@mongosh/async-rewriter2

next-gen async-rewriter

This package contains babel plugins that transpile code in a way that allows implicitly awaiting selecting Promises.

Motivation

The predecessor of this package uses a symbol-table-based approach, in which it uses static analysis to keep track of which function calls would end up needing an implicit await in front of it. This is brittle and strongly limits the set of JS features that could be used in the shell, as well as the ways in which one could interact with API object programmatically.

Therefore, this package drops the static analysis part and focuses entirely on transforming the code in a way that allows all 'interesting' work to happen at runtime. It is fully stateless and enables removing any symbol table tracking. It’s also easier to introduce support for more built-in JS functions that take callbacks this way, as replacing them with a polyfill (that is also transformed) does the trick here. As such, special-casing calls to e.g. .forEach() is no longer necessary either.

Downsides to this approach are that it’s not currently taking into account situations in which implicitly async functions cannot be used (e.g. class constructors or synchronous generator functions), that we don’t error for conflicting API usage (e.g. top-level variables with names like db) and that error messages may end up referring to odd locations in the code (at least without support for source maps).

Idea

We (ab-)use the fact that async functions execute fully synchronously until they reach their first await expression, and the fact that we can determine which Promises need awaiting by marking them as such using decorators on the API surface.

The transformation takes place in two main steps.

Step one: IIFE wrapping

The input code is wrapped in an IIFE. For example:

function foo() { return db.test.find(); }
class A {}
foo()

is converted into roughly:

var A;

function foo() {
  return db.test.find();
}

(() => {
  A = class A {};
  return foo();
})();

Note how identifiers remain accessible in the outside environment, including top-level functions being hoisted to the outside.

Step two: Async function wrapping

We perform three operations:

1. We give all shorthand arrow functions statement bodies. This is necessary for the other steps to work.
2. We turn all input functions into async functions and generate non-async wrappers for them. We keep track of the execution state of the ’inner’ async function when it is called, and forward synchronous results synchronously.
3. We add checks for most expressions inside the ‘inner’ function, which conditionally uses await based on whether the result of the expression has a specific Symbol property. This Symbol property is set by functions in the API whose results should be implicitly awaited.

This does result in a significant increase in code size. For example,

(() => {
  return db.test.find().toArray();
})();

(which is the result of db.test.find().toArray() after Step 1) would be turned into code looking like the following (some adjustments have been made for readability).

(() => {
  // Keep a copy of the original source code for Function.prototype.toString.
  '<async_rewriter>(() => {\n  return db.test.find().toArray();\n})</>';
  const _syntheticPromise = Symbol.for("@@mongosh.syntheticPromise");

  function _markSyntheticPromise(p) {
    return Object.defineProperty(p, _syntheticPromise, {
      value: true
    });
  }

  function _isp(p) { // '_isSyntheticPromise' would be way too long here
    return p && p[_syntheticPromise];
  }

  function _demangleError(err) {
    // ... fix up the error message in 'err' using the original source code ...
  }

  let _functionState = "sync",
      _synchronousReturnValue,
      _ex;

  const _asynchronousReturnValue = (async () => {
    try {
      // All return statements are decorated with
      // `return (_synchronousReturnValue = ..., _functionState === 'async' ? _synchronousReturnValue : null)`
      // The function state check is here that, if we are returning synchronously,
      // we know that we are going to discard the value of `_asynchronousReturnValue`,
      // which is not what we want if the return value happens to be a rejected
      // Promise (because Node.js print a warning in that case).
      return (
        _synchronousReturnValue = (
          // Most expressions are wrapped in ('original source', _ex = ..., _isp(_ex) ? await _ex : _ex)
          _ex = ('db.test.find()',
            _ex = ('db.test',
              _ex = ('db',
                _ex = db, _isp(_ex) ? await _ex : _ex
              ).test, _isp(_ex) ? await _ex : _ex
            ).find(), _isp(_ex) ? await _ex : _ex
          ).toArray()
          , _isp(_ex) ? await _ex : _ex
        ),
        _functionState === 'async' ? _synchronousReturnValue : null);
    } catch (err) {
      err = _demangleError(err);
      if (_functionState === "sync") {
        // Forward synchronous exceptions.
        _synchronousReturnValue = err;
        _functionState = "threw";
      } else {
        // If we are already asynchronous, just return a rejected Promise as usual.
        throw err;
      }
    } finally {
      // If we did not throw here, we returned. Tell the caller that.
      if (_functionState !== "threw") {
        _functionState = "returned";
      }
    }
  })();

  if (_functionState === "returned") {
    return _synchronousReturnValue;
  } else if (_functionState === "threw") {
    throw _synchronousReturnValue;
  }

  _functionState = "async";
  // Since this was originally a non-async function, mark this as something
  // that should implicitly be awaited.
  return _markSyntheticPromise(_asynchronousReturnValue);
})();

API

import AsyncWriter from '@mongosh/async-rewriter2';
const transpiledCodeString = new AsyncWriter().process(inputCodeString);