Asynchronous Code

Emscripten supports two ways (Asyncify and JSPI) that let synchronous C or C++ code interact with asynchronous JavaScript. This allows things like:

  • A synchronous call in C that yields to the event loop, which allows browser events to be handled.

  • A synchronous call in C that waits for an asynchronous operation in JS to complete.

In general the two options are very similar, but rely on different underlying mechanisms to work.

  • Asyncify - Asyncify automatically transforms your compiled code into a form that can be paused and resumed, and handles pausing and resuming for you, so that it is asynchronous (hence the name “Asyncify”) even though you wrote it in a normal synchronous way. This works in most environments, but can cause the Wasm output to be much larger.

  • JSPI (experimental) - Uses the VM’s support for JavaScript Promise Integration (JSPI) for interacting with async JavaScript. The code size will remain the same, but support for this feature is still experimental.

For more on Asyncify see the Asyncify introduction blogpost for general background and details of how it works internally (you can also view this talk about Asyncify). The following expands on the Emscripten examples from that post.

Sleeping / yielding to the event loop

Let’s begin with the example from that blogpost:

// example.cpp
#include <emscripten.h>
#include <stdio.h>

// start_timer(): call JS to set an async timer for 500ms
EM_JS(void, start_timer, (), {
  Module.timer = false;
  setTimeout(function() {
    Module.timer = true;
  }, 500);
});

// check_timer(): check if that timer occurred
EM_JS(bool, check_timer, (), {
  return Module.timer;
});

int main() {
  start_timer();
  // Continuously loop while synchronously polling for the timer.
  while (1) {
    if (check_timer()) {
      printf("timer happened!\n");
      return 0;
    }
    printf("sleeping...\n");
    emscripten_sleep(100);
  }
}

You can compile that using either -sASYNCIFY or -sJSPI

emcc -O3 example.cpp -s<ASYNCIFY or JSPI>

Note

It’s very important to optimize (-O3 here) when using Asyncify, as unoptimized builds are very large.

And you can run it with

nodejs a.out.js

Or with JSPI

nodejs --experimental-wasm-stack-switching a.out.js

You should then see something like this:

sleeping...
sleeping...
sleeping...
sleeping...
sleeping...
timer happened!

The code is written with a straightforward loop, which does not exit while it is running, which normally would not allow async events to be handled by the browser. With Asyncify/JSPI, those sleeps actually yield to the browser’s main event loop, and the timer can happen!

Making async Web APIs behave as if they were synchronous

Aside from emscripten_sleep and the other standard sync APIs Asyncify supports, you can also add your own functions. To do so, you must create a JS function that is called from Wasm (since Emscripten controls pausing and resuming the Wasm from the JS runtime).

One way to do that is with a JS library function. Another is to use EM_ASYNC_JS, which we’ll use in this next example:

// example.c
#include <emscripten.h>
#include <stdio.h>

EM_ASYNC_JS(int, do_fetch, (), {
  out("waiting for a fetch");
  const response = await fetch("a.html");
  out("got the fetch response");
  // (normally you would do something with the fetch here)
  return 42;
});

int main() {
  puts("before");
  do_fetch();
  puts("after");
}

In this example the async operation is a fetch, which means we need to wait for a Promise. While that operation is async, note how the C code in main() is completely synchronous!

To run this example, first compile it with

emcc example.c -O3 -o a.html -s<ASYNCIFY or JSPI>

To run this, you must run a local webserver and then browse to http://localhost:8000/a.html. You will see something like this:

before
waiting for a fetch
got the fetch response
after

That shows that the C code only continued to execute after the async JS completed.

Ways to use Asyncify APIs in older engines

If your target JS engine doesn’t support the modern async/await JS syntax, you can desugar the above implementation of do_fetch to use Promises directly with EM_JS and Asyncify.handleAsync instead:

EM_JS(int, do_fetch, (), {
  return Asyncify.handleAsync(function () {
    out("waiting for a fetch");
    return fetch("a.html").then(function (response) {
      out("got the fetch response");
      // (normally you would do something with the fetch here)
      return 42;
    });
  });
});

When using this form, the compiler doesn’t statically know that do_fetch is asynchronous anymore. Instead, you must tell the compiler that do_fetch() can do an asynchronous operation using ASYNCIFY_IMPORTS, otherwise it won’t instrument the code to allow pausing and resuming (see more details later down):

emcc example.c -O3 -o a.html -sASYNCIFY -sASYNCIFY_IMPORTS=do_fetch

Finally, if you can’t use Promises either, you can desugar the example to use Asyncify.handleSleep, which will pass a wakeUp callback to your function implementation. When this wakeUp callback is invoked, the C/C++ code will resume:

EM_JS(int, do_fetch, (), {
  return Asyncify.handleSleep((wakeUp) => {
    out("waiting for a fetch");
    fetch("a.html").then(function (response) {
      out("got the fetch response");
      // (normally you would do something with the fetch here)
      wakeUp(42);
    });
  });
});

Note that when using this form, you can’t return a value from the function itself. Instead, you need to pass it as an argument to the wakeUp callback and propagate it by returning the result of Asyncify.handleSleep in do_fetch itself.

More on ASYNCIFY_IMPORTS

As in the above example, you can add JS functions that do an async operation but look synchronous from the perspective of C. If you don’t use EM_ASYNC_JS, it’s vital to add such methods to ASYNCIFY_IMPORTS. That list of imports is the list of imports to the Wasm module that the Asyncify instrumentation must be aware of. Giving it that list tells it that all other JS calls will not do an async operation, which lets it not add overhead where it isn’t needed.

Note

If the import is not inside env the full path must be specified, for example, ASYNCIFY_IMPORTS=wasi_snapshot_preview1.fd_write

Asyncify with Dynamic Linking

If you want to use Asyncify in dynamic libraries, those methods which are imported from other linked modules (and that will be on the stack in an async operation) should be listed in ASYNCIFY_IMPORTS.

// sleep.cpp
#include <emscripten.h>

extern "C" void sleep_for_seconds() {
  emscripten_sleep(100);
}

In the side module, you can compile sleep.cpp in the ordinal emscripten dynamic linking manner:

emcc sleep.cpp -O3 -o libsleep.wasm -sASYNCIFY -sSIDE_MODULE
// main.cpp
#include <emscripten.h>

extern "C" void sleep_for_seconds();

int main() {
  sleep_for_seconds();
  return 0;
}

In the main module, the compiler doesn’t statically know that sleep_for_seconds is asynchronous. Therefore, you must add sleep_for_seconds to the ASYNCIFY_IMPORTS list.

emcc main.cpp libsleep.wasm -O3 -sASYNCIFY -sASYNCIFY_IMPORTS=sleep_for_seconds -sMAIN_MODULE

Usage with Embind

If you’re using Embind for interaction with JavaScript and want to await a dynamically retrieved Promise, you can call an await() method directly on the val instance:

val my_object = /* ... */;
val result = my_object.call<val>("someAsyncMethod").await();

In this case you don’t need to worry about ASYNCIFY_IMPORTS or JSPI_IMPORTS, since it’s an internal implementation detail of val::await and Emscripten takes care of it automatically.

Note that when using Embind exports, Asyncify and JSPI behave differently. When Asyncify is used with Embind and the code is invoked from JavaScript, then the function will return a Promise if the export calls any suspending functions, otherwise the result will be returned synchronously. However, with JSPI, the parameter emscripten::async() must be used to mark the function as asynchronous and the export will always return a Promise regardless if the export suspended.

#include <emscripten/bind.h>
#include <emscripten.h>

static int delayAndReturn(bool sleep) {
  if (sleep) {
    emscripten_sleep(0);
  }
  return 42;
}

EMSCRIPTEN_BINDINGS(example) {
  // Asyncify
  emscripten::function("delayAndReturn", &delayAndReturn);
  // JSPI
  emscripten::function("delayAndReturn", &delayAndReturn, emscripten::async());
}

Build with

emcc -O3 example.cpp -lembind -s<ASYNCIFY or JSPI>

Then invoke from JavaScript (using Asyncify)

let syncResult = Module.delayAndReturn(false);
console.log(syncResult); // 42
console.log(await syncResult); // also 42 because `await` is no-op

let asyncResult = Module.delayAndReturn(true);
console.log(asyncResult); // Promise { <pending> }
console.log(await asyncResult); // 42

In contrast to JavaScript async functions which always return a Promise, the return value is determined at run time, and a Promise is only returned if Asyncify calls are encountered (such as emscripten_sleep(), val::await(), etc).

If the code path is undetermined, the caller may either check if the returned value is an instanceof Promise or simply await on the returned value.

When using JSPI the return values will always be a Promise as seen below

let syncResult = Module.delayAndReturn(false);
console.log(syncResult); // Promise { <pending> }
console.log(await syncResult); // 42

let asyncResult = Module.delayAndReturn(true);
console.log(asyncResult); // Promise { <pending> }
console.log(await asyncResult); // 42

Usage with ccall

To make use of an Asyncify-using Wasm export from Javascript, you can use the Module.ccall function and pass async: true to its call options object. ccall will then return a Promise, which will resolve with the result of the function once the computation completes.

In this example, a function “func” is called which returns a Number.

Module.ccall("func", "number", [], [], {async: true}).then(result => {
  console.log("js_func: " + result);
});

Differences Between Asyncify and JSPI

Besides using different underlying mechanisms, Asyncify and JSPI also handle async imports and exports differently. Asyncify will automatically determine what exports will become async based on what could potentially call an an async import (ASYNCIFY_IMPORTS). However, with JSPI, the async imports and exports must be explicitly set using JSPI_IMPORTS and JSPI_EXPORTS settings.

Note

<JSPI/ASYNCIFY>_IMPORTS and JSPI_EXPORTS aren’t needed when using various helpers mentioned above such as: EM_ASYNC_JS, Embind’s Async support, ccall, etc…

Optimizing Asyncify

Note

This section does not apply to JSPI.

As mentioned earlier, unoptimized builds with Asyncify can be large and slow. Build with optimizations (say, -O3) to get good results.

Asyncify adds overhead, both code size and slowness, because it instruments code to allow unwinding and rewinding. That overhead is usually not extreme, something like 50% or so. Asyncify achieves that by doing a whole-program analysis to find functions need to be instrumented and which do not - basically, which can call something that reaches one of ASYNCIFY_IMPORTS. That analysis avoids a lot of unnecessary overhead, however, it is limited by indirect calls, since it can’t tell where they go - it could be anything in the function table (with the same type).

If you know that indirect calls are never on the stack when unwinding, then you can tell Asyncify to ignore indirect calls using ASYNCIFY_IGNORE_INDIRECT.

If you know that some indirect calls matter and others do not, then you can provide a manual list of functions to Asyncify:

  • ASYNCIFY_REMOVE is a list of functions that do not unwind the stack. As Asyncify processes the call tree, functions in this list will be removed, and neither they nor their callers will be instrumented (unless their callers need to be instrumented for other reasons.)

  • ASYNCIFY_ADD is a list of functions that do unwind the stack, and will be processed like the imports. This is mostly useful if you use ASYNCIFY_IGNORE_INDIRECT but want to also mark some additional functions that need to unwind. If the ASYNCIFY_PROPAGATE_ADD setting is disabled however, then this list will only be added after the whole-program analysis. If ASYNCIFY_PROPAGATE_ADD is disabled then you must also add their callers, their callers’ callers, and so on.

  • ASYNCIFY_ONLY is a list of the only functions that can unwind the stack. Asyncify will instrument exactly those and no others.

You can enable the ASYNCIFY_ADVISE setting, which will tell the compiler to output which functions it is currently instrumenting and why. You can then determine whether you should add any functions to ASYNCIFY_REMOVE or whether it would be safe to enable ASYNCIFY_IGNORE_INDIRECT. Note that this phase of the compiler happens after many optimization phases, and several functions maybe be inlined already. To be safe, run it with -O0.

For more details see settings.js. Note that the manual settings mentioned here are error-prone - if you don’t get things exactly right, your application can break. If you don’t absolutely need maximal performance, it’s usually ok to use the defaults.

Potential problems

Stack overflows (Asyncify)

If you see an exception thrown from an asyncify_* API, then it may be a stack overflow. You can increase the stack size with the ASYNCIFY_STACK_SIZE option.

Reentrancy

While waiting on an asynchronous operation browser events can happen. That is often the point of using Asyncify, but unexpected events can happen too. For example, if you just want to pause for 100ms then you can call emscripten_sleep(100), but if you have any event listeners, say for a keypress, then if a key is pressed the handler will fire. If that handler calls into compiled code, then it can be confusing, since it starts to look like coroutines or multithreading, with multiple executions interleaved.

It is not safe to start an async operation while another is already running. The first must complete before the second begins.

Such interleaving may also break assumptions in your codebase. For example, if a function uses a global and assumes nothing else can modify it until it returns, but if that function sleeps and an event causes other code to change that global, then bad things can happen.

Starting to rewind with compiled code on the stack (Asyncify)

The examples above show wakeUp() being called from JS (after a callback, typically), and without any compiled code on the stack. If there were compiled code on the stack, then that could interfere with properly rewinding and resuming execution, in confusing ways, and therefore an assertion will be thrown in a build with ASSERTIONS.

(Specifically, the problem there is that while rewinding will work properly, if you later unwind again, that unwinding will also unwind through that extra compiled code that was on the stack - causing a later rewind to behave badly.)

A simple workaround you may find useful is to do a setTimeout of 0, replacing wakeUp() with setTimeout(wakeUp, 0);. That will run wakeUp in a later callback, when nothing else is on the stack.

Migrating from older Asyncify APIs

If you have code uses the old Emterpreter-Async API, or the old Asyncify, then almost everything should just work when you replace -sEMTERPRETIFY usage with -sASYNCIFY. In particular all the things like emscripten_wget should just work as they did before.

Some minor differences include:

  • The Emterpreter had “yielding” as a concept, but it isn’t needed in Asyncify. You can replace emscripten_sleep_with_yield() calls with emscripten_sleep().

  • The internal JS API is different. See notes above on Asyncify.handleSleep(), and see src/library_async.js for more examples.