View on GitHub


Explanatory writing

Compiler Manual

Welcome to the compiler manual. It details the compiler implementation and its philosophy. Because it focusses on implementation, it’s necessarily out-of-date and incomplete. Also, in this version, I totally make guesses about things I’m not sure about.


The basic data structures used by the compiler (and elsewhere) are Node, Symbol and Type.


The parser produces nodes in the form of an Abstract Syntax Tree (AST). Nodes track syntax kind, location, and children. For example, a + 1 produces a BinaryExpression with three children: an Identifier(a), a PlusToken and a LiteralExpression(1).


The binder produces symbols in the form of a symbol table. A symbol tracks the declaration(s) of an identifier. The declaration is a Node. In the example below, when the binder encounters let a = 1, it creates a symbol a. Then when the checker wants to find the type of a * 2, it can request the symbol for the identifier a, go to a’s declaration and look for an initial value (1, in this case) or a type annotation.

let a = 1
a * 2


The checker produces types. Lots of these types have their own caches, but types aren’t part of some general structure. Types fall into three categories: primitive types, structured types and type parameters. (Type parameters could be called type variables.)

  1. Primitive types: number, string, etc.
  2. Structured types: object types, classes, unions, intersections, etc.
  3. Type parameters: um … type variables that are parameter to functions or other types?

Primitives don’t track much of anything. They are mainly used for identity. If you call getApparentType on numberType, you get globalNumberType, which is an object type that represents Number. Number has methods like toFixed et al.

Structured types are types that have or potentially could have properties. Classes, interfaces and type literals are pretty obvious; their properties are the ones declared in their respective declarations.

Structured types like unions, intersections, index types and spread types represent a type that will have properties … once the type parameters that they contain are instantiated. For example T & { __brand } will definitely have properties once T is provided. At that point the newer structured types collapse into a normal object type, so something like { ...{ a: string } } immediately converts to just { a: string }.

If you need the currently known properties of a structured type, you can call getApparentType. getApparentType(T & { __brand }) gives you just __brand.

Note that structured types also have call/construct signatures and numeric/string index signatures. These 4 things are stored and checked separately from the properties.

TODO: Discuss type parameters.


The compiler implementation is organised into phases: scanner, parser, binder, checker, and transformer. Each phase has more or less one file except the transformer, which puts each transformation in its own file (so, for example, the ESNext transform is in transforms/esnext.ts).

This means that big phases have big files, especially the checker. And each phase is one big namespace with lots of functions inside. In fact, the files are so big, with so many functions, that the best way to work is just to search for the function name you need and then jump from definition to usage and back again. Unfortunately this means that it’s difficult to get started in an area you don’t yet understand because you’ll have to guess the names of the functions that implement the thing you want to work on. I hope that this document will help a bit with that problem.


The scanner reads individual characters and produces tokens. It can be configured to ignore spaces and to ignore comments. The default is to ignore both.


The scanner can also produce leading and trailing comments. However, the definition is not quite right. Sometimes what appear to be leading comments are parsed as trailing comments. Unfortunately, I didn’t take notes when I tried to fix the code, but here’s what I remember. Sometimes getLeadingCommentRanges will not return what appears to be a leading comment. In this case you will have to call getTrailingCommentRanges as well. For example, getJSDocCommentRanges has a list of nodes that need to call both. It just concats the results.

The core algorithm that is broken is in iterateCommentRange. At one point I fixed it, but it churned the emitter and services code so much that I discarded the fix.


It’s a recursive descent parser. It’s pretty resilient, so if you search for functions matching the thing you want to change, you can probably get away with just adding the code to parse. There aren’t any surprises in the general implementation style here.

TODO: More here later.


The parser parses JSDoc half-heartedly. You have to explicitly call addJSDocComment after calling finishNode — it’s not part of finishNode itself because so many nodes can not have JSDoc on them.

Then once it starts parsing it uses a completely different tokeniser, nextJSDocToken(), from the normal parser. (It doesn’t ignore whitespace and returns way fewer SyntaxKinds). And it parses everything, even things that exist in Typescript, into completely separate AST nodes like JSDocArrayType for parsing T[]. There are a few exceptions to this: type literals call back into the normal parser.

JSDoc is stored in the jsDocComments property of the node that it’s associated with. Each JSDoc then contains a top-level comment and a list of tags. There are various kinds of tags, but they all have an associated string comment as well. For example:

 * simple function
 * @returns nothing
function f() { }

Attaches a JSDoc to f with comment = "simple function" and a JSDocReturnTag that has comment = "nothing".

Since JSDoc supports non-local annotations — parameter information usually occurs on a @param tag on the containing function, for example — retrieval becomes complicated. Typescript chooses to stitch the information back together when it’s needed, usually while the checker is running. More on that in the checker section.


It gathers symbols from files. There’s one created when you start the compiler, and that’s it. It creates symbols for declarations so that the checker can stick types on them.

TODO: This is woefully incomplete.

TODO: It also sets up some data structures that the checker uses, like control flow.

TODO: It also marks nodes with their language for the emitter. For example, all types are typescript syntax, exponentiation is ES2016 syntax and class is ES2015 syntax.


The checker makes sure the types are right. ‘Nuff said!

Actually, it’s 20,000 lines long, and does almost everything that’s not syntactic. Surprisingly, a checker gets created every time the language service requests information because it tries to present an immutable interface. This works because the checker is very lazy.

Here’s a table of contents:

  1. Type references and instantiation
  2. Control Flow
  3. Type inference
  4. JSDoc
  5. Other topics

Type references

The life cycle of a generic type is complicated. TODO: Fill in and correct the details here.

First, when checking a generic declaration like class A<T> { a: T }, the compiler creates an InterfaceType with a property named typeParameters. Next, when checking a declaration that uses the generic type, like let an: A<number>, the compiler creates a TypeReference, which has a property named typeParameters ([T]) and a property named typeArguments ([number]).

Up to this point, little work has been done. Nothing has caused the checker to go look for members of A<number> or their types. Just a couple of types have been created with only a few properties on each. The actual work doesn’t start until some code refers to members of an.

For example, when you write return an.a, the checker calls checkPropertyAccessExpression. This first gets the type of an, which creates the type reference A<number> that I just talked about. Next it asks for the type of property a on A<number> with getPropertyOfType. This kicks off the real work.

  1. getPropertyOfType(A<number>, 'a') calls
  2. resolveStructuredTypeMembers(A<number>) calls
  3. resolveTypeReferenceMembers calls resolveObjectTypeMembers which then calls
  4. resolveObjectTypeMembers calls
  5. createTypeMapper to create a function that takes T -> number.
  6. Then it calls createInstantiatedSymbolTable with the mapper for each symbol (and call/construct/index signatures).

which calls instantiateSymbol on each symbol in A<T>. This doesn’t actually call the type mapper, it just creates a Symbol with the Instantiated flag that has a pointer to the original symbol and the mapper. However, this means that A<number> has a members table of its own symbols now. We just haven’t figured out what the types of the members are yet.

Now checkPropertyAccessExpression can get a symbol for A<number>.a, it calls getTypeOfSymbol on it. This kicks off the second half of the work:

  1. getTypeOfSymbol calls getTypeOfInstantiatedSymbol, which first calls
  2. getTypeOfSymbol on A<T>.a, which checks the declaration of a to return T
  3. then getTypeOfInstantiatedSymbol calls
  4. instantiateType, which calls mapper(T)

Which finally returns number, since mapper got created back in step 5 to map T -> number. Finally, we know that an.a has the type number.

Notice that we don’t calculate the type of any other members of A<number> (if there were any). They would have to go through steps (7) to (10) to produce their type.

Control Flow

The binder sets up backpointers for control flow analysis using only syntactic information. Then the checker uses these backpointers to trace control flow from usage to declaration in getFlowTypeOfReference. Without control flow, the type of a variable would always be the type it was declared with. With control flow, usage can narrow the type of a variable. That means that the checker now has two concepts: declared type and flow type (or assigned type, or computed type). The computed type should always be a narrowing of the declared type (except for instanceof).

Control flow analysis works by, at each reference of a variable, starting with the declared type and walking back up the control flow chain looking at uses of the variable that might change the type. For example, the assignment sn = 12 in

let sn: string | number
sn = 12

Makes the flow type of sn: number on the last line.

Control flow analysis runs on each reference to a variable. There is some caching but I believe it’s active only within a call to getFlowTypeOfReference. So it’s safe to assume each call to getFlowTypeOfReference is independent.

Note: ‘declared type’ here is different than the type/declared type distinction that corresponds to static/instance type of classes, and really only makes sense there. getDeclaredTypeOfSymbol and friends refer to this distinction, but work on any type.

Contextual typing

Control flow can run into circularities whenever it calls back into the normal type-checking code. The normal checking code may end up requesting the control flow type, which may iterate through control flow nodes until it requests the type of a node, which may end up requesting the control flow type….

For example, when you assign to an evolving array, control flow analysis needs to know the type of the expression being inserted into the array:

let x = []
x[0] = { p: 11 }

However, to find the type of this particular expression, an object literal, the checker first asks for the contextual type. The contextual type is the element type of x. And the element type of x is … determined by control flow analysis. Oops.

Type Inference

Type inference is basically the hairiest thing the compiler does. Here are some notes from my latest brush with it; I make no claim that these are complete, just useful.

Type inference is a triple-nested loop. The outer loop checks overloads in order. The middle loop initially ignores context-sensitive parameters, then enables them one at a time. And the inner loop infers from parameters one at a time (skipping context-sensitive parameters that haven’t been enabled yet).

The outer two loops are in chooseOverload, called from inside resolveCall and the inner loop is in inferTypeArguments.

The actual inference rules are in inferFromTypes and are pretty straightforward. Basically, inference recurs through various types until it reaches a target that is a type parameter. Then it adds whatever the source is to the list of inferences. Each inference is added with a specific priority that is determined by context, such as InferencePriority.ReturnType which is used when looking at the return type of a function. The algorithm looks a bit like the assignability algorith, but is more permissive — for example, it often ignores problems that would make isTypeRelatedTo return false and continues looking for more inferences.

After the structure of the types has been explored, the algorithm looks at the collected inferences and discards everything but the highest priority inferences. For example, if any normal inferences (inference.priority === 0) are found, like from arguments to a function, then inferences from the return type of that same function (inference.priority === InferencePriority.ReturnType) are completely ignored. To get the final inferred type, getInferredType calls getCommonSupertype (unless it was asked to infer a return type, in which case it calls getUnionType instead).

Contextual Typing

Contextual typing is especially tricky because it can fix type parameters. When the compiler decides that a parameter is contextually typed and tries to determine the contextual type, it must also fix the type parameters. That’s because the contextual type needs to be the instantiated type. But getting the instantiated type means that the compiler has to know how to map the type parameter to the instantiated type — which is exactly the task of type inference!

What’s more, neither assigning a contextual type nor fixing a type parameter is reversible. So once the contextual type has forced us to fix a type parameter, we’re stuck with both decisions.

That’s why contextual types are always processed last; the other parameters already have types, so type inference is not destructive. But type inference of contextually typed things actually causes the types to be added to them permanently.

Other Notes


JSDoc is handled half-heartedly in the checker as well. That’s largely for historical reasons; the checker didn’t care about JSDoc at all until Typescript was able to handle Javascript because it didn’t contribute types or even symbols to the compilation. (That’s right: JSDoc types in Typescript are ignored. It only uses the documentation.)

That’s different now because JSDoc is the only way to specify types in Javascript. But the checker still handles it in a very ad-hoc way — everywhere you might get some information from a type annotation, you also need to remember to call getJSDocType if you are checking a Javascript file.

Alas, since JSDoc stores information non-locally, retrieving type information for a specific node is not that easy. The worst example is parameters: type information can be stored as a comment for the parameter, for the function or for the variable that the function is assigned to:

var /** @param {number} x - a */f =
  /** @param {number} x - b */ function (/** c */x) {

Typescript delays the work needed to stitch together this information until it’s needed (as usual), hoping that (1) it won’t be needed and (2) it won’t be that hard to retrieve in the common case and (3) that once cached it will be needed again.

Fortunately, getJSDocType conceals this work from you pretty well. If you’re in the checker, that’s probably all you’ll need, except perhaps related functions like getJSDocReturnTag (to get the return type of a Javascript function) or getJSDocTemplateTag (to get the type parameters of a Javascript function).

Other Topics


Optionality is represented two different ways. The original way is a flag on Symbol. The second way is as a union type that includes undefined.

If you are running with strictNullChecks: true then adding ‘?’ after a property does two things: it marks the symbol as optional and it adds | undefined to the type of the property. You have to remember to do these two things yourself whenever you synthesise a property inside the compiler.

Conversely, the compiler actually throws away | undefined if strictNullChecks: false. So you can’t use just maybeTypeOfKind(t, TypeFlags.Undefined) to check for optionality. You also have to check symbol.flags & SymbolFlags.Optional.

Literal widening

Literal widening is different from nullable widening. It doesn’t happen in the same place, and it widens a subtype (the literal type) to a supertype (string, number or boolean). Nullable widening widens undefined and null to any. It probably wouldn’t exist if null and undefined had been denotable types before 2.0.

You can widen literals using two different functions, getWidenedLiteralType and getBaseTypeOfLiteralType.

The difference is that getBaseTypeOfLiteralType unconditionally widens whereas getWidenedLiteralType only widens if the literal type is fresh. For full discussion of freshness, see the PR that implemented it.

Briefly, literal types start fresh if they arise from a literal expression. Fresh literal types widen when they are assigned to a non-cost/readonly location. Non-fresh literal types do not widen. For example,

const fresh = 1
type notFresh = 1
let f = fresh // f is mutable, so '1' widens to 'number'
let i: notFresh = 1 // even though i is mutable, it has type '1'

Own-enumerability in spread/rest

Object spread and rest are supposed to skip non-own, non-enumerable members. However, the compiler doesn’t track these two properties. So it fakes them. The most common non-own non-enumerable member is a class method, so those are removed.


The transformer recently replaced the emitter, which translated TypeScript to JavaScript. The transformer transforms TypeScript or JavaScript (various versions) to JavaScript (various versions) using various module systems. The input and output are basically both trees from the same AST type, just using different features.

The emitter is now a fairly small AST printer that can print any TypeScript or JavaScript AST.


The binder sets up information for the transformer since it makes a complete pass over the AST anyway. Primarily, it marks individual nodes with their associated dialect. In the example below, ...x is marked as ES2015, and any[] and void are marked as Typescript.

function f(...x: any[]): void {

Then each transformer runs on the AST. The Typescript transformer mostly just throws away annotations. The ES2015 transformation actually has to emit ES5 code that implements the rest parameter. The transforms progressively convert Typescript into older and older Javascript dialects. The transformer will run until the AST is valid ES3, or it will stop earlier if the target is something newer like ES2015. Then the emitter pretty-prints the AST.

TODO: This leaves out module transformers.

Branded Types

Two branded types exist within the compiler, Path and __String. These are both branded strings with different shapes and different purposes.

Path is used to represent an absolute, normalized, canonical folder path or file name. If you have a Path, it should already be all of the above. You may get a path from a path-looking string via toPath, and a Path may be safely used anywhere a string is requested.

__String is used to represent a string whose leading underscore have been escaped (if present) or a string representing an internal compiler symbol name, such as "__call". To escape a string in this way, call escapeLeadingUnderscores. To unescape such a string, call unescapeLeadingUnderscores. Strings of this kind are used to represent symbol names and identifier text, to allow for internal symbol names to cohabitate the same symbol table as normal members without being in conflict. The brand on this type is structured such that it is castable and comparable with normal strings, but a normal string cannot be used in its place and it can not be used in place of a normal string (to prevent escaped text from leaking out via diagnostic messages and the like). The union member enabling this is the member intersected with void, which makes it incompatible with anything asking for a string.


I don’t know anything about services. And besides, this readme is for the compiler.

Other files

This is an unsorted inventory. Some of these deserve an in-depth explanation, or at least a big overview of what they mean.

  1. core
  2. utilities
  3. program
  4. scanner
  5. sourcemap
  6. sys
  7. tsc
  8. types
  9. commandLineParser
  10. declarationEmitter