Repository navigation
RFC: Generator functions #8934
Description
Summary
Add generator functions to ReScript: gen functions that compile to JavaScript function*, with yield, yield*, and async gen for async generators. They produce the Generator.t<'yield, 'return, 'next> and AsyncGenerator.t values the standard library already models (#8355).
let range = gen (start, stop) => {
for i in start to stop - 1 {
yield i
}
}
// range: (int, int) => Generator.t<int, unit, 'next>
for x of range(0, 3)->Generator.asIterable {
Console.log(x)
}Motivation
- JavaScript parity. Generators are the last major JavaScript control-flow construct ReScript cannot express. Consuming them already works (
for..of,for await..of, theGeneratorandIteratormodules); producing them requires%raw. - Libraries built on generators. Structured concurrency and effect systems use generators as their programming model: Effection, redux-saga, Effect. None of them can be used from ReScript without writing the generator functions in JavaScript.
- Lazy sequences. Iterators over trees, paginated APIs and infinite sequences are most naturally written as generators.
The groundwork is in place: the iterator and generator types (#8355), for..of and for await..of, break/continue (#8348), and try..finally (#8931), which generators need because cancellation (return() on a generator, an early exit from for..of, halting an Effection task) reaches the generator only through finally blocks.
Syntax
let numbers = gen () => {
yield 1
yield 2
}
let all = gen () => {
yield* numbers() // delegates to another generator
yield 3
}
let lines = async gen file => {
let handle = await open_(file)
try {
while !handle->eof {
yield await handle->readLine
}
} finally {
await handle->close
}
}genfunctions.genis a contextual prefix for arrow functions, likeasync:gen (x) => …,gen x => …,gen () => …, andasync gen (x) => ….genstays an ordinary identifier everywhere else, so existing code that usesgenas a name keeps working.gen (x) => …andgen x => …are syntax errors today.yield eyields a value.yield ()yieldsunit(JavaScript's bareyield).yieldbinds like JavaScript's:yield a + bisyield (a + b).yield* edelegates to another generator, as in JavaScript.yieldis a keyword only insidegenbodies, and an ordinary identifier elsewhere, including in normal functions nested in a generator, as in JavaScript. This keepsyield*unambiguous (it can't be read as a multiplication) without reserving a new keyword globally.
The arrow lookahead for async and gen must not apply in switch-guard position. Today | Some(x) if async(x) => 1 fails to parse, because async(x) => is taken as an async arrow. With gen this would break real code, since gen is a common function name. A function is never valid in a guard (guards are bool), so the parser can skip the lookahead there. That also fixes the async case (#8933).
Typing
Generator functions
A gen function's body is typed in a generator context holding a yield type 'y and a resume type 'n:
- Function: if the body has type
'r, with yield type'yand resume type'nin scope, thengen (x: 'a) => bodyhas type'a => Generator.t<'y, 'r, 'n>. - Yield: inside a generator with yield type
'yand resume type'n, ife : 'y, thenyield e : 'n. The value of ayieldexpression is what the consumer passes tonext(v)(Generator.nextValue). - Delegation: in the same context, if
g : Generator.t<'y, 'r2, 'n>, thenyield* g : 'r2. The inner generator's yield and resume types must match the outer ones; its return value becomes the value ofyield*. - Plain iterables: if
e : iterable<'y>ore : array<'y>, thenyield* e : unit. That's what built-in iterables return in JavaScript (undefined). - Outside a generator:
yieldandyield*are errors.
The context is entered by gen functions and reset by every other function boundary, the same way break/continue scoping works. It is not reset by finally blocks: yield inside finally is allowed (JavaScript allows it, and Effection relies on it for asynchronous cleanup), while break and continue out of finally stay rejected (#8931).
The return type is the type of the body's final expression, unit for a generator that just ends. When early return (#8656) is added, return e in a generator will also need e : 'r.
Async generators
async gen functions have type 'a => AsyncGenerator.t<'y, 'r, 'n>. Inside them, await is allowed, and yield e takes e : 'y, the same rule as in sync generators. To yield the result of a promise, write yield await p.
Caveat: in JavaScript, yield v in an async generator waits for v if it is a promise. A value of type promise<_> yielded directly would therefore reach the consumer already resolved, while the types say it's a promise. This is the same nested-promise problem promise<promise<_>> has today. Making it an error to yield a value whose type is a promise inside async gen would close it.
Why yield* matters
Every yield in a generator shares one resume type 'n. Generators that only produce values ('n is never used) are not affected. Generators that receive typed answers, such as sagas, would need 'n to be a union of all possible answers. yield* avoids that, because its value is the inner generator's return type, which is new for each call. That's the pattern typed-redux-saga uses in TypeScript, and the one Effection's whole API is built on:
// Effection bindings
type effect // internal; users never construct it
type operation<'a> = Generator.t<effect, 'a, unknown>
@module("effection") external sleep: int => operation<unit> = "sleep"
@module("effection") external spawn: (unit => operation<'a>) => operation<task<'a>> = "spawn"
@module("effection") external run: (unit => operation<'a>) => task<'a> = "run"
let main = run(gen () => {
let task = yield* spawn(gen () => {
yield* sleep(1000)
"done"
})
yield* Task.join(task) // string
})operation is an alias of Generator.t, although Effection's operations are iterables rather than generators. The binding can't express that difference yet (see Future extensions).
Compilation
-
Lambda.
gensets ageneratorflag infunction_attribute, next toasync.yieldandyield*become primitives, which, likeawait, always count as having side effects. -
Optimizations. Generator functions are never inlined, and their self-calls are never turned into loops. Calling a generator function creates a generator object rather than running the body.
-
JavaScript. Generator functions are emitted as
function*(async:async function*), never as arrows, since JavaScript has no generator arrows. -
yieldandawaitmust never move into a function the compiler creates. One such case exists today and has to be fixed first: partial application puts its arguments inside the new closure. Withhused more than once,let h = add(await p, ...)
compiles to
let h = (extra, extra$1) => add(await p, extra, extra$1);, anawaitinside a non-async arrow, which is a JavaScript syntax error. It also evaluates the argument on every call ofh. Arguments should be let-bound outside the closure, as the optional-argument path already does (Partial application moves its arguments into the generated closure #8932). Ayieldthere would fail the same way.
Editor support (hover, completion), gentype (the TypeScript mapping for Generator.t exists since #8355) and reanalyze need the usual updates. Reanalyze's termination analysis must accept while true { yield x }, and its exception analysis has to decide whether yield can throw (Generator.throwError throws into the generator at a yield).
Future extensions
These are not part of this proposal and can be added later without breaking anything.
Typed resume values (@yieldResult)
Libraries whose runtime answers each yielded effect with a value of a known type, like redux-saga, could mark their effect type:
@yieldResult
type effect<'res>
@module("redux-saga/effects")
external call: ('args => promise<'res>, 'args) => effect<'res> = "call"
let saga = gen () => {
let user = yield call(fetchUser, id) // user: user
}With the marker, yield e : 'a whenever e : effect<'a>, the way await types promise<'a>. The library's runtime is trusted to answer correctly, just as the JavaScript runtime is trusted to resolve a promise<'a> with an 'a. Without the marker, libraries write one helper generator per effect kind that casts the resume value, and users delegate with yield*.
Iterables with a return type
iterable<'a> carries only the yield type, so yield* over a library iterable that isn't a generator (Effection's operations, Effect's effects) can't be given its result type honestly. A three-parameter iterable type, like TypeScript 5.6's Iterable<T, TReturn, TNext>, with Generator.t, IterableIterator.t and IteratorObject.t as private abbreviations of it, would allow yield* on any such iterable.
for..of over generators without asIterable
for x of g with g: Generator.t<…> currently fails ("expected iterable<'a>"), because Generator.t is a private abbreviation of iterable<'yield>. An implicit coercion in for..of (and in yield*'s plain-iterable rule) would remove the need for ->Generator.asIterable. This is independent of generator functions and could ship first.
Open questions
- The name of the
@yieldResultattribute. - Whether yielding a promise inside
async genshould be an error. - Whether reanalyze's exception analysis should treat
yieldas possibly throwing.
Plan
- Prerequisites, each a separate PR:
- Partial application moves its arguments into the generated closure #8932: let-bind partial-application arguments outside the generated closure. This fixes
awaitin partial applications today. - Switch guard calling a function named async fails to parse #8933: skip the arrow lookahead in switch guards. This fixes
if async(x) =>today. - Optionally: implicit coercion to
iterableinfor..of.
- Partial application moves its arguments into the generated closure #8932: let-bind partial-application arguments outside the generated closure. This fixes
- Generator functions: syntax, typing, Lambda and JS output, tests, tooling.
- Async generators.
- Later, as needed:
@yieldResult, iterables with a return type.
Activity
jfrolich commentedon Oct 21, 2026
I really like this proposal: it fills the last big gap in JavaScript parity, and the typing rules hold up well in practice. I've been using an Effect-style library for ReScript (effect.res, not public yet) as a practical test case, since
Effect.gen+yield*is how Effect code is written. The shared yield type is exactly what makes error and service rows merge acrossyield*s. Two things would make a big difference though, and I think both matter for generators in general, not just for effect-style libraries.How effect.res types effects (errors and required services are polymorphic variant rows):
1. Variance annotations on
Generator.t/AsyncGenerator.tThe stdlib currently declares:
Without annotations, the parameters of this private abbreviation are invariant. In general, that's wrong for a producer type, and it's inconsistent with the rest of the stdlib:
iterable<+'a>,promise<+'a>andlist<+'a>are covariant. TypeScript models it asGenerator<out T, out TReturn, in TNext>. Yields and the return value only flow out of a generator, and resume values only flow in.Generator.tcan be coerced toiterable<'y>, but aGenerator.t<[#A], _, _>can't be coerced toGenerator.t<[#A | #B], _, _>, while the equivalentiterablecoercion works.Generator.tinherits the invariance.In effect.res terms, a shared effect used by two programs with different error sets fails:
Reduced repro, mirroring the stdlib declaration:
With
type t<+'y, +'r, -'n>inG, this compiles. So:and the same for
AsyncGenerator.t.2. Move "iterables with a return type" into v1, and base
yield*on itThe RFC's advice to alias
Generator.tfor things that aren't generators is a bad idea in general, for three reasons.It's an unsound type.
Generator.tpromisesnext,returnandthrow. Effection operations, Effect values, and any custom iterable with a result only promise[Symbol.iterator]. With the alias, this type-checks and crashes:The same goes for every binding written that way while the extension is pending. Effect TS also makes
OptionandResultyieldable, so it's not just one type per library.Choosing the rule by the operand's type breaks principality. A generator gives
'r2, an iterable or array givesunit, and the choice depends on what the type checker knows at that point. In any helper that takes the thing to delegate to as a parameter, the operand's type isn't known yet:So the inferred type depends on the order of type checking. Code can type-check in one shape and fail after an innocent refactor (moving a definition, removing an annotation). That's the same class of problem OCaml's type-directed disambiguation has, which is why it warns under
-principal."Can be added later without breaking anything" holds for the language, not the ecosystem. Every library that follows the advice and aliases
Generator.tnow has to change its public types when the three-parameter iterable lands. TypeScript went through the same transition (Iterable<T, TReturn, TNext>,BuiltinIteratorReturnin 5.6). Starting with the final shape avoids it.A single rule fixes all three:
yield* erequirese: iterable<'y, 'r, 'n>(or a dedicated abstract type) and has type'r. Arrays and plain iterables are iterables with'r = unit, andGenerator.tis a private abbreviation of it. Thenis honest about what an effect is, and
operation<'a>can do the same.Smaller notes
(~x=compute(), ()) => ...putscompute()inside the body). In a generator they would then run at the firstnext(), not when the function is called as in JavaScript. For example, ingen (~cursor=Cursor.start(), ()) => ...the start cursor would be computed lazily. Either evaluate defaults (and parameter patterns) before creating the generator object, or document the difference.yielduses JavaScript's low precedence (yield a + b=yield (a + b)), whileawait p + 1is(await p) + 1today. Worth spelling out thatyield* Db.query(id)->Effect.retry(policy)=yield* (Db.query(id)->Effect.retry(policy)), andyield await p + 1=yield ((await p) + 1).'yand'n, so a mismatch is reported far from its cause. For example,yield userinstead ofyield* Effect.succeed(user)would show up as a type error on some otheryield*. The error should point at bothyieldexpressions, like mismatchedswitchbranches.let loadConfig = Effect.gen(gen () => ...)is an application, so it only stays polymorphic when the library type is covariant. Another reason for (1), and maybe worth a sentence near the Effection example.