-, |, and ^ of dictkeys are implemented as: PyObject *result = PySet_New(self); // Call set.difference_update, set.update, set.symmetric_difference_update with other. PySet_New(iterable) has optimized step for iterable is dict. But since iterable is dictkeys, PyIter_Next() is called for all elements in the dict. We can pass dict instead of dictkey object to PySet_New. $ ./python -m pyperf timeit -o patched.json -s 'k = dict.fromkeys("abcdefghijklmnopqrstuvwxyz").keys(); s={1,2,3}' -- 'k | {1,2,3}' $ ./python -m pyperf compare_to master.json patched.json Mean +- std dev: [master] 778 ns +- 17 ns -> [patched] 550 ns +- 24 ns: 1.42x faster (-29%)

How does it work with dict subclasses?

> How does it work with dict subclasses? PySet_New(iterable) uses fast path only when `PyDict_CheckExact(iterable)` is true. So there is no change for dict subclasses.

But what if a dict subclass overrides __iter__? You can get different result if set(d) != set(d.keys()).

Consider testing PyDict_CheckExact.

done.

 New changeset 6cbc84fb99acb33dd659d7adb29a20adbe62b74a by Inada Naoki in branch 'master': bpo-38613: Optimize set operations of dict keys. (GH-16961) https://github.com/python/cpython/commit/6cbc84fb99acb33dd659d7adb29a20adbe62b74a