Step-driven valuation hides Libmarpa’s grammar rewrites from the application, and is quite efficient. Libmarpa knows which rules are sequences. Libmarpa optimizes stack manipulations based on this knowledge. Long sequences are very common in practical grammars. For these, the stack manipulations suggested by Libmarpa’s step-driven valuator will be significantly faster than the traditional stack evaluation algorithm.
Step-driven evalution has another advantage. To illustrate this, consider what is a very common case: The semantics are implemented in a higher-level language, using callbacks. If Libmarpa did not use step-driven valuation, it would need to provide for this case. But for generality, Libmarpa would have to deal in C callbacks. Therefore, a middle layer would have to create C language wrappers for the callbacks in the higher level language.
The implementation that results is this: The higher level language would need to wrap each callback in C. When calling Libmarpa, it would pass the wrappered callback. Libmarpa would then need to call the C language “wrappered” callback. Next, the wrapper would call the higher-level language callback. The return value, which would be data native to the higher-level language, would need to be passed to the C language wrapper, which will need to make arrangements for it to be based back to the higher-level language when appropriate.
A setup like this is not terribly efficient. And exception handling across language boundaries would be very tricky. But neither of these is the worst problem.
Callbacks are hard to debug. Wrappered callbacks are even worse. Calls made across language boundaries are harder yet to debug. In the system described above, by the time a return value is finally consumed, a language boundary will have been crossed four times.
How do programmers deal with difficulties like this? Usually, by doing the absolute minimum possible in the callbacks. A horrific debugging enviroment can become a manageable one if there is next to no code to be debugged. And this can be accomplished by doing as much as possible in pre- and post-processing.
In essence, callbacks force applications to do most of the programming via side effects. One need not be a functional programming purist to find this a very undesirable style of design to force on an application. But the ability to debug can make the difference between code that does work and code that does not. Unfairly or not, code is rarely considered well-designed when it does not work.
So, while step-driven valuation seems a roundabout approach, it is simpler and more direct than the likely alternatives. And there is something to be said for pushing semantics up to the higher levels — they can be expected to know more about it.
These advantages of step-driven valuation are strictly in the context of a low-level interface. The author is under no illusion that direct use of Libmarpa’s valuator will be found satisfactory by most application programmers, even those using the C language. The author certainly avoids using step-driven valuation directly. Libmarpa’s valuator is intended to be used via an upper layer, one which does know about semantics.