Move stackref buffer to per-interpreter. Check for C recursion limit in testcapi

Author: Fidget-Spinner

Modules/_testcapi/vectorcall.c

@@ -98,7 +98,15 @@ _testcapi_pyobject_vectorcall_impl(PyObject *module, PyObject *func,
         PyErr_SetString(PyExc_TypeError, "kwnames must be None or a tuple");
         return NULL;
     }
-    return PyObject_Vectorcall(func, stack, nargs, kwnames);
+    PyObject *res;
+    // The CPython interpreter does not guarantee that vectorcalls are
+    // checked for recursion limit. It's thus up to the C extension themselves to check.
+    if (Py_EnterRecursiveCall("in _testcapi.pyobject_vectorcall")) {
+        return NULL;
+    }
+    res = PyObject_Vectorcall(func, stack, nargs, kwnames);
+    Py_LeaveRecursiveCall();
+    return res;
 }
 
 static PyObject *

Python/ceval.c

@@ -1168,6 +1168,13 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
         return NULL;
     }
 
+    /* +1 because vectorcall might use -1 to write self */
+    /* gh-138115: This must not be in individual cases for
+       non-tail-call interpreters, as it results in excessive
+       stack usage in some compilers.
+    */
+    PyObject *STACKREF_SCRATCH[MAX_STACKREF_SCRATCH+1];
+
     /* Local "register" variables.
      * These are cached values from the frame and code object.  */
     _Py_CODEUNIT *next_instr;

Python/ceval_macros.h

@@ -428,8 +428,7 @@ do { \
 #define MAX_STACKREF_SCRATCH 10
 
 #define STACKREFS_TO_PYOBJECTS(ARGS, ARG_COUNT, NAME) \
-    /* +1 because vectorcall might use -1 to write self */ \
-    PyObject *NAME##_temp[MAX_STACKREF_SCRATCH+1]; \
+    PyObject **NAME##_temp = (PyObject **)&STACKREF_SCRATCH; \
     PyObject **NAME = _PyObjectArray_FromStackRefArray(ARGS, ARG_COUNT, NAME##_temp + 1);
 
 #define STACKREFS_TO_PYOBJECTS_CLEANUP(NAME) \