Doug's Compiler Corner

Originally posted on 2024-02-10 20:00:00 +0000

Last updated on 2024-06-15 05:10:50 +0000

Swift for C++ Practitioners, Part 2: Reference Types & Optionals

In our first episode, we talked a lot about value types. In general, you should prefer value types in Swift whenever you can.

However, Object-Oriented Programming (OOP) is a reasonable approach when you need your objects to have identity and to form relationships amongst each other in arbitrary ways. For these cases, Swift has classes with support for inheritance, overriding, and subtyping. A class is defined in Swift with---you guessed it---the class keyword:

class Person {
  var name: String
  var age: Int
  
  init(name: String, age: Int) {
    self.name = name
    self.age = age
  }
  
  func greet() {
    print("Hello \(name)!")
  }
}

Classes look a lot like structs, with all the same basic ingredients: stored properties ("non-static data members" in C++ parlance), initializers, methods, etc. Each of these ingredients works basically the same way, so if you know how to write a struct, you're pretty close to writing a class. In fact, you could take nearly any struct from my value-types post and use the class keyword, and it would probably compile! However, even though it will compile, it will behave radically differently, because...

Classes have reference semantics

Let's say we have a little Temperature struct:

struct Temperature {
  var fahrenheit: Double
  
  init(fahrenheit: Double) {
    self.fahrenheit = fahrenheit
  }
}

And some sensible-looking code:

var temp = Temperature(fahrenheit: 70)
home.thermostat.temperature = temp

temp.fahrenheit = 400
home.oven.temperature = temp
home.oven.bake()

Fine, right?

If we change that struct to a class, everything will still compile fine... but it's going to get really, really hot in here. When Temperature is a class, it has reference semantics: there is one instance temp created in the first line, and that instance is shared between the thermostat and the oven, so changing the temperature on one affects the other.

There are times when reference semantics are the right way to model your problem, and classes are great for those. Temperature is an example of something that should remain a struct, because it is intrinsically a value. The key distinguishing characteristics for choosing a class is when you need identity and sharing, such that two different places in your code need to be able to refer to the same entity, and see the effects of each other's changes to that entity.

Coming from C++, there are other clues as to whether to choose struct or class. If in C++ you would always pass around the type by pointer or reference (whether a raw pointer or a smart pointer like std::shared_ptr), or ever comparing the address of the instances, or if you're going to be deleting the copy constructor, then you probably have reference semantics and should use a class in Swift. C++ features like virtual functions or dynamic_cast are also indicators of reference semantics, but they aren't as strong: these could be merely implementation mechanisms for something else.

Most importantly, in C++ we often use inheritance to get implementations for free in a new type, even without virtual functions. The Curiously Recurring Template Pattern (CRTP) is one such pattern that uses C++ inheritance even in types that have value semantics. This pattern does not translate well into Swift, and there are alternative ways to express this kind of polymorphism through the generics system and so-called "protocol extensions". The fact that one would use inheritance in C++ to model a problem is not a good indicator that one should use classes in Swift. Instead, it all comes back to reference semantics and identity.

Inheritance & overriding

Swift classes allow single inheritance, with overriding of methods. Let's create a subclass of Person for employees of some company:

class Employee: Person {
  let badgeNumber: Int
  
  init(name: String, age: Int, badgeNumber: Int) {
    self.badgeNumber = badgeNumber
    super.init(name: name, age: age)
  }
  
  override func greet() {
    print("Hello \(name), your badge number is \(badgeNumber)!")
  }
}

The base class (which we call the superclass in Swift) is specified after the colon. There can be at most one superclass, which keeps Objected-Oriented hierarchies a lot more simple and direct. The expressivity gap with C++'s support for multiple inheritance is mostly filled by the generics system, which we'll get to later (sorry, I keep saying that).

All inheritance is effectively public. There's no notion of private inheritance, and protected doesn't even exist in Swift. I'm going to ignore Swift's access control for a while longer, and note that when all of your code is in one Swift module, you often don't think about access control, and it's fine.

Our Employee type has an initializer and a greet method. The greet method overrides the one declared in Person, as indicated by the override keyword. Class members are effectively virtual by default, although one can explicitly mark a specific class or class member as final to cut off all subclassing or overriding for that entity for good. Subtyping and virtual dispatch works as one would expect:

let person: Person = Employee(name: "Doug", age: 39, badgeNumber: 1)
person.greet() // prints "Hello Doug, your badge number is 1!"

Member initialization is... backwards?

If you looked at the body of that initializer for Employee, it might have freaked you out a little bit:

   init(name: String, age: Int, badgeNumber: Int) {
    self.badgeNumber = badgeNumber
    super.init(name: name, age: age)
  }

Notice how we initialized the data members of our subclass (or "derived class" in C++ parlance) first, and then called the superclass initializer. Weird, right?

This is absolutely not how C++ works: C++ initializes the base classes first, and then the members of the derived class, and then executes the code in the constructor of the derived class. But the C++ model has some weird behaviors with virtual function calls in constructors that fall out of it. Let's see what happens if we write the above classes in C++ and try to greet the person in the constructor:

class Person {
public:
  std::string name;
  int age;
  
  Person(const std::string &name, int age) : name(name), age(age) {
    greet()
  }
  
  virtual void greet() {
    std::cout << "Hello " << name << "!" << std::endl;
  }
}

class Employee: public Person {
  int badgeNumber;
  
  Employee(const std::string &name, int age, int badgeNumber) 
    : Person(name, age), badgeNumber(badgeNumber) { }
  
  void greet() override {
    std::cout << "Hello " << name << ", your badge number is " << badgeNumber << "!" << std::endl;
  }
}

Person *person = new Employee("Doug", 39, 1) // prints "Hello Doug!"
person->greet(); // prints "Hello Doug, your badge number is 1!"

Weird-er, right? The problem with calling virtual functions in constructors is that the base class constructor doesn't have a fully-formed object of the subclass type, because the members of the derived class haven't been initialized yet, and it would be undefined behavior to access them. So it's not safe for the Person constructor to treat the object as the Employee instance it will eventually be. Instead, it treats the object as an instance of Person, calling Person::greet instead of Employee::greet, causing the difference in behavior above. This is also the reason that it's possible to get errors at runtime due to calls to abstract virtual functions in C++, even though the language prevents you from creating an instance of an abstract type. Supporting this weird semantics is also awful for C++ compiler writers; read this if you'd like to feel bad for them.

Okay, back to Swift. If we go put a call to greet in the initializer for Person, we'll get consistent output:

Hello Doug, your badge number is 1
Hello Doug, your badge number is 1

The "backwards" initialization required by Swift, where you must initialize all of your own data members before calling your superclass initializer (via super.init), gives an important guarantee: the superclass initializer knows that all of it's subclasses have already initialized all of their own data members before it ever runs. Follow that logic all the way up to the root class's initializer (which has no superclass), and we know that once that initializer has initialized all of its own data members, the object is fully-initialized with its final type. Definite Initialization then makes self available (which happens after the super.init call in non-root classes), so you can call any overridden method in an initializer and get a consistent result. So the Swift approach, while weird at first, provides more consistent semantics with less syntax than C++, and doesn't require any kind of "default" state for data members. Sometimes weird is actually good, eh?

Deinitializers

Unlike structs and enums, an instance of a class has a well-defined point at which it is destroyed (when it is no longer used). At this point, one can execute cleanup code in the deinitializer. A Swift deinitializer is like a C++ destructor, except it's written as deinit like this:

// in Person
deinit {
  print("\(name) has expired")
}

Deinitializers are never overridden in the sense that methods are. Rather, all of the deinitializers in the class hierarchy are executed from the most-derived class to the root class, and then all of the data members are destroyed.

Downcasting

Let's say you have a Person and you want to check whether it's actually an instance of Employee. In C++, you would use dynamic_cast to perform the cast dynamically. In Swift, we use as?, like this:

func checkIfEmployee(person: Person) -> Bool {
  if let employee = person as? Employee {
    print("Yes, employee badge number is #\(employee.badgeNumber)")
    return true
  }
  
  return false
}

The way to read this is that we are attempting to downcast the person instance into an Employee. If it succeeds, we'll enter the body of the if let with the variable employee bound to the same instance as person, but with type Employee. If it fails, the if body doesn't run.

Now, sometimes we know based on other invariants in the program that a particular person instance must be an Employee. In C++, you would use a static_cast here (or a dynamic_cast on a reference to the type). In Swift, you can use as!, e.g.,

print("Employee badge number is #\((person as! Employee).badgeNumber)")

Swift is going to check this cast at runtime, and halt the program with an error message if person is not, in fact, an Employee. The as! cast should be used only in rare cases where you're unable to express what you need through the type system, and other invariants ensure that it will always succeed. Just like we always do with static_cast in C++, right? RIGHT?

Optionals

In bundling together the as? cast and the if let in the example above, I glossed over what's actually happening here. Specifically, the as? can be used anywhere, and it produces a value of optional type. If we were to write:

let maybeEmployee = person as? Employee

The type of maybeEmployee is Employee?, where the question mark implies that might have an Employee, or we might have nothing. If you've used the C++ std::optional, it's the same idea, but with syntactic sugar in Swift to make it more ergonomic. if let is one of the ways to check whether there's a value inside an optional:

if let employee = maybeEmployee {
  print("Yes, employee badge number is #\(employee.badgeNumber)")
} else {
  // employee is not available
}

Here, the if let is checking whether the optional value to the right-hand side of = actually contains a value. If so, it pulls that value out and puts it into the variable declared to the left of the =. Otherwise, the body of the if doesn't run, but the else block (if present) would be executed. The else block does not have access to the employee variable at all.

Options for accessing optionals

if let is the most popular way to extract a value out of an optional, but it isn't the only one. There is also the dual to if, called guard, which helps with the "early returns" style of programming. We could refactor the code above to look like this:

guard let employee = maybeEmployee else {
  // employee is not available
  return false
}

// employee is an Employee instance. Use it for the rest of the function
print("Yes, employee badge number is #\(employee.badgeNumber)")
return true

A guard statement checks that its conditions are true, which can involving introducing new variables. It also requires an else block that must return, so there is no way to fall out of the else block and into later code. Early returns are surprisingly controversial in C++, I think because you have to mentally dig for the return along all of the paths in the early-returning if. Swift's guard addresses those concerns by giving this pattern a specific keyword up front (guard), making the condition a positive one (you pass the guard if this condition holds), and ensuring that if the condition fails you always do an early return. Enough about guard; back to optionals, shall we?

The third way one can look into an optional is with a switch, because the ? syntax is actually just synactic sugar for a use of the generic Optional type that's defined in the Swift standard library. It looks like this:

enum Optional<Wrapped> {
  case none
  case some(Wrapped)
}

The none case means there is no value, and the some case means that there is a value of type Wrapped. The angle brackets are for generics, as one would expect, declaring that Optional has a type parameter (akin to a template type parameter in C++). Given a value of optional type, you can switch on the two cases:

switch maybeEmployee {
  case .none:
    return false
  
  case .some(let employee):
    print("Yes, employee badge number is #\(employee.badgeNumber)")
    return true
}

Optionals for dictionary access

Optionals can be formed from any type, and are useful for APIs that might return "no result". One great example of this is accessing elements within a Swift Dictionary, which is like a std::unordered_map, by looking for an element with a specific key. Let's say we create a dictionary that maps from badge numbers to the employee with that badge number, like this:

// Create a dictionary from (badgeNumber, employee) pairs
let employeesByBadgeNumber: [Int: Employee] = .init(uniqueKeysWithValues: employees.map { 
  employee in (employee.badgeNumber, employee)
})

The use of closures will be explained in a separate post, but if you're used to C++ lambdas, it probably looks familiar already. The [Int: Employee] syntax is sugar for a Dictionary mapping from Int keys to Employee values. It mirrors the syntax of dictionary literals, i.e., [1: doug] is a dictionary containing a single entry mapping 1 to the doug instance.

We can look for the employee with a given badge number by subscripting this dictionary. The result of the subscript is going to be an Employee?, i.e., it's the employee if one with that badge number was found, or nil otherwise, which works very nicely with if let:

if let employee = employeesByBadgeNumber[17] {
	employee.greet()
}

again, the type system is helping here: the obvious way to find the value associated with a key is to use subscript, and it returns an optional so you're sure to handle the case where the value is missing. If you find yourself longing for the insert-a-default-if-nothing-is-there behavior of C++'s std::map and std::unordered_map, there's a special subscript just for you that also takes a default. You could use it like this, say, to count word frequency in a list of words:

var frequency: [String: Int] = [:]
for word in words {
  frequency[word, default: 0] += 1
}

The subscript that takes a default value produces a non-optional value, because an entry [word: 0] will be added if one wasn't already present. So, we can increment the frequency of each word directly.

Optional chaining

Optionals are really common in Swift code, so the language has a number of affordances to make them easier to work with. For example, earlier we wanted to greet an employee with the given badge number, so we wrote this:

if let employee = employeesByBadgeNumber[17] {
	employee.greet()
}

This code is fine. It will work. But it's a little bit verbose, with the whole if-let dance to create a new variable. Instead, one can use optional chaining to call the greet method only when there is an employee, like this:

employeesByBadgeNumber[17]?.greet()

The ? in this case means that we're doing optional chaining. It is followed by accesses into the instance (i.e., .greet()) that only happen when there is a value inside the optional. You can even chain multiple optional accesses together. For example, let's imagine that our Employee class added a manager data member, i.e.,

var manager: Employee? = nil

We could get the name of the employee 17's manager like this:

let managerName = employeesByBadgeNumber[17]?.manager?.name
// >----------------------------------------^--------^---->
//                                          1        2

Evaluation is left-to-right in Swift, so you can read this as looking in the dictionary for an employee with badge number 17. If that Employee? result contains an Employee (i.e., it is not nil) at the point marked #1, then we access the manager of that employee. If the manager data member contains an Employee at the point marked #2, then we access the name of that manager instance.

But what is the type of the managerName variable? Well, we know that Employee.name is a String, but of course at either #1 or #2 we might have encountered a nil and had to stop evaluation. We can't just leave managerName undefined (gasp!), so instead it gets type String?, where it either stores the result of completing the optional chain or it nil if a nil was encountered along the way.

Optionals replace the dreaded NULL

In Swift, an instance of class type can never be NULL. If you have a value employee of type Employee, there is always a valid instance there. Instead of NULL pointers, Swift uses optionals: Employee? is either a value employee, or it's nil. But unlike the C++ equivalent of Employee*, Swift has static affordances to make it easy to use optionals correctly. That means no NULL pointer dereferences, no defensive checks against NULL when you're not sure. Unless you're one of the few using a not_null smart pointer type), it may surprise you at just how liberating it is to stop thinking about the special case of NULL everywhere and let the type checker support you.

Aside: How big is an optional?

One last little fun bit of trivial: if I run this C++ code, what values should we expect to get?

class Employee { };

int main() {
  std::cout << sizeof(Employee*) << std::endl;
  std::cout << sizeof(std::optional<Employee*>) << std::endl;
}

On my system, I get 8 and 16, respectively. Using std::optional doubles the storage needed!

Now what happens with the Swift equivalent code?

class Employee { }

print(MemoryLayout<Employee>.size)
print(MemoryLayout<Employee?>.size)

I get 8 for both! Swift is doing an important space optimization behind the scenes: because an Employee must refer to a valid address, and the integral value of 0 is not a valid address, Swift will use the integer representation 0 to mean the .none of the optional (which corresponds to nil) and all other values to mean the .some case of the optional. Therefore, it can use a single pointer's width of storage to capture an optional value. This is generally good for memory usage, but it's also really important for interoperability with C++, because a C++ pointer value (that can be NULL) can be treated as a Swift optional, and vice-versa.

Now, if you were to create an optional integer, Int?, you would see that the optional takes more space than a single Int. That's because all possible bit representations in an Int are valid values of an Int, so we add an extra byte (from 8 to 9 on a 64-bit machine) to tell whether we're in the .none or the .some case.

But the thing is... spare bits and holes in bit representations are everywhere, and Swift will dig to try to find them. For example, let's add a Contractor class and create an enum with three cases:

enum Payee {
  case employee(Employee)
  case contractor(Contractor)
  case myself
}

How big are Payee and Payee?, based on what we've seen before?. In fact, both are the same 8 bytes on a 64-bit system. It's not magic, it's how computers work: a class like Employee or Contractor is going to be at least 8-byte-aligned, because of it's object header. That means that a valid reference to a class (basically, a pointer) will always have the lowest three bits be zero: anything else would be a misaligned pointer. So between those two cases, we have three common low bits we can use for whatever we want, and the third case (myself) has no data. Three bits is plenty of space to store a discriminator that says whether a given Payee instance is an employee, contractor, or myself: we need only mask off the upper bits to figure out what case we're in, and then mask off the lower bits to get a proper Employee or Contractor pointer when we know which one it is. Cool, right?

That explains Payee, but what about the optional Payee? size? The principle is the same: Payee only needed two of the three low bits to represent its three cases. That leaves one glorious extra bit to distinguish between the some and none cases, so Payee? still needs only 8 bytes.

You can do all of these optimization tricks in C++ with template metaprogramming. Clang uses LLVM's PointerUnion class template extensively to get this same kind of memory savings, but it's a whole lot nicer when the language just does it for you and you don't have to think about it.

Automatic reference counting

By now you've probably noted the lack of delete or free or any other kind of explicit deallocation. Swift provides automatic memory management using reference counting: a new instance of class type is allocated with a reference count of 1. Anything that needs to hold onto the object will increment its reference count to do so, and decrement the reference count when it is no longer needed. When the reference count hits zero, the object will be de-initialized and the memory freed.

The C++ standard library has a reference-counting smart pointer template, std::shared_ptr. If you've used shared_ptr, you already have a sense of how reference counting works in practice, and you can think of Swift's classes as using roughly the same semantic model but eliminating the syntactic overhead by bringing it into the language. Swift also has an analogue to std::weak_ptr, which is a non-owning reference to something that may be kept alive by other shared_ptr instances.

However, the shared_ptr analogy is somewhat flawed because std::shared_ptr makes a number of compromises so that it can be used to point to any type. A shared_ptr is often two pointers large, because it has to separate out the "control block" (which handles the reference count) from the raw object pointer. This isn't great for efficiency, and also makes the get() operation particularly dangerous: once you've extracted the raw object pointer, there's no way to get back to the safety of a shared_ptr unless you've also used std::enable_shared_from_this. A better comparison to Swift would be boost::intrusive_ptr, because the Swift reference count is embedded within the object itself. That means you're only passing around a single pointer to the object rather than the two pointers often needed for a std::shared_ptr, and reference-counting operations have better locality with the object.

Weak references

One problem with reference counting is the potential for reference cycles: if object a holds a reference to object b, and object b holds a reference to object a, then these two objects will keep each other alive even after the rest of the program has forgotten about them, resulting in a leak:

class A {
  var b: B?
}

class B {
  var a: A?
}

This problem exists in all reference counting implementations, whether it's the C++ std::shared_ptr or Python. Python has a built-in garbage collector to find these reference cycles and arbitrarily break them to free memory, but such an approach requires a fairly heavy runtime. Swift's approach is more like std::shared_ptr: the programmer is responsible for ensuring there are no reference cycles, and Swift provides tools to identify and deal with them when they happen.

Weak references are the primary mechanism for breaking cycles. A weak pointer points at an object but does not keep it alive. Weak references are introduced with the weak modifier on a variable, e.g.,

class B {
  weak var a: A?
}

A weak variable must have optional type, because the object it points to might go away at any point. If the object it points to is destroyed, then the weak reference will get the value nil. Code that works with weak references will naturally "promote" the weak reference to a strong reference when working with it, because it's the same operation as determining whether any other optional has a value in it:

if let a = b.a {
  // a has type A, a strong reference that keeps the object alive
}

This is akin to the lock() operation of a std::weak_ptr returning a possibly-NULL std::shared_ptr, but benefits from all of the conveniences for optional types that Swift offers.

Why reference counting?

We opted for reference counting in Swift because we felt it hit the sweet spot for automatic memory management.

Unlike more traditional garbage collectors, reference counting is deterministic: when your reference count goes to zero, your object goes away. There's no delay until the garbage collector runs where you are using memory than you should be, nor is there ever a need to "pause" the program to run the garbage collector. Moreover, reference counting can be implemented with a tiny runtime footprint---just a couple of atomic operations---vs. a more elaborate garbage collector runtime. It also places no restrictions on the program beyond the need to correctly balance out "retains" and "releases", so you can freely mix reference-counted code with other code in the system (say, in other languages). Finally, reference counting is locally optimizable: if you can prove that an object is kept alive within a given scope, you don't need to perform any reference counting within that region of the code. That means you can optimize away all of the reference counts in a particularly performance-sensitive part of the code while not having to change the basic model used by the same classes elsewhere in the program.

At the other end of the spectrum from a traditional garbage collector are systems with unique ownership, such as std::unique_ptr. This approach requires no overhead for automatic memory management, but makes it hard to express any data structure that isn't tree-like in nature. Reference counting can express arbitrary graphs of objects directly, making it more flexible than models based on unique ownership.

Coming from C++, the "automatic" part of Swift's automatic reference counting can be frustrating. The optimizer might not remove all of the reference counting that you want it to, and where in C++ you might just grab the raw pointer when your smart pointers aren't optimizing away, Swift makes it a little harder. My advice here is to profile your code: not some micro-benchmark that does no work, but your actual code. If reference-counting is causing performance issues, swift_retain and swift_release will light up in your profiler and you can work on optimizing those hot spots. Swift has some tools for helping with this, such as unowned(unsafe) references (which are effectively raw pointers) as well as ways to be deliberate about ownership, such as the consume operation. In my experience, most of the time the reference counting isn't in the way, and when it is a problem, one can optimize locally, usually without having to dip into unsafe code.

Wrap-up and what's next?

We dove into Swift's support for Object-Oriented Programming, which is based on classes with single inheritance and method overriding. We went over initialization (weird, but good) and de-initialization, as well as down casting.

That led us into the world of optionals, and how Swift uses them to eliminate NULL pointers from the language (Tony Hoare's self-proclaimed billion-dollar mistake). Optionals are a huge part of working with Swift, even if you stick to the value-semantic world of structs and enums, so it's worth getting a feel for the syntactic constructs that make them easy to work with.

We're going to take a break on introducing new kinds of types: structs, enums, and classes form the core of Swift's semantic model. Next, we're going to talk a bit about code organization in Swift, and how extensions take all of the fun out of the "free function or member function?" debate.

Tagged with: