Pranesh Srinivasan

Data. Hacks. Opinions.


C++11 introduces unique_ptr a wonderful way to manage the most common pointer use cases without losing sanity. While any pointer can point only to one object at a time, the unique_ptr construct further enforces only one pointer to the object at any point of time; i.e, a bijective mapping between pointers and objects. For example,

std::unique_ptr<Foo> foo_ptr(new Foo());

is equivalent to saying that foo_ptr is the only pointer that points to the object we passed into the constructor. The dereference and member access methods are overriden so that foo_ptr looks like, sounds like and behaves like a normal pointer. One can say *foo_ptr and foo_ptr->method() as expected. Also being consistent with denoting ownership, a unique_ptr, it cannot be copied, only moved.1

In short, a unique_ptr is

  • cheap since it has very minimal overhead
  • lightweight as it does not need to reference count (an object can only have one valid unique pointer at any point of time)
  • and avoids memory leaks since it automatically frees the owned raw pointer when it goes out of scope via the destructor.

Even more useful, I think, is that they constrain ownership thereby avoiding the reader to reference count2 as she reads through the code.

Return values work as expected

The trick to using, understanding and writing effective code with unique_ptrs is to realize that a lot of moving happens under the hood. Copying or assigning a unique pointer is disallowed, but (roughly speaking), the new C++11 move semantics automatically kick in for rvalue copys.

As a result, returning unique pointers work as expected since C++ move semantics converts the rvalue copy into a move:

unique_ptr<Foo> FooFactory(ConfigObject C) {
      unique_ptr<Foo> foo_ptr;
      foo_ptr.reset(new Foo());
      // read config object
      // set desired properties on Foo
      return foo;

int main() {
  unique_ptr<Foo> foo_ptr = FooFactory(ConfigObject()); // Works!

Ownership is made explicit

When a function accepts a pointer, questions of ownership arise. For instance, imagine a function

void ColorifyFoo(Foo* uncolored_foo);

Without peeking into the definition of ColorifyFoo, it is difficult to say if ColorifyFoo were to own the pointer. It gets even more difficult to predict if this were a class:

class FooColorifier{
        FooColorifier(Foo *uncolored_foo); // Is uncolored_foo owned?

It is almost impossible to tell whether the FooColorifier class owns foo without looking into what the class does with it. This is where unique pointers shine. They explicitly prohibit such call sites! Trying to call a unique_pointer version of the above function:

void ColorifyFoo(unique_ptr<Foo> uncolored_foo);

in a manner like the following:

unique_ptr<Foo> foo_ptr(new Foo());
ColorifyFoo(foo_ptr);  // compiler error!

would throw up a compiler error! One would have to explicitly move the pointer by calling it such:

ColorifyFoo(move(foo_ptr));  // works!
// now foo_ptr is NULL

This makes ownership clear to the reader! Of course, if you only wanted to pass a reference to ColorifyFoo so that the pointer is not moved, you should, as you guessed right, pass a reference:

void ReferenceVersionOfColorifyFoo(const Foo& uncolored_foo);

// callsite like the following:
unique_ptr<Foo> foo_ptr(new Foo());

Usage with containers

Allowing containers to use pointers to their data rather than objects themselves can significantly increase the speed of move-heavy operations like sorting3. However, just using containers with raw pointers is unsafe in terms of memory. For instance, passing said containers around leads to keeping track of who owns the container to appropriately manage said memory.

Due to rvalue references and move semantics, C++11 enables unique_ptr to be effortlessly used with containers. This is where they go above and beyond auto_ptr. Their usage with containers is as expected:

vector<unique_ptr<Foo>> vector_of_foos;
unique_ptr<Foo> foo_ptr = new Foo();

vector_of_foos.push_back(foo_ptr);  // Throws compiler error!

vector_of_foos.push_back(move(foo_ptr));  // Works!
vector_of_foos.emplace_back(new Foo());  // Also works!

Move-heavy operations like sort, shuffle and reverse can really benefit (sometimes even upto 10x) via a judicious use of unique_ptr.

The headfake

In reality, the unique_ptr is actually a very good headfake. Sure, they help manage memory, but their real benefit is for the reader4! Consistent usage of unique_ptr makes ownership extremely clear.

Pointers have historically been semantically overloaded to handle both ownership as well as cheap read copies. Separating the two use cases by the usage of unique_ptr5 to serve ownership and the usage of const references to allow for cheap read copies removes the ambiguity and improves programmer-read time. Programmer-read time is, of course, the most important time constraint of all!

  1. None of this is radically new. auto_ptr was the first attempt at thinking about RIAA pointer semantics, but it did not work well with containers. C++11 fixes this with the addition of move semantics.
  2. I’ve often joked that readers of programs have to have execution models simultaneously operating in their heads. With programming languages that do not automatically manage memory such as C / C++, they also need to reference count pointers and references. IMO, this is the source of the increased programmer productivity that Spolsky refers to in his API War post.
  3. And by a smaller amount, even lookup operations like searches due to cache coherence.
  4. One of the most frequently recurring readers is often a future version of the author!
  5. And it’s sibling shared_ptr for the multiple ownership case.