Doug's Compiler Corner

Originally posted on 2024-04-15 03:09:00 +0000

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

Swift for C++ Practitioners, Part 6: Error Handling

C++ has a few different mechanisms for handling errors. One is std::expected, where the return type of a function is either the expected value or an error value. Another is exceptions, which come with a whole host of downsides (we'll get there in a moment). Personally, I find that I end up hand-rolling something more like std::expected in C++ more often than not, and the experience isn't great.

Swift takes some of the syntax of C++ exceptions (throw and catch), but the underlying philosophy of error handling in Swift is a bit different from C++. To get to it, let's start by looking at some of the issues with C++ exceptions, as they will help inform the choices made in Swift.

Issues with C++ exceptions

Herb Sutter did quite an excellent job of laying out the issues with C++ exceptions in P0709. Rather than come up with my own formulation from scratch, I'll quote directly from his abstract and refine from there:

  • §4.1: C++ projects commonly ban exceptions, because today’s dynamic exception types violate the zero- overhead principle, and do not have statically boundable space and time costs. In particular, throw requires dynamic allocation and catch of a type requires RTTI.
  • §4.2: Programs bugs are not recoverable run-time errors and so should not be reported as exceptions or error codes.
  • §4.3: Allocation failure is not like other recoverable run-time errors and should be treated separately.
  • §4.5: Some users don’t use exceptions because exceptional control flow is invisible.

I think Herb missed one overarching problem, which is that C++ got the defaults wrong with respect to exceptions. In C++, a function is assumed to throw unless it is explicitly marked noexcept(or throw() in the pre-C++11 beforetimes). This means that nearly every C++ function out there can throw, because only the most diligent C++ developer is going to carefully audit each function to make sure it is marked noexcept, and suffer the unexpected std::terminate call if they got it wrong. The problem with getting this default wrong is that it undercuts everything else you'd like to fix with the C++ exceptions design. The space and time costs for exceptions (Herb's §4.1) might be perfectly acceptable if you only paid for them where you used exceptions, but because of throwing-by-default, you're paying those costs across the entire code base even if only a narrow portion of it actually uses exceptions. Additionally, you can't change the language to make exceptional control flow visible (Herb's §4.5) when the default is "everything throws", because exceptional control flow is everywhere in C++. And while you could eliminate std::bad_alloc (Herb's §4.3) to reduce the number of places that could automatically throw, it barely makes a dent in the user experience when everything is already assumed to throw.

Error handling is hard: it's hard to anticipate what failures can occur, hard to figure out how to respond appropriately to get back to a reasonable state after an error, and hard to simulate those errors in a test to be sure you got it right. A language needs to help you identify and deal with errors, and most-everything-can-throw does the opposite.

Swift's approach to error handling

Swift's error handling model is similar in spirit to C++: an error is thrown with throw, can propagate through multiple stack frames (tearing down local values and running defer blocks along the way), and is eventually caught with a catch block. Nearly any type can be thrown: the only constraint is that the type must conform to the Error protocol. By default, errors in Swift are type-erased to any Error; see my prior post on type erasure if you want to learn more about type erasure.

The differences are mainly in the defaults, but they make a world of difference. A Swift function cannot throw unless it explicitly specifies that it throws, and memory allocation failure is not handled via a thrown error (Herb's §4.3), so the vast majority of Swift functions are non-throwing.

Thrown errors are checked at compile time: a throwing operation must be marked via the try keyword to make all of the control flow in code explicit (Herb's §4.5), and the thrown error must either be handled (e.g., with a catch block) or the enclosing function must be marked throws to allow the error to propagate out of it. The key goal here is that there should be no surprises in error handling: you know where errors originate and how they propagate, and while you can write bad error-handling code, it's hard to forget to write it at all, and you only do so when things can actually fail.

An example

Let's see Swift's error handling in action by writing a simple function to parse a string into an integer. Because I love generic programming, we're going to make it generic over any fixed-with integer type by turning it into an extension of the FixedWidthInteger protocol. But first, let's think about what can go wrong: the easiest thing that can go wrong is that one of the characters in the string isn't a digit. Let's capture that case in a new Error-conforming enum that will describe errors that occur during integer parsing:

enum IntegerParseError: Error {
  case nonDigitCharacter(String, index: String.Index)
}

We haven't talked about the String type much, so allow me a slight digression...

Digression: A proper discussion of String could fill up its own post, so here's the short version: String is a fully Unicode-aware string type and always contains a valid Unicode string. It's generally best to form and manipulate them with string literals, string interpolations, or other high-level operations. For lower-level operations you can treat them as a collection of Character instances. Character is about as far from the C++ char type as you can get, because it captures the notion of a grapheme cluster, which is the nearest approximation Unicode has to what a human would consider a single character on screen. This covers everything from simple ASCII characters to composed multi-byte sequences like flag emoji (🇺🇦) and family emoji (👩‍👩‍👦‍👦).

So, let's build our first iteration of a parsing function that takes a string and converts it to any fixed-width integer:

extension FixedWidthInteger {
  init(parsing string: String) throws {
    self = 0
    for index in string.indices {
      let char = string[index]
      guard let digit = char.wholeNumberValue else {
        throw IntegerParseError.nonDigitCharacter(string, index: index)
      }

      self = self * 10 + Self(exactly: digit)!
    }
  }
}

I chose to make this an initializer, so that it can be used as (e.g.) Int8(parsing: string). The code itself walks through each of the indices in the string, extracting each character, turning it into a whole number value, and accumulating the result into self. If at any point it finds a character that isn't a whole number value, it will throw an instance of IntegerParseError that describes what the problem is.

There is no error-handling code within this initializer, so it must be marked as throws to indicate that an error can propagate out of a call to the initializer. The Swift compiler will produce an error if the throws is missing.

Now, when we use this initializer, we will always need to acknowledge the error. For example, let's try to parse "1+23" into an Int8:

let value = try Int8(parsing: "1+23")

The try must be present to indicate that an operation to its right can throw. Just as any code containing a throw must account for the error (by handling it or being throws itself), code containing a try must account for the error. This makes control flow due to errors explicit. Note that you can cover multiple throw sites with a single try if you'd like. For example:

let (first, second) = try (Int8(parsing: firstString), Int8(parsing: secondString))

is equivalent to

let (first, second) = (try Int8(parsing: firstString), try Int8(parsing: secondString))

It's a matter of taste. My personal style leans toward the former, because if there's any kind of side effect in a subexpression that would need different control flow, I'm going to break that out into a separate statement anyway.

do..catch blocks

So, we've seen that we can mark a function as throws to allow an error to propagate out of it, so how we can handle an error when we don't want it to propagate? In Swift, it's a do..catch block:

do {
  let value = try Int8(parsing: "1 + 23")
  print("Integer value is \(value)")
} catch let error {
  fatalError("Invalid input string: \(error)")
}

Any errors thrown within the body of the do part of a do..catch block will be checked against each of the catch blocks, much like C++'s try..catch. In this example, we have a single catch block that catches any error. The error itself will be named error and have type any Error.

There can be multiple catch clauses, and they use the same pattern-matching syntax as switch statements, so we can match a particular error case if we want:

let value: Int8?
do {
  value = try Int8(parsing: "1 + 23")
} catch IntegerParseError.nonDigitCharacter(let string, index: let index) {
  print(#"Non-digit character "\#(string[index])" found in string "\#(string)"."#)
  value = nil
} catch {
  print("Unhandled error \(error)")
	throw error
}

Now when an IntegerParserError.nonDigitCharacter error is thrown, it will be matched against the first catch, with the string and index values of the error bound to the stored values.

Aside: I snuck another minor Swift feature into that print statement, raw string literals, which are written with #"..."#. Inside, double-quotes don't finish the raw string literal unless followed by a #, and the \ is not an escaping character unless followed by the #. If you really want to use "# or \# inside the raw string literal, you can even add more #s around the outside double-quotes, where you need a matching number of #s to make the " or # significant, e.g., ###" not escaped \"# until we get here "###.

Any errors that don't match that first catch clause will fall into the second catch clause. As a little shortcut, if the catch clause doesn't specify anything to match, it's equivalent to let error, i.e., it introduces local error value of type any Error.

Note that this second catch block throws error again: Swift doesn't have (nor need) a special "rethrows" syntax, because errors are just values. So if we want to "rethrow", we can throw the error we got. Or we can package it up---there's no need for special mechanisms like std::exception_ptr in C++.

try?

There's one last shorthand syntax to get out of the way before we switch gears a bit. If you find yourself trying an operation and producing either a value or nil, like this:

let value: Int8?
do {
  value = try Int8(parsing: string)
} catch {
  value = nil
}

Then you can use the try? shorthand to simplify the code to:

let value = try? Int8(parsing: string)

The try? expression performs the operation to its right and, if successful, returns its value. If the operation throws an error, then the try? expression returns nil. The type of a try? is an optional of the expression to its right, so in this case, we get an Int?. I'm not a huge fan of this feature, because it loses information about the specific error that was thrown, but it can be convenient when throwing together some Swift code to get a task done once or twice.

Throwing closures

Just like functions, closures can throw. Such closures can explicitly be marked with throws if you provide a parameter list, like this:

func commaSeparatedIntegers(string: String) throws -> [Int8] {
  // split String into [Substring] instances at the given separator
  let strings = string.split(separator: ",") 
  return try strings.map { (substring: String) throws in
    try UInt8(parsing: String(substring))
  }
}

A few things to note: the split(separator:) operation on a collection turns an array of subcollections that are separated by the Equatable element provided by the separator argument. In this example, we then map over those substrings (a std::transform in C++ STL-speak) to parse each of them into an integer. Because the initializer of the UInt8 throws, we must mark the closure's return expression with try. Now, when giving a throwing closure, the map operation itself also throws, so we need to mark the map call with a try as well. Therefore, all of the exceptional control flow (out of the closure, and out of the map call) is indicated in the program.

Now, we don't have to mark a closure as throwing when it throws, because it can be inferred. A closure that contains a throw or a try (that isn't swallowed up in a do..catch / try? / try!) is known to throw, so the above can be written more succinctly as:

func commaSeparatedIntegers(string: String) throws -> [Int8] {
  // split String into [Substring] instances at the given separator
  let strings = string.split(separator: ",") 
  return try strings.map { 
    try UInt8(parsing: String($0))
  }
}

I'm also using the $0 shorthand for "the first closure argument", because this closure is so small that it's not worth naming the argument. The presence of try implies that the closure throws.

We're going to come back to map and it's "it throws if its input closure throws" behavior a bit later, when we talk about typed throws.

Program bugs are not exceptional conditions

We've talked a lot about the mechanics of error handling in Swift, so now let's talk some philosophy. Swift takes the viewpoint that errors are exceptional conditions that the program has to deal with, but program bugs are not exceptional conditions. When Swift detects a program bug, such as an out-of-bounds index into an array or an attempt to force an optional value containing nil (via ! or as!), it will immediately halt the program. It will look like a program crash, but should be accompanied by a message indicating what went wrong.

The Swift philosophy, echoed by Herb's §4.2, is that a program should not attempt to recover from a programmer error. Programmer errors are likely to mean broken invariants that will cause more problems later, and a program that's limping on after a program error is ripe for security exploits. So when things go wrong, Swift halts the program to prevent further damage.

Now, you don't have to buy in to Swift's philosophy here. You can go ahead and add an Array subscript that throws when out of bounds, or a cast operation that throws when it fails, and generally turn every bad thing that can happen into a thrown error. You'll be going against the grain of the Swift standard library and general community, but it's possible. But know that I will know, and I will judge you :).

try!

There are some times when you're using a general operation that can throw an error, but because of where you are in the program, you know it can't fail. Perhaps you're working from data that's baked into the program elsewhere, or are in program startup where things can't go wrong yet. As a corollary to try?, there's a try! operation that produces a fatal error if an error is thrown from its subexpression. For example, if we do:

let value = try! Int8(parsing: "1+23")

The type of the try! expression is the type of its subexpression; here, that's Int8. If an error is thrown from the subexpression, the program will halt with an error message something like this:

parsing/parsing.swift:94: Fatal error: 'try!' expression unexpectedly raised an error: parsing.IntegerParseError.nonDigitCharacter("12+3", index: Swift.String.Index(_rawBits: 65799))

You can think of try! as having about the same semantics as (try? <subexpression>)!, but providing a better error message when things go wrong. Unlike try?, I actually do like try!: when the program invariants say there can't be an error here, it's a lot more convenient than a do..catch with a fatalError in the catch block.

Aside: Some Swift style guides prohibit the use of the various ! operations, whether they are ! to force-unwrap optional, as! to force-cast, or try! to assert that an operation cannot fail. These constructs exist in Swift for a reason, and are appropriate to use when the surrounding program invariants ensure that the operation cannot fail, but you weren't able to express those invariants in the type system. The consistent use of ! is meant to say "be wary", not "run away screaming."

Preconditions and assertions

A careful programmer will document the invariants of data structures and the preconditions of any functions, so that misuses due to programmer error are caught quickly. To support this checking consistently, Swift provides the precondition function, which takes a Bool condition, an optional error message string, and optional file/line information (that will be automatically filled in for you). For example, a bounds check on an array subscript will use a precondition like this:

precondition(index >= 0 && index < count, "out-of-bounds array index")

If that Bool expression evaluates false, the program will halt immediately and display the error message, line, and column. The actual declaration of precondition shows off a few tiny features that Swift library developers like to use, that I otherwise wouldn't have gotten to, so here it is:

public func precondition(
  _ condition: @autoclosure () -> Bool,
  _ message: @autoclosure () -> StaticString = StaticString(),
  file: StaticString = #file, line: UInt = #line
)

The first argument is the Bool condition, but it has function type and this @autoclosure thing. What's going on? Well, an "autoclosure" is a fun way in which a function can defer computation of one of its arguments. In essence, when we wrote the expression index >= 0 && index < count, Swift packaged that up into a closure

{ index >= 0 && index < count }

and passed that closure into the precondition function. We can evaluate the condition by calling condition() in the body of the function. Note that the message is also an @autoclosure: we won't call the closure unless the condition has failed, so we don't need to form the string instance except along the error path. Auto-closures actually came into being because we wanted to be able to express the short-circuiting behavior of the logical && operation in the library. The implementation looks like this:

extension Bool {
  public static func &&(lhs: Bool, rhs: @autoclosure () -> Bool) -> Bool {
    lhs ? rhs() : false
  }
}

Autoclosures are also used in Swift's assert function, which ensures that a particular condition holds but only in debug builds. In release builds, the assert function folds away, never evaluating the condition. Use asserts for checking invariants that are too expensive in release builds, but can help aid debugging when things go wrong:

assert(self.isSorted, "insert operation must maintain sortedness property")

Both precondition and assert have defaulted file and line arguments. If you don't pass in a value, they take on the default values #file and #line, respectively. These are built-in macros that produce the file and line at the call site, which allows precondition and assert to point at where the failing precondition/assertion occurred in the source code.

Arithmetic overflow is a program bug

In the C++ community, we've spent a lot of time fretting over what to do about arithmetic overflow. It's been undefined behavior since the dawn of time, and undefined behavior is bad, so there's been a push to define it somehow. The most reasonable answer is to define it as basically every system out there implements it, which is two's complement wrapping. That's better, but it's not great: arithmetic overflow that is guaranteed to wrap can open up security vulnerabilities if an attacker can manager to wrap a buffer index to get access to other data they shouldn't.

Swift's answer is to define arithmetic overflow as a programmer error, and trap when it occurs. You can see the effect of this by trying to pass a string for a too-large integer into our parsing initializer:

let value = try Int8(parsing: "155")

The program will trap (crash) when trying to multiply 15 * 10 as an Int8. This is not great for our parsing function, so let's fix it!

Swift has two ways of dealing with arithmetic overflow programmatically. The first is a set of "wrapping" arithmetic operators with the & prefix that wrap their results according to two's complement arithmetic. If we replace the * and + in our init(parsing:) implementation with &* and &+, respectively, the crash will go away! Yay! Except that the answers are going to be wrong, because try Int8(parsing: "155") will now produce the result -101. Use the wrapping arithmetic operations when the algorithm you're implementing is designed for two's complement arithmetic, not as a "please make it not crash" hammer.

The second way of dealing with overflow is through the *ReportingOverflow family of methods on FixedWidthInteger, which form the appropriate operation and return the partial value along with a flag that indicates whether it overflowed or not. This allows us to detect an react to the overflow. Let's do that, and report the overflow via a thrown error. We can represent the failure condition with a new case in the IntegerParseError enum:

case overflow(String, any Numeric.Type)

Now, instead of writing self * 10, we can use the multipliedReportingOverflow(by:) operation, like this:

var overflow: Bool
(self, overflow) = self.multipliedReportingOverflow(by: 10)
if overflow {
   throw IntegerParseError.overflow(string, Self.self)
}

It's a little verbose, but it works. I bet someone could write a nice throwing wrapper around a FixedWidthInteger type that makes this kind of code easier to write, but for now I'm going to go the brute-force method and also addition counterpart. The final integer-parsing operation is as follows:

extension FixedWidthInteger {
  init(parsing string: String) throws {
    self = 0
    for index in string.indices {
      let char = string[index]

      guard let digit = char.wholeNumberValue else {
        throw IntegerParseError.nonDigitCharacter(string, index: index)
      }

      var overflow: Bool
      (self, overflow) = self.multipliedReportingOverflow(by: 10)
      if overflow {
         throw IntegerParseError.overflow(string, Self.self)
      }

      (self, overflow) = self.addingReportingOverflow(Self(digit))
      if overflow {
         throw IntegerParseError.overflow(string, Self.self)
      }
    }
  }
}

Now if we try Int8(parsing: "155") we'll get a thrown error that describes the failure: the value "155" doesn't fit into the integer type. We can even make the error nice and human-readable:

extension IntegerParseError: CustomStringConvertible {
  var description: String {
    switch self {
    case .nonDigitCharacter(let string, let index):
      "non-digit character '\(string[index])' at index \(index) when converting '\(string)' to an integer"

    case .overflow(let string, let type):
      "overflow converting '\(string)' to integer type '\(type)'"
    }
  }
}

so we end up with this:

overflow converting '155' to integer type 'Int8'

It's definitely worth making your error types provide nice, human-readable error messages, because they're likely to propagate up to some part of the system that can't do anything better than present the error to the user. And that message is a whole lot nicer for most users than the default rendering parsing.IntegerParseError.overflow("155", Swift.Int8).

Typed throws

Thus far, all the functions and closures we've seen have used "untyped" thows, where all errors are type-erased to any Error. Untyped errors are a good default: you can still pick out a specific error type using pattern-matching in a catch in those (generally rare) cases where you must, but most error-handling code just deals with any kind of error opaquely and safely backs out of the failed operation.

There are a few cases where untyped throws are insufficient, though:

  1. In performance-critical or very low-level code that cannot afford the runtime cost of type erasure or perhaps cannot allocate at all along the error path. This includes Embedded Swift, which doesn't allow for type erasure in any form. (Note that this restriction is captured by Herb's §4.1).
  2. In generic code that will only propagate errors that are introduced by operations on its generic inputs, such as the map operation throwing an error only when its closure parameter throws.

In such cases, it can be valuable to specify the exact type of the error that can be thrown. Swift allows this with typed throws, which allows one to specify a concrete thrown error type in parentheses after the throws keyword. Our parsing code could adopt typed throws like this:

init(parsing string: String) throws(IntegerParseError) { ... }

Now, this initializer may only throw or propagate errors of the type IntegerParseError. If you really want, you can take advantage of the type inference this affords for throws statements to omit the IntegerParseError on them, e.g.,

throw .nonDigitCharacter(string, index: index)

For callers, the knowledge that this initializer can only throw IntegerParseError gives tighter bounds for do..catch blocks. For example, this code knows that the caught error is of type IntegerParseError:

do {
  let value = try Int8(parsing: string)
  print("Value is \(value)")
} catch {
  // error has type `IntegerParseError`
}

That's the basics of typed throws: you can specify the type of a thrown error, which of course must conform to the Error protocol. This becomes part of the interface contract, so the implementation must only every throw this type, and clients are guaranteed to only ever see a thrown error of this type. It's all enforced statically, so you can't get it wrong. However, you can come to regret being specific about the thrown error type, if you later change the implementation in a manner that can produce new kinds of errors.

Warning: As of the time of this writing (April, 2024), typed throws has been accepted into the upcoming Swift 6.0 and is available in nightly snapshots, but has not yet made it into a Swift release. While it's not terribly likely, it's possible that details may change.

Untyped throws

Okay, so now we have untyped throws, and we have typed throws, so how do these relate? Fortunately, this is easy: untyped throws (spelled throws) is the same thing as throwing any Error (spelled throws(any Error)); it's effectively just a shorthand. A function that throws some type E can be converted to a function that throws any Error. That's it; nothing weird here. Just some syntactic sugar and an implicit conversion.

Non-throwing functions and the Never type

A much more interesting question is how typed throws relates to non-throwing functions. For that, we need to introduce another special Swift type called Never. There's no real equivalent to Never in C++, so I'll describe it from first principles. Please try not to equate it to anything in C++, because it'll create more confusion: it may sound a little like void, but it's not that. It may sound a little like an incomplete type, but it's not that. Enough preamble, what is it?

Never is a type for which there can never be an instance, ever. Formally, Never is an enum with no cases, which is also called an uninhabited enum:

enum Never { /* empty */ }

Now, because there can never be an instance of type Never, any computation that produces a value of type Never must be unreachable. Dead code. An impossible state.

Never came into Swift as a more composable way to indicate that a function could never return. C++11 has the noreturn attribute to say that a function could never return, but it's an attribute tacked on to a function, and doesn't get any real checking. In Swift, we give that function the result type Never:

func explode() -> Never {
  fatalError("BOOM!")
}

A Never-returning function must not return along any path. If any path could return or fall off the end of the function, the compiler will emit an error. The explode() function above is known to be correct because it's calling fatalError, which is defined in the Standard Library to return Never:

public func fatalError(
  _ message: @autoclosure () -> String = String(),
  file: StaticString = #file, line: UInt = #line
) -> Never

Using the Never type instead of an attribute has really nice properties, because it interacts nicely with type inference. If you form a closure that calls fatalError, it will be inferred to have result type Never, so we know that calls to that closure never return. And if you call a generic function that returns a T, and T is inferred to Never, you know that the generic function won't return.

Let's get back to throwing. If a function isn't marked with throws, then it doesn't throw. One might say it... never... throws. Joking aside, this means that a function that throws(Never) is equivalent to one that does not throw. So typed throws is actually a generalization of throwing functions, where throws(any Error) is "can throw anything" and throws(Never) is "does not throw", and you can have any throws(Concrete) in between.

Generic error propagation

A bit earlier, we had this call to map:

   return try strings.map { 
    try UInt8(parsing: String($0))
  }

and I noted that map propagates the error from the closure out. Now that we have typed throws, we can express the map operation generically:

extension Collection {
  func map<T, E>(body: (Element) throws(E) -> T) throws(E) -> [T] {
    var result: [T] = []
    for element in self {
      result.append(try body(element))
    }
    return result
  }
}

The basic idea is that we're calling body for each element in the collection, and appending the result to the result array to be returned at the end. Easy.

Now, body can throw an error of type E. When it does, that error propagates out of map, which is also defined to throw type E. E, in this case, is a generic parameter that is inferred from the call site. Everything type-checks in the map definition because the only throwing call site produces an E.

Now let's look at the client. In our original map call

   return try strings.map { 
    try UInt8(parsing: String($0))
  }

the closure will be inferred to throw an error (IntegerParseError) based on the call to init(parsing:). The generic argument for E will, therefore, be inferred to be IntegerParseError, so the whole map call throws IntegerParseError. What about a non-throwing case?

return strings.map { 
    (try? UInt8(parsing: String($0))) ?? UInt8(0) // turn failures into 0s
  }

Here, the closure does not throw (because the error from init(parsing:) is swallowed by the try?), so it has a thrown error type of Never. The generic argument for E is inferred to Never, so the call to map is known not to throw... and we don't need a try outside of the call to map.

Operations like map that are expressed in terms of generic thrown error types are only considered to throw when the thrown error type is inferred to something other than Never, allowing them to act as either throwing or not depending on context. This abstraction over thrown error types means we don't have to duplicate generic implementations to account for throwing vs. non-throwing. Never may feel odd, but it's quite powerful.

Historical note: Prior to the introduction of typed throws, a different feature called rethrows captured the notion that a particular function would only throw when one of the closure arguments passed into it throws. This feature was used for algorithms like map, and is expected to rapidly fall out of favor (and probably be deprecated) once typed throws is in widespread use.

The Result type

Swift has a standard type Result type for packaging up either a successful result, or a failure condition, and is intended to be used in much the same way as std::expected. Personally, I've found a lot less of a need for Result now that Swift has async/await (the subject of a later post, I promise), but you may see it in Swift code and it's good for exposition. Result is an enum with two cases:

enum Result<Success, Failure: Error> {
  case success(Success)
  case failure(Failure)
}

A function that returns Result<T, E> is morally equivalent to one that returns a T and throws(E). Indeed, one can convert in both directions fairly readily. To go from a function (or closure) to a Result instance, use this handy initializer:

extension Result {
  init(catching body: () throws(Failure) -> Success) {
    do {
      self = .success(try body())      
    } catch {
      self = .failure(error)
    }
  }
}

In other words: run the body. If it succeeds, put the successful value into self via the success case. If it fails, it will have thrown an instance of type Failure, so the catch block will initialize self via the failure case. Either way, self is fully initialized at the end!

To go the other way, there's a get operation:

extension Result {
  func get() throws(Failure) -> Success {
    switch self {
      case .success(let value): return success
      case .failure(let error): throw error
    }
  }
}

A call to get will either get the success value, or throw failure. If you prefer to use property access, we can write that with a throwing computed property:

extension Result {
  var value: Success {
    get throws(Failure) {
      return get()
    }
  }
}

Result illustrates another one of the great things about Never: if you have a type Result<Int, Never>, then it's just a wrapper around an Int. If you call get() or use value, it's known not to throw (because Failure is Never), so you don't need a try. All of the generic code just works, because it's impossible to ever end up in the failure case dynamically to trigger a thrown error.

Wrap-up

Error handling is one of the parts of Swift I adore: it's purposeful and opinioned, the pieces all compose well, and it helps take a tricky aspect of programming and makes it easier through useful application of a static type system. For C++ practitioners, there's a lot to be familiar with, because the basic model is effectively the same: thrown errors propagate up until they hit a matching catch. However, the defaults have been re-oriented to make exceptional conditions uncommon, static checking has been made sound to prevent accidents, and exceptional control flow has been made explicit to aid in understanding code.

And although it's not baked into the language design per se, Swift has definite opinions about when to throw errors and when to trap. Programmer errors should trap as soon as they are detected to prevent them from causing bad behavior, including security issues. Other exceptional conditions that prevent the normal flow of the problem should be reported as errors, preserving enough detail to provide a useful error message to be logged for the programmer or reported to the end user. I believe that following this philosophy, and being properly supported by the language in handling errors, leads to better-quality code in the long run.

Tagged with: