Jank, a Clojure dialect targeting native via LLVM, added a custom SSA-based IR to enable Clojure-semantic optimizations, cutting fibonacci runtime from 5,522ms toward JVM’s 200ms.
Key Takeaways
The new IR operates at Clojure’s semantic level (var derefs, dynamic calls, persistent data structures) rather than LLVM IR’s low-level primitives, enabling higher-level optimization passes.
IR is SSA-based and CFG-structured, stored as C++ data structures but renderable to Clojure EDN for debugging; no round-trip serialization yet.
Before IR optimizations, jank took 5,522ms on (fibonacci 35) vs Clojure JVM’s 200ms on OpenJDK 21 – a 27x gap the IR pipeline is designed to close.
First optimization target is inlining arithmetic: Clojure JVM inlines math via metadata; jank is reimplementing this properly after removing a prior hacky solution.
The IR took six weeks to design and implement, replacing AST-based C++ codegen; no optimization passes run yet – the merge prioritized pipeline stability over optimization depth.
