[MaxDesiatov]

A ~Copyable @JSClass struct stores a noncopyable JSUniqueRef, and generated bindings take these wrappers borrowing. In the new Imported object benchmarks, creating or returning a noncopyable wrapper takes 15-26% less time than a copyable wrapper on main.

Median ns per iteration:

┌──────────────────────────┬──────────────────┬─────────────┬────────┐
│           Case           │ Copyable on main │ Noncopyable │ Change │
├──────────────────────────┼──────────────────┼─────────────┼────────┤
│ create and drop          │ 71.1             │ 52.5        │ -26.2% │
├──────────────────────────┼──────────────────┼─────────────┼────────┤
│ create, call and drop    │ 80.5             │ 62.8        │ -22.0% │
├──────────────────────────┼──────────────────┼─────────────┼────────┤
│ return and drop          │ 71.3             │ 58.4        │ -18.1% │
├──────────────────────────┼──────────────────┼─────────────┼────────┤
│ return and pass          │ 82.8             │ 70.7        │ -14.7% │
├──────────────────────────┼──────────────────┼─────────────┼────────┤
│ pass an existing wrapper │ 16.4             │ 16.4        │ -0.3%  │
└──────────────────────────┴──────────────────┴─────────────┴────────┘

Arrays of JSUniqueRef-values are bridged as UniqueArray, which is made conditionally available when using Swift 6.4. No dictionary counterpart is available, those either require users to manually add and bridge using swift-collections dependency that has UniqueDictionary, or bridging those as usual copyable values backed by copyable Dictionary.

[MaxDesiatov]

cc @sliemeobn

[kateinoigakukun]

Can you split PRs into series, one for adding JSUniqueRef, one for supporting @JSClass struct C: ~Copyable? I haven't checked the details of BridgeJS and TS2Swift changes but it seems like containing a lot of the new things.

Also allowing the JSObject -> JSUniqueRef conversion looks like a soundness hole? The opposite conversion (JSUniqueRef -> JSObject) should be allowed, but the "takeover" from JSObject to JSUniqueRef looks wrong because we cannot ban the further JSObject usage after the takeover?