GH-123516: Improve JIT memory consumption by invalidating cold executors (GH-124443)

Co-authored-by: Bénédikt Tran <10796600+picnixz@users.noreply.github.com>

Python/bytecodes.c

@@ -4836,6 +4836,14 @@ dummy_func(
             assert(((_PyExecutorObject *)executor)->vm_data.valid);
         }
 
+        tier2 op(_MAKE_WARM, (--)) {
+            current_executor->vm_data.warm = true;
+            // It's okay if this ends up going negative.
+            if (--tstate->interp->trace_run_counter == 0) {
+                _Py_set_eval_breaker_bit(tstate, _PY_EVAL_JIT_INVALIDATE_COLD_BIT);
+            }
+        }
+
         tier2 op(_FATAL_ERROR, (--)) {
             assert(0);
             Py_FatalError("Fatal error uop executed.");

Python/ceval_gil.c

@@ -1289,6 +1289,12 @@ _Py_HandlePending(PyThreadState *tstate)
         _Py_RunGC(tstate);
     }
 
+    if ((breaker & _PY_EVAL_JIT_INVALIDATE_COLD_BIT) != 0) {
+        _Py_unset_eval_breaker_bit(tstate, _PY_EVAL_JIT_INVALIDATE_COLD_BIT);
+        _Py_Executors_InvalidateCold(tstate->interp);
+        tstate->interp->trace_run_counter = JIT_CLEANUP_THRESHOLD;
+    }
+
     /* GIL drop request */
     if ((breaker & _PY_GIL_DROP_REQUEST_BIT) != 0) {
         /* Give another thread a chance */

Python/executor_cases.c.h

@@ -5435,6 +5435,15 @@
             break;
         }
 
+        case _MAKE_WARM: {
+            current_executor->vm_data.warm = true;
+            // It's okay if this ends up going negative.
+            if (--tstate->interp->trace_run_counter == 0) {
+                _Py_set_eval_breaker_bit(tstate, _PY_EVAL_JIT_INVALIDATE_COLD_BIT);
+            }
+            break;
+        }
+
         case _FATAL_ERROR: {
             assert(0);
             Py_FatalError("Fatal error uop executed.");

Python/optimizer.c

@@ -565,6 +565,7 @@ translate_bytecode_to_trace(
             code->co_firstlineno,
             2 * INSTR_IP(initial_instr, code));
     ADD_TO_TRACE(_START_EXECUTOR, 0, (uintptr_t)instr, INSTR_IP(instr, code));
+    ADD_TO_TRACE(_MAKE_WARM, 0, 0, 0);
     uint32_t target = 0;
 
     for (;;) {
@@ -1194,6 +1195,9 @@ make_executor_from_uops(_PyUOpInstruction *buffer, int length, const _PyBloomFil
     executor->jit_code = NULL;
     executor->jit_side_entry = NULL;
     executor->jit_size = 0;
+    // This is initialized to true so we can prevent the executor
+    // from being immediately detected as cold and invalidated.
+    executor->vm_data.warm = true;
     if (_PyJIT_Compile(executor, executor->trace, length)) {
         Py_DECREF(executor);
         return NULL;
@@ -1659,4 +1663,42 @@ _Py_Executors_InvalidateAll(PyInterpreterState *interp, int is_invalidation)
     }
 }
 
+void
+_Py_Executors_InvalidateCold(PyInterpreterState *interp)
+{
+    /* Walk the list of executors */
+    /* TO DO -- Use a tree to avoid traversing as many objects */
+    PyObject *invalidate = PyList_New(0);
+    if (invalidate == NULL) {
+        goto error;
+    }
+
+    /* Clearing an executor can deallocate others, so we need to make a list of
+     * executors to invalidate first */
+    for (_PyExecutorObject *exec = interp->executor_list_head; exec != NULL;) {
+        assert(exec->vm_data.valid);
+        _PyExecutorObject *next = exec->vm_data.links.next;
+
+        if (!exec->vm_data.warm && PyList_Append(invalidate, (PyObject *)exec) < 0) {
+            goto error;
+        }
+        else {
+            exec->vm_data.warm = false;
+        }
+
+        exec = next;
+    }
+    for (Py_ssize_t i = 0; i < PyList_GET_SIZE(invalidate); i++) {
+        _PyExecutorObject *exec = (_PyExecutorObject *)PyList_GET_ITEM(invalidate, i);
+        executor_clear(exec);
+    }
+    Py_DECREF(invalidate);
+    return;
+error:
+    PyErr_Clear();
+    Py_XDECREF(invalidate);
+    // If we're truly out of memory, wiping out everything is a fine fallback
+    _Py_Executors_InvalidateAll(interp, 0);
+}
+
 #endif /* _Py_TIER2 */

Python/optimizer_cases.c.h

@@ -2381,6 +2381,10 @@
             break;
         }
 
+        case _MAKE_WARM: {
+            break;
+        }
+
         case _FATAL_ERROR: {
             break;
         }

Python/pystate.c

@@ -660,6 +660,7 @@ init_interpreter(PyInterpreterState *interp,
 #ifdef _Py_TIER2
     (void)_Py_SetOptimizer(interp, NULL);
     interp->executor_list_head = NULL;
+    interp->trace_run_counter = JIT_CLEANUP_THRESHOLD;
 #endif
     if (interp != &runtime->_main_interpreter) {
         /* Fix the self-referential, statically initialized fields. */