Feature or enhancement

Theobject_lookup_special microbenchmark inTools/ftscalingbench/ftscalingbench.py currently doesn't scale well and is indicative of a broader FT performance issue that we should fix. The benchmark just callsround() from multiple threads concurrently:

The issue is thatround() calls_PyObject_LookupSpecial(number, &_Py_ID(__round__)), which increments the reference count of the returned function (i.e., offloat.round). The underlying function supports deferred reference counting, but_PyObject_LookupSpecial and_PyType_LookupRef do not take advantage of it.

For the FT build, we also need some extra support in order to safely use_PyStackRef inbuiltin_round_impl, because it's important that all_PyStackRefs are visible to the GC. To support this, we can add a singly linked list of active_PyStackRefs to_PyThreadStateImpl.

Thestruct _PyCStackRef implements this linked list pointer + a_PyStackRef. In the GIL-enabled build, there's no linked list and it's essentially the same as_PyStackRef.

// A stackref that can be stored in a regular C local variable and be visible// to the GC in the free threading build.// Used in combination with _PyThreadState_PushCStackRef().typedefstruct_PyCStackRef {_PyStackRefref;#ifdefPy_GIL_DISABLEDstruct_PyCStackRef*next;#endif}_PyCStackRef;struct_PyThreadStateImpl {    ...// Linked list (stack) of active _PyCStackRefstruct_PyCStackRef*c_stack_refs;    ...}staticinlinevoid_PyThreadState_PushCStackRef(PyThreadState*tstate,_PyCStackRef*ref) { ... }staticinlinevoid_PyThreadState_PopCStackRef(PyThreadState*tstate,_PyCStackRef*ref) { ... }

cc@mpage@markshannon

Linked PRs

• gh-131586: Avoid refcount contention in some "special" calls #131588
• gh-131586: Avoid refcount contention in context managers #131851