Python on iOS README

Authors:	Russell Keith-Magee (2023-11)

This document provides a quick overview of some iOS specific features in the Python distribution.

These instructions are only needed if you're planning to compile Python for iOS yourself. Most users should not need to do this. If you're looking to experiment with writing an iOS app in Python, tools such as BeeWare's Briefcase and Kivy's Buildozer will provide a much more approachable user experience.

Compilers for building on iOS

Building for iOS requires the use of Apple's Xcode tooling. It is strongly recommended that you use the most recent stable release of Xcode. This will require the use of the most (or second-most) recently released macOS version, as Apple does not maintain Xcode for older macOS versions. The Xcode Command Line Tools are not sufficient for iOS development; you need a full Xcode install.

If you want to run your code on the iOS simulator, you'll also need to install an iOS Simulator Platform. You should be prompted to select an iOS Simulator Platform when you first run Xcode. Alternatively, you can add an iOS Simulator Platform by selecting an open the Platforms tab of the Xcode Settings panel.

iOS specific arguments to configure

--enable-framework[=DIR]

This argument specifies the location where the Python.framework will be installed. If DIR is not specified, the framework will be installed into a subdirectory of the iOS/Frameworks folder.

This argument must be provided when configuring iOS builds. iOS does not support non-framework builds.
--with-framework-name=NAME

Specify the name for the Python framework; defaults to Python.

Use this option with care!

Unless you know what you're doing, changing the name of the Python framework on iOS is not advised. If you use this option, you won't be able to run the make testios target without making signficant manual alterations, and you won't be able to use any binary packages unless you compile them yourself using your own framework name.

Building Python on iOS

ABIs and Architectures

iOS apps can be deployed on physical devices, and on the iOS simulator. Although the API used on these devices is identical, the ABI is different - you need to link against different libraries for an iOS device build (iphoneos) or an iOS simulator build (iphonesimulator).

Apple uses the XCframework format to allow specifying a single dependency that supports multiple ABIs. An XCframework is a wrapper around multiple ABI-specific frameworks that share a common API.

iOS can also support different CPU architectures within each ABI. At present, there is only a single supported architecture on physical devices - ARM64. However, the simulator supports 2 architectures - ARM64 (for running on Apple Silicon machines), and x86_64 (for running on older Intel-based machines).

To support multiple CPU architectures on a single platform, Apple uses a "fat binary" format - a single physical file that contains support for multiple architectures. It is possible to compile and use a "thin" single architecture version of a binary for testing purposes; however, the "thin" binary will not be portable to machines using other architectures.

Building a single-architecture framework

The Python build system will create a Python.framework that supports a single ABI with a single architecture. Unlike macOS, iOS does not allow a framework to contain non-library content, so the iOS build will produce a bin and lib folder in the same output folder as Python.framework. The lib folder will be needed at runtime to support the Python library.

If you want to use Python in a real iOS project, you need to produce multiple Python.framework builds, one for each ABI and architecture. iOS builds of Python must be constructed as framework builds. To support this, you must provide the --enable-framework flag when configuring the build. The build also requires the use of cross-compilation. The minimal commands for building Python for the ARM64 iOS simulator will look something like:

$ export PATH="$(pwd)/iOS/Resources/bin:/usr/bin:/bin:/usr/sbin:/sbin:/Library/Apple/usr/bin"
$ ./configure \
      --enable-framework \
      --host=arm64-apple-ios-simulator \
      --build=arm64-apple-darwin \
      --with-build-python=/path/to/python.exe
$ make
$ make install

In this invocation:

iOS/Resources/bin has been added to the path, providing some shims for the compilers and linkers needed by the build. Xcode requires the use of xcrun to invoke compiler tooling. However, if xcrun is pre-evaluated and the result passed to configure, these results can embed user- and version-specific paths into the sysconfig data, which limits the portability of the compiled Python. Alternatively, if xcrun is used as the compiler, it requires that compiler variables like CC include spaces, which can cause significant problems with many C configuration systems which assume that CC will be a single executable.

To work around this problem, the iOS/Resources/bin folder contains some wrapper scripts that present as simple compilers and linkers, but wrap underlying calls to xcrun. This allows configure to use a CC definition without spaces, and without user- or version-specific paths, while retaining the ability to adapt to the local Xcode install. These scripts are included in the bin directory of an iOS install.

These scripts will, by default, use the currently active Xcode installation. If you want to use a different Xcode installation, you can use xcode-select to set a new default Xcode globally, or you can use the DEVELOPER_DIR environment variable to specify an Xcode install. The scripts will use the default iphoneos/iphonesimulator SDK version for the select Xcode install; if you want to use a different SDK, you can set the IOS_SDK_VERSION environment variable. (e.g, setting IOS_SDK_VERSION=17.1 would cause the scripts to use the iphoneos17.1 and iphonesimulator17.1 SDKs, regardless of the Xcode default.)

The path has also been cleared of any user customizations. A common source of bugs is for tools like Homebrew to accidentally leak macOS binaries into an iOS build. Resetting the path to a known "bare bones" value is the easiest way to avoid these problems.
--host is the architecture and ABI that you want to build, in GNU compiler triple format. This will be one of:
- arm64-apple-ios for ARM64 iOS devices.
- arm64-apple-ios-simulator for the iOS simulator running on Apple Silicon devices.
- x86_64-apple-ios-simulator for the iOS simulator running on Intel devices.
--build is the GNU compiler triple for the machine that will be running the compiler. This is one of:
- arm64-apple-darwin for Apple Silicon devices.
- x86_64-apple-darwin for Intel devices.
/path/to/python.exe is the path to a Python binary on the machine that will be running the compiler. This is needed because the Python compilation process involves running some Python code. On a normal desktop build of Python, you can compile a python interpreter and then use that interpreter to run Python code. However, the binaries produced for iOS won't run on macOS, so you need to provide an external Python interpreter. This interpreter must be the same version as the Python that is being compiled. To be completely safe, this should be the exact same commit hash. However, the longer a Python release has been stable, the more likely it is that this constraint can be relaxed - the same micro version will often be sufficient.
The install target for iOS builds is slightly different to other platforms. On most platforms, make install will install the build into the final runtime location. This won't be the case for iOS, as the final runtime location will be on a physical device.

However, you still need to run the install target for iOS builds, as it performs some final framework assembly steps. The location specified with --enable-framework will be the location where make install will assemble the complete iOS framework. This completed framework can then be copied and relocated as required.

For a full CPython build, you also need to specify the paths to iOS builds of the binary libraries that CPython depends on (XZ, BZip2, LibFFI and OpenSSL). This can be done by defining the LIBLZMA_CFLAGS, LIBLZMA_LIBS, BZIP2_CFLAGS, BZIP2_LIBS, LIBFFI_CFLAGS, and LIBFFI_LIBS environment variables, and the --with-openssl configure option. Versions of these libraries pre-compiled for iOS can be found in this repository.

By default, Python will be compiled with an iOS deployment target (i.e., the minimum supported iOS version) of 12.0. To specify a different deployment target, provide the version number as part of the --host argument - for example, --host=arm64-apple-ios15.4-simulator would compile an ARM64 simulator build with a deployment target of 15.4.

Merge thin frameworks into fat frameworks

Once you've built a Python.framework for each ABI and and architecture, you must produce a "fat" framework for each ABI that contains all the architectures for that ABI.

The iphoneos build only needs to support a single architecture, so it can be used without modification.

If you only want to support a single simulator architecture, (e.g., only support ARM64 simulators), you can use a single architecture Python.framework build. However, if you want to create Python.xcframework that supports all architectures, you'll need to merge the iphonesimulator builds for ARM64 and x86_64 into a single "fat" framework.

The "fat" framework can be constructed by performing a directory merge of the content of the two "thin" Python.framework directories, plus the bin and lib folders for each thin framework. When performing this merge:

The pure Python standard library content is identical for each architecture, except for a handful of platform-specific files (such as the sysconfig module). Ensure that the "fat" framework has the union of all standard library files.
Any binary files in the standard library, plus the main libPython3.X.dylib, can be merged using the lipo tool, provide by Xcode:
```
$ lipo -create -output module.dylib path/to/x86_64/module.dylib path/to/arm64/module.dylib
```
The header files will be indentical on both architectures, except for pyconfig.h. Copy all the headers from one platform (say, arm64), rename pyconfig.h to pyconfig-arm64.h, and copy the pyconfig.h for the other architecture into the merged header folder as pyconfig-x86_64.h. Then copy the iOS/Resources/pyconfig.h file from the CPython sources into the merged headers folder. This will allow the two Python architectures to share a common pyconfig.h header file.

At this point, you should have 2 Python.framework folders - one for iphoneos, and one for iphonesimulator that is a merge of x86+64 and ARM64 content.

Merge frameworks into an XCframework

Now that we have 2 (potentially fat) ABI-specific frameworks, we can merge those frameworks into a single XCframework.

The initial skeleton of an XCframework is built using:

xcodebuild -create-xcframework -output Python.xcframework -framework path/to/iphoneos/Python.framework -framework path/to/iphonesimulator/Python.framework

Then, copy the bin and lib folders into the architecture-specific slices of the XCframework:

cp path/to/iphoneos/bin Python.xcframework/ios-arm64
cp path/to/iphoneos/lib Python.xcframework/ios-arm64

cp path/to/iphonesimulator/bin Python.xcframework/ios-arm64_x86-64-simulator
cp path/to/iphonesimulator/lib Python.xcframework/ios-arm64_x86-64-simulator

Note that the name of the architecture-specific slice for the simulator will depend on the CPU architecture that you build.

Then, add symbolic links to "common" platform names for each slice:

ln -si ios-arm64 Python.xcframework/iphoneos
ln -si ios-arm64_x86-64-simulator Python.xcframework/iphonesimulator

You now have a Python.xcframework that can be used in a project.

Testing Python on iOS

The iOS/testbed folder that contains an Xcode project that is able to run the iOS test suite. This project converts the Python test suite into a single test case in Xcode's XCTest framework. The single XCTest passes if the test suite passes.

To run the test suite, configure a Python build for an iOS simulator (i.e., --host=arm64-apple-ios-simulator or --host=x86_64-apple-ios-simulator ), specifying a framework build (i.e. --enable-framework). Ensure that your PATH has been configured to include the iOS/Resources/bin folder and exclude any non-iOS tools, then run:

$ make all
$ make install
$ make testios

This will:

Build an iOS framework for your chosen architecture;
Finalize the single-platform framework;
Make a clean copy of the testbed project;
Install the Python iOS framework into the copy of the testbed project; and
Run the test suite on an "iPhone SE (3rd generation)" simulator.

While the test suite is running, Xcode does not display any console output. After showing some Xcode build commands, the console output will print Testing started, and then appear to stop. It will remain in this state until the test suite completes. On a 2022 M1 MacBook Pro, the test suite takes approximately 12 minutes to run; a couple of extra minutes is required to boot and prepare the iOS simulator.

On success, the test suite will exit and report successful completion of the test suite. No output of the Python test suite will be displayed.

On failure, the output of the Python test suite will be displayed. This will show the details of the tests that failed.

Debugging test failures

The easiest way to diagnose a single test failure is to open the testbed project in Xcode and run the tests from there using the "Product > Test" menu item.

Running specific tests

As the test suite is being executed on an iOS simulator, it is not possible to pass in command line arguments to configure test suite operation. To work around this limitation, the arguments that would normally be passed as command line arguments are configured as a static string at the start of the XCTest method - (void)testPython in iOSTestbedTests.m. To pass an argument to the test suite, add a a string to the argv defintion. These arguments will be passed to the test suite as if they had been passed to python -m test at the command line.

Disabling automated breakpoints

By default, Xcode will inserts an automatic breakpoint whenever a signal is raised. The Python test suite raises many of these signals as part of normal operation; unless you are trying to diagnose an issue with signals, the automatic breakpoints can be inconvenient. However, they can be disabled by creating a symbolic breakpoint that is triggered at the start of the test run.

Select "Debug > Breakpoints > Create Symbolic Breakpoint" from the Xcode menu, and populate the new brewpoint with the following details:

Name: IgnoreSignals
Symbol: UIApplicationMain
Action: Add debugger commands for: - process handle SIGINT -n true -p true -s false - process handle SIGUSR1 -n true -p true -s false - process handle SIGUSR2 -n true -p true -s false - process handle SIGXFSZ -n true -p true -s false
Check the "Automatically continue after evaluating" box.

All other details can be left blank. When the process executes the UIApplicationMain entry point, the breakpoint will trigger, run the debugger commands to disable the automatic breakpoints, and automatically resume.

Uh oh!