Mistakes in programming language design

http://beza1e1.tuxen.de/articles/proglang_mistakes.html

by SlyShy • 16 years ago

39 44 16 years ago

beza1e1.tuxen.de

singular4• 16 years ago

I totally disagree on the parser-unfriendly syntax point (no. 1). C++ is a particularly bad example in that code cannot be definitely parsed without semantic information, however I think limiting yourself to a language which can be expressed as LALR/LL(1) is pretty brain dead - you limit yourself to what a given algorithm can express clearly rather than what is clearest to the programmer. It's really actually pretty hard to make a language "naturally" LALR anyway, at least without hacks of some kind. C# definitely isn't LALR, there are many features which get in the way of LALR-ness, such as generics which are, however, useful.

I do agree on every other point, however!

joe_the_user4• 16 years ago

And I completely agree with you here.

The progress in programming language design is going to come from eliminating human unfriendly syntax. The closer you can get to something that let's someone express their intentions without restrictions, the better. The computer is, uh, servant of a human.

There are many barriers in the way of programming languages or computers in general adapting to human functioning. But whenever someone takes the line that says humans must-adopt-to-computers, it is a fail. And that fail is going to be swept away by the next technology which just works.

antichaos2• 16 years ago

Programming languages force programmers to think clearly and express ideas unambiguously. It's not a bad thing we should avoid; it helps us write robust software.

brazzy• 16 years ago

The progress in programming language design is going to come from eliminating human unfriendly syntax. The closer you can get to something that let's someone express their intentions without restrictions, the better. The computer is, uh, servant of a human.

And a servant isn't very useful if he can't understand its master. Ignoring machine-friendliness in language design just leads to worse and more buggy tools (including compilers).

whenever someone takes the line that says humans must-adopt-to-computers, it is a fail. And that fail is going to be swept away by the next technology which just works.

Except the "just works" is far more likely to be actually achieved if it's easy to make it work as well as easy to understand.

Besides, "human-friendly vs. computer-friendly" is a false dichtomy. There are certainly things that are good for one and bad for the other, but language features can also be good for both, or very good for one while incurring very small disadvantages to the other.

gte910h• 16 years ago

I completely agree with the parser friendly syntax (and disagree with you).

The issue is that tools are important. C++ is famously tool unfriendly, when it in fact, has the most need for tools to verify it.

From my understanding of the generics problem, generics do not get in the way of LALR requirements. You just have to implement them slightly differently than C# does to make the tokens for the generic types more identifiable. So you can do generics, just they can't look like C++ makes them look.

shadowfox• 16 years ago

It would be interesting to see how often code is written for the language vs how often a parser is written for the language before tuning the syntax for one or the other.

Also I suspect that a sufficiently well designed parser api, even if it internally uses quirky non LALR parsing techniques, will get a lot of tools going without the need for rewriting a parser each time.

singular4• 16 years ago

That is not the only problem I had with writing a LALR C# grammar. Anything even slightly touching anything else that is used for other stuff caused endless shift/reduce reduce/reduce conflicts. Not to mention I had to totally mangle the intuitive representation of a C# grammar to fit LALR.

Forcing tool writers to have to do the sort of shit I did, even without generics, is not the way forward.

The alternative of a fully LALR-able language is not necessarily desirable either - 'I want to add a feature which makes my language far more usable, oh hang on, that screws the LALR. Oh well, let's scrap that then'.

How about looking at LALR alternatives?!

antichaos2• 16 years ago

C++ is a pertinent example of terrible syntax design. Not only does the complexity make it a nightmare to write a standard compliant compiler, it also slows down development because you need much non-contextual information to read code.

binaryfinery4• 16 years ago

Yeah, Ok. http://www.youtube.com/watch#!v=v9kTVZiJ3Uc

"A keyboard? How quaint!"

Sure, it would be great if you could just describe what it is you are after, except for two problems:

1. Programmers out of a job. 2. How good are you at describing what you want? In English? You think an AI will be able to do it better than a lawyer?

What has this got to do with your point? Well, until that time, your productivity is going to be determined by the quality of your tools. You are going to need to iterate, test and refactor. To do that you need good tools, and good tools are hard to write for non-easily parsed languages. Compare the refactoring tools for C++ (limited) and Objective-C (ha ha) vs C# or Java. My productivity is double, tripple in Java and C# because I'm not spending my time moving characters around files - I'm programming.

http://blog.binaryfinery.com/?p=250

chc• 16 years ago

You're comparing four relatively similar C-family languages — it's not too surprising that tool support is the biggest difference here. Throw in Ruby or F# here if you really want to compare language quality vs. tool support.

jmillikin4• 16 years ago

> Dynamic languages are fast enough to implement internet services and outgrow the demeaning term "scripting language".

This is never going to happen, because "scripting language" is used as a slur or insult, rather than a term with any usefully defined meaning. It's not possible for Python to "outgrow" being a scripting language as long as that's used as a condescending shorthand for "doesn't look like C++".

Consider Ruby. How many applications do you know of which use Ruby for scripting? Or Python -- I think my system has two applications which can be scripted in Python (Gimp and Blender), but many dozens of applications written in Python. If Ruby and Python are scripting languages, then so are Smalltalk, Haskell, Boo, or dozens of other high-level languages. And yet, you'll never hear of these being called "scripting languages" because nobody has an axe to grind against them.

------------

Unrelated to that, Haskell does have pointers (including null pointers) -- you can use them just like pointers in C/C++.

j_baker2• 16 years ago

"This is never going to happen, because "scripting language" is used as a slur or insult, rather than a term with any usefully defined meaning."

This isn't necessarily true. See Ousterhout's dichotomy: http://home.pacbell.net/ouster/scripting.html

silentbicycle6• 16 years ago

I agree with you, but I think jmillikin's point is when most programmers hear "scripting language", they think "dumbed-down programming language" rather than "cleanly separated layers of design".

jmillikin4• 16 years ago

That's a really wonderful link, and I wish more people used the definition it advocates. I've one minor complaint, in that it conflates typeless and dynamically-typed languages, but considering when it was written that's not a big deal -- strong, dynamically-typed languages probably didn't exist in any significant way.

Unfortunately, almost every use of "scripting language" I've ever heard is more along the lines of:

"We can't write our GUI frontend in Python^! Scripting languages are too slow for rendering advanced graphics."

Usually uttered by somebody who's been writing C++ for 20 years and refuses to touch anything more recent.

^ Substitute Boo, Haskell, Scala, Clojure, or anything else without enough curly braces

j_baker2• 16 years ago

I don't necessarily agree with #2. Yeah, SML has formalized syntax, but it's also a frozen language. It simply isn't practical to have completely formalized semantics for a language like Python or Ruby that are still actively being changed. Nor is it necessarily always possible to eliminate all implementation-specific quirks.

In short, it's a good thing to have a language have few quirks, but there is such thing as overspecifying.

silentbicycle6• 16 years ago

These things don't need to be taken to their logical extremes, though - handling most of the semantic edge cases is still an improvement, even if it isn't exhaustive.

Emphasizing well-defined semantics could also be a push to keeping the language small and clean.

RodgerTheGreat• 16 years ago

It's possible that I'm missing the thrust of #0, but I don't see how an object-oriented language completely without the concept of a "null pointer" could be used to construct elementary data structures like linked-lists and trees without resorting to special classes as terminators. It would be just as clumsy as using null pointers in the first place.

If "not-null" references were provided in addition to normal references, the effect would be pretty similar to the common usage of "final" fields in Java- compiler-enforced initialization.

Groxx2• 16 years ago

I was wondering that too, so I looked up the linked "Cyclone" language. If I'm understanding what I read there, the entire #0 is pretty misleading.

Basically, it boils down to adding a "qualifier" to the language, so you can mark pointers as nullable / never-null, so the system doesn't have to check.[1] You can also mark them as "thin" (no pointer arithmetic) and "fat" (safer pointer arithmetic), all of which is accomplished by putting a "@nullable" / "@notnull" / etc. qualifier on the pointer when you declare it (I think). @nullable does a check before each access, @notnull does not.

So, it's C + super-static-types to make C somewhat safer. At the cost of making the language significantly more complex. But still not safe, as this is put there by the programmer - people make mistakes. And C programmers seem especially vulnerable to making things unsafe "for speed"... what's to stop an overzealous optimizer from marking more pointers as @notnull?

[1] http://cyclone.thelanguage.org/wiki/Pointers

ekiru• 16 years ago

If the compile can't statically prove that a pointer can't be null, any attempt to convert it to *@notnull inserts a runtime assertion that it isn't null, according to the website.

silentbicycle6• 16 years ago

> It would be just as clumsy as using null pointers in the first place.

When you combine this with type-inference, it's actually really convenient. Check out the Maybe monad in Haskell or the option type in OCaml - nullable types are boxed in a variant type of None | Some 'a, you're reminded to check for both possibilities in the few places where it actually matters, and you can quit wasting time and bloating your code writing null tests, forever.

You can define special sentinel subclasses of things in OO languages, but there's usually quite a bit of boilerplate compared to just adding a variant type in ML (and the compiler still can't guarantee that nulls will never occur), so people don't usually bother.

Chris_Newton2• 16 years ago

There are alternatives to making references nullable by default.

For example, functional programming languages often support algebraic data types and pattern matching. Creating the equivalent of a nullable value is one possible application of these features. Typically, the type system would make sure you can't get to the underlying value without first checking that your data is non-null via pattern matching, so the equivalent of derefencing a null pointer simply isn't possible without generating a compile-time error.

Note that this idea is independent of the underlying type, so there's no particular reason you couldn't have a type system that could wrap an OO class type in an algebraic data type. Unfortunately, for now, these ideas seem to live in different worlds: I know of no mainstream OO language that has anything close to the power of algebraic data types and pattern matching that is commonly available in functional programming languages. But this is one way you could have an OO-friendly language with powerful data structures and without relying on the idea of a nullable reference.

tel• 16 years ago

This was the first thing I thought of. Addition types with pattern matching (which enable Maybes) both provide safe options for these data types. An example Haskell type that keeps the surface idea of nulls (via Maybe having an object) might be something like:

  List a :: Cons a (Maybe (List a))
  Tree a :: Branch (Maybe (Tree a)) (Maybe (Tree a))

The type system removes the concept of 'dereferencing nulls' while maintaining that if the pointer isn't null it still must be the (polymorphic) type a.

Of course, the simpler implementation unfolds the addition type inside Maybe and is more like:

  List a :: Cons a (List a) | End
  Tree a :: Branch (Tree a) (Tree a) | Leaf a

silentbicycle6• 16 years ago

Pattern matching seems to mix really poorly with conventional OO subtyping / inheritance, though multi-methods seem like a good synthesis. (I haven't really used CLOS enough to decide, though.)

InclinedPlane• 16 years ago

Welcome to the idea of the "null object". It's an object which satisfies all of the use cases of the null value but doesn't cause you to litter nearly every line of your code with defensive null checks to avoid your program randomly barfing. When done right it's a far more elegant solution than "null".

grogers• 16 years ago

Linked lists are a particuarly poor example though. It is just as easy to make them without null pointers as with null pointers:

struct list_node { list_node *next; };

struct list { list_node sentinel; };

IE. A circular linked list. In C++ speak, begin() is sentinel.next, end() is sentinel. In an empty list, sentinel.next points to &sentinel.

It might seem more clumsy, but it isn't really any more work. When you actually do need the equivalent of a null pointer you can use an option type to specify that you either want nothing or a valid pointer.

Mistakes in programming language design

Related Stories