Swift 4 way to fake existentials with protocols and `Self` - swift

There are hundreds of solutions to fake existentials in Swift with protocols and Self, but they mostly refer to Swift 2 and the bright future that Swift 3 might bring...
Now Swift 4 is out, with nice additions for Generics. But I couldn't find any suggestions how to fit that into the missing existentials problem.
Any Ideas how to solve this the Swift 4 way?
Example:
import UIKit
protocol Bla {
func compare(other: Self)
}
extension CGFloat : Bla {
func compare(other: CGFloat) {
print("Extended CGFloat")
}
}
extension UIEdgeInsets : Bla {
func compare(other: UIEdgeInsets) {
print("Extended UIEdgeInsets")
}
}
/* Possible, but what if we want to CGFloat _and_ UIEdgeInsets inside here?
Well, that would _not_ work! */
class Possible<T: Bla> {
var array: [T]!
}
/* This is open to everything...
And therefore fails to compile, no dynamic type info at runtime I guess. */
class Fail {
var array: [Bla]!
}
// Works, but I don't need that.
let list = Possible<CGFloat>()
// I need that:
let list = Fail()
let f: CGFloat = 1.23
let edge = UIEdgeInsets()
list.array.append(f)
list.array.append(edge)

Fundamentally, it can't be done. If you could have:
class Fail {
var array: [Bla]!
}
You could then try to write code like this:
func compareAll(foo: Fail)
{
for x in foo.array
{
x.compare(other: y)
}
}
What type is y? The protocol says it has to be the same type as the object adopting the protocol but you don't know the type of x until runtime. There is no way to write that code with y being both a UIEdgeInsets and a CGFloat at the same time.
I think you could make it work by removing the dependency on Self by making compare generic. Your protocol would look like this:
protocol Bla {
func compare<T: Bla>(other: T)
}
Implementations of compare would have to test the type of other and cast to the right type.
extension CGFloat: Bla
{
func compare<T: Bla>(other: T)
{
if let casted = other as? CGFloat
{
// do whatever
}
}
}
I think this approach is better than using type erasure (see Daniel Hall's answer) for a couple of reasons:
There's no overhead of having to wrap everything in a wrapper object or the indirection of compare.
The compare function can work when other is an arbitrary type, not just the same type of self if it makes sense e.g.
func compare<T: Bla>(other: T)
{
if let casted = other as? CGFloat
{
// do whatever
}
else if let casted = other as? UIEdgeInsets
{
// Do something else
}
}
Of course, if you are simply given the protocol and you can't change it, type erasure is your only option.

This is typically handled with a type eraser, for example:
struct AnyBla: Bla {
private let compareClosure: (Any) -> ()
func compare(other: AnyBla) {
compareClosure(other)
}
init<T: Bla>(_ bla: T) {
compareClosure = {
if let other = $0 as? T {
bla.compare(other: other)
} else {
print("Can't compare type \(type(of: bla)) with \(type(of: $0))")
}
}
}
}
Then you would change your array to any array of the type eraser, i.e.
class Fail {
var array: [AnyBla]!
}
Then you can accomplish what you want to do like this:
let list = Fail()
let f: CGFloat = 1.23
let edge = UIEdgeInsets()
list.array.append(AnyBla(f))
list.array.append(AnyBla(edge))

Swift 4 didn't make any generics changes that address this issue. For the full roadmap, see the Generics Manifesto. That said, the protocol you're describing is equivalent to Equatable, and so would have all the same problems. There is no current roadmap to create what you're describing. The closest is discussed at the very end of the Manifesto, in the section called "Opening existentials."
As you've described the problem, it isn't clear how it would be useful, so I suspect your actual problem is different. You should go back to your actual program and consider what its real requirements are. For example, if your real requirement is "an array of CGFloat or UIEdgeInsets" that is absolutely buildable in Swift. "An array" translates to [T], and "OR" translates to enum (Swift's "sum type" in type theory parlance). So what you mean is:
enum Element {
case scalar(CGFloat)
case insets(UIEdgeInsets)
}
let list: [Element]
That is completely different than "an array of things that can be compared to themselves." If, on the other hand, you really mean "an array of value-like things," then that also is pretty easy to build in Swift with NSValue (and its subclass, NSNumber):
class Things {
var array: [NSValue] = []
func append(_ value: CGFloat) {
array.append(NSNumber(value: Double(value)))
}
func append(_ insets: UIEdgeInsets) {
array.append(NSValue(uiEdgeInsets: insets))
}
}
let list = Things()
let f: CGFloat = 1.23
let edge = UIEdgeInsets()
list.append(f)
list.append(edge)
list.array[0] == list.array[0] // true
list.array[0] == list.array[1] // false

Related

Swift Generics and Protocols

Given the below, this will throw a compile time error about the protocol not being able to adhere to itself and only struct/enum can adhere to the protocol. This seems to defeat the purpose of being able to use protocols in generics. I'm trying to understand why this doesn't work, but if I remove the generic and just put the protocol where 'Z' is everything is fine. It seems antithetical to what protocols and generics should be allowed for.
**Edit for question clarity: I need to take a type of Any that can be cast to a dictionary of [String:MyProtocol] and pass it into the method printEm. printEm must use the generic as it will be instantiating instances of the class Z.
protocol MyProtocol {
init()
var whoAmI:String { get }
}
func genericPassing(unknownThing:Any) {
let knownThing = unknownThing as? [String:MyProtocol]
if(knownThing != nil){
self.printEm(knownThing)
}
}
func printEm<Z:MyProtocol>(theThings:[String:Z]) {
let zCollection:[Z] = []
for thing in theThings {
print(thing.whoAmI)
zCollection.append(Z())
}
}
**Edited printEm to illustrate why generic is needed.
** Edit for more complex code. The two primary requirements are the use of a generic to call Z() and the ability to take an Any and somehow type check it and/or cast it so that it can be used in a genericized method.
private func mergeUpdates<Z:RemoteDataSyncable>(source:inout Z, updates:[WritableKeyPath<Z, Any>:Any]) throws {
for key in updates.keys {
let value = updates[key]!
let valueDict = value as? [String:[WritableKeyPath<RemoteDataSyncable, Any>:Any]]
if(valueDict != nil) {
var currentValueArray = source[keyPath: key] as? [RemoteDataSyncable]
if(currentValueArray != nil) {
self.mergeUpdates(source: &currentValueArray!, updates: valueDict!)
}
else {
throw SyncError.TypeError
}
}
else {
source[keyPath: key] = value
}
}
}
private func mergeUpdates<Z:RemoteDataSyncable>(source:inout [Z], updates:[String:[WritableKeyPath<Z,Any>:Any]]) {
for key in updates.keys {
var currentObject = source.first { syncable -> Bool in
return syncable.identifier == key
}
if(currentObject != nil) {
try! self.mergeUpdates(source: &currentObject!, updates: updates[key]!)
}
else {
var newSyncable = Z()
try! self.mergeUpdates(source: &newSyncable, updates: updates[key]!)
source.append(newSyncable)
}
}
}
This is a perfect example of why protocols do not conform to themselves. In your code Z is MyProtocol, so Z() is MyProtocol(). How would that work? What type is it? How big is it? You then can't put them into a [Z], since they might be different types.
You mean to pass arbitrary MyProtocols and call init on the type of each element:
func printEm(theThings:[String: MyProtocol]) {
var zCollection:[MyProtocol] = []
for thing in theThings.values {
print(thing.whoAmI)
zCollection.append(type(of: thing).init())
}
}
When I suggested using closures, this is the kind of thing I mean. This Updater can accept arbitrary ReferenceWritableKeyPaths, and when you pass it Any update value, it'll assign it to every key path that can accept it. That's kind of useless, but shows the technique. (Keep in mind that the updater is retaining the object, so that may be a problem that you need to address.)
class Updater {
private(set) var updaters: [(Any) -> ()] = []
func add<Root, Value>(keyPath: ReferenceWritableKeyPath<Root, Value>, on root: Root) {
updaters.append { value in
if let value = value as? Value {
root[keyPath: keyPath] = value
}
}
}
func updateAll(with value: Any) {
for updater in updaters {
updater(value)
}
}
}
class Client {
var updateMe: Int = 0
}
let client = Client()
let updater = Updater()
updater.add(keyPath: \.updateMe, on: client)
updater.updateAll(with: 3)
client.updateMe // 3
The key lesson in this code is that the generic types on the add are erased (hidden) by the (Any) -> () closure at compile-time. And the runtime as? check is done inside that closure, where the types are all known.

Using generic methods in protocols Swift

I believe I have some misunderstanding of how generics work. I have the protocol:
protocol CommandProtocol {
func execute<T>() -> T
func unExecute<T>() -> T
}
And a class that conforms to it:
class CalculatorCommand: CommandProtocol {
...
func execute<String>() -> String {
return calculator.performOperation(operator: `operator`, with: operand) as! String
}
func unExecute<Double>() -> Double {
return calculator.performOperation(operator: undo(operator: `operator`), with: operand) as! Double
}
...
}
The calculator.performOperation() method actually returns Double, but here I just try to play with generics so I replace return type from Double to String.
After that, I have a class which invokes these methods:
class Sender {
...
// MARK: - Public methods
func undo() -> Double {
if current > 0 {
current -= 1
let command = commands[current]
return command.unExecute()
}
return 0
}
func redo() -> Double? {
if current < commands.count {
let command = commands[current]
current += 1
let value: Double = command.execute()
print(type(of: value))
return command.execute()
}
return nil
}
...
}
In the undo() method everything works as expected (one thing that I did not understand fully is how Swift really knows whether the unExecute value will return Double or not, or compiler infers it based on the undo() return type?)
But in the redo() method, I am calling the execute() method which returns String, but the method expects Double, so I thought that my program would crash, but not, it works totally fine as if execute() method returns Double.
Please, could someone explain to me what exactly happens under the cover of this code? Thank you in advance.
You are correct that you misunderstand generics. First, let's look at this protocol:
protocol CommandProtocol {
func execute<T>() -> T
func unExecute<T>() -> T
}
This says "no matter what type the caller requests, this function will return that type." That's impossible to successfully implement (by "successfully" I mean "correctly returns a value in all cases without crashing"). According this protocol, I'm allowed to write the following code:
func run(command: CommandProtocol) -> MyCustomType {
let result: MyCustomType = command.execute()
return result
}
There's no way to write an execute that will actually do that, no matter what MyCustomType is.
Your confusion is compounded by a subtle syntax mistake:
func execute<String>() -> String {
This does not mean "T = String," which is what I think you expect it to mean. It creates a type variable called String (that has nothing to do with Swift's String type), and it promises to return it. when you later write as! String, that means "if this values isn't compatible with the type requested (not "a string" but whatever was requested by the caller), then crash.
The tool that behaves closer to what you want here is an associated type. You meant to write this:
protocol CommandProtocol {
associatedType T
func execute() -> T
func unExecute() -> T
}
But this almost certainly won't do what you want. For example, with that, it's impossible to have an array of commands.
Instead what you probably want is a struct:
struct Command {
let execute: () -> Void
let undo: () -> Void
}
You then make Commands by passing closures that do what you want:
let command = Command(execute: { self.value += 1 },
undo: { self.value -= 1 })
Alternately, since this is a calculator, you could do it this way:
struct Command {
let execute: (Double) -> Double
let undo: (Double) -> Double
}
let command = Command(execute: { $0 + 1 }, undo: { $0 - 1 })
Then your caller would look like:
value = command.execute(value)
value = command.undo(value)
You think this returns a Swift.Double, but no. This code is no different than using T instead of Double. Swift does not require the names of generic placeholders to match what you put in a protocol.
func unExecute<Double>() -> Double {
return calculator.performOperation(operator: undo(operator: `operator`), with: operand) as! Double
}
You're not actually looking for generic methods. You want this, instead.
protocol CommandProtocol {
associatedtype ExecuteValue
associatedtype UnExecuteValue
func execute() -> ExecuteValue
func unExecute() -> UnExecuteValue
}

Declare a Swift protocol which has a property return value CollectionType<Int>?

Is something like
protocol A {
var intCollection: CollectionType<Int> { get }
}
or
protocol A {
typealias T: CollectionType where T.Generator.Element == Int
var intCollection: T
}
possible in Swift 2.1?
Update for Swift 4
Swift 4 now support this feature! read more in here
Not as a nested protocol, but it's fairly straightforward using the type erasers (the "Any" structs).
protocol A {
var intCollection: AnyRandomAccessCollection<Int> { get }
}
This is actually often quite convenient for return values because the caller usually doesn't care so much about the actual type. You just have to throw a return AnyRandomAccessCollection(resultArray) at the end of your function and it all just works. Lots of stdlib now returns Any erasers. For the return value problem, it's almost always the way I recommend. It has the nice side effect of making A concrete, so it's much easier to work with.
If you want to keep the CollectionType, then you need to restrict it at the point that you create a function that needs it. For example:
protocol A {
typealias IntCollection: CollectionType
var intCollection: IntCollection { get }
}
extension A where IntCollection.Generator.Element == Int {
func sum() -> Int {
return intCollection.reduce(0, combine: +)
}
}
This isn't ideal, since it means you can have A with the wrong kind of collection type. They just won't have a sum method. You also will find yourself repeating that "where IntCollection.Generator.Element == Int" in a surprising number of places.
In my experience, it is seldom worth this effort, and you quickly come back to Arrays (which are the dominant CollectionType anyway). But when you need it, these are the two major approaches. That's the best we have today.
You can't do this upright as in your question, and there exists several thread here on SO on the subject of using protocols as type definitions, with content that itself contains Self or associated type requirements (result: this is not allowed). See e.g. the link provided by Christik, or thread Error using associated types and generics.
Now, for you example above, you could do the following workaround, however, perhaps mimicing the behaviour you're looking for
protocol A {
typealias MyCollectionType
typealias MyElementType
func getMyCollection() -> MyCollectionType
func printMyCollectionType()
func largestValue() -> MyElementType?
}
struct B<U: Comparable, T: CollectionType where T.Generator.Element == U>: A {
typealias MyCollectionType = T
typealias MyElementType = U
var myCollection : MyCollectionType
init(coll: MyCollectionType) {
myCollection = coll
}
func getMyCollection() -> MyCollectionType {
return myCollection
}
func printMyCollectionType() {
print(myCollection.dynamicType)
}
func largestValue() -> MyElementType? {
guard var largestSoFar = myCollection.first else {
return nil
}
for item in myCollection {
if item > largestSoFar {
largestSoFar = item
}
}
return largestSoFar
}
}
So you can implement blueprints for your generic collection types in you protocol A, and implement these blueprints in the "interface type" B, which also contain the actual collection as a member property. I have taken the largestValue() method above from here.
Example usage:
/* Examples */
var myArr = B<Int, Array<Int>>(coll: [1, 2, 3])
var mySet = B<Int, Set<Int>>(coll: [10, 20, 30])
var myRange = B<Int, Range<Int>>(coll: 5...10)
var myStrArr = B<String, Array<String>>(coll: ["a", "c", "b"])
myArr.printMyCollectionType() // Array<Int>
mySet.printMyCollectionType() // Set<Int>
myRange.printMyCollectionType() // Range<Int>
myStrArr.printMyCollectionType() // Array<String>
/* generic T type constrained to protocol 'A' */
func printLargestValue<T: A>(coll: T) {
print(coll.largestValue() ?? "Empty collection")
}
printLargestValue(myArr) // 3
printLargestValue(mySet) // 30
printLargestValue(myRange) // 10
printLargestValue(myStrArr) // c

How to constraint generic type to another generic type in Swift?

I'd like to do something like this:
class Config<T> {
func configure(x:T)
// constraint B to be subclass of A
class func apply<A,B:A>(c:Config<A>, to:B) {
c.configure(to)
}
}
So later, for example, I can apply a Config to a UILabel:
class RedViewConfig<T:UIView> : Config<T> {
func configure(x:T) {
x.backgroundColor = .redColor();
}
}
let label = UILabel()
Config.apply(RedViewConfig(), to:label)
Or extend Config classes:
class RedLabelConfig<T:UILabel> : RedViewConfig<T> {
func configure(x:T) {
super.configure(x)
x.textColor = .redColor();
}
}
Config.apply(RedLabelConfig(), to:label)
I tried to do it, but I couldn't constraint classes. So I tried with protocols and associated types, but when subclassing I found problems (like this) when overriding the associated type.
Do you actually need the generic parameter B? If your argument to: was typed as A as well, it could be any subtype of A. Like such:
class View {}
class LabelView : View {}
class Config<T> {
func configure(x:T) { print ("Configured: \(x)") }
}
func applyConfig<A> (c:Config<A>, to:A) {
c.configure(to)
}
applyConfig(Config<View>(), to: LabelView())
Classes make this way too complicated. Inheritance is almost always a bad idea in Swift if you can possibly avoid it.
Structs, though closer, still make this a bit over-complicated and restrictive.
Really, these configurators are just functions. They take a thing and they do something to it, returning nothing. They're just T -> Void. Let's build a few of those.
func RedViewConfig(view: UIView) { view.backgroundColor = .redColor() }
func VisibleConfig(view: UIView) { view.hidden = false }
And we can use them pretty easily:
let label = UILabel()
VisibleConfig(label)
We can compose them (like super, but without the baggage) if their types are compatible:
func RedLabelConfig(label: UILabel) {
RedViewConfig(label)
label.textColor = .redColor()
}
We can pass them around in data structures, and the compiler will apply the right covariance for us:
let configs = [RedLabelConfig, VisibleConfig]
// [UILabel -> ()]
// This has correctly typed visibleConfig as taking `UILabel`,
// even though visibleConfig takes `UIView`
// And we can apply them
for config in configs { config(label) }
Now if we want other syntaxes, we can build those pretty easily too. Something more like your original:
func applyConfig<T>(f: T -> Void, to: T) {
f(to)
}
applyConfig(VisibleConfig, to: label)
or even closer to your original:
struct Config {
static func apply<T>(config: T -> Void, to: T) { config(to) }
}
Config.apply(VisibleConfig, to: label)
The point is that just using functions here makes everything very flexible without adding any of the complexity of class inheritance or even structs.

Using a Type Variable in a Generic

I have this question except for Swift. How do I use a Type variable in a generic?
I tried this:
func intType() -> Int.Type {
return Int.self
}
func test() {
var t = self.intType()
var arr = Array<t>() // Error: "'t' is not a type". Uh... yeah, it is.
}
This didn't work either:
var arr = Array<t.Type>() // Error: "'t' is not a type"
var arr = Array<t.self>() // Swift doesn't seem to even understand this syntax at all.
Is there a way to do this? I get the feeling that Swift just doesn't support it and is giving me somewhat ambiguous error messages.
Edit: Here's a more complex example where the problem can't be circumvented using a generic function header. Of course it doesn't make sense, but I have a sensible use for this kind of functionality somewhere in my code and would rather post a clean example instead of my actual code:
func someTypes() -> [Any.Type] {
var ret = [Any.Type]()
for (var i = 0; i<rand()%10; i++) {
if (rand()%2 == 0){ ret.append(Int.self) }
else {ret.append(String.self) }
}
return ret
}
func test() {
var ts = self.someTypes()
for t in ts {
var arr = Array<t>()
}
}
Swift's static typing means the type of a variable must be known at compile time.
In the context of a generic function func foo<T>() { ... }, T looks like a variable, but its type is actually known at compile time based on where the function is called from. The behavior of Array<T>() depends on T, but this information is known at compile time.
When using protocols, Swift employs dynamic dispatch, so you can write Array<MyProtocol>(), and the array simply stores references to things which implement MyProtocol — so when you get something out of the array, you have access to all functions/variables/typealiases required by MyProtocol.
But if t is actually a variable of kind Any.Type, Array<t>() is meaningless since its type is actually not known at compile time. (Since Array is a generic struct, the compiler needs know which type to use as the generic parameter, but this is not possible.)
I would recommend watching some videos from WWDC this year:
Protocol-Oriented Programming in Swift
Building Better Apps with Value Types in Swift
I found this slide particularly helpful for understanding protocols and dynamic dispatch:
There is a way and it's called generics. You could do something like that.
class func foo() {
test(Int.self)
}
class func test<T>(t: T.Type) {
var arr = Array<T>()
}
You will need to hint the compiler at the type you want to specialize the function with, one way or another. Another way is with return param (discarded in that case):
class func foo() {
let _:Int = test()
}
class func test<T>() -> T {
var arr = Array<T>()
}
And using generics on a class (or struct) you don't need the extra param:
class Whatever<T> {
var array = [T]() // another way to init the array.
}
let we = Whatever<Int>()
jtbandes' answer - that you can't use your current approach because Swift is statically typed - is correct.
However, if you're willing to create a whitelist of allowable types in your array, for example in an enum, you can dynamically initialize different types at runtime.
First, create an enum of allowable types:
enum Types {
case Int
case String
}
Create an Example class. Implement your someTypes() function to use these enum values. (You could easily transform a JSON array of strings into an array of this enum.)
class Example {
func someTypes() -> [Types] {
var ret = [Types]()
for _ in 1...rand()%10 {
if (rand()%2 == 0){ ret.append(.Int) }
else {ret.append(.String) }
}
return ret
}
Now implement your test function, using switch to scope arr for each allowable type:
func test() {
let types = self.someTypes()
for type in types {
switch type {
case .Int:
var arr = [Int]()
arr += [4]
case .String:
var arr = [String]()
arr += ["hi"]
}
}
}
}
As you may know, you could alternatively declare arr as [Any] to mix types (the "heterogenous" case in jtbandes' answer):
var arr = [Any]()
for type in types {
switch type {
case .Int:
arr += [4]
case .String:
arr += ["hi"]
}
}
print(arr)
I would break it down with the things you already learned from the first answer. I took the liberty to refactor some code. Here it is:
func someTypes<T>(t: T.Type) -> [Any.Type] {
var ret = [Any.Type]()
for _ in 0..<rand()%10 {
if (rand()%2 == 0){ ret.append(T.self) }
else {
ret.append(String.self)
}
}
return ret
}
func makeArray<T>(t: T) -> [T] {
return [T]()
}
func test() {
let ts = someTypes(Int.self)
for t in ts {
print(t)
}
}
This is somewhat working but I believe the way of doing this is very unorthodox. Could you use reflection (mirroring) instead?
Its possible so long as you can provide "a hint" to the compiler about the type of... T. So in the example below one must use : String?.
func cast<T>(_ value: Any) -> T? {
return value as? T
}
let inputValue: Any = "this is a test"
let casted: String? = cast(inputValue)
print(casted) // Optional("this is a test")
print(type(of: casted)) // Optional<String>
Why Swift doesn't just allow us to let casted = cast<String>(inputValue) I'll never know.
One annoying scenerio is when your func has no return value. Then its not always straightford to provide the necessary "hint". Lets look at this example...
func asyncCast<T>(_ value: Any, completion: (T?) -> Void) {
completion(value as? T)
}
The following client code DOES NOT COMPILE. It gives a "Generic parameter 'T' could not be inferred" error.
let inputValue: Any = "this is a test"
asyncCast(inputValue) { casted in
print(casted)
print(type(of: casted))
}
But you can solve this by providing a "hint" to compiler as follows:
asyncCast(inputValue) { (casted: String?) in
print(casted) // Optional("this is a test")
print(type(of: casted)) // Optional<String>
}