Modules

Modules

Modules in Julia help organize code into coherent units. They are delimited syntactically inside module NameOfModule ... end, and have the following features:

Modules are separate namespaces, each introducing a new global scope. This is useful, because it allows the same name to be used for different functions or global variables without conflict, as long as they are in separate modules.
Modules have facilities for detailed namespace management: each defines a set of names it exports, and can import names from other modules with using and import (we explain these below).
Modules can be precompiled for faster loading, and may contain code for runtime initialization.

Typically, in larger Julia packages you will see module code organized into files, eg

module SomeModule
# export, using, import statements are usually here; we discuss these below
include("file1.jl")
include("file2.jl")
end

Files and file names are mostly unrelated to modules; modules are associated only with module expressions. One can have multiple files per module, and multiple modules per file. include behaves as if the contents of the source file were evaluated in the global scope of the including module. In this chapter, we use short and simplified examples, so we won’t use include.

The recommended style is not to indent the body of the module, since that would typically lead to whole files being indented. Also, it is common to use UpperCamelCase for module names (just like types), and use the plural form if applicable, especially if the module contains a similarly named identifier, to avoid name clashes. For example,

module FastThings
struct FastThing
    ...
end
end

Namespace management

Namespace management refers to the facilities the language offers for making names in a module available in other modules. We discuss the related concepts and functionality below in detail.

Qualified names

Names for functions, variables and types in the global scope like sin, ARGS, and UnitRange always belong to a module, called the parent module, which can be found interactively with parentmodule, for example

julia> parentmodule(UnitRange)
Base

One can also refer to these names outside their parent module by prefixing them with their module, eg Base.UnitRange. This is called a qualified name. The parent module may be accessible using a chain of submodules like Base.Math.sin, where Base.Math is called the module path. Due to syntactic ambiguities, qualifying a name that contains only symbols, such as an operator, requires inserting a colon, e.g. Base.:+. A small number of operators additionally require parentheses, e.g. Base.:(==).

If a name is qualified, then it is always accessible, and in case of a function, it can also have methods added to it by using the qualified name as the function name.

Within a module, a variable name can be “reserved” without assigning to it by declaring it as global x. This prevents name conflicts for globals initialized after load time. The syntax M.x = y does not work to assign a global in another module; global assignment is always module-local.

Export lists

Names (referring to functions, types, global variables, and constants) can be added to the export list of a module with export: these are the symbols that are imported when using the module. Typically, they are at or near the top of the module definition so that readers of the source code can find them easily, as in

julia> module NiceStuff
       export nice, DOG
       struct Dog end      # singleton type, not exported
       const DOG = Dog()   # named instance, exported
       nice(x) = "nice $x" # function, exported
       end;

but this is just a style suggestion — a module can have multiple export statements in arbitrary locations.

It is common to export names which form part of the API (application programming interface). In the above code, the export list suggests that users should use nice and DOG. However, since qualified names always make identifiers accessible, this is just an option for organizing APIs: unlike other languages, Julia has no facilities for truly hiding module internals.

Also, some modules don’t export names at all. This is usually done if they use common words, such as derivative, in their API, which could easily clash with the export lists of other modules. We will see how to manage name clashes below.

Standalone using and import

Possibly the most common way of loading a module is using ModuleName. This loads the code associated with ModuleName, and brings

the module name
and the elements of the export list into the surrounding global namespace.

Technically, the statement using ModuleName means that a module called ModuleName will be available for resolving names as needed. When a global variable is encountered that has no definition in the current module, the system will search for it among variables exported by ModuleName and use it if it is found there. This means that all uses of that global within the current module will resolve to the definition of that variable in ModuleName.

To load a module from a package, the statement using ModuleName can be used. To load a module from a locally defined module, a dot needs to be added before the module name like using .ModuleName.

To continue with our example,

julia> using .NiceStuff

would load the above code, making NiceStuff (the module name), DOG and nice available. Dog is not on the export list, but it can be accessed if the name is qualified with the module path (which here is just the module name) as NiceStuff.Dog.

Importantly, using ModuleName is the only form for which export lists matter at all.

In contrast,

julia> import .NiceStuff

brings only the module name into scope. Users would need to use NiceStuff.DOG, NiceStuff.Dog, and NiceStuff.nice to access its contents. Usually, import ModuleName is used in contexts when the user wants to keep the namespace clean. As we will see in the next section import .NiceStuff is equivalent to using .NiceStuff: NiceStuff.

You can combine multiple using and import statements of the same kind in a comma-separated expression, e.g.

julia> using LinearAlgebra, Statistics

using and import with specific identifiers, and adding methods

When using ModuleName: or import ModuleName: is followed by a comma-separated list of names, the module is loaded, but only those specific names are brought into the namespace by the statement. For example,

julia> using .NiceStuff: nice, DOG

will import the names nice and DOG.

Importantly, the module name NiceStuff will not be in the namespace. If you want to make it accessible, you have to list it explicitly, as

julia> using .NiceStuff: nice, DOG, NiceStuff

Julia has two forms for seemingly the same thing because only import ModuleName: f allows adding methods to f without a module path. That is to say, the following example will give an error:

julia> using .NiceStuff: nice
julia> struct Cat end
julia> nice(::Cat) = "nice 😸"
ERROR: invalid method definition in Main: function NiceStuff.nice must be explicitly imported to be extended
Stacktrace:
 [1] top-level scope
   @ none:0
 [2] top-level scope
   @ none:1

This error prevents accidentally adding methods to functions in other modules that you only intended to use.

There are two ways to deal with this. You can always qualify function names with a module path:

julia> using .NiceStuff
julia> struct Cat end
julia> NiceStuff.nice(::Cat) = "nice 😸"

Alternatively, you can import the specific function name:

julia> import .NiceStuff: nice
julia> struct Cat end
julia> nice(::Cat) = "nice 😸"
nice (generic function with 2 methods)

Which one you choose is a matter of style. The first form makes it clear that you are adding a method to a function in another module (remember, that the imports and the method definition may be in separate files), while the second one is shorter, which is especially convenient if you are defining multiple methods.

Once a variable is made visible via using or import, a module may not create its own variable with the same name. Imported variables are read-only; assigning to a global variable always affects a variable owned by the current module, or else raises an error.

Renaming with as

An identifier brought into scope by import or using can be renamed with the keyword as. This is useful for working around name conflicts as well as for shortening names. For example, Base exports the function name read, but the CSV.jl package also provides CSV.read. If we are going to invoke CSV reading many times, it would be convenient to drop the CSV. qualifier. But then it is ambiguous whether we are referring to Base.read or CSV.read:

julia> read;
julia> import CSV: read
WARNING: ignoring conflicting import of CSV.read into Main

Renaming provides a solution:

julia> import CSV: read as rd

Imported packages themselves can also be renamed:

import BenchmarkTools as BT

as works with using only when a single identifier is brought into scope. For example using CSV: read as rd works, but using CSV as C does not, since it operates on all of the exported names in CSV.

Mixing multiple using and import statements

When multiple using or import statements of any of the forms above are used, their effect is combined in the order they appear. For example,

julia> using .NiceStuff         # exported names and the module name
julia> import .NiceStuff: nice  # allows adding methods to unqualified functions

would bring all the exported names of NiceStuff and the module name itself into scope, and also allow adding methods to nice without prefixing it with a module name.

Handling name conflicts

Consider the situation where two (or more) packages export the same name, as in

julia> module A
       export f
       f() = 1
       end
A
julia> module B
       export f
       f() = 2
       end
B

The statement using .A, .B works, but when you try to call f, you get a warning

julia> using .A, .B
julia> f
WARNING: both B and A export "f"; uses of it in module Main must be qualified
ERROR: UndefVarError: `f` not defined

Here, Julia cannot decide which f you are referring to, so you have to make a choice. The following solutions are commonly used:

Simply proceed with qualified names like A.f and B.f. This makes the context clear to the reader of your code, especially if f just happens to coincide but has different meaning in various packages. For example, degree has various uses in mathematics, the natural sciences, and in everyday life, and these meanings should be kept separate.
Use the as keyword above to rename one or both identifiers, eg
```
julia> using .A: f as f
julia> using .B: f as g
```
would make B.f available as g. Here, we are assuming that you did not use using A before, which would have brought f into the namespace.
When the names in question do share a meaning, it is common for one module to import it from another, or have a lightweight “base” package with the sole function of defining an interface like this, which can be used by other packages. It is conventional to have such package names end in ...Base (which has nothing to do with Julia’s Base module).

Default top-level definitions and bare modules

Modules automatically contain using Core, using Base, and definitions of the eval and include functions, which evaluate expressions/files within the global scope of that module.

If these default definitions are not wanted, modules can be defined using the keyword baremodule instead (note: Core is still imported). In terms of baremodule, a standard module looks like this:

baremodule Mod
using Base
eval(x) = Core.eval(Mod, x)
include(p) = Base.include(Mod, p)
...
end

If even Core is not wanted, a module that imports nothing and defines no names at all can be defined with Module(:YourNameHere, false, false) and code can be evaluated into it with @eval or Core.eval:

julia> arithmetic = Module(:arithmetic, false, false)
Main.arithmetic
julia> @eval arithmetic add(x, y) = $(+)(x, y)
add (generic function with 1 method)
julia> arithmetic.add(12, 13)
25

Standard modules

There are three important standard modules:

Core contains all functionality “built into” the language.
Base contains basic functionality that is useful in almost all cases.
Main is the top-level module and the current module, when Julia is started.

Standard library modules

By default Julia ships with some standard library modules. These behave like regular Julia packages except that you don’t need to install them explicitly. For example, if you wanted to perform some unit testing, you could load the Test standard library as follows:

using Test

Submodules and relative paths

Modules can contain submodules, nesting the same syntax module ... end. They can be used to introduce separate namespaces, which can be helpful for organizing complex codebases. Note that each module introduces its own scope, so submodules do not automatically “inherit” names from their parent.

It is recommended that submodules refer to other modules within the enclosing parent module (including the latter) using relative module qualifiers in using and import statements. A relative module qualifier starts with a period (.), which corresponds to the current module, and each successive . leads to the parent of the current module. This should be followed by modules if necessary, and eventually the actual name to access, all separated by .s.

Consider the following example, where the submodule SubA defines a function, which is then extended in its “sibling” module:

julia> module ParentModule
       module SubA
       export add_D  # exported interface
       const D = 3
       add_D(x) = x + D
       end
       using .SubA  # brings `add_D` into the namespace
       export add_D # export it from ParentModule too
       module SubB
       import ..SubA: add_D # relative path for a “sibling” module
       struct Infinity end
       add_D(x::Infinity) = x
       end
       end;

You may see code in packages, which, in a similar situation, uses

julia> import .ParentModule.SubA: add_D

However, this operates through code loading, and thus only works if ParentModule is in a package. It is better to use relative paths.

Note that the order of definitions also matters if you are evaluating values. Consider

module TestPackage
export x, y
x = 0
module Sub
using ..TestPackage
z = y # ERROR: UndefVarError: `y` not defined
end
y = 1
end

where Sub is trying to use TestPackage.y before it was defined, so it does not have a value.

For similar reasons, you cannot use a cyclic ordering:

module A
module B
using ..C # ERROR: UndefVarError: `C` not defined
end
module C
using ..B
end
end

Module initialization and precompilation

Large modules can take several seconds to load because executing all of the statements in a module often involves compiling a large amount of code. Julia creates precompiled caches of the module to reduce this time.

Precompiled module files (sometimes called “cache files”) are created and used automatically when import or using loads a module. If the cache file(s) do not yet exist, the module will be compiled and saved for future reuse. You can also manually call Base.compilecache(Base.identify_package(“modulename”)) to create these files without loading the module. The resulting cache files will be stored in the compiled subfolder of DEPOT_PATH[1]. If nothing about your system changes, such cache files will be used when you load the module with import or using.

Precompilation cache files store definitions of modules, types, methods, and constants. They may also store method specializations and the code generated for them, but this typically requires that the developer add explicit precompile directives or execute workloads that force compilation during the package build.

However, if you update the module’s dependencies or change its source code, the module is automatically recompiled upon using or import. Dependencies are modules it imports, the Julia build, files it includes, or explicit dependencies declared by include_dependency(path) in the module file(s).

For file dependencies, a change is determined by examining whether the modification time (mtime) of each file loaded by include or added explicitly by include_dependency is unchanged, or equal to the modification time truncated to the nearest second (to accommodate systems that can’t copy mtime with sub-second accuracy). It also takes into account whether the path to the file chosen by the search logic in require matches the path that had created the precompile file. It also takes into account the set of dependencies already loaded into the current process and won’t recompile those modules, even if their files change or disappear, in order to avoid creating incompatibilities between the running system and the precompile cache. Finally, it takes account of changes in any compile-time preferences.

If you know that a module is not safe to precompile (for example, for one of the reasons described below), you should put __precompile__(false) in the module file (typically placed at the top). This will cause Base.compilecache to throw an error, and will cause using / import to load it directly into the current process and skip the precompile and caching. This also thereby prevents the module from being imported by any other precompiled module.

You may need to be aware of certain behaviors inherent in the creation of incremental shared libraries which may require care when writing your module. For example, external state is not preserved. To accommodate this, explicitly separate any initialization steps that must occur at runtime from steps that can occur at compile time. For this purpose, Julia allows you to define an __init__() function in your module that executes any initialization steps that must occur at runtime. This function will not be called during compilation (--output-*). Effectively, you can assume it will be run exactly once in the lifetime of the code. You may, of course, call it manually if necessary, but the default is to assume this function deals with computing state for the local machine, which does not need to be – or even should not be – captured in the compiled image. It will be called after the module is loaded into a process, including if it is being loaded into an incremental compile (--output-incremental=yes), but not if it is being loaded into a full-compilation process.

In particular, if you define a function __init__() in a module, then Julia will call __init__() immediately after the module is loaded (e.g., by import, using, or require) at runtime for the first time (i.e., __init__ is only called once, and only after all statements in the module have been executed). Because it is called after the module is fully imported, any submodules or other imported modules have their __init__ functions called before the __init__ of the enclosing module.

Two typical uses of __init__ are calling runtime initialization functions of external C libraries and initializing global constants that involve pointers returned by external libraries. For example, suppose that we are calling a C library libfoo that requires us to call a foo_init() initialization function at runtime. Suppose that we also want to define a global constant foo_data_ptr that holds the return value of a void *foo_data() function defined by libfoo – this constant must be initialized at runtime (not at compile time) because the pointer address will change from run to run. You could accomplish this by defining the following __init__ function in your module:

const foo_data_ptr = Ref{Ptr{Cvoid}}(0)
function __init__()
    ccall((:foo_init, :libfoo), Cvoid, ())
    foo_data_ptr[] = ccall((:foo_data, :libfoo), Ptr{Cvoid}, ())
    nothing
end

Notice that it is perfectly possible to define a global inside a function like __init__; this is one of the advantages of using a dynamic language. But by making it a constant at global scope, we can ensure that the type is known to the compiler and allow it to generate better optimized code. Obviously, any other globals in your module that depends on foo_data_ptr would also have to be initialized in __init__.

Constants involving most Julia objects that are not produced by ccall do not need to be placed in __init__: their definitions can be precompiled and loaded from the cached module image. This includes complicated heap-allocated objects like arrays. However, any routine that returns a raw pointer value must be called at runtime for precompilation to work (Ptr objects will turn into null pointers unless they are hidden inside an isbits object). This includes the return values of the Julia functions @cfunction and pointer.

Dictionary and set types, or in general anything that depends on the output of a hash(key) method, are a trickier case. In the common case where the keys are numbers, strings, symbols, ranges, Expr, or compositions of these types (via arrays, tuples, sets, pairs, etc.) they are safe to precompile. However, for a few other key types, such as Function or DataType and generic user-defined types where you haven’t defined a hash method, the fallback hash method depends on the memory address of the object (via its objectid) and hence may change from run to run. If you have one of these key types, or if you aren’t sure, to be safe you can initialize this dictionary from within your __init__ function. Alternatively, you can use the IdDict dictionary type, which is specially handled by precompilation so that it is safe to initialize at compile-time.

When using precompilation, it is important to keep a clear sense of the distinction between the compilation phase and the execution phase. In this mode, it will often be much more clearly apparent that Julia is a compiler which allows execution of arbitrary Julia code, not a standalone interpreter that also generates compiled code.

Other known potential failure scenarios include:

Global counters (for example, for attempting to uniquely identify objects). Consider the following code snippet:
```
mutable struct UniquedById
    myid::Int
    let counter = 0
        UniquedById() = new(counter += 1)
    end
end
```
while the intent of this code was to give every instance a unique id, the counter value is recorded at the end of compilation. All subsequent usages of this incrementally compiled module will start from that same counter value.

Note that objectid (which works by hashing the memory pointer) has similar issues (see notes on Dict usage below).

One alternative is to use a macro to capture @__MODULE__ and store it alone with the current counter value, however, it may be better to redesign the code to not depend on this global state.
Associative collections (such as Dict and Set) need to be re-hashed in __init__. (In the future, a mechanism may be provided to register an initializer function.)
Depending on compile-time side-effects persisting through load-time. Example include: modifying arrays or other variables in other Julia modules; maintaining handles to open files or devices; storing pointers to other system resources (including memory);

Creating accidental “copies” of global state from another module, by referencing it directly instead of via its lookup path. For example, (in global scope):

#mystdout = Base.stdout #= will not work correctly, since this will copy Base.stdout into this module =#
# instead use accessor functions:
getstdout() = Base.stdout #= best option =#
# or move the assignment into the runtime:
__init__() = global mystdout = Base.stdout #= also works =#

Several additional restrictions are placed on the operations that can be done while precompiling code to help the user avoid other wrong-behavior situations:

Calling eval to cause a side-effect in another module. This will also cause a warning to be emitted when the incremental precompile flag is set.
global const statements from local scope after __init__() has been started (see issue #12010 for plans to add an error for this)
Replacing a module is a runtime error while doing an incremental precompile.

A few other points to be aware of:

No code reload / cache invalidation is performed after changes are made to the source files themselves, (including by Pkg.update), and no cleanup is done after Pkg.rm
The memory sharing behavior of a reshaped array is disregarded by precompilation (each view gets its own copy)
Expecting the filesystem to be unchanged between compile-time and runtime e.g. @__FILE__/source_path() to find resources at runtime, or the BinDeps @checked_lib macro. Sometimes this is unavoidable. However, when possible, it can be good practice to copy resources into the module at compile-time so they won’t need to be found at runtime.
WeakRef objects and finalizers are not currently handled properly by the serializer (this will be fixed in an upcoming release).
It is usually best to avoid capturing references to instances of internal metadata objects such as Method, MethodInstance, MethodTable, TypeMapLevel, TypeMapEntry and fields of those objects, as this can confuse the serializer and may not lead to the outcome you desire. It is not necessarily an error to do this, but you simply need to be prepared that the system will try to copy some of these and to create a single unique instance of others.

It is sometimes helpful during module development to turn off incremental precompilation. The command line flag --compiled-modules={yes|no} enables you to toggle module precompilation on and off. When Julia is started with --compiled-modules=no the serialized modules in the compile cache are ignored when loading modules and module dependencies. More fine-grained control is available with --pkgimages=no, which suppresses only native-code storage during precompilation. Base.compilecache can still be called manually. The state of this command line flag is passed to Pkg.build to disable automatic precompilation triggering when installing, updating, and explicitly building packages.

You can also debug some precompilation failures with environment variables. Setting JULIA_VERBOSE_LINKING=true may help resolve failures in linking shared libraries of compiled native code. See the Developer Documentation part of the Julia manual, where you will find further details in the section documenting Julia’s internals under “Package Images”.