alloc sans `alloc`

[ds84182]

Given that the Allocator trait is in core, shouldn't the various allocatable collections be available in core as well?

e.g. Vec<T, A> and Box<T, A> in core without the fns that require A = GlobalAlloc (so no Vec::new, only Vec::new_in).

This would solve #123.

[CAD97]

This unfortunately effectively requires solving at least inference-affecting defaults in practice, which is a very difficult problem.

We could potentially break coherence within the std facade to accomplish this, but we've generally been hesitant to do so beyond what already exists, at least without first developing a way for non-std crates to do so.

[ds84182]

maybe it's worthwhile experimenting with it for just Box right now? Box-like types provided by third party crates can't be moved out of, so there's actually a benefit to this. Also it's the only stable way to call dyn FnOnce.

[ds84182]

Coming back to this, I feel the future vision of alloc should ideally be "core + default allocator".

I have a project where I actually don't want to have a global allocator, but Box, Vec, and Arc are extremely useful. I have a process-local heap and a cross-process shared memory heap. My global allocator is process-local, so if anyone makes the mistake of putting a Box or Arc (using the default allocator) in shared memory it could trigger an invalid free in another process. It's really not ideal and I don't have tools to guard against it.