I have a few projects that use boost::shared_ptr or std::shared_ptr extensively (I can convert to either implementation soon enough, if there is a good answer to this question for one, but not the other). The Boost implementation uses Boost.Assert to avoid returning in the case of encountering an empty (NULL) pointer in operator* or operator-> at runtime; while the libc++ implementation seems to lack any check.
While of course the validity of a shared_ptr should be checked before use, a large, mixed-paradigm codebase leads me to want to try an exception-throwing variation; as most of the code is relatively exception-aware and will at most fail to a high-level but resumable state, rather than std::terminate() or segfault.
How should I best customise these accessors while maintaining the robustness of shared_ptr? It seems that encapsulating shared_ptr in a throwing_shared_ptr may be the best option, but I’m wary of breaking the magic. Am I best off copying the Boost source and just changing the ASSERTs to an appropriate throw statement?
The actual type name used everywhere for the appropriate smart_ptr<T> type is a typedef expanded from a macro; i.e. ForwardDeclarePtr(Class) expands to something like:
class Class;
typedef boost::smart_ptr<Class> ClassPtr;
Everything passes, takes, or stores a ClassPtr – so I can replace the underlying type pretty freely; and I suspect this alleviates the potential slicing/hiding issue.
There’s really no “magic” in
std::shared_ptr<T>that would be removed if you wrapped it inside a custom class that would throw an exception when dereferencing aNULLshared pointer. So I don’t see why that approach wouldn’t work, as long as your new wrapper-class follows all the semantics of thestd::shared_ptr<T>type.BTW, you could also take a slightly different approach, and that is create a wrapper-class that simply won’t allow others to pass
NULLpointers to the wrappedstd::shared_ptr<T>data-member in the first-place. Basically it would be a class that would enforce thestd::make_shared<T>idiom in its constructor. I’m not sure, based on the workings of your code if this is possible, but it’s another way to circumvent the problem using a RAII approach rather than throwing exceptions.