[3.7] bpo-27987: align PyGC_Head to alignof(long double) (GH-13335)

[methane]

Reverts #13569

#13318 fixed the issue happened by merging #13335.

https://bugs.python.org/issue27987

[gpshead]

I'm going to test this out with PGO and ubsan builds first. I agree that the obmalloc alignment leading to the ubsan issues is probably fixed by #13318 at this point.

What I'm missing is why this is even necessary at this point given the obmalloc alignment change? What does this change actually cause to happen?

[methane]

What I'm missing is why this is even necessary at this point given the obmalloc alignment change? What does this change actually cause to happen?

Alignment is all or nothing. If this is not aligned, no need to align pointer returned by obmalloc.
See this figure.

#13318 make pointer PyObject_Malloc returned (a) aligned to 16 byte. Now PyGC_Head *gc is aligned to 16 byte.

This pull request make sizeof(PyGC_Head) (b) aligned to 16 byte. Now PyObject *o = (PyObject*)(gc+1) is aligned to 16 byte.

C compiler align (c) part automatically. Now member of PyObject is aligned.

[gpshead]

thanks, that was helpful :) also, testing this on an ubsan build it shows a notable improvement. the two changes together work.