Languages versus Implementations

In the world of programming languages, there is a clear distinction between the language itself, and the implementation of the language. There is only one abstract Python, but there are many concrete implementations of Python. Only one Haskell, but several implementations thereof.

However, this is a false dichotomy. Each implementation is necessarily different from the others, mainly because implementing an identical specification in identical fashion is boring and doesn’t solve any problems.

Outside of those implementations we make as a learning process, we programmers typically create new implementations of existing languages in search of better performance or better user experience. I don’t expect anyone to agree with me that changing the performance characteristics of a language is the same as a change to the language, but that is what I believe.

More importantly, the implementors of a language will often try to make their implementation stand out as more productive, more useful to the programmer, by adding extensions. Whether or not these extensions are intended for widespread adoption by other implementations, and whether you call them extensions or pragmas or whatever else, the effect is the same. A program that is accepted by one implementation is either not accepted by another, or it runs differently under the different implementations.

The moment this happens, the two implementations are implementing different languages—with a lot of overlap, certainly, but nevertheless distinct. This divergence happens inadvertently as well: because of bugs, one implementation will likely fail or produce incorrect output for a program that another implementation handles just fine, and vice versa. These are still, by the same reasoning, implementing different languages.

So next time you go to accuse someone of conflating the theoretical and the practical, consider that they may simply be ignoring theory for the sake of practicality.