Similar to http://bugs.python.org/issue14509, Python 3.3 conflates Py_DEBUG and non-NDEBUG builds, creating build failures when building with 'CFLAGS=-UNDEBUG ./configure --without-pydebug'. (assert statements are only compiled out when NDEBUG is set, not when Py_DEBUG is unset.) This patch fixes the two root causes.

The patch changes unicodeobject.c and obmalloc.c
Thomas, is the problem new to 3.3 (and the new unicode implementation)?

Yes, I already had the same kinds of failures fixed for 3.2, before; this only affects 3.3 and later.

I added _PyUnicode_CheckConsistency() to Python 3.3 to check the implementation of the PEP 393, this function should not be called in release mode, that's just it is not defined in release mode.

> that's just it is not defined in release mode.

that's *why*

2013/3/15 STINNER Victor <report@bugs.python.org>:
>
> STINNER Victor added the comment:
>
> I added _PyUnicode_CheckConsistency() to Python 3.3 to check the implementation of the PEP 393, this function should not be called in release mode, that's just it is not defined in release mode.
>
> ----------
>
> _______________________________________
> Python tracker <report@bugs.python.org>
> <http://bugs.python.org/issue17411>
> _______________________________________

Unfortunately there is no "release mode". There's Py_DEBUG mode, and the absence of Py_DEBUG. And there's NDEBUG, and the absence of NDEBUG, which controls the assert macro. Py_DEBUG unsets NDEBUG, but *not* setting Py_DEBUG doesn't *set* NDEBUG (nor should it.)

In the Google build environment, where we occasionally embed Python and also build extensions that use Python and Python itself in slightly different ways, we have build modes that don't set Py_DEBUG but don't set NDEBUG -- so we keep the asserts, but don't set Py_DEBUG (which would change the ABI.) I suggest that this is a valid expectation :) If something is used in assert statements (and _PyUnicode_CheckConsistency is, quite a lot) it needs to be defined whenever NDEBUG is not defined, not just when Py_DEBUG is defined. The patch makes it a macro that calls PyUnicode_Check() in that case, because that's what unicodeobject.c itself does.

In the case of obmalloc, the asserts are calling functions that are only defined when PYMALLOC_DEBUG is set, so it seems obvious that that's the guard that should surround their use.