Julia Language Types


  • immutable MyType; field; field; end
  • type MyType; field; field; end


Types are key to Julia's performance. An important idea for performance is type stability, which occurs when the type a function returns only depends on the types, not the values, of its arguments.

Dispatching on Types

On Julia, you can define more than one method for each function. Suppose we define three methods of the same function:

foo(x) = 1
foo(x::Number) = 2
foo(x::Int) = 3

When deciding what method to use (called dispatch), Julia chooses the more specific method that matches the types of the arguments:

julia> foo('one')

julia> foo(1.0)

julia> foo(1)

This facilitates polymorphism. For instance, we can easily create a linked list by defining two immutable types, named Nil and Cons. These names are traditionally used to describe an empty list and a non-empty list, respectively.

abstract LinkedList
immutable Nil <: LinkedList end
immutable Cons <: LinkedList

We will represent the empty list by Nil() and any other lists by Cons(first, rest), where first is the first element of the linked list and rest is the linked list consisting of all remaining elements. For example, the list [1, 2, 3] will be represented as

julia> Cons(1, Cons(2, Cons(3, Nil())))

Is the list empty?

Suppose we want to extend the standard library's isempty function, which works on a variety of different collections:

julia> methods(isempty)
# 29 methods for generic function "isempty":
isempty(v::SimpleVector) at essentials.jl:180
isempty(m::Base.MethodList) at reflection.jl:394

We can simply use the function dispatch syntax, and define two additional methods of isempty. Since this function is from the Base module, we have to qualify it as Base.isempty in order to extend it.

Base.isempty(::Nil) = true
Base.isempty(::Cons) = false

Here, we did not need the argument values at all to determine whether the list is empty. Merely the type alone suffices to compute that information. Julia allows us to omit the names of arguments, keeping only their type annotation, if we need not use their values.

We can test that our isempty methods work:

julia> using Base.Test

julia> @test isempty(Nil())
Test Passed
  Expression: isempty(Nil())

julia> @test !isempty(Cons(1, Cons(2, Cons(3, Nil()))))
Test Passed
  Expression: !(isempty(Cons(1,Cons(2,Cons(3,Nil())))))

and indeed the number of methods for isempty have increased by 2:

julia> methods(isempty)
# 31 methods for generic function "isempty":
isempty(v::SimpleVector) at essentials.jl:180
isempty(m::Base.MethodList) at reflection.jl:394

Clearly, determining whether a linked list is empty or not is a trivial example. But it leads up to something more interesting:

How long is the list?

The length function from the standard library gives us the length of a collection or certain iterables. There are many ways to implement length for a linked list. In particular, using a while loop is likely fastest and most memory-efficient in Julia. But premature optimization is to be avoided, so let's suppose for a second that our linked list need not be efficient. What's the simplest way to write a length function?

Base.length(::Nil) = 0
Base.length(xs::Cons) = 1 + length(xs.rest)

The first definition is straightforward: an empty list has length 0. The second definition is also easy to read: to count the length of a list, we count the first element, then count the length of the rest of the list. We can test this method similarly to how we tested isempty:

julia> @test length(Nil()) == 0
Test Passed
  Expression: length(Nil()) == 0
   Evaluated: 0 == 0

julia> @test length(Cons(1, Cons(2, Cons(3, Nil())))) == 3
Test Passed
  Expression: length(Cons(1,Cons(2,Cons(3,Nil())))) == 3
   Evaluated: 3 == 3

Next steps

This toy example is pretty far from implementing all of the functionality that would be desired in a linked list. It is missing, for instance, the iteration interface. However, it illustrates how dispatch can be used to write short and clear code.

Immutable Types

The simplest composite type is an immutable type. Instances of immutable types, like tuples, are values. Their fields cannot be changed after they are created. In many ways, an immutable type is like a Tuple with names for the type itself and for each field.

Singleton types

Composite types, by definition, contain a number of simpler types. In Julia, this number can be zero; that is, an immutable type is allowed to contain no fields. This is comparable to the empty tuple ().

Why might this be useful? Such immutable types are known as "singleton types", as only one instance of them could ever exist. The values of such types are known as "singleton values". The standard library Base contains many such singleton types. Here is a brief list:

  • Void, the type of nothing. We can verify that Void.instance (which is special syntax for retrieving the singleton value of a singleton type) is indeed nothing.
  • Any media type, such as MIME"text/plain", is a singleton type with a single instance, MIME("text/plain").
  • The Irrational{:π}, Irrational{:e}, Irrational{:φ}, and similar types are singleton types, and their singleton instances are the irrational values π = 3.1415926535897..., etc.
  • The iterator size traits Base.HasLength, Base.HasShape, Base.IsInfinite, and Base.SizeUnknown are all singleton types.
  • In version 0.5 and later, each function is a singleton instance of a singleton type! Like any other singleton value, we can recover the function sin, for example, from typeof(sin).instance.

Because they contain nothing, singleton types are incredibly lightweight, and they can frequently be optimized away by the compiler to have no runtime overhead. Thus, they are perfect for traits, special tag values, and for things like functions that one would like to specialize on.

To define a singleton type,

julia> immutable MySingleton end

To define custom printing for the singleton type,

julia> Base.show(io::IO, ::MySingleton) = print(io, "sing")

To access the singleton instance,

julia> MySingleton.instance

Often, one assigns this to a constant:

julia> const sing = MySingleton.instance

Wrapper types

If zero-field immutable types are interesting and useful, then perhaps one-field immutable types are even more useful. Such types are commonly called "wrapper types" because they wrap some underlying data, providing an alternative interface to said data. An example of a wrapper type in Base is String. We will define a similar type to String, named MyString. This type will be backed by a vector (one-dimensional array) of bytes (UInt8).

First, the type definition itself and some customized showing:

immutable MyString <: AbstractString

function Base.show(io::IO, s::MyString)
    print(io, "MyString: ")
    write(io, s.data)

Now our MyString type is ready for use! We can feed it some raw UTF-8 data, and it displays as we like it to:

julia> MyString([0x48,0x65,0x6c,0x6c,0x6f,0x2c,0x20,0x57,0x6f,0x72,0x6c,0x64,0x21])
MyString: Hello, World!

Obviously, this string type needs a lot of work before it becomes as usable as the Base.String type.

True composite types

Perhaps most commonly, many immutable types contain more than one field. An example is the standard library Rational{T} type, which contains two fieds: a num field for the numerator, and a den field for the denominator. It is fairly straightforward to emulate this type design:

immutable MyRational{T}
    MyRational(n, d) = (g = gcd(n, d); new(n÷g, d÷g))
MyRational{T}(n::T, d::T) = MyRational{T}(n, d)

We have successfully implemented a constructor that simplifies our rational numbers:

julia> MyRational(10, 6)