This PR speeds up auto-completion by a factor of ~3.5x through various
performance improvements in the language server. On one machine, with
`import Mathlib`, completing `i` used to take 3200ms and now instead
yields a result in 920ms.
Specifically, the following improvements are made:
- The watchdog process no longer de-serializes and re-serializes most
messages from the file worker before passing them on to the user - a
fast partial de-serialization procedure is now used to determine whether
the message needs to be de-serialized in full or not.
- `escapePart` is optimized to perform better on ASCII strings that do
not need escaping.
- `Json.compress` is optimized to allocate fewer objects.
- A faster JSON compression specifically for completion responses is
implemented that skips allocating `Json` altogether.
- The JSON compression has been moved to the task where we convert a
request response to `Json` so that converting to a string won't block
the output task of the FileWorker and so the `Json` value is not marked
as multi-threaded when we compress is, which drastically increases the
cost of reference-counting.
- The JSON representation of the `data?` field of each completion item
is optimized.
- Both the completion kind and the set of completion tags for each
imported completion item is now cached.
- The filtering of duplicate completion items is optimized.
Other adjustments:
- `LT UInt8` and `LE UInt8` are moved to Prelude so that they can be
used in `Init.Meta` for the name part escaping fast path.
- `Array.usize` is exposed since it was marked as `@[simp]`.
This PR removes uses of `Lean.RBMap` in Lean itself.
Furthermore some massaging of the import graph is done in order to avoid
having `Std.Data.TreeMap.AdditionalOperations` (which is quite
expensive) be the critical path for a large chunk of Lean. In particular
we can build `Lean.Meta.Simp` and `Lean.Meta.Grind` without it thanks to
these changes.
We did previously not conduct this change as `Std.TreeMap` was not
outperforming `Lean.RBMap` yet, however this has changed with the new
code generator.
This PR fixes a bug where the unknown identifier code actions wouldn't
work correctly for some unknown identifier error spans and adjusts
several unknown identifier spans to actually end on the identifier in
question.
The following additional adjustments are made:
- The fallback mechanism of the unknown identifier code actions is
removed, since it could produce severely incorrect suggestions for
unknown identifier errors on fields.
- A performance bug when using the code action to import all unknown
identifiers is fixed.
- A bug that occurs when the elaborator produces multiple overlapping
completion infos is fixed.
- A bug in the snapshot selection that could cause it to wait for
snapshots in snapshots with non-canonical syntax is fixed.
- Some invariants of the snapshot tree are documented.
- The snapshot tree formatting is adjusted to display the final info
tree again.
This PR fixes a number of bugs related to the handling of the source
search path in the language server, where deleting files could cause
several features to stop functioning and both untitled files and files
that don't exist on disc could have conflicting module names.
In detail, it makes the following adjustments:
- The URI <-> module name conversion was adjusted to produce no name
collisions.
- File URIs in the search path yield a module name relative to the
search path, as before.
- File URIs not in the search path, non-file URIs and non-`.lean` files
yield a `«external:<full uri>»` module name.
- To avoid the issue of the URI -> module name conversion failing when a
file is deleted from disc, we now cache the result of this conversion in
the watchdog and the file worker when the file is first opened.
- All of the URI <-> module name conversions now consistently go through
`Server.documentUriFromModule?` and `moduleFromDocumentUri` to ensure
that we don't have minor deviations for this conversion all over the
place.
- The threading of the source search path through the file worker (from
`lake setup-file`) is removed. It turns out that `lake serve` already
sets the correct source search path in the environment, so we can just
always use the search path from the environment.
- Since we can now answer more requests that need the .ileans in
untitled files, a lot of the tests that test 'Go to definition' needed
to be adjusted so that they use the information from the watchdog, not
the file worker. As we load references asynchronously, this PR adds an
internal `$/lean/waitForILeans` request that tests can use to wait for
all .ilean files to be loaded and for the ilean references from the file
worker for the current document version to be finalized.
- As part of this PR, we noticed that the .ileans aren't available in
the NixOS setup, so @Kha adjusted the Nix CI to fix this.
### Breaking changes
- `Server.documentUriFromModule` has been renamed to
`Server.documentUriFromModule?` and doesn't take a `SearchPath` argument
anymore, as the `SearchPath` is now computed from the `LEAN_SRC_PATH`
environment variable. It has also been moved from `Lean.Server.GoTo` to
`Lean.Server.Utils`.
- `Server.moduleFromDocumentUri` does not take a `SearchPath` argument
anymore and won't return an `Option` anymore. It has also been moved
from `Lean.Server.GoTo` to `Lean.Server.Utils`.
- The `System.SearchPath.searchModuleNameOfUri` function has been
removed. It is recommended to use `Server.moduleFromDocumentUri`
instead.
- The `initSrcSearchPath` function has been renamed to
`getSrcSearchPath` and has been moved from `Lean.Util.Paths` to
`Lean.Util.Path`. It also doesn't need to take a `pkgSearchPath`
argument anymore.
---------
Co-authored-by: Sebastian Ullrich <sebasti@nullri.ch>
This PR adds support for code actions that resolve 'unknown identifier'
errors by either importing the missing declaration or by changing the
identifier to one from the environment.
<details>
<summary>Demo (Click to open)</summary>
</details>
Specifically, the following kinds of code actions are added by this PR,
all of which are triggered on 'unknown identifier' errors:
- A code action to import the module containing the identifier at the
text cursor position.
- A code action to change the identifier at the text cursor position to
one from the environment.
- A source action to import the modules for all unambiguous identifiers
in the file.
### Details
When clicking on an identifier with an 'unknown identifier' diagnostic,
after a debounce delay of 1000ms, the language server looks up the
(potentially partial) identifier at the position of the cursor in the
global reference data structure by fuzzy-matching against all
identifiers and collects the 10 closest matching entries. This search
accounts for open namespaces at the position of the cursor, including
the namespace of the type / expected type when using dot notation. The
10 closest matching entries are then offered to the user as code
actions:
- If the suggested identifier is not contained in the environment, a
code action that imports the module that the identifier is contained in
and changes the identifier to the suggested one is offered. The
suggestion is inserted in a "minimal" manner, i.e. by accounting for
open namespaces.
- If the suggested identifier is contained in the environment, a code
action that only changes the identifier to the suggested one is offered.
- If the suggested identifier is not contained in the environment and
the suggested identifier is a perfectly unambiguous match, a source
action to import all unambiguous in the file is offered.
The source action to import all unambiguous identifiers can also always
be triggered by right-clicking in the document and selecting the 'Source
Action...' entry.
At the moment, for large projects, the search for closely matching
identifiers in the global reference data structure is still a bit slow.
I hope to optimize it next quarter.
### Implementation notes
- Since the global reference data structure is in the watchdog process,
whereas the elaboration information is in the file worker process, this
PR implements support for file worker -> watchdog requests, including a
new `$/lean/queryModule` request that can be used by the file worker to
request global identifier information.
- To identify 'unknown identifier' errors, several 'unknown identifier'
errors in the elaborator are tagged with a new tag.
- The debounce delay of 1000ms is necessary because VS Code will
re-request code actions while editing an unknown identifier and also
while hovering over the identifier.
- We also implement cancellation for these 'unknown identifier' code
actions. Once the file worker responds to the request as having been
cancelled, the watchdog cancels its computation of all corresponding
file worker -> watchdog requests, too.
- Aliases (i.e. `export`) are currently not accounted for. I've found
that we currently don't handle them correctly in auto-completion, too,
so we will likely add support for this later when fixing the
corresponding auto-completion issue.
- The new code actions added by this request support incrementality.
This PR ensures info tree users such as linters and request handlers
have access to info subtrees created by async elab task by introducing
API to leave holes filled by such tasks.
**Breaking change**: other metaprogramming users of
`Command.State.infoState` may need to call `InfoState.substituteLazy` on
it manually to fill all holes.
This PR fixes several inlay hint race conditions that could result in a
violation of the monotonic progress assumption, introduced in #7149.
Specifically:
- In rare circumstances, it could happen that stateful LSP requests were
executed out-of-order with their `didChange` handlers, as both requests
and the `didChange` handlers waited on `lake setup-file` to complete,
with the latter running those handlers in a dedicated task afterwards.
This meant that a request could be added to the stateful LSP handler
request queue before the corresponding `didChange` call that actually
came before it. This PR resolves this issue by folding the task that
waits for `lake setup-file` into the `RequestContext`, which ensures
that we only need to wait for it when actually executing the request
handler.
- While #7164 fixed the monotonic progress assertion violation that was
caused by `$/cancelRequest`, it did not account for our internal notion
of silent request cancellation in stateful LSP requests, which we use to
cancel the inlay hint edit delay when VS Code fails to emit a
`$/cancelRequest` notification. This issue is resolved by always
producing the full finished prefix of the command snapshot queue, even
on cancellation. Additionally, this also fixes an issue where in the
same circumstances, the language server could produce an empty inlay
hint response when a request was cancelled by our internal notion of
silent request cancellation.
- For clients that use `fullChange` `didChange` notifications (e.g. not
VS Code), we would get several aspects of stateful LSP request
`didChange` state handling wrong, which is also addressed by this PR.
This PR ensures that all tasks in the language server either use
dedicated tasks or reuse an existing thread from the thread pool. This
ensures that elaboration tasks cannot prevent language server tasks from
being scheduled. This is especially important with parallelism right
around the corner and elaboration becoming more likely to starve the
language server of computation, which could drive up language server
latencies significantly on machines with few cores.
Specifically, all language server tasks are refactored to use a new thin
`ServerTask` API wrapper with a single "costly" vs "cheap" dimension,
where costly tasks are always scheduled as dedicated tasks, and cheap
tasks are always made to either run on the calling thread or to reuse
the thread of the task being mapped on by using the `sync` flag.
ProofWidgets4 adaption PR:
https://github.com/leanprover-community/ProofWidgets4/pull/106
### Other changes
- This PR makes several tasks dedicated that weren't dedicated before,
and uses `sync := true` for some others. The rules for this are
described in the module docstring of `ServerTask.lean`.
- Most notably, the reporting task in the file worker was *not* a
dedicated task before this PR, which could easily lead to thread pool
starvation on successive changes. It also did not support cancellation.
This PR ensures that it does.
### Breaking changes
- `RequestTask` and the request-oriented snapshot API are refactored to
use `ServerTask` instead of `Task`. All functions in `Task` have close
analogues in `ServerTask`, and functions on `RequestTask` now need to
distinguish between whether a `map` or a `bind` is cheap or costly. This
affects all downstream users of `RequestM`, e.g. tools that extend the
language server with their own requests, or some users of the RPC
mechanism.
- The following unused functions of the `AsyncList` API have been
deleted: `append`, `unfoldAsync`, `getAll`, `waitHead?`, `cancel`
This PR adds language server support for request cancellation to the
following expensive requests: Code actions, auto-completion, document
symbols, folding ranges and semantic highlighting. This means that when
the client informs the language server that a request is stale (e.g.
because it belongs to a previous state of the document), the language
server will now prematurely cancel the computation of the response in
order to reduce the CPU load for requests that will be discarded by the
client anyways.
This PR fixes a bug where the goal state selection would sometimes
select incomplete incremental snapshots on whitespace, leading to an
incorrect "no goals" response. Fixes#6594, a regression that was
originally introduced in 4.11.0 by #4727.
The fundamental cause of #6594 was that the snapshot selection would
always select the first snapshot with a range that contains the cursor
position. For tactics, whitespace had to be included in this range.
However, in the test case of #6594, this meant that the snapshot
selection would also sometimes pick a snapshot before the cursor that
still contains the cursor in its whitespace, but which also does not
necessarily contain all the information needed to produce a correct goal
state. Specifically, at the `InfoTree`-level, when the cursor is in
whitespace, we distinguish competing goal states by their level of
indentation. The snapshot selection did not have access to this
information, so it necessarily had to do the wrong thing in some cases.
This PR fixes the issue by adjusting the snapshot selection for goals to
explicitly account for whitespace and indentation, and refactoring the
language processor architecture to thread enough information through to
the snapshot selection so that it can decide which snapshots to use
without having to force too many tasks, which would destroy
incrementality in goal state requests.
Specifically, this PR makes the following adjustments:
- Refactor `SnapshotTask` to contain both a `Syntax` and a `Range`.
Before, `SnapshotTask`s had a single range that was used both for
displaying file progress information and for selecting snapshots in
server requests. For most snapshots, this range did not include
whitespace, though for tactics it did. Now, the `reportingRange` field
of `SnapshotTask` is intended exclusively for reporting file progress
information, and the `Syntax` is used for selecting snapshots in server
requests. Importantly, the `Syntax` contains the full range information
of the snapshot, i.e. its regular range and its range including
whitespace.
- Adjust all call-sites of `SnapshotTask` to produce a reasonable
`Syntax`.
- Adjust the goal snapshot selection to account for whitespace and
indentation, as the `InfoTree` goal selection does.
- Fix a bug in the snapshot tree tracing that would cause it to render
the `Info` of a snapshot at the wrong location when `trace.Elab.info`
was also set.
This PR is based on #6329.
This PR implements a number of refinements for the auto-implicit inlay
hints implemented in #6768.
Specifically:
- In #6768, there was a bug where the inlay hint edit delay could
accumulate on successive edits, which meant that it could sometimes take
much longer for inlay hints to show up. This PR implements the basic
infrastructure for request cancellation and implements request
cancellation for semantic tokens and inlay hints to resolve the issue.
With this edit delay bug fixed, it made more sense to increase the edit
delay slightly from 2000ms to 3000ms.
- In #6768, we applied the edit delay to every single inlay hint request
in order to reduce the amount of inlay hint flickering. This meant that
the edit delay also had a significant effect on how far inlay hints
would lag behind the file progress bar. This PR adjusts the edit delay
logic so that it only affects requests sent directly after a
corresponding `didChange` notification. Once the edit delay is used up,
all further semantic token requests are responded to without delay, so
that the only latency that affects how far the inlay hints lag behind
the progress bar is how often we emit refresh requests and how long VS
Code takes to respond to them.
- For inlay hints, refresh requests are now emitted 500ms after a
response to an inlay hint request, not 2000ms, which means that after
the edit delay, inlay hints should only lag behind the progress bar by
about up to 500ms. This is justifiable for inlay hints because the
response should be much smaller than e.g. is the case for semantic
tokens.
- In #6768, 'Restart File' did not prompt a refresh, but it does now.
- VS Code does not immediately remove old inlay hints from the document
when they are applied. In #6768, this meant that inlay hints would
linger around for a bit once applied. To mitigate this issue, this PR
adjusts the inlay hint edit delay logic to identify edits sent from the
client as being inlay hint applications, and sets the edit delay to 0ms
for the inlay hint requests following it. This means that inlay hints
are now applied immediately.
- In #6768, hovering over single-letter auto-implicit inlay hints was a
bit finicky because VS Code uses the regular cursor icon on inlay hints,
not the thin text cursor icon, which means that it is easy to put the
cursor in the wrong spot. We now add the separation character (` ` or
`{`) preceding an auto-implicit to the hover range as well, which makes
hovering over inlay hints much smoother.
This PR adds preliminary support for inlay hints, as well as support for
inlay hints that denote the auto-implicits of a function. Hovering over
an auto-implicit displays its type and double-clicking the auto-implicit
inserts it into the text document.
This PR is an extension of #3910.
### Known issues
- In VS Code, when inserting an inlay hint, the inlay hint may linger
for a couple of seconds before it disappears. This is a defect of the VS
Code implementation of inlay hints and cannot adequately be resolved by
us.
- When making a change to the document, it may take a couple of seconds
until the inlay hints respond to the change. This is deliberate and
intended to reduce the amount of inlay hint flickering while typing. VS
Code has a mechanism of its own for this, but in my experience it is
still far too sensitive without additional latency.
- Inserting an auto-implicit inlay hint that depends on an auto-implicit
meta-variable causes a "failed to infer binder type" error. We can't
display these meta-variables in the inlay hint because they don't have a
user-displayable name, so it is not clear how to resolve this problem.
- Inlay hints are currently always resolved eagerly, i.e. we do not
support the `textDocument/inlayHint/resolve` request yet. Implementing
support for this request is future work.
### Other changes
- Axioms did not support auto-implicits due to an oversight in the
implementation. This PR ensures they do.
- In order to reduce the amount of inlay hint flickering when making a
change to the document, the language server serves old inlay hints for
parts of the file that have not been processed yet. This requires LSP
request handler state (that sometimes must be invalidated on
`textDocument/didChange`), so this PR introduces the notion of a
stateful LSP request handler.
- The partial response mechanism that we use for semantic tokens, where
we simulate incremental LSP responses by periodically emitting refresh
requests to the client, is generalized to accommodate both inlay hints
and semantic tokens. Additionally, it is made more robust to ensure that
we never emit refresh requests while a corresponding request is in
flight, which causes VS Code to discard the respond of the request, as
well as to ensure that we keep prompting VS Code to send another request
if it spuriously decides not to respond to one of our refresh requests.
- The synthetic identifier of an `example` had the full declaration as
its (non-canonical synthetic) range. Since we need a reasonable position
for the identifier to insert an inlay hint for the auto-implicits of an
`example`, we change the (canonical synthetic) range of the synthetic
identifier to that of the `example` keyword.
- The semantic highlighting request handling is moved to a separate
file.
### Breaking changes
- The semantic highlighting request handler is not a pure request
handler anymore, but a stateful one. Notably, this means that clients
that extend the semantic highlighting of the Lean language server with
the `chainLspRequestHandler` function must now use the
`chainStatefulLspRequestHandler` function instead.
This PR allows the dot ident notation to resolve to the current
definition, or to any of the other definitions in the same mutual block.
Existing code that uses dot ident notation may need to have `nonrec`
added if the ident has the same name as the definition.
Closes#6601
This PR enables tactic completion in the whitespace of a tactic proof
and adds tactic docstrings to the completion menu.
Future work:
- A couple of broken tactic completions: This is due to tactic
completion now using @david-christiansen's `Tactic.Doc.allTacticDocs` to
obtain the tactic docstrings and should be fixed soon.
- Whitespace tactic completion in tactic combinators: This requires
changing the syntax of tactic combinators to produce a syntax node that
makes it clear that a tactic is expected at the given position.
Closes#1651.
This PR enables the use of incrementality for completion in tactic
blocks. Consider the following example:
```lean
example : True := by
have : True := T
sleep 10000
```
Before this PR, in order to respond to a completion request after `T`,
`sleep 10000` has to complete first since the command must be fully
elaborated. After this PR, the completion request is responded to
immediately.
After each tactic step, we save the info tree created by it together
with an appropriate info tree context that makes it stand-alone (which
we already did before to some degree, see `Info.updateContext?`). Then,
in the adjusted request handlers, we first search for a snapshot task
containing the required position, if so wait on it, and if it yielded an
info tree, use it to answer the request, or else continue searching and
waiting, falling back to the full info tree, which should be unchanged
by this PR.
The definition header does *not* report info trees early as in general
it is not stand-alone in the tactic sense but may contain e.g.
metavariables solved by the body in which case we do want to show the
ultimate state as before. This could be refined in the future in case
there are no unsolved mvars.
The adjusted request handlers are exactly the ones waited on together by
the info view, so they all have to be adjusted to have any effect on the
UX. Further request handlers may be adjusted in the future.
No new tests as "replies early" is not something we can test with our
current framework but the existing test suite did help in uncovering
functional regressions.
This PR addresses several performance issues in the auto-completion
implementation. It also fixes a number of smaller bugs related to
auto-completion.
In a file with `import Mathlib`, the performance of various kinds of
completions has improved as follows:
- Completing `C`: 49000ms -> 1400ms
- Completing `Cat`: 14300ms -> 1000ms
- Completing `x.` for `x : Nat`: 3700ms -> 220ms
- Completing `.` for an expected type of `Nat`: 11000ms -> 180ms
The following bugs have been fixed as well:
- VS Code never used our custom completion order. Now, the server fuzzy
completion score decides the order that completions appear in.
- Dot auto-completion for private types did not work at all. It does
now.
- Completing `.<identifier>` (where the expected type is used to infer
the namespace) did not filter by the expected type and instead displayed
all matching constants in the respective namespace. Now, it uses the
expected type for filtering. Note that this is not perfect because
sub-namespaces are technically correct completions as well (e.g.
`.Foo.foobar`). Implementing this is future work.
- Completing `.` was often not possible at all. Now, as long as the `.`
is not used in a bracket (where it may be used for the anonymous lambda
feature, e.g. `(. + 1)`), it triggers the correct completion.
- Fixes#3228.
- The auto-completion in `#check` commands would always try to complete
identifiers using the full declaration name (including namespaces) if it
could be resolved. Now it simply uses the identifier itself in case
users want to complete this identifier to another identifier.
## Details
Regarding completion performance, I have more ideas on how to improve it
further in the future.
Other changes:
- The feature that completions with a matching expected type are sorted
to the top of the server-side ordering was removed. This was never
enabled in VS Code because it would use its own completion item order
and when testing it I found it to be more confusing than useful.
- In the server-side ordering, we would always display keywords at the
top of the list. They are now displayed according to their fuzzy match
score as well.
The following approaches have been used to improve performance:
- Pretty-printing the type for every single completion made up a
significant amount of the time needed to compute the completions. We now
do not pretty-print the type for every single completion that is offered
to the user anymore. Instead, the language server now supports
`completionItem/resolve` requests to compute the type lazily when the
user selects a completion item.
- Note that we need to keep the amount of properties that we compute in
a resolve request to a minimum. When the server receives the resolve
request, the document state may have changed from the state it was in
when the initial auto-completion request was received. LSP doesn't tell
us when it will stop sending resolve requests, so we cannot keep this
state around, as we would have to keep it around forever.
LSP's solution for this dilemma is to have servers send all the state
they need to compute a response to a resolve request to the client as
part of the initial auto completion response (which then sends it back
as part of the resolve request), but this is clearly infeasible for all
real language servers where the amount of state needed to resolve a
request is massive.
This means that the only practical solution is to use the current state
to compute a response to the resolve request, which may yield an
incorrect result. This scenario can especially occur when using
LiveShare where the document is edited by another person while cycling
through available completions.
- Request handlers can now specify a "header caching handler" that is
called after elaborating the header of a file. Request handlers can use
this caching handler to compute caches for information stored in the
header. The auto-completion uses this to pre-compute non-blacklisted
imported declarations, which in turn allow us to iterate only over
non-blacklisted imported declarations where we would before iterate over
all declarations in the environment. This is significant because
blacklisted declarations make up about 4/5 of all declarations.
- Dot completion now looks up names modulo private prefixes to figure
out whether a declaration is in the namespace of the type to the left of
the dot instead of first stripping the private prefix from the name and
then comparing it. This has the benefit that we do not need to scan the
full name in most cases.
This PR also adds a couple of regression tests for fixed bugs, but *no
benchmarks*. We will add these in the future when we add proper support
for benchmarking server interaction sessions to our benchmarking
architecture.
All tests that were broken by producing different completion output
(empty `detail` field, added `sortText?` and `data?` fields) have been
manually checked by me to be still correct before replacing their
expected output.
