InjectionIII.app Project

Yes, HotReloading for Swift

Chinese language README: 中文集成指南

Code injection allows you to update the implementation of functions and any method of a class, struct or enum incrementally in the iOS simulator without having to perform a full rebuild or restart your application. This saves the developer a significant amount of time tweaking code or iterating over a design. Effectively it changes Xcode from being a "source editor" to being a "program editor" where source changes are not just saved to disk but into your running program directly.

How to use it

Setting up your projects to use injection is now as simple as downloading one of the github releases of the app or from the Mac App Store and adding the code below somewhere in your app to be executed on startup (it is no longer necessary to actually run the app itself).

#if DEBUG
Bundle(path: "/Applications/InjectionIII.app/Contents/Resources/iOSInjection.bundle")?.load()
//for tvOS:
Bundle(path: "/Applications/InjectionIII.app/Contents/Resources/tvOSInjection.bundle")?.load()
//Or for macOS:
Bundle(path: "/Applications/InjectionIII.app/Contents/Resources/macOSInjection.bundle")?.load()
#endif

It's also important to add the options -Xlinker and -interposable (without double quotes and on separate lines) to the "Other Linker Flags" of targets in your project (for the Debug configuration only) to enable "interposing" (see the explanation below).

After that, when you run your app in the simulator you should see a message saying a file watcher has started for your home directory and, whenever you save a source file in the current project it should report it has been injected. This means all places that formerly called the old implementation will have been updated to call the latest version of your code.

It's not quite as simple as that as to see results on the screen immediately the new code needs to have actually been called. For example, if you inject a view controller it needs to force a redisplay. To resolve this problem, classes can implement an @objc func injected() method which will be called after the class has been injected to perform any update to the display. One technique you can use is to include the following code somewhere in your program:

#if DEBUG
extension UIViewController {
    @objc func injected() {
        viewDidLoad()
    }
}
#endif

Another solution to this problem is "hosting" using the Inject Swift Package introduced by this blog post.

What injection can't do

You can't inject changes to how data is laid out in memory i.e. you cannot add, remove or reorder properties with storage. For non-final classes this also applies to adding or removing methods as the vtable used for dispatch is itself a data structure which must not change over injection. Injection also can't work out what pieces of code need to be re-executed to update the display as discussed above. Also, don't get carried away with access control. private properties and methods can't be injected directly, particularly in extensions as they are not a global interposable symbol. They generally inject indirectly as they can only be acessed inside the file being injected but this can cause confusion. Finally, Injection doesn't cope well with source files being added/renamed/deleted during injection. You may need to build and relaunch your app or even close and reopen your project to clear out old Xcode build logs.

Injection of SwiftUI

SwiftUI is, if anything, better suited to injection than UIKit as it has specific mechanisms to update the display but you need to make a couple changes to each View struct you want to inject. To force redraw the simplest way is to add a property that observes when an injection has occurred:

    @ObserveInjection var forceRedraw

This property wrapper is available in either the HotSwiftUI or Inject Swift Package. It essentially contains an @Published integer your views observe that increments with each injection. You can use one of the following to make one of these packages available throughout your project:

@_exported import HotSwiftUI
or
@_exported import Inject

The second change you need to make for reliable SwiftUI injection is to "erase the return type" of the body property by wrapping it in AnyView using the .enableInjection() method extending View in these packages. This is because, as you add or remove SwiftUI elements it can change the concrete return type of the body property which amounts to a memory layout change that may crash. In summary, the tail end of each body should always look like this:

    var body: some View {
    	 VStack or whatever {
        // Your SwiftUI code...
        }
        .enableInjection()
    }

    @ObserveInjection var redraw

You can leave these modifications in your production code as, for a Release build they optimise out to a no-op.

Injection on an iOS, tvOS or visionOS device

This can work but you will need to actually run one of the github 4.8.0+ releases of the InjectionIII.app, set a user default to opt-in and restart the app.

$ defaults write com.johnholdsworth.InjectionIII deviceUnlock any

Then, instead of loading the injection bundles run this script in a "Build Phase": (You will also need to turn off the project build setting "User Script Sandboxing")

RESOURCES=/Applications/InjectionIII.app/Contents/Resources
if [ -f "$RESOURCES/copy_bundle.sh" ]; then
    "$RESOURCES/copy_bundle.sh"
fi

and, in your application execute the following code on startup:

    #if DEBUG
    if let path = Bundle.main.path(forResource:
            "iOSInjection", ofType: "bundle") ??
        Bundle.main.path(forResource:
            "macOSInjection", ofType: "bundle") {
        Bundle(path: path)!.load()
    }
    #endif

Once you have switched to this configuaration it will also work when using the simulator. Consult the README of the HotReloading project for details on how to debug having your program connect to the InjectionIII.app over Wi-Fi. You will also need to select the project directory for the file watcher manually from the pop-down menu.

Injection on macOS

It works but you need to temporarily turn off the "app sandbox" and "library validation" under the "hardened runtime" during development so it can dynamically load code. In order to avoid codesigning problems, use the new copy_bundle.sh script as detailed in the instructions for injection on real devices above.

How it works

Injection has worked various ways over the years, starting out using the "Swizzling" apis for Objective-C but is now largely built around a feature of Apple's linker called "interposing" which provides a solution for any Swift method or computed property of any type.

When your code calls a function in Swift, it is generally "statically dispatched", i.e. linked using the "mangled symbol" of the function being called. Whenever you link your application with the "-interposable" option however, an additional level of indirection is added where it finds the address of all functions being called through a section of writable memory. Using the operating system's ability to load executable code and the fishhook library to "rebind" the call it is therefore possible to "interpose" new implementations of any function and effectively stitch them into the rest of your program at runtime. From that point it will perform as if the new code had been built into the program.

Injection uses the FSEventSteam api to watch for when a source file has been changed and scans the last Xcode build log for how to recompile it and links a dynamic library that can be loaded into your program. Runtime support for injection then loads the dynamic library and scans it for the function definitions it contains which it then "interposes" into the rest of the program. This isn't the full story as the dispatch of non-final class methods uses a "vtable" (think C++ virtual methods) which also has to be updated but the project looks after that along with any legacy Objective-C "swizzling".

If you are interested knowing more about how injection works the best source is either my book Swift Secrets or the new, start-over reference implementation in the InjectionLite Swift Package. For more information about "interposing" consult this blog post or the README of the fishhook project. For more information about the organisation of the app itself, consult ROADMAP.md.

A bit of terminology

Getting injection to work has three components. A FileWatcher, the code to recompile any changed files and build a dynamic library that can be loaded and the injection code itself which stitches the new versions of your code into the app while it's running. How these three components are combined gives rise to the number of ways injection can be used.

"Injection classic" is where you download one of the binary releases from github and run the InjectionIII.app. You then load one of the bundles inside that app into your program as shown above in the simulator. In this configuration, the file watcher and source recompiling is done inside the app and the bundle connects to the app using a socket to know when a new dynamic library is ready to be loaded.

"App Store injection" This version of the app is sandboxed and while the file watcher still runs inside the app, the recompiling and loading is delegated to be performed inside the simulator. This can create problems with C header files as the simulator uses a case sensitive file system to be a faithful simulation of a real device.

"HotReloading injection" was where you are running your app on a device and because you cannot load a bundle off your Mac's filesystem on a real phone you add the HotReloading Swift Package to your project (during development only!) which contains all the code that would normally be in the bundle to perform the dynamic loading. This requires that you use one of the un-sandboxed binary releases. It has also been replaced by the copy_bundle.sh script described above.

"Standalone injection". This was the most recent evolution of the project where you don't run the app itself anymore but simply load one of the injection bundles and the file watcher, re-compilation and injection are all performed inside the simulator. By default this watches for changes to any Swift file inside your home directory though you can change this using the environment variable INJECTION_DIRECTORIES.

InjectionLite is a start-over minimal implementation of standalone injection for reference. Just add this Swift package and you should be able to inject in the simulator.

InjectionNext is a currently experimental version of Injection that should be faster and more reliable for large projects. It integrates into a debugging flag of Xcode to find out how to recompile files to avoid parsing build logs.

All these variations require you to add the "-Xlinker -interposble" linker flags for a Debug build or you will only be able to inject non-final methods of classes.

Further information

Consult the old README which if anything contained simply "too much information" including the various environment variables you can use for customisation. A few examples:

With an INJECTION_TRACE environment variable, injecting any file will add logging of all calls to functions and methods in the file along with their argument values as an aid to debugging.

A little known feature of InjectionIII is that provided you have run the tests for your app at some point you can inject an individual XCTest class and have if run immediately – reporting if it has failed each time you modify it.

Acknowledgements:

This project includes code from rentzsch/mach_inject, erwanb/MachInjectSample, davedelong/DDHotKey and acj/TimeLapseBuilder-Swift under their respective licenses.

The App Tracing functionality uses the OliverLetterer/imp_implementationForwardingToSelector trampoline implementation via the SwiftTrace project under an MIT license.

SwiftTrace uses the very handy https://github.com/facebook/fishhook. See the project source and header file included in the app bundle for licensing details.

This release includes a very slightly modified version of the excellent canviz library to render "dot" files in an HTML canvas which is subject to an MIT license. The changes are to pass through the ID of the node to the node label tag (line 212), to reverse the rendering of nodes and the lines linking them (line 406) and to store edge paths so they can be coloured (line 66 and 303) in "canviz-0.1/canviz.js".

It also includes CodeMirror JavaScript editor for the code to be evaluated using injection under an MIT license.

The fabulous app icon is thanks to Katya of pixel-mixer.com.

$Date: 2024/06/30