pythonGH-132508: Use tagged integers on the evaluation stack for the last instruction offset (pythonGH-132545)

Author: markshannon

Include/internal/pycore_interpframe.h

@@ -18,6 +18,7 @@ extern "C" {
     ((int)((IF)->instr_ptr - _PyFrame_GetBytecode((IF))))
 
 static inline PyCodeObject *_PyFrame_GetCode(_PyInterpreterFrame *f) {
+    assert(!PyStackRef_IsNull(f->f_executable));
     PyObject *executable = PyStackRef_AsPyObjectBorrow(f->f_executable);
     assert(PyCode_Check(executable));
     return (PyCodeObject *)executable;

Include/internal/pycore_stackref.h

@@ -63,11 +63,13 @@ extern void _Py_stackref_associate(PyInterpreterState *interp, PyObject *obj, _P
 
 static const _PyStackRef PyStackRef_NULL = { .index = 0 };
 
-#define PyStackRef_None ((_PyStackRef){ .index = 1 } )
-#define PyStackRef_False ((_PyStackRef){ .index = 2 })
-#define PyStackRef_True ((_PyStackRef){ .index = 3 })
+// Use the first 3 even numbers for None, True and False.
+// Odd numbers are reserved for (tagged) integers
+#define PyStackRef_None ((_PyStackRef){ .index = 2 } )
+#define PyStackRef_False ((_PyStackRef){ .index = 4 })
+#define PyStackRef_True ((_PyStackRef){ .index = 6 })
 
-#define LAST_PREDEFINED_STACKREF_INDEX 3
+#define INITIAL_STACKREF_INDEX 8
 
 static inline int
 PyStackRef_IsNull(_PyStackRef ref)
@@ -96,6 +98,7 @@ PyStackRef_IsNone(_PyStackRef ref)
 static inline PyObject *
 _PyStackRef_AsPyObjectBorrow(_PyStackRef ref, const char *filename, int linenumber)
 {
+    assert((ref.index & 1) == 0);
     _Py_stackref_record_borrow(ref, filename, linenumber);
     return _Py_stackref_get_object(ref);
 }
@@ -132,31 +135,45 @@ _PyStackRef_FromPyObjectImmortal(PyObject *obj, const char *filename, int linenu
 }
 #define PyStackRef_FromPyObjectImmortal(obj) _PyStackRef_FromPyObjectImmortal(_PyObject_CAST(obj), __FILE__, __LINE__)
 
+static inline bool
+PyStackRef_IsTaggedInt(_PyStackRef ref)
+{
+    return (ref.index & 1) == 1;
+}
+
 static inline void
 _PyStackRef_CLOSE(_PyStackRef ref, const char *filename, int linenumber)
 {
+    if (PyStackRef_IsTaggedInt(ref)) {
+        return;
+    }
     PyObject *obj = _Py_stackref_close(ref, filename, linenumber);
     Py_DECREF(obj);
 }
 #define PyStackRef_CLOSE(REF) _PyStackRef_CLOSE((REF), __FILE__, __LINE__)
 
+
 static inline void
 _PyStackRef_XCLOSE(_PyStackRef ref, const char *filename, int linenumber)
 {
     if (PyStackRef_IsNull(ref)) {
         return;
     }
-    PyObject *obj = _Py_stackref_close(ref, filename, linenumber);
-    Py_DECREF(obj);
+    _PyStackRef_CLOSE(ref, filename, linenumber);
 }
 #define PyStackRef_XCLOSE(REF) _PyStackRef_XCLOSE((REF), __FILE__, __LINE__)
 
 static inline _PyStackRef
 _PyStackRef_DUP(_PyStackRef ref, const char *filename, int linenumber)
 {
-    PyObject *obj = _Py_stackref_get_object(ref);
-    Py_INCREF(obj);
-    return _Py_stackref_create(obj, filename, linenumber);
+    if (PyStackRef_IsTaggedInt(ref)) {
+        return ref;
+    }
+    else {
+        PyObject *obj = _Py_stackref_get_object(ref);
+        Py_INCREF(obj);
+        return _Py_stackref_create(obj, filename, linenumber);
+    }
 }
 #define PyStackRef_DUP(REF) _PyStackRef_DUP(REF, __FILE__, __LINE__)
 
@@ -210,8 +227,40 @@ _PyStackRef_FromPyObjectNewMortal(PyObject *obj, const char *filename, int linen
 
 extern int PyStackRef_Is(_PyStackRef a, _PyStackRef b);
 
+extern bool PyStackRef_IsTaggedInt(_PyStackRef ref);
+
+extern intptr_t PyStackRef_UntagInt(_PyStackRef ref);
+
+extern _PyStackRef PyStackRef_TagInt(intptr_t i);
+
+extern bool
+PyStackRef_IsNullOrInt(_PyStackRef ref);
+
 #else
 
+#define Py_INT_TAG 3
+
+static inline bool
+PyStackRef_IsTaggedInt(_PyStackRef i)
+{
+    return (i.bits & Py_INT_TAG) == Py_INT_TAG;
+}
+
+static inline _PyStackRef
+PyStackRef_TagInt(intptr_t i)
+{
+    assert(Py_ARITHMETIC_RIGHT_SHIFT(intptr_t, (i << 2), 2) == i);
+    return (_PyStackRef){ .bits = ((((uintptr_t)i) << 2) | Py_INT_TAG) };
+}
+
+static inline intptr_t
+PyStackRef_UntagInt(_PyStackRef i)
+{
+    assert((i.bits & Py_INT_TAG) == Py_INT_TAG);
+    intptr_t val = (intptr_t)i.bits;
+    return Py_ARITHMETIC_RIGHT_SHIFT(intptr_t, val, 2);
+}
+
 
 #ifdef Py_GIL_DISABLED
 
@@ -232,6 +281,8 @@ static const _PyStackRef PyStackRef_NULL = { .bits = Py_TAG_DEFERRED};
 #define PyStackRef_IsTrue(ref) (PyStackRef_AsPyObjectBorrow(ref) == Py_True)
 #define PyStackRef_IsFalse(ref) (PyStackRef_AsPyObjectBorrow(ref) == Py_False)
 
+#define PyStackRef_IsNullOrInt(stackref) (PyStackRef_IsNull(stackref) || PyStackRef_IsTaggedInt(stackref))
+
 static inline PyObject *
 PyStackRef_AsPyObjectBorrow(_PyStackRef stackref)
 {
@@ -451,6 +502,7 @@ PyStackRef_RefcountOnObject(_PyStackRef ref)
 static inline PyObject *
 PyStackRef_AsPyObjectBorrow(_PyStackRef ref)
 {
+    assert(!PyStackRef_IsTaggedInt(ref));
     return BITS_TO_PTR_MASKED(ref);
 }
 
@@ -587,6 +639,12 @@ PyStackRef_CLOSE(_PyStackRef ref)
 }
 #endif
 
+static inline bool
+PyStackRef_IsNullOrInt(_PyStackRef ref)
+{
+    return PyStackRef_IsNull(ref) || PyStackRef_IsTaggedInt(ref);
+}
+
 static inline void
 PyStackRef_CLOSE_SPECIALIZED(_PyStackRef ref, destructor destruct)
 {
@@ -726,7 +784,7 @@ _Py_TryXGetStackRef(PyObject **src, _PyStackRef *out)
 // Like Py_VISIT but for _PyStackRef fields
 #define _Py_VISIT_STACKREF(ref)                                         \
     do {                                                                \
-        if (!PyStackRef_IsNull(ref)) {                                  \
+        if (!PyStackRef_IsNullOrInt(ref)) {                             \
             int vret = _PyGC_VisitStackRef(&(ref), visit, arg);         \
             if (vret)                                                   \
                 return vret;                                            \

Misc/NEWS.d/next/Core_and_Builtins/2025-04-15-10-09-49.gh-issue-132508.zVe3iI.rst

@@ -0,0 +1,3 @@
+Uses tagged integers on the evaluation stack to represent the instruction
+offsets when reraising an exception. This avoids the need to box the integer
+which could fail in low memory conditions.

Objects/frameobject.c

@@ -1813,15 +1813,16 @@ frame_lineno_set_impl(PyFrameObject *self, PyObject *value)
         start_stack = pop_value(start_stack);
     }
     while (start_stack > best_stack) {
+        _PyStackRef popped = _PyFrame_StackPop(self->f_frame);
         if (top_of_stack(start_stack) == Except) {
             /* Pop exception stack as well as the evaluation stack */
-            PyObject *exc = PyStackRef_AsPyObjectBorrow(_PyFrame_StackPop(self->f_frame));
+            PyObject *exc = PyStackRef_AsPyObjectBorrow(popped);
             assert(PyExceptionInstance_Check(exc) || exc == Py_None);
             PyThreadState *tstate = _PyThreadState_GET();
             Py_XSETREF(tstate->exc_info->exc_value, exc == Py_None ? NULL : exc);
         }
         else {
-            PyStackRef_XCLOSE(_PyFrame_StackPop(self->f_frame));
+            PyStackRef_XCLOSE(popped);
         }
         start_stack = pop_value(start_stack);
     }

Python/bytecodes.c

@@ -1384,16 +1384,7 @@ dummy_func(
 
             assert(oparg >= 0 && oparg <= 2);
             if (oparg) {
-                PyObject *lasti = PyStackRef_AsPyObjectBorrow(values[0]);
-                if (PyLong_Check(lasti)) {
-                    frame->instr_ptr = _PyFrame_GetBytecode(frame) + PyLong_AsLong(lasti);
-                    assert(!_PyErr_Occurred(tstate));
-                }
-                else {
-                    _PyErr_SetString(tstate, PyExc_SystemError, "lasti is not an int");
-                    Py_DECREF(exc);
-                    ERROR_NO_POP();
-                }
+                frame->instr_ptr = _PyFrame_GetBytecode(frame) + PyStackRef_UntagInt(values[0]);
             }
             assert(exc && PyExceptionInstance_Check(exc));
             _PyErr_SetRaisedException(tstate, exc);
@@ -3472,7 +3463,7 @@ dummy_func(
             if (tb == NULL) {
                 tb = Py_None;
             }
-            assert(PyStackRef_LongCheck(lasti));
+            assert(PyStackRef_IsTaggedInt(lasti));
             (void)lasti; // Shut up compiler warning if asserts are off
             PyObject *stack[5] = {NULL, PyStackRef_AsPyObjectBorrow(exit_self), exc, val_o, tb};
             int has_self = !PyStackRef_IsNull(exit_self);
@@ -5378,11 +5369,8 @@ dummy_func(
             }
             if (lasti) {
                 int frame_lasti = _PyInterpreterFrame_LASTI(frame);
-                PyObject *lasti = PyLong_FromLong(frame_lasti);
-                if (lasti == NULL) {
-                    goto exception_unwind;
-                }
-                _PyFrame_StackPush(frame, PyStackRef_FromPyObjectSteal(lasti));
+                _PyStackRef lasti = PyStackRef_TagInt(frame_lasti);
+                _PyFrame_StackPush(frame, lasti);
             }
 
             /* Make the raw exception data

Python/ceval.c

@@ -146,6 +146,10 @@ dump_item(_PyStackRef item)
         printf("<NULL>");
         return;
     }
+    if (PyStackRef_IsTaggedInt(item)) {
+        printf("%" PRId64, (int64_t)PyStackRef_UntagInt(item));
+        return;
+    }
     PyObject *obj = PyStackRef_AsPyObjectBorrow(item);
     if (obj == NULL) {
         printf("<nil>");

Python/executor_cases.c.h

@@ -547,6 +547,7 @@ _PyGC_VisitStackRef(_PyStackRef *ref, visitproc visit, void *arg)
     // This is a bit tricky! We want to ignore stackrefs with embedded
     // refcounts when computing the incoming references, but otherwise treat
     // them like normal.
+    assert(!PyStackRef_IsTaggedInt(*ref));
     if (!PyStackRef_RefcountOnObject(*ref) && (visit == visit_decref)) {
         return 0;
     }
@@ -560,7 +561,9 @@ _PyGC_VisitFrameStack(_PyInterpreterFrame *frame, visitproc visit, void *arg)
     _PyStackRef *ref = _PyFrame_GetLocalsArray(frame);
     /* locals and stack */
     for (; ref < frame->stackpointer; ref++) {
-        _Py_VISIT_STACKREF(*ref);
+        if (!PyStackRef_IsTaggedInt(*ref)) {
+            _Py_VISIT_STACKREF(*ref);
+        }
     }
     return 0;
 }
@@ -1495,8 +1498,11 @@ mark_stacks(PyInterpreterState *interp, PyGC_Head *visited, int visited_space, b
             objects_marked += move_to_reachable(func, &reachable, visited_space);
             while (sp > locals) {
                 sp--;
+                if (PyStackRef_IsNullOrInt(*sp)) {
+                    continue;
+                }
                 PyObject *op = PyStackRef_AsPyObjectBorrow(*sp);
-                if (op == NULL || _Py_IsImmortal(op)) {
+                if (_Py_IsImmortal(op)) {
                     continue;
                 }
                 if (_PyObject_IS_GC(op)) {

Python/gc.c

@@ -265,7 +265,7 @@ frame_disable_deferred_refcounting(_PyInterpreterFrame *frame)
 
     frame->f_funcobj = PyStackRef_AsStrongReference(frame->f_funcobj);
     for (_PyStackRef *ref = frame->localsplus; ref < frame->stackpointer; ref++) {
-        if (!PyStackRef_IsNull(*ref) && PyStackRef_IsDeferred(*ref)) {
+        if (!PyStackRef_IsNullOrInt(*ref) && PyStackRef_IsDeferred(*ref)) {
             *ref = PyStackRef_AsStrongReference(*ref);
         }
     }
@@ -420,7 +420,7 @@ gc_visit_heaps(PyInterpreterState *interp, mi_block_visit_fun *visitor,
 static inline void
 gc_visit_stackref(_PyStackRef stackref)
 {
-    if (PyStackRef_IsDeferred(stackref) && !PyStackRef_IsNull(stackref)) {
+    if (PyStackRef_IsDeferred(stackref) && !PyStackRef_IsNullOrInt(stackref)) {
         PyObject *obj = PyStackRef_AsPyObjectBorrow(stackref);
         if (_PyObject_GC_IS_TRACKED(obj) && !gc_is_frozen(obj)) {
             gc_add_refs(obj, 1);
@@ -817,7 +817,7 @@ gc_abort_mark_alive(PyInterpreterState *interp,
 static int
 gc_visit_stackref_mark_alive(gc_mark_args_t *args, _PyStackRef stackref)
 {
-    if (!PyStackRef_IsNull(stackref)) {
+    if (!PyStackRef_IsNullOrInt(stackref)) {
         PyObject *op = PyStackRef_AsPyObjectBorrow(stackref);
         if (gc_mark_enqueue(op, args) < 0) {
             return -1;
@@ -1706,6 +1706,7 @@ _PyGC_VisitStackRef(_PyStackRef *ref, visitproc visit, void *arg)
     // This is a bit tricky! We want to ignore deferred references when
     // computing the incoming references, but otherwise treat them like
     // regular references.
+    assert(!PyStackRef_IsTaggedInt(*ref));
     if (!PyStackRef_IsDeferred(*ref) ||
         (visit != visit_decref && visit != visit_decref_unreachable))
     {