Three new books, Go Optimizations 101, Go Details & Tips 101 and Go Generics 101 are published now. It is most cost-effective to buy all of them through this book bundle in the Leanpub book store.

Go Value Copy Costs

Value copying happens frequently in Go programming. Values assignments, argument passing and channel value send operations are all value copying involved. This article will talk about the copy costs of values of all kinds of types.

Value Sizes

The size of a value means how many bytes the direct part of the value will occupy in memory. The indirect underlying parts of a value don't contribute to the size of the value.

In Go, if the types of two values belong to the same kind, and the type kind is not string kind, interface kind, array kind and struct kind, then the sizes of the two value are always equal.

In fact, for the standard Go compiler/runtime, the sizes of two string values are also always equal. The same relation is for the sizes of two interface values.

Up to present (Go Toolchain 1.23.n), for the standard Go compiler (and gccgo), values of a specified type always have the same value size. So, often, we call the size of a value as the size of the type of the value.

The size of an array type depends on the element type size and the length of the array type. The array type size is the product of the size of the array element type and the array length.

The size of a struct type depends on all of the sizes and the order of its fields. For there may be some padding bytes being inserted between two adjacent struct fields to guarantee certain memory address alignment requirements of these fields, so the size of a struct type must be not smaller than (and often larger than) the sum of the respective type sizes of its fields.

The following table lists the value sizes of all kinds of types (for the standard Go compiler v1.23.n). In the table, one word means one native word, which is 4 bytes on 32-bit architectures and 8 bytes on 64-bit architectures.

Kinds of Types Value Size Required by Go Specification
bool 1 byte not specified
int8, uint8 (byte) 1 byte 1 byte
int16, uint16 2 bytes 2 bytes
int32 (rune), uint32, float32 4 bytes 4 bytes
int64, uint64, float64, complex64 8 bytes 8 bytes
complex128 16 bytes 16 bytes
int, uint 1 word architecture dependent, 4 bytes on 32-bit architectures and 8 bytes on 64-bit architectures
uintptr 1 word large enough to store the uninterpreted bits of a pointer value
string 2 words not specified
pointer (safe or unsafe) 1 word not specified
slice 3 words not specified
map 1 word not specified
channel 1 word not specified
function 1 word not specified
interface 2 words not specified
struct (the sum of sizes of all fields) + (the number of padding bytes) the size of a struct type is zero if it contains no fields that have a size greater than zero
array (element value size) * (array length) the size of an array type is zero if its element type has zero size

Value Copy Costs

Generally speaking, the cost to copy a value is proportional to the size of the value. However, value sizes are not the only factor determining value copy costs. Different CPU models and compiler versions may specially optimize value copying for values with specific sizes.

In practice, we can view struct values with less than 5 fields and with sizes not larger than four native words as small-size values. The costs of copying small-size values are small.

For the standard Go compiler, except values of large-size struct and array types, other types in Go are all small-size types.

To avoid large value copy costs in argument passing and channel value send and receive operations, we should try to avoid using large-size struct and array types as function and method parameter types (including method receiver types) and channel element types. We can use pointer types whose base types are large-size types instead for such scenarios.

On the other hand, we should also consider the fact that too many pointers will increase the pressure of garbage collectors at run time. So whether large-size struct and array types or their corresponding pointer types should be used relies on specific circumstances.

Generally, in practice, we seldom use pointer types whose base types are slice types, map types, channel types, function types, string types and interface types. The costs of copying values of these assumed base types are very small.

We should also try to avoid using the two-iteration-variable forms to iterate array and slice elements if the element types are large-size types, for each element value will be copied to the second iteration variable in the iteration process.

The following is an example which benchmarks different ways to iterate slice elements.
package main

import "testing"

type S [12]int64
var sX = make([]S, 1000)
var sY = make([]S, 1000)
var sZ = make([]S, 1000)
var sumX, sumY, sumZ int64

func Benchmark_Loop(b *testing.B) {
	for i := 0; i < b.N; i++ {
		sumX = 0
		for j := 0; j < len(sX); j++ {
			sumX += sX[j][0]
		}
	}
}

func Benchmark_Range_OneIterVar(b *testing.B) {
	for i := 0; i < b.N; i++ {
		sumY = 0
		for j := range sY {
			sumY += sY[j][0]
		}
	}
}

func Benchmark_Range_TwoIterVar(b *testing.B) {
	for i := 0; i < b.N; i++ {
		sumZ = 0
		for _, v := range sZ {
			sumZ += v[0]
		}
	}
}
Run the benchmarks in the directory of the test file, we will get a result similar to:
Benchmark_Loop-4             424342 2708 ns/op
Benchmark_Range_OneIterVar-4 407905 2808 ns/op
Benchmark_Range_TwoIterVar-4 214860 3915 ns/op

We can find that the efficiency of the two-iteration-variable form is much lower than the other two. But please note that, some compilers might make special optimizations to remove the performance differences between these forms. The above benchmark result is based on the standard Go compiler v1.23.n.


(more articles ↡)

The Go 101 project is hosted on Github. Welcome to improve Go 101 articles by submitting corrections for all kinds of mistakes, such as typos, grammar errors, wording inaccuracies, description flaws, code bugs and broken links.

If you would like to learn some Go details and facts every serveral days, please follow Go 101's official Twitter account @zigo_101.

The digital versions of this book are available at the following places:
Tapir, the author of Go 101, has been on writing the Go 101 series books and maintaining the go101.org website since 2016 July. New contents will be continually added to the book and the website from time to time. Tapir is also an indie game developer. You can also support Go 101 by playing Tapir's games (made for both Android and iPhone/iPad):
Individual donations via PayPal are also welcome.

Articles in this book: