Inside STL: The shared_ptr constructor vs make_shared
There are two ways to create a new object that is controlled by a shared_ptr.
// From a raw pointer auto p = std::shared_ptr<S>(new S()); // Via make_shared auto p = std::make_shared<S>();
They result in two different memory layouts.
In the first case, you manually created a new S object, and then passed a pointer to it to the shared_ptr constructor. The shared_ptr adopts the raw pointer and creates a control block to monitor its lifetime. When the last shared pointer destructs, the Dispose() method deletes the pointer you passed in.¹ When the last shared or weak pointer destructs, the Delete() method deletes the control block.
| p | ||||
| object | • | → |
S
|
freed separately |
| control | • | → |
control_block
|
|
In the second case, you let the make_ function create the S object, and in practice, what it does is create a single memory allocation that consists of a control block stacked on top of an S object. This time, when the last shared pointer destructs, the Dispose() method runs the S destructor, but the memory isn’t freed yet. Only when the last shared or weak pointer destructs does the Delete() method get called to free the entire memory block.
| p | ||||
| object | • | control_block | freed as a unit | |
| ⤮ | ||||
| control | • | S | ||
The two memory layouts have their own pros and cons.
| Two allocations | Single allocation | |
|---|---|---|
| Last shared pointer destructs |
Object destructs Object memory freed |
Object destructs Object memory not freed |
| Last shared or weak pointer destructs |
Control block destructs Control block freed |
Control block destructs Combo block freed |
| Locality | Worse | Better |
| Straggler weak pointer | Control block lingers (Object memory freed already) |
Entire combo block lingers |
The single-allocation version has better memory locality since the control block is kept right next to the managed object.
On the other hand, with the single-allocation version, a straggler weak pointer (a weak pointer which lives for a long time after the last shared pointer has destructed) prevents the entire combo memory block from being freed. By comparison, only the control block remains in memory with the two-allocation version.
If your weak pointers exist to break circular references, then you won’t have stragglers because they will go away when the object graph destructs. Similarly, if your weak pointers are in event handlers, then those weak pointers won’t be stragglers if you are careful to unregister the event handlers at destruction. The stragglers come into play if you retain weak references in, say, a cache or other long-lifetime storage, and even then, they cause a problem only if sizeof(S) is significant or if you have a lot of them.
Next time, we’ll look at make_shared‘s close friend, std::.
¹ More specifically, the Deleter object deletes the pointer you passed in. The default deleter uses the delete operator to delete the pointer.
Light
Dark
7 comments
So std::make_shared is really only appropriate with small classes and classes that point to their content, like std::vector? News to me, I thought it was good practice to use it everywhere!
My third-to-last paragraph says the opposite! If you like the improved locality, and your weak pointers don’t outlive your strong pointers for long enough that the extra memory usage is concerning, then go ahead and reap the locality gains of the combined allocation.
It’s not only locality, but less overhead (heap memory block header of some sort for each allocation) and ultimately less heap fragmentation.
“[if] your weak pointers don’t outlive your strong pointers for long enough that the extra memory usage is concerning”
But if I wasn’t concerned, and I’m using a fat class, I wouldn’t be using weak_ptr … except in situations where it’s necessary to break an ownership cycle in a graph as you said. That’s the only situation I can think of where I’m not concerned. So appropriate with small classes… or when using weak_ptr to break the ownership cycle.
I think the gist of my statement is that it’s surprising that std::make_shared has this caveat, given its intended use. But it’s certainly *sometimes* a valid optimisation.
Not the first time that the standard library “forces” a not-always-valid optimization on you. Another is
std::vector<bool>.There are other scenarios beyond “breaking cycles” where you have weak pointers that will not outlive the main object by long, so “except for cycle-breaking” may be too strict.
For common use cases, this is typically less of a concern, but there are other performance implications to using std::make_shared. By allocating the control block in memory right before the object S, members of S might get “pushed” to another cache line. This might introduce a measurable performance drawback in the application.