Inferred type parameter does not satisfy constraint

Bug Report

🔎 Search Terms

generic constraint, inference, invalid, specified, manually, explicitly, invalid

🕗 Version & Regression Information

• This changed between versions 3.3 and 3.5

⏯ Playground Link

Playground link with relevant code

💻 Code

declare function f<T, U extends T>(cb: (u: U) => T): U

const v = f(u => "x");
// const v: unknown <-- wait, what?
// function f<string, unknown>(cb: (u: unknown) => string): unknown

f<string, unknown>(u => "x"); // error, of course
// -----> ~~~~~~~
// Type 'unknown' does not satisfy the constraint 'string'.

🙁 Actual behavior

In TS3.5+, the type parameter T is inferred to be string, and U is inferred to be unknown in the first call to f(). But then U extends T is not met; you couldn't even manually specify T and U like this when calling f(), as shown in the second call to f().

(In TS3.3, T and U are both inferred as {}.)

🙂 Expected behavior

Either I'd expect both T and U to be string, or I'd expect an error where the compiler says it cannot find a suitable candidate for U. Presumably the compiler has not inferred T before it needs to infer U, and so it constrains U to T's constraint which is implicitly unknown, and then falls back to that. But it never circles back to make sure that such a fallback actually works for U extends T.

I came upon this while looking for an answer to this SO question. This behavior is weird, right? I haven't found an existing issue for it.

Reproduziere das verlinkte TypeScript Playground-Beispiel und vergleiche das gemeldete Verhalten mit TypeScript 3.3 und 3.5+. Beginne damit, die Inferenz generischer Typen und die Constraint-Prüfung für die gezeigten Funktionsaufrufe nachzuverfolgen. Als abgeschlossen gilt die Aufgabe, wenn der Compiler die inferierte U extends T-Constraint konsistent behandelt und eine Regressionstestabdeckung für das ausgewählte erwartete Verhalten vorhanden ist.