Get tos caching working with exit-at-end. Add some more tracing info

Author: markshannon

Include/internal/pycore_stackref.h

@@ -310,6 +310,16 @@ PyStackRef_IsValid(_PyStackRef ref)
     return ref.bits >= Py_INT_TAG;
 }
 
+static inline bool
+PyStackRef_IsWrapped(_PyStackRef ref)
+{
+#ifdef Py_DEBUG
+    return (ref.bits & Py_TAG_INVALID) != 0 && !PyStackRef_IsError(ref);
+#else
+    Py_FatalError("Cannot determine if value is wrapped if Py_DEBUG is not set");
+#endif
+}
+
 static inline bool
 PyStackRef_IsTaggedInt(_PyStackRef i)
 {

Include/internal/pycore_uop_metadata.h

@@ -5349,6 +5349,9 @@ dummy_func(
         tier2 op(_START_EXECUTOR, (executor/4 --)) {
 #ifndef _Py_JIT
             current_executor = (_PyExecutorObject*)executor;
+#endif
+#ifdef Py_DEBUG
+            assert(STACK_LEVEL() + current_cached_values == oparg);
 #endif
             assert(tstate->jit_exit == NULL || tstate->jit_exit->executor == current_executor);
             tstate->current_executor = (PyObject *)executor;

Python/bytecodes.c

@@ -164,26 +164,23 @@ dump_item(_PyStackRef item)
         printf("%" PRId64, (int64_t)PyStackRef_UntagInt(item));
         return;
     }
-    if (PyStackRef_IsValid(item)) {
-        PyObject *obj = PyStackRef_AsPyObjectBorrow(item);
-        if (obj == NULL) {
-            printf("<nil>");
-            return;
-        }
-        // Don't call __repr__(), it might recurse into the interpreter.
-        printf("<%s at %p>", Py_TYPE(obj)->tp_name, (void *)obj);
+    if (PyStackRef_IsWrapped(item)) {
+        void *ptr = PyStackRef_Unwrap(item);
+        printf("Wrapped(pointer %p)", ptr);
+        return;
     }
-    else {
-        /* Already handled NULL */
-        if (PyStackRef_IsError(item)) {
-            printf("ERROR");
-        }
-        else {
-            // Wrapped item
-            void *ptr = PyStackRef_Unwrap(item);
-            printf("Wrapped(pointer %p)", ptr);
-        }
+    if (PyStackRef_IsError(item)) {
+        printf("ERROR");
+        return;
+    }
+    assert(PyStackRef_IsValid(item));
+    PyObject *obj = PyStackRef_AsPyObjectBorrow(item);
+    if (obj == NULL) {
+        printf("<nil>");
+        return;
     }
+    // Don't call __repr__(), it might recurse into the interpreter.
+    printf("<%s at %p>", Py_TYPE(obj)->tp_name, (void *)obj);
 }
 
 static void
@@ -214,11 +211,21 @@ dump_stack(_PyInterpreterFrame *frame, _PyStackRef *stack_pointer)
         }
         printf("]\n");
     }
-    fflush(stdout);
     PyErr_SetRaisedException(exc);
     _PyFrame_GetStackPointer(frame);
 }
 
+static void
+dump_cache_item(_PyStackRef cache, int position, int depth)
+{
+    if (position < depth) {
+        dump_item(cache);
+    }
+    else {
+        printf("---");
+    }
+}
+
 static void
 lltrace_instruction(_PyInterpreterFrame *frame,
                     _PyStackRef *stack_pointer,
@@ -1185,6 +1192,13 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
 #ifdef Py_DEBUG
         if (frame->lltrace >= 3) {
             dump_stack(frame, stack_pointer);
+            printf("    cache=[");
+            dump_cache_item(_tos_cache0, 0, current_cached_values);
+            printf(", ");
+            dump_cache_item(_tos_cache1, 1, current_cached_values);
+            printf(", ");
+            dump_cache_item(_tos_cache2, 2, current_cached_values);
+            printf("]\n");
             if (next_uop->opcode == _START_EXECUTOR_r00) {
                 printf("%4d uop: ", 0);
             }
@@ -1193,6 +1207,7 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
             }
             _PyUOpPrint(next_uop);
             printf("\n");
+            fflush(stdout);
         }
 #endif
         next_uop++;

Python/ceval.c

@@ -212,6 +212,9 @@ GETITEM(PyObject *v, Py_ssize_t i) {
 #define WITHIN_STACK_BOUNDS() \
    (frame->owner == FRAME_OWNED_BY_INTERPRETER || (STACK_LEVEL() >= 0 && STACK_LEVEL() <= STACK_SIZE()))
 
+#define WITHIN_STACK_BOUNDS_WITH_CACHE() \
+   (frame->owner == FRAME_OWNED_BY_INTERPRETER || (STACK_LEVEL() >= 0 && (STACK_LEVEL() + current_cached_values) <= STACK_SIZE()))
+
 /* Data access macros */
 #define FRAME_CO_CONSTS (_PyFrame_GetCode(frame)->co_consts)
 #define FRAME_CO_NAMES  (_PyFrame_GetCode(frame)->co_names)