<exception>: Consider changing exception_ptr to have the size of a single pointer

Hi!

I'm wondering if it's not too late to request an ABI breaking change to shrink the size of the STL's std::exception_ptr class type down to match that of a single pointer. Currently, it uses a std::shared_ptr to manage the lifetime of the stored exception object, and this naturally requires two pointer-sized variables for the std::shared_ptr's state. By changing the internal representation of std::exception_ptr over to an intrusive reference-counted pointer implementation, it would be possible to make sizeof(std::exception_ptr) == sizeof(void*).

The reason this would be valuable is it makes implementing both P1028 and P0709 much easier, in particular type-erasing a std::exception_ptr object and storing it in a std::system_code, which uses a single std::intptr_t for state. Both of these proposals, while well received, are still a long way off and may never land in their present form. However, there is an open source implementation and other libraries using the same technique (the company I work has its own implementation of P1028 used for exceptionless futures/promises and senders/receivers). With std::exception_ptr being larger than std::uintptr_t as it currently is in the Microsoft STL, the only way to do this type-erasure is to allocate additional memory, which would be nice to avoid inside of error handling code paths.

Both libc++ and libstdc++ appear to use the intrusive-reference counting approach for std::exception_ptr and have sizeof(std::exception_ptr) == sizeof(void*), so you would also be hitting parity with these libraries if the change were to be done.

vNext note: Resolving this issue will require breaking binary compatibility. We won't be able to accept pull requests for this issue until the vNext branch is available. See #169 for more information.

Start by reading the vNext note and issue #169, then compare the exception_ptr representation described here with the approaches used by libc++ and libstdc++. Done means evaluating the ABI-breaking size change for vNext; the issue states that pull requests cannot be accepted until that branch is available.