I have a small C++ program where I create two objects of a Person class. This class has char *m_szFirstName and char *m_szLastName for data.
I then I assign one object to another, causing both object's data members to point to same location.
In the destructor, I delete what memory I had allocated for the first object and assign NULL values to the pointers. Something like this.
if (m_szFirstName!= NULL)
{
delete [] m_szFirstName;
m_szFirstName = NULL;
}
Then when I go to delete memory for the second object, the check for NULL does not work and when I delete memory, I get a crash. From debugger, it show that my pointer is not NULL. It has 0xfeee.
I am aware that memory was already deleted before and should not be delete. However, I don't know how to check whether I should delete memory or not.
Reason for Crash:
You should follow the Rule of Three to avoid this problem of dangling pointers.
If you need to explicitly declare either the destructor, copy constructor or copy assignment operator yourself, you probably need to explicitly declare all three of them.
In your case You do not define a copy assignment operator thus leading to shallow copy of the pointer.
Suggested Solution:
If you can use std::string instead of char * just simply use std::string, it has the first and foremost preference over any kind of silly pointer stuff.
You can avoid all the funky pointer stuff by using std::string.
If you can't read on and the following suggestion applies to any class pointer member in general.
Note that the ideal solution here is to not use raw pointers at all, Whenever you use raw pointers you are forced to manually manage the resources acquired by them, it is always difficult and error prone to manually manage resources.So the onus is to avoid it.
To do so, You should use a Smart pointer which will manage the dynamic memory of the pointer implicitly.Using a smart pointer will ensure that the dynamic memory is implicitly released after usage & you do not have to manually manage it.
The scenario you have is the very reason that in C++ you should rely on RAII rather than manual resource management & using a Smart pointer is the way to go about in your case.
Caveat:
Note that I restrained myself from suggesting which smart pointer to use because the choice rather depends on the ownership and lifetime of the elements involved, which is not clear from the data provided in the Question.So I will suggest reading,
Which kind of pointer do I use when?
to make your choice of the smart pointer to use.
With
if (m_szFirstName!= NULL)
{
delete [] m_szFirstName;
m_szFirstName = NULL;
}
You only set m_szFirstName to point to NULL, not m_szLastName. This means you have to have some way to keep track of the fact they point to the same location. Is there a reason they point to the same location? Could you copy the name instead of pointing the pointers to the same place?
If you need the two pointers to shared the same data, I would have a look at std::tr1::shared_ptr, which will solve this issue for you by keeping track of the number of references and deleting when the number of references reaches 0.
Don't delete it again, if (m_szFirstName == m_szLastName).
But this will give you a memory leak (when you assign one pointer to other).
When you have two pointers pointing to the same location (after you assigned the first one to the second one), you have two pointers pointing at the same address. Deleting one frees the memory pointed to by both of them. But setting one to NULL doesn't alter the other pointer. The same happens if you have two integers, for example.
int a = 3;
int b = a;
Now if you run
a = 0;
the value of b doesn't change. The same way your second pointer doesn't change when you alter the first one (but when you alter the memory pointed to by either pointer, you can see the effect through the other as well).
Your problem is a classic C/C++ problem which is known as "Dangling Pointer" problem. Dereferencing the dangling pointer has resulted in the crash. The problem is about reference counting. Once you assign the same memory to the second pointer then the reference count should be 2. So if you delete one pointer reference count should become 1 and your memory should not be deallocated or freed unless count is 0. At 0 it can be garbage collected.
Now there are good answers above to solve your problem. Since you are using C++ so you can use something like an auto_ptr (OR shared_ptr). They provide what I had mentioned above, the reference counting stuff and even you need not worry about deleting or freeing your objects, these classes would take care. They work on simething known as RAII pattern where a destructor is auto-magically called when the object on the stack goes out of scope.
Just stop setting pointers to NULL when you have deleted the object. As you can see, it just leads to pain. You cannot assume that because the pointer is not-NULL, it has not been deleted already.
You can use any sensible pattern you want to avoid this problem. For example, Boost's shared_ptr is a great choice.
Related
Is there any way to distinguish two following situations at run time:
double ptr * = new double(3.14159);
double variable = 3.14159
double * testPtr_1 = ptr;
double * testPtr_2 = &variable;
delete testPtr_1 // fine...
delete testPtr_2 // BIG RUN TIME ERROR !!!
I have find myself in situation in with I need to call delete operator for some unknown pointer. The pointer can point to anywhere (to a "local" variable or to dynamically allocated variable).
How can I find out where my "unknown" pointer points, and therefore choose when to and when not to call operator delete on it
EDIT:
Ok I see that everyone is pointing me to the smart pointers, but what if I am trying to write my own set of smart pointers (that is The reason behind my question)?
There is no way to test if a pointer is pointing to a memory area that would be valid to delete. Moreover,
There is no way to tell between pointers that must be freed with delete vs. delete[],
There is no way to tell between the pointers that have been freed and pointers that have not been freed,
There is no way to tell among pointers to an automatic variable, pointers to static variable, and pointers to dynamically allocated blocks.
The approach that you should take is tracking allocations/deallocations by some other means, such as storing flags along with your pointers. However, this is rather tedious: a much better practice is to switch to smart pointers, which would track resources for you.
You need to set some better coding practices for yourself (or for your project).
Especially since most platforms have, at the very least, a C++11-compliant compiler, there's no reason not to be using the following paradigm:
Raw Pointers (T*) should ONLY be used as non-owning pointers. If you receive a T* as the input for a function or constructor, you should assume you have no responsibility for deleting it. If you have an instance or local variable that is a T*, you should assume you have no responsibility for deleting it.
Unique Pointers (std::unique_ptr<T>) should be used as single-ownership pointers, and in general, these should be your default go-to choice for any situation where you need to dynamically allocate memory. std::make_unique<T>() should be preferred for creating any kind of Unique Pointer, as this prevents you from ever seeing the raw pointer in use, and it prevents issues like you described in your original post.
Shared Pointers (std::shared_ptr<T> and std::weak_ptr<T>) should ONLY be used in situations where it is logically correct to have multiple owners of an object. These situations occur less often than you think, by the way! std::make_shared<T>() is the preferred method of creating Shared Pointers, for the same reasons as std::make_unique, and also because std::make_shared can perform some optimizations on the allocations, improving performance.
Vectors (std::vector<T>) should be used in situations where you need to allocate multiple objects into heap space, the same as if you called new T[size]. There's no reason to use pointers at all except in very exotic situations.
It should go without saying that you need to take my rules of "ONLY do 'x'" with a grain of salt: Occasionally, you will have to break those rules, and you might be in a situation where you need a different set of rules. But for 99% of use-cases, those rules are correct and will best convey the semantics you need to prevent memory leaks and properly reason about the behavior of your code.
You cannot.
Avoid raw pointers and use smart pointers, particularly std::unique_ptr. It conveys clearly who is responsible for deleting the object, and the object will be deleted when the std::unique_ptr goes out of scope.
When creating objects, avoid using new. Wrap them in a smart pointer directly and do not take addresses of anything to wrap it in a smart pointer. This way, all raw pointers will never need freeing and all smart pointers will get cleaned up properly when their time has come.
Okay, some things you can distinguish in a very platform-specific, implementation-defined manner. I won’t go into details here, because it’s essentially insane to do (and, again, depends on the platform and implementation), but you are asking for it.
Distinguish local, global and heap variables. This is actually possible on many modern architectures, simply because those three are different ranges of the address space. Global variables live in the data section (as defined by the linker and run-time loader), local variables on the stack (usually at the end of the address space) and heap variables live in memory obtained during run-time (usually not at the end of the address space and of course not overlapping the data and code sections, a.k.a. "mostly everything else"). The memory allocator knows which range that is and can tell you details about the blocks in there, see below.
Detect already-freed variables: you can ask the memory allocator that, possibly by inspecting its state. You can even find out when a pointer points into a allocated region and then find out the block to which it belongs. This is however probably computationally expensive to do.
Distinguishing heap and stack is a bit tricky. If your stack grows large and your program is running long and some piece of heap has been returned to the OS, it is possible that an address which formerly belonged to the heap now belongs to the stack (and the opposite may be possible too). So as I mentioned, it is insane to do this.
You can't reliably. This is why owning raw pointers are dangerous, they do not couple the lifetime to the pointer but instead leave it up to you the programmers to know all the things that could happen and prepare for them all.
This is why we have smart pointers now. These pointers couple the life time to the pointer which means the pointer is only deleted once it is no longer in use anywhere. This makes dealing with pointer much more manageable.
The cpp core guildlines suggests that a raw pointer should never be deleted as it is just a view. You are just using it like a reference and it's lifetime is managed by something else.
Ok I see that everyone is pointing me to the smart pointers, but what if I am trying to write my own set of smart pointers (that is The reason behind my question)?
In that case do like the standard smart pointers do and take a deleter which you default to just using delete. That way if the user of the class wants to pass in a pointer to a stack object they can specify a do nothing deleter and you smart pointer will use that and, well, do nothing. This puts the onus on the person using the smart pointer to tell the pointer how to delete what it points to. Normally they will never need to use something other than the default but if they happen to use a custom allocator and need to use a custom deallocator they can do so using this method.
Actually you can. But memory overhead occurs.
You overload new and delete operator and then keep track of allocations and store it somewhere(void *)
#include<iostream>
#include<algorithm>
using namespace std;
void** memoryTrack=(void **)malloc(sizeof(void *)*100); //This will store address of newly allocated memory by new operator(really malloc)
int cnt=0;//just to count
//New operator overloaded
void *operator new( size_t stAllocateBlock ) {
cout<<"in new";
void *ptr = malloc(stAllocateBlock); //Allocate memory using malloc
memoryTrack[cnt] = ptr;//Store it in our memoryTrack
cnt++; //Increment counter
return ptr; //return address generated by malloc
}
void display()
{
for(int i=0;i<cnt;i++)
cout<<memoryTrack[i]<<endl;
}
int main()
{
double *ptr = new double(3.14159);
double variable = 3.14159;
double * testPtr_1 = ptr;
double * testPtr_2 = &variable;
delete testPtr_1; // fine...
delete testPtr_2;
return 0;
}
Now the most important function(You will have to work on this because it is not complete)
void operator delete( void *pvMem )
{
//Just printing the address to be searched in our memoryTrack
cout<<pvMem<<endl;
//If found free the memory
if(find(memoryTrack,memoryTrack+cnt,pvMem)!=memoryTrack+cnt)
{
//cout<<*(find(memoryTrack,memoryTrack+cnt,pvMem));
cout<<"Can be deleted\n";
free (pvMem);
//After this make that location of memoryTrack as NULL
//Also keep track of indices that are NULL
//So that you can insert next address there
//Or better yet implement linked list(Sorry was too lazy to do)
}
else
cout<<"Don't delete memory that was not allocated by you\n";
}
Output
in new
0xde1360
0xde1360
Can be deleted
0xde1360
0x7ffe4fa33f08
Dont delete memory that was not allocated by you
0xde1360
Important Node
This is just basics and just code to get you started
Open for others to edit and make necessary changes/optimization
Cannot use STL, they use new operator(if some can implement them please do,it would help to reduce and optimize the code)
I'd like to illustrate one thing I've seen:
class Something {};
void do_something(const Something *p_sth) {}
Then:
do_something(new Something());
should cause a memory leak because when you call new you should also always call delete, right? Would this be a good solution?
void do_something(const Something *p_sth)
{
delete p_sth;
}
Or is it better to use references &? I also find out that smart pointers can solve this, so delete isn't required (it seems as a good thing but I have never used it before). I just want to know what's the best solution for this to avoid the memory leak. Thanks
*Thank you all for your answers. It helped me to clear up few things. I'm also sorry for the code I posted as it was maybe too general.
Your suggestion of assuming the ownership of the pointer and deleting the object has problems.
That behaviour is not idiomatic in c++. Programmers expect that they must delete every object that they allocate with new. If a user of your function expects that they're responsible for deleting the object whose address they pass to the function, then your solution breaks apart. Also, it prevents using your function with objects that must keep existing after the call has ended:
Something* s = new s();
do_something(s);
s.foo() // oops, object was deleted, I can't use it anymore
delete s; // oops, double delete because I'm for some reason not responsible for deleting the object that I allocated
Your solution also prevents using automatically and statically allocated objects with the function.
Something s;
do_something(&s); //oops, do_something assumes that object is dynamically allocated
All of these caveats would have to be documented to the user of the function.
A raw pointer without deleting inside the function has none of these problems. Managing the callers memory should really not be the responsibility of your function. Doing that would break the single responsibility principle. There's nothing wrong with raw pointer parameters when you don't want to transfer or share ownership. If you do want to imply changes in ownership, then use smart pointers which were designed exactly for that.
Smart pointers don't have some of the above problems, but they still prevent using the function with automatic and static objects.
A reference parameter is in many cases ideal. It communicates the caller that they're still responsible for the object. As a bonus, The lack of need for addressof operator allows slightly nicer syntax. Sure, the caller may still forget to manage their memory, but as I pointed out, that shouldn't be your responsibility.
References have one potential drawback. They can't be null. If you don't need null, then it's actually an advantage.
Which solution is ideal, depends on what you need. Following is not true for all corner cases, but should hold for most common cases:
If you want to modify the object, then pass a reference.
Unless you need null, in which case use a pointer
If you just want to read the object, then
If the object is copyable, small (size less than or equal to word), doesn't cointain pointers to dynamic objects and not polymorphic, then pass by value
Otherwise or if you don't know those things because you're writing a template, pass a const reference
Unless you need null, in which case use a pointer
If you want to If you want to transfer ownership, then use unique_ptr
If you want that ownership to be shared, then use shared_ptr
Best is to use a smart pointer
class Something {};
void do_something(std::shared_ptr<Something> p_sth)
{
...
}
That way the ownership is clear when you look at the prototype as well as you get an automatic delete when you leave scope.
I just want to know what's the best solution for this to avoid the memory leak. Thanks
The best solution to ensure there are no memory leaks would be to use std::shared_ptrwhich is a smart pointer which, as soon as it does not have any references, deletes itself. This pointer is best if you have more than on reference to the same pointer. Use std::unique_ptr if there is only one reference at a given time.
This will prevent memory leaks and I also prefer using smart pointers rather than standard pointers as they are very powerful. Also, retaining to the question:
Or is it better to use references &?
Use references wherever you need to reference the object, if you delete the reference, being a smart pointer, it will delete the pointer as well (Being there are no other references to that pointer)
I hope this answers your questions :)
Prefer to avoid the use of raw pointers.
If you do not need a heap allocation and can use a stack allocated variable, then prefer to pass by reference (or even pass by value if an appropriate move constructor is in place, but that's a topic for another day).
If you need a heap allocation (for polymorphism, dependency injection, information hiding, transfer of ownership etc.) then determine what the ownership semantics will be for that object and use the appropriate smart pointer to manage those semantics.
If all else fails and you must use a raw pointer (perhaps you are dealing with a legacy interface or a C interface, or something similar) then again, clearly define your ownership semantics and document them to make it clear who is responsible for deleting the heap allocated object, etc.
If you must have a raw pointer to a stack allocated object then document that you will not be deleting the pointer and be careful about documenting the lifetime of the pointer to avoid accessing the object after it has gone out of scope.
Cleanest solution would be to avoid the pointer completely.
And yes, a reference is a good idea.
void do_something(const Something &p_sth) {}
Then you can declare your 'something' on the Stack.
void function_that_does_something()
{
Something mySomething;
do_something( mySomething );
}
It is always best to let the compiler do the cleanup for you.
I always wondered why automatic setting of the pointer to NULL after delete is not part of the standard. If this gets taken care of then many of the crashes due to an invalid pointer would not occur. But having said that I can think of couple of reasons why the standard would have restricted this:
Performance:
An additional instruction could slow down the delete performance.
Could it be because of const pointers.
Then again standard could have done something for this special case I guess.
Does anyone know exact reasons for not allowing this?
Stroustrup himself answers. An excerpt:
C++ explicitly allows an
implementation of delete to zero out
an lvalue operand, and I had hoped
that implementations would do that,
but that idea doesn't seem to have
become popular with implementers.
But the main issue he raises is that delete's argument need not be an lvalue.
First, setting to null would require a memory stored variable. It's true, that you usually have a pointer in a variable but sometimes you might want to delete an object at a just calculated address. That would be impossible with "nullifying" delete.
Then comes performance. You might have written code in such a way that the pointer will go out of scope immediately after delete is done. Filling it with null is just a waste of time. And C++ is a language with "don't need it? then you don't have to pay for it" ideology.
If you need safety there's a wide range of smart pointers at you service or you can write your own - better and smarter.
You can have multiple pointers pointing to that memory. It would create a false sense of security if the pointer you specified for the delete got set to null, but all the other pointers did not. A pointer is nothing more than an address, a number. It might as well be an int with a dereference operation. My point is you would have to also scan every single pointer to find those that are referencing the same memory you just deleted, and null them out as well. It would be computationally intense to scan all the pointers for that address and null them out, because the language is not designed for that. (Although some other languages structure their references to accomplish a similar goal in a different way.)
A pointer can be saved in more than one variable, setting one of these to NULL would still leave invalid pointers in the other variables. So you don't really gain much, you are more likely creating a false sense of security.
Besides of that, you can create your own function that does what you want:
template<typename T>
void deleten(T *&ptr) {
delete ptr;
ptr = NULL;
}
Because there isn't really any need to, and because it would require delete taking pointer-to-pointer rather than just pointer.
delete is used mostly in destructors, in which case setting a member to NULL is pointless. A few lines later, at the closing }, the member no longer exists. In assignment operators, a delete is typically followed by an assignment anyway.
Also, it would render the following code illegal:
T* const foo = new T;
delete foo;
Here's another reason; suppose delete does set its argument to NULL:
int *foo = new int;
int *bar = foo;
delete foo;
Should bar get set to NULL? Can you generalize this?
If you have an array of pointers, and your second action is to delete the empty array, then there is no point setting each value to null when the memory is about to be freed. If you want it to be null.. write null to it :)
C++ allows you to define your own operator new and delete so that for instance they would use your own pool allocator. If you do this then it is possible to use new and delete with things that are not strictly addresses but say indexes in your pool array. In this context the value of NULL (0) might have a legal meaning (referring to the first item in the pool).
So having delete set NULL automatically to its argument doesn't always have the meaning of - set the value to an invalid value. The invalid value may not always be NULL.
Philosophy of C++ is "pay for it only if you use it". I think it may answer your question.
Also sometimes you could have your own heap which will recover deleted memory.. or sometimes pointer not owned by any variable. Or pointer stored in few variables - it possible zero just one of them.
As you can see it have many issues and possible problems.
Setting the pointer to NULL automatically would not solve most of the issues with bad pointer usage. The only crash it would avoid is if you try to delete it twice. What if you call a member function on such a pointer? It would still crash (assuming that it accesses member variables). C++ does not restrict you from calling any function on NULL pointers, nor should it do that from performance point of view.
I see people giving weird answers to this question.
ptr = NULL;
How can such a simple statement cause performance delay?
Another answer is saying that we can have multiple pointers pointing to the same
memory location. Surely we can. In this case delete operation on one pointer would make only that pointer NULL (if delete was making pointer NULL) and the other pointer would be non-NULL and pointing to memory location which is free.
The solution for this should have been that user should delete all pointers pointing to same location. Internally it should check if memory is already freed than don't free. Only make the pointer NULL.
Stroustrup could have designed delete to work in this manner. He thought programmers would take care of this. So he ignored.
Below is an exception proof approach of assignment operator shared by Scott Meyer. Can anybody let me know that is it safe to delete the raw pointer?
int *orig =m_p;
m_p=new int (*obj.m_p);
delete orig;
The problem with the question is that you can't tell what m_p contains initially. If its value has been obtained by some &intVar, then the answer is: NO. If the value has been produced by new but has already been deleted and not set to NULL, then the answer is NO again. If the value is undefined, the answer is another NO.
That leaves a possibilitity where it is safe.
If you cannot do without a raw pointer
Yes you have to delete it and it is safe. You can do one more thing, which is setting it to 0 after deleting it because Setting them to 0 after deleting is a good practice, except in destructors
(because the object instance that contains the pointers is going away anyway). The standard allows deletes of NULL
pointers (it does nothing), and any good OS with memory management will cause your program to crash on the
first dereference (read or write) of a NULL pointer. So now it's up to your requirements that how you go for it
If you can do without a raw pointer
Raw pointer is not recommended to be used, there are other alternatives which you could go for like you can Use a managed pointer class instead of raw pointers. std::auto_ptr<>, boost::scoped_ptr<>, boost::shared_ptr<>,
boost::weak_ptr<>, boost::intrusive_ptr<>, std::unique_ptr<> (C++0x). Even on multi-million line projects with tons of memory allocation this is used and even they need to very rarely do delete anything; all the pointers are managed.
If I have a pointer pointing to a specific memory address at the heap. I would like this same pointer to point to another memory address, should I first delete the pointer? But, in this case am I actually deleting the pointer or just breaking the reference (memory address) the pointer is pointing at?
So, in other words, if I delete a pointer, does this mean it doesn't exist any more? Or, it is there, but not pointing to where it was?
The syntax of delete is a bit misleading. When you write
T* ptr = /* ... */
delete ptr;
You are not deleting the variable ptr. Instead, you are deleting the object that ptr points at. The value of ptr is unchanged and it still points where it used to, so you should be sure not to dereference it without first reassigning it.
There is no requirement that you delete a pointer before you reassign it. However, you should ensure that if you are about to reassign a pointer in a way that causes you to lose your last reference to the object being pointed at (for example, if this pointer is the only pointer in the program to its pointee), then you should delete it to ensure that you don't leak memory.
One technique many C++ programmers use to simplify the logic for when to free memory is to use smart pointers, objects that overload the operators necessary to mimic a pointer and that have custom code that executes automatically to help keep track of resources. The new C++0x standard, for example, will provide a shared_ptr and unique_ptr type for this purpose. shared_ptr acts like a regular pointer, except that it keeps track of how many shared_ptrs there are to a resource. When the last shared_ptr to a resource changes where it's pointing (either by being reassigned or by being destroyed), it then frees the resource. For example:
{
shared_ptr<int> myPtr(new int);
*myPtr = 137;
{
shared_ptr<int> myOtherPtr = myPtr;
*myPtr = 42;
}
}
Notice that nowhere in this code is there a call to delete to match the call to new! This is because the shared_ptr is smart enough to notice when the last pointer stops pointing to the resource.
There are a few idiosyncrasies to be aware of when using smart pointers, but they're well worth the time investment to learn about. You can write much cleaner code once you understand how they work.
When you delete a pointer you release the memory allocated to the pointed to object. So if you just want your pointer to point to a new memory location you should not delete the pointer. But if you want to destroy the object currently it is pointing to and then then point to a different object then you should delete the pointer.
xtofl, while being funny, is a bit correct.
If YOU 'new' a memory address, then you should delete it, otherwise leave it alone. Maybe you are thinking too much about it, but you think about it like this. Yea, the memory is always there, but if you put a fence around it, you need to take the fence down, or no one else can you it.
When you call delete you mark the memory pointed to by the pointer as free - the heap takes ownership of it and can reuse it, that's all, the pointer itself is usually unchanged.
What to do with the pointer depends on what you want. If you no longer need that memory block - use delete to free the block. If you need it later - store the address somewhere where you can retrieve it later.
To answer your question directly, this has been asked before. delete will delete what your pointer points to, but there was a suggestion by Bjarne Stroustrup that the value of the pointer itself can no longer be relied upon, particularly if it is an l-value. However that does not affect the ability to reassign it so this would be valid:
for( p = first; p != last; ++p )
{
delete p;
}
if you are iterating over an array of pointers, all of which had been allocated with new.
Memory management in C++ is best done with a technique called RAII, "resource acquisition is initialization".
What that actually means is that at the time you allocate the resource you immediately take care of its lifetime, i.e. you "manage" it by putting it inside some object that will delete it for you when it's no longer required.
shared_ptr is a technique commonly used where the resource will be used in many places and you do not know for certain which will be the last one to "release" it, i.e. no longer require it.
shared_ptr is often used in other places simply for its semantics, i.e. you can copy and assign them easily enough.
There are other memory management smart pointers, in particular std::auto_ptr which will be superceded by unique_ptr, and there is also scoped_ptr. weak_ptr is a methodology to be able to obtain a shared_ptr if one exists somewhere, but not holding a reference yourself. You call "lock()" which gives you a shared_ptr to memory or a NULL one if all the current shared pointers have gone.
For arrays, you would not normally use a smart pointer but simply use vector.
For strings you would normally use the string class rather than think of it as a vector of char.
In short, you do not "delete a pointer", you delete whatever the pointer points to.
This is a classical problem: If you delete it, and someone else is pointing to it, they will read garbage (and most likely crash your application). On the other hand, if you do not and this was the last pointer, your application will leak memory.
In addition, a pointer may point to thing that were not originally allocated by "new", e.g. a static variable, an object on the stack, or into the middle of another object. In all those cases you're not allowed to delete whatever the pointer points to.
Typically, when designing an application, you (yes, you) have to decide which part of the application owns a specific object. It, and only it, should delete the objects when it is done with it.