vak | Что такое Rc<T>

Продолжим разбираться с умными указателями в Rust. Глянем на тип Rc<T>. Название расшифровывается как Reference Counted, что намекает. Спросим grok.com. Выходит почти тот же shared_ptr, с двумя отличиями: он не может оказаться NULL, и счетчик простой, а не std::atomic. Сгенерённый ним исходник смотрите в файле: rc.cpp

Question: How Rc<T> would look if implemented in C++?

Let’s re-implement Rust’s Rc<T> in C++ - a fascinating exercise since Rc<T> introduces reference-counted shared ownership, contrasting with the unique ownership of Box<T>. In Rust, Rc<T> (short for "Reference Counted") allows multiple owners of a heap-allocated T, tracking the number of references via a counter. When the last Rc is dropped, the T is deallocated. This is akin to C++’s std::shared_ptr<T>, but we’ll craft a custom version to mirror Rust’s Rc<T> behavior, focusing on its single-threaded nature (unlike Arc<T> for multi-threading).

What Rc<T> Does in Rust

Allocation: Places a T on the heap with a reference count.
Shared Ownership: Multiple Rc<T> instances can point to the same T, incrementing the count.
Cloning: Rc::clone creates a new reference, bumping the count.
Deallocation: When the last Rc is dropped (count hits 0), T is destroyed and memory freed.
No Mutability: Rc<T> alone doesn’t allow mutation (use RefCell<T> or Cell<T> for that).
Non-Atomic: Single-threaded, unlike Arc<T>.

C++ Implementation

We’ll create an Rc<T> class in C++ that:

Uses a control block to store the reference count and T together.
Supports copying (to increment the count) and RAII cleanup.
Mimics Rust’s immutability by providing const access (mutation would need a separate mechanism).
Avoids atomic operations, matching Rc’s single-threaded design.

Here’s the code:

#include <cassert>
#include <utility>
#include <iostream>

// Control block to hold the reference count and the value
template <typename T>
struct RcControlBlock {
    size_t count;
    T value;

    template <typename... Args>
    RcControlBlock(Args&&... args) : count(1), value(std::forward<Args>(args)...) {}
};

template <typename T>
class Rc {
private:
    RcControlBlock<T>* block;

    // Private constructor from raw block (for internal use)
    explicit Rc(RcControlBlock<T>* b) : block(b) {}

    void increment() {
        if (block) {
            ++block->count;
        }
    }

    void decrement() {
        if (block && --block->count == 0) {
            delete block;
            block = nullptr;
        }
    }

public:
    // Constructor: Allocate and construct T
    template <typename... Args>
    explicit Rc(Args&&... args) : block(new RcControlBlock<T>(std::forward<Args>(args)...)) {}

    // Destructor: Decrease count, cleanup if last reference
    ~Rc() {
        decrement();
    }

    // Copy constructor: Share ownership
    Rc(const Rc& other) : block(other.block) {
        increment();
    }

    // Copy assignment: Share ownership
    Rc& operator=(const Rc& other) {
        if (this != &other) {
            decrement();           // Drop old reference
            block = other.block;   // Take new reference
            increment();
        }
        return *this;
    }

    // Move constructor: Transfer pointer, no count change
    Rc(Rc&& other) noexcept : block(other.block) {
        other.block = nullptr; // Leave moved-from empty
    }

    // Move assignment: Transfer pointer, no count change
    Rc& operator=(Rc&& other) noexcept {
        if (this != &other) {
            decrement();           // Drop old reference
            block = other.block;   // Take new reference
            other.block = nullptr; // Leave moved-from empty
        }
        return *this;
    }

    // Accessors (const only, mimicking Rust's immutability)
    const T& operator*() const {
        assert(block != nullptr);
        return block->value;
    }

    const T* operator->() const {
        assert(block != nullptr);
        return &block->value;
    }

    // Clone method (like Rc::clone)
    Rc clone() const {
        return Rc(*this); // Uses copy constructor
    }

    // Get reference count (for debugging)
    size_t use_count() const {
        return block ? block->count : 0;
    }
};

// Factory function (like Rc::new)
template <typename T, typename... Args>
Rc<T> make_rc(Args&&... args) {
    return Rc<T>(std::forward<Args>(args)...);
}

This Rc<T> captures Rust’s essence - shared, immutable, single-threaded ownership.

Breaking It Down

1. Control Block

Rust: Rc<T> internally uses a struct like RcBox { ref_count: NonZeroUsize, value: T }.
C++: RcControlBlock<T> holds count and value. Allocated once, shared by all Rc instances.
Why: Keeps the count and data together, ensuring atomic-like management (though non-atomic here).

2. Allocation

Rust: Rc::new allocates via the global allocator, initializes T, sets count to 1.
C++: Constructor uses new RcControlBlock<T>(args...) with perfect forwarding.
Why: C++’s new handles both allocation and construction, like Rust.

3. Shared Ownership

Rust: Cloning via Rc::clone increments the count.
C++: Copy constructor and clone() call increment(), sharing the block.
Why: Mimics Rust’s shared ownership without C++’s atomic ops (unlike shared_ptr).

4. Deallocation

Rust: Drop decrements the count; if 0, deallocates.
C++: Destructor calls decrement(); if count hits 0, deletes the block.
Why: RAII ensures cleanup, matching Rust’s scope-based freeing.

5. Move Semantics

Rust: Moves are cheap pointer transfers.
C++: Move constructor/assignment transfers block, nulls the source, no count change.
Why: Optimizes transfer without affecting reference count.

6. Access

Rust: Deref to &T (immutable).
C++: Const operator* and operator-> enforce immutability.
Why: Reflects Rc<T>’s read-only nature (mutation needs RefCell).

Differences from std::shared_ptr<T>

1. Thread Safety

Rc<T>: Non-atomic, single-threaded (uses usize for count).
std::shared_ptr<T>: Atomic ref counting, thread-safe (uses std::atomic).
Our Rc<T>: Non-atomic, like Rust’s Rc, cheaper but unsafe across threads.

2. Deleter

Rc<T>: Fixed deallocation (global allocator).
std::shared_ptr<T>: Custom deleters via constructor.
Our Rc<T>: Fixed delete, no customization.

3. Weak Pointers

Rc<T>: Pairs with Weak<T> for non-owning refs (separate count).
std::shared_ptr<T>: Pairs with std::weak_ptr<T> (separate weak count).
Our Rc<T>: No Weak equivalent; we’d need another class and a weak count.

4. Allocation

Rc<T>: Single allocation for count and T.
std::shared_ptr<T>: Often two allocations (control block separate from T), unless make_shared is used.
Our Rc<T>: Single allocation in RcControlBlock, like Rust.

5. Nullability

Rc<T>: Always points to a valid T.
std::shared_ptr<T>: Can be null.
Our Rc<T>: Non-null by design, like Rust.

6. Overhead

Rc<T>: sizeof(*mut T) (pointer size).
std::shared_ptr<T>: Often larger (pointer + control block pointer or atomic overhead).
Our Rc<T>: sizeof(T*), minimal like Rust.

Rust Internals Peek

Rust’s Rc<T> is roughly:

struct RcBox<T: ?Sized> {
    strong: usize, // Reference count
    value: T,
}

pub struct Rc<T: ?Sized> {
    ptr: *mut RcBox<T>,
}

impl<T> Rc<T> {
    pub fn new(value: T) -> Rc<T> {
        let ptr = unsafe { alloc::alloc(Layout::for_value(&RcBox { strong: 1, value })) };
        // Initialize...
        Rc { ptr }
    }
}

impl<T: ?Sized> Clone for Rc<T> {
    fn clone(&self) -> Rc<T> {
        unsafe { (*self.ptr).strong += 1; }
        Rc { ptr: self.ptr }
    }
}

impl<T: ?Sized> Drop for Rc<T> {
    fn drop(&mut self) {
        unsafe {
            (*self.ptr).strong -= 1;
            if (*self.ptr).strong == 0 {
                alloc::dealloc(self.ptr as *mut u8, Layout::for_value(&*self.ptr));
            }
        }
    }
}

Our C++ version mirrors this single-allocation, non-atomic design.

Пятнадцать байтов на стек от конца

Что такое Rc<T>

Что такое Rc<T>

What Rc<T> Does in Rust

C++ Implementation

Breaking It Down

Differences from std::shared_ptr<T>

Rust Internals Peek

Профиль

Метки

Посетители