BO3: Workload-Specific Compilation Pipelines for Multi-Level Compilers (CASES 2026 Artifact)

Artifact for the CASES 2026 (ESWEEK) paper "BO3: Workload-Specific Compilation Pipelines for Multi-Level Compilers."

BO3 (Beyond O3) is a legality-aware three-stage reinforcement-learning framework for automated pass ordering in IREE, an MLIR-based ML compiler. This archive contains the complete implementation: the Gymnasium environment with GraphSAGE SSA-graph observations and dynamic legality masking, the C++/pybind11 MLIR toolchain backend, Stage-I backend-knob screening results, Stage-II Maskable-PPO training and Stage-III candidate-selection scripts, random-rollout and beam-search baselines, ablation configurations, ONNX export scripts for all nine evaluated workload instances (AlexNet, ResNet-50, DenseNet, BERT-Base, ViT-224, YOLOv5, UNet-256/384/640), full test suites, and the accepted paper.

It reproduces the paper's claims, including a 5.66x geometric-mean speedup over the default IREE -O3 pipeline on a single pinned x86 CPU core.

Start with ARTIFACT.md. Requirements, installation, badge claims, and experiment workflows are documented in REQUIREMENTS.md, INSTALLATION.md, and STATUS.md.