| Contents |
|---|
| Swift 5.1 |
| Swift 5.0 |
| Swift 4.2 |
| Swift 4.1 |
| Swift 4.0 |
| Swift 3.1 |
| Swift 3.0 |
| Swift 2.2 |
| Swift 2.1 |
| Swift 2.0 |
| Swift 1.2 |
| Swift 1.1 |
| Swift 1.0 |
Self can now be used inside member functions and for function arguments of structs and enums to refer to the containing type.
Enum cases can now be matched against an optional enum without requiring a ‘?’ at the end of the pattern.
enum Foo { case zero, one } let foo: Foo? = .zero switch foo { case .zero: break case .one: break case .none: break }
weak and unowned variables can now be used inside types that declare Equatable or Hashable conformance.
An @autoclosure closure can now be a typealias.
class Foo { typealias FooClosure = () -> String func fooFunction(closure: @autoclosure FooClosure) {} }
Functions marked with @objc can now return Self
@objc func returnDynamicSelf() -> Self { return self }
Assigning ‘.none’ to an optional enum which also has a ‘none’ case or comparing such an enum with ‘.none’ will now warn. Such expressions create an ambiguity because the compiler chooses Optional.none over Foo.none.
enum Foo { case none } // Assigned Optional.none instead of Foo.none let foo: Foo? = .none // Comparing with Optional.none instead of Foo.none let isEqual = foo == .none
The compiler will provide a warning along with a fix-it to replace ‘.none’ with ‘Optional.none’ or ‘Foo.none’ to resolve the ambiguity.
Key path expressions can now include references to tuple elements.
Single-parameter functions accepting values of type Any are no longer preferred over other functions.
func foo(_: Any) { print("Any") } func foo<T>(_: T) { print("T") } foo(0) // prints "Any" in Swift < 5.1, "T" in Swift 5.1
Array and ContiguousArray now have init(unsafeUninitializedCapacity:initializingWith:), which provides access to the array's uninitialized storage.
Add new entries to the top of this section, not here!
The standard library now contains a Result type for manually propagating errors.
enum Result<Success, Failure: Error> { case success(Success) case failure(Failure) }
This type serves a complementary role to that of throwing functions and initializers. Use Result in situations where automatic error propagation or try-catch blocks are undesirable, such as in asynchronous code or when accumulating the results of successive error-producing operations.
Error now conforms to itself. This allows for the use of Error itself as the argument for a generic parameter constrained to Error.
Swift 3 mode has been removed. Supported values for the -swift-version flag are 4, 4.2, and 5.
String interpolation has been overhauled to improve its performance, clarity, and efficiency.
Note that the old _ExpressibleByStringInterpolation protocol has been removed; any code making use of this protocol will need to be updated for the new design. An #if compiler block can be used to conditionalize code between 4.2 and 5.0, for example:
#if compiler(<5.0) extension MyType : _ExpressibleByStringInterpolation { ... } #else extension MyType : ExpressibleByStringInterpolation { ... } #endif
If T conforms to one of the ExpressibleBy* protocols and literal is a literal expression, then T(literal) will construct a literal of type T using the corresponding protocol, rather than calling a constructor member of T with a value of the protocol's default literal type.
For example, expressions like UInt64(0xffff_ffff_ffff_ffff) are now valid, where previously they would overflow the default integer literal type of Int.
In Swift 5 mode, try? with an expression of Optional type will flatten the resulting Optional, instead of returning an Optional of an Optional.
In Swift 5 mode, @autoclosure parameters can no longer be forwarded to @autoclosure arguments in another function call. Instead, you must explicitly call the function value with (); the call itself is wrapped inside an implicit closure, guaranteeing the same behavior as in Swift 4 mode.
Example:
func foo(_ fn: @autoclosure () -> Int) {} func bar(_ fn: @autoclosure () -> Int) { foo(fn) // Incorrect, `fn` can't be forwarded and has to be called foo(fn()) // Ok }
Single-element labeled tuple expressions, for example (label: 123), were allowed in some contexts but often resulted in surprising, inconsistent behavior that varied across compiler releases. They are now completely disallowed.
Note that single-element labeled types, for example var x: (label: Int), have already been prohibited since Swift 3.
In Swift 5 mode, a class method returning Self can no longer be overridden with a method returning a non-final concrete class type. Such code is not type safe and will need to be updated.
For example,
class Base { class func factory() -> Self { ... } } class Derived : Base { class override func factory() -> Derived { ... } }
In Swift 5 mode, the type of self in a convenience initializer of a non-final class is now the dynamic Self type, and not the concrete class type.
Protocols can now constrain their conforming types to those that subclasses a given class. Two equivalent forms are supported:
protocol MyView : UIView { ... } protocol MyView where Self : UIView { ... }
Note that Swift 4.2 accepted the second form, but it was not fully implemented and could sometimes crash at compile time or run time.
Extension binding now supports extensions of nested types which themselves are defined inside extensions. Previously this could fail with some declaration orders, producing spurious “undeclared type” errors.
Exclusive memory access is now enforced at runtime by default in optimized (-O/-Osize) builds. Programs that violate exclusivity will trap at runtime with an “overlapping access” diagnostic message. This can be disabled via a command line flag: -enforce-exclusivity=unchecked, but doing so may result in undefined behavior.
Runtime violations of exclusivity typically result from simultaneous access of class properties, global variables (including variables in top-level code), or variables captured by escaping closures.
The @dynamicCallable attribute enables nominal types to be “callable” via a simple syntactic sugar. The primary use case is dynamic language interoperability.
Toy example:
@dynamicCallable struct ToyCallable { func dynamicallyCall(withArguments: [Int]) {} func dynamicallyCall(withKeywordArguments: KeyValuePairs<String, Int>) {} } let x = ToyCallable() x(1, 2, 3) // desugars to `x.dynamicallyCall(withArguments: [1, 2, 3])` x(label: 1, 2) // desugars to `x.dynamicallyCall(withKeywordArguments: ["label": 1, "": 2])`
In Swift 5 mode, attempting to declare a static property with the same name as a nested type is now always correctly rejected. Previously, it was possible to perform such a redeclaration in an extension of a generic type.
For example:
struct Foo<T> {} extension Foo { struct i {} // compiler error: Invalid redeclaration of 'i' // (prior to Swift 5, this did not produce an error) static var i: Int { return 0 } }
In Swift 5 mode, when casting an optional value to a generic placeholder type, the compiler will be more conservative with the unwrapping of the value. The result of such a cast now more closely matches the result you would get in a non-generic context.
For example:
func forceCast<U>(_ value: Any?, to type: U.Type) -> U { return value as! U } let value: Any? = 42 print(forceCast(value, to: Any.self)) // prints: Optional(42) // (prior to Swift 5, this would print: 42) print(value as! Any) // prints: Optional(42)
Key paths now support the \.self keypath, which is a WritableKeyPath that refers to its entire input value:
let id = \Int.self var x = 2 print(x[keyPath: id]) // prints 2 x[keyPath: id] = 3 print(x[keyPath: id]) // prints 3
The DictionaryLiteral type has been renamed to KeyValuePairs. A typealias preserves the old name for compatibility.
Default arguments are now printed in SourceKit-generated interfaces for Swift modules, instead of just using a placeholder default.
unowned and unowned(unsafe) variables now support Optional types.
Designated initializers with variadic parameters are now correctly inherited in subclasses.
Extensions of concrete subclasses of generic classes can now contain @objc members.
Complex recursive type definitions involving classes and generics that would previously cause deadlocks at run time are now fully supported.
In Swift 5 mode, when setting a property from within its own didSet or willSet observer, the observer will now only avoid being recursively called if the property is set on self (either implicitly or explicitly).
For example:
class Node { var children = [Node]() var depth: Int = 0 { didSet { if depth < 0 { // Will not recursively call didSet, as setting depth on self (same // with `self.depth = 0`). depth = 0 } // Will call didSet for each of the children, as we're not setting the // property on self (prior to Swift 5, this did not trigger property // observers to be called again). for child in children { child.depth = depth + 1 } } } }
Add new entries to the top of this section, not here!
The standard library now provides a unified set of randomization functionality. Integer types, floating point types, and Bool all introduce a new static method that creates a random value.
let diceRoll = Int.random(in: 1 ... 6) let randomUnit = Double.random(in: 0 ..< 1) let randomBool = Bool.random()
There are also additions to select a random element from a collection or shuffle its contents.
let greetings = ["hey", "hello", "hi", "hola"] let randomGreeting = greetings.randomElement()! // This returns an Optional let newGreetings = greetings.shuffled() // ["hola", "hi", "hey", "hello"]
Core to the randomization functionality is a new RandomNumberGenerator protocol. The standard library defines its own random number generator called SystemRandomNumberGenerator which is backed by a secure and thread-safe random number generator on each platform. All the randomization functions have a using: parameter that take a RandomNumberGenerator that users can pass in their own random number generator.
struct MersenneTwister: RandomNumberGenerator { func next() -> UInt64 { // implementation } } var mt = MersenneTwister() let diceRoll = Int.random(in: 1 ... 6, using: &mt)
The new CaseIterable protocol describes types which have a static “allCases” property that is used to describe all of the cases of the type. Swift will synthesize this “allCases” property for enums that have no associated values. For example:
enum Suit: CaseIterable { case heart case club case diamond case spade } print(Suit.allCases) // prints [Suit.heart, Suit.club, Suit.diamond, Suit.spade]
Protocol conformances are now able to be synthesized in extensions in the same file as the type definition, allowing automatic synthesis of conditional conformances to Hashable, Equatable and Codable (both Encodable and Decodable). For instance, if there is a generic wrapper type that can only be Equatable when its wrapped type is also Equatable, the == method can be automatically constructed by the compiler:
struct Generic<Param> { var property: Param } extension Generic: Equatable where Param: Equatable {} // Automatically synthesized inside the extension: // static func ==(lhs: Generic, rhs: Generic) -> Bool { // return lhs.property == rhs.property // }
Code that wants to be as precise as possible should generally not conditionally conform to Codable directly, but rather its two constituent protocols Encodable and Decodable, or else one can only (for instance) decode a Generic<Param> if Param is Encodable in addition to Decodable, even though Encodable is likely not required:
// Unnecessarily restrictive: extension Generic: Codable where Param: Codable {} // More precise: extension Generic: Encodable where Param: Encodable {} extension Generic: Decodable where Param: Decodable {}
Finally, due to Decodable having an init requirement, it is not possible to conform to Decodable in an extension of a non-final class: such a class needs to have any inits from protocols be required, which means they need to be in the class definition.
ImplicitlyUnwrappedOptional<T> is now an unavailable typealias of Optional<T>. Declarations annotated with ! have the type Optional<T>. If an expression involving one of these values will not compile successfully with the type Optional<T>, it is implicitly unwrapped, producing a value of type T.
In some cases this change will cause code that previously compiled to need to be adjusted. Please see this blog post for more information.
The standard library now uses a high-quality, randomly seeded, universal hash function, represented by the new public Hasher struct.
“Random seeding” varies the result of hashValue on each execution of a Swift program, improving the reliability of the standard library's hashed collections such as Set and Dictionary. In particular, random seeding enables better protection against (accidental or deliberate) hash-flooding attacks.
This change fulfills a long-standing prophecy in Hashable's documentation:
Hash values are not guaranteed to be equal across different executions of your program. Do not save hash values to use during a future execution.
As a consequence of random seeding, the elements in Set and Dictionary values may have a different order on each execution. This may expose some bugs in existing code that accidentally relies on repeatable ordering.
Additionally, the Hashable protocol now includes an extra function requirement, hash(into:). The new requirement is designed to be much easier to implement than the old hashValue property, and it generally provides better hashing. To implement hash(into:), simply feed the exact same components of your type that you compare in Equatable's == implementation to the supplied Hasher:
struct Foo: Hashable { var a: String? var b: [Int] var c: [String: Int] static func ==(lhs: Foo, rhs: Foo) -> Bool { return lhs.a == rhs.a && lhs.b == rhs.b && lhs.c == rhs.c } func hash(into hasher: inout Hasher) { hasher.combine(a) hasher.combine(b) hasher.combine(c) } }
Automatic synthesis for Hashable (SE-0185) has been updated to generate hash(into:) implementations. For example, the == and hash(into:) implementations above are equivalent to the ones synthesized by the compiler, and can be removed without changing the meaning of the code.
Synthesis has also been extended to support deriving hashValue from hash(into:), and vice versa. Therefore, code that only implements hashValue continues to work in Swift 4.2. This new compiler functionality works for all types that can implement Hashable, including classes.
Note that these changes don‘t affect Foundation’s hashing interface. Classes that subclass NSObject should override the hash property, like before.
In certain controlled environments, such as while running particular tests, it may be helpful to selectively disable hash seed randomization, so that hash values and the order of elements in Set/Dictionary values remain consistent across executions. You can disable hash seed randomization by defining the environment variable SWIFT_DETERMINISTIC_HASHING with the value of 1. The Swift runtime looks at this variable during process startup and, if it is defined, replaces the random seed with a constant value.
The behavior of .description and .debugDescription for floating-point numbers has been changed. Previously these unconditionally printed a fixed number of decimal digits (e.g. 15 and 17 for Double, respectively). They now print exactly as many digits as are needed for the resulting string to convert back to the original source value, and no more. For more details, see the original bug report and the linked pull request.
Various function-like declarations can now be marked as @inlinable, making their bodies available for optimizations from other modules.
Inlinable function bodies must only reference public declarations, unless the referenced declaration is marked as @usableFromInline.
Note that the presence of the attribute itself does not force inlining or any other optimization to be performed, nor does it have any effect on optimizations performed within a single module.
The C long double type is now imported as Float80 on i386 and x86_64 macOS and Linux. The tgmath functions in the Darwin and glibc modules now support Float80 as well as Float and Double. Several tgmath functions have been made generic over [Binary]FloatingPoint so that they will automatically be available for any conforming type.
The standard library types Optional, Array, ArraySlice, ContiguousArray, Dictionary, Range, and ClosedRange now conform to the Hashable protocol when their element or bound types (as the case may be) conform to Hashable. This makes synthesized Hashable implementations available for types that include stored properties of these types.
Custom compile-time warnings or error messages can be emitted using the #warning(_:) and #error(_:) directives.
#warning("this is incomplete") #if MY_BUILD_CONFIG && MY_OTHER_BUILD_CONFIG #error("MY_BUILD_CONFIG and MY_OTHER_BUILD_CONFIG cannot both be set") #endif
Public classes may now have internal required initializers. The rule for required initializers is that they must be available everywhere the class can be subclassed, but previously we said that required initializers on public classes needed to be public themselves. (This limitation is a holdover from before the introduction of the open/public distinction in Swift 3.)
C macros containing casts are no longer imported to Swift if the type in the cast is unavailable or deprecated, or produces some other diagnostic when referenced. (These macros were already only imported under very limited circumstances with very simple values, so this is unlikely to affect real-world code.)
Runtime query of conditional conformances is now implemented. Therefore, a dynamic cast such as value as? P, where the dynamic type of value conditionally conforms to P, will succeed when the conditional requirements are met.
Compile-time testing for the existence and importability of modules is now implemented as a build configuration test. The canImport test allows the development of features that require a possibly-failing import declaration across multiple platforms.
#if canImport(UIKit) import UIKit class MyView : UIView {} #elseif canImport(AppKit) import AppKit class MyView : NSView {} #else class MyView : CustomView {} #endif
If an initializer is declared in a different module from a struct, it must use self.init(…) or self = … before returning or accessing self. Failure to do so will produce a warning in Swift 4 and an error in Swift 5. This is to keep a client app from accidentally depending on a library's implementation details, and matches an existing restriction for classes, where cross-module initializers must be convenience initializers.
This will most commonly affect code that extends a struct imported from C. However, most imported C structs are given a zeroing no-argument initializer, which can be called as self.init() before modifying specific properties.
Swift library authors who wish to continue allowing initialization on a per-member basis should explicitly declare a public memberwise initializer for clients in other modules to use.
The standard library now defines the conformances of Optional, Array, Dictionary, and Set to Encodable and Decodable as conditional conformances, available only when their type parameters conform to Encodable or Decodable, respectively.
Index types for most standard library collections now conform to Hashable. These indices can now be used in key-path subscripts and hashed collections:
let s = "Hashable" let p = \String.[s.startIndex] s[keyPath: p] // "H"
SE-0143 The standard library types Optional, Array, ArraySlice, ContiguousArray, and Dictionary now conform to the Equatable protocol when their element types conform to Equatable. This allows the == operator to compose (e.g., one can compare two values of type [Int : [Int?]?] with ==), as well as use various algorithms defined for Equatable element types, such as index(of:).
SE-0157 is implemented. Associated types can now declare “recursive” constraints, which require that the associated type conform to the enclosing protocol. The standard library protocols have been updated to make use of recursive constraints. For example, the SubSequence associated type of Sequence follows the enclosing protocol:
protocol Sequence { associatedtype Element associatedtype SubSequence: Sequence where SubSequence.Element == Element, SubSequence.SubSequence == SubSequence // ... } protocol Collection: Sequence where Self.SubSequence: Collection { // ... }
As a result, a number of new constraints have been introduced into the standard library protocols:
Indices associated type have the same traversal requirements as its enclosing protocol, e.g., Collection.Indices conforms to Collection, BidirectionalCollection.Indices conforms to BidirectionalCollection, and so onNumeric.Magnitude conform to NumericSubSequence types for lazy filter and map*Indexable protocolsSE-0161 is fully implemented. KeyPaths now support subscript, optional chaining, and optional force-unwrapping components.
It is no longer valid to use the ownership keywords weak and unowned for property declarations in protocols. These keywords are meaningless and misleading when used in a protocol as they don't have any effect.
In Swift 3 and 4 mode the following example will produce a warning with a fix-it to remove the keyword. In Swift 5 mode and above an error will be produced.
class A {} protocol P { weak var weakVar: A? { get set } unowned var unownedVar: A { get set } }
Structs and enums that declare a conformance to Equatable/Hashable now get an automatically synthesized implementation of ==/hashValue. For structs, all stored properties must be Equatable/Hashable. For enums, all enum cases with associated values must be Equatable/Hashable.
public struct Point: Hashable { public let x: Int public let y: Int public init(x: Int, y: Int) { self.x = x self.y = y } } Point(3, 0) == Point(0, 3) // false Point(3, 0) == Point(3, 0) // true Point(3, 0).hashValue // -2942920663782199421 public enum Token: Hashable { case comma case identifier(String) case number(Int) } Token.identifier("x") == .number(5) // false Token.identifier("x") == .identifier("x") // true Token.number(50).hashValue // -2002318238093721609
If you wish to provide your own implementation of ==/hashValue, you still can; a custom implementation will replace the one synthesized by the compiler.
The standard library's Dictionary and Set types have some new features. You can now create a new dictionary from a sequence of keys and values, and merge keys and values into an existing dictionary.
let asciiTable = Dictionary(uniqueKeysWithValues: zip("abcdefghijklmnopqrstuvwxyz", 97...)) // ["w": 119, "n": 110, "u": 117, "v": 118, "x": 120, "q": 113, ...] let vegetables = ["tomato", "carrot", "onion", "onion", "carrot", "onion"] var vegetableCounts = Dictionary(zip(vegetables, repeatElement(1, count: Int.max)), uniquingKeysWith: +) vegetableCounts.merge([("tomato", 1)], uniquingKeysWith: +) // ["tomato": 2, "carrot": 2, "onion": 3]
Filtering a set or a dictionary now results in the same type. You can also now transform just the values of a dictionary, keeping the same keys, using the mapValues(_:) method.
let vowels: Set<Character> = ["a", "e", "i", "o", "u"] let asciiVowels = asciiTable.filter({ vowels.contains($0.key) }) asciiVowels["a"] // 97 asciiVowels["b"] // nil let asciiHexTable = asciiTable.mapValues({ "0x" + String($0, radix: 16) }) // ["w": "0x77", "n": "0x6e", "u": "0x75", "v": "0x76", "x": "0x78", ...]
When using a key as a dictionary subscript, you can now supply a default value to be returned if the key is not present in the dictionary.
for veg in ["tomato", "cauliflower"] { vegetableCounts[veg, default: 0] += 1 } // ["tomato": 3, "carrot": 2, "onion": 3, "cauliflower": 1]
Use the new init(grouping:by:) initializer to convert an array or other sequence into a dictionary, grouped by a particular trait.
let buttons = // an array of button instances let buttonsByStatus = Dictionary(grouping: buttons, by: { $0.isEnabled }) // How many enabled buttons? print("Enabled:", buttonsByStatus[true]?.count ?? 0)
Additionally, dictionaries and sets now have a visible capacity property and a reserveCapacity(_:) method similar to arrays, and a dictionary's keys and values properties are represented by specialized collections.
SE-0161 is partially implemented. Swift now natively supports key path objects for properties. Similar to KVC key path strings in Cocoa, key path objects allow a property to be referenced independently of accessing it from a value:
struct Point { var x, y: Double } let x = \Point.x let y = \Point.y let p = Point(x: 3, y: 4) p[keyPath: x] // gives 3 p[keyPath: y] // gives 4
Core Foundation types implicitly conform to Hashable (and Equatable), using CFHash and CFEqual as the implementation. This change applies even to “Swift 3 mode”, so if you were previously adding this conformance yourself, use #if swift(>=3.2) to restrict the extension to Swift 3.1 and below. (SR-2388)
Protocol composition types can now contain one or more class type terms, forming a class-constrained protocol composition.
For example:
protocol Paintable { func paint() } class Canvas { var origin: CGPoint } class Wall : Canvas, Paintable { func paint() { ... } } func render(_: Canvas & Paintable) { ... } render(Wall())
Note that class-constrained protocol compositions can be written and used in both Swift 3 and Swift 4 mode.
Generated headers for Swift APIs will map class-constrained protocol compositions to Objective-C protocol-qualified class types in both Swift 3 and Swift 4 mode (for instance, NSSomeClass & SomeProto & OtherProto in Swift becomes NSSomeClass <SomeProto, OtherProto> in Objective-C).
Objective-C APIs which use protocol-qualified class types differ in behavior when imported by a module compiled in Swift 3 mode and Swift 4 mode. In Swift 3 mode, these APIs will continue to import as protocol compositions without a class constraint (eg, SomeProto & OtherProto).
In Swift 4 mode, protocol-qualified class types import as class-constrained protocol compositions, for a more faithful mapping of APIs from Objective-C to Swift.
Note that the current implementation of class-constrained protocol compositions lacks three features outlined in the Swift evolution proposal:
In the evolution proposal, a class-constrained is permitted to contain two different classes as long as one is a superclass of the other. The current implementation only allows multiple classes to appear in the composition if they are identical.
In the evolution proposal, associated type and class inheritance clauses are generalized to allow class-constrained protocol compositions. The current implementation does not allow this.
In the evolution proposal, protocol inheritance clauses are allowed to contain a class, placing a requirement that all conforming types are a subclass of the given class. The current implementation does not allow this.
These missing aspects of the proposal can be introduced in a future release without breaking source compatibility with existing code.
Protocols and associated types can now contain where clauses that provide additional restrictions on associated types. For example:
protocol StringRepresentable: RawRepresentable where RawValue == String { } protocol RawStringWrapper { associatedtype Wrapped: RawRepresentable where Wrapper.RawValue == String }
In Swift 4 mode, a declaration is inferred to be @objc where it is required for semantic consistency of the programming model. Specifically, it is inferred when:
@objc declaration@objc protocol@IBAction, @IBOutlet, @IBInspectable, @GKInspectable, or @NSManagedAdditionally, in Swift 4 mode, dynamic declarations that don't have @objc inferred based on the rules above will need to be explicitly marked @objc.
Swift 3 compatibility mode retains the more-permissive Swift 3 rules for inference of @objc within subclasses of NSObject. However, the compiler will emit warnings about places where the Objective-C entry points for these inference cases are used, e.g., in a #selector or #keyPath expression, via messaging through AnyObject, or direct uses in Objective-C code within a mixed project. The warnings can be silenced by adding an explicit @objc. Uses of these entrypoints that are not statically visible to the compiler can be diagnosed at runtime by setting the environment variable SWIFT_DEBUG_IMPLICIT_OBJC_ENTRYPOINT to a value between 1 and 3 and testing the application. See the migration discussion in SE-0160.
Slicing a raw buffer no longer results in the same raw buffer type. Specifically, Unsafe[Mutable]BufferPointer.SubSequence now has type [Mutable]RandomAccessSlice<Unsafe[Mutable]RawBufferPointer>. Therefore, indexing into a raw buffer slice is no longer zero-based. This is required for raw buffers to fully conform to generic Collection. Changing the slice type resulted in the following behavioral changes:
Passing a region within buffer to another function that takes a buffer can no longer be done via subscript:
Incorrect: takesRawBuffer(buffer[i..<j])
This now requires explicit initialization, using a rebasing: initializer, which converts from a slice to a zero-based Unsafe[Mutable]RawBufferPointer:
Correct: takesRawBuffer(UnsafeRawBufferPointer(rebasing: buffer[i..<j]))
Subscript assignment directly from a buffer no longer compiles:
Incorrect: buffer[n..<m] = smaller_buffer
This now requires creation of a slice from the complete source buffer:
Correct: buffer[n..<m] = smaller_buffer.suffix(from: 0)
UnsafeRawBufferPointer's slice type no longer has a nonmutating subscript setter. So assigning into a mutable let buffer no longer compiles:
let slice = buffer[n..<m] slice[i..<j] = buffer[k..<l]
The assigned buffer slice now needs to be a var.
var slice = buffer[n..<m] slice[i..<j] = buffer[k..<l]
Covariant method overrides are now fully supported, fixing many crashes and compile-time assertions when defining or calling such methods. Examples:
class Bed {} class Nook : Bed {} class Cat<T> { func eat(snack: T) {} func play(game: String) {} func sleep(where: Nook) {} } class Dog : Cat<(Int, Int)> { // 'T' becomes concrete override func eat(snack: (Int, Int)) {} // 'game' becomes optional override func play(game: String?) {} // 'where' becomes a superclass override func sleep(where: Bed) {} }
Subscript declarations can now be defined to have generic parameter lists. Example:
extension JSON { subscript<T>(key: String) -> T? where T : JSONConvertible { // ... } }
In Swift 4 mode, Swift's type system properly distinguishes between functions that take one tuple argument, and functions that take multiple arguments.
More types of C macros which define integer constants are supported by the importer. Specifically the +, -, *, /, ^, >>, ==, <, <=, >, >= operators are now recognized, and the previously-supported <<, &&, ||, &, | operators always look through importable macros on each side of the operator. Logical AND and OR macros (&& and ||) are now imported as Boolean constants, rather than integers of value 0 or 1.
#define HIGHER (5 + 5) #define THE_EDGE (INT64_MAX - 1) #define FORTY_TWO (6 * 9) #define SPLIT (THE_EDGE / FORTY_TWO) #define HALF_AND_HALF (UINT64_MAX ^ UINT32_MAX) #define SMALL (BITWIDTH == 32) #define TINY (BITWIDTH <= 16) #define LIMITED (SMALL || TINY) // now imported as Bool.
Adds a new family of conversion initializers to all numeric types that either complete successfully without loss of information or return nil.
Swift will now warn when an NSObject subclass attempts to override the class initialize method. Swift doesn't guarantee that references to class names trigger Objective-C class realization if they have no other side effects, leading to bugs when Swift code attempted to override initialize.
C functions that “return twice” are no longer imported into Swift. Instead, they are explicitly made unavailable, so attempting to reference them will result in a compilation error.
Examples of functions that “return twice” include vfork and setjmp. These functions change the control flow of a program in ways that that Swift has never supported. For example, definitive initialization of variables, a core Swift language feature, could not be guaranteed when these functions were used.
Swift code that references these functions will no longer compile. Although this could be considered a source-breaking change, it's important to note that any use of these functions would have most likely crashed at runtime. Now, the compiler will prevent them from being used in the first place.
Indirect fields from C structures and unions are now always imported, while they previously weren't imported if they belonged to a union. This is done by naming anonymous fields. For example:
typedef struct foo_t { union { int a; double b; }; } foo_t;
Get imported as:
struct foo_t { struct __Unnamed_union___Anonymous_field0 { var a : Int { get set } var b : Double { get set } } var __Anonymous_field0 : foo_t.__Unnamed_union___Anonymous_field0 // a and b are computed properties accessing the content of __Anonymous_field0 var a : Int { get set } var b : Double { get set } }
Since new symbols are exposed from imported structure/unions, this may conflict with existing code that extended C types in order to provide their own accessors to the indirect fields.
The withoutActuallyEscaping function from SE-0103 has been implemented. To pass off a non-escaping closure to an API that formally takes an @escaping closure, but which is used in a way that will not in fact escape it in practice, use withoutActuallyEscaping to get an escapable copy of the closure and delimit its expected lifetime. For example:
func doSimultaneously(_ f: () -> (), and g: () -> (), on q: DispatchQueue) { // DispatchQueue.async normally has to be able to escape its closure // since it may be called at any point after the operation is queued. // By using a barrier, we ensure it does not in practice escape. withoutActuallyEscaping(f) { escapableF in withoutActuallyEscaping(g) { escapableG in q.async(escapableF) q.async(escapableG) q.sync(flags: .barrier) {} } } // `escapableF` and `escapableG` must be dequeued by the point // `withoutActuallyEscaping` returns. }
The old workaround of using unsafeBitCast to cast to an @escaping type is not guaranteed to work in future versions of Swift, and will now raise a warning.
Nested types may now appear inside generic types, and nested types may have their own generic parameters:
struct OuterNonGeneric { struct InnerGeneric<T> {} } struct OuterGeneric<T> { struct InnerNonGeneric {} struct InnerGeneric<T> {} } extension OuterNonGeneric.InnerGeneric {} extension OuterGeneric.InnerNonGeneric {} extension OuterGeneric.InnerGeneric {}
Constrained extensions allow same-type constraints between generic parameters and concrete types. This enables you to create extensions, for example, on Array with Int elements:
extension Array where Element == Int { }
The Sequence protocol adds two new members prefix(while:) and drop(while:) for common utility. prefix(while:) requests the longest subsequence satisfying a predicate. drop(while:) requests the remaining subsequence after dropping the longest subsequence satisfying a predicate.
The functions sizeof(), strideof(), and alignof() have been removed. Memory layout properties for a type T are now spelled MemoryLayout<T>.size, MemoryLayout<T>.stride, and MemoryLayout<T>.alignment, respectively.
The functions sizeofValue(), strideofValue(), and alignofValue() have been renamed to MemoryLayout.size(ofValue:), MemoryLayout.stride(ofValue:), and MemoryLayout.alignment(ofValue:).
The functions isUniquelyReferenced() and isUniquelyReferencedNonObjC() have been removed. Call the function isKnownUniquelyReferenced() instead.
Classes using isUniquelyReferenced() needed to inherit from NonObjectiveCBase. The NonObjectiveCBase class has been removed.
The method ManagedBufferPointer.holdsUniqueReference has been renamed to ManagedBufferPointer.isUniqueReference.
// old class SwiftKlazz : NonObjectiveCBase {} expectTrue(isUniquelyReferenced(SwiftKlazz())) var managedPtr : ManagedBufferPointer = ... if !managedPtr.holdsUniqueReference() { print("not unique") } // new class SwiftKlazz {} expectTrue(isKnownUniquelyReferenced(SwiftKlazz())) var managedPtr : ManagedBufferPointer = ... if !managedPtr.isUniqueReference() { print("not unique") }
Initializers on Int and UInt that accept an ObjectIdentifier must now use an explicit bitPattern label.
let x: ObjectIdentifier = ... // old let u = UInt(x) let i = Int(x) // new let u = UInt(bitPattern: x) let i = Int(bitPattern: x)
The collection methods partition() and partition(isOrderedBefore:) have been removed from Swift. They are replaced by the method partition(by:) which takes a unary predicate.
Calls to the partition() method can be replaced by the following code.
// old let p = c.partition() // new let p = c.first.flatMap({ first in c.partition(by: { $0 >= first }) }) ?? c.startIndex
Closure parameters are now non-escaping by default and do not require @noescape annotation. Use @escaping to indicate that a closure parameter can escape. @autoclosure(escaping) is now spelled @autoclosure @escaping. @noescape and @autoclosure(escaping) are deprecated.
To clarify their roles, *LiteralConvertible protocols have been renamed to ExpressibleBy*Literal. The protocol requirements are unchanged.
An Unsafe[Mutable]RawPointer type has been introduced. It replaces Unsafe[Mutable]Pointer<Void>. Conversion from UnsafePointer<T> to UnsafePointer<U> has been disallowed. Unsafe[Mutable]RawPointer provides an API for untyped memory access, and an API for binding memory to a type. Binding memory allows for safe conversion between pointer types.
For detailed instructions on how to migrate your code to the new API refer to the UnsafeRawPointer migration guide. See also: See bindMemory(to:capacity:), assumingMemoryBound(to:), and withMemoryRebound(to:capacity:).
The dynamicType keyword has been removed from Swift. It's replaced by a new primitive function type(of:). Existing code using the .dynamicType member to retrieve the type of an expression should migrate to this new primitive. Code using .dynamicType in conjunction with sizeof should migrate to the MemoryLayout structure introduced by SE-0101.
The following two methods were added to FloatingPoint:
func rounded(_ rule: FloatingPointRoundingRule) -> Self mutating func round( _ rule: FloatingPointRoundingRule)
These methods bind the IEEE 754 roundToIntegral operations. They provide the functionality of the C / C++ round(), ceil(), floor(), and trunc() functions along with other rounding operations.
Following onto SE-0113 and SE-0067, the following Darwin.C and glibc module mathematical operations now operate on any type conforming to FloatingPoint: fabs, sqrt, fma, remainder, fmod, ceil, floor, round, and trunc.
See also: the changes associated with SE-0067.
The FloatingPoint protocol has been expanded to include most IEEE 754 required operations. A number of useful properties have been added to the protocol, representing quantities like the largest finite value or the smallest positive normal value (these correspond to the macros such as FLT_MAX defined in C).
While almost all of the changes are additive, four changes impact existing code:
The % operator is no longer available for FloatingPoint types. It was difficult to use correctly and its semantics did not match those of the corresponding integer operation. This made it something of an attractive nuisance. The new method formTruncatingRemainder(dividingBy:) provides the old semantics if they are needed.
The static property .NaN has been renamed .nan.
The static property .quietNaN was redundant and has been removed. Use .nan instead.
The predicate isSignaling has been renamed isSignalingNaN.
See also: the changes associated with SE-0113.
Argument labels have been removed from Swift function types. They are now part of the name of a function, subscript, or initializer. Calls to a function or initializer, and subscript uses, still require argument labels as they always have:
func doSomething(x: Int, y: Int) { } doSomething(x: 0, y: 0) // argument labels are required
Unapplied references to functions or initializers no longer carry argument labels. For example:
let f = doSomething(x:y:) // inferred type is now (Int, Int) -> Void
Explicitly-written function types can no longer carry argument labels. You can still provide parameter names for documentation purposes using the ‘_’ in the argument label position:
typealias CompletionHandler = (token: Token, error: Error?) -> Void // error: function types cannot have argument labels typealias CompletionHandler = (_ token: Token, _ error: Error?) -> Void // okay: names are for documentation purposes
The access level formerly known as private is now called fileprivate. A Swift 3 declaration marked private can no longer be accessed outside its lexical scope (essentially its enclosing curly braces {}). A private declaration at the top level of a file can be accessed anywhere within the same file, as it could in Swift 2.
The standard library introduces the AnyHashable type for use in hashed heterogeneous collections. Untyped NSDictionary and NSSet Objective-C APIs now import as [AnyHashable: Any] and Set<AnyHashable>.
Swift removes the @noreturn attribute on function declarations and replaces the attribute with an empty Never type:
@noreturn func fatalError(msg: String) { ... } // old func fatalError(msg: String) -> Never { ... } // new func performOperation<T>(continuation: @noreturn T -> ()) { ... } // old func performOperation<T>(continuation: T -> Never) { ... } // new
Swift now imports Objective-C id APIs as Any. In Swift 2, id imported as AnyObject. Swift also imports untyped NSArray and NSDictionary as [Any] and [AnyHashable: Any], respectively.
Swift eliminates implicit bridging conversions. Use as to force the conversion from a Swift value type to its corresponding object. For example, use string as NSString. Use as AnyObject to convert a Swift value to its boxed id representation.
Collection subtype conversions and dynamic casts now work with protocol types:
protocol P {}; extension Int: P {} var x: [Int] = [1, 2, 3] var p: [P] = x var x2 = p as! [Int]
The hasPrefix and hasSuffix functions now consider the empty string to be a prefix and suffix of all strings.
Some non-failable UnicodeScalar initializers now return an Optional. When a UnicodeScalar cannot be constructed, these initializers return nil.
// Old var string = "" let codepoint: UInt32 = 55357 // Invalid let ucode = UnicodeScalar(codepoint) // Program crashes here. string.append(ucode)
The updated initializers allow users to write code that safely works around invalid codepoints, like this example:
// New var string = "" let codepoint: UInt32 = 55357 // Invalid if let ucode = UnicodeScalar(codepoint) { string.append(ucode) } else { // do something else }
Swift removes the protocol<...> composition construct and introduces an infix type operator & in its place.
let a: Foo & Bar let b = value as? A & B & C func foo<T : Foo & Bar>(x: T) { ... } func bar(x: Foo & Bar) { ... } typealias G = GenericStruct<Foo & Bar>
Swift previously defined the empty protocol composition (the Any type) as protocol<>. This definition has been removed from the standard library. The Any keyword behavior remains unchanged.
Swift permits you to define operators within types or their extensions. For example:
struct Foo: Equatable { let value: Int static func ==(lhs: Foo, rhs: Foo) -> Bool { return lhs.value == rhs.value } }
You must declare these operators as static (or, within a class, class final) and they must use the same signature as their global counterparts. As part of this change, protocol-declared operator requirements must be declared static explicitly:
protocol Equatable { static func ==(lhs: Self, rhs: Self) -> Bool }
Note: The type checker performance optimization described by SE-0091 is not yet implemented.
Condition clauses in if, guard, and while statements now use a more regular syntax. Each pattern or optional binding must be prefixed with case or let respectively, and all conditions are separated by , instead of where.
// before if let a = a, b = b where a == b { } // after if let a = a, let b = b, a == b { }
The NSError type now bridges to the Swift Error protocol type (formerly ErrorProtocol in Swift 3, ErrorType in Swift 2) in Objective-C APIs. NSError now bridges like other Objective-C types, e.g., NSString bridges to String.
For example, the UIApplicationDelegate method applicate(_:didFailToRegisterForRemoteNotificationsWithError:) previously accepted an NSError argument:
optional func application(_ application: UIApplication, didFailToRegisterForRemoteNotificationsWithError error: NSError)
Now it accepts an Error argument:
optional func application(_ application: UIApplication, didFailToRegisterForRemoteNotificationsWithError error: Error)
Error types imported from Cocoa[Touch] maintain all of the information in the corresponding NSError. You no longer catch let as NSError to extract, for example, the user-info dictionary.
Specific error types now contain typed accessors for their common user-info keys. For example:
catch let error as CocoaError where error.code == .fileReadNoSuchFileError { print("No such file: \(error.url)") }
Swift-defined error types can now provide localized error descriptions by adopting the new LocalizedError protocol, e.g.,
extension HomeworkError : LocalizedError { var errorDescription: String? { switch self { case .forgotten: return NSLocalizedString("I forgot it") case .lost: return NSLocalizedString("I lost it") case .dogAteIt: return NSLocalizedString("The dog ate it") } } }
New RecoverableError and CustomNSError protocols allow additional control over the handling of the error.
Function parameters with defaulted arguments are specified in declaration order. Call sites must now supply those arguments using that order:
func requiredArguments(a: Int, b: Int, c: Int) {} func defaultArguments(a: Int = 0, b: Int = 0, c: Int = 0) {} requiredArguments(a: 0, b: 1, c: 2) requiredArguments(b: 0, a: 1, c: 2) // error defaultArguments(a: 0, b: 1, c: 2) defaultArguments(b: 0, a: 1, c: 2) // error
Labeled parameters with default arguments may still be elided, so long as included arguments follow declaration order:
defaultArguments(a: 0) // ok defaultArguments(b: 1) // ok defaultArguments(c: 2) // ok defaultArguments(a: 1, c: 2) // ok defaultArguments(b: 1, c: 2) // ok defaultArguments(c: 1, b: 2) // error
Traps from force-unwrapping nil Optionals now show the source location of the force unwrap operator.
Slice types add a base property that allows public readonly access to their base collections.
Nested generic functions may now capture bindings from the environment, for example:
func outer<T>(t: T) -> T { func inner<U>(u: U) -> (T, U) { return (t, u) } return inner(u: (t, t)).0 }
Initializers are now inherited even if the base class or derived class is generic:
class Base<T> { let t: T init(t: T) { self.t = t } } class Derived<T> : Base<T> { // init(t: T) is now synthesized to call super.init(t: t) }
“Move where clause to end of declaration” is now implemented. This change allows you to write where clauses after a declaration signature and before its body. For example, before this change was implemented, you'd write:
func anyCommonElements<T : SequenceType, U : SequenceType where T.Generator.Element: Equatable, T.Generator.Element == U.Generator.Element> (lhs: T, _ rhs: U) -> Bool { ... }
Now the where clause appears just before the body:
func anyCommonElements<T : SequenceType, U : SequenceType>(lhs: T, _ rhs: U) -> Bool where T.Generator.Element: Equatable, T.Generator.Element == U.Generator.Element { ... }
The old form is currently accepted for compatibility. It will eventually be rejected.
“Allow (most) keywords in member references” is implemented. This change allows the use of members after a dot without backticks, e.g. “foo.default”, even though default is a keyword for switch statements.
Objective-C lightweight generic classes are now imported as generic types in Swift. Some limitations apply because Objective-C generics are not represented at runtime:
When an ObjC generic class is used in a checked as?, as!, or is cast, the generic parameters are not checked at runtime. The cast succeeds if the operand is an instance of the ObjC class, regardless of parameters.
let x = NSFoo<NSNumber>(value: NSNumber(integer: 0)) let y: AnyObject = x let z = y as! NSFoo<NSString> // Succeeds
Swift subclasses can only inherit from an ObjC generic class when its generic parameters are fully specified.
// Error: Can't inherit ObjC generic class with unbound parameter T class SwiftFoo1<T>: NSFoo<T> { } // OK: Can inherit ObjC generic class with specific parameters class SwiftFoo2<T>: NSFoo<NSString> { }
Swift can extend ObjC generic classes but the extensions cannot be constrained, and definitions inside the extension don‘t have access to the class’s generic parameters.
extension NSFoo { // Error: Can't access generic param T func foo() -> T { return T() } } // Error: extension can't be constrained extension NSFoo where T: NSString { }
Foundation container classes NS[Mutable]Array, NS[Mutable]Set, and NS[Mutable]Dictionary are still imported as nongeneric classes for the time being.
Enum elements can no longer be accessed as instance members in instance methods.
id in id *) are no longer assumed to always be _Nullable even if annotated otherwise._Null_unspecified on a pointee type still imports as Optional.The types UnsafePointer, UnsafeMutablePointer, AutoreleasingUnsafeMutablePointer, OpaquePointer, Selector, and Zone (formerly NSZone) now represent non-nullable pointers, i.e. pointers that are never nil. A nullable pointer is now represented using Optional, e.g. UnsafePointer<Int>? For types imported from C, non-object pointers (such as int *) now have their nullability taken into account.
One possible area of difficulty is passing a nullable pointer to a function that uses C variadics. Swift will not permit this directly. As a workaround, use the following idiom to pass a pointer-sized integer value instead:
unsafeBitCast(nullablePointer, to: Int.self)
Function parameters adopt consistent labeling across all function parameters. With this update, first parameter declarations match the existing behavior of the second and later parameters. This change makes the language simpler.
Functions that were written and called as follows:
func foo(x: Int, y: Int) {} foo(1, y: 2) func bar(a a: Int, b: Int) {} bar(a: 3, b: 4)
Are now written as follows with the same behavior at call sites:
func foo(_ x: Int, y: Int) {} foo(1, y: 2) func bar(a: Int, b: Int) {} bar(a: 3, b: 4)
Comments are now treated as whitespace when determining whether an operator is prefix, postfix, or binary. For example, these now work:
if /*comment*/!foo { ... } 1 +/*comment*/2
Comments can no longer appear between a unary operator and its argument.
foo/* comment */! // no longer works
Parse errors resulting from this change can be resolved by moving the comment outside the expression.
The location of the inout attribute moves to after the colon (:) and before the parameter type.
func foo(inout x: Int) { }
will now be written as:
func foo(x: inout Int) { }
let is no longer accepted as a parameter attribute for functions. The compiler provides a fixit to remove it from the function declaration.
var is no longer accepted as a parameter attribute for functions. The compiler provides a fixit to create a shadow copy in the function body.
func foo(var x: Int) { }
will now be written as:
func foo(x: Int) { var x = x }
The “none” members of imported NS_OPTIONS option sets are marked as unavailable when they are imported. Use [] to make an empty option set, instead of a None member.
Adds the ability to declare variables in multiple patterns in cases.
Allows the Clang importer to import ObjC symbols using substantially different Swift-like naming paradigms:
swift_name, allowing arbitrary C and Objective-C entity import names. This adds fine-grained control over the import process.documentForURL(_: NSURL) becomes document(for: URL)). Selectors are guaranteed to never be empty, to be transformed into Swift keywords, to be vacuously named (like get, set, with, for). Additional pruning rules preserve readability and sense.nil, option sets to [], and NSDictionary parameters to [:].) First argument labels are added for defaulted arguments.is, and read as assertions on the receiver.compare(_:) -> NSComparisonResult automatically import as Comparable.Attributes change from using = in parameters lists to using :, aligning with function call syntax.
// before @available(*, unavailable, renamed="MyRenamedProtocol") // after @available(*, unavailable, renamed: "MyRenamedProtocol")
Generic typealiases are now supported. For example:
typealias StringDictionary<T> = Dictionary<String, T> typealias IntFunction<T> = (T) -> Int typealias MatchingTriple<T> = (T, T, T) typealias BackwardTriple<T1, T2, T3> = (T3, T2, T1)
etc.
The @noescape attribute is extended to be a more general type attribute. You can now declare values of @noescape function type, e.g. in manually curried function signatures. You can now also declare local variables of @noescape type, and use @noescape in typealiases. For example, this is now valid code:
func apply<T, U>(@noescape f: T -> U, @noescape g: (@noescape T -> U) -> U) -> U { return g(f) }
The #line directive (which resets the logical source location for diagnostics and debug information) is renamed to #sourceLocation.
Curried function syntax (with successive parenthesized groups of arguments) is removed, and now produces a compile-time error. Use chained functional return types instead.
// Before public func project(function f: FunctionType)(p0: CGPoint, p1: CGPoint)(x: CGFloat) -> CGPoint // After public func project(function f: FunctionType) -> (p0: CGPoint, p1: CGPoint) -> (x: CGFloat) -> CGPoint
Generic signatures can now contain superclass requirements with generic parameter types, for example:
func f<Food : Chunks<Meat>, Meat : Molerat>(f: Food, m: Meat) {}
Section markers are created in ELF binaries through special objects during link time. These objects allow for the deletion of swift.ld and the use of non-BFD linkers. A new argument to swiftc is provided to select the linker used, and the gold linker is set as the default for arm-based platforms.
Catch blocks in rethrows functions may now throw errors. For example:
func process(f: () throws -> Int) rethrows -> Int { do { return try f() } catch is SomeError { throw OtherError() } }
Throwing closure arguments of a rethrowing function may now be optional. For example:
func executeClosureIfNotNil(closure: (() throws -> Void)?) rethrows { try closure?() }
The Objective-C selectors for the getter or setter of a property can now be referenced with #selector. For example:
let sel1 = #selector(getter: UIView.backgroundColor) // sel1 has type Selector let sel2 = #selector(setter: UIView.backgroundColor) // sel2 has type Selector
A key-path can now be formed with #keyPath. For example:
person.valueForKeyPath(#keyPath(Person.bestFriend.lastName))
Associated types in protocols can now be specified with a new associatedtype declaration, to replace the use of typealias:
protocol P { associatedtype Ty }
The typealias keyword is still allowed (but deprecated and produces a warning) in Swift 2.2. This warning will become an error in Swift 3.0.
Curried function syntax has been deprecated, and is slated to be removed in Swift 3.0.
The ++ and -- operators have been deprecated, and are slated to be removed in Swift 3.0. As a replacement, please use x += 1 on integer or floating point types, and x = x.successor() on Index types.
The implicit tuple splat behavior in function application has been deprecated and will be removed in Swift 3.0. For example, this code:
func foo(a : Int, b : Int) { ... } let x = (1, b: 2) foo(x) // Warning, deprecated.
should move to being written as:
foo(x.0, x.b)
New #file, #line, #column, and #function expressions have been introduced to replace the existing __FILE__, __LINE__, __COLUMN__, and __FUNCTION__ symbols. The __FILE__-style symbols have been deprecated, and will be removed in Swift 3.0.
The operator identifier lexer grammar has been revised to simplify the rules for operators that start with a dot (“.”). The new rule is that an operator that starts with a dot may contain other dots in it, but operators that start with some other character may not contain dots. For example:
x....foo --> "x" "...." "foo" x&%^.foo --> "x" "&%^" ".foo"
This eliminates a special case for the ..< operator, folding it into a simple and consistent rule.
The “C-style for loop”, which is spelled for init; comparison; increment {} has been deprecated and is slated for removal in Swift 3.0.
Three new doc comment fields, namely - keyword:, - recommended: and - recommendedover:, allow Swift users to cooperate with code completion engine to deliver more effective code completion results. The - keyword: field specifies concepts that are not fully manifested in declaration names. - recommended: indicates other declarations are preferred to the one decorated; to the contrary, - recommendedover: indicates the decorated declaration is preferred to those declarations whose names are specified.
Designated class initializers declared as failable or throwing may now return nil or throw an error, respectively, before the object has been fully initialized. For example:
class Widget : Gadget { let complexity: Int init(complexity: Int, elegance: Int) throws { if complexity > 3 { throw WidgetError.TooComplex } self.complexity = complexity try super.init(elegance: elegance) } }
All slice types now have removeFirst() and removeLast() methods.
ArraySlice.removeFirst() now preserves element indices.
Global anyGenerator() functions have been changed into initializers on AnyGenerator, making the API more intuitive and idiomatic. They have been deprecated in Swift 2.2, and will be removed in Swift 3.0.
Closures appearing inside generic types and generic methods can now be converted to C function pointers, as long as no generic type parameters are referenced in the closure's argument list or body. A conversion of a closure that references generic type parameters now produces a diagnostic instead of crashing.
Anonymously-typed members of C structs and unions can now be accessed from Swift. For example, given the following struct ‘Pie’, the ‘crust’ and ‘filling’ members are now imported:
struct Pie { struct { bool crispy; } crust; union { int fruit; } filling; }
Since Swift does not support anonymous structs, these fields are imported as properties named crust and filling having nested types named Pie.__Unnamed_crust and Pie.__Unnamed_filling.
Argument labels and parameter names can now be any keyword except var, let, or inout. For example:
NSURLProtectionSpace(host: "somedomain.com", port: 443, protocol: "https", realm: "Some Domain", authenticationMethod: "Basic")
would previously have required protocol to be surrounded in back-ticks.
Tuples (up to arity 6) whose elements are all Comparable or Equatable now implement the full set of comparison/equality operators. The comparison operators are defined in terms of lexicographical order.
The @objc(SomeName) attribute is now supported on enums and enum cases to rename the generated Objective-C declaration.
When referencing a function or initializer, one can provide the complete name, including argument labels. For example:
let fn1 = someView.insertSubview(_:at:) let fn2 = someView.insertSubview(_:aboveSubview:) let buttonFactory = UIButton.init(type:)
There is a new build configuration function, #if swift(>=x.y), which tests if the current Swift language version is at least x.y. This allows you to conditionally compile code for multiple language versions in the same file, even with different syntax, by deactivating parsing in inactive code blocks. For example:
#if swift(>=2.2) // Only this code will be parsed in Swift 3.0 func foo(x: Int) -> (y: Int) -> () {} #else // This code is ignored entirely. func foo(x: Int)(y: Int) {} #endif
The Objective-C selector of a Swift method can now be determined directly with the #selector expression, e.g.,:
let sel = #selector(insertSubview(_:aboveSubview:)) // sel has type Selector
Along with this change, the use of string literals as selectors has been deprecated, e.g.,
let sel: Selector = "insertSubview:aboveSubview:"
Generally, such string literals should be replaced with uses of #selector, and the compiler will provide Fix-Its that use #selector. In cases where this is not possible (e.g., when referring to the getter of a property), one can still directly construct selectors, e.g.:
let sel = Selector("propertyName")
Note that the compiler is now checking the string literals used to construct Selectors to ensure that they are well-formed Objective-C selectors and that there is an @objc method with that selector.
Enums imported from C now automatically conform to the Equatable protocol, including a default implementation of the == operator. This conformance allows you to use C enum pattern matching in switch statements with no additional code. (17287720)
The NSNumber.unsignedIntegerValue property now has the type UInt instead of Int, as do other methods and properties that use the NSUInteger type in Objective-C and whose names contain unsigned... Most other uses of NSUInteger in system frameworks are imported as Int as they were in Xcode 7. (19134055)
Field getters and setters are now created for named unions imported from C. In addition, an initializer with a named parameter for the field is provided. For example, given the following Objective-C typedef:
typedef union IntOrFloat { int intField; float floatField; } IntOrFloat;
Importing this typedef into Swift generates the following interface:
struct IntOrFloat { var intField: Int { get set } init(intField: Int) var floatField: Float { get set } init(floatField: Float) }
(19660119)
Bitfield members of C structs are now imported into Swift. (21702107)
The type dispatch_block_t now refers to the type @convention(block) () -> Void, as it did in Swift 1.2. This change allows programs using dispatch_block_create to work as expected, solving an issue that surfaced in Xcode 7.0 with Swift 2.0.
Note: Converting to a Swift closure value and back is not guaranteed to preserve the identity of a dispatch_block_t. (22432170)
Editing a file does not trigger a recompile of files that depend upon it if the edits only modify declarations marked private. (22239821)
Expressions interpolated in strings may now contain string literals. For example, My name is \(attributes["name"]!) is now a valid expression. (14050788)
Error messages produced when the type checker cannot solve its constraint system continue to improve in many cases.
For example, errors in the body of generic closures (for instance, the argument closure to map) are much more usefully diagnosed. (18835890)
Conversions between function types are supported, exhibiting covariance in function result types and contravariance in function parameter types. For example, it is legal to assign a function of type Any -> Int to a variable of type String -> Any. (19517003)
New defer statement. This statement runs cleanup code when the scope is exited, which is particularly useful in conjunction with the new error handling model. For example:
func xyz() throws { let f = fopen("x.txt", "r") defer { fclose(f) } try foo(f) // f is closed if an error is propagated. let f2 = fopen("y.txt", "r") defer { fclose(f2) } try bar(f, f2) // f2 is closed, then f is closed if an error is propagated. } // f2 is closed, then f is closed on a normal path
(17302850)
Printing values of certain enum types shows the enum case and payload, if any. This is not supported for @objc enums or certain enums with multiple payloads. (18334936)
You can specify availability information on your own declarations with the @available() attribute.
For example:
@available(iOS 8.0, OSX 10.10, *) func startUserActivity() -> NSUserActivity { ... }
This code fragment indicates that the startUserActivity() method is available on iOS 8.0+, on OS X v10.10+, and on all versions of any other platform. (20938565)
A new @nonobjc attribute is introduced to selectively suppress ObjC export for instance members that would otherwise be @objc. (16763754)
Nongeneric classes may now inherit from generic classes. (15520519)
Public extensions of generic types are now permitted.
public extension Array { ... }
(16974298)
Enums now support multiple generic associated values, for example:
enum Either<T, U> { case Left(T), Right(U) }
(15666173)
Protocol extensions: Extensions can be written for protocol types. With these extensions, methods and properties can be added to any type that conforms to a particular protocol, allowing you to reuse more of your code. This leads to more natural caller side “dot” method syntax that follows the principle of “fluent interfaces” and that makes the definition of generic code simpler (reducing “angle bracket blindness”). (11735843)
Protocol default implementations: Protocols can have default implementations for requirements specified in a protocol extension, allowing “mixin” or “trait” like patterns.
Availability checking. Swift reports an error at compile time if you call an API that was introduced in a version of the operating system newer than the currently selected deployment target.
To check whether a potentially unavailable API is available at runtime, use the new #available() condition in an if or guard statement. For example:
if #available(iOS 8.0, OSX 10.10, *) { // Use Handoff APIs when available. let activity = NSUserActivity(activityType:"com.example.ShoppingList.view") activity.becomeCurrent() } else { // Fall back when Handoff APIs not available. }
(14586648)
Native support for C function pointers: C functions that take function pointer arguments can be called using closures or global functions, with the restriction that the closure must not capture any of its local context. For example, the standard C qsort function can be invoked as follows:
var array = [3, 14, 15, 9, 2, 6, 5] qsort(&array, array.count, sizeofValue(array[0])) { a, b in return Int32(UnsafePointer<Int>(a).memory - UnsafePointer<Int>(b).memory) } print(array)
(16339559)
Error handling. You can create functions that throw, catch, and manage errors in Swift.
Using this capability, you can surface and deal with recoverable errors, such as “file-not-found” or network timeouts. Swift's error handling interoperates with NSError. (17158652)
Testability: Tests of Swift 2.0 frameworks and apps are written without having to make internal routines public.
Use @testable import {ModuleName} in your test source code to make all public and internal routines usable. The app or framework target needs to be compiled with the Enable Testability build setting set to Yes. The Enable Testability build setting should be used only in your Debug configuration, because it prohibits optimizations that depend on not exporting internal symbols from the app or framework. (17732115)
if statements can be labeled, and labeled break statements can be used as a jump out of the matching if statement.
An unlabeled break does not exit the if statement. It exits the enclosing loop or switch statement, or it is illegal if none exists. (19150249)
A new x? pattern can be used to pattern match against optionals as a synonym for .Some(x). (19382878)
Concatenation of Swift string literals, including across multiple lines, is now a guaranteed compile-time optimization, even at -Onone. (19125926)
Nested functions can now recursively reference themselves and other nested functions. (11266246)
Improved diagnostics:
Failable convenience initializers are allowed to return nil before calling self.init.
Designated initializers still must initialize all stored properties before returning nil; this is a known limitation. (20193929)
A new readLine() function has been added to the standard library. (15911365)
SIMD Support: Clang extended vectors are imported and usable in Swift.
This capability enables many graphics and other low-level numeric APIs (for example, simd.h) to be usable in Swift.
New guard statement: This statement allows you to model an early exit out of a scope.
For example:
guard let z = bar() else { return } use(z)
The else block is required to exit the scope (for example, with return, throw, break, continue, and so forth) or end in a call to a @noreturn function. (20109722)
Improved pattern matching: switch/case pattern matching is available to many new conditional control flow statements, including if/case, while/case, guard/case, and for-in/case. for-in statements can also have where clauses, which combine to support many of the features of list comprehensions in other languages.
A new do statement allows scopes to be introduced with the do statement.
For example:
do { //new scope do { //another scope } }
A new keyword try? has been added to Swift.
try? attempts to perform an operation that may throw. If the operation succeeds, the result is wrapped in an optional; if it fails (that is, if an error is thrown), the result is nil and the error is discarded.
For example:
func produceGizmoUsingTechnology() throws -> Gizmo { ... } func produceGizmoUsingMagic() throws -> Gizmo { ... } if let result = try? produceGizmoUsingTechnology() { return result } if let result = try? produceGizmoUsingMagic() { return result } print("warning: failed to produce a Gizmo in any way") return nil
try? always adds an extra level of Optional to the result type of the expression being evaluated. If a throwing function's normal return type is Int?, the result of calling it with try? will be Int??, or Optional<Optional<Int>>. (21692467)
Type names and enum cases now print and convert to String without qualification by default. debugPrint or String(reflecting:) can still be used to get fully qualified names. For example:
enum Fruit { case Apple, Banana, Strawberry } print(Fruit.Apple) // "Apple" debugPrint(Fruit.Apple) // "MyApp.Fruit.Apple")
(21788604)
C typedefs of block types are imported as typealiass for Swift closures.
The primary result of this is that typedefs for blocks with a parameter of type BOOL are imported as closures with a parameter of type Bool (rather than ObjCBool as in the previous release). This matches the behavior of block parameters to imported Objective-C methods. (22013912)
The type Boolean in MacTypes.h is imported as Bool in contexts that allow bridging between Swift and Objective-C types.
In cases where the representation is significant, Boolean is imported as a distinct DarwinBoolean type, which is BooleanLiteralConvertible and can be used in conditions (much like the ObjCBool type). (19013551)
Fields of C structs that are marked __unsafe_unretained are presented in Swift using Unmanaged.
It is not possible for the Swift compiler to know if these references are really intended to be strong (+1) or unretained (+0). (19790608)
The NS_REFINED_FOR_SWIFT macro can be used to move an Objective-C declaration aside to provide a better version of the same API in Swift, while still having the original implementation available. (For example, an Objective-C API that takes a Class could offer a more precise parameter type in Swift.)
The NS_REFINED_FOR_SWIFT macro operates differently on different declarations:
init methods will be imported with the resulting Swift initializer having __ prepended to its first external parameter name.
// Objective-C - (instancetype)initWithClassName:(NSString *)name NS_REFINED_FOR_SWIFT;
// Swift init(__className: String)
Other methods will be imported with __ prepended to their base name.
// Objective-C - (NSString *)displayNameForMode:(DisplayMode)mode NS_REFINED_FOR_SWIFT;
// Swift func __displayNameForMode(mode: DisplayMode) -> String
Subscript methods will be treated like any other methods and will not be imported as subscripts.
Other declarations will have __ prepended to their name.
// Objective-C @property DisplayMode mode NS_REFINED_FOR_SWIFT;
// Swift var __mode: DisplayMode { get set }
(20070465)
Xcode provides context-sensitive code completions for enum elements and option sets when using the shorter dot syntax. (16659653)
The NSManaged attribute can be used with methods as well as properties, for access to Core Data's automatically generated Key-Value-Coding-compliant to-many accessors.
@NSManaged var employees: NSSet @NSManaged func addEmployeesObject(employee: Employee) @NSManaged func removeEmployeesObject(employee: Employee) @NSManaged func addEmployees(employees: NSSet) @NSManaged func removeEmployees(employees: NSSet)
These can be declared in your NSManagedObject subclass. (17583057)
The grammar has been adjusted so that lines beginning with ‘.’ are always parsed as method or property lookups following the previous line, allowing for code formatted like this to work:
foo .bar .bas = 68000
It is no longer possible to begin a line with a contextual static member lookup (for example, to say .staticVar = MyType()). (20238557)
Code generation for large struct and enum types has been improved to reduce code size. (20598289)
Nonmutating methods of structs, enums, and protocols may now be partially applied to their self parameter:
let a: Set<Int> = [1, 2, 3] let b: [Set<Int>] = [[1], [4]] b.map(a.union) // => [[1, 2, 3], [1, 2, 3, 4]]
(21091944)
Swift documentation comments recognize a new top-level list item: - Throws: ...
This item is used to document what errors can be thrown and why. The documentation appears alongside parameters and return descriptions in Xcode QuickHelp. (21621679)
Unnamed parameters now require an explicit _: to indicate that they are unnamed. For example, the following is now an error:
func f(Int) { }
and must be written as:
func f(_: Int) { }
This simplifies the argument label model and also clarifies why cases like func f((a: Int, b: Int)) do not have parameters named a and b. (16737312)
It is now possible to append a tuple to an array. (17875634)
The ability to refer to the 0 element of a scalar value (producing the scalar value itself) has been removed. (17963034)
Variadic parameters can now appear anywhere in the parameter list for a function or initializer. For example:
func doSomethingToValues(values: Int... , options: MyOptions = [], fn: (Int) -> Void) { ... }
(20127197)
Generic subclasses of Objective-C classes are now supported. (18505295)
If an element of an enum with string raw type does not have an explicit raw value, it will default to the text of the enum's name. For example:
enum WorldLayer : String { case Ground, BelowCharacter, Character }
is equivalent to:
enum WorldLayer : String { case Ground = "Ground" case BelowCharacter = "BelowCharacter" case Character = "Character" }
(15819953)
The performSelector family of APIs is now available for Swift code. (17227475)
When delegating or chaining to a failable initializer (for example, with self.init(...) or super.init(...)), one can now force-unwrap the result with !. For example:
extension UIImage { enum AssetIdentifier: String { case Isabella case William case Olivia } convenience init(assetIdentifier: AssetIdentifier) { self.init(named: assetIdentifier.rawValue)! } }
(18497407)
Initializers can now be referenced like static methods by referring to .init on a static type reference or type object. For example:
let x = String.init(5) let stringType = String.self let y = stringType.init(5) let oneTwoThree = [1, 2, 3].map(String.init).reduce("", combine: +)
.init is still implicit when constructing using a static type, as in String(5). .init is required when using dynamic type objects or when referring to the initializer as a function value. (21375845)
Enums and cases can now be marked indirect, which causes the associated value for the enum to be stored indirectly, allowing for recursive data structures to be defined. For example:
enum List<T> { case Nil indirect case Cons(head: T, tail: List<T>) } indirect enum Tree<T> { case Leaf(T) case Branch(left: Tree<T>, right: Tree<T>) }
(21643855)
Formatting for Swift expression results has changed significantly when using po or expr -O. Customization that was introduced has been refined in the following ways:
The formatted summary provided by either debugDescription or description methods will always be used for types that conform to CustomDebugStringConvertible or CustomStringConvertible respectively. When neither conformance is present, the type name is displayed and reference types also display the referenced address to more closely mimic existing behavior for Objective-C classes.
Value types such as enums, tuples, and structs display all members indented below the summary by default, while reference types will not. This behavior can be customized for all types by implementing CustomReflectable.
These output customizations can be bypassed by using p or expr without the -O argument to provide a complete list of all fields and their values. (21463866)
Properties and methods using Unmanaged can now be exposed to Objective-C. (16832080)
Applying the @objc attribute to a class changes that class‘s compile-time name in the target’s generated Objective-C header as well as changing its runtime name. This applies to protocols as well. For example:
// Swift @objc(MyAppDelegate) class AppDelegate : NSObject, UIApplicationDelegate { // ... }
// Objective-C @interface MyAppDelegate : NSObject <UIApplicationDelegate> // ... @end
(17469485)
Collections containing types that are not Objective-C compatible are no longer considered Objective-C compatible types themselves.
For example, previously Array<SwiftClassType> was permitted as the type of a property marked @objc; this is no longer the case. (19787270)
Generic subclasses of Objective-C classes, as well as nongeneric classes that inherit from such a class, require runtime metadata instantiation and cannot be directly named from Objective-C code.
When support for generic subclasses of Objective-C classes was first added, the generated Objective-C bridging header erroneously listed such classes, which, when used, could lead to incorrect runtime behavior or compile-time errors. This has been fixed.
The behavior of the @objc attribute on a class has been clarified such that applying @objc to a class which cannot appear in a bridging header is now an error.
Note that this change does not result in a change of behavior with valid code because members of a class are implicitly @objc if any superclass of the class is an @objc class, and all @objc classes must inherit from NSObject. (21342574)
The performance of -Onone (debug) builds has been improved by using prespecialized instances of generics in the standard library. It produces significantly faster executables in debug builds in many cases, without impacting compile time. (20486658)
AnyObject and NSObject variables that refer to class objects can be cast back to class object types. For example, this code succeeds:
let x: AnyObject = NSObject.self let y = x as! NSObject.Type
Arrays, dictionaries, and sets that contain class objects successfully bridge with NSArray, NSDictionary, and NSSet as well. Objective-C APIs that provide NSArray<Class> * objects, such as -[NSURLSessionConfiguration protocolClasses], now work correctly when used in Swift. (16238475)
print() and reflection via Mirrors is able to report both the current case and payload for all enums with multiple payload types. The only remaining enum types that do not support reflection are @objc enums and enums imported from C. (21739870)
Enum cases with payloads can be used as functions. For example:
enum Either<T, U> { case Left(T), Right(U) } let lefts: [Either<Int, String>] = [1, 2, 3].map(Either.Left) let rights: [Either<Int, String>] = ["one", "two", "three"].map(Either.Right)
(19091028)
ExtensibleCollectionType has been folded into RangeReplaceableCollectionType. In addition, default implementations have been added as methods, which should be used instead of the free Swift module functions related to these protocols. (18220295)
The standard library moved many generic global functions (such as map, filter, and sort) to be methods written with protocol extensions. This allows those methods to be pervasively available on all sequence and collection types and allowed the removal of the global functions.
Deprecated enum elements no longer affect the names of nondeprecated elements when an Objective-C enum is imported into Swift. This may cause the Swift names of some enum elements to change. (17686122)
All enums imported from C are RawRepresentable, including those not declared with NS_ENUM or NS_OPTIONS. As part of this change, the value property of such enums has been renamed rawValue. (18702016)
Swift documentation comments use a syntax based on the Markdown format, aligning them with rich comments in playgrounds.
Outermost list items are interpreted as special fields and are highlighted in Xcode QuickHelp.
There are two methods of documenting parameters: parameter outlines and separate parameter fields. You can mix and match these forms as you see fit in any order or continuity throughout the doc comment. Because these are parsed as list items, you can nest arbitrary content underneath them.
Parameter outline syntax:
- Parameters: - x: ... - y: ...
Separate parameter fields:
- parameter x: ... - parameter y: ...
Documenting return values:
- returns: ...
Other special fields are highlighted in QuickHelp, as well as rendering support for all of Markdown. (20180161)
The CFunctionPointer<T -> U> type has been removed. C function types are specified using the new @convention(c) attribute. Like other function types, @convention(c) T -> U is not nullable unless made optional. The @objc_block attribute for specifying block types has also been removed and replaced with @convention(block).
Methods and functions have the same rules for parameter names. You can omit providing an external parameter name with _. To further simplify the model, the shorthand # for specifying a parameter name has been removed, as have the special rules for default arguments.
// Declaration func printFunction(str: String, newline: Bool) func printMethod(str: String, newline: Bool) func printFunctionOmitParameterName(str: String, _ newline: Bool) // Call printFunction("hello", newline: true) printMethod("hello", newline: true) printFunctionOmitParameterName("hello", true)
(17218256)
NS_OPTIONS types get imported as conforming to the OptionSetType protocol, which presents a set-like interface for options. Instead of using bitwise operations such as:
// Swift 1.2: object.invokeMethodWithOptions(.OptionA | .OptionB) object.invokeMethodWithOptions(nil) if options @ .OptionC == .OptionC { // .OptionC is set }
Option sets support set literal syntax, and set-like methods such as contains:
object.invokeMethodWithOptions([.OptionA, .OptionB]) object.invokeMethodWithOptions([]) if options.contains(.OptionC) { // .OptionC is set }
A new option set type can be written in Swift as a struct that conforms to the OptionSetType protocol. If the type specifies a rawValue property and option constants as static let constants, the standard library will provide default implementations of the rest of the option set API:
struct MyOptions: OptionSetType { let rawValue: Int static let TuringMachine = MyOptions(rawValue: 1) static let LambdaCalculus = MyOptions(rawValue: 2) static let VonNeumann = MyOptions(rawValue: 4) } let churchTuring: MyOptions = [.TuringMachine, .LambdaCalculus]
(18069205)
Type annotations are no longer allowed in patterns and are considered part of the outlying declaration. This means that code previously written as:
var (a : Int, b : Float) = foo()
needs to be written as:
var (a, b) : (Int, Float) = foo()
if an explicit type annotation is needed. The former syntax was ambiguous with tuple element labels. (20167393)
The do/while loop is renamed to repeat/while to make it obvious whether a statement is a loop from its leading keyword.
In Swift 1.2:
do { ... } while <condition>
In Swift 2.0:
repeat { ... } while <condition>
(20336424)
forEach has been added to SequenceType. This lets you iterate over elements of a sequence, calling a body closure on each. For example:
(0..<10).forEach { print($0) }
This is very similar to the following:
for x in 0..<10 { print(x) }
But take note of the following differences:
Unlike for-in loops, you can't use break or continue to exit the current call of the body closure or skip subsequent calls.
Also unlike for-in loops, using return in the body closure only exits from the current call to the closure, not any outer scope, and won't skip subsequent calls.
(18231840)
The Word and UWord types have been removed from the standard library; use Int and UInt instead. (18693488)
Most standard library APIs that take closures or @autoclosure parameters now use rethrows. This allows the closure parameters to methods like map and filter to throw errors, and allows short-circuiting operators like &&, ||, and ?? to work with expressions that may produce errors. (21345565)
SIMD improvements: Integer vector types in the simd module now only support unchecked arithmetic with wraparound semantics using the &+, &-, and &* operators, in order to better support the machine model for vectors. The +, -, and * operators are unavailable on integer vectors, and Xcode automatically suggests replacing them with the wrapping operators.
Code generation for vector types in the simd module has been improved to better utilize vector hardware, leading to dramatically improved performance in many cases. (21574425)
All CollectionType objects are now sliceable. SequenceType now has a notion of SubSequence, which is a type that represents only some of the values but in the same order. For example, the ArraySubSequence type is ArraySlice, which is an efficient view on the Array type's buffer that avoids copying as long as it uniquely references the Array from which it came.
The following free Swift functions for splitting/slicing sequences have been removed and replaced by method requirements on the SequenceType protocol with default implementations in protocol extensions. CollectionType has specialized implementations, where possible, to take advantage of efficient access of its elements.
/// Returns the first `maxLength` elements of `self`, /// or all the elements if `self` has fewer than `maxLength` elements. prefix(maxLength: Int) -> SubSequence /// Returns the last `maxLength` elements of `self`, /// or all the elements if `self` has fewer than `maxLength` elements. suffix(maxLength: Int) -> SubSequence /// Returns all but the first `n` elements of `self`. dropFirst(n: Int) -> SubSequence /// Returns all but the last `n` elements of `self`. dropLast(n: Int) -> SubSequence /// Returns the maximal `SubSequence`s of `self`, in order, that /// don't contain elements satisfying the predicate `isSeparator`. split(maxSplits maxSplits: Int, allowEmptySlices: Bool, @noescape isSeparator: (Generator.Element) -> Bool) -> [SubSequence]
The following convenience extension is provided for split:
split(separator: Generator.Element, maxSplit: Int, allowEmptySlices: Bool) -> [SubSequence]
Also, new protocol requirements and default implementations on CollectionType are now available:
/// Returns `self[startIndex..<end]` prefixUpTo(end: Index) -> SubSequence /// Returns `self[start..<endIndex]` suffixFrom(start: Index) -> SubSequence /// Returns `prefixUpTo(position.successor())` prefixThrough(position: Index) -> SubSequence
(21663830)
The print and debugPrint functions are improved:
Both functions have become variadic, and you can print any number of items with a single call.
separator: String = " " was added so you can control how the items are separated.
terminator: String = "\n" replaced appendNewline: bool = true. With this change, print(x, appendNewline: false) is expressed as print(x, terminator: "").
For the variants that take an output stream, the argument label toStream was added to the stream argument.
The println function from Swift 1.2 has been removed. (21788540)
For consistency and better composition of generic code, ArraySlice indices are no longer always zero-based but map directly onto the indices of the collection they are slicing and maintain that mapping even after mutations.
Before:
var a = Array(0..<10) var s = a[5..<10] s.indices // 0..<5 s[0] = 111 s // [111, 6, 7, 8, 9] s.removeFirst() s.indices // 1..<5
After:
var a = Array(0..<10) var s = a[5..<10] s.indices // 5..<10 s[5] = 99 s // [99, 6, 7, 8, 9] s.removeFirst() s.indices // 6..<10
Rather than define variants of collection algorithms that take explicit subrange arguments, such as a.sortSubrangeInPlace(3..<7), the Swift standard library provides “slicing,” which composes well with algorithms. This enables you to write a[3..<7].sortInPlace(), for example. With most collections, these algorithms compose naturally.
For example, before this change was incorporated:
extension MyIntCollection { func prefixThroughFirstNegativeSubrange() -> SubSequence { // Find the first negative element let firstNegative = self.indexOf { $0 < 0 } ?? endIndex // Slice off non-negative prefix let startsWithNegative = self.suffixFrom(firstNegative) // Find the first non-negative position in the slice let end = startsWithNegative.indexOf { $0 >= 0 } ?? endIndex return self[startIndex..<end] } }
The above code would work for any collection of Ints unless the collection is an Array<Int>. Unfortunately, when array slice indices are zero-based, the last two lines of the method need to change to:
let end = startsWithNegative.indexOf { $0 >= 0 } ?? startsWithNegative.endIndex return self[startIndex..<end + firstNegative]
These differences made working with slices awkward, error-prone, and nongeneric.
After this change, Swift collections start to provide a guarantee that, at least until there is a mutation, slice indices are valid in the collection from which they were sliced, and refer to the same elements. (21866825)
The method RangeReplaceableCollectionType.extend() was renamed to appendContentsOf(), and the splice() method was renamed to insertContentsOf(). (21972324)
find has been renamed to indexOf(), sort has been renamed to sortInPlace(), and sorted() becomes sort().
String.toInt() has been renamed to a failable Int(String) initializer, since initialization syntax is the preferred style for type conversions.
String no longer conforms to SequenceType in order to prevent non-Unicode correct sequence algorithms from being prominently available on String. To perform grapheme-cluster-based, UTF-8-based, or UTF-16-based algorithms, use the .characters, .utf8, and .utf16 projections respectively.
Generic functions that declare type parameters not used within the generic function's type produce a compiler error. For example:
func foo<T>() { } // error: generic parameter 'T' is not used in function signature
The Dictionary.removeAtIndex(_:) method now returns the key-value pair being removed as a two-element tuple (rather than returning Void). Similarly, the Set.removeAtIndex(_:) method returns the element being removed. (20299881)
Generic parameters on types in the Swift standard library have been renamed to reflect the role of the types in the API. For example, Array<T> became Array<Element>, UnsafePointer<T> became UnsafePointer<Memory>, and so forth. (21429126)
The SinkType protocol and SinkOf struct have been removed from the standard library in favor of (T) -> () closures. (21663799)
The notions of guaranteed conversion and “forced failable” conversion are now separated into two operators. Forced failable conversion now uses the as! operator. The ! makes it clear to readers of code that the cast may fail and produce a runtime error. The as operator remains for upcasts (e.g. someDerivedValue as Base) and type annotations (0 as Int8) which are guaranteed to never fail. (19031957)
Immutable (let) properties in struct and class initializers have been revised to standardize on a general “lets are singly initialized but never reassigned or mutated” model. Previously, they were completely mutable within the body of initializers. Now, they are only allowed to be assigned to once to provide their value. If the property has an initial value in its declaration, that counts as the initial value for all initializers. (19035287)
The implicit conversions from bridged Objective-C classes (NSString/NSArray/NSDictionary) to their corresponding Swift value types (String/Array/Dictionary) have been removed, making the Swift type system simpler and more predictable.
This means that the following code will no longer work:
import Foundation func log(s: String) { println(x) } let ns: NSString = "some NSString" // okay: literals still work log(ns) // fails with the error // "'NSString' is not convertible to 'String'"
In order to perform such a bridging conversion, make the conversion explicit with the as keyword:
log(ns as String) // succeeds
Implicit conversions from Swift value types to their bridged Objective-C classes are still permitted. For example:
func nsLog(ns: NSString) { println(ns) } let s: String = "some String" nsLog(s) // okay: implicit conversion from String to NSString is permitted
Note that these Cocoa types in Objective-C headers are still automatically bridged to their corresponding Swift type, which means that code is only affected if it is explicitly referencing (for example) NSString in a Swift source file. It is recommended you use the corresponding Swift types (for example, String) directly unless you are doing something advanced, like implementing a subclass in the class cluster. (18311362)
The @autoclosure attribute is now an attribute on a parameter, not an attribute on the parameter's type.
Where before you might have used:
func assert(predicate : @autoclosure () -> Bool) {...} you now write this as: func assert(@autoclosure predicate : () -> Bool) {...}
(15217242)
The @autoclosure attribute on parameters now implies the new @noescape attribute.
Curried function parameters can now specify argument labels.
For example:
func curryUnnamed(a: Int)(_ b: Int) { return a + b } curryUnnamed(1)(2) func curryNamed(first a: Int)(second b: Int) -> Int { return a + b } curryNamed(first: 1)(second: 2)
(17237268)
Swift now detects discrepancies between overloading and overriding in the Swift type system and the effective behavior seen via the Objective-C runtime.
For example, the following conflict between the Objective-C setter for property in a class and the method setProperty in its extension is now diagnosed:
class A : NSObject { var property: String = "Hello" // note: Objective-C method 'setProperty:' // previously declared by setter for // 'property' here } extension A { func setProperty(str: String) { } // error: method 'setProperty' // redeclares Objective-C method //'setProperty:' } Similar checking applies to accidental overrides in the Objective-C runtime: class B : NSObject { func method(arg: String) { } // note: overridden declaration // here has type '(String) -> ()' } class C : B { func method(arg: [String]) { } // error: overriding method with // selector 'method:' has incompatible // type '([String]) -> ()' } as well as protocol conformances: class MyDelegate : NSObject, NSURLSessionDelegate { func URLSession(session: NSURLSession, didBecomeInvalidWithError: Bool){ } // error: Objective-C method 'URLSession:didBecomeInvalidWithError:' // provided by method 'URLSession(_:didBecomeInvalidWithError:)' // conflicts with optional requirement method // 'URLSession(_:didBecomeInvalidWithError:)' in protocol // 'NSURLSessionDelegate' }
(18391046, 18383574)
The precedence of the Nil Coalescing Operator (??) has been raised to bind tighter than short-circuiting logical and comparison operators, but looser than as conversions and range operators. This provides more useful behavior for expressions like:
if allowEmpty || items?.count ?? 0 > 0 {...}
The &/ and &% operators were removed, to simplify the language and improve consistency.
Unlike the &+, &-, and &* operators, these operators did not provide two's-complement arithmetic behavior; they provided special case behavior for division, remainder by zero, and Int.min/-1. These tests should be written explicitly in the code as comparisons if needed. (17926954)
Constructing a UInt8 from an ASCII value now requires the ascii keyword parameter. Using non-ASCII unicode scalars will cause this initializer to trap. (18509195)
The C size_t family of types are now imported into Swift as Int, since Swift prefers sizes and counts to be represented as signed numbers, even if they are non-negative.
This change decreases the amount of explicit type conversion between Int and UInt, better aligns with sizeof returning Int, and provides safer arithmetic properties. (18949559)
Classes that do not inherit from NSObject but do adopt an @objc protocol will need to explicitly mark those methods, properties, and initializers used to satisfy the protocol requirements as @objc.
For example:
@objc protocol SomethingDelegate { func didSomething() } class MySomethingDelegate : SomethingDelegate { @objc func didSomething() { ... } }
Dynamic casts (as!, as? and is) now work with Swift protocol types, so long as they have no associated types. (18869156)
Adding conformances within a Playground now works as expected.
For example:
struct Point { var x, y: Double } extension Point : Printable { var description: String { return "(\(x), \(y))" } } var p1 = Point(x: 1.5, y: 2.5) println(p1) // prints "(1.5, 2.5)"
Imported NS_ENUM types with undocumented values, such as UIViewAnimationCurve, can now be converted from their raw integer values using the init(rawValue:) initializer without being reset to nil. Code that used unsafeBitCast as a workaround for this issue can be written to use the raw value initializer.
For example:
let animationCurve = unsafeBitCast(userInfo[UIKeyboardAnimationCurveUserInfoKey].integerValue, UIViewAnimationCurve.self) can now be written instead as: let animationCurve = UIViewAnimationCurve(rawValue: userInfo[UIKeyboardAnimationCurveUserInfoKey].integerValue)!
(19005771)
Negative floating-point literals are now accepted as raw values in enums. (16504472)
Unowned references to Objective-C objects, or Swift objects inheriting from Objective-C objects, no longer cause a crash if the object holding the unowned reference is deallocated after the referenced object has been released. (18091547)
Variables and properties with observing accessors no longer require an explicit type if it can be inferred from the initial value expression. (18148072)
Generic curried functions no longer produce random results when fully applied. (18988428)
Comparing the result of a failed NSClassFromString lookup against nil now behaves correctly. (19318533)
Subclasses that override base class methods with co- or contravariance in Optional types no longer cause crashes at runtime.
For example:
class Base { func foo(x: String) -> String? { return x } } class Derived: Base { override func foo(x: String?) -> String { return x! } }
(19321484)
Swift now supports building targets incrementally, i.e. not rebuilding every Swift source file in a target when a single file is changed.
The incremental build capability is based on a conservative dependency analysis, so you may still see more files rebuilding than absolutely necessary. If you find any cases where a file is not rebuilt when it should be, please file a bug report. Running Clean on your target afterwards should allow you to complete your build normally. (18248514)
A new Set data structure is included which provides a generic collection of unique elements with full value semantics. It bridges with NSSet, providing functionality analogous to Array and Dictionary. (14661754)
The if-let construct has been expanded to allow testing multiple optionals and guarding conditions in a single if (or while) statement using syntax similar to generic constraints:
if let a = foo(), b = bar() where a < b, let c = baz() { }
This allows you to test multiple optionals and include intervening boolean conditions, without introducing undesirable nesting (for instance, to avoid the optional unwrapping “pyramid of doom”).
Further, if-let now also supports a single leading boolean condition along with optional binding let clauses. For example:
if someValue > 42 && someOtherThing < 19, let a = getOptionalThing() where a > someValue { }
(19797158, 19382942)
The if-let syntax has been extended to support a single leading boolean condition along with optional binding let clauses.
For example:
if someValue > 42 && someOtherThing < 19, let a = getOptionalThing() where a > someValue { }
(19797158)
let constants have been generalized to no longer require immediate initialization. The new rule is that a let constant must be initialized before use (like a var), and that it may only be initialized: not reassigned or mutated after initialization. This enables patterns such as:
let x: SomeThing if condition { x = foo() } else { x = bar() } use(x)
which formerly required the use of a var, even though there is no mutation taking place. (16181314)
static methods and properties are now allowed in classes (as an alias for class final). You are now allowed to declare static stored properties in classes, which have global storage and are lazily initialized on first access (like global variables). Protocols now declare type requirements as static requirements instead of declaring them as class requirements. (17198298)
Type inference for single-expression closures has been improved in several ways:
Void return types can now be used in Void contexts.Swift enums can now be exported to Objective-C using the @objc attribute. @objc enums must declare an integer raw type, and cannot be generic or use associated values. Because Objective-C enums are not namespaced, enum cases are imported into Objective-C as the concatenation of the enum name and case name.
For example, this Swift declaration:
// Swift @objc enum Bear: Int { case Black, Grizzly, Polar }
imports into Objective-C as:
// Objective-C typedef NS_ENUM(NSInteger, Bear) { BearBlack, BearGrizzly, BearPolar };
(16967385)
Objective-C language extensions are now available to indicate the nullability of pointers and blocks in Objective-C APIs, allowing your Objective-C APIs to be imported without ImplicitlyUnwrappedOptional. (See items below for more details.) (18868820)
Swift can now partially import C aggregates containing unions, bitfields, SIMD vector types, and other C language features that are not natively supported in Swift. The unsupported fields will not be accessible from Swift, but C and Objective-C APIs that have arguments and return values of these types can be used in Swift. This includes the Foundation NSDecimal type and the GLKit GLKVector and GLKMatrix types, among others. (15951448)
Imported C structs now have a default initializer in Swift that initializes all of the struct's fields to zero.
For example:
import Darwin var devNullStat = stat() stat("/dev/null", &devNullStat)
If a structure contains fields that cannot be correctly zero initialized (i.e. pointer fields marked with the new __nonnull modifier), this default initializer will be suppressed. (18338802)
New APIs for converting among the Index types for String, String.UnicodeScalarView, String.UTF16View, and String.UTF8View are available, as well as APIs for converting each of the String views into Strings. (18018911)
Type values now print as the full demangled type name when used with println or string interpolation.
toString(Int.self) // prints "Swift.Int" println([Float].self) // prints "Swift.Array<Swift.Float>" println((Int, String).self) // prints "(Swift.Int, Swift.String)"
(18947381)
A new @noescape attribute may be used on closure parameters to functions. This indicates that the parameter is only ever called (or passed as an @noescape parameter in a call), which means that it cannot outlive the lifetime of the call. This enables some minor performance optimizations, but more importantly disables the self. requirement in closure arguments. This enables control-flow-like functions to be more transparent about their behavior. In a future beta, the standard library will adopt this attribute in functions like autoreleasepool().
func autoreleasepool(@noescape code: () -> ()) { pushAutoreleasePool() code() popAutoreleasePool() }
(16323038)
Performance is substantially improved over Swift 1.1 in many cases. For example, multidimensional arrays are algorithmically faster in some cases, unoptimized code is much faster in many cases, and many other improvements have been made.
The diagnostics emitted for expression type check errors are greatly improved in many cases. (18869019)
Type checker performance for many common expression kinds has been greatly improved. This can significantly improve build times and reduces the number of “expression too complex” errors. (18868985)
The @autoclosure attribute has a second form, @autoclosure(escaping), that provides the same caller-side syntax as @autoclosure but allows the resulting closure to escape in the implementation.
For example:
func lazyAssertion(@autoclosure(escaping) condition: () -> Bool, message: String = "") { lazyAssertions.append(condition) // escapes } lazyAssertion(1 == 2, message: "fail eventually")
(19499207)
Whole Module Optimization build setting or by using the swiftc command line tool with the flag -whole-module-optimization. (18603795)flatMap was added to the standard library. flatMap is the function that maps a function over something and returns the result flattened one level. flatMap has many uses, such as to flatten an array:
[[1,2],[3,4]].flatMap { $0 }
or to chain optionals with functions:
[[1,2], [3,4]].first.flatMap { find($0, 1) }
(19881534)
The function zip was added. It joins two sequences together into one sequence of tuples. (17292393)
utf16Count is removed from String. Instead use count on the UTF16 view of the String.
For example:
count(string.utf16)
(17627758)
Class methods and initializers that satisfy protocol requirements now properly invoke subclass overrides when called in generic contexts. For example:
protocol P { class func foo() } class C: P { class func foo() { println("C!") } } class D: C { override class func foo() { println("D!") } } func foo<T: P>(x: T) { x.dynamicType.foo() } foo(C()) // Prints "C!" foo(D()) // Used to incorrectly print "C!", now prints "D!"
(18828217)
Values of type Any can now contain values of function type. (16406907)
Documentation for the standard library (displayed in quick help and in the synthesized header for the Swift module) is improved. (16462500)
Class properties don't need to be marked final to avoid O(n) mutations on value semantic types. (17416120)
Casts can now be performed between CF types (such as CFString, CGImage, and SecIdentity) and AnyObject. Such casts will always succeed at run-time. For example:
var cfStr: CFString = ... var obj: AnyObject = cfStr as AnyObject var cfStr = obj as CFString
(18088474)
HeapBuffer<Value, Element>, HeapBufferStorage<Value, Element>, and OnHeap<Value> were never really useful, because their APIs were insufficiently public. They have been replaced with a single class, ManagedBuffer<Value, Element>. See also the new function isUniquelyReferenced(x) which is often useful in conjunction with ManagedBuffer.
The Character enum has been turned into a struct, to avoid exposing its internal implementation details.
The countElements function has been renamed count, for better consistency with our naming conventions.
Mixed-sign addition and subtraction operations, that were unintentionally allowed in previous versions, now cause a compilation error.
OS X apps can now apply the @NSApplicationMain attribute to their app delegate class in order to generate an implicit main for the app. This works like the @UIApplicationMain attribute for iOS apps.
Objective-C init and factory methods are now imported as failable initializers when they can return nil. In the absence of information about a potentially-nil result, an Objective-C init or factory method will be imported as init!.
As part of this change, factory methods that have NSError** parameters, such as +[NSString stringWithContentsOfFile:encoding:error:], will now be imported as (failable) initializers, e.g.,
init?(contentsOfFile path: String, encoding: NSStringEncoding, error: NSErrorPointer)
Nested classes explicitly marked @objc will now properly be included in a target's generated header as long as the containing context is also (implicitly or explicitly) @objc. Nested classes not explicitly marked @objc will never be printed in the generated header, even if they extend an Objective-C class.
All of the *LiteralConvertible protocols, as well as StringInterpolationConvertible, now use initializers for their requirements rather than static methods starting with convertFrom. For example, IntegerLiteralConvertible now has the following initializer requirement:
init(integerLiteral value: IntegerLiteralType)
Any type that previously conformed to one of these protocols will need to replace its convertFromXXX static methods with the corresponding initializer.
Initializers can now fail by returning nil. A failable initializer is declared with init? (to return an explicit optional) or init! (to return an implicitly-unwrapped optional). For example, you could implement String.toInt as a failable initializer of Int like this:
extension Int { init?(fromString: String) { if let i = fromString.toInt() { // Initialize self = i } else { // Discard self and return 'nil'. return nil } } }
The result of constructing a value using a failable initializer then becomes optional:
if let twentytwo = Int(fromString: "22") { println("the number is \(twentytwo)") } else { println("not a number") }
In the current implementation, struct and enum initializers can return nil at any point inside the initializer, but class initializers can only return nil after all of the stored properties of the object have been initialized and self.init or super.init has been called. If self.init or super.init is used to delegate to a failable initializer, then the nil return is implicitly propagated through the current initializer if the called initializer fails.
The RawRepresentable protocol that enums with raw types implicitly conform to has been redefined to take advantage of failable initializers. The fromRaw(RawValue) static method has been replaced with an initializer init?(rawValue: RawValue), and the toRaw() method has been replaced with a rawValue property. Enums with raw types can now be used like this:
enum Foo: Int { case A = 0, B = 1, C = 2 } let foo = Foo(rawValue: 2)! // formerly 'Foo.fromRaw(2)!' println(foo.rawValue) // formerly 'foo.toRaw()'
+. Use String(c1) + String(c2) instead.@objc protocol types using a as [C], type checking is now deferred until the moment each element is accessed. Because bridging conversions from NSArray are equivalent to force-casts from [NSArray], this makes certain Array round-trips through Objective-C code O(1) instead of O(N).RawOptionSetType now implements BitwiseOperationsType, so imported NS_OPTIONS now support the bitwise assignment operators |=, &=, and ^=. It also no longer implements BooleanType; to check if an option set is empty, compare it to nil.
Types implementing BitwiseOperationsType now automatically support the bitwise assignment operators |=, &=, and ^=.
Optionals can now be coalesced with default values using the ?? operator. ?? is a short-circuiting operator that takes an optional on the left and a non-optional expression on the right. If the optional has a value, its value is returned as a non-optional; otherwise, the expression on the right is evaluated and returned:
var sequence: [Int] = [] sequence.first ?? 0 // produces 0, because sequence.first is nil sequence.append(22) sequence.first ?? 0 // produces 22, the value of sequence.first
The optional chaining ? operator can now be mutated through, like !. The assignment and the evaluation of the right-hand side of the operator are conditional on the presence of the optional value:
var sequences = ["fibonacci": [1, 1, 2, 3, 4], "perfect": [6, 28, 496]] sequences["fibonacci"]?[4]++ // Increments element 4 of key "fibonacci" sequences["perfect"]?.append(8128) // Appends to key "perfect" sequences["cubes"]?[3] = 3*3*3 // Does nothing; no "cubes" key
Note that optional chaining still flows to the right, so prefix increment operators are not included in the chain, so this won't type-check:
++sequences["fibonacci"]?[4] // Won't type check, can't '++' Int?
The swift command line interface is now divided into an interactive driver swift, and a batch compiler swiftc:
swift [options] input-file [program-arguments] Runs the script 'input-file' immediately, passing any program-arguments to the script. Without any input files, invokes the repl. swiftc [options] input-filenames The familiar swift compiler interface: compiles the input-files according to the mode options like -emit-object, -emit-executable, etc.
For greater clarity and explicitness when bypassing the type system, reinterpretCast has been renamed unsafeBitCast, and it has acquired a (required) explicit type parameter. So
let x: T = reinterpretCast(y)
becomes
let x = unsafeBitCast(y, T.self)
Because their semantics were unclear, the methods asUnsigned (on the signed integer types) and asSigned (on the unsigned integer types) have been replaced. The new idiom is explicit construction of the target type using the bitPattern: argument label. So,
myInt.asUnsigned()
has become
UInt(bitPattern: myInt)
To better follow Cocoa naming conventions and to encourage immutability, The following pointer types were renamed:
| Old Name | New Name |
|---|---|
UnsafePointer<T> | UnsafeMutablePointer<T> |
ConstUnsafePointer<T> | UnsafePointer<T> |
AutoreleasingUnsafePointer<T> | AutoreleasingUnsafeMutablePointer<T> |
Note that the meaning of UnsafePointer has changed from mutable to immutable. As a result, some of your code may fail to compile when assigning to an UnsafePointer.memory property. The fix is to change your UnsafePointer<T> into an UnsafeMutablePointer<T>.
The optional unwrapping operator x! can now be assigned through, and mutating methods and operators can be applied through it:
var x: Int! = 0 x! = 2 x!++ // Nested dictionaries can now be mutated directly: var sequences = ["fibonacci": [1, 1, 2, 3, 0]] sequences["fibonacci"]![4] = 5 sequences["fibonacci"]!.append(8)
The @auto_closure attribute has been renamed to @autoclosure.
There is a new dynamic declaration modifier. When applied to a method, property, subscript, or initializer, it guarantees that references to the declaration are always dynamically dispatched and never inlined or devirtualized, and that the method binding can be reliably changed at runtime. The implementation currently relies on the Objective-C runtime, so dynamic can only be applied to @objc-compatible declarations for now. @objc now only makes a declaration visible to Objective-C; the compiler may now use vtable lookup or direct access to access (non-dynamic) @objc declarations.
class Foo { // Always accessed by objc_msgSend dynamic var x: Int // Accessed by objc_msgSend from ObjC; may be accessed by vtable // or by static reference in Swift @objc var y: Int // Not exposed to ObjC (unless Foo inherits NSObject) var z: Int }
dynamic enables KVO, proxying, and other advanced Cocoa features to work reliably with Swift declarations.
Clang submodules can now be imported:
import UIKit.UIGestureRecognizerSubclass
The numeric optimization levels -O[0-3] have been removed in favor of the named levels -Onone and -O.
The -Ofast optimization flag has been renamed to -Ounchecked. We will accept both names for now and remove -Ofast in a later build.
An initializer that overrides a designated initializer from its superclass must be marked with the override keyword, so that all overrides in the language consistently require the use of override. For example:
class A { init() { } } class B : A { override init() { super.init() } }
Required initializers are now more prominent in several ways. First, a (non-final) class that conforms to a protocol that contains an initializer requirement must provide a required initializer to satisfy that requirement. This ensures that subclasses will also conform to the protocol, and will be most visible with classes that conform to NSCoding:
class MyClass : NSObject, NSCoding { required init(coder aDecoder: NSCoder!) { /*... */ } func encodeWithCoder(aCoder: NSCoder!) { /* ... */ } }
Second, because required places a significant requirement on all subclasses, the required keyword must be placed on overrides of a required initializer:
class MySubClass : MyClass { var title: String = "Untitled" required init(coder aDecoder: NSCoder!) { /*... */ } override func encodeWithCoder(aCoder: NSCoder!) { /* ... */ } }
Finally, required initializers can now be inherited like any other initializer:
class MySimpleSubClass : MyClass { } // inherits the required init(coder:).
Access control has been implemented.
public declarations can be accessed from any module.internal declarations (the default) can be accessed from within the current module.private declarations can be accessed only from within the current file.There are still details to iron out here, but the model is in place. The general principle is that an entity cannot be defined in terms of another entity with less accessibility.
Along with this, the generated header for a framework will only include public declarations. Generated headers for applications will include public and internal declarations.
CGFloat is now a distinct floating-point type that wraps either a Float (on 32-bit architectures) or a Double (on 64-bit architectures). It provides all of the same comparison and arithmetic operations of Float and Double, and can be created using numeric literals.
The immediate mode swift -i now works for writing #! scripts that take command line arguments. The -i option to the swift driver must now come at the end of the compiler arguments, directly before the input filename. Any arguments that come after -i and the input filename are treated as arguments to the interpreted file and forwarded to Process.arguments.
Type inference for for..in loops has been improved to consider the sequence along with the element pattern. For example, this accepts the following loops that were previously rejected:
for i: Int8 in 0..<10 { } for i: Float in 0.0...10.0 { }
Introduced the new BooleanLiteralConvertible protocol, which allows user-defined types to support Boolean literals. true and false are now Boolean constants and keywords.
The @final, @lazy, @required and @optional attributes are now considered to be declaration modifiers - they no longer require (or allow) an @ sign.
The @prefix, @infix, and @postfix attributes have been changed to declaration modifiers, so they are no longer spelled with an @ sign. Operator declarations have been rearranged from operator prefix + to prefix operator + for consistency.
C function pointer types are now imported as CFunctionPointer<T>, where T is a Swift function type. CFunctionPointer and COpaquePointer can be explicitly constructed from one another, but they do not freely convert, nor is CFunctionPointer compatible with Swift closures.
Example: int (*)(void) becomes CFunctionPointer<(Int) -> Void>.
The interop model for pointers in C APIs has been simplified. Most code that calls C functions by passing arrays, UnsafePointers, or the addresses of variables with &x does not need to change. However, the CConstPointer and CMutablePointer bridging types have been removed, and functions and methods are now imported as and overridden by taking UnsafePointer and ConstUnsafePointer directly. Void pointers are now imported as (Const)UnsafePointer<Void>; COpaquePointer is only imported for opaque types now.
Array types are now spelled with the brackets surrounding the element type. For example, an array of Int is written as:
var array: [Int]
Dictionary types can now be spelled with the syntax [K : V], where K is the key type and V is the value type. For example:
var dict: [String : Int] = ["Hello" : 1, "World" : 2]
The type [K : V] is syntactic sugar for Dictionary<K, V>; nothing else has changed.
The @IBOutlet attribute no longer implicitly (and invisibly) changes the type of the declaration it is attached to. It no longer implicitly makes variables be an implicitly unwrapped optional and no longer defaults them to weak.
The \x, \u and \U escape sequences in string literals have been consolidated into a single and less error prone \u{123456} syntax.
The half-open range operator has been renamed from .. to ..< to reduce confusion. The ..< operator is precedented in Groovy (among other languages) and makes it much more clear that it doesn't include the endpoint.
Class objects such as NSObject.self can now be converted to AnyObject and used as object values.
Objective-C protocol objects such as NSCopying.self can now be used as instances of the Protocol class, such as in APIs such as XPC.
Arrays now have full value semantics: both assignment and initialization create a logically-distinct object
The sort function and array method modify the target in-place. A new sorted function and array method are non-mutating, creating and returning a new collection.
sort, map, filter, and reduce methods on Arrays accept trailing closures:
let a = [5, 6, 1, 3, 9] a.sort{ $0 > $1 } println(a) // [9, 6, 5, 3, 1] println(a.map{ $0 * 2 }) // [18, 12, 10, 6, 2] println(a.map{ $0 * 2 }.filter{ $0 < 10}) // [6, 2] println(a.reduce(1000){ $0 + $1 }) // 1024 (no kidding)
A lazy map() function in the standard library works on any Sequence. Example:
class X { var value: Int init(_ value: Int) { self.value = value println("created X(\(value))") } } // logically, this sequence is X(0), X(1), X(2), ... X(50) let lazyXs = map(0..50){ X($0) } // Prints "created X(...)" 4 times for x in lazyXs { if x.value == 4 { break } }
There's a similar lazy filter() function:
// 0, 10, 20, 30, 40 let tens = filter(0..50) { $0 % 10 == 0 } let tenX = map(tens){ X($0) } // 5 lazy Xs let tenXarray = Array(tenX) // Actually creates those Xs
Weak pointers of classbound protocol type work now.
IBOutlets now default to weak pointers with implicit optional type (T!).
NSArray* parameters and result types of Objective-C APIs are now imported as AnyObject[]!, i.e., an implicitly unwrapped optional array storing AnyObject values. For example, NSView's constraints property
@property (readonly) NSArray *constraints;
is now imported as
var constraints: AnyObject[]!
Note that one can implicitly convert between an AnyObject[] and an NSArray (in both directions), so (for example) one can still explicitly use NSArray if desired:
var array: NSArray = view.constraints
Swift arrays bridge to NSArray similarly to the way Swift strings bridge to NSString.
ObjCMutablePointer has been renamed AutoreleasingUnsafePointer.
UnsafePointer (and AutoreleasingUnsafePointer)'s set() and get() have been replaced with a property called memory.
Previously you would write:
val = p.get() p.set(val)
Now you write:
val = p.memory p.memory = val
Removed shorthand x as T!; instead use (x as T)!
x as T! now means “x as implicitly unwrapped optional”.Range operators .. and ... have been switched.
1..3 now means 1,21...3 now means 1,2,3The pound sign (#) is now used instead of the back-tick (`) to mark an argument name as a keyword argument, e.g.,
func moveTo(#x: Int, #y: Int) { ... } moveTo(x: 5, y: 7)
Objective-C factory methods are now imported as initializers. For example, NSColor's +colorWithRed:green:blue:alpha becomes
init(red: CGFloat, green: CGFloat, blue: CGFloat, alpha: CGFloat)
which allows an NSColor to be created as, e.g.,
NSColor(red: 0.5, green: 0.25, blue: 0.25, alpha: 0.5)
Factory methods are identified by their kind (class methods), name (starts with words that match the words that end the class name), and result type (instancetype or the class type).
Objective-C properties of some CF type are no longer imported as Unmanaged.
REPL mode now uses LLDB, for a greatly-expanded set of features. The colon prefix now treats the rest of the line as a command for LLDB, and entering a single colon will drop you into the debugging command prompt. Most importantly, crashes in the REPL will now drop you into debugging mode to see what went wrong.
If you do have a need for the previous REPL, pass -integrated-repl.
In a UIKit-based application, you can now eliminate your ‘main.swift’ file and instead apply the @UIApplicationMain attribute to your UIApplicationDelegate class. This will cause the main entry point to the application to be automatically generated as follows:
UIApplicationMain(argc, argv, nil, NSStringFromClass(YourApplicationDelegate.self))
If you need nontrivial logic in your application entry point, you can still write out a main.swift. Note that @UIApplicationMain and main.swift are mutually exclusive.
weak pointers now work with implicitly unchecked optionals, enabling usecases where you don't want to ! every use of a weak pointer. For example:
weak var myView : NSView!
of course, they still work with explicitly checked optionals like NSView?
Dictionary subscripting now takes/returns an optional type. This allows querying a dictionary via subscripting to gracefully fail. It also enables the idiom of removing values from a dictionary using dict[key] = nil. As part of this, deleteKey is no longer available.
Stored properties may now be marked with the @lazy attribute, which causes their initializer to be evaluated the first time the property is touched instead of when the enclosing type is initialized. For example:
func myInitializer() -> Int { println("hello\n"); return 42 } class MyClass { @lazy var aProperty = myInitializer() } var c = MyClass() // doesn't print hello var tmp = c.aProperty // prints hello on first access tmp = c.aProperty // doesn't print on subsequent loads. c = MyClass() // doesn't print hello c.aProperty = 57 // overwriting the value prevents it from ever running
Because lazy properties inherently rely on mutation of the property, they cannot be lets. They are currently also limited to being members of structs and classes (they aren't allowed as local or global variables yet) and cannot be observed with willSet/didSet yet.
Closures can now specify a capture list to indicate with what strength they want to capture a value, and to bind a particular field value if they want to.
Closure capture lists are square-bracket delimited and specified before the (optional) argument list in a closure. Each entry may be specified as weak or unowned to capture the value with a weak or unowned pointer, and may contain an explicit expression if desired. Some examples:
takeClosure { print(self.title) } // strong capture takeClosure { [weak self] in print(self!.title) } // weak capture takeClosure { [unowned self] in print(self.title) } // unowned capture
You can also bind arbitrary expression to named values in the capture list. The expression is evaluated when the closure is formed, and captured with the specified strength. For example:
// weak capture of "self.parent" takeClosure { [weak tmp = self.parent] in print(tmp!.title) }
The full form of a closure can take a signature (an argument list and optionally a return type) if needed. To use either the capture list or the signature, you must specify the context sensitive in keyword. Here is a (weird because there is no need for unowned) example of a closure with both:
myNSSet.enumerateObjectsUsingBlock { [unowned self] (obj, stop) in self.considerWorkingWith(obj) }
The word with is now removed from the first keyword argument name if an initialized imported from Objective-C. For example, instead of building UIColor as:
UIColor(withRed: r, green: g, blue: b, alpha: a)
it will now be:
UIColor(red: r, green: g, blue: b, alpha: a)
Dictionary can be bridged to NSDictionary and vice versa:
NSDictionary has an implicit conversion to Dictionary<NSObject, AnyObject>. It bridges in O(1), without memory allocation.
Dictionary<K, V> has an implicit conversion to NSDictionary. Dictionary<K, V> bridges to NSDictionary iff both K and V are bridged. Otherwise, a runtime error is raised.
Depending on K and V the operation can be O(1) without memory allocation, or O(N) with memory allocation.
Single-quoted literals are no longer recognized. Use double-quoted literals and an explicit type annotation to define Characters and UnicodeScalars:
var ch: Character = "a" var us: UnicodeScalar = "a"
The use of keyword arguments is now strictly enforced at the call site. For example, consider this method along with a call to it:
class MyColor { func mixColorWithRed(red: Float, green: Float, blue: Float) { /* ... */ } } func mix(color: MyColor, r: Float, g: Float, b: Float) { color.mixColorWithRed(r, g, b) }
The compiler will now complain about the missing green: and blue: labels, with a Fix-It to correct the code:
color.swift:6:24: error: missing argument labels 'green:blue:' in call
color.mixColorWithRed(r, g, b)
^
green: blue:
The compiler handles missing, extraneous, and incorrectly-typed argument labels in the same manner. Recall that one can make a parameter a keyword argument with the back-tick or remove a keyword argument with the underscore.
class MyColor { func mixColor(`red: Float, green: Float, blue: Float) { /* ... */ } func mixColorGuess(red: Float, _ green: Float, _ blue: Float) { /* ... */ } } func mix(color: MyColor, r: Float, g: Float, b: Float) { color.mixColor(red: r, green: g, blue: b) // okay: all keyword arguments color.mixColorGuess(r, g, b) // okay: no keyword arguments }
Arguments cannot be re-ordered unless the corresponding parameters have default arguments. For example, given:
func printNumber(`number: Int, radix: Int = 10, separator: String = ",") { }
The following three calls are acceptable because only the arguments for defaulted parameters are re-ordered relative to each other:
printNumber(number: 256, radix: 16, separator: "_") printNumber(number: 256, separator: "_") printNumber(number: 256, separator: ",", radix: 16)
However, this call:
printNumber(separator: ",", radix: 16, number: 256)
results in an error due to the re-ordering:
printnum.swift:7:40: error: argument 'number' must precede argument 'separator'
printNumber(separator: ",", radix: 16, number: 256)
~~~~~~~~~~~~~~ ^ ~~~
; can no longer be used to demarcate an empty case in a switch statement, use break instead.
The compiler's ability to diagnose many common kinds of type check errors has improved. (expression does not type-check has been retired.)
Ranges can be formed with floating point numbers, e.g. 0.0 .. 100.0.
Convenience initializers are now spelled as convenience init instead of with the -> Self syntax. For example:
class Foo { init(x : Int) {} // designated initializer convenience init() { self.init(42) } // convenience initializer }
You still cannot declare designated initializers in extensions, only convenience initializers are allowed.
Reference types using the CoreFoundation runtime are now imported as class types. This means that Swift will automatically manage the lifetime of a CFStringRef the same way that it manages the lifetime of an NSString.
In many common cases, this will just work. Unfortunately, values are returned from CF-style APIs in a wide variety of ways, and unlike Objective-C methods, there simply isn‘t enough consistency for Swift to be able to safely apply the documented conventions universally. The framework teams have already audited many of the most important CF-style APIs, and those APIs should be imported without a hitch into Swift. For all the APIs which haven’t yet been audited, we must import return types using the Unmanaged type. This type allows the programmer to control exactly how the object is passed.
For example:
// CFBundleGetAllBundles() returns an Unmanaged<CFArrayRef>. // From the documentation, we know that it returns a +0 value. let bundles = CFBundleGetAllBundles().takeUnretainedValue() // CFRunLoopCopyAllModes() returns an Unmanaged<CFArrayRef>. // From the documentation, we know that it returns a +1 value. let modes = CFRunLoopCopyAllModes(CFRunLoopGetMain()).takeRetainedValue()
You can also use Unmanaged types to pass and return objects indirectly, as well as to generate unbalanced retains and releases if you really require them.
The API of the Unmanaged type is still in flux, and your feedback would be greatly appreciated.
The @NSManaged attribute can be applied to the properties of an NSManagedObject subclass to indicate that they should be handled by CoreData:
class Employee : NSManagedObject { @NSManaged var name: String @NSManaged var department: Department }
The @weak and @unowned attributes have become context sensitive keywords instead of attributes. To declare a weak or unowned pointer, use:
weak var someOtherWindow : NSWindow? unowned var someWindow : NSWindow
... with no @ on the weak/unowned.
Swift now supports a #elseif form for build configurations, e.g.:
#if os(OSX) typealias SKColor = NSColor #elseif os(iOS) typealias SKColor = UIColor #else typealias SKColor = Green #endif
You can now use the true and false constants in build configurations, allowing you to emulate the C idioms of #if 0 (but spelled #if false).
break now breaks out of switch statements.
It is no longer possible to specify @mutating as an attribute, you may only use it as a keyword, e.g.:
struct Pair { var x, y : Int mutating func nuke() { x = 0; y = 0 } }
The former @!mutating syntax used to mark setters as non-mutating is now spelled with the nonmutating keyword. Both mutating and nonmutating are context sensitive keywords.
NSLog is now available from Swift code.
The parser now correctly handles expressions like var x = Int[]() to create an empty array of integers. Previously you'd have to use syntax like Array<Int>() to get this. Now that this is all working, please prefer to use Int[] consistently instead of Array<Int>.
Character is the new character literal type:
var x = 'a' // Infers 'Character' type
You can force inference of UnicodeScalar like this:
var scalar: UnicodeScalar = 'a'
Character type represents a Unicode extended grapheme cluster (to put it simply, a grapheme cluster is what users think of as a character: a base plus any combining marks, or other cases explained in Unicode Standard Annex #29).
Loops and switch statements can now carry labels, and you can break/continue to those labels. These use conventional C-style label syntax, and should be dedented relative to the code they are in. An example:
func breakContinue(x : Int) -> Int { Outer: for a in 0..1000 { Switch: switch x { case 42: break Outer case 97: continue Outer case 102: break Switch case 13: continue // continue always works on loops. case 139: break // break will break out of the switch (but see below) } } }
We are changing the behavior of break to provide C-style semantics, to allow breaking out of a switch statement. Previously, break completely ignored switches so that it would break out of the nearest loop. In the example above, case 139 would break out of the Outer loop, not the Switch.
In order to avoid breaking existing code, we're making this a compile time error instead of a silent behavior change. If you need a solution for the previous behavior, use labeled break.
This error will be removed in a week or two.
Cocoa methods and properties that are annotated with the NS_RETURNS_INNER_POINTER attribute, including -[NSData bytes] and -[{NS,UI}Color CGColor], are now safe to use and follow the same lifetime extension semantics as ARC.
Enabling/disabling of asserts
assert(condition, msg)
is enabled/disabled dependent on the optimization level. In debug mode at -O0 asserts are enabled. At higher optimization levels asserts are disabled and no code is generated for them. However, asserts are always type checked even at higher optimization levels.
Alternatively, assertions can be disabled/enabled by using the frontend flag -assert-config Debug, or -assert-config Release.
Added optimization flag -Ofast. It disables all assertions (assert), and runtime overflow and type checks.
The “selector-style” function and initializer declaration syntax is being phased out. For example, this:
init withRed(red: CGFloat) green(CGFloat) blue(CGFloat) alpha(CGFloat)
will now be written as:
init(withRed red: CGFloat, green: CGFloat, blue: CGFloat, alpha: CGFloat)
For each parameter, one can have both an argument API name (i.e., withRed, which comes first and is used at the call site) and an internal parameter name that follows it (i.e. red, which comes second and is used in the implementation). When the two names are the same, one can simply write the name once and it will be used for both roles (as with green, blue, and alpha above). The underscore (_) can be used to mean “no name”, as when the following function/method:
func murderInRoom(room:String) withWeapon(weapon: String)
is translated to:
func murderInRoom(_ room: String, withWeapon weapon: String)
The compiler now complains when it sees the selector-style syntax and will provide Fix-Its to rewrite to the newer syntax.
Note that the final form of selector syntax is still being hammered out, but only having one declaration syntax, which will be very close to this, is a known.
Stored properties can now be marked with the @NSCopying attribute, which causes their setter to be synthesized with a copy to copyWithZone:. This may only be used with types that conform to the NSCopying protocol, or option types thereof. For example:
@NSCopying var myURL : NSURL
This fills the same niche as the (copy) attribute on Objective-C properties.
Optional variables and properties are now default-initialized to nil:
class MyClass { var cachedTitle: String? // "= nil" is implied }
@IBOutlet has been improved in a few ways:
IBOutlets can now be @unchecked optional.
An IBOutlet declared as non-optional, i.e.,
@IBOutlet var button: NSButton
will be treated as an @unchecked optional. This is considered to be the best practice way to write an outlet, unless you want to explicitly handle the null case - in which case, use NSButton? as the type. Either way, the = nil that was formerly required is now implicit.
The precedence of is and as is now higher than comparisons, allowing the following sorts of things to be written without parens:
if x is NSButton && y is NSButtonCell { ... } if 3/4 as Float == 6/8 as Float { ... }
Objective-C blocks are now transparently bridged to Swift closures. You never have to write @objc_block when writing Objective-C-compatible methods anymore. Block parameters are now imported as unchecked optional closure types, allowing nil to be passed.
Dictionary changes:
Elements are now tuples, so you can write
for (k, v) in d { // ... }
keys and values properties, which are Collections projecting the corresponding aspect of each element. Dictionary indices are usable with their keys and values properties, so:
for i in indices(d) { let (k, v) = d[i] assert(k == d.keys[i]) assert(v == d.values[i]) }
Semicolon can be used as a single no-op statement in otherwise empty cases in switch statements:
switch x { case 1, 2, 3: print("x is 1, 2 or 3") default: ; }
override is now a context sensitive keyword, instead of an attribute:
class Base { var property: Int { return 0 } func instanceFunc() {} class func classFunc() {} } class Derived : Base { override var property: Int { return 1 } override func instanceFunc() {} override class func classFunc() {} }
Prefix splitting for imported enums has been revised again due to feedback:
10_4), leave one more word in the result than would otherwise be there (Behavior10_4).k prefix (for constant) and the enum doesn't, the k should not be considered when finding the common prefix.NSSomethingOptions) and the enumerator names use the singular form (NSSomethingOptionMagic), this is considered a matching prefix (but only if nothing follows the plural).Cocoa APIs that take pointers to plain C types as arguments now get imported as taking the new CMutablePointer<T> and CConstPointer<T> types instead of UnsafePointer<T>. These new types allow implicit conversions from Swift inout parameters and from Swift arrays:
let rgb = CGColorSpaceCreateDeviceRGB() // CGColorRef CGColorCreate(CGColorSpaceRef, const CGFloat*); let white = CGColorCreate(rgb, [1.0, 1.0, 1.0]) var s = 0.0, c = 0.0 // void sincos(double, double*, double*); sincos(M_PI/2, &s, &c)
Pointers to pointers to ObjC classes, such as NSError**, get imported as ObjCMutablePointer<NSError?>. This type doesn't work with arrays, but accepts inouts or nil:
var error: NSError? let words = NSString.stringWithContentsOfFile("/usr/share/dict/words", encoding: .UTF8StringEncoding, error: &error)
Void pointer parameters can be passed an array or inout of any type:
// + (NSData*)dataWithBytes:(const void*)bytes length:(NSUInteger)length; let data = NSData.dataWithBytes([1.5, 2.25, 3.125], length: sizeof(Double.self) * 3) var fromData = [0.0, 0.0, 0.0] // - (void)getBytes:(void*)bytes length:(NSUInteger)length; data.getBytes(&fromData, length: sizeof(Double.self) * 3)
Note that we don't know whether an API reads or writes the C pointer, so you need to explicitly initialize values (like s and c above) even if you know that the API overwrites them.
This pointer bridging only applies to arguments, and only works with well- behaved C and ObjC APIs that don't keep the pointers they receive as arguments around or do other dirty pointer tricks. Nonstandard use of pointer arguments still requires UnsafePointer.
Objective-C pointer types now get imported by default as the @unchecked T? optional type. Swift class types no longer implicitly include nil.
A value of @unchecked T? can be implicitly used as a value of T. Swift will implicitly cause a reliable failure if the value is nil, rather than introducing undefined behavior (as in Objective-C ivar accesses or everything in C/C++) or silently ignoring the operation (as in Objective-C message sends).
A value of @unchecked T? can also be implicitly used as a value of T?, allowing you explicitly handle the case of a nil value. For example, if you would like to just silently ignore a message send a la Objective-C, you can use the postfix ? operator like so:
fieldsForKeys[kHeroFieldKey]?.setEditable(true)
This design allows you to isolate and handle nil values in Swift code without requiring excessive “bookkeeping” boilerplate to use values that you expect to be non-nil.
For now, we will continue to import C pointers as non-optional UnsafePointer and C*Pointer types; that will be evaluated separately.
We intend to provide attributes for Clang to allow APIs to opt in to importing specific parameters, return types, etc. as either the explicit optional type T? or the simple non-optional type T.
The “separated” call syntax, i.e.,
NSColor.colorWithRed(r) green(g) blue(b) alpha(a) UIColor.init withRed(r) green(g) blue(b) alpha(a)
is being removed. The compiler will now produce an error and provide Fix-Its to rewrite calls to the “keyword-argument” syntax:
NSColor.colorWithRed(r, green: g, blue: b, alpha: a) UIColor(withRed: r, green:g, blue:b, alpha: a)
The objc attribute now optionally accepts a name, which can be used to provide the name for an entity as seen in Objective-C. For example:
class MyType { var enabled: Bool { @objc(isEnabled) get { // ... } } }
The @objc attribute can be used to name initializers, methods, getters, setters, classes, and protocols.
Methods, properties and subscripts in classes can now be marked with the @final attribute. This attribute prevents overriding the declaration in any subclass, and provides better performance (since dynamic dispatch is avoided in many cases).
Attributes on declarations are no longer comma separated.
Old syntax:
@_silgen_name("foo"), @objc func bar() {}
New syntax:
@_silgen_name("foo") @objc
The , was vestigial when the attribute syntax consisted of bracket lists.
switch now always requires a statement after a case or default.
Old syntax:
switch x { case .A: case .B(1): println(".A or .B(1)") default: // Ignore it. }
New syntax:
switch x { case .A, .B(1): println(".A or .B(1)") default: () // Ignore it. }
The following syntax can be used to introduce guard expressions for patterns inside the case:
switch x { case .A where isFoo(), .B(1) where isBar(): ... }
Observing properties can now @override properties in a base class, so you can observe changes that happen to them.
class MyAwesomeView : SomeBasicView { @override var enabled : Bool { didSet { println("Something changed") } } ... }
Observing properties still invoke the base class getter/setter (or storage) when accessed.
An as cast can now be forced using the postfix ! operator without using parens:
class B {} class D {} let b: B = D() // Before let d1: D = (b as D)! // After let d2: D = b as D!
Casts can also be chained without parens:
// Before let b2: B = (((D() as B) as D)!) as B // After let b3: B = D() as B as D! as B
as can now be used in switch cases to match the result of a checked cast:
func printHand(hand: Any) { switch hand { case 1 as Int: print("ace") case 11 as Int: print("jack") case 12 as Int: print("queen") case 13 as Int: print("king") case let numberCard as Int: print("\(numberCard)") case let (a, b) as (Int, Int) where a == b: print("two ") printHand(a) print("s") case let (a, b) as (Int, Int): printHand(a) print(" and a ") printHand(b) case let (a, b, c) as (Int, Int, Int) where a == b && b == c: print("three ") printHand(a) print("s") case let (a, b, c) as (Int, Int, Int): printHand(a) print(", ") printHand(b) print(", and a ") printHand(c) default: print("unknown hand") } } printHand(1, 1, 1) // prints "three aces" printHand(12, 13) // prints "queen and a king"
Enums and option sets imported from C/Objective-C still strip common prefixes, but the name of the enum itself is now taken into consideration as well. This keeps us from dropping important parts of a name that happen to be shared by all members.
// NSFileManager.h typedef NS_OPTIONS(NSUInteger, NSDirectoryEnumerationOptions) { NSDirectoryEnumerationSkipsSubdirectoryDescendants = 1UL << 0, NSDirectoryEnumerationSkipsPackageDescendants = 1UL << 1, NSDirectoryEnumerationSkipsHiddenFiles = 1UL << 2 } NS_ENUM_AVAILABLE(10_6, 4_0);
// Swift let opts: NSDirectoryEnumerationOptions = .SkipsPackageDescendants
init methods in Objective-C protocols are now imported as initializers. To conform to NSCoding, you will now need to provide
init withCoder(aDecoder: NSCoder) { ... }
rather than
func initWithCoder(aDecoder: NSCoder) { ... }
When a class provides no initializers of its own but has default values for all of its stored properties, it will automatically inherit all of the initializers of its superclass. For example:
class Document { var title: String init() -> Self { self.init(withTitle: "Default title") } init withTitle(title: String) { self.title = title } } class VersionedDocument : Document { var version = 0 // inherits 'init' and 'init withTitle:' from Document }
When one does provide a designated initializer in a subclass, as in the following example:
class SecureDocument : Document { var key: CryptoKey init withKey(key: CryptoKey) -> Self { self.init(withKey: key, title: "Default title") } init withKey(key: CryptoKey) title(String) { self.key = key super.init(withTitle: title) } }
the compiler emits Objective-C method stubs for all of the designated initializers of the parent class that will abort at runtime if called, and which indicate which initializer needs to be implemented. This provides memory safety for cases where an Objective-C initializer (such as -[Document init] in this example) appears to be inherited, but isn't actually implemented.
nil may now be used as a Selector value. This allows calls to Cocoa methods that accept nil selectors.
[] and [:] can now be used as the empty array and dictionary literal, respectively. Because these carry no information about their element types, they may only be used in a context that provides this information through type inference (e.g. when passing a function argument).
Properties defined in classes are now dynamically dispatched and can be overridden with @override. Currently @override only works with computed properties overriding other computed properties, but this will be enhanced in coming weeks.
The didSet accessor of an observing property now gets passed in the old value, so you can easily implement an action for when a property changes value. For example:
class MyAwesomeView : UIView { var enabled : Bool = false { didSet(oldValue): if oldValue != enabled { self.needsDisplay = true } } ... }
The implicit argument name for set and willSet property specifiers has been renamed from (value) to (newValue). For example:
var i : Int { get { return 42 } set { // defaults to (newValue) instead of (value) print(newValue) } }
The magic identifier __FUNCTION__ can now be used to get the name of the current function as a string. Like __FILE__ and __LINE__, if __FUNCTION__ is used as a default argument, the function name of the caller is passed as the argument.
func malkovich() { println(__FUNCTION__) } malkovich() // prints "malkovich" func nameCaller(name: String = __FUNCTION__) -> String { return name } func foo() { println(nameCaller()) // prints "foo" } func foo(x: Int) bar(y: Int) { println(nameCaller()) // prints "foo:bar:" }
At top level, __FUNCTION__ gives the module name:
println(nameCaller()) // prints your module name
Selector-style methods can now be referenced without applying arguments using member syntax foo.bar:bas:, for instance, to test for the availability of an optional protocol method:
func getFrameOfObjectValueForColumn(ds: NSTableViewDataSource, tableView: NSTableView, column: NSTableColumn, row: Int) -> AnyObject? { if let getObjectValue = ds.tableView:objectValueForTableColumn:row: { return getObjectValue(tableView, column, row) } return nil }
The compiler now warns about cases where a variable is inferred to have AnyObject, AnyClass, or () type, since type inference can turn a simple mistake (e.g. failing to cast an AnyObject when you meant to) into something with ripple effects. Here is a simple example:
t.swift:4:5: warning: variable 'fn' inferred to have type '()', which may be unexpected
var fn = abort()
^
t.swift:4:5: note: add an explicit type annotation to silence this warning
var fn = abort()
^
: ()
If you actually did intend to declare a variable of one of these types, you can silence this warning by adding an explicit type (indicated by the Fixit). See rdar://15263687 and rdar://16252090 for more rationale.
x.type has been renamed to x.dynamicType, and you can use type as a regular identifier again.
C macros that expand to a single constant string are now imported as global constants. Normal string literals are imported as CString; NSString literals are imported as String.
All values now have a self property, exactly equivalent to the value itself:
let x = 0 let x2 = x.self
Types also have a self property that is the type object for that type:
let theClass = NSObject.self let theObj = theClass()
References to type names are now disallowed outside of a constructor call or member reference; to get a type object as a value, T.self is required. This prevents the mistake of intending to construct an instance of a class but forgetting the parens and ending up with the class object instead:
let x = MyObject // oops, I meant MyObject()... return x.description() // ...and I accidentally called +description // instead of -description
Initializers are now classified as designated initializers, which are responsible for initializing the current class object and chaining via super.init, and convenience initializers, which delegate to another initializer and can be inherited. For example:
class A { var str: String init() -> Self { // convenience initializer self.init(withString: "hello") } init withString(str: String) { // designated initializer self.str = str } }
When a subclass overrides all of its superclass's designated initializers, the convenience initializers are inherited:
class B { init withString(str: String) { // designated initializer super.init(withString: str) } // inherits A.init() }
Objective-C classes that provide NS_DESIGNATED_INITIALIZER annotations will have their init methods mapped to designated initializers or convenience initializers as appropriate; Objective-C classes without NS_DESIGNATED_INITIALIZER annotations have all of their init methods imported as designated initializers, which is safe (but can be verbose for subclasses). Note that the syntax and terminology is still somewhat in flux.
Initializers can now be marked as required with an attribute, meaning that every subclass is required to provide that initializer either directly or by inheriting it from a superclass. To construct
class View { @required init withFrame(frame: CGRect) { ... } } func buildView(subclassObj: View.Type, frame: CGRect) -> View { return subclassObj(withFrame: frame) } class MyView : View { @required init withFrame(frame: CGRect) { super.init(withFrame: frame) } } class MyOtherView : View { // error: must override init withFrame(CGRect). }
Properties in Objective-C protocols are now correctly imported as properties. (Previously the getter and setter were imported as methods.)
Simple enums with no payloads, including NS_ENUMs imported from Cocoa, now implicitly conform to the Equatable and Hashable protocols. This means they can be compared with the == and != operators and can be used as Dictionary keys:
enum Flavor { case Lemon, Banana, Cherry } assert(Flavor.Lemon == .Lemon) assert(Flavor.Banana != .Lemon) struct Profile { var sweet, sour: Bool } let flavorProfiles: Dictionary<Flavor, Profile> = [ .Lemon: Profile(sweet: false, sour: true ), .Banana: Profile(sweet: true, sour: false), .Cherry: Profile(sweet: true, sour: true ), ] assert(flavorProfiles[.Lemon].sour)
val has been removed. Long live let!
Values whose names clash with Swift keywords, such as Cocoa methods or properties named class, protocol, type, etc., can now be defined and accessed by wrapping reserved keywords in backticks to suppress their builtin meaning:
let `class` = 0 let `type` = 1 let `protocol` = 2 println(`class`) println(`type`) println(`protocol`) func foo(Int) `class`(Int) {} foo(0, `class`: 1)
The override attribute is now required when overriding a method, property, or subscript from a superclass. For example:
class A { func foo() { } } class B : A { @override func foo() { } // 'override' is required here }
We're renaming val back to let. The compiler accepts both for this week, next week it will just accept let. Please migrate your code this week, sorry for the back and forth on this.
Swift now supports #if, #else and #endif blocks, along with target configuration expressions, to allow for conditional compilation within declaration and statement contexts.
Target configurations represent certain static information about the compile-time build environment. They are implicit, hard-wired into the compiler, and can only be referenced within the conditional expression of an #if block.
Target configurations are tested against their values via a pseudo-function invocation expression, taking a single argument expressed as an identifier. The argument represents certain static build-time information.
There are currently two supported target configurations: os, which can have the values OSX or iOS arch, which can have the values i386, x86_64, arm and arm64
Within the context of an #if block's conditional expression, a target configuration expression can evaluate to either true or false.
For example:
#if arch(x86_64) println("Building for x86_64") #else println("Not building for x86_64") #endif class C { #if os(OSX) func foo() { // OSX stuff goes here } #else func foo() { // non-OSX stuff goes here } #endif }
The conditional expression of an #if block can be composed of one or more of the following expression types:
!&& or ||For example:
#if os(iOS) && !arch(I386) ... #endif
Note that #if/#else/#endif blocks do not constitute a preprocessor, and must form valid and complete expressions or statements. Hence, the following produces a parser error:
class C { #if os(iOS) func foo() {} } #else func bar() {} func baz() {} } #endif
Also note that “active” code will be parsed, typechecked and emitted, while “inactive” code will only be parsed. This is why code in an inactive #if or #else block will produce parser errors for malformed code. This allows the compiler to detect basic errors in inactive regions.
This is the first step to getting functionality parity with the important subset of the C preprocessor. Further refinements are planned for later.
Swift now has both fully-closed ranges, which include their endpoint, and half-open ranges, which don't.
(swift) for x in 0...5 { print(x) } ; print('\n') // half-open range 01234 (swift) for x in 0..5 { print(x) } ; print('\n') // fully-closed range 012345
Property accessors have a new brace-based syntax, instead of using the former “label like” syntax. The new syntax is:
var computedProperty: Int { get { return _storage } set { _storage = value } } var implicitGet: Int { // This form still works. return 42 } var storedPropertyWithObservingAccessors: Int = 0 { willSet { ... } didSet { ... } }
Properties and subscripts now work in protocols, allowing you to do things like:
protocol Subscriptable { subscript(idx1: Int, idx2: Int) -> Int { get set } var prop: Int { get } } func foo(s: Subscriptable) { return s.prop + s[42, 19] }
These can be used for generic algorithms now as well.
The syntax for referring to the type of a type, T.metatype, has been changed to T.Type. The syntax for getting the type of a value, typeof(x), has been changed to x.type.
DynamicSelf is now called Self; the semantics are unchanged.
destructor has been replaced with deinit, to emphasize that it is related to init. We will refer to these as deinitializers. We've also dropped the parentheses, i.e.:
class MyClass { deinit { // release any resources we might have acquired, etc. } }
Class methods defined within extensions of Objective-C classes can now refer to self, including using instancetype methods. As a result, NSMutableString, NSMutableArray, and NSMutableDictionary objects can now be created with their respective literals, i.e.,
var dict: NSMutableDictionary = ["a" : 1, "b" : 2]
The Stream protocol has been renamed back to Generator, which is precedented in other languages and causes less confusion with I/O streaming.
The type keyword was split into two: static and class. One can define static functions and static properties in structs and enums like this:
struct S { static func foo() {} static var bar: Int = 0 } enum E { static func foo() {} }
class keyword allows one to define class properties and class methods in classes and protocols:
class C { class func foo() {} class var bar: Int = 0 } protocol P { class func foo() {} class var bar: Int = 0 }
When using class and static in the extension, the choice of keyword depends on the type being extended:
extension S { static func baz() {} } extension C { class func baz() {} }
The let keyword is no longer recognized. Please move to val.
The standard library has been renamed to Swift (instead of swift) to be more consistent with other modules on our platforms.
NSInteger and other types that are layout-compatible with Swift standard library types are now imported directly as those standard library types.
Optional types now support a convenience method named “cache” to cache the result of a closure. For example:
class Foo { var _lazyProperty: Int? var property: Int { return _lazyProperty.cache { computeLazyProperty() } } }
We are experimenting with a new message send syntax. For example:
SKAction.colorizeWithColor(SKColor.whiteColor()) colorBlendFactor(1.0) duration(0.0)
When the message send is too long to fit on a single line, subsequent lines must be indented from the start of the statement or declaration. For example, this is a single message send:
SKAction.colorizeWithColor(SKColor.whiteColor()) colorBlendFactor(1.0) duration(0.0)
while this is a message send to colorizeWithColor: followed by calls to colorBlendFactor and duration (on self or to a global function):
SKAction.colorizeWithColor(SKColor.whiteColor()) colorBlendFactor(1.0) // call to 'colorBlendFactor' duration(0.0) // call to 'duration'
We are renaming the let keyword to val. The let keyword didn‘t work out primarily because it is not a noun, so “defining a let” never sounded right. We chose val over const and other options because var and val have similar semantics (making syntactic similarity useful), because const has varied and sordid connotations in C that we don’t want to bring over, and because we don't want to punish the “preferred” case with a longer keyword.
For migration purposes, the compiler now accepts let and val as synonyms, let will be removed next week.
Selector arguments in function arguments with only a type are now implicitly named after the selector chunk that contains them. For example, instead of:
func addIntsWithFirst(first : Int) second(second : Int) -> Int { return first+second }
you can now write:
func addIntsWithFirst(first : Int) second(Int) -> Int { return first+second }
if you want to explicitly want to ignore an argument, it is recommended that you continue to use the _ to discard it, as in:
func addIntsWithFirst(first : Int) second(_ : Int) -> Int {...}
The @inout attribute in argument lists has been promoted to a context-sensitive keyword. Where before you might have written:
func swap<T>(a : @inout T, b : @inout T) { (a, b) = (b, a) }
You are now required to write:
func swap<T>(inout a : T, inout b : T) { (a, b) = (b, a) }
We made this change because inout is a fundamental part of the type system, which attributes are a poor match for. The inout keyword is also orthogonal to the var and let keywords (which may be specified in the same place), so it fits naturally there.
The @mutating attribute (which can be used on functions in structs, enums, and protocols) has been promoted to a context-sensitive keyword. Mutating struct methods are now written as:
struct SomeStruct { mutating func f() {} }
Half-open ranges (those that don't include their endpoint) are now spelled with three .s instead of two, for consistency with Ruby.
(swift) for x in 0...5 { print(x) } ; print('\n') // new syntax 01234
Next week, we'll introduce a fully-closed range which does include its endpoint. This will provide:
(swift) for x in 0..5 { print(x) } ; print('\n') // coming soon 012345
These changes are being released separately so that users have a chance to update their code before its semantics changes.
Objective-C properties with custom getters/setters are now imported into Swift as properties. For example, the Objective-C property
@property (getter=isEnabled) BOOL enabled;
was previously imported as getter (isEnabled) and setter (setEnabled) methods. Now, it is imported as a property (enabled).
didSet/willSet properties may now have an initial value specified:
class MyAwesomeView : UIView { var enabled : Bool = false { // Initial value. didSet: self.needsDisplay = true } ... }
they can also be used as non-member properties now, e.g. as a global variable or a local variable in a function.
Objective-C instancetype methods are now imported as methods that return Swift's DynamicSelf type. While DynamicSelf is not generally useful for defining methods in Swift, importing to it eliminates the need for casting with the numerous instancetype APIs, e.g.,
let tileNode: SKSpriteNode = SKSpriteNode.spriteNodeWithTexture(tileAtlas.textureNamed("tile\(tileNumber).png"))!
becomes
let tileNode = SKSpriteNode.spriteNodeWithTexture(tileAtlas.textureNamed("tile\(tileNumber).png"))
DynamicSelf will become more interesting in the coming weeks.
if and while statements can now conditionally bind variables. If the condition of an if or while statement is a let declaration, then the right-hand expression is evaluated as an Optional value, and control flow proceeds by considering the binding to be true if the Optional contains a value, or false if it is empty, and the variables are available in the true branch. This allows for elegant testing of dynamic types, methods, nullable pointers, and other Optional things:
class B : NSObject {} class D : B { func foo() { println("we have a D") } } var b: B = D() if let d = b as D { d.foo() } var id: AnyObject = D() if let foo = id.foo { foo() }
When referring to a member of an AnyObject (or AnyClass) object and using it directly (such as calling it, subscripting, or accessing a property on it), one no longer has to write the ? or !. The run-time check will be performed implicitly. For example:
func doSomethingOnViews(views: NSArray) { for view in views { view.updateLayer() // no '!' needed } }
Note that one can still test whether the member is available at runtime using ?, testing the optional result, or conditionally binding a variable to the resulting member.
The swift command line tool can now create executables and libraries directly, just like Clang. Use swift main.swift to create an executable and swift -emit-library -o foo.dylib foo.swift to create a library.
Object files emitted by Swift are not debuggable on their own, even if you compiled them with the -g option. This was already true if you had multiple files in your project. To produce a debuggable Swift binary from the command line, you must compile and link in a single step with swift, or pass object files AND swiftmodule files back into swift after compilation. (Or use Xcode.)
import will no longer import other source files, only built modules.
The current directory is no longer implicitly an import path. Use -I . if you have modules in your current directory.
Properties in structs and classes may now have willSet: and didSet: observing accessors defined on them:
For example, where before you may have written something like this in a class:
class MyAwesomeView : UIView { var _enabled : Bool // storage var enabled : Bool { // computed property get: return _enabled set: _enabled = value self.needDisplay = true } ... }
you can now simply write:
class MyAwesomeView : UIView { var enabled : Bool { // Has storage & observing methods didSet: self.needDisplay = true } ... }
Similarly, if you want notification before the value is stored, you can use willSet, which gets the incoming value before it is stored:
var x : Int { willSet(value): // value is the default and may be elided, as with set: println("changing from \(x) to \(value)") didSet: println("we've got a value of \(x) now.\n") }
The willSet/didSet observers are triggered on any store to the property, except stores from init(), destructors, or from within the observers themselves.
Overall, a property now may either be “stored” (the default), “computed” (have a get: and optionally a set: specifier), or an observed (willSet/didSet) property. It is not possible to have a custom getter or setter on an observed property, since they have storage.
Two known-missing bits are:
Because of the first one, for now, you need to explicitly store an initial value to the property in your init() method.
Objective-C properties with custom getter or setter names are (temporarily) not imported into Swift; the getter and setter will be imported individually as methods instead. Previously, they would appear as properties within the Objective-C class, but attempting to use the accessor with the customized name would result in a crash.
The long-term fix is tracked as (rdar://problem/15877160).
Computed ‘type’ properties (that is, properties of types, rather than of values of the type) are now permitted on classes, on generic structs and enums, and in extensions. Stored ‘type’ properties in these contexts remain unimplemented.
The implementation of stored ‘type’ properties is tracked as (rdar://problem/15915785) (for classes) and (rdar://problem/15915867) (for generic types).
The following command-line flags have been deprecated in favor of new spellings. The old spellings will be removed in the following week's build:
| Old Spelling | New Spelling |
|---|---|
-emit-llvm | -emit-ir |
-triple | -target |
-serialize-diagnostics | -serialize-diagnostics-path |
Imported NS_OPTIONS types now have a default initializer which produces a value with no options set. They can also be initialized to the empty set with nil. These are equivalent:
var x = NSMatchingOptions() var y: NSMatchingOptions = nil
The swift binary no longer has an SDK set by default. Instead, you must do one of the following:
-sdk /path/to/sdkSDKROOT in your environmentswift through xcrun, which sets SDKROOT for youlet declarations can now be used as struct/class properties. A let property is mutable within init(), and immutable everywhere else.
class C { let x = 42 let y : Int init(y : Int) { self.y = y // ok, self.y is mutable in init() } func test() { y = 42 // error: 'y' isn't mutable } }
The immutability model for structs and enums is complete, and arguments are immutable by default. This allows the compiler to reject mutations of temporary objects, catching common bugs. For example, this is rejected:
func setTo4(a : Double[]) { a[10] = 4.0 // error: 'a' isn't mutable } ... setTo4(someArray)
since a is semantically a copy of the array passed into the function. The proper fix in this case is to mark the argument is @inout, so the effect is visible in the caller:
func setTo4(a : @inout Double[]) { a[10] = 4.0 // ok: 'a' is a mutable reference } ... setTo4(&someArray)
Alternatively, if you really just want a local copy of the argument, you can mark it var. The effects aren't visible in the caller, but this can be convenient in some cases:
func doStringStuff(var s : String) { s += "foo" print(s) }
Objective-C instance variables are no longer imported from headers written in Objective-C. Previously, they would appear as properties within the Objective-C class, but trying to access them would result in a crash. Additionally, their names can conflict with property names, which confuses the Swift compiler, and there are no patterns in our frameworks that expect you to access a parent or other class's instance variables directly. Use properties instead.
The NSObject protocol is now imported under the name NSObjectProtocol (rather than NSObjectProto).
Improved deallocation of Swift classes that inherit from Objective-C classes: Swift destructors are implemented as -dealloc methods that automatically call the superclass's -dealloc. Stored properties are released right before the object is deallocated (using the same mechanism as ARC), allowing properties to be safely used in destructors.
Subclasses of NSManagedObject are now required to provide initial values for each of their stored properties. This permits initialization of these stored properties directly after +alloc to provide memory safety with CoreData's dynamic subclassing scheme.
let declarations are continuing to make slow progress. Curried and selector-style arguments are now immutable by default, and let declarations now get proper debug information.
The static keyword changed to type. One can now define “type functions” and “type variables” which are functions and variables defined on a type (rather than on an instance of the type), e.g.,
class X { type func factory() -> X { ... } type var version: Int }
The use of static was actively misleading, since type methods on classes are dynamically dispatched (the same as Objective-C + methods).
Note that type is a context-sensitive keyword; it can still be used as an identifier.
Strings have a new native UTF-16 representation that can be converted back and forth to NSString at minimal cost. String literals are emitted as UTF-16 for string types that support it (including Swift's String).
Initializers can now delegate to other initializers within the same class by calling self.init. For example:
class A { } class B : A { var title: String init() { // note: cannot access self before delegating self.init(withTitle: "My Title") } init withTitle(title: String) { self.title = title super.init() } }
Objective-C protocols no longer have the Proto suffix unless there is a collision with a class name. For example, UITableViewDelegate is now imported as UITableViewDelegate rather than UITableViewDelegateProto. Where there is a conflict with a class, the protocol will be suffixed with Proto, as in NSObject (the class) and NSObjectProto (the protocol).
Happy New Year
Division and remainder arithmetic now trap on overflow. Like with the other operators, one can use the “masking” alternatives to get non-trapping behavior. The behavior of the non-trapping masking operators is defined:
x &/ 0 == 0 x &% 0 == 0 SIGNED_MIN_FOR_TYPE &/ -1 == -1 // i.e. Int8: -0x80 / -1 == -0x80 SIGNED_MIN_FOR_TYPE &% -1 == 0
Protocol conformance checking for @mutating methods is now implemented: an @mutating struct method only fulfills a protocol requirement if the protocol method was itself marked @mutating:
protocol P { func nonmutating() @mutating func mutating() } struct S : P { // Error, @mutating method cannot implement non-@mutating requirement. @mutating func nonmutating() {} // Ok, mutating allowed, but not required. func mutating() {} }
As before, class methods never need to be marked @mutating (and indeed, they aren't allowed to be marked as such).
Merry Christmas
The setters of properties on value types (structs/enums) are now @mutating by default. To mark a setter non-mutating, use the @!mutating attribute.
Compiler inserts calls to super.init() into the class initializers that do not call any initializers explicitly.
A map method with the semantics of Haskell's fmap was added to Array<T>. Map applies a function f: T->U to the values stored in the array and returns an Array. So,
(swift) func names(x: Int[]) -> String[] { return x.map { "<" + String($0) + ">" } } (swift) names(Array<Int>()) // r0 : String[] = [] (swift) names([3, 5, 7, 9]) // r1 : String[] = ["<3>", "<5>", "<7>", "<9>"]
Global variables and static properties are now lazily initialized on first use. Where you would use dispatch_once to lazily initialize a singleton object in Objective-C, you can simply declare a global variable with an initializer in Swift. Like dispatch_once, this lazy initialization is thread safe.
Unlike C++ global variable constructors, Swift global variables and static properties now never emit static constructors (and thereby don‘t raise build warnings). Also unlike C++, lazy initialization naturally follows dependency order, so global variable initializers that cross module boundaries don’t have undefined behavior or fragile link order dependencies.
Swift has the start of an immutability model for value types. As part of this, you can now declare immutable value bindings with a new let declaration, which is semantically similar to defining a get-only property:
let x = foo() print(x) // ok x = bar() // error: cannot modify an immutable value swap(&x, &y) // error: cannot pass an immutable value as @inout parameter x.clear() // error: cannot call mutating method on immutable value getX().clear() // error: cannot mutate a temporary
In the case of bindings of class type, the bound object itself is still mutable, but you cannot change the binding.
let r = Rocket() r.blastOff() // Ok, your rocket is mutable. r = Rocket() // error: cannot modify an immutable binding.
In addition to the let declaration itself, self on classes, and a few other minor things have switched to immutable bindings.
A pivotal part of this is that methods of value types (structs and enums) need to indicate whether they can mutate self - mutating methods need to be disallowed on let values (and get-only property results, temporaries, etc) but non-mutating methods need to be allowed. The default for a method is that it does not mutate self, though you can opt into mutating behavior with a new @mutating attribute:
struct MyWeirdCounter { var count : Int func empty() -> Bool { return count == 0 } @mutating func reset() { count = 0 } ... } let x = MyWeirdCounter() x.empty() // ok x.reset() // error, cannot mutate immutable 'let' value
One missing piece is that the compiler does not yet reject mutations of self in a method that isn't marked @mutating. That will be coming soon. Related to methods are properties. Getters and setters can be marked mutating as well:
extension MyWeirdCounter { var myproperty : Int { get: return 42 @mutating set: count = value*2 } }
The intention is for setters to default to mutating, but this has not been implemented yet. There is more to come here.
A map method with the semantics of Haskell's fmap was added to Optional<T>. Map applies a function f: T->U to any value stored in an Optional<T>, and returns an Optional<U>. So,
(swift) func nameOf(x: Int?) -> String? { return x.map { "<" + String($0) + ">" } } (swift) (swift) var no = nameOf(.None) // Empty optional in... // no : String? = <unprintable value> (swift) no ? "yes" : "no" // ...empty optional out // r0 : String = "no" (swift) (swift) nameOf(.Some(42)) // Non-empty in // r1 : String? = <unprintable value> (swift) nameOf(.Some(42))! // Non-empty out // r2 : String = "<42>"
Cocoa types declared with the NS_OPTIONS macro are now available in Swift. Like NS_ENUM types, their values are automatically shortened based on the common prefix of the value names in Objective-C, and the name can be elided when type context provides it. They can be used in if statements using the &, |, ^, and ~ operators as in C:
var options: NSJSONWritingOptions = .PrettyPrinted if options & .PrettyPrinted { println("pretty-printing enabled") }
We haven't yet designed a convenient way to author NS_OPTIONS-like types in Swift.
Objective-C id is now imported as AnyObject (formerly known as DynamicLookup), Objective-C Class is imported as AnyClass.
The casting syntax x as T now permits both implicit conversions (in which case it produces a value of type T) and for runtime-checked casts (in which case it produces a value of type T? that will be .Some(casted x) on success and .None on failure). An example:
func f(x: AnyObject, y: NSControl) { var view = y as NSView // has type 'NSView' var maybeView = x as NSView // has type NSView? }
The precedence levels of binary operators has been redefined, with a much simpler model than C‘s. This is with a goal to define away classes of bugs such as those caught by Clang’s -Wparentheses warnings, and to make it actually possible for normal humans to reason about the precedence relationships without having to look them up.
We ended up with 6 levels, from tightest binding to loosest:
exponentiative: <<, >> multiplicative: *, /, %, & additive: +, -, |, ^ comparative: ==, !=, <, <=, >=, > conjunctive: && disjunctive: ||
The Enumerable protocol has been renamed Sequence.
The Char type has been renamed UnicodeScalar. The preferred unit of string fragments for users is called Character.
Initialization semantics for classes, structs and enums init methods are now properly diagnosed by the compiler. Instance variables now follow the same initialization rules as local variables: they must be defined before use. The initialization model requires that all properties with storage in the current class be initialized before super.init is called (or, in a root class, before any method is called on self, and before the final return).
For example, this will yield an error:
class SomeClass : SomeBase { var x : Int init() { // error: property 'self.x' not initialized at super.init call super.init() } }
A simple fix for this is to change the property definition to var x = 0, or to explicitly assign to it before calling super.init().
Relatedly, the compiler now diagnoses incorrect calls to super.init(). It validates that any path through an initializer calls super.init() exactly once, that all ivars are defined before the call to super.init, and that any uses which require the entire object to be initialized come after the super.init call.
Type checker performance has improved considerably (but we still have much work to do here).
Slice<T> has been folded into Array<T> and T[] is just sugar for Array<T>.Unreachable code warning has been added:
var y: Int = 1 if y == 1 { // note: condition always evaluates to true return y } return 1 // warning: will never be executed
Overflows on integer type conversions are now detected at runtime and, when dealing with constants, at compile time:
var i: Int = -129 var i8 = Int8(i) // error: integer overflows when converted from 'Int' to 'Int8' var si = Int8(-1) var ui = UInt8(si) // error: negative integer cannot be converted to unsigned type 'UInt8'
def keyword was changed back to func.
Objective-C-compatible protocols can now contain optional requirements, indicated by the @optional attribute:
@class_protocol @objc protocol NSWobbling { @optional def wobble() }
A class that conforms to the NSWobbling protocol above can (but does not have to) implement wobble. When referring to the wobble method for a value of type NSWobbling (or a value of generic type that is bounded by NSWobbling), the result is an optional value indicating whether the underlying object actually responds to the given selector, using the same mechanism as messaging id. One can use ! to assume that the method is always there, ? to chain the optional, or conditional branches to handle each case distinctly:
def tryToWobble(w : NSWobbling) { w.wobble() // error: cannot call a value of optional type w.wobble!() // okay: calls -wobble, but fails at runtime if not there w.wobble?() // okay: calls -wobble only if it's there, otherwise no-op if w.wobble { // okay: we know -wobble is there } else { // okay: we know -wobble is not there } }
Enums from Cocoa that are declared with the NS_ENUM macro are now imported into Swift as Swift enums. Like all Swift enums, the constants of the Cocoa enum are scoped as members of the enum type, so the importer strips off the common prefix of all of the constant names in the enum when forming the Swift interface. For example, this Objective-C declaration:
typedef NS_ENUM(NSInteger, NSComparisonResult) { NSOrderedAscending, NSOrderedSame, NSOrderedDescending, };
shows up in Swift as:
enum NSComparisonResult : Int { case Ascending, Same, Descending }
The enum cases can then take advantage of type inference from context. In Objective-C, you would write:
NSNumber *foo = [NSNumber numberWithInt: 1]; NSNumber *bar = [NSNumber numberWithInt: 2]; switch ([foo compare: bar]) { case NSOrderedAscending: NSLog(@"ascending\n"); break; case NSOrderedSame: NSLog(@"same\n"); break; case NSOrderedDescending: NSLog(@"descending\n"); break; }
In Swift, this becomes:
var foo: NSNumber = 1 var bar: NSNumber = 2 switch foo.compare(bar) { case .Ascending: println("ascending") case .Same: println("same") case .Descending: println("descending") }
Work has begun on implementing static properties. Currently they are supported for nongeneric structs and enums.
struct Foo { static var foo: Int = 2 } enum Bar { static var bar: Int = 3 } println(Foo.foo) println(Bar.bar)
func keyword was changed to def.
Implicit conversions are now allowed from an optional type T? to another optional type U? if T is implicitly convertible to U. For example, optional subclasses convert to their optional base classes:
class Base {} class Derived : Base {} var d: Derived? = Derived() var b: Base? = d
Type inference for variables has been improved, allowing any variable to have its type inferred from its initializer, including global and instance variables:
class MyClass { var size = 0 // inferred to Int } var name = "Swift"
Additionally, the arguments of a generic type can also be inferred from the initializer:
// infers Dictionary<String, Int> var dict: Dictionary = ["Hello": 1, "World": 2]
Missing return statement from a non-Void function is diagnosed as an error.
Vector<T> has been replaced with Array<T>. This is a complete rewrite to use value-semantics and copy-on-write behavior. The former means that you never need to defensively copy again (or remember to attribute a property as “copy”) and the latter yields better performance than defensive copying. Dictionary<T> is next.
switch can now pattern-match into structs and classes, using the syntax case Type(property1: pattern1, property2: pattern2, ...):.
struct Point { var x, y: Double } struct Size { var w, h: Double } struct Rect { var origin: Point; var size: Size } var square = Rect(Point(0, 0), Size(10, 10)) switch square { case Rect(size: Size(w: var w, h: var h)) where w == h: println("square") case Rect(size: Size(w: var w, h: var h)) where w > h: println("long rectangle") default: println("tall rectangle") }
Currently only stored properties (“ivars” in ObjC terminology) are supported by the implementation.
Array and dictionary literals allow an optional trailing comma:
var a = [1, 2,] var d = ["a": 1, "b": 2,]
Unlike in Objective-C, objects of type id in Swift do not implicitly convert to any class type. For example, the following code is ill-formed:
func getContentViewBounds(window : NSWindow) -> NSRect { var view : NSView = window.contentView() // error: 'id' doesn't implicitly convert to NSView return view.bounds() }
because contentView() returns an id. One can now use the postfix ! operator to allow an object of type id to convert to any class type, e.g.,
func getContentViewBounds(window : NSWindow) -> NSRect { var view : NSView = window.contentView()! // ok: checked conversion to NSView return view.bounds() }
The conversion is checked at run-time, and the program will fail if the object is not an NSView. This is shorthand for
var view : NSView = (window.contentView() as NSView)!
which checks whether the content view is an NSView (via the as NSView). That operation returns an optional NSView (written NSView?) and the ! operation assumes that the cast succeeded, i.e., that the optional has a value in it.
The unconditional checked cast syntax x as! T has been removed. Many cases where conversion from id is necessary can now be handled by postfix ! (see above). Fully general unconditional casts can still be expressed using as and postfix ! together, (x as T)!.
The old “square bracket” attribute syntax has been removed.
Overflows on construction of integer and floating point values from integer literals that are too large to fit the type are now reported by the compiler. Here are some examples:
var x = Int8(-129) // error: integer literal overflows when stored into 'Int8' var y: Int = 0xFFFF_FFFF_FFFF_FFFF_F // error: integer literal overflows when stored into 'Int'
Overflows in constant integer expressions are also reported by the compiler.
var x: Int8 = 125 var y: Int8 = x + 125 // error: arithmetic operation '125 + 125' (on type 'Int8') results in // an overflow
Division by zero in constant expressions is now detected by the compiler:
var z: Int = 0 var x = 5 / z // error: division by zero
Generic structs with type parameters as field types are now fully supported.
struct Pair<T, U> { var first: T var second: U }
Autorelease pools can now be created using the autoreleasepool function.
autoreleasepool { // code }
Note that the wrapped code is a closure, so constructs like break and continue and return do not behave as they would inside an Objective-C @autoreleasepool statement.
Enums can now declare a “raw type”, and cases can declare “raw values”, similar to the integer underlying type of C enums:
// Declare the underlying type as in Objective-C or C++11, with // ': Type' enum AreaCode : Int { // Assign explicit values to cases with '=' case SanFrancisco = 415 case EastBay = 510 case Peninsula = 650 case SanJose = 408 // Values are also assignable by implicit auto-increment case Galveston // = 409 case Baltimore // = 410 }
This introduces fromRaw and toRaw methods on the enum to perform conversions from and to the raw type:
/* As if declared: extension AreaCode { // Take a raw value, and produce the corresponding enum value, // or None if there is no corresponding enum value static func fromRaw(raw:Int) -> AreaCode? // Return the corresponding raw value for 'self' func toRaw() -> Int } */ AreaCode.fromRaw(415) // => .Some(.SanFrancisco) AreaCode.fromRaw(111) // => .None AreaCode.SanJose.toRaw() // => 408
Raw types are not limited to integer types--they can additionally be character, floating-point, or string values:
enum State : String { case CA = "California" case OR = "Oregon" case WA = "Washington" } enum SquareRootOfInteger : Float { case One = 1.0 case Two = 1.414 case Three = 1.732 case Four = 2.0 }
Raw types are currently limited to simple C-like enums with no payload cases. The raw values are currently restricted to simple literal values; expressions such as 1 + 1 or references to other enum cases are not yet supported. Raw values are also currently required to be unique for each case in an enum.
Enums with raw types implicitly conform to the RawRepresentable protocol, which exposes the fromRaw and toRaw methods to generics:
protocol RawRepresentable { typealias RawType static func fromRaw(raw: RawType) -> Self? func toRaw() -> RawType }
Attribute syntax has been redesigned (see (rdar://10700853) and (rdar://14462729)) so that attributes now precede the declaration and use the @ character to signify them. Where before you might have written:
func [someattribute=42] foo(a : Int) {}
you now write:
@someattribute=42 func foo(a : Int) {}
This flows a lot better (attributes don't push the name for declarations away), and means that square brackets are only used for array types, collection literals, and subscripting operations.
The for loop now uses the Generator protocol instead of the Enumerator protocol to iterate a sequence. This protocol looks like this:
protocol Generator { typealias Element func next() -> Element? }
The single method next() advances the generator and returns an Optional, which is either .Some(value), wrapping the next value out of the underlying sequence, or .None to signal that there are no more elements. This is an improvement over the previous Enumerator protocol because it eliminates the separate isEmpty() query and better reflects the semantics of ephemeral sequences like un-buffered input streams.
The [byref] attribute has been renamed to [inout]. When applied to a logical property, the getter is invoked before a call and the setter is applied to write back the result. inout conveys this better and aligns with existing Objective-C practice better.
[inout] arguments can now be captured into closures. The semantics of a inout capture are that the captured variable is an independent local variable of the callee, and the inout is updated to contain the value of that local variable at function exit.
In the common case, most closure arguments do not outlive the duration of their callee, and the observable behavior is unchanged. However, if the captured variable outlives the function, you can observe this. For example, this code:
func foo(x : [inout] Int) -> () -> Int { func bar() -> Int { x += 1 return x } // Call 'bar' once while the inout is active. bar() return bar } var x = 219 var f = foo(&x) // x is updated to the value of foo's local x at function exit. println("global x = \(x)") // These calls only update the captured local 'x', which is now independent // of the inout parameter. println("local x = \(f())") println("local x = \(f())") println("local x = \(f())") println("global x = \(x)")
will print:
global x = 220 local x = 221 local x = 222 local x = 223 global x = 220
In no case will you end up with a dangling pointer or other unsafe construct.
x as T now performs a checked cast to T?, producing .Some(t) if the cast succeeds, or .None if the cast fails.
The ternary expression (x ? y : z) now requires whitespace between the first expression and the question mark. This permits ? to be used as a postfix operator.
A significant new piece of syntactic sugar has been added to ease working with optional values. The ? postfix operator is analogous to !, but instead of asserting on None, it causes all the following postfix operators to get skipped and return None.
In a sense, this generalizes (and makes explicit) the Objective-C behavior where message sends to nil silently produce the zero value of the result.
For example, this code
object?.parent.notifyChildEvent?(object!, .didExplode)
first checks whether object has a value; if so, it drills to its parent and checks whether that object implements the notifyChildEvent method; if so, it calls that method. (Note that we do not yet have generalized optional methods.)
This code:
var titleLength = object?.title.length
checks whether object has a value and, if so, asks for the length of its title. titleLength will have type Int?, and if object was missing, the variable will be initialized to None.
Objects with type id can now be used as the receiver of property accesses and subscript operations to get (but not set) values. The result is of optional type. For example, for a variable obj of type id, the expression
obj[0]
will produce a value of type id, which will either contain the result of the message send objectAtIndexedSubscript(0) (wrapped in an optional type) or, if the object does not respond to objectAtIndexedSubscript:, an empty optional. The same approach applies to property accesses.
_ can now be used not only in var bindings, but in assignments as well, to ignore elements of a tuple assignment, or to explicitly ignore values.
var a = (1, 2.0, 3) var x = 0, y = 0 _ = a // explicitly load and discard 'a' (x, _, y) = a // assign a.0 to x and a.2 to y
The union keyword has been replaced with enum. Unions and enums are semantically identical in swift (the former just has data associated with its discriminators) and enum is the vastly more common case. For more rationale, please see docs/proposals/Enums.rst
The Optional type T? is now represented as an enum:
enum Optional<T> { case None case Some(T) }
This means that, in addition to the existing Optional APIs, it can be pattern-matched with switch:
var x : X?, y : Y? switch (x, y) { // Both are present case (.Some(var a), .Some(var b)): println("both") // One is present case (.Some, .None): case (.None, .Some): println("one") // Neither is present case (.None, .None): println("neither") }
Enums now allow multiple cases to be declared in a comma-separated list in a single case declaration:
enum Color { case Red, Green, Blue }
The Objective-C id and Class types now support referring to methods declared in any class or protocol without a downcast. For example, given a variable sender of type id, one can refer to -isEqual: with:
sender.isEqual
The actual object may or may not respond to -isEqual, so this expression returns result of optional type whose value is determined via a compiler-generated -respondsToSelector send. When it succeeds, the optional contains the method; when it fails, the optional is empty.
To safely test the optional, one can use, e.g.,
var senderIsEqual = sender.isEqual if senderIsEqual { // this will never trigger an "unrecognized selector" failure var equal = senderIsEqual!(other) } else { // sender does not respond to -isEqual: }
When you know that the method is there, you can use postfix ! to force unwrapping of the optional, e.g.,
sender.isEqual!(other)
This will fail at runtime if in fact sender does not respond to -isEqual:. We have some additional syntactic optimizations planned for testing an optional value and handling both the success and failure cases concisely. Watch this space.
Weak references now always have optional type. If a weak variable has an explicit type, it must be an optional type:
var [weak] x : NSObject?
If the variable is not explicitly typed, its type will still be inferred to be an optional type.
There is now an implicit conversion from T to T?.
Constructor syntax has been improved to align better with Objective-C's init methods. The constructor keyword has been replaced with init, and the selector style of declaration used for func declarations is now supported. For example:
class Y : NSObject { init withInt(i : Int) string(s : String) { super.init() // call superclass initializer } }
One can use this constructor to create a Y object with, e.g.,
Y(withInt:17, string:"Hello")
Additionally, the rules regarding the selector corresponding to such a declaration have been revised. The selector for the above initializer is initWithInt:string:; the specific rules are described in the documentation.
Finally, Swift initializers now introduce Objective-C entry points, so a declaration such as:
class X : NSObject { init() { super.init() } }
Overrides NSObject's -init method (which it calls first) as well as introducing the ‘allocating’ entry point so that one can create a new X instance with the syntax X().
Variables in top-level code (i.e. scripts, but not global variables in libraries) that lack an initializer now work just like local variables: they must be explicitly assigned-to sometime before any use, instead of being default constructed. Instance variables are still on the TODO list.
Generic unions with a single payload case and any number of empty cases are now implemented, for example:
union Maybe<T> { case Some(T) case None } union Tristate<T> { case Initialized(T) case Initializing case Uninitialized }
Generic unions with multiple payload cases are still not yet implemented.
The implementation now supports partial application of class and struct methods:
(swift) class B { func foo() { println("B") } } (swift) class D : B { func foo() { println("D") } } (swift) var foo = B().foo // foo : () -> () = <unprintable value> (swift) foo() B (swift) foo = D().foo (swift) foo() D
Support for partial application of Objective-C class methods and methods in generic contexts is still incomplete.
Local variable declarations without an initializer are no longer implicitly constructed. The compiler now verifies that they are initialized on all paths leading to a use of the variable. This means that constructs like this are now allowed:
var p : SomeProtocol if whatever { p = foo() } else { p = bar() }
where before, the compiler would reject the definition of p saying that it needed an initializer expression.
Since all local variables must be initialized before use, simple things like this are now rejected as well:
var x : Int print(x)
The fix is to initialize the value on all paths, or to explicitly default initialize the value in the declaration, e.g. with var x = 0 or with var x = Int() (which works for any default-constructible type).
The implementation now supports unions containing protocol types and weak reference types.
The type annotation syntax, x as T, has been removed from the language. The checked cast operations x as! T and x is T still remain.
this has been renamed to self. Similarly, This has been renamed to Self.
Swift now supports unions. Unlike C unions, Swift's union is type-safe and always knows what type it contains at runtime. Union members are labeled using case declarations; each case may have a different set of types or no type:
union MaybeInt { case Some(Int) case None } union HTMLTag { case A(href:String) case IMG(src:String, alt:String) case BR }
Each case with a type defines a static constructor function for the union type. case declarations without types become static members:
var br = HTMLTag.BR var a = HTMLTag.A(href:"http://www.apple.com/") // 'HTMLTag' scope deduced for '.IMG' from context var img : HTMLTag = .IMG(src:"http://www.apple.com/mac-pro.png", alt:"The new Mac Pro")
Cases can be pattern-matched using switch:
switch tag { case .BR: println("<br>") case .IMG(var src, var alt): println("<img src=\"\(escape(src))\" alt=\"\(escape(alt))\">") case .A(var href): println("<a href=\"\(escape(href))\">") }
Due to implementation limitations, recursive unions are not yet supported.
Swift now supports autolinking, so importing frameworks or Swift libraries should no longer require adding linker flags or modifying your project file.
Swift now supports weak references by applying the [weak] attribute to a variable declaration.
(swift) var x = NSObject() // x : NSObject = <NSObject: 0x7f95d5804690> (swift) var [weak] w = x // w : NSObject = <NSObject: 0x7f95d5804690> (swift) w == nil // r2 : Bool = false (swift) x = NSObject() (swift) w == nil // r3 : Bool = true
Swift also supports a special form of weak reference, called [unowned], for references that should never be nil but are required to be weak to break cycles, such as parent or sibling references. Accessing an [unowned] reference asserts that the reference is still valid and implicitly promotes the loaded reference to a strong reference, so it does not need to be loaded and checked for nullness before use like a true [weak] reference.
class Parent { var children : Array<Child> func addChild(c:Child) { c.parent = this children.append(c) } } class Child { var [unowned] parent : Parent }
Numeric literals can now use underscores as separators. For example:
var billion = 1_000_000_000 var crore = 1_00_00_000 var MAXINT = 0x7FFF_FFFF_FFFF_FFFF var SMALLEST_DENORM = 0x0.0000_0000_0000_1p-1022
Types conforming to protocols now must always declare the conformance in their inheritance clause.
The build process now produces serialized modules for the standard library, greatly improving build times.
Arithmetic operators +, -, *, and / on integer types now do overflow checking and trap on overflow. A parallel set of masking operators, &+, &-, &*, and &/, are defined to perform two's complement wrapping arithmetic for all signed and unsigned integer types.
Debugger support. Swift has a -g command line switch that turns on debug info for the compiled output. Using the standard lldb debugger this will allow single-stepping through Swift programs, printing backtraces, and navigating through stack frames; all in sync with the corresponding Swift source code. An unmodified lldb cannot inspect any variables.
Example session:
$ echo 'println("Hello World")' >hello.swift
$ swift hello.swift -c -g -o hello.o
$ ld hello.o "-dynamic" "-arch" "x86_64" "-macosx_version_min" "10.9.0" \
-framework Foundation lib/swift/libswift_stdlib_core.dylib \
lib/swift/libswift_stdlib_posix.dylib -lSystem -o hello
$ lldb hello
Current executable set to 'hello' (x86_64).
(lldb) b top_level_code
Breakpoint 1: where = hello`top_level_code + 26 at hello.swift:1, addre...
(lldb) r
Process 38592 launched: 'hello' (x86_64)
Process 38592 stopped
* thread #1: tid = 0x1599fb, 0x0000000100000f2a hello`top_level_code + ...
frame #0: 0x0000000100000f2a hello`top_level_code + 26 at hello.shi...
-> 1 println("Hello World")
(lldb) bt
* thread #1: tid = 0x1599fb, 0x0000000100000f2a hello`top_level_code + ...
frame #0: 0x0000000100000f2a hello`top_level_code + 26 at hello.shi...
frame #1: 0x0000000100000f5c hello`main + 28
frame #2: 0x00007fff918605fd libdyld.dylib`start + 1
frame #3: 0x00007fff918605fd libdyld.dylib`start + 1
Also try s, n, up, down.
Swift now has a switch statement, supporting pattern matching of multiple values with variable bindings, guard expressions, and range comparisons. For example:
func classifyPoint(point:(Int, Int)) { switch point { case (0, 0): println("origin") case (_, 0): println("on the x axis") case (0, _): println("on the y axis") case (var x, var y) where x == y: println("on the y = x diagonal") case (var x, var y) where -x == y: println("on the y = -x diagonal") case (-10..10, -10..10): println("close to the origin") case (var x, var y): println("length \(sqrt(x*x + y*y))") } }
Swift has a new closure syntax. The new syntax eliminates the use of pipes. Instead, the closure signature is written the same way as a function type and is separated from the body by the in keyword. For example:
sort(fruits) { (lhs : String, rhs : String) -> Bool in return lhs > rhs }
When the types are omitted, one can also omit the parentheses, e.g.,
sort(fruits) { lhs, rhs in lhs > rhs }
Closures with no parameters or that use the anonymous parameters ($0, $1, etc.) don't need the in, e.g.,
sort(fruits) { $0 > $1 }
nil can now be used without explicit casting. Previously, nil had type NSObject, so one would have to write (e.g.) nil as! NSArray to create a nil NSArray. Now, nil picks up the type of its context.
POSIX.EnvironmentVariables and swift.CommandLineArguments global variables were merged into a swift.Process variable. Now you can access command line arguments with Process.arguments. In order to access environment variables add import POSIX and use Process.environmentVariables.