http://www.ai.mit.edu/~gregs/ll1-discuss-archive-html/msg04323.html

Copyright © 2003, Scott McKay , all rights reserved (based on a post to lll-discuss and wikified by nik gaffney )

About 8 or 9 months ago, there was a thread on “how to design a language”. I offered some thought, other people chipped in, and I volunteered to collate the results. This is as good an excuse as any to post what's there so far. I never really finished it, so it's unintentionally a bit folksy…

Here are some questions a prospective language designer should ask himself when starting the designing a programming language.

What need are you trying to fill? Don't fall into the trap of “a scripting language”, because they always turn into general-purpose languages.

What's the metaphor? Even though you might not be trying to build a “pure” language, it's worth having a model for the core language, such as “imperative, block-structured” (C), “object oriented” (Smalltalk), “generic object orientation” (Lisp), “functional” (ML), “lazy” (Haskell), “logic” (Prolog), “production system” (OPS5), etc. These different core models influence the “natural” styles of program development in different languages even if the set of available facilities is similar. They also help define which late-arriving features will “fit” and which will be warts. [Jerry Jackson]

How many programming paradigms does your language support? How tightly are they integrated? Which other paradigms can you integrate with the built-in facilities? How natural is the syntax of user-defined extensions? Many problems are much better suited to some non-standard programming model than to the usual object-oriented/functional approaches. For example, constraint languages allow a very concise description (and efficient solution) of many optimization problems. Dylan supports functional and object-oriented programming in a tightly integrated manner, but it offers no support for non-deterministic programming, constraint-solving, etc., and not much support to add them to the language. If you have first-class continuations in the language you can add one additional programming model that requires non-standard control flow, but, in general, different extensions based on call/cc don't work together. [Matthias Holzl]

Is high performance an issue? This says something about whether you want to implement an interpreted, a VM-based, or a natively compiled language.

Is high programmer productivity an issue? How important is this with respect to performance? This decision can affect how you store values and do function calling.

How portable across platforms do you want the language to be? This will relates to whether you want to compile to a VM or to machine code, and to how well you support native libraries. It will also affect library design for such things as graphics and GUI tools. [Anton van Straaten]

Do you want easily distributed executable code, i.e., do you you want to allow code to be easily transmitted across networks and run elsewhere, as Java does? Do you want to provide built-in support for remote execution, like RPC/CORBA/RMI? If you are writing for a VM, this can simplify some of these issue considerably. [Anton van Straaten]

What about debuggability? If you plan to compile it, you need to think about how to store debugging information.

How do you want to bootstrap it? This, too, says something about what kind of back-end you might build. Perhaps you build a tiny VM in C, then compile to C. This way, you avoid fun but time-consuming work on code generation for modern super-scalar hardware, register allocation, etc.

Do you want to be able to catch type errors early or late? That says something about your type system (whether you require that all types be statically declared at compile-time, or allow them to be dynamic, or have a hybrid scheme like Dylan does). In addition to the obvious effect on performance, this decision will affects your memory model in that completely static systems do not require tags or boxing.

Will variables be associated with explicit type declarations? - If yes, will these type declarations be required or optional? - If optional, will the language use inferencing to supply unspecified types, or simply use an all-purpose type (like Object or 'any')? [Anton van Straaten]

Will the language have any run-time type discrimination/checking at all, or will types be completely statically determined? Some languages considered statically typed still do some run-time checking, such as Java. [Anton van Straaten]

Will any type checking happen at compile-time? Some languages with explicit type declarations don't always check types at compile-time, such as old Visual Basic. [Anton van Straaten]

If you allow type declarations, you will want to think about whether you want parameterized types. If you go whole hog with, say, F-bounded polymorphism, you can get performance and type safety and ease of use, but it's hard to get this exactly right.

What about namespaces? Do you want to have a simple scheme as in Java, where classes, namespaces, and files are roughly equivalent? Lisp-style packages? Dylan-style modules and libraries? Within a single first-class namespace, how many second-class namespaces are there? Java has 7 or 8: class names, function names, local variable names, slot names, etc. Common Lisp has at least 3 (function, variable, and class names). Dylan and Scheme have one, which greatly simplifies things at a small loss of generality which can usually be worked around with name conventions.

What about encapsulation? Do you want to do information-hiding on a per-class basis as in C++ and Java, or on a “module” basis as in Dylan?

Is your language a functional language (that is, without side-effects)? If so, is it an almost-functional language or a true pure functional language? Or is there a functional core with some sort of machinery for isolating side-effects, like monads do in Haskell?

What kind of evaluation semantics does the language have? Eager as in most languages, or lazy as in Haskell?

Is your language purely lexical or do you offer dynamic variables (or, more generally, access to the dynamic environment) as well? Dynamic binding allows you to introduce local state for the duration of a computation without side effects and without adding additional parameters. [Matthias Holzl]

Are there different semantics for “pointer-ish” and “non-pointer-ish” values, like in C? Or is everything a first-class object reference, like in Lisp? Having multiple ways of referencing values can make the user mode much more complicated. On the other hand, making everything be object references can require boxing and/or tagging schemes that make your compiler and FFI more complex.

How do you want to pass arguments to functions? By name as in Algol? By value or by reference as in C? By object reference like Lisp does? Is there more than one convention in the language?

Do you want first-class functions? What about lexical closures? First-class continuations? The answer to those questions will tell you things about heap- and stack-allocation, and will also tell you how important it might be to do a continuation-based compiler. It also tells you how hard your compiler has to work to avoid consing environments unnecessarily. Lots of sophisticated language designers go with simple closures and avoid full continuations, because full-scale environment capture is hard to do well.

Does your language have an unwind-protect like facility? When you design a new language it is tempting to include call/cc because it allows you to do define many common (and uncommon) control structures. On the other hand you want to have a facility that allows you to reliably relinquish resources after you are done. If you simply try to combine call/cc and unwind-protect, you immediately get the “impenetrable shield vs. unstoppable force” problem in your language. Possible solutions include: no call/cc, weakened unwind-protect, different semantics for call/cc. [Matthias Holzl]

How do you handle conditions/errors? Return codes or signalling? Do you have an unwinding-only model like C++/Java or do you allow restarts like Dylan/CL? If you do the latter do you separate conditions and restarts like Common Lisp or do unify them like Dylan? These questions are important, because every programming language has to deal with error conditions, and in many cases the unwinding model is used simply because the language designer is not aware of any other possibilities. [Matthias Holzl]

Do you want the language to be “object-oriented” at all, given a broad definition of OO that includes the spectrum from single inheritance single receiver languages as in Java to multiple inheritance multiple receiver languages as in CLOS? Do you want to provide genericity through some sort of template scheme?

Here is how Jonathan Rees has characterized the very fuzzy term “OO”:

1. Encapsulation – the ability to hide the implementation of a type
2. Protection – the inability of the client of a type to detect its implementation, guaranteeing that any changes to an implementation that preserve the behavior of the interface will not break any clients. This also gives some measure of “security”, because things like passwords can't leak out. )
3. Ad hoc polymorphism – functions and data structures with parameters that can take on values of many different types.
4. Parametric polymorphism – functions and data structures that parameterize over arbitrary values, such as “a list of anything”). ML and Lisp both have this. Java doesn't quite because of its non-Object primitive types.
5. Everything is an object – all values are objects. True in Dylan, but not in Java because of its primitive types.
6. “All you can do is send a message” (AYCDISAM) = Actors model – there is no direct manipulation of objects, only communication with (or invocation of) them. The presence of fields in Java violates this.
7. Specification inheritance = subtyping – there are distinct types known to the language with the property that a value of one type is as good as a value of another for the purposes of type correctness. An example is Java interface inheritance.
8. Implementation inheritance/reuse – having written one pile of code, a similar pile (such as a superset) can be generated in a controlled manner, that is the code doesn't have to be copied and edited. A limited and peculiar kind of abstraction. (E.g. Java class inheritance.)
9. Sum-of-product-of-function pattern – objects are, in effect, restricted to be functions that take as first argument a distinguished method key argument that is drawn from a finite set of simple names. Some people say Lisp is OO, meaning {3,4,5,7}. Some people say Java is OO, meaning {1,2,3,7,8,9}. E is supposed to be more OO than Java because it has {1,2,3,4,5,7,9} and almost has 6; 8 (subclassing) is seen as antagonistic to E's goals and not necessary for OO. The conventional Simula 67-like pattern of class and instance will get you {1,3,7,9}, which many people take as a definition for OO. [Jonathan Rees]

If the language is object-oriented, do you want it to be class-based or prototype-based? [Steve Dekorte]

got an object system, do you want it to have first-class objects that exist in the run-time? Should the object system extend to include all the way to the primitive types, or do you want to special-case those like Java does? Do you want a Smalltalk/Java-style single receiver object orientation, or a CLOS-style multi-method generic function dispatch? If the former, do you need some sort of static overloading like C++ has? If the latter and performance is important, do you need some sort of Dylan-style “sealing” so that you can do some compile-time optimizations? Do you want single inheritance, single inheritance with interfaces, multiple inheritance, or a hybrid single inheritance with mixins? If you've got a more static type system, you'll need to deal with casts. Do you additionally want auto-conversion?

If you've got an object system, how much of a meta-object system do you want to expose? Do you want it to be purely reflective, or more than that? In Dylan, we separated 'make' from 'initialize', which was a good idea, but do you also want to separate out 'allocate', so that you have control over where an object is created, e.g., in a “persistent memory” pool that might be back-ended by a database?

Do you need hairy CLOS-style method combination, or is a simpler style like we did in Dylan enough? Do you care about what Gregor Kiczales calls “aspects”, which might change your decision?

A more general question that relates to the object system, the meta-object system, and a different dimension of the bootstrapping question is: do you want to implement a language which provides a bunch of predefined and fixed constructs (such as an object system) or do you want to provide a layered language that implements such constructs in terms of lower-level features in the language? The former is probably easier, but the latter can allow very flexible customization, which tends to be traded off against standardization. Note that even a language with a powerful built-in meta-object system won't necessarily allow you to replace that object system with something else, for example, unless the language supports that sort of thing. [Anton van Straaten]

How do you want to do memory management, manual or automatic (GC)?

Do you want to support threading? Do you want to roll your own threads or use OS threads? Do you want to support massive concurrency like Erlang does? The answers to those questions will tell you about aspects of the run-time, memory allocation/GC, and performance. Oh yeah, it also tells you if you can actually take advantage of the multiple processors sitting in most of the machines we all have. Do you want Java-style synchronization where it is built in to objects, or should that be handled orthogonally?

If you have threads and continuations, how do they relate to each other?

How well do you want to be able to integrate with native libraries? This decision affects your memory model, how you plan to represent run-time type info, how function call/return works, how signalling works, etc. By “memory model”, I also mean to include what sorts of objects are boxed or tagged. (Opinion: the Harlqn/FunO Dylan compiler got it wrong – I think we should have boxed everything, and then concentrated our efforts on box/unbox optimizations. This would have hugely simplified FFI issues.) Good integration with native code probably means that you will end up using a conservative collector, and that will effect the semantics of “finalization” (if you have it).

Do you want to be able to return multiple values? How about &rest arguments? These affect function call/return, tail-call elimination, and stack vs. heap allocation optimizations.

What's your order of evaluation in expressions? This affects what sort of optimizations can be safely done.

What compilation model do you want? Lots of include files like C[++]? Lots of “packages” like Java? Whole-worlds like Lisp? Separate libraries like Dylan? This affects a lot of things, not least of which is the ability to deliver small applications. It also informs the design of your core run-time.

Is the core run-time tiny like Scheme's? Small like Dylan's? Huge like Common Lisp's? If you like the Common Lisp model, it's worth looking at EuLisp to see how to re-package it in a more layered way.

Even in a small run-time, you need to get the basic types right. Are your numeric types “closed” (that is, do they include reals – rationals and irrationals – and complex numbers)? Are your string and character types rich enough to model Unicode?

Think hard about collections. How do the following relate to each other: sets, tables, vectors, arrays, lists, sequences, ranges? In Dylan, we decided too late having the tail of a list be a “cons” was maybe not such a great idea; what about that? How do your collections interact with your threading model?

Think hard about iteration, especially over collections. If all collections obey a uniform iteration protocol, it means that you can do things like 'for e in c …'. Note that if iterators are done in a first-class way, this has performance implications that your compiler needs to worry about.

Do you want some sort of security model built into the language? What sort of model do you want to use? A simple “checker” like the Java VM uses, or a more sophisticated capability-based model.

What syntax do you want? Parentheses unaccountably give lots of people hives, but S-expressions make a lot of things much simpler. Infix syntax is quite nice when it's done well, but you've got to get the “kernel” of that exactly right if you want your infix macro system ever to be usable. If you decide on S-expressions, should they be represented as lists and conses, or do you want a first-class object for that?

Do you want to allow syntactic extensions (macros)? Lisp-style macros? Dylan-style pattern-matching non-procedural hygienic macros? Scheme-style 'syntax-case' pattern-matching procedural hygienic macros? This says a lot about the syntax of your language, and it also says a lot about the model you choose for compile-time evaluation environments.

Contributors

• Andrew Cooke
• Steve Dekorte (*)
• Matthias Holzl (*)
• Jerry Jackson (*)
• Jonathan Rees (*)
• Anton van Straaten (*)

(*) Contributions marked above