Following PR adds several asserts for checking that sq_length and mq_length either return non-negative result or raise an exception.

One assert already was in PySequence_GetItem().

The changes appear consistent within themselves (ie, no apparent typos), and the one for builtin len appears to fix a bug that would allow negative length.  Can tests be written in Python that fail without the patch?   Or would mis-behaving classes have to be written in C?

Mis-behaving classes have to be written in C. If the __len__ method in Python class returns a negative value, ValueError is raised in slot_sq_length().

PySequence_GetItem() already contains an assertion that sq_length returns negative value only when an exception is set. The patch just extends this to other cases of calling sq_length and mq_length.

The alternate approach is raising SystemError rather than crashing in debug build on assertion. Since this check is used in exceptional case (when a raised exception is expected) it doesn't slow down working code.

Please don't add more code to critical paths.  

There is no demonstrated need for an additional check here.  The current code has worked well for millions of users for many years.

Also, there are too many checks in common paths and some of them are redundant.  We need less of this, not more.

assert() is no-op in release build and that are not critical parts.

Actually the patch decreases the number of checks in release build. Runtime check PyErr_Occurred() is moved into assert() in builtin_len().

New changeset 813f943c592cf225871b99cffc99304c8cbbee40 by Serhiy Storchaka in branch 'master':
bpo-29838: Add asserts for checking results of sq_length and mq_length slots. (#700)
https://github.com/python/cpython/commit/813f943c592cf225871b99cffc99304c8cbbee40