rust/doc/tutorial.md

75 KiB
Raw Blame History

% Rust Language Tutorial

Introduction

Rust is a programming language with a focus on type safety, memory safety, concurrency and performance. It is intended for writing large-scale, high-performance software while preventing several classes of common errors. Rust has a sophisticated memory model that encourages efficient data structures and safe concurrency patterns, forbidding invalid memory accesses that would otherwise cause segmentation faults. It is statically typed and compiled ahead of time.

As a multi-paradigm language, Rust supports writing code in procedural, functional and object-oriented styles. Some of its pleasant high-level features include:

  • Pattern matching and algebraic data types (enums). As popularized by functional languages, pattern matching on ADTs provides a compact and expressive way to encode program logic.
  • Type inference. Type annotations on local variable declarations are optional.
  • Task-based concurrency. Rust uses lightweight tasks that do not share memory.
  • Higher-order functions. Rust's efficient and flexible closures are heavily relied on to provide iteration and other control structures
  • Parametric polymorphism (generics). Functions and types can be parameterized over type variables with optional trait-based type constraints.
  • Trait polymorphism. Rust's type system features a unique combination of type classes and object-oriented interfaces.

Scope

This is an introductory tutorial for the Rust programming language. It covers the fundamentals of the language, including the syntax, the type system and memory model, and generics. Additional tutorials cover specific language features in greater depth.

It assumes the reader is familiar with the basic concepts of programming, and has programmed in one or more other languages before. It will often make comparisons to other languages, particularly those in the C family.

Conventions

Throughout the tutorial, words that indicate language keywords or identifiers defined in example code are displayed in code font.

Code snippets are indented, and also shown in a monospaced font. Not all snippets constitute whole programs. For brevity, we'll often show fragments of programs that don't compile on their own. To try them out, you might have to wrap them in fn main() { ... }, and make sure they don't contain references to things that aren't actually defined.

Warning: Rust is a language under heavy development. Notes about potential changes to the language, implementation deficiencies, and other caveats appear offset in blockquotes.

Getting started

The Rust compiler currently must be built from a tarball, unless you are on Windows, in which case using the installer is recommended.

Since the Rust compiler is written in Rust, it must be built by a precompiled "snapshot" version of itself (made in an earlier state of development). As such, source builds require a connection to the Internet, to fetch snapshots, and an OS that can execute the available snapshot binaries.

Snapshot binaries are currently built and tested on several platforms:

  • Windows (7, Server 2008 R2), x86 only
  • Linux (various distributions), x86 and x86-64
  • OSX 10.6 ("Snow Leopard") or 10.7 ("Lion"), x86 and x86-64

You may find that other platforms work, but these are our "tier 1" supported build environments that are most likely to work.

Note: Windows users should read the detailed getting started notes on the wiki. Even when using the binary installer the Windows build requires a MinGW installation, the precise details of which are not discussed in this tutorial.

To build from source you will also need the following prerequisite packages:

  • g++ 4.4 or clang++ 3.x
  • python 2.6 or later (but not 3.x)
  • perl 5.0 or later
  • gnu make 3.81 or later
  • curl

Assuming you're on a relatively modern *nix system and have met the prerequisites, something along these lines should work.

$ wget http://dl.rust-lang.org/dist/rust-0.4.tar.gz
$ tar -xzf rust-0.4.tar.gz
$ cd rust-0.4
$ ./configure
$ make && make install

You may need to use sudo make install if you do not normally have permission to modify the destination directory. The install locations can be adjusted by passing a --prefix argument to configure. Various other options are also supported, pass --help for more information on them.

When complete, make install will place several programs into /usr/local/bin: rustc, the Rust compiler; rustdoc, the API-documentation tool, and cargo, the Rust package manager.

Compiling your first program

Rust program files are, by convention, given the extension .rs. Say we have a file hello.rs containing this program:

fn main() {
    io::println("hello? yes, this is rust");
}

If the Rust compiler was installed successfully, running rustc hello.rs will produce an executable called hello (or hello.exe on Windows) which, upon running, will likely do exactly what you expect (unless you are on Windows, in which case what it does is subject to local weather conditions).

Note: That may or may not be hyperbole, but there are some 'gotchas' to be aware of on Windows. First, the MinGW environment must be set up perfectly. Please read the wiki. Second, rustc may need to be referred to as rustc.exe. It's a bummer, I know, and I am so very sorry.

The Rust compiler tries to provide useful information when it runs into an error. If you modify the program to make it invalid (for example, by changing io::println to some nonexistent function), and then compile it, you'll see an error message like this:

hello.rs:2:4: 2:16 error: unresolved name: io::print_with_unicorns
hello.rs:2     io::print_with_unicorns("hello? yes, this is rust");
               ^~~~~~~~~~~~~~~~~~~~~~~

In its simplest form, a Rust program is a .rs file with some types and functions defined in it. If it has a main function, it can be compiled to an executable. Rust does not allow code that's not a declaration to appear at the top level of the file—all statements must live inside a function. Rust programs can also be compiled as libraries, and included in other programs. The extern mod std directive that appears at the top of many examples imports the standard library, described in more detail later on.

Editing Rust code

There are vim highlighting and indentation scripts in the Rust source distribution under src/etc/vim/. There is an emacs mode under src/etc/emacs/ called rust-mode, but do read the instructions included in that directory. In particular, if you are running emacs 24, then using emacs's internal package manager to install rust-mode is the easiest way to keep it up to date. There is also a package for Sublime Text 2, available both standalone and through Sublime Package Control, and support for Kate under src/etc/kate.

There is ctags support via src/etc/ctags.rust, but many other tools and editors are not provided for yet. If you end up writing a Rust mode for your favorite editor, let us know so that we can link to it.

Syntax Basics

Assuming you've programmed in any C-family language (C++, Java, JavaScript, C#, or PHP), Rust will feel familiar. Code is arranged in blocks delineated by curly braces; there are control structures for branching and looping, like the familiar if and while; function calls are written myfunc(arg1, arg2); operators are written the same and mostly have the same precedence as in C; comments are again like C.

The main surface difference to be aware of is that the condition at the head of control structures like if and while do not require paretheses, while their bodies must be wrapped in brackets. Single-statement, bracket-less bodies are not allowed.

# fn recalibrate_universe() -> bool { true }
fn main() {
    /* A simple loop */
    loop {
        // A tricky calculation
        if recalibrate_universe() {
            return;
        }
    }
}

The let keyword introduces a local variable. Variables are immutable by default, so let mut can be used to introduce a local variable that can be reassigned.

let hi = "hi";
let mut count = 0;

while count < 10 {
    io::println(hi);
    count += 1;
}

Although Rust can almost always infer the types of local variables, you can specify a variable's type by following it with a colon, then the type name.

let monster_size: float = 57.8;
let imaginary_size = monster_size * 10.0;
let monster_size: int = 50;

Local variables may shadow earlier declarations, as in the previous example in which monster_size is first declared as a float then a second monster_size is declared as an int. If you were to actually compile this example though, the compiler will see that the second monster_size is unused, assume that you have made a mistake, and issue a warning. For occasions where unused variables are intentional, their name may be prefixed with an underscore to silence the warning, like let _monster_size = 50;.

Rust identifiers follow the same rules as C; they start with an alphabetic character or an underscore, and after that may contain any sequence of alphabetic characters, numbers, or underscores. The preferred style is to begin function, variable, and module names with a lowercase letter, using underscores where they help readability, while writing types in camel case.

let my_variable = 100;
type MyType = int;     // some built-in types are _not_ camel case

Expression syntax

Though it isn't apparent in all code, there is a fundamental difference between Rust's syntax and predecessors like C. Many constructs that are statements in C are expressions in Rust, allowing code to be more concise. For example, you might write a piece of code like this:

# let item = "salad";
let price;
if item == "salad" {
    price = 3.50;
} else if item == "muffin" {
    price = 2.25;
} else {
    price = 2.00;
}

But, in Rust, you don't have to repeat the name price:

# let item = "salad";
let price =
    if item == "salad" {
        3.50
    } else if item == "muffin" {
        2.25
    } else {
        2.00
    };

Both pieces of code are exactly equivalent—they assign a value to price depending on the condition that holds. Note that there are not semicolons in the blocks of the second snippet. This is important; the lack of a semicolon after the last statement in a braced block gives the whole block the value of that last expression.

Put another way, the semicolon in Rust ignores the value of an expression. Thus, if the branches of the if had looked like { 4; }, the above example would simply assign () (nil or void) to price. But without the semicolon, each branch has a different value, and price gets the value of the branch that was taken.

In short, everything that's not a declaration (let for variables, fn for functions, et cetera) is an expression, including function bodies.

fn is_four(x: int) -> bool {
   // No need for a return statement. The result of the expression
   // is used as the return value.
   x == 4
}

If all those things are expressions, you might conclude that you have to add a terminating semicolon after every statement, even ones that are not traditionally terminated with a semicolon in C (like while). That is not the case, though. Expressions that end in a block only need a semicolon if that block contains a trailing expression. while loops do not allow trailing expressions, and if statements tend to only have a trailing expression when you want to use their value for something—in which case you'll have embedded it in a bigger statement.

# fn foo() -> bool { true }
# fn bar() -> bool { true }
# fn baz() -> bool { true }
// `let` is not an expression, so it is semi-colon terminated;
let x = foo();

// When used in statement position, bracy expressions do not
// usually need to be semicolon terminated
if x {
    bar();
} else {
    baz();
} // No semi-colon

// Although, if `bar` and `baz` have non-nil return types, and
// we try to use them as the tail expressions, rustc will
// make us terminate the expression.
if x {
    bar()
} else {
    baz()
}; // Semi-colon to ignore non-nil block type

// An `if` embedded in `let` again requires a semicolon to terminate
// the `let` statement
let y = if x { foo() } else { bar() };

This may sound intricate, but it is super-useful and will grow on you.

Types

The basic types include the usual boolean, integral, and floating-point types.


() Nil, the type that has only a single value bool Boolean type, with values true and false int, uint Machine-pointer-sized signed and unsigned integers i8, i16, i32, i64 Signed integers with a specific size (in bits) u8, u16, u32, u64 Unsigned integers with a specific size float The largest floating-point type efficiently supported on the target machine f32, f64 Floating-point types with a specific size char A Unicode character (32 bits)


These can be combined in composite types, which will be described in more detail later on (the Ts here stand for any other type, while N should be a literal number):


[T * N] Vector (like an array in other languages) with N elements [mut T * N] Mutable vector with N elements (T1, T2) Tuple type; any arity above 1 is supported &T, ~T, @T Pointer types


Some types can only be manipulated by pointer, never directly. For instance, you cannot refer to a string (str); instead you refer to a pointer to a string (@str, ~str, or &str). These dynamically-sized types consist of:


fn(a: T1, b: T2) -> T3 Function types str String type (in UTF-8) [T] Vector with unknown size (also called a slice) [mut T] Mutable vector with unknown size


In function types, the return type is specified with an arrow, as in the type fn() -> bool or the function declaration fn foo() -> bool { }. For functions that do not return a meaningful value, you can optionally write -> (), but usually the return annotation is simply left off, as in fn main() { ... }.

Types can be given names or aliases with type declarations:

type MonsterSize = uint;

This will provide a synonym, MonsterSize, for unsigned integers. It will not actually create a new, incompatible type—MonsterSize and uint can be used interchangeably, and using one where the other is expected is not a type error.

To create data types which are not synonyms, struct and enum can be used. They're described in more detail below, but they look like this:

enum HidingPlaces {
   Closet(uint),
   UnderTheBed(uint)
}

struct HeroicBabysitter {
   bedtime_stories: uint,
   sharpened_stakes: uint
}

struct BabysitterSize(uint);  // a single-variant struct
enum MonsterSize = uint;      // a single-variant enum

Literals

Integers can be written in decimal (144), hexadecimal (0x90), and binary (0b10010000) base. Each integral type has a corresponding literal suffix that can be used to indicate the type of a literal: i for int, u for uint, and i8 for the i8 type, etc.

In the absense of an integer literal suffix, Rust will infer the integer type based on type annotations and function signatures in the surrounding program. In the absence of any type information at all, Rust will assume that an unsuffixed integer literal has type int.

let a = 1;       // a is an int
let b = 10i;     // b is an int, due to the 'i' suffix
let c = 100u;    // c is a uint
let d = 1000i32; // d is an i32

Floating point numbers are written 0.0, 1e6, or 2.1e-4. Without a suffix, the literal is assumed to be of type float. Suffixes f32 (32-bit) and f64 (64-bit) can be used to create literals of a specific type.

The nil literal is written just like the type: (). The keywords true and false produce the boolean literals.

Character literals are written between single quotes, as in 'x'. Just as in C, Rust understands a number of character escapes, using the backslash character, such as \n, \r, and \t. String literals, written between double quotes, allow the same escape sequences. Rust strings may contain newlines.

Operators

Rust's set of operators contains very few surprises. Arithmetic is done with *, /, %, +, and - (multiply, divide, remainder, plus, minus). - is also a unary prefix operator that does negation. As in C, the bit operators >>, <<, &, |, and ^ are also supported.

Note that, if applied to an integer value, ! flips all the bits (like ~ in C).

The comparison operators are the traditional ==, !=, <, >, <=, and >=. Short-circuiting (lazy) boolean operators are written && (and) and || (or).

For type casting, Rust uses the binary as operator. It takes an expression on the left side and a type on the right side and will, if a meaningful conversion exists, convert the result of the expression to the given type.

let x: float = 4.0;
let y: uint = x as uint;
assert y == 4u;

The main difference with C is that ++ and -- are missing, and that the logical bitwise operators have higher precedence — in C, x & 2 > 0 means x & (2 > 0), but in Rust, it means (x & 2) > 0, which is more likely what a novice expects.

Syntax extensions

Syntax extensions are special forms that are not built into the language, but are instead provided by the libraries. To make it clear to the reader when a syntax extension is being used, the names of all syntax extensions end with !. The standard library defines a few syntax extensions, the most useful of which is fmt!, a sprintf-style text formatter that is expanded at compile time.

fmt! supports most of the directives that printf supports, but will give you a compile-time error when the types of the directives don't match the types of the arguments.

# let mystery_object = ();

io::println(fmt!("%s is %d", "the answer", 43));

// %? will conveniently print any type
io::println(fmt!("what is this thing: %?", mystery_object));

You can define your own syntax extensions with the macro system, which is out of scope of this tutorial.

Control structures

Conditionals

We've seen if pass by a few times already. To recap, braces are compulsory, an optional else clause can be appended, and multiple if/else constructs can be chained together:

if false {
    io::println("that's odd");
} else if true {
    io::println("right");
} else {
    io::println("neither true nor false");
}

The condition given to an if construct must be of type boolean (no implicit conversion happens). If the arms return a value, this value must be of the same type for every arm in which control reaches the end of the block:

fn signum(x: int) -> int {
    if x < 0 { -1 }
    else if x > 0 { 1 }
    else { return 0 }
}

Pattern matching

Rust's match construct is a generalized, cleaned-up version of C's switch construct. You provide it with a value and a number of arms, each labelled with a pattern, and the code will attempt to match each pattern in order. For the first one that matches, the arm is executed.

# let my_number = 1;
match my_number {
  0     => io::println("zero"),
  1 | 2 => io::println("one or two"),
  3..10 => io::println("three to ten"),
  _     => io::println("something else")
}

There is no 'falling through' between arms, as in C—only one arm is executed, and it doesn't have to explicitly break out of the construct when it is finished.

The part to the left of the arrow => is called the pattern. Literals are valid patterns and will match only their own value. The pipe operator (|) can be used to assign multiple patterns to a single arm. Ranges of numeric literal patterns can be expressed with two dots, as in M..N. The underscore (_) is a wildcard pattern that matches everything.

The patterns in an match arm are followed by a fat arrow, =>, then an expression to evaluate. Each case is separated by commas. It's often convenient to use a block expression for each case, in which case the commas are optional.

# let my_number = 1;
match my_number {
  0 => { io::println("zero") }
  _ => { io::println("something else") }
}

match constructs must be exhaustive: they must have an arm covering every possible case. For example, if the arm with the wildcard pattern was left off in the above example, the typechecker would reject it.

A powerful application of pattern matching is destructuring, where you use the matching to get at the contents of data types. Remember that (float, float) is a tuple of two floats:

fn angle(vector: (float, float)) -> float {
    let pi = float::consts::pi;
    match vector {
      (0f, y) if y < 0f => 1.5 * pi,
      (0f, y) => 0.5 * pi,
      (x, y) => float::atan(y / x)
    }
}

A variable name in a pattern matches everything, and binds that name to the value of the matched thing inside of the arm block. Thus, (0f, y) matches any tuple whose first element is zero, and binds y to the second element. (x, y) matches any tuple, and binds both elements to a variable.

Any match arm can have a guard clause (written if EXPR), which is an expression of type bool that determines, after the pattern is found to match, whether the arm is taken or not. The variables bound by the pattern are available in this guard expression.

You've already seen simple let bindings, but let is a little fancier than you've been led to believe. It too supports destructuring patterns. For example, you can say this to extract the fields from a tuple, introducing two variables, a and b.

# fn get_tuple_of_two_ints() -> (int, int) { (1, 1) }
let (a, b) = get_tuple_of_two_ints();

Let bindings only work with irrefutable patterns, that is, patterns that can never fail to match. This excludes let from matching literals and most enum variants.

Loops

while produces a loop that runs as long as its given condition (which must have type bool) evaluates to true. Inside a loop, the keyword break can be used to abort the loop, and loop can be used to abort the current iteration and continue with the next.

let mut cake_amount = 8;
while cake_amount > 0 {
    cake_amount -= 1;
}

loop is the preferred way of writing while true:

let mut x = 5;
loop {
    x += x - 3;
    if x % 5 == 0 { break; }
    io::println(int::str(x));
}

This code prints out a weird sequence of numbers and stops as soon as it finds one that can be divided by five.

For more involved iteration, such as going over the elements of a collection, Rust uses higher-order functions. We'll come back to those in a moment.

Basic datatypes

The core datatypes of Rust are structs, enums (tagged unions, algebraic data types), and tuples. They are immutable by default.

struct Point { x: float, y: float }

enum Shape {
    Circle(Point, float),
    Rectangle(Point, Point)
}

Structs

Rust struct types must be declared before they are used using the struct syntax: struct Name { field1: T1, field2: T2 [, ...] }, where T1, T2, ... denote types. To construct a struct, use the same syntax, but leave off the struct; for example: Point { x: 1.0, y: 2.0 }.

Structs are quite similar to C structs and are even laid out the same way in memory (so you can read from a Rust struct in C, and vice-versa). The dot operator is used to access struct fields (mypoint.x).

Fields that you want to mutate must be explicitly marked mut.

struct Stack {
    content: ~[int],
    mut head: uint
}

With a value of such a type, you can do mystack.head += 1. If mut were omitted from the type, such an assignment would result in a type error.

Structs can be destructured in match patterns. The basic syntax is Name {fieldname: pattern, ...}:

# struct Point { x: float, y: float }
# let mypoint = Point { x: 0.0, y: 0.0 };
match mypoint {
    Point { x: 0.0, y: yy } => { io::println(yy.to_str());                     }
    Point { x: xx,  y: yy } => { io::println(xx.to_str() + " " + yy.to_str()); }
}

In general, the field names of a struct do not have to appear in the same order they appear in the type. When you are not interested in all the fields of a struct, a struct pattern may end with , _ (as in Name {field1, _}) to indicate that you're ignoring all other fields.

Enums

Enums are datatypes that have several alternate representations. For example, consider the type shown earlier:

# struct Point { x: float, y: float }
enum Shape {
    Circle(Point, float),
    Rectangle(Point, Point)
}

A value of this type is either a Circle, in which case it contains a Point struct and a float, or a Rectangle, in which case it contains two Point structs. The run-time representation of such a value includes an identifier of the actual form that it holds, much like the 'tagged union' pattern in C, but with better ergonomics.

The above declaration will define a type Shape that can be used to refer to such shapes, and two functions, Circle and Rectangle, which can be used to construct values of the type (taking arguments of the specified types). So Circle(Point {x: 0f, y: 0f}, 10f) is the way to create a new circle.

Enum variants need not have type parameters. This, for example, is equivalent to a C enum:

enum Direction {
    North,
    East,
    South,
    West
}

This will define North, East, South, and West as constants, all of which have type Direction.

When an enum is C-like, that is, when none of the variants have parameters, it is possible to explicitly set the discriminator values to an integer value:

enum Color {
  Red = 0xff0000,
  Green = 0x00ff00,
  Blue = 0x0000ff
}

If an explicit discriminator is not specified for a variant, the value defaults to the value of the previous variant plus one. If the first variant does not have a discriminator, it defaults to 0. For example, the value of North is 0, East is 1, etc.

When an enum is C-like the as cast operator can be used to get the discriminator's value.

There is a special case for enums with a single variant. These are used to define new types in such a way that the new name is not just a synonym for an existing type, but its own distinct type. If you say:

enum GizmoId = int;

That is a shorthand for this:

enum GizmoId { GizmoId(int) }

Enum types like this can have their content extracted with the dereference (*) unary operator:

# enum GizmoId = int;
let my_gizmo_id: GizmoId = GizmoId(10);
let id_int: int = *my_gizmo_id;

For enum types with multiple variants, destructuring is the only way to get at their contents. All variant constructors can be used as patterns, as in this definition of area:

# type Point = {x: float, y: float};
# enum Shape { Circle(Point, float), Rectangle(Point, Point) }
fn area(sh: Shape) -> float {
    match sh {
        Circle(_, size) => float::consts::pi * size * size,
        Rectangle({x, y}, {x: x2, y: y2}) => (x2 - x) * (y2 - y)
    }
}

Like other patterns, a lone underscore ignores individual fields. Ignoring all fields of a variant can be written Circle(*). As in their introductory form, nullary enum patterns are written without parentheses.

# type Point = {x: float, y: float};
# enum Direction { North, East, South, West }
fn point_from_direction(dir: Direction) -> Point {
    match dir {
        North => {x:  0f, y:  1f},
        East  => {x:  1f, y:  0f},
        South => {x:  0f, y: -1f},
        West  => {x: -1f, y:  0f}
    }
}

Tuples

Tuples in Rust behave exactly like structs, except that their fields do not have names (and can thus not be accessed with dot notation). Tuples can have any arity except for 0 or 1 (though you may consider nil, (), as the empty tuple if you like).

let mytup: (int, int, float) = (10, 20, 30.0);
match mytup {
  (a, b, c) => log(info, a + b + (c as int))
}

Functions and methods

We've already seen several function definitions. Like all other static declarations, such as type, functions can be declared both at the top level and inside other functions (or modules, which we'll come back to later). They are introduced with the fn keyword, the type of arguments are specified following colons and the return type follows the arrow.

fn line(a: int, b: int, x: int) -> int {
    return a * x + b;
}

The return keyword immediately returns from the body of a function. It is optionally followed by an expression to return. A function can also return a value by having its top level block produce an expression.

fn line(a: int, b: int, x: int) -> int {
    a * x + b
}

Functions that do not return a value are said to return nil, (), and both the return type and the return value may be omitted from the definition. The following two functions are equivalent.

fn do_nothing_the_hard_way() -> () { return (); }

fn do_nothing_the_easy_way() { }

Ending the function with a semicolon like so is equivalent to returning ().

fn line(a: int, b: int, x: int) -> int { a * x + b  }
fn oops(a: int, b: int, x: int) -> ()  { a * x + b; }

assert 8  == line(5, 3, 1);
assert () == oops(5, 3, 1);

Methods are like functions, except that they are defined for a specific 'self' type (like 'this' in C++). Calling a method is done with dot notation, as in my_vec.len(). Methods may be defined on most Rust types with the impl keyword. As an example, lets define a draw method on our Shape enum.

# fn draw_circle(p: Point, f: float) { }
# fn draw_rectangle(p: Point, p: Point) { }
struct Point {
    x: float,
    y: float
}

enum Shape {
    Circle(Point, float),
    Rectangle(Point, Point)
}

impl Shape {
    fn draw() {
        match self {
            Circle(p, f) => draw_circle(p, f),
            Rectangle(p1, p2) => draw_rectangle(p1, p2)
        }
    }
}

let s = Circle(Point { x: 1f, y: 2f }, 3f);
s.draw();

This defines an implementation for Shape containing a single method, draw. In most most respects the draw method is defined like any other function, with the exception of the name self. self is a special value that is automatically defined in each method, referring to the value being operated on. If we wanted we could add additional methods to the same impl, or multiple impls for the same type. We'll discuss methods more in the context of traits and generics.

Note: The method definition syntax will change to require declaring the self type explicitly, as the first argument.

The Rust memory model

At this junction let's take a detour to explain the concepts involved in Rust's memory model. We've seen some of Rust's pointer sigils (@, ~, and &) float by in a few examples, and we aren't going to get much further without explaining them. Rust has a very particular approach to memory management that plays a significant role in shaping the "feel" of the language. Understanding the memory landscape will illuminate several of Rust's unique features as we encounter them.

Rust has three competing goals that inform its view of memory:

  • Memory safety: memory that is managed by and is accessible to the Rust language must be guaranteed to be valid; under normal circumstances it must be impossible for Rust to trigger a segmentation fault or leak memory
  • Performance: high-performance low-level code must be able to employ a number of allocation strategies; low-performance high-level code must be able to employ a single, garbage-collection-based, heap allocation strategy
  • Concurrency: Rust must maintain memory safety guarantees, even for code running in parallel

How performance considerations influence the memory model

Most languages that offer strong memory safety guarantees rely upon a garbage-collected heap to manage all of the objects. This approach is straightforward both in concept and in implementation, but has significant costs. Languages that follow this path tend to aggressively pursue ways to ameliorate allocation costs (think the Java Virtual Machine). Rust supports this strategy with managed boxes: memory allocated on the heap whose lifetime is managed by the garbage collector.

By comparison, languages like C++ offer very precise control over where objects are allocated. In particular, it is common to put them directly on the stack, avoiding expensive heap allocation. In Rust this is possible as well, and the compiler will use a clever pointer lifetime analysis to ensure that no variable can refer to stack objects after they are destroyed.

How concurrency considerations influence the memory model

Memory safety in a concurrent environment involves avoiding race conditions between two threads of execution accessing the same memory. Even high-level languages often require programmers to correctly employ locking to ensure that a program is free of races.

Rust starts from the position that memory cannot be shared between tasks. Experience in other languages has proven that isolating each task's heap from the others is a reliable strategy and one that is easy for programmers to reason about. Heap isolation has the additional benefit that garbage collection must only be done per-heap. Rust never "stops the world" to reclaim memory.

Complete isolation of heaps between tasks would, however, mean that any data transferred between tasks must be copied. While this is a fine and useful way to implement communication between tasks, it is also very inefficient for large data structures. Because of this, Rust also employs a global exchange heap. Objects allocated in the exchange heap have ownership semantics, meaning that there is only a single variable that refers to them. For this reason, they are referred to as owned boxes. All tasks may allocate objects on the exchange heap, then transfer ownership of those objects to other tasks, avoiding expensive copies.

Boxes and pointers

In contrast to a lot of modern languages, aggregate types like structs and enums are not represented as pointers to allocated memory in Rust. They are, as in C and C++, represented directly. This means that if you let x = Point {x: 1f, y: 1f};, you are creating a struct on the stack. If you then copy it into a data structure, the whole struct is copied, not just a pointer.

For small structs like Point, this is usually more efficient than allocating memory and going through a pointer. But for big structs, or those with mutable fields, it can be useful to have a single copy on the heap, and refer to that through a pointer.

Rust supports several types of pointers. The safe pointer types are @T for managed boxes allocated on the local heap, ~T, for uniquely-owned boxes allocated on the exchange heap, and &T, for borrowed pointers, which may point to any memory, and whose lifetimes are governed by the call stack.

All pointer types can be dereferenced with the * unary operator.

Note

: You may also hear managed boxes referred to as 'shared boxes' or 'shared pointers', and owned boxes as 'unique boxes/pointers'. Borrowed pointers are sometimes called 'region pointers'. The preferred terminology is as presented here.

Managed boxes

Managed boxes are pointers to heap-allocated, garbage collected memory. Creating a managed box is done by simply applying the unary @ operator to an expression. The result of the expression will be boxed, resulting in a box of the right type. Copying a shared box, as happens during assignment, only copies a pointer, never the contents of the box.

let x: @int = @10; // New box
let y = x; // Copy of a pointer to the same box

// x and y both refer to the same allocation. When both go out of scope
// then the allocation will be freed.

Any type that contains managed boxes or other managed types is considered managed. Managed types are the only types that can construct cyclic data structures in Rust, such as doubly-linked lists.

// A linked list node
struct Node {
    mut next: MaybeNode,
    mut prev: MaybeNode,
    payload: int
}

enum MaybeNode {
    SomeNode(@Node),
    NoNode
}

let node1 = @Node { next: NoNode, prev: NoNode, payload: 1 };
let node2 = @Node { next: NoNode, prev: NoNode, payload: 2 };
let node3 = @Node { next: NoNode, prev: NoNode, payload: 3 };

// Link the three list nodes together
node1.next = SomeNode(node2);
node2.prev = SomeNode(node1);
node2.next = SomeNode(node3);
node3.prev = SomeNode(node2);

Managed boxes never cross task boundaries.

Note: managed boxes are currently reclaimed through reference counting and cycle collection, but we will switch to a tracing garbage collector eventually.

Owned boxes

In contrast to managed boxes, owned boxes have a single owning memory slot and thus two owned boxes may not refer to the same memory. All owned boxes across all tasks are allocated on a single exchange heap, where their uniquely owned nature allows them to be passed between tasks efficiently.

Because owned boxes are uniquely owned, copying them involves allocating a new owned box and duplicating the contents. Copying owned boxes is expensive so the compiler will complain if you do so without writing the word copy.

let x = ~10;
let y = x; // error: copying a non-implicitly copyable type

If you really want to copy a unique box you must say so explicitly.

let x = ~10;
let y = copy x;

let z = *x + *y;
assert z == 20;

This is where the 'move' operator comes in. It is similar to copy, but it de-initializes its source. Thus, the owned box can move from x to y, without violating the constraint that it only has a single owner (if you used assignment instead of the move operator, the box would, in principle, be copied).

let x = ~10;
let y = move x;

let z = *x + *y; // would cause an error: use of moved variable: `x`

Owned boxes, when they do not contain any managed boxes, can be sent to other tasks. The sending task will give up ownership of the box, and won't be able to access it afterwards. The receiving task will become the sole owner of the box.

Note: this discussion of copying vs moving does not account for the "last use" rules that automatically promote copy operations to moves. Last use is expected to be removed from the language in favor of explicit moves.

Borrowed pointers

Rust borrowed pointers are a general purpose reference/pointer type, similar to the C++ reference type, but guaranteed to point to valid memory. In contrast to owned pointers, where the holder of a unique pointer is the owner of the pointed-to memory, borrowed pointers never imply ownership. Pointers may be borrowed from any type, in which case the pointer is guaranteed not to outlive the value it points to.

As an example, consider a simple struct type, Point:

struct Point {
    x: float, y: float
}

We can use this simple definition to allocate points in many ways. For example, in this code, each of these three local variables contains a point, but allocated in a different place:

# struct Point { x: float, y: float }
let on_the_stack : Point  =  Point {x: 3.0, y: 4.0};
let shared_box   : @Point = @Point {x: 5.0, y: 1.0};
let unique_box   : ~Point = ~Point {x: 7.0, y: 9.0};

Suppose we wanted to write a procedure that computed the distance between any two points, no matter where they were stored. For example, we might like to compute the distance between on_the_stack and shared_box, or between shared_box and unique_box. One option is to define a function that takes two arguments of type point—that is, it takes the points by value. But this will cause the points to be copied when we call the function. For points, this is probably not so bad, but often copies are expensive or, worse, if there are mutable fields, they can change the semantics of your program. So wed like to define a function that takes the points by pointer. We can use borrowed pointers to do this:

# struct Point { x: float, y: float }
# fn sqrt(f: float) -> float { 0f }
fn compute_distance(p1: &Point, p2: &Point) -> float {
    let x_d = p1.x - p2.x;
    let y_d = p1.y - p2.y;
    sqrt(x_d * x_d + y_d * y_d)
}

Now we can call compute_distance() in various ways:

# struct Point{ x: float, y: float };
# let on_the_stack : Point  =  Point {x: 3.0, y: 4.0};
# let shared_box   : @Point = @Point {x: 5.0, y: 1.0};
# let unique_box   : ~Point = ~Point {x: 7.0, y: 9.0};
# fn compute_distance(p1: &Point, p2: &Point) -> float { 0f }
compute_distance(&on_the_stack, shared_box);
compute_distance(shared_box, unique_box);

Here the & operator is used to take the address of the variable on_the_stack; this is because on_the_stack has the type Point (that is, a struct value) and we have to take its address to get a value. We also call this borrowing the local variable on_the_stack, because we are created an alias: that is, another route to the same data.

In the case of the boxes shared_box and unique_box, however, no explicit action is necessary. The compiler will automatically convert a box like @point or ~point to a borrowed pointer like &point. This is another form of borrowing; in this case, the contents of the shared/unique box is being lent out.

Whenever a value is borrowed, there are some limitations on what you can do with the original. For example, if the contents of a variable have been lent out, you cannot send that variable to another task, nor will you be permitted to take actions that might cause the borrowed value to be freed or to change its type. This rule should make intuitive sense: you must wait for a borrowed value to be returned (that is, for the borrowed pointer to go out of scope) before you can make full use of it again.

For a more in-depth explanation of borrowed pointers, read the borrowed pointer tutorial.

Vectors and strings

Vectors are a contiguous section of memory containing zero or more values of the same type. Like other types in Rust, vectors can be stored on the stack, the local heap, or the exchange heap. Borrowed pointers to vectors are also called 'slices'.

enum Crayon {
    Almond, AntiqueBrass, Apricot,
    Aquamarine, Asparagus, AtomicTangerine,
    BananaMania, Beaver, Bittersweet,
    Black, BlizzardBlue, Blue
}

// A fixed-size stack vector
let stack_crayons: [Crayon * 3] = [Almond, AntiqueBrass, Apricot];

// A borrowed pointer to stack allocated vector
let stack_crayons: &[Crayon] = &[Aquamarine, Asparagus, AtomicTangerine];

// A local heap (managed) vector of crayons
let local_crayons: @[Crayon] = @[BananaMania, Beaver, Bittersweet];

// An exchange heap (owned) vector of crayons
let exchange_crayons: ~[Crayon] = ~[Black, BlizzardBlue, Blue];

The + operator means concatenation when applied to vector types.

# enum Crayon { Almond, AntiqueBrass, Apricot,
#               Aquamarine, Asparagus, AtomicTangerine,
#               BananaMania, Beaver, Bittersweet };

let my_crayons = ~[Almond, AntiqueBrass, Apricot];
let your_crayons = ~[BananaMania, Beaver, Bittersweet];

// Add two vectors to create a new one
let our_crayons = my_crayons + your_crayons;

// += will append to a vector, provided it leves
// in a mutable slot
let mut my_crayons = move my_crayons;
my_crayons += your_crayons;

Note: The above examples of vector addition use owned vectors. Some operations on slices and stack vectors are not well supported yet, owned vectors are often the most usable.

Indexing into vectors is done with square brackets:

# enum Crayon { Almond, AntiqueBrass, Apricot,
#               Aquamarine, Asparagus, AtomicTangerine,
#               BananaMania, Beaver, Bittersweet };
# fn draw_scene(c: Crayon) { }
let crayons: [Crayon * 3] = [BananaMania, Beaver, Bittersweet];
match crayons[0] {
    Bittersweet => draw_scene(crayons[0]),
    _ => ()
}

The elements of a vector inherit the mutability of the vector, and as such individual elements may not be reassigned when the vector lives in an immutable slot.

# enum Crayon { Almond, AntiqueBrass, Apricot,
#               Aquamarine, Asparagus, AtomicTangerine,
#               BananaMania, Beaver, Bittersweet };
let crayons: ~[Crayon] = ~[BananaMania, Beaver, Bittersweet];

crayons[0] = Apricot; // ERROR: Can't assign to immutable vector

Moving it into a mutable slot makes the elements assignable.

# enum Crayon { Almond, AntiqueBrass, Apricot,
#               Aquamarine, Asparagus, AtomicTangerine,
#               BananaMania, Beaver, Bittersweet };
let crayons: ~[Crayon] = ~[BananaMania, Beaver, Bittersweet];

// Put the vector into a mutable slot
let mut mutable_crayons = move crayons;

// Now it's mutable to the bone
mutable_crayons[0] = Apricot;

This is a simple example of Rust's dual-mode data structures, also referred to as freezing and thawing.

Strings are implemented with vectors of [u8], though they have a distinct type. They support most of the same allocation options as vectors, though the string literal without a storage sigil, e.g. "foo" is treated differently than a comparable vector ([foo]). Whereas plain vectors are stack-allocated fixed-length vectors, plain strings are region pointers to read-only memory. Strings are always immutable.

// A plain string is a slice to read-only (static) memory
let stack_crayons: &str = "Almond, AntiqueBrass, Apricot";

// The same thing, but with the `&`
let stack_crayons: &str = &"Aquamarine, Asparagus, AtomicTangerine";

// A local heap (managed) string
let local_crayons: @str = @"BananMania, Beaver, Bittersweet";

// An exchange heap (owned) string
let exchange_crayons: ~str = ~"Black, BlizzardBlue, Blue";

Both vectors and strings support a number of useful methods, defined in core::vec and core::str. Here are some examples.

# use io::println;
# enum Crayon {
#     Almond, AntiqueBrass, Apricot,
#     Aquamarine, Asparagus, AtomicTangerine,
#     BananaMania, Beaver, Bittersweet
# }
# fn unwrap_crayon(c: Crayon) -> int { 0 }
# fn eat_crayon_wax(i: int) { }
# fn store_crayon_in_nasal_cavity(i: uint, c: Crayon) { }
# fn crayon_to_str(c: Crayon) -> ~str { ~"" }

let crayons = &[Almond, AntiqueBrass, Apricot];

// Check the length of the vector
assert crayons.len() == 3;
assert !crayons.is_empty();

// Iterate over a vector, obtaining a pointer to each element
for crayons.each |crayon| {
    let delicious_crayon_wax = unwrap_crayon(*crayon);
    eat_crayon_wax(delicious_crayon_wax);
}

// Map vector elements
let crayon_names = crayons.map(|v| crayon_to_str(*v));
let favorite_crayon_name = crayon_names[0];

// Remove whitespace from before and after the string
let new_favorite_crayon_name = favorite_crayon_name.trim();

if favorite_crayon_name.len() > 5 {
   // Create a substring
   println(favorite_crayon_name.substr(0, 5));
}

Closures

Named functions, like those we've seen so far, may not refer to local variables declared outside the function - they do not "close over their environment". For example, you couldn't write the following:

let foo = 10;

fn bar() -> int {
   return foo; // `bar` cannot refer to `foo`
}

Rust also supports closures, functions that can access variables in the enclosing scope.

# use println = io::println;
fn call_closure_with_ten(b: fn(int)) { b(10); }

let captured_var = 20;
let closure = |arg| println(fmt!("captured_var=%d, arg=%d", captured_var, arg));

call_closure_with_ten(closure);

Closures begin with the argument list between bars and are followed by a single expression. The types of the arguments are generally omitted, as is the return type, because the compiler can almost always infer them. In the rare case where the compiler needs assistance though, the arguments and return types may be annotated.

# type mygoodness = fn(~str) -> ~str; type what_the = int;
let bloop = |well, oh: mygoodness| -> what_the { fail oh(well) };

There are several forms of closure, each with its own role. The most common, called a stack closure, has type fn& and can directly access local variables in the enclosing scope.

let mut max = 0;
(~[1, 2, 3]).map(|x| if *x > max { max = *x });

Stack closures are very efficient because their environment is allocated on the call stack and refers by pointer to captured locals. To ensure that stack closures never outlive the local variables to which they refer, they can only be used in argument position and cannot be stored in structures nor returned from functions. Despite the limitations stack closures are used pervasively in Rust code.

Managed closures

When you need to store a closure in a data structure, a stack closure will not do, since the compiler will refuse to let you store it. For this purpose, Rust provides a type of closure that has an arbitrary lifetime, written fn@ (boxed closure, analogous to the @ pointer type described earlier).

A managed closure does not directly access its environment, but merely copies out the values that it closes over into a private data structure. This means that it can not assign to these variables, and will not 'see' updates to them.

This code creates a closure that adds a given string to its argument, returns it from a function, and then calls it:

extern mod std;

fn mk_appender(suffix: ~str) -> fn@(~str) -> ~str {
    return fn@(s: ~str) -> ~str { s + suffix };
}

fn main() {
    let shout = mk_appender(~"!");
    io::println(shout(~"hey ho, let's go"));
}

This example uses the long closure syntax, fn@(s: ~str) ..., making the fact that we are declaring a box closure explicit. In practice boxed closures are usually defined with the short closure syntax introduced earlier, in which case the compiler will infer the type of closure. Thus our managed closure example could also be written:

fn mk_appender(suffix: ~str) -> fn@(~str) -> ~str {
    return |s| s + suffix;
}

Owned closures

Owned closures, written fn~ in analogy to the ~ pointer type, hold on to things that can safely be sent between processes. They copy the values they close over, much like managed closures, but they also 'own' them—meaning no other code can access them. Owned closures are used in concurrent code, particularly for spawning tasks.

Closure compatibility

A nice property of Rust closures is that you can pass any kind of closure (as long as the arguments and return types match) to functions that expect a fn(). Thus, when writing a higher-order function that wants to do nothing with its function argument beyond calling it, you should almost always specify the type of that argument as fn(), so that callers have the flexibility to pass whatever they want.

fn call_twice(f: fn()) { f(); f(); }
call_twice(|| { ~"I am an inferred stack closure"; } );
call_twice(fn&() { ~"I am also a stack closure"; } );
call_twice(fn@() { ~"I am a managed closure"; });
call_twice(fn~() { ~"I am a owned closure"; });
fn bare_function() { ~"I am a plain function"; }
call_twice(bare_function);

Note: Both the syntax and the semantics will be changing in small ways. At the moment they can be unsound in multiple scenarios, particularly with non-copyable types.

Do syntax

The do expression is syntactic sugar for use with functions which take a closure as a final argument, because closures in Rust are so frequently used in combination with higher-order functions.

Consider this function which iterates over a vector of integers, passing in a pointer to each integer in the vector:

fn each(v: &[int], op: fn(v: &int)) {
   let mut n = 0;
   while n < v.len() {
       op(&v[n]);
       n += 1;
   }
}

The reason we pass in a pointer to an integer rather than the integer itself is that this is how the actual each() function for vectors works. Using a pointer means that the function can be used for vectors of any type, even large structs that would be impractical to copy out of the vector on each iteration. As a caller, if we use a closure to provide the final operator argument, we can write it in a way that has a pleasant, block-like structure.

# fn each(v: &[int], op: fn(v: &int)) { }
# fn do_some_work(i: int) { }
each(&[1, 2, 3], |n| {
    debug!("%i", *n);
    do_some_work(*n);
});

This is such a useful pattern that Rust has a special form of function call that can be written more like a built-in control structure:

# fn each(v: &[int], op: fn(v: &int)) { }
# fn do_some_work(i: int) { }
do each(&[1, 2, 3]) |n| {
    debug!("%i", *n);
    do_some_work(*n);
}

The call is prefixed with the keyword do and, instead of writing the final closure inside the argument list it is moved outside of the parenthesis where it looks visually more like a typical block of code.

do is often used for task spawning.

use task::spawn;

do spawn() || {
    debug!("I'm a task, whatever");
}

That's nice, but look at all those bars and parentheses - that's two empty argument lists back to back. Wouldn't it be great if they weren't there?

# use task::spawn;
do spawn {
   debug!("Kablam!");
}

Empty argument lists can be omitted from do expressions.

For loops

Most iteration in Rust is done with for loops. Like do, for is a nice syntax for doing control flow with closures. Additionally, within a for loop, break, loop, and return work just as they do with while and loop.

Consider again our each function, this time improved to break early when the iteratee returns false:

fn each(v: &[int], op: fn(v: &int) -> bool) {
   let mut n = 0;
   while n < v.len() {
       if !op(&v[n]) {
           break;
       }
       n += 1;
   }
}

And using this function to iterate over a vector:

# use each = vec::each;
# use println = io::println;
each(&[2, 4, 8, 5, 16], |n| {
    if *n % 2 != 0 {
        println(~"found odd number!");
        false
    } else { true }
});

With for, functions like each can be treated more like builtin looping structures. When calling each in a for loop, instead of returning false to break out of the loop, you just write break. To skip ahead to the next iteration, write loop.

# use each = vec::each;
# use println = io::println;
for each(&[2, 4, 8, 5, 16]) |n| {
    if *n % 2 != 0 {
        println(~"found odd number!");
        break;
    }
}

As an added bonus, you can use the return keyword, which is not normally allowed in closures, in a block that appears as the body of a for loop — this will cause a return to happen from the outer function, not just the loop body.

# use each = vec::each;
fn contains(v: &[int], elt: int) -> bool {
    for each(v) |x| {
        if (*x == elt) { return true; }
    }
    false
}

for syntax only works with stack closures.

Note: This is, essentially, a special loop protocol: the keywords break, loop, and return work, in varying degree, with while, loop, do, and for constructs.

Generics

Throughout this tutorial, we've been defining functions that act only on specific data types. With type parameters we can also define functions whose arguments represent generic types, and which can be invoked with a variety of types. Consider a generic map function.

fn map<T, U>(vector: &[T], function: fn(v: &T) -> U) -> ~[U] {
    let mut accumulator = ~[];
    for vec::each(vector) |element| {
        accumulator.push(function(element));
    }
    return accumulator;
}

When defined with type parameters, as denoted by <T, U>, this function can be applied to any type of vector, as long as the type of function's argument and the type of the vector's content agree with each other.

Inside a generic function, the names of the type parameters (capitalized by convention) stand for opaque types. You can't look inside them, but you can pass them around. Note that instances of generic types are often passed by pointer. For example, the parameter function() is supplied with a pointer to a value of type T and not a value of type T itself. This ensures that the function works with the broadest set of types possible, since some types are expensive or illegal to copy and pass by value.

Generic type, struct, and enum declarations follow the same pattern:

# use std::map::HashMap;
type Set<T> = HashMap<T, ()>;

struct Stack<T> {
    elements: ~[mut T]
}

enum Maybe<T> {
    Just(T),
    Nothing
}

These declarations produce valid types like Set<int>, Stack<int> and Maybe<int>.

Generic functions in Rust are compiled to very efficient runtime code through a process called monomorphisation. This is a fancy way of saying that, for each generic function you call, the compiler generates a specialized version that is optimized specifically for the argument types. In this respect Rust's generics have similar performance characteristics to C++ templates.

Traits

Within a generic function the operations available on generic types are very limited. After all, since the function doesn't know what types it is operating on, it can't safely modify or query their values. This is where traits come into play. Traits are Rust's most powerful tool for writing polymorphic code. Java developers will see in them aspects of Java interfaces, and Haskellers will notice their similarities to type classes.

As motivation, let us consider copying in Rust. Perhaps surprisingly, the copy operation is not defined for all Rust types. In particular, types with user-defined destructors cannot be copied, either implicitly or explicitly, and neither can types that own other types containing destructors (the actual mechanism for defining destructors will be discussed elsewhere).

This complicates handling of generic functions. If you have a type parameter T, can you copy values of that type? In Rust, you can't, and if you try to run the following code the compiler will complain.

// This does not compile
fn head_bad<T>(v: &[T]) -> T {
    v[0] // error: copying a non-copyable value
}

We can tell the compiler though that the head function is only for copyable types with the Copy trait.

// This does
fn head<T: Copy>(v: &[T]) -> T {
    v[0]
}

This says that we can call head on any type T as long as that type implements the Copy trait. When instantiating a generic function, you can only instantiate it with types that implement the correct trait, so you could not apply head to a type with a destructor.

While most traits can be defined and implemented by user code, three traits are automatically derived and implemented for all applicable types by the compiler, and may not be overridden:

  • Copy - Types that can be copied, either implicitly, or using the copy expression. All types are copyable unless they are classes with destructors or otherwise contain classes with destructors.

  • Send - Sendable (owned) types. All types are sendable unless they contain managed boxes, managed closures, or otherwise managed types. Sendable types may or may not be copyable.

  • Const - Constant (immutable) types. These are types that do not contain mutable fields.

Note: These three traits were referred to as 'kinds' in earlier iterations of the language, and often still are.

Declaring and implementing traits

A trait consists of a set of methods, without bodies, or may be empty, as is the case with Copy, Send, and Const. For example, we could declare the trait Printable for things that can be printed to the console, with a single method:

trait Printable {
    fn print();
}

Traits may be implemented for specific types with impls. An impl that implements a trait includes the name of the trait at the start of the definition, as in the following impls of Printable for int and ~str.

# trait Printable { fn print(); }
impl int: Printable {
    fn print() { io::println(fmt!("%d", self)) }
}

impl ~str: Printable {
    fn print() { io::println(self) }
}

# 1.print();
# (~"foo").print();

Methods defined in an implementation of a trait may be called just as any other method, using dot notation, as in 1.print(). Traits may themselves contain type parameters. A trait for generalized sequence types might look like the following:

trait Seq<T> {
    fn len() -> uint;
    fn iter(b: fn(v: &T));
}

impl<T> ~[T]: Seq<T> {
    fn len() -> uint { vec::len(self) }
    fn iter(b: fn(v: &T)) {
        for vec::each(self) |elt| { b(elt); }
    }
}

The implementation has to explicitly declare the type parameter that it binds, T, before using it to specify its trait type. Rust requires this declaration because the impl could also, for example, specify an implementation of Seq<int>. The trait type -- appearing after the colon in the impl -- refers to a type, rather than defining one.

The type parameters bound by a trait are in scope in each of the method declarations. So, re-declaring the type parameter T as an explicit type parameter for len -- in either the trait or the impl -- would be a compile-time error.

Within a trait definition, self is a special type that you can think of as a type parameter. An implementation of the trait for any given type T replaces the self type parameter with T. Simply, in a trait, self is a type, and in an impl, self is a value. The following trait describes types that support an equality operation:

// In a trait, `self` refers to the type implementing the trait
trait Eq {
  fn equals(other: &self) -> bool;
}

// In an impl, self refers to the value of the receiver
impl int: Eq {
  fn equals(other: &int) -> bool { *other == self }
}

Notice that in the trait definition, equals takes a self type argument, whereas, in the impl, equals takes an int type argument, and uses self as the name of the receiver (analogous to the this pointer in C++).

Bounded type parameters and static method dispatch

Traits give us a language for talking about the abstract capabilities of types, and we can use this to place bounds on type parameters, so that we can then operate on generic types.

# trait Printable { fn print(); }
fn print_all<T: Printable>(printable_things: ~[T]) {
    for printable_things.each |thing| {
        thing.print();
    }
}

By declaring T as conforming to the Printable trait (as we earlier did with Copy), it becomes possible to call methods from that trait on values of that type inside the function. It will also cause a compile-time error when anyone tries to call print_all on an array whose element type does not have a Printable implementation.

Type parameters can have multiple bounds by separating them with spaces, as in this version of print_all that makes copies of elements.

# trait Printable { fn print(); }
fn print_all<T: Printable Copy>(printable_things: ~[T]) {
    let mut i = 0;
    while i < printable_things.len() {
        let copy_of_thing = printable_things[0];
        copy_of_thing.print();
    }
}

Method calls to bounded type parameters are statically dispatched, imposing no more overhead than normal function invocation, so are the preferred way to use traits polymorphically.

This usage of traits is similar to Haskell type classes.

Casting to a trait type and dynamic method dispatch

The above allows us to define functions that polymorphically act on values of a single unknown type that conforms to a given trait. However, consider this function:

# type Circle = int; type Rectangle = int;
# impl int: Drawable { fn draw() {} }
# fn new_circle() -> int { 1 }

trait Drawable { fn draw(); }

fn draw_all<T: Drawable>(shapes: ~[T]) {
    for shapes.each |shape| { shape.draw(); }
}

# let c: Circle = new_circle();
# draw_all(~[c]);

You can call that on an array of circles, or an array of squares (assuming those have suitable Drawable traits defined), but not on an array containing both circles and squares. When such behavior is needed, a trait name can alternately be used as a type.

# trait Drawable { fn draw(); }
fn draw_all(shapes: ~[@Drawable]) {
    for shapes.each |shape| { shape.draw(); }
}

In this example there is no type parameter. Instead, the @Drawable type is used to refer to any managed box value that implements the Drawable trait. To construct such a value, you use the as operator to cast a value to a trait type:

# type Circle = int; type Rectangle = bool;
# trait Drawable { fn draw(); }
# fn new_circle() -> Circle { 1 }
# fn new_rectangle() -> Rectangle { true }
# fn draw_all(shapes: ~[Drawable]) {}

impl @Circle: Drawable { fn draw() { ... } }

impl @Rectangle: Drawable { fn draw() { ... } }

let c: @Circle = @new_circle();
let r: @Rectangle = @new_rectangle();
draw_all(~[c as @Drawable, r as @Drawable]);

Note that, like strings and vectors, trait types have dynamic size and may only be used via one of the pointer types. In turn, the impl is defined for @Circle and @Rectangle instead of for just Circle and Rectangle. Other pointer types work as well.

# type Circle = int; type Rectangle = int;
# trait Drawable { fn draw(); }
# impl int: Drawable { fn draw() {} }
# fn new_circle() -> int { 1 }
# fn new_rectangle() -> int { 2 }
// A managed trait instance
let boxy: @Drawable = @new_circle() as @Drawable;
// An owned trait instance
let owny: ~Drawable = ~new_circle() as ~Drawable;
// A borrowed trait instance
let stacky: &Drawable = &new_circle() as &Drawable;

Note: Other pointer types actually do not work here. This is an evolving corner of the language.

Method calls to trait types are dynamically dispatched. Since the compiler doesn't know specifically which functions to call at compile time it uses a lookup table (vtable) to decide at runtime which method to call.

This usage of traits is similar to Java interfaces.

Modules and crates

The Rust namespace is divided into modules. Each source file starts with its own module.

Local modules

The mod keyword can be used to open a new, local module. In the example below, chicken lives in the module farm, so, unless you explicitly import it, you must refer to it by its long name, farm::chicken.

#[legacy_exports]
mod farm {
    fn chicken() -> ~str { ~"cluck cluck" }
    fn cow() -> ~str { ~"mooo" }
}
fn main() {
    io::println(farm::chicken());
}

Modules can be nested to arbitrary depth.

Crates

The unit of independent compilation in Rust is the crate. Libraries tend to be packaged as crates, and your own programs may consist of one or more crates.

When compiling a single .rs file, the file acts as the whole crate. You can compile it with the --lib compiler switch to create a shared library, or without, provided that your file contains a fn main somewhere, to create an executable.

It is also possible to include multiple files in a crate. For this purpose, you create a .rc crate file, which references any number of .rs code files. A crate file could look like this:

#[link(name = "farm", vers = "2.5", author = "mjh")];
#[crate_type = "lib"];
mod cow;
mod chicken;
mod horse;

Compiling this file will cause rustc to look for files named cow.rs, chicken.rs, horse.rs in the same directory as the .rc file, compile them all together, and, depending on the presence of the crate_type = "lib" attribute, output a shared library or an executable. (If the line #[crate_type = "lib"]; was omitted, rustc would create an executable.)

The #[link(...)] part provides meta information about the module, which other crates can use to load the right module. More about that later.

To have a nested directory structure for your source files, you can nest mods in your .rc file:

mod poultry {
    mod chicken;
    mod turkey;
}

The compiler will now look for poultry/chicken.rs and poultry/turkey.rs, and export their content in poultry::chicken and poultry::turkey. You can also provide a poultry.rs to add content to the poultry module itself.

The compiler then builds the crate as a platform-specific shared library or executable which can be distributed.

Using other crates

Having compiled a crate that contains the #[crate_type = "lib"] attribute, you can use it in another crate with a use directive. We've already seen extern mod std in several of the examples, which loads in the standard library.

use directives can appear in a crate file, or at the top level of a single-file .rs crate. They will cause the compiler to search its library search path (which you can extend with -L switch) for a Rust crate library with the right name.

It is possible to provide more specific information when using an external crate.

extern mod myfarm (name = "farm", vers = "2.7");

When a comma-separated list of name/value pairs is given after use, these are matched against the attributes provided in the link attribute of the crate file, and a crate is only used when the two match. A name value can be given to override the name used to search for the crate. So the above would import the farm crate under the local name myfarm.

Our example crate declared this set of link attributes:

#[link(name = "farm", vers = "2.5", author = "mjh")];

The version does not match the one provided in the use directive, so unless the compiler can find another crate with the right version somewhere, it will complain that no matching crate was found.

The core library

A set of basic library routines, mostly related to built-in datatypes and the task system, are always implicitly linked and included in any Rust program.

This library is documented here.

A minimal example

Now for something that you can actually compile yourself. We have these two files:

// world.rs
#[link(name = "world", vers = "1.0")];
fn explore() -> ~str { ~"world" }
// main.rs
extern mod world;
fn main() { io::println(~"hello " + world::explore()); }

Now compile and run like this (adjust to your platform if necessary):

> rustc --lib world.rs  # compiles libworld-94839cbfe144198-1.0.so
> rustc main.rs -L .    # compiles main
> ./main
"hello world"

Importing

When using identifiers from other modules, it can get tiresome to qualify them with the full module path every time (especially when that path is several modules deep). Rust allows you to import identifiers at the top of a file, module, or block.

extern mod std;
use io::println;
fn main() {
    println(~"that was easy");
}

It is also possible to import just the name of a module (use std::list;, then use list::find), to import all identifiers exported by a given module (use io::*), or to import a specific set of identifiers (use math::{min, max, pi}).

Rust uses different namespaces for modules, types, and values. You can also rename an identifier when importing using the = operator:

use prnt = io::println;

Exporting

By default, a module exports everything that it defines. This can be restricted with export directives at the top of the module or file.

mod enc {
    export encrypt, decrypt;
    const SUPER_SECRET_NUMBER: int = 10;
    fn encrypt(n: int) -> int { n + SUPER_SECRET_NUMBER }
    fn decrypt(n: int) -> int { n - SUPER_SECRET_NUMBER }
}

This defines a rock-solid encryption algorithm. Code outside of the module can refer to the enc::encrypt and enc::decrypt identifiers just fine, but it does not have access to enc::super_secret_number.

Resolution

The resolution process in Rust simply goes up the chain of contexts, looking for the name in each context. Nested functions and modules create new contexts inside their parent function or module. A file that's part of a bigger crate will have that crate's context as its parent context.

Identifiers can shadow each other. In this program, x is of type int:

type MyType = ~str;
fn main() {
    type MyType = int;
    let x: MyType = 17;
}

An use directive will only import into the namespaces for which identifiers are actually found. Consider this example:

mod foo {
   fn bar() {}
}

fn main() {
    let bar = 10;

    {
        use foo::bar;
        let quux = bar;
        assert quux == 10;
    }
}

When resolving the type name bar in the quux definition, the resolver will first look at local block context for baz. This has an import named bar, but that's function, not a value, So it continues to the baz function context and finds a value named bar defined there.

Normally, multiple definitions of the same identifier in a scope are disallowed. Local variables defined with let are an exception to this—multiple let directives can redefine the same variable in a single scope. When resolving the name of such a variable, the most recent definition is used.

fn main() {
    let x = 10;
    let x = x + 10;
    assert x == 20;
}

This makes it possible to rebind a variable without actually mutating it, which is mostly useful for destructuring (which can rebind, but not assign).

What next?

Now that you know the essentials, check out any of the additional tutorials on individual topics.

There is further documentation on the wiki, including articles about unit testing in Rust, documenting and packaging Rust code, and a discussion of the attributes used to apply metada to code.