- Overview
- Usage
- Creating a Package
- Defining Dependencies
- Publishing a Package
- Requiring System Libraries
- Packaging Legacy Code
- Handling Version-specific Logic
- Editing a Package
- Resolving Versions (Package.resolved file)
- Setting the Swift Tools Version
- Testing
- Running
- Setting the Build Configuration
- Depending on Apple Modules
- Creating C Language Targets
- Using Shell Completion Scripts
- Package manifest specification
- Packages and continuous integration
Simply put: a package is a git repository with semantically versioned tags,
that contains Swift sources and a Package.swift
manifest file at its root.
A library package contains code which other packages can use and depend on. To
get started, create a directory and run swift package init
:
$ mkdir MyPackage
$ cd MyPackage
$ swift package init # or swift package init --type library
$ swift build
$ swift test
This will create the directory structure needed for a library package with a target and the corresponding test target to write unit tests. A library package can contain multiple targets as explained in Target Format Reference.
SwiftPM can create native binaries which can be executed from the command line. To get started:
$ mkdir MyExecutable
$ cd MyExecutable
$ swift package init --type executable
$ swift build
$ swift run
Hello, World!
This creates the directory structure needed for executable targets. Any target
can be turned into a executable target if there is a main.swift
file present in
its sources. The complete reference for layout is
here.
SwiftPM can generate boilerplate for custom macros:
$ mkdir MyMacro
$ cd MyMacro
$ swift package init --type macro
$ swift build
$ swift run
The value 42 was produced by the code "a + b"
This creates a package with a .macro
type target with its required dependencies
on swift-syntax, a library .target
containing the macro's code, and an .executableTarget
and .testTarget
for
running the macro. The sample macro, StringifyMacro
, is documented in the Swift
Evolution proposal for Expression Macros
and the WWDC Write Swift macros
video. See further documentation on macros in The Swift Programming Language book.
To depend on a package, define the dependency and the version in the manifest of
your package, and add a product from that package as a dependency, e.g., if
you want to use https://github.com/apple/example-package-playingcard as
a dependency, add the GitHub URL in the dependencies of Package.swift
:
import PackageDescription
let package = Package(
name: "MyPackage",
dependencies: [
.package(url: "https://github.com/apple/example-package-playingcard.git", from: "3.0.4"),
],
targets: [
.target(
name: "MyPackage",
dependencies: ["PlayingCard"]
),
.testTarget(
name: "MyPackageTests",
dependencies: ["MyPackage"]
),
]
)
Now you should be able to import PlayingCard
in the MyPackage
target.
To publish a package, create and push a semantic version tag:
$ git init
$ git add .
$ git remote add origin [github-URL]
$ git commit -m "Initial Commit"
$ git tag 1.0.0
$ git push origin master --tags
Now other packages can depend on version 1.0.0 of this package using the github url. An example of a published package can be found here: https://github.com/apple/example-package-fisheryates
You can link against system libraries using the package manager. To do so, you'll
need to add a special target
of type .systemLibrary
, and a module.modulemap
for each system library you're using.
Let's see an example of adding libgit2 as a dependency to an executable target.
Create a directory called example
, and initialize it as a package that
builds an executable:
$ mkdir example
$ cd example
example$ swift package init --type executable
Edit the Sources/example/main.swift
so it consists of this code:
import Clibgit
let options = git_repository_init_options()
print(options)
To import Clibgit
, the package manager requires that the libgit2 library has
been installed by a system packager (eg. apt
, brew
, yum
, nuget
, etc.). The
following files from the libgit2 system-package are of interest:
/usr/local/lib/libgit2.dylib # .so on Linux
/usr/local/include/git2.h
Note: the system library may be located elsewhere on your system, such as:
-
/usr/
, or/opt/homebrew/
if you're using Homebrew on an Apple Silicon Mac. -
C:\vcpkg\installed\x64-windows\include
on Windows, if you're usingvcpkg
. On most Unix-like systems, you can usepkg-config
to lookup where a library is installed:example$ pkg-config --cflags libgit2 -I/usr/local/libgit2/1.6.4/include
First, let's define the target
in the package description:
// swift-tools-version: 5.8
// The swift-tools-version declares the minimum version of Swift required to build this package.
import PackageDescription
let package = Package(
name: "example",
targets: [
// systemLibrary is a special type of build target that wraps a system library
// in a target that other targets can require as their dependency.
.systemLibrary(
name: "Clibgit",
pkgConfig: "libgit2",
providers: [
.brew(["libgit2"]),
.apt(["libgit2-dev"])
]
)
]
)
Note: For Windows-only packages pkgConfig
should be omitted as
pkg-config
is not expected to be available. If you don't want to use the
pkgConfig
parameter you can pass the path of a directory containing the
library using the -L
flag in the command line when building your package
instead.
example$ swift build -Xlinker -L/usr/local/lib/
Next, create a directory Sources/Clibgit
in your example
project, and
add a module.modulemap
and the header file to it:
module Clibgit [system] {
header "git2.h"
link "git2"
export *
}
The header file should look like this:
// git2.h
#pragma once
#include <git2.h>
Note: Avoiding specifying an absolute path to git2.h
provided
by the library in the module.modulemap
. Doing so will break compatibility of
your project between machines that may use a different file system layout or
install libraries to different paths.
The convention we hope the community will adopt is to prefix such modules with
C
and to camelcase the modules as per Swift module name conventions. Then the community is free to name another module simplylibgit
which contains more “Swifty” function wrappers around the raw C interface.
The example
directory structure should look like this now:
.
├── Package.swift
└── Sources
├── Clibgit
│ ├── git2.h
│ └── module.modulemap
└── main.swift
At this point, your system library target is fully defined, and you can now use
that target as a dependency in other targets in your Package.swift
, like this:
import PackageDescription
let package = Package(
name: "example",
targets: [
.executableTarget(
name: "example",
// example executable requires "Clibgit" target as its dependency.
// It's a systemLibrary target defined below.
dependencies: ["Clibgit"],
path: "Sources"
),
// systemLibrary is a special type of build target that wraps a system library
// in a target that other targets can require as their dependency.
.systemLibrary(
name: "Clibgit",
pkgConfig: "libgit2",
providers: [
.brew(["libgit2"]),
.apt(["libgit2-dev"])
]
)
]
)
Now if we type swift build
in our example app directory we will create an
executable:
example$ swift build
…
example$ .build/debug/example
git_repository_init_options(version: 0, flags: 0, mode: 0, workdir_path: nil, description: nil, template_path: nil, initial_head: nil, origin_url: nil)
example$
Let’s see another example of using IJG’s JPEG library from an executable, which has some caveats.
Create a directory called example
, and initialize it as a package that builds
an executable:
$ mkdir example
$ cd example
example$ swift package init --type executable
Edit the Sources/main.swift
so it consists of this code:
import CJPEG
let jpegData = jpeg_common_struct()
print(jpegData)
Install the JPEG library, on macOS you can use Homebrew package manager: brew install jpeg
.
jpeg
is a keg-only formula, meaning it won't be linked to /usr/local/lib
,
and you'll have to link it manually at build time.
Just like in the previous example, run mkdir Sources/CJPEG
and add the
following module.modulemap
:
module CJPEG [system] {
header "shim.h"
header "/usr/local/opt/jpeg/include/jpeglib.h"
link "jpeg"
export *
}
Create a shim.h
file in the same directory and add #include <stdio.h>
in
it.
$ echo '#include <stdio.h>' > shim.h
This is because jpeglib.h
is not a correct module, that is, it does not contain
the required line #include <stdio.h>
. Alternatively, you can add #include <stdio.h>
to the top of jpeglib.h to avoid creating the shim.h
file.
Now to use the CJPEG package we must declare our dependency in our example
app’s Package.swift
:
import PackageDescription
let package = Package(
name: "example",
targets: [
.executableTarget(
name: "example",
dependencies: ["CJPEG"],
path: "Sources"
),
.systemLibrary(
name: "CJPEG",
providers: [
.brew(["jpeg"])
])
]
)
Now if we type swift build
in our example app directory we will create an
executable:
example$ swift build -Xlinker -L/usr/local/jpeg/lib
…
example$ .build/debug/example
jpeg_common_struct(err: nil, mem: nil, progress: nil, client_data: nil, is_decompressor: 0, global_state: 0)
example$
We have to specify the path where the libjpeg is present using -Xlinker
because
there is no pkg-config file for it. We plan to provide a solution to avoid passing
the flag in the command line.
Some system packages provide multiple libraries (.so
and .dylib
files). In
such cases you should add all the libraries to that Swift modulemap package’s
.modulemap
file:
module CFoo [system] {
header "/usr/local/include/foo/foo.h"
link "foo"
export *
}
module CFooBar [system] {
header "/usr/include/foo/bar.h"
link "foobar"
export *
}
module CFooBaz [system] {
header "/usr/include/foo/baz.h"
link "foobaz"
export *
}
foobar
and foobaz
link to foo
; we don’t need to specify this information
in the module-map because the headers foo/bar.h
and foo/baz.h
both include
foo/foo.h
. It is very important however that those headers do include their
dependent headers, otherwise when the modules are imported into Swift the
dependent modules will not get imported automatically and link errors will
happen. If these link errors occur for consumers of a package that consumes your
package, the link errors can be especially difficult to debug.
Module maps must contain absolute paths, thus they are not cross-platform. We intend to provide a solution for this in the package manager. In the long term, we hope that system libraries and system packagers will provide module maps and thus this component of the package manager will become redundant.
Notably the above steps will not work if you installed JPEG and JasPer with
Homebrew since the files will be installed to /usr/local
on
Intel Macs, or /opt/homebrew on Apple silicon Macs. For now adapt the paths,
but as said, we plan to support basic relocations like these.
Version the module maps semantically. The meaning of semantic version is less clear here, so use your best judgement. Do not follow the version of the system library the module map represents; version the module map(s) independently.
Follow the conventions of system packagers; for example, the debian package for
python3 is called python3, as there is not a single package for python and
python is designed to be installed side-by-side. Were you to make a module map
for python3 you should name it CPython3
.
At this time you will need to make another module map package to represent system packages that are built with optional dependencies.
For example, libarchive
optionally depends on xz
, which means it can be
compiled with xz
support, but it is not required. To provide a package that
uses libarchive with xz you must make a CArchive+CXz
package that depends on
CXz
and provides CArchive
.
You may be working with code that builds both as a package and not. For example, you may be packaging a project that also builds with Xcode.
In these cases, you can use the preprocessor definition SWIFT_PACKAGE
to
conditionally compile code for Swift packages.
In your source file:
#if SWIFT_PACKAGE
import Foundation
#endif
The package manager is designed to support packages which work with a variety of Swift project versions, including both the language and the package manager version.
In most cases, if you want to support multiple Swift versions in a package you should do so by using the language-specific version checks available in the source code itself. However, in some circumstances this may become unmanageable, specifically, when the package manifest itself cannot be written to be Swift version agnostic (for example, because it optionally adopts new package manager features not present in older versions).
The package manager has support for a mechanism to allow Swift version-specific customizations for the both package manifest and the package versions which will be considered.
The tags which define the versions of the package available for clients to use
can optionally be suffixed with a marker in the form of @swift-3
. When the
package manager is determining the available tags for a repository, if
a version-specific marker is available which matches the current tool version,
then it will only consider the versions which have the version-specific
marker. Conversely, version-specific tags will be ignored by any non-matching
tool version.
For example, suppose the package Foo
has the tags [1.0.0, 1.2.0@swift-3, 1.3.0]
. If version 3.0 of the package manager is evaluating the available
versions for this repository, it will only ever consider version 1.2.0
.
However, version 4.0 would consider only 1.0.0
and 1.3.0
.
This feature is intended for use in the following scenarios:
-
A package wishes to maintain support for Swift 3.0 in older versions, but newer versions of the package require Swift 4.0 for the manifest to be readable. Since Swift 3.0 will not know to ignore those versions, it would fail when performing dependency resolution on the package if no action is taken. In this case, the author can re-tag the last versions which supported Swift 3.0 appropriately.
-
A package wishes to maintain dual support for Swift 3.0 and Swift 4.0 at the same version numbers, but this requires substantial differences in the code. In this case, the author can maintain parallel tag sets for both versions.
It is not expected that the packages would ever use this feature unless absolutely necessary to support existing clients. Specifically, packages should not adopt this syntax for tagging versions supporting the latest GM Swift version.
The package manager supports looking for any of the following marked tags, in order of preference:
MAJOR.MINOR.PATCH
(e.g.,[email protected]
)MAJOR.MINOR
(e.g.,[email protected]
)MAJOR
(e.g.,1.2.0@swift-3
)
The package manager will additionally look for a version-specific marked
manifest version when loading the particular version of a package, by searching
for a manifest in the form of [email protected]
. The set of markers
looked for is the same as for version-specific tag selection.
This feature is intended for use in cases where a package wishes to maintain compatibility with multiple Swift project versions, but requires a substantively different manifest file for this to be viable (e.g., due to changes in the manifest API).
It is not expected the packages would ever use this feature unless absolutely necessary to support existing clients. Specifically, packages should not adopt this syntax for tagging versions supporting the latest GM Swift version.
In case the current Swift version doesn't match any version-specific manifest, the package manager will pick the manifest with the most compatible tools version. For example, if there are three manifests:
Package.swift
(tools version 3.0)
[email protected]
(tools version 4.0)
[email protected]
(tools version 4.2)
The package manager will pick Package.swift
on Swift 3, [email protected]
on
Swift 4, and [email protected]
on Swift 4.2 and above because its tools
version will be most compatible with future version of the package manager.
Swift package manager supports editing dependencies, when your work requires
making a change to one of your dependencies (for example, to fix a bug, or add
a new API). The package manager moves the dependency into a location under the
Packages/
directory where it can be edited.
For the packages which are in the editable state, swift build
will always use
the exact sources in this directory to build, regardless of their state, Git
repository status, tags, or the tag desired by dependency resolution. In other
words, this will just build against the sources that are present. When an
editable package is present, it will be used to satisfy all instances of that
package in the dependency graph. It is possible to edit all, some, or none of
the packages in a dependency graph, without restriction.
Editable packages are best used to do experimentation with dependency code, or to create and submit a patch in the dependency owner's repository (upstream). There are two ways to put a package in editable state:
$ swift package edit Foo --branch bugFix
This will create a branch called bugFix
from the currently resolved version and
put the dependency Foo
in the Packages/
directory.
$ swift package edit Foo --revision 969c6a9
This is similar to the previous version, except that the Package Manager will leave the dependency at a detached HEAD on the specified revision.
Note: If the branch or revision option is not provided, the Package Manager will checkout the currently resolved version on a detached HEAD.
Once a package is in an editable state, you can navigate to the directory
Packages/Foo
to make changes, build and then push the changes or open a pull
request to the upstream repository.
You can end editing a package using unedit
command:
$ swift package unedit Foo
This will remove the edited dependency from Packages/
and put the originally
resolved version back.
This command fails if there are uncommitted changes or changes which are not
pushed to the remote repository. If you want to discard these changes and
unedit, you can use the --force
option:
$ swift package unedit Foo --force
This feature allows overriding a dependency with a local checkout on the filesystem. This checkout is completely unmanaged by the package manager and will be used as-is. The only requirement is that the package name in the overridden checkout should not change. This is extremely useful when developing multiple packages in tandem or when working on packages alongside an application.
The command to attach (or create) a local checkout is:
$ swift package edit <package name> --path <path/to/dependency>
For example, if Foo
depends on Bar
and you have a checkout of Bar
at
/workspace/bar
:
foo$ swift package edit Bar --path /workspace/bar
A checkout of Bar
will be created if it doesn't exist at the given path. If
a checkout exists, package manager will validate the package name at the given
path and attach to it.
The package manager will also create a symlink in the Packages/
directory to the
checkout path.
Use unedit command to stop using the local checkout:
$ swift package unedit <package name>
# Example:
$ swift package unedit Bar
The package manager records the result of dependency resolution in a
Package.resolved
file in the top-level of the package, and when this file is
already present in the top-level, it is used when performing dependency
resolution, rather than the package manager finding the latest eligible version
of each package. Running swift package update
updates all dependencies to the
latest eligible versions and updates the Package.resolved
file accordingly.
Resolved versions will always be recorded by the package manager. Some users may
choose to add the Package.resolved file to their package's .gitignore file. When
this file is checked in, it allows a team to coordinate on what versions of the
dependencies they should use. If this file is gitignored, each user will
separately choose when to get new versions based on when they run the swift package update
command, and new users will start with the latest eligible
version of each dependency. Either way, for a package which is a dependency of
other packages (e.g., a library package), that package's Package.resolved
file
will not have any effect on its client packages.
The swift package resolve
command resolves the dependencies, taking into
account the current version restrictions in the Package.swift
manifest and
Package.resolved
resolved versions file, and issuing an error if the graph
cannot be resolved. For packages which have previously resolved versions
recorded in the Package.resolved
file, the resolve command will resolve to
those versions as long as they are still eligible. If the resolved version's file
changes (e.g., because a teammate pushed a new version of the file) the next
resolve command will update packages to match that file. After a successful
resolve command, the checked out versions of all dependencies and the versions
recorded in the resolved versions file will match. In most cases the resolve
command will perform no changes unless the Package.swift
manifest or
Package.resolved
file have changed.
Most SwiftPM commands will implicitly invoke the swift package resolve
functionality before running, and will cancel with an error if dependencies
cannot be resolved.
The tools version declares the minimum version of the Swift tools required to
use the package, determines what version of the PackageDescription API should
be used in the Package.swift
manifest, and determines which Swift language
compatibility version should be used to parse the Package.swift
manifest.
When resolving package dependencies, if the version of a dependency that would normally be chosen specifies a Swift tools version which is greater than the version in use, that version of the dependency will be considered ineligible and dependency resolution will continue with evaluating the next-best version. If no version of a dependency (which otherwise meets the version requirements from the package dependency graph) supports the version of the Swift tools in use, a dependency resolution error will result.
The Swift tools version is specified by a special comment in the first line of
the Package.swift
manifest. To specify a tools version, a Package.swift
file
must begin with the string // swift-tools-version:
, followed by a version
number specifier.
The version number specifier follows the syntax defined by semantic versioning
2.0, with an amendment that the patch version component is optional and
considered to be 0 if not specified. The semver
syntax allows for an optional
pre-release version component or build version component; those components will
be completely ignored by the package manager currently.
After the version number specifier, an optional ;
character may be present;
it, and anything else after it until the end of the first line, will be ignored
by this version of the package manager, but is reserved for the use of future
versions of the package manager.
Some Examples:
// swift-tools-version:3.1
// swift-tools-version:3.0.2
// swift-tools-version:4.0
The following Swift tools version commands are supported:
-
Report tools version of the package:
$ swift package tools-version
-
Set the package's tools version to the version of the tools currently in use:
$ swift package tools-version --set-current
-
Set the tools version to a given value:
$ swift package tools-version --set <value>
Use the swift test
tool to run the tests of a Swift package. For more information on
the test tool, run swift test --help
.
Use the swift run [executable [arguments...]]
tool to run an executable product of a Swift
package. The executable's name is optional when running without arguments and when there
is only one executable product. For more information on the run tool, run
swift run --help
.
SwiftPM allows two build configurations: Debug (default) and Release.
By default, running swift build
will build in its debug configuration.
Alternatively, you can also use swift build -c debug
. The build artifacts are
located in a directory called debug
under the build folder. A Swift target is built
with the following flags in debug mode:
-Onone
: Compile without any optimization.-g
: Generate debug information.-enable-testing
: Enable the Swift compiler's testability feature.
A C language target is built with the following flags in debug mode:
-O0
: Compile without any optimization.-g
: Generate debug information.
To build in release mode, type swift build -c release
. The build artifacts
are located in directory named release
under the build folder. A Swift target is
built with following flags in release mode:
-O
: Compile with optimizations.-whole-module-optimization
: Optimize input files (per module) together instead of individually.
A C language target is built with following flags in release mode:
-O2
: Compile with optimizations.
You can pass more flags to the C, C++, or Swift compilers in three different ways:
- Command-line flags passed to these tools: flags like
-Xcc
or-Xswiftc
are used to pass C or Swift flags to all targets, as shown with-Xlinker
above. - Target-specific flags in the manifest: options like
cSettings
orswiftSettings
are used for fine-grained control of compilation flags for particular targets. - A destination JSON file: once you have a set of working command-line flags that you
want applied to all targets, you can collect them in a JSON file and pass them in through
extra-cc-flags
andextra-swiftc-flags
with--destination example.json
. Take a look atUtilities/build_ubuntu_cross_compilation_toolchain
for an example.
One difference is that C flags passed in the -Xcc
command-line or manifest's cSettings
are supplied to the Swift compiler too for convenience, but extra-cc-flags
aren't.
Swift Package Manager includes a build system that can build for macOS and Linux.
Xcode 11 integrates with libSwiftPM
to provide support for iOS, watchOS, and tvOS platforms.
To build your package with Xcode from command line you can use
xcodebuild
.
An example invocation would be:
xcodebuild -scheme Foo -destination 'generic/platform=iOS'
where Foo
would be the name of the library product you're trying to build. You can
get the full list of available schemes for you SwiftPM package with xcodebuild -list
.
You can get the list of available destinations for a given scheme with this invocation:
xcodebuild -showdestinations -scheme Foo
C language targets are similar to Swift targets, except that the C language
libraries should contain a directory named include
to hold the public headers.
To allow a Swift target to import a C language target, add a target in the manifest file. Swift Package Manager will automatically generate a modulemap for each C language library target for these 3 cases:
-
If
include/Foo/Foo.h
exists andFoo
is the only directory under the include directory, and the include directory contains no header files, theninclude/Foo/Foo.h
becomes the umbrella header. -
If
include/Foo.h
exists andinclude
contains no other subdirectory, theninclude/Foo.h
becomes the umbrella header. -
Otherwise, the
include
directory becomes an umbrella directory, which means that all headers under it will be included in the module.
In case of complicated include
layouts or headers that are not compatible with
modules, a custom module.modulemap
can be provided in the include
directory.
For executable targets, only one valid C language main file is allowed, e.g., it
is invalid to have main.c
and main.cpp
in the same target.
SwiftPM ships with completion scripts for both Bash and ZSH. These files should be generated in order to use them.
Use the following commands to install the Bash completions to ~/.swift-package-complete.bash
and automatically load them using your ~/.bash_profile
file.
swift package completion-tool generate-bash-script > ~/.swift-package-complete.bash
echo -e "source ~/.swift-package-complete.bash\n" >> ~/.bash_profile
source ~/.swift-package-complete.bash
Alternatively, add the following commands to your ~/.bash_profile
file to directly load completions:
# Source Swift completion
if [ -n "`which swift`" ]; then
eval "`swift package completion-tool generate-bash-script`"
fi
Use the following commands to install the ZSH completions to ~/.zsh/_swift
. You can chose a different folder, but the filename should be _swift
. This will also add ~/.zsh
to your $fpath
using your ~/.zshrc
file.
mkdir ~/.zsh
swift package completion-tool generate-zsh-script > ~/.zsh/_swift
echo -e "fpath=(~/.zsh \$fpath)\n" >> ~/.zshrc
compinit