Arrays, Slices and Maps in Go

Each value of the three kinds of types is used to store a collection of element values. The types of all the elements stored in a container value are identical. The identical type is called the element type of (the container type of) the container value.

Each element in a container has an associated key. An element value can be accessed or modified through its associated key. The key types of map types must be comparable types. The key types of array and slice types are all the built-in type int. The keys of the elements of an array or slice are non-negative integers which mark the positions of these elements in the array or slice. The non-negative integer keys are often called indexes.

Each container value has a length property, which indicates how many elements are stored in that container. The valid range of the integer keys of an array or slice value is from zero (inclusive) to the length (exclusive) of the array or slice. For each value of a map type, the key values of that map value can be an arbitrary value of the key type of the map type.

There are also many differences between the three kinds of container types. Most of the differences originate from the differences between the value memory layouts of the three kinds of types. From the last article, value parts, we learned that an array value consists of only one direct part, however a slice or map value may have an underlying part, which is referenced by the direct part of the slice or map value.

Elements of an array or a slice are both stored contiguously in a continuous memory segment. For an array, the continuous memory segment hosts the direct part of the array. For a slice, the continuous memory segment hosts the underlying indirect part of the slice. The map implementation of the standard Go compiler/runtime adopts the hashtable algorithm. So all elements of a map are also stored in an underlying continuous memory segment, but they may be not contiguous. There may be many holes (gaps) within the continuous memory segment. Another common map implementation algorithm is the binary tree algorithm. Whatever algorithm is used, the keys associated with the elements of a map are also stored in (the underlying parts of) the map.

We can access an element through its key. The time complexities of element accesses on all container values are all O(1), though, generally map element accesses are several times slower than array and slice element accesses. But maps have two advantages over arrays and slices:

• the key types of maps can be any comparable types.
• maps consume much less memory than arrays and slices if most elements are zero values.

From the last article, we have learned that the underlying parts of a value will not get copied when the value is copied. In other words, if a value has underlying parts, a copy of the value will share the underlying parts with the value. This is the root reason of many behavior differences between array and slice/map values. These behavior differences will be introduced below.

The literal representations of the three kinds of unnamed container types:

• array types: [N]T
• slice types: []T
• map types: map[K]T

where

• T is an arbitrary type. It specifies the element type of a container type. Only values of the specified element type can be stored as element values of values of the container type.
• N must be a non-negative integer constant. It specifies the number of elements stored in any value of an array type, and it can be called the length of the array type. This means the length of an array type is the inherent part of the array type. For example, [5]int and [8]int are two distinct array types.
• K is an arbitrary comparable type. It specifies the key type of a map type. Most types in Go are comparable, incomparable types are listed here.

Here are some container type literal representation examples:

const Size = 32

type Person struct {
	name string
	age  int
}

/* Array types */

[5]string
[Size]int
// Element type is a slice type: []byte
[16][]byte
// Element type is a struct type: Person
[100]Person

/* Slice types *

[]bool
[]int64
// Element type is a map type: map[int]bool
[]map[int]bool
// Element type is a pointer type: *int
[]*int

/* Map types */

map[string]int
map[int]bool
// Element type is an array type: [6]string
map[int16][6]string
// Element type is a slice type: []string
map[bool][]string
// Element type is a pointer type: *int8,
// and key type is a struct type.
map[struct{x int}]*int8

The sizes of all slice types are the same. The sizes of all map types are also the same. The size of an array type depends on its length and the size of its element type. The size of a zero-length array type or an array type with a zero-size element type is zero.

Like struct values, container values can also be represented with composite literals, T{...}, where T denotes container type (except the zero values of slice and map types). Here are some examples:

// An array value containing four bool values.
[4]bool{false, true, true, false}

// A slice value which contains three words.
[]string{"break", "continue", "fallthrough"}

// A map value containing some key-value pairs.
map[string]int{"C": 1972, "Python": 1991, "Go": 2009}

Each key-element pair between the braces of a map composite literal is also called an entry.

There are several variants for array and slice composite literals:

// The followings slice composite literals
// are equivalent to each other.
[]string{"break", "continue", "fallthrough"}
[]string{0: "break", 1: "continue", 2: "fallthrough"}
[]string{2: "fallthrough", 1: "continue", 0: "break"}
[]string{2: "fallthrough", 0: "break", "continue"}

// The followings array composite literals
// are equivalent to each other.
[4]bool{false, true, true, false}
[4]bool{0: false, 1: true, 2: true, 3: false}
[4]bool{1: true, true}
[4]bool{2: true, 1: true}
[...]bool{false, true, true, false}
[...]bool{3: false, 1: true, true}

In the last two literals, the ...s mean we want to let compilers deduce the lengths for the corresponding array values.

From the above examples, we know that element indexes (keys) are optional in array and slice composite literals. In an array or slice composite literal,

• if an index is present, it is not needed to be a typed value of the key type int, but it must be a non-negative constant representable as a value of type int. And if it is typed, its type must be a basic integer type.
• an element which index is absent uses the previous element's index plus one as its index.
• if the index of the first element is absent, its index is zero.

The keys of entries in a map literal must not be absent, they can be non-constants.

var a uint = 1
var _ = map[uint]int {a : 123} // okay

// The following two lines fail to compile,
// for "a" is not a constant key/index.
var _ = []int{a: 100}  // error
var _ = [5]int{a: 100} // error

Constant keys in one specific composite literal can't be duplicate.

We have learned that struct composite literals can be taken addresses directly before. Container composite literals have no exceptions here.

Example:

package main

import "fmt"

func main() {
	pm := &map[string]int{"C": 1972, "Go": 2009}
	ps := &[]string{"break", "continue"}
	pa := &[...]bool{false, true, true, false}
	fmt.Printf("%T\n", pm) // *map[string]int
	fmt.Printf("%T\n", ps) // *[]string
	fmt.Printf("%T\n", pa) // *[4]bool
}

If a composite literal nested many other composite literals, then those nested composited literals can simplified to the form {...}.

For example, the slice value literal

// A slice value of a type whose element type is
// *[4]byte. The element type is a pointer type
// whose base type is [4]byte. The base type is
// an array type whose element type is "byte".
var heads = []*[4]byte{
	&[4]byte{'P', 'N', 'G', ' '},
	&[4]byte{'G', 'I', 'F', ' '},
	&[4]byte{'J', 'P', 'E', 'G'},
}

can be simplified to

var heads = []*[4]byte{
	{'P', 'N', 'G', ' '},
	{'G', 'I', 'F', ' '},
	{'J', 'P', 'E', 'G'},
}

The array value literal in the following example

type language struct {
	name string
	year int
}

var _ = [...]language{
	language{"C", 1972},
	language{"Python", 1991},
	language{"Go", 2009},
}

can be simplified to

var _ = [...]language{
	{"C", 1972},
	{"Python", 1991},
	{"Go", 2009},
}

And the map value literal in the following example

type LangCategory struct {
	dynamic bool
	strong  bool
}

// A value of map type whose key type is
// a struct type and whose element type
// is another map type "map[string]int".
var _ = map[LangCategory]map[string]int{
	LangCategory{true, true}: map[string]int{
		"Python": 1991,
		"Erlang": 1986,
	},
	LangCategory{true, false}: map[string]int{
		"JavaScript": 1995,
	},
	LangCategory{false, true}: map[string]int{
		"Go":   2009,
		"Rust": 2010,
	},
	LangCategory{false, false}: map[string]int{
		"C": 1972,
	},
}

can be simplified to

var _ = map[LangCategory]map[string]int{
	{true, true}: {
		"Python": 1991,
		"Erlang": 1986,
	},
	{true, false}: {
		"JavaScript": 1995,
	},
	{false, true}: {
		"Go":   2009,
		"Rust": 2010,
	},
	{false, false}: {
		"C": 1972,
	},
}

Please notes, in the above several examples, the comma following the last item in each composite literal can't be omitted. Please read the line break rules in Go for more information later.

Example:

package main

import "fmt"

func main() {
	var a [16]byte
	var s []int
	var m map[string]int

	fmt.Println(a == a)   // true
	fmt.Println(m == nil) // true
	fmt.Println(s == nil) // true
	fmt.Println(nil == map[string]int{}) // false
	fmt.Println(nil == []int{})          // false

	// The following lines fail to compile.
	/*
	_ = m == m
	_ = s == s
	_ = m == map[string]int(nil)
	_ = s == []int(nil)
	var x [16][]int
	_ = x == x
	var y [16]map[int]bool
	_ = y == y
	*/
}

Example:

package main

import "fmt"

func main() {
	var a [5]int
	fmt.Println(len(a), cap(a)) // 5 5
	var s []int
	fmt.Println(len(s), cap(s)) // 0 0
	s, s2 := []int{2, 3, 5}, []bool{}
	fmt.Println(len(s), cap(s))   // 3 3
	fmt.Println(len(s2), cap(s2)) // 0 0
	var m map[int]bool
	fmt.Println(len(m)) // 0
	m, m2 := map[int]bool{1: true, 0: false}, map[int]int{}
	fmt.Println(len(m), len(m2)) // 2 0
}

The length and capacity of each slice shown in the above specified example value are equal. This is not true for every slice value. We will use some slices whose respective lengths and capacities are not equal in the following sections.

The element associated to key k stored in a container value v is represented with the element indexing syntax form v[k].

For a use of v[k], assume v is an array or slice,

• if k is a constant, then it must satisfy the requirements described above for the indexes in container composite literals. In addition, if v is an array, the k must be smaller than the length of the array.
• if k is a non-constant value, it must be a value of any basic integer type. In addition, it must be larger than or equal to zero and smaller than len(v), otherwise, a run-time panic will occur.
• if v is a nil slice, a run-time panic will occur.

For a use of v[k], assume v is a map, then k must be assignable to values of the element type of the map type, and

• if k is an interface value whose dynamic type is incomparable, a panic will occur at run time.
• if v[k] is used as a destination value in an assignment and v is a nil map, a panic will occur at run time.
• if v[k] is used to retrieve the element value corresponding key k in map v, then no panics will occur, even if v is a nil map. (Assume the evaluation of k will not panic.)
• if v[k] is used to retrieve the element value corresponding key k in map v, and the map v doesn't contain an entry with key k, v[k] results in a zero value of the element type of the corresponding map type of v. Generally, v[k] is viewed as a single-value expression. However, when v[k] is used as the only source value expression in an assignment, it can be viewed as a multi-value expression and result a second optional untyped boolean value, which indicates whether or not the map v contains an entry with key k.

An example of container element accesses and modifications:

package main

import "fmt"

func main() {
	a := [3]int{-1, 0, 1}
	s := []bool{true, false}
	m := map[string]int{"abc": 123, "xyz": 789}
	fmt.Println (a[2], s[1], m["abc"])    // retrieve
	a[2], s[1], m["abc"] = 999, true, 567 // modify
	fmt.Println (a[2], s[1], m["abc"])    // retrieve

	n, present := m["hello"]
	fmt.Println(n, present, m["hello"]) // 0 false 0
	n, present = m["abc"]
	fmt.Println(n, present, m["abc"]) // 567 true 567
	m = nil
	fmt.Println(m["abc"]) // 0

	// The two lines fail to compile.
	/*
	_ = a[3]  // index 3 out of bounds
	_ = s[-1] // index must be non-negative
	*/

	// Each of the following lines can cause a panic.
	_ = a[n]         // panic: index out of range
	_ = s[n]         // panic: index out of range
	m["hello"] = 555 // panic: assign to entry in nil map
}

To understand slice types and values better and explain slices easier, we need to have an impression on the internal structure of slice types. From the last article, value parts, we learned that the internal structure of slice types defined by the standard Go compiler/runtime is like

type _slice struct {
	elements unsafe.Pointer // referencing underlying elements
	len      int            // length
	cap      int            // capacity
}

The internal structure definitions used by other compilers/runtimes implementations may be not the exact same but would be similar. The following explanations are based on the official slice implementation.

The above shown internal structure explains the memory layouts of the direct parts of slice values. The len field of the direct part of a slice indicates the length of the slice, and the cap field indicates the capacity of the slice. The following picture depicts one possible memory layout of a slice value.

Although the underlying memory segment which hosts the elements of a slice may be very large, the slice may be only aware of a sub-segment of the memory segment. For example, in the above picture, the slice is only aware of the middle grey sub-segment of the whole memory segment.

For the slice depicted in the above picture, the elements from index len to index cap (exclusive) don't belong to the elements of the slice. They are just some redundant element slots for the depicted slice, but they may be effective elements of other slices or another array.

The next section will explain how to append elements to a base slice and yield a new slice by using the built-in append function. The result slice of an append function call may share starting elements with the base slice or not, depending on the capacity (and length) of the base slice and how many elements are appended.

When the slice is used as the base slice in an append function call,

• if the number of appended elements is larger than the number of the redundant element slots of the base slice, a new underlying memory segment will be allocated for the result slice, thus the result slice and the base slice will not share any elements.
• otherwise, no new underlying memory segments will be allocated for the result slice, and the elements of the base slice also belong to the elements of the result slice. In other words, the two slices share some elements and all of their elements are hosted on the same underlying memory segment.

The section after next will show a picture which describes both of the two possible cases in appending slice elements.

There are more routes which lead to the elements of two slices are hosted on the same underlying memory segment. Such as assignments and the below to be introduced subslice operations.

Note, generally, we can't modify the three fields of a slice value individually, except through the reflection and unsafe ways. In other words, generally, to modify a slice value, its three fields must be modified together. Generally, this is achieved by assigning another slice value (of the same slice type) to the slice which needs to be modified.

If a map is assigned to another map, then the two maps will share all (underlying) elements. Appending elements into (or deleting elements from) one map will reflect on the other map.

Like map assignments, if a slice is assigned to another slice, they will share all (underlying) elements. Their respective lengths and capacities equal to each other. However, if the length/capacity of one slice changes later, the change will not reflect on the other slice.

When an array is assigned to another array, all the elements are copied from the source one to the destination one. The two arrays don't share any elements.

Example:

package main

import "fmt"

func main() {
	m0 := map[int]int{0:7, 1:8, 2:9}
	m1 := m0
	m1[0] = 2
	fmt.Println(m0, m1) // map[0:2 1:8 2:9] map[0:2 1:8 2:9]

	s0 := []int{7, 8, 9}
	s1 := s0
	s1[0] = 2
	fmt.Println(s0, s1) // [2 8 9] [2 8 9]

	a0 := [...]int{7, 8, 9}
	a1 := a0
	a1[0] = 2
	fmt.Println(a0, a1) // [7 8 9] [2 8 9]
}

The syntax of appending a key-element pair (an entry) to a map is the same as the syntax of modifying a map element. For example, for a non-nil map value m, the following line

m[k] = e

put the key-element pair (k, e) into the map m if m doesn't contain an entry with key k, or modify the element value associated with k if m contains an entry with key k.

There is a built-in delete function which is used to delete an entry from a map. For example, the following line will delete the entry with key k from the map m. If the map m doesn't contain an entry with key k, it is a no-op, no panics will occur, even if m is a nil map.

delete(m, k)

An example shows how to append (put) entries to and delete entries from maps:

package main

import "fmt"

func main() {
	m := map[string]int{"Go": 2007}
	m["C"] = 1972     // append
	m["Java"] = 1995  // append
	fmt.Println(m)    // map[C:1972 Go:2007 Java:1995]
	m["Go"] = 2009    // modify
	delete(m, "Java") // delete
	fmt.Println(m)    // map[C:1972 Go:2009]
}

Please note, before Go 1.12, the entry print order of a map is unspecified.

Array elements can neither be appended nor deleted, though elements of addressable arrays can be modified.

We can use the built-in append function to append multiple elements into a base slice and result a new slice. The result new slice contains the elements of the base slice and the appended elements. Please note, the base slice is not modified by the append function call. Surely, if we expect (and often in practice), we can assign the result slice to the base slice to modify the base slice.

There is not a built-in way to delete an element from a slice. We must use the append function and the subslice feature introduced below together to achieve this goal. Slice element deletions and insertions will be demoed in the below more slice manipulations section. Here, the following example only shows how to use the append function.

An example showing how to use the append function:

package main

import "fmt"

func main() {
	s0 := []int{2, 3, 5}
	fmt.Println(s0, cap(s0)) // [2 3 5] 3
	s1 := append(s0, 7)      // append one element
	fmt.Println(s1, cap(s1)) // [2 3 5 7] 6
	s2 := append(s1, 11, 13) // append two elements
	fmt.Println(s2, cap(s2)) // [2 3 5 7 11 13] 6
	s3 := append(s0)         // <=> s3 := s0
	fmt.Println(s3, cap(s3)) // [2 3 5] 3
	s4 := append(s0, s0...)  // double s0 as s4
	fmt.Println(s4, cap(s4)) // [2 3 5 2 3 5] 6

	s0[0], s1[0] = 99, 789
	fmt.Println(s2[0], s3[0], s4[0]) // 789 99 2
}

Note, the built-in append function is a variadic function. It has two parameters, the second one is a variadic parameter.

Variadic functions will be explained in the article after next. Currently, we only need to know that there are two manners to pass variadic arguments. In the above example, line 8, line 10 and line 12 use one manner and line 14 uses the other manner. For the former manner, we can pass zero or more element values as the variadic arguments. For the latter manner, we must pass a slice as the only variadic argument and which must be followed by three dots .... We can learn how to call variadic functions from the the article after next.

In the above example, line 14 is equivalent to

	s4 := append(s0, s0[0], s0[1], s0[2])

line 8 is equivalent to

	s1 := append(s0, []int{7}...)

and line 10 is equivalent to

	s2 := append(s1, []int{11, 13}...)

For the three-dot ... manner, the append function doesn't require the variadic argument must be a slice with the same type as the first slice argument, but their element types must be identical. In other words, the two argument slices must share the same underlying type.

In the above program,

• the append call at line 8 will allocate a new underlying memory segment for slice s1, for slice s0 doesn't have enough redundant element slots to store the new appended element. The same situation is for the append call at line 14.
• the append call at line 10 will not allocate a new underlying memory segment for slice s2, for slice s1 has enough redundant element slots to store the new appended elements.

So, s1 and s2 share some elements, s0 and s3 share all elements, and s4 doesn't share elements with others. The following picture depicted the statuses of these slices at the end of the above program.

Please note that, when an append call allocate a new underlying memory segment for the result slice, the capacity of the result slice is compiler dependent. For the standard Go compiler, if the capacity of the base slice is small, the capacity of the result slice will be at least the double of the base slice, to avoid allocating underlying memory segments frequently when the result slice is used as the base slices in later possible append calls.

As mentioned above, we can assign the result slice to the base slice in an append call to append elements into the base slice. For example,

package main

import "fmt"

func main() {
	var s = append([]string(nil), "array", "slice")
	fmt.Println(s)      // [array slice]
	fmt.Println(cap(s)) // 2
	s = append(s, "map")
	fmt.Println(s)      // [array slice map]
	fmt.Println(cap(s)) // 4
	s = append(s, "channel")
	fmt.Println(s)      // [array slice map channel]
	fmt.Println(cap(s)) // 4
}

The first argument of an append function call can't be an untyped nil (up to Go 1.20).

Besides using composite literals to create map and slice values, we can also use the built-in make function to create map and slice values. The built-in make function can't be used to create array values.

BTW, the built-in make function can also be used to create channels, which will be explained in the article channels in Go later.

Assume M is a map type and n is an integer, we can use the following two forms to create new maps of type M.

make(M, n)
make(M)

The first form creates a new empty map which is allocated with enough space to hold at least n entries without reallocating memory again. The second form only takes one argument, in which case a new empty map with enough space to hold a small number of entries without reallocating memory again. The small number is compiler dependent.

Note: the second argument n may be negative or zero, in which case it will be ignored.

Assume S is a slice type, length and capacity are two non-negative integers, length is not larger than capacity, we can use the following two forms to create new slices of type S. (The types of length and capacity are not required to be identical.)

make(S, length, capacity)
make(S, length)

The first form creates a new slice with the specified length and capacity. The second form only takes two arguments, in which case the capacity of the new created slice is the same as its length.

All the elements in the result slice of a make function call are initialized as the zero value (of the slice element type).

An example on how to use the built-in make function to create maps and slices:

package main

import "fmt"

func main() {
	// Make new maps.
	fmt.Println(make(map[string]int)) // map[]
	m := make(map[string]int, 3)
	fmt.Println(m, len(m)) // map[] 0
	m["C"] = 1972
	m["Go"] = 2009
	fmt.Println(m, len(m)) // map[C:1972 Go:2009] 2

	// Make new slices.
	s := make([]int, 3, 5)
	fmt.Println(s, len(s), cap(s)) // [0 0 0] 3 5
	s = make([]int, 2)
	fmt.Println(s, len(s), cap(s)) // [0 0] 2 2
}

From the article pointers in Go, we learned that we can also call the built-in new function to allocate a value of any type and get a pointer which references the allocated value. The allocated value is a zero value of its type. For this reason, it is a nonsense to use new function to create map and slice values.

It is not totally a nonsense to allocate a zero value of an array type with the built-in new function. However, it is seldom to do this in practice, for it is more convenient to use composite literals to allocate arrays.

Example:

package main

import "fmt"

func main() {
	m := *new(map[string]int)   // <=> var m map[string]int
	fmt.Println(m == nil)       // true
	s := *new([]int)            // <=> var s []int
	fmt.Println(s == nil)       // true
	a := *new([5]bool)          // <=> var a [5]bool
	fmt.Println(a == [5]bool{}) // true
}

Following are some facts about the addressabilities of container elements.

• Elements of addressable array values are also addressable. Elements of unaddressable array values are also unaddressable. The reason is each array value only consists of one direct part.
• Elements of any slice value are always addressable, whether or not that slice value is addressable. This is because the elements of a slice are stored in the underlying (indirect) value part of the slice and the underlying part is always hosted on an allocated memory segment.
• Elements of map values are always unaddressable. Please read this FAQ item for reasons.

For example:

package main

import "fmt"

func main() {
	a := [5]int{2, 3, 5, 7}
	s := make([]bool, 2)
	pa2, ps1 := &a[2], &s[1]
	fmt.Println(*pa2, *ps1) // 5 false
	a[2], s[1] = 99, true
	fmt.Println(*pa2, *ps1) // 99 true
	ps0 := &[]string{"Go", "C"}[0]
	fmt.Println(*ps0) // Go

	m := map[int]bool{1: true}
	_ = m
	// The following lines fail to compile.
	/*
	_ = &[3]int{2, 3, 5}[0]
	_ = &map[int]bool{1: true}[1]
	_ = &m[1]
	*/
}

Unlike most other unaddressable values, which direct parts can not be modified, the direct part of a map element values can be modified, but can only be modified (overwritten) as a whole. For most kinds of element types, this is not a big issue. However, if the element type of map type is an array type or struct type, things become some counter-intuitive.

From the last article, value parts, we learned that each of struct and array values only consists of one direct part. So

• if the element type of a map is a struct type, we can not individually modify each field of an element (which is a struct) of the map.
• if the element type of a map is an array type, we can not individually modify each element of an element (which is an array) of the map.

Example:

package main

import "fmt"

func main() {
	type T struct{age int}
	mt := map[string]T{}
	mt["John"] = T{age: 29} // modify it as a whole
	ma := map[int][5]int{}
	ma[1] = [5]int{1: 789} // modify it as a whole

	// The following two lines fail to compile,
	// for map elements can be modified partially.
	/*
	ma[1][1] = 123      // error
	mt["John"].age = 30 // error
	*/

	// Accesses are okay.
	fmt.Println(ma[1][1])       // 789
	fmt.Println(mt["John"].age) // 29
}

To make any expected modification work in the above example, the corresponding map element should be saved in a temporary variable firstly, then the temporary variable is modified as needed, in the end the corresponding map element is overwritten by the temporary variable. For example,

package main

import "fmt"

func main() {
	type T struct{age int}
	mt := map[string]T{}
	mt["John"] = T{age: 29}
	ma := map[int][5]int{}
	ma[1] = [5]int{1: 789}

	t := mt["John"] // a temporary copy
	t.age = 30
	mt["John"] = t // overwrite it back

	a := ma[1] // a temporary copy
	a[1] = 123
	ma[1] = a // overwrite it back

	fmt.Println(ma[1][1], mt["John"].age) // 123 30
}

Note: the just mentioned limit might be lifted later.

We can derive a new slice from another (base) slice or a base addressable array by using the subslice syntax forms (Go specification calls them as slice syntax forms). The process is also often called as reslicing. The elements of the derived slice and the base array or slice are hosted on the same memory segment. In other words, the derived slice and the base array or slice may share some contiguous elements.

There are two subslice syntax forms (baseContainer is an array or slice):

baseContainer[low : high]       // two-index form
baseContainer[low : high : max] // three-index form

The two-index form is equivalent to

baseContainer[low : high : cap(baseContainer)]

So the two-index form is a special case of the three-index form. The two-index form is used much more popularly than the three-index form in practice.

Note, the three-index form is only supported since Go 1.2.

In a subslice expression, the low, high and max indexes must satisfy the following relation requirements.

// two-index form
0 <= low <= high <= cap(baseContainer)

// three-index form
0 <= low <= high <= max <= cap(baseContainer)

Indexes not satisfying these requirements may make the subslice expression fail to compile at compile time or panic at run time, depending on the base container type kind and whether or not the indexes are constants.

Note,

• the low and high indexes can be both larger than len(baseContainer), as long as the above relations are all satisfied. But the two indexes must not be larger than cap(baseContainer).
• a subslice expression will not cause a panic if baseContainer is a nil slice and all indexes used in the expression are zero. The result slice derived from a nil slice is still a nil slice.

The length of the result derived slice is equal to high - low, and the capacity of the result derived slice is equal to max - low. The length of a derived slice may be larger than the base container, but the capacity will never be larger than the base container.

In practice, for simplicity, we often omitted some indexes in subslice syntax forms. The omission rules are:

• if the low index is equal to zero, it can be omitted, either for two-index or three-index forms.
• if the high is equal to len(baseContainer), it can be omitted, but only for two-index forms.
• the max can never be omitted in three-index forms.

For example, the following expressions are equivalent.

baseContainer[0 : len(baseContainer)]
baseContainer[: len(baseContainer)]
baseContainer[0 :]
baseContainer[:]
baseContainer[0 : len(baseContainer) : cap(baseContainer)]
baseContainer[: len(baseContainer) : cap(baseContainer)]

An example of using subslice syntax forms:

package main

import "fmt"

func main() {
	a := [...]int{0, 1, 2, 3, 4, 5, 6}
	s0 := a[:]     // <=> s0 := a[0:7:7]
	s1 := s0[:]    // <=> s1 := s0
	s2 := s1[1:3]  // <=> s2 := a[1:3]
	s3 := s1[3:]   // <=> s3 := s1[3:7]
	s4 := s0[3:5]  // <=> s4 := s0[3:5:7]
	s5 := s4[:2:2] // <=> s5 := s0[3:5:5]
	s6 := append(s4, 77)
	s7 := append(s5, 88)
	s8 := append(s7, 66)
	s3[1] = 99
	fmt.Println(len(s2), cap(s2), s2) // 2 6 [1 2]
	fmt.Println(len(s3), cap(s3), s3) // 4 4 [3 99 77 6]
	fmt.Println(len(s4), cap(s4), s4) // 2 4 [3 99]
	fmt.Println(len(s5), cap(s5), s5) // 2 2 [3 99]
	fmt.Println(len(s6), cap(s6), s6) // 3 4 [3 99 77]
	fmt.Println(len(s7), cap(s7), s7) // 3 4 [3 4 88]
	fmt.Println(len(s8), cap(s8), s8) // 4 4 [3 4 88 66]
}

The following picture depicts the final memory layouts of the array and slice values used in the above example.

From the picture, we know that the elements of slice s7 and s8 are hosted on a different underlying memory segment than the other containers. The elements of the other slices are hosted on the same memory segment hosting the array a.

Please note that, subslice operations may cause kind-of memory leaking. For example, half of the memory allocated for the return slice of a call to the following function will be wasted unless the returned slice becomes unreachable (if no other slices share the underlying element memory segment with the returned slice).

func f() []int {
	s := make([]int, 10, 100)
	return s[50:60]
}

Please note that, in the above function, the lower index (50) is larger than the length (10) of s, which is allowed.

Since Go 1.17, a slice may be converted to an array pointer. In such a conversion, if the length of the base array type of the pointer type is larger than the length of the slice, a panic occurs.

An example:

package main

type S []int
type A [2]int
type P *A

func main() {
	var x []int
	var y = make([]int, 0)
	var x0 = (*[0]int)(x) // okay, x0 == nil
	var y0 = (*[0]int)(y) // okay, y0 != nil
	_, _ = x0, y0

	var z = make([]int, 3, 5)
	var _ = (*[3]int)(z) // okay
	var _ = (*[2]int)(z) // okay
	var _ = (*A)(z)      // okay
	var _ = P(z)         // okay

	var w = S(z)
	var _ = (*[3]int)(w) // okay
	var _ = (*[2]int)(w) // okay
	var _ = (*A)(w)      // okay
	var _ = P(w)         // okay

	var _ = (*[4]int)(z) // will panic
}

Since Go 1.20, a slice may be converted to an array. In such a conversion, if the length of the array type is larger than the length of the slice, a panic occurs. The slice and the result array don't share any element.

An example:

package main

import "fmt"

func main() {
	var s = []int{0, 1, 2, 3}
	var a = [3]int(s[1:])
	s[2] = 9
	fmt.Println(s) // [0 1 9 3]
	fmt.Println(a) // [1 2 3]
	
	_ = [3]int(s[:2]) // panic
}

We can use the built-in copy function to copy elements from one slice to another, the types of the two slices are not required to be identical, but their element types must be identical. In other words, the two argument slices must share the same underlying type. The first parameter of the copy function is the destination slice and the second one is the source slice. The two parameters can overlap some elements. copy function returns the number of elements copied, which will be the smaller one of the lengths of the two parameters.

With the help of the subslice syntax, we can use the copy function to copy elements between two arrays or between an array and a slice.

An example:

package main

import "fmt"

func main() {
	type Ta []int
	type Tb []int
	dest := Ta{1, 2, 3}
	src := Tb{5, 6, 7, 8, 9}
	n := copy(dest, src)
	fmt.Println(n, dest) // 3 [5 6 7]
	n = copy(dest[1:], dest)
	fmt.Println(n, dest) // 2 [5 5 6]

	a := [4]int{} // an array
	n = copy(a[:], src)
	fmt.Println(n, a) // 4 [5 6 7 8]
	n = copy(a[:], a[2:])
	fmt.Println(n, a) // 2 [7 8 7 8]
}

Note, as a special case, the built-in copy function can be used to copy bytes from a string to a byte slice.

Neither of the two arguments of a copy function call can be an untyped nil value (up to Go 1.20).

In Go, keys and elements of a container value can be iterated with the following syntax:

for key, element = range aContainer {
	// use key and element ...
}

where for and range are two keywords, key and element are called iteration variables. If aContainer is a slice or an array (or an array pointer, see below), then the type of key must be built-in type int.

The assignment sign = can be a short variable declaration sign :=, in which case the two iteration variables are both two new declared variables which are only visible within the for-range code block body, if aContainer is a slice or an array (or an array pointer), then the type of key is deduced as int.

Like the traditional for loop block, each for-range loop block creates two code blocks, an implicit one and an explicit one which is formed by using {}. The explicit one is nested in the implicit one.

Like for loop blocks, break and continue statements can also be used in for-range loop blocks,

Example:

package main

import "fmt"

func main() {
	m := map[string]int{"C": 1972, "C++": 1983, "Go": 2009}
	for lang, year := range m {
		fmt.Printf("%v: %v \n", lang, year)
	}

	a := [...]int{2, 3, 5, 7, 11}
	for i, prime := range a {
		fmt.Printf("%v: %v \n", i, prime)
	}

	s := []string{"go", "defer", "goto", "var"}
	for i, keyword := range s {
		fmt.Printf("%v: %v \n", i, keyword)
	}
}

The form for-range code block syntax has several variants:

// Ignore the key iteration variable.
for _, element = range aContainer {
	// ...
}

// Ignore the element iteration variable.
for key, _ = range aContainer {
	element = aContainer[key]
	// ...
}

// The element iteration variable is omitted.
// This form is equivalent to the last one.
for key = range aContainer {
	element = aContainer[key]
	// ...
}

// Ignore both the key and element iteration variables.
for _, _ = range aContainer {
	// This variant is not much useful.
}

// Both the key and element iteration variables are
// omitted. This form is equivalent to the last one.
for range aContainer {
	// This variant is not much useful.
}

Iterating over nil maps or nil slices is allowed. Such iterations are no-ops.

Some details about iterations over maps are listed here.

• For a map, the entry order in an iteration is not guaranteed to be the same as the next iteration, even if the map is not modified between the two iterations. By Go specification, the order is unspecified (kind-of randomized).
• If a map entry (a key-element pair) which has not yet been reached is removed during an iteration, then the entry will not iterated in the same iteration for sure.
• If a map entry is created during an iteration, that entry may be iterated during the same iteration, or not.

If it is promised that there are no other goroutines manipulating a map m, then the following code is guaranteed to clear all entries stored in the map m:

for key := range m {
	delete(m, key)
}

Surely, array and slice elements can also be iterated by using the traditional for loop block:

for i := 0; i < len(anArrayOrSlice); i++ {
	element := anArrayOrSlice[i]
	// ...
}

Since version 1.11, the standard Go compiler has specially optimized the above loop block so that the loop block runs much faster than it looks.

For a for-range loop block

for key, element = range aContainer {...}

there are three important facts.

1. The ranged container is a copy of aContainer. Please note, only the direct part of aContainer is copied. The container copy is anonymous, so there are no ways to modify it.
   • If the aContainer is an array, then the modifications made on the array elements during the iteration will not be reflected to the iteration variables. The reason is that the copy of the array doesn't share elements with the array.
   • If the aContainer is a slice or map, then the modifications made on the slice or map elements during the iteration will be reflected to the iteration variables. The reason is that the clone of the slice (or map) shares all elements (entries) with the slice (or map).
2. A key-element pair of the copy of aContainer will be assigned (copied) to the iteration variable pair at each iteration step, so the modifications made on the direct parts of the iteration variables will not be reflected to the elements (and keys for maps) stored in aContainer. (For this fact, and as using for-range loop blocks is the only way to iterate map keys and elements, it is recommended not to use large-size types as map key and element types, to avoid large copy burdens.)
3. All key-element pairs will be assigned to the same iteration variable pair.

An example which proves the first and second facts.

package main

import "fmt"

func main() {
	type Person struct {
		name string
		age  int
	}
	persons := [2]Person {{"Alice", 28}, {"Bob", 25}}
	for i, p := range persons {
		fmt.Println(i, p)

		// This modification has no effects on
		// the iteration, for the ranged array
		// is a copy of the persons array.
		persons[1].name = "Jack"

		// This modification has not effects on
		// the persons array, for p is just a
		// copy of a copy of one persons element.
		p.age = 31
	}
	fmt.Println("persons:", &persons)
}

The output:

0 {Alice 28}
1 {Bob 25}
persons: &[{Alice 28} {Jack 25}]

If we change the array in the above to a slice, then the modification on the slice during the iteration has effects on the iteration, but the modification on the iteration variable still has no effects on the slice.

...

	// A slice.
	persons := []Person {{"Alice", 28}, {"Bob", 25}}
	for i, p := range persons {
		fmt.Println(i, p)

		// Now this modification has effects
		// on the iteration.
		persons[1].name = "Jack"

		// This modification still has not
		// any real effects.
		p.age = 31
	}
	fmt.Println("persons:", &persons)
}

The output becomes to:

0 {Alice 28}
1 {Jack 25}
persons: &[{Alice 28} {Jack 25}]

An example to prove the second and third facts.

package main

import "fmt"

func main() {
	langs := map[struct{ dynamic, strong bool }]map[string]int{
		{true, false}:  {"JavaScript": 1995},
		{false, true}:  {"Go": 2009},
		{false, false}: {"C": 1972},
	}
	// The key type and element type of this map
	// are both pointer types. Some weird, just
	// for education purpose.
	m0 := map[*struct{ dynamic, strong bool }]*map[string]int{}
	for category, langInfo := range langs {
		m0[&category] = &langInfo
		// This following line has no effects on langs.
		category.dynamic, category.strong = true, true
	}
	for category, langInfo := range langs {
		fmt.Println(category, langInfo)
	}

	m1 := map[struct{ dynamic, strong bool }]map[string]int{}
	for category, langInfo := range m0 {
		m1[*category] = *langInfo
	}
	// m0 and m1 both contain only one entry.
	fmt.Println(len(m0), len(m1)) // 1 1
	fmt.Println(m1) // map[{true true}:map[C:1972]]
}

As mentioned above, the entry iteration order is randomized, so the order of the first three lines of the output of the above program may be not same as the following one.

{false true} map[Go:2009]
{false false} map[C:1972]
{true false} map[JavaScript:1995]
1 1
map[{true true}:map[Go:2009]]

The cost of a slice or map assignment is small, but the cost of an array assignment is large if the size of the array type is large. So, generally, it is not a good idea to range over a large array. We can range over a slice derived from the array, or range over a pointer to the array (see the next section for details).

For an array or slice, if the size of its element type is large, then, generally, it is also not a good idea to use the second iteration variable to store the iterated element at each loop step. For such arrays and slices, we should use the one-iteration-variable for-range loop variant or the traditional for loop to iterate their elements. In the following example, the loop in function fa is much less efficient than the loop in function fb.

type Buffer struct {
	start, end int
	data       [1024]byte
}

func fa(buffers []Buffer) int {
	numUnreads := 0
	for _, buf := range buffers {
		numUnreads += buf.end - buf.start
	}
	return numUnreads
}

func fb(buffers []Buffer) int {
	numUnreads := 0
	for i := range buffers {
		numUnreads += buffers[i].end - buffers[i].start
	}
	return numUnreads
}

In many scenarios, we can use a pointer to an array as the array.

We can range over a pointer to an array to iterate the elements of the array. For arrays with large lengths, this way is much more efficient, for copying a pointer is much more efficient than copying a large-size array. In the following example, the two loop blocks are equivalent and both are efficient.

package main

import "fmt"

func main() {
	var a [100]int

	// Copying a pointer is cheap.
	for i, n := range &a {
		fmt.Println(i, n)
	}

	// Copying a slice is cheap.
	for i, n := range a[:] {
		fmt.Println(i, n)
	}
}

If the second iteration in a for-range loop is neither ignored nor omitted, then range over a nil array pointer will panic. In the following example, each of the first two loop blocks will print five indexes, however, the last one will produce a panic.

package main

import "fmt"

func main() {
	var p *[5]int // nil

	for i, _ := range p { // okay
		fmt.Println(i)
	}

	for i := range p { // okay
		fmt.Println(i)
	}

	for i, n := range p { // panic
		fmt.Println(i, n)
	}
}

Array pointers can also used to index array elements. Indexing array elements through a nil array pointer produces a panic.

package main

import "fmt"

func main() {
	a := [5]int{2, 3, 5, 7, 11}
	p := &a
	p[0], p[1] = 17, 19
	fmt.Println(a) // [17 19 5 7 11]
	p = nil
	_ = p[0] // panic
}

We can also derive slices from array pointers. Deriving slices from a nil array pointer produce a panic.

package main

import "fmt"

func main() {
	pa := &[5]int{2, 3, 5, 7, 11}
	s := pa[1:3]
	fmt.Println(s) // [3 5]
	pa = nil
	s = pa[0:0] // panic
	// Should this line execute, it also panics.
	_ = (*[0]byte)(nil)[:]
}

We can also pass array pointers as the arguments of the built-in len and cap functions. Nil array pointer arguments for the two functions will not produce panics.

var pa *[5]int // == nil
fmt.Println(len(pa), cap(pa)) // 5 5

Assume t0 is a literal presentation of the zero value of type T, and a is an array or slice which element type is T, then the standard Go compiler will translate the following one-iteration-variable for-range loop block

for i := range a {
	a[i] = t0
}

to an internal memclr call, generally which is faster than resetting each element one by one.

The optimization was adopted in the standard Go compiler 1.5.

Since Go Toolchain 1.19, the optimization also works if the ranged value is an array pointer.

If the argument passed to a built-in function len or cap function call is an array or an array pointer value, then the call is evaluated at compile time and the result of the call is a typed constant with type as the built-in type int. The result can be bound to named constants.

Example:

package main

import "fmt"

var a [5]int
var p *[7]string

// N and M are both typed constants.
const N = len(a)
const M = cap(p)

func main() {
	fmt.Println(N) // 5
	fmt.Println(M) // 7
}

Above has mentioned, generally, the length and capacity of a slice value can't be modified individually. A slice value can only be overwritten as a whole by assigning another slice value to it. However, we can modify the length and capacity of a slice individually by using reflections. Reflection will be explained in a later article in detail.

Example:

package main

import (
	"fmt"
	"reflect"
)

func main() {
	s := make([]int, 2, 6)
	fmt.Println(len(s), cap(s)) // 2 6

	reflect.ValueOf(&s).Elem().SetLen(3)
	fmt.Println(len(s), cap(s)) // 3 6

	reflect.ValueOf(&s).Elem().SetCap(5)
	fmt.Println(len(s), cap(s)) // 3 5
}

The second argument passed to the reflect.SetLen function must not be larger than the current capacity of the argument slice s. The second argument passed to the reflect.SetCap function must not be smaller than the current length of the argument slice s and larger than the current capacity of the argument slice s. Otherwise, a panic will occur.

The reflection way is very inefficient, it is slower than a slice assignment.

For the current Go verison (1.20), the simplest way to clone a slice is:

sClone := append(s[:0:0], s...)

We can also use the following line instead, but comparing with the above way, it has one imperfection. If the source slice s is blank non-nil slice, then the result slice is nil.

sClone := append([]T(nil), s...)

The above append ways have one drawback. They both might allocate (then initialize) more elements than needed. We can also use the following ways to avoid this drawback:

// The two-line make+copy way.
sClone := make([]T, len(s))
copy(sClone, s)

// Or the following one-line make+append way.
// As of Go Toolchain v1.20, this way is a bit
// slower than the above two-line make+copy way.
sClone := append(make([]T, 0, len(s)), s...)

The "make" ways have a drawback that if s is a nil slice, then they both result a non-nil blank slice. However, in practice, we often needn't to pursue the perfection. If we do, then three more lines are needed:

var sClone []T
if s != nil {
	sClone = make([]T, len(s))
	copy(sClone, s)
}

Above has mentioned that the elements a slice are stored contiguously in memory and there are no gaps between any two adjacent elements of the slice. So when a slice element is removed,

• if the element order must be preserved, then each of the subsequent elements followed the removed elements must be moved forwards.
• if the element order doesn't need to be preserved, then we can move the last elements in the slice to the removed indexes.

In the following example, assume from and to are two legal indexes, from is not larger than to, and the to index is exclusive.

// way 1 (preserve element orders):
s = append(s[:from], s[to:]...)

// way 2 (preserve element orders):
s = s[:from + copy(s[from:], s[to:])]

// Don't preserve element orders:
if n := to-from; len(s)-to < n {
	copy(s[from:to], s[to:])
} else {
	copy(s[from:to], s[len(s)-n:])
}
s = s[:len(s)-(to-from)]

If the slice elements reference other values, we should reset tail elements (on the just freed-up slots) to avoid memory leaking.

// "len(s)+to-from" is the old slice length.
temp := s[len(s):len(s)+to-from]
for i := range temp {
	temp[i] = t0
}

As mentioned above, the for-range loop code block will be optimized as a memclr call by the standard Go compiler.

Deleting one element is similar to, and also simpler than, deleting a segment of elements.

In the following example, assume i the index of the element to be removed and i is a legal index.

// Way 1 (preserve element orders):
s = append(s[:i], s[i+1:]...)

// Way 2 (preserve element orders):
s = s[:i + copy(s[i:], s[i+1:])]

// There will be len(s)-i-1 elements being
// copied in either of the above two ways.

// Don't preserve element orders:
s[i] = s[len(s)-1]
s = s[:len(s)-1]

If the slice elements contain pointers, then after the deletion action, we should reset the last element of the old slice value to avoid memory leaking:

s[len(s):len(s)+1][0] = t0
// or
s[:len(s)+1][len(s)] = t0

Sometimes, we may need to delete slice elements by some conditions.

// Assume T is a small-size type.
func DeleteElements(s []T, keep func(T) bool, clear bool) []T {
	//result := make([]T, 0, len(s))
	result := s[:0] // without allocating a new slice
	for _, v := range s {
		if keep(v) {
			result = append(result, v)
		}
	}
	if clear { // avoid memory leaking
		temp := s[len(result):]
		for i := range temp {
			// t0 is a zero value literal of T.
			temp[i] = t0
		}
	}
	return result
}

Please note, if T is not a small-size type, then generally we should try to avoid using T as function parameter types and using two-iteration-variable for-range block form to iterate slices with element types as T.

Assume the insertion position is a legal index i and elements is the slice whose elements are to be inserted.

// One-line implementation:
s = append(s[:i], append(elements, s[i:]...)...)

// The one-line way always copies s[i:] twice and
// might make at most two allocations.
// The following verbose way always copies s[i:]
// once and will only make at most one allocation.
// However, as of Go Toolchain v1.20, the "make"
// call will clear partial of just allocated elements,
// which is actually unnecessary. So the verbose
// way is more efficient than the one-line way
// only for small slices now.

if cap(s) >= len(s) + len(elements) {
	s = s[:len(s)+len(elements)]
	copy(s[i+len(elements):], s[i:])
	copy(s[i:], elements)
} else {
	x := make([]T, 0, len(elements)+len(s))
	x = append(x, s[:i]...)
	x = append(x, elements...)
	x = append(x, s[i:]...)
	s = x
}

// Push:
s = append(s, elements...)

// Unshift (insert at the beginning):
s = append(elements, s...)

Assume the pushed or popped element is e and slice s has at least one element.

// Pop front (shift):
s, e = s[1:], s[0]
// Pop back:
s, e = s[:len(s)-1], s[len(s)-1]
// Push front
s = append([]T{e}, s...)
// Push back:
s = append(s, e)

Please note, using append to insert elements is often inefficient, for all the elements after the insertion position need to be move backwards, and a larger space needs to be allocated to store all the elements if the free capacity is insufficient. These might be not serious problems for small slices, but for large slices, they often are. So if the number of the elements needing to be moved is large, it is best to use a linked list to do the insertions.

• Go 101에 대해 - 이 책이 쓰여진 이유
• 감사의 말

• Go 소개 - Go를 배우는 가치
• Go 툴체인 - Go 프로그램을 컴파일하고 실행하는 방법

• Go 코드에 익숙해지기
  • 소스 코드 요소 소개
  • 키워드와 식별자
  • 기본 자료형과 기본 값 리터럴
  • 상수와 변수 - 무형성(untyped) 값과 자료형 추론 소개를 포함
  • 일반 연산자 - 더 많은 자료형 추론 규칙 소개를 포함
  • 함수 선언과 호출
  • 코드 패키지와 패키지 들여오기
  • 표현식, 구문과 단순 구문
  • 기본 흐름 제어
  • 고루틴, 지연된 함수 호출과 패닉/복구

• Go 자료형 체계
  • Go 자료형 체계 개요 - Go 프로그래밍 숙달을 위해 반드시 읽어봐야 하는
  • 포인터
  • 구조체
  • 변수 - Go 변수에 대한 더 깊은 이해
  • 배열, 슬라이스와 맵 - 1급 객체 컨테이너 자료형
  • 문자열
  • 함수 - 함수 자료형과 값, 가변 인자 함수
  • 채널 - Go에서 동시성 동기화를 하는 방법
  • 메서드
  • 인터페이스 - 리플렉션과 다형성을 하는 값 상자
  • 자료형 임베딩 - 자료형을 확장하는 방법
  • 자료형에 안전하지 않는 포인터
  • 제네릭 - 합성 자료형의 사용과 읽는 법
  • 리플렉션 - reflect 표준 패키지

• 특별 주제
  • 개행 규칙
  • 지연된 함수 호출 더 알아보기
  • 패닉/복구 사용 사례
  • 패닉/복구 메커니즘에 대한 고찰 - 함수 호출 종료 단계를 포함
  • 코드 블록과 식별자 스코프
  • 표현식 평가 순서
  • Go의 값 복사 비용
  • 경계 검사 제거(BCE)

• 동시성 프로그래밍
  • 동시성 동기화 개요
  • 채널 사용 사례
  • 채널을 깔끔하게 닫는 방법
  • 기타 동시성 동기화 기술 - sync 표준 패키지
  • 원자적 연산 - sync/atomic 표준 패키지
  • Go의 메모리 순서 보장
  • 흔히들 저지르는 동시성 프로그래밍 실수들

• 메모리 관련
  • 메모리 블록
  • 메모리 레이아웃
  • 메모리 누수 시나리오

• 일부 요약
  • 몇 가지 간단한 요약
  • Go의 nil
  • 값 변환, 할당과 비교 규칙
  • 구문/의미 예외
  • Go Details 101
  • Go FAQ 101
  • Go Tips 101

• 더 많은 토픽