PyInterpreterState_New(), PyThreadState_New(): use malloc/free directly.
This appears to finish repairs for SF bug 1041645.
This is a critical bugfix.
diff --git a/Misc/NEWS b/Misc/NEWS
index e279dca..87c6f37 100644
--- a/Misc/NEWS
+++ b/Misc/NEWS
@@ -136,6 +136,16 @@
C API
-----
+- The C API calls ``PyInterpreterState_New()`` and ``PyThreadState_New()``
+ are two of the very few advertised as being safe to call without holding
+ the GIL. However, this wasn't true in a debug build, as bug 1041645
+ demonstrated. In a debug build, Python redirects the ``PyMem`` family
+ of calls to Python's small-object allocator, to get the benefit of
+ its extra debugging capabilities. But Python's small-object allocator
+ isn't threadsafe, relying on the GIL to avoid the expense of doing its
+ own locking. ``PyInterpreterState_New()`` and ``PyThreadState_New()``
+ call the platform ``malloc()`` directly now, regardless of build type.
+
- PyLong_AsUnsignedLong[Mask] now support int objects as well.
- SF patch #998993: ``PyUnicode_DecodeUTF8Stateful`` and