Skip to content

Latest commit

 

History

History
3672 lines (2986 loc) · 141 KB

README.md

File metadata and controls

3672 lines (2986 loc) · 141 KB
Raw

Language design

STATUS: Up-to-date on 09-Aug-2022, including proposals up through #1327.

Table of contents

Introduction

This documentation describes the design of the Carbon language, and the rationale for that design. This documentation is an overview of the Carbon project in its current state, written for the builders of Carbon and for those interested in learning more about Carbon.

This document is not a complete programming manual, and, nor does it provide detailed and comprehensive justification for design decisions. These descriptions are found in linked dedicated designs.

This document is provisional

This document includes much that is provisional or placeholder. This means that the syntax used, language rules, standard library, and other aspects of the design have things that have not been decided through the Carbon process. This preliminary material fills in gaps until aspects of the design can be filled in. Features that are provisional have been marked as such on a best-effort basis.

Tour of the basics

Here is a simple function showing some Carbon code:

import Math;

// Returns the smallest factor of `n` > 1, and
// whether `n` itself is prime.
fn SmallestFactor(n: i32) -> (i32, bool) {
  let limit: i32 = Math.Sqrt(n) as i32;
  var i: i32 = 2;
  while (i <= limit) {
    let remainder: i32 = n % i;
    if (remainder == 0) {
      Carbon.Print("{0} is a factor of {1}", i, n);
      return (i, false);
    }
    if (i == 2) {
      i = 3;
    } else {
      // Skip even numbers once we get past `2`.
      i += 2;
    }
  }
  return (n, true);
}

Carbon is a language that should feel familiar to C++ and C developers. This example has familiar constructs like imports, comments, function definitions, typed arguments, and expressions. Statements and declarations are terminated with a ; or something in curly braces {...}.

A few other features that are unlike C or C++ may stand out. First, declarations start with introducer keywords. fn introduces a function declaration, and var introduces a variable declaration.

The example starts with an import declaration. Carbon imports are more like C++ modules than textual inclusion during preprocessing using #include. The import declaration imports a library from a package. It must appear at the top of a Carbon source file, the first thing after the optional package declaration. Libraries can optionally be split into api and implementation files, like C++'s header and source files but without requiring a source file in any cases. This declaration from the example:

import Math;

imports the default library from package Math. The names from this library are accessible as members of Math, like Math.Sqrt. The Carbon.Print function comes from the Carbon package's prelude library which is imported by default. Unlike C++, the namespaces of different packages are kept separate, so there are no name conflicts.

Carbon comments must be on a line by themselves starting with //:

// Returns the smallest factor of `n` > 1, and
// whether `n` itself is prime.
...
      // Skip even numbers once we get past `2`.

A function definition consists of:

  • the fn keyword introducer,
  • the function's name,
  • a parameter list in round parens (...),
  • an optional -> and return type, and
  • a body inside curly braces {...}.
fn SmallestFactor(n: i32) -> (i32, bool) {
  ...
      return (i, false);
  ...
  return (n, true);
}

The body of the function is an ordered sequence of statements and declarations. Function execution ends when it reaches a return statement or the end of the function body. return statements can also specify an expression whose value is returned.

Here i32 refers to a signed integer type, with 32 bits, and bool is the boolean type. Carbon also has floating-point types like f32 and f64, and string types.

A variable declaration has three parts:

  • the var keyword introducer,
  • the name followed by a : and a type, declared the same way as a parameter in a function signature, and
  • an optional initializer.
  var i: i32 = 2;

You can modify the value of a variable with an assignment statement:

      i = 3;
      ...
      ++i;
      ...
      i += 2;

Constants are declared with the let keyword introducer. The syntax parallels variable declarations except the initializer is required:

  let limit: i32 = Math.Sqrt(n) as i32;
  ...
    let remainder: i32 = n % i;

The initializer Math.Sqrt(n) as i32 is an expression. It first calls the Math.Sqrt function with n as the argument. Then, the as operator casts the floating-point return value to i32. Lossy conversions like that must be done explicitly.

Other expressions include n % i, which applies the binary % modulo operator with n and i as arguments, and remainder == 0, which applies the == comparison operator producing a bool result. Expression return values are ignored when expressions are used as statements, as in this call to the Carbon.Print function:

      Carbon.Print("{0} is a factor of {1}", i, n);

Function calls consist of the name of the function followed by the comma-separated argument list in round parentheses (...).

Control flow statements, including if, while, for, break, and continue, change the order that statements are executed, as they do in C++:

  while (i <= limit) {
    ...
    if (remainder == 0) {
      ...
    }
    if (i == 2) {
      ...
    } else {
      ...
    }
  }

Every code block in curly braces {...} defines a scope. Names are visible from their declaration until the end of innermost scope containing it. So remainder in the example is visible until the curly brace } that closes the while.

The example function uses a tuple, a composite type, to return multiple values. Both tuple values and types are written using a comma-separated list inside parentheses. So (i, false) and (n, true) are tuple values, and (i32, bool) is their type.

Struct types are similar, except their members are referenced by name instead of position. The example could be changed to use structs instead as follows:

// Return type of `{.factor: i32, .prime: bool}` is a struct
// with an `i32` field named `.factor`, and a `bool` field
// named `.prime`.
fn SmallestFactor(n: i32) -> {.factor: i32, .prime: bool} {
  ...
    if (remainder == 0) {
      // Return a struct value.
      return {.factor = i, .prime = false};
    }
  ...
  // Return a struct value.
  return {.factor = n, .prime = true};
}

Code and comments

All source code is UTF-8 encoded text. Comments, identifiers, and strings are allowed to have non-ASCII characters.

var résultat: String = "Succès";

Comments start with two slashes // and go to the end of the line. They are required to be the only non-whitespace on the line.

// Compute an approximation of π

References:

Build modes

The behavior of the Carbon compiler depends on the build mode:

  • In a development build, the priority is diagnosing problems and fast build time.
  • In a performance build, the priority is fastest execution time and lowest memory usage.
  • In a hardened build, the first priority is safety and second is performance.

References: Safety strategy

Types are values

Expressions compute values in Carbon, and these values are always strongly typed much like in C++. However, an important difference from C++ is that types are themselves modeled as values; specifically, compile-time-constant values. This has a number of consequences:

  • Names for types are in the same namespace shared with functions, variables, namespaces, and so on.
  • The grammar for writing a type is the expression grammar, not a separate grammar for types. As a result, Carbon doesn't use angle brackets <...> in types, since < and > are used for comparison in expressions.
  • Function call syntax is used to specify parameters to a type, like HashMap(String, i64).

Some values, such as () and {}, may even be used as types, but only act like types when they are in a type position, like after a : in a variable declaration or the return type after a -> in a function declaration. Any expression in a type position must be a constant or symbolic value so the compiler can resolve whether the value can be used as a type. This also puts limits on how much operators can do different things for types. This is good for consistency, but is a significant restriction on Carbon's design.

Primitive types

Primitive types fall into the following categories:

  • the boolean type bool,
  • signed and unsigned integer types,
  • IEEE-754 floating-point types, and
  • string types.

These are made available through the prelude.

bool

The type bool is a boolean type with two possible values: true and false. The names bool, true, and false are keywords. Comparison expressions produce bool values. The condition arguments in control-flow statements, like if and while, and if-then-else conditional expressions take bool values.

References:

Integer types

The signed-integer type with bit width N may be written iN, as long as N is a positive multiple of 8. For example, i32 is a signed 32-bit integer. Signed-integer overflow is a programming error:

  • In a development build, overflow will be caught immediately when it happens at runtime.
  • In a performance build, the optimizer can assume that such conditions don't occur. As a consequence, if they do, the behavior of the program is not defined.
  • In a hardened build, overflow does not result in undefined behavior. Instead, either the program will be aborted, or the arithmetic will evaluate to a mathematically incorrect result, such as a two's complement result or zero.

The unsigned-integer types may be written uN, with N a positive multiple of 8. Unsigned integer types wrap around on overflow; we strongly advise that they are not used except when those semantics are desired. These types are intended for bit manipulation or modular arithmetic as often found in hashing, cryptography, and PRNG use cases. Values which can never be negative, like sizes, but for which wrapping does not make sense should use signed integer types.

Identifiers of the form iN and uN are type literals, resulting in the corresponding type.

Not all operations will be supported for all bit sizes. For example, division may be limited to integers of at most 128 bits due to LLVM limitations.

Open question: Bit-field (1, 2) support will need some way to talk about non-multiple-of-eight-bit integers, even though Carbon will likely not support pointers to those types.

References:

Integer literals

Integers may be written in decimal, hexadecimal, or binary:

  • 12345 (decimal)
  • 0x1FE (hexadecimal)
  • 0b1010 (binary)

Underscores (_) may be used as digit separators. Numeric literals are case-sensitive: 0x, 0b must be lowercase, whereas hexadecimal digits must be uppercase. Integer literals never contain a ..

Unlike in C++, literals do not have a suffix to indicate their type. Instead, numeric literals have a type derived from their value, and can be implicitly converted to any type that can represent that value.

References:

Floating-point types

Floating-point types in Carbon have IEEE-754 semantics, use the round-to-nearest rounding mode, and do not set any floating-point exception state. They are named with a type literals, consisting of f and the number of bits, which must be a multiple of 8. These types will always be available: f16, f32, and f64. Other sizes may be available, depending on the platform, such as f80, f128, or f256.

Carbon also supports the BFloat16 format, a 16-bit truncation of a "binary32" IEEE-754 format floating point number.

References:

Floating-point literals

Floating-point types along with user-defined types may initialized from real-number literals. Decimal and hexadecimal real-number literals are supported:

  • 123.456 (digits on both sides of the .)
  • 123.456e789 (optional + or - after the e)
  • 0x1.Ap123 (optional + or - after the p)

As with integer literals, underscores (_) may be used as digit separators. Real-number literals always have a period (.) and a digit on each side of the period. When a real-number literal is interpreted as a value of a floating-point type, its value is the representable real number closest to the value of the literal. In the case of a tie, the nearest value whose mantissa is even is selected.

References:

String types

Note: This is provisional, no design for string types has been through the proposal process yet.

There are two string types:

  • String - a byte sequence treated as containing UTF-8 encoded text.
  • StringView - a read-only reference to a byte sequence treated as containing UTF-8 encoded text.

There is an implicit conversion from String to StringView.

References:

String literals

String literals may be written on a single line using a double quotation mark (") at the beginning and end of the string, as in "example".

Multi-line string literals, called block string literals, begin and end with three single quotation marks ('''), and may have a file type indicator after the first '''.

// Block string literal:
var block: String = '''
    The winds grow high; so do your stomachs, lords.
    How irksome is this music to my heart!
    When such strings jar, what hope of harmony?
    I pray, my lords, let me compound this strife.
        -- History of Henry VI, Part II, Act II, Scene 1, W. Shakespeare
    ''';

The indentation of a block string literal's terminating line is removed from all preceding lines.

Strings may contain escape sequences introduced with a backslash (\). Raw string literals are available for representing strings with \s and "s.

References:

Values, objects, and expressions

Carbon has both abstract values and concrete objects. Carbon values are things like 42, true, and i32 (a type value). Carbon objects have storage where values can be read and written. Storage also allows taking the address of an object in memory in Carbon.

References:

Expression categories

A Carbon expression produces a value, references an object, or initializes an object. Every expression has a category, similar to C++:

  • Value expressions produce abstract, read-only values that cannot be modified or have their address taken.
  • Reference expressions refer to objects with storage where a value may be read or written and the object's address can be taken.
  • Initializing expressions which require storage to be provided implicitly when evaluating the expression. The expression then initializes an object in that storage. These are used to model function returns, which can construct the returned value directly in the caller's storage.

Expressions in one category can be converted to any other category when needed. The primitive conversion steps used are:

  • Value binding converts a reference expression into a value expression.
  • Direct initialization converts a value expression into an initializing expression.
  • Copy initialization converts a reference expression into an initializing expression.
  • Temporary materialization converts an initializing expression into a reference expression.

References:

Value phases

Value expressions are also broken down into three value phases:

  • A constant has a value known at compile time, and that value is available during type checking, for example to use as the size of an array. These include literals (integer, floating-point, string), concrete type values (like f64 or Optional(i32*)), expressions in terms of constants, and values of template parameters.
  • A symbolic value has a value that will be known at the code generation stage of compilation when monomorphization happens, but is not known during type checking. This includes checked-generic parameters, and type expressions with checked-generic arguments, like Optional(T*).
  • A runtime value has a dynamic value only known at runtime.

Carbon will automatically convert a constant to a symbolic value, or any value to a runtime value:

graph TD;
    A(constant)-->B(symbolic value)-->C(runtime value);
    D(reference expression)-->C;
Loading

Constants convert to symbolic values and to runtime values. Symbolic values will generally convert into runtime values if an operation that inspects the value is performed on them. Runtime values will convert into constants or to symbolic values if constant evaluation of the runtime expression succeeds.

Note: Conversion of runtime values to other phases is provisional.

Composite types

Tuples

A tuple is a fixed-size collection of values that can have different types, where each value is identified by its position in the tuple. An example use of tuples is to return multiple values from a function:

fn DoubleBoth(x: i32, y: i32) -> (i32, i32) {
  return (2 * x, 2 * y);
}

Breaking this example apart:

  • The return type is a tuple of two i32 types.
  • The expression uses tuple syntax to build a tuple of two i32 values.

Both of these are expressions using the tuple syntax (<expression>, <expression>). The only difference is the type of the tuple expression: one is a tuple of types, the other a tuple of values. In other words, a tuple type is a tuple of types.

The components of a tuple are accessed positionally, so element access uses subscript syntax, but the index must be a compile-time constant:

fn DoubleTuple(x: (i32, i32)) -> (i32, i32) {
  return (2 * x[0], 2 * x[1]);
}

Tuple types are structural.

Note: This is provisional, no design for tuples has been through the proposal process yet. Many of these questions were discussed in dropped proposal #111.

References: Tuples

Struct types

Carbon also has structural types whose members are identified by name instead of position. These are called structural data classes, also known as a struct types or structs.

Both struct types and values are written inside curly braces ({...}). In both cases, they have a comma-separated list of members that start with a period (.) followed by the field name.

  • In a struct type, the field name is followed by a colon (:) and the type, as in: {.name: String, .count: i32}.
  • In a struct value, called a structural data class literal or a struct literal, the field name is followed by an equal sign (=) and the value, as in {.key = "Joe", .count = 3}.

References:

Pointer types

The type of pointers-to-values-of-type-T is written T*. Carbon pointers do not support pointer arithmetic; the only pointer operations are:

There are no null pointers in Carbon. To represent a pointer that may not refer to a valid object, use the type Optional(T*).

Future work: Perhaps Carbon will have stricter pointer provenance or restrictions on casts between pointers and integers.

References:

Arrays and slices

The type of an array of holding 4 i32 values is written [i32; 4]. There is an implicit conversion from tuples to arrays of the same length as long as every component of the tuple may be implicitly converted to the destination element type. In cases where the size of the array may be deduced, it may be omitted, as in:

var i: i32 = 1;
// `[i32;]` equivalent to `[i32; 3]` here.
var a: [i32;] = (i, i, i);

Elements of an array may be accessed using square brackets ([...]), as in a[i]:

a[i] = 2;
Carbon.Print(a[0]);

TODO: Slices

Note: This is provisional, no design for arrays has been through the proposal process yet. See pending proposal #1928: Arrays.

Expressions

Expressions describe some computed value. The simplest example would be a literal number like 42: an expression that computes the integer value 42.

Some common expressions in Carbon include:

When an expression appears in a context in which an expression of a specific type is expected, implicit conversions are applied to convert the expression to the target type.

References:

Declarations, Definitions, and Scopes

Declarations introduce a new name and say what that name represents. For some kinds of entities, like functions, there are two kinds of declarations: forward declarations and definitions. For those entities, there should be exactly one definition for the name, and at most one additional forward declaration that introduces the name before it is defined, plus any number of declarations in a match_first block. Forward declarations can be used to separate interface from implementation, such as to declare a name in an api file that is defined in an impl file. Forward declarations also allow entities to be used before they are defined, such as to allow cyclic references. A name that has been declared but not defined is called incomplete, and in some cases there are limitations on what can be done with an incomplete name. Within a definition, the defined name is incomplete until the end of the definition is reached, but is complete in the bodies of member functions because they are parsed as if they appeared after the definition.

A name is valid until the end of the innermost enclosing scope. There are a few kinds of scopes:

  • the outermost scope, which includes the whole file,
  • scopes that are enclosed in curly braces ({...}), and
  • scopes that encompass a single declaration.

For example, the names of the parameters of a function or class are valid until the end of the declaration. The name of the function or class itself is visible until the end of the enclosing scope.

References:

Patterns

Note: This is provisional, no design for patterns has been through the proposal process yet.

A pattern says how to receive some data that is being matched against. There are two kinds of patterns:

  • Refutable patterns can fail to match based on the runtime value being matched.
  • Irrefutable patterns are guaranteed to match, so long as the code type-checks.

In the introduction, function parameters, variable var declarations, and constant let declarations use a "name : type" construction. That construction is an example of an irrefutable pattern, and in fact any irrefutable pattern may be used in those positions. match statements can include both refutable patterns and irrefutable patterns.

References:

Binding patterns

The most common irrefutable pattern is a binding pattern, consisting of a new name, a colon (:), and a type. It binds the matched value of that type to that name. It can only match values that may be implicitly converted to that type. A underscore (_) may be used instead of the name to match a value but without binding any name to it.

Binding patterns default to let bindings. The var keyword is used to make it a var binding.

  • A let binding binds a name to a value, so the name can be used as a value expression. This means the value cannot be modified, and its address generally cannot be taken.
  • A var binding creates an object with dedicated storage, and so the name can be used as a reference expression which can be modified and has a stable address.

A let-binding may be implemented as an alias for the original value (like a const reference in C++), or it may be copied from the original value (if it is copyable), or it may be moved from the original value (if it was a temporary). The Carbon implementation's choice among these options may be indirectly observable, for example through side effects of the destructor, copy, and move operations, but the program's correctness must not depend on which option the Carbon implementation chooses.

A generic binding uses :! instead of a colon (:) and can only match constant or symbolic values, not run-time values.

The keyword auto may be used in place of the type in a binding pattern, as long as the type can be deduced from the type of a value in the same declaration.

Destructuring patterns

There are also irrefutable destructuring patterns, such as tuple destructuring. A tuple destructuring pattern looks like a tuple of patterns. It may only be used to match tuple values whose components match the component patterns of the tuple. An example use is:

// `Bar()` returns a tuple consisting of an
// `i32` value and 2-tuple of `f32` values.
fn Bar() -> (i32, (f32, f32));

fn Foo() -> i64 {
  // Pattern in `var` declaration:
  var (p: i64, _: auto) = Bar();
  return p;
}

The pattern used in the var declaration destructures the tuple value returned by Bar(). The first component pattern, p: i64, corresponds to the first component of the value returned by Bar(), which has type i32. This is allowed since there is an implicit conversion from i32 to i64. The result of this conversion is assigned to the name p. The second component pattern, _: auto, matches the second component of the value returned by Bar(), which has type (f32, f32).

Refutable patterns

Additional kinds of patterns are allowed in match statements, that may or may not match based on the runtime value of the match expression:

  • An expression pattern is an expression, such as 42, whose value must be equal to match.
  • A choice pattern matches one case from a choice type, as described in the choice types section.
  • A dynamic cast pattern is tests the dynamic type, as described in inheritance.

See match for examples of refutable patterns.

References:

Name-binding declarations

There are two kinds of name-binding declarations:

  • constant declarations, introduced with let, and
  • variable declarations, introduced with var.

There are no forward declarations of these; all name-binding declarations are definitions.

Constant let declarations

A let declaration matches an irrefutable pattern to a value. In this example, the name x is bound to the value 42 with type i64:

let x: i64 = 42;

Here x: i64 is the pattern, which is followed by an equal sign (=) and the value to match, 42. The names from binding patterns are introduced into the enclosing scope.

References:

Variable var declarations

A var declaration is similar, except with var bindings, so x here is a reference expression for an object with storage and an address, and so may be modified:

var x: i64 = 42;
x = 7;

Variables with a type that has an unformed state do not need to be initialized in the variable declaration, but do need to be assigned before they are used.

References:

auto

If auto is used as the type in a var or let declaration, the type is the static type of the initializer expression, which is required.

var x: i64 = 2;
// The type of `y` is inferred to be `i64`.
let y: auto = x + 3;
// The type of `z` is inferred to be `bool`.
var z: auto = (y > 1);

References:

Global constants and variables

Constant let declarations may occur at a global scope as well as local and member scopes. However, there are currently no global variables.

Note: The semantics of global constant declarations and absence of global variable declarations is currently provisional.

We are exploring several different ideas for how to design less bug-prone patterns to replace the important use cases programmers still have for global variables. We may be unable to fully address them, at least for migrated code, and be forced to add some limited form of global variables back. We may also discover that their convenience outweighs any improvements afforded.

Functions

Functions are the core unit of behavior. For example, this is a forward declaration of a function that adds two 64-bit integers:

fn Add(a: i64, b: i64) -> i64;

Breaking this apart:

  • fn is the keyword used to introduce a function.
  • Its name is Add. This is the name added to the enclosing scope.
  • The parameter list in parentheses ((...)) is a comma-separated list of irrefutable patterns.
  • It returns an i64 result. Functions that return nothing omit the -> and return type.

You would call this function like Add(1, 2).

A function definition is a function declaration that has a body block instead of a semicolon:

fn Add(a: i64, b: i64) -> i64 {
  return a + b;
}

The names of the parameters are in scope until the end of the definition or declaration. The parameter names in a forward declaration may be omitted using _, but must match the definition if they are specified.

References:

Parameters

The bindings in the parameter list default to let bindings, and so the parameter names are treated as value expressions. This is appropriate for input parameters. This binding will be implemented using a pointer, unless it is legal to copy and copying is cheaper.

If the var keyword is added before the binding, then the arguments will be copied (or moved from a temporary) to new storage, and so can be mutated in the function body. The copy ensures that any mutations will not be visible to the caller.

Use a pointer parameter type to represent an input/output parameter, allowing a function to modify a variable of the caller's. This makes the possibility of those modifications visible: by taking the address using & in the caller, and dereferencing using * in the callee.

Outputs of a function should prefer to be returned. Multiple values may be returned using a tuple or struct type.

References:

auto return type

If auto is used in place of the return type, the return type of the function is inferred from the function body. It is set to common type of the static type of arguments to the return statements in the function. This is not allowed in a forward declaration.

// Return type is inferred to be `bool`, the type of `a > 0`.
fn Positive(a: i64) -> auto {
  return a > 0;
}

References:

Blocks and statements

A block is a sequence of statements. A block defines a scope and, like other scopes, is enclosed in curly braces ({...}). Each statement is terminated by a semicolon or block. Expressions, assignments and var and let are valid statements.

Statements within a block are normally executed in the order they appear in the source code, except when modified by control-flow statements.

The body of a function is defined by a block, and some control-flow statements have their own blocks of code. These are nested within the enclosing scope. For example, here is a function definition with a block of statements defining the body of the function, and a nested block as part of a while statement:

fn Foo() {
  Bar();
  while (Baz()) {
    Quux();
  }
}

References:

Control flow

Blocks of statements are generally executed sequentially. Control-flow statements give additional control over the flow of execution and which statements are executed.

Some control-flow statements include blocks. Those blocks will always be within curly braces {...}.

// Curly braces { ... } are required.
if (condition) {
  ExecutedWhenTrue();
} else {
  ExecutedWhenFalse();
}

This is unlike C++, which allows control-flow constructs to omit curly braces around a single statement.

References:

if and else

if and else provide conditional execution of statements. An if statement consists of:

  • An if introducer followed by a condition in parentheses. If the condition evaluates to true, the block following the condition is executed, otherwise it is skipped.
  • This may be followed by zero or more else if clauses, whose conditions are evaluated if all prior conditions evaluate to false, with a block that is executed if that evaluation is to true.
  • A final optional else clause, with a block that is executed if all conditions evaluate to false.

For example:

if (fruit.IsYellow()) {
  Carbon.Print("Banana!");
} else if (fruit.IsOrange()) {
  Carbon.Print("Orange!");
} else {
  Carbon.Print("Vegetable!");
}

This code will:

  • Print Banana! if fruit.IsYellow() is true.
  • Print Orange! if fruit.IsYellow() is false and fruit.IsOrange() is true.
  • Print Vegetable! if both of the above return false.

References:

Loops

References: Loops

while

while statements loop for as long as the passed expression returns true. For example, this prints 0, 1, 2, then Done!:

var x: i32 = 0;
while (x < 3) {
  Carbon.Print(x);
  ++x;
}
Carbon.Print("Done!");

References:

for

for statements support range-based looping, typically over containers. For example, this prints each String value in names:

for (var name: String in names) {
  Carbon.Print(name);
}

References:

break

The break statement immediately ends a while or for loop. Execution will continue starting from the end of the loop's scope. For example, this processes steps until a manual step is hit (if no manual step is hit, all steps are processed):

for (var step: Step in steps) {
  if (step.IsManual()) {
    Carbon.Print("Reached manual step!");
    break;
  }
  step.Process();
}

References:

continue

The continue statement immediately goes to the next loop of a while or for. In a while, execution continues with the while expression. For example, this prints all non-empty lines of a file, using continue to skip empty lines:

var f: File = OpenFile(path);
while (!f.EOF()) {
  var line: String = f.ReadLine();
  if (line.IsEmpty()) {
    continue;
  }
  Carbon.Print(line);
}

References:

return

The return statement ends the flow of execution within a function, returning execution to the caller.

// Prints the integers 1 .. `n` and then
// returns to the caller.
fn PrintFirstN(n: i32) {
  var i: i32 = 0;
  while (true) {
    i += 1;
    if (i > n) {
      // None of the rest of the function is
      // executed after a `return`.
      return;
    }
    Carbon.Print(i);
  }
}

If the function returns a value to the caller, that value is provided by an expression in the return statement. For example:

fn Sign(i: i32) -> i32 {
  if (i > 0) {
    return 1;
  }
  if (i < 0) {
    return -1;
  }
  return 0;
}

Assert(Sign(-3) == -1);

References:

returned var

To avoid a copy when returning a variable, add a returned prefix to the variable's declaration and use return var instead of returning an expression, as in:

fn MakeCircle(radius: i32) -> Circle {
  returned var c: Circle;
  c.radius = radius;
  // `return c` would be invalid because `returned` is in use.
  return var;
}

This is instead of the "named return value optimization" of C++.

References:

match

match is a control flow similar to switch of C and C++ and mirrors similar constructs in other languages, such as Swift. The match keyword is followed by an expression in parentheses, whose value is matched against the case declarations, each of which contains a refutable pattern, in order. The refutable pattern may optionally be followed by an if expression, which may use the names from bindings in the pattern.

The code for the first matching case is executed. An optional default block may be placed after the case declarations, it will be executed if none of the case declarations match.

An example match is:

fn Bar() -> (i32, (f32, f32));

fn Foo() -> f32 {
  match (Bar()) {
    case (42, (x: f32, y: f32)) => {
      return x - y;
    }
    case (p: i32, (x: f32, _: f32)) if (p < 13) => {
      return p * x;
    }
    case (p: i32, _: auto) if (p > 3) => {
      return p * Pi;
    }
    default => {
      return Pi;
    }
  }
}

Note: This is provisional, no design for match statements has been through the proposal process yet.

References:

User-defined types

Classes

Nominal classes, or just classes, are a way for users to define their own data structures or record types.

This is an example of a class definition:

class Widget {
  var x: i32;
  var y: i32;
  var payload: String;
}

Breaking this apart:

  • This defines a class named Widget. Widget is the name added to the enclosing scope.
  • The name Widget is followed by curly braces ({...}) containing the class body, making this a definition. A forward declaration would instead have a semicolon(;).
  • Those braces delimit the class' scope.
  • Fields, or instances variables, are defined using var declarations. Widget has two i32 fields (x and y), and one String field (payload).

The order of the field declarations determines the fields' memory-layout order.

Classes may have other kinds of members beyond fields declared in its scope:

Within the scope of a class, the unqualified name Self can be used to refer to the class itself.

Members of a class are accessed using the dot (.) notation, so given an instance dial of type Widget, dial.payload refers to its payload field.

Both structural data classes and nominal classes are considered class types, but they are commonly referred to as "structs" and "classes" respectively when that is not confusing. Like structs, classes refer to their members by name. Unlike structs, classes are nominal types.

References:

Assignment

There is an implicit conversions defined between a struct literal and a class type with the same fields, in any scope that has access to all of the class' fields. This may be used to assign or initialize a variable with a class type, as in:

var sprocket: Widget = {.x = 3, .y = 4, .payload = "Sproing"};
sprocket = {.x = 2, .y = 1, .payload = "Bounce"};

References:

Class functions and factory functions

Classes may also contain class functions. These are functions that are accessed as members of the type, like static member functions in C++, as opposed to methods that are members of instances. They are commonly used to define a function that creates instances. Carbon does not have separate constructors like C++ does.

class Point {
  // Class function that instantiates `Point`.
  // `Self` in class scope means the class currently being defined.
  fn Origin() -> Self {
    return {.x = 0, .y = 0};
  }
  var x: i32;
  var y: i32;
}

Note that if the definition of a function is provided inside the class scope, the body is treated as if it was defined immediately after the outermost class definition. This means that members such as the fields will be considered declared even if their declarations are later in the source than the class function.

The returned var feature can be used if the address of the instance being created is needed in a factory function, as in:

class Registered {
  fn Make() -> Self {
    returned var result: Self = {...};
    StoreMyPointerSomewhere(&result);
    return var;
  }
}

This approach can also be used for types that can't be copied or moved.

References:

Methods

Class type definitions can include methods:

class Point {
  // Method defined inline
  fn Distance[self: Self](x2: i32, y2: i32) -> f32 {
    var dx: i32 = x2 - self.x;
    var dy: i32 = y2 - self.y;
    return Math.Sqrt(dx * dx + dy * dy);
  }
  // Mutating method declaration
  fn Offset[addr self: Self*](dx: i32, dy: i32);

  var x: i32;
  var y: i32;
}

// Out-of-line definition of method declared inline
fn Point.Offset[addr self: Self*](dx: i32, dy: i32) {
  self->x += dx;
  self->y += dy;
}

var origin: Point = {.x = 0, .y = 0};
Assert(Math.Abs(origin.Distance(3, 4) - 5.0) < 0.001);
origin.Offset(3, 4);
Assert(origin.Distance(3, 4) == 0.0);

This defines a Point class type with two integer data members x and y and two methods Distance and Offset:

  • Methods are defined as class functions with a self parameter inside square brackets [...] before the regular explicit parameter list in parens (...).
  • Methods are called using the member syntax, origin.Distance(...) and origin.Offset(...).
  • Distance computes and returns the distance to another point, without modifying the Point. This is signified using [self: Self] in the method declaration.
  • origin.Offset(...) does modify the value of origin. This is signified using [addr self: Self*] in the method declaration. Since calling this method requires taking the non-const address of origin, it may only be called on reference expressions.
  • Methods may be declared lexically inline like Distance, or lexically out of line like Offset.

References:

Inheritance

The philosophy of inheritance support in Carbon is to focus on use cases where inheritance is a good match, and use other features for other cases. For example, mixins for implementation reuse and generics for separating interface from implementation. This allows Carbon to move away from multiple inheritance, which doesn't have as efficient of an implementation strategy.

Classes by default are final, which means they may not be extended. A class may be declared as allowing extension using either the base class or abstract class introducer instead of class. An abstract class is a base class that may not itself be instantiated.

base class MyBaseClass { ... }

Either kind of base class may be extended to get a derived class. Derived classes are final unless they are themselves declared base or abstract. Classes may only extend a single class. Carbon only supports single inheritance, and will use mixins instead of multiple inheritance.

base class MiddleDerived {
  extend base: MyBaseClass;
  ...
}
class FinalDerived {
  extend base: MiddleDerived;
  ...
}
// ❌ Forbidden: class Illegal { extend base: FinalDerived; ... }
// may not extend `FinalDerived` since not declared `base` or `abstract`.

A base class may define virtual methods. These are methods whose implementation may be overridden in a derived class. By default methods are non-virtual, the declaration of a virtual method must be prefixed by one of these three keywords:

  • A method marked virtual has a definition in this class but not in any base.
  • A method marked abstract does not have a definition in this class, but must have a definition in any non-abstract derived class.
  • A method marked impl has a definition in this class, overriding any definition in a base class.

A pointer to a derived class may be cast to a pointer to one of its base classes. Calling a virtual method through a pointer to a base class will use the overriding definition provided in the derived class. Base classes with virtual methods may use run-time type information in a match statement to dynamically test whether the dynamic type of a value is some derived class, as in:

var base_ptr: MyBaseType* = ...;
match (base_ptr) {
  case dyn p: MiddleDerived* => { ... }
}

For purposes of construction, a derived class acts like its first field is called base with the type of its immediate base class.

class MyDerivedType {
  extend base: MyBaseType;
  fn Make() -> MyDerivedType {
    return {.base = MyBaseType.Make(), .derived_field = 7};
  }
  var derived_field: i32;
}

Abstract classes can't be instantiated, so instead they should define class functions returning partial Self. Those functions should be marked protected so they may only be used by derived classes.

abstract class AbstractClass {
  protected fn Make() -> partial Self {
    return {.field_1 = 3, .field_2 = 9};
  }
  // ...
  var field_1: i32;
  var field_2: i32;
}
// ❌ Error: can't instantiate abstract class
var abc: AbstractClass = ...;

class DerivedFromAbstract {
  extend base: AbstractClass;
  fn Make() -> Self {
    // AbstractClass.Make() returns a
    // `partial AbstractClass` that can be used as
    // the `.base` member when constructing a value
    // of a derived class.
    return {.base = AbstractClass.Make(),
            .derived_field = 42 };
  }

  var derived_field: i32;
}

References:

Access control

Class members are by default publicly accessible. The private keyword prefix can be added to the member's declaration to restrict it to members of the class or any friends. A private virtual or private abstract method may be implemented in derived classes, even though it may not be called.

Friends may be declared using a friend declaration inside the class naming an existing function or type. Unlike C++, friend declarations may only refer to names resolvable by the compiler, and don't act like forward declarations.

protected is like private, but also gives access to derived classes.

References:

Destructors

A destructor for a class is custom code executed when the lifetime of a value of that type ends. They are defined with the destructor keyword followed by either [self: Self] or [addr self: Self*] (as is done with methods) and the block of code in the class definition, as in:

class MyClass {
  destructor [self: Self] { ... }
}

or:

class MyClass {
  // Can modify `self` in the body.
  destructor [addr self: Self*] { ... }
}

The destructor for a class is run before the destructors of its data members. The data members are destroyed in reverse order of declaration. Derived classes are destroyed before their base classes.

A destructor in an abstract or base class may be declared virtual like with methods. Destructors in classes derived from one with a virtual destructor must be declared with the impl keyword prefix. It is illegal to delete an instance of a derived class through a pointer to a base class unless the base class is declared virtual or impl. To delete a pointer to a non-abstract base class when it is known not to point to a value with a derived type, use UnsafeDelete.

References:

const

For every type MyClass, there is the type const MyClass such that:

  • The data representation is the same, so a MyClass* value may be implicitly converted to a (const MyClass)*.
  • A const MyClass reference expression may automatically convert to a MyClass value expression, the same way that a MyClass reference expression can.
  • If member x of MyClass has type T, then member x of const MyClass has type const T.
  • While all of the member names in MyClass are also member names in const MyClass, the effective API of a const MyClass reference expression is a subset of MyClass, because only addr methods accepting a const Self* will be valid.

Note that const binds more tightly than postfix-* for forming a pointer type, so const MyClass* is equal to (const MyClass)*.

This example uses the definition of Point from the "methods" section:

var origin: Point = {.x = 0, .y = 0};

// ✅ Allowed conversion from `Point*` to
// `const Point*`:
let p: const Point* = &origin;

// ✅ Allowed conversion of `const Point` reference expression
// to `Point` value expression.
let five: f32 = p->Distance(3, 4);

// ❌ Error: mutating method `Offset` excluded
// from `const Point` API.
p->Offset(3, 4);

// ❌ Error: mutating method `AssignAdd.Op`
// excluded from `const i32` API.
p->x += 2;

References:

Unformed state

Types indicate that they support unformed states by implementing a particular interface, otherwise variables of that type must be explicitly initialized when they are declared.

An unformed state for an object is one that satisfies the following properties:

  • Assignment from a fully formed value is correct using the normal assignment implementation for the type.
  • Destruction must be correct using the type's normal destruction implementation.
  • Destruction must be optional. The behavior of the program must be equivalent whether the destructor is run or not for an unformed object, including not leaking resources.

A type might have more than one in-memory representation for the unformed state, and those representations may be the same as valid fully formed values for that type. For example, all values are legal representations of the unformed state for any type with a trivial destructor like i32. Types may define additional initialization for the hardened build mode. For example, this causes integers to be set to 0 when in unformed state in this mode.

Any operation on an unformed object other than destruction or assignment from a fully formed value is an error, even if its in-memory representation is that of a valid value for that type.

References:

Move

Carbon will allow types to define if and how they are moved. This can happen when returning a value from a function or by using the move operator ~x. This leaves x in an unformed state and returns its old value.

Note: This is provisional. The move operator was discussed but not proposed in accepted proposal #257: Initialization of memory and variables.

Mixins

Mixins allow reuse with different trade-offs compared to inheritance. Mixins focus on implementation reuse, such as might be done using CRTP or multiple inheritance in C++.

TODO: The design for mixins is still under development. The details here are provisional. The mixin use case was included in accepted proposal #561: Basic classes: use cases, struct literals, struct types, and future work.

Choice types

A choice type is a tagged union, that can store different types of data in a storage space that can hold the largest. A choice type has a name, and a list of cases separated by commas (,). Each case has a name and an optional parameter list.

choice IntResult {
  Success(value: i32),
  Failure(error: String),
  Cancelled
}

The value of a choice type is one of the cases, plus the values of the parameters to that case, if any. A value can be constructed by naming the case and providing values for the parameters, if any:

fn ParseAsInt(s: String) -> IntResult {
  var r: i32 = 0;
  for (c: i32 in s) {
    if (not IsDigit(c)) {
      // Equivalent to `IntResult.Failure(...)`
      return .Failure("Invalid character");
    }
    // ...
  }
  return .Success(r);
}

Choice type values may be consumed using a match statement:

match (ParseAsInt(s)) {
  case .Success(value: i32) => {
    return value;
  }
  case .Failure(error: String) => {
    Display(error);
  }
  case .Cancelled => {
    Terminate();
  }
}

They can also represent an enumerated type, if no additional data is associated with the choices, as in:

choice LikeABoolean { False, True }

References:

Names

Names are introduced by declarations and are valid until the end of the scope in which they appear. Code may not refer to names earlier in the source than they are declared. In executable scopes such as function bodies, names declared later are not found. In declarative scopes such as packages, classes, and interfaces, it is an error to refer to names declared later, except that inline class member function bodies are parsed as if they appeared after the class.

A name in Carbon is formed from a sequence of letters, numbers, and underscores, and starts with a letter. We intend to follow Unicode's Annex 31 in selecting valid identifier characters, but a concrete set of valid characters has not been selected yet.

References:

Files, libraries, packages

  • Files are grouped into libraries, which are in turn grouped into packages.
  • Libraries are the granularity of code reuse through imports.
  • Packages are the unit of distribution.

Each library must have exactly one api file. This file includes declarations for all public names of the library. Definitions for those declarations must be in some file in the library, either the api file or an impl file.

Every package has its own namespace. This means libraries within a package need to coordinate to avoid name conflicts, but not across packages.

References:

Package declaration

Files start with an optional package declaration, consisting of:

  • optionally, the package keyword followed by an identifier specifying the package name,
  • optionally, the library keyword followed by a string with the library name,
  • either api or impl, and
  • a terminating semicolon (;).

For example:

// Package name is `Geometry`.
// Library name is "Shapes".
// This file is an `api` file, not an `impl` file.
package Geometry library "Shapes" api;

Parts of this declaration may be omitted:

  • If the package keyword is not specified, as in library "Widgets" api;, the file contributes to the Main package. No other package may import from the Main package, and it cannot be named explicitly.

  • If the library keyword is not specified, as in package Geometry api;, this file contributes to the default library.

  • If both keywords are omitted, the package declaration must be omitted entirely. In this case, the file is an impl file belonging to the default library of the Main package, which implicitly has an empty api file. This library is used to define the entry point for the program, and tests and smaller examples may choose to reside entirely within this library. No other library can import this library even from within the default package.

If the default library of the Main package contains a function named Run, that function is the program entry point. Otherwise, the program's entry point may be defined in another language, such as by defining a C++ main function.

Note: Valid signatures for the entry point have not yet been decided.

References:

Imports

After the package declaration, files may include import declarations. The import keyword is followed by the package name, library followed by the library name, or both. If the library is omitted, the default library for that package is imported.

All import declarations must appear before all other non-package declarations in the file.

Same-package imports

The package name must be omitted when importing a library from the current package.

// Import the "Vertex" library from the package containing this file.
import library "Vertex";

The import library ... syntax adds all the public top-level names within the given library to the top-level scope of the current file as private names, and similarly for names in namespaces.

Every impl file automatically imports the api file for its library. Attempting to perform an import of the current library is invalid.

package MyPackage library "Widgets" impl;

// ❌ Error, this import is performed implicitly.
import MyPackage library "Widgets";

The default library for a package does not have a string name, and is instead named with the default keyword.

// Import the default library from the same package.
import library default;

It is an error to use the import library default; syntax in the Main package.

Cross-package imports

When the package name is specified, the import declaration imports a library from another package.

package MyPackage impl;

// Import the "Vector" library from the `LinearAlgebra` package.
import LinearAlgebra library "Vector";

// Import the default library from the `ArbitraryPrecision` package.
import ArbitraryPrecision;

The syntax import PackageName ... introduces the name PackageName as a private name naming the given package. Importing additional libraries from that package makes additional members of PackageName visible.

It is an error to specify the name of the current package. The package name must be omitted when importing from the same package.

It is an error to specify library default in a package-qualified import. Instead, omit the library portion of the declaration.

It is an error to specify the package name Main. Libraries in the Main package can only be imported from within that package.

References:

Name visibility

The names visible from an imported library are determined by these rules:

  • Declarations in an api file are by default public, which means visible to any file that imports that library. This matches class members, which are also default public.
  • A private prefix on a declaration in an api file makes the name library private. This means the name is visible in the file and all impl files for the same library.
  • The visibility of a name is determined by its first declaration, considering api files before impl files. The private prefix is only allowed on the first declaration.
  • A name declared in an impl file and not the corresponding api file is file private, meaning visible in just that file. Its first declaration must be marked with a private prefix. TODO: This needs to be finalized in a proposal to resolve inconsistency between #665 and #1136.
  • Private names don't conflict with names outside the region they're private to: two different libraries can have different private names foo without conflict, but a private name conflicts with a public name in the same scope.

At most one api file in a package transitively used in a program may declare a given name public.

References:

Package scope

The top-level scope in a file is the scope of the package. This means:

  • Within this scope (and its sub-namespaces), all visible names from the same package appear. This includes names from the same file, names from the api file of a library when inside an impl file, and names from imported libraries of the same package.
  • In scopes where package members might have a name conflict with something else, the syntax package.Foo can be used to name the Foo member of the current package.

In this example, the names F and P are used in a scope where they could mean two different things, and qualifications are needed to disambiguate:

import P;
fn F();
class C {
  fn F();
  class P {
    fn H();
  }
  fn G() {
    // ❌ Error: ambiguous whether `F` means
    // `package.F` or `package.C.F`.
    F();
   // ✅ Allowed: fully qualified
    package.F();
    package.C.F();
    // ✅ Allowed: unambiguous
    C.F();
    // ❌ Error: ambiguous whether `P` means
    // `package.P` or `package.C.P`.
    P.H();
    // ✅ Allowed
    package.P.H();
    package.C.P.H();
    C.P.H();
  }
}

References:

Namespaces

A namespace declaration defines a name that may be used as a prefix of names declared afterward. When defining a member of a namespace, other members of that namespace are considered in scope and may be found by name lookup without the namespace prefix. In this example, package P defines some of its members inside a namespace N:

package P api;

// Defines namespace `N` within the current package.
namespace N;

// Defines namespaces `M` and `M.L`.
namespace M.L;

fn F();
// ✅ Allowed: Declares function `G` in namespace `N`.
private fn N.G();
// ❌ Error: `Bad` hasn't been declared.
fn Bad.H();

fn J() {
  // ❌ Error: No `package.G`
  G();
}

fn N.K() {
  // ✅ Allowed: Looks in both `package` and `package.N`.
  // Finds `package.F` and `package.N.G`.
  F();
  G();
}

// ✅ Allowed: Declares function `R` in namespace `M.L`.
fn M.L.R();
// ✅ Allowed: Declares function `Q` in namespace `M`.
fn M.Q();

Another package importing P can refer to the public members of that namespace by prefixing with the package name P followed by the namespace:

import P;

// ✅ Allowed: `F` is public member of `P`.
P.F();
// ❌ Error: `N.G` is a private member of `P`.
P.N.G();
// ✅ Allowed: `N.K` is public member of `P`.
P.N.K();
// ✅ Allowed: `M.L.R` is public member of `P`.
P.M.L.R();
// ✅ Allowed: `M.Q` is public member of `P`.
P.M.Q();

References:

Naming conventions

Our naming conventions are:

  • For idiomatic Carbon code:
    • UpperCamelCase will be used when the named entity cannot have a dynamically varying value. For example, functions, namespaces, or compile-time constant values. Note that virtual methods are named the same way to be consistent with other functions and methods.
    • lower_snake_case will be used when the named entity's value won't be known until runtime, such as for variables.
  • For Carbon-provided features:
    • Keywords and type literals will use lower_snake_case.
    • Other code will use the conventions for idiomatic Carbon code.

References:

Aliases

alias declares a name as equivalent to another name, for example:

alias NewName = SomePackage.OldName;

Note that the right-hand side of the equal sign (=) is a name not a value, so alias four = 4; is not allowed. This allows alias to work with entities like namespaces, which aren't values in Carbon.

This can be used during an incremental migration when changing a name. For example, alias would allow you to have two names for a data field in a class while clients were migrated between the old name and the new name.

class MyClass {
  var new_name: String;
  alias old_name = new_name;
}

var x: MyClass = {.new_name = "hello"};
Carbon.Assert(x.old_name == "hello");

Another use is to include a name in a public API. For example, alias may be used to include a name from an interface implementation as a member of a class or named constraint, possibly renamed:

class ContactInfo {
  impl as Printable;
  impl as ToPrinterDevice;
  alias PrintToScreen = Printable.Print;
  alias PrintToPrinter = ToPrinterDevice.Print;
  ...
}

References:

Name lookup

The general principle of Carbon name lookup is that we look up names in all relevant scopes, and report an error if the name is found to refer to more than one different entity. So Carbon requires disambiguation by adding qualifiers instead of doing any shadowing of names. For an example, see the "package scope" section.

Unqualified name lookup walks the semantically-enclosing scopes, not only the lexically-enclosing ones. So when a lookup is performed within fn MyNamespace.MyClass.MyNestedClass.MyFunction(), we will look in MyNestedClass, MyClass, MyNamespace, and the package scope, even when the lexically-enclosing scope is the package scope. This means that the definition of a method will look for names in the class' scope even if it is written lexically out of line:

class C {
  fn F();
  fn G();
}
fn C.G() {
  // ✅ Allowed: resolves to `package.C.F`.
  F();
}

Member name lookup follows a similar philosophy. If a checked-generic type parameter is known to implement multiple interfaces due to a constraint using & or where clauses, member name lookup into that type will look in all of the interfaces. If it is found in multiple, the name must be disambiguated by qualifying using compound member access (1, 2). A template-generic type parameter performs look up into the caller's type in addition to the constraint.

Carbon also rejects cases that would be invalid if all declarations in the file, including ones appearing later, were visible everywhere, not only after their point of appearance:

class C {
  fn F();
  fn G();
}
fn C.G() {
  F();
}
// Error: use of `F` in `C.G` would be ambiguous
// if this declaration was earlier.
fn F();

References:

Name lookup for common types

Common types that we expect to be used universally will be provided for every file are made available as if there was a special "prelude" package that was imported automatically into every api file. Dedicated type literal syntaxes like i32 and bool refer to types defined within this package, based on the "all APIs are library APIs" principle.

TODO: Prelude provisionally imports the Carbon package which includes common facilities, like Print and the interfaces for operator overloading.

References:

Generics

Generics allow Carbon constructs like functions and classes to be written with compile-time parameters and apply generically to different types using those parameters. For example, this Min function has a type parameter T that can be any type that implements the Ordered interface.

fn Min[T:! Ordered](x: T, y: T) -> T {
  // Can compare `x` and `y` since they have
  // type `T` known to implement `Ordered`.
  return if x <= y then x else y;
}

var a: i32 = 1;
var b: i32 = 2;
// `T` is deduced to be `i32`
Assert(Min(a, b) == 1);
// `T` is deduced to be `String`
Assert(Min("abc", "xyz") == "abc");

Since the T type parameter is in the deduced parameter list in square brackets ([...]) before the explicit parameter list in parentheses ((...)), the value of T is determined from the types of the explicit arguments instead of being passed as a separate explicit argument.

References: TODO: Revisit

Checked and template parameters

The :! indicates that the T parameter is generic, and therefore bound at compile time. Generic parameters may either be checked or template, and default to checked.

"Checked" here means that the body of Min is type checked when the function is defined, independent of the specific type values T is instantiated with, and name lookup is delegated to the constraint on T (Ordered in this case). This type checking is equivalent to saying the function would pass type checking given any type T that implements the Ordered interface. Subsequent calls to Min only need to check that the deduced type value of T implements Ordered.

The parameter could alternatively be declared to be a template generic parameter by prefixing with the template keyword, as in template T:! type.

fn Convert[template T:! type](source: T, template U:! type) -> U {
  var converted: U = source;
  return converted;
}

fn Foo(i: i32) -> f32 {
  // Instantiates with the `T` implicit argument set to `i32` and the `U`
  // explicit argument set to `f32`, then calls with the runtime value `i`.
  return Convert(i, f32);
}

A template parameter can still use a constraint. The Min example could have been declared as:

fn Min[template T:! Ordered](x: T, y: T) -> T {
  return if x <= y then x else y;
}

Carbon templates follow the same fundamental paradigm as C++ templates: they are instantiated when called, resulting in late type checking, duck typing, and lazy binding.

One difference from C++ templates, Carbon template instantiation is not controlled by the SFINAE rule of C++ (1, 2) but by explicit if clauses evaluated at compile-time. The if clause is at the end of the declaration, and the condition can only use constant values known at type-checking time, including template parameters.

class Array(template T:! type, template N:! i64)
    if N >= 0 and N < MaxArraySize / sizeof(T);

Member lookup into a template type parameter is done in the actual type value provided by the caller, in addition to any constraints. This means member name lookup and type checking for anything dependent on the template parameter can't be completed until the template is instantiated with a specific concrete type. When the constraint is just type, this gives semantics similar to C++ templates. Constraints can then be added incrementally, with the compiler verifying that the semantics stay the same. Once all constraints have been added, removing the word template to switch to a checked parameter is safe.

The value phase of a checked parameter is a symbolic value whereas the value phase of a template parameter is constant.

Although checked generics are generally preferred, templates enable translation of code between C++ and Carbon, and address some cases where the type checking rigor of generics are problematic.

References:

Interfaces and implementations

Interfaces specify a set of requirements that a types might satisfy. Interfaces act both as constraints on types a caller might supply and capabilities that may be assumed of types that satisfy that constraint.

interface Printable {
  // Inside an interface definition `Self` means
  // "the type implementing this interface".
  fn Print[self: Self]();
}

In addition to function requirements, interfaces can contain:

Types only implement an interface if there is an explicit impl declaration that they do. Simply having a Print function with the right signature is not sufficient.

class Circle {
  var radius: f32;

  extend impl as Printable {
    fn Print[self: Self]() {
      Carbon.Print("Circle with radius: {0}", self.radius);
    }
  }
}

In this case, Print is a member of Circle. Interfaces may also be implemented externally, which means the members of the interface are not direct members of the type. Those methods may still be called using compound member access syntax (1, 2) to qualify the name of the member, as in x.(Printable.Print)(). External implementations don't have to be in the same library as the type definition, subject to the orphan rule (1, 2) for coherence.

Interfaces and implementations may be forward declared by replacing the definition scope in curly braces ({...}) with a semicolon.

References:

Combining constraints

A function can require calling types to implement multiple interfaces by combining them using an ampersand (&):

fn PrintMin[T:! Ordered & Printable](x: T, y: T) {
  // Can compare since type `T` implements `Ordered`.
  if (x <= y) {
    // Can call `Print` since type `T` implements `Printable`.
    x.Print();
  } else {
    y.Print();
  }
}

The body of the function may call functions that are in either interface, except for names that are members of both. In that case, use the compound member access syntax (1, 2) to qualify the name of the member, as in:

fn DrawTies[T:! Renderable & GameResult](x: T) {
  if (x.(GameResult.Draw)()) {
    x.(Renderable.Draw)();
  }
}

References:

Associated types

An associated type is a type member of an interface whose value is determined by the implementation of that interface for a specific type. These values are set to compile-time values in implementations, and so use the :! generic syntax inside a let declaration without an initializer. This allows types in the signatures of functions in the interface to vary. For example, an interface describing a stack might use an associated type to represent the type of elements stored in the stack.

interface StackInterface {
  let ElementType:! Movable;
  fn Push[addr self: Self*](value: ElementType);
  fn Pop[addr self: Self*]() -> ElementType;
  fn IsEmpty[addr self: Self*]() -> bool;
}

Then different types implementing StackInterface can specify different type values for the ElementType member of the interface using a where clause:

class IntStack {
  extend impl as StackInterface where .ElementType = i32 {
    fn Push[addr self: Self*](value: i32);
    // ...
  }
}

class FruitStack {
  extend impl as StackInterface where .ElementType = Fruit {
    fn Push[addr self: Self*](value: Fruit);
    // ...
  }
}

References:

Generic entities

Many Carbon entities, not just functions, may be made generic by adding checked or template parameters.

Generic Classes

Classes may be defined with an optional explicit parameter list. All parameters to a class must be generic, and so defined with :!, either with or without the template prefix. For example, to define a stack that can hold values of any type T:

class Stack(T:! type) {
  fn Push[addr self: Self*](value: T);
  fn Pop[addr self: Self*]() -> T;

  var storage: Array(T);
}

var int_stack: Stack(i32);

In this example:

  • Stack is a type parameterized by a type T.
  • T may be used within the definition of Stack anywhere a normal type would be used.
  • Array(T) instantiates generic type Array with its parameter set to T.
  • Stack(i32) instantiates Stack with T set to i32.

The values of type parameters are part of a type's value, and so may be deduced in a function call, as in this example:

fn PeekTopOfStack[T:! type](s: Stack(T)*) -> T {
  var top: T = s->Pop();
  s->Push(top);
  return top;
}

// `int_stack` has type `Stack(i32)`, so `T` is deduced to be `i32`.
PeekTopOfStack(&int_stack);

References:

Generic choice types

Choice types may be parameterized similarly to classes:

choice Result(T:! type, Error:! type) {
  Success(value: T),
  Failure(error: Error)
}

Generic interfaces

Interfaces are always parameterized by a Self type, but in some cases they will have additional parameters.

interface AddWith(U:! type);

Interfaces without parameters may only be implemented once for a given type, but a type can have distinct implementations of AddWith(i32) and AddWith(BigInt).

Parameters to an interface determine which implementation is selected for a type, in contrast to associated types which are determined by the implementation of an interface for a type.

References:

Generic implementations

An impl declaration may be parameterized by adding forall [generic parameter list] after the impl keyword introducer, as in:

impl forall [T:! Printable] Vector(T) as Printable;
impl forall [Key:! Hashable, Value:! type]
    HashMap(Key, Value) as Has(Key);
impl forall [T:! Ordered] T as PartiallyOrdered;
impl forall [T:! ImplicitAs(i32)] BigInt as AddWith(T);
impl forall [U:! type, T:! As(U)]
    Optional(T) as As(Optional(U));

Generic implementations can create a situation where multiple impl definitions apply to a given type and interface query. The specialization rules pick which definition is selected. These rules ensure:

  • Implementations have coherence, so the same implementation is always selected for a given query.
  • Libraries will work together as long as they pass their separate checks.
  • A generic function can assume that some impl will be successfully selected if it can see an impl that applies, even though another more specific impl may be selected.

Implementations may be marked final to indicate that they may not be specialized, subject to some restrictions.

References:

Other features

Carbon generics have a number of other features, including:

  • Named constraints may be used to disambiguate when combining two interfaces that have name conflicts. Named constraints may be implemented and otherwise used in place of an interface.
  • Template constraints are a kind of named constraint that can contain structural requirements. For example, a template constraint could match any type that has a function with a specific name and signature without any explicit declaration that the type implements the constraint. Template constraints may only be used as requirements for template parameters.
  • An adapter type is a type with the same data representation as an existing type, so you may cast between the two types, but can implement different interfaces or implement interfaces differently.
  • Additional requirements can be placed on the associated types of an interface using where constraints.
  • Implied constraints allows some constraints to be deduced and omitted from a function signature.
  • Dynamic erased types can hold any value with a type implementing an interface, and allows the functions in that interface to be called using dynamic dispatch, for some interfaces marked "dyn-safe". Note: Provisional.
  • Variadics supports variable-length parameter lists. Note: Provisional.

References:

Generic type equality and observe declarations

Determining whether two types must be equal in a generic context is in general undecidable, as has been shown in Swift.

To make compilation fast, the Carbon compiler will limit its search to a depth of 1, only identifying types as equal if there is an explicit declaration that they are equal in the code, such as in a where constraint. There will be situations where two types must be equal as the result of combining these facts, but the compiler will return a type error since it did not realize they are equal due to the limit of the search. An observe...== declaration may be added to describe how two types are equal, allowing more code to pass type checking.

An observe declaration showing types are equal can increase the set of interfaces the compiler knows that a type implements. It is also possible that knowing a type implements one interface implies that it implements another, from an interface requirement or generic implementation. An observe...impls declaration may be used to observe that a type implements an interface.

References:

Operator overloading

Uses of an operator in an expression is translated into a call to a method of an interface. For example, if x has type T and y has type U, then x + y is translated into a call to x.(AddWith(U).Op)(y). So overloading of the + operator is accomplished by implementing interface AddWith(U) for type T. In order to support implicit conversion of the first operand to type T and the second argument to type U, add the like keyword to both types in the impl declaration, as in:

impl like T as AddWith(like U) where .Result = V {
  // `Self` is `T` here
  fn Op[self: Self](other: U) -> V { ... }
}

When the operand types and result type are all the same, this is equivalent to implementing the Add interface:

impl T as Add {
  fn Op[self: Self](other: Self) -> Self { ... }
}

The interfaces that correspond to each operator are given by:

  • Arithmetic:
    • -x: Negate
    • x + y: Add or AddWith(U)
    • x - y: Sub or SubWith(U)
    • x * y: Mul or MulWith(U)
    • x / y: Div or DivWith(U)
    • x % y: Mod or ModWith(U)
  • Bitwise and shift operators:
    • ^x: BitComplement
    • x & y: BitAnd or BitAndWith(U)
    • x | y: BitOr or BitOrWith(U)
    • x ^ y: BitXor or BitXorWith(U)
    • x << y: LeftShift or LeftShiftWith(U)
    • x >> y: RightShift or RightShiftWith(U)
  • Comparison:
  • Conversion:
  • Indexing:
  • TODO: Dereference: *p
  • TODO: Move: ~x
  • TODO: Function call: f(4)

The logical operators can not be overloaded.

Operators that result in reference expressions, such as dereferencing *p and indexing a[3], have interfaces that return the address of the value. Carbon automatically dereferences the pointer to form the reference expression.

Operators that can take multiple arguments, such as function calling operator f(4), have a variadic parameter list.

Whether and how a value supports other operations, such as being copied, swapped, or set into an unformed state, is also determined by implementing corresponding interfaces for the value's type.

References:

Common type

There are some situations where the common type for two types is needed:

  • A conditional expression like if c then t else f returns a value with the common type of t and f.

  • If there are multiple parameters to a function with a type parameter, it will be set to the common type of the corresponding arguments, as in:

    fn F[T:! type](x: T, y: T);

// Calls `F` with `T` set to the
// common type of `G()` and `H()`:
F(G(), H());

  • The inferred return type of a function with auto return type is the common type of its return statements.

The common type is specified by implementing the CommonTypeWith interface:

// Common type of `A` and `B` is `C`.
impl A as CommonTypeWith(B) where .Result = C { }

The common type is required to be a type that both types have an implicit conversion to.

References:

Bidirectional interoperability with C and C++

Interoperability, or interop, is the ability to call C and C++ code from Carbon code and the other way around. This ability achieves two goals:

  • Allows sharing a code and library ecosystem with C and C++.
  • Allows incremental migration to Carbon from C and C++.

Carbon's approach to interop is most similar to Java/Kotlin interop, where the two languages are different, but share enough of runtime model that data from one side can be used from the other. For example, C++ and Carbon will use the same memory model.

The design for interoperability between Carbon and C++ hinges on:

  1. The ability to interoperate with a wide variety of code, such as classes/structs and templates, not just free functions.
  2. A willingness to expose the idioms of C++ into Carbon code, and the other way around, when necessary to maximize performance of the interoperability layer.
  3. The use of wrappers and generic programming, including templates, to minimize or eliminate runtime overhead.

This feature will have some restrictions; only a subset of Carbon APIs will be available to C++ and a subset of C++ APIs will be available to Carbon.

  • To achieve simplification in Carbon, its programming model will exclude some rarely used and complex features of C++. For example, there will be limitations on multiple inheritance.
  • C or C++ features that compromise the performance of code that don't use that feature, like RTTI and exceptions, are in particular subject to revision in Carbon.

References:

Goals

The goals for interop include:

References:

Non-goals

The non-goals for interop include:

References:

Importing and #include

Note: This is provisional, no design for importing C++ has been through the proposal process yet.

A C++ library header file may be imported into Carbon using an import declaration of the special Cpp package.

// like `#include "circle.h"` in C++
import Cpp library "circle.h";

This adds the names from circle.h into the Cpp namespace. If circle.h defines some names in a namespace shapes { ... } scope, those will be found in Carbon's Cpp.shapes namespace.

In the other direction, Carbon packages can export a header file to be #included from C++ files.

// like `import Geometry` in Carbon
#include "geometry.carbon.h"

Generally Carbon entities will be usable from C++ and C++ entities will be usable from Carbon. This includes types, function, and constants. Some entities, such as Carbon interfaces, won't be able to be translated directly.

C and C++ macros that are defining constants will be imported as constants. Otherwise, C and C++ macros will be unavailable in Carbon. C and C++ typedefs would be translated into type constants, as if declared using a let.

Carbon functions and types that satisfy some restrictions may be annotated as exported to C as well, like C++'s extern "C" marker.

ABI and dynamic linking

Note: This reflects goals and plans. No specific design for the implementation has been through the proposal process yet.

Carbon itself will not have a stable ABI for the language as a whole, and most language features will be designed around not having any ABI stability. Instead, we expect to add dedicated language features that are specifically designed to provide an ABI-stable boundary between two separate parts of a Carbon program. These ABI-resilient language features and API boundaries will be opt-in and explicit. They may also have functionality restrictions to make them easy to implement with strong ABI resilience.

When interoperating with already compiled C++ object code or shared libraries, the C++ interop may be significantly less feature rich than otherwise. This is an open area for us to explore, but we expect to require re-compiling C++ code in order to get the full ergonomic and performance benefits when interoperating with Carbon. For example, recompilation lets us ensure Carbon and C++ can use the same representation for key vocabulary types.

However, we expect to have full support for the C ABI when interoperating with already-compiled C object code or shared libraries. We expect Carbon's bridge code functionality to cover similar use cases as C++'s extern "C" marker in order to provide full bi-directional support here. The functionality available across this interop boundary will of course be restricted to what is expressible in the C ABI, and types may need explicit markers to have guaranteed ABI compatibility.

References:

Operator overloading

Note: This is provisional, no design for this has been through the proposal process yet.

Operator overloading is supported in Carbon, but is done by implementing an interface instead of defining a method or nonmember function as in C++.

Carbon types implementing an operator overload using an interface should get the corresponding operator overload in C++. So implementing ModWith(U) in Carbon for a type effectively implements operator% in C++ for that type. This also works in the other direction, so C++ types implementing an operator overload are automatically considered to implement the corresponding Carbon interface. So implementing operator% in C++ for a type also implements interface ModWith(U) in Carbon. However, there may be edge cases around implicit conversions or overload selection that don't map completely into Carbon.

In some cases, the operation might be written differently in the two languages. In those cases, they are matched according to which operation has the most similar semantics rather than using the same symbols. For example, the ^x operation and BitComplement interface in Carbon corresponds to the ~x operation and operator~ function in C++. Similarly, the ImplicitAs(U) Carbon interface corresponds to implicit conversions in C++, which can be written in multiple different ways. Other C++ customization points like swap will correspond to a Carbon interface, on a case-by-case basis.

Some operators will only exist or be overridable in C++, such as logical operators or the comma operator. In the unlikely situation where those operators need to be overridden for a Carbon type, that can be done with a nonmember C++ function.

Carbon interfaces with no C++ equivalent, such as CommonTypeWith(U), may be implemented for C++ types externally in Carbon code. To satisfy the orphan rule (1, 2), each C++ library will have a corresponding Carbon wrapper library that must be imported instead of the C++ library if the Carbon wrapper exists. TODO: Perhaps it will automatically be imported, so a wrapper may be added without requiring changes to importers?

Templates

Note: This is provisional, no design for this has been through the proposal process yet.

Carbon supports both checked and template generics. This provides a migration path for C++ template code:

  • C++ template -> Carbon template: This involves migrating the code from C++ to Carbon. If that migration is faithful, the change should be transparent to callers.
  • -> Carbon template with constraints: Constraints may be added one at a time. Adding a constraint never changes the meaning of the code as long as it continues to compile. Compile errors will point to types for which an implementation of missing interfaces is needed. A temporary template implementation of that interface can act as a bridge during the transition.
  • -> Carbon checked generic: Once all callers work after all constraints have been added, the template parameter may be switched to a checked generic.

Carbon will also provide direct interop with C++ templates in many ways:

  • Ability to call C++ templates and use C++ templated types from Carbon.
  • Ability to instantiate a C++ template with a Carbon type.
  • Ability to instantiate a Carbon generic with a C++ type.

We expect the best interop in these areas to be based on a Carbon-provided C++ toolchain. However, even when using Carbon's generated C++ headers for interop, we will include the ability where possible to use a Carbon generic from C++ as if it were a C++ template.

Standard types

Note: This is provisional, no design for this has been through the proposal process yet.

The Carbon integer types, like i32 and u64, are considered equal to the corresponding fixed-width integer types in C++, like int32_t and uint64_t, provided by <stdint.h> or <cstdint>. The basic C and C++ integer types like int, char, and unsigned long are available in Carbon inside the Cpp namespace given an import Cpp; declaration, with names like Cpp.int, Cpp.char, and Cpp.unsigned_long. C++ types are considered different if C++ considers them different, so C++ overloads are resolved the same way. Carbon conventions for implicit conversions between integer types apply here, allowing them whenever the numerical value for all inputs may be preserved by the conversion.

Other C and C++ types are equal to Carbon types as follows:

C or C++ Carbon
bool bool
float f32
double f64
T* Optional(T*)
T[4] [T; 4]

Further, C++ reference types like T& will be translated to T* in Carbon, which is Carbon's non-null pointer type.

Carbon will work to have idiomatic vocabulary view types for common data structures, like std::string_view and std::span, map transparently between C++ and the Carbon equivalents. This will include data layout so that even pointers to these types translate seamlessly, contingent on a suitable C++ ABI for those types, potentially by re-compiling the C++ code with a customized ABI. We will also explore how to expand coverage to similar view types in other libraries.

However, Carbon's containers will be distinct from the C++ standard library containers in order to maximize our ability to improve performance and leverage language features like checked generics in their design and implementation.

Where possible, we will also try to provide implementations of Carbon's standard library container interfaces for the relevant C++ container types so that they can be directly used with generic Carbon code. This should allow generic code in Carbon to work seamlessly with both Carbon and C++ containers without performance loss or constraining the Carbon container implementations. In the other direction, Carbon containers will satisfy C++ container requirements, so templated C++ code can operate directly on Carbon containers as well.

Inheritance

Carbon has single inheritance allowing C++ classes using inheritance to be migrated. The data representation will be consistent so that Carbon classes may inherit from C++ classes, and the other way around, even with virtual methods.

C++ multiple inheritance and CRTP will be migrated using a combination of Carbon features. Carbon mixins support implementation reuse and Carbon interfaces allow a type to implement multiple APIs. However, there may be limits on the degree of interop available with multiple inheritance across the C++ <-> Carbon boundaries.

Carbon dyn-safe interfaces may be exported to C++ as an abstract base class. The reverse operation is also possible using a proxy object implementing a C++ abstract base class and holding a pointer to a type implementing the corresponding interface.

References:

Enums

TODO

Unfinished tales

Note: Everything in this section is provisional and forward looking.

Safety

Carbon's premise is that C++ users can't give up performance to get safety. Even if some isolated users can make that tradeoff, they share code with performance-sensitive users. Any path to safety must preserve performance of C++ today. This rules out garbage collection, and many other options. The only well understood mechanism of achieving safety without giving up performance is compile-time safety. The leading example of how to achieve this is Rust.

The difference between Rust's approach and Carbon's is that Rust starts with safety and Carbons starts with migration. Rust supports interop with C, and there is ongoing work to improve the C++-interop story and develop migration tools. However, there is a large gap in programming models between the two languages, generally requiring a revision to the architecture. So, thus far the common pattern in the Rust community is to "rewrite it in Rust" (1, 2, 3). Carbon's approach is to focus on migration from C++, including seamless interop, and then incrementally improve safety.

The first impact on Carbon's design to support its safety strategy are the necessary building blocks for this level of compile-time safety. We look at existing languages like Rust and Swift to understand what fundamental capabilities they ended up needing. The two components that stand out are:

  • Expanded type system that includes more semantic information.
  • More pervasive use of type system abstractions (typically generics).

For migrating C++ code, we also need the ability to add features and migrate code to use those new features incrementally and over time. This requires designing the language with evolution baked in on day one. This impacts a wide range of features:

  • At the lowest level, a simple and extensible syntax and grammar.
  • Tools and support for adding and removing APIs.
  • Scalable migration strategies, including tooling support.

Rust shows the value of expanded semantic information in the type system such as precise lifetimes. This is hard to do in C++ since it has too many kinds of references and pointers, which increases the complexity in the type system multiplicatively. Carbon is attempting to compress C++'s type variations into just values and pointers.

Rust also shows the value of functions parameterized by lifetimes. Since lifetimes are only used to establish safety properties of the code, there is no reason to pay the cost of monomorphization for those parameters. So we need a generics system that can reason about code before it is instantiated, unlike C++ templates.

In conclusion, there are two patterns in how Carbon diverges from C++:

  • Simplify and removing things to create space for new safety features. This trivially requires breaking backwards compatibility.
  • Re-engineer foundations to model and enforce safety. This has complex and difficulty in C++ without first simplifying the language.

This leads to Carbon's incremental path to safety:

  • Keep your performance, your existing codebase, and your developers.
  • Adopt Carbon through a scalable, tool-assisted migration from C++.
  • Address initial, easy safety improvements starting day one.
  • Shift the Carbon code onto an incremental path towards memory safety over the next decade.

References: Safety strategy

Lifetime and move semantics

TODO:

Metaprogramming

TODO: References need to be evolved. Needs a detailed design and a high level summary provided inline.

Carbon provides metaprogramming facilities that look similar to regular Carbon code. These are structured, and do not offer arbitrary inclusion or preprocessing of source text such as C and C++ do.

References: Metaprogramming

Pattern matching as function overload resolution

TODO: References need to be evolved. Needs a detailed design and a high level summary provided inline.

References: Pattern matching

Error handling

For now, Carbon does not have language features dedicated to error handling, but we would consider adding some in the future. At this point, errors are represented using choice types like Result and Optional.

This is similar to the story for Rust, which started using Result, then added ? operator for convenience, and is now considering (1, 2) adding more.

Execution abstractions

Carbon provides some higher-order abstractions of program execution, as well as the critical underpinnings of such abstractions.

Abstract machine and execution model

TODO:

Lambdas

TODO:

Co-routines

TODO:

Concurrency

TODO: