Let's Go

Introduction

This document is a tutorial introduction to the basics of the Go systems programming language, intended for programmers familiar with C or C++. It is not a comprehensive guide to the language; at the moment the document closest to that is the draft specification:

    http://go/go/doc/go_spec.html

    http://go/go/doc/go_setup.html

    /doc/progs

Hello, World

Let's start in the usual way:

 
01    package main

03    import fmt "fmt"  // Package implementing formatted I/O.

05    func main() {
06        fmt.Printf("Hello, world; or Καλημέρα κόσμε; or こんにちは 世界\n");
07    }

Every Go source file declares, using a package statement, which package it's part of. The main package's main function is where the program starts running (after any initialization). It may also import other packages to use their facilities. This program imports the package fmt to gain access to our old, now capitalized and package-qualified friend, fmt.Printf.

Function declarations are introduced with the func keyword.

Notice that string constants can contain Unicode characters, encoded in UTF-8. Go is defined to accept UTF-8 input. Strings are arrays of bytes, usually used to store Unicode strings represented in UTF-8.

The comment convention is the same as in C++:

    /* ... */
    // ...

Echo

Next up, here's a version of the Unix utility echo(1):

 
01    package main

03    import (
04        "os";
05        "flag";
06    )

08    var n_flag = flag.Bool("n", false, "don't print final newline")

10    const (
11        kSpace = " ";
12        kNewline = "\n";
13    )

15    func main() {
16        flag.Parse();   // Scans the arg list and sets up flags
17        var s string = "";
18        for i := 0; i < flag.NArg(); i++ {
19            if i > 0 {
20                s += kSpace
21            }
22            s += flag.Arg(i)
23        }
24        if !*n_flag {
25            s += kNewline
26        }
27        os.Stdout.WriteString(s);
28    }

This program is small but it's doing a number of new things. In the last example, we saw func introducing a function. The keywords var, const, and type (not used yet) also introduce declarations, as does import. Notice that we can group declarations of the same sort into parenthesized, semicolon-separated lists if we want, as on lines 3-6 and 10-13. But it's not necessary to do so; we could have said

    const Space = " "
    const Newline = "\n"

This program imports the "os" package to access its Stdout variable, of type *os.File. The import statement is actually a declaration: in its general form, as used in our ``hello world'' program, it names the identifier (fmt) that will be used to access members of the package imported from the file ("fmt"), found in the current directory or in a standard location. In this program, though, we've dropped the explicit name from the imports; by default, packages are imported using the name defined by the imported package, which by convention is of course the file name itself. Our ``hello world'' program could have said just import "fmt".

You can specify your own import names if you want but it's only necessary if you need to resolve a naming conflict.

Given os.Stdout we can use its WriteString method to print the string.

Having imported the flag package, line 8 creates a global variable to hold the value of echo's -n flag. The variable n_flag has type *bool, pointer to bool.

In main.main, we parse the arguments (line 16) and then create a local string variable we will use to build the output.

The declaration statement has the form

    var s string = "";

Go tries to be terse, and this declaration could be shortened. Since the string constant is of type string, we don't have to tell the compiler that. We could write

    var s = "";

    s := "";

 
18        for i := 0; i < flag.NArg(); i++ {

The flag package has parsed the arguments and left the non-flag arguments in a list that can be iterated over in the obvious way.

The Go for statement differs from that of C in a number of ways. First, it's the only looping construct; there is no while or do. Second, there are no parentheses on the clause, but the braces on the body are mandatory. The same applies to the if and switch statements. Later examples will show some other ways for can be written.

The body of the loop builds up the string s by appending (using +=) the flags and separating spaces. After the loop, if the -n flag is not set, it appends a newline, and then writes the result.

Notice that main.main is a niladic function with no return type. It's defined that way. Falling off the end of main.main means ''success''; if you want to signal an erroneous return, call

    os.Exit(1)

An Interlude about Types

Go has some familiar types such as int and float, which represent values of the ''appropriate'' size for the machine. It also defines specifically-sized types such as int8, float64, and so on, plus unsigned integer types such as uint, uint32, etc. These are distinct types; even if int and int32 are both 32 bits in size, they are not the same type. There is also a byte synonym for uint8, which is the element type for strings.

Speaking of string, that's a built-in type as well. Strings are immutable values -- they are not just arrays of byte values. Once you've built a string value, you can't change it, although of course you can change a string variable simply by reassigning it. This snippet from strings.go is legal code:

 
07        s := "hello";
08        if s[1] != 'e' { os.Exit(1) }
09        s = "good bye";
10        var p *string = &s;
11        *p = "ciao";

However the following statements are illegal because they would modify a string value:

    s[0] = 'x';
    (*p)[1] = 'y';

Yes, there are pointers. However, Go simplifies their use a little; read on.

Arrays are declared like this:

    var array_of_int [10]int;

The size of the array is part of its type; however, one can declare a slice variable, to which one can assign a pointer to any array with the same element type or - much more commonly - a slice expression of the form a[low : high], representing the subarray indexed by low through high-1. Slices look a lot like arrays but have no explicit size ([] vs. [10]) and they reference a segment of an underlying, often anonymous, regular array. Multiple slices can share data if they represent pieces of the same array; multiple arrays can never share data.

Slices are actually much more common in Go programs than regular arrays; they're more flexible, have reference semantics, and are efficient. What they lack is the precise control of storage layout of a regular array; if you want to have a hundred elements of an array stored within your structure, you should use a regular array.

When passing an array to a function, you almost always want to declare the formal parameter to be a slice. When you call the function, take the address of the array and Go will automatically create (efficiently) a slice reference and pass that.

Using slices one can write this function (from sum.go):

 
05    func sum(a []int) int {   // returns an int
06        s := 0;
07        for i := 0; i < len(a); i++ {
08            s += a[i]
09        }
10        return s
11    }

and invoke it like this:

 
15        s := sum(&[3]int{1,2,3});  // a slice of the array is passed to sum

Note how the return type (int) is defined for sum() by stating it after the parameter list. The expression [3]int{1,2,3} -- a type followed by a brace-bounded expression -- is a constructor for a value, in this case an array of 3 ints. Putting an & in front gives us the address of a unique instance of the value. We pass the pointer to sum() by (automatically) promoting it to a slice.

If you are creating a regular array but want the compiler to count the elements for you, use ... as the array size:

    s := sum(&[...]int{1,2,3});

    s := sum([]int{1,2,3});

    m := map[string] int {"one":1 , "two":2}

An Interlude about Allocation

Most types in Go are values. If you have an int or a struct or an array, assignment copies the contents of the object. To allocate something on the stack, just declare a variable. To allocate it on the heap, use new(), which returns a pointer to the allocated storage.

    type T struct { a, b int }
    var t *T = new(T);

    t := new(T);

    m := make(map[string] int);

    var m map[string] int;

Note that new(T) returns type *T while make(T) returns type T. If you (mistakenly) allocate a reference object with new(), you receive a pointer to an uninitialized reference, equivalent to declaring an uninitialized variable and taking its address.

An Interlude about Constants

Although integers come in lots of sizes in Go, integer constants do not. There are no constants like 0ll or 0x0UL. Instead, integer constants are evaluated as ideal, large-precision values that can overflow only when they are assigned to an integer variable with too little precision to represent the value.

    const hard_eight = (1 << 100) >> 97  // legal

    var a uint64 = 0  // a has type uint64, value 0
    a := uint64(0)    // equivalent; use a "conversion"
    i := 0x1234       // i gets default type: int
    var j int = 1e6   // legal - 1000000 is representable in an int
    x := 1.5          // a float
    i3div2 := 3/2     // integer division - result is 1
    f3div2 := 3./2.   // floating point division - result is 1.5

An I/O Package

Next we'll look at a simple package for doing file I/O with the usual sort of open/close/read/write interface. Here's the start of file.go:

 
01    package file

03    import (
04        "os";
05        "syscall";
06    )

08    type File struct {
09        fd      int;  // file descriptor number
10        name    string; // file name at Open time
11    }

The first line declares the name of the package -- file -- and then we import two packages. The os package hides the differences between various operating systems to give a consistent view of files and so on; here we're only going to use its error handling utilities and reproduce the rudiments of its file I/O.

The other item is the low-level, external syscall package, which provides a primitive interface to the underlying operating system's calls.

Next is a type definition: the type keyword introduces a type declaration, in this case a data structure called File. To make things a little more interesting, our File includes the name of the file that the file descriptor refers to.

Because File starts with a capital letter, the type is available outside the package, that is, by users of the package. In Go the rule about visibility of information is simple: if a name (of a top-level type, function, method, constant, variable, or of a structure field) is capitalized, users of the package may see it. Otherwise, the name and hence the thing being named is visible only inside the package in which it is declared. This is more than a convention; the rule is enforced by the compiler. In Go, the term for publicly visible names is ''exported''.

In the case of File, all its fields are lower case and so invisible to users, but we will soon give it some exported, upper-case methods.

First, though, here is a factory to create them:

 
13    func newFile(fd int, name string) *File {
14        if fd < 0 {
15            return nil
16        }
17        return &File{fd, name}
18    }

This returns a pointer to a new File structure with the file descriptor and name filled in. This code uses Go's notion of a ''composite literal'', analogous to the ones used to build maps and arrays, to construct a new heap-allocated object. We could write

    n := new(File);
    n.fd = fd;
    n.name = name;
    return n

We can use the factory to construct some familiar, exported variables of type *File:

 
20    var (
21        Stdin  = newFile(0, "/dev/stdin");
22        Stdout = newFile(1, "/dev/stdout");
23        Stderr = newFile(2, "/dev/stderr");
24    )

The newFile function was not exported because it's internal. The proper, exported factory to use is Open:

 
26    func Open(name string, mode int, perm int) (file *File, err os.Error) {
27        r, e := syscall.Open(name, mode, perm);
28        if e != 0 {
29            err = os.Errno(e);
30        }
31        return newFile(r, name), err
32    }

There are a number of new things in these few lines. First, Open returns multiple values, an File and an error (more about errors in a moment). We declare the multi-value return as a parenthesized list of declarations; syntactically they look just like a second parameter list. The function syscall.Open also has a multi-value return, which we can grab with the multi-variable declaration on line 27; it declares r and e to hold the two values, both of type int64 (although you'd have to look at the syscall package to see that). Finally, line 28 returns two values: a pointer to the new File and the error. If syscall.Open fails, the file descriptor r will be negative and NewFile will return nil.

About those errors: The os library includes a general notion of an error string, maintaining a unique set of errors throughout the program. It's a good idea to use its facility in your own interfaces, as we do here, for consistent error handling throughout Go code. In Open we use a conversion to os.Errno to translate Unix's integer errno value into an error value, which will be stored in a unique instance of type os.Error.

Now that we can build Files, we can write methods for them. To declare a method of a type, we define a function to have an explicit receiver of that type, placed in parentheses before the function name. Here are some methods for *File, each of which declares a receiver variable file.

 
34    func (file *File) Close() os.Error {
35        if file == nil {
36            return os.EINVAL
37        }
38        e := syscall.Close(file.fd);
39        file.fd = -1;  // so it can't be closed again
40        if e != 0 {
41            return os.Errno(e);
42        }
43        return nil
44    }

46    func (file *File) Read(b []byte) (ret int, err os.Error) {
47        if file == nil {
48            return -1, os.EINVAL
49        }
50        r, e := syscall.Read(file.fd, b);
51        if e != 0 {
52            err = os.Errno(e);
53        }
54        return int(r), err
55    }

57    func (file *File) Write(b []byte) (ret int, err os.Error) {
58        if file == nil {
59            return -1, os.EINVAL
60        }
61        r, e := syscall.Write(file.fd, b);
62        if e != 0 {
63            err = os.Errno(e);
64        }
65        return int(r), err
66    }

68    func (file *File) String() string {
69        return file.name
70    }

There is no implicit this and the receiver variable must be used to access members of the structure. Methods are not declared within the struct declaration itself. The struct declaration defines only data members. In fact, methods can be created for any type you name, such as an integer or array, not just for structs. We'll see an example with arrays later.

The String method is so called because of printing convention we'll describe later.

The methods use the public variable os.EINVAL to return the (os.Error version of the) Unix error code EINVAL. The os library defines a standard set of such error values.

We can now use our new package:

 
01    package main

03    import (
04        "./file";
05        "fmt";
06        "os";
07    )

09    func main() {
10        hello := []byte{'h', 'e', 'l', 'l', 'o', ',', ' ', 'w', 'o', 'r', 'l', 'd', '\n'};
11        file.Stdout.Write(hello);
12        file, err := file.Open("/does/not/exist",  0,  0);
13        if file == nil {
14            fmt.Printf("can't open file; err=%s\n",  err.String());
15            os.Exit(1);
16        }
17    }

The import of ''./file'' tells the compiler to use our own package rather than something from the directory of installed packages.

Finally we can run the program:

    % helloworld3
    hello, world
    can't open file; err=No such file or directory
    % 

Rotting cats

Building on the file package, here's a simple version of the Unix utility cat(1), progs/cat.go:

 
01    package main

03    import (
04        "./file";
05        "flag";
06        "fmt";
07        "os";
08    )

10    func cat(f *file.File) {
11        const NBUF = 512;
12        var buf [NBUF]byte;
13        for {
14            switch nr, er := f.Read(&buf); true {
15            case nr < 0:
16                fmt.Fprintf(os.Stderr, "error reading from %s: %s\n", f.String(), er.String());
17                os.Exit(1);
18            case nr == 0:  // EOF
19                return;
20            case nr > 0:
21                if nw, ew := file.Stdout.Write(buf[0:nr]); nw != nr {
22                    fmt.Fprintf(os.Stderr, "error writing from %s: %s\n", f.String(), ew.String());
23                }
24            }
25        }
26    }

28    func main() {
29        flag.Parse();   // Scans the arg list and sets up flags
30        if flag.NArg() == 0 {
31            cat(file.Stdin);
32        }
33        for i := 0; i < flag.NArg(); i++ {
34            f, err := file.Open(flag.Arg(i), 0, 0);
35            if f == nil {
36                fmt.Fprintf(os.Stderr, "can't open %s: error %s\n", flag.Arg(i), err);
37                os.Exit(1);
38            }
39            cat(f);
40            f.Close();
41        }
42    }

By now this should be easy to follow, but the switch statement introduces some new features. Like a for loop, an if or switch can include an initialization statement. The switch on line 14 uses one to create variables nr and er to hold the return values from f.Read(). (The if on line 21 has the same idea.) The switch statement is general: it evaluates the cases from top to bottom looking for the first case that matches the value; the case expressions don't need to be constants or even integers, as long as they all have the same type.

Since the switch value is just true, we could leave it off -- as is also the situation in a for statement, a missing value means true. In fact, such a switch is a form of if-else chain. While we're here, it should be mentioned that in switch statements each case has an implicit break.

Line 21 calls Write() by slicing the incoming buffer, which is itself a slice. Slices provide the standard Go way to handle I/O buffers.

Now let's make a variant of cat that optionally does rot13 on its input. It's easy to do by just processing the bytes, but instead we will exploit Go's notion of an interface.

The cat() subroutine uses only two methods of f: Read() and String(), so let's start by defining an interface that has exactly those two methods. Here is code from progs/cat_rot13.go:

 
22    type reader interface {
23        Read(b []byte) (ret int, err os.Error);
24        String() string;
25    }

Any type that implements the two methods of reader -- regardless of whatever other methods the type may also contain -- is said to implement the interface. Since file.File implements these methods, it implements the reader interface. We could tweak the cat subroutine to accept a reader instead of a *file.File and it would work just fine, but let's embellish a little first by writing a second type that implements reader, one that wraps an existing reader and does rot13 on the data. To do this, we just define the type and implement the methods and with no other bookkeeping, we have a second implementation of the reader interface.

 
27    type rotate13 struct {
28        source    reader;
29    }

31    func newRotate13(source reader) *rotate13 {
32        return &rotate13{source}
33    }

35    func (r13 *rotate13) Read(b []byte) (ret int, err os.Error) {
36        r, e := r13.source.Read(b);
37        for i := 0; i < r; i++ {
38            b[i] = rot13(b[i])
39        }
40        return r, e
41    }

43    func (r13 *rotate13) String() string {
44        return r13.source.String()
45    }
46    // end of rotate13 implementation

(The rot13 function called on line 38 is trivial and not worth reproducing.)

To use the new feature, we define a flag:

 
10    var rot13_flag = flag.Bool("rot13", false, "rot13 the input")

and use it from within a mostly unchanged cat() function:

 
48    func cat(r reader) {
49        const NBUF = 512;
50        var buf [NBUF]byte;

52        if *rot13_flag {
53            r = newRotate13(r)
54        }
55        for {
56            switch nr, er := r.Read(&buf); {
57            case nr < 0:
58                fmt.Fprintf(os.Stderr, "error reading from %s: %s\n", r.String(), er.String());
59                os.Exit(1);
60            case nr == 0:  // EOF
61                return;
62            case nr > 0:
63                nw, ew := file.Stdout.Write(buf[0:nr]);
64                if nw != nr {
65                    fmt.Fprintf(os.Stderr, "error writing from %s: %s\n", r.String(), ew.String());
66                }
67            }
68        }
69    }

(We could also do the wrapping in main and leave cat() mostly alone, except for changing the type of the argument; consider that an exercise.) Lines 52 through 55 set it all up: If the rot13 flag is true, wrap the reader we received into a rotate13 and proceed. Note that the interface variables are values, not pointers: the argument is of type reader, not *reader, even though under the covers it holds a pointer to a struct.

Here it is in action:

    % echo abcdefghijklmnopqrstuvwxyz | ./cat
    abcdefghijklmnopqrstuvwxyz
    % echo abcdefghijklmnopqrstuvwxyz | ./cat --rot13
    nopqrstuvwxyzabcdefghijklm
    % 

Fans of dependency injection may take cheer from how easily interfaces allow us to substitute the implementation of a file descriptor.

Interfaces are a distinct feature of Go. An interface is implemented by a type if the type implements all the methods declared in the interface. This means that a type may implement an arbitrary number of different interfaces. There is no type hierarchy; things can be much more ad hoc, as we saw with rot13. The type file.File implements reader; it could also implement a writer, or any other interface built from its methods that fits the current situation. Consider the empty interface

    type Empty interface {}

Every type implements the empty interface, which makes it useful for things like containers.

Sorting

Interfaces provide a simple form of polymorphism since they completely separate the definition of what an object does from how it does it, allowing distinct implementations to be represented at different times by the same interface variable.

As an example, consider this simple sort algorithm taken from progs/sort.go:

 
09    func Sort(data SortInterface) {
10        for i := 1; i < data.Len(); i++ {
11            for j := i; j > 0 && data.Less(j, j-1); j-- {
12                data.Swap(j, j-1);
13            }
14        }
15    }

The code needs only three methods, which we wrap into SortInterface:

 
03    type SortInterface interface {
04        Len() int;
05        Less(i, j int) bool;
06        Swap(i, j int);
07    }

We can apply Sort to any type that implements Len, Less, and Swap. The sort package includes the necessary methods to allow sorting of arrays of integers, strings, etc.; here's the code for arrays of int

 
29    type IntArray []int

31    func (p IntArray) Len() int            { return len(p); }
32    func (p IntArray) Less(i, j int) bool  { return p[i] < p[j]; }
33    func (p IntArray) Swap(i, j int)       { p[i], p[j] = p[j], p[i]; }


36    type FloatArray []float

38    func (p FloatArray) Len() int            { return len(p); }
39    func (p FloatArray) Less(i, j int) bool  { return p[i] < p[j]; }
40    func (p FloatArray) Swap(i, j int)       { p[i], p[j] = p[j], p[i]; }


43    type StringArray []string

45    func (p StringArray) Len() int            { return len(p); }
46    func (p StringArray) Less(i, j int) bool  { return p[i] < p[j]; }
47    func (p StringArray) Swap(i, j int)       { p[i], p[j] = p[j], p[i]; }


50    // Convenience wrappers for common cases

52    func SortInts(a []int)        { Sort(IntArray(a)); }
53    func SortFloats(a []float)    { Sort(FloatArray(a)); }
54    func SortStrings(a []string)  { Sort(StringArray(a)); }


57    func IntsAreSorted(a []int) bool       { return IsSorted(IntArray(a)); }
58    func FloatsAreSorted(a []float) bool   { return IsSorted(FloatArray(a)); }
59    func StringsAreSorted(a []string) bool { return IsSorted(StringArray(a)); }

Here we see methods defined for non-struct types. You can define methods for any type you define and name in your package.

And now a routine to test it out, from progs/sortmain.go. This uses a function in the sort package, omitted here for brevity, to test that the result is sorted.

 
08    func ints() {
09        data := []int{74, 59, 238, -784, 9845, 959, 905, 0, 0, 42, 7586, -5467984, 7586};
10        a := sort.IntArray(data);
11        sort.Sort(a);
12        if !sort.IsSorted(a) {
13            panic()
14        }
15    }

If we have a new type we want to be able to sort, all we need to do is to implement the three methods for that type, like this:

 
26    type day struct {
27        num        int;
28        short_name string;
29        long_name  string;
30    }

32    type dayArray struct {
33        data []*day;
34    }

36    func (p *dayArray) Len() int            { return len(p.data); }
37    func (p *dayArray) Less(i, j int) bool  { return p.data[i].num < p.data[j].num; }
38    func (p *dayArray) Swap(i, j int)       { p.data[i], p.data[j] = p.data[j], p.data[i]; }

Printing

The examples of formatted printing so far have been modest. In this section we'll talk about how formatted I/O can be done well in Go.

We've seen simple uses of the package fmt, which implements Printf, Fprintf, and so on. Within the fmt package, Printf is declared with this signature:

    Printf(format string, v ...) (n int, errno os.Error)

For example, in C each format must correspond to the type of its argument. It's easier in many cases in Go. Instead of %llud you can just say %d; Printf knows the size and signedness of the integer and can do the right thing for you. The snippet

 
06        var u64 uint64 = 1<<64-1;
07        fmt.Printf("%d %d\n", u64, int64(u64));

prints

    18446744073709551615 -1

 
10        type T struct { a int; b string };
11        t := T{77, "Sunset Strip"};
12        a := []int{1, 2, 3, 4};
13        fmt.Printf("%v %v %v\n", u64, t, a);

is

    18446744073709551615 {77 Sunset Strip} [1 2 3 4]

 
14        fmt.Print(u64, " ", t, " ", a, "\n");
15        fmt.Println(u64, t, a);

If you have your own type you'd like Printf or Print to format, just give it a String() method that returns a string. The print routines will examine the value to inquire whether it implements the method and if so, use it rather than some other formatting. Here's a simple example.

 
05    type testType struct { a int; b string }

07    func (t *testType) String() string {
08        return fmt.Sprint(t.a) + " " + t.b
09    }

11    func main() {
12        t := &testType{77, "Sunset Strip"};
13        fmt.Println(t)
14    }

Since *T has a String() method, the default formatter for that type will use it and produce the output

    77 Sunset Strip

Another feature of Printf is that the format %T will print a string representation of the type of a value, which can be handy when debugging polymorphic code.

It's possible to write full custom print formats with flags and precisions and such, but that's getting a little off the main thread so we'll leave it as an exploration exercise.

You might ask, though, how Printf can tell whether a type implements the String() method. Actually what it does is ask if the value can be converted to an interface variable that implements the method. Schematically, given a value v, it does this:

    type Stringer interface {
        String() string
    }

    s, ok := v.(Stringer);  // Test whether v implements "String()"
    if ok {
        result = s.String()
    } else {
        result = default_output(v)
    }

In this snippet the name Stringer follows the convention that we add [e]r to interfaces describing simple method sets like this.

One last wrinkle. To complete the suite, besides Printf etc. and Sprintf etc., there are also Fprintf etc. Unlike in C, Fprintf's first argument is not a file. Instead, it is a variable of type io.Writer, which is an interface type defined in the io library:

    type Writer interface {
        Write(p []byte) (n int, err os.Error);
    }

Prime numbers

Now we come to processes and communication -- concurrent programming. It's a big subject so to be brief we assume some familiarity with the topic.

A classic program in the style is the prime sieve of Eratosthenes. It works by taking a stream of all the natural numbers and introducing a sequence of filters, one for each prime, to winnow the multiples of that prime. At each step we have a sequence of filters of the primes so far, and the next number to pop out is the next prime, which triggers the creation of the next filter in the chain.

Here's a flow diagram; each box represents a filter element whose creation is triggered by the first number that flowed from the elements before it.

To create a stream of integers, we use a Go channel, which, borrowing from CSP's descendants, represents a communications channel that can connect two concurrent computations. In Go, channel variables are references to a run-time object that coordinates the communication; as with maps and slices, use make to create a new channel.

Here is the first function in progs/sieve.go:

 
05    // Send the sequence 2, 3, 4, ... to channel 'ch'.
06    func generate(ch chan int) {
07        for i := 2; ; i++ {
08            ch <- i  // Send 'i' to channel 'ch'.
09        }
10    }

The generate function sends the sequence 2, 3, 4, 5, ... to its argument channel, ch, using the binary communications operator <-. Channel operations block, so if there's no recipient for the value on ch, the send operation will wait until one becomes available.

The filter function has three arguments: an input channel, an output channel, and a prime number. It copies values from the input to the output, discarding anything divisible by the prime. The unary communications operator <- (receive) retrieves the next value on the channel.

 
12    // Copy the values from channel 'in' to channel 'out',
13    // removing those divisible by 'prime'.
14    func filter(in, out chan int, prime int) {
15        for {
16            i := <-in;  // Receive value of new variable 'i' from 'in'.
17            if i % prime != 0 {
18                out <- i  // Send 'i' to channel 'out'.
19            }
20        }
21    }

The generator and filters execute concurrently. Go has its own model of process/threads/light-weight processes/coroutines, so to avoid notational confusion we'll call concurrently executing computations in Go goroutines. To start a goroutine, invoke the function, prefixing the call with the keyword go; this starts the function running in parallel with the current computation but in the same address space:

    go sum(huge_array); // calculate sum in the background

    ch := make(chan int);
    go sum(huge_array, ch);
    // ... do something else for a while
    result := <-ch;  // wait for, and retrieve, result

 
24    func main() {
25        ch := make(chan int);  // Create a new channel.
26        go generate(ch);  // Start generate() as a goroutine.
27        for {
28            prime := <-ch;
29            fmt.Println(prime);
30            ch1 := make(chan int);
31            go filter(ch, ch1, prime);
32            ch = ch1
33        }
34    }

Line 25 creates the initial channel to pass to generate, which it then starts up. As each prime pops out of the channel, a new filter is added to the pipeline and its output becomes the new value of ch.

The sieve program can be tweaked to use a pattern common in this style of programming. Here is a variant version of generate, from progs/sieve1.go:

 
06    func generate() chan int {
07        ch := make(chan int);
08        go func(){
09            for i := 2; ; i++ {
10                ch <- i
11            }
12        }();
13        return ch;
14    }

This version does all the setup internally. It creates the output channel, launches a goroutine internally using a function literal, and returns the channel to the caller. It is a factory for concurrent execution, starting the goroutine and returning its connection.

The function literal notation (lines 8-12) allows us to construct an anonymous function and invoke it on the spot. Notice that the local variable ch is available to the function literal and lives on even after generate returns.

The same change can be made to filter:

 
17    func filter(in chan int, prime int) chan int {
18        out := make(chan int);
19        go func() {
20            for {
21                if i := <-in; i % prime != 0 {
22                    out <- i
23                }
24            }
25        }();
26        return out;
27    }

The sieve function's main loop becomes simpler and clearer as a result, and while we're at it let's turn it into a factory too:

 
29    func sieve() chan int {
30        out := make(chan int);
31        go func() {
32            ch := generate();
33            for {
34                prime := <-ch;
35                out <- prime;
36                ch = filter(ch, prime);
37            }
38        }();
39        return out;
40    }

Now main's interface to the prime sieve is a channel of primes:

 
42    func main() {
43        primes := sieve();
44        for {
45            fmt.Println(<-primes);
46        }
47    }

Multiplexing

With channels, it's possible to serve multiple independent client goroutines without writing an actual multiplexer. The trick is to send the server a channel in the message, which it will then use to reply to the original sender. A realistic client-server program is a lot of code, so here is a very simple substitute to illustrate the idea. It starts by defining a request type, which embeds a channel that will be used for the reply.

 
05    type request struct {
06        a, b    int;
07        replyc  chan int;
08    }

The server will be trivial: it will do simple binary operations on integers. Here's the code that invokes the operation and responds to the request:

 
10    type binOp func(a, b int) int

12    func run(op binOp, req *request) {
13        reply := op(req.a, req.b);
14        req.replyc <- reply;
15    }

Line 10 defines the name binOp to be a function taking two integers and returning a third.

The server routine loops forever, receiving requests and, to avoid blocking due to a long-running operation, starting a goroutine to do the actual work.

 
17    func server(op binOp, service chan *request) {
18        for {
19            req := <-service;
20            go run(op, req);  // don't wait for it
21        }
22    }

We construct a server in a familiar way, starting it up and returning a channel to connect to it:

 
24    func startServer(op binOp) chan *request {
25        req := make(chan *request);
26        go server(op, req);
27        return req;
28    }

Here's a simple test. It starts a server with an addition operator, and sends out lots of requests but doesn't wait for the reply. Only after all the requests are sent does it check the results.

 
30    func main() {
31        adder := startServer(func(a, b int) int { return a + b });
32        const N = 100;
33        var reqs [N]request;
34        for i := 0; i < N; i++ {
35            req := &reqs[i];
36            req.a = i;
37            req.b = i + N;
38            req.replyc = make(chan int);
39            adder <- req;
40        }
41        for i := N-1; i >= 0; i-- {   // doesn't matter what order
42            if <-reqs[i].replyc != N + 2*i {
43                fmt.Println("fail at", i);
44            }
45        }
46        fmt.Println("done");
47    }

One annoyance with this program is that it doesn't exit cleanly; when main returns there are a number of lingering goroutines blocked on communication. To solve this, we can provide a second, quit channel to the server:

 
28    func startServer(op binOp) (service chan *request, quit chan bool) {
29        service = make(chan *request);
30        quit = make(chan bool);
31        go server(op, service, quit);
32        return service, quit;
33    }

It passes the quit channel to the server function, which uses it like this:

 
17    func server(op binOp, service chan *request, quit chan bool) {
18        for {
19            select {
20            case req := <-service:
21                go run(op, req);  // don't wait for it
22            case <-quit:
23                return;
24            }
25        }
26    }

Inside server, a select statement chooses which of the multiple communications listed by its cases can proceed. If all are blocked, it waits until one can proceed; if multiple can proceed, it chooses one at random. In this instance, the select allows the server to honor requests until it receives a quit message, at which point it returns, terminating its execution.

All that's left is to strobe the quit channel at the end of main:

 
36        adder, quit := startServer(func(a, b int) int { return a + b });

 
51        quit <- true;

There's a lot more to Go programming and concurrent programming in general but this quick tour should give you some of the basics.