Particle.news
Download on the App Store

Google Brings AutoFDO to the Android Kernel for Faster, More Efficient Phones

The effort steers LLVM optimizations with lab-built workload profiles to target the hottest kernel paths.

Overview

  • Google has begun delivering AutoFDO-optimized builds to the android16-6.12 and android15-6.6 kernel branches, with support planned for android17-6.18.
  • Profiles are synthesized in a controlled lab by exercising the top 100 apps on Pixel devices using simpleperf and ARM ETE/TRBE tracing, showing about 85% similarity to internal fleet patterns.
  • Early results reported by Google include roughly 4.3% faster cold app launches and about 2.1% quicker boot times, with kernel tests indicating similar gains.
  • A verification pipeline aggregates traces, converts them to AutoFDO format, trims cold functions, analyzes profile-to-binary changes, and runs targeted benchmarks before release.
  • To protect stability, AutoFDO adjusts compiler heuristics rather than kernel logic, applies a conservative-by-default policy, and is slated to expand to GKI modules and vendor DDK-built drivers.