Feasibility of automatic cycle breaker for `std::shared_ptr` - c++

Question

C++11 introduced reference-counted smart pointers, std::shared_ptr. Being reference counted, these pointers are unable to automatically reclaim cyclic data structures. However, automatic collection of reference cycles was shown to be possible, for example by Python and PHP. To distinguish this technique from garbage collection, the rest of the question will refer to it as cycle breaking.
Given that there seem to be no proposals to add equivalent functionality to C++, is there a fundamental reason why a cycle breaker similar to the ones already deployed in other languages wouldn't work for std::shared_ptr?
Note that this question doesn't boil down to "why isn't there a GC for C++", which has been asked before. A C++ GC normally refers to a system that automatically manages all dynamically allocated objects, typically implemented using some form of Boehm's conservative collector. It has been pointed out that such a collector is not a good match for RAII. Since a garbage collector primarily manages memory, and might not even be called until there is a memory shortage, and C++ destructors manage other resources, relying on the GC to run destructors would introduce non-determinism at best and resource starvation at worst. It has also bee pointed out that a full-blown GC is largely unnecessary in the presence of the more explicit and predictable smart pointers.
However, a library-based cycle breaker for smart pointers (analogous to the one used by reference-counted interpreters) would have important differences from a general-purpose GC:
It only cares about objects managed through shared_ptr. Such objects already participate in shared ownership, and thus have to handle delayed destructor invocation, whose exact timing depends on ownership structure.
Due to its limited scope, a cycle breaker is unconcerned with patterns that break or slow down Boehm GC, such as pointer masking or huge opaque heap blocks that contain an occasional pointer.
It can be opt-in, like std::enable_shared_from_this. Objects that don't use it don't have to pay for the additional space in the control block to hold the cycle breaker metadata.
A cycle breaker doesn't require a comprehensive list of "root" objects, which is hard to obtain in C++. Unlike a mark-sweep GC which finds all live objects and discards the rest, a cycle breaker only traverses objects that can form cycles. In existing implementations, the type needs to provide help in the form of a function that enumerates references (direct or indirect) to other objects that can participate in a cycle.
It relies on regular "destroy when reference count drops to zero" semantics to destroy cyclic garbage. Once a cycle is identified, the objects that participate in it are requested to clear their strongly-held references, for example by calling reset(). This is enough to break the cycle and would automatically destroy the objects. Asking the objects to provide and clear its strongly-held references (on request) makes sure that the cycle breaker does not break encapsulation.
Lack of proposals for automatic cycle breaking indicates that the idea was rejected for practical or philosophical reasons. I am curious as what the reasons are. For completeness, here are some possible objections:
"It would introduce non-deterministic destruction of cyclic shared_ptr objects." If the programmer were in control of the cycle breaker's invocation, it would not be non-deterministic. Also, once invoked, the cycle breaker's behavior would be predictable - it would destroy all currently known cycles. This is akin to how shared_ptr destructor destroys the underlying object once its reference count drops to zero, despite the possibility of this causing a "non-deterministic" cascade of further destructions.
"A cycle breaker, just like any other form of garbage collection, would introduce pauses in program execution." Experience with runtimes that implement this feature shows that the pauses are minimal because the GC only handles cyclic garbage, and all other objects are reclaimed by reference counting. If the cycle detector is never invoked automatically, the cycle breaker's "pause" could be a predictable consequence of running it, similar to how destroying a large std::vector might run a large number of destructors. (In Python, the cyclic gc is run automatically, but there is API to disable it temporarily in code sections where it is not needed. Re-enabling the GC later will pick up all cyclic garbage created in the meantime.)
"A cycle breaker is unnecessary because cycles are not that frequent and they can be easily avoided using std::weak_ptr." Cycles in fact turn up easily in many simple data structures - e.g. a tree where children have a back-pointer to the parent, or a doubly-linked list. In some cases, cycles between heterogenous objects in complex systems are formed only occasionally with certain patterns of data and are hard to predict and avoid. In some cases it is far from obvious which pointer to replace with the weak variant.

Answer 1

There are a number of issues to be discussed here, so I've rewritten my post to better condense this information.
Automatic cycle detection
Your idea is to have a circle_ptr smart pointer (I know you want to add it to shared_ptr, but it's easier to talk about a new type to compare the two). The idea is that, if the type that the smart pointer is bound to derives from some cycle_detector_mixin, this activates automatic cycle detection.
This mixin also requires that the type implement an interface. It must provide the ability to enumerate all of the circle_ptr instances directly owned by that instance. And it must provide the means to invalidate one of them.
I submit that this is a highly impractical solution to this problem. It is excessively fragile and requires immense amounts of manual work from the user. And therefore, it is not appropriate for inclusion in the standard library. And here are some reasons why.
Determinism and cost
"It would introduce non-deterministic destruction of cyclic shared_ptr objects." Cycle detection only happens when a shared_ptr's reference count drops to zero, so the programmer is in control of when it happens. It would therefore not be non-deterministic. Its behavior would be predictable - it would destroy all currently known cycles from that pointer. This is akin to how shared_ptr destructor destroys the underlying object once its reference count drops to zero, despite the possibility of this causing a "non-deterministic" cascade of further destructions.
This is true, but not in a helpful way.
There is a substantial difference between the determinism of regular shared_ptr destruction and the determinism of what you suggest. Namely: shared_ptr is cheap.
shared_ptr's destructor does an atomic decrement, followed by a conditional test to see if the value was decremented to zero. If so, a destructor is called and memory is freed. That's it.
What you suggest makes this more complicated. Worst-case, every time a circle_ptr is destroyed, the code will have to walk through data structures to determine if there's a cycle. Most of the time, cycles won't exist. But it still has to look for them, just to make sure. And it must do so every single time you destroy a circle_ptr.
Python et. al. get around this problem because they are built into the language. They are able to see everything that's going on. And therefore, they can detect when a pointer is assigned at the time those assignments are made. In this way, such systems are constantly doing small amounts of work to build up cyclic chains. Once a reference goes away, it can look at its data structures and take action if that creates a cyclical chain.
But what you're suggesting is a library feature, not a language feature. And library types can't really do that. Or rather, they can, but only with help.
Remember: an instance of circle_ptr cannot know the subobject it is a member of. It cannot automatically transform a pointer to itself into a pointer to its owning class. And without that ability, it cannot update the data structures in the cycle_detector_mixin that owns it if it is reassigned.
Now, it could manually do this, but only with help from its owning instance. Which means that circle_ptr would need a set of constructors that are given a pointer to its owning instance, which derives from cycle_detector_mixin. And then, its operator= would be able to inform its owner that it has been updated. Obviously, the copy/move assignment would not copy/move the owning instance pointer.
Of course, this requires the owning instance to give a pointer to itself to every circle_ptr that it creates. In every constructor&function that creates circle_ptr instances. Within itself and any classes it owns which are not also managed by cycle_detection_mixin. Without fail. This creates a degree of fragility in the system; manual effort must be expended for each circle_ptr instance owned by a type.
This also requires that circle_ptr contain 3 pointer types: a pointer to the object you get from operator*, a pointer to the actual managed storage, and a pointer to that instance's owner. The reason that the instance must contain a pointer to its owner is that it is per-instance data, not information associated with the block itself. It is the instance of circle_ptr that needs to be able to tell its owner when it is rebound, so the instance needs that data.
And this must be static overhead. You can't know when a circle_ptr instance is within another type and when it isn't. So every circle_ptr, even those that don't use the cycle detection features, must bear this 3 pointer cost.
So not only does this require a large degree of fragility, it's also expensive, bloating the type's size by 50%. Replacing shared_ptr with this type (or more to the point, augmenting shared_ptr with this functionality) is just not viable.
On the plus side, you no longer need users who derive from cycle_detector_mixin to implement a way to fetch the list of circle_ptr instances. Instead, you have the class register itself with the circle_ptr instances. This allows circle_ptr instances that could be cyclic to talk directly to their owning cycle_detector_mixin.
So there's something.
Encapsulation and invariants
The need to be able to tell a class to invalidate one of its circle_ptr objects fundamentally changes the way the class can interact with any of its circle_ptr members.
An invariant is some state that a piece of code assumes is true because it should be logically impossible for it to be false. If you check that a const int variable is > 0, then you have established an invariant for later code that this value is positive.
Encapsulation exists to allow you to be able to build invariants within a class. Constructors alone can't do it, because external code could modify any values that the class stores. Encapsulation allows you to prevent external code from making such modifications. And therefore, you can develop invariants for various data stored by the class.
This is what encapsulation is for.
With a shared_ptr, it is possible to build an invariant around the existence of such a pointer. You can design your class so that the pointer is never null. And therefore, nobody has to check for it being null.
That's not the case with circle_ptr. If you implement the cycle_detector_mixin, then your code must be able to handle the case of any of those circle_ptr instances becoming null. Your destructor therefore cannot assume that they are valid, nor can any code that your destructor calls make that assumption.
Your class therefore cannot establish an invariant with the object pointed to by circle_ptr. At least, not if it's part of a cycle_detector_mixin with its associated registration and whatnot.
You can argue that your design does not technically break encapsulation, since the circle_ptr instances can still be private. But the class is willingly giving up encapsulation to the cycle detection system. And therefore, the class can no longer ensure certain kinds of invariants.
That sounds like breaking encapsulation to me.
Thread safety
In order to access a weak_ptr, the user must lock it. This returns a shared_ptr, which ensures that the object will remain alive (if it still was). Locking is an atomic operation, just like reference incrementing/decrementing. So this is all thread-safe.
circle_ptrs may not be very thread safe. It may be possible for a circle_ptr to become invalid from another thread, if the other thread released the last non-circular reference to it.
I'm not entirely sure about this. It may be that such circumstances only appear if you've already had a data race on the object's destruction, or are using a non-owning reference. But I'm not sure that your design can be thread safe.
Virulence factors
This idea is incredibly viral. Every other type where cyclic references can happen must implement this interface. It's not something you can put on one type. In order to get the benefits, every type that could participate in a cyclical reference must use it. Consistently and correctly.
If you try to make circle_ptr require that the object it manages implement cycle_detector_mixin, then you make it impossible to use such a pointer with any other type. It wouldn't be a replacement of (or augmentation for) shared_ptr. So there is no way for a compiler to help detect accidental misuse.
Sure, there are accidental misuses of make_shared_from_this that cannot be detected by compilers. However, that is not a viral construct. It is therefore only a problem for those who need this feature. By contrast, the only way to get a benefit from cycle_detector_mixin is to use it as comprehensively as possible.
Equally importantly, because this idea is so viral, you will be using it a lot. And therefore, you are far more likely to encounter the multiple-inheritance problem than users of make_shared_from_this. And that's not a minor issue. Especially since cycle_detector_mixin will likely use static_cast to access the derived class, so you won't be able to use virtual inheritance.
Summation
So here is what you must do, without fail, in order to detect cycles, none of which the compiler will verify:
Every class participating in a cycle must be derived from cycle_detector_mixin.
Anytime a cycle_detector_mixin-derived class constructs a circle_ptr instance within itself (either directly or indirectly, but not within a class that itself derives from cycle_detector_mixin), pass a pointer to yourself to that cycle_ptr.
Don't assume that any cycle_ptr subobject of a class is valid. Possibly even to the extent of becoming invalid within a member function thanks to threading issues.
And here are the costs:
Cycle-detecting data structures within cycle_detector_mixin.
Every cycle_ptr must be 50% bigger, even the ones that aren't used for cycle detection.
Misconceptions about ownership
Ultimately, I think this whole idea comes down to a misconception about what shared_ptr is actually for.
"A cycle detector is unnecessary because cycles are not that frequent and they can be easily avoided using std::weak_ptr." Cycles in fact turn up easily in many simple data structures - e.g. a tree where children have a back-pointer to the parent, or a doubly-linked list. In some cases, cycles between heterogenous objects in complex systems are formed only occasionally with certain patterns of data and are hard to predict and avoid. In some cases it is far from obvious which pointer to replace with the weak variant.
This is a very common argument for general-purpose GC. The problem with this argument is that it usually makes an assumption about the use of smart pointers that just isn't valid.
To use a shared_ptr means something. If a class stores a shared_ptr, that represents that the class has ownership of that object.
So explain this: why does a node in a linked list need to own both the next and previous nodes? Why does a child node in a tree need to own its parent node? Oh, they need to be able to reference the other nodes. But they do not need to control the lifetime of them.
For example, I would implement a tree node as an array of unique_ptr to their children, with a single pointer to the parent. A regular pointer, not a smart pointer. After all, if the tree is constructed correctly, the parent will own its children. So if a child node exists, it's parent node must exist; the child cannot exist without having a valid parent.
With a double linked list, I might have the left pointer be a unique_ptr, with the right being a regular pointer. Or vice-versa; one way is no better than the other.
Your mentality seems to be that we should always be using shared_ptr for things, and just let the automatic system work out how to deal with the problems. Whether it's circular references or whatever, just let the system figure it out.
That's not what shared_ptr is for. The goal of smart pointers is not that you don't think about ownership anymore; it's that you can express ownership relationships directly in code.
Overall
How is any of this an improvement over using weak_ptr to break cycles? Instead of recognizing when cycles might happen and doing extra work, you now do a bunch of extra work everywhere. Work that is exceedingly fraglile; if you do it wrong, you're no better off than if you missed a place where you should have used weak_ptr. Only it's worse, because you probably think your code is safe.
The illusion of safety is worse than no safety at all. At least the latter makes you careful.
Could you implement something like this? Possibly. Is it an appropriate type for the standard library? No. It's just too fragile. You must implement it correctly, at all times, in all ways, everywhere that cycles might appear... or you get nothing.
Authoritative references
There can be no authoritative references for something that was never proposed, suggested, or even imagined for standardization. Boost has no such type, and such constructs were never even considered for boost::shared_ptr. Even the very first smart pointer paper (PDF) never considered the possibility. The subject of expanding shared_ptr to automatically be able to handle cycles through some manual effort has never been discussed even on the standard proposal forums where far stupider ideas have been deliberated.
The closest to a reference I can provide is this paper from 1994 about a reference-counted smart pointer. This paper basically talks about making the equivalent of shared_ptr and weak_ptr part of the language (this was in the early days; they didn't even think it was possible to write a shared_ptr that allowed casting a shared_ptr<T> to a shared_ptr<U> when U is a base of T). But even so, it specifically says that cycles would not be collected. It doesn't spend much time on why not, but it does state this:
However, cycles of collected objects with clean-up
functions are problematic. If A and B are reachable from
each other, then destroying either one first will violate
the ordering guarantee, leaving a dangling pointer. If the
collector breaks the cycle arbitrarily, programmers would
have no real ordering guarantee, and subtle, time-dependent
bugs could result. To date, no one has devised a safe,
general solution to this problem [Hayes 92].
This is essentially the encapsulation/invariant issue I pointed out: making a pointer member of a type invalid breaks an invariant.
So basically, few people have even considered the possibility, and those few who did quickly discarded it as being impractical. If you truly believe that they're wrong, the single best way to prove it is by implementing it yourself. Then propose it for standardization.

Answer 2

std::weak_ptr is the solution to this problem. Your worry about
a tree where children have a back-pointer to the parent
can be solved by using raw pointers as the back-pointer. You have no worry of leakage if you think about it.
and your worry about
doubly-linked list
is solved by std::weak_ptr or a raw one.

Answer 3

I believe that the answer to your question is that, contrary to what you claim, there is no efficient way to automatically handle cyclic references. Checking for cycles must be carried out every time a "shared_ptr" is destroyed. On the other hand, introducing any deferring mechanism will inevitably result in a undetermined behavior.

Answer 4

The shared_ptr was not made for automatic reclamation of circular references. It existed in the boost library for some time before being copied to STL. It is a class that attaches a reference counter to any c++ object - be it an array, a class, or an int. It is a relatively lightweight and self-sufficient class. It does not know what it contains, with exception that it knows a deleter function to call when needed.
Al this cycle resolution requires too much heavy code. If you like GC, you can use another language, that was designed for GC from the beginning. Bolting it on via STL would look ugly. A language extension as in C++/CLI would be much nicer.

Answer 5

By reference counting what you ask for is impossible. In order to identify a circle one would have to hold identification of the references to your object. That is easy in memory managed languages since the virtual machine knows who references whom.
In c++ you can only do that by holding a list of references in the circular pointer of e.g. UUID that identifies the object referencing your resources. This would imply that the uuid is somehow passed into the structure when the object is acquired, or that the pointer has access to that resources internals.
These now become implementation specific, since you require a different pointer interface e.g copy and assignment could not be implemented as raw pointers, and demand from every platform to have a uuid source, which cannot be the case for every system. You could of course provide the memory address as a uuid .
Still to overcome the copy , and proper assignment without having a specialized assign method would probably require a single source that allocates references. This cannot be embedded in the language, but may be implemented for a specific application as global registry.
Apart from that, copying such a larger shared pointer would incurr larger performance impact, since during those operations on would have to make lookups for adding , removing, or resolving cycles. Since , doing cycle detection in a graph, from a complexity point of view, would require to traverse the graph registered and apply DFS with backtracking, which is at least proportional to the size of references, I don't see how all these do not scream GC.