Check frame back-pointer before we clear it. CPython can simply assert the identity in genobject.c because it always owns a frame itself for the current coroutine. Since Cython does not have frames here and uses the last frame of whatever traceback it currently owns, it's not clear if we can really give any guarantees better than best-effort here.

Move exception state cleanup into generator body code to allow a distinction between normal yields and yields in except blocks. They require different handling according to what CPython does.

Add a script that concatenates the code sections in CPython's ceval.c in the execution sequence given by the byte code of a Python function.

Keep direct closure of generators and coroutines intact during cleanup by disabling their tp_clear() as they still need their closure for handling the final GeneratorExit call.

See #1731.

Include deprecation warning from CPython (3.3+) in __Pyx_PyNumber_IntOrLong() and avoid inlining error handling code.

Streamline __Pyx_PyNumber_IntOrLong() with direct slot calls instead of C-API function calls that only do what we already do on our side already.

Specialise coercion from bytes to char* (and friends) to avoid going through several type checks in the generic "__Pyx_PyObject_AsStringAndSize()" function.

Split uncommon case from unicode/bytes.endswith() implementation to suggest inlining the normal case.

Fix reference leak in the Py2.x __Pyx_ReleaseBuffer() implementation for types with an overridden __releasebuffer__ method.

Closes #1638.

Fix more tests: Correct the source character position of name nodes to improve error reporting (previously reported the character position *behind* the name).

Correct the source character position of name nodes to improve error reporting (previously reported the character position *behind* the name).

Closes #1824.

Repair "range(enum)" iteration in C++ which (rightfully) disallows arbitrary assignments between enum+int.

Revert "Re-enable failure checks for Py3.2 in travis as that version is correctly supported (again)."

This reverts commit 835cf4c1.

The travis test runs seem to be unreliable for some reason.