For composable nested parallelism, I think spawn and sync like Julia has would be great. If there is a variant of spawn that asserts there is no concurrency is required for the child task, I'd imagine the runtime can do more optimizations (Julia doesn't have it but there was an experimental PR [1]). But concurrency sounds like a very basic thing in JavaScript so I'm not sure if it fits well in the language. But I don't have experience in serious JavaScript programming so I don't know what fits best there. I'd imagine you can build some data parallel libraries already with web workers API.
