creating an observable from C callback functions

[therealkenc]

I cannot figure out the correct way to create an observable from a C style emitter with callback functions. Example shape:

typedef void (*StringCallback)(const char* s);
typedef void (*DoneCallback)();
// we don't know how this function is implemented, only that it calls onstring()
// a number of times and ondone() once, almost certainly from another thread.
void do_async_thing(StringCallback onstring, DoneCallback ondone);

I have a solution of sorts, but it is almost certainly "wrong" as it involves creating a dynamic object and cleaning up. Rather than lead with my incorrect approach, better to just ask flatly how to turn the above into an observable to which I can subscribe. The key point in the setup above is that all the "state" is maintained inside the black box which is do_async_thing(). It takes care of whatever state it needs to emit C strings, I don't. I am sure the solution is "obvious" once I see it, but I've tried at length and appear to be missing something fundamental. TIA.

[AlexInLog]

Hi!

Yeah, pretty interesting task. Due to c-style function pointer is stateless (i mean, we can't pass lambda with some capture) and it doesn't provide extra user_object (for example, void(const char* s, void* user_object)) to be used to have some state inside, there is only one possible thing - use global variables/functions. Like this:

https://rpp.godbolt.org/z/5azdnh67h

There are 2 options, not sure which one is more appropriate for your needs. First one looks much more straightforward, second one is more "reactive" (call to do_async_thing happens only after first ever subscription)

If it is not suitable solution, feel free to provide some additional details/requirements =)

[therealkenc]

Thank-you so much for taking a look. I can give you a void* context on the api -- I was trying to make the example simpler by omitting but I think doing that obscured rather than simplified. Using a global publish_subject is kind of like a new() with only one slot.

Here is what I have:

https://rpp.godbolt.org/z/9MxsYhGhK

Crux:

auto create_observable() {
    auto onstring = [](const void* ctx, const char* s) {
        subject_from_ctx(ctx)->get_observer().on_next(s);
    };
    auto ondone = [](const void* ctx) {
        auto subject = subject_from_ctx(ctx);
        subject->get_observer().on_completed();
        delete subject;  // <--- this I don't love
    };
    auto subject = new rpp::subjects::publish_subject<const char*>();
    do_async_thing(subject, onstring, ondone);
    return subject->get_observable();
}

It is so "awesome" that it causes clang to crash. GCC will eat it.

The crash is bad enough, but from a pattern standpoint, creating a new subject just so I can delete it feels like I am doing it wong.

[AlexInLog]

In this case you can do it even much easier
https://rpp.godbolt.org/z/jcjz4K5xP

In short: we are doing raw detached copy of original observer instead of providing extra indirection via subject

But there is difference in behavior:

• you are starting do_async_thing for each new observer instead of "sharing" same do_async_thing for observers using same observable.
• To achieve previous behavior you can add | rpp::operators::publish() | rpp::operators::ref_count() - actually it works in same way, but more elegant - it makes subscription to original observable during first ever subscription and then share this observable for any future subscriptions.
• if you still need to start do_async_thing even before first ever subscription ,then you can call .connect() on result of | rpp::operators::publish (it returns connectable observable)

So, i mean this one:

const auto connectable = observable | rpp::ops::publish();
connectable.connect(); // there actually happens subscription of "nothing" to your observable and it starts do_async_thing

// then subscribe as regular
connectable.subscribe(....);

or

const auto resulting_observable = observable | rpp::ops::publish() | rpp::ops::ref_count();

resulting_observable.subscribe(....); // this one initiates do_async_thing
resulting_observable.subscribe(....); // this one just connects to previous do_async_thing

BTW: looks like clang is crashing due to using latches in unexpected way

[therealkenc]

Thank you! I can work with this.

BTW: looks like clang is crashing due to using latches in unexpected way

Interesting, I did not expect that was the problem. Is capturing not allowed (or not well formed)? Using your setup:

    std::string huh = "huh";
    create_observable()
    | rpp::ops::as_blocking()
    | rpp::ops::subscribe(
        [](const std::string& s) { 
            std::cout << "got: " << s << std::endl; 
        },
        [=]() {
            std::cout << huh << " completed" << std::endl;
        });
    std::cout << "and we're done" << std::endl;

This takes down clang too, though I can't say I understand why.

[AlexInLog]

Yeah, looks interesting.. have to check it later

[AlexInLog]

Made attemp to fix.

[AlexInLog]

Fixed on godbolt
https://rpp.godbolt.org/z/5jrGcq4os

(compilation results of exactly same source are cached, so, have to modify it in anyway - atleast add space at any place)