When to use reinterpret_cast?

The C++ standard guarantees the following:

static_casting a pointer to and from void* preserves the address. That is, in the following, a, b and c all point to the same address:

int* a = new int();
void* b = static_cast<void*>(a);
int* c = static_cast<int*>(b);

reinterpret_cast only guarantees that if you cast a pointer to a different type, and then reinterpret_cast it back to the original type, you get the original value. So in the following:

int* a = new int();
void* b = reinterpret_cast<void*>(a);
int* c = reinterpret_cast<int*>(b);

a and c contain the same value, but the value of b is unspecified. (in practice it will typically contain the same address as a and c, but that's not specified in the standard, and it may not be true on machines with more complex memory systems.)

For casting to and from void*, static_cast should be preferred.

One case when reinterpret_cast is necessary is when interfacing with opaque data types. This occurs frequently in vendor APIs over which the programmer has no control. Here's a contrived example where a vendor provides an API for storing and retrieving arbitrary global data:

// vendor.hpp
typedef struct _Opaque * VendorGlobalUserData;
void VendorSetUserData(VendorGlobalUserData p);
VendorGlobalUserData VendorGetUserData();

To use this API, the programmer must cast their data to VendorGlobalUserData and back again. static_cast won't work, one must use reinterpret_cast:

// main.cpp
#include "vendor.hpp"
#include <iostream>
using namespace std;

struct MyUserData {
    MyUserData() : m(42) {}
    int m;
};

int main() {
    MyUserData u;

        // store global data
    VendorGlobalUserData d1;
//  d1 = &u;                                          // compile error
//  d1 = static_cast<VendorGlobalUserData>(&u);       // compile error
    d1 = reinterpret_cast<VendorGlobalUserData>(&u);  // ok
    VendorSetUserData(d1);

        // do other stuff...

        // retrieve global data
    VendorGlobalUserData d2 = VendorGetUserData();
    MyUserData * p = 0;
//  p = d2;                                           // compile error
//  p = static_cast<MyUserData *>(d2);                // compile error
    p = reinterpret_cast<MyUserData *>(d2);           // ok

    if (p) { cout << p->m << endl; }
    return 0;
}

Below is a contrived implementation of the sample API:

// vendor.cpp
static VendorGlobalUserData g = 0;
void VendorSetUserData(VendorGlobalUserData p) { g = p; }
VendorGlobalUserData VendorGetUserData() { return g; }

The short answer: If you don't know what reinterpret_cast stands for, don't use it. If you will need it in the future, you will know.

Full answer:

Let's consider basic number types.

When you convert for example int(12) to unsigned float (12.0f) your processor needs to invoke some calculations as both numbers has different bit representation. This is what static_cast stands for.

On the other hand, when you call reinterpret_cast the CPU does not invoke any calculations. It just treats a set of bits in the memory like if it had another type. So when you convert int* to float* with this keyword, the new value (after pointer dereferecing) has nothing to do with the old value in mathematical meaning (ignoring the fact that it is undefined behavior to read this value).

Be aware that reading or modifying values after reinterprt_cast'ing are very often Undefined Behavior. In most cases, you should use pointer or reference to std::byte (starting from C++17) if you want to achieve the bit representation of some data, it is almost always a legal operation. Other "safe" types are char and unsigned char, but I would say it shouldn't be used for that purpose in modern C++ as std::byte has better semantics.

Example: It is true that reinterpret_cast is not portable because of one reason - byte order (endianness). But this is often surprisingly the best reason to use it. Let's imagine the example: you have to read binary 32bit number from file, and you know it is big endian. Your code has to be generic and works properly on big endian (e.g. some ARM) and little endian (e.g. x86) systems. So you have to check the byte order. It is well-known on compile time so you can write constexpr function: You can write a function to achieve this:

/*constexpr*/ bool is_little_endian() {
  std::uint16_t x=0x0001;
  auto p = reinterpret_cast<std::uint8_t*>(&x);
  return *p != 0;
}

Explanation: the binary representation of x in memory could be 0000'0000'0000'0001 (big) or 0000'0001'0000'0000 (little endian). After reinterpret-casting the byte under p pointer could be respectively 0000'0000 or 0000'0001. If you use static-casting, it will always be 0000'0001, no matter what endianness is being used.

EDIT:

_{In the first version I made example function is_little_endian to be constexpr. It compiles fine on the newest gcc (8.3.0) but the standard says it is illegal. The clang compiler refuses to compile it (which is correct).}

The meaning of reinterpret_cast is not defined by the C++ standard. Hence, in theory a reinterpret_cast could crash your program. In practice compilers try to do what you expect, which is to interpret the bits of what you are passing in as if they were the type you are casting to. If you know what the compilers you are going to use do with reinterpret_cast you can use it, but to say that it is portable would be lying.

For the case you describe, and pretty much any case where you might consider reinterpret_cast, you can use static_cast or some other alternative instead. Among other things the standard has this to say about what you can expect of static_cast (§5.2.9):

An rvalue of type “pointer to cv void” can be explicitly converted to a pointer to object type. A value of type pointer to object converted to “pointer to cv void” and back to the original pointer type will have its original value.

So for your use case, it seems fairly clear that the standardization committee intended for you to use static_cast.

One use of reinterpret_cast is if you want to apply bitwise operations to (IEEE 754) floats. One example of this was the Fast Inverse Square-Root trick:

https://en.wikipedia.org/wiki/Fast_inverse_square_root#Overview_of_the_code

It treats the binary representation of the float as an integer, shifts it right and subtracts it from a constant, thereby halving and negating the exponent. After converting back to a float, it's subjected to a Newton-Raphson iteration to make this approximation more exact:

float Q_rsqrt( float number )
{
    long i;
    float x2, y;
    const float threehalfs = 1.5F;

    x2 = number * 0.5F;
    y  = number;
    i  = * ( long * ) &y;                       // evil floating point bit level hacking
    i  = 0x5f3759df - ( i >> 1 );               // what the deuce? 
    y  = * ( float * ) &i;
    y  = y * ( threehalfs - ( x2 * y * y ) );   // 1st iteration
//  y  = y * ( threehalfs - ( x2 * y * y ) );   // 2nd iteration, this can be removed

    return y;
}

This was originally written in C, so uses C casts, but the analogous C++ cast is the reinterpret_cast.

Here is a variant of Avi Ginsburg's program which clearly illustrates the property of reinterpret_cast mentioned by Chris Luengo, flodin, and cmdLP: that the compiler treats the pointed-to memory location as if it were an object of the new type:

#include <iostream>
#include <string>
#include <iomanip>
using namespace std;

class A
{
public:
    int i;
};

class B : public A
{
public:
    virtual void f() {}
};

int main()
{
    string s;
    B b;
    b.i = 0;
    A* as = static_cast<A*>(&b);
    A* ar = reinterpret_cast<A*>(&b);
    B* c = reinterpret_cast<B*>(ar);
    
    cout << "as->i = " << hex << setfill('0')  << as->i << "\n";
    cout << "ar->i = " << ar->i << "\n";
    cout << "b.i   = " << b.i << "\n";
    cout << "c->i  = " << c->i << "\n";
    cout << "\n";
    cout << "&(as->i) = " << &(as->i) << "\n";
    cout << "&(ar->i) = " << &(ar->i) << "\n";
    cout << "&(b.i) = " << &(b.i) << "\n";
    cout << "&(c->i) = " << &(c->i) << "\n";
    cout << "\n";
    cout << "&b = " << &b << "\n";
    cout << "as = " << as << "\n";
    cout << "ar = " << ar << "\n";
    cout << "c  = " << c  << "\n";
    
    cout << "Press ENTER to exit.\n";
    getline(cin,s);
}

Which results in output like this:

as->i = 0
ar->i = 50ee64
b.i   = 0
c->i  = 0

&(as->i) = 00EFF978
&(ar->i) = 00EFF974
&(b.i)   = 00EFF978
&(c->i)  = 00EFF978

&b = 00EFF974
as = 00EFF978
ar = 00EFF974
c  = 00EFF974
Press ENTER to exit.

It can be seen that the B object is built in memory as B-specific data first, followed by the embedded A object. The static_cast correctly returns the address of the embedded A object, and the pointer created by static_cast correctly gives the value of the data field. The pointer generated by reinterpret_cast treats b's memory location as if it were a plain A object, and so when the pointer tries to get the data field it returns some B-specific data as if it were the contents of this field.

One use of reinterpret_cast is to convert a pointer to an unsigned integer (when pointers and unsigned integers are the same size):

int i; unsigned int u = reinterpret_cast<unsigned int>(&i);

You could use reinterprete_cast to check inheritance at compile time.
Look here: Using reinterpret_cast to check inheritance at compile time

template <class outType, class inType>
outType safe_cast(inType pointer)
{
    void* temp = static_cast<void*>(pointer);
    return static_cast<outType>(temp);
}

I tried to conclude and wrote a simple safe cast using templates. Note that this solution doesn't guarantee to cast pointers on a functions.

First you have some data in a specific type like int here:

int x = 0x7fffffff://==nan in binary representation

Then you want to access the same variable as an other type like float: You can decide between

float y = reinterpret_cast<float&>(x);

//this could only be used in cpp, looks like a function with template-parameters

or

float y = *(float*)&(x);

//this could be used in c and cpp

BRIEF: it means that the same memory is used as a different type. So you could convert binary representations of floats as int type like above to floats. 0x80000000 is -0 for example (the mantissa and exponent are null but the sign, the msb, is one. This also works for doubles and long doubles.

OPTIMIZE: I think reinterpret_cast would be optimized in many compilers, while the c-casting is made by pointerarithmetic (the value must be copied to the memory, cause pointers couldn't point to cpu- registers).

NOTE: In both cases you should save the casted value in a variable before cast! This macro could help:

#define asvar(x) ({decltype(x) __tmp__ = (x); __tmp__; })

Quick answer: use static_cast if it compiles, otherwise resort to reinterpret_cast.

Read the FAQ! Holding C++ data in C can be risky.

In C++, a pointer to an object can be converted to void * without any casts. But it's not true the other way round. You'd need a static_cast to get the original pointer back.