Declaration emit reveals paths within dependency that were not referred to in the source file

[mheiber]

Context:

• Generating declaration files

Declaration emit reveals paths within dependency that were not referred to in the source file:

export declare const instance: import("dependency/internal-types").TheInterface;

This is a problem because changing paths within a dependency can break a dependent package.

TypeScript Version: 3.9.0-dev.20200212

Search Terms:

• inline import declaration
• declaration external relative path

Code

node_modules/@types/dependency/entrypoint.d.ts

import { TheInterface } from "./internal-types";
export { TheInterface };
export declare function getInstance(): TheInterface;

node_modules/@types/dependency/internal-types.d.ts

export interface TheInterface {}

node_modules/@types/dependency/index.d.ts

// empty - doesn't matter for this example

example.ts

import { getInstance } from "dependency/entrypoint";
export const instance = getInstance();

src/app/tsconfig.json

{
 "compilerOptions": {
 "declaration": true,
 }
}

Expected behavior:

example.d.ts

export declare const instance: import("dependency/entrypoint").TheInterface;

Actual behavior:

example.d.ts

export declare const instance: import("dependency/internal-types").TheInterface;

It seems that part of the solution could involve having the compiler avoid using relative paths in types (import("<relativepath>").<typename>) if the relativepath is outside the project.

Repro Repo: https://github.com/mheiber/repro-rel-import-inlining

Related Issues:

• Declaration emit should not inline type definitions - Declaration emit should not inline type definitions #37151
• Proposal: Bundling TS module type definitions - Proposal: Bundling TS module type definitions #4433

Issue written with the help of: Robin Ricard (@rricard), Rob Palmer (@robpalme) and Kubilay Kahveci (@mkubilayk)

[rricard]

When writing up that issue we also found out a pretty interesting effect.

If you were to move the internal-types.d.ts into a sub-directory in the dependency and have entrypoint load it:

node_modules/@types/dependency/entrypoint.d.ts

import { TheInterface } from "./int/internal-types";
export { TheInterface };
export declare function getInstance(): TheInterface;

We do get the wanted declaration output:

example.d.ts

export declare const instance: import("dependency/entrypoint").TheInterface;

This is great because that means that TypeScript has a notion of some files are meant to be imported and others do not.

At the moment it seems like everything in the root of the dependency is considered importable externally while subdirectories are not. Actually the rule seems to be, whoever is the least deep gets it:

checker.ts:5014

 function sortByBestName(a: number, b: number) {
 const specifierA = parentSpecifiers[a];
 const specifierB = parentSpecifiers[b];
 if (specifierA && specifierB) {
 const isBRelative = pathIsRelative(specifierB);
 if (pathIsRelative(specifierA) === isBRelative) {
 // Both relative or both non-relative, sort by number of parts
 return moduleSpecifiers.countPathComponents(specifierA) - moduleSpecifiers.countPathComponents(specifierB);
 }
 if (isBRelative) {
 // A is non-relative, B is relative: prefer A
 return -1;
 }
 // A is relative, B is non-relative: prefer B
 return 1;
 }
 return 0;
 }

#27340 / 7a71887 by Wesley Wigham (@weswigham)

Is that behavior intentional, in which case we can use it as a workaround, or is this arbitrary behavior that can change?

[rricard]

Additionally the example here leverages node_modules but we do not use it in our actual system. Instead we use compilerOptions.paths, the same issue and the same rootDir/subdir effect is observable with paths. If you are interested in an example with paths I can also provide one.

[weswigham]

Preferring he shortest (in path segments) absolute path we find is intended behavior, yes.

[mheiber]

Wesley Wigham (@weswigham) thanks for getting back to us. Is the "prefer the shortest (in path segments) path" behavior that is likely to not change in future releases?

If so, we can use it to hack around the problem we're seeing:

• we can move all non-entrypoints of our package into a subdirectory

Ideally, there would be a non-hacky userspace solution, but the subdirectorification trick could work in the short term.

[weswigham]

It is the only heuristic we currently use to sort all found import paths. The only other heuristic I could see us maybe using in the future is combining it with a "longest alias chain to source" (minus cycles) to break ties, but calculating that would be crazy expensive compared to just this (which is essentially a string comparison and loosely tracks the same thing in conventional project structures). Since we have no interest in making the already sluggish declaration emit slower, we're unlikely to change it anytime soon, I think. And even then, we'd still be using segment count as the initial sort and filter, so if you were relying on it, you'd be fine.

[robpalme]

Wesley Wigham (@weswigham) One additional heuristic that might help solve this problem in future would be to favour explicit package entrypoints as designated in pkg.json "exports".

That would ensure generated declaration files use the same encapsulation boundaries defined by the source code's modules & packages.

[weswigham]

When we add support for the new esm and cjs resolvers in node, yeah, probably.