"pyo" macro from gdbinit (see #3631) uses _PyObject_Dump() to display
an object. This function calls (indirectly) string_print() to display
one line of text. But if the GIL is released (I guess it's the GIL or
is it called the "thread state"?), gdb crashs Python:
   object  : Fatal Python error: PyEval_SaveThread: NULL tstate
   Program received signal SIGABRT, Aborted.
   0xffffe410 in __kernel_vsyscall ()
Workaround: ensure GIL before Py_BEGIN_ALLOW_THREADS... That sounds
ugly but it works :-) So I propose to enable it in debug mode (#ifdef
Py_DEBUG) with a patch.
I guess that the issue is very specific to (gdb) debugging and should
not affect normal usage of Python. That's why I choosed to enable it
only in debug mode.

I once fell into the same issue, but the patch should modify
_PyObject_Dump(), around the call to PyObject_Print.
string_print() is not the only function called by _PyObject_Dump, by far...
And beware that many people routinely run python in "debug mode" outside
gdb, for example during development.
That's why it's better to modify _PyObject_Dump only; we can be sure
that it only affects a debugger session.

Oh! I have a better idea: why not patching _PyObject_Dump() instead of
string_print() :-) So here is a new patch.

Can anyone review my new patch?

The patch is fine.
I don't know if it can make the 2.6 release, but it is very simple, and
affect only a function used in debugger macros.

I would love to have this patch, along with those of #3631 and #3610,
included in Misc as diffs. This would make it easier to get the improved
functionality in a new development box, besides allowing distributions
to apply them at will.

Can anyone apply this patch? Or explain me why it can not be applied?

Committed in r67802.