Notes on numerics in Swift

Concrete integer types, part 1


The Swift standard library provides ten concrete integer types, all defined as value types wrapping LLVM primitive types from the Builtin module:

public struct UInt8 /* ... */ {
  /* ... */
  public var _value: Builtin.Int8
  /* ... */

Four signed types are of explicit bit width (Int8, Int16, Int32, Int64), as are four unsigned types (UInt8, UInt16, UInt32, UInt64). LLVM does support 128-bit integer types on all platforms, but this support is not surfaced by the Swift standard library.

One signed type, Int, and one unsigned type, UInt, each has bit width equal to the platform’s native word size.

It is possible to determine the bit width of Int or UInt via the static or instance property named bitWidth, but there is no platform condition available in Swift to evaluate pointer bit width at compile time.

Standard library integer types simultaneously model integers and the sequences of bits that are used to represent them. In other words, some functions (such as those that perform basic arithmetic) operate on the integer value that is represented, while other functions operate on the two’s complement binary representation of those values.

All built-in standard library integer types use two’s complement representation internally for signed values. However, standard library protocols don’t preclude a hypothetical arbitrary-width integer type (“BigInt”) from using sign-and-magnitude or another representation internally, as long as bit shifting and bitwise operations notionally manipulate the two’s complement representation of any values.

The basic arithmetic infix operators +, -, *, /, %, and the prefix operator -, have behavior that will be largely familiar to users of other “C family” languages. In Swift, however, these operators trap on integer overflow. It can sometimes be overlooked that -T.min overflows if T is a signed (and fixed-width) integer type. As we will discuss, other functions are available that provide alternative overflow behavior.

Each of the basic arithmetic operators have a corresponding mutating (in-place) counterpart. For the infix operators, those are the assignment operators +=, -=, *=, /=, and %=. For prefix operator -, the mutating counterpart is spelled as the instance method negate(). These functions also trap on integer overflow.

Integer literals

An integer value can be represented in Swift source code as an integer literal. The same value can be written in any of four different bases; a prefix is used to indicate bases other than 10:

let x = 42        // Decimal (base 10).
let y = 0b101010  // Binary (base 2).
let z = 0o52      // Octal (base 8).
let a = 0x2a      // Hexadecimal (base 16).

In Swift, leading zeros don’t affect the base of an integer literal or its represented value. Underscores (_) can be used to group digits in numeric literals (e.g., 100_000_000); they too have no effect on the value that is represented.

In many other “C family” languages, leading 0 is used as an octal prefix. That is, numeric constants written with a leading 0 are interpreted in base 8. This has been a source of unintended error and confusion, and Swift does not perpetuate the convention.

Negative values can be represented by prepending the hyphen-minus character (-). This is considered to be part of the integer literal. In other words, the expression -42 is lexed as a single value, not as a call to the prefix operator - with 42 as its operand.

By contrast, the expression -(42) is lexed as a call to the prefix operator -. The distinction can be observed when working with floating-point types that support distinct representations for +0.0 and -0.0:

let x: Double = -0
x.sign // .plus

let y: Double = -(0)
y.sign // .minus

In the first example, a built-in 2048-bit integer is initialized to the value -0, but integer types do not support signed zero and the sign is ignored. Then, based on the type annotation, the integer value is converted to a positive floating-point value.

In the second example, a built-in 2048-bit integer is initialized to the value 0. Based on the type annotation, the value is converted to a positive floating-point value, then the floating-point prefix operator - is called to change the sign.

Note that the Swift standard library defines the prefix operator + for symmetry but does not consider a prepended + to be part of an integer literal. Therefore, +42 is lexed as a call to the prefix operator + with 42 as its operand.

In spite of these considerations, the member expression dot operator (.) binds more tightly than both the prefix operator - and the prepended literal -:

-(42.trailingZeroBitCount) // -1
(-42).trailingZeroBitCount //  1
-42.trailingZeroBitCount   // -1

Type inference

In Swift, literals have no type of their own. Instead, the type checker attempts to infer the type of a literal expression based on other available information such as explicit type annotations:

let x: Int8 = 42

Besides using an explicit type annotation, the type coercion operator as (which is to be distinguished from dynamic cast operators as?, as!, and is) can be used to provide information for type inference:

let x = 42 as Int8

In the absence of other available information, the inferred type of a literal expression defaults to IntegerLiteralType, which is a type alias for Int unless it is shadowed by the user:

typealias IntegerLiteralType = Int32
let x = 42
type(of: x) // Int32

The following caveat applies to current versions of Swift. It will not be applicable after changes described in SE-0213: Integer initialization via coercion, which was implemented in July 2018, are included in a future Swift release.

A frequent misunderstanding found even in the Swift project itself concerns the use of a type conversion initializer to indicate the desired type of a literal expression. For example:

// Avoid writing such code.
let x = Int8(42)

This usage frequently gives the intended result, but the function call does not provide information for type inference. Instead, this statement creates an instance of type IntegerLiteralType (which again, by default, is a type alias for Int) with the value 42, then converts this value to Int8.

The distinction can be demonstrated as follows:

let x = 32768 as Int16
// Causes a compile time error:
// integer literal '32768' overflows when stored into 'Int16'

let i = 32768 as Int
let y = Int16(i)
// Causes a **runtime** error:
// Not enough bits to represent a signed value

let z = Int16(32768)
// Causes a **runtime** error:
// Not enough bits to represent a signed value

Differences in diagnostics might be uninteresting to many users. However, the same misunderstanding with floating-point types can produce different results due to unintended rounding error:

let a = 3.14159265358979323846 as Float80
// 3.14159265358979323851

let b = Float80(3.14159265358979323846)
// 3.141592653589793116

Conversions among integer types

Five different initializers are available to convert between standard library integer types. A value source of type T can be converted to a value of type U as follows:

  1. U(source)
    Converts the given value if it can be represented exactly as a value of type U.
    Otherwise, a runtime error occurs.

  2. U(exactly: source)
    Failable initializer.
    Converts the given value if it can be represented exactly as a value of type U.
    Otherwise, returns nil.

  3. U(clamping: source)
    Converts the given value to the closest representable value of type U.
    If source > U.max, then the result is U.max.
    If source < U.min, then the result is U.min.

  4. U(truncatingIfNeeded: source)
    Creates a new value of type U from the binary representation in memory of source.
    When T and U are not of the same bit width, the binary representation of source is truncated or sign-extended as necessary.

  5. U(bitPattern: source)
    Available only for conversion between signed and unsigned types of explicitly the same bit width.
    Creates a new value of type U with the same binary representation in memory as that of source.

A note about init(truncatingIfNeeded:)

In previous versions of Swift, the same initializer was named init(extendingOrTruncating:). It was renamed to emphasize the potentially lossy semantics of truncation over the lossless semantics of sign-extension.

However, sign-extension produces what may be an unexpected result in one scenario: given a value source of type T, if source < 0, T.bitWidth < U.bitWidth, and U is an unsigned type, the result of U(truncatingIfNeeded: source) is equivalent to 0 &- U(truncatingIfNeeded: -source).

Concrete integer types, part 2

27 February–5 March 2018
Updated 18 August 2018