Streaming regexp matching

[Isiah Meadows]

Not yet, but if I were, it wouldn't be a subclass (it's not
necessary). But the key trouble implementing this is that I'd have to
reimplement the regexp matching logic entirely from scratch if I want
remotely reasonable runtime complexity. As for why:

- If you ignore backreferences (which IIUC makes JS regexps go from
regular to context-sensitive recognizers), it's *possible* to
implement partial matching using regexp rewriting, but only because JS
doesn't have any other arcane enough features to make it infeasible.
- There is no possible way to extract duplicate matches without
rewriting the regexp entirely and using `regexp.exec`, and the
complexity of the logic to do this generally I suspect is NP-complete,
maybe PSPACE-complete, and in either case, certainly infeasible when
backreferences enter the picture.
- It's technically possible to refactor a string for streaming, but I
then lose the ability to discern a match from a non-match. I also have
the same issue as per above WRT duplicate matches and partial matching
with complexity issues.
- If you were to combine the three, rewriting the regexp generally to
support streaming matches, including duplicates, I suspect would
likely be PSPACE-complete because it seems *very* similar to the first
problem listed here [1].

\* Backreferences bring the grammatical complexity of JS regexps from
I'm pretty sure regular to context-sensitive.

Or in other words, either you control the raw matching logic yourself,
or the polyfill runtime complexity could get absurd for this proposal.
And implementing a pushdown state machine-based regexp engine + parser
in JS isn't exactly something I'm willing to prototype for a strawman.

[1]: https://en.wikipedia.org/wiki/PSPACE-complete#Regular_expressions_and_automata

-----

Isiah Meadows
contact at isiahmeadows.com
www.isiahmeadows.com


On Mon, Jul 30, 2018 at 2:44 PM, Jordan Harband <ljharb at gmail.com> wrote:
> Have you tried to implement this as a RegExp subclass, overriding all the
> necessary extension points?
>
> On Mon, Jul 30, 2018 at 11:39 AM, Isiah Meadows <isiahmeadows at gmail.com>
> wrote:
>>
>> There's two things I've found that need suspendable matching:
>>
>> 1. Partially matching against a string, which is useful with
>> interactive form validation and such.
>> 2. Pattern matching and replacement over a character stream, which is
>> useful for things like matching against files without loading the
>> entire thing into memory or easier filtering of requests.
>>
>> Also, it'd be nice if there was a facility to get *all* matches,
>> including duplicate group matches. This is often useful for simple
>> parsing, where if such support existed, you could just use a Kleene
>> star instead of the standard `exec` loops (which admittedly get old).
>>
>> And finally, we could avoid setting regexp globals here. That would
>> speed up the matcher quite a bit.
>>
>> So, here's my proposal:
>>
>> - `regexp.matcher() -> matcher` - Create a streaming regexp matcher.
>> - `matcher.consume(codePoint, charSize?) -> result | undefined` -
>> Consume a Unicode code point or `-1` if no more characters exist, and
>> return a match result, `undefined` if no match occurred. `charSize` is
>> the number of bytes represented by `codePoint` (default: 1-2 if `/u`
>> is set, 1 otherwise), so it can work with other encodings flexibly.
>> - `matcher.nextPossibleStart -> number` - The next possible start the
>> matcher could have, for more effective buffering and stream
>> management. This is implementation-defined, but it *must* be be `-1`
>> after the matcher completes, and it *must* be within [0, N) otherwise,
>> where N is the next returned match.
>> - `result.group -> string | number | undefined` - Return the group
>> index/name of the current match, or `undefined` if it's just issuing a
>> match of the global regexp.
>> - `result.start -> number` - Return the matched value's start index.
>> - `result.end -> number` - Return the matched value's end index.
>> - This does *not* modify any globals or regexp instance members. It
>> only reads `regexp.lastIndex` on creation. (It doesn't operate on
>> strings, so it shouldn't return any it doesn't already have.)
>>
>> Most RegExp methods could similarly be built using this as a base: if
>> they work on strings, they can iterate their code points.
>>
>> As for the various concerns:
>>
>> - Partial matching is just iterating a string's character codes and
>> seeing if the matcher ever returned non-`undefined`.
>> - Streaming pattern matching is pretty obvious from just reading the API.
>> - Getting all matches is just iterating the string and returning an
>> object with all the groups + strings it matched.
>>
>> So WDYT?
>>
>> /cc Mathias Bynens, since I know you're involved in this kind of
>> text-heavy stuff.
>>
>> -----
>>
>> Isiah Meadows
>> contact at isiahmeadows.com
>> www.isiahmeadows.com
>> _______________________________________________
>> es-discuss mailing list
>> es-discuss at mozilla.org
>> https://mail.mozilla.org/listinfo/es-discuss
>
>