I pushed a fix for the test failure that I caused by getting greedy with the pointer() => nullptr cleanup. It revealed a pre-existing bug in the old (VS 2010-era) Dev10_722102_shared_ptr_nullptr test. That contained a fancy_pointer being given to unique_ptr which totally failed to meet the Cpp17NullablePointer requirements. Fixes:

• Include <cstddef> for nullptr_t (the test is already using namespace std;) and change the other headers for consistency.
• Cpp17NullablePointer requires implicit construction from nullptr, which previously wouldn't compile. Add a constructor and a data member initializer T* ptr_{nullptr};.
• We also need to provide assignment from nullptr - I can never remember whether this will invoke implicit construction and whether that's hazardous in corner cases, so I added an assignment operator for simplicity.
• We also need to be equality/inequality comparable, with other fancy pointers and nullptr. Provide all forms (not bothering to drop any for C++20 mode).

With this, the test should pass, and I believe it meets the requirements now.

• Cpp17NullablePointer requires value-initialized pointers to be null, but instead this was garbage. Fix this with a data member initializer: T* ptr_{nullptr};

Value-init of this class previously resulted in zero-initialization since the default constructor was not user-provided. (I don't know if there's a name for this "trivial default-init but deterministic value-init" idiom.)

• We also need to provide assignment from nullptr - I can never remember whether this will invoke implicit construction and whether that's hazardous in corner cases, so I added an assignment operator for simplicity.

Implicit construction, yes, and the most common failure is trying to assign convertible to convertible to X, violating the "at most one user-defined conversion" rule.

Value-init of this class previously resulted in zero-initialization since the default constructor was not user-provided.

Ah, you are indeed correct. I updated my description accordingly, thanks. (Yay, the code was only broken in N - 1 ways.)

Thanks for improving vector performance - the STL's most important data structure! This should have the bonus effect of increasing compiler throughput (as it was reported to us by @xiangfan-ms on the compiler front-end team). 🚀 🚀