Incorrect inference of arrow function with intersected mapped type argument under strictFunctionTypes

[smoogly]

TypeScript Version: 3.6.0-dev.20190801

Search Terms: inference intersection type strictFunctionTypes

Code

type MappedIntersection<T extends {}> = { [P in keyof T]: string } & {};
type MappedRaw<T extends {}> = { [P in keyof T]: string };

// Case 1: Intersection + no arrow = succeed
type ToInferNoArrowIntersection<T extends {}> = {
    fn(params: MappedIntersection<T>): void;
};

type NoArrowIntersection<T extends ToInferNoArrowIntersection<any>> = void;
declare const case1: NoArrowIntersection<{
    fn: (params: MappedIntersection<{a: 1}>) => string,
}>;


// Case 2: No intersection + arrow = succeed
type ToInferArrowNoIntersection<T extends {}> = {
    fn: (params: MappedRaw<T>) => void;
};

type ArrowNoIntersection<T extends ToInferArrowNoIntersection<any>> = void;
declare const case2: ArrowNoIntersection<{
    fn: (params: MappedRaw<{a: 1}>) => void,
}>;


// Case 3: Intersection + arrow = fail
type ToInferArrowIntersection<T extends {}> = {
    fn: (params: MappedIntersection<T>) => void;
};

type ArrowIntersection<T extends ToInferArrowIntersection<any>> = void;
declare const case3: ArrowIntersection<{ // Fails here
    fn: (params: MappedIntersection<{a: 1}>) => void,
}>;

Expected behavior:
Expected case 3 to succeed.

Actual behavior:
Case 3 produces:

 Type '{ fn: (params: { a: string; }) => void; }' does not satisfy the constraint 'ToInferArrowIntersection<any>'.
  Types of property 'fn' are incompatible.
    Type '(params: { a: string; }) => void' is not assignable to type '(params: { [x: string]: string; }) => void'.
      Types of parameters 'params' and 'params' are incompatible.
        Property 'a' is missing in type '{ [x: string]: string; }' but required in type '{ a: string; }'.

Seems relevant that compilation succeeds without strictFunctionTypes

Playground Link: Playground

Related Issues:
#29123 (comment)
#31081

[jack-williams]

Case 1 correctly produces no error because fn is a method and therefore related by bivariantly.

Case 3 should produce an error because under strict function types the function inputs are related in reverse, and a string indexer is not related to the object with property a.

declare const x: { [x: string]: string; };
const b: { a: string } = x; // error

It appears to be Case 2 that is questionable, and is most likely due to variance measurements. I can correctly synthesise the error by using identical but distinct type aliases.

type MappedIntersection<T extends {}> = { [P in keyof T]: string } & {};
type MappedRaw1<T extends {}> = { [P in keyof T]: string };
type MappedRaw2<T extends {}> = { [P in keyof T]: string };

// Case 2: No intersection + arrow = succeed
type ToInferArrowNoIntersection<T extends {}> = {
    fn: (params: MappedRaw1<T>) => void;
};

type ArrowNoIntersection<T extends ToInferArrowNoIntersection<any>> = void;
declare const case2: ArrowNoIntersection<{
    fn: (params: MappedRaw2<{a: 1}>) => void, // error
}>;

So what seems to be going on is that in Case 2 the params arguments are being related by their type parameters: this succeeds because any is related to everything. In Case 3 it appears that the intersection causes variance checking to be skipped and a structural check is used instead, which correctly detects the error.

[jack-williams]

Minimal repro:

declare const a: MappedIntersection<any>;
declare const b: MappedRaw1<any>;
const zMeasure: MappedRaw1<{x: string}> = b; // no error
const zStructural: MappedRaw2<{x: string}> = b; // error
const zDifferent: MappedRaw2<{x: string}> = a; // error

[weswigham]

Does #32225 fix the issue, I wonder?

[jack-williams]

Yeah I wasn't sure if that was merged, but it seems to target the problem!

[smoogly]

I might be wrong somewhere, but aren't Case 1 and Case 3 supposed to be identical?

I was under impression that fn(): void and fn: () => void describe the same thing.

[weswigham]

Sorta. Outside strict, yes, but within strict mode the method syntax is used as a hint for looser typechecking.

[smoogly]

Does it means that by design third case will always fail?

Does that also mean that sometime in the future first case will also be failing?

I have trouble understanding why describing same underlying JS might fail or succeed based on a choice of syntax.

[jack-williams]

@weswigham Seems to be an issue with unreliable markers rather than unmeasurable, if my understanding is correct. Optionality modifiers work ok, but readonly (or none) do not.

type MappedRaw1RO<T extends {}> = { readonly [P in keyof T]: string };
type MappedRaw2RO<T extends {}> = { readonly [P in keyof T]: string };

declare const b: MappedRaw1RO<any>;
const zMeasure: MappedRaw1RO<{x: string}> = b; // ***no error***
const zStructural: MappedRaw2RO<{x: string}> = b; // ***error***

type MappedRaw1OPT<T extends {}> = { [P in keyof T]-?: string };
type MappedRaw2OPT<T extends {}> = { [P in keyof T]-?: string };

declare const b2: MappedRaw1OPT<any>;
const zMeasure2: MappedRaw1OPT<{x: string}> = b; // ***error***
const zStructural2: MappedRaw2OPT<{x: string}> = b; // ***error***

[smoogly]

@jack-williams, if Case 3 indeed should produce an error, check out the case where type MappedIntersection<T extends {}> = MappedRaw<T> & {}; in initial submission. Because this gets Case 3 to compile successfully.

[fatcerberus]

@smoogly The only fully typesafe way to relate two functions is to relate their parameters contravariantly, i.e. function A is assignable to function type B only if all of B's parameters are assignable to A's. This is because the parameters represent data that flows in the opposite direction over the =: assignment typically moves data right to left, but calling the aliased function will send the arguments left to right over that same assignment. --strictFunctionTypes enforces this for functions and lambdas, but not methods.

Methods are related bivariantly as a tradeoff, because much of the public JS API (especially the DOM API) have fundamentally unsound designs that would become very difficult to use if the "correct" variance were enforced for methods. For example, Array should be invariant to be fully typesafe, but in practice is covariant because that reflects the way idiomatic JS code is written.

[jack-williams]

@smoogly That compiles for the same reason case two compiles: the use of the alias enables variance measurements. The type MappedRaw is never expanded and the checker just ensures that the type arguments are related.

[smoogly]

Gentlemen, thank you for setting my mind straight about assignability.
My expectation was wrong about Case 3 based on habit of use in ts@2.8.4.

[weswigham]

#32225 doesn't fix this :(