Common Go Patterns
This chapter is from the book
This chapter is from the book
This chapter is from the book
The first step to fluent use of any programming language is understanding the design patterns and idioms that are commonly used. Learning the syntax is only the first step to learning how to think in the language, akin to learning vocabulary and basic grammar in a natural language. People speaking a second language often make very amusing mistakes by literally translating idioms from their first language.
Programming languages are no different. If you’ve read C++ code written by Java programmers, or vice versa, then you’ve probably encountered this. Just translating an approach that you would use in one language into another will work (as long as both languages are equally expressive), but it will usually give horrible code. Design patterns in a programming language are like the idioms of a natural language. Some work in a lot of languages, others don’t. Quite often, you will find that design patterns in one language work around a missing feature. For example resource acquisition is initialization (RAII) is a common C++ idiom, yet makes no sense in a garbage collected language because object lifetimes are not related to their scopes, and better techniques (such as Go’s defer statement) exist solve the same problem. Go, like every other language, has a set of common patterns that are not necessarily applicable elsewhere.
Zero Initialization
5 type Logger struct {
6 out*os.File
7 }
8
9 func (l Logger) Log(s string) {
10 out := l.out
11 if (nil == out) {
12 out = os.Stderr
13 }
14 fmt.Fprintf(out, "%s [%d]: %s\n", os.Args[0],
os.Getpid(), s)
15 }
16
17 func (l*Logger) SetOutput(out*os.File) {
18 l.out = out
19 }
From: log.go
One of the important concepts in Go is the zero value. When you declare a new variable, or when you create a value with the new() built-in function, it is initialized to the zero value for the type.
As its name implies, the zero value is the value when all of the memory used by the type is filled with zeros. It is common for Go data types to be expected to work with their zero value, without any further initialization. For example, the zero value for a Go mutex is an unlocked mutex, you just need to create the memory for it and it’s ready to use. Similarly, the zero value for an arbitrary-precision integer in Go represents the value zero.
In other languages, the two-stage creation patternis common. This separates the allocation and initialization of objects into two explicit steps. In Go, there is no support for explicitly managing memory. If you declare a local variable and then take its address, or declare a pointer and use new() to create an object that it points to, the compiler is likely to generate the same code. The way in which you declare an object is a hint to the compiler, not an instruction. There is therefore little point in supporting two-stage creation in Go.
The second stage is also often redundant. An initializer that takes no arguments should not have to be stated. The fact that it can be commonly leads to bugs.
A concrete example of this is the POSIX thread API mutex. On FreeBSD, this is a pointer and a NULL value will be implicitly initialized. With the GNU/Linux implementation, it is a structure, and using an uninitialized version has undefined behavior. The compiler, however, has no way of knowing the difference between an initialized and an uninitialized mutex, so it will not give any warnings. A program that forgets to initialize the mutex can compile without any warnings, even at the highest warning level, and may work sometimes, but will fail unpredictably on some platforms.
In Go, this kind of bug is very rare. The common structures all use the zero initialization pattern, which means that you can always use a newly created instance of them immediately. The only time that you need to explicitly initialize one is when you want something other than the default behavior.
The same is true of other types. A pointer in Go is always initialized to nil, unless explicitly initialized to point to a valid object. In contrast, a pointer in C declared as a local value can contain any value. It may point to a live variable, or to an invalid memory address.
The Go approach has several advantages. First, there is no need for the compiler to perform complex flow analysis to warn you that a variable might be used uninitialized: there is no such thing as an uninitialized variable. This sounds simple, but determining if a variable may be used before being initialized is nontrivial, and modern C compilers still don’t always get it right. Secondly, it simplifies code. You only ever explicitly initialize a variable when you want to then use the value that you assigned to it.
You should aim to support this pattern in any Go structures that you define. Usually, this is easy. If your structure contains other structures as fields, and these structures support this pattern, then you get support for free.
If you are relying on other values, then it can be more complex. This is especially true with pointers: pointers do not support the zero initialization pattern and if you call a method on a nil pointer then you will get a crash.
The example at the start of this section shows one way of implementing the zero initialization pattern for structures that contain pointers. This example defines a structure for generating log messages and sending them to a file. The zero structure implicitly uses the standard error channel, rather than a file stored in the structure.
There are two things to note about this. The first is that we are not setting the out field in the structure to anything, if it is not set, we are just using a different value. There are two reasons for this. The first is largely aesthetic: it lets us tell the difference between a Logger that is writing to the standard output because it has been explicitly set to use the standard output, and one that is using it implicitly. In this particular example, it’s not very important, but in other cases it can be.
The other reason is that it means that this method does not need to take a pointer to the structure. This is quite important because of the how it relates to the Go type system. If you call a method via an interface, then methods that accept a pointer are only callable if the interface variable contains a pointer. If the interface variable contains a value, then methods that take a pointer will not be callable. It’s therefore good style in Go to only require methods to take a pointer when they modify the structure.