The compiler consists of the following stages: PARSER: sieve-parser.c, sieve-lexer.c Parses the scriptfile and produces an abstract syntax tree for it (sieve-ast.c). VALIDATOR: sieve-validator.c Performs contextual analysis on the ast produced by the parser. This checks for the validity of commands, tests and arguments. Also, the ast is decorated with any context data acquired during the process. This context is used by the last compiler stage. GENERATOR: sieve-generator.c This last compiler stage uses a visitor pattern to wander through the ast and produces sieve byte code (sieve-binary.c). The resulting (in-memory) binary can be fed to the interpreter for execution: INTERPRETER: sieve-interpreter.c The interpreter executes the byte code and produces a sieve_result object. This result is no more than just a collection of actions to be performed. During execution, action commands add actions to the result. Duplates and conflicts between actions are handled in this execution phase. RESULT: sieve-result.c sieve-actions.c When the result is to be executed, it needs no further checking, as the validity of the result was verified during interpretation already. The result's actions are executed in a transaction-like atomic manner. If one of the actions fails, the whole transaction is rolled back meaning that either everything succeeds or everything fails. This is only possible to some extent: transmitted responses can of course not be rolled back. However, these are executed in the commit phase, meaning that they will only be performed if all other actions were successful. Debugging: BINARY-DUMPER: sieve-code-dumper.c sieve-binary-dumper.c A loaded binary can be dumped to a stream in human-readable form using the binary-dumper. The binary-dumper displays information on all the blocks that the binary consists off. Program code blocks are dumped using the code-dumper. It's implementation is similar to the interpreter, with the exception that it performs no actions and just sequentially wanders through the byte code printing instructions along the way. The term human-readable is a bit optimistic though; currently, the presented data looks like an assembly language.