Two years ago I wrote about TypeScript compiler APIs. That investigation, and discussions around that blog post and elsewhere pointed to some deficiencies in the exposed API. In particular, constructing a TypeScript AST directly (not from source code) and manipulating the AST, and emitting or pretty-printing TypeScript from an existing AST, were not directly supported.

I’ve been following this Github issue for some time, and was pleased to hear that TypeScript 2.2 exposes some AST node factories and the AST pretty-printer and just released 2.3 exposes a transformation API. That issue has a rather concise example of how to put these pieces together, here I will go through this in a little more depth.

There has recently been further progress in this area and I explore the newly exposed API. Examples available in this repository.

Ecosystem update

Looking back at that earlier blog post, it was evident that the TypeScript compiler and ecosystem has moved on. Again I was initially in the position of using features that haven’t made it into an official TypeScript release yet, but rather than having to build a local distribution of the current TypeScript master, I was able to just npm install typescript@next and get a nightly build, to build with or bundle with my demo code.

Various changes have been made to the compiler API, in particular some of the interfaces I discussed before have changed a little, in particular getNamedDeclarations is no longer exposed (though you can easily enough search for them with ts.forEachChild).

Despite the restricted API until recently provided by the TypeScript compiler, various projects have made extensive use of the TypeScript compiler API. For example, TSLint rules are written as TS AST visitors. The TSTypeInfo library has provided an AST wrapper for information/manipulation. With the changes in the TypeScript compiler API, this is now superseded by ts-simple-ast which aims to add a convenience layer over the compiler AST (though currently is a work in progress).

Pretty-printing

A simple thing to start with, but one which enables various other tooling possibilities, a pretty-printer has been exposed to convert a TypeScript API into TypeScript source. This should enable transformation and program generation scenarios:

import * as ts from 'typescript';
const printer: ts.Printer = ts.createPrinter();
const sourceFile: ts.SourceFile = ts.createSourceFile(
  'test.ts', 'const x  :  number = 42', ts.ScriptTarget.ES2015, true, ts.ScriptKind.TS
);
console.log(printer.printFile(sourceFile))

Prints

const x: number = 42;

As well as printFile a printNode is available to print particular parts of an AST.

Node factories

A major limitation previously was that there was no way to construct a TypeScript AST directly, it could only be parsed in from source file. Now factory functions have been exposed for creating (and updating) AST nodes. IntelliSense starting create shows the list of options (there are many more on scrolling down):

ts.create IntelliSense

Here’s an example of creating and printing a TypeScript AST for a simple expression, and printing it out:

const lit = ts.createAdd(ts.createLiteral(42), ts.createLiteral("foo"));

const sourceFile: ts.SourceFile = 
    ts.createSourceFile('test.ts', '', ts.ScriptTarget.ES2015, true, ts.ScriptKind.TS);
console.log(ts.createPrinter().printNode(ts.EmitHint.Expression, lit, sourceFile) );

(Note that for some reason a source file is required to print a node, even if not really relevant).

Or as another example, (x: number) => 42:

const lit = ts.createArrowFunction([], [], [
    ts.createParameter([], [], null, 'x', null, ts.createTypeReferenceNode('number', []))
], null, null, ts.createLiteral(42));

A little verbose, perhaps

Transforming the AST

The brief example linked earlier hints at the transformation API:

const result: ts.TransformationResult<ts.SourceFile> = ts.transform(
  sourceFile, [ transformer ]
);

const transformedSourceFile: ts.SourceFile = result.transformed[0];

But what is a transformer? The type of ts.transform is as follows:

function transform<T extends Node>(source: T | T[],
    transformers: TransformerFactory<T>[],
    compilerOptions?: CompilerOptions): TransformationResult<T>;
type TransformerFactory<T extends Node> =
    (context: TransformationContext) => Transformer<T>;
type Transformer<T extends Node> = (node: T) => T;

So at the root of it, we have a simple function from Node to Node. I previously mentioned ts.forEachChild, which can be used to build up a traversal of the AST, the sibling ts.visitEachChild can be used to build up a traversal modifying the AST.

function visitEachChild<T extends Node>(node: T,
    visitor: Visitor,
    context: TransformationContext): T;

The important thing to note here is that visitEachChild iterates over the children, but returns us the (possibly updated) node it is called on. First lets build up the stupidest transformer which simply logs the types of nodes it encounters:

const transformer = <T extends ts.Node>(context: ts.TransformationContext) =>
        (rootNode: T) => {
    function visit(node: ts.Node): ts.Node {
        console.log("Visiting " + ts.SyntaxKind[node.kind]);
        return ts.visitEachChild(node, visit, context);
    }
    return ts.visitNode(rootNode, visit);
};

If I call this with input let x = n + 42; I get the result:

Visiting SourceFile
Visiting VariableStatement
Visiting VariableDeclarationList
Visiting VariableDeclaration
Visiting Identifier
Visiting BinaryExpression
Visiting Identifier
Visiting FirstLiteralToken

Transformation example: constant folding

We’ll apply this to a simple tree transformation, eliminating arithmetic expressions on constants to constants, i.e. replacing 2 + 2 with 4.

const transformer = <T extends ts.Node>(context: ts.TransformationContext) =>
        (rootNode: T) => {
    function visit(node: ts.Node): ts.Node {
        if (node.kind === ts.SyntaxKind.BinaryExpression) {
            const binary = node as ts.BinaryExpression;
            if (binary.left.kind === ts.SyntaxKind.NumericLiteral
              && binary.right.kind === ts.SyntaxKind.NumericLiteral) {
                const left = binary.left as ts.NumericLiteral;
                const leftVal = parseFloat(left.text);
                const right = binary.right as ts.NumericLiteral;
                const rightVal = parseFloat(right.text);
                if (binary.operatorToken.kind === ts.SyntaxKind.PlusToken) {
                    return ts.createLiteral(leftVal + rightVal);
                }
            }
        }
        return ts.visitEachChild(node, visit, context);
    }
    return ts.visitNode(rootNode, visit);
};

This seems to work! When run with input var x = 2 + 2;, it outputs var x = 4;. However, when run with var x = 1 + 2 + 3, it outputs var x = 3 + 3;. Run it a second time and it would simplify again, but what we want is a bottom-up recursive traversal. The fix is simple, first we transform the node by visiting the children, then construct the results (also catering for another couple of cases):

const transformer = <T extends ts.Node>(context: ts.TransformationContext) =>
        (rootNode: T) => {
    function visit(node: ts.Node): ts.Node {
        node = ts.visitEachChild(node, visit, context);
        if (node.kind === ts.SyntaxKind.BinaryExpression) {
            const binary = node as ts.BinaryExpression;
            if (binary.left.kind === ts.SyntaxKind.NumericLiteral
              && binary.right.kind === ts.SyntaxKind.NumericLiteral) {
                const left = binary.left as ts.NumericLiteral;
                const leftVal = parseFloat(left.text);
                const right = binary.right as ts.NumericLiteral;
                const rightVal = parseFloat(right.text);
                switch (binary.operatorToken.kind) {
                    case ts.SyntaxKind.PlusToken:
                        return ts.createLiteral(leftVal + rightVal);
                    case ts.SyntaxKind.AsteriskToken:
                        return ts.createLiteral(leftVal * rightVal);
                    case ts.SyntaxKind.MinusToken:
                        return ts.createLiteral(leftVal - rightVal);
                }
            }
        }
        return node;
    }
    return ts.visitNode(rootNode, visit);
};

I wish I could say I gave the earlier example for pedagogical purposes, but I really did fall into this trap.

Putting it together

Transformations can be plugged in in a few places, the transform method used above gives a simple API for operating on nodes (which might not even typecheck), but transformations can also be plugged into the emit process. There are also some subtleties to understand around the transforms API, for example onSubstituteNode:

This hook provides just-in-time substitution during the final print phase, and is primarily used to substitute a relatively small subset of nodes to avoid an expensive full walk of the tree during every transformation

Last words

I’d very much like to be able to generate a .d.ts file, declarations generated according to some external data source. Currently it seems this is not supported via this API; the dts-dom library gives a way to generate .d.ts files.

Examples above, and an updated version of the in-memory CompilerHost which can be useful in playing around with these APIs, are available in this github repo.