François Pottier

Hi,

I did a part of the incremental parsing in Menhir and the whole recovery aspect. I can try to explain a bit.
My goal was Merlin (https://github.com/ocaml/merlin/), not research so there is no paper covering the work I am about to describe. Also as of today only incrementality and error message generation are part of upstream version of Menhir, but the rest should come soon.

# Incrementality, part I

The notion of incrementality that comes builtin with Menhir is slightly weaker than what you are looking for.

With Menhir, the parser state is reified and control of the parsing is given back to the user. The important point here is the departure from a Bison-like interface. The user of the parser is handled a (pure) abstract object that represents the state of the parsing.
In regular parsing, this means we can store a snapshot of the parsing for each token, and resume from the first token that has changed (effectively sharing the prefix).
But on the side, we can also run arbitrary analysis on a parser (for error message generation, recovery, syntactic completion, or more incrementality...).

# Incrementality, part II

Sharing prefix was good enough for our use case (parsing is not a bottleneck in the pipeline). But it turns out that a trivial extension to the parser can also solve your case.

Using the token stream and the LR automaton, you can structure the tokens as a tree:

- starts with a root node annotated by the state of the automaton

- store each token consumed…

> The problem with bottom-up parsers of any flavor, LR included, is that when you encounter a syntax error there is no context information, or context is difficult to infer, which makes creating a sensible error message less fun than poking yourself in the eye repeatedly with a sharp stick.

Both LL and LR parsers share a fundamental flaw for error reporting; they are both prefix parsers. They will parse the input until it fails to be a valid prefix of the language represented by the grammar. However, it is rare that this coincides with the place where the actual error was made.

The difference in information available here at the end of the valid prefix is less about the difference between LL and LR parsing and more about the difference between LL and LR grammars. An LR parser parsing an LL grammar is able to recover the same information as an LL parser at the point of error. And for non-LL grammars, it is easy to extend an LR parser with additional context information by splitting states (see http://gallium.inria.fr/~fpottier/publis/fpottier-reachability-cc2016.pdf).

Parsing algorithms that do not produce prefix parsers (e.g. precedence parsing) are often able to produce more logical error messages with less work, because they continue making actions after the input is no longer a valid prefix, and this can amount to gathering more information about the error. For the same reason, LR-family parsers that perform default reductions (e.g. LALR parsers, or any minimal state…

Details about the "incremental" mode are listed in the documentation PDF[0] at section 9.2

Here are the first couple of paragraphs:

"In this API, control is inverted. The parser does not have access to the lexer. Instead, when the parser needs the next token, it stops and returns its current state to the user. The user is then responsible for obtaining this token (typically by invoking the lexer) and resuming the parser from that state. The directory demos/calc-incremental contains a demo that illustrates the use of the incremental API.

This API is “incremental” in the sense that the user has access to a sequence of the intermediate states of the parser. Assuming that semantic values are immutable, a parser state is a persistent data structure: it can be stored and used multiple times, if desired. This enables applications such as “live parsing”, where a buffer is continuously parsed while it is being edited. The parser can be re-started in the middle of the buffer whenever the user edits a character. Because two successive parser states share most of their data in memory, a list of n successive parser states occupies only O(n) space in memory."

There does not appear to be a specific mention of having the partial AST available.

[0] - linked from their front page and available at http://gallium.inria.fr/~fpottier/menhir/manual.pdf

I think the `#coding:` approach is promising. You have to fit your parsing/transpilation into a pretty small time budget (which is a fun challenge on its own) to keep the startup time hit negligible. The source code gets reparsed serval times in certain cases, e.g. when printing a stack trace, so its a good idea to have some sort of caching mechanism.

As far as parsing goes, if you want to stick with python I've had good success with pyparsing [1], otherwise I have a strong preference to do language-related things in OCaml with menhir [2]. I've toyed with wrapping the Python parser in an OCaml library with decent success [3]. But, of course, unless you're optimizing for fun, it's probably a good idea to stick with Python.

Another, weirder, approach is creating a language that happens to be able to be parsed by the same grammar as python, then using a decorator (or similar) to get the source code or the ast to reparse and transpile. As an example you could have something like `a <- b` be the syntax for an actor model message receive dsl (which is valid python).

[1] https://github.com/pyparsing/pyparsing
[2] http://gallium.inria.fr/~fpottier/menhir/
[3] https://github.com/georgek42/pyparser

https://arxiv.org/abs/1803.02796 has a good intro referencing other work. The paper doesn't go into much detail on the borrow checking side of things, but that author is both thoroughly steeped in the linear types research tradition and knows Rust and mentions it a bunch.

The fact that that paper talks about the affine types far more than the lifetimes obliquely illustrates my point in my view, I'm not sure if it's been stated more directly.

https://gallium.inria.fr/~fpottier/slides/fpottier-2007-05-linear-bestiary.pdf has some nice history. It wish there is a newer version for things that have been published since, in case the issue is more resolved and I'm out of date.

So in Rust today type inference strategies are less the point this hinges on. It's more about abolishing the difference between intra-procedural control flow / data flow / whatever, and their inter-procedural equivalents. Similarly, if we had an annotated version of the MIR which was borrow-check correct under composition that would also be good. NLL is great, and starts to clarify what borrow checking is, but running it in "whole" functions makes it to what clear it is a compositional analysis.

Some people have mentioned parser generators, but so far nobody has mentioned Menhir (http://gallium.inria.fr/~fpottier/menhir/).

It's an LR(1) parser generator for OCaml and Coq with a lot of extremely interesting features, such as genuinely good debugging support for grammars and the ability of generating error messages by example.

What this means is that after you write down your grammar, Menhir will give you examples of all the possible syntax errors that could occur. You can then write error messages for each case and get a parser with built-in error reporting for syntax errors.

This works a lot better than you'd think and I really wonder why nobody else implements this feature. Or for that matter, why they're not advertising it on the webpage! If you want to know more, look in the manual, section 11.