Hermes Hacking: Demystifying JavaScript Engines

Lexing turns JavaScript source into tokens, removing whitespace and comments and converting string and number syntax into values. Parsing organizes those tokens into an abstract syntax tree, with nodes for functions, calls, identifiers, and literals. Pietruszewski notes that JavaScript syntax can be ambiguous, so parsing uses context to distinguish cases such as division and regular expressions.

An AST interpreter evaluates each node, working through expressions such as `1 + 2 * 3` and looking up variables in an execution context. A practical interpreter also needs to handle function declarations, hoisting, nested scopes, and the call stack. Pietruszewski uses `const a = 1` followed by `console.log(a + 2)` to show how declarations and calls are evaluated.

A bytecode interpreter first compiles the AST into a flat sequence of simple instructions, then executes them on a virtual machine with registers and locals. Pietruszewski says this layout can improve performance because execution follows a compact instruction buffer instead of jumping among AST objects in memory. Hermes also uses an intermediate representation, applies optimization passes, and generates bytecode ahead of execution.

Pietruszewski describes mark-and-sweep collection: the engine follows references from roots such as globals, event handlers, and stack locals, then frees objects it cannot reach. Collection may pause execution, so the hard engineering problem is reducing pauses and doing more work concurrently. He points to Hermes documentation and a WebKit post by Filip Pizło for deeper reading.