Jevopt

A C/C++ compiler driver that combines Clang's optimiser with Jev for function-call inlining decisions. At every discretionary call site, Jev chooses whether to inline using the current LLVM IR, the original source, the build context and a small set of structural facts.

The README's headline is 58.2% code-size reduction on Statemate from the Embench benchmark, with a dashboard shipped with the released jevopt-embench data. Because Jev's outputs are stochastic, reruns can produce different results and may not match the published figures.

uv run jevopt compile --source-dir <source_directory> --output-dir <output_directory> [--verify]. Output includes the Clang -Oz reference binary, the Jev-guided binary and result.json with both sizes and the complete decision trace. Project-specific options pass through repeatable --compile-arg and --link-arg flags.

It runs both binaries and compares their exit code and stdout. If the source directory contains expected.stdout, both outputs must also match it. The separate passthrough command returns LLVM's own decisions and must be byte-for-byte identical to regular Clang, making it an integration check.

Linux and a pinned LLVM 21 toolchain: Python 3.12+ with uv, Clang, Clang++, LLVM Size, LLD and LLVM development headers, plus CMake and Ninja and a TypeSafe API key added to .env. LTO is not supported because the necessary advisor support only runs per translation unit.

Clone Embench 1.0 into third_party/embench-iot, then run uv run benchmarks/run_embench.py --output-dir runs/jevopt-embench-reproduction. The runner refuses to overwrite an existing result and verifies every output binary; logs from the original run live in runs/jevopt-embench.

The live dashboard is at ramneet-singh.github.io/jevopt and publishes to GitHub Pages on pushes to main. You can also open dashboard/index.html directly, including from a downloaded copy, or serve it with python -m http.server 8000 --directory dashboard. Export other runs with python -m jevopt.dashboard.