std::move doesn't move anything: A deep dive into Value Categories

The Problem: When “Optimization” Makes Things Slower

Let’s start with something that trips up even experienced developers. You write what looks like perfectly reasonable C++ code:

struct HeavyObject {
    std::string data;

    HeavyObject(HeavyObject&& other) : data(std::move(other.data)) {}

    HeavyObject(const HeavyObject& other) : data(other.data) {}
    
    HeavyObject(const char* s) : data(s) {}
};

std::vector<HeavyObject> createData() {
    std::vector<HeavyObject> data;
    // ... populate data ...
    return data;
}

void processData() {
    auto result = createData();
}

This code works. It compiles. It runs. But depending on how you’ve implemented your types, it might be performing thousands of expensive copy operations instead of cheap moves without you realizing it.

Here’s what’s happening behind the scenes: When your std::vector needs to grow beyond its reserved capacity, it allocates new memory and moves all elements from the old memory to the new memory. But here’s the catch, if your move constructor isn’t marked with the noexcept keyword, the compiler won’t use it at all. Instead, it falls back to copying every single element.

Why? Because std::vector needs to maintain what’s called the “strong exception guarantee.” This is a fancy way of saying: if something goes wrong during reallocation, your original vector should be left completely untouched. If copies throw an exception during reallocation, no problem, the original vector is still intact. But if moves throw an exception, some elements might have already been moved, leaving your original vector in a corrupted state.

So the standard library plays it safe: if your move constructor might throw (because you didn’t mark it noexcept), containers just copy everything instead. That “optimization” you thought you were getting? It’s not happening.

And here’s where things get interesting: std::move won’t magically fix this problem. In fact, if you use it incorrectly, you’ll make things worse. Let me show you why.

The Mechanics: What is std::move Really?

Here’s the truth that might surprise you: std::move doesn’t actually move anything. Not a single byte of memory changes location when you call std::move. it’s one of the most misleading named functions in the C++ standard library.

So what does it acctually do? Let’s look at the real implementation from the standard library (this is from libstdc++, but other standard libraries have similar implementations):

template<typename _Tp>
constexpr typename std::remove_reference<_Tp>::type&&
move(_Tp&& __t) noexcept
{
    return static_cast<typename std::remove_reference<_Tp>::type&&>(__t);
}

If you’re looking at this and thinking “that’s just a cast!”, you’re absolutely right. That’s all it is. std::move takes whatever you pass to it, strips off any reference qualifiers (the std::remove_reference part), adds && to make it an rvalue reference, and then performs a static_cast to that type, That’s the entire function.

Let me put this in simpler terms: std::move is like putting a sign on your object “I’m done with this, you can take its stuff.” The acctual taking of the stuff happens later, when some other code sees that sign. Specifically, that ‘sign’ (the rvalue reference type) tells the compiler to select the Move Constructor instead of the Copy Constructor.

Understanding Value Categories: The Foundation Everything Builds On

Now, to really understand why std::move behaves this way, we need to dive into the concept of Value Categories in C++. thsi is the way to understanding not just std::move, but move semantics in general.

In modern C++, every expression has a Value Category, which determines how that expression can be used. Think of these categories as answering two questions about an expression:

1. Does it have an identity? (Can I take its address?)
2. Can I move from it? (Can I steal its resources?)

Based on these questions, C++ classifies expressions into several categories. We will focus on the three main ones:

graph TD;
    Expression-->GL[glvalue<br/>'has identity'];
    Expression-->R[rvalue<br/>'moveable'];
    GL-->L[lvalue<br/>'I have a name<br/>and an address'];
    GL-->X[xvalue<br/>'I'm about to die<br/>take my stuff!'];
    R-->X;
    R-->PR[prvalue<br/>'I'm a temporary<br/>with no name'];

    style X stroke:#4ade80,stroke-width:2px

Let me break these down:

• lvalue (left value): This is the most familiar category. An lvalue is anything that has a name and occupies a specific location in memory. You can take its address with the & operator. When you write:

int x = 5;
std::string name = "Alice";

Both x and name are lvalues. They have names, they have addresses, they stick around beyond the current expression. You can think of them as “things with permanent addresses.”

• prvalue (pure rvalue): This category represents temporary values that do not have a persistent identity. They are typically created during expressions and are meant to be used immediately. For example:

42              // literal, a prvalue
5 + 3           // result of addition, a prvalue
std::string("hello")  // temporary object, a prvalue

These values don’t have names or addresses you can take. They exist only for the duration of the expression in which they are created. They’re like “things passing through.”

• xvalue (expiring value): This is the special one, and it’s what std::move creates. An xvalue is something that still has an identity (you can refer to it, it has a name), but we’re treating it as if it’s about to expire. It’s like saying “this thing technically still exists, but I’m done with it, so treat it like a temporary.”

When you write std::move(name), you’re not moving name. You’re converting the lvalue name into an xvalue. You’re changing how the compiler sees name. The actual variable name doesn’t go anywhere. Its contents don’t move. All that happens is the compiler now treats this expression as “this object is expiring, you can pillage its resources.”

This is why I said std::move is just a cast. It changes the Value Category of the expression, not the object itself by casting from one value category to another(xvalue). The actual moving of resources happens later, when the Move Constructor or Move Assignment Operator is invoked on that xvalue.

Think of it like this: std::move is putting a “FREE STUFF” sign on your object. The actual taking of the stuff happens when someone who knows how to move sees that sign and acts on it.

The Naive Approach: Three Mistakes That Will Hurt Your Performance

Now that we understand what std::move really does, let’s talk about how people often misuse it, leading to worse performance instead of better.

Mistake 1: return std::move(local_var): Breaking the Compiler’s Best Optimization

This is probably the single most common misuse of std::move.

std::string createString() {
    std::string result = "expensive data";
    // ... do lots of work building up result ...
    return std::move(result);  // DON'T DO THIS!
}

You might think:“Hey, I’m returning a local variable, so I should use std::move to avoid a copy!” But actually, you’re making things worse. Let me explain why.

Modern C++ compilers have an optimization called NRVO (Named Return Value Optimization). Here’s how it works: when you return a local variable, the compiler is allowed to construct that variable directly into the memory location where the caller expects the return value to be. This means no copy or move is needed at all. The object is constructed in place. In other words, instead of:

1. Creating result in the function’s stack frame.
2. Copying or moving result to the caller’s context.
3. Destroying result in the function’s stack frame.

The compiler can optimize it to just:

1. Construct result directly in the caller’s context.

This eliminates both the move and the destruction of the local object. It’s better than moving, it’s zero operations instead of one.

But here’s the catch: NRVO has rules. For the optimization to work, the return statement must return a variable by name. When you write return std::move(result);, you’re no longer returning result by name. Instead, you’re returning an xvalue created by std::move.

SO the compiler says “Well, I can’t do NRVO because this isn’t just a name beig returned.” Now you’re stuck with a move operation. You’ve traded a zero-cost operation (NRVO) for one operation (move). This is called pessimization (the opposite of an optimization).

“But wait” you might say, “isn’t a move cheap?” Yes, moving is usually cheaper than copying. But zero operations are cheaper than one operation. And for types where moving isn’t trivial (maybe your string has a small string optimization, or other complexities), you might be forcing extra work that the compiler could have avoided entirely.

The fix is simple: just return the variable by name:

std::string createString() {
    std::string result = "expensive data";
    // ... do lots of work ...
    return result;  // Correct - enables NRVO, or moves if NRVO can't happen
}

The compiler will do NRVO if it can. If it can’t (maybe because of complex control flow), it will automatically treat the return as an rvalue and move for you. You don’t need to do anything.

The rule: Never use std::move on a return statement with a local variable. The compiler is smarter than you think.

Mistake 2: const T obj; std::move(obj): The Silent Copy

This one is hidden and hurts:

void process() {
    const std::vector<int> data = getData();
    consume(std::move(data));  // Surprise: this COPIES, doesn't move!
}

Let me walk you through why this fails. When you mark something const, you’re telling the compiler “this object’s state won’t change.” But moving is fundamentally about changing state, it’s about taking the resources from an object and transferring them to another. The source object’s state absolutely must change (usually, it becomes empty or null).

So what happens when you try to move from a const object? Let’s look at what the compiler sees:

1. std::move(data) returns const std::vector<int>&& (a const rvalue reference)
2. The move constructor signature is vector(vector&& other) (takes a non-const rvalue reference)
3. A const T&& cannot bind to a T&& parameter (you can’t remove const in a conversion)
4. But a const T&& can bind to a const T& parameter (the copy constructor)
5. So the compiler uses the copy constructor instead

You wrote code that looks like it’s moving, but the compiler silently falls back to copying because it can’t legally move from a const object. All those expensive copies you were trying to avoid are still happening.

Here’s what’s really happening:

// You write:
consume(std::move(data));

// The compiler effectively sees:
// Can't call consume(vector&&) because data is const
// Falls back to consume(const vector&)
// This calls the COPY constructor

This one of the most dangerous bugs because there’s no warning or error. Your code compiles and runs, but it’s not doing what you think it is.

The rule: Never use std::move on const objects. If something is const, you can’t move from it by definition. Moving requires modifying the source object, and const means “cannot modify.”

Mistake 3: Using the Object After Move: Playing with Fire

Here’s the third common mistake:

std::string name = "Alice";
std::string movedName = std::move(name);
std::cout << name << std::endl;  // What happens here?

The C++ standard says that after you move from a standard library object, it’s in a “valid but unspecified state.” Let me unpack what this cryptic phrase means.

“Valid” means the object still maintains its class invariants. For std::string, this means:

• Its internal pointers aren’t dangling
• Its size is consistent with its capacity
• You can safely call its destructor
• You can call methods that don’t have preconditions

“Unspecified” means you don’t know what the value is. Maybe name is now empty. Maybe it still contains “Alice”. Maybe it contains something completely different. The standard doesn’t say, and implementations vary.

Here’s what you can safely do with a moved-from object:

• Destroy it (it will clean up properly)
• Assign to it (name = "Bob" is fine)
• Call methods with no preconditions (name.empty(), name.clear())

Here’s what you should not do:

• Read its value (std::cout << name)
• Call methods with preconditions (name[0] or name.back() assume the string isn’t empty)
• Make any assumptions about its state

In practice, most standard library implementations do leave moved-from objects in a predictable state (strings are usually empty, vectors are usually empty), but this is not guaranteed. Code that relies on this is non-portable and technically undefined behavior.

The mental model I use: treat a moved-from object as if you’d just destroyed it. It’s still technically alive (so you can assign to it), but its value is gone. It’s like a person who’s been mind-wiped, still physically there, but everything that made them “them” is gone.

The rule: After calling std::move on an object, don’t use that object again except to assign a new value to it or destroy it. Treat it as effectively dead.

Implementing Move Semantics Correctly

Now that we’ve covered what to not do, let’s talk about doing it right. If you’re implementing a class that manages resources (memory, file handles, network connections, etc.), you need to implement move semantics. And there’s a well-established pattern for doing this correctly.

The Rule of Five

In modern C++, if you need to implement one of these special member functions, you usually need to implement all five:

1. Destructor
2. Copy Constructor
3. Copy Assignment Operator
4. Move Constructor
5. Move Assignment Operator

This is called the “Rule of Five” (it used to be the “Rule of Three” before move semantics existed). Let me show you a complete, correct implementation, and then we’ll break down each piece:

class Resource {
private:
    int* data;
    size_t size;

public:
    // Constructor
    Resource(size_t n) : data(new int[n]), size(n) {
        std::cout << "Constructing Resource with " << n << " elements\n";
    }
    
    // Destructor
    ~Resource() {
        std::cout << "Destroying Resource\n";
        delete[] data;
    }
    
    // Copy constructor: creates a deep copy
    Resource(const Resource& other) 
        : data(new int[other.size]), size(other.size) {
        std::cout << "Copy constructing Resource\n";
        std::copy(other.data, other.data + size, data);
    }
    
    // Copy assignment operator: creates a deep copy
    Resource& operator=(const Resource& other) {
        std::cout << "Copy assigning Resource\n";
        if (this != &other) {  // Protect against self-assignment
            delete[] data;
            size = other.size;
            data = new int[size];
            std::copy(other.data, other.data + size, data);
        }
        return *this;
    }
    
    // Move constructor: transfers ownership
    Resource(Resource&& other) noexcept
        : data(std::exchange(other.data, nullptr)),
          size(std::exchange(other.size, 0)) {
        std::cout << "Move constructing Resource\n";
    }
    
    // Move assignment operator: transfers ownership
    Resource& operator=(Resource&& other) noexcept {
        std::cout << "Move assigning Resource\n";
        if (this != &other) {  // Protect against self-assignment
            delete[] data;
            data = std::exchange(other.data, nullptr);
            size = std::exchange(other.size, 0);
        }
        return *this;
    }
};

Let me walk through each of these, because each one servers a specific purpose:

The Constructor and Destructor are straightforward: the constructor allocates resources, the destructor frees them. This is basic RAII (Resource Acquisition Is Initialization), the resource’s lifetime is tied to the object’s lifetime.

The Copy Constructor and Copy Assignment do exactly what you’d expect: they create a completely independent copy of the resource. If one Resource owns a chunk of memory, copying it creates a new Resource that owns a completely different chunk of memory with the same contents. Neither object affects the other after copying.

The Move Constructor and Move Assignment are where things get interesting. Instead of creating a new resource, they steal the resource from the source object. Let’s zoom in on the move constructor:

Resource(Resource&& other) noexcept
    : data(std::exchange(other.data, nullptr)),
      size(std::exchange(other.size, 0)) {
    std::cout << "Move constructing Resource\n";
}

Understanding std::exchange: The Clean Way to Move

Notice that we’re using std::exchange here. This is a utility function from <utility> that does two things in one operation:

1. It returns the current value of the first argument.
2. It sets the first argument to the second argument.

So std::exchange(other.data, nullptr) means:

• Get the current value of other.data (the pointer to the resource).
• Set other.data to nullptr (indicating that other no longer owns the resource).
• Return the original pointer value.

This is perfect for implementing moves because we’re doing exactly what moving requires:

1. Taking ownership of the resource from other.
2. Leaving other in a valid, empty state (so its destructor won’t free the resource we just took).

We could write this without std::exchange:

Resource(Resource&& other) noexcept
    : data(other.data), size(other.size) {
    other.data = nullptr;
    other.size = 0;
}

But std::exchange is cleaner and more explicit about what’s happening. It’s the idiomatic way to implement moves in modern C++.

Here’s what’s happening visually:

graph LR;
    subgraph Before Move
    A[Object A<br/>data: 0x1234<br/>size: 1000] -->|owns| Heap[Heap Memory<br/>0x1234<br/>1000 ints];
    B[Object B<br/>data: null<br/>size: 0];
    end
    subgraph After Move
    A2[Object A<br/>data: null<br/>size: 0];
    B2[Object B<br/>data: 0x1234<br/>size: 1000] -->|owns| Heap2[Heap Memory<br/>0x1234<br/>1000 ints];
    end
    style Heap fill:#4ade80,stroke:#fff,color:#000
    style Heap2 fill:#4ade80,stroke:#fff,color:#000

The memory itself doesn’t move. The pointers swap. Object B takes ownership of the heap memory, and Object A is left in a safe empty state.

The Critical Importance of noexcept

Now let’s talk about why both move operations are marked noexcept. This keyword is not optional, it’s absolutely critical for performance.

Remember earlier when I mentioned that std::vector won’t use your move constructor during reallocation unless it’s noexcept? Here’s why:

When std::vector grows, it needs to maintain the strong exception guarantee. This means if something goes wrong during reallocation, the original vector must remain completely unchanged. Let’s think through the scenarios:

Scenario 1: Using copy constructors

• Create new memory block
• Copy element 1 to new memory (might throw)
• If it throws, delete new memory, original vector is untouched
• Copy element 2 to new memory (might throw)
• If it throws, destroy element 1 in new memory, delete new memory, original is untouched
• Continue…

No matter when an exception is thrown, we can clean up and the original vector is fine.

Scenario 2: Using move constructors that may throw

• Create new memory block
• Move element 1 to new memory (might throw)
• If it throws, element 1 is now in an unspecified state
• Move element 2 to new memory (might throw)
• If it throws, element 1 was already moved (possibly corrupted), element 2 is now corrupted
• We cannot restore the original state!

If moves can throw, we can’t guarantee the original vector remains valid if an exception occurs during reallocation. So std::vector makes a choice: if your move constructor might throw (not marked noexcept), it uses copies instead to maintain the guarantee.

The practical impact: if you forget noexcept on your move constructor, every time your vector grows, it will copy all elements instead of moving them. For a vector of complex objects, this could mean millions of unnecessary allocations.

The rule: Always, always, always mark move constructors and move assignment operators noexcept unless you have an exceptional reason not to (and you almost never do).

std::move vs std::forward: Two Tools for Different Jobs

Now that you understand std::move, let’s briefly distinguish it from its cousin, std::forward. These are both cast functions that deal with value categories, but they serve different purposes.

std::move is unconditional. It always converts its argument to an rvalue reference, regardless of what you pass in:

template<typename T>
void process(T&& arg) {
    // arg is an lvalue here (it has a name!)
    consume(std::move(arg));  // Always passes an rvalue to consume
}

Even though arg has && in its type, once it has a name, it’s an lvalue. That’s the rule: if it has a name, it’s an lvalue. So inside process, arg is an lvalue, and we use std::move to convert it back to an rvalue for consumption.

std::forward is conditional. It preserves the value category of its argument. This is used for perfect forwarding in templates:

template<typename T>
void wrapper(T&& arg) {
    // If arg was originally an lvalue, pass it as an lvalue
    // If arg was originally an rvalue, pass it as an rvalue
    process(std::forward<T>(arg));
}

The key difference is that std::forward remembers what arg was when it was passed to wrapper, and maintains that. If you called wrapper(x) with an lvalue x, std::forward<T>(arg) produces an lvalue. If you called wrapper(std::move(x)) with an rvalue, std::forward<T>(arg) produces an rvalue.

Here’s when to use each:

• Use std::move when you know you want to move from something, and you’re done with it
• Use std::forward only in template functions with forwarding reference (T&& where T is a template parameter), when you want to pass arguments along while preserving whether they were lvalues or rvalues

In 99% of normal code, you’ll use std::move. std::forward is mainly for library authors and framework code that’s trying to wrap other functions perfectly.

Modern C++ Context: How Move Semantics Have Evolved

Move semantics were introduced in C++11, but they’ve continued to evolve in subsequent standards. Let’s look at some key developments that affect how you write modern C++.

C++14: constexpr Move

In C++11, std::move was a runtime-only operation. C++14 changed this by marking std::move as constexpr.

This might seem like a small detail, after all, std::move is just a cast, but it was the first step toward enabling complex move-heavy logic (like sorting or swapping) to happen entirely at compile time.

C++17: Mandatory Copy Elision

Before C++17, copy elision (including RVO and NRVO) was an allowed optimization, but not required. The compiler could do it, but didn’t have to. This changed in C++17.

When you return a prvalue (a pure temporary), the compiler is now required to construct the object directly in the caller’s slot:

std::string create() {
    return std::string("hello");  // Guaranteed no copy/move since C++17
}

The object is constructed directly in the caller’s memory. Always. This isn’t an optimization that might happen, it’s required by the standard.

This is different from NRVO (Named Return Value Optimization), which is still optional. When you return a local variable with a name:

std::string create() {
    std::string result = "hello";
    return result;  // NRVO still optional, but compilers usually do it
}

The compiler is allowed to construct result directly in the caller’s memory, but it’s not required to. In practice, modern compilers do this optimization reliably, but it’s technically not guaranteed by the standard.

The takeaway: returning temporaries is guaranteed efficient in C++17 and later. Don’t try to “optimize” with std::move.

C++20: Moving “Heap” Memory at Compile Time

This is where things get wild. While std::move was constexpr since C++14, you couldn’t actually do much with it regarding standard containers because you couldn’t allocate memory at compile time.

C++20 introduced constexpr dynamic allocation. This means std::vector and std::string can now be used (and moved!) inside constexpr functions.

constexpr int sum_data() {
    std::vector<int> data = {1, 2, 3};
    std::vector<int> moved_data = std::move(data); // Valid in C++20!
    
    int sum = 0;
    for(int i : moved_data) sum += i;
    return sum;
}

// The vector is created, moved, and destroyed entirely at compile time
constexpr int result = sum_data();

This enables more sophisticated compile-time programming. You can now write constexpr functions that move objects around, and the entire computation happens at compile time.

C++23: Move-Only Function Wrapper

C++23 introduced std::move_only_function, which is like std::function but can hold non-copyable types:

// Before C++23, this didn't work because std::function requires copyability:
// std::function<void()> func = [ptr = std::make_unique<int>(42)]() {
//     std::cout << *ptr;
// };  // Error: unique_ptr is not copyable!

// C++23 solution:
std::move_only_function<void()> func = [ptr = std::make_unique<int>(42)]() {
    std::cout << *ptr;
};  // Works! func can only be moved, not copied

This is particularly useful for callbacks and handlers that need to own unique resources.

The Future: Trivial Relocation (Still in Development)

There’s an ongoing standardization effort around “trivial relocation” that’s worth knowing about, even though it’s not standardized yet. The basic idea is that for many types, moving an object is equivalent to just copying its bytes and forgetting about the original.

Consider what happens today when a std::vector<std::string> needs to grow (reallocate)

1. For each string, call the move constructor (copies the pointer, nulls the old pointer)
2. For each string in the old memory, call the destructor (checks if pointer is null, does nothing)

That’s lot of function calls. But conceptually, we could just:

1. memcpy the entire block of strings to new memory
2. Forget about the old memory (no destructors needed, they’re all null anyway)

This is called “trivial relocation”, the type can be relocated via a simple byte copy.

Two proposals are competing for this feature:

• P1144 (by Arthur O’Dwyer): Aligns with how libraries like Folly, BDE, and Qt already implement this
• P2786 (by Giuseppe D’Angelo and others): Was merged into the Working Draft at the Hagenberg 2025 meeting, but remains controversial

The controversy stems from differences in semantics and interface. P2786 was merged despite concerns from implementers that its semantics don’t match existing practice. Many major library maintainers prefer P1144’s design.

Why does this matter to you? If/when trivial relocation is standardized, operations like vector reallocation could become dramatically faster for suitable types, potentially orders of magnitude faster for large containers. But the exact details of how to opt in and what guarantees you get are still being decided.

Benchmarks: The Performance Impact of Getting This Right

Let’s look at some concrete numbers to understand the performance implications. I ran benchmarks on x86_64 machine with GCC 13.3.0, optimization level -O3, measuring operations on a vector containing 10’000 custom objects:

Test 1: The Cost of “Safe” Operations

Moving vs. Copying a vector of 10,000 heavy objects.

Test 2: The Return Value Myth

A common “optimization” is wrapping return values in std::move. Does it help?

They are identical.

Test 3: The Cost of Reallocation

The result is a 10x slowdown.

• Compiler explorer link for the benchmark code: https://godbolt.org/z/95d5PerdE

Let me put these numbers in perspective:

Move vs Copy: A correctly implemented move is approximately 7,250 times faster than a copy for this workload. That’s not a typo, 7,250×. The copy requires allocating and copying 1 million objects. The move just swaps a few pointers.

NRVO vs Move: Modern compilers (like GCC 15) are smart enough to construct the return value directly in the caller’s stack frame (Named Return Value Optimization). Adding std::move here doesn’t make it faster, at best, it does nothing. At worst, it prevents the compiler from performing full elision.

The const mistake: Moving from const has exactly the same performance as copying because that’s what it does. The compiler silently chooses the copy constructor, and you get no warning. This is why profiling is so important, code that looks optimal might be doing something completely different.

To put this in real-world terms: If you’re building a data processing pipeline that moves data between stages, getting move semantics wrong could turn a millisecond operation into a 7-millisecond operation. At scale, that’s the difference between processing 1,000 requests per second and processing ~140 requests per second.

Conclusion: The Mental Model for std::move

Let’s bring this all together into a mental model you can use when writing code.

Think of std::move as a promise to the compiler: “I’m done with this object. You can safely pillage its resources.” It’s not an instruction to move, it’s permission to move. The actual moving happens when that permission is acted upon by a move constructor or move assignment operator.

When you write std::move(x), you’re changing how the compiler thinks about x in that expression. You’re converting it from an lvalue (something with a persistent name and address) to an xvalue (something that’s expiring and can be cannibalized). The variable x itself doesn’t go anywhere. Its contents don’t move. It’s still sitting there in memory. You’ve just changed its category in the type system.

The actual movement, the transfer of resources, happens when that xvalue expression is used to construct or assign to another object. That’s when the move constructor or move assignment operator runs, and that’s when the pointers get swapped, the ownership gets transferred, and the source gets nulled out.

Your Practical Checklist

Here’s what you need to remember when writing production C++ code:

1. Don’t use std::move on return values: The compiler will do RVO or NRVO if possible, or automatically move if not. Adding std::move prevents RVO and makes things slower.

2. Always mark move constructors and move assignment operators noexcept: Without this, standard containers won’t use your move operations during reallocation. They’ll copy instead, killing your performance.

3. Never call std::move on const objects: You’ll silently get copies instead of moves, and the compiler won’t warn you. If something is const, it cannot be moved from.

4. Use std::exchange in move implementations: It’s the cleanest, most idiomatic way to implement moves. It makes the ownership transfer explicit and obvious.

5. Don’t use moved-from objects: Except to assign new values or destroy them. Treat them as effectively dead. Their value is gone, even if they’re technically still alive.

6. Reserve std::forward for templates only: In normal code, use std::move. Only use std::forward when implementing perfect forwarding in template functions with universal references.

Real-World Example: Building a Move-Aware Container

Let’s work through a complete real-world example to tie everything together. We’ll build a simple dynamic array class that properly implements move semantics, and I’ll explain every decision along the way.

#include <iostream>
#include <algorithm>
#include <utility>
#include <stdexcept>

template<typename T>
class DynamicArray {
private:
    T* data_;
    size_t size_;
    size_t capacity_;
    
    // Helper to grow capacity when needed
    void reserve_more() {
        size_t new_capacity = capacity_ == 0 ? 1 : capacity_ * 2;
        T* new_data = new T[new_capacity];
        
        // Move existing elements to new memory
        for (size_t i = 0; i < size_; ++i) {
            new_data[i] = std::move(data_[i]);
        }
        
        delete[] data_;
        data_ = new_data;
        capacity_ = new_capacity;
    }

public:
    // Constructor
    DynamicArray() : data_(nullptr), size_(0), capacity_(0) {
        std::cout << "Default constructor\n";
    }
    
    // Destructor
    ~DynamicArray() {
        std::cout << "Destructor (size=" << size_ << ")\n";
        delete[] data_;
    }
    
    // Copy constructor - deep copy
    DynamicArray(const DynamicArray& other)
        : data_(new T[other.capacity_]),
          size_(other.size_),
          capacity_(other.capacity_) {
        std::cout << "Copy constructor (copying " << size_ << " elements)\n";
        std::copy(other.data_, other.data_ + size_, data_);
    }
    
    // Copy assignment
    DynamicArray& operator=(const DynamicArray& other) {
        std::cout << "Copy assignment (copying " << other.size_ << " elements)\n";
        if (this != &other) {
            // Create new data first (strong exception guarantee)
            T* new_data = new T[other.capacity_];
            std::copy(other.data_, other.data_ + other.size_, new_data);
            
            // Only after success do we modify our state
            delete[] data_;
            data_ = new_data;
            size_ = other.size_;
            capacity_ = other.capacity_;
        }
        return *this;
    }
    
    // Move constructor - transfer ownership
    DynamicArray(DynamicArray&& other) noexcept
        : data_(std::exchange(other.data_, nullptr)),
          size_(std::exchange(other.size_, 0)),
          capacity_(std::exchange(other.capacity_, 0)) {
        std::cout << "Move constructor (transferred " << size_ << " elements)\n";
    }
    
    // Move assignment
    DynamicArray& operator=(DynamicArray&& other) noexcept {
        std::cout << "Move assignment (transferred " << other.size_ << " elements)\n";
        if (this != &other) {
            // Clean up our current resources
            delete[] data_;
            
            // Take ownership of other's resources
            data_ = std::exchange(other.data_, nullptr);
            size_ = std::exchange(other.size_, 0);
            capacity_ = std::exchange(other.capacity_, 0);
        }
        return *this;
    }
    
    // Add element
    void push_back(const T& value) {
        if (size_ == capacity_) {
            reserve_more();
        }
        data_[size_++] = value;
    }
    
    // Add element (move version)
    void push_back(T&& value) {
        if (size_ == capacity_) {
            reserve_more();
        }
        data_[size_++] = std::move(value);
    }
    
    // Access
    T& operator[](size_t index) {
        if (index >= size_) throw std::out_of_range("Index out of range");
        return data_[index];
    }
    
    const T& operator[](size_t index) const {
        if (index >= size_) throw std::out_of_range("Index out of range");
        return data_[index];
    }
    
    size_t size() const { return size_; }
    size_t capacity() const { return capacity_; }
};

Now let’s use this class and see exactly what happens:

int main() {
    std::cout << "=== Creating array1 ===\n";
    DynamicArray<std::string> array1;
    array1.push_back("Hello");
    array1.push_back("World");
    
    std::cout << "\n=== Copy construction (array2) ===\n";
    DynamicArray<std::string> array2 = array1;  // Calls copy constructor
    
    std::cout << "\n=== Move construction (array3) ===\n";
    DynamicArray<std::string> array3 = std::move(array1);  // Calls move constructor
    // array1 is now in moved-from state - don't use it!
    
    std::cout << "\n=== Creating array4 for assignment ===\n";
    DynamicArray<std::string> array4;
    
    std::cout << "\n=== Copy assignment ===\n";
    array4 = array2;  // Calls copy assignment
    
    std::cout << "\n=== Move assignment ===\n";
    array4 = std::move(array3);  // Calls move assignment
    // array3 is now in moved-from state
    
    std::cout << "\n=== Function returning by value ===\n";
    auto make_array = []() {
        DynamicArray<std::string> temp;
        temp.push_back("Temporary");
        return temp;  // RVO or automatic move
    };
    
    DynamicArray<std::string> array5 = make_array();
    
    std::cout << "\n=== Destructors will be called ===\n";
    return 0;
}

Let’s walk through what happens at each step:

Step 1: Creating array1

Default constructor

Simple construction. The array starts with null pointer, zero size, zero capacity.

Step 2: Copy construction

Copy constructor (copying 2 elements)

When we write DynamicArray<std::string> array2 = array1, we explicitly want a copy. The copy constructor allocates new memory and copies each element. Both array1 and array2 now own independent resources.

Step 3: Move construction

Move constructor (transferred 2 elements)

Here’s where it gets interesting. std::move(array1) converts array1 to an xvalue. The move constructor sees this and instead of allocating new memory and copying, it just:

• Takes array1’s data pointer
• Takes array1’s size and capacity
• Sets array1’s pointer to nullptr and size/capacity to 0

No memory allocation. No copying. Just pointer swaps. Array3 now owns what array1 used to own, and array1 is left in a valid empty state.

Step 4: Copy assignment

Copy assignment (copying 2 elements)

Array4 already exists (it was default constructed), so we’re using assignment, not construction. The copy assignment allocates new memory, copies the data, then swaps it in. Note that we allocate the new data before deleting the old data, this provides the strong exception guarantee. If allocation fails, array4 remains unchanged.

Step 5: Move assignment

Move assignment (transferred 2 elements)

Similar to move construction, but we first need to clean up array4’s existing resources (if any) before taking ownership of array3’s resources. Again, no allocation, no copying, just pointer swaps.

Step 6: Function return

Default constructor
(possibly) Move constructor (transferred 1 elements)

Let’s talk about why I wrote “(possibly)”.

If you compile this code with high optimizations enabled (like -O3 in GCC/Clang or /O2 in MSVC), you likely won’t see the “Move constructor” line printed at all. You’ll just see the “Default constructor”.

This is due to NRVO (Named Return Value Optimization). The compiler is smart enough to realize that temp inside the lambda and array5 outside the lambda are logically the same object. It elides (skips) the move entirely and constructs the data directly in its final destination.

However, if you compile in Debug mode (where optimizations are turned off to make debugging easier), or if the function logic is too complex for the compiler to figure out, NRVO might not happen.

In that case, C++ guarantees the next best thing: Automatic Move. The compiler implicitly treats the returned local variable as an rvalue. It sees return temp; but treats it as return std::move(temp);

The Takeaway: You get the best of both worlds. Best case: zero cost (NRVO). Worst case: low cost (Move). You never pay the cost of a deep copy here.

• Compiler explorer link for the example: https://godbolt.org/z/j4fbxEcWs

• Code source of the exemple in Github: https://github.com/ttheghost/SimpleDynamicArray

What This Example Teaches Us

This example demonstrates several key principles:

1. The Rule of Five in action: We implemented all five special member functions. If we’d implemented just some of them, we could get surprising behavior. For example, if we implemented a destructor but not a copy constructor, the compiler-generated copy constructor would do a shallow copy, and we’d get double-deletion bugs.

2. noexcept on moves: Notice both move operations are marked noexcept. This is crucial. Without it, if someone puts our DynamicArray into a std::vector, that vector won’t use our move operations during reallocation.

3. Using moved-from objects: After std::move(array1) and std::move(array3), those objects are in moved-from states. They still exist (they haven’t been destroyed), but their resources have been transferred. Their destructors will still run, but they’re destructing empty objects (nullptr is safe to delete).

4. Strong exception guarantee: Notice in the copy assignment operator, we allocate and copy to new memory before we delete the old memory. This way, if the allocation or copy throws, our object remains unchanged. This is a common pattern in exception-safe C++.

5. Overloading for rvalues: Notice we have two versions of push_back, one takes const T& (for lvalues), one takes T&& (for rvalues). When you call push_back with a temporary, the rvalue overload is selected, and we can move the temporary into the array. When you call it with a named variable, the lvalue overload is selected, and we copy.

Performance Pitfalls: Where Move Semantics Go Wrong

Let’s look at some subtle performance issues that can occur even when you think you’re using move semantics correctly.

Pitfall 1: Small String Optimization (SSO) Makes Moves Not Free

Modern standard library implementations use Small String Optimization for std::string. Strings below a certain size (typically 15-23 characters) are stored directly in the string object, not on the heap.

std::string small = "Hi";         // Stored inline, no heap allocation
std::string large = "This is a much longer string that definitely needs heap allocation";

std::string moved_small = std::move(small);  // Copies the inline buffer
std::string moved_large = std::move(large);  // Just swaps pointer

When you move a small string, you’re essentially copying a small fixed-size buffer. This is still faster than heap allocation, but it’s not “free” like moving a large string. The moved-from small string is still set to a valid state (usually empty).

Lesson: Moves aren’t always free. For small objects stored inline, moving is essentially copying. This is still fine, it’s by design, but it’s important to understand what’s actually happening.

Pitfall 2: Forgetting to Move in Loops

std::vector<std::string> source = getLargeStrings();
std::vector<std::string> dest;

for (const auto& s : source) {  // const reference = can't move
    dest.push_back(s);  // Always copies
}

The const here prevents moving. Even if you wanted to move from source, you can’t because const references can’t bind to rvalue references. The correct version:

for (auto& s : source) {  // Non-const reference
    dest.push_back(std::move(s));  // Now we can move
}
// Note: source's strings are now in moved-from state (probably empty)

But be aware: after this loop, source still exists and still contains strings, but they’re all in moved-from states (typically empty). If you’re truly done with source, this is fine. If you need to use source again, this is a bug.

Better approach if you want to consume source:

std::vector<std::string> dest = std::move(source);  // Just move the whole vector
// source is now empty, dest owns everything

This moves the vector itself (just a few pointer swaps), not the individual strings. Much more efficient.

Pitfall 3: Moving Through Multiple Layers

void process(std::string s) {  // Takes by value
    consume(s);  // Oops! Copying, not moving
}

std::string data = "important";
process(std::move(data));  // We moved into process, but process copies to consume

The move gets us into process efficiently, but then we copy to consume. If we want the move to propagate through:

void process(std::string s) {
    consume(std::move(s));  // Now we move to consume
}

This is safe because s is passed by value, we own a copy and can do whatever we want with it. After moving from s, we don’t use it again (the function ends).

Lesson: Moves don’t propagate automatically. Each function call is a new opportunity to either move or copy. Be explicit with std::move when you’re done with a variable.

Pitfall 4: Accidental Copies in Return Statements

std::pair<std::string, std::string> getData() {
    std::string a = "first";
    std::string b = "second";
    return {a, b};  // Copies! Not moves!
}

The braced initializer list {a, b} creates a temporary pair, copying a and b into it. Then that temporary is moved or copy-elided to the return value. We paid for two copies we didn’t need.

Better:

std::pair<std::string, std::string> getData() {
    std::string a = "first";
    std::string b = "second";
    return {std::move(a), std::move(b)};  // Now we move into the pair
}

This is one of the rare cases where using std::move in a return statement is correct, because we’re not moving the return value itself, we’re moving elements into an object we’re constructing for return.

Move Semantics in Inheritance

When you have inheritance, move semantics need to be handled carefully:

class Base {
    std::string base_data_;
public:
    Base(Base&& other) noexcept
        : base_data_(std::move(other.base_data_)) {}
};

class Derived : public Base {
    std::string derived_data_;
public:
    Derived(Derived&& other) noexcept
        : Base(std::move(other)),  // Must explicitly move base
              derived_data_(std::move(other.derived_data_)) {}
};

Notice Base(std::move(other)) in the Derived move constructor. Even though other is an rvalue reference, when used as an expression, it’s an lvalue (it has a name!). Without std::move, we’d call the Base copy constructor, not move constructor.

Critical point: Inside the derived move constructor, other is an lvalue, even though its type is Derived&&. This is the “named rvalue reference is an lvalue” rule that confuses everyone at first.

Final Thoughts: The Philosophy of Move Semantics

Let me leave you with the deeper insight behind move semantics. Before C++11, C++ had a fundamental problem: you had to choose between efficiency and safety.

Efficiency: Pass by pointer, manually manage ownership, risk leaks and dangling pointers.

Safety: Use deep copying everywhere, accept the performance cost.

Move semantics give us a third option: transfer ownership safely and efficiently. They let us write code that’s as safe as deep copying (the compiler tracks everything, no manual management) but as fast as pointer manipulation (no deep copies, just pointer swaps).

The key insight is that many objects have a “dead man walking” state, they’re about to be destroyed, or we’re done using them. Move semantics let us recover value from these doomed objects. Instead of letting their resources die with them, we transfer those resources to objects that will keep using them.

std::move is the explicit marker for this transfer. It’s you telling the compiler: “I’m done with this object. It’s walking dead. Take its organs and give them to someone who needs them.”

This is why the mental model matters more than the syntax. Move semantics aren’t just a performance trick, they’re a fundamental shift in how we think about object lifetime and resource ownership in C++. Understanding them deeply makes you not just faster at writing C++, but better at reasoning about resource management in any language.

Further Reading and Resources

If you want to dive even deeper into move semantics and related topics, here are some excellent resources:

Official Documentation

• cppreference: std::move - Complete reference with examples
• cppreference: Value categories - Detailed explanation of lvalues, rvalues, xvalues, etc.
• cppreference: Copy elision - When and how compilers eliminate copies and moves
• cppreference: Move constructors - The full specification

Standards Proposals

• P1144: Object relocation - Arthur O’Dwyer’s proposal for trivial relocatability
• P2786: Trivial relocatability for C++26 - Alternative approach that was merged into the Working Draft

Articles and Talks

• Understanding when not to std::move (Red Hat Developer Blog) - Focused on the common mistakes
• “Back to Basics: Move Semantics” - Multiple talks from CppCon explaining these concepts from different angles

Books

• “Effective Modern C++” by Scott Meyers - Items 23-25 cover move semantics in detail
• “C++ Move Semantics - The Complete Guide” by Nicolai Josuttis - Entire book dedicated to this topic

Remember: move semantics are a tool, not a goal. The goal is to write correct, maintainable, and efficient code. Move semantics help achieve that goal, but only when used appropriately. Sometimes a copy is exactly what you need. Sometimes the compiler eliminates the operation entirely. Understanding when and how to use moves—and when not to—is what separates good C++ code from great C++ code.