	</ol></p>
				890
				891	<div class="rationale"><p>Rationale: this is all admittedly
				892	convoluted. In an ideal world, we would see that a local variable is
				893	being passed to an out-parameter and retroactively modify its type to
				894	be <tt>__autoreleasing</tt> rather than <tt>__strong</tt>. This would
				895	be remarkably difficult and not always well-founded under the C type
				896	system. However, it was judged unacceptably invasive to require
				897	programmers to write <tt>__autoreleasing</tt> on all the variables
				898	they intend to use for out-parameters. This was the least bad
				899	solution.</p></div>
				900
				901	</div>
				902
				903	<div id="ownership.restrictions.records">
				904	<h1>Ownership-qualified fields of structs and unions</h1>
				905
				906	<p>A program is ill-formed if it declares a member of a C struct or
				907	union to have a nontrivially ownership-qualified type.</p>
				908
				909	<div class="rationale"><p>Rationale: the resulting type would be
				910	non-POD in the C++ sense, but C does not give us very good language
				911	tools for managing the lifetime of aggregates, so it is more
				912	convenient to simply forbid them. It is still possible to manage this
				913	with a <tt>void*</tt> or an <tt>__unsafe_unretained</tt>
				914	object.</p></div>
				915
				916	<p>This restriction does not apply in Objective-C++. However,
				917	nontrivally ownership-qualified types are considered non-POD: in C++0x
				918	terms, they are not trivially default constructible, copy
				919	constructible, move constructible, copy assignable, move assignable,
				920	or destructible. It is a violation of C++ One Definition Rule to use
				921	a class outside of ARC that, under ARC, would have an
				922	ownership-qualified member.</p>
				923
				924	<div class="rationale"><p>Rationale: unlike in C, we can express all
				925	the necessary ARC semantics for ownership-qualified subobjects as
				926	suboperations of the (default) special member functions for the class.
				927	These functions then become non-trivial. This has the non-obvious
				928	repercussion that the class will have a non-trivial copy constructor
				929	and non-trivial destructor; if it wouldn't outside of ARC, this means
				930	that objects of the type will be passed and returned in an
				931	ABI-incompatible manner.</p></div>
				932
				933	</div>
				934
				935	</div>
				936
				937	<div id="ownership.inference">
				938	<h1>Ownership inference</h1>
				939
				940	<div id="ownership.inference.variables">
				941	<h1>Objects</h1>
				942
				943	<p>If an object is declared with retainable object owner type, but
				944	without an explicit ownership qualifier, its type is implicitly
				945	adjusted to have <tt>__strong</tt> qualification.</p>
				946
				947	<p>As a special case, if the object's base type is <tt>Class</tt>
				948	(possibly protocol-qualified), the type is adjusted to
				949	have <tt>__unsafe_unretained</tt> qualification instead.</p>
				950
				951	</div>
				952
				953	<div id="ownership.inference.indirect_parameters">
				954	<h1>Indirect parameters</h1>
				955
				956	<p>If a function or method parameter has type <tt>T*</tt>, where
				957	<tt>T</tt> is an ownership-unqualified retainable object pointer type,
				958	then:</p>
				959
				960	<ul>
				961	<li>if <tt>T</tt> is <tt>const</tt>-qualified or <tt>Class</tt>, then
				962	it is implicitly qualified with <tt>__unsafe_unretained</tt>;</li>
				963	<li>otherwise, it is implicitly qualified
				964	with <tt>__autoreleasing</tt>.</li>
				965	</ul>
				966	</p>
				967
				968	<div class="rationale"><p>Rationale: <tt>__autoreleasing</tt> exists
				969	mostly for this case, the Cocoa convention for out-parameters. Since
				970	a pointer to <tt>const</tt> is obviously not an out-parameter, we
				971	instead use a type more useful for passing arrays. If the user
				972	instead intends to pass in a <em>mutable</em> array, inferring
				973	<tt>__autoreleasing</tt> is the wrong thing to do; this directs some
				974	of the caution in the following rules about writeback.</p></div>
				975
				976	<p>Such a type written anywhere else would be ill-formed by the
				977	general rule requiring ownership qualifiers.</p>
				978
				979	<p>This rule does not apply in Objective-C++ if a parameter's type is
				980	dependent in a template pattern and is only <em>instantiated</em> to
				981	a type which would be a pointer to an unqualified retainable object
				982	pointer type. Such code is still ill-formed.</p>
				983
				984	<div class="rationale"><p>Rationale: the convention is very unlikely
				985	to be intentional in template code.</p></div>
				986
				987	</div> <!-- ownership.inference.indirect_parameters -->
Douglas Gregor	e559ca1	2011-06-17 22:11:49 +0000	[diff] [blame]	988
				989	<div id="ownership.inference.template_arguments">
				990	<h1>Template arguments</h1>
				991
				992	<p>If a template argument for a template type parameter is an
				993	retainable object owner type that does not have an explicit ownership
				994	qualifier, it is adjusted to have <tt>__strong</tt>
Douglas Gregor	54fb28a	2011-06-17 22:19:27 +0000	[diff] [blame]	995	qualification. This adjustment occurs regardless of whether the
Douglas Gregor	e559ca1	2011-06-17 22:11:49 +0000	[diff] [blame]	996	template argument was deduced or explicitly specified. </p>
				997
				998	<div class="rationale"><p>Rationale: <tt>__strong</tt> is a useful default for containers (e.g., <tt>std::vector<id></tt>), 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.</p></div>
				999
				1000	</div> <!-- ownership.inference.template_arguments -->
John McCall	8246702	2011-06-15 21:21:53 +0000	[diff] [blame]	1001	</div> <!-- ownership.inference -->
				1002	</div> <!-- ownership -->
				1003
Douglas Gregor	e559ca1	2011-06-17 22:11:49 +0000	[diff] [blame]	1004
John McCall	8246702	2011-06-15 21:21:53 +0000	[diff] [blame]	1005	<div id="family">
				1006	<h1>Method families</h1>
				1007
				1008	<p>An Objective-C method may fall into a <span class="term">method
				1009	family</span>, which is a conventional set of behaviors ascribed to it
				1010	by the Cocoa conventions.</p>
				1011
				1012	<p>A method is in a certain method family if:
				1013	<ul>
				1014	<li>it has a <tt>objc_method_family</tt> attribute placing it in that
				1015	family; or if not that,</li>
				1016	<li>it does not have an <tt>objc_method_family</tt> attribute placing
				1017	it in a different or no family, and</li>
				1018	<li>its selector falls into the corresponding selector family, and</li>
				1019	<li>its signature obeys the added restrictions of the method family.</li>
				1020	</ul></p>
				1021
				1022	<p>A selector is in a certain selector family if, ignoring any leading
				1023	underscores, the first component of the selector either consists
				1024	entirely of the name of the method family or it begins with that name
				1025	followed by a character other than a lowercase letter. For
				1026	example, <tt>_perform:with:</tt> and <tt>performWith:</tt> would fall
				1027	into the <tt>perform</tt> family (if we recognized one),
				1028	but <tt>performing:with</tt> would not.</p>
				1029
				1030	<p>The families and their added restrictions are:</p>
				1031
				1032	<ul>
				1033	<li><tt>alloc</tt> methods must return a retainable object pointer type.</li>
				1034	<li><tt>copy</tt> methods must return a retainable object pointer type.</li>
				1035	<li><tt>mutableCopy</tt> methods must return a retainable object pointer type.</li>
				1036	<li><tt>new</tt> methods must return a retainable object pointer type.</li>
				1037	<li><tt>init</tt> methods must be instance methods and must return an
				1038	Objective-C pointer type. Additionally, a program is ill-formed if it
				1039	declares or contains a call to an <tt>init</tt> method whose return
				1040	type is neither <tt>id</tt> nor a pointer to a super-class or
				1041	sub-class of either the declaring class, if the method was declared on
				1042	a class, or the static receiver type of the call, if it was declared
				1043	on a protocol.</p>
				1044
				1045	<div class="rationale"><p>Rationale: there are a fair number of existing
				1046	methods with <tt>init</tt>-like selectors which nonetheless don't
				1047	follow the <tt>init</tt> conventions. Typically these are either
				1048	accidental naming collisions or helper methods called during
				1049	initialization. Because of the peculiar retain/release behavior
				1050	of <tt>init</tt> methods, it's very important not to treat these
				1051	methods as <tt>init</tt> methods if they aren't meant to be. It was
				1052	felt that implicitly defining these methods out of the family based on
				1053	the exact relationship between the return type and the declaring class
				1054	would much too subtle and fragile. Therefore we identify a small
				1055	number of legitimate-seeming return types and call everything else an
				1056	error. This serves the secondary purpose of encouraging programmers
				1057	not to accidentally give methods names in the <tt>init</tt> family.</p></div>
				1058	</li>
				1059	</ul>
				1060
				1061	<p>A program is ill-formed if a method's declarations,
				1062	implementations, and overrides do not all have the same method
				1063	family.</p>
				1064
				1065	<div id="family.attribute">
				1066	<h1>Explicit method family control</h1>
				1067
				1068	<p>A method may be annotated with the <tt>objc_method_family</tt>
				1069	attribute to precisely control which method family it belongs to. If
				1070	a method in an <tt>@implementation</tt> does not have this attribute,
				1071	but there is a method declared in the corresponding <tt>@interface</tt>
				1072	that does, then the attribute is copied to the declaration in the
				1073	<tt>@implementation</tt>. The attribute is available outside of ARC,
				1074	and may be tested for with the preprocessor query
				1075	<tt>__has_attribute(objc_method_family)</tt>.</p>
				1076
				1077	<p>The attribute is spelled
				1078	<tt>__attribute__((objc_method_family(<i>family</i>)))</tt>.
				1079	If <i>family</i> is <tt>none</tt>, the method has no family, even if
				1080	it would otherwise be considered to have one based on its selector and
				1081	type. Otherwise, <i>family</i> must be one
				1082	of <tt>alloc</tt>, <tt>copy</tt>, <tt>init</tt>,
				1083	<tt>mutableCopy</tt>, or <tt>new</tt>, in which case the method is
				1084	considered to belong to the corresponding family regardless of its
				1085	selector. It is an error if a method that is explicitly added to a
				1086	family in this way does not meet the requirements of the family other
				1087	than the selector naming convention.</p>
				1088
				1089	<div class="rationale"><p>Rationale: the rules codified in this document
				1090	describe the standard conventions of Objective-C. However, as these
				1091	conventions have not heretofore been enforced by an unforgiving
				1092	mechanical system, they are only imperfectly kept, especially as they
				1093	haven't always even been precisely defined. While it is possible to
				1094	define low-level ownership semantics with attributes like
				1095	<tt>ns_returns_retained</tt>, this attribute allows the user to
				1096	communicate semantic intent, which of use both to ARC (which, e.g.,
				1097	treats calls to <tt>init</tt> specially) and the static analyzer.</p></div>
				1098	</div>
				1099
				1100	<div id="family.semantics">
				1101	<h1>Semantics of method families</h1>
				1102
				1103	<p>A method's membership in a method family may imply non-standard
				1104	semantics for its parameters and return type.</p>
				1105
				1106	<p>Methods in the <tt>alloc</tt>, <tt>copy</tt>, <tt>mutableCopy</tt>,
				1107	and <tt>new</tt> families — that is, methods in all the
				1108	currently-defined families except <tt>init</tt> — transfer
				1109	ownership of a +1 retain count on their return value to the calling
				1110	function, as if they were implicitly annotated with
				1111	the <tt>ns_returns_retained</tt> attribute. However, this is not true
				1112	if the method has either of the <tt>ns_returns_autoreleased</tt> or
				1113	<tt>ns_returns_not_retained</tt> attributes.</p>
				1114
				1115	<div id="family.semantics.init">
				1116	<h1>Semantics of <tt>init</tt></h1>
				1117
				1118	<p>Methods in the <tt>init</tt> family must be transferred ownership
				1119	of a +1 retain count on their <tt>self</tt> parameter, exactly as if
				1120	the method had the <tt>ns_consumes_self</tt> attribute, and must
				1121	transfer ownership of a +1 retain count on their return value, exactly
				1122	as if they method had the <tt>ns_returns_retained</tt> attribute.
				1123	Neither of these may be altered through attributes.</p>
				1124
				1125	<p>A call to an <tt>init</tt> method with a receiver that is either
				1126	<tt>self</tt> (possibly parenthesized or casted) or <tt>super</tt> is
				1127	called a <span class="term">delegate init call</span>. It is an error
				1128	for a delegate init call to be made except from an <tt>init</tt>
				1129	method, and excluding blocks within such methods.</p>
				1130
				1131	<p>The variable <tt>self</tt> is mutable in an <tt>init</tt> method
				1132	and is implicitly qualified as <tt>__strong</tt>. However, a program
				1133	is ill-formed, no diagnostic required, if it alters <tt>self</tt>
				1134	except to assign it the immediate result of a delegate init call. It
				1135	is an error to use the previous value of <tt>self</tt> after the
				1136	completion of a delegate init call.</p>
				1137
				1138	<p>A program is ill-formed, no diagnostic required, if it causes two
				1139	or more calls to <tt>init</tt> methods on the same object, except that
				1140	each <tt>init</tt> method invocation may perform at most one
				1141	delegate init call.</p>
				1142
John McCall	f3d08a6	2011-06-18 07:31:30 +0000	[diff] [blame^]	1143	</div> <!-- family.semantics.init -->
John McCall	8246702	2011-06-15 21:21:53 +0000	[diff] [blame]	1144
				1145	<div id="family.semantics.result_type">
				1146	<h1>Related result types</h1>
				1147
				1148	<p>Certain methods are candidates to have <span class="term">related
				1149	result types</span>:</p>
				1150	<ul>
				1151	<li>class methods in the <tt>alloc</tt> and <tt>new</tt> method families</li>
				1152	<li>instance methods in the <tt>init</tt> family</li>
				1153	<li>the instance method <tt>self</tt></li>
				1154	<li>outside of ARC, the instance methods <tt>retain</tt> and <tt>autorelease</tt></li>
				1155	</ul>
				1156
				1157	<p>If the formal result type of such a method is <tt>id</tt> or
				1158	protocol-qualified <tt>id</tt>, or a type equal to the declaring class
				1159	or a superclass, then it is said to have a related result type. In
				1160	this case, when invoked in an explicit message send, it is assumed to
				1161	return a type related to the type of the receiver:</p>
				1162
				1163	<ul>
				1164	<li>if it is a class method, and the receiver is a class
				1165	name <tt>T</tt>, the message send expression has type <tt>T*</tt>;
				1166	otherwise</li>
				1167	<li>if it is an instance method, and the receiver has type <tt>T</tt>,
				1168	the message send expression has type <tt>T</tt>; otherwise</li>
				1169	<li>the message send expression has the normal result type of the
				1170	method.</li>
				1171	</ul>
				1172
				1173	<p>This is a new rule of the Objective-C language and applies outside
				1174	of ARC.</p>
				1175
				1176	<div class="rationale"><p>Rationale: ARC's automatic code emission is
				1177	more prone than most code to signature errors, i.e. errors where a
				1178	call was emitted against one method signature, but the implementing
				1179	method has an incompatible signature. Having more precise type
				1180	information helps drastically lower this risks, as well as catching
				1181	a number of latent bugs.</p></div>
				1182
				1183	</div> <!-- family.semantics.result_type -->
				1184	</div> <!-- family.semantics -->
				1185	</div> <!-- family -->
				1186
				1187	<div id="optimization">
				1188	<h1>Optimization</h1>
				1189
				1190	<p>ARC applies aggressive rules for the optimization of local
				1191	behavior. These rules are based around a core assumption of
				1192	<span class="term">local balancing</span>: that other code will
				1193	perform retains and releases as necessary (and only as necessary) for
				1194	its own safety, and so the optimizer does not need to consider global
				1195	properties of the retain and release sequence. For example, if a
				1196	retain and release immediately bracket a call, the optimizer can
				1197	delete the retain and release on the assumption that the called
				1198	function will not do a constant number of unmotivated releases
				1199	followed by a constant number of <q>balancing</q> retains, such that
				1200	the local retain/release pair is the only thing preventing the called
				1201	function from ending up with a dangling reference.</p>
				1202
				1203	<p>The optimizer assumes that when a new value enters local control,
				1204	e.g. from a load of a non-local object or as the result of a function
				1205	call, it is instaneously valid. Subsequently, a retain and release of
				1206	a value are necessary on a computation path only if there is a use of
				1207	that value before the release and after any operation which might
				1208	cause a release of the value (including indirectly or non-locally),
				1209	and only if the value is not demonstrably already retained.</p>
				1210
				1211	<p>The complete optimization rules are quite complicated, but it would
				1212	still be useful to document them here.</p>
				1213
				1214	</div>
				1215
				1216	<div id="misc">
				1217	<h1>Miscellaneous</h1>
				1218
John McCall	f3d08a6	2011-06-18 07:31:30 +0000	[diff] [blame^]	1219	<div id="misc.special_methods">
				1220	<h1>Special methods</h1>
				1221
				1222	<div id="misc.special_methods.retain">
				1223	<h1>Memory management methods</h1>
				1224
				1225	<p>A program is ill-formed if it contains a method definition, message
				1226	send, or <tt>@selector</tt> expression for any of the following
				1227	selectors:
				1228	<ul>
				1229	<li><tt>autorelease</tt></li>
				1230	<li><tt>release</tt></li>
				1231	<li><tt>retain</tt></li>
				1232	<li><tt>retainCount</tt></li>
				1233	</ul>
				1234	</p>
				1235
				1236	<div class="rationale"><p>Rationale: <tt>retainCount</tt> is banned
				1237	because ARC robs it of consistent semantics. The others were banned
				1238	after weighing three options for how to deal with message sends:</p>
				1239
				1240	<p><b>Honoring</b> them would work out very poorly if a programmer
				1241	naively or accidentally tried to incorporate code written for manual
				1242	retain/release code into an ARC program. At best, such code would do
				1243	twice as much work as necessary; quite frequently, however, ARC and
				1244	the explicit code would both try to balance the same retain, leading
				1245	to crashes. The cost is losing the ability to perform <q>unrooted</q>
				1246	retains, i.e. retains not logically corresponding to a strong
				1247	reference in the object graph.</p>
				1248
				1249	<p><b>Ignoring</b> them would badly violate user expectations about their
				1250	code. While it <em>would</em> make it easier to develop code simultaneously
				1251	for ARC and non-ARC, there is very little reason to do so except for
				1252	certain library developers. ARC and non-ARC translation units share
				1253	an execution model and can seamlessly interoperate. Within a
				1254	translation unit, a developer who faithfully maintains their code in
				1255	non-ARC mode is suffering all the restrictions of ARC for zero
				1256	benefit, while a developer who isn't testing the non-ARC mode is
				1257	likely to be unpleasantly surprised if they try to go back to it.</p>
				1258
				1259	<p><b>Banning</b> them has the disadvantage of making it very awkward
				1260	to migrate existing code to ARC. The best answer to that, given a
				1261	number of other changes and restrictions in ARC, is to provide a
				1262	specialized tool to assist users in that migration.</p>
				1263
				1264	<p>Implementing these methods was banned because they are too integral
				1265	to the semantics of ARC; many tricks which worked tolerably under
				1266	manual reference counting will misbehave if ARC performs an ephemeral
				1267	extra retain or two. If absolutely required, it is still possible to
				1268	implement them in non-ARC code, for example in a category; the
				1269	implementations must obey the <a href="#objects.retains">semantics</a>
				1270	laid out elsewhere in this document.</p>
				1271
				1272	</div>
				1273	</div> <!-- misc.special_methods.retain -->
				1274
				1275	<div id="misc.special_methods.dealloc">
				1276	<h1><tt>dealloc</tt></h1>
				1277
				1278	<p>A program is ill-formed if it contains a message send
				1279	or <tt>@selector</tt> expression for the selector <tt>dealloc</tt>.</p>
				1280
				1281	<div class="rationale"><p>Rationale: there are no legitimate reasons
				1282	to call <tt>dealloc</tt> directly.</p></div>
				1283
				1284	<p>A class may provide a method definition for an instance method
				1285	named <tt>dealloc</tt>. This method will be called after the final
				1286	<tt>release</tt> of the object but before it is deallocated or any of
				1287	its instance variables are destroyed. The superclass's implementation
				1288	of <tt>dealloc</tt> will be called automatically when the method
				1289	returns.</p>
				1290
				1291	<div class="rationale"><p>Rationale: even though ARC destroys instance
				1292	variables automatically, there are still legitimate reasons to write
				1293	a <tt>dealloc</tt> method, such as freeing non-retainable resources.
				1294	Failing to call <tt>[super dealloc]</tt> in such a method is nearly
				1295	always a bug. Sometimes, the object is simply trying to prevent
				1296	itself from being destroyed, but <tt>dealloc</tt> is really far too
				1297	late for the object to be raising such objections. Somewhat more
				1298	legitimately, an object may have been pool-allocated and should not be
				1299	deallocated with <tt>free</tt>; for now, this can only be supported
				1300	with a <tt>dealloc</tt> implementation outside of ARC. Such an
				1301	implementation must be very careful to do all the other work
				1302	that <tt>NSObject</tt>'s <tt>dealloc</tt> would, which is outside the
				1303	scope of this document to describe.</p></div>
				1304
				1305	</div>
				1306
				1307	</div> <!-- misc.special_methods -->
				1308
John McCall	8246702	2011-06-15 21:21:53 +0000	[diff] [blame]	1309	<div id="autoreleasepool">
				1310	<h1><tt>@autoreleasepool</tt></h1>
				1311
				1312	<p>To simplify the use of autorelease pools, and to bring them under
				1313	the control of the compiler, a new kind of statement is available in
				1314	Objective-C. It is written <tt>@autoreleasepool</tt> followed by
				1315	a <i>compound-statement</i>, i.e. by a new scope delimited by curly
				1316	braces. Upon entry to this block, the current state of the
				1317	autorelease pool is captured. When the block is exited normally,
				1318	whether by fallthrough or directed control flow (such
				1319	as <tt>return</tt> or <tt>break</tt>), the autorelease pool is
				1320	restored to the saved state, releasing all the objects in it. When
				1321	the block is exited with an exception, the pool is not drained.</p>
				1322
John McCall	f3d08a6	2011-06-18 07:31:30 +0000	[diff] [blame^]	1323	<p><tt>@autoreleasepool</tt> may be used in non-ARC translation units,
				1324	with equivalent semantics.</p>
				1325
John McCall	8246702	2011-06-15 21:21:53 +0000	[diff] [blame]	1326	<p>A program is ill-formed if it refers to the
				1327	<tt>NSAutoreleasePool</tt> class.</p>
				1328
				1329	<div class="rationale"><p>Rationale: autorelease pools are clearly
				1330	important for the compiler to reason about, but it is far too much to
				1331	expect the compiler to accurately reason about control dependencies
				1332	between two calls. It is also very easy to accidentally forget to
				1333	drain an autorelease pool when using the manual API, and this can
				1334	significantly inflate the process's high-water-mark. The introduction
				1335	of a new scope is unfortunate but basically required for sane
				1336	interaction with the rest of the language. Not draining the pool
				1337	during an unwind is apparently required by the Objective-C exceptions
				1338	implementation.</p></div>
				1339
				1340	</div> <!-- autoreleasepool -->
				1341
				1342	<div id="misc.self">
				1343	<h1><tt>self</tt></h1>
				1344
				1345	<p>The <tt>self</tt> parameter variable of an Objective-C method is
				1346	never actually retained by the implementation. It is undefined
				1347	behavior, or at least dangerous, to cause an object to be deallocated
				1348	during a message send to that object. To make this
				1349	safe, <tt>self</tt> is implicitly <tt>const</tt> unless the method is
				1350	in the <a href="#family.semantics.init"><tt>init</tt> family</a>.</p>
				1351
				1352	<div class="rationale"><p>Rationale: the cost of
				1353	retaining <tt>self</tt> in all methods was found to be prohibitive, as
				1354	it tends to be live across calls, preventing the optimizer from
				1355	proving that the retain and release are unnecessary — for good
				1356	reason, as it's quite possible in theory to cause an object to be
				1357	deallocated during its execution without this retain and release.
				1358	Since it's extremely uncommon to actually do so, even unintentionally,
				1359	and since there's no natural way for the programmer to remove this
				1360	retain/release pair otherwise (as there is for other parameters by,
				1361	say, making the variable <tt>__unsafe_unretained</tt>), we chose to
				1362	make this optimizing assumption and shift some amount of risk to the
				1363	user.</p></div>
				1364
				1365	</div> <!-- misc.self -->
				1366
				1367	<div id="misc.enumeration">
				1368	<h1>Fast enumeration iteration variables</h1>
				1369
				1370	<p>If a variable is declared in the condition of an Objective-C fast
				1371	enumeration loop, and the variable has no explicit ownership
				1372	qualifier, then it is qualified with <tt>const __strong</tt> and
				1373	objects encountered during the enumeration are not actually
				1374	retained.</p>
				1375
				1376	<div class="rationale"><p>Rationale: this is an optimization made
				1377	possible because fast enumeration loops promise to keep the objects
				1378	retained during enumeration, and the collection itself cannot be
				1379	synchronously modified. It can be overridden by explicitly qualifying
				1380	the variable with <tt>__strong</tt>, which will make the variable
				1381	mutable again and cause the loop to retain the objects it
				1382	encounters.</div>
				1383
				1384	</div>
				1385
				1386	<div id="misc.blocks">
				1387	<h1>Blocks</h1>
				1388
				1389	<p>The implicit <tt>const</tt> capture variables created when
				1390	evaluating a block literal expression have the same ownership
				1391	semantics as the local variables they capture. The capture is
				1392	performed by reading from the captured variable and initializing the
				1393	capture variable with that value; the capture variable is destroyed
				1394	when the block literal is, i.e. at the end of the enclosing scope.</p>
				1395
				1396	<p>The <a href="#ownership.inference">inference</a> rules apply
				1397	equally to <tt>__block</tt> variables, which is a shift in semantics
				1398	from non-ARC, where <tt>__block</tt> variables did not implicitly
				1399	retain during capture.</p>
				1400
				1401	<p><tt>__block</tt> variables of retainable object owner type are
				1402	moved off the stack by initializing the heap copy with the result of
				1403	moving from the stack copy.</tt></p>
				1404
				1405	<p>With the exception of retains done as part of initializing
				1406	a <tt>__strong</tt> parameter variable or reading a <tt>__weak</tt>
				1407	variable, whenever these semantics call for retaining a value of
				1408	block-pointer type, it has the effect of a <tt>Block_copy</tt>. The
				1409	optimizer may remove such copies when it sees that the result is
				1410	used only as an argument to a call.</p>
				1411
				1412	</div> <!-- misc.blocks -->
				1413
				1414	<div id="misc.exceptions">
				1415	<h1>Exceptions</h1>
				1416
				1417	<p>By default in Objective C, ARC is not exception-safe for normal
				1418	releases:
				1419	<ul>
				1420	<li>It does not end the lifetime of <tt>__strong</tt> variables when
				1421	their scopes are abnormally terminated by an exception.</li>
				1422	<li>It does not perform releases which would occur at the end of
				1423	a full-expression if that full-expression throws an exception.</li>
				1424	</ul>
				1425
				1426	<p>A program may be compiled with the option
				1427	<tt>-fobjc-arc-exceptions</tt> in order to enable these, or with the
				1428	option <tt>-fno-objc-arc-exceptions</tt> to explicitly disable them,
				1429	with the last such argument <q>winning</q>.</p>
				1430
				1431	<div class="rationale"><p>Rationale: the standard Cocoa convention is
				1432	that exceptions signal programmer error and are not intended to be
				1433	recovered from. Making code exceptions-safe by default would impose
				1434	severe runtime and code size penalties on code that typically does not
				1435	actually care about exceptions safety. Therefore, ARC-generated code
				1436	leaks by default on exceptions, which is just fine if the process is
				1437	going to be immediately terminated anyway. Programs which do care
				1438	about recovering from exceptions should enable the option.</p></div>
				1439
				1440	<p>In Objective-C++, <tt>-fobjc-arc-exceptions</tt> is enabled by
				1441	default.</p>
				1442
				1443	<div class="rationale"><p>Rationale: C++ already introduces pervasive
				1444	exceptions-cleanup code of the sort that ARC introduces. C++
				1445	programmers who have not already disabled exceptions are much more
				1446	likely to actual require exception-safety.</p></div>
				1447
				1448	<p>ARC does end the lifetimes of <tt>__weak</tt> objects when an
				1449	exception terminates their scope unless exceptions are disabled in the
				1450	compiler.</p>
				1451
				1452	<div class="rationale"><p>Rationale: the consequence of a
				1453	local <tt>__weak</tt> object not being destroyed is very likely to be
				1454	corruption of the Objective-C runtime, so we want to be safer here.
				1455	Of course, potentially massive leaks are about as likely to take down
				1456	the process as this corruption is if the program does try to recover
				1457	from exceptions.</p></div>
				1458
				1459	</div> <!-- misc.exceptions -->
				1460
				1461	</div> <!-- misc -->
				1462	</div> <!-- root -->
				1463	</body>
				1464	</html>