Starting with AGP 9.2.0, R8 now optimizes most `Atomic*FieldUpdater` calls into direct `Unsafe` variants, eliminating costly reflective safety checks at runtime. This change delivers a 2x–4x speedup on common atomic operations and has a major impact on `kotlinx.atomicfu`, the library underpinning `kotlinx.coroutines`. As a result, launching and cancelling coroutines in Android apps — including Jetpack Compose's `LaunchedEffect` — is up to 2x faster. The post explains the root cause (ART does not JIT-optimize `AtomicReferenceFieldUpdater` the way the JVM does), walks through benchmarks confirming the overhead, and details how R8 performs the optimization in three phases: instrumentation (adding an offset field), replacement (rewriting call sites to use `Unsafe` directly), and clean-up (removing dead updater fields and initializers). Apps targeting API 36 may also benefit from similar optimizations being added natively to ART's JIT. Upgrading to AGP 9.2.0 is all that's needed to get the improvement.
