Speculative Optimizations without Fear

Just-in-time (JIT) language implementations make heavy use of speculative optimizations, based on assumptions that might become invalid. This requires the implementation to support a bailout mechanism to undo invalidated optimizations. The interaction between assumptions and optimizations is a major source of complexity for JIT implementers: how should optimizations preserve bailout information? What are the trade-offs when adding more assumptions? We demonstrate how to reason about correctness of speculative optimization and deoptimization. We present sourir, an intermediate representation with explicit assumption checkpoints, designed to formally study speculative optimizations in JIT systems. Sourir models core difficulties of speculative optimization, such as interleaving assumptions and optimizations, and the interaction with inlining. Our formalization stays at the IR level, thus abstracting away orthogonal aspects of JIT implementations, such as dynamic code generation and self mutation. We describe a set of common optimizations (constant folding, unreachable code elimination, and function inlining) and prove them correct in presence of speculative assumptions. CCS Concepts: • Software and its engineering→ Just-in-time compilers;