In NumPy we use `_Py_HashDouble` for our floating point hashing needs and that seems like a pretty good candiate for the unstable API to me: https://github.com/python/cpython/issues/91744#issuecomment-1111990057 and https://github.com/python/cpython/issues/101101#issue-1536645235. However, the function is currently removed rather than made unstable (to my understanding). While it should probably be easy enough to vendor what we needed, I would prefer the unstable API tier, it gives a better sense of having the required identical hashing. NumPy issue: https://github.com/numpy/numpy/issues/25035 _Originally reported by @ngoldbaum in https://github.com/python/cpython/issues/111418#issuecomment-1785929387_ <!-- gh-linked-prs --> ### Linked PRs * gh-112063 * gh-112095 * gh-112096 * gh-112098 * gh-112449 * gh-112476 * gh-112647 * gh-113115 <!-- /gh-linked-prs -->
