Inconsistent `import()` promise order for TLA rejections vs fulfillments

[nicolo-ribaudo]

Consider the test case below.

There are two modules (a and b), with a depending on b. b uses top-level await. We dynamically import a and b, and then settle the promise that b is waiting on.

// entrypoint.js

import { p1, pA, pB } from "./setup.js";

let logs = [];

const importsP = Promise.all([
  import("./a.js").finally(() => logs.push(1)).catch(() => {}),
  import("./b.js").finally(() => logs.push(2)).catch(() => {}),
])

// Wait for evaluation of both graphs with entyrpoints in A and B to start before
// settling the promise that B is blocked on.
Promise.all([pA.promise, pB.promise]).then(p1.resolve);

importsP.then(() => {
  print(logs);
});

// setup.js
export const p1 = Promise.withResolvers();
export const pA = Promise.withResolvers();
export const pB = Promise.withResolvers();

// a.js
import "./a-sentinel.js";
import "./b.js";

// b.js
import "./b-sentinel.js";
import { p1 } from "./setup.js";
await p1.promise;

// a-sentinel.js
import { pA } from "./setup.js";
pA.resolve();

// b-sentinel.js
import { pB } from "./setup.js";
pB.resolve();

• If we fullfill p1 (p1.resolve()), then the logs order will be 2,1 (b's promise settles first). This is because AsyncModuleExecutionFulfilled(module) (where module is b) first resolves module.[[TopLevelCapability]] and then recurses to its parents.
• If we reject p1 (p1.reject()), then the logs order will be 1,2 (a's promise settles first). This is because AsyncModuleExecutionRejected(module) first recurses to module's parents and then rejects module.[[TopLevelCapability]].

Obviously the order for the .resolve() case cannot change (because a could itself use TLA), but the order for .reject() can be changed because there is no code execution between when b errors and all its ancestors get rejected.

My eshost is broken, but I tried in Deno and Bun and neither follows the spec:

• Deno logs 2,1 in both cases
• Bun logs 1,2 in both cases

I propose that we change the spec to match what Deno/V8 is doing.

This is not covered by test262, so even if we decide to keep the current behavior we need to add a test.

[syg]

Deno logs 2,1 in both cases

In d8 I see 2,1 for resolve and 1,2 for reject. Indeed, SourceTextModule::AsyncModuleExecutionReject is verbatim from the spec.

[nicolo-ribaudo]

Weird, I just tested also in Chrome (with an index.html that just contains <script type="module" src="./index.js"></script>) and I still see 2,1 in both cases.

And in Firefox I see 2,1 for resolve and 1,2 for reject.

[nicolo-ribaudo]

Oh and Node.js matches what you see in d8. I guess there is some extra layer at the embedder level that affects it with V8.