Comment by chubot | Hacker News Reader

Comment by chubot | Hacker News Reader

Parent Full thread

I will be immodest and say that I reached that stage a long time ago. Because, for better or worse, I learned CS before I really learned to program :-/ I did program in middle/high school, but not very often, and I was bad it. I was better at reasoning about data structures than debugging, and I think the latter skill is "table stakes" before calling yourself a programmer.

As a concrete example of thinking outside a particular language, My shell Oil is 24K lines of code that does most of what bash does at 140K LOC, written not in Python, but:

1. OPy, a subset of Python 2 -- http://www.oilshell.org/blog/2018/03/04.html#faq

2. Zephyr ASDL -- http://www.oilshell.org/cross-ref.html?tag=zephyr-asdl#zephy...

3. a mathematical dialect of regular expressions (that runs under both Python's re engine and re2c's dialect, via translation).

In other words, it's a composition of DSLs, independent of any language.

I started a translator from OPy (typed with MyPy) to C++. It partially works but doesn't translate the whole codebase yet. If it works then I will have fully achieved this "language abstraction" goal.

It's analogous to how TeX was implemented by Knuth. TeX is not implemented in Pascal and compiled with a Pascal compiler. It's implemented in H-Pascal, and abstract subset of Pascal. Then it's translated to C and compiled with a C compiler!

https://news.ycombinator.com/item?id=16526151

In case it isn't obvious, this makes the style of code VERY different and more abstract than what you see in every other codebase.

This approach has downsides -- namely that the shell is way too slow right now. Hopefully the translation will fix that. Either way, I can definitely claim that the ideas and architecture are completely separate from any programming language. I've documented a lot of these ideas on the blog:

http://www.oilshell.org/blog/

Interesting, is there a name for this subset pattern? I write in subsets to. Basically simplify the language slightly to make it easier to build ad hoc parsers and code generators for the problem domain. Linting on steroids, perhaps.

I don't call it this, but maybe:

https://en.wikipedia.org/wiki/Language-oriented_programming

i.e. composing different DSLs to solve a problem. The codebase is separated into parts and each part may be expressed in a different language.

I tend to call it "metaprogramming", which might be vague but is captures the spirit IMO. Metaprogramming how you implement bespoke DSLs. I categorize textual code generation as the most basic form of metaprogramming, i.e. programming where your data is code.

It's one level removed -- rather than talking about the problem, you're talking about the tools/language/construct you're using to solve the problem.

And yes I would say the other downside is that it's easy to go off the deep end :) But I think that certain problems are difficult and you need some leverage to solve them. For example, writing a bash-compatible shell would be extremely repetitive otherwise. It's like a dozen different ways of groveling through backslashes and braces one-by-one.

I'm a big fan of Language Oriented Programming and Racket, but I guess here what I am talking about is an "in moderation" flavor of that, where you are not building new languages by addition, but rather by subtraction (sticking to a smaller subset of an existing language). It's certainly the same idea, but this one is more of a pragmatic approach that hopefully helps you avoid falling into unanticipated complexity traps that you encounter when building a new language from the bottom up.

That's true -- I don't know of a name for it. There should be one!

Someday I would like to define the subset of HTML I actually use... HTML is messy with lots of implementation quirks. And plenty of tools operate on it in ad hoc ways.

It would be nice to have some notion of correctness for those tools, and defining a subset seems like the best way to do that.

Anytime there's a pattern in language design I don't know the name for I turn to my monoid-loving coworker next to me and ask what the horribly complicated term for it is in Haskell.

For now how about "Chop" Programming? "Choose Half" Oriented Programming. Or "Super" programming for "Subtract Unwanted Programming Elements Rigorously".

HTML5 is supposed to be a specification of HTML, with conformance tests and everything.

Have you got an article where you talk a bit about point #3 - "a mathematical dialect of regular expressions". The words all make sense, but I cannot picture it :-). Your Oil posts always appear up on lobste.rs shortly after they're posted and I thoroughly enjoy them, but I must have missed a few!

I don't know if this is relevant to what Chubot's doing but Brzozowski’s Derivatives of Regular Expressions are pretty mathy. (Disclosure, I wrote this) http://joypy.osdn.io/notebooks/Derivatives_of_Regular_Expres...

That's cool -- I have a big page of notes on derivatives and have thought about replacing my usage of re2c with them (re2c being automata-based).

One benefit of that is aesthetic -- re2c is a 30K line piece of C code itself. The other benefit is practical -- it would be nice to "hoist" it up to the Oil language level, so shell users can use efficiently compiled regexes.

And I filed this issue to try derivatives awhile ago:

https://github.com/oilshell/oil/issues/197

There is a possible performance win, but I'm not sure if it makes sense. If you see anything interesting there I'd love to chat about it! (contact info in profile)

-----

What I mean by "math" is either automata-based methods or derivatives. In contrast, Perl-style backtracking engines aren't math.

To me, the "magic" of "regular languages" is that they give you non-determinism for free. The equivalence between NFAs and DFAs isn't obvious, and it's useful in practice.

I've taken advantage of this free nondeterminism in my huge shell lexer. (See my other reply for details)

I'm enamored of dRE (as I call it) but pragmatically it seems to me you could stick with re2c.

I think it's interesting to generalize dRE to general control-flow (not just parsing.)

Are you aware of Abstract State Machines? https://en.wikipedia.org/wiki/Abstract_state_machine

I've discussed related topics but haven't had the chance to address it directly. I'm glad someone is paying attention :)

Concretely, the first two links in this post show (old versions of) frontend/lex.py and osh-lex.re2c.h. TODO for me: put up the latest versions, as well as the huge C file with state machines that re2c eventually generates.

When Are Lexer Modes Useful? http://www.oilshell.org/blog/2017/12/17.html

It works like this:

    ( ) frontend/lex.py (a bunch of Python regexes, has some "metaprogramming") ->
    (+) frontend/lex_gen.py ->
    ( ) osh-lex.re2c.h (re2c input file) ->
    (+) re2c ->
    ( ) osh-lex.h (state machines in C, i.e. DFA as a big  switch/goto)

where the (+) nodes are compilers, and the ( ) are source code files.

The reason I call this "a mathematical dialect" is because the same regular expressions run under Python's re engine (a backtracking engine) and as native code via re2c, an automata-based compiler.

If you scroll toward the bottom of this doc there's a useful table:

Regex Theory and Practice http://www.oilshell.org/share/05-31-pres.html

One side is "Perl-style regexes", which Python's engine is based on. The other side is "regular languages". Regular expressions were always mathematical, but the name got taken over by programmers to mean something different, so I call them "regular languages" or this "mathematical dialect".

These articles explain the difference,

https://swtch.com/~rsc/regexp/

but they're very long and a lot of people still don't understand the difference. https://news.ycombinator.com/item?id=20311630

(which is understandable since it mainly comes up in performance corner cases, and when you want to compile regexes, which most people don't do)

I should write about this, but the lexer is one of the more solid pieces of the project. That is, it's "done" for now, and I need help with all the other parts, so I prioritize writing about those parts!!!

More lexing posts: http://www.oilshell.org/blog/tags.html?tag=lexing#lexing

Reply on news.ycombinator.com