Announcing TypeScript 5.5 RC
Today we are excited to announce the availability of the release candidate of TypeScript 5.5.
To get started using the RC, you can get it through NuGet, or through npm with the following command:
npm install -D typescript@rc
Here’s a quick list of what’s new in TypeScript 5.5!
- Inferred Type Predicates
- Control Flow Narrowing for Constant Indexed Accesses
- Type Imports in JSDoc
- Regular Expression Syntax Checking
- Support for New ECMAScript
Set
Methods - Isolated Declarations
- The
${configDir}
Template Variable for Configuration Files - Consulting
package.json
Dependencies for Declaration File Generation - Editor and Watch-Mode Reliability Improvements
- Performance and Size Optimizations
- Easier API Consumption from ECMAScript Modules
- The
transpileDeclaration
API - Notable Behavioral Changes
What’s New Since the Beta?
Since the beta, we’ve made a few changes that we wanted to call out.
For one, we added support for ECMAScript’s new Set
methods. Additionally, we’ve adjusted the behavior of TypeScript’s new regular expression checking to be slightly more lenient, while still erroring on questionable escapes that are only allowed per ECMAScript’s Annex B.
We’ve also added and documented even more performance optimizations: notably, skipped checking in transpileModule
and optimizations in how TypeScript filters contextual types.
These optimizations can lead to faster build and iteration time in many common scenarios.
Inferred Type Predicates
This section was written by Dan Vanderkam, who implemented this feature in TypeScript 5.5. Thanks Dan!
TypeScript’s control flow analysis does a great job of tracking how the type of a variable changes as it moves through your code:
interface Bird {
commonName: string;
scientificName: string;
sing(): void;
}
// Maps country names -> national bird.
// Not all nations have official birds (looking at you, Canada!)
declare const nationalBirds: Map<string, Bird>;
function makeNationalBirdCall(country: string) {
const bird = nationalBirds.get(country); // bird has a declared type of Bird | undefined
if (bird) {
bird.sing(); // bird has type Bird inside the if statement
} else {
// bird has type undefined here.
}
}
By making you handle the undefined
case, TypeScript pushes you to write more robust code.
In the past, this sort of type refinement was more difficult to apply to arrays. This would have been an error in all previous versions of TypeScript:
function makeBirdCalls(countries: string[]) {
// birds: (Bird | undefined)[]
const birds = countries
.map(country => nationalBirds.get(country))
.filter(bird => bird !== undefined);
for (const bird of birds) {
bird.sing(); // error: 'bird' is possibly 'undefined'.
}
}
This code is perfectly fine: we’ve filtered all the undefined
values out of the list.
But TypeScript hasn’t been able to follow along.
With TypeScript 5.5, the type checker is fine with this code:
function makeBirdCalls(countries: string[]) {
// birds: Bird[]
const birds = countries
.map(country => nationalBirds.get(country))
.filter(bird => bird !== undefined);
for (const bird of birds) {
bird.sing(); // ok!
}
}
Note the more precise type for birds
.
This works because TypeScript now infers a type predicate for the filter
function.
You can see what’s going on more clearly by pulling it out into a standalone function:
// function isBirdReal(bird: Bird | undefined): bird is Bird
function isBirdReal(bird: Bird | undefined) {
return bird !== undefined;
}
bird is Bird
is the type predicate.
It means that, if the function returns true
, then it’s a Bird
(if the function returns false
then it’s undefined
).
The type declarations for Array.prototype.filter
know about type predicates, so the net result is that you get a more precise type and the code passes the type checker.
TypeScript will infer that a function returns a type predicate if these conditions hold:
- The function does not have an explicit return type or type predicate annotation.
- The function has a single
return
statement and no implicit returns. - The function does not mutate its parameter.
- The function returns a
boolean
expression that’s tied to a refinement on the parameter.
Generally this works how you’d expect. Here’s a few more examples of inferred type predicates:
// const isNumber: (x: unknown) => x is number
const isNumber = (x: unknown) => typeof x === 'number';
// const isNonNullish: <T>(x: T) => x is NonNullable<T>
const isNonNullish = <T,>(x: T) => x != null;
Previously, TypeScript would have just inferred that these functions return boolean
.
It now infers signatures with type predicates like x is number
or x is NonNullable<T>
.
Type predicates have "if and only if" semantics.
If a function returns x is T
, then it means that:
- If the function returns
true
thenx
is has typeT
. - If the function returns
false
thenx
does not have typeT
.
If you’re expecting a type predicate to be inferred but it’s not, then you may be running afoul of the second rule. This often comes up with "truthiness" checks:
function getClassroomAverage(students: string[], allScores: Map<string, number>) {
const studentScores = students
.map(student => allScores.get(student))
.filter(score => !!score);
return studentScores.reduce((a, b) => a + b) / studentScores.length;
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// error: Object is possibly 'undefined'.
}
TypeScript did not infer a type predicate for score => !!score
, and rightly so: if this returns true
then score
is a number
.
But if it returns false
, then score
could be either undefined
or a number
(specifically, 0
).
This is a real bug: if any student got a zero on the test, then filtering out their score will skew the average upwards.
Fewer will be above average and more will be sad!
As with the first example, it’s better to explicitly filter out undefined
values:
function getClassroomAverage(students: string[], allScores: Map<string, number>) {
const studentScores = students
.map(student => allScores.get(student))
.filter(score => score !== undefined);
return studentScores.reduce((a, b) => a + b) / studentScores.length; // ok!
}
A truthiness check will infer a type predicate for object types, where there’s no ambiguity.
Remember that functions must return a boolean
to be a candidate for an inferred type predicate: x => !!x
might infer a type predicate, but x => x
definitely won’t.
Explicit type predicates continue to work exactly as before. TypeScript will not check whether it would infer the same type predicate. Explicit type predicates ("is") are no safer than a type assertion ("as").
It’s possible that this feature will break existing code if TypeScript now infers a more precise type than you want. For example:
// Previously, nums: (number | null)[]
// Now, nums: number[]
const nums = [1, 2, 3, null, 5].filter(x => x !== null);
nums.push(null); // ok in TS 5.4, error in TS 5.5
The fix is to tell TypeScript the type that you want using an explicit type annotation:
const nums: (number | null)[] = [1, 2, 3, null, 5].filter(x => x !== null);
nums.push(null); // ok in all versions
For more information, check out the implementing pull request and Dan’s blog post about implementing this feature.
Control Flow Narrowing for Constant Indexed Accesses
TypeScript is now able to narrow expressions of the form obj[key]
when both obj
and key
are effectively constant.
function f1(obj: Record<string, unknown>, key: string) {
if (typeof obj[key] === "string") {
// Now okay, previously was error
obj[key].toUpperCase();
}
}
In the above, neither obj
nor key
are ever mutated, so TypeScript can narrow the type of obj[key]
to string
after the typeof
check.
For more information, see the implementing pull request here.
Type Imports in JSDoc
Today, if you want to import something only for type-checking in a JavaScript file, it is cumbersome.
JavaScript developers can’t simply import a type named SomeType
if it’s not there at runtime.
// ./some-module.d.ts
export interface SomeType {
// ...
}
// ./index.js
import { SomeType } from "./some-module"; // ❌ runtime error!
/**
* @param {SomeType} myValue
*/
function doSomething(myValue) {
// ...
}
SomeType
won’t exist at runtime, so the import will fail.
Developers can instead use a namespace import instead.
import * as someModule from "./some-module";
/**
* @param {someModule.SomeType} myValue
*/
function doSomething(myValue) {
// ...
}
But ./some-module
is still imported at runtime – which might also not be desirable.
To avoid this, developers typically had to use import(...)
types in JSDoc comments.
/**
* @param {import("./some-module").SomeType} myValue
*/
function doSomething(myValue) {
// ...
}
If you wanted to reuse the same type in multiple places, you could use a typedef
to avoid repeating the import.
/**
* @typedef {import("./some-module").SomeType} SomeType
*/
/**
* @param {SomeType} myValue
*/
function doSomething(myValue) {
// ...
}
This helps with local uses of SomeType
, but it gets repetitive for many imports and can be a bit verbose.
That’s why TypeScript now supports a new @import
comment tag that has the same syntax as ECMAScript imports.
/** @import { SomeType } from "some-module" */
/**
* @param {SomeType} myValue
*/
function doSomething(myValue) {
// ...
}
Here, we used named imports. We could also have written our import as a namespace import.
/** @import * as someModule from "some-module" */
/**
* @param {someModule.SomeType} myValue
*/
function doSomething(myValue) {
// ...
}
Because these are just JSDoc comments, they don’t affect runtime behavior at all.
We would like to extend a big thanks to Oleksandr Tarasiuk who contributed this change!
Regular Expression Syntax Checking
Until now, TypeScript has typically skipped over most regular expressions in code. This is because regular expressions technically have an extensible grammar and TypeScript never made any effort to compile regular expressions to earlier versions of JavaScript. Still, this meant that lots of common problems would go undiscovered in regular expressions, and they would either turn into errors at runtime, or silently fail.
But TypeScript now does basic syntax checking on regular expressions!
let myRegex = /@robot(\s+(please|immediately)))? do some task/;
// ~
// error!
// Unexpected ')'. Did you mean to escape it with backslash?
This is a simple example, but this checking can catch a lot of common mistakes. In fact, TypeScript’s checking goes slightly beyond syntactic checks. For instance, TypeScript can now catch issues around backreferences that don’t exist.
let myRegex = /@typedef \{import\((.+)\)\.([a-zA-Z_]+)\} \3/u;
// ~
// error!
// This backreference refers to a group that does not exist.
// There are only 2 capturing groups in this regular expression.
The same applies to named capturing groups.
let myRegex = /@typedef \{import\((?<importPath>.+)\)\.(?<importedEntity>[a-zA-Z_]+)\} \k<namedImport>/;
// ~~~~~~~~~~~
// error!
// There is no capturing group named 'namedImport' in this regular expression.
TypeScript’s checking is now also aware of when certain RegExp features are used when newer than your target version of ECMAScript. For example, if we use named capturing groups like the above in an ES5 target, we’ll get an error.
let myRegex = /@typedef \{import\((?<importPath>.+)\)\.(?<importedEntity>[a-zA-Z_]+)\} \k<importedEntity>/;
// ~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~
// error!
// Named capturing groups are only available when targeting 'ES2018' or later.
The same is true for certain regular expression flags as well.
Note that TypeScript’s regular expression support is limited to regular expression literals.
If you try calling new RegExp
with a string literal, TypeScript will not check the provided string.
We would like to thank GitHub user graphemecluster who iterated a ton with us to get this feature into TypeScript.
Support for New ECMAScript Set
Methods
TypeScript 5.5 declares new proposed methods for the ECMAScript Set
type.
Some of these methods, like union
, intersection
, difference
, and symmetricDifference
, take another Set
and return a new Set
as the result.
The other methods, isSubsetOf
, isSupersetOf
, and isDisjointFrom
, take another Set
and return a boolean
.
None of these methods mutate the original Set
s.
Here’s a quick example of how you might use these methods and how they behave:
let fruits = new Set(["apples", "bananas", "pears", "oranges"]);
let applesAndBananas = new Set(["apples", "bananas"]);
let applesAndOranges = new Set(["apples", "oranges"]);
let oranges = new Set(["oranges"]);
let emptySet = new Set();
////
// union
////
// Set(4) {'apples', 'bananas', 'pears', 'oranges'}
console.log(fruits.union(oranges));
// Set(3) {'apples', 'bananas', 'oranges'}
console.log(applesAndBananas.union(oranges));
////
// intersection
////
// Set(2) {'apples', 'bananas'}
console.log(fruits.intersection(applesAndBananas));
// Set(0) {}
console.log(applesAndBananas.intersection(oranges));
// Set(1) {'apples'}
console.log(applesAndBananas.intersection(applesAndOranges));
////
// difference
////
// Set(3) {'apples', 'bananas', 'pears'}
console.log(fruits.difference(oranges));
// Set(2) {'pears', 'oranges'}
console.log(fruits.difference(applesAndBananas));
// Set(1) {'bananas'}
console.log(applesAndBananas.difference(applesAndOranges));
////
// symmetricDifference
////
// Set(2) {'bananas', 'oranges'}
console.log(applesAndBananas.symmetricDifference(applesAndOranges)); // no apples
////
// isDisjointFrom
////
// true
console.log(applesAndBananas.isDisjointFrom(oranges));
// false
console.log(applesAndBananas.isDisjointFrom(applesAndOranges));
// true
console.log(fruits.isDisjointFrom(emptySet));
// true
console.log(emptySet.isDisjointFrom(emptySet));
////
// isSubsetOf
////
// true
console.log(applesAndBananas.isSubsetOf(fruits));
// false
console.log(fruits.isSubsetOf(applesAndBananas));
// false
console.log(applesAndBananas.isSubsetOf(oranges));
// true
console.log(fruits.isSubsetOf(fruits));
// true
console.log(emptySet.isSubsetOf(fruits));
////
// isSupersetOf
////
// true
console.log(fruits.isSupersetOf(applesAndBananas));
// false
console.log(applesAndBananas.isSupersetOf(fruits));
// false
console.log(applesAndBananas.isSupersetOf(oranges));
// true
console.log(fruits.isSupersetOf(fruits));
// false
console.log(emptySet.isSupersetOf(fruits));
We’d like to thank Kevin Gibbons who not only co-championed the feature in ECMAScript, but also provided the declarations for Set
, ReadonlySet
, and ReadonlySetLike
in TypeScript!
Isolated Declarations
This section was co-authored by Rob Palmer who supported the design of isolated declarations.
Declaration files (a.k.a. .d.ts
files) describe the shape of existing libraries and modules to TypeScript.
This lightweight description includes the library’s type signatures and excludes implementation details such as the function bodies.
They are published so that TypeScript can efficiently check your usage of a library without needing to analyse the library itself.
Whilst it is possible to handwrite declaration files, if you are authoring typed code, it’s much safer and simpler to let TypeScript generate them automatically from source files using --declaration
.
The TypeScript compiler and its APIs have always had the job of generating declaration files; however, there are some use-cases where you might want to use other tools, or where the traditional build process doesn’t scale.
Use-case: Faster Declaration Emit Tools
Imagine if you wanted to create a faster tool to generate declaration files, perhaps as part of a publishing service or a new bundler. Whilst there is a thriving ecosystem of blazing fast tools that can turn TypeScript into JavaScript, the same is not true for turning TypeScript into declaration files. The reason is that TypeScript’s inference allows us to write code without explicitly declaring types, meaning declaration emit can be complex.
Let’s consider a simple example of a function that adds two imported variables.
// util.ts
export let
export let two = "2";
// add.ts
import { one, two } from "./util";
export function add() { return one + two; }
Even if the only thing we want to do is generate add.d.ts
, TypeScript needs to crawl into another imported file (util.ts
), infer that the type of one
and two
are strings, and then calculate that the +
operator on two strings will lead to a string
return type.
// add.d.ts
export declare function add(): string;
While this inference is important for the developer experience, it means that tools that want to generate declaration files would need to replicate parts of the type-checker including inference and the ability to resolve module specifiers to follow the imports.
Use-case: Parallel Declaration Emit and Parallel Checking
Imagine if you had a monorepo containing many projects and a multi-core CPU that just wished it could help you check your code faster. Wouldn’t it be great if we could check all those projects at the same time by running each project on a different core?
Unfortunately we don’t have the freedom to do all the work in parallel. The reason is that we have to build those projects in dependency order, because each project is checking against the declaration files of their dependencies. So we must build the dependency first to generate the declaration files. TypeScript’s project references feature works the same way, building the set of projects in "topological" dependency order.
As an example, if we have two projects called backend
and frontend
, and they both depend on a project called core
, TypeScript can’t start type-checking either frontend
or backend
until core
has been built and its declaration files have been generated.
In the above graph, you can see that we have a bottleneck.
Whilst we can build frontend
and backend
in parallel, we need to first wait for core
to finish building before either can start.
How could we improve upon this?
Well, if a fast tool could generate all those declaration files for core
in parallel, TypeScript then could immediately follow that by type-checking core
, frontend
, and backend
also in parallel.
Solution: Explicit Types!
The common requirement in both use-cases is that we need a cross-file type-checker to generate declaration files. Which is a lot to ask from the tooling community.
As a more complex example, if we want a declaration file for the following code…
import { add } from "./add";
const x = add();
export function foo() {
return x;
}
…we would need to generate a signature for foo
.
Well that requires looking at the implementation of foo
.
foo
just returns x
, so getting the type of x
requires looking at the implementation of add
.
But that might require looking at the implementation of add
‘s dependencies, and so on.
What we’re seeing here is that generating declaration files requires a whole lot of logic to figure out the types of different places that might not even be local to the current file.
Still, for developers looking for fast iteration time and fully parallel builds, there is another way of thinking about this problem. A declaration file only requires the types of the public API of a module – in other words, the types of the things that are exported. If, controversially, developers are willing to explicitly write out the types of the things they export, tools could generate declaration files without needing to look at the implementation of the module – and without reimplementing a full type-checker.
This is where the new --isolatedDeclarations
option comes in.
--isolatedDeclarations
reports errors when a module can’t be reliably transformed without a type-checker.
More plainly, it makes TypeScript report errors if you have a file that isn’t sufficiently annotated on its exports.
That means in the above example, we would see an error like the following:
export function foo() {
// ~~~
// error! Function must have an explicit
// return type annotation with --isolatedDeclarations.
return x;
}
Why are errors desirable?
Because it means that TypeScript can
- Tell us up-front whether other tools will have issues with generating declaration files
- Provide a quick fix to help add these missing annotations.
This mode doesn’t require annotations everywhere though. For locals, these can be ignored, since they don’t affect the public API. For example, the following code would not produce an error:
import { add } from "./add";
const x = add("1", "2"); // no error on 'x', it's not exported.
export function foo(): string {
return x;
}
There are also certain expressions where the type is "trivial" to calculate.
// No error on 'x'.
// It's trivial to calculate the type is 'number'
export let x = 10;
// No error on 'y'.
// We can get the type from the return expression.
export function y() {
return 20;
}
// No error on 'z'.
// The type assertion makes it clear what the type is.
export function z() {
return Math.max(x, y()) as number;
}
Using isolatedDeclarations
isolatedDeclarations
requires that either the declaration
or composite
flags are also set.
Note that isolatedDeclarations
does not change how TypeScript performs emit – just how it reports errors.
Importantly, and similar to isolatedModules
, enabling the feature in TypeScript won’t immediately bring about the potential benefits discussed here.
So please be patient and look forward to future developments in this space.
Keeping tool authors in mind, we should also recognize that today, not all of TypeScript’s declaration emit can be easily replicated by other tools wanting to use it as a guide.
That’s something we’re actively working on improving.
On top of this, isolated declarations are still a new feature, and we’re actively working on improving the experience.
Some scenarios, like using computed property declarations in classes and object literals, are not yet supported under isolatedDeclarations
.
Keep an eye on this space, and feel free to provide us with feedback.
We also feel it is worth calling out that isolatedDeclarations
should be adopted on a case-by-case basis.
There are some developer ergonomics that are lost when using isolatedDeclarations
, and thus it may not be the right choice if your setup is not leveraging the two scenarios mentioned earlier.
For others, the work on isolatedDeclarations
has already uncovered many optimizations and opportunities to unlock different parallel build strategies.
In the meantime, if you’re willing to make the trade-offs, we believe isolatedDeclarations
can be a powerful tool to speed up your build process once external tooling becomes available.
Credit
Work on isolatedDeclarations
has been a long-time collaborative effort between the TypeScript team and the infrastructure and tooling teams within Bloomberg and Google.
Individuals like Hana Joo from Google who implemented the quick fix for isolated declaration errors (more on that soon), as well as Ashley Claymore, Jan Kühle, Lisa Velden, Rob Palmer, and Thomas Chetwin have been involved in discussion, specification, and implementation for many months.
But we feel it is specifically worth calling out the tremendous amount of work provided by Titian Cernicova-Dragomir from Bloomberg.
Titian has been instrumental in driving the implementation of isolatedDeclarations
and has been a contributor to the TypeScript project for years prior.
While the feature involved many changes, you can see the core work for Isolated Declarations here.
The ${configDir}
Template Variable for Configuration Files
It’s common in many codebases to reuse a shared tsconfig.json
file that acts as a "base" for other configuration files.
This is done by using the extends
field in a tsconfig.json
file.
{
"extends": "../../tsconfig.base.json",
"compilerOptions": {
"outDir": "./dist"
}
}
One of the issues with this is that all paths in the tsconfig.json
file are relative to the location of the file itself.
This means that if you have a shared tsconfig.base.json
file that is used by multiple projects, relative paths often won’t be useful in the derived projects.
For example, imagine the following tsconfig.base.json
:
{
"compilerOptions": {
"typeRoots": [
"./node_modules/@types"
"./custom-types"
],
"outDir": "dist"
}
}
If author’s intent was that every tsconfig.json
that extends this file should
- output to a
dist
directory relative to the derivedtsconfig.json
, and - have a
custom-types
directory relative to the derivedtsconfig.json
,
then this would not work.
The typeRoots
paths would be relative to the location of the shared tsconfig.base.json
file, not the project that extends it.
Each project that extends this shared file would need to declare its own outDir
and typeRoots
with identical contents.
This could be frustrating and hard to keep in sync between projects, and while the example above is using typeRoots
, this is a common problem for paths
and other options.
To solve this, TypeScript 5.5 introduces a new template variable ${configDir}
.
When ${configDir}
is written in certain path fields of a tsconfig.json
or jsconfig.json
files, this variable is substituted with the containing directory of the configuration file in a given compilation.
This means that the above tsconfig.base.json
could be rewritten as:
{
"compilerOptions": {
"typeRoots": [
"${configDir}/node_modules/@types"
"${configDir}/custom-types"
],
"outDir": "${configDir}/dist"
}
}
Now, when a project extends this file, the paths will be relative to the derived tsconfig.json
, not the shared tsconfig.base.json
file.
This makes it easier to share configuration files across projects and ensures that the configuration files are more portable.
If you intend to make a tsconfig.json
file extendable, consider if a ./
should instead be written with ${configDir}
.
For more information, see the proposal issue and the implementing pull request.
Consulting package.json
Dependencies for Declaration File Generation
Previously, TypeScript would often issue an error message like
The inferred type of "X" cannot be named without a reference to "Y". This is likely not portable. A type annotation is necessary.
This was often due to TypeScript’s declaration file generation finding itself in the contents of files that were never explicitly imported in a program.
Generating an import to such a file could be risky if the path ended up being relative.
Still, for codebases with explicit dependencies in the dependencies
(or peerDependencies
and optionalDependencies
) of a package.json
, generating such an import should be safe under certain resolution modes.
So in TypeScript 5.5, we’re more lenient when that’s the case, and many occurrences of this error should disappear.
See this pull request for more details on the change.
Editor and Watch-Mode Reliability Improvements
TypeScript has either added some new functionality or fixed existing logic that makes --watch
mode and TypeScript’s editor integration feel more reliable.
That should hopefully translate to fewer TSServer/editor restarts.
Correctly Refresh Editor Errors in Configuration Files
TypeScript can generate errors for tsconfig.json
files;
however, those errors are actually generated from loading a project, and editors typically don’t directly request those errors for tsconfig.json
files.
While this sounds like a technical detail, it means that when all errors issued in a tsconfig.json
are fixed, TypeScript doesn’t issue a new fresh empty set of errors, and users are left with stale errors unless they reload their editor.
TypeScript 5.5 now intentionally issues an event to clear these out. See more here.
Better Handling for Deletes Followed by Immediate Writes
Instead of overwriting files, some tools will opt to delete them and then create new files from scratch.
This is the case when running npm ci
, for instance.
While this can be efficient for those tools, it can be problematic for TypeScript’s editor scenarios where deleting a watched might dispose of it and all of its transitive dependencies. Deleting and creating a file in quick succession could lead to TypeScript tearing down an entire project and then rebuilding it from scratch.
TypeScript 5.5 now has a more nuanced approach by keeping parts of a deleted project around until it picks up on a new creation event.
This should make operations like npm ci
work a lot better with TypeScript.
See more information on the approach here.
Symlinks are Tracked in Failed Resolutions
When TypeScript fails to resolve a module, it will still need to watch for any failed lookup paths in case the module is added later. Previously this was not done for symlinked directories, which could cause reliability issues in monorepo-like scenarios when a build occurred in one project but was not witnessed in the other. This should be fixed in TypeScript 5.5, and means you won’t need to restart your editor as often.
Project References Contribute to Auto-Imports
Auto-imports no longer requires at least one explicit import to dependent projects in a project reference setup.
Instead, auto-import completions should just work across anything you’ve listed in the references
field of your tsconfig.json
.
See more on the implementing pull request.
Performance and Size Optimizations
Monomorphized Objects in Language Service and Public API
In TypeScript 5.0, we ensured that our Node
and Symbol
objects had a consistent set of properties with a consistent initialization order.
Doing so helps reduce polymorphism in different operations, which allows runtimes to fetch properties more quickly.
By making this change, we witnessed impressive speed wins in the compiler; however, most of these changes were performed on internal allocators for our data structures. The language service, along with TypeScript’s public API, uses a different set of allocators for certain objects. This allowed the TypeScript compiler to be a bit leaner, as data used only for the language service would never be used in the compiler.
In TypeScript 5.5, the same monomorphization work has been done for the language service and public API. What this means is that your editor experience, and any build tools that use the TypeScript API, will get a decent amount faster. In fact, in our benchmarks, we’ve seen a 5-8% speedup in build times when using the public TypeScript API’s allocators, and language service operations getting 10-20% faster. While this does imply an increase in memory, we believe that tradeoff is worth it and hope to find ways to reduce that memory overhead. Things should feel a lot snappier now.
For more information, see the change here.
Monomorphized Control Flow Nodes
In TypeScript 5.5, nodes of the control flow graph have been monomorphized so that they always hold a consistent shape. By doing so, check times will often be reduced by about 1%.
Optimizations on our Control Flow Graph
In many cases, control flow analysis will traverse nodes that don’t provide any new information. We observed that in the absence of any early termination or effects in the antecedents (or "dominators") of certain nodes meant that those nodes could always be skipped over. As such, TypeScript now constructs its control flow graphs to take advantage of this by linking to an earlier node that does provide interesting information for control flow analysis. This yields a flatter control flow graph, which can be more efficient to traverse. This optimization has yielded modest gains, but with up to 2% reductions in build time on certain codebases.
You can read more here.
Skipped Checking in transpileModule
and transpileDeclaration
TypeScript’s transpileModule
API can be used for compiling a single TypeScript file’s contents into JavaScript.
Similarly, the transpileDeclaration
API (see below) can be used to generate a declaration file for a single TypeScript file.
One of the issues with these APIs is that TypeScript internally would perform a full type-checking pass over the entire contents of the file before emitting the output.
This was necessary to collect certain information which would later be used for the emit phase.
In TypeScript 5.5, we’ve found a way to avoid performing a full check, only lazily collecting this information as necessary, and transpileModule
and transpileDeclaration
both enable this functionality by default.
As a result, tools that integrate with with these APIs, like ts-loader with transpileOnly
and ts-jest, should see a noticeable speedup.
In our testing, we generally witness around a 2x speed-up in build time using transpileModule
.
TypeScript Package Size Reduction
Further leveraging our transition to modules in 5.0, we’ve significantly reduced TypeScript’s overall package size by making tsserver.js
and typingsInstaller.js
import from a common API library instead of having each of them produce standalone bundles.
This reduces TypeScript’s size on disk from 30.2 MB to 20.4 MB, and reduces its packed size from 5.5 MB to 3.7 MB!
Node Reuse in Declaration Emit
As part of the work to enable isolatedDeclarations
, we’ve substantially improved how often TypeScript can directly copy your input source code when producing declaration files.
For example, let’s say you wrote
export const strBool: string | boolean = "hello";
export const boolStr: boolean | string = "world";
Note that the union types are equivalent, but the order of the union is different. When emitting the declaration file, TypeScript has two equivalent output possibilities.
The first is to use a consistent canonical representation for each type:
export const strBool: string | boolean;
export const boolStr: string | boolean;
The second is to re-use the type annotations exactly as written:
export const strBool: string | boolean;
export const boolStr: boolean | string;
The second approach is generally preferable for a few reasons:
- Many equivalent representations still encode some level of intent that is better to preserve in the declaration file
- Producing a fresh representation of a type can be somewhat expensive, so avoiding is better
- User-written types are usually shorter than generated type representations
In 5.5, we’ve greatly improved the number of places where TypeScript can correctly identify places where it’s safe and correct to print back types exactly as they were written in the input file. Many of these cases are invisible performance improvements – TypeScript would generate fresh sets of syntax nodes and serialize them into a string. Instead, TypeScript can now operate over the original syntax nodes directly, which is much cheaper and faster.
Caching Contextual Types from Discriminated Unions
When TypeScript asks for the contextual type of an expression like an object literal, it will often encounter a union type. In those cases, TypeScript tries to filter out members of the union based on known properties with well known values (i.e. discriminant properties). This work can be fairly expensive, especially if you end up with an object consisting of many many properties. In TypeScript 5.5, much of the computation is cached once so that TypeScript doesn’t need to recompute it for every property in the object literal. Performing this optimization shaved 250ms off of compiling the TypeScript compiler itself.
Easier API Consumption from ECMAScript Modules
Previously, if you were writing an ECMAScript module in Node.js, named imports were not available from the typescript
package.
import { createSourceFile } from "typescript"; // ❌ error
import * as ts from "typescript";
ts.createSourceFile // ❌ undefined???
ts.default.createSourceFile // ✅ works - but ugh!
This is because cjs-module-lexer did not recognize the pattern of TypeScript’s generated CommonJS code. This has been fixed, and users can now use named imports from the TypeScript npm package with ECMAScript modules in Node.js.
import { createSourceFile } from "typescript"; // ✅ works now!
import * as ts from "typescript";
ts.createSourceFile // ✅ works now!
For more information, see the change here.
The transpileDeclaration
API
TypeScript’s API exposes a function called transpileModule
.
It’s intended to make it easy to compile a single file of TypeScript code.
Because it doesn’t have access to an entire program, the caveat is that it may not produce the right output if the code violates any errors under the isolatedModules
option.
In TypeScript 5.5, we’ve added a new similar API called transpileDeclaration
.
This API is similar to transpileModule
, but it’s specifically designed to generate a single declaration file based on some input source text.
Just like transpileModule
, it doesn’t have access to a full program, and a similar caveat applies: it only generates an accurate declaration file if the input code is free of errors under the new isolatedDeclarations
option.
If desired, this function can be used to parallelize declaration emit across all files under isolatedDeclarations
mode.
For more information, see the implementation here.
Notable Behavioral Changes
This section highlights a set of noteworthy changes that should be acknowledged and understood as part of any upgrade. Sometimes it will highlight deprecations, removals, and new restrictions. It can also contain bug fixes that are functionally improvements, but which can also affect an existing build by introducing new errors.
Disabling Features Deprecated in TypeScript 5.0
TypeScript 5.0 deprecated the following options and behaviors:
charset
target: ES3
importsNotUsedAsValues
noImplicitUseStrict
noStrictGenericChecks
keyofStringsOnly
suppressExcessPropertyErrors
suppressImplicitAnyIndexErrors
out
preserveValueImports
prepend
in project references- implicitly OS-specific
newLine
To continue using the deprecated options above, developers using TypeScript 5.0 and other more recent versions have had to specify a new option called ignoreDeprecations
with the value "5.0"
.
In TypeScript 5.5, these options no longer have any effect. To help with a smooth upgrade path, you may still specify them in your tsconfig, but these will be an error to specify in TypeScript 6.0. See also the Flag Deprecation Plan which outlines our deprecation strategy.
More information around these deprecation plans is available on GitHub, which contains suggestions in how to best adapt your codebase.
lib.d.ts
Changes
Types generated for the DOM may have an impact on type-checking your codebase. For more information, see the DOM updates for TypeScript 5.5.
Respecting File Extensions and package.json
in Other Module Modes
Before Node.js implemented support for ECMAScript modules in v12, there was never a good way for TypeScript to know whether .d.ts
files it found in node_modules
represented JavaScript files authored as CommonJS or ECMAScript modules.
When the vast majority of npm was CommonJS-only, this didn’t cause many problems – if in doubt, TypeScript could just assume that everything behaved like CommonJS.
Unfortunately, if that assumption was wrong it could allow unsafe imports:
// node_modules/dep/index.d.ts
export declare function doSomething(): void;
// index.ts
// Okay if "dep" is a CommonJS module, but fails if
// it's an ECMAScript module - even in bundlers!
import dep from "dep";
dep.doSomething();
In practice, this didn’t come up very often.
But in the years since Node.js started supporting ECMAScript modules, the share of ESM on npm has grown.
Fortunately, Node.js also introduced a mechanism that can help TypeScript determine if a file is an ECMAScript module or a CommonJS module: the .mjs
and .cjs
file extensions and the package.json
"type"
field.
TypeScript 4.7 added support for understanding these indicators, as well as authoring .mts
and .cts
files;
however, TypeScript would only read those indicators under --module node16
and --module nodenext
, so the unsafe import above was still a problem for anyone using --module esnext
and --moduleResolution bundler
, for example.
To solve this, TypeScript 5.5 reads and stores module format information encoded by file extensions and package.json
"type"
in all module
modes, and uses it to resolve ambiguities like the one in the example above in all modes (except for amd
, umd
, and system
).
A secondary effect of respecting this format information is that the format-specific TypeScript file extensions (.mts
and .cts
) or an explicitly set package.json "type"
in your own project will override your --module
option if it’s set to commonjs
or es2015
through esnext
.
Previously, it was technically possible to produce CommonJS output into a .mjs
file or vice versa:
// main.mts
export default "oops";
// $ tsc --module commonjs main.mts
// main.mjs
Object.defineProperty(exports, "__esModule", { value: true });
exports.default = "oops";
Now, .mts
files (or .ts
files in scope of a package.json
with "type": "module"
) never emit CommonJS output, and .cts
files (or .ts
files in scope of a package.json with "type": "commonjs"
) never emit ESM output.
More details are available on the change here.
Stricter Parsing for Decorators
Since TypeScript originally introduced support for decorators, the specified grammar for the proposal has been tightened up. TypeScript is now stricter about what forms it allows. While rare, existing decorators may need to be parenthesized to avoid errors.
class DecoratorProvider {
decorate(...args: any[]) { }
}
class D extends DecoratorProvider {
m() {
class C {
@super.decorate // ❌ error
method1() { }
@(super.decorate) // ✅ okay
method2() { }
}
}
}
See more information on the change here.
undefined
is No Longer a Definable Type Name
TypeScript has always disallowed type alias names that conflict with built-in types:
// Illegal
type null = any;
// Illegal
type number = any;
// Illegal
type object = any;
// Illegal
type any = any;
Due to a bug, this logic didn’t also apply to the built-in type undefined
.
In 5.5, this is now correctly identified as an error:
// Now also illegal
type undefined = any;
Bare references to type aliases named undefined
never actually worked in the first place.
You could define them, but you couldn’t use them as an unqualified type name.
export type undefined = string;
export const m: undefined = "";
// ^
// Errors in 5.4 and earlier - the local definition of 'undefined' was not even consulted.
For more information, see the change here.
Simplified Reference Directive Declaration Emit
When producing a declaration file, TypeScript would synthesize a reference directive when it believed one was required. For example, all Node.js modules are declared ambiently, so cannot be loaded by module resolution alone. A file like:
import path from "path";
export const myPath = path.parse(__filename);
Would emit a declaration file like:
/// <reference types="node" />
import path from "path";
export declare const myPath: path.ParsedPath;
Even though the reference directive never appeared in the original source.
Similarly, TypeScript also removed reference directives that it did not believe needed to be a part of the output.
For example, let’s imagine we had a reference directive to jest
;
however, imagine the reference directive isn’t necessary to generate the declaration file.
TypeScript would simply drop it.
So in the following example:
/// <reference types="jest" />
import path from "path";
export const myPath = path.parse(__filename);
TypeScript would still emit:
/// <reference types="node" />
import path from "path";
export declare const myPath: path.ParsedPath;
In the course of working on isolatedDeclarations
, we realized that this logic was untenable for anyone attempting to implement a declaration emitter without type checking or using more than a single file’s context.
This behavior is also hard to understand from a user’s perspective; whether or not a reference directive appeared in the emitted file seems inconsistent and difficult to predict unless you understand exactly what’s going on during typechecking.
To prevent declaration emit from being different when isolatedDeclarations
was enabled, we knew that our emit needed to change.
Through experimentation, we found that nearly all cases where TypeScript synthesized reference directives were just to pull in node
or react
.
These are cases where the expectation is that a downstream user already references those types through tsconfig.json "types"
or library imports, so no longer synthesizing these reference directives would be unlikely to break anyone.
It’s worth noting that this is already how it works for lib.d.ts
; TypeScript doesn’t synthesize a reference to lib="es2015"
when a module exports a WeakMap
, instead assuming that a downstream user will have included that as part of their environment.
For reference directives that had been written by library authors (not synthesized), further experimentation showed that nearly all were removed, never showing up in the output. Most reference directives that were preserved were broken and likely not intended to be preserved.
Given those results, we decided to greatly simplfy reference directives in declaration emit in TypeScript 5.5. A more consistent strategy will help library authors and consumers have better control of their declaration files.
Reference directives are no longer synthesized.
User-written reference directives are no longer preserved, unless annotated with a new preserve="true"
attribute.
Concretely, an input file like:
/// <reference types="some-lib" preserve="true" />
/// <reference types="jest" />
import path from "path";
export const myPath = path.parse(__filename);
will emit:
/// <reference types="some-lib" preserve="true" />
import path from "path";
export declare const myPath: path.ParsedPath;
Adding preserve="true"
is backwards compatible with older versions of TypeScript as unknown attributes are ignored.
This change also improved performance; in our benchmarks, the emit stage saw a 1-4% improvement in projects with declaration emit enabled.
What’s Next?
At this point, we anticipate very few changes to TypeScript 5.5 apart from critical bug fixes to the compiler and minor bug fixes to the language service. In the next few weeks, we’ll be releasing the first stable version of TypeScript 5.5. Keep an eye on our iteration plan for target release dates and more if you need to coordinate around that.
Otherwise, our main focus is on developing TypeScript 5.6, and we’ll have the iteration plan available in the coming days (including scheduled release dates). On top of that, we make it easy to use nightly builds of TypeScript on npm, and there is an extension to use those nightly releases in Visual Studio Code.
So give the RC or our nightlies a try, and let us know how it goes!
Happy Hacking!
– Daniel Rosenwasser and the TypeScript Team
1 comment
Congratulations on the new features! 😀
A minor thing I noticed, shouldn’t the “Since the beta” link point to the beta article instead of this one?