introduce coroutines to the language

[zerbina]

Summary

Add first-class coroutines to the language. A coroutine is a stack-less, resumeable procedure, which is similar to .closure iterators, but with a lower-level and more flexible interface.

The main inspiration for this feature is the CPS library.

Important: this PR is a work-in-progress. It exists for the purpose of tracking progress and to have something concrete that discussion can focus on. Nothing about the specification nor implementation is final, not even the name "coroutine".

Details

Current status:

• a basic specification exists
• the compiler implements a subset of the specification

The specification evolved from an earlier draft implementation, which can be found here. For turning procedures into resumable procedures, the existing closureiters machinery is re-used.

---

To-Do

• iterate on the specification until a release candidate emerges
• implement the specification (can start concurrently with iterating on the specification)

[saem]

Is the var parameter restriction because we don't know how to handle them in the environment? Because technically we can just treat it as a ptr to the that location (out parameter), although perhaps we're inheriting this limitation from closure iterators (haven't poked at those much).

[zerbina]

Disallowing var is not so much a limitation with closure iterators, rather it's for safety. Capturing a var as a ptr is indeed possible, but then it's very easy to create a situation of dangling pointers. Consider:

proc coro(x: var int) {.coroutine.} =
  x = 1
  echo x

proc test(): Coroutine[void] =
  var y = 1
  result = launch coro(y)

# upon returning from `test`, the `y` local is gone, so reading/writing from/to
# `x` would be a use-after-(stack)-free  
resume(test())

Callsite escape analysis doesn't help, since the coroutine could save the instance in some global, for example.

[saem]

Oh, so we couldn't require var T params be treated as lent T?

[zerbina]

Due to refs shared-owner ship semantics, and Coroutine beings ref, it would not work, yeah.

If Coroutine were a unique-ownership type (e.g., something akin to C++'s std::unique_ptr), then storing the var T params as lent T could work, but the borrow checker would need to understand such unique-ownership type.

[saem]

Hmm, even if a coroutine callable was proc coro(self: sink Coroutine): Coroutine[T], sink doesn't work because it might be consumed (not strong enough) and that doesn't play well with the fact that we need to return it. I guess we'd have to have {.noalias.} become first class, as opposed to restrict in codgen.

[zerbina]

I don't think it's a good idea, but we could introduce a CoroutinePtr, which is a heap-allocated object that has value semantics. Making this type as ergonomic to use as a ref type would be quite tricky, however.

[saem]

I agree, I don't think it'd be a good idea. If anything that capability should develop separately (guided by a few more use cases) and if successful used here. Otherwise, I'm guessing there are better solutions.

[blackmius]

add tests where you passing lambda functions into coroutine, tests calling coroutine inside other coroutine (or you designed it to call launch in all cases?), and running lambdas inside coroutines that returns other coroutine

proc a(cb: proc()) {.coroutine.}
proc b() {.coroutine} =
  b() # recursively
proc c(coro: proc() {.coroutine.}) {.coroutine.}

[blackmius]

declaring self explicitly is better imo

proc coro(self: Coroutine) {.coroutine.} =
  echo typeof(self)
  echo self.status

[disruptek]

Agreed; if we're not going to have colorless functions, we should go all the way into color. Otherwise, it's just confusing to both programmers and metaprogrammers.

[saem]

I don't know if it is, proc c(a: int) {.coroutine.} describes at least two things, one is the coroutine constructor, and the other is all the fields (parameter list) that are part of the backing closure environment. The body of c assumes both self and an unpacked a.

We could have proc c(self: Coroutine, a: int) {.coroutine.}, but besides being redundant, it only provides an optional rename of the self parameter and specification of an alternative Coroutine base type. With that said:

1. self's type isn't Coroutine, that's actually the base type
2. renaming self is likely not helpful for most readers (misfeature)
3. we can do named parameter shenanigans?

If you really want to be explicit, then it'd be more like this:

proc coro(#[constructor params go here]#) {.coroutine.} =
  proc (self: Coroutine) =
    echo typeof(self)
    echo self.status
  # `coroutine` pragma makes the return type a `void proc closure/Coroutine[void]`

[zerbina]

I had considered this, but I'm not sure if it's better. The first parameter having special meaning seems inconsistent with the rest of the language, where parameter position doesn't have special meaning.

I think it's also somewhat confusing for the programmer:

proc coro(self: Coroutine) {.coroutine.} =
  discard

# why is there no argument passed to the coroutine constructor?
discard launch(coro())

Edit: this boils down to what @saem said, but I missed said comment due to not refreshing the page

[zerbina]

if we're not going to have colorless functions

I'm not sure I understand this remark. The feature as currently specified doesn't introduce "color", in that the coroutines can be launched and resumed from everywhere, without tainting the caller.

[blackmius]

also how try/except/finally and defer will work in coroutines?

[disruptek]

I'd prefer to reduce/eliminate the status use; it feels like API that doesn't actually add much other than explicit boilerplate. Having to check the status of a coroutine feels like it makes what should be a simple function call into a fiddly instrument not unlike recovering an Option or Result value.

When should I choose a coroutine over a closure iterator and why?

[disruptek]

What does it mean that an apparently immutable coroutine can change state?
Why do we need a launch operator?
Why can't we simply call instance() to resume the coroutine and recover the argument of a yield?
What purpose does returning the coroutine from resume serve?

[saem]

What purpose does returning the coroutine from resume serve?

The point of returning a coroutine is to allow one coroutine call to "forever" yield control to another coroutine in its stead. This allows two way communication between coroutines, where it's not so much caller-callee, but "symmetric".

I recommend the wikipedia article on coroutines, the "Definition and Types" section is really good for mapping the design space (stackless/full, a/symmetric, and first-class vs constrained), and then the comparison section with subroutines illustrates the peering/symmetric relationship between a producer consumer pair.

[zerbina]

What does it mean that an apparently immutable coroutine can change state?

I'm not sure I understand the question. A couroutine instance is not immutable, or at least I don't consider them to be.

Why do we need a launch operator?

To construct/create an instance of a coroutine without blocking use of the coroutine(...) standalone invocation syntax for other purposes. At present, coroutine(...) is expanded to trampoline(launch coroutine()).

Why can't we simply call instance() to resume the coroutine and recover the argument of a yield?

Having an explicit resume relies on less less-tested compiler features, but I agree that having a proc ()[T: Coroutine](c: T): T that's an alias for resume would make sense.

[disruptek]

Agreed; if we're not going to have colorless functions, we should go all the way into color. Otherwise, it's just confusing to both programmers and metaprogrammers.

[disruptek]

Why is this preferable to simply raising the exception in the code path where it's running?

[saem]

Because presently the exception is doing double duty for exceptions + cancellation. At the moment, I think cancellation needs its own thing, but that needs syntax/some operator and still jive with exceptions, as those can be raised by calls subordinate to the coroutine.

What should be possible is assuming there is an executing coroutine (coro) and it encounters an error: it should be able to signal to the caller than an error occurred, the caller should be able to recover, and then it can resume execution of coro.

[disruptek]

It would be a deep disappointment to have to trampoline coroutines implemented in the compiler.

[zerbina]

Hm, why? In the earlier revision I linked, the trampoline action was fixed and inserted by the compiler, but that took away control from the programmer while unnecessarily complicating the compiler.

I'm imagining the responsibilities to be the following:

• the compiler provides a low-level but still easy to use and build upon interface for interacting with coroutines
• some sugar and often useful functionality (such as a general trampoline) is provided by the standard library
• libraries can then build their own abstractions on-top, if wanted

In short, the compiler does the heavy lifting and exposes a low-level interface; sugar and other convenience is provided by the standard (or external) library code. In my mind, this should help with keeping the compiler itself lean(er) and make further evolution of the feature easier.

[disruptek]

In that case, why can't it accept mutable arguments?

[disruptek]

I feel this should evaporate with the explicit first argument defining the coroutine's type.

[saem]

@zerbina did you consider/think of an alternative approach to cancellation, I'm trying to sort out how we might do:

1. coroutine: signals error
2. caller: recovers and resume
3. coroutine: resumes post-recovery

I imagine that'd be pretty key for enabling something like effect handling.

It could also be that raise is a coroutine entirely giving up (irrecoverable) vs "something else" where a coroutine needs input/something to continue or it's cancelled ('graceful' cancellation). This distinction might be entirely off base.

[zerbina]

did you consider/think of an alternative approach to cancellation

I did consider an explicit cancel, but the way I imagined it seemed redundant with raising an exception, so I left it out.

So, with the current specification, one can do two things:

• raise, and not handle, an exception from within the coroutine (terminal; coroutine instance cannot be resumed afterwards)
• suspend and expect the caller to perform some action (like changing some coroutine instance state)

I think what you're saying, please correct me if I'm misunderstanding, is that coroutines should provide a built-in middle-ground, where the coroutine instance can signal that it can be resumed, but some action must be taken first; not doing so would raise a CoroutineError. How this could look like, I'm not sure yet.

---

Aside: whichever solution we pick, I think an escaping exception should abort the coroutine. My reasoning is that I think the programmer and compiler should be able reason about control-flow within a coroutine the same as they do everywhere else (minus the suspending, of course, but that should ideally be more or less opaque).

Put differently, I think that it should be possible to turn a normal procedure into a coroutine without having to change anything else, and the body continues to behave exactly the same.

If a coroutine can continue after a raise (or a normal call that raised), then that's no longer the above no longer holds, and whether raise is a terminator for structured control-flow becomes context and run-time dependent.

In short, I think raise should be "entirely giving up".

[saem]

So, with the current specification, one can do two things:

• raise, and not handle, an exception from within the coroutine (terminal; coroutine instance cannot be resumed afterwards)
• suspend and expect the caller to perform some action (like changing some coroutine instance state)

I think what you're saying, please correct me if I'm misunderstanding, is that coroutines should provide a built-in middle-ground, where the coroutine instance can signal that it can be resumed, but some action must be taken first; not doing so would raise a CoroutineError. How this could look like, I'm not sure yet.

Yup, I think a middle ground makes sense, to share my motivation for having a middle ground, I want to see if we can drop the status field, ideally more, and I believe the middle ground will be required in fulfilling that.

I'll pick up dropping the state field below the section break.

Aside: whichever solution we pick, I think an escaping exception should abort the coroutine. My reasoning is that I think the programmer and compiler should be able reason about control-flow within a coroutine the same as they do everywhere else (minus the suspending, of course, but that should ideally be more or less opaque).

I also agree that raising exceptions, and their impact on control flow, should remain as expected (hah!) and therefore predictable based on existing reasoning/intuition.

Put differently, I think that it should be possible to turn a normal procedure into a coroutine without having to change anything else, and the body continues to behave exactly the same.

This is a good property, because any background CPS transform, or the reverse, should be equivalent -- otherwise we've got bigger problems. 😬

If a coroutine can continue after a raise (or a normal call that raised), then that's no longer the above no longer holds, and whether raise is a terminator for structured control-flow becomes context and run-time dependent.

In short, I think raise should be "entirely giving up".

Yup, on the same page; raise is an 'abort', instead of a 'fail/error'.

---

Dropping the state Field

I'd like to drop the state field, for possibly the same reasons @disruptek suggested removing it, mine are:

• it reduces the amount of mutable state
• the state of a coroutine is a derivation of it's internal control flow, so the field is really a cache than the reality
• it invites the temptation of fiddling with it
• smaller coroutine instances; moar fast!

Instead, at least at a low level, a coroutine should yields those state values (statically knowable within a transformed coroutine definition body). For completion, where completion might have a non-void return value a finishedWithResult state that informs the receiver that a 'result' field both exists and it contains a value of type T for a Coroutine[T]. This path might allow us to shrink CoroutineBase of other fields as well.

Otherwise, where state needs to be queried later, that burden can be carried by any runner, out of band, or by extended/more stateful coroutine instance types.

[saem]

The wikipedia article on coroutines does a good job outlining the design space. Except for stackfull vs stackless, where the decision is sorta made for us (it's stackless), symmetric vs asymmetric and constrained vs first-class are still open for discussion.

I think we should aim for symmetric coroutines, because without that I believe we end up with generators (a limited form), which IIUC we already have via first class iterators. Major open questions here:

• call vs pass/yield: how do we call (pass control with the expectation of control coming back) vs yield/pass to (transfer control with no expectation of getting control back) a coroutine -- this is of course mostly within coroutines
• success vs cancellation, and recovery: if a computation (coroutine) fails, how do we recover and resume? how does this mesh with exceptions?

Constrained vs First-Class, I don't have a good intuition about it, I lean towards first-class where one can manipulate coroutine objects, but I also understand this could wreak havoc with analysis/safety properties, not to mention future composition and extension.

---

Quick explainer of terms, but reading the 'Definition and Types' section (it's very short) is likely much better:

• Stackfull vs Stackless: if any ol' proc can yield, then it's stack full, otherwise stackless (IIUC limitation of closure iterators means we're stackless)
• Constrained vs First-Class: whether coroutines are objects that can freely be manipulated by the programmer (caveat: flexibility now means less flexibility later)
• Asymmetric vs Symmetric: If a coroutine can choose where to yield control to, then it's symmetric, allowing for one coroutine to setup another, yield directly to it, with no expectation that it's a call (give control, and then expect it back)

[zerbina]

Thank you for the overview, @saem!

My current opinions/thoughts on the major questions you raised:

I think we should aim for symmetric coroutines

I agree.

• call vs pass/yield: how do we call (pass control with the expectation of control coming back) vs yield/pass to (transfer control with no expectation of getting control back) a coroutine -- this is of course mostly within coroutines

I do think both should be possible. That is, one should be able to yield to another coroutine (without any expectation of getting control back; at the moment), while also being able to invoke another coroutine (with the expectation of getting control back).

For what it's worth, the current specification allows for both:

proc coro() {.coroutine.} =
  let other = launch coro2()
  resume(other)
  # ^^ works just like it would outside of a coroutine, i.e., it's guaranteed the `coro`
  # gets control back
  suspend(other)
  # ^^ yield to other coroutine. No guarantees that `coro` ever continues

• success vs cancellation, and recovery: if a computation (coroutine) fails, how do we recover and resume? how does this mesh with exceptions?

I don't have a fully thought-through opinion on this, but, building on what I said earlier, I do think that a coroutine should not be able to resume after an unhandled exception escapes it.

A mock example of how I think handling an exception and continuing the coroutine afterwards could look like:

proc coro() {.coroutine.} =
  cancel CatchableError.newException("error")
  echo "here"

var instance = launch coro()
resume(instance)
# sugar implemented in the standard library could provide a more convenient
# abstraction for this kind of handling
if instance.status == csAborted:
  instance.handle(proc (e: sink Exception): bool =
    echo "handled: ", e.msg
    result = true # error was handled
  )
  # if returning 'false', then the `cancel` call within the coroutine would `raise`
  # the exception
# now the instance can resume
resume(instance)

Constrained vs First-Class

My opinion is that it they should be constrained, with coroutine objects consisting of internal, not freely modifiable state, and state that can be freely modified by the programmer (the custom coroutine types provide this).

• (IIUC limitation of closure iterators means we're stackless)

It depends on how stackful coroutines should work. For example, they could be (and without being visible to the programmer) emulated by promoting every routine that yields to a coroutine to a coroutine. However, such coroutines would only be "stackful" in terms of semantics, not in the sense that there's a "real" stack memory backing them.

As a minor aside, (experimental) stackful coroutine support did exist as a library solution at one point, but it was removed (refer to #45).

[zerbina]

Thanks for taking interest and providing feedback, @blackmius.

add tests where you passing lambda functions into coroutine

The current set of tests are meant as specification tests, which describe what works and what doesn't. If something isn't explicitly tested for or mentioned, then that means it's not special and should just work.

tests calling coroutine inside other coroutine (or you designed it to call launch in all cases?)

Either you'd launch one and then yield/resume, or you'd use the trampoline feature, yeah. The current set of tests should already covers this.

and running lambdas inside coroutines that returns other coroutine

Closures returning coroutines work just like any other closure. As for a proc () {.coroutine.} type, I'm not entirely sure what they should represent, nor whether adding them is a good idea.

Should it be the type of the internal launch procedure for a coroutine with the given signature? Should it be something else? How to use the type?

They could work like this:

proc coro() {.coroutine.} =
  discard

let c = coro
# `c` can now be used in all places where `coro` can be used
var instance = launch c()
c() # invokes `trampoline`

also how try/except/finally and defer will work in coroutines?

Early implementation limitations aside (closureiters has some issues), those constructs would work the same as everywhere else. For me, a guiding principle is that the programmer should be able to reason about the body of a coroutine in much the same way as they would do for any other routine.

[disruptek]

Is this thread about coroutines or is it about something else which is currently named coroutine?

[zerbina]

Is this thread about coroutines or is it about something else which is currently named coroutine?

I think Coroutine is an adequate name for the feature as it's currently specified. What name do you think would describe/fit the current behaviour better?

[saem]

• call vs pass/yield: how do we call (pass control with the expectation of control coming back) vs yield/pass to (transfer control with no expectation of getting control back) a coroutine -- this is of course mostly within coroutines

I do think both should be possible. That is, one should be able to yield to another coroutine (without any expectation of getting control back; at the moment), while also being able to invoke another coroutine (with the expectation of getting control back).

... snip ...

• success vs cancellation, and recovery: if a computation (coroutine) fails, how do we recover and resume? how does this mesh with exceptions?

I don't have a fully thought-through opinion on this, but, building on what I said earlier, I do think that a coroutine should not be able to resume after an unhandled exception escapes it.

The above facilities might provide the means to allow hooks that raise, because we should be able to handle the raised exception within a coroutine with a try/except all within the hook, then pass control back to the hook.

As a minor aside, (experimental) stackful coroutine support did exist as a library solution at one point, but it was removed (refer to #45).

I forgot about this, we can restore it at some point, although not as a dialect.

[saem]

Is this thread about coroutines or is it about something else which is currently named coroutine?

Fair point, they really look like coroutines, and if we get them symmetric then we're pretty much at one-shot/delimited continuations. I'm just having a really hard time getting to the essence of what we're after without establishing a definition of the properties we want it to exhibit, and along the way sort out an interface that allows the user to best exploits those properties.

[blackmius]

@zerbina i suggested such tests because there were problems with these cases in CPS and async/await implementations.

Closures returning coroutines work just like any other closure. As for a proc () {.coroutine.} type, I'm not entirely sure what they should represent, nor whether adding them is a good idea.

just like regular closures they represents either callback or function returning value, but now they can suspend and be awaited with value. and needs for composition

here are examples

# coroutine callback (also as closure)
proc httpServe(handler: proc(req: Request, res: Response)  {.coroutine.}) {.coroutine.}
httpServe(proc (req: Request, res: Response) {.coroutine.} =
  resume launch res.end("hello", 200)
)

# coroutine closure
proc socks5Req(url: string, auth: proc (sock: Socket): bool {.coroutine.}): Socks5 {.coroutine.} =
  let sock = newSocket(url)
  # initiate socks5 protocol
  if not resume(launch auth(sock)):
    # wait authenticated
    raise Socks5AuthException()
  return Socks5(sock)

# maybe some pipeline operations makings requests what needs to await values to continue

maybe it is not proc () {.coroutine.} but proc (): Coroutine[void]

[zerbina]

just like regular closures they represents either callback or function returning value, but now they can suspend and be awaited with value. and needs for composition

Okay, so if I understand you correctly, you argue that proc () {.coroutine.} should be a first-class coroutine (a Coroutine represents a coroutine instance), which makes sense, and overlaps with what I suggested above. Having them seems like a good idea, yeah, I'll write the specification tests.

One note about your example code: if not resume(launch auth(sock)) won't work, at least with the current specification, since resume always returns a (.discardable) Coroutine. Assuming that an applicable trampoline overload exists, you'd have to use if not auth(sock) (which the compiler expands to if not trampoline(launch auth(sock)).

[saem]

if proc foo() = discard is a proc def, then shouldn't proc bar() {.coroutine.} = discard be a coroutine def, although a better term might be coroutine-constructor?

[zerbina]

The current terminology I'm using (it changed a bit since writing the comment):

• proc c() {.coroutine.} = discard is a coroutine definition
• c is the name of a coroutine
• the c() in launch c() is a coroutine instance construction
• proc() {.coroutine.} is a coroutine type
• launch returns a coroutine instance
• with let v = c, v stores a coroutine value (coroutine alone probably suffices)

[blackmius]

there seems to be a lack of transitions from child to parent
like mom field in CPS at least i dont get how coroutine can return control back to its parent after several suspendings without storing parent ref in self

[blackmius]

proc a() {.coroutine.} =
  echo 2
  suspend()
  echo 3
proc b() {.coroutine.} =
  echo 1
  a()
  echo 4
b()

with current trampoline implementation this will work because there are not real suspends. i think it is more like closure iterators right now but with only single return at the end (so even less usefull)

but if we save a coroutine somewhere and will try to continue later it will lost its parent because it lost its stack on which it all was worked before

var wait: Coroutine

proc a() {.coroutine.} =
  echo 2
  wait = self
  suspend()
  echo 3
proc b() {.coroutine.} =
  echo 1
  launch a()
  echo 4
launch b()
resume wait

[zerbina]

there seems to be a lack of transitions from child to parent like mom field in CPS

Correct, at the moment, the basic instance type has no concept of coroutine instance relationship.

at least i dont get how coroutine can return control back to its parent after several suspendings without storing parent ref in self

It is currently possible to implement tail calls, which I think is what you're after, by using a custom instance type that stores the parent instance. If knowing the parent coroutine instance (if any) is something that's necessary often enough, then I think it would make sense to have a mom-like field part of Coroutine[T].

with current trampoline implementation this will work because there are not real suspends

Yep, but keep in mind that the trampoline implementation is only a showcase implementation. It's there to have a working example of how an implementation could look like.

In addition, the trampoline is looked up in the context of a(), meaning that it's possible to override/overload the default trampoline (I'm still not sure whether there should be an always-present default trampoline at all).

[saem]

at least i dont get how coroutine can return control back to its parent after several suspendings without storing parent ref in self

It is currently possible to implement tail calls, which I think is what you're after, by using a custom instance type that stores the parent instance. If knowing the parent coroutine instance (if any) is something that's necessary often enough, then I think it would make sense to have a mom-like field part of Coroutine[T].

Hmm, if the current implementation doesn't remember the parent, and the sorta unconventional control flow, at least based on coroutine "standards", would a potentially better name for this more stripped down concept be suspendable. It's a suspendable routine, as in it can suspend it's execution, but that's really the only guarantee it provides. It doesn't have the state (i.e.: mom) to honour returning control back to the parent.

I'm sorta borrowing terminology from Kotlin here. Don't mean for this to be out of the blue, been thinking about it for a while and this occurred to me just now. I might be putting too hard a requirement on how coroutines should behave, but given that subroutines are meant to be a special case of coroutine, I think the parent requirement seems reasonable (subroutines take control from and return control to the parent).

[Varriount]

Keep in mind that symmetric coroutines can be implemented via asymmetric coroutines, and vice-versa.

[blackmius]

suspend lifted differently than yield, maybe it is better to check these cases too

proc a() {.coroutine.} =
  for i in 1..5:
    echo i
    suspend(self)
  var q = 5
  while q > 0:
    suspend(self)
    echo q
    q -= 1
trampoline a()

proc b(a: int) {.coroutine.} =
  if a > 2:
    suspend(self)
    echo 1
  case a:
  of 1:
    suspend(self)
    echo 2
  else:
    suspend(self)

proc c() {.coroutine.} =
  defer:
    suspend(self)
    echo 4
  try:
    suspend(self)
    echo 1
    raise newException("a")
  except:
    echo 2
    suspend(self)
  finally:
    echo 3
    suspend(self)

[zerbina]

@blackmius: Thank you for testing the branch! I've reduced the test you provided to the problematic parts and added them to the test suite.

suspend lifted differently than yield

suspend is lowered to yield; they're not transformed differently. The problem was that not all temporaries were properly lifted into the environment.

[zerbina]

On Matrix, the suggestion was made to have suspend() mean the same as suspend(self). I think this is a good idea, so I've changed the specification for suspend and updated the implementation accordingly.

[zerbina]

Regarding whether the built-in coroutine should know about continuations (i.e., the coroutine to continue with on exit if the current one), after spending some time on thinking about it, my current opinion is that it should.

Upsides

• a baseline coroutines is more powerful, making them compose better
• less reliance on abstraction (could also be seen as a downside)
• the name "coroutine" would be more fitting (in their present shape, @saem suggested that "suspendable procedure" might be a better name for coroutines)
• no new hook (for providing the continuation) is required

Downsides

• slightly more code in the compiler
• a coroutine instance is larger by default (by one ref)
• the continuation is managed by the runtime/compiler, so there's somewhat less freedom for custom abstractions

---

What's still missing either way is a tail routine, which allows for tail-calling a coroutine. This would be similar to suspend(other), but with the important difference that:

• control doesn't get passed back to the caller of resume
• exceptions and return value unpacking happen in the coroutine

A practical demonstration:

proc other(): int {.coroutine.} =
  echo "2"
  result = 3

proc coro() {.coroutine.} =
  echo "1"
  echo tail(other())
  echo "4"

let c = resume coro()
doAssert c.status == csPending
# echoes 1 2 3 4

If continuation support is built into coroutines, this, more or less, work as is. If going the library route, the standard library needs to provide a Continuation instance type, which other would have to use.

[zerbina]

Regarding the interface, I've been thinking about whether implicit coroutine constructors should be removed. At the moment, the signature of a coroutine describes the constructor, not the coroutine itself:

proc coro(a, b: int): int {.coroutine.} =
  discard

# coro is the *constructor* for a coroutine instance of type `Coroutine[int]`, not
# the coroutine itself  

What I've been thinking about is making the constructor explicit, by re-introducing launch:

proc coro(self: Coroutine) {.coroutine.} =
  discard

resume(launch coro)

# for capturing parameters, a dedicate constructor procedure is required
proc construct(x, y: int): Coroutine[void] =
  launch(proc (self: Coroutine) {.coroutine.} =
    echo x, y
  )

resume(construct(1, 2))

The Upsides

• there's less magic going on, making the code easier to reason about for the programmer
• the self parameter becomes explicit, which would allow for removing the .coroutine pragma parameter
• the implementation in the compiler becomes simpler

In addition, removing implicit coroutine constructors might allow for the following to work, but there are some open questions regarding how it'd be implemented:

proc coro(self: Coroutine, a, b: int) {.coroutine.} =
  echo (a, b)
  suspend()
  echo (a, b)

let c = launch coro
resume c, 1, 2 # echoes "(1, 2)"
resume c, 3, 4 # echoes "(3, 4)"

The Downsides

• capturing parameters becomes less efficient, at least with how lambda-lifting currently works
• it's not possible to turn a procedure into a coroutine by just adding the .coroutine pragma
• macros cannot support being coroutines

The overlap (in terms of interface) between closure iterators and coroutines would also increase significantly.

[saem]

Regarding the interface, I've been thinking about whether implicit coroutine constructors should be removed. At the moment, the signature of a coroutine describes the constructor, not the coroutine itself

I briefly thought about this too, this is where my head went, sharing for sharing's sake, I don't consider this any sort of final interface, but merely showing how I was thinking about it if trending towards a more explicit interface.

proc cc(a: int): Coroutine[int] =
  ## `cc` is the coroutine constructor
  result.coroutine: # template to lift/assign `c` to the constructed coroutine
    proc c(self: Coroutine) =
      ## `c` is the "computation-sequence"/state-machine definition
      echo a
      self.suspend()
      self.result = a

let coro = cc(10)
resume coro  # first state of `c`
resume coro  # second state of `c`

I think the more explicit style of the declaration is better, as it's quite honest wrt what a coroutine is; but the calling/consumption side of the interface might leave something to be desired. An alternative approach for calling could be:

let coro = launch(cc(10))
resume coro
resume coro

# where `cc(10)` could then be construct + run to the end

[zerbina]

I briefly thought about this too, this is where my head went, sharing for sharing's sake, I don't consider this any sort of final interface, but merely showing how I was thinking about it if trending towards a more explicit interface.

Thank you for sharing, I appreciate it! I do quite like the idea of having a magic routine/operator that performs the lifting/assignment; it's very close to what currently happens internally, but within user-space. It's very honest, like you said, and there's a lot less hidden processing, symbol introduction, etc. going on, which I personally consider a significant plus.

If going with explicit constructors, I think an important design/interface decision is whether coroutine procedures (for the lack of better name) should be define-able everywhere, or, going with the interface from your idea, only as name-less argument to the coroutine magic.

Also, without explicit constructors, in the ideal case, I think the .coroutine pragma should not be a built-in pragma, as it would make the name available for custom, programmer-defined macro/template pragmas. Whether it's really possible to not need the built-in pragma, I'm not sure yet.

An alternative approach for calling could be:

let coro = launch(cc(10))
resume coro
resume coro

# where `cc(10)` could then be construct + run to the end

Being able to "call" a coroutine as if it were a normal procedure is certainly nice, and it's why, with the early draft, there was the "coroutine constructor invocation outside of launch calls expands to trampoline cc(...)" rule.

However, without implicit coroutine constructors, I'm not sure how it could work, since the explicit constructor would just return a Coroutine[T] value. What could work is designating coroutine constructors as such, likely via a pragma, in order to allow the compiler to special-case them, though I'm still undecided whether that's a good idea or not.

[saem]

I'll take a peek at Zig, they were focused on async/await, but they did address running the same routine in cooperative mode or simply linearly with blocking.

Maybe there are some ideas there.

[saem]

Suggested change

	## An anonymous routine be turned into a coroutine too.
	## An anonymous routine can be turned into a coroutine too.

[saem]

• the name "coroutine" would be more fitting (in their present shape, @saem suggested that "suspendable procedure" might be a better name for coroutines)

I think this is a connotative vs denotative meaning issue. As far as coroutines have an "official definition", they really aren't more than a suspendable routine (denotative). The modern expectation and understanding extends that to blur the line by including the continuation (connotative).

I'm less bullish on combining the two concepts by default. Suspendable routines being called that is a concession to the now somewhat popularized understanding of coroutine (we don't have to concede). While keeping the continuation passing separate let's us better explore structured concurrency and continuation context. Otherwise I suspect we'll end up adding another field or two (maybe that's not an issue?). Also, I think we lose an important bit of generality, it just adds suspendability to any routine.

I might be overly romanticizing the simplicity of only layering in suspension support, and it might be practically entirely useless. I consider unstructured concurrency that way, it all sounds straightforward, but it becomes untenable so fast that it's sorta pointless.

I'll admit I'm on the fence about it, with a preference of not mixing the two because that might be easier to change later.