ChatGPT解决这个技术问题 Extra ChatGPT

When should static_cast, dynamic_cast, const_cast, and reinterpret_cast be used?

What are the proper uses of:

static_cast

dynamic_cast

const_cast

reinterpret_cast

(type)value (C-style cast)

type(value) (function-style cast)

How does one decide which to use in which specific cases?

For some useful concrete examples of using different kind of casts, you can check the first answer on a similar question in this other topic.
You can find really good answers for your question above. But I would like to put one more point here, @e.James "There is nothing that these new c++ cast operators can do and c style cast cannot. These are added more or less for the better code readability."
@BreakBadSP The new casts are not only for better code readability. They are there to make it harder to do do dangerous things, like casting away const or casting pointers instead of their values. static_cast has much less possibilities to do something dangerous than a c style cast!

2
24 revs, 18 users 67%

static_cast is the first cast you should attempt to use. It does things like implicit conversions between types (such as int to float, or pointer to void*), and it can also call explicit conversion functions (or implicit ones). In many cases, explicitly stating static_cast isn't necessary, but it's important to note that the T(something) syntax is equivalent to (T)something and should be avoided (more on that later). A T(something, something_else) is safe, however, and guaranteed to call the constructor.

static_cast can also cast through inheritance hierarchies. It is unnecessary when casting upwards (towards a base class), but when casting downwards it can be used as long as it doesn't cast through virtual inheritance. It does not do checking, however, and it is undefined behavior to static_cast down a hierarchy to a type that isn't actually the type of the object.

const_cast can be used to remove or add const to a variable; no other C++ cast is capable of removing it (not even reinterpret_cast). It is important to note that modifying a formerly const value is only undefined if the original variable is const; if you use it to take the const off a reference to something that wasn't declared with const, it is safe. This can be useful when overloading member functions based on const, for instance. It can also be used to add const to an object, such as to call a member function overload.

const_cast also works similarly on volatile, though that's less common.

dynamic_cast is exclusively used for handling polymorphism. You can cast a pointer or reference to any polymorphic type to any other class type (a polymorphic type has at least one virtual function, declared or inherited). You can use it for more than just casting downwards – you can cast sideways or even up another chain. The dynamic_cast will seek out the desired object and return it if possible. If it can't, it will return nullptr in the case of a pointer, or throw std::bad_cast in the case of a reference.

dynamic_cast has some limitations, though. It doesn't work if there are multiple objects of the same type in the inheritance hierarchy (the so-called 'dreaded diamond') and you aren't using virtual inheritance. It also can only go through public inheritance - it will always fail to travel through protected or private inheritance. This is rarely an issue, however, as such forms of inheritance are rare.

reinterpret_cast is the most dangerous cast, and should be used very sparingly. It turns one type directly into another — such as casting the value from one pointer to another, or storing a pointer in an int, or all sorts of other nasty things. Largely, the only guarantee you get with reinterpret_cast is that normally if you cast the result back to the original type, you will get the exact same value (but not if the intermediate type is smaller than the original type). There are a number of conversions that reinterpret_cast cannot do, too. It's used primarily for particularly weird conversions and bit manipulations, like turning a raw data stream into actual data, or storing data in the low bits of a pointer to aligned data.

C-style cast and function-style cast are casts using (type)object or type(object), respectively, and are functionally equivalent. They are defined as the first of the following which succeeds:

const_cast

static_cast (though ignoring access restrictions)

static_cast (see above), then const_cast

reinterpret_cast

reinterpret_cast, then const_cast

It can therefore be used as a replacement for other casts in some instances, but can be extremely dangerous because of the ability to devolve into a reinterpret_cast, and the latter should be preferred when explicit casting is needed, unless you are sure static_cast will succeed or reinterpret_cast will fail. Even then, consider the longer, more explicit option.

C-style casts also ignore access control when performing a static_cast, which means that they have the ability to perform an operation that no other cast can. This is mostly a kludge, though, and in my mind is just another reason to avoid C-style casts.


dynamic_cast is only for polymorphic types. you only need to use it when you're casting to a derived class. static_cast is certainly the first option unless you specifically need dynamic_cast's functinoality. It's not some miraculous silver-bullet "type-checking cast" in general.
Great answer! One quick remark: static_cast might be necessary to cast up the hierarchy in case you have a Derived*& to cast into Base*&, since double pointers/references don't automagically cast up the hierarchy. I came across such (frankly, not common) situation two minutes ago. ;-)
*"no other C++ cast is capable of removing const (not even reinterpret_cast)"... really? What about reinterpret_cast<int *>(reinterpret_cast<uintptr_t>(static_cast<int const *>(0)))?
I think an important detail missing above is that dynamic_cast has a run-time performance penalty compared to static or reinterpret_cast. This is important, e.g. in real-time software.
May be worth mentioning that reinterpret_cast is often the weapon of choice when dealing with an API's set of opaque data types
M
Mateen Ulhaq

Use dynamic_cast for converting pointers/references within an inheritance hierarchy.

Use static_cast for ordinary type conversions.

Use reinterpret_cast for low-level reinterpreting of bit patterns. Use with extreme caution.

Use const_cast for casting away const/volatile. Avoid this unless you are stuck using a const-incorrect API.


Be careful with dynamic_cast. It relies on RTTI and this will not work as expected across shared libraries boundaries. Simply because you build executable and shared library independly in there is no standardized way to sync RTTI across different builds. For this reason in Qt library there exists qobject_cast<> which uses the QObject type info for checking types.
R
RamblingMad

(A lot of theoretical and conceptual explanation has been given above)

Below are some of the practical examples when I used static_cast, dynamic_cast, const_cast, reinterpret_cast.

(Also referes this to understand the explaination : http://www.cplusplus.com/doc/tutorial/typecasting/)

static_cast :

OnEventData(void* pData)

{
  ......

  //  pData is a void* pData, 

  //  EventData is a structure e.g. 
  //  typedef struct _EventData {
  //  std::string id;
  //  std:: string remote_id;
  //  } EventData;

  // On Some Situation a void pointer *pData
  // has been static_casted as 
  // EventData* pointer 

  EventData *evtdata = static_cast<EventData*>(pData);
  .....
}

dynamic_cast :

void DebugLog::OnMessage(Message *msg)
{
    static DebugMsgData *debug;
    static XYZMsgData *xyz;

    if(debug = dynamic_cast<DebugMsgData*>(msg->pdata)){
        // debug message
    }
    else if(xyz = dynamic_cast<XYZMsgData*>(msg->pdata)){
        // xyz message
    }
    else/* if( ... )*/{
        // ...
    }
}

const_cast :

// *Passwd declared as a const

const unsigned char *Passwd


// on some situation it require to remove its constness

const_cast<unsigned char*>(Passwd)

reinterpret_cast :

typedef unsigned short uint16;

// Read Bytes returns that 2 bytes got read. 

bool ByteBuffer::ReadUInt16(uint16& val) {
  return ReadBytes(reinterpret_cast<char*>(&val), 2);
}

The theory of some of the other answers are good, but still confusing, seeing these examples after reading the other answers really makes them all make sense. That is without the examples, I was still unsure, but with them, I am now sure about what the other answers mean.
About the last usage of reinterpret_cast: isn't this the same as using static_cast<char*>(&val) ?
@LorenzoBelli Of course not. Did you try it? The latter is not valid C++ and blocks compilation. static_cast only works between types with defined conversions, visible relation by inheritance, or to/from void *. For everything else, there are other casts. reinterpret cast to any char * type is permitted to allow reading the representation of any object - and one of the only cases where that keyword is useful, not a rampant generator of implementation-/undefined behaviour. But this isn't considered a 'normal' conversion, so isn't allowed by the (usually) very conservative static_cast.
reinterpret_cast is pretty common when you are working with system software such as databases. Most cases you write your own page manager which has no idea about what is the data type stored in the page and just returns a void pointer. Its up to the higher levels to do a reinterpret cast and infer it as whatever they want.
The first example is dangerous, in that it assumes good behavior on the part of the caller (to always pass a pointer to an actual EventData object and nothing else). Unfortunately I don't think there's any practical way to type-check a void pointer in any meaningful way. Ideally the argument would be strongly-typed. Just some observations; not a critique of the answer.
A
Andrew Truckle

It might help if you know little bit of internals...

static_cast

C++ compiler already knows how to convert between scaler types such as float to int. Use static_cast for them.

When you ask compiler to convert from type A to B, static_cast calls B's constructor passing A as param. Alternatively, A could have a conversion operator (i.e. A::operator B()). If B doesn't have such constructor, or A doesn't have a conversion operator, then you get compile time error.

Cast from A* to B* always succeeds if A and B are in inheritance hierarchy (or void) otherwise you get compile error.

Gotcha: If you cast base pointer to derived pointer but if actual object is not really derived type then you don't get error. You get bad pointer and very likely a segfault at runtime. Same goes for A& to B&.

Gotcha: Cast from Derived to Base or viceversa creates new copy! For people coming from C#/Java, this can be a huge surprise because the result is basically a chopped off object created from Derived.

dynamic_cast

dynamic_cast uses runtime type information to figure out if cast is valid. For example, (Base*) to (Derived*) may fail if pointer is not actually of derived type.

This means, dynamic_cast is very expensive compared to static_cast!

For A* to B*, if cast is invalid then dynamic_cast will return nullptr.

For A& to B& if cast is invalid then dynamic_cast will throw bad_cast exception.

Unlike other casts, there is runtime overhead.

const_cast

While static_cast can do non-const to const it can't go other way around. The const_cast can do both ways.

One example where this comes handy is iterating through some container like set which only returns its elements as const to make sure you don't change its key. However if your intent is to modify object's non-key members then it should be ok. You can use const_cast to remove constness.

Another example is when you want to implement T& SomeClass::foo() as well as const T& SomeClass::foo() const. To avoid code duplication, you can apply const_cast to return value of one function from another.

reinterpret_cast

This basically says that take these bytes at this memory location and think of it as given object.

For example, you can load 4 bytes of float to 4 bytes of int to see how bits in float looks like.

Obviously, if data is not correct for the type, you may get segfault.

There is no runtime overhead for this cast.


I added the conversion operator information, but there are a few other things that should be fixed as well and I don't feel that comfortable updating this too much. Items are: 1. If you cast base pointer to derived pointer but if actual object is not really derived type then you don't get error. You get bad pointer and segfault at runtime. You get UB which may result in a segfault at runtime if you're lucky. 2. Dynamic casts can also be used in cross casting. 3. Const casts can result in UB in some cases. Using mutable may be a better choice to implement logical constness.
@Adrian you are correct in all count. The answer is written for folks at more or less beginner level and I didn't wanted to overwhelm them with all other complications that comes with mutable, cross casting etc.
@Shital Shah "Cast from Derived to Base or viceversa creates new copy! For people coming from C#/Java, this can be a huge surprise because the result is basically a chopped off object created from Derived." Could you please show a simple example code to make it easier to understand? Thanks.
What about std::bit_cast ?
t
talnicolas

Does this answer your question?

I have never used reinterpret_cast, and wonder whether running into a case that needs it isn't a smell of bad design. In the code base I work on dynamic_cast is used a lot. The difference with static_cast is that a dynamic_cast does runtime checking which may (safer) or may not (more overhead) be what you want (see msdn).


I have used reintrepret_cast for one purpose -- getting the bits out of a double (same size as long long on my platform).
reinterpret_cast is needed e.g. for working with COM objects. CoCreateInstance() has output parameter of type void** (the last parameter), in which you will pass your pointer declared as e.g. "INetFwPolicy2* pNetFwPolicy2". To do that, you need to write something like reinterpret_cast(&pNetFwPolicy2) .
Perhaps there is a different approach, but I use reinterpret_cast to extract pieces of data out of an array. For instance if I have a char* containing a big buffer full of packed binary data that I need to move through and get individual primitives of varying types. Something like this: template<class ValType> unsigned int readValFromAddress(char* addr, ValType& val) { /*On platforms other than x86(_64) this could do unaligned reads, which could be bad*/ val = (*(reinterpret_cast<ValType*>(addr))); return sizeof(ValType); }
I have never used reinterpret_cast, there aren't very many uses for it.
Personally I have only ever seen reinterpret_cast used for one reason. I've seen raw object data stored to a "blob" datatype in a database, then when the data is retrieved from the database, reinterpret_cast is used to turn this raw data into the object.
L
Lii

In addition to the other answers so far, here is unobvious example where static_cast is not sufficient so that reinterpret_cast is needed. Suppose there is a function which in an output parameter returns pointers to objects of different classes (which do not share a common base class). A real example of such function is CoCreateInstance() (see the last parameter, which is in fact void**). Suppose you request particular class of object from this function, so you know in advance the type for the pointer (which you often do for COM objects). In this case you cannot cast pointer to your pointer into void** with static_cast: you need reinterpret_cast<void**>(&yourPointer).

In code:

#include <windows.h>
#include <netfw.h>
.....
INetFwPolicy2* pNetFwPolicy2 = nullptr;
HRESULT hr = CoCreateInstance(__uuidof(NetFwPolicy2), nullptr,
    CLSCTX_INPROC_SERVER, __uuidof(INetFwPolicy2),
    //static_cast<void**>(&pNetFwPolicy2) would give a compile error
    reinterpret_cast<void**>(&pNetFwPolicy2) );

However, static_cast works for simple pointers (not pointers to pointers), so the above code can be rewritten to avoid reinterpret_cast (at a price of an extra variable) in the following way:

#include <windows.h>
#include <netfw.h>
.....
INetFwPolicy2* pNetFwPolicy2 = nullptr;
void* tmp = nullptr;
HRESULT hr = CoCreateInstance(__uuidof(NetFwPolicy2), nullptr,
    CLSCTX_INPROC_SERVER, __uuidof(INetFwPolicy2),
    &tmp );
pNetFwPolicy2 = static_cast<INetFwPolicy2*>(tmp);

Wouldn't it work something like &static_cast<void*>(pNetFwPolicy2) instead of static_cast<void**>(&pNetFwPolicy2)?
C
Ciro Santilli Путлер Капут 六四事

static_cast vs dynamic_cast vs reinterpret_cast internals view on a downcast/upcast

In this answer, I want to compare these three mechanisms on a concrete upcast/downcast example and analyze what happens to the underlying pointers/memory/assembly to give a concrete understanding of how they compare.

I believe that this will give a good intuition on how those casts are different:

static_cast: does one address offset at runtime (low runtime impact) and no safety checks that a downcast is correct.

dyanamic_cast: does the same address offset at runtime like static_cast, but also and an expensive safety check that a downcast is correct using RTTI. This safety check allows you to query if a base class pointer is of a given type at runtime by checking a return of nullptr which indicates an invalid downcast. Therefore, if your code is not able to check for that nullptr and take a valid non-abort action, you should just use static_cast instead of dynamic cast. If an abort is the only action your code can take, maybe you only want to enable the dynamic_cast in debug builds (-NDEBUG), and use static_cast otherwise, e.g. as done here, to not slow down your fast runs.

reinterpret_cast: does nothing at runtime, not even the address offset. The pointer must point exactly to the correct type, not even a base class works. You generally don't want this unless raw byte streams are involved.

Consider the following code example:

main.cpp

#include <iostream>

struct B1 {
    B1(int int_in_b1) : int_in_b1(int_in_b1) {}
    virtual ~B1() {}
    void f0() {}
    virtual int f1() { return 1; }
    int int_in_b1;
};

struct B2 {
    B2(int int_in_b2) : int_in_b2(int_in_b2) {}
    virtual ~B2() {}
    virtual int f2() { return 2; }
    int int_in_b2;
};

struct D : public B1, public B2 {
    D(int int_in_b1, int int_in_b2, int int_in_d)
        : B1(int_in_b1), B2(int_in_b2), int_in_d(int_in_d) {}
    void d() {}
    int f2() { return 3; }
    int int_in_d;
};

int main() {
    B2 *b2s[2];
    B2 b2{11};
    D *dp;
    D d{1, 2, 3};

    // The memory layout must support the virtual method call use case.
    b2s[0] = &b2;
    // An upcast is an implicit static_cast<>().
    b2s[1] = &d;
    std::cout << "&d           " << &d           << std::endl;
    std::cout << "b2s[0]       " << b2s[0]       << std::endl;
    std::cout << "b2s[1]       " << b2s[1]       << std::endl;
    std::cout << "b2s[0]->f2() " << b2s[0]->f2() << std::endl;
    std::cout << "b2s[1]->f2() " << b2s[1]->f2() << std::endl;

    // Now for some downcasts.

    // Cannot be done implicitly
    // error: invalid conversion from ‘B2*’ to ‘D*’ [-fpermissive]
    // dp = (b2s[0]);

    // Undefined behaviour to an unrelated memory address because this is a B2, not D.
    dp = static_cast<D*>(b2s[0]);
    std::cout << "static_cast<D*>(b2s[0])            " << dp           << std::endl;
    std::cout << "static_cast<D*>(b2s[0])->int_in_d  " << dp->int_in_d << std::endl;

    // OK
    dp = static_cast<D*>(b2s[1]);
    std::cout << "static_cast<D*>(b2s[1])            " << dp           << std::endl;
    std::cout << "static_cast<D*>(b2s[1])->int_in_d  " << dp->int_in_d << std::endl;

    // Segfault because dp is nullptr.
    dp = dynamic_cast<D*>(b2s[0]);
    std::cout << "dynamic_cast<D*>(b2s[0])           " << dp           << std::endl;
    //std::cout << "dynamic_cast<D*>(b2s[0])->int_in_d " << dp->int_in_d << std::endl;

    // OK
    dp = dynamic_cast<D*>(b2s[1]);
    std::cout << "dynamic_cast<D*>(b2s[1])           " << dp           << std::endl;
    std::cout << "dynamic_cast<D*>(b2s[1])->int_in_d " << dp->int_in_d << std::endl;

    // Undefined behaviour to an unrelated memory address because this
    // did not calculate the offset to get from B2* to D*.
    dp = reinterpret_cast<D*>(b2s[1]);
    std::cout << "reinterpret_cast<D*>(b2s[1])           " << dp           << std::endl;
    std::cout << "reinterpret_cast<D*>(b2s[1])->int_in_d " << dp->int_in_d << std::endl;
}

Compile, run and disassemble with:

g++ -ggdb3 -O0 -std=c++11 -Wall -Wextra -pedantic -o main.out main.cpp
setarch `uname -m` -R ./main.out
gdb -batch -ex "disassemble/rs main" main.out

where setarch is used to disable ASLR to make it easier to compare runs.

Possible output:

&d           0x7fffffffc930
b2s[0]       0x7fffffffc920
b2s[1]       0x7fffffffc940
b2s[0]->f2() 2
b2s[1]->f2() 3
static_cast<D*>(b2s[0])            0x7fffffffc910
static_cast<D*>(b2s[0])->int_in_d  1
static_cast<D*>(b2s[1])            0x7fffffffc930
static_cast<D*>(b2s[1])->int_in_d  3
dynamic_cast<D*>(b2s[0])           0
dynamic_cast<D*>(b2s[1])           0x7fffffffc930
dynamic_cast<D*>(b2s[1])->int_in_d 3
reinterpret_cast<D*>(b2s[1])           0x7fffffffc940
reinterpret_cast<D*>(b2s[1])->int_in_d 32767

Now, as mentioned at: https://en.wikipedia.org/wiki/Virtual_method_table in order to support the virtual method calls efficiently, supposing that the memory data structures of B1 is of form:

B1:
  +0: pointer to virtual method table of B1
  +4: value of int_in_b1

and B2 is of form:

B2:
  +0: pointer to virtual method table of B2
  +4: value of int_in_b2

then memory data structure of D has to look something like:

D:
  +0: pointer to virtual method table of D (for B1)
  +4: value of int_in_b1
  +8: pointer to virtual method table of D (for B2)
 +12: value of int_in_b2
 +16: value of int_in_d

The key fact is that the memory data structure of D contains inside it memory structure identical to that of B1 and B2, i.e.:

+0 looks exactly like a B1, with the B1 vtable for D followed by int_in_b1

+8 looks exactly like a B2, with the B2 vtable for D followed by int_in_b2

Therefore we reach the critical conclusion:

an upcast or downcast only needs to shift the pointer value by a value known at compile time

This way, when D gets passed to the base type array, the type cast actually calculates that offset and points something that looks exactly like a valid B2 in memory, except that this one has the vtable for D instead of B2, and therefore all virtual calls work transparently.

E.g.:

b2s[1] = &d;

simply needs to get the address of d + 8 to reach the corresponding B2-like data structure.

Now, we can finally get back to type casting and the analysis of our concrete example.

From the stdout output we see:

&d           0x7fffffffc930
b2s[1]       0x7fffffffc940

Therefore, the implicit static_cast done there did correctly calculate the offset from the full D data structure at 0x7fffffffc930 to the B2 like one which is at 0x7fffffffc940. We also infer that what lies between 0x7fffffffc930 and 0x7fffffffc940 is likely be the B1 data and vtable.

Then, on the downcast sections, it is now easy to understand how the invalid ones fail and why:

static_cast(b2s[0]) 0x7fffffffc910: the compiler just went up 0x10 at compile time bytes to try and go from a B2 to the containing D But because b2s[0] was not a D, it now points to an undefined memory region. The disassembly is: 49 dp = static_cast(b2s[0]); 0x0000000000000fc8 <+414>: 48 8b 45 d0 mov -0x30(%rbp),%rax 0x0000000000000fcc <+418>: 48 85 c0 test %rax,%rax 0x0000000000000fcf <+421>: 74 0a je 0xfdb 0x0000000000000fd1 <+423>: 48 8b 45 d0 mov -0x30(%rbp),%rax 0x0000000000000fd5 <+427>: 48 83 e8 10 sub $0x10,%rax 0x0000000000000fd9 <+431>: eb 05 jmp 0xfe0 0x0000000000000fdb <+433>: b8 00 00 00 00 mov $0x0,%eax 0x0000000000000fe0 <+438>: 48 89 45 98 mov %rax,-0x68(%rbp) so we see that GCC does: check if pointer is NULL, and if yes return NULL otherwise, subtract 0x10 from it to reach the D which does not exist

check if pointer is NULL, and if yes return NULL

otherwise, subtract 0x10 from it to reach the D which does not exist

dynamic_cast(b2s[0]) 0: C++ actually found that the cast was invalid and returned nullptr! There is no way this can be done at compile time, and we will confirm that from the disassembly: 59 dp = dynamic_cast(b2s[0]); 0x00000000000010ec <+706>: 48 8b 45 d0 mov -0x30(%rbp),%rax 0x00000000000010f0 <+710>: 48 85 c0 test %rax,%rax 0x00000000000010f3 <+713>: 74 1d je 0x1112 0x00000000000010f5 <+715>: b9 10 00 00 00 mov $0x10,%ecx 0x00000000000010fa <+720>: 48 8d 15 f7 0b 20 00 lea 0x200bf7(%rip),%rdx # 0x201cf8 <_ZTI1D> 0x0000000000001101 <+727>: 48 8d 35 28 0c 20 00 lea 0x200c28(%rip),%rsi # 0x201d30 <_ZTI2B2> 0x0000000000001108 <+734>: 48 89 c7 mov %rax,%rdi 0x000000000000110b <+737>: e8 c0 fb ff ff callq 0xcd0 <__dynamic_cast@plt> 0x0000000000001110 <+742>: eb 05 jmp 0x1117 0x0000000000001112 <+744>: b8 00 00 00 00 mov $0x0,%eax 0x0000000000001117 <+749>: 48 89 45 98 mov %rax,-0x68(%rbp) First there is a NULL check, and it returns NULL if th einput is NULL. Otherwise, it sets up some arguments in the RDX, RSI and RDI and calls __dynamic_cast. I don't have the patience to analyze this further now, but as others said, the only way for this to work is for __dynamic_cast to access some extra RTTI in-memory data structures that represent the class hierarchy. It must therefore start from the B2 entry for that table, then walk this class hierarchy until it finds that the vtable for a D typecast from b2s[0]. This is why dynamic cast is potentially expensive! Here is an example where a one liner patch converting a dynamic_cast to a static_cast in a complex project reduced runtime by 33%!.

reinterpret_cast(b2s[1]) 0x7fffffffc940 this one just believes us blindly: we said there is a D at address b2s[1], and the compiler does no offset calculations. But this is wrong, because D is actually at 0x7fffffffc930, what is at 0x7fffffffc940 is the B2-like structure inside D! So trash gets accessed. We can confirm this from the horrendous -O0 assembly that just moves the value around: 70 dp = reinterpret_cast(b2s[1]); 0x00000000000011fa <+976>: 48 8b 45 d8 mov -0x28(%rbp),%rax 0x00000000000011fe <+980>: 48 89 45 98 mov %rax,-0x68(%rbp)

Related questions:

When should static_cast, dynamic_cast, const_cast and reinterpret_cast be used?

How is dynamic_cast implemented

Downcasting using the 'static_cast' in C++

Tested on Ubuntu 18.04 amd64, GCC 7.4.0.


T
Timmy_A

While other answers nicely described all differences between C++ casts, I would like to add a short note why you should not use C-style casts (Type) var and Type(var).

For C++ beginners C-style casts look like being the superset operation over C++ casts (static_cast<>(), dynamic_cast<>(), const_cast<>(), reinterpret_cast<>()) and someone could prefer them over the C++ casts. In fact C-style cast is the superset and shorter to write.

The main problem of C-style casts is that they hide developer real intention of the cast. The C-style casts can do virtually all types of casting from normally safe casts done by static_cast<>() and dynamic_cast<>() to potentially dangerous casts like const_cast<>(), where const modifier can be removed so the const variables can be modified and reinterpret_cast<>() that can even reinterpret integer values to pointers.

Here is the sample.

int a=rand(); // Random number.

int* pa1=reinterpret_cast<int*>(a); // OK. Here developer clearly expressed he wanted to do this potentially dangerous operation.

int* pa2=static_cast<int*>(a); // Compiler error.
int* pa3=dynamic_cast<int*>(a); // Compiler error.

int* pa4=(int*) a; // OK. C-style cast can do such cast. The question is if it was intentional or developer just did some typo.

*pa4=5; // Program crashes.

The main reason why C++ casts were added to the language was to allow a developer to clarify his intentions - why he is going to do that cast. By using C-style casts which are perfectly valid in C++ you are making your code less readable and more error prone especially for other developers who didn't create your code. So to make your code more readable and explicit you should always prefer C++ casts over C-style casts.

Here is a short quote from Bjarne Stroustrup's (the author of C++) book The C++ Programming Language 4th edition - page 302.

This C-style cast is far more dangerous than the named conversion operators because the notation is harder to spot in a large program and the kind of conversion intended by the programmer is not explicit.


Upvoting due to referencing Stroustrup's quote. Hard to find these days especially that we often instead heard it from very smart people instead of the man himself.
p
pkthapa

To understand, let's consider below code snippet:

struct Foo{};
struct Bar{};

int main(int argc, char** argv)
{
    Foo* f = new Foo;

    Bar* b1 = f;                              // (1)
    Bar* b2 = static_cast<Bar*>(f);           // (2)
    Bar* b3 = dynamic_cast<Bar*>(f);          // (3)
    Bar* b4 = reinterpret_cast<Bar*>(f);      // (4)
    Bar* b5 = const_cast<Bar*>(f);            // (5)

    return 0;
}

Only line (4) compiles without error. Only reinterpret_cast can be used to convert a pointer to an object to a pointer to an any unrelated object type.

One this to be noted is: The dynamic_cast would fail at run-time, however on most compilers it will also fail to compile because there are no virtual functions in the struct of the pointer being casted, meaning dynamic_cast will work with only polymorphic class pointers.

When to use C++ cast:

Use static_cast as the equivalent of a C-style cast that does value conversion, or when we need to explicitly up-cast a pointer from a class to its superclass.

Use const_cast to remove the const qualifier.

Use reinterpret_cast to do unsafe conversions of pointer types to and from integer and other pointer types. Use this only if we know what we are doing and we understand the aliasing issues.


The provided snippet is a bad example. While I agree that, indeed, it compiles. The When listing is vaguely correct but mostly filled with opinions insufficient to fathom the required granularity.
A
Adrian

Nice feature of reinterpret_cast, not mentioned in the other answers, is that it allows us to create a sort of void* pointer for function types. Normally, for object types one uses static_cast to retrieve the original type of a pointer stored in void*:

  int i = 13;
  void *p = &i;
  auto *pi = static_cast<int*>(p);

For functions, we must use reinterpret_cast twice:

#include<iostream>

using any_fcn_ptr_t = void(*)();


void print(int i)
{
   std::cout << i <<std::endl;
}

int main()
{     
  //Create type-erased pointer to function:
  auto any_ptr = reinterpret_cast<any_fcn_ptr_t>(&print);
  
  //Retrieve the original pointer:
  auto ptr = reinterpret_cast< void(*)(int) >(any_ptr);
  
  ptr(7);
}

With reinterpret_cast we can even get a similar sort-of-void* pointer for pointers to member functions.

As with plain void* and static_cast, C++ guarantees that ptr points to print function (as long as we pass the correct type to reinterpret_cast).