dispatch2: Add fn data_create() which copies a slice

[MarijnS95]

Fixes #699

TODO

• Add a variant that takes ownership over a Box (or T?) and calls its destructor (EDIT: This requires a raw Block pointer, which we can probably get via a Deref on GlobalBlock).
• Make regenerated Metal code import/use the correct dispatch_data_t type (preferably taking &DispatchObject and calling .as_raw()?).
• Ensure calling set_finalizer() is still a valid thing to do?
• Can the header translator parse #define or are we expected to redefine them ourselves?

[madsmtm]

Add a variant that takes ownership over a Box (or T?) and calls its destructor (EDIT: This requires a raw Block pointer, which we can probably get via a Deref on GlobalBlock).

Yeah, could be useful. A variant which takes &'static [u8] could also be, and so could taking Arc<[u8]>.

Maybe it could even be generic over T: AsRef<[u8]> + 'static? Though a difficulty here is in thread-safety, I'm not sure, but I think + Send would also be required (maybe depending on the queue the destructor is run on)?

Hmm... It actually seems like the block passed to dispatch_data_create doesn't take any parameters, so it would have to keep a pointer to the original data inside the block? That seems odd? Maybe I don't understand this API enough?

• Make regenerated Metal code import/use the correct dispatch_data_t type (preferably taking &DispatchObject and calling .as_raw()?).

See #681; I intend to start work on that soon, that should resolve this kind of issue.

Can the header translator parse #define or are we expected to redefine them ourselves?

See #609.

[MarijnS95]

Hmm... It actually seems like the block passed to dispatch_data_create doesn't take any parameters, so it would have to keep a pointer to the original data inside the block? That seems odd? Maybe I don't understand this API enough?

That's actually useful for DSTs like [u8] which need a fat pointer with size to reconstruct the size. moveing a raw pointer into a new closure solves this naturally (but we do need to allocate a new block for each call, which could be circumvented if the API returned a pointer plus size).

I'll experiment with your suggestions for AsRef<[u8]> + 'static later, that does sound great! Let me invesigate the threading on these queues, I'm not too familiar with them yet.

• Make regenerated Metal code import/use the correct dispatch_data_t type (preferably taking &DispatchObject and calling .as_raw()?).

See #681; I intend to start work on that soon, that should resolve this kind of issue.

Cool yeah, I see that adds a trait AsRawDispatchObject to get the raw object pointer, but see no generator changes: are those still needed to support converting the object on the fly?

I got the header-translator to run and the output tries to use the low-level FFI bindings imported from use dispatch2::*; but simple typedefs are (obviously) not publicly reexported from the dispatch2 root: https://github.com/MarijnS95/objc2-generated/compare/mtl-dispatch

[madsmtm]

Sorry, still haven't gotten around to finishing the dispatch2 work. It's on the milestone for the next patch release though, so hopefully I'll get it done this month.