1 % The Rust Language Tutorial
5 Rust is a programming language with a focus on type safety, memory
6 safety, concurrency and performance. It is intended for writing
7 large-scale, high-performance software that is free from several
8 classes of common errors. Rust has a sophisticated memory model that
9 encourages efficient data structures and safe concurrency patterns,
10 forbidding invalid memory accesses that would otherwise cause
11 segmentation faults. It is statically typed and compiled ahead of
14 As a multi-paradigm language, Rust supports writing code in
15 procedural, functional and object-oriented styles. Some of its
16 pleasant high-level features include:
18 * **Type inference.** Type annotations on local variable declarations
20 * **Safe task-based concurrency.** Rust's lightweight tasks do not share
21 memory, instead communicating through messages.
22 * **Higher-order functions.** Efficient and flexible closures provide
23 iteration and other control structures
24 * **Pattern matching and algebraic data types.** Pattern matching on
25 Rust's enumeration types (a more powerful version of C's enums,
26 similar to algebraic data types in functional languages) is a
27 compact and expressive way to encode program logic.
28 * **Polymorphism.** Rust has type-parametric functions and
29 types, type classes and OO-style interfaces.
33 This is an introductory tutorial for the Rust programming language. It
34 covers the fundamentals of the language, including the syntax, the
35 type system and memory model, generics, and modules. [Additional
36 tutorials](#what-next?) cover specific language features in greater
39 This tutorial assumes that the reader is already familiar with one or
40 more languages in the C family. Understanding of pointers and general
41 memory management techniques will help.
45 Throughout the tutorial, language keywords and identifiers defined in
46 example code are displayed in `code font`.
48 Code snippets are indented, and also shown in a monospaced font. Not
49 all snippets constitute whole programs. For brevity, we'll often show
50 fragments of programs that don't compile on their own. To try them
51 out, you might have to wrap them in `fn main() { ... }`, and make sure
52 they don't contain references to names that aren't actually defined.
54 > *Warning:* Rust is a language under ongoing development. Notes
55 > about potential changes to the language, implementation
56 > deficiencies, and other caveats appear offset in blockquotes.
60 There are two ways to install the Rust compiler: by building from source or
61 by downloading prebuilt binaries or installers for your platform. The
62 [install page][rust-install] contains links to download binaries for both
63 the nightly build and the most current Rust major release. For Windows and
64 OS X, the install page provides links to native installers.
66 > *Note:* Windows users should read the detailed
67 > [Getting started][wiki-start] notes on the wiki. Even when using
68 > the binary installer, the Windows build requires a MinGW installation,
69 > the precise details of which are not discussed here.
71 For Linux and OS X, the install page provides links to binary tarballs.
72 To install the Rust compiler from the from a binary tarball, download
73 the binary package, extract it, and execute the `install.sh` script in
74 the root directory of the package.
76 To build the Rust compiler from source, you will need to obtain the source through
77 [Git][git] or by downloading the source package from the [install page][rust-install].
79 Since the Rust compiler is written in Rust, it must be built by
80 a precompiled "snapshot" version of itself (made in an earlier state
81 of development). The source build automatically fetches these snapshots
82 from the Internet on our supported platforms.
84 Snapshot binaries are currently built and tested on several platforms:
86 * Windows (7, 8, Server 2008 R2), x86 only
87 * Linux (2.6.18 or later, various distributions), x86 and x86-64
88 * OSX 10.7 (Lion) or greater, x86 and x86-64
90 You may find that other platforms work, but these are our "tier 1"
91 supported build environments that are most likely to work.
93 [wiki-start]: https://github.com/rust-lang/rust/wiki/Note-getting-started-developing-Rust
94 [git]: https://github.com/rust-lang/rust.git
95 [rust-install]: http://www.rust-lang.org/install.html
97 To build from source you will also need the following prerequisite
100 * g++ 4.7 or clang++ 3.x
101 * python 2.6 or later (but not 3.x)
103 * gnu make 3.81 or later
106 If you've fulfilled those prerequisites, something along these lines
110 $ curl -O http://static.rust-lang.org/dist/rust-nightly.tar.gz
111 $ tar -xzf rust-nightly.tar.gz
114 $ make && make install
117 You may need to use `sudo make install` if you do not normally have
118 permission to modify the destination directory. The install locations
119 can be adjusted by passing a `--prefix` argument to
120 `configure`. Various other options are also supported: pass `--help`
121 for more information on them.
123 When complete, `make install` will place several programs into
124 `/usr/local/bin`: `rustc`, the Rust compiler, and `rustdoc`, the
125 API-documentation tool.
127 [tarball]: http://static.rust-lang.org/dist/rust-nightly.tar.gz
128 [win-exe]: http://static.rust-lang.org/dist/rust-nightly-install.exe
130 ## Compiling your first program
132 Rust program files are, by convention, given the extension `.rs`. Say
133 we have a file `hello.rs` containing this program:
140 > *Note:* An identifier followed by an exclamation point, like
141 > `println!`, is a macro invocation. Macros are explained
142 > [later](#syntax-extensions); for now just remember to include the
145 If the Rust compiler was installed successfully, running `rustc
146 hello.rs` will produce an executable called `hello` (or `hello.exe` on
147 Windows) which, upon running, will likely do exactly what you expect.
149 The Rust compiler tries to provide useful information when it encounters an
150 error. If you introduce an error into the program (for example, by changing
151 `println!` to some nonexistent macro), and then compile it, you'll see
152 an error message like this:
155 hello.rs:2:5: 2:24 error: macro undefined: 'print_with_unicorns'
156 hello.rs:2 print_with_unicorns!("hello?");
160 In its simplest form, a Rust program is a `.rs` file with some types
161 and functions defined in it. If it has a `main` function, it can be
162 compiled to an executable. Rust does not allow code that's not a
163 declaration to appear at the top level of the file: all statements must
164 live inside a function. Rust programs can also be compiled as
165 libraries, and included in other programs, even ones not written in Rust.
169 There are vim highlighting and indentation scripts in the Rust source
170 distribution under `src/etc/vim/`. There is an emacs mode under
171 `src/etc/emacs/` called `rust-mode`, but do read the instructions
172 included in that directory. In particular, if you are running emacs
173 24, then using emacs's internal package manager to install `rust-mode`
174 is the easiest way to keep it up to date. There is also a package for
175 Sublime Text 2, available both [standalone][sublime] and through
176 [Sublime Package Control][sublime-pkg], and support for Kate
177 under `src/etc/kate`.
179 A community-maintained list of available Rust tooling is [on the
180 wiki][wiki-packages].
182 There is ctags support via `src/etc/ctags.rust`, but many other
183 tools and editors are not yet supported. If you end up writing a Rust
184 mode for your favorite editor, let us know so that we can link to it.
186 [sublime]: http://github.com/jhasse/sublime-rust
187 [sublime-pkg]: http://wbond.net/sublime_packages/package_control
191 Assuming you've programmed in any C-family language (C++, Java,
192 JavaScript, C#, or PHP), Rust will feel familiar. Code is arranged
193 in blocks delineated by curly braces; there are control structures
194 for branching and looping, like the familiar `if` and `while`; function
195 calls are written `myfunc(arg1, arg2)`; operators are written the same
196 and mostly have the same precedence as in C; comments are again like C;
197 module names are separated with double-colon (`::`) as with C++.
199 The main surface difference to be aware of is that the condition at
200 the head of control structures like `if` and `while` does not require
201 parentheses, while their bodies *must* be wrapped in
202 braces. Single-statement, unbraced bodies are not allowed.
205 # mod universe { pub fn recalibrate() -> bool { true } }
209 // A tricky calculation
210 if universe::recalibrate() {
217 The `let` keyword introduces a local variable. Variables are immutable by
218 default. To introduce a local variable that you can re-assign later, use `let
226 println!("count is {}", count);
231 Although Rust can almost always infer the types of local variables, you can
232 specify a variable's type by following it in the `let` with a colon, then the
233 type name. Static items, on the other hand, always require a type annotation.
237 static MONSTER_FACTOR: f64 = 57.8;
238 let monster_size = MONSTER_FACTOR * 10.0;
239 let monster_size: int = 50;
242 Local variables may shadow earlier declarations, as in the previous example:
243 `monster_size` was first declared as a `f64`, and then a second
244 `monster_size` was declared as an `int`. If you were to actually compile this
245 example, though, the compiler would determine that the first `monster_size` is
246 unused and issue a warning (because this situation is likely to indicate a
247 programmer error). For occasions where unused variables are intentional, their
248 names may be prefixed with an underscore to silence the warning, like `let
249 _monster_size = 50;`.
251 Rust identifiers start with an alphabetic
252 character or an underscore, and after that may contain any sequence of
253 alphabetic characters, numbers, or underscores. The preferred style is to
254 write function, variable, and module names with lowercase letters, using
255 underscores where they help readability, while writing types in camel case.
258 let my_variable = 100;
259 type MyType = int; // primitive types are _not_ camel case
262 ## Expressions and semicolons
264 Though it isn't apparent in all code, there is a fundamental
265 difference between Rust's syntax and predecessors like C.
266 Many constructs that are statements in C are expressions
267 in Rust, allowing code to be more concise. For example, you might
268 write a piece of code like this:
271 # let item = "salad";
275 } else if item == "muffin" {
282 But, in Rust, you don't have to repeat the name `price`:
285 # let item = "salad";
289 } else if item == "muffin" {
296 Both pieces of code are exactly equivalent: they assign a value to
297 `price` depending on the condition that holds. Note that there
298 are no semicolons in the blocks of the second snippet. This is
299 important: the lack of a semicolon after the last statement in a
300 braced block gives the whole block the value of that last expression.
302 Put another way, the semicolon in Rust *ignores the value of an expression*.
303 Thus, if the branches of the `if` had looked like `{ 4; }`, the above example
304 would simply assign `()` (unit or void) to `price`. But without the semicolon, each
305 branch has a different value, and `price` gets the value of the branch that
308 In short, everything that's not a declaration (declarations are `let` for
309 variables; `fn` for functions; and any top-level named items such as
310 [traits](#traits), [enum types](#enums), and static items) is an
311 expression, including function bodies.
314 fn is_four(x: int) -> bool {
315 // No need for a return statement. The result of the expression
316 // is used as the return value.
321 ## Primitive types and literals
323 There are general signed and unsigned integer types, `int` and `uint`,
324 as well as 8-, 16-, 32-, and 64-bit variants, `i8`, `u16`, etc.
325 Integers can be written in decimal (`144`), hexadecimal (`0x90`), octal (`0o70`), or
326 binary (`0b10010000`) base. Each integral type has a corresponding literal
327 suffix that can be used to indicate the type of a literal: `i` for `int`,
328 `u` for `uint`, `i8` for the `i8` type.
330 In the absence of an integer literal suffix, Rust will infer the
331 integer type based on type annotations and function signatures in the
332 surrounding program. In the absence of any type information at all,
333 Rust will assume that an unsuffixed integer literal has type
337 let a = 1; // `a` is an `int`
338 let b = 10i; // `b` is an `int`, due to the `i` suffix
339 let c = 100u; // `c` is a `uint`
340 let d = 1000i32; // `d` is an `i32`
343 There are two floating-point types: `f32`, and `f64`.
344 Floating-point numbers are written `0.0`, `1e6`, or `2.1e-4`.
345 Like integers, floating-point literals are inferred to the correct type.
346 Suffixes `f32`, and `f64` can be used to create literals of a specific type.
348 The keywords `true` and `false` produce literals of type `bool`.
350 Characters, the `char` type, are four-byte Unicode codepoints,
351 whose literals are written between single quotes, as in `'x'`.
352 Just like C, Rust understands a number of character escapes, using the backslash
353 character, such as `\n`, `\r`, and `\t`. String literals,
354 written between double quotes, allow the same escape sequences, and do no
355 other processing, unlike languages such as PHP or shell.
357 On the other hand, raw string literals do not process any escape sequences.
358 They are written as `r##"blah"##`, with a matching number of zero or more `#`
359 before the opening and after the closing quote, and can contain any sequence of
360 characters except their closing delimiter. More on strings
361 [later](#vectors-and-strings).
363 The unit type, written `()`, has a single value, also written `()`.
367 Rust's set of operators contains very few surprises. Arithmetic is done with
368 `*`, `/`, `%`, `+`, and `-` (multiply, quotient, remainder, add, and subtract). `-` is
369 also a unary prefix operator that negates numbers. As in C, the bitwise operators
370 `>>`, `<<`, `&`, `|`, and `^` are also supported.
372 Note that, if applied to an integer value, `!` flips all the bits (bitwise
375 The comparison operators are the traditional `==`, `!=`, `<`, `>`,
376 `<=`, and `>=`. Short-circuiting (lazy) boolean operators are written
377 `&&` (and) and `||` (or).
379 For compile-time type casting, Rust uses the binary `as` operator. It takes
380 an expression on the left side and a type on the right side and will, if a
381 meaningful conversion exists, convert the result of the expression to the
382 given type. Generally, `as` is only used with the primitive numeric types or
383 pointers, and is not overloadable. [`transmute`][transmute] can be used for
384 unsafe C-like casting of same-sized types.
388 let y: uint = x as uint;
392 [transmute]: http://doc.rust-lang.org/std/mem/fn.transmute.html
396 *Syntax extensions* are special forms that are not built into the language,
397 but are instead provided by the libraries. To make it clear to the reader when
398 a name refers to a syntax extension, the names of all syntax extensions end
399 with `!`. The standard library defines a few syntax extensions, the most
400 useful of which is [`format!`][fmt], a `sprintf`-like text formatter that you
401 will often see in examples, and its related family of macros: `print!`,
402 `println!`, and `write!`.
404 `format!` draws syntax from Python, but contains many of the same principles
405 that [printf][pf] has. Unlike printf, `format!` will give you a compile-time
406 error when the types of the directives don't match the types of the arguments.
409 // `{}` will print the "default format" of a type
410 println!("{} is {}", "the answer", 43i);
417 # let mystery_object = ();
418 // `{:?}` will conveniently print any type,
419 // but requires the `debug` crate to be linked in
420 println!("what is this thing: {:?}", mystery_object);
424 [pf]: http://en.cppreference.com/w/cpp/io/c/fprintf
425 [fmt]: http://doc.rust-lang.org/std/fmt/
427 You can define your own syntax extensions with the macro system. For details,
428 see the [macro tutorial][macros]. Note that macro definition is currently
429 considered an unstable feature.
435 We've seen `if` expressions a few times already. To recap, braces are
436 compulsory, an `if` can have an optional `else` clause, and multiple
437 `if`/`else` constructs can be chained together:
441 println!("that's odd");
445 println!("neither true nor false");
449 The condition given to an `if` construct *must* be of type `bool` (no
450 implicit conversion happens). If the arms are blocks that have a
451 value, this value must be of the same type for every arm in which
452 control reaches the end of the block:
455 fn signum(x: int) -> int {
464 Rust's `match` construct is a generalized, cleaned-up version of C's
465 `switch` construct. You provide it with a value and a number of
466 *arms*, each labelled with a pattern, and the code compares the value
467 against each pattern in order until one matches. The matching pattern
468 executes its corresponding arm.
473 0 => println!("zero"),
474 1 | 2 => println!("one or two"),
475 3..10 => println!("three to ten"),
476 _ => println!("something else")
480 Unlike in C, there is no "falling through" between arms: only one arm
481 executes, and it doesn't have to explicitly `break` out of the
482 construct when it is finished.
484 A `match` arm consists of a *pattern*, then a fat arrow `=>`, followed
485 by an *action* (expression). Each case is separated by commas. It is
486 often convenient to use a block expression for each case, in which case
487 the commas are optional as shown below. Literals are valid patterns and
488 match only their own value. A single arm may match multiple different
489 patterns by combining them with the pipe operator (`|`), so long as every
490 pattern binds the same set of variables. Ranges of numeric literal
491 patterns can be expressed with two dots, as in `M..N`. The underscore
492 (`_`) is a wildcard pattern that matches any single value. (`..`) is a
493 different wildcard that can match one or more fields in an `enum` variant.
498 0 => { println!("zero") }
499 _ => { println!("something else") }
503 `match` constructs must be *exhaustive*: they must have an arm
504 covering every possible case. For example, the typechecker would
505 reject the previous example if the arm with the wildcard pattern was
508 A powerful application of pattern matching is *destructuring*:
509 matching in order to bind names to the contents of data types.
511 > *Note:* The following code makes use of tuples (`(f64, f64)`) which
512 > are explained in section 5.3. For now you can think of tuples as a list of
517 fn angle(vector: (f64, f64)) -> f64 {
518 let pi = f64::consts::PI;
520 (0.0, y) if y < 0.0 => 1.5 * pi,
521 (0.0, _) => 0.5 * pi,
522 (x, y) => (y / x).atan()
527 A variable name in a pattern matches any value, *and* binds that name
528 to the value of the matched value inside of the arm's action. Thus, `(0.0,
529 y)` matches any tuple whose first element is zero, and binds `y` to
530 the second element. `(x, y)` matches any two-element tuple, and binds both
531 elements to variables. `(0.0,_)` matches any tuple whose first element is zero
532 and does not bind anything to the second element.
534 A subpattern can also be bound to a variable, using `variable @ pattern`. For
540 a @ 0..20 => println!("{} years old", a),
541 _ => println!("older than 21")
545 Any `match` arm can have a guard clause (written `if EXPR`), called a
546 *pattern guard*, which is an expression of type `bool` that
547 determines, after the pattern is found to match, whether the arm is
548 taken or not. The variables bound by the pattern are in scope in this
549 guard expression. The first arm in the `angle` example shows an
550 example of a pattern guard.
552 You've already seen simple `let` bindings, but `let` is a little
553 fancier than you've been led to believe. It, too, supports destructuring
554 patterns. For example, you can write this to extract the fields from a
555 tuple, introducing two variables at once: `a` and `b`.
558 # fn get_tuple_of_two_ints() -> (int, int) { (1, 1) }
559 let (a, b) = get_tuple_of_two_ints();
562 Let bindings only work with _irrefutable_ patterns: that is, patterns
563 that can never fail to match. This excludes `let` from matching
564 literals and most `enum` variants.
568 `while` denotes a loop that iterates as long as its given condition
569 (which must have type `bool`) evaluates to `true`. Inside a loop, the
570 keyword `break` aborts the loop, and `continue` aborts the current
571 iteration and continues with the next.
574 let mut cake_amount = 8;
575 while cake_amount > 0 {
580 `loop` denotes an infinite loop, and is the preferred way of writing `while true`:
586 if x % 5 == 0 { break; }
591 This code prints out a weird sequence of numbers and stops as soon as
592 it finds one that can be divided by five.
594 There is also a for-loop that can be used to iterate over a range of numbers:
597 for n in range(0u, 5) {
602 The snippet above prints integer numbers under 5 starting at 0.
604 More generally, a for loop works with anything implementing the `Iterator` trait.
605 Data structures can provide one or more methods that return iterators over
606 their contents. For example, strings support iteration over their contents in
616 The snippet above prints the characters in "Hello" vertically, adding a new
617 line after each character.
624 Rust struct types must be declared before they are used using the `struct`
625 syntax: `struct Name { field1: T1, field2: T2 [, ...] }`, where `T1`, `T2`,
626 ... denote types. To construct a struct, use the same syntax, but leave off
627 the `struct`: for example: `Point { x: 1.0, y: 2.0 }`.
629 Structs are quite similar to C structs and are even laid out the same way in
630 memory (so you can read from a Rust struct in C, and vice-versa). Use the dot
631 operator to access struct fields, as in `mypoint.x`.
640 Structs have "inherited mutability", which means that any field of a struct
641 may be mutable, if the struct is in a mutable slot.
643 With a value (say, `mypoint`) of such a type in a mutable location, you can do
644 `mypoint.y += 1.0`. But in an immutable location, such an assignment to a
645 struct without inherited mutability would result in a type error.
648 # struct Point { x: f64, y: f64 }
649 let mut mypoint = Point { x: 1.0, y: 1.0 };
650 let origin = Point { x: 0.0, y: 0.0 };
652 mypoint.y += 1.0; // `mypoint` is mutable, and its fields as well
653 origin.y += 1.0; // ERROR: assigning to immutable field
656 `match` patterns destructure structs. The basic syntax is
657 `Name { fieldname: pattern, ... }`:
660 # struct Point { x: f64, y: f64 }
661 # let mypoint = Point { x: 0.0, y: 0.0 };
663 Point { x: 0.0, y: yy } => println!("{}", yy),
664 Point { x: xx, y: yy } => println!("{} {}", xx, yy)
668 In general, the field names of a struct do not have to appear in the same
669 order they appear in the type. When you are not interested in all
670 the fields of a struct, a struct pattern may end with `, ..` (as in
671 `Name { field1, .. }`) to indicate that you're ignoring all other fields.
672 Additionally, struct fields have a shorthand matching form that simply
673 reuses the field name as the binding name.
676 # struct Point { x: f64, y: f64 }
677 # let mypoint = Point { x: 0.0, y: 0.0 };
679 Point { x, .. } => println!("{}", x)
685 Enums are datatypes with several alternate representations. A simple `enum`
686 defines one or more constants, all of which have the same type:
697 Each variant of this enum has a unique and constant integral discriminator
698 value. If no explicit discriminator is specified for a variant, the value
699 defaults to the value of the previous variant plus one. If the first variant
700 does not have a discriminator, it defaults to 0. For example, the value of
701 `North` is 0, `East` is 1, `South` is 2, and `West` is 3.
703 When an enum has simple integer discriminators, you can apply the `as` cast
704 operator to convert a variant to its discriminator value as an `int`:
707 # #[deriving(Show)] enum Direction { North }
708 println!( "{} => {}", North, North as int );
711 It is possible to set the discriminator values to chosen constant values:
721 Variants do not have to be simple values; they may be more complex:
724 # struct Point { x: f64, y: f64 }
727 Rectangle(Point, Point)
731 A value of this type is either a `Circle`, in which case it contains a
732 `Point` struct and a f64, or a `Rectangle`, in which case it contains
733 two `Point` structs. The run-time representation of such a value
734 includes an identifier of the actual form that it holds, much like the
735 "tagged union" pattern in C, but with better static guarantees.
737 This declaration defines a type `Shape` that can refer to such shapes, and two
738 functions, `Circle` and `Rectangle`, which can be used to construct values of
739 the type. To create a new Circle, write `Circle(Point { x: 0.0, y: 0.0 },
742 All of these variant constructors may be used as patterns. The only way to
743 access the contents of an enum instance is the destructuring of a match. For
748 # struct Point {x: f64, y: f64}
749 # enum Shape { Circle(Point, f64), Rectangle(Point, Point) }
750 fn area(sh: Shape) -> f64 {
752 Circle(_, size) => f64::consts::PI * size * size,
753 Rectangle(Point { x, y }, Point { x: x2, y: y2 }) => (x2 - x) * (y2 - y)
758 Use a lone `_` to ignore an individual field. Ignore all fields of a variant
759 like: `Circle(..)`. Nullary enum patterns are written without parentheses:
762 # struct Point { x: f64, y: f64 }
763 # enum Direction { North, East, South, West }
764 fn point_from_direction(dir: Direction) -> Point {
766 North => Point { x: 0.0, y: 1.0 },
767 East => Point { x: 1.0, y: 0.0 },
768 South => Point { x: 0.0, y: -1.0 },
769 West => Point { x: -1.0, y: 0.0 }
774 Enum variants may also be structs. For example:
777 # #![feature(struct_variant)]
779 # struct Point { x: f64, y: f64 }
780 # fn square(x: f64) -> f64 { x * x }
782 Circle { center: Point, radius: f64 },
783 Rectangle { top_left: Point, bottom_right: Point }
785 fn area(sh: Shape) -> f64 {
787 Circle { radius: radius, .. } => f64::consts::PI * square(radius),
788 Rectangle { top_left: top_left, bottom_right: bottom_right } => {
789 (bottom_right.x - top_left.x) * (top_left.y - bottom_right.y)
796 > *Note:* This feature of the compiler is currently gated behind the
797 > `#[feature(struct_variant)]` directive. More about these directives can be
798 > found in the manual.
802 Tuples in Rust behave exactly like structs, except that their fields do not
803 have names. Thus, you cannot access their fields with dot notation. Tuples
804 can have any arity (number of elements) except for 0 (though you may consider
805 unit, `()`, as the empty tuple if you like).
808 let mytup: (int, int, f64) = (10, 20, 30.0);
810 (a, b, c) => println!("{}", a + b + (c as int))
816 Rust also has _tuple structs_, which behave like both structs and tuples,
817 except that, unlike tuples, tuple structs have names (so `Foo(1, 2)` has a
818 different type from `Bar(1, 2)`), and tuple structs' _fields_ do not have
824 struct MyTup(int, int, f64);
825 let mytup: MyTup = MyTup(10, 20, 30.0);
827 MyTup(a, b, c) => println!("{}", a + b + (c as int))
831 <a name="newtype"></a>
833 There is a special case for tuple structs with a single field, which are
834 sometimes called "newtypes" (after Haskell's "newtype" feature). These are
835 used to define new types in such a way that the new name is not just a
836 synonym for an existing type but is rather its own distinct type.
842 Types like this can be useful to differentiate between data that have
843 the same underlying type but must be used in different ways.
847 struct Centimeters(int);
850 The above definitions allow for a simple way for programs to avoid
851 confusing numbers that correspond to different units. Their integer
852 values can be extracted with pattern matching:
855 # struct Inches(int);
856 let length_with_unit = Inches(10);
857 let Inches(integer_length) = length_with_unit;
858 println!("length is {} inches", integer_length);
863 We've already seen several function definitions. Like all other static
864 declarations, such as `type`, functions can be declared both at the
865 top level and inside other functions (or in modules, which we'll come
866 back to [later](#crates-and-the-module-system)). The `fn` keyword introduces a
867 function. A function has an argument list, which is a parenthesized
868 list of `name: type` pairs separated by commas. An arrow `->`
869 separates the argument list and the function's return type.
872 fn line(a: int, b: int, x: int) -> int {
877 The `return` keyword immediately returns from the body of a function. It
878 is optionally followed by an expression to return. A function can
879 also return a value by having its top-level block produce an
883 fn line(a: int, b: int, x: int) -> int {
888 It's better Rust style to write a return value this way instead of
889 writing an explicit `return`. The utility of `return` comes in when
890 returning early from a function. Functions that do not return a value
891 are said to return unit, `()`, and both the return type and the return
892 value may be omitted from the definition. The following two functions
896 fn do_nothing_the_hard_way() -> () { return (); }
898 fn do_nothing_the_easy_way() { }
901 Ending the function with a semicolon like so is equivalent to returning `()`.
904 fn line(a: int, b: int, x: int) -> int { a * x + b }
905 fn oops(a: int, b: int, x: int) -> () { a * x + b; }
907 assert!(8 == line(5, 3, 1));
908 assert!(() == oops(5, 3, 1));
911 As with `match` expressions and `let` bindings, function arguments support
912 pattern destructuring. Like `let`, argument patterns must be irrefutable,
913 as in this example that unpacks the first value from a tuple and returns it.
916 fn first((value, _): (int, f64)) -> int { value }
921 A *destructor* is a function responsible for cleaning up the resources used by
922 an object when it is no longer accessible. Destructors can be defined to handle
923 the release of resources like files, sockets and heap memory.
925 Objects are never accessible after their destructor has been called, so no
926 dynamic failures are possible from accessing freed resources. When a task
927 fails, destructors of all objects in the task are called.
929 The `box` operator performs memory allocation on the heap:
933 // an integer allocated on the heap
936 // the destructor frees the heap memory as soon as `y` goes out of scope
939 Rust includes syntax for heap memory allocation in the language since it's
940 commonly used, but the same semantics can be implemented by a type with a
945 Rust formalizes the concept of object ownership to delegate management of an
946 object's lifetime to either a variable or a task-local garbage collector. An
947 object's owner is responsible for managing the lifetime of the object by
948 calling the destructor, and the owner determines whether the object is mutable.
950 Ownership is recursive, so mutability is inherited recursively and a destructor
951 destroys the contained tree of owned objects. Variables are top-level owners
952 and destroy the contained object when they go out of scope.
955 // the struct owns the objects contained in the `x` and `y` fields
956 struct Foo { x: int, y: Box<int> }
959 // `a` is the owner of the struct, and thus the owner of the struct's fields
960 let a = Foo { x: 5, y: box 10 };
962 // when `a` goes out of scope, the destructor for the `Box<int>` in the struct's
965 // `b` is mutable, and the mutability is inherited by the objects it owns
966 let mut b = Foo { x: 5, y: box 10 };
970 If an object doesn't contain any non-Send types, it consists of a single
971 ownership tree and is itself given the `Send` trait which allows it to be sent
972 between tasks. Custom destructors can only be implemented directly on types
973 that are `Send`, but non-`Send` types can still *contain* types with custom
974 destructors. Example of types which are not `Send` are [`Gc<T>`][gc] and
975 [`Rc<T>`][rc], the shared-ownership types.
977 [gc]: http://doc.rust-lang.org/std/gc/struct.Gc.html
978 [rc]: http://doc.rust-lang.org/std/rc/struct.Rc.html
980 # Implementing a linked list
982 An `enum` is a natural fit for describing a linked list, because it can express
983 a `List` type as being *either* the end of the list (`Nil`) or another node
984 (`Cons`). The full definition of the `Cons` variant will require some thought.
988 Cons(...), // an incomplete definition of the next element in a List
989 Nil // the end of a List
993 The obvious approach is to define `Cons` as containing an element in the list
994 along with the next `List` node. However, this will generate a compiler error.
997 // error: illegal recursive enum type; wrap the inner value in a box to make it
1000 Cons(u32, List), // an element (`u32`) and the next node in the list
1005 This error message is related to Rust's precise control over memory layout, and
1006 solving it will require introducing the concept of *boxing*.
1010 A value in Rust is stored directly inside the owner. If a `struct` contains
1011 four `u32` fields, it will be four times as large as a single `u32`.
1014 use std::mem::size_of; // bring `size_of` into the current scope, for convenience
1023 assert_eq!(size_of::<Foo>(), size_of::<u32>() * 4);
1032 assert_eq!(size_of::<Bar>(), size_of::<u32>() * 16);
1035 Our previous attempt at defining the `List` type included an `u32` and a `List`
1036 directly inside `Cons`, making it at least as big as the sum of both types. The
1037 type was invalid because the size was infinite!
1039 An *owned box* (`Box`, located in the `std::owned` module) uses a dynamic memory
1040 allocation to provide the invariant of always being the size of a pointer,
1041 regardless of the contained type. This can be leveraged to create a valid `List`
1047 Cons(u32, Box<List>),
1052 Defining a recursive data structure like this is the canonical example of an
1053 owned box. Much like an unboxed value, an owned box has a single owner and is
1054 therefore limited to expressing a tree-like data structure.
1056 Consider an instance of our `List` type:
1060 # Cons(u32, Box<List>),
1063 let list = Cons(1, box Cons(2, box Cons(3, box Nil)));
1066 It represents an owned tree of values, inheriting mutability down the tree and
1067 being destroyed along with the owner. Since the `list` variable above is
1068 immutable, the whole list is immutable. The memory allocation itself is the
1069 box, while the owner holds onto a pointer to it:
1072 List box List box List box List box
1073 +--------------+ +--------------+ +--------------+ +----------+
1074 list -> | Cons | 1 | | -> | Cons | 2 | | -> | Cons | 3 | | -> | Nil |
1075 +--------------+ +--------------+ +--------------+ +----------+
1078 > *Note:* the above diagram shows the logical contents of the enum. The actual
1079 > memory layout of the enum may vary. For example, for the `List` enum shown
1080 > above, Rust guarantees that there will be no enum tag field in the actual
1081 > structure. See the language reference for more details.
1083 An owned box is a common example of a type with a destructor. The allocated
1084 memory is cleaned up when the box is destroyed.
1088 Rust uses a shallow copy for parameter passing, assignment and returning from
1089 functions. Passing around the `List` will copy only as deep as the pointer to
1090 the box rather than doing an implicit heap allocation.
1094 # Cons(u32, Box<List>),
1097 let xs = Cons(1, box Cons(2, box Cons(3, box Nil)));
1098 let ys = xs; // copies `Cons(u32, pointer)` shallowly
1101 Rust will consider a shallow copy of a type with a destructor like `List` to
1102 *move ownership* of the value. After a value has been moved, the source
1103 location cannot be used unless it is reinitialized.
1107 # Cons(u32, Box<List>),
1113 // attempting to use `xs` will result in an error here
1117 // `xs` can be used again
1120 A destructor call will only occur for a variable that has not been moved from,
1121 as it is only called a single time.
1124 Avoiding a move can be done with the library-defined `clone` method:
1128 let y = x.clone(); // `y` is a newly allocated box
1129 let z = x; // no new memory allocated, `x` can no longer be used
1132 The `clone` method is provided by the `Clone` trait, and can be derived for
1133 our `List` type. Traits will be explained in detail [later](#traits).
1138 Cons(u32, Box<List>),
1142 let x = Cons(5, box Nil);
1145 // `x` can still be used!
1149 // and now, it can no longer be used since it has been moved
1152 The mutability of a value may be changed by moving it to a new owner:
1156 let mut s = r; // box becomes mutable
1158 let t = s; // box becomes immutable
1161 A simple way to define a function prepending to the `List` type is to take
1166 Cons(u32, Box<List>),
1170 fn prepend(xs: List, value: u32) -> List {
1175 xs = prepend(xs, 1);
1176 xs = prepend(xs, 2);
1177 xs = prepend(xs, 3);
1180 However, this is not a very flexible definition of `prepend` as it requires
1181 ownership of a list to be passed in rather than just mutating it in-place.
1185 The obvious signature for a `List` equality comparison is the following:
1188 fn eq(xs: List, ys: List) -> bool { /* ... */ }
1191 However, this will cause both lists to be moved into the function. Ownership
1192 isn't required to compare the lists, so the function should take *references*
1196 fn eq(xs: &List, ys: &List) -> bool { /* ... */ }
1199 A reference is a *non-owning* view of a value. A reference can be obtained with the `&` (address-of)
1200 operator. It can be dereferenced by using the `*` operator. In a pattern, such as `match` expression
1201 branches, the `ref` keyword can be used to bind to a variable name by-reference rather than
1202 by-value. A recursive definition of equality using references is as follows:
1206 # Cons(u32, Box<List>),
1209 fn eq(xs: &List, ys: &List) -> bool {
1210 // Match on the next node in both lists.
1212 // If we have reached the end of both lists, they are equal.
1213 (&Nil, &Nil) => true,
1214 // If the current elements of both lists are equal, keep going.
1215 (&Cons(x, box ref next_xs), &Cons(y, box ref next_ys))
1216 if x == y => eq(next_xs, next_ys),
1217 // If the current elements are not equal, the lists are not equal.
1222 let xs = Cons(5, box Cons(10, box Nil));
1223 let ys = Cons(5, box Cons(10, box Nil));
1224 assert!(eq(&xs, &ys));
1227 > *Note:* Rust doesn't guarantee [tail-call](http://en.wikipedia.org/wiki/Tail_call) optimization,
1228 > but LLVM is able to handle a simple case like this with optimizations enabled.
1230 ## Lists of other types
1232 Our `List` type is currently always a list of 32-bit unsigned integers. By
1233 leveraging Rust's support for generics, it can be extended to work for any
1236 The `u32` in the previous definition can be substituted with a type parameter:
1238 > *Note:* The following code introduces generics, which are explained in a
1239 > [dedicated section](#generics).
1243 Cons(T, Box<List<T>>),
1248 The old `List` of `u32` is now available as `List<u32>`. The `prepend`
1249 definition has to be updated too:
1253 # Cons(T, Box<List<T>>),
1256 fn prepend<T>(xs: List<T>, value: T) -> List<T> {
1261 Generic functions and types like this are equivalent to defining specialized
1262 versions for each set of type parameters.
1264 Using the generic `List<T>` works much like before, thanks to type inference:
1268 # Cons(T, Box<List<T>>),
1271 # fn prepend<T>(xs: List<T>, value: T) -> List<T> {
1272 # Cons(value, box xs)
1274 let mut xs = Nil; // Unknown type! This is a `List<T>`, but `T` can be anything.
1275 xs = prepend(xs, 10); // Here the compiler infers `xs`'s type as `List<int>`.
1276 xs = prepend(xs, 15);
1277 xs = prepend(xs, 20);
1280 The code sample above demonstrates type inference making most type annotations optional. It is
1281 equivalent to the following type-annotated code:
1285 # Cons(T, Box<List<T>>),
1288 # fn prepend<T>(xs: List<T>, value: T) -> List<T> {
1289 # Cons(value, box xs)
1291 let mut xs: List<int> = Nil::<int>;
1292 xs = prepend::<int>(xs, 10);
1293 xs = prepend::<int>(xs, 15);
1294 xs = prepend::<int>(xs, 20);
1297 In declarations, the language uses `Type<T, U, V>` to describe a list of type
1298 parameters, but expressions use `identifier::<T, U, V>`, to disambiguate the
1301 ## Defining list equality with generics
1303 Generic functions are type-checked from the definition, so any necessary properties of the type must
1304 be specified up-front. Our previous definition of list equality relied on the element type having
1305 the `==` operator available, and took advantage of the lack of a destructor on `u32` to copy it
1306 without a move of ownership.
1308 We can add a *trait bound* on the `PartialEq` trait to require that the type implement the `==` operator.
1309 Two more `ref` annotations need to be added to avoid attempting to move out the element types:
1313 # Cons(T, Box<List<T>>),
1316 fn eq<T: PartialEq>(xs: &List<T>, ys: &List<T>) -> bool {
1317 // Match on the next node in both lists.
1319 // If we have reached the end of both lists, they are equal.
1320 (&Nil, &Nil) => true,
1321 // If the current elements of both lists are equal, keep going.
1322 (&Cons(ref x, box ref next_xs), &Cons(ref y, box ref next_ys))
1323 if x == y => eq(next_xs, next_ys),
1324 // If the current elements are not equal, the lists are not equal.
1329 let xs = Cons('c', box Cons('a', box Cons('t', box Nil)));
1330 let ys = Cons('c', box Cons('a', box Cons('t', box Nil)));
1331 assert!(eq(&xs, &ys));
1334 This would be a good opportunity to implement the `PartialEq` trait for our list type, making the `==` and
1335 `!=` operators available. We'll need to provide an `impl` for the `PartialEq` trait and a definition of the
1336 `eq` method. In a method, the `self` parameter refers to an instance of the type we're implementing
1341 # Cons(T, Box<List<T>>),
1344 impl<T: PartialEq> PartialEq for List<T> {
1345 fn eq(&self, ys: &List<T>) -> bool {
1346 // Match on the next node in both lists.
1348 // If we have reached the end of both lists, they are equal.
1349 (&Nil, &Nil) => true,
1350 // If the current elements of both lists are equal, keep going.
1351 (&Cons(ref x, box ref next_xs), &Cons(ref y, box ref next_ys))
1352 if x == y => next_xs == next_ys,
1353 // If the current elements are not equal, the lists are not equal.
1359 let xs = Cons(5i, box Cons(10i, box Nil));
1360 let ys = Cons(5i, box Cons(10i, box Nil));
1361 // The methods below are part of the PartialEq trait,
1362 // which we implemented on our linked list.
1363 assert!(xs.eq(&ys));
1364 assert!(!xs.ne(&ys));
1366 // The PartialEq trait also allows us to use the shorthand infix operators.
1367 assert!(xs == ys); // `xs == ys` is short for `xs.eq(&ys)`
1368 assert!(!(xs != ys)); // `xs != ys` is short for `xs.ne(&ys)`
1373 The most common use case for owned boxes is creating recursive data structures
1374 like a binary search tree. Rust's trait-based generics system (covered later in
1375 the tutorial) is usually used for static dispatch, but also provides dynamic
1376 dispatch via boxing. Values of different types may have different sizes, but a
1377 box is able to *erase* the difference via the layer of indirection they
1380 In uncommon cases, the indirection can provide a performance gain or memory
1381 reduction by making values smaller. However, unboxed values should almost
1382 always be preferred when they are usable.
1384 Note that returning large unboxed values via boxes is unnecessary. A large
1385 value is returned via a hidden output parameter, and the decision on where to
1386 place the return value should be left to the caller:
1389 fn foo() -> (u64, u64, u64, u64, u64, u64) {
1393 let x = box foo(); // allocates a box, and writes the integers directly to it
1396 Beyond the properties granted by the size, an owned box behaves as a regular
1397 value by inheriting the mutability and lifetime of the owner:
1400 let x = 5; // immutable
1401 let mut y = 5; // mutable
1404 let x = box 5; // immutable
1405 let mut y = box 5; // mutable
1406 *y += 2; // the `*` operator is needed to access the contained value
1412 owned boxes, where the holder of an owned box is the owner of the pointed-to
1413 memory, references never imply ownership - they are "borrowed".
1414 You can borrow a reference to
1415 any object, and the compiler verifies that it cannot outlive the lifetime of
1418 As an example, consider a simple struct type, `Point`:
1427 We can use this simple definition to allocate points in many different
1428 ways. For example, in this code, each of these local variables
1429 contains a point, but allocated in a different location:
1432 # struct Point { x: f64, y: f64 }
1433 let on_the_stack : Point = Point { x: 3.0, y: 4.0 };
1434 let on_the_heap : Box<Point> = box Point { x: 7.0, y: 9.0 };
1437 Suppose we want to write a procedure that computes the distance
1438 between any two points, no matter where they are stored. One option is
1439 to define a function that takes two arguments of type point—that is,
1440 it takes the points by value. But this will cause the points to be
1441 copied when we call the function. For points, this is probably not so
1442 bad, but often copies are expensive. So we’d like to define a function
1443 that takes the points by pointer. We can use references to do this:
1446 # struct Point { x: f64, y: f64 }
1447 fn compute_distance(p1: &Point, p2: &Point) -> f64 {
1448 let x_d = p1.x - p2.x;
1449 let y_d = p1.y - p2.y;
1450 (x_d * x_d + y_d * y_d).sqrt()
1454 Now we can call `compute_distance()` in various ways:
1457 # struct Point{ x: f64, y: f64 };
1458 # let on_the_stack : Point = Point { x: 3.0, y: 4.0 };
1459 # let on_the_heap : Box<Point> = box Point { x: 7.0, y: 9.0 };
1460 # fn compute_distance(p1: &Point, p2: &Point) -> f64 { 0.0 }
1461 compute_distance(&on_the_stack, on_the_heap);
1464 Here the `&` operator is used to take the address of the variable
1465 `on_the_stack`; this is because `on_the_stack` has the type `Point`
1466 (that is, a struct value) and we have to take its address to get a
1467 reference. We also call this _borrowing_ the local variable
1468 `on_the_stack`, because we are creating an alias: that is, another
1469 route to the same data.
1471 In the case of `owned_box`, however, no
1472 explicit action is necessary. The compiler will automatically convert
1473 a box `box point` to a reference like
1474 `&point`. This is another form of borrowing; in this case, the
1475 contents of the owned box are being lent out.
1477 Whenever a value is borrowed, there are some limitations on what you
1478 can do with the original. For example, if the contents of a variable
1479 have been lent out, you cannot send that variable to another task, nor
1480 will you be permitted to take actions that might cause the borrowed
1481 value to be freed or to change its type. This rule should make
1482 intuitive sense: you must wait for a borrowed value to be returned
1483 (that is, for the reference to go out of scope) before you can
1484 make full use of it again.
1486 For a more in-depth explanation of references and lifetimes, read the
1487 [references and lifetimes guide][lifetimes].
1491 Lending an &-pointer to an object freezes the pointed-to object and prevents
1492 mutation—even if the object was declared as `mut`. `Freeze` objects have
1493 freezing enforced statically at compile-time. An example of a non-`Freeze` type
1494 is [`RefCell<T>`][refcell].
1499 let y = &x; // `x` is now frozen. It cannot be modified or re-assigned.
1501 // `x` is now unfrozen again
1505 [refcell]: http://doc.rust-lang.org/std/cell/struct.RefCell.html
1507 # Dereferencing pointers
1509 Rust uses the unary star operator (`*`) to access the contents of a
1510 box or pointer, similarly to C.
1516 let sum = *owned + *borrowed;
1519 Dereferenced mutable pointers may appear on the left hand side of
1520 assignments. Such an assignment modifies the value that the pointer
1524 let mut owned = box 10;
1527 let borrowed = &mut value;
1529 *owned = *borrowed + 100;
1530 *borrowed = *owned + 1000;
1533 Pointers have high operator precedence, but lower precedence than the
1534 dot operator used for field and method access. This precedence order
1535 can sometimes make code awkward and parenthesis-filled.
1538 # struct Point { x: f64, y: f64 }
1539 # enum Shape { Rectangle(Point, Point) }
1540 # impl Shape { fn area(&self) -> int { 0 } }
1541 let start = box Point { x: 10.0, y: 20.0 };
1542 let end = box Point { x: (*start).x + 100.0, y: (*start).y + 100.0 };
1543 let rect = &Rectangle(*start, *end);
1544 let area = (*rect).area();
1547 To combat this ugliness the dot operator applies _automatic pointer
1548 dereferencing_ to the receiver (the value on the left-hand side of the
1549 dot), so in most cases, explicitly dereferencing the receiver is not necessary.
1552 # struct Point { x: f64, y: f64 }
1553 # enum Shape { Rectangle(Point, Point) }
1554 # impl Shape { fn area(&self) -> int { 0 } }
1555 let start = box Point { x: 10.0, y: 20.0 };
1556 let end = box Point { x: start.x + 100.0, y: start.y + 100.0 };
1557 let rect = &Rectangle(*start, *end);
1558 let area = rect.area();
1561 You can write an expression that dereferences any number of pointers
1562 automatically. For example, if you feel inclined, you could write
1563 something silly like
1566 # struct Point { x: f64, y: f64 }
1567 let point = &box Point { x: 10.0, y: 20.0 };
1568 println!("{:f}", point.x);
1571 The indexing operator (`[]`) also auto-dereferences.
1573 # Vectors and strings
1575 A vector is a contiguous block of memory containing zero or more values of the
1576 same type. Rust also supports vector reference types, called slices, which are
1577 a view into a block of memory represented as a pointer and a length.
1579 Strings are represented as vectors of `u8`, with the guarantee of containing a
1580 valid UTF-8 sequence.
1582 Fixed-size vectors are an unboxed block of memory, with the element length as
1583 part of the type. A fixed-size vector owns the elements it contains, so the
1584 elements are mutable if the vector is mutable. Fixed-size strings do not exist.
1587 // A fixed-size vector
1588 let numbers = [1i, 2, 3];
1589 let more_numbers = numbers;
1591 // The type of a fixed-size vector is written as `[Type, ..length]`
1592 let five_zeroes: [int, ..5] = [0, ..5];
1595 A unique vector is dynamically sized, and has a destructor to clean up
1596 allocated memory on the heap. A unique vector owns the elements it contains, so
1597 the elements are mutable if the vector is mutable.
1600 use std::string::String;
1602 // A dynamically sized vector (unique vector)
1603 let mut numbers = vec![1i, 2, 3];
1607 // The type of a unique vector is written as `Vec<int>`
1608 let more_numbers: Vec<int> = numbers.move_iter().map(|i| i+1).collect();
1610 // The original `numbers` value can no longer be used, due to move semantics.
1612 let mut string = String::from_str("fo");
1613 string.push_char('o');
1616 Slices are similar to fixed-size vectors, but the length is not part of the
1617 type. They simply point into a block of memory and do not have ownership over
1622 let xs = &[1, 2, 3];
1624 // Slices have their type written as `&[int]`
1625 let ys: &[int] = xs;
1627 // Other vector types coerce to slices
1628 let three = [1, 2, 3];
1629 let zs: &[int] = three;
1631 // An unadorned string literal is an immutable string slice
1632 let string = "foobar";
1634 // A string slice type is written as `&str`
1635 let view: &str = string.slice(0, 3);
1638 Mutable slices also exist, just as there are mutable references. However, there
1639 are no mutable string slices. Strings are a multi-byte encoding (UTF-8) of
1640 Unicode code points, so they cannot be freely mutated without the ability to
1644 let mut xs = [1, 2, 3];
1645 let view = xs.mut_slice(0, 2);
1648 // The type of a mutable slice is written as `&mut [T]`
1649 let ys: &mut [int] = &mut [1, 2, 3];
1652 Square brackets denote indexing into a slice or fixed-size vector:
1655 # enum Crayon { Almond, AntiqueBrass, Apricot,
1656 # Aquamarine, Asparagus, AtomicTangerine,
1657 # BananaMania, Beaver, Bittersweet };
1658 # fn draw_scene(c: Crayon) { }
1659 let crayons: [Crayon, ..3] = [BananaMania, Beaver, Bittersweet];
1661 Bittersweet => draw_scene(crayons[0]),
1666 A slice or fixed-size vector can be destructured using pattern matching:
1669 let numbers: &[int] = &[1, 2, 3];
1670 let score = match numbers {
1673 [a, b] => a * 6 + b * 4,
1674 [a, b, c, ..rest] => a * 5 + b * 3 + c * 2 + rest.len() as int
1678 Both vectors and strings support a number of useful [methods](#methods),
1679 defined in [`std::vec`], [`std::slice`], and [`std::str`].
1681 [`std::vec`]: std/vec/index.html
1682 [`std::slice`]: std/slice/index.html
1683 [`std::str`]: std/str/index.html
1685 # Ownership escape hatches
1687 Ownership can cleanly describe tree-like data structures, and references provide non-owning pointers. However, more flexibility is often desired and Rust provides ways to escape from strict
1688 single parent ownership.
1690 The standard library provides the `std::rc::Rc` pointer type to express *shared ownership* over a
1691 reference counted box. As soon as all of the `Rc` pointers go out of scope, the box and the
1692 contained value are destroyed.
1697 // A fixed-size array allocated in a reference-counted box
1698 let x = Rc::new([1i, 2, 3, 4, 5, 6, 7, 8, 9, 10]);
1699 let y = x.clone(); // a new owner
1700 let z = x; // this moves `x` into `z`, rather than creating a new owner
1702 assert!(*z == [1i, 2, 3, 4, 5, 6, 7, 8, 9, 10]);
1704 // the variable is mutable, but not the contents of the box
1705 let mut a = Rc::new([10, 9, 8, 7, 6, 5, 4, 3, 2, 1]);
1709 A garbage collected pointer is provided via `std::gc::Gc`, with a task-local garbage collector
1710 having ownership of the box. It allows the creation of cycles, and the individual `Gc` pointers do
1711 not have a destructor.
1716 // A fixed-size array allocated in a garbage-collected box
1717 let x = box(GC) [1i, 2, 3, 4, 5, 6, 7, 8, 9, 10];
1718 let y = x; // does not perform a move, unlike with `Rc`
1721 assert!(*z == [1i, 2, 3, 4, 5, 6, 7, 8, 9, 10]);
1724 With shared ownership, mutability cannot be inherited so the boxes are always immutable. However,
1725 it's possible to use *dynamic* mutability via types like `std::cell::Cell` where freezing is handled
1726 via dynamic checks and can fail at runtime.
1728 The `Rc` and `Gc` types are not sendable, so they cannot be used to share memory between tasks. Safe
1729 immutable and mutable shared memory is provided by the `sync::arc` module.
1733 Named functions, like those we've seen so far, may not refer to local
1734 variables declared outside the function: they do not close over their
1735 environment (sometimes referred to as "capturing" variables in their
1736 environment). For example, you couldn't write the following:
1741 // `fun` cannot refer to `x`
1742 fn fun() -> () { println!("{}", x); }
1745 A _closure_ does support accessing the enclosing scope; below we will create
1746 2 _closures_ (nameless functions). Compare how `||` replaces `()` and how
1747 they try to access `x`:
1752 // `fun` is an invalid definition
1753 fn fun () -> () { println!("{}", x) } // cannot capture from enclosing scope
1754 let closure = || -> () { println!("{}", x) }; // can capture from enclosing scope
1756 // `fun_arg` is an invalid definition
1757 fn fun_arg (arg: int) -> () { println!("{}", arg + x) } // cannot capture
1758 let closure_arg = |arg: int| -> () { println!("{}", arg + x) }; // can capture
1760 // Requires a type because the implementation needs to know which `+` to use.
1761 // In the future, the implementation may not need the help.
1763 fun(); // Still won't work
1764 closure(); // Prints: 3
1766 fun_arg(7); // Still won't work
1767 closure_arg(7); // Prints: 10
1770 Closures begin with the argument list between vertical bars and are followed by
1771 a single expression. Remember that a block, `{ <expr1>; <expr2>; ... }`, is
1772 considered a single expression: it evaluates to the result of the last
1773 expression it contains if that expression is not followed by a semicolon,
1774 otherwise the block evaluates to `()`, the unit value.
1776 In general, return types and all argument types must be specified
1777 explicitly for function definitions. (As previously mentioned in the
1778 [Functions section](#functions), omitting the return type from a
1779 function declaration is synonymous with an explicit declaration of
1780 return type unit, `()`.)
1783 fn fun (x: int) { println!("{}", x) } // this is same as saying `-> ()`
1784 fn square(x: int) -> uint { (x * x) as uint } // other return types are explicit
1786 // Error: mismatched types: expected `()` but found `uint`
1787 fn badfun(x: int) { (x * x) as uint }
1790 On the other hand, the compiler can usually infer both the argument
1791 and return types for a closure expression; therefore they are often
1792 omitted, since both a human reader and the compiler can deduce the
1793 types from the immediate context. This is in contrast to function
1794 declarations, which require types to be specified and are not subject
1795 to type inference. Compare:
1798 // `fun` as a function declaration cannot infer the type of `x`, so it must be provided
1799 fn fun (x: int) { println!("{}", x) }
1800 let closure = |x | { println!("{}", x) }; // infers `x: int`, return type `()`
1802 // For closures, omitting a return type is *not* synonymous with `-> ()`
1803 let add_3 = |y | { 3i + y }; // infers `y: int`, return type `int`.
1805 fun(10); // Prints 10
1806 closure(20); // Prints 20
1807 closure(add_3(30)); // Prints 33
1809 fun("String"); // Error: mismatched types
1811 // Error: mismatched types
1812 // inference already assigned `closure` the type `|int| -> ()`
1816 In cases where the compiler needs assistance, the arguments and return
1817 types may be annotated on closures, using the same notation as shown
1818 earlier. In the example below, since different types provide an
1819 implementation for the operator `*`, the argument type for the `x`
1820 parameter must be explicitly provided.
1823 // Error: the type of `x` must be known to be used with `x * x`
1824 let square = |x | -> uint { (x * x) as uint };
1827 In the corrected version, the argument type is explicitly annotated,
1828 while the return type can still be inferred.
1831 let square_explicit = |x: int| -> uint { (x * x) as uint };
1832 let square_infer = |x: int| { (x * x) as uint };
1834 println!("{}", square_explicit(20)); // 400
1835 println!("{}", square_infer(-20)); // 400
1838 There are several forms of closure, each with its own role. The most
1839 common, called a _stack closure_, has type `||` and can directly
1840 access local variables in the enclosing scope.
1844 let f = |x: int| if x > max { max = x };
1845 for x in [1, 2, 3].iter() {
1850 Stack closures are very efficient because their environment is
1851 allocated on the call stack and refers by pointer to captured
1852 locals. To ensure that stack closures never outlive the local
1853 variables to which they refer, stack closures are not
1854 first-class. That is, they can only be used in argument position; they
1855 cannot be stored in data structures or returned from
1856 functions. Despite these limitations, stack closures are used
1857 pervasively in Rust code.
1861 Owned closures, written `proc`,
1862 hold on to things that can safely be sent between
1863 processes. They copy the values they close over,
1864 but they also own them: that is, no other code can access
1865 them. Owned closures are used in concurrent code, particularly
1866 for spawning [tasks][tasks].
1868 Closures can be used to spawn tasks.
1869 A practical example of this pattern is found when using the `spawn` function,
1870 which starts a new task.
1873 use std::task::spawn;
1875 // proc is the closure which will be spawned.
1877 println!("I'm a new task")
1881 ## Closure compatibility
1883 Rust closures have a convenient subtyping property: you can pass any kind of
1884 closure (as long as the arguments and return types match) to functions
1885 that expect a `||`. Thus, when writing a higher-order function that
1886 only calls its function argument, and does nothing else with it, you
1887 should almost always declare the type of that argument as `||`. That way,
1888 callers may pass any kind of closure.
1891 fn call_twice(f: ||) { f(); f(); }
1892 let closure = || { "I'm a closure, and it doesn't matter what type I am"; };
1893 fn function() { "I'm a normal function"; }
1894 call_twice(closure);
1895 call_twice(function);
1898 > *Note:* Both the syntax and the semantics will be changing
1899 > in small ways. At the moment they can be unsound in some
1900 > scenarios, particularly with non-copyable types.
1904 Methods are like functions except that they always begin with a special argument,
1906 which has the type of the method's receiver. The
1907 `self` argument is like `this` in C++ and many other languages.
1908 Methods are called with dot notation, as in `my_vec.len()`.
1910 _Implementations_, written with the `impl` keyword, can define
1911 methods on most Rust types, including structs and enums.
1912 As an example, let's define a `draw` method on our `Shape` enum.
1915 # fn draw_circle(p: Point, f: f64) { }
1916 # fn draw_rectangle(p: Point, p: Point) { }
1924 Rectangle(Point, Point)
1930 Circle(p, f) => draw_circle(p, f),
1931 Rectangle(p1, p2) => draw_rectangle(p1, p2)
1936 let s = Circle(Point { x: 1.0, y: 2.0 }, 3.0);
1940 This defines an _implementation_ for `Shape` containing a single
1941 method, `draw`. In most respects the `draw` method is defined
1942 like any other function, except for the name `self`.
1944 The type of `self` is the type on which the method is implemented,
1945 or a pointer thereof. As an argument it is written either `self`,
1946 `&self`, or `~self`.
1947 A caller must in turn have a compatible pointer type to call the method.
1950 # fn draw_circle(p: Point, f: f64) { }
1951 # fn draw_rectangle(p: Point, p: Point) { }
1952 # struct Point { x: f64, y: f64 }
1954 # Circle(Point, f64),
1955 # Rectangle(Point, Point)
1958 fn draw_reference(&self) { /* ... */ }
1959 fn draw_owned(~self) { /* ... */ }
1960 fn draw_value(self) { /* ... */ }
1963 let s = Circle(Point { x: 1.0, y: 2.0 }, 3.0);
1965 (&s).draw_reference();
1966 (box s).draw_owned();
1970 Methods typically take a reference self type,
1971 so the compiler will go to great lengths to convert a callee
1975 # fn draw_circle(p: Point, f: f64) { }
1976 # fn draw_rectangle(p: Point, p: Point) { }
1977 # struct Point { x: f64, y: f64 }
1979 # Circle(Point, f64),
1980 # Rectangle(Point, Point)
1983 # fn draw_reference(&self) { /* ... */ }
1984 # fn draw_owned(~self) { /* ... */ }
1985 # fn draw_value(self) { /* ... */ }
1987 # let s = Circle(Point { x: 1.0, y: 2.0 }, 3.0);
1988 // As with typical function arguments, owned pointers
1989 // are automatically converted to references
1991 (box s).draw_reference();
1993 // Unlike typical function arguments, the self value will
1994 // automatically be referenced ...
1997 // ... and dereferenced
1998 (& &s).draw_reference();
2000 // ... and dereferenced and borrowed
2001 (&box s).draw_reference();
2004 Implementations may also define standalone (sometimes called "static")
2005 methods. The absence of a `self` parameter distinguishes such methods.
2006 These methods are the preferred way to define constructor functions.
2010 fn area(&self) -> f64 { /* ... */ }
2011 fn new(area: f64) -> Circle { /* ... */ }
2015 To call such a method, just prefix it with the type name and a double colon:
2018 use std::f64::consts::PI;
2019 struct Circle { radius: f64 }
2021 fn new(area: f64) -> Circle { Circle { radius: (area / PI).sqrt() } }
2023 let c = Circle::new(42.5);
2028 Throughout this tutorial, we've been defining functions that act only
2029 on specific data types. With type parameters we can also define
2030 functions whose arguments have generic types, and which can be invoked
2031 with a variety of types. Consider a generic `map` function, which
2032 takes a function `function` and a vector `vector` and returns a new
2033 vector consisting of the result of applying `function` to each element
2037 fn map<T, U>(vector: &[T], function: |v: &T| -> U) -> Vec<U> {
2038 let mut accumulator = Vec::new();
2039 for element in vector.iter() {
2040 accumulator.push(function(element));
2046 When defined with type parameters, as denoted by `<T, U>`, this
2047 function can be applied to any type of vector, as long as the type of
2048 `function`'s argument and the type of the vector's contents agree with
2051 Inside a generic function, the names of the type parameters
2052 (capitalized by convention) stand for opaque types. All you can do
2053 with instances of these types is pass them around: you can't apply any
2054 operations to them or pattern-match on them. Note that instances of
2055 generic types are often passed by pointer. For example, the parameter
2056 `function()` is supplied with a pointer to a value of type `T` and not
2057 a value of type `T` itself. This ensures that the function works with
2058 the broadest set of types possible, since some types are expensive or
2059 illegal to copy and pass by value.
2061 Generic `type`, `struct`, and `enum` declarations follow the same pattern:
2064 type Set<T> = std::collections::HashMap<T, ()>;
2077 These declarations can be instantiated to valid types like `Set<int>`,
2078 `Stack<int>`, and `Option<int>`.
2080 The last type in that example, `Option`, appears frequently in Rust code.
2081 Because Rust does not have null pointers (except in unsafe code), we need
2082 another way to write a function whose result isn't defined on every possible
2083 combination of arguments of the appropriate types. The usual way is to write
2084 a function that returns `Option<T>` instead of `T`.
2087 # struct Point { x: f64, y: f64 }
2088 # enum Shape { Circle(Point, f64), Rectangle(Point, Point) }
2089 fn radius(shape: Shape) -> Option<f64> {
2091 Circle(_, radius) => Some(radius),
2092 Rectangle(..) => None
2097 The Rust compiler compiles generic functions very efficiently by
2098 *monomorphizing* them. *Monomorphization* is a fancy name for a simple
2099 idea: generate a separate copy of each generic function at each call site,
2100 a copy that is specialized to the argument
2101 types and can thus be optimized specifically for them. In this
2102 respect, Rust's generics have similar performance characteristics to
2107 Within a generic function—that is, a function parameterized by a
2108 type parameter, say, `T`—the operations we can do on arguments of
2109 type `T` are quite limited. After all, since we don't know what type
2110 `T` will be instantiated with, we can't safely modify or query values
2111 of type `T`. This is where _traits_ come into play. Traits are Rust's
2112 most powerful tool for writing polymorphic code. Java developers will
2113 see them as similar to Java interfaces, and Haskellers will notice
2114 their similarities to type classes. Rust's traits give us a way to
2115 express *bounded polymorphism*: by limiting the set of possible types
2116 that a type parameter could refer to, they expand the number of
2117 operations we can safely perform on arguments of that type.
2119 As motivation, let us consider copying of values in Rust. The `clone`
2120 method is not defined for values of every type. One reason is
2121 user-defined destructors: copying a value of a type that has a
2122 destructor could result in the destructor running multiple times.
2123 Therefore, values of types that have destructors cannot be copied
2124 unless we explicitly implement `clone` for them.
2126 This complicates handling of generic functions.
2127 If we have a function with a type parameter `T`,
2128 can we copy values of type `T` inside that function?
2130 and if we try to run the following code the compiler will complain.
2133 // This does not compile
2134 fn head_bad<T>(v: &[T]) -> T {
2135 v[0] // error: copying a non-copyable value
2139 However, we can tell the compiler
2140 that the `head` function is only for copyable types.
2141 In Rust, copyable types are those that _implement the `Clone` trait_.
2142 We can then explicitly create a second copy of the value we are returning
2143 by calling the `clone` method:
2147 fn head<T: Clone>(v: &[T]) -> T {
2152 The bounded type parameter `T: Clone` says that `head`
2153 can be called on an argument of type `&[T]` for any `T`,
2154 so long as there is an implementation of the
2155 `Clone` trait for `T`.
2156 When instantiating a generic function,
2157 we can only instantiate it with types
2158 that implement the correct trait,
2159 so we could not apply `head` to a vector whose elements are of some type
2160 that does not implement `Clone`.
2162 While most traits can be defined and implemented by user code,
2163 three traits are automatically derived and implemented
2164 for all applicable types by the compiler,
2165 and may not be overridden:
2167 * `Send` - Sendable types.
2169 unless they contain references.
2171 * `Share` - Types that are *threadsafe*
2172 These are types that are safe to be used across several threads with access to
2173 a `&T` pointer. `Mutex<T>` is an example of a *sharable* type with internal mutable data.
2175 * `'static` - Non-borrowed types.
2176 These are types that do not contain any data whose lifetime is bound to
2177 a particular stack frame. These are types that do not contain any
2178 references, or types where the only contained references
2179 have the `'static` lifetime. (For more on named lifetimes and their uses,
2180 see the [references and lifetimes guide][lifetimes].)
2182 > *Note:* These built-in traits were referred to as 'kinds' in earlier
2183 > iterations of the language, and often still are.
2185 Additionally, the `Drop` trait is used to define destructors. This
2186 trait provides one method called `drop`, which is automatically
2187 called when a value of the type that implements this trait is
2188 destroyed, either because the value went out of scope or because the
2189 garbage collector reclaimed it.
2196 impl Drop for TimeBomb {
2197 fn drop(&mut self) {
2198 for _ in range(0, self.explosivity) {
2205 It is illegal to call `drop` directly. Only code inserted by the compiler
2208 ## Declaring and implementing traits
2210 At its simplest, a trait is a set of zero or more _method signatures_.
2211 For example, we could declare the trait
2212 `Printable` for things that can be printed to the console,
2213 with a single method signature:
2221 We say that the `Printable` trait _provides_ a `print` method with the
2222 given signature. This means that we can call `print` on an argument
2223 of any type that implements the `Printable` trait.
2225 Rust's built-in `Send` and `Share` types are examples of traits that
2226 don't provide any methods.
2228 Traits may be implemented for specific types with [impls]. An impl for
2229 a particular trait gives an implementation of the methods that
2230 trait provides. For instance, the following impls of
2231 `Printable` for `int` and `String` give implementations of the `print`
2237 # trait Printable { fn print(&self); }
2238 impl Printable for int {
2239 fn print(&self) { println!("{}", *self) }
2242 impl Printable for String {
2243 fn print(&self) { println!("{}", *self) }
2247 # ("foo".to_string()).print();
2250 Methods defined in an impl for a trait may be called just like
2251 any other method, using dot notation, as in `1.print()`.
2253 ## Default method implementations in trait definitions
2255 Sometimes, a method that a trait provides will have the same
2256 implementation for most or all of the types that implement that trait.
2257 For instance, suppose that we wanted `bool`s and `f32`s to be
2258 printable, and that we wanted the implementation of `print` for those
2259 types to be exactly as it is for `int`, above:
2262 # trait Printable { fn print(&self); }
2263 impl Printable for f32 {
2264 fn print(&self) { println!("{}", *self) }
2267 impl Printable for bool {
2268 fn print(&self) { println!("{}", *self) }
2275 This works fine, but we've now repeated the same definition of `print`
2276 in three places. Instead of doing that, we can simply include the
2277 definition of `print` right in the trait definition, instead of just
2278 giving its signature. That is, we can write the following:
2285 // Default method implementation
2286 fn print(&self) { println!("{:?}", *self) }
2289 impl Printable for int {}
2291 impl Printable for String {
2292 fn print(&self) { println!("{}", *self) }
2295 impl Printable for bool {}
2297 impl Printable for f32 {}
2300 # ("foo".to_string()).print();
2306 Here, the impls of `Printable` for `int`, `bool`, and `f32` don't
2307 need to provide an implementation of `print`, because in the absence
2308 of a specific implementation, Rust just uses the _default method_
2309 provided in the trait definition. Depending on the trait, default
2310 methods can save a great deal of boilerplate code from having to be
2311 written in impls. Of course, individual impls can still override the
2312 default method for `print`, as is being done above in the impl for
2315 ## Type-parameterized traits
2317 Traits may be parameterized by type variables. For example, a trait
2318 for generalized sequence types might look like the following:
2322 fn length(&self) -> uint;
2325 impl<T> Seq<T> for Vec<T> {
2326 fn length(&self) -> uint { self.len() }
2330 The implementation has to explicitly declare the type parameter that
2331 it binds, `T`, before using it to specify its trait type. Rust
2332 requires this declaration because the `impl` could also, for example,
2333 specify an implementation of `Seq<int>`. The trait type (appearing
2334 between `impl` and `for`) *refers* to a type, rather than
2337 The type parameters bound by a trait are in scope in each of the
2338 method declarations. So, re-declaring the type parameter
2339 `T` as an explicit type parameter for `length`, in either the trait or
2340 the impl, would be a compile-time error.
2342 Within a trait definition, `Self` is a special type that you can think
2343 of as a type parameter. An implementation of the trait for any given
2344 type `T` replaces the `Self` type parameter with `T`. The following
2345 trait describes types that support an equality operation:
2348 // In a trait, `self` refers to the self argument.
2349 // `Self` refers to the type implementing the trait.
2351 fn equals(&self, other: &Self) -> bool;
2354 // In an impl, `self` refers just to the value of the receiver
2355 impl PartialEq for int {
2356 fn equals(&self, other: &int) -> bool { *other == *self }
2360 Notice that in the trait definition, `equals` takes a
2361 second parameter of type `Self`.
2362 In contrast, in the `impl`, `equals` takes a second parameter of
2363 type `int`, only using `self` as the name of the receiver.
2365 Just as in type implementations, traits can define standalone (static)
2366 methods. These methods are called by prefixing the method name with the trait
2367 name and a double colon. The compiler uses type inference to decide which
2368 implementation to use.
2371 use std::f64::consts::PI;
2372 trait Shape { fn new(area: f64) -> Self; }
2373 struct Circle { radius: f64 }
2374 struct Square { length: f64 }
2376 impl Shape for Circle {
2377 fn new(area: f64) -> Circle { Circle { radius: (area / PI).sqrt() } }
2379 impl Shape for Square {
2380 fn new(area: f64) -> Square { Square { length: area.sqrt() } }
2384 let c: Circle = Shape::new(area);
2385 let s: Square = Shape::new(area);
2388 ## Bounded type parameters and static method dispatch
2390 Traits give us a language for defining predicates on types, or
2391 abstract properties that types can have. We can use this language to
2392 define _bounds_ on type parameters, so that we can then operate on
2396 # trait Printable { fn print(&self); }
2397 fn print_all<T: Printable>(printable_things: Vec<T>) {
2398 for thing in printable_things.iter() {
2404 Declaring `T` as conforming to the `Printable` trait (as we earlier
2405 did with `Clone`) makes it possible to call methods from that trait
2406 on values of type `T` inside the function. It will also cause a
2407 compile-time error when anyone tries to call `print_all` on a vector
2408 whose element type does not have a `Printable` implementation.
2410 Type parameters can have multiple bounds by separating them with `+`,
2411 as in this version of `print_all` that copies elements.
2414 # trait Printable { fn print(&self); }
2415 fn print_all<T: Printable + Clone>(printable_things: Vec<T>) {
2417 while i < printable_things.len() {
2418 let copy_of_thing = printable_things.get(i).clone();
2419 copy_of_thing.print();
2425 Method calls to bounded type parameters are _statically dispatched_,
2426 imposing no more overhead than normal function invocation, so are
2427 the preferred way to use traits polymorphically.
2429 This usage of traits is similar to Haskell type classes.
2431 ## Trait objects and dynamic method dispatch
2433 The above allows us to define functions that polymorphically act on
2434 values of a single unknown type that conforms to a given trait.
2435 However, consider this function:
2438 # type Circle = int; type Rectangle = int;
2439 # impl Drawable for int { fn draw(&self) {} }
2440 # fn new_circle() -> int { 1 }
2441 trait Drawable { fn draw(&self); }
2443 fn draw_all<T: Drawable>(shapes: Vec<T>) {
2444 for shape in shapes.iter() { shape.draw(); }
2446 # let c: Circle = new_circle();
2447 # draw_all(vec![c]);
2450 You can call that on a vector of circles, or a vector of rectangles
2451 (assuming those have suitable `Drawable` traits defined), but not on
2452 a vector containing both circles and rectangles. When such behavior is
2453 needed, a trait name can alternately be used as a type, called
2457 # trait Drawable { fn draw(&self); }
2458 fn draw_all(shapes: &[Box<Drawable>]) {
2459 for shape in shapes.iter() { shape.draw(); }
2463 In this example, there is no type parameter. Instead, the `Box<Drawable>`
2464 type denotes any owned box value that implements the `Drawable` trait.
2465 To construct such a value, you use the `as` operator to cast a value
2469 # type Circle = int; type Rectangle = bool;
2470 # trait Drawable { fn draw(&self); }
2471 # fn new_circle() -> Circle { 1 }
2472 # fn new_rectangle() -> Rectangle { true }
2473 # fn draw_all(shapes: &[Box<Drawable>]) {}
2475 impl Drawable for Circle { fn draw(&self) { /* ... */ } }
2476 impl Drawable for Rectangle { fn draw(&self) { /* ... */ } }
2478 let c: Box<Circle> = box new_circle();
2479 let r: Box<Rectangle> = box new_rectangle();
2480 draw_all([c as Box<Drawable>, r as Box<Drawable>]);
2483 We omit the code for `new_circle` and `new_rectangle`; imagine that
2484 these just return `Circle`s and `Rectangle`s with a default size. Note
2485 that, like strings and vectors, objects have dynamic size and may
2486 only be referred to via one of the pointer types.
2487 Other pointer types work as well.
2488 Casts to traits may only be done with compatible pointers so,
2489 for example, an `&Circle` may not be cast to a `Box<Drawable>`.
2492 # type Circle = int; type Rectangle = int;
2493 # trait Drawable { fn draw(&self); }
2494 # impl Drawable for int { fn draw(&self) {} }
2495 # fn new_circle() -> int { 1 }
2496 # fn new_rectangle() -> int { 2 }
2498 let owny: Box<Drawable> = box new_circle() as Box<Drawable>;
2499 // A borrowed object
2500 let stacky: &Drawable = &new_circle() as &Drawable;
2503 Method calls to trait types are _dynamically dispatched_. Since the
2504 compiler doesn't know specifically which functions to call at compile
2505 time, it uses a lookup table (also known as a vtable or dictionary) to
2506 select the method to call at runtime.
2508 This usage of traits is similar to Java interfaces.
2510 There are some built-in bounds, such as `Send` and `Share`, which are properties
2511 of the components of types. By design, trait objects don't know the exact type
2512 of their contents and so the compiler cannot reason about those properties.
2514 You can instruct the compiler, however, that the contents of a trait object must
2515 acribe to a particular bound with a trailing colon (`:`). These are examples of
2522 fn sendable_foo(f: Box<Foo + Send>) { /* ... */ }
2523 fn shareable_bar<T: Share>(b: &Bar<T> + Share) { /* ... */ }
2526 When no colon is specified (such as the type `Box<Foo>`), it is inferred that the
2527 value ascribes to no bounds. They must be added manually if any bounds are
2528 necessary for usage.
2530 Builtin kind bounds can also be specified on closure types in the same way (for
2531 example, by writing `fn:Send()`), and the default behaviours are the same as
2532 for traits of the same storage class.
2534 ## Trait inheritance
2536 We can write a trait declaration that _inherits_ from other traits, called _supertraits_.
2537 Types that implement a trait must also implement its supertraits.
2539 we can define a `Circle` trait that inherits from `Shape`.
2542 trait Shape { fn area(&self) -> f64; }
2543 trait Circle : Shape { fn radius(&self) -> f64; }
2546 Now, we can implement `Circle` on a type only if we also implement `Shape`.
2549 use std::f64::consts::PI;
2550 # trait Shape { fn area(&self) -> f64; }
2551 # trait Circle : Shape { fn radius(&self) -> f64; }
2552 # struct Point { x: f64, y: f64 }
2553 # fn square(x: f64) -> f64 { x * x }
2554 struct CircleStruct { center: Point, radius: f64 }
2555 impl Circle for CircleStruct {
2556 fn radius(&self) -> f64 { (self.area() / PI).sqrt() }
2558 impl Shape for CircleStruct {
2559 fn area(&self) -> f64 { PI * square(self.radius) }
2563 Notice that methods of `Circle` can call methods on `Shape`, as our
2564 `radius` implementation calls the `area` method.
2565 This is a silly way to compute the radius of a circle
2566 (since we could just return the `radius` field), but you get the idea.
2568 In type-parameterized functions,
2569 methods of the supertrait may be called on values of subtrait-bound type parameters.
2570 Refering to the previous example of `trait Circle : Shape`:
2573 # trait Shape { fn area(&self) -> f64; }
2574 # trait Circle : Shape { fn radius(&self) -> f64; }
2575 fn radius_times_area<T: Circle>(c: T) -> f64 {
2576 // `c` is both a Circle and a Shape
2577 c.radius() * c.area()
2581 Likewise, supertrait methods may also be called on trait objects.
2584 use std::f64::consts::PI;
2585 # trait Shape { fn area(&self) -> f64; }
2586 # trait Circle : Shape { fn radius(&self) -> f64; }
2587 # struct Point { x: f64, y: f64 }
2588 # struct CircleStruct { center: Point, radius: f64 }
2589 # impl Circle for CircleStruct { fn radius(&self) -> f64 { (self.area() / PI).sqrt() } }
2590 # impl Shape for CircleStruct { fn area(&self) -> f64 { PI * square(self.radius) } }
2591 # fn square(x: f64) -> f64 { x * x }
2593 let concrete = box CircleStruct{center:Point{x:3.0,y:4.0},radius:5.0};
2594 let mycircle: Box<Circle> = concrete as Box<Circle>;
2595 let nonsense = mycircle.radius() * mycircle.area();
2598 > *Note:* Trait inheritance does not actually work with objects yet
2600 ## Deriving implementations for traits
2602 A small number of traits in `std` and `extra` can have implementations
2603 that can be automatically derived. These instances are specified by
2604 placing the `deriving` attribute on a data type declaration. For
2605 example, the following will mean that `Circle` has an implementation
2606 for `PartialEq` and can be used with the equality operators, and that a value
2607 of type `ABC` can be randomly generated and converted to a string:
2612 #[deriving(PartialEq)]
2613 struct Circle { radius: f64 }
2615 #[deriving(Rand, Show)]
2616 enum ABC { A, B, C }
2619 // Use the Show trait to print "A, B, C."
2620 println!("{}, {}, {}", A, B, C);
2624 The full list of derivable traits is `PartialEq`, `Eq`, `PartialOrd`,
2625 `Ord`, `Encodable`, `Decodable`, `Clone`,
2626 `Hash`, `Rand`, `Default`, `Zero`, `FromPrimitive` and `Show`.
2628 # Crates and the module system
2630 Rust's module system is very powerful, but because of that also somewhat complex.
2631 Nevertheless, this section will try to explain every important aspect of it.
2635 In order to speak about the module system, we first need to define the medium it exists in:
2637 Let's say you've written a program or a library, compiled it, and got the resulting binary.
2638 In Rust, the content of all source code that the compiler directly had to compile in order to end up with
2639 that binary is collectively called a 'crate'.
2641 For example, for a simple hello world program your crate only consists of this code:
2646 println!("Hello world!");
2650 A crate is also the unit of independent compilation in Rust: `rustc` always compiles a single crate at a time,
2651 from which it produces either a library or an executable.
2653 Note that merely using an already compiled library in your code does not make it part of your crate.
2655 ## The module hierarchy
2657 For every crate, all the code in it is arranged in a hierarchy of modules starting with a single
2658 root module. That root module is called the 'crate root'.
2660 All modules in a crate below the crate root are declared with the `mod` keyword:
2663 // This is the crate root
2666 // This is the body of module 'farm' declared in the crate root.
2668 fn chicken() { println!("cluck cluck"); }
2669 fn cow() { println!("mooo"); }
2672 // Body of module 'barn'
2674 fn hay() { println!("..."); }
2679 println!("Hello farm!");
2683 As you can see, your module hierarchy is now three modules deep: There is the crate root, which contains your `main()`
2684 function, and the module `farm`. The module `farm` also contains two functions and a third module `barn`,
2685 which contains a function `hay`.
2687 ## Paths and visibility
2689 We've now defined a nice module hierarchy. But how do we access the items in it from our `main` function?
2690 One way to do it is to simply fully qualifying it:
2694 fn chicken() { println!("cluck cluck"); }
2699 println!("Hello chicken!");
2701 ::farm::chicken(); // Won't compile yet, see further down
2705 The `::farm::chicken` construct is what we call a 'path'.
2707 Because it's starting with a `::`, it's also a 'global path', which qualifies
2708 an item by its full path in the module hierarchy relative to the crate root.
2710 If the path were to start with a regular identifier, like `farm::chicken`, it
2711 would be a 'local path' instead. We'll get to them later.
2713 Now, if you actually tried to compile this code example, you'll notice that you
2714 get a `function 'chicken' is private` error. That's because by default, items
2715 (`fn`, `struct`, `static`, `mod`, ...) are private.
2717 To make them visible outside their containing modules, you need to mark them
2718 _public_ with `pub`:
2722 pub fn chicken() { println!("cluck cluck"); }
2723 pub fn cow() { println!("mooo"); }
2728 println!("Hello chicken!");
2729 ::farm::chicken(); // This compiles now
2733 Visibility restrictions in Rust exist only at module boundaries. This
2734 is quite different from most object-oriented languages that also
2735 enforce restrictions on objects themselves. That's not to say that
2736 Rust doesn't support encapsulation: both struct fields and methods can
2737 be private. But this encapsulation is at the module level, not the
2740 Fields are _private_ by default, and can be made _public_ with
2745 # pub type Chicken = int;
2746 # struct Human(int);
2747 # impl Human { pub fn rest(&self) { } }
2748 # pub fn make_me_a_farm() -> Farm { Farm { chickens: vec![], farmer: Human(0) } }
2750 chickens: Vec<Chicken>,
2755 fn feed_chickens(&self) { /* ... */ }
2756 pub fn add_chicken(&self, c: Chicken) { /* ... */ }
2759 pub fn feed_animals(farm: &Farm) {
2760 farm.feed_chickens();
2765 let f = make_me_a_farm();
2766 f.add_chicken(make_me_a_chicken());
2767 farm::feed_animals(&f);
2770 // This wouldn't compile because both are private:
2771 // `f.feed_chickens();`
2772 // `let chicken_counter = f.chickens.len();`
2774 # fn make_me_a_farm() -> farm::Farm { farm::make_me_a_farm() }
2775 # fn make_me_a_chicken() -> farm::Chicken { 0 }
2778 Exact details and specifications about visibility rules can be found in the Rust
2781 ## Files and modules
2783 One important aspect of Rust's module system is that source files and modules are not the same thing. You define a module hierarchy, populate it with all your definitions, define visibility, maybe put in a `fn main()`, and that's it.
2785 The only file that's relevant when compiling is the one that contains the body
2786 of your crate root, and it's only relevant because you have to pass that file
2787 to `rustc` to compile your crate.
2789 In principle, that's all you need: You can write any Rust program as one giant source file that contains your
2790 crate root and everything else in `mod ... { ... }` declarations.
2792 However, in practice you usually want to split up your code into multiple
2793 source files to make it more manageable. Rust allows you to move the body of
2794 any module into its own source file. If you declare a module without its body,
2795 like `mod foo;`, the compiler will look for the files `foo.rs` and `foo/mod.rs`
2796 inside some directory (usually the same as of the source file containing the
2797 `mod foo;` declaration). If it finds either, it uses the content of that file
2798 as the body of the module. If it finds both, that's a compile error.
2800 To move the content of `mod farm` into its own file, you can write:
2803 // `main.rs` - contains body of the crate root
2804 mod farm; // Compiler will look for `farm.rs` and `farm/mod.rs`
2807 println!("Hello farm!");
2813 // `farm.rs` - contains body of module 'farm' in the crate root
2814 pub fn chicken() { println!("cluck cluck"); }
2815 pub fn cow() { println!("mooo"); }
2818 pub fn hay() { println!("..."); }
2823 In short, `mod foo;` is just syntactic sugar for `mod foo { /* content of <...>/foo.rs or <...>/foo/mod.rs */ }`.
2825 This also means that having two or more identical `mod foo;` declarations
2826 somewhere in your crate hierarchy is generally a bad idea,
2827 just like copy-and-paste-ing a module into multiple places is a bad idea.
2828 Both will result in duplicate and mutually incompatible definitions.
2830 When `rustc` resolves these module declarations, it starts by looking in the
2831 parent directory of the file containing the `mod foo` declaration. For example,
2832 given a file with the module body:
2845 The compiler will look for these files, in this order:
2852 src/animals/fish/mod.rs
2854 src/animals/mammals/humans.rs
2855 src/animals/mammals/humans/mod.rs
2858 Keep in mind that identical module hierarchies can still lead to different path
2859 lookups depending on how and where you've moved a module body to its own file.
2860 For example, if you move the `animals` module into its own file:
2869 // `src/animals.rs` or `src/animals/mod.rs`
2876 ...then the source files of `mod animals`'s submodules can either be in the same directory as the animals source file or in a subdirectory of its directory. If the animals file is `src/animals.rs`, `rustc` will look for:
2883 src/mammals/humans.rs
2884 src/mammals/humans/mod.rs
2887 If the animals file is `src/animals/mod.rs`, `rustc` will look for:
2892 src/animals/fish/mod.rs
2894 src/animals/mammals/humans.rs
2895 src/animals/mammals/humans/mod.rs
2899 These rules allow you to write small modules consisting of single source files which can live in the same directory as well as large modules which group submodule source files in subdirectories.
2901 If you need to override where `rustc` will look for the file containing a
2902 module's source code, use the `path` compiler directive. For example, to load a
2903 `classified` module from a different file:
2906 #[path="../../area51/alien.rs"]
2910 ## Importing names into the local scope
2912 Always referring to definitions in other modules with their global
2913 path gets old really fast, so Rust has a way to import
2914 them into the local scope of your module: `use`-statements.
2916 They work like this: At the beginning of any module body, `fn` body, or any other block
2917 you can write a list of `use`-statements, consisting of the keyword `use` and a __global path__ to an item
2918 without the `::` prefix. For example, this imports `cow` into the local scope:
2922 # mod farm { pub fn cow() { println!("I'm a hidden ninja cow!") } }
2923 # fn main() { cow() }
2926 The path you give to `use` is per default global, meaning relative to the crate root,
2927 no matter how deep the module hierarchy is, or whether the module body it's written in
2928 is contained in its own file. (Remember: files are irrelevant.)
2930 This is different from other languages, where you often only find a single import construct that combines the semantic
2931 of `mod foo;` and `use`-statements, and which tend to work relative to the source file or use an absolute file path
2932 - Ruby's `require` or C/C++'s `#include` come to mind.
2934 However, it's also possible to import things relative to the module of the `use`-statement:
2935 Adding a `super::` in front of the path will start in the parent module,
2936 while adding a `self::` prefix will start in the current module:
2940 # pub fn some_parent_item(){ println!("...") }
2942 use super::some_parent_item;
2943 use self::some_child_module::some_item;
2944 # pub fn bar() { some_parent_item(); some_item() }
2945 # pub mod some_child_module { pub fn some_item() {} }
2950 Again - relative to the module, not to the file.
2952 Imports are also shadowed by local definitions:
2953 For each name you mention in a module/block, `rust`
2954 will first look at all items that are defined locally,
2955 and only if that results in no match look at items you brought in
2956 scope with corresponding `use` statements.
2959 # // FIXME: Allow unused import in doc test
2962 # mod farm { pub fn cow() { println!("Hidden ninja cow is hidden.") } }
2963 fn cow() { println!("Mooo!") }
2966 cow() // resolves to the locally defined `cow()` function
2970 To make this behavior more obvious, the rule has been made that `use`-statement always need to be written
2971 before any declaration, like in the example above. This is a purely artificial rule introduced
2972 because people always assumed they shadowed each other based on order, despite the fact that all items in rust are
2973 mutually recursive, order independent definitions.
2975 One odd consequence of that rule is that `use` statements also go in front of any `mod` declaration,
2976 even if they refer to things inside them:
2981 pub fn cow() { println!("Moooooo?") }
2987 This is what our `farm` example looks like with `use` statements:
2995 pub fn chicken() { println!("cluck cluck"); }
2996 pub fn cow() { println!("mooo"); }
2999 pub fn hay() { println!("..."); }
3004 println!("Hello farm!");
3006 // Can now refer to those names directly:
3013 And here an example with multiple files:
3016 // `a.rs` - crate root
3029 pub fn foo() { println!("Foo!"; }
3034 pub fn bar() { println!("Bar!"); }
3037 There also exist two short forms for importing multiple names at once:
3039 1. Explicit mention multiple names as the last element of an `use` path:
3042 use farm::{chicken, cow};
3044 # pub fn cow() { println!("Did I already mention how hidden and ninja I am?") }
3045 # pub fn chicken() { println!("I'm Bat-chicken, guardian of the hidden tutorial code.") }
3047 # fn main() { cow(); chicken() }
3050 2. Import everything in a module with a wildcard:
3053 # #![feature(globs)]
3056 # pub fn cow() { println!("Bat-chicken? What a stupid name!") }
3057 # pub fn chicken() { println!("Says the 'hidden ninja' cow.") }
3059 # fn main() { cow(); chicken() }
3062 > *Note:* This feature of the compiler is currently gated behind the
3063 > `#![feature(globs)]` directive. More about these directives can be found in
3066 However, that's not all. You can also rename an item while you're bringing it into scope:
3069 use egg_layer = farm::chicken;
3070 # mod farm { pub fn chicken() { println!("Laying eggs is fun!") } }
3078 In general, `use` creates a local alias:
3079 An alternate path and a possibly different name to access the same item,
3080 without touching the original, and with both being interchangeable.
3082 ## Reexporting names
3084 It is also possible to reexport items to be accessible under your module.
3086 For that, you write `pub use`:
3090 pub use self::barn::hay;
3092 pub fn chicken() { println!("cluck cluck"); }
3093 pub fn cow() { println!("mooo"); }
3096 pub fn hay() { println!("..."); }
3107 Just like in normal `use` statements, the exported names
3108 merely represent an alias to the same thing and can also be renamed.
3110 The above example also demonstrate what you can use `pub use` for:
3111 The nested `barn` module is private, but the `pub use` allows users
3112 of the module `farm` to access a function from `barn` without needing
3113 to know that `barn` exists.
3115 In other words, you can use it to decouple a public api from its internal implementation.
3119 So far we've only talked about how to define and structure your own crate.
3121 However, most code out there will want to use preexisting libraries,
3122 as there really is no reason to start from scratch each time you start a new project.
3124 In Rust terminology, we need a way to refer to other crates.
3126 For that, Rust offers you the `extern crate` declaration:
3130 // `num` ships with Rust (much like `extra`; more details further down).
3133 // The rational number '1/2':
3134 let one_half = ::num::rational::Ratio::new(1i, 2);
3138 A statement of the form `extern crate foo;` will cause `rustc` to search for the crate `foo`,
3139 and if it finds a matching binary it lets you use it from inside your crate.
3141 The effect it has on your module hierarchy mirrors aspects of both `mod` and `use`:
3143 - Like `mod`, it causes `rustc` to actually emit code:
3144 The linkage information the binary needs to use the library `foo`.
3146 - But like `use`, all `extern crate` statements that refer to the same library are interchangeable,
3147 as each one really just presents an alias to an external module (the crate root of the library
3148 you're linking against).
3150 Remember how `use`-statements have to go before local declarations because the latter shadows the former?
3151 Well, `extern crate` statements also have their own rules in that regard:
3152 Both `use` and local declarations can shadow them, so the rule is that `extern crate` has to go in front
3153 of both `use` and local declarations.
3155 Which can result in something like this:
3161 use num::rational::Ratio;
3164 pub fn dog() { println!("woof"); }
3169 let a_third = Ratio::new(1i, 3);
3173 It's a bit weird, but it's the result of shadowing rules that have been set that way because
3174 they model most closely what people expect to shadow.
3176 ## Crate metadata and settings
3178 For every crate you can define a number of metadata items, such as link name, version or author.
3179 You can also toggle settings that have crate-global consequences. Both mechanism
3180 work by providing attributes in the crate root.
3182 For example, Rust uniquely identifies crates by their link metadata, which includes
3183 the link name and the version. It also hashes the filename and the symbols in a binary
3184 based on the link metadata, allowing you to use two different versions of the same library in a crate
3187 Therefore, if you plan to compile your crate as a library, you should annotate it with that information:
3190 # #![allow(unused_attribute)]
3193 # #![crate_type = "lib"]
3194 #![crate_id = "farm#2.5"]
3200 You can also specify crate id information in a `extern crate` statement. For
3201 example, these `extern crate` statements would both accept and select the
3206 extern crate farm = "farm#2.5";
3207 extern crate my_farm = "farm";
3210 Other crate settings and metadata include things like enabling/disabling certain errors or warnings,
3211 or setting the crate type (library or executable) explicitly:
3214 # #![allow(unused_attribute)]
3218 // This crate is a library ("bin" is the default)
3219 #![crate_id = "farm#2.5"]
3220 #![crate_type = "lib"]
3222 // Turn on a warning
3223 #[warn(non_camel_case_types)]
3227 ## A minimal example
3229 Now for something that you can actually compile yourself.
3231 We define two crates, and use one of them as a library in the other.
3234 # #![allow(unused_attribute)]
3236 #![crate_id = "world#0.42"]
3238 # mod secret_module_to_make_this_test_run {
3239 pub fn explore() -> &'static str { "world" }
3246 fn main() { println!("hello {}", world::explore()); }
3249 Now compile and run like this (adjust to your platform if necessary):
3252 $ rustc --crate-type=lib world.rs # compiles libworld-<HASH>-0.42.rlib
3253 $ rustc main.rs -L . # compiles main
3258 Notice that the library produced contains the version in the file name
3259 as well as an inscrutable string of alphanumerics. As explained in the previous paragraph,
3260 these are both part of Rust's library versioning scheme. The alphanumerics are
3261 a hash representing the crate's id.
3263 ## The standard library and the prelude
3265 While reading the examples in this tutorial, you might have asked yourself where all
3266 those magical predefined items like `range` are coming from.
3268 The truth is, there's nothing magical about them: They are all defined normally
3269 in the `std` library, which is a crate that ships with Rust.
3271 The only magical thing that happens is that `rustc` automatically inserts this line into your crate root:
3277 As well as this line into every module body:
3280 use std::prelude::*;
3283 The role of the `prelude` module is to re-export common definitions from `std`.
3285 This allows you to use common types and functions like `Option<T>` or `range`
3286 without needing to import them. And if you need something from `std` that's not in the prelude,
3287 you just have to import it with an `use` statement.
3289 For example, it re-exports `range` which is defined in `std::iter::range`:
3292 use iter_range = std::iter::range;
3295 // `range` is imported by default
3296 for _ in range(0u, 10) {}
3298 // Doesn't hinder you from importing it under a different name yourself
3299 for _ in iter_range(0u, 10) {}
3301 // Or from not using the automatic import.
3302 for _ in ::std::iter::range(0u, 10) {}
3306 Both auto-insertions can be disabled with an attribute if necessary:
3309 # #![allow(unused_attribute)]
3310 // In the crate root:
3315 # #![allow(unused_attribute)]
3317 #![no_implicit_prelude]
3320 See the [API documentation][stddoc] for details.
3322 [stddoc]: std/index.html
3326 Now that you know the essentials, check out any of the additional
3327 guides on individual topics.
3329 * [Pointers][pointers]
3330 * [Lifetimes][lifetimes]
3331 * [Tasks and communication][tasks]
3333 * [The foreign function interface][ffi]
3334 * [Containers and iterators][container]
3335 * [Documenting Rust code][rustdoc]
3336 * [Testing Rust code][testing]
3337 * [The Rust Runtime][runtime]
3339 There is further documentation on the [wiki], however those tend to be even more out of date as this document.
3341 [pointers]: guide-pointers.html
3342 [lifetimes]: guide-lifetimes.html
3343 [tasks]: guide-tasks.html
3344 [macros]: guide-macros.html
3345 [ffi]: guide-ffi.html
3346 [container]: guide-container.html
3347 [testing]: guide-testing.html
3348 [runtime]: guide-runtime.html
3349 [rustdoc]: rustdoc.html
3350 [wiki]: https://github.com/rust-lang/rust/wiki/Docs
3352 [wiki-packages]: https://github.com/rust-lang/rust/wiki/Doc-packages,-editors,-and-other-tools