Generic parameter constraint on class crashes the compiler - swift

This is a simplified form of some swift3 code:
class GenericListViewModel<CellViewModel> {
let cells: [CellViewModel]
required init(cells: [CellViewModel]) {
self.cells = cells
}
}
class ViewController<CellViewModel, ListViewModel: GenericListViewModel<CellViewModel>> {
var viewModel: ListViewModel
init(cellViewModels: [CellViewModel]) {
viewModel = ListViewModel(cells: cellViewModels)
}
}
The compiler crashes with the following error:
While emitting IR SIL function #_TFC4Xxxx14ViewControllercfT14cellViewModelsGSax__GS0_xq__ for 'init'
at /.../GenericStuff.swift:22:5
Am is missing something, or is this a Swift compiler bug?
Edit:
I reported this here https://bugs.swift.org/browse/SR-3315 and it looks like it's fixed in current swift master branch.

You're pushing the system too hard with inheritance. Generics based on subclasses of other generics tends to break the compiler's brain, and is usually not really what you meant anyway. (That said: there is never an excuse for the compiler crashing, so you should absolutely open a bugreport.)
Do you really mean to subclass GenericListViewModel and then parameterize ViewController on that precise subclass? This seems very over-complicated and I'm not seeing how you would get any actual value out of it (since you can't rely on any additional methods added to your subclasses, and you already have dynamic dispatch). You're using both subclasses and generics to solve the same problem.
What you likely mean is that there's a CellViewModel and you want GenericListViewModel<CellViewModel> to wrap that, and aren't thinking about subclasses at all.
So, assuming you don't really mean to parameterize this specifically, let inheritance do its job. ListViewModel should be typealias, not a type parameter:
class ViewController<CellViewModel> {
typealias ListViewModel = GenericListViewModel<CellViewModel>
var viewModel: ListViewModel
init(cellViewModels: [CellViewModel]) {
viewModel = ListViewModel(cells: cellViewModels)
}
}
Now it's fine. That said, do you really need the view model to be a reference type? View models often don't need identity themselves (unless you're observing them with KVO). They may wrap a reference type, but as an adapter, a value type is often fine. Assuming this is true for you, then this can and should be simplified to a struct:
struct GenericListViewModel<CellViewModel> {
let cells: [CellViewModel]
}
class ViewController<CellViewModel> {
typealias ListViewModel = GenericListViewModel<CellViewModel>
var viewModel: ListViewModel
init(cellViewModels: [CellViewModel]) {
viewModel = ListViewModel(cells: cellViewModels)
}
}
To your goals of "custom logic like filtering the model, or keeping some other state specific to each controller," I would be very careful of using subclasses for this. It sounds like you're tempted to mix too much functionality into a single type. First, think about how you'd call your code the way you're thinking about it. ListViewModel isn't constrained by the init call, so you can't use type-inference. You'll have to initialize it like:
let vc: ViewController<SomeCellModel, GenericListViewModelSubclass<SomeCellModel>> = ViewController(cells: cells)
That's pretty hideous and is fighting all the things Swift wants to help you with. Since you want to be able to pass in the ListViewModel type, let's just pass it in. This is what protocols are for, not classes.
protocol CellViewModelProviding {
associatedtype CellViewModel
var cells: [CellViewModel] { get }
}
class ViewController<ListViewModel: CellViewModelProviding> {
var viewModel: ListViewModel
init(listViewModel: ListViewModel) {
viewModel = listViewModel
}
}
Now we can create different providers.
// A more standard name for your GenericListViewModel
struct AnyListViewModel<CellViewModel>: CellViewModelProviding {
let cells: [CellViewModel]
}
struct FilteredListViewModel<CellViewModel>: CellViewModelProviding {
var cells: [CellViewModel] {
return unfilteredCells.filter(predicate)
}
var unfilteredCells: [CellViewModel]
var predicate: (CellViewModel) -> Bool
}
Now we can use it with:
let vc = ViewController(listViewModel: AnyListViewModel(cells: [1,2,3]))
let vc2 = ViewController(listViewModel: FilteredListViewModel(unfilteredCells: [1,2,3],
predicate: { $0 % 2 == 0 }))
So that's pretty nice, but we could do better. It's kind of annoying to have to wrap our cells up in an AnyListViewModel in the normal case. We could probably create a factory method to get around this, but yuck. The better answer is to make use of the power of AnyListViewModel to be a type eraser. This is going to get a little more advanced, so if you're happy with the above solution, you can stop, but let's walk through it because it's really powerful and flexible if you need it.
First, we convert AnyListViewModel into a full type eraser that can accept either another view list model, or just an array.
struct AnyListViewModel<CellViewModel>: CellViewModelProviding {
private let _cells: () -> [CellViewModel]
var cells: [CellViewModel] { return _cells() }
init(cells: [CellViewModel]) {
_cells = { cells }
}
init<ListViewModel: CellViewModelProviding>(_ listViewModel: ListViewModel)
where ListViewModel.CellViewModel == CellViewModel {
_cells = { listViewModel.cells }
}
}
Now ViewController doesn't have to care what kind of ListViewModel is passed. It can turn anything into an AnyListViewModel and work with that.
class ViewController<CellViewModel> {
var viewModel: AnyListViewModel<CellViewModel>
init<ListViewModel: CellViewModelProviding>(listViewModel: ListViewModel)
where ListViewModel.CellViewModel == CellViewModel {
viewModel = AnyListViewModel(listViewModel)
}
init(cells: [CellViewModel]) {
viewModel = AnyListViewModel(cells: cells)
}
}
OK, that's cool, but it's not a huge improvement. Well, let's rebuild FilteredListViewModel and see what that gets us.
struct FilteredListViewModel<CellViewModel>: CellViewModelProviding {
var cells: [CellViewModel] {
return listViewModel.cells.filter(predicate)
}
private var listViewModel: AnyListViewModel<CellViewModel>
var predicate: (CellViewModel) -> Bool
// We can lift any other listViewModel
init<ListViewModel: CellViewModelProviding>(filtering listViewModel: ListViewModel,
withPredicate predicate: #escaping (CellViewModel) -> Bool)
where ListViewModel.CellViewModel == CellViewModel {
self.listViewModel = AnyListViewModel(listViewModel)
self.predicate = predicate
}
// Or, just for convenience, we can handle the simple [cell] case
init(filtering cells: [CellViewModel], withPredicate predicate: #escaping (CellViewModel) -> Bool) {
self.init(filtering: AnyListViewModel(cells: cells), withPredicate: predicate)
}
}
This is where things get powerful. We've said FilteredListViewModel can take some cells and filter them, sure. But it can also filter any other view list model.
let someList = AnyListViewModel(cells: [1,2,3])
let evenList = FilteredListViewModel(filtering: someList, withPredicate: { $0 % 2 == 0 })
So now you can chain things together. You could glue together filtering with sorting or something that modified the cells or whatever. You don't need one super-specialized subclass that does everything you need. You can click together simpler pieces to build complex solutions.

Related

Swift memory conflict where it should not happen

I am working on a SwiftUI project, where I use the MVVM-architecture.
When changing a View-model object property from the SwiftUI view, it causes a memory conflict crash in the view-model object.
The error is of the type: Simultaneous accesses to 0x600003591b48, but modification requires exclusive access.
In steps, here is what happens:
View-model property is changed from view
View-model property changes model property
Model property notifies about changes
View-model receives change notification
View-model access model object
Crash occur due to memory conflict
Relevant code snippets are seen below. Xcode project is a standard SwiftUI project.
The error will happen, after first clicking the add button, and then the modify button.
If the "update" code is moved into the "receiveValue" closure, the error will not occur. Likewise, the error will not occur, if the View-model class is made non-generic.
To my best knowledge, the code is all-right, so I suspect it is a compiler problem. But I am not sure.
import Foundation
import SwiftUI
import Combine
struct ContentView: View {
#ObservedObject var item: ViewModel<Model> = ViewModel<Model>()
var body: some View {
VStack {
Button("Add", action: { item.add(model:Model()) })
Button("Modify", action: { item.selected.toggle() })
}
}
}
protocol ModelType {
var objectDidChange: ObservableObjectPublisher { get }
var selected: Bool { get set }
}
class Model: ModelType {
let objectDidChange = ObservableObjectPublisher()
var selected = false {
didSet {
objectDidChange.send()
}
}
}
class ViewModel<Model:ModelType>: ObservableObject {
var selected = false {
didSet {
model.selected = selected
}
}
func add(model: Model) {
self.model = model
cancellable = model.objectDidChange.sink(receiveValue: { _ in
self.update()
})
}
private var model: Model! = nil
private var cancellable: AnyCancellable? = nil
func update() {
// Crash log: Simultaneous accesses to 0x600003591b48, but modification requires exclusive access.
print("update \(model.selected)")
}
}
Short version: require AnyObject for ModelType.
Long version:
You're trying to read from self.model while you're in the middle of setting self.model. When you say "If the "update" code is moved into the "receiveValue" closure, the error will not occur," this isn't quite correct. I expect what you mean is you wrote this:
cancellable = model.objectDidChange.sink(receiveValue: { _ in
print("update \(model.selected)")
})
And that worked, but that's completely different code. model in this case is the local variable, not the property self.model. You'll get the same crash if you write it this way:
cancellable = model.objectDidChange.sink(receiveValue: { _ in
print("update \(self.model.selected)")
})
The path that gets you here is:
ViewModel.selected.didSet
WRITE to Model.selected <---
Model.selected.didSet
(observer-closure)
ViewModel.update
READ from ViewModel.model <---
This is a read and write to the same value, and that violates exclusive access. Note that the "value" in question is "the entire ViewModel value," not ViewModel.selected. You can show this by changing the update function to:
print("update \(model!)")
You'll get the same crash.
So why does this work when you take out the generic? Because this particularly strict version of exclusivity only applies to value types (like structs). It doesn't apply to classes. So when this is concrete, Swift knows viewModel is a class, and that's ok. (The why behind this difference a bit complex, so I suggest reading the proposal that explains it.)
When you make this generic, Swift has to be very cautious. It doesn't know that Model is a class, so it applies stricter rules. You can fix this by promising that it's a class:
protocol ModelType: AnyObject { ... }

Referring to properties of containing class when using internal structs in swift

I'm refactoring a project to use MVVM and using protocols to ensure that my view models have a consistent structure. This works fine for defining public properties relating to input and output (which are based on internal structs) but defining actions in the same way is proving problemmatic as, currently, they are defined as closures which have to refer to view model properties. If I use the same approach as I have to input and output properties, I don't think I can access properties of the containing instance.
Example:
protocol ViewModelType {
associatedtype Input
associatedtype Output
associatedtype Action
}
final class MyViewModel: ViewModelType {
struct Input { var test: String }
struct Output { var result: String }
struct Action {
lazy var createMyAction: Action<String, Void> = { ... closure to generate Action which uses a MyViewModel property }
}
var input: Input
var output: Output
var action: Action
}
It's not a deal breaker if I can't do it, but I was curious as I can't see any way of getting access to the parent's properties.
Answer to your question
Let's begin with a note that createMyAction: Action<String, Void> refers to the type (struct) named Action as if it was a generic, but you have not declared it as such and will thus not work.
And to answer your question of the nested struct Action can refer its outer class MyViewModel - yes you can refer static properties, like this:
struct Foo {
struct Bar {
let biz = Foo.buz
}
static let buz = "buz"
}
let foobar = Foo.Bar()
print(foobar.biz)
But you should probably avoid such circular references. And I will omit any ugly hack that might be able to achive such a circular reference on non static properties (would probably involve mutable optional types). It is a code smell.
Suggestion for MVVM
Sounds like you would like to declare Action as a function? I'm using this protocol myself:
protocol ViewModelType {
associatedtype Input
associatedtype Output
func transform(input: Input) -> Output
}
Originally inspired by SergDort's CleanArchitecture.
You can prepare an instance of input (containing Observables) from the UIViewController and call the transform function and then map the Output of transform (being Observabless) to update the GUI.
So this code assumes you have basic Reactive knowledge. As for Observables you can chose between RxSwift or ReactiveSwift - yes their names are similar.
If you are comfortable with Rx, it is an excellent way of achieving a nice MVVM architecture with simple async updates of the GUI. In the example below, you will find the type Driver which is documented here, but the short explanation is that is what you want to use for input from views and input to views, since it updates the views on the GUI thread and it is guaranteed to not error out.
CleanArchitecture contains e.g. PostsViewModel :
final class PostsViewModel: ViewModelType {
struct Input {
let trigger: Driver<Void>
let createPostTrigger: Driver<Void>
let selection: Driver<IndexPath>
}
struct Output {
let fetching: Driver<Bool>
let posts: Driver<[PostItemViewModel]>
let createPost: Driver<Void>
let selectedPost: Driver<Post>
let error: Driver<Error>
}
private let useCase: PostsUseCase
private let navigator: PostsNavigator
init(useCase: PostsUseCase, navigator: PostsNavigator) {
self.useCase = useCase
self.navigator = navigator
}
func transform(input: Input) -> Output {
let activityIndicator = ActivityIndicator()
let errorTracker = ErrorTracker()
let posts = input.trigger.flatMapLatest {
return self.useCase.posts()
.trackActivity(activityIndicator)
.trackError(errorTracker)
.asDriverOnErrorJustComplete()
.map { $0.map { PostItemViewModel(with: $0) } }
}
let fetching = activityIndicator.asDriver()
let errors = errorTracker.asDriver()
let selectedPost = input.selection
.withLatestFrom(posts) { (indexPath, posts) -> Post in
return posts[indexPath.row].post
}
.do(onNext: navigator.toPost)
let createPost = input.createPostTrigger
.do(onNext: navigator.toCreatePost)
return Output(fetching: fetching,
posts: posts,
createPost: createPost,
selectedPost: selectedPost,
error: errors)
}
}

Passing a value type as reference in Swift

I have my model implemented as structs in Swift 3.0. Several of these structs have delegates that should be able to modify the model depending on the user's actions.
However, when I pass the struct to the delegate method, it gets copied.
How do you solve this? Can you force the compiler to pass this struct as a reference, or the only option is to use a class?
structs are always passed by value. The whole point of using a struct is to have it behave as a value type. If you need delegation (which usually implies mutable state), you should be using a class.
If you really really need to, you could force pass-by-reference by using an inout parameter, but that is not recommended in general. You could also use a box type to simulate passing by reference. But, in general, you should just use a class if you need reference behavior.
The whole point of using struct in the first place is that this is desirable behavior. It preserves the immutability of the data. inout can achieve this, but it's not recommended in the general case.
protocol Delegate {
func callback(_ oldValue: Int) -> Int
}
struct IncrementerDelegate: Delegate {
let step: Int
func callback(_ oldValue: Int) -> Int {
return oldValue + step
}
}
struct Model {
var i = 0
}
class Controller {
var model = Model()
var delegate: Delegate
init(delegate: Delegate) {
self.delegate = delegate
}
// To be called externally, such as by a button
func doSomething() {
// Delegate determains new value, but it's still the
// controller's job to perform the mutation.
model.i = delegate.callback(model.i)
}
}
let delegate = IncrementerDelegate(step: 5)
let controller = Controller(delegate: delegate)
print(controller.model.i)
controller.doSomething() // simulate button press
print(controller.model.i)
protocol CrappyDelegate {
func callback(_ model: inout Model)
}
struct CrappyIncrementerDelegate: CrappyDelegate {
let step: Int
func callback(_ model: inout Model) {
model.i = 9999999
// Just hijacked the models value,
// and the controller can't do anything about it.
}
}
class VulnerableController {
var model = Model()
var delegate: CrappyDelegate
init(delegate: CrappyDelegate) {
self.delegate = delegate
}
// To be called externally, such as by a button
func doSomething() {
// Controller leaks entire model, and has no control over what happens to it
delegate.callback(&model)
}
}
let crappyDelegate = CrappyIncrementerDelegate(step: 5)
let vulnerableController = VulnerableController(delegate: crappyDelegate)
print(controller.model.i)
controller.doSomething() // simulate button press
print(controller.model.i) // model hijacked
If you want to pass by reference, you should generally use a class not a struct.
https://developer.apple.com/library/content/documentation/Swift/Conceptual/Swift_Programming_Language/ClassesAndStructures.html states:
You can use both classes and structures to define custom data types to
use as the building blocks of your program’s code.
However, structure instances are always passed by value, and class
instances are always passed by reference. This means that they are
suited to different kinds of tasks. As you consider the data
constructs and functionality that you need for a project, decide
whether each data construct should be defined as a class or as a
structure.

Simple observable struct with RxSwift?

I'm trying to come up with a simple observable object in Swift and thought to use RxSwift. I couldn't find a simple example to do something like this:
protocol PropertyObservable {
typealias PropertyType
var propertyChanged: Event<(PropertyType, Any)> { get }
}
class Car: PropertyObservable {
typealias PropertyType = CarProperty
let propertyChanged = Event<(CarProperty, Any)>()
dynamic var miles: Int = 0 {
didSet {
propertyChanged.raise(.Miles, oldValue as Any)
}
}
dynamic var name: String = "Turbo" {
didSet {
propertyChanged.raise(.Name, oldValue as Any)
}
}
}
The above is pure Swift solution for observables from this blog post; I really like how it's a protocol-based solution and not invasive. In my case, I have an object in my project where each property is set asynchronously under the hood (bluetooth device). So I need to observe/subscribe to the changes instead of getting/setting the properties in real-time.
I keep hearing RxSwift will do just that and more. However, I can't find a simple example to match above and beginning to think RxSwift is overkill for my need? Thanks for any help.
Easiest way to quickly make this observable with RxSwift would probably be to use the RxSwift class Variable (all code here is untested off the top of my head):
import RxSwift
class Car {
var miles = Variable<Int>(0)
var name = Variable<String>("Turbo")
}
This enables you to observe the values by subscribing to them:
let disposeBag = DisposeBag()
let car = Car
car.name.asObservable()
.subscribeNext { name in print("Car name changed to \(name)") }
.addToDisposeBag(disposeBag) // Make sure the subscription disappears at some point.
Now you've lost the old value in each event. There are of course numerous ways to solve this, the RxSwifty way would probably be to add a scan operation to your element sequence, which works a lot like reduce does on a normal Array:
car.name.asObservable()
.scan(seed: ("", car.name.value)) { (lastEvent, newElement) in
let (_, oldElement) = lastEvent
return (oldElement, newElement)
}
.subscribeNext { (old, new) in print("Car name changed from \(old) to \(new)") }
.addToDisposeBag(disposeBag)

EXC_BAD_ACCESS using protocol composition

I want to configure an object with multiple presentables; I've created protocols for them; so that I can combine them into a concrete presentable.
protocol Presentable {
}
protocol TextPresentable: Presentable {
var text:String { get }
}
protocol ImagePresentable: Presentable {
var image:String { get }
var images:[String] { get }
}
The concrete struct:
struct ConcretePresentable: TextPresentable, ImagePresentable {
var text:String { return "Text" }
var image:String { return "Image" }
var images:[String] { return ["A", "B"] }
}
The next thing would be to have a presenter, and I would check if the passed in presenter is actually valid:
typealias TextAndImagePresentable = protocol<TextPresentable, ImagePresentable>
struct ConcretePresenter {
func configureWithPresentable(presentable: Presentable) {
guard let textAndImagePresentable = presentable as? TextAndImagePresentable else {
return
}
print(textAndImagePresentable.text)
print(textAndImagePresentable.image)
print(textAndImagePresentable.images)
}
}
To configure:
let concretePresentable = ConcretePresentable()
let concretePresenter = ConcretePresenter()
concretePresenter.configureWithPresentable(concretePresentable)
It goes fine if I run this in a playground, with all of the code in the same place. However, once I put this in a project and split it up into multiple files (ConcretePresenter.swift, ConcretePresentable.swift holding the concrete structs and Presentable.swift which holds the protocols), I get a EXC_BAD_ACCESS.
Why does that happen?
As a disclaimer, I don't necessarily find this answer very satisfying, but it does work.
So, once I noticed that the text & image properties were being returned in each others places (the value for text is being returned by the image property and vice versa), I figured the problem had something to do with what Swift is doing with managing pointers here.
So, out of curiosity, I wanted to add a truly scalar value to the protocols. I added a value property as an Int to the TextPresentable protocol:
protocol Presentable {}
protocol TextPresentable: Presentable {
var text:String { get }
var value: Int { get }
}
protocol ImagePresentable: Presentable {
var image:String { get }
var images:[String] { get }
}
And then I set up the concrete implementation to return some known value. Here, we're returning 0.
struct ConcretePresentable: TextPresentable, ImagePresentable {
var text:String { return "SomeText" }
var value: Int { return 0 }
var image:String { return "SomeImage" }
var images:[String] { return ["A", "B"] }
}
After running this code, we still get the same crash, but I notice that value, which really shouldn't have a problem printing 0 is instead printing some very large number: 4331676336. This isn't right at all.
I also changed images from an array to a dictionary to see if the error persists--it does. It seems the crash is related to collections and not specific to arrays.
From here, I tried some other things.
I tried making ConcretePresentable a class rather than a struct.
class ConcretePresentable: TextPresentable, ImagePresentable
That resulted in the same behavior.
I tried making ConcretePresentable conform to the typealias rather than the protocols independently:
struct ConcretePresentable: TextAndImagePresentable
That resulted in the same behavior.
I tried doing both of the aforementioned at once:
class ConcretePresentable: TextAndImagePresentable
Yet still the same behavior.
I did come up with one way to make it work though. Make a protocol that conforms to the two protocols in your typealias and make ConcretePresentable conform to that:
protocol TextAndImagePresentable: TextPresentable, ImagePresentable {}
struct ConcretePresentable: TextAndImagePresentable {
// ...
}
The problem here is that if you don't explicitly make ConcretePresentable conform to the protocol, it will fail the guard let even if it does conform to TextPresentable and ImagePresentable.
I asked about this on Swift Users and filled a bug.
Confirmed to be a bug in the compiler:
https://bugs.swift.org/browse/SR-4477
Fixed by Joe Groff now:
Joe Groff added a comment - 2 hours ago
Merged. Should be fixed in future snapshots.
struct uses value semantics and so properties are copied. Swift should have reported this as an error since you are trying to inherit from two protocols which derive from the same base protocol. In classes this will work but in struct it wont because of value semantics for struct. In case you decide to add a variable to Presentable protocol Swift would be confused which ones to bring into the struct. From TextPresentable or ImagePresentable
You should use #protocol Presentable : class and then convert ConcretePresentable to class to fix this.
protocol Presentable : class {
}
class ConcretePresenter {...
The problem is in the cast of the Presentable to TextAndImagePresentable. The guard let succeed, but creates an invalid value (I don't know exactly why).
One way to check it, is look to the console on the execution of the commands:
print(textAndImagePresentable.text)
print(textAndImagePresentable.image)
print(textAndImagePresentable.images)
That will print:
Image
Text
Program ended with exit code: 9
One way to avoid it is to change your method signature to avoid the casting:
func configureWithPresentable(presentable: TextAndImagePresentable) {
print(presentable.text)
print(presentable.image)
print(presentable.images)
}
And, in my opinion, since nothing will happen if the presentable do not conform to both protocols, makes more sense to delimiter it on the method signature.