This release contains a new inlining policy trained with a new Imitation Learning based algorithm. An implementation of the algorithm is checked in under compiler_opt/rl/imitation_learning. The policy was trained with an initial set of baseline policies consisting of the previous inlining model and the greedy policy that is always part of the compiler. Using inlining-Oz-v1.1 as starting point, we observe and additional 0.7% size reduction on a proprietary set of targets that are part of a cloud systems infrastructure. We also observe significant size reduction on some external facing Android targets.
There is also, specifically, a separate, Chrome-trained model as well, trained using the same methodology, which was trained starting with the PPO-trained baseline previously used there.