gh-103615: Use local events for opcode tracing (GH-109472)

* Use local monitoring for opcode trace

* Remove f_opcode_trace_set

* Add test for setting f_trace_opcodes after settrace

Python/instrumentation.c

@@ -1833,6 +1833,23 @@ _PyMonitoring_SetLocalEvents(PyCodeObject *code, int tool_id, _PyMonitoringEvent
     return 0;
 }
 
+int
+_PyMonitoring_GetLocalEvents(PyCodeObject *code, int tool_id, _PyMonitoringEventSet *events)
+{
+    assert(0 <= tool_id && tool_id < PY_MONITORING_TOOL_IDS);
+    PyInterpreterState *interp = _PyInterpreterState_GET();
+    if (check_tool(interp, tool_id)) {
+        return -1;
+    }
+    if (code->_co_monitoring == NULL) {
+        *events = 0;
+        return 0;
+    }
+    _Py_LocalMonitors *local = &code->_co_monitoring->local_monitors;
+    *events = get_local_events(local, tool_id);
+    return 0;
+}
+
 /*[clinic input]
 module monitoring
 [clinic start generated code]*/

Python/legacy_tracing.c

@@ -117,6 +117,35 @@ sys_profile_call_or_return(
     Py_RETURN_NONE;
 }
 
+int
+_PyEval_SetOpcodeTrace(
+    PyFrameObject *frame,
+    bool enable
+) {
+    assert(frame != NULL);
+    assert(PyCode_Check(frame->f_frame->f_executable));
+
+    PyCodeObject *code = (PyCodeObject *)frame->f_frame->f_executable;
+    _PyMonitoringEventSet events = 0;
+
+    if (_PyMonitoring_GetLocalEvents(code, PY_MONITORING_SYS_TRACE_ID, &events) < 0) {
+        return -1;
+    }
+
+    if (enable) {
+        if (events & (1 << PY_MONITORING_EVENT_INSTRUCTION)) {
+            return 0;
+        }
+        events |= (1 << PY_MONITORING_EVENT_INSTRUCTION);
+    } else {
+        if (!(events & (1 << PY_MONITORING_EVENT_INSTRUCTION))) {
+            return 0;
+        }
+        events &= (~(1 << PY_MONITORING_EVENT_INSTRUCTION));
+    }
+    return _PyMonitoring_SetLocalEvents(code, PY_MONITORING_SYS_TRACE_ID, events);
+}
+
 static PyObject *
 call_trace_func(_PyLegacyEventHandler *self, PyObject *arg)
 {
@@ -130,6 +159,12 @@ call_trace_func(_PyLegacyEventHandler *self, PyObject *arg)
                         "Missing frame when calling trace function.");
         return NULL;
     }
+    if (frame->f_trace_opcodes) {
+        if (_PyEval_SetOpcodeTrace(frame, true) != 0) {
+            return NULL;
+        }
+    }
+
     Py_INCREF(frame);
     int err = tstate->c_tracefunc(tstate->c_traceobj, frame, self->event, arg);
     Py_DECREF(frame);
@@ -230,11 +265,14 @@ sys_trace_instruction_func(
                         "Missing frame when calling trace function.");
         return NULL;
     }
-    if (!frame->f_trace_opcodes) {
+    PyThreadState *tstate = _PyThreadState_GET();
+    if (!tstate->c_tracefunc || !frame->f_trace_opcodes) {
+        if (_PyEval_SetOpcodeTrace(frame, false) != 0) {
+            return NULL;
+        }
         Py_RETURN_NONE;
     }
     Py_INCREF(frame);
-    PyThreadState *tstate = _PyThreadState_GET();
     int err = tstate->c_tracefunc(tstate->c_traceobj, frame, self->event, Py_None);
     frame->f_lineno = 0;
     Py_DECREF(frame);
@@ -531,9 +569,15 @@ _PyEval_SetTrace(PyThreadState *tstate, Py_tracefunc func, PyObject *arg)
             (1 << PY_MONITORING_EVENT_PY_UNWIND) | (1 << PY_MONITORING_EVENT_PY_THROW) |
             (1 << PY_MONITORING_EVENT_STOP_ITERATION) |
             (1 << PY_MONITORING_EVENT_EXCEPTION_HANDLED);
-        if (tstate->interp->f_opcode_trace_set) {
-            events |= (1 << PY_MONITORING_EVENT_INSTRUCTION);
+
+        PyFrameObject* frame = PyEval_GetFrame();
+        if (frame->f_trace_opcodes) {
+            int ret = _PyEval_SetOpcodeTrace(frame, true);
+            if (ret != 0) {
+                return ret;
+            }
         }
     }
+
     return _PyMonitoring_SetEvents(PY_MONITORING_SYS_TRACE_ID, events);
 }

Python/pystate.c

@@ -708,7 +708,6 @@ init_interpreter(PyInterpreterState *interp,
         /* Fix the self-referential, statically initialized fields. */
         interp->dtoa = (struct _dtoa_state)_dtoa_state_INIT(interp);
     }
-    interp->f_opcode_trace_set = false;
 
     interp->_initialized = 1;
     return _PyStatus_OK();
@@ -958,7 +957,6 @@ interpreter_clear(PyInterpreterState *interp, PyThreadState *tstate)
         interp->code_watchers[i] = NULL;
     }
     interp->active_code_watchers = 0;
-    interp->f_opcode_trace_set = false;
     // XXX Once we have one allocator per interpreter (i.e.
     // per-interpreter GC) we must ensure that all of the interpreter's
     // objects have been cleaned up at the point.