Partial Evaluation in Truffle

Truffle uses partial evaluation plus dynamic speculation and runtime profiling to turn interpreters into highly optimized machine code without a separate, language-specific compiler. By treating stable interpreter state and guest-program structure as compilation constants, Truffle specializes interpreter code to particular programs and removes interpreter indirection. Partial evaluation distinguishes static from dynamic inputs, producing a residual program specialized to the static values; this reduces runtime work and enables further compiler optimizations. The approach simplifies language implementation by deriving optimized code directly from the interpreter, while requiring interpreter code to be written to expose stable values and to handle specialization, generalization, and invalidation correctly.

Applied to interpreters, guest programs (ASTs or bytecode) are often stable and become PE constants. Truffle exemplifies this with an AST interpreter where a concrete AddNode with LoadArgumentNode children becomes monomorphic, gets inlined, and has final fields propagated transitively into constants, leaving code equivalent to a compiled addition of arguments. At runtime Truffle profiles CallTargets (RootNodes), marks hot targets for compilation, partially evaluates the interpreter with respect to a call target, and hands the specialized Java code to the Graal compiler to produce optimized machine code.