The exception thrown by Node.js's fs.statSync function contains a stack
trace that can be expensive to compute. Since this exception isn't used
by fileSystemEntryExists, we can safely set Error.stackTraceLimit to 0
without a change in behavior.
---
A significant performance improvement was noticed with this change while
profiling tsserver on packages within a proprietary monorepo.
Specifically, my team saw high self time percentages for Node.js's
uvException and handleErrorFromBinding internal functions. These
functions are executed within fs.statSync when it fails to find the
given path.
https://user-images.githubusercontent.com/906558/90183227-220cb800-dd81-11ea-8d61-f41f89481f46.png
fs.statSync: https://github.com/nodejs/node/blob/v14.4.0/lib/fs.js#L1030-L1037
handleErrorFromBinding: https://github.com/nodejs/node/blob/v14.4.0/lib/internal/fs/utils.js#L254-L269
uvException: https://github.com/nodejs/node/blob/v14.4.0/lib/internal/errors.js#L390-L443
## Measurements
After adding Error.stackTraceLimit = 0, we saw:
- For a large configured project with 12,565 files, tsserver reached the
projectLoadingFinish event 48.78% faster. (~46.786s vs ~31.447s)
- For a medium project with 7,064 files, tsserver was 25.75% faster.
(~20.897s vs ~16.618s)
- For a small project with 796 files, tsserver was only a negligible
3.00% faster. (~3.545s vs ~3.442)
Measurements were taken on macOS 10.15.6, Node.js 14.4.0, and a recent
master commit of TypeScript (610fa28d). The average of 3 runs before and
after this change were taken.
I would normally include .cpuprofile and isolate-*-*-*.log files, but
can't post them publicly in this case. If there's any other summaries
the TypeScript team would be curious about I can report them.
## fs.statSync Misses
Within our monorepo, the fs.statSync misses were mostly searches for
alternative file extensions of module imports.
- For node_modules imports, a lot of .ts/.tsx lookups failed until the
.d.ts file was found.
- Within projects with a lot of JSX files, .ts files were looked for
before finding the .tsx version.
- In the medium scale project mentioned above, a total of 38,515
non-existent files were queried during createProgram.
* Add full implemention of Map and Set to shims
* Update default Map interface
* Remove WeakMap/WeakSet
* Add tests for set shim
* Update most usages of Map<K, true> to Set
* PR Feedback
* Fix lint issues
* Change key in fsWatchCallback
* Simpler shim, more tests
* Fix typo in collection shim
* If there is no changes to folder structure when watching directories recursively, send the updates to fileNames only
Fixes#37994
* Ignore excluded directories from wild card watching
* Testcase showing that renaming file with non sync directory watcher displays correct error in the end
Testcase for #38684
* Create different watch options in compiler options
* Thread through the new watch options
* Actually use the options passed through for watch strategy
* Support delay on updating child directory watches
* Make watchOptions separate from compilerOptions
* Support passing watch options from command line
* Handle displaying of watchOptions
* Remove redundant checker functions, use patterns more friendly to modules
* Also use a helper for localizedDiagnosticMessages
* Move types into same file as consts
* Accept slightly changed api baseline
* Whitespace!
Checking for directory existence is expensive and frequently indicates
success. Instead of pre-emptively creating the directory containing a
file to be written, attempt to create the file and only do the directory
scaffolding if the write fails.
Appears to reduce file write time by 10-20% for a file-I/O heavy partner
build.
Thanks to @rbuckton for the suggestion!
* Add support for capturing cpu profiles into tsc itself
* Accept baseline for new compiler option in showConfig
* Fix lints
* Support profiling build mode, only ever have one live profiling session
* Minor modification to enable/disable semaphore, accept re-cased baseline
* Add pid into autognerated cpuprofile path
* Rename to fix case
* Sanitize filepaths in emitted cpuprofile for easier adoption by enterprise people, add inspector to browser field
* Update LKG
* Add @types/node version bump
* Small paatches/casts to be compatible with latest node
* Accept API baseline update
* Make internal NodeBuffer compatable with latest Buffer
* Why do we even have an internal buffer type
* Sync up internal buffer type better
* Fix lint
* Readd mroe missing Buffer methods
