tree-sitter-d

D Grammar for Tree Sitter

What is this?

This is a D grammar for Tree Sitter.

As of this writing, I believe it fully supports D 2.108.

I believe this is the most accurate and complete machine readable grammar for D, as of this writing. In fact it might be the only one that can claim full, or nearly full, conformance to D 2.108, except that a limit of 5 levels of nesting is provided for nesting comments (/+ nested /+ comment +/ +/).

We have a large body of code at Weka, and that corpus was used to test and verify this parser. As of this writing, it passes cleanly, and I am not aware of any outstanding exceptions where this grammar fails.

Additionally this grammar has been tested with the DMD source code, as well as the test cases for DMD.

I hope that this work will be able to be useful not just in the context of Tree Sitter, but as a starting port for a more formal and accurate grammar which reflects the actual behavior of the DMD reference compiler.

Grammar Deviations

This has been used to parse the test corpus from DMD itself. A few failures in that case come in a couple of forms:

• Grammar or errors that are in code not caught by DMD, because it is in code that is not compiled (such as templates that are not instantiated). This grammar doesn't know about instantiation or constructs that are tucked behind conditional compilation, and verifies the entire body of the source document. I consider this a feature, not a bug.

• Deprecated use of the former body keyword (it is no longer a keyword.) If your source file has this problem, change body to do. You could also just delete the word, as it is entirely optional in the syntax where it appears.

• A pathological case with #line directives with a multi-line comment on in the middle of the directive. Nothing real emits such a busted syntax, and fixing it would require significant changes to the lexer, for absolutely zero real world benefit. (This was fixed in newer versions of D.)

• Inline assembler is essentially treated as a token stream with no real validation. As this is compiler and CPU specific, it doesn't make a lot of sense to try to add that here. (Use of DMD's inline assembler is not widely used, even within the D community, as it's limited to x86.)

• Use of enum as return type is now removed, following D 2.105. Please use auto if that creates an error in old code. (Not actually a deviation, for users of newer editions of D.)

In some areas this grammar is stricter than what is formally specified on the D website. It has some rules that make the grammar reject constructs that would appear legal in the D grammar, but actually are rejected by the compiler during semantic analysis.

I do not believe that this increase in tightness will cause any correct programs to be rejected, and it reduces some of the ambiguities in the language. This mainly consists of rejection of the use of expression lists (i.e. "expression1, expression2, ...") in contexts where only a single value is legal. (For example, you cannot use a comma expression in a return statement.)

Acknowledgements

While the resources available online - both the grammar that is part of the DGrammar project and what is specified on the website - have numerous problems, they still afforded an excellent starting point for this work. Without those resources, this would have been impossible, or at least a very much more difficult project.

Additionally the D community on Discord has been helpful in understanding the grammar and were patient with a number of my queries about potential language constructions, many of which were utterly non-sensical.

Challenges and Acknowledgements

D is an incredibly complex (some might say too much so) language, and it lacks an accurate formal grammar, and the machine readable grammars that exist (DGrammar) are buggy or out of date. Additionally, the specification is also not up to date, and has conflicts and inaccuracies that do not not reflect the actual language.

Making this work, with a grammar that is loosely specified as a recursive descent grammar with Tree Sitters GLR(1) style is no mean feat. There are many conflicts, and ambiguities abound in the actual grammar itself.

As a result this was generated mostly by hand, reading the grammar specification, and experimenting heavily.

In doing this work, it was easy to identify a number of language constructs that some might argue that the language could easily do without, without losing any of it's real capability.

Additionally this work has some ugly hacks that I wish it did not -- much of which is almost surely a result of my insufficient understanding of Tree Sitter or D or both. I welcome contributions or suggestions by those more knowledgable to improve this.

Future Directions

• It would be nice to get this project moved either into either the D or the Tree-Sitter communities. That would likely facilitate serendipitous discovery by the folks most likely to benefit from it.

• Additional queries. Local queries, etc. Again, contributions are welcome!

• Improvements to the indentation or highlighting queries.

• DDOC support. Arguably DDOC is a language unto itself.

• Extended test coverage. A lot of test cases are here, but we could really do with a much richer corpus. This is fairly tedious, but using the -u flag with tree-sitter test can be useful.

• Also highlight test coverage!

Performance

Trial runs of Tree Sitter using this grammar to parse the enter corpus of files in the DMD compiler (excluding the negative tests), ran in just 1.94 seconds (1.86 user, 0.07 system).

This involved parsing 3094 files, containing 834,637 lines of source code.

This is an average of 627 us per file, or 2.3 us per line of code.

Put another way, this parses at a rate of about 430K lines of code per second.

This test was performed on a 2020 MacBook Air with an M1 processor and 16GB RAM.