ScriptEvaluation semantics for ShadowRealm

[mhofman]

It looks like PerformShadowRealmEval is based off PerformEval, which means there is no way to have the ScriptEvaluation semantics in ShadowRealms, namely the behavior of GlobalDeclarationInstantiation to create global object properties for var and function declarations.

I'm wondering if this option was ever discussed. While it'd be easy to add a new API later (e.g. ShadowRealm.prototype.executeScript()) with the ScriptEvaluation semantics, I also don't think there'd be harm in changing the semantics of PerformShadowRealmEval before shipping. It's always possible to go from ScriptEvaluation to PerformEval (const evaluate = shadowRealm.executeScript('eval')), but not the other way around.

[caridy]

Let's discuss this next week at SES meeting. @erights and @nicolo-ribaudo might have opinions here as well.

[ptomato]

Tomorrow's there is no SES meeting due to TG3. I've put it on the agenda for 2024-02-21.

[nicolo-ribaudo]

We discussed this during today's SES meeting -- we didn't come to a conclusion, but there were to points in favor of the two different options.

To recap, the observable difference is that realm.evaluate("var foo = 1") would create a global configurable property under PerformEval semantics, and a global non-configurable property under ScriptEvaluation semantics.

• PerformEval has the advantage that usually code written assuming a non-configurable property will also work if that property is configurable, while the other way around would break.
• ScriptEvaluation has the advantage that today the majority of script code is being executed through <script> tags and not through eval(), so it matches the most common way of running code.

In addition to this, @mhofman notes that const evaluate = shadowRealm.executeScript('eval') doesn't actually work when considering TrustedTypes, because that eval would be behind the callable boundary and thus it can only accept strings (while shadowRealm.executeScript(trustedTypesObject) would work).

[nicolo-ribaudo]

Also, I realized that there is another difference we didn't explicitly talk about during the SES meeting (but maybe it's a consequence of the configurability of properties). The following code:

realm.evaluate("var a = 1");
realm.evaluate("let a = 2");

will work under PerformEval semantics, and throw under ScriptEvaluation semantics.

Test cases:

<!-- ScriptEvaluation -->
<script>var a = 1;</script>
<script>let a = 2;</script>

<!-- PerformEval -->
<script>
  eval("var a = 1;");
  eval("let a = 1;");
</script>