Errored modules have an empty [[CycleRoot]]

[nicolo-ribaudo]

This PR fixes #3582 by defaulting [[CycleRoot]] to the module itself in case of error.

In case the module is effectively on a cycle, this would be lying, but in practice it doesn't matter: when a module has an error, the error is immediately propagated to all its ancestors, so the real cycle root has the same error state.

I originally tried to set [[CycleRoot]] properly, which requires more complexity. This code that sets the states of the modules in stack in case of error, needs to:

• first, iterate the stack reversed, propagating setting [[DFSAncestorIndex]] of each module to min([[DFSAncestorIndex]] of module, [[DFSAncestorIndex]] of its child)
• then, when iterating in order, it needs to set [[CycleRoot]] to the module itself unless the [[DFSAncestorIndex]] is the same as the previous module, in which case it copies the [[CycleRoot]] from it.

While this complex approach is more correct, I discarded it because it's still not 100% correct. In this graph, assuming that E throws, we will have not traversed F at that point so we don't know that D should actually have [[CycleRoot]] set to C.

In https://github.com/nicolo-ribaudo/ecma262/tree/cycleroot-fix-ugly I have an alternative fix if you prefer, that instead of ensuring that [[CycleRoot]] is always set it acknowledges that in this specific case it will be ~empty~.

[syg]

Editors' call: We prefer the alternative fix at 09d32a3 for clarity, since it immediately answers the question of what [[CycleRoot]] is used for for the error case.

@guybedford Can we ask a favor for you to review the alternative fix being equivalent to this one?

[guybedford]

@syg sure I'd be happy to do a careful review on this. Will aim to follow-up soon.

[guybedford]

I have left a review at 09d32a3#r167698635:

100% agreed this is a nicer approach than eagerly setting the cycle root, and more closely matches the intent.

That said, while the first two assertions are well-defined semantics, for the last assertion on _module_.[[TopLevelCapability]]:

1. I don't fully follow that this is true - surely it isn't the case if there was a transitive dependency error, and that dependency is later imported directly?
2. Even if it is true, it seems superfluous to the underlying semantic being verified so still may be better to remove.

[nicolo-ribaudo]

(Guy's comment is for this line)

[nicolo-ribaudo]

You are right, that last part is trivially false. Updated.

[github-actions]

The rendered spec for this PR is available at https://tc39.es/ecma262/pr/3583.