On Tue, 11 Dec 2012 03:05:19 +0100 (CET) gregory.p.smith <python-checkins@python.org> wrote:
Using 'long double' to force this structure to be worst case aligned is no longer required as of Python 2.5+ when the gc_refs changed from an int (4 bytes) to a Py_ssize_t (8 bytes) as the minimum size is 16 bytes.
The use of a 'long double' triggered a warning by Clang trunk's Undefined-Behavior Sanitizer as on many platforms a long double requires 16-byte alignment but the Python memory allocator only guarantees 8 byte alignment.
So our code would allocate and use these structures with technically improper alignment. Though it didn't matter since the 'dummy' field is never used. This silences that warning.
Spelunking into code history, the double was added in 2001 to force better alignment on some platforms and changed to a long double in 2002 to appease Tru64. That issue should no loner be present since the upgrade from int to Py_ssize_t where the minimum structure size increased to 16 (unless anyone knows of a platform where ssize_t is 4 bytes?)
What?? Every 32-bit platform has a 4 bytes ssize_t (and size_t).
We can probably get rid of the double and this union hack all together today. That is a slightly more invasive change that can be left for later.
How do you suggest to get rid of it? Some platforms still have strict alignment rules and we must enforce that PyObjects (*) are always aligned to the largest possible alignment, since a PyObject-derived struct can hold arbitrary C types. (*) GC-enabled PyObjects, anyway. Others will be naturally aligned thanks to the memory allocator. What's more, I think you shouldn't be doing this kind of change in a bugfix release. It might break compiled C extensions since you are changing some characteristics of object layout (although you would probably only break those extensions which access the GC header, which is probably not many of them). Resource consumption improvements generally go only into the next feature release. Regards Antoine.