In November of 2013, Alan Dipert gave a lightning talk at Clojure/conj about gherkin, a Lisp implemented in bash. His presentation led me to ask myself the question of whether a Lisp could be created using the GNU Make macro language. As you have probably guessed, the answer to that question is yes.
Interestingly, the current pedagogical/educational purpose of mal happened due to a semantic naming accident (naming is such a fraught task in computer science). If I am remembering correctly, the name "mal" original meant "MAke Lisp". I do not remember precisely why I continued to create more implementations, apart from the fact that it was a fun challenge, but after the make implementation, many of the others were relatively easy. At some point during that process, I realized that the multiple implementations and incremental steps (which was originally just for my own clarity) was a useful learning tool and so the "mal" name became a double entendre for "Make, A Lisp" and "make-a-lisp" (and eventually just the latter given that the make implementation is now just a small part of the whole).
The split between code that goes in steps and code that goes into other files is not completely arbitrary (a bit arbitrary, but not completely). My rule of thumb is something like this: if the code is specific and necessary for implementing a Lisp then it belongs in the step files. If the purpose of the code is for implementing new dynamic data-types/objects and the functions or methods that operate on those types, then it goes in separate files.
If the target language has types and functions that resemble mal types, then those files tend to be very small or non-existent. Examples:
- the mal implementation has no types, reader, printer files and has a trivial core file (just to hoist underlying functions)
- the Clojure implementation has no types file and fairly trivial reader and printer files (just to modify the Clojure reader/writer slightly) and a fairly trivial core file
- ruby types and the functions that operate on them are very "Lispy" so the Ruby types file and core file are very small.
The env file is somewhat more arbitrary, however, it is a self-contained module that is implemented early and changes very little after that, so I decided to separate it. Also, for languages that have hierarchical maps/dictionaries (e.g. Javascript objects/prototype chain), you do not necessarily need an env file.
Another way of summarizing this answer is that the step files represent the core of what makes something a Lisp, the rest of the modules are just language specific details (they may be the harder than the Lisp part, but that is due to the nature of the target language not because of Lisp functionality per se).
There is no single consistent rule that I have used to determine which functionality goes in which step and the arrangement has changed numerous times since the beginning of the project. There are several different goals that I try and balance in determining which functionality goes into which step:
-
Optimize Lisp learning: I want developers who are unfamiliar with Lisp to be able to use the project and guide to learn about Lisp without becoming overwhelmed. In many Lisp introductions, concepts like quoting and homoiconicity (i.e. a user exposed eval function) are introduced early. But these are fairly foreign to most other languages so they are introduced in later steps in mal. I also try to not to concentrate too many Lisp concepts in a single step. So many steps contain one or two Lisp concepts plus some core function additions that support those concepts.
-
Optimize implementation language learning (equal-ish step sizing): I try to structure the steps so that the target implementation can be learned incrementally. This goal is the one that has caused me to refactor the steps the most. Different languages have different areas that they optimize and make simple for the developer. For example, in Java (prior to 8) and PostScript creating the equivalent of anonymous functions and function closures is painful. In other languages, function closures are trivial, but IO and error handling are tedious when you are first learning the language (I am looking at you Haskell). So this goal is really about trying to balance step size across multiple languages.
-
Practical results early and continuous feedback: it is a scientific fact that many small rewards are more motivating than a single large reward (citation intentionally omitted, get a small reward by googling it yourself). Each step in mal adds new functionality that can actually be exercised by the implementor and, just as importantly, easily tested.
Also, the step structure of mal/make-a-lisp is not perfect. It never will be perfect, but there are some areas that could be improved. The most glaring problem is that step1 is on the heavy/large size because in most languages you have to implement a good portion of the reader/printer before you can begin using/testing the step. The compromise I have settled on for now is to put extra detail in the process guide for step1 and to be clear that many of the types are deferrable until later. But I am always open to suggestions.
Absolutely! I want mal to have a idiomatic implementation in every programming language.
Here are a few guidelines for getting your implementation accepted into the main repository:
-
Your implementation needs to be complete enough to self-host. This means that all the mandatory tests should pass in both direct and self-hosted modes:
make "test^[IMPL_NAME]" make MAL_IMPL=[IMPL_NAME] "test^mal"
You do not need to pass the final optional tests for stepA that are marked as optional and not needed for self-hosting.
-
Your implementation should follow the existing mal steps and structure: Lisp-centric code (eval, eval_ast, quasiquote, macroexpand) in the step files, other code in reader, printer, env, and core files. See code layout rationale above. I encourage you to create implementations that take mal in new directions for your own learning and experimentation, but for it to be included in the main repository I ask that it follows the steps and structure.
-
Your implementation should stick as much as possible to the accepted idioms and conventions in that language. Try to create an implementation that will not make an expert in that language say "Woah, that's a strange way of doing things". And on that topic, I make no guarantees that the existing implementations are particularly idiomatic in their target languages (improvements are welcome). However, if it is clear to me that your implementation is not idiomatic in a given language then I will probably ask you to improve it first.
-
If you are creating a new implementation for an existing implementation (or somebody beats you to the punch while you are working on it), there is still a chance I will merge your implementation. If you can make a compelling argument that your implementation is more idiomatic or significantly better than the existing implementation then I may replace the existing one. However, if your approach is different or unique from the existing implementation, there is still a good chance I will merge your implementation side-by-side with the existing one. In that case I will add your github username as a suffix to the language implementation directory. At the very least, even if I decide not to merge your implementation, I am certainly willing to link to you implementation once it is completed.
-
You do not need to implement line editing (i.e. readline) functionality for your implementation, however, it is a nice convenience for users of your implementation and I personally find it saves a lot of time when I am creating a new implementation to have line edit support early in the process.
Good questions that either don't have answer or need more detail