[Paper link] [Slides link] [Code link] [ICML link]

We proposed a simple yet effective pipeline for pass ordering for program size reduction.

• Key ideas
  • We don’t search good passes sequentially
  • Directly find a universal core set of pass sequences (termed coreset)
  • Make decision on top of this different action space to avoid the challenge of sparse reward
• The coreset
  • Optimize most programs
  • Require only an upfront search cost
• The policy
  • Learn to predict the relative performance of coreset sequences
  • Apply the optimal coreset sequence to compile new programs at inference

In summary, we first identify a small set (termed coreset) of pass sequences that generally optimize the size of most programs. Then, a policy is learned to pick the optimal sequences by predicting the normalized values of the pass sequences in the coreset.

The policy can be as simple as a multi-layer perceptron (MLP) that leverages the Autophase feature of a program. It can also be a graph neural network (GNN) that exploits the ProGraML graph of the program to predict the optimal pass sequence.