Fix edge cases that prevent completion #1181

faldor20 · 2023-09-17T18:21:31Z

I encountered a few cases where completion doesn't work correctly. Most obviously:

let a=[1]|>Li

This will not complete properly.
This was because the position_prefix code had a bunch of edge cases it didn't account for. I have added a whole collection of tests for these edge cases, and included two different proposed replacements for this prefix code, plus a benchmark of the difference between the old and new solutions.

I'm quite new to ocaml overall so i wanted to just run this by you all to get some feedback before i clean it up and settle on a solution.

After mucking around with this for many hours I think that the best solution would be to actually run the lexer over the tiny snippet and then use that token information to make an informed decision about what to feed back to merlin.

One key thing that occured to me is that currently none of these solution covers very odd edge cases like:

[1]|>List.
ma
[1]|> List. ma

but use lexer would solve this problem. though I'm honestly not sure if merlin would even deal with such a thing.

The solution here is to change all whitespace to spaces for ease of regex matching(all whitespace is equivelent semantically) and then remove all spaces from the prefix that's passed to merlin.

faldor20 · 2023-09-20T21:33:04Z

┌─────────────────┬────────────┬─────────┬────────────┐
│ Name            │   Time/Run │ mWd/Run │ Percentage │
├─────────────────┼────────────┼─────────┼────────────┤
│ regex           │   916.98ns │  90.00w │     44.62% │
│ regex_long      │ 1_168.30ns │  97.00w │     56.85% │
│ regex_only      │   392.91ns │  51.00w │     19.12% │
│ regex_only_long │ 2_055.10ns │  50.99w │    100.00% │
│ old             │   269.50ns │  22.00w │     13.11% │
│ old_long        │   260.33ns │  22.00w │     12.67% │
└─────────────────┴────────────┴─────────┴────────────┘

these are the results of running it on my laptop.
This could get a bit if we are happy to reverse the list still. but it does make the regex a little weird, because it's backwards.
This is that reversed version:

┌─────────────────┬──────────┬─────────┬────────────┐
│ Name            │ Time/Run │ mWd/Run │ Percentage │
├─────────────────┼──────────┼─────────┼────────────┤
│ regex           │ 652.22ns │  96.00w │     72.12% │
│ regex_long      │ 904.39ns │ 103.00w │    100.00% │
│ regex_only      │ 128.21ns │  51.00w │     14.18% │
│ regex_only_long │ 169.68ns │  52.00w │     18.76% │
│ old             │ 236.27ns │  22.00w │     26.12% │
│ old_long        │ 235.92ns │  22.00w │     26.09% │
└─────────────────┴──────────┴─────────┴────────────┘

I think it's worth it to use the reversed approach

rgrinberg · 2023-09-21T12:56:01Z

ocaml-lsp-server/test/e2e/__tests__/textDocument-completion.test.ts

@@ -91,6 +91,395 @@ describe_opt("textDocument/completion", () => {
    expect(items).toMatchObject([{ label: "StringLabels" }]);
  });

+  it("can start completion after operator with space", async () => {


The typescript test suite is deprecated. Look into e2e-new for the new OCaml tests.

I didn't really want to migrate the whole test suite in this PR and i didn't see any existing completion tests.

Would it be okay to just merge this with the new tests in the old suite and then bundle them in with a larger effort to convert all the completion tests to the new e2e suite?

If you don't have time, I can do it myself. I'm afraid that the larger effort to convert everything to the new test suite is never going to come. So for now, we're just not going to let the typescript test suite grow.

rgrinberg · 2023-09-21T12:56:58Z

I'm curious for why you added the benchmarks, anything in particular you expect to be slow here?

ocaml-lsp-server/src/compl.ml

ocaml-lsp-server/src/compl.mli

ocaml-lsp-server/src/prefix_parser.ml

ocaml-lsp-server/bench/ocaml_lsp_bench.ml

rgrinberg · 2023-09-21T13:01:07Z

ocaml-lsp-server/src/prefix_parser.ml

+
+(** matches let%lwt and let* style expressions. See here:https://v2.ocaml.org/manual/bindingops.html *)
+let monadic_bind =
+  Re.compile @@ Re.Posix.re {|(let|and)([$&*+\-/=>@^|.]|(%[a-zA-Z0-9_']*))$|}


Can you use the combinators from Re instead of Re.Posix please.

How set are you on this? I far far prefer to use standard regex because
a) it's something universal
b) it means it's easy to test and debug using tools like https://regexr.com/
But if you think there are good reasons to use the combinators I can convert it

I think the main issue with the way you've written it is that you're using capturing groups that you don't actually need. That's a little confusing and also slows down the matching.

Also, I think that if you used the combinators, constructing the "reverse" regex would be much simpler.

faldor20 · 2023-09-21T13:03:15Z

@rgrinberg Well, no not particularly. But for code that runs almost every keypress I think speed is key. I wanted to validate that I wasn't causing any meaningful slowdown

faldor20 · 2023-09-21T14:33:06Z

Also I've switched to the reversed implementation because its performance is more predictable and writing some regexes backwards isn't that onerous

faldor20 · 2023-09-21T14:43:41Z

As for the benchmarks, I understand they may not make sense in this PR and you may want a different benchmarking lib etc. But I do think it makes sense to have something setup at some point. Even just for testing little things to ensure that new changes aren't causing perf regressions. I'm sure some people out there can always accurately guess the performance effect of a change just by looking at it, but I think microbenchmarks are a very good thing to have around.

An example of this is that i find this server mostly very performant(kudos, the startup time is super impressive) but every now and again it gets into a weird state or i open a particular file that causes major slowdown. I'd love to find out why

ocaml-lsp-server/src/prefix_parser.ml

ocaml-lsp-server/bench/dune

ocaml-lsp-server/src/testing.ml

rgrinberg · 2023-09-23T08:59:19Z

As for the benchmarks, I understand they may not make sense in this PR and you may want a different benchmarking lib etc. But I do think it makes sense to have something setup at some point. Even just for testing little things to ensure that new changes aren't causing perf regressions. I'm sure some people out there can always accurately guess the performance effect of a change just by looking at it, but I think microbenchmarks are a very good thing to have around.

I don't mind the benchmarking library you chose. It's just that there's so much overhead to getting completions that I doubt that running a regex is going to make any difference. Anyway, you can keep them.

An example of this is that i find this server mostly very performant(kudos, the startup time is super impressive) but every now and again it gets into a weird state or i open a particular file that causes major slowdown. I'd love to find out why

Unfortunately most of the those performance problems are inherited from merlin itself.

converted to forward regex

faldor20 · 2023-10-10T13:22:05Z

Okay this is all good to go, all the old and new completion tests have been migrated and the ground work for migrating the others has been established

I used chatgpt for a lot of the migration, I think with a subscription and gpt4 it could be done very very quickly

faldor20 · 2023-10-16T19:12:54Z

Okay, I have no idea why the formatter decided that it would be a good idea to format the result of that expect test, should be fixed now. Sorry this has been a bit of a mess, I'm actually traveling right now, so my work on this has been a bit erratic and distracted.

faldor20 · 2023-10-18T20:04:24Z

Okay I actually have no idea why the build failed, I didn't change the doc_to_md file, at all.
I'll try to check more tomorrow when I at my laptop.

faldor20 · 2023-10-18T20:09:28Z

Ahh I see, looks like I was hit with the issue fixed here #1204 could I get a rerun of CI @rgrinberg?

rgrinberg · 2023-10-23T17:27:07Z

dune-project

@@ -59,6 +59,7 @@ possible and does not make any assumptions about IO.
  (odoc-parser (and (>= 2.0.0) (< 2.3.0)))
  (ppx_expect (and (>= v0.15.0) :with-test))
  (ocamlformat (and :with-test (= 0.24.1)))
+  (ppx_bench (and :with-test (>= 0.16.0)))


Why do we need this for the tests?

rgrinberg · 2023-11-27T01:48:59Z

Thanks for the thorough PR and apologies for such a late response. I've merged it in #1212 with a minor tweaks.

The benchmarks had to be disabled in CI because there's no good way to report their results and they just slow things down with the unnecessary dependencies otherwise.

faldor20 · 2023-11-28T13:45:02Z

All good, sorry I didn't quite have the time and expertise to finish that last little bit. Glad it got done :)

CHANGES: ## Fixes - Fix missing super & subscripts in markdown documentation. (ocaml/ocaml-lsp#1170) - Do not invoke dune at all if `--fallback-read-dot-merlin` flag is on. (ocaml/ocaml-lsp#1173) - Fix semantic highlighting of infix operators that contain '.'. (ocaml/ocaml-lsp#1186) - Disable highlighting unit as an enum member to fix comment highlighting bug. (ocaml/ocaml-lsp#1185) - Improve type-on-hover and type-annotate efficiency by only formatting the type of the first enclosing. (ocaml/ocaml-lsp#1191, ocaml/ocaml-lsp#1196) - Fix the encoding of URI's to match how vscode does it (ocaml/ocaml-lsp#1197) - Fix parsing of completion prefixes (ocaml/ocaml-lsp#1181) ## Features - Compatibility with Odoc 2.3.0, with support for the introduced syntax: tables, and "codeblock output" (ocaml/ocaml-lsp#1184) - Display text of references in doc strings (ocaml/ocaml-lsp#1166) - Add mark/remove unused actions for open, types, for loop indexes, modules, match cases, rec, and constructors (ocaml/ocaml-lsp#1141) - Offer auto-completion for the keyword `in` (ocaml/ocaml-lsp#1217)

faldor20 added 2 commits September 18, 2023 18:37

Tests passing

96376d1

Added support for whitespace in completion

9bb112a

The solution here is to change all whitespace to spaces for ease of regex matching(all whitespace is equivelent semantically) and then remove all spaces from the prefix that's passed to merlin.

faldor20 force-pushed the master branch from 1dc3296 to 9bb112a Compare September 18, 2023 08:37

faldor20 added 3 commits September 18, 2023 18:46

Addded ability to have whitespace before or after the dot and test.

0b1920e

finished cleanup and added support for monadic bind.

cfde1e0

removed pointless regex

f6a031c

reversed implimentation

bfddc86