IssueTracker

During migration from Android Gradle Plugin 3.4.2 to 4.0.1 our team noticed a strange side effect. Consider the following example:

class Sample {
    private val manager: ConfigurationManager = ConfigurationManager()
    private var listener: ConfigurationHandler? = null

    fun run() {
        listener = InnerHandler().also(manager::execute)
    }
    
    private inner class InnerHandler : ConfigurationHandler {
        override fun invoke() {
            println(this@Sample.hashCode())
        }
    }
}

Let's take a look at the bytecode.

• R8 bundled with AGP 3.4.2: https://pastebin.com/5KV4fxdn
• R8 bundled with AGP 4.0.1: https://pastebin.com/YyePLedQ

The main difference between those outputs is that R8 used in AGP 4.0.1 inlines listener field while R8 bundled with AGP 3.4.2 doesn't. It becomes an issue if we rely on that field to be present. For example, consider the following implementation of ConfigurationHandler:

class ConfigurationManager {
    private val handlers = mutableListOf<WeakReference<ConfigurationHandler>>()

    fun execute(handler: ConfigurationHandler) {
        handlers.add(WeakReference(handler))
        executeNetworkCall()
    }
    
    fun executeNetworkCall() {
        // network related code
        handlers.forEach { it.get()?.invoke() }
    }
}

As you can see, if we remove the private field(listener) there is a chance that GC will collect WeakReference. That kind of behavior is unexpected and can lead to some hard to reproduce bugs.

Currently, we are using specific proguard rule to avoid that behavior:

-keepclassmembers class * {
    private <fields>;
}

Which is working fine, even though it seems a bit broad because it also keeps fully unused fields.

There is a reproducible sample attached.