Announcing code annotations for SourceHut July 8, 2019 on Drew DeVault's blog

Today I’m happy to announce that code annotations are now available for SourceHut! These allow you to decorate your code with arbitrary links and markdown. The end result looks something like this:

NOTICE: Annotations were ultimately removed from sourcehut.

The annotations shown here are sourced from a JSON file which you can generate and upload during your CI process. It looks something like this:

{
  "98bc0394a2f15171fb113acb5a9286a7454f22e7": [
    {
      "type": "markdown",
      "lineno": 33,
      "title": "1 reference",
      "content": "- [../main.c:123](https://example.org)"
    },
    {
      "type": "link",
      "lineno": 38,
      "colno": 7,
      "len": 15,
      "to": "#L6"
    },
    ...

You can probably infer from this that annotations are very powerful. Not only can you annotate your code’s semantic elements to your heart’s content, but you can also do exotic things we haven’t thought of yet, for every programming language you can find a parser for.

I’ll be going into some detail on the thought process that went into this feature’s design and implementation in a moment, but if you’re just excited and want to try it out, here are a few interesting annotated repos to browse:

• ~sircmpwn/scdoc: man page generator (C)
• ~sircmpwn/aerc: TUI email client (Go)
• ~mcf/cproc: C compiler (C)

And here are the docs for generating your own: annotations on git.sr.ht. Currently annotators are available for C and Go, and I intend to write another for Python. For the rest, I’ll be relying on the community to put together annotators for their favorite programming languages, and to help me expand on the ones I’ve built.

Design

A lot of design thought went into this feature, but I knew one thing from the outset: I wanted to make a generic system that users could use to annotate their source code in any manner they chose. My friend Andrew Kelley (of Zig fame) once expressed to me his frustration with GitHub’s refusal to implement syntax highlighting for “small” languages, citing a shortage of manpower. It’s for this reason that it’s important to me that SourceHut’s open-source platform allows users large and small to volunteer to build the perfect integration for their needs - I don’t scale alone¹.

To get a head start for the most common use-cases - scanning source files and linking references and definitions together - the best approach was unclear. I spent a lot of time studying ctags, for example, which supports a huge set of programming languages, but unfortunately only finds definitions. I thought about combining this with another approach for finding references, but the only generic library with lots of parsers I’m aware of is Pygments, and I didn’t necessarily want to bring Python into every user’s CI process if they weren’t already using it. That approach would also make it more difficult to customize the annotations for each language. Other options I considered were cscope and gtags, but the former doesn’t have many programming languages supported (making the tradeoff questionable), and the latter just uses Pygments anyway.

So I decided: I’m going to write my own annotators for each language. Or at least the languages I use the most:

• C, because I like it but also because scdoc is the demo repo shown on the SourceHut marketing page.
• Python, because SourceHut is largely written in Python and using it to browse itself would be cool.
• Go, because parts of SourceHut are written in it but also because I use it a lot for my own projects. I also knew that Go had at least some first-class support for working with its AST - and boy was I in for a surprise.

With these initial languages decided, let’s turn to the implementations.

Annotating C code

I began with the C annotator, because I knew it would be the most difficult. There does not exist any widely available standalone C parsing library to provide C programs with access to an AST. There’s LLVM, but I have a deeply held belief that programming language compiler and introspection tooling should be implemented in the language itself. So, I set about to write a C parser from scratch.

Or, almost from scratch. There exist two standard POSIX tools for writing compilers with: lex and yacc, which are respectively a lexer generator and a compiler compiler. Additionally, there are pre-fab lex and yacc files which mostly implement the C11 standard grammar. However, C is not a context-free language, so additional work was necessary to track typedefs and use them to change future tokens emitted by the scanner. A little more work was also necessary for keeping track of line and column numbers in the lexer. Overall, however, this was relatively easy, and in less than a day’s work I had a fully functional C11 parser.

However, my celebration was short-lived as I started to feed my parser C programs from the wild. The GNU C Compiler, GCC, implements many C extensions, and their use, while inadvisable, is extremely common. Not least of the offenders is glibc, and thus running my parser on any system with glibc headers installed would likely immediately run into syntax errors. GCC’s extensions are not documented in the form of an addendum to the C specification, but rather as end-user documentation and a 15 million lines-of-code compiler for you to reverse engineer. It took me almost a week of frustration to get a parser which worked passably on a large subset of the C programs found in the wild, and I imagine I’ll be dealing with GNU problems for years to come. Please don’t use C extensions, folks.

In any case, the result now works fairly well for a lot of programs, and I have plans on expanding it to integrate more nicely with build systems like meson. Check out the code here: annotatec. The features of the C annotator include:

• Annotating function definitions with a list of files/linenos which call them
• Linking function calls to the definition of that function

In the future I intend to add support for linking to external symbols as well - for example, linking to the POSIX spec for functions specified by POSIX, or to the Linux man pages for Linux calls. It would also be pretty cool to support linking between related projects, so that wlroots calls in sway can be linked to their declarations in the wlroots repo.

Annotating Go code

The Go annotator was far easier. I started over my morning cup of coffee today and I was finished with the basics by lunch. Go has a bunch of support in the standard library for parsing and analyzing Go programs - I was very impressed:

• go/ast
• go/scanner
• go/token
• go/types

To support Go 1.12’s go modules, the experimental (but good enough) packages module is available as well. All of this is nicely summarized by a lovely document in the golang examples repository. The type checker is also available as a library, something which is less common even among languages with parsers-as-libraries, and allows for many features which would be very difficult without it. Nice work, Go!

The resulting annotator clocks in at just over 250 lines of code - compare that to the C annotator’s ~1,300 lines of C, lex, and yacc source code. The Go annotator is more featureful, too, it can:

• Link function calls to their definitions, and in reverse
• Link method calls to their definitions, and in reverse
• Link variables to their definitions, even in other files
• Link to godoc for symbols defined in external packages

I expect a lot more to be possible in the future. It might get noisy if you turn everything on, so each annotation type is gated behind a command line flag.

Displaying annotations

Displaying these annotations required a bit more effort than I would have liked, but the end result is fairly clean and reusable. Since SourceHut uses Pygments for syntax highlighting, I ended up writing a custom Formatter based on the existing Pygments HtmlFormatter. The result is the AnnotationFormatter, which splices annotations into the highlighted code. One downside of this approach is that it works at the token level - a more sophisticated implementation will be necessary for annotations that span more than a single token. Annotations are fairly expensive to render, so the rendered HTML is stowed in Redis.

The future?

I intend to write a Python annotator soon, and I’ll be relying on the community to build more. If you’re looking for a fun weekend hack and a chance to learn more about your favorite programming language, this’d be a great project. The format for annotations on SourceHut is also pretty generalizable, so I encourage other code forges to reuse it so that our annotators are useful on every code hosting platform.

builds.sr.ht will also soon grow first-class support for making these annotators available to your build process, as well as for making an OAuth token available (ideally with a limited set of permissions) to your build environment. Rigging up an annotator is a bit involved today (though the docs help), and streamlining that process will be pretty helpful. Additionally, this feature is only available for git.sr.ht, though it should generalize to hg.sr.ht fairly easily and I hope we’ll see it available there soon.

I’m also looking forward to seeing more novel use-cases for annotation. Can we indicate code coverage by coloring a gutter alongside each line of code? Can we link references to ticket numbers in the comments to your bug tracker? If you have any cool ideas, I’m all ears. Here’s that list of cool annotated repos to browse again, if you made it this far and want to check them out:

• ~sircmpwn/scdoc: man page generator (C)
• ~sircmpwn/aerc: TUI email client (Go)
• ~mcf/cproc: C compiler (C)