The accurate number of digits is (sizeof(x) * 8 + PyLong_SHIFT - 1) / PyLong_SHIFT or (sizeof(x) * 8 - 1) / PyLong_SHIFT + 1.

There is a similar code in _PyLong_AsUnsignedLongLongMask().

Changes of Modules/_testcapimodule.c look not related to the implementation of _PyLong_AsUnsignedLongMask.

Needed some results of microbenchmarking. The patch speeds up _PyLong_AsUnsignedLongLongMask() for very large integers, but in most cases this function is called for not such larger integers, and the patch adds additional check. We should ensure that this doesn't slow down the function.