Skip to content

Latest commit

 

History

History
664 lines (510 loc) · 26.3 KB

0253-callable.md

File metadata and controls

664 lines (510 loc) · 26.3 KB

Callable values of user-defined nominal types

Introduction

This proposal introduces "statically" callable values to Swift. Callable values are values that define function-like behavior and can be called using function call syntax. In contrast to dynamically callable values introduced in SE-0216, this feature supports statically declared arities, argument labels, and parameter types, and is not constrained to primary type declarations.

In a nutshell, values that have a method whose base name is callAsFunction (referred to as a "callAsFunction method" for the rest of this proposal) can be called like a function. The function call syntax forwards arguments to the corresponding callAsFunction method.

struct Adder {
    var base: Int
    func callAsFunction(_ x: Int) -> Int {
        return base + x
    }
}

let add3 = Adder(base: 3)
add3(10) // => 13

Motivation

Currently, in Swift, only a few kinds of values are syntactically callable:

  • Values with function types.
  • Type names (e.g. T can be called like T(...), which is desugared to T.init(...)).
  • Values with a @dynamicCallable type.

However, call-syntax can also be useful for other values, primarily those that behave like functions. This includes:

  • Values that represent functions: mathematical functions, function expressions, etc.
  • Values that have one main use and want to provide a simple call-syntax interface: neural network layers, parsers, efficient bound closures, etc.

Here are some concrete sources of motivation.

Values representing functions

Values of some nominal types exactly represent functions: in the mathematical sense (a mapping from inputs to outputs), or in the context of programming languages.

Here are some examples:

/// Represents a polynomial function, e.g. `2 + 3x + 4x²`.
struct Polynomial {
    /// Represents the coefficients of the polynomial, starting from power zero.
    let coefficients: [Float]
}

Since these types represent functions, naturally they can be applied to inputs. However, currently in Swift, the "function application" functionality must be defined as a method.

extension Polynomial {
    func evaluated(at input: Float) -> Float {
        var result: Float = 0
        for (i, c) in coefficients.enumerated() {
            result += c * pow(input, Float(i))
        }
        return result
    }
}

let f = Polynomial(coefficients: [2, 3, 4])
print(f.evaluated(at: 2)) // => 24

The mathematical notation for function application is simply output = f(input). Using subscript methods achieve a similar application syntax f[x], but subscripts and square brackets typically connote "indexing into a collection", which is not the behavior here.

extension Polynomial {
    subscript(input: Float) -> Float {
        ...
    }
}
let f = Polynomial(coefficients: [2, 3, 4])
// Subscript syntax, may be confusing.
print(f[2]) // => 24

The proposed feature enables the same call syntax as the mathematical notation:

extension Polynomial {
    func callAsFunction(_ input: Float) -> Float {
        ...
    }
}
let f = Polynomial(coefficients: [2, 3, 4])
// Call syntax.
print(f(2)) // => 24

Bound closures

Variable-capturing closures can be modeled explicitly as structs that store the bound variables. This representation is more performant and avoids the type-erasure of closure contexts.

// Represents a nullary function capturing a value of type `T`.
struct BoundClosure<T> {
    var function: (T) -> Void
    var value: T

    func callAsFunction() { return function(value) }
}

let x = "Hello world!"
let closure = BoundClosure(function: { print($0) }, value: x)
closure() // prints "Hello world!"

A call syntax sugar would enable BoundClosure instances to be applied like normal functions.

Nominal types with one primary method

Some nominal types have a "primary method" that performs their main use. For example:

  • Calculators calculate: calculator.calculating(query).
  • Parsers parse: parser.parsing(text).
  • Neural network layers apply to inputs: layer.applied(to: input).
  • Types representing functions apply to arguments: function.applied(to: arguments).

Types that have a primary method usually call that method frequently. Thus, it may be desirable to sugar applications of the main method with call syntax to reduce noise.

Let's explore neural network layers and string parsers in detail.

Neural network layers

Machine learning models often represent a function that contains an internal state called "trainable parameters", and the function takes an input and predicts the output. In code, models are often represented as a data structure that stores trainable parameters, and a method that defines the transformation from an input to an output in terms of these trained parameters. Here's an example:

struct Perceptron {
    var weight: Vector<Float>
    var bias: Float

    func applied(to input: Vector<Float>) -> Float {
        return weight • input + bias
    }
}

Stored properties weight and bias are considered as trainable parameters, and are used to define the transformation from model inputs to model outputs. Models can be trained , during which parameters like weight are updated, thus changing the behavior of applied(to:). When a model is used, the call site looks just like a function call.

let model: Perceptron = ...
let ŷ = model.applied(to: x)

Many deep learning models are composed of layers, or layers of layers. In the definition of those models, repeated calls to applied(to:) significantly complicate the look of the program and reduce the clarity of the resulting code.

struct Model {
    var conv = Conv2D<Float>(filterShape: (5, 5, 3, 6))
    var maxPool = MaxPool2D<Float>(poolSize: (2, 2), strides: (2, 2))
    var flatten = Flatten<Float>()
    var dense = Dense<Float>(inputSize: 36 * 6, outputSize: 10)

    func applied(to input: Tensor<Float>) -> Tensor<Float> {
        return dense.applied(to: flatten.applied(to: maxPool.applied(to: conv.applied(to: input))))
    }
}

These repeated calls to applied(to:) harm clarity and makes code less readable. If model could be called like a function, which it mathematically represents, the definition of Model becomes much shorter and more concise. The proposed feature promotes clear usage by omitting needless words.

struct Model {
    var conv = Conv2D<Float>(filterShape: (5, 5, 3, 6))
    var maxPool = MaxPool2D<Float>(poolSize: (2, 2), strides: (2, 2))
    var flatten = Flatten<Float>()
    var dense = Dense<Float>(inputSize: 36 * 6, outputSize: 10)

    func callAsFunction(_ input: Tensor<Float>) -> Tensor<Float> {
        // Call syntax.
        return dense(flatten(maxPool(conv(input))))
    }
}

let model: Model = ...
let ŷ = model(x)

There are more ways to further simplify model definitions, but making models callable like functions is a good first step.

Domain specific languages

DSL constructs like string parsers represent functions from inputs to outputs. Parser combinators are often implemented as higher-order functions operating on parser values, which are themselves data structures—some implementations store closures, while some other efficient implementations store an expression tree. They all have an "apply"-like method that performs an application of the parser (i.e. parsing).

struct Parser<Output> {
    // Stored state...

    func applied(to input: String) throws -> Output {
        // Using the stored state...
    }

    func many() -> Parser<[Output]> { ... }
    func many<T>(separatedBy separator: Parser<T>) -> Parser<[Output]> { ... }
}

When using a parser, one would need to explicitly call applied(to:), but that is a bit cumbersome. Since parsers are like functions, it would be cleaner if parsers themself were callable.

func callAsFunction(_ input: String) throws -> Output {
    // Using the stored state...
}
let sexpParser: Parser<Expression> = ...
// Call syntax.
let sexp = sexpParser("(+ 1 2)")

A static counterpart to @dynamicCallable

SE-0216 introduced user-defined dynamically callable values. In its alternatives considered section, it was requested that we design and implement the "static callable" version of this proposal in conjunction with the dynamic version proposed. See its pitch thread for discussions about "static callables".

Prior art

Many languages offer the call syntax sugar:

Unifying compound types and nominal types

A long term goal with the type system is to unify compound types (e.g. function types and tuple types) and nominal types, to allow compound types to conform to protocols and have members. When function types can have members, it will be most natural for them to have a call method, which can help unify the compiler's type checking rules for call expressions.

Proposed design

We propose to introduce a syntactic sugar for values that have an instance method whose base name is call (a call method).

struct Adder {
    var base: Int
    func callAsFunction(_ x: Int) -> Int {
        return base + x
    }
}

Values that have a callAsFunction method can be called like a function, forwarding arguments to the callAsFunction method.

let add3 = Adder(base: 3)
add3(10) // => 13

Note: There are many alternative syntaxes for marking "call-syntax delegate methods". These are listed and explored in the "Alternatives considered" section.

Detailed design

callAsFunction methods

Instance methods whose base name is callAsFunction will be recognized as an implementation that makes a value of the enclosing type "callable" like a function.

When type-checking a call expression, the type checker will try to resolve the callee. Currently, the callee can be a value with a function type, a type name, or a value of a @dynamicCallable type. This proposal adds a fourth kind of a callee: a value with a matching callAsFunction method.

struct Adder {
    var base: Int

    func callAsFunction(_ x: Int) -> Int {
        return base + x
    }

    func callAsFunction(_ x: Float) -> Float {
        return Float(base) + x
    }

    func callAsFunction<T>(_ x: T, bang: Bool) throws -> T where T: BinaryInteger {
        if bang {
            return T(Int(exactly: x)! + base)
        } else {
            return T(Int(truncatingIfNeeded: x) + base)
        }
    }
}
let add1 = Adder(base: 1)
add1(2) // => 3
try add1(4, bang: true) // => 5

When type-checking fails, error messages look like those for function calls. When there is ambiguity, the compiler will show relevant callAsFunction method candidates.

add1("foo")
// error: cannot invoke 'add1' with an argument list of type '(String)'
// note: overloads for functions named 'callAsFunction' exist with these partially matching parameter lists: (Float), (Int)
add1(1, 2, 3)
// error: cannot invoke 'add1' with an argument list of type '(Int, Int, Int)'

Direct reference to callAsFunction

Since a callAsFunction method is a normal method, one can refer to a callAsFunction method using its declaration name and get a closure where self is captured. This is exactly how method references work today.

let add1 = Adder(base: 1)
let f1: (Int) -> Int = add1.callAsFunction
let f2: (Float) -> Float = add1.callAsFunction(_:)
let f3: (Int, Bool) throws -> Int = add1.callAsFunction(_:bang:)

When the type is also @dynamicCallable

A type can both have callAsFunction methods and be declared with @dynamicCallable. When type-checking a call expression, the type checker will first try to resolve the call to a function or initializer call, then a callAsFunction method call, and finally a dynamic call.

Implementation

The implementation is very simple and non-invasive: less than 200 lines of code in the type checker that performs lookup and expression rewrite.

let add1 = Adder(base: 1)
add1(0) // Rewritten to `add1.callAsFunction(0)` after type checking.

Source compatibility

This is a strictly additive proposal with no source-breaking changes.

Effect on ABI stability

This is a strictly additive proposal with no ABI-breaking changes.

Effect on API resilience

This has no impact on API resilience which is not already captured by other language features.

Future directions

Implicit conversions to function

A value cannot be implicitly converted to a function when the destination function type matches the type of the callAsFunction method. Since callAsFunction methods are normal methods, you can refer to them directly via .callAsFunction and get a function.

Implicit conversions impact the entire type system and require runtime support to work with dynamic casts; thus, further exploration is necessary for a formal proposal. This base proposal is self-contained; incremental proposals involving conversion can come later.

let h: (Int) -> Int = add1

A less controversial future direction is to support explicit conversion via as:

let h = add1 as (Int) -> Int

Function type as a constraint

On the pitch thread, Joe Groff brought up the possibility of allowing function types to be used as conformance constraints. Performance-minded programmers can define custom closure types where the closure context is not fully type-erased.

struct BoundClosure<T, F: (T) -> ()>: () -> () {
  var function: F
  var value: T

  func callAsFunction() { return function(value) }
}

let f = BoundClosure({ print($0) }, x) // instantiates BoundClosure<(underlying type of closure), Int>
f() // invokes call on BoundClosure

In this design, the function type constraint behaves like a protocol that requires a callAsFunction method whose parameter types are the same as the function type's parameter types.

Alternatives considered

Alternative names for call-syntax delegate methods

In addition to the word call in callAsFunction, there are other words that can be used to denote the function call syntax. The most common ones are apply and invoke as they are used to declare call-syntax delegate methods in other programming languages.

Both apply and invoke are good one-syllable English words that are technically correct, but we feel there are two concerns with these names:

  • They are officially completely new terminology to Swift. In the Functions chapter of The Swift Programming Language book, there is no mention of "apply" or "invoke" anywhere. Function calls are officially called "function calls".

  • They do not work very well with Swift's API naming conventions. According to Swift API Design Guidelines - Strive for Fluent Usage, functions should be named according to their side-effects.

    Those with side-effects should read as imperative verb phrases, e.g., print(x), x.sort(), x.append(y).

    Both apply and invoke are clearly imperative verbs. If call-syntax delegate methods must be named apply or invoke, their declarations and direct references will almost certainly read like a mutating function while they may not be.

    In contrast, call is both a noun and a verb. It is perfectly suited for describing the precise functionality while not having a strong implication about the function's side-effects.

    call

    • v. Cause (a subroutine) to be executed.
    • n. A command to execute a subroutine.

After the second round of proposal review, the core team accepted the proposal while revising the function base name to callFunction. The revision invoked a significant amount of bikeshedding on the thread. After considering the feedback, the core team decided to further revise the proposal to name the call-syntax delegate method callAsFunction.

Alternative ways to declare call-syntax delegate methods

Create a new declaration kind like subscript and init

Declarations that are associated with special invocation syntax often have their own declaration kind. For example, subscripts are implemented with a subscript declaration, and initialization calls are implemented with an init declaration. Since the function call syntax is first-class, one direction is to make the declaration be as first-class as possible.

struct Adder {
    var base: Int
    call(_ x: Int) -> Int {
        return base + x
    }
}

This alternative is in fact what's proposed in the first revision of this proposal, which got returned for revision.

Use unnamed func declarations to mark call-syntax delegate methods

struct Adder {
    var base: Int
    // Option: `func` with literally no name.
    func(_ x: Int) -> Int { ... }

    // Option: `func` with an underscore at the base name position.
    func _(_ x: Int) -> Int
    
    // Option: `func` with a `self` keyword at the base name position.
    func self(_ x: Int) -> Int

    // Option: `call` method modifier on unnamed `func` declarations.
    // Makes unnamed `func` less weird and clearly states "call".
    call func(_ x: Int) -> Int { ... }
}

This approach represents call-syntax delegate methods as unnamed func declarations instead of func callAsFunction.

One option is to use func(...) without an identifier name. Since the word "call" does not appear, it is less clear that this denotes a call-syntax delegate method. Additionally, it's not clear how direct references would work: the proposed design of referencing callAsFunction methods via foo.call is clear and consistent with the behavior of init declarations.

To make unnamed func(...) less weird, one option is to add a call declaration modifier: call func(...). The word call appears in both this option and the proposed design, clearly conveying "call-syntax delegate method". However, declaration modifiers are currently also treated as keywords, so with both approaches, parser changes to ensure source compatibility are necessary. call func(...) requires additional parser changes to allow func to sometimes not be followed by a name. The authors lean towards callAsFunction methods for simplicity and uniformity.

Use an attribute to mark call-syntax delegate methods

struct Adder {
    var base: Int
    @callDelegate
    func addingWithBase(_ x: Int) -> Int {
        return base + x
    }
}

This approach achieves a similar effect as callAsFunction methods, except that it allows call-syntax delegate methods to have a custom name and be directly referenced by that name. This is useful for types that want to make use of the call syntax sugar, but for which the name "call" does not accurately describe the callable functionality.

However, there are two concerns.

  • First, we feel that using a @callableMethod method attribute is more noisy, as many callable values do not need a special name for its call-syntax delegate methods.

  • Second, custom names often involve argument labels that form a phrase with the base name in order to be idiomatic. The grammaticality will be lost in the call syntax when the base name disappears.

    struct Layer {
    ...
    @callDelegate
    func applied(to x: Int) -> Int { ... }
}

let layer: Layer = ...
layer.applied(to: x) // Grammatical.
layer(to: x)         // Broken.

In contrast, standardizing on a specific name defines these problems away and makes this feature easier to use.

For reference: Other languages with callable functionality typically require call-syntax delegate methods to have a particular name (e.g. def __call__ in Python, def apply in Scala).

Use a type attribute to mark types with call-syntax delegate methods

@staticCallable // alternative name `@callable`; similar to `@dynamicCallable`
struct Adder {
    var base: Int
    // Informal rule: all methods with a particular name (e.g. `func callAsFunction`) are deemed call-syntax delegate methods.
    //
    // `StringInterpolationProtocol` has a similar informal requirement for
    // `func appendInterpolation` methods.
    // https://github.com/swiftlang/swift-evolution/blob/master/proposals/0228-fix-expressiblebystringinterpolation.md#proposed-solution
    func callAsFunction(_ x: Int) -> Int {
        return base + x
    }
}

We feel this approach is not ideal because a marker type attribute is not particularly meaningful. The call-syntax delegate methods of a type are what make values of that type callable - a type attribute means nothing by itself. There's also an unfortunate edge case that must be explicitly handled: if a @staticCallable type defines no call-syntax delegate methods, an error must be produced.

After the first round of review, the core team also did not think a type-level attribute is necessary.

After discussion, the core team doesn't think that a type level attribute is necessary, and there is no reason to limit this to primal type declarations - it is fine to add callable members (or overloads) in extensions, just as you can add subscripts to a type in extensions today.

Use a Callable protocol to represent callable types

// Compiler-known `Callable` marker protocol.
struct Adder: Callable {
    var base: Int
    // Informal rule: all methods with a particular name (e.g. `func callAsFunction`) are deemed call-syntax delegate methods.
    func callAsFunction(_ x: Int) -> Int {
        return base + x
    }
}

We feel this approach is not ideal for the same reasons as the marker type attribute. A marker protocol by itself is not meaningful and the name for call-syntax delegate methods is informal. Additionally, protocols should represent particular semantics, but call-syntax behavior has no inherent semantics.

Also allow static/class callAsFunction methods

Static callAsFunction methods could in theory look like initializers at the call site.

extension Adder {
    static func callAsFunction(base: Int) -> Int {
        ...
    }
    static func callAsFunction(_ x: Int) -> Int {
        ...
    }
}
Adder(base: 3) // error: ambiguous static member; do you mean `init(base:)` or `call(base:)`?
Adder(3) // okay, returns an `Int`, but it looks really like an initializer that returns an `Adder`.

This is an interesting direction, but parentheses followed by a type identifier often connote initialization and it is not source-compatible. We believe this would make call sites look very confusing.

Unify callable functionality with @dynamicCallable

Both @dynamicCallable and the proposed callAsFunction methods involve syntactic sugar related to function applications. However, the rules of the sugar are different, making unification difficult. In particular, @dynamicCallable provides a special sugar for argument labels that is crucial for usability.

// Let `PythonObject` be a `@dynamicMemberLookup` type with callable functionality.
let np: PythonObject = ...
// `PythonObject` with `@dynamicCallable.
np.random.randint(-10, 10, dtype: np.float)
// `PythonObject` with `callAsFunction` methods. The empty strings are killer.
np.random.randint(["": -10, "": 10, "dtype": np.float])