This is somewhat of a caricature. If one's compilation strategy were to build an abstract syntax tree of the whole source program before emitting any object code, then things would go as Steve describes. In reality, TMG wasn't used this way. For one thing, ASTs needed too much memory. In TMG one could emit code while parsing. Typically code was generated on a statement-by-statement basis. This limited the backtracking, so even a naive "syntax error" diagnostic could be localized to the statement level. Long-distance connections, such as the branch labels in a nest of if statements, could nevertheless be realized recursively. Thus, in actual use, a TMG "grammar" was partly a parsing program and partly abstract specification. The balance of viewpoints was left to the discretion of the grammar's author. Yacc swung the pendulum toward the abstract. Doug