Skip to content

Latest commit

 

History

History
1061 lines (855 loc) · 33.6 KB

0030-property-behavior-decls.md

File metadata and controls

1061 lines (855 loc) · 33.6 KB

Property Behaviors

Introduction

There are property implementation patterns that come up repeatedly. Rather than hardcode a fixed set of patterns into the compiler, we should provide a general "property behavior" mechanism to allow these patterns to be defined as libraries.

Swift Evolution Discussion
Review

Motivation

We've tried to accommodate several important patterns for properties with targeted language support, but this support has been narrow in scope and utility. For instance, Swift 1 and 2 provide lazy properties as a primitive language feature, since lazy initialization is common and is often necessary to avoid having properties be exposed as Optional. Without this language support, it takes a lot of boilerplate to get the same effect:

classFoo{ // lazy var foo = 1738 privatevar_foo:Int?varfoo:Int{get{iflet value = _foo {return value }letinitialValue=1738 _foo = initialValue return initialValue }set{ _foo = newValue }}}

Building lazy into the language has several disadvantages. It makes the language and compiler more complex and less orthogonal. It's also inflexible; there are many variations on lazy initialization that make sense, but we wouldn't want to hardcode language support for all of them. For instance, some applications may want the lazy initialization to be synchronized, but lazy only provides single-threaded initialization. The standard implementation of lazy is also problematic for value types. A lazy getter must be mutating, which means it can't be accessed from an immutable value. Inline storage is also suboptimal for many memoization tasks, since the cache cannot be reused across copies of the value. A value-oriented memoized property implementation might look very different, using a class instance to store the cached value out-of-line in order to avoid mutation of the value itself.

There are important property patterns outside of lazy initialization. It often makes sense to have "delayed", once-assignable-then-immutable properties to support multi-phase initialization:

classFoo{letimmediatelyInitialized="foo"var_initializedLater:String? // We want initializedLater to present like a non-optional 'let' to user code; // it can only be assigned once, and can't be accessed before being assigned. varinitializedLater:String{get{return _initializedLater! }set{assert(_initializedLater ==nil) _initializedLater = newValue }}}

Implicitly-unwrapped optionals allow this in a pinch, but give up a lot of safety compared to a non-optional 'let'. Using IUO for multi-phase initialization gives up both immutability and nil-safety.

We also have other application-specific property features like didSet/willSet that add language complexity for limited functionality. Beyond what we've baked into the language already, there's a seemingly endless set of common property behaviors, including synchronized access, copying, and various kinds of proxying, all begging for language attention to eliminate their boilerplate.

Proposed solution

I suggest we allow for property behaviors to be implemented within the language. A var declaration can specify its behaviors in square brackets after the keyword:

var[lazy] foo =1738

which implements the property foo in a way described by the property behavior declaration for lazy:

varbehavior lazy<Value>: Value {var value: Value?=nil initialValue mutating get {iflet value = value {return value }letinitial= initialValue value = initial return initial }set{ value = newValue }}

Property behaviors can control the storage, initialization, and access of affected properties, obviating the need for special language support for lazy, observers, and other special-case property features.

Examples

Before describing the detailed design, I'll run through some examples of potential applications for behaviors.

Lazy

The current lazy property feature can be reimplemented as a property behavior.

// Property behaviors are declared using the `var behavior` keyword cluster. publicvarbehavior lazy<Value>: Value { // Behaviors can declare storage that backs the property. privatevarvalue:Value? // Behaviors can bind the property's initializer expression with an // `initialValue` property declaration. initialValue // Behaviors can declare initialization logic for the storage. // (Stored properties can also be initialized in-line.) init(){ value =nil} // Inline initializers are also supported, so `var value: Value? = nil` // would work equivalently. // Behaviors can declare accessors that implement the property. mutating get {iflet value = value {return value }letinitial= initialValue value = initial return initial }set{ value = newValue }}

Properties declared with the lazy behavior are backed by the Optional-typed storage and accessors from the behavior:

var[lazy] x =1738 // Allocates an Int? behind the scenes, inited to nil print(x) // Invokes the `lazy` getter, initializing the property x =679 // Invokes the `lazy` setter

Delayed Initialization

A property behavior can model "delayed" initialization behavior, where the DI rules for properties are enforced dynamically rather than at compile time. This can avoid the need for implicitly-unwrapped optionals in multi-phase initialization. We can implement both a mutable variant, which allows for reassignment like a var:

publicvarbehavior delayedMutable<Value>: Value {privatevarvalue:Value?=nilget{guardlet value = value else{fatalError("property accessed before being initialized")}return value }set{ value = newValue }}

and an immutable variant, which only allows a single initialization like a let:

publicvarbehavior delayedImmutable<Value>: Value {privatevarvalue:Value?=nilget{guardlet value = value else{fatalError("property accessed before being initialized")}return value } // Perform an initialization, trapping if the // value is already initialized. set{iflet _ = value {fatalError("property initialized twice")} value = initialValue }}

This enables multi-phase initialization, like this:

classFoo{var[delayedImmutable] x: Int init(){ // We don't know "x" yet, and we don't have to set it }func initializeX(x:Int){self.x = x // Will crash if 'self.x' is already initialized }func getX()->Int{return x // Will crash if 'self.x' wasn't initialized }}

Property Observers

A property behavior can also approximate the built-in behavior of didSet/willSet observers, by declaring support for custom accessors:

publicvarbehavior observed<Value>: Value { initialValue var value = initialValue // A behavior can declare accessor requirements, the implementations of // which must be provided by property declarations using the behavior. // The behavior may provide a default implementation of the accessors, in // order to make them optional. // The willSet accessor, invoked before the property is updated. The // default does nothing. mutating accessor willSet(newValue: Value){} // The didSet accessor, invoked before the property is updated. The // default does nothing. mutating accessor didSet(oldValue: Value){}get{return value }set{willSet(newValue)letoldValue= value value = newValue didSet(oldValue)}}

A common complaint with didSet/willSet is that the observers fire on every write, not only ones that cause a real change. A behavior that supports a didChange accessor, which only gets invoked if the property value really changed to a value not equal to the old value, can be implemented as a new behavior:

publicvarbehavior changeObserved<Value:Equatable>:Value{ initialValue varvalue= initialValue mutating accessor didChange(oldValue: Value){}get{return value }set{letoldValue= value value = newValue if oldValue != newValue {didChange(oldValue)}}}

For example:

var[changeObserved] x: Int =1{didChange{print("\(oldValue) => \(x)")}} x =1 // Prints nothing x =2 // Prints 1 => 2

(Note that, like didSet/willSet today, neither behavior implementation will observe changes through class references that mutate a referenced class instance without changing the reference itself. Also, as currently proposed, behaviors would force the property to be initialized in-line, which is not acceptable for instance properties. That's a limitation that can be lifted by future extensions.)

Synchronized Property Access

Objective-C supports atomic properties, which take a lock on get and set to synchronize accesses to a property. This is occasionally useful, and it can be brought to Swift as a behavior. The real implementation of atomic properties in ObjC uses a global bank of locks, but for illustrative purposes (and to demonstrate referring to self) I'll use a per-object lock instead:

// A class that owns a mutex that can be used to synchronize access to its // properties. publicprotocolSynchronizable:class{func withLock<R>(@noescape body:()->R)->R} // Behaviors can refer to a property's containing type using // the implicit `Self` generic parameter. Constraints can be // applied using a 'where' clause, like in an extension. publicvarbehavior synchronized<Value where Self: Synchronizable>:Value{ initialValue varvalue:Value= initialValue get{returnself.withLock{return value }}set{self.withLock{ value = newValue }}}

NSCopying

Many Cocoa classes implement value-like objects that require explicit copying. Swift currently provides an @NSCopying attribute for properties to give them behavior like Objective-C's @property(copy), invoking the copy method on new objects when the property is set. We can turn this into a behavior:

publicvarbehavior copying<Value:NSCopying>:Value{ initialValue // Copy the value on initialization. varvalue:Value= initialValue.copy()get{return value }set{ // Copy the value on reassignment. value = newValue.copy()}}

This is a small sampling of the possibilities of behaviors. Let's look at the proposed design in detail:

Detailed design

Property behavior declarations

A property behavior declaration is introduced by the var behavior contextual keyword cluster. The declaration is designed to resemble the syntax of a property using the behavior:

property-behavior-decl ::= attribute* decl-modifier* 'var' 'behavior' identifier // behavior name generic-signature? ':' type '{' property-behavior-member-decl* '}'

Inside the behavior declaration, standard initializer, property, method, and nested type declarations are allowed, as are core accessor declarations —get and set. Accessor requirement declarations and initial value requirement declarations are also recognized contextually within the declaration:

property-behavior-member-decl ::= decl property-behavior-member-decl ::= accessor-decl // get, set property-behavior-member-decl ::= accessor-requirement-decl property-behavior-member-decl ::= initial-value-requirement-decl

Bindings within Behavior Declarations

Inside a behavior declaration, self is implicitly bound to the value that contains the property instantiated using this behavior. For a freestanding property at global or local scope, this will be the empty tuple (), and for a static or class property, this will be the metatype. Within the behavior declaration, the type of self is abstract and represented by the implicit generic type parameter Self. Constraints can be placed on Self in the generic signature of the behavior, to make protocol members available on self:

protocolFungible{typealias Fungus func funge()->Fungus}varbehavior runcible<Value where Self:Fungible, Self.Fungus ==Value>:Value{get{returnself.funge()}}

Lookup within self is not implicit within behaviors and must always be explicit, since unqualified lookup refers to the behavior's own members. self is immutable except in mutating methods, where it is considered an inout parameter unless the Self type has a class constraint. self cannot be accessed within inline initializers of the behavior's storage or in init declarations, since these may run during the container's own initialization phase.

Definitions within behaviors can refer to other members of the behavior by unqualified lookup, or if disambiguation is necessary, by qualified lookup on the behavior's name (since self is already taken to mean the containing value):

varbehavior foo<Value>: Value {var x: Int init(){ x =1738}mutatingfunc update(x:Int){ foo.x = x // Disambiguate reference to behavior storage }}

If the behavior includes accessor requirement declarations, then the declared accessor names are bound as functions with labeled arguments:

varbehavior fakeComputed<Value>: Value { accessor get()-> Value mutating accessor set(newValue: Value)get{returnget()}set{set(newValue: newValue)}}

Note that the behavior's own core accessor implementations get { ... } and set { ... } are not referenceable this way.

If the behavior includes an initial value requirement declaration, then the identifier initialValue is bound as a get-only computed property that evaluates the initial value expression for the property

Properties and Methods in Behaviors

Behaviors may include property and method declarations. Any storage produced by behavior properties is expanded into the containing scope of a property using the behavior.

varbehavior runcible<Value>: Value {var x: Int =0lety:String=""...}var[runcible] a: Int // expands to: var`a.[runcible].x`:Intlet`a.[runcible].y`:Stringvara:Int{...}

For public behaviors, this is inherently fragile, so adding or removing storage is a breaking change. Resilience can be achieved by using a resilient type as storage. The instantiated properties must also be of types that are visible to potential users of the behavior, meaning that public behaviors must use storage with types that are either public or internal-with-availability, similar to the restrictions on inlineable functions.

Method and computed property implementations have only immutable access to self and their storage by default, unless they are mutating. (As with computed properties, setters are mutating by default unless explicitly marked nonmutating).

init in Behaviors

The storage of a behavior must be initialized, either by inline initialization, or by an init declaration within the initializer:

varbehavior inlineInitialized<Value>: Value {var x: Int =0 // initialized inline ...}var behavior initInitialized<Value>: Value {varx:Intinit(){ x =0}}

Behaviors can contain at most one init declaration, which must take no parameters. This init declaration cannot take a visibility modifier; it is always as visible as the behavior itself. Neither inline initializers nor init declaration bodies may reference self, since they will be executed during the initialization of a property's containing value.

Initial Value Requirement Declaration

An initial value requirement declaration specifies that a behavior requires any property declared using the behavior to be declared with an initial value expression.

initial-value-requirement-decl ::= 'initialValue'

The initial value expression from the property declaration is coerced to the property's type and bound to the initialValue identifier in the scope of the behavior. Loading from initialValue behaves like a get-only computed property, evaluating the expression every time it is loaded:

varbehavior evalTwice<Value>: Value { initialValue get { // Evaluate the initial value twice, for whatever reason. _ = initialValue return initialValue }}var[evalTwice] test:()=print("test") // Prints "test" twice _ = evalTwice

A property declared with a behavior must have an initial value expression if and only if the behavior has an initial value requirement.

Accessor Requirement Declarations

An accessor requirement declaration specifies that a behavior requires any property declared using the behavior to provide an accessor implementation. An accessor requirement declaration is introduced by the contextual accessor keyword:

accessor-requirement-decl ::= attribute* decl-modifier* 'accessor' identifier function-signature function-body?

An accessor requirement declaration looks like, and serves a similar role to, a function requirement declaration in a protocol. A property using the behavior must supply an implementation for each of its accessor requirements that don't have a default implementation. The accessor names (with labeled arguments) are bound as functions within the behavior declaration:

// Reinvent computed properties varbehavior foobar<Value>: Value { accessor foo()-> Value mutating accessor bar(bas: Value) get {returnfoo()}set{bar(bas: newValue)}}var[foobar] foo: Int {foo{return0}bar{ // Parameter gets the name 'bas' from the accessor requirement // by default, as with built-in accessors today. print(bas)}}var[foobar] bar: Int {foo{return0}bar(myNewValue){ // Parameter name can be overridden as well print(myNewValue)}}

Accessor requirements can be made optional by specifying a default implementation:

// Reinvent property observers varbehavior observed<Value>: Value { // Requirements initialValue mutating accessor willSet(newValue: Value){ // do nothing by default }mutating accessor didSet(oldValue: Value){ // do nothing by default } // Implementation init(){ value = initialValue }get{return value }set{willSet(newValue: newValue)letoldValue= value value = newValue didSet(oldValue: oldValue)}}

Accessor requirements cannot take visibility modifiers; they are always as visible as the behavior itself.

Like methods, accessors are not allowed to mutate the storage of the behavior or self unless declared mutating. Mutating accessors can only be invoked by the behavior from other mutating contexts.

Core Accessor Declarations

The behavior implements the property by defining its core accessors, get and optionally set. If a behavior only provides a getter, it produces read-only properties; if it provides both a getter and setter, it produces mutable properties (though properties that instantiate the behavior may still control the visibility of their setters). It is an error if a behavior declaration does not provide at least a getter.

Using Behaviors in Property Declarations

Property declarations gain the ability to instantiate behavior, with arbitrary accessors:

property-decl ::= attribute* decl-modifier* core-property-decl core-property-decl ::= ('var' | 'let') behavior? pattern-binding ((',' pattern-binding)+ | accessors)? behavior ::= '[' visibility? decl-ref ']' pattern-binding ::= var-pattern (':' type)? inline-initializer? inline-initializer ::= '=' expr accessors ::= '{' accessor+ '}' | brace-stmt // see notes about disambiguation accessor ::= decl-modifier* decl-ref accessor-args? brace-stmt accessor-args ::= '(' identifier (',' identifier)* ')'

For example:

publicvar[behavior] prop: Int {accessor1{body()} behavior.accessor2(arg){body()}}

If multiple properties are declared in the same declaration, the behavior apply to every declared property. let properties cannot yet use behaviors.

If the behavior requires accessors, the implementations for those accessors are taken from the property's accessor declarations, matching by name. To support future composition of behaviors, the accessor definitions can use qualified names behavior.accessor. If an accessor requirement takes parameters, but the definition in for the property does not explicitly name parameters, the parameter labels from the behavior's accessor requirement declaration are implicitly bound by default.

varbehavior foo<Value>: Value { accessor bar(arg: Int)...}var[foo] x: Int {bar{print(arg)} // `arg` implicitly bound }var[foo] x: Int {bar(myArg){print(myArg)} // `arg` explicitly bound to `myArg` }

If any accessor definition in the property does not match up to a behavior requirement, it is an error.

The shorthand for get-only computed properties is only allowed for computed properties that use no behaviors. Any property that uses behaviors with accessors must name all those accessors explicitly.

If a property with behaviors declares an inline initializer, the initializer expression is captured as the implementation of a computed, get-only property which is bound to the behavior's initializer requirement. If the behavior does not have a behavior requirement, then it is an error to use an inline initializer expression. Conversely, it is an error not to provide an initializer expression to a behavior that requires one.

Properties cannot be declared using behaviors inside protocols.

Under this proposal, even if a property with a behavior has an initial value expression, the type is always required to be explicitly declared. Behaviors also do not allow for out-of-line initialization of properties. Both of these restrictions can be lifted by future extensions; see the Future directions section below.

Impact on existing code

By itself, this is an additive feature that doesn't impact existing code. However, with some of the future directions suggested, it can potentially obsolete lazy, willSet/didSet, and @NSCopying as hardcoded language features. We could grandfather these in, but my preference would be to phase them out by migrating them to library-based property behavior implementations. (Removing them should be its own separate proposal, though.)

Alternatives considered

Using a protocol (formal or not) instead of a new declaration

A previous iteration of this proposal used an informal instantiation protocol for property behaviors, desugaring a behavior into function calls, so that:

var[lazy] foo =1738

would act as sugar for something like this:

var`foo.[lazy]`=lazy(var:Int.self, initializer:{1738})varfoo:Int{get{return`foo.[lazy]`[varIn:self, initializer:{1738}]}set{`foo.[lazy]`[varIn:self, initializer:{1738}]= newValue }}

There are a few disadvantages to this approach:

  • Behaviors would pollute the namespace, potentially with multiple global functions and/or types.
  • In practice, it would require every behavior to be implemented using a new (usually generic) type, which introduces runtime overhead for the type's metadata structures.
  • The property behavior logic ends up less clear, being encoded in unspecialized language constructs.
  • Determining the capabilities of a behavior relied on function overload resolution, which can be fiddly, and would require a lot of special case diagnostic work to get good, property-oriented error messages out of.
  • Without severely complicating the informal protocol, it would be difficult to support eager vs. deferred initializers, or allow mutating access to self concurrently with the property's own storage without violating inout aliasing rules. The code generation for standalone behavior decls can hide this complexity.

Making property behaviors a distinct declaration undeniably increases the language size, but the demand for something like behaviors is clearly there. In return for a new declaration, we get better namespacing, more efficient code generation, clearer, more descriptive code for their implementation, and more expressive power with better diagnostics. I argue that the complexity can pay for itself, today by eliminating several special-case language features, and potentially in the future by generalizing to other kinds of behaviors (or being subsumed by an all-encompassing macro system). For instance, a future func behavior could conceivably provide Python decorator-like behavior for transforming function bodies.

"Template"-style behavior declaration syntax

John McCall proposed a "template"-like syntax for property behaviors, used in a previous revision of this proposal:

behaviorvar[lazy] name: Value =initialValue{...}

It's appealing from a declaration-follows-use standpoint, and provides convenient places to slot in name, type, and initial value bindings. However, this kind of syntax is unprecedented in Swift, and in initial review, was not popular.

Declaration syntax for properties using behaviors

Alternatives to the proposed var [behavior] propertyName syntax include:

  • A different set of brackets, var (behavior) propertyName or var {behavior} propertyName. Parens have the problem of being ambiguous with a tuple var declaration, requiring lookahead to resolve. Square brackets also work better with other declarations behaviors could be extended to apply to in the future, such as subscripts or functions
  • An attribute, such as @behavior(lazy) or behavior(lazy) var. This is the most conservative answer, but is clunky.
  • Use the behavior function name directly as an attribute, so that e.g. @lazy works.
  • Use a new keyword, as in var x: T by behavior.
  • Something on the right side of the colon, such as var x: lazy(T). To me this reads like lazy(T) is a type of some kind, which it really isn't.

Future directions

The functionality proposed here is quite broad, so to attempt to minimize the review burden of the initial proposal, I've factored out several aspects for separate consideration:

Behaviors for immutable let properties

Since we don't have an effects system (yet?), let behavior implementations have the potential to invalidate the immutability assumptions expected of let properties, and it would be the programmer's responsibility to maintain them. We don't support computed lets for the same reason, so I suggest leaving lets out of property behaviors for now. let behaviors could be added in the future when we have a comprehensive design for immutable computed properties and/or functions.

Type inference of properties with behaviors

There are subtle issues with inferring the type of a property using a behavior when the behavior introduces constraints on the property type. If you have something like this:

varbehavior uint16only:UInt16{...}var[uint16only] x =1738

there are two, and possibly more, ways to define what happens:

  • We type-check the initializer expression in isolation before resolving behaviors. In this case, 1738 would type-check by defaulting to Int, and then we'd raise an error instantiating the uint16only behavior, which requires a property to have type UInt16.
  • We apply the behaviors before type-checking the initializer expression, introducing generic constraints on the contextual type of the initializer. In this case, applying the uint16only behavior would constrain the contextual type of the initializer to UInt16, and we'd successfully type-check the literal as a UInt16.

There are merits and downsides to both approaches. To allow these issues to be given proper consideration, I'm subsetting them out by proposing to first require that properties with behaviors always declare an explicit type.

Composing behaviors

It is useful to be able to compose behaviors, for instance, to have a lazy property with observers that's also synchronized. Relatedly, it is useful for subclasses to be able to override their inherited properties by applying behaviors over the base class implementation, as can be done with didSet and willSet today. Linear composition can be supported by allowing behaviors to stack, each referring to the underlying property beneath it by super or some other magic binding. However, this form of composition can be treacherous, since it allows for "incorrect" compositions of behaviors. One of lazy • synchronized or synchronized • lazy is going to do the wrong thing. This possibility can be handled somewhat by allowing certain compositions to be open-coded; John McCall has suggested that every composition ought to be directly implemented as an entirely distinct behavior. That of course has an obvious exponential explosion problem; it's infeasible to anticipate and hand-code every useful combination of behaviors. These issues deserve careful separate consideration, so I'm leaving behavior composition out of this initial proposal.

Deferred evaluation of initialization expressions

This proposal does not suggest changing the allowed operations inside initialization expressions; in particular, an initialization of an instance property may not refer to self or other instance properties or methods, due to the potential for the expression to execute before the value is fully initialized:

structFoo{vara=1varb= a // Not allowed varc=foo() // Not allowed func foo(){}}

This is inconvenient for behaviors like lazy that only ever evaluate the initial value expression after the true initialization phase has completed, and where it's desirable to reference self to lazily initialize. Behaviors could be extended to annotate the initializer as "deferred", which would allow the initializer expression to refer to self, while preventing the initializer expression from being evaluated at initialization time. (If we consider behaviors to be essentially always fragile, this could be inferred from the behavior implementation.)

Out-of-line initialization with behaviors

This proposal also does not allow for behaviors that support out-of-line initialization, as in:

func foo(){ // Out-of-line local variable initialization var[behavior] x: Int x =1}structFoo{var[behavior] y: Int init(){ // Out-of-line instance property initialization y =1}}

This is a fairly serious limitation for instance properties. There are a few potential approaches we can take. One is to allow a behavior's init logic to take an out-of-line initialization as a parameter, either directly or by having a different constraint on the initializer requirement that only allows it to be referred to from init (the opposite of "deferred"). It can also be supported indirectly by linear behavior composition, if the default root super behavior for a stack of properties defaults to a plain old stored property, which can then follow normal initialization rules. This is similar to how didSet/willSet behave today. However, this would not allow behaviors to change the initialization behavior in any way.

Binding the name of a property using the behavior

There are a number of clever things you can do with the name of a property if it can be referenced as a string, such as using it to look up a value in a map, to log, or to serialize. We could conceivably support a name requirement declaration:

varbehavior echo<Value:StringLiteralConvertible>:Value{name: String get{return name }}var[echo] echo: String print(echo) // => echo

Overloading behaviors

It may be useful for behaviors to be overloadable, for instance to give a different implementation to computed and stored variants of a concept:

// A behavior for stored properties... varbehavior foo<Value>: Value { initialValue var value: Value = initialValue get {...}set{...}} // Same behavior for computed properties... varbehavior foo<Value>: Value { initialValue accessor get()-> Value accessor set(newValue: Value)get{...}set{...}}

We could resolve overloads by accessors, type constraints on Value, and/or initializer requirements. However, determining what this overload signature should be, and also the exciting interactions with type inference from initializer expressions, should be a separate discussion.

Accessing "out-of-band" behavior members

It is useful to add out-of-band operations to a property that aren't normal members of its formal type, for instance, to clear a lazy property to be recomputed later, or to reset a property to an implementation-defined default value. This is useful, but it complicates the design of the feature. Aside from the obvious surface-level concerns of syntax for accessing these members, this also exposes behaviors as interface rather than purely an implementation detail, meaning their interaction with resilience, protocols, class inheritance, and other abstractions needs to be designed. It's also a fair question whether out-of- band members should be tied to behaviors at all--it could be useful to design out-of-band members as an independent feature independent with behaviors.

close