Generic not inferred from contextual types

[OliverJAsh]

TypeScript Version: 2.9.2

Search Terms: generic infer contextual type

Code

const baz: (n: number) => number = (
    // param `n` is correctly inferred to be `number` :-)
    n => n
);

const identity = <T>(t: T) => t;

const bar: (n: number) => number = identity(
    // param `n` is inferred as `any`, expected `number` :-(
    x => x
);

Expected behavior: TypeScript should infer the generic T in identity from the contextual type.

Actual behavior: TypeScript infers the generic T in identity as any!

Playground Link:

Related Issues:

[Hotell]

hmm I don't think this is valid.

What you are after is following:

// inferred bar 👉 (x: number) => number
const bar = identity((x: number) => x)

by doing

const bar: (n: number) => number = identity(
    x => x
);

you're telling compiler that bar should be type of (n: number) => number, but your are assigning to it identity<any>((x:any): any => x) which is inferred to type (x:any) => any and therefore you got correct error

[OliverJAsh]

I thought TypeScript would use bars type annotation as a "contextual type" and use it to infer the generic to identity as (n: number) => number. 🤔

[Hotell]

I might be wrong but I don't think so, that would mean that Generics can flow deeply to function calls.

but again explicitly annotating a variable is expressing what type should be on the right side of assignment, not to infer

[ahejlsberg]

This is working as intended. Function and arrow expression parameters without type annotations are assigned contextual types only when the contextual type is known (before instantiation) to be a function type. In your example the contextual type is simply T, so we infer from the argument (x: any) => any to T. Had the contextual type been something like (x: T) => T, we would fix the inferences we've made so far for T (which would include inferences from the return type) and then assign that type to the x parameter.

[Hotell]

yup what @ahejlsberg and (myself partially) said :)

[OliverJAsh]

Thanks for the explanation @ahejlsberg and @Hotell.

To justify why I would like this with a real world use case:

const numberPromise = Promise.resolve(1);

numberPromise.then(
  // x is inferred as number, no type annotation required
  x => x + 1
);

// Later, when I need to expand on this function, if I use `pipe` I'm
// forced to add a type annotation to the parameter when it wasn't
// needed before.

numberPromise.then(pipe(
  // x is inferred as any, requiring me to add a type
  // annotation to the param that wasn't needed before
  x => x + 1,
  x => x * 2,
));

numberPromise.then(pipe(
  (x: number) => x + 1,
  x => x * 2,
));

[ahejlsberg]

@OliverJAsh How is pipe defined?

[OliverJAsh]

@ahejlsberg I'm using flow from Lodash, which is the same idea as pipe: https://github.com/DefinitelyTyped/DefinitelyTyped/blob/64104e9760065bd498a3d62d98bb05ea26500dc2/types/lodash/common/util.d.ts#L209

[Hotell]

Hmm interesting, those definitions might be wrong ? this works

declare const pipe: <A, R1, R2>(a: (arg: A) => R1, b: (args: R1) => R2) => (resolved: A) => R2

// $Expect Promise<number>
const numberPromise = Promise.resolve(1)

numberPromise.then(
  // $Expect (x:number)=>number
  (x) => x + 1
)

// $Expect Promise<number>
const result = numberPromise.then(
  pipe(
    // $Expect (x:number)=>number
    (x) => x + 1,
    // $Expect (x:number)=>number
    (x) => x * 2
  )
)

[ahejlsberg]

The issue appears to be that the flow wall of overloads starts with a series of declarations that have zero parameters for the first function. In the process of matching against those, we assign type any to the x parameter of the lambda (because there is no matching contextual parameter). However, once we've made that assignment it sticks and for reasons of consistency we can't undo it. It all works if you eliminate all but the one-parameter-first-function declarations of flow. Not sure there is much we can do to change this.

[OliverJAsh]

Ahh, that's good to know, so we can at least workaround it by patching the flow types to remove that overload (which we don't use).

I realised my original example was off the mark. Here is a correct example which does show the actual problem, as you described:

// https://github.com/DefinitelyTyped/DefinitelyTyped/blob/64104e9760065bd498a3d62d98bb05ea26500dc2/types/lodash/common/util.d.ts#L209
declare const pipe: {
    // 0-argument first function
    <R1, R2>(f1: () => R1, f2: (a: R1) => R2): () => R2;
    // 1-argument first function
    <A1, R1, R2>(f1: (a1: A1) => R1, f2: (a: R1) => R2): (a1: A1) => R2;
}

const bar: (n: number) => number = pipe(
  // param `n` is inferred as `any`, expected `number` :-(
  x => x + 1,
  x => x * 2
);

If you remove the first overload, the param n is inferred as expected.

[OliverJAsh]

I guess this can be labelled as "won't fix"?