GH-108866: Guarantee forward progress in executors. (GH-113006)

Author: markshannon

Include/cpython/optimizer.h

@@ -32,7 +32,7 @@ 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. */
-    struct _PyInterpreterFrame *(*execute)(struct _PyExecutorObject *self, struct _PyInterpreterFrame *frame, PyObject **stack_pointer);
+    _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;

Include/internal/pycore_opcode_metadata.h

@@ -24,7 +24,7 @@ typedef struct {
     _PyUOpInstruction trace[1];
 } _PyUOpExecutorObject;
 
-_PyInterpreterFrame *_PyUOpExecute(
+_Py_CODEUNIT *_PyUOpExecute(
     _PyExecutorObject *executor,
     _PyInterpreterFrame *frame,
     PyObject **stack_pointer);

Include/internal/pycore_uops.h

@@ -0,0 +1,3 @@
+Change the API and contract of ``_PyExecutorObject`` to return the
+next_instr pointer, instead of the frame, and to always execute at least one
+instruction.

Misc/NEWS.d/next/Core and Builtins/2023-12-12-04-53-19.gh-issue-108866.xbJ-9a.rst

@@ -2352,20 +2352,17 @@ dummy_func(
 
             PyCodeObject *code = _PyFrame_GetCode(frame);
             _PyExecutorObject *executor = (_PyExecutorObject *)code->co_executors->executors[oparg&255];
-            int original_oparg = executor->vm_data.oparg | (oparg & 0xfffff00);
-            JUMPBY(1-original_oparg);
-            frame->instr_ptr = next_instr;
             Py_INCREF(executor);
             if (executor->execute == _PyUOpExecute) {
                 current_executor = (_PyUOpExecutorObject *)executor;
                 GOTO_TIER_TWO();
             }
-            frame = executor->execute(executor, frame, stack_pointer);
-            if (frame == NULL) {
-                frame = tstate->current_frame;
+            next_instr = executor->execute(executor, frame, stack_pointer);
+            frame = tstate->current_frame;
+            if (next_instr == NULL) {
                 goto resume_with_error;
             }
-            goto enter_tier_one;
+            stack_pointer = _PyFrame_GetStackPointer(frame);
         }
 
         replaced op(_POP_JUMP_IF_FALSE, (unused/1, cond -- )) {

Python/bytecodes.c

@@ -167,6 +167,7 @@ _PyOptimizer_BackEdge(_PyInterpreterFrame *frame, _Py_CODEUNIT *src, _Py_CODEUNI
     }
     _PyOptimizerObject *opt = interp->optimizer;
     _PyExecutorObject *executor = NULL;
+    /* Start optimizing at the destination to guarantee forward progress */
     int err = opt->optimize(opt, code, dest, &executor, (int)(stack_pointer - _PyFrame_Stackbase(frame)));
     if (err <= 0) {
         assert(executor == NULL);
@@ -247,14 +248,13 @@ PyTypeObject _PyCounterExecutor_Type = {
     .tp_methods = executor_methods,
 };
 
-static _PyInterpreterFrame *
+static _Py_CODEUNIT *
 counter_execute(_PyExecutorObject *self, _PyInterpreterFrame *frame, PyObject **stack_pointer)
 {
     ((_PyCounterExecutorObject *)self)->optimizer->count++;
     _PyFrame_SetStackPointer(frame, stack_pointer);
-    frame->instr_ptr = ((_PyCounterExecutorObject *)self)->next_instr;
     Py_DECREF(self);
-    return frame;
+    return ((_PyCounterExecutorObject *)self)->next_instr;
 }
 
 static int
@@ -891,7 +891,7 @@ uop_optimize(
 /* Dummy execute() function for UOp Executor.
  * The actual implementation is inlined in ceval.c,
  * in _PyEval_EvalFrameDefault(). */
-_PyInterpreterFrame *
+_Py_CODEUNIT *
 _PyUOpExecute(_PyExecutorObject *executor, _PyInterpreterFrame *frame, PyObject **stack_pointer)
 {
     Py_FatalError("Tier 2 is now inlined into Tier 1");