Type inference/narrowing lost after assignment

[sgoll]

TypeScript Version: 3.1

Search Terms: type inference, type guard, narrowing, lost, assignment

Code

let a: unknown = 'x';

if (typeof a === 'string') {
  // a inferred as `string`
  a = a.substr(0, 5);  // (method) String.substr(from: number, length?: number): string
  // a inferred as `unknown`
  a.length;            // Failure: Object is of type 'unknown'.
}

Expected behavior: This should compile without an error.

Actual behavior: Line 7 fails with: Object is of type 'unknown'.

Playground Link: https://www.typescriptlang.org/play/index.html#src=let%20a%3A%20unknown%20%3D%20'x'%3B%0D%0A%0D%0Aif%20(typeof%20a%20%3D%3D%3D%20'string')%20%7B%0D%0A%20%20%2F%2F%20a%20inferred%20as%20%60string%60%0D%0A%20%20a%20%3D%20a.substr(0%2C%205)%3B%0D%0A%20%20%2F%2F%20a%20inferred%20as%20%60unknown%60%0D%0A%20%20a.length%3B%0D%0A%7D%0D%0A

Related Issues: #18840, #19955, #26673

[weswigham]

@DanielRosenwasser We do not currently narrow non-union types on assignment. For example, given

declare class Foo {
    x: string;
}

declare class Bar extends Foo {
    y: string;
}

declare var x: Foo;

if (x instanceof Bar) {
    x.y;
    x = new Bar();
    x.y;
}

we error on the second x.y as well.

Did we want to change this?

[aleclarson]

👍 Can we expect this in 3.3.0, or no guarantees yet? Here's my use case.

[danielrentz]

Another related simple example:

function reverse1(a: ArrayLike<number>): number[] {
    return Array.from(a).reverse(); // works!
}

function reverse2(a: ArrayLike<number>): number[] {
    a = Array.from(a); // makes "a" a real Array
    return a.reverse(); // error; TS thinks "a" is still of type ArrayLike
}