Comment by chubot | Hacker News Reader

Comment by chubot | Hacker News Reader

Parent Full thread

(author here)

Well, a key point is that it doesn't have to be an entire Python interpreter. It just has to be the subset that Oil uses. (This applies to both both Python syntax and semantics.)

I also forked the 8K LOC bytecode compiler -- I didn't write it from scratch. That's the point of "Cobbling together a Python interpreter".

So for a cartoonish view, compare these implementation strategies.

1. Clone bash in C, which (charitably) would involve writing around 80K lines of C code from scratch, given that bash is 160K lines. You can probably do better by not writing it over 30 years :-)

2. Write 16K lines of Python, fork a bytecode compiler in Python (8K lines), and then either fork CPython's VM or write your own.

The post implies that #2 is probably easier than #1. I haven't done the last part of #2, so I could be wrong. I think it's an interesting experiment to find out whether you can write a bespoke compiler/VM for a specific program in a high level language.

-----

It's a little bit like TeX and Pascal. I discovered awhile ago that that TeX is written in an abstract subset of Pascal. Apparently the version you use on most Linux distros is compiled to machine code via a Pascal-to-C translator. I'm not sure if that translator was written specifically for TeX or not, but if anyone knows, please chime in!

But I have seen enough of Knuth's code to know that he does not rely on implementation details. That is roughly the case with Oil and Python. As I mention in the post, Oil is written in Python+ASDL, not Python. And I would even call it Python + ASDL + some abstract regex subset. That subset could be called the intersection of Python regexes + re2c regexes, e.g. at the end of this comment: https://news.ycombinator.com/item?id=16525378

Basically when writing Oil, I try to be clear about the algorithms, and not think in terms of implementation details of a particular platform. This is possible for a shell because it's mostly string handling, and the only libraries it uses are a handful of libc calls (which are 40 years old, standardized, and well understood.)

> Clone bash in C, which (charitably) would involve writing around 80K lines of C code from scratch, given that bash is 160K lines.

Or, third option, you could have used a language like Rust that would allow you to write using a high-level programming language, and yet still you could dip down to low-level code whenever needed for performance reasons.

Don't get me wrong, I'm a big Python fan for certain types of projects. And now with Python 3 and asyncio, Python is great for writing I/O-bound code. But it's not ideal, even if you confine your use to a Python subset, for a project for which parsing is a critical part.

Finally, you may prove me wrong, but even using a subset of Python, I don't think you're going to eke out the performance you're looking for using CPython's VM. I say this simply because it's not a very performance-oriented VM (partly because of the bytecode instruction set but also because it's stack- instead of register-based, and some other architectural issues). So then you're back to writing your own VM from scratch -- a significant undertaking.

In any case, I wish you luck. Oil Shell's a cool project, so I hope it succeeds.

I find Rust more interesting than I did at the start of the project, since I ended up using algebraic data types all over the place, which Rust has. And it doesn't have the runtime issues I mentioned with Go.

But the compile times are a dealbreaker. As mentioned in the post, if I had a C++ compiler in my edit-run cycle, I would have never have gotten the project done. From what I understand having a Rust compiler in there would be even worse.

I can understand why you think Python is not ideal, and I listed many reasons in the post why it's not. But that is why I used a name OPy -- OPy is purposefully diverging Python. See the Tex analogy.

I agree it's an experiment. I didn't make any claims in the post; I specifically mentioned that I would like to prove a point.

I agree that I will need to change the Python VM. The speedups can be unlimited in that case that I write an entirely new VM (which is easier than rewriting the very flexible and capable Python VM from scratch). Another thing I didn't mention: it is probably possible to make the shell VM and OPy VM converge. So Oil's runtime could be faster than any other shell, because every shell is a tree interpreter and not a bytecode VM. (The parser might still be slower, but I don't expect that to be an issue.)

----

EDIT: On the other hand, I doubt you can write bash in Rust in less than say 50K to 80K lines of code. I think Rust will still be at least 3x more verbose than Python. It inherently expresses more, so this is fundamental. So even if compile times were not an issue, I still probably wouldn't choose Rust.

Also, build dependencies matter a lot for a shell, since it's used to bootstrap embedded systems. I conjecture that shells run all sorts of systems that Rust programs have never run on. Rust isn't as portable as C, or C++.

> I conjecture that shells run all sorts of systems that Rust programs have never run on.

Rust now runs on all kinds of embedded systems, including MSP, most current ARM architectures, MIPS, PowerPC, and even (via forked version of Rust soon to be merged) AVR. So it can run on most, if not all, major architectures that shells run on.

Since you asked: Yes, the Pascal-to-C translator was written specifically for TeX.

To be more precise (and add more detail than anyone asked for!):

- TeX is not written in Pascal itself, but in a literate-programming system called WEB [1,2] (a language basically created for writing the TeX program in!), which lets you mix Pascal source code with TeX documentation, and lets you write your program as a “web” of “sections”, each independently understandable, and referring to other sections (with macros etc). The “tangle” program converts the WEB source code to Pascal source code.

- The Pascal used in the TeX program is a limited subset [3] of Pascal (and specifically of what Knuth called “Pascal-H”, the Pascal implementation that was available to him on the DEC PDP-10 system used at Stanford SAIL).

- It is not exactly true that he does not rely on implementation details, but he's conscious about where he does so, and he's included a list of all such places in the index under “system dependencies” [4] — in most places he gives suggestions on what could be done if the compiler didn't work the same way.

- In the early days, when Pascal was the most common language available at the places TeX was usually run (universities, research labs, etc.), the Pascal code (generated by Tangle) was directly compiled into a Pascal program and run.

- With the rise of C (into the wider world outside Bell Labs), there arose both hand-translations of the WEB (or Pascal) source code, and programs to do this automatically.

- Today, major TeX distributions have their own Pascal(WEB)-to-C converters, written specifically for the TeX (and METAFONT) program. For example, TeX Live uses web2c[5], MiKTeX uses its own “C4P”[6], and even the more obscure distributions like KerTeX[7] have their own WEB/Pascal-to-C translators. One interesting project is web2w[8,9], which translates the TeX program from WEB (the Pascal-based literate programming system) to CWEB (the C-based literate programming system).

- The only exception I'm aware of (that does not translate WEB or Pascal to C) is the TeX-GPC distribution [10,11,12], which makes only the changes needed to get the TeX program running with a modern Pascal compiler (GPC, GNU Pascal).

[1]: http://literateprogramming.com/knuthweb.pdf

[2]: http://mirrors.ctan.org/info/knuth/webman.pdf

[3]: http://texdoc.net/texmf-dist/doc/generic/knuth/tex/tex.pdf#p... (section 3)

[4]: http://texdoc.net/texmf-dist/doc/generic/knuth/tex/tex.pdf#p... (entry “system dependencies”)

[5]: https://tug.org/web2c/

[6]: http://tug.org/interviews/schenk.html

[7]: https://tex.stackexchange.com/questions/111332/how-to-compil...

[8]: https://w3-o.cs.hm.edu/~ruckert/web2w/index.html

[9]: https://ctan.org/pkg/web2w

[10]: http://www.tex.ac.uk/FAQ-sysunix.html

[11]: http://mirrors.ctan.org/systems/unix/tex-gpc/README

[12]: http://mirrors.ctan.org/systems/unix/tex-gpc/tex.pdf

Thanks for the details! It is amazing to me that TeX has lasted so long. I would conjecture that this abstract style and the resulting portability are a big part of it.

I didn't know the name Pascal-H -- that is analogous to "OPy" in my mind. Although OPy I expect it to diverge from Python for efficiency reasons -- that's the whole point. It will probably look superficially like Python for a long time.

Why are there so many Pascal(WEB)-to-C converters? Why wouldn't they all use the same one? I guess I don't understand what TeX distribution is. Are there multiple maintainers? I would have thought they would have converged over time.

(FWIW the "system dependencies" docs are exactly what I meant. Most programs are littered with implementation details; there is no clear separation between them and their environment. But it sounds like Knuth is highly aware of them, and localizes them within the code. Obviously you need system dependencies to get anything done.)

I don't know why everyone doesn't use the same converter! I haven't looked into any of them in detail… As for what a TeX distribution is: it's all the stuff (beyond the source code of TeX itself, which is maintained solely by Knuth) needed to get TeX and related systems in a usable form. This includes the to-C converter and build scripts, for TeX and for other programs/extensions like pdfTeX, the major macro packages like LaTeX and “format files” for them, font information for commonly used fonts, lots of “packages” written by lots of people over the years, hyphenation rules for languages, other programs like makeindex etc., DVI viewers and converters, TeX editors — the entire ecosystem that a typical TeX user may want. (Some distributions more minimal than others… TeX Live and MiKTeX are probably the two main ones today.)

Reply on news.ycombinator.com