GH-113860: All executors are now defined in terms of micro ops. Convert counter executor to use uops. (GH-113864)

9 files changed, 125 insertions, 139 deletions

diff --git a/Include/cpython/optimizer.h b/Include/cpython/optimizer.h
index d521eac..f077da7 100644
--- a/Include/cpython/optimizer.h
+++ b/Include/cpython/optimizer.h
@@ -31,8 +31,6 @@ typedef struct {
 
 typedef struct _PyExecutorObject {
     PyObject_VAR_HEAD
-    /* WARNING: execute consumes a reference to self. This is necessary to allow executors to tail call into each other. */
-    _Py_CODEUNIT *(*execute)(struct _PyExecutorObject *self, struct _PyInterpreterFrame *frame, PyObject **stack_pointer);
     _PyVMData vm_data; /* Used by the VM, but opaque to the optimizer */
     /* Data needed by the executor goes here, but is opaque to the VM */
 } _PyExecutorObject;
@@ -52,6 +50,12 @@ typedef struct _PyOptimizerObject {
     /* Data needed by the optimizer goes here, but is opaque to the VM */
 } _PyOptimizerObject;
 
+/** Test support **/
+typedef struct {
+    _PyOptimizerObject base;
+    int64_t count;
+} _PyCounterOptimizerObject;
+
 PyAPI_FUNC(int) PyUnstable_Replace_Executor(PyCodeObject *code, _Py_CODEUNIT *instr, _PyExecutorObject *executor);
 
 PyAPI_FUNC(void) PyUnstable_SetOptimizer(_PyOptimizerObject* optimizer);
diff --git a/Include/internal/pycore_opcode_metadata.h b/Include/internal/pycore_opcode_metadata.h
index 7d39e4b..a9d698d 100644
--- a/Include/internal/pycore_opcode_metadata.h
+++ b/Include/internal/pycore_opcode_metadata.h
@@ -1009,7 +1009,7 @@ const struct opcode_metadata _PyOpcode_opcode_metadata[268] = {
     [END_ASYNC_FOR] = { true, INSTR_FMT_IX, HAS_ERROR_FLAG | HAS_ESCAPES_FLAG },
     [END_FOR] = { true, INSTR_FMT_IX, 0 },
     [END_SEND] = { true, INSTR_FMT_IX, 0 },
-    [ENTER_EXECUTOR] = { true, INSTR_FMT_IB, HAS_ARG_FLAG | HAS_EVAL_BREAK_FLAG | HAS_ERROR_FLAG | HAS_ESCAPES_FLAG },
+    [ENTER_EXECUTOR] = { true, INSTR_FMT_IB, HAS_ARG_FLAG | HAS_EVAL_BREAK_FLAG },
     [EXIT_INIT_CHECK] = { true, INSTR_FMT_IX, HAS_ERROR_FLAG | HAS_ESCAPES_FLAG },
     [EXTENDED_ARG] = { true, INSTR_FMT_IB, HAS_ARG_FLAG },
     [FORMAT_SIMPLE] = { true, INSTR_FMT_IX, HAS_ERROR_FLAG | HAS_ESCAPES_FLAG },
diff --git a/Include/internal/pycore_uop_ids.h b/Include/internal/pycore_uop_ids.h
index 4a9a00b..b3b3632 100644
--- a/Include/internal/pycore_uop_ids.h
+++ b/Include/internal/pycore_uop_ids.h
@@ -231,7 +231,9 @@ extern "C" {
 #define _SAVE_RETURN_OFFSET 378
 #define _INSERT 379
 #define _CHECK_VALIDITY 380
-#define MAX_UOP_ID 380
+#define _LOAD_CONST_INLINE_BORROW 381
+#define _INTERNAL_INCREMENT_OPT_COUNTER 382
+#define MAX_UOP_ID 382
 
 #ifdef __cplusplus
 }
diff --git a/Include/internal/pycore_uop_metadata.h b/Include/internal/pycore_uop_metadata.h
index 300bd3b..ab498e9 100644
--- a/Include/internal/pycore_uop_metadata.h
+++ b/Include/internal/pycore_uop_metadata.h
@@ -203,6 +203,8 @@ const uint16_t _PyUop_Flags[MAX_UOP_ID+1] = {
     [_EXIT_TRACE] = HAS_DEOPT_FLAG,
     [_INSERT] = HAS_ARG_FLAG,
     [_CHECK_VALIDITY] = HAS_DEOPT_FLAG,
+    [_LOAD_CONST_INLINE_BORROW] = 0,
+    [_INTERNAL_INCREMENT_OPT_COUNTER] = 0,
 };
 
 const char *const _PyOpcode_uop_name[MAX_UOP_ID+1] = {
@@ -303,6 +305,7 @@ const char *const _PyOpcode_uop_name[MAX_UOP_ID+1] = {
     [_INIT_CALL_BOUND_METHOD_EXACT_ARGS] = "_INIT_CALL_BOUND_METHOD_EXACT_ARGS",
     [_INIT_CALL_PY_EXACT_ARGS] = "_INIT_CALL_PY_EXACT_ARGS",
     [_INSERT] = "_INSERT",
+    [_INTERNAL_INCREMENT_OPT_COUNTER] = "_INTERNAL_INCREMENT_OPT_COUNTER",
     [_IS_NONE] = "_IS_NONE",
     [_IS_OP] = "_IS_OP",
     [_ITER_CHECK_LIST] = "_ITER_CHECK_LIST",
@@ -328,6 +331,7 @@ const char *const _PyOpcode_uop_name[MAX_UOP_ID+1] = {
     [_LOAD_ATTR_WITH_HINT] = "_LOAD_ATTR_WITH_HINT",
     [_LOAD_BUILD_CLASS] = "_LOAD_BUILD_CLASS",
     [_LOAD_CONST] = "_LOAD_CONST",
+    [_LOAD_CONST_INLINE_BORROW] = "_LOAD_CONST_INLINE_BORROW",
     [_LOAD_DEREF] = "_LOAD_DEREF",
     [_LOAD_FAST] = "_LOAD_FAST",
     [_LOAD_FAST_AND_CLEAR] = "_LOAD_FAST_AND_CLEAR",
diff --git a/Include/internal/pycore_uops.h b/Include/internal/pycore_uops.h
index 153884f..eb10002 100644
--- a/Include/internal/pycore_uops.h
+++ b/Include/internal/pycore_uops.h
@@ -24,11 +24,6 @@ typedef struct {
     _PyUOpInstruction trace[1];
 } _PyUOpExecutorObject;
 
-_Py_CODEUNIT *_PyUOpExecute(
-    _PyExecutorObject *executor,
-    _PyInterpreterFrame *frame,
-    PyObject **stack_pointer);
-
 #ifdef __cplusplus
 }
 #endif
diff --git a/Python/bytecodes.c b/Python/bytecodes.c
index e1a6a25..f53ddae 100644
--- a/Python/bytecodes.c
+++ b/Python/bytecodes.c
@@ -2366,16 +2366,8 @@ dummy_func(
             _PyExecutorObject *executor = (_PyExecutorObject *)code->co_executors->executors[oparg&255];
             if (executor->vm_data.valid) {
                 Py_INCREF(executor);
-                if (executor->execute == _PyUOpExecute) {
-                    current_executor = (_PyUOpExecutorObject *)executor;
-                    GOTO_TIER_TWO();
-                }
-                next_instr = executor->execute(executor, frame, stack_pointer);
-                frame = tstate->current_frame;
-                if (next_instr == NULL) {
-                    goto resume_with_error;
-                }
-                stack_pointer = _PyFrame_GetStackPointer(frame);
+                current_executor = (_PyUOpExecutorObject *)executor;
+                GOTO_TIER_TWO();
             }
             else {
                 code->co_executors->executors[oparg & 255] = NULL;
@@ -4066,6 +4058,16 @@ dummy_func(
             DEOPT_IF(!current_executor->base.vm_data.valid);
         }
 
+        op(_LOAD_CONST_INLINE_BORROW, (ptr/4 -- value)) {
+            value = ptr;
+        }
+
+        /* Internal -- for testing executors */
+        op(_INTERNAL_INCREMENT_OPT_COUNTER, (opt --)) {
+            _PyCounterOptimizerObject *exe = (_PyCounterOptimizerObject *)opt;
+            exe->count++;
+        }
+
 
 // END BYTECODES //
 
diff --git a/Python/executor_cases.c.h b/Python/executor_cases.c.h
index 14fb3a0..ea4caa9 100644
--- a/Python/executor_cases.c.h
+++ b/Python/executor_cases.c.h
@@ -3397,4 +3397,22 @@
             break;
         }
 
+        case _LOAD_CONST_INLINE_BORROW: {
+            PyObject *value;
+            PyObject *ptr = (PyObject *)CURRENT_OPERAND();
+            value = ptr;
+            stack_pointer[0] = value;
+            stack_pointer += 1;
+            break;
+        }
+
+        case _INTERNAL_INCREMENT_OPT_COUNTER: {
+            PyObject *opt;
+            opt = stack_pointer[-1];
+            _PyCounterOptimizerObject *exe = (_PyCounterOptimizerObject *)opt;
+            exe->count++;
+            stack_pointer += -1;
+            break;
+        }
+
 #undef TIER_TWO
diff --git a/Python/generated_cases.c.h b/Python/generated_cases.c.h
index 8226d82..e693e3e 100644
--- a/Python/generated_cases.c.h
+++ b/Python/generated_cases.c.h
@@ -2380,16 +2380,8 @@
             _PyExecutorObject *executor = (_PyExecutorObject *)code->co_executors->executors[oparg&255];
             if (executor->vm_data.valid) {
                 Py_INCREF(executor);
-                if (executor->execute == _PyUOpExecute) {
-                    current_executor = (_PyUOpExecutorObject *)executor;
-                    GOTO_TIER_TWO();
-                }
-                next_instr = executor->execute(executor, frame, stack_pointer);
-                frame = tstate->current_frame;
-                if (next_instr == NULL) {
-                    goto resume_with_error;
-                }
-                stack_pointer = _PyFrame_GetStackPointer(frame);
+                current_executor = (_PyUOpExecutorObject *)executor;
+                GOTO_TIER_TWO();
             }
             else {
                 code->co_executors->executors[oparg & 255] = NULL;
diff --git a/Python/optimizer.c b/Python/optimizer.c
index ad5b499..28e12db 100644
--- a/Python/optimizer.c
+++ b/Python/optimizer.c
@@ -212,27 +212,6 @@ PyUnstable_GetExecutor(PyCodeObject *code, int offset)
     return NULL;
 }
 
-/** Test support **/
-
-
-typedef struct {
-    _PyOptimizerObject base;
-    int64_t count;
-} _PyCounterOptimizerObject;
-
-typedef struct {
-    _PyExecutorObject executor;
-    _PyCounterOptimizerObject *optimizer;
-    _Py_CODEUNIT *next_instr;
-} _PyCounterExecutorObject;
-
-static void
-counter_dealloc(_PyCounterExecutorObject *self) {
-    _Py_ExecutorClear((_PyExecutorObject *)self);
-    Py_DECREF(self->optimizer);
-    PyObject_Free(self);
-}
-
 static PyObject *
 is_valid(PyObject *self, PyObject *Py_UNUSED(ignored))
 {
@@ -244,84 +223,6 @@ static PyMethodDef executor_methods[] = {
     { NULL, NULL },
 };
 
-PyTypeObject _PyCounterExecutor_Type = {
-    PyVarObject_HEAD_INIT(&PyType_Type, 0)
-    .tp_name = "counting_executor",
-    .tp_basicsize = sizeof(_PyCounterExecutorObject),
-    .tp_itemsize = 0,
-    .tp_flags = Py_TPFLAGS_DEFAULT | Py_TPFLAGS_DISALLOW_INSTANTIATION,
-    .tp_dealloc = (destructor)counter_dealloc,
-    .tp_methods = executor_methods,
-};
-
-static _Py_CODEUNIT *
-counter_execute(_PyExecutorObject *self, _PyInterpreterFrame *frame, PyObject **stack_pointer)
-{
-    ((_PyCounterExecutorObject *)self)->optimizer->count++;
-    _PyFrame_SetStackPointer(frame, stack_pointer);
-    Py_DECREF(self);
-    return ((_PyCounterExecutorObject *)self)->next_instr;
-}
-
-static int
-counter_optimize(
-    _PyOptimizerObject* self,
-    PyCodeObject *code,
-    _Py_CODEUNIT *instr,
-    _PyExecutorObject **exec_ptr,
-    int Py_UNUSED(curr_stackentries)
-)
-{
-    _PyCounterExecutorObject *executor = (_PyCounterExecutorObject *)_PyObject_New(&_PyCounterExecutor_Type);
-    if (executor == NULL) {
-        return -1;
-    }
-    executor->executor.execute = counter_execute;
-    Py_INCREF(self);
-    executor->optimizer = (_PyCounterOptimizerObject *)self;
-    executor->next_instr = instr;
-    *exec_ptr = (_PyExecutorObject *)executor;
-    _PyBloomFilter empty;
-    _Py_BloomFilter_Init(&empty);
-    _Py_ExecutorInit((_PyExecutorObject *)executor, &empty);
-    return 1;
-}
-
-static PyObject *
-counter_get_counter(PyObject *self, PyObject *args)
-{
-    return PyLong_FromLongLong(((_PyCounterOptimizerObject *)self)->count);
-}
-
-static PyMethodDef counter_optimizer_methods[] = {
-    { "get_count", counter_get_counter, METH_NOARGS, NULL },
-    { NULL, NULL },
-};
-
-PyTypeObject _PyCounterOptimizer_Type = {
-    PyVarObject_HEAD_INIT(&PyType_Type, 0)
-    .tp_name = "Counter optimizer",
-    .tp_basicsize = sizeof(_PyCounterOptimizerObject),
-    .tp_itemsize = 0,
-    .tp_flags = Py_TPFLAGS_DEFAULT | Py_TPFLAGS_DISALLOW_INSTANTIATION,
-    .tp_methods = counter_optimizer_methods,
-    .tp_dealloc = (destructor)PyObject_Del,
-};
-
-PyObject *
-PyUnstable_Optimizer_NewCounter(void)
-{
-    _PyCounterOptimizerObject *opt = (_PyCounterOptimizerObject *)_PyObject_New(&_PyCounterOptimizer_Type);
-    if (opt == NULL) {
-        return NULL;
-    }
-    opt->base.optimize = counter_optimize;
-    opt->base.resume_threshold = INT16_MAX;
-    opt->base.backedge_threshold = 0;
-    opt->count = 0;
-    return (PyObject *)opt;
-}
-
 ///////////////////// Experimental UOp Optimizer /////////////////////
 
 static void
@@ -381,7 +282,7 @@ PySequenceMethods uop_as_sequence = {
 PyTypeObject _PyUOpExecutor_Type = {
     PyVarObject_HEAD_INIT(&PyType_Type, 0)
     .tp_name = "uop_executor",
-    .tp_basicsize = sizeof(_PyUOpExecutorObject) - sizeof(_PyUOpInstruction),
+    .tp_basicsize = offsetof(_PyUOpExecutorObject, trace),
     .tp_itemsize = sizeof(_PyUOpInstruction),
     .tp_flags = Py_TPFLAGS_DEFAULT | Py_TPFLAGS_DISALLOW_INSTANTIATION,
     .tp_dealloc = (destructor)uop_dealloc,
@@ -843,7 +744,6 @@ make_executor_from_uops(_PyUOpInstruction *buffer, _PyBloomFilter *dependencies)
         dest--;
     }
     assert(dest == -1);
-    executor->base.execute = _PyUOpExecute;
     _Py_ExecutorInit((_PyExecutorObject *)executor, dependencies);
 #ifdef Py_DEBUG
     char *python_lltrace = Py_GETENV("PYTHON_LLTRACE");
@@ -899,15 +799,6 @@ uop_optimize(
     return 1;
 }
 
-/* Dummy execute() function for UOp Executor.
- * The actual implementation is inlined in ceval.c,
- * in _PyEval_EvalFrameDefault(). */
-_Py_CODEUNIT *
-_PyUOpExecute(_PyExecutorObject *executor, _PyInterpreterFrame *frame, PyObject **stack_pointer)
-{
-    Py_FatalError("Tier 2 is now inlined into Tier 1");
-}
-
 static void
 uop_opt_dealloc(PyObject *self) {
     PyObject_Free(self);
@@ -937,6 +828,84 @@ PyUnstable_Optimizer_NewUOpOptimizer(void)
     return (PyObject *)opt;
 }
 
+static void
+counter_dealloc(_PyUOpExecutorObject *self) {
+    PyObject *opt = (PyObject *)self->trace[0].operand;
+    Py_DECREF(opt);
+    uop_dealloc(self);
+}
+
+PyTypeObject _PyCounterExecutor_Type = {
+    PyVarObject_HEAD_INIT(&PyType_Type, 0)
+    .tp_name = "counting_executor",
+    .tp_basicsize = offsetof(_PyUOpExecutorObject, trace),
+    .tp_itemsize = sizeof(_PyUOpInstruction),
+    .tp_flags = Py_TPFLAGS_DEFAULT | Py_TPFLAGS_DISALLOW_INSTANTIATION,
+    .tp_dealloc = (destructor)counter_dealloc,
+    .tp_methods = executor_methods,
+};
+
+static int
+counter_optimize(
+    _PyOptimizerObject* self,
+    PyCodeObject *code,
+    _Py_CODEUNIT *instr,
+    _PyExecutorObject **exec_ptr,
+    int Py_UNUSED(curr_stackentries)
+)
+{
+    _PyUOpInstruction buffer[3] = {
+        { .opcode = _LOAD_CONST_INLINE_BORROW, .operand = (uintptr_t)self },
+        { .opcode = _INTERNAL_INCREMENT_OPT_COUNTER },
+        { .opcode = _EXIT_TRACE, .target = (uint32_t)(instr - _PyCode_CODE(code)) }
+    };
+    _PyBloomFilter empty;
+    _Py_BloomFilter_Init(&empty);
+    _PyExecutorObject *executor = make_executor_from_uops(buffer, &empty);
+    if (executor == NULL) {
+        return -1;
+    }
+    Py_INCREF(self);
+    Py_SET_TYPE(executor, &_PyCounterExecutor_Type);
+    *exec_ptr = executor;
+    return 1;
+}
+
+static PyObject *
+counter_get_counter(PyObject *self, PyObject *args)
+{
+    return PyLong_FromLongLong(((_PyCounterOptimizerObject *)self)->count);
+}
+
+static PyMethodDef counter_optimizer_methods[] = {
+    { "get_count", counter_get_counter, METH_NOARGS, NULL },
+    { NULL, NULL },
+};
+
+PyTypeObject _PyCounterOptimizer_Type = {
+    PyVarObject_HEAD_INIT(&PyType_Type, 0)
+    .tp_name = "Counter optimizer",
+    .tp_basicsize = sizeof(_PyCounterOptimizerObject),
+    .tp_itemsize = 0,
+    .tp_flags = Py_TPFLAGS_DEFAULT | Py_TPFLAGS_DISALLOW_INSTANTIATION,
+    .tp_methods = counter_optimizer_methods,
+    .tp_dealloc = (destructor)PyObject_Del,
+};
+
+PyObject *
+PyUnstable_Optimizer_NewCounter(void)
+{
+    _PyCounterOptimizerObject *opt = (_PyCounterOptimizerObject *)_PyObject_New(&_PyCounterOptimizer_Type);
+    if (opt == NULL) {
+        return NULL;
+    }
+    opt->base.optimize = counter_optimize;
+    opt->base.resume_threshold = INT16_MAX;
+    opt->base.backedge_threshold = 0;
+    opt->count = 0;
+    return (PyObject *)opt;
+}
+
 
 /*****************************************
  *        Executor management