[WIP] gh-106608: Have _PyUOpExecutorObject use a variable-length array

Include/internal/pycore_uops.h

@@ -10,22 +10,36 @@ extern "C" {
 
 #define _Py_UOP_MAX_TRACE_LENGTH 32
 
+PyAPI_DATA(PyTypeObject) UOpExecutor_Type;
+
 typedef struct {
     uint32_t opcode;
     uint32_t oparg;
     uint64_t operand;  // A cache entry
 } _PyUOpInstruction;
 
 typedef struct {
+    PyObject_VAR_HEAD
     _PyExecutorObject base;
-    _PyUOpInstruction trace[_Py_UOP_MAX_TRACE_LENGTH];  // TODO: variable length
+    _PyUOpInstruction trace[1];
 } _PyUOpExecutorObject;
 
 _PyInterpreterFrame *_PyUopExecute(
     _PyExecutorObject *executor,
     _PyInterpreterFrame *frame,
     PyObject **stack_pointer);
 
+/* Cast argument to _PyUOpExecutorObject* type. */
+#define _PyUOpExecutor_CAST(op) \
+    _Py_CAST(_PyUOpExecutorObject*, (op))
+
+static inline Py_ssize_t PyUOpExecutor_GET_SIZE(PyObject *op) {
+    _PyUOpExecutorObject *executor = _PyUOpExecutor_CAST(op);
+    return Py_SIZE(executor);
+}
+#define PyUOpExecutor_GET_SIZE(op) PyUOpExecutor_GET_SIZE(_PyObject_CAST(op))
+PyAPI_FUNC(PyObject *) PyUOpExecutor_New(_PyUOpInstruction trace[], Py_ssize_t size);
+
 #ifdef __cplusplus
 }
 #endif

Python/executor.c

@@ -30,6 +30,30 @@
 #undef ENABLE_SPECIALIZATION
 #define ENABLE_SPECIALIZATION 0
 
+static _PyUOpExecutorObject *
+PyUOpExecutor_alloc(Py_ssize_t size)
+{
+    _PyUOpExecutorObject *op = PyObject_GC_NewVar(_PyUOpExecutorObject, &UOpExecutor_Type, size);
+    if (op == NULL) {
+        return NULL;
+    }
+    return op;
+}
+
+PyObject *
+PyUOpExecutor_New(_PyUOpInstruction trace[], Py_ssize_t size)
+{
+    _PyUOpExecutorObject *op;
+    op = PyUOpExecutor_alloc(size);
+    if (op == NULL) {
+        return NULL;
+    }
+    for (Py_ssize_t i = 0; i < size; i++) {
+        op->trace[i] = trace[i];
+    }
+    _PyObject_GC_TRACK(op);
+    return (PyObject *) op;
+}
 
 _PyInterpreterFrame *
 _PyUopExecute(_PyExecutorObject *executor, _PyInterpreterFrame *frame, PyObject **stack_pointer)

Python/optimizer.c

@@ -320,13 +320,7 @@ uop_name(int index) {
 static Py_ssize_t
 uop_len(_PyUOpExecutorObject *self)
 {
-    int count = 0;
-    for (; count < _Py_UOP_MAX_TRACE_LENGTH; count++) {
-        if (self->trace[count].opcode == 0) {
-            break;
-        }
-    }
-    return count;
+    return PyUOpExecutor_GET_SIZE(self);
 }
 
 static PyObject *
@@ -365,7 +359,7 @@ PySequenceMethods uop_as_sequence = {
     .sq_item = (ssizeargfunc)uop_item,
 };
 
-static PyTypeObject UOpExecutor_Type = {
+PyTypeObject UOpExecutor_Type = {
     PyVarObject_HEAD_INIT(&PyType_Type, 0)
     .tp_name = "uop_executor",
     .tp_basicsize = sizeof(_PyUOpExecutorObject),
@@ -699,15 +693,8 @@ uop_optimize(
         return trace_length;
     }
     OBJECT_STAT_INC(optimization_traces_created);
-    _PyUOpExecutorObject *executor = PyObject_New(_PyUOpExecutorObject, &UOpExecutor_Type);
-    if (executor == NULL) {
-        return -1;
-    }
+    _PyUOpExecutorObject *executor = (_PyUOpExecutorObject *) PyUOpExecutor_New(trace, trace_length);
     executor->base.execute = _PyUopExecute;
-    memcpy(executor->trace, trace, trace_length * sizeof(_PyUOpInstruction));
-        if (trace_length < _Py_UOP_MAX_TRACE_LENGTH) {
-            executor->trace[trace_length].opcode = 0;  // Sentinel
-        }
     *exec_ptr = (_PyExecutorObject *)executor;
     return 1;
 }