Tracing GCs fundamentally look at what is still reachable, while ref counting tracks deadness. They are actually the “yin-and-yang” of each other, with different tradeoffs, tracing being more practical on current hardware/usage pattern.
There is no practical difference in whether a tracing GC can reach an object from multiple place, or just a single one - but there might be some other analogue feature in tracing corresponding to 0-1 only ref count.
> There is no practical difference in whether a tracing GC can reach an object from multiple place, or just a single one
The real gain would be deterministically run finalizers in this case. Many languages have looked at C++'s RAII and realized they need a deterministic way to do things like close file handles. As such they already have all the parts in place.
It doesn't gain anything else though as the tracer still needs to go through all the pointers in the 0-1 reference count as they might point to something that is sometimes shared but currently is not (and thus the tracer wouldn't reach that memory otherwise - if the memory is never shared it could also be marked 0-1).
I'm not convinced it is a good idea though. Manual memory management is hard. I write C++ every day, std::unique_ptr makes my life much easier, but it has limits (shared_ptr is in my experience a sign you don't understand your memory and so I will have to figure out some rare bug when you stored a reference to the underlying thing). If you always have garbage collection it is easy as you don't have to worry (though you can run into problems: either over allocating when a reference would work; or storing a reference in some data structure after the data isn't needed). If you never have garbage collection you know you need to think about object lifetimes. The 0-1 makes it too easy to have a situation where you didn't think enough and now 0-1 memory cleaned up even though it is still reached.
I don't see how this could be implemented in a safe and performant way - either you check for existing references at runtime, or you risk some use-after-free bug. But perhaps your project is already in a GC language and you're happy with that, but just want to optimise GC for this critical component. And we already have the concept of "unsafe" code blocks in many languages.
Does anything like this exist? I Googled "smart pointers in Java" but just got a bunch of mid-quality answers where people explained that I'm stupid for even asking this and they're unnecessary because Java manages its own memory. But hasn't someone smarter thought of this?