* Add a GH Action to file a new issue if we go a week without seeing a typescript-error-deltas issue
* Don't use the search terms in the title of the bug report
* Use binary search in file system cache
Previously, we were linear searching a linear number of times, resulting in too many `toLowerCaseFileName` calls on Windows.
* Use SortedArray types for clarity
* Use insertSorted after making it return a flat
* Drop redundant undefined
* Correct copy-paste error
* Add a simple queue implementation with better performance than `Array.shift`
This lets us clean up the hack introduced in #49581
* Correct typo
Co-authored-by: Mateusz Burzyński <mateuszburzynski@gmail.com>
Co-authored-by: Mateusz Burzyński <mateuszburzynski@gmail.com>
* Tune FAR aggregation
In making the work queue management more intelligible, we centralized the redundancy check at dequeue time. As a result, the queue tends to get very large (~1.6M items for `SyntaxKind` in this repo) and dequeuing via `shift` is too slow to do that many times. This change makes a few tweaks:
1. Use `Project` identity for de-duping and only maintain a set of keys for `loadAncestorProjectTree`
2. Attempt to filter prior to insertion
3. Use `splice` if many consecutive work queue items will be discarded.
On my box, this cuts FAR for `SyntaxKind` in parser.ts from 38 minutes to 20 seconds (we could do better, but effectively decided not to optimize this worst case scenario).
* Port #48978 forward to the revised FAR code
The old regression test didn't catch the corresponding bug in the new code because it hit the fast path where there's only a single project.
* Drop spurious non-null assertion
* Handle the case where the FAR results map is empty
Searching exactly one project does not imply the presence of exactly one references list.
Fixes#49217
* Use emptyArray
Co-authored-by: Sheetal Nandi <shkamat@microsoft.com>
Co-authored-by: Sheetal Nandi <shkamat@microsoft.com>
* Remove unnecessary members of HostCache
* Standardize on calling compiler host members in preparation for adding caching at that layer
* Call changeCompilerHostLikeToUseCache to cache existence checks
* Drop now-redundant HostCache
* Don't make directoryExists caching contingent on createDirectory
* Clear compilerHost rather than tracking state
* Clean up FAR and RenameLocations
This change had two goals:
1. Make the code easier to understand, primarily by simplifying the callback structure and minimizing side-effects
2. Improve performance by reducing repeated work, both FAR searches of individual projects and default tsconfig searches
This implementation attempts to preserve the merging order found in the original code (someone less relevant in the present state of using syntactic isDefinition).
* Stop enforcing search and aggregation order
...in preparation for implementing isDefinition explicitly.
Also restore convention of referring to `DocumentPosition`s as "locations".
* Introduce LanguageService.updateIsDefinitionOfReferencedSymbols
...to allow use of the checker when computing isDefinition across projects.
* Update baselines
* Tidy diff
* De-dup simplified results
* Baseline cross-project isDefinition results
* Move de-duping upstream to fix Full output
* Add server baseline test to confirm searches are not repeated
* Manually merge #48758
* Update baseline for newer fix to #48963
* Don't report isDefinition unless the starting node is a declaration
* Drop isDefinition everywhere it isn't specifically needed
* Fix tsserver tests
* Update shim comment
* Update baselines
* Add tests for isDefinition
* Update doc comment
* Clear isDefinition from all references if the first one lacks it
* Convert referenceToClass to baselining
* Delete stray marker from findAllRefsInsideWithBlock
* Bulk convert tests that just call verify.singleReferenceGroup
* Update tests with duplicate names
* Bulk update verify.singleReferenceGroup tests with BOMs
* Semi-automatically convert FAR tests with explicit range code
* Convert findAllReferencesDynamicImport2 to baselines
It should be adjusted as in `getReferencedSymbolsForNode`. The baseline changes show that it was incorrectly returning false at keywords preceding definitions.
* Trace project creation, loading, and updateGraph
* Drop generic event tracing
* Make argument names more consistent
* Trace diagnostics to make steps easier to interpret
* Fill an unexplained gap in updateGraph
* Move updateGraph tracing into base type
* Fill the gaps in updateGraph
They were using the same helpers as FAR and rename, but they actually behave differently. Decoupling them helps avoid some unnecessary work and will make it easier to clean up FAR in the future.
Scoping NavTo results to a single project (rather than, e.g. all loaded projects) makes the behavior more logical and the implementation simpler.
TestRunner was already using it, but upstream projects (e.g. Compiler) were not, causing TestRunner to re-parse and re-bind all upstream files in (at least) editor scenarios), slowing down initial project load.
In local testing, this cut a find-all-refs call in checker.ts from 5s to 2.5s.
Walking up the tree to find the enclosing SourceFile would distort the
timing too much so, instead, we attach a Path in the binder.
At present, the path is determined retroactively by walking up the call
stack in the trace visualizer, but this is both inconvenient and
routinely inaccurate (checking an expression in one file may require
checking an expression in another file and there's no way to determine
from the trace where this transition occurred).
* Clean up createMap
* Delete dead createMapFromTemplate
* Clean up assertDefined
* Delete dead assertEachDefined
* Delete dead createUnderscoreEscapedMap
* Delete dead hasEntries
* Delete dead ReadonlyNodeSet, NodeSet, ReadonlyNodeMap, NodeMap
* Use updated SyntaxKind names
* Update API baselines
* Pass absolute path to directoryExists
As pointed out by @gretzkiy, the `directoryExists` call added to
`matchFiles` in #44710 should have been passing the absolute path (since
the current directory might not match `currentDirectory`).
* Add test, simplify/clarify/fix matchFiles and friends
Co-authored-by: Andrew Branch <andrew@wheream.io>
* Make getSourceOfProjectReferenceRedirect take a Path
* Add useCaseSensitiveFileNames to ModuleResolutionHost
...so that path comparisons can use it during module resolution.
* Re-enable realpathSync.native for case-insensitive file systems
When a class declaration lacks a name, don't throw an exception when
producing the display parts (e.g. for QuickInfo).
Remaining issues:
1. The name shows as "__missing", the name of the underlying symbol,
rather than "(Missing)", as it is for the corresponding function
declaration case (because the parse constructs a missing identifier
node for the function declaration).
2. "(Missing)" is hard-coded, rather than being a localizable resource
string.
3. When an anonymous class declaration is a default export, the
corresponding symbol is named "default", resulting in the confusing
display string "class default".
Since display parts are built using existing `symbolToString`
functionality, it wasn't clear whether detecting special symbol names
and replacing them with user-friendly strings could be done without
breaking other functionality.
Similarly, changing the shape of the parse tree seemed riskier than the
problem justified (the user experience is just not getting QuickInfo for
the incomplete declaration, which seems acceptable).
* Add id and version to ManyToManyPathMap
...so that unchanged maps can be recognized without having to examine
their contents.
* Track cache version on BuilderState
In practice, `updateExportedFilesMapFromCache` is called repeatedly
without the cache changing in between. When this occurs, there's no
need to update the `BuilderState` (this was already the net effect, but
it took a long time to determine that no work was required).
* Fix typo in comment
* Apply disableReferencedProjectLoad to getOriginalLocationEnsuringConfiguredProject
* Reuse previously computed values and refine comments
* Add baselines for test matrix
* Make export-module and reference maps invertible
Right now, we're enumerating all the entries to find out which keys map
to a corresponding value. By maintaining a two-way map, we can convert
this linear search into a map lookup and skip allocation of many, many
iterator results.
* Fix lint error
* Add some explanatory comments
* Rename, drop type parameters, and add readonly variant
* Simplify member list
* Fold non-exporting behavior into custom map type
* Make handleDtsMayChangeOf void-returning
Right now, it always returns false. This seems important, since
otherwise it would stop graph traversals prematurely. It took me a
while to figure that out though and I thought it might be clearer if it
were simply void-returning.
I've kept the behavior the same, except in
`forEachReferencingModulesOfExportOfAffectedFile`, where it seemed like
never enqueueing new references was a mistake.
* Make forEachFileAndExportsOfFile void-returning
As far as I can tell, it could only return `false`.
* Cache parsed path mapping patterns
If a project has many of them (e.g. 1800), parsing the patterns
repeatedly can take up a lot of time.
* Move cache to ConfigFileSpecs
* Inline constants
* Simplify cache access
* Kick out of normalizePath if there's nothing to do
...using `relativePathSegmentRegExp`.
Bonus: use a regex to handle "/./" to avoid splitting and joining in a
common case.
When building the compiler, for example, it looks like ~95% of arguments
to `normalizePath` do not require any normalization.
* Check normalization before and after . cleanup
* Also cleanup leading ./
* Make normalizeSlashes a no-op there are no bad slashes
On Windows, there will probably be a negligible slowdown, iterating over
the pre-slash prefix of each unnormalized path (though we might come out
ahead if paths are normalized more than once).
On *nix, this saves work - 1.8s -> 0.4s in the project I'm
investigating.
* Reuse already-computed index
PR #42556 was a nice optimization that dramatically sped up comparisons of discriminated unions. Unfortunately, the cost of determining whether a union is discriminated can be prohibitively high. In particular, an internal team with a very large repo saw their type count double and their memory usage increase from 6GB to 9GB, breaking their build. This changes splits the difference by not trying to compute the property types of intersection types - a notoriously slow operation.
* Skip up-to-date checks during force build
Save work by not checking timestamps or parsing the buildinfo file.
Retain correctness checks (like input file existence).
Suppress project status, which was confusing anyway.
* Print a message to indicate that --force is causing the rebuild
* Don't bother updating buildInfoChecked
* Add missing force check
* Don't use _fs.realpathSync.native on windows, a semi-revert of #41292 (#43348)
We're planning a real fix for TS 4.3, but port the workaround from 4.2
so the beta doesn't have this bug.
(cherry picked from commit e462dfa347)
* Un-reverse condition
Co-authored-by: Orta Therox <ortam@microsoft.com>
* Tracing: dump more info about substitution types
* Tracing: dump more info about reverse-mapped types
* Tracing: dump more info about destructuring types
* Tracing: dump more info about evolving-array types
* Tracing: dump more info about tuple types
* Tracing: dump more info about type references
* Fix lint errors
* Tracing: extract getLocation helper
* Test that forceConsistentCasingInFileNames does not apply to Windows drive roots
* Add file symlink baselines
* Add directory symlink baselines
* Update test to retain its meaning
Its purpose is (apparently) to demonstrate that
forceConsistenCasingInFileNames can interact badly with synthesized
react imports. Since the casing of the synthesized import has changed,
also modify the casing of the explicit reference to still conflict.
* Make VFSWithWatch.realpath use the path on disk
* Update VFS realpathSync to behave like realpathSync.native
* Use fs.realpathSync.native when available
In local measurements of an Office project, we saw initial project
loading get 5% faster on Windows and 13% faster on Linux. The only
identified behavioral change is that it restores the case used on disk,
whereas realpathSync retains the input lowercase.
* Rename SortedMap.getKeyAndValue to getEntry
* Detect comparisons between large unions or intersections
If their multiplied size is greater than 1E6 (chosen based on the repro
in #41517), then we'll expend a large amount of time and memory
comparing them, so record a trace event.
Related to #41517
* Make an exception for primitive union comparisons
* Address PR feedback
* Pick up baseline change from master
* Eliminate diagnostic and only trace
* Don't check reportErrors
* Use stricter types for tracing event arguments
In local development, I've routinely passed the wrong local and ended up
having JSON.stringify throw.
* Make the type recursive
Courtesy of @rbuckton
* Fix lint error
Future versions of node will be able to return undefined, rather than
allocating and throwing an exception, when a file is not found.
See https://github.com/nodejs/node/pull/33716
* Add tracing support to tsserver
Read the `TSS_TRACE` environment variable to determine which directory
trace files should be written to.
Notable changes from tsc tracing:
1) Drop all tracepoints that depend on type IDs
2) Write output to trace.PID.json
3) New, server-specific events (request/response, cancellation, etc)
* Drop try-finally blocks that aren't strictly necessary
* Fix lint error
* Trace background work (for diagnostics)
* Move try-finally blocks into session so tsc doesn't use them
* Add missing try-finally
* Use consistent capitalization
* Inline canPop call where underlying variable is available
* Clarify comments
* Include PID in build-mode file names
* Introduce more efficient popAll function
* Trace throwIfCancellationRequested rather than isCancellationRequested
* Remove unnecessary try-finally blocks
* Add a command-line argument for consistency with logging
* Fix rebase issues
* Address PR feedback
* Rename completionEvents to eventStack
* Drop assertStackEmpty as hard-to-maintain and marginally valuable
* Rename stepCancellation to stepCanceledEarly
* Rename stepEarlyCancellation to stepCanceled and use flag instead
* Check correct variable on exit
* Delete lib/diagnosticMessages.generated.json
The values are baked in - only the localized json files are needed in
the LKG.
* Delete lib/lib.esnext files no longer found in built/local
* Copy typesMap.json in produceLKG.ts
It doesn't change very often and there's a hardcoded backup, but this
seems like a booby trap.
* Remove unused import
* Handle undefined in indent helper
Telemetry shows that it's called with undefined (probably `stderr` in an
error scenario?).
* Add undefined to parameter type
* Add test for #35011
When searching for a default configured project, stop at
`node_modules`.
* Be more explicit about inferred projects
* Move test into tsserver/projects.ts
* Use existing helpers to simplify tests
This addresses issue #28975
(https://github.com/microsoft/TypeScript/issues/28975).
When providing a value as a type argument, we can suggest a more specific
error message: "Did you mean to use typeof T?"
adds error message
WIP: Detect error
WIP: progress
updated tests
janky implementation
adds test coverage around literal types being unaffected
refactor out isIdentifierATypeArgument function
adds test case for type alias
adds test case for nested type arguments
fixes linting errors
merge master into branch to overwrite changes
changes value as type error message
This suggests 'typeof T' as a potential alternative when we give an error
about using value T as a type.
remove stale tests from old change
Co-authored-by: John Patterson <john@johnppatterson.com>
* Optionally stop looking for the default configured project at
node_modules
* Make stopping at node_modules non-optional
* Generalize and simplify the change - node_modules files don't have default configured projects
Add a response property indicating how much of the elapsed time (from
`onMessage` to `doOutput`) was spent in `updateGraph` calls. If there's
no `updateGraph` call, the property is undefined, to save space (with
the downside that it's harder to tell whether a given telemetry event
could have had the property).
Fixes#34774
Andrew Casey
