Type inference in conditional types

[ahejlsberg]

This PR introduces the ability to use type inference in conditional types (#21316), enabling pattern matching for types. For example, the following extracts the return type of a function type:

typeReturnType<T>=Textends(...args: any[])=> infer R ? R : T;

Within the extends clause of a conditional type, it is now possible to have infer declarations that introduce a type variable to be inferred. Such inferred type variables may be referenced in the true branch of the conditional type. It is possible to have multiple infer locations for the same type variable.

A conditional type T extends U ? X : Y is either resolved to X or Y, or deferred because the condition depends on one or more type variables. Whether to resolve or defer is determined as follows:

• First, given types T' and U' that are instantiations of T and U where all occurrences of type parameters are replaced with any, if T' is not assignable to U', the conditional type is resolved to Y. Intuitively, if the most permissive instantiation of T is not assignable to the most permissive instantiation of U, we know that no instantiation will be and we can just resolve to Y.
• Next, for each type variable introduced by an infer declaration within U collect a set of candidate types by inferring from T to U (using the same inference algorithm as type inference for generic functions). For a given infer type variable V, if any candidates were inferred from co-variant positions, the type inferred for V is a union of those candidates. Otherwise, if any candidates were inferred from contra-variant positions, the type inferred for V is an intersection of those candidates. Otherwise, the type inferred for V is never.
• Then, given a type T'' that is an instantiation of T where all infer type variables are replaced with the types inferred in the previous step, if T'' is definitely assignable to U, the conditional type is resolved to X. The definitely assignable relation is the same as the regular assignable relation, except that type variable constraints are not considered. Intuitively, when a type is definitely assignable to another type, we know that it will be assignable for all instantiations of those types.
• Otherwise, the condition depends on one or more type variables and the conditional type is deferred.

Conditional types can be nested to form a sequence of pattern matches that are evaluated in order:

typeUnpacked<T>=Textends(infer U)[] ? U : Textends(...args: any[])=> infer U ? U : TextendsPromise<infer U> ? U : T;typeT0=Unpacked<string>;// stringtypeT1=Unpacked<string[]>;// stringtypeT2=Unpacked<()=>string>;// stringtypeT3=Unpacked<Promise<string>>;// stringtypeT4=Unpacked<Promise<string>[]>;// Promise<string>typeT5=Unpacked<Unpacked<Promise<string>[]>>;// string

Note that is not possible for a conditional type to recursively reference itself, as might be desired in the Unpacked<T> case above. We're still considering ways in which to implement this.

The following example demonstrates how multiple candidates for the same type variable in co-variant positions causes a union type to be inferred:

typeFoo<T>=Textends{a: infer U,b: infer U} ? U : never;typeT10=Foo<{a: string,b: string}>;// stringtypeT11=Foo<{a: string,b: number}>;// string | number

Likewise, multiple candidates for the same type variable in contra-variant positions causes an intersection type to be inferred:

typeBar<T>=Textends{a: (x: infer U)=>void,b: (x: infer U)=>void} ? U : never;typeT20=Bar<{a: (x: string)=>void,b: (x: string)=>void}>;// stringtypeT21=Bar<{a: (x: string)=>void,b: (x: number)=>void}>;// string & number

When inferring from a type with multiple call signatures (such as the type of an overloaded function), inferences are made from the last signature (which, presumably, is the most permissive catch-all case). It is not possible to perform overload resolution based on a list of argument types (this would require us to support typeof for arbitrary expressions, as suggested in #6606, or something similar).

declarefunctionfoo(x: string): number;declarefunctionfoo(x: number): string;declarefunctionfoo(x: string|number): string|number;typeT30=ReturnType<typeoffoo>;// string | number

It is not possible to use infer declarations in constraint clauses for regular type parameters:

typeReturnType<Textends(...args: any[])=> infer R>=R;// Error, not supported

However, much the same effect can be obtained by erasing the type variables in the constraint and instead specifying a conditional type:

typeAnyFunction=(...args: any[])=>any;typeReturnType<TextendsAnyFunction>=Textends(...args: any[])=> infer R ? R : never;

[MadaraUchiha]

@simonbuchan You don't need infer for that,

declarefunctionzip<A,B>(a: ReadonlyArray<A>,b: ReadonlyArray<B>): Array<[A,B]>;

You really only need infer when A is part of another type you don't particularly know about, for example, you have T which could be a U[], and you need U.

[simonbuchan]

The issue is more about how you would describe infer so that users don't expect to be able to do that. (Or maybe decide that alternate syntax is fine and implement it, I dunno.)

Basically: infer T means something like "introduce a new type parameter that is inferred to the contextual type apon instantiation", but that definition is too wide.

Here's another case, where the user is trying to emulate the flow * type:

// Must be a promise, but I can't be bothered writing what kindletfoo: Promise<infer T>=asyncMethod();

[benstevens48]

I actually suggested this more general use of infer (with the syntax =T instead of infer T) some time ago in #13949. It will be interesting to see if this is implemented at some point. The current implementation is already very useful though, thanks!

[masaeedu]

[KiaraGrouwstra]

@jcalz: I can't tell what JS function you're typing there, but now that infer is in, I tried to run your type. It doesn't seem to be working yet though. I'm getting this:

typeEverything={foo: (x: {a: string;})=>void;bar: (x: {b: number;})=>void;baz: (x: {c: boolean;})=>void;}

.. so apparently these Everything instances don't directly have a / b / c.

[jcalz]

@tycho01

Hmm, I just checked 2.8.0-dev.20180211 and it gives:

typeEverything={a: string;}&{b: number;}&{c: boolean;}

which is indeed the intersection type I was aiming for. Not sure why you see something else.

[MeirionHughes]

a welcome addition; but don't you think ReturnType<T> is confusing? I would expect it to return the actual typescript Type of some function T. Perhaps ReturnPrimitive<T> would allow for a future ReturnType that is a type evaluator can be used for currying? I guess in the future this current ReturnType implementation could be extended though?

[masaeedu]

@MeirionHughesReturnType does "return" the return type of a TypeScript function type. Perhaps you're talking about a runtime representation of a type, which at least currently is against the stated goals of the project (see #1573).

[MeirionHughes]

@masaeedu Yeah, my apologies - I was confusing this one with 21316.

[ericanderson]

Any updates on recursively handling this?

[KiaraGrouwstra]

@ericanderson#21613 (comment)

[MeirionHughes]

I'm trying to do a promisify and it seems to die if you give it an overloaded method.

The single case works perfectly fine:

typePromisify<T>=Textends(cb: (v: infer V)=>void)=>void ? ()=>Promise<V> : Textends(a: infer A,cb: (v: infer V)=>void)=>void ? (a: A)=>Promise<V> : Textends(a: infer A,b: infer B,cb: (v: infer V)=>void)=>void ? (a: A,b: B)=>Promise<V> : T;typeFoo=((a: string,cb: (v)=>void)=>void)letbar: Promisify<Foo>;

overload it and it dies completely:

typeFoo=((a: string,cb: (v)=>void)=>void)&((a: string,b: number,cb: (v)=>void)=>void);

bug? already reported?

never mind - its mentioned it isn't supported - still, it shows nothing (first post says it should be picking the last signature). So maybe it is a bug?

[MeirionHughes]

@ahejlsberg this limitation on the function overloads is super annoying (not to sound ungrateful: this pr is awesome) - I can't seem to find a way to "mask" the input function (pick from the overloads) either.

You mention you choose the last function from the available ones. Would it not be possible to try each until you arrive at a non-never result?

[donaldpipowitch]

As ReturnType was added as a built-in helper are there plans to add Unpacked as well?