TOS caching

Author: markshannon

Include/internal/pycore_interpframe.h

@@ -43,7 +43,7 @@ static inline PyFunctionObject *_PyFrame_GetFunction(_PyInterpreterFrame *f) {
 }
 
 static inline _PyStackRef *_PyFrame_Stackbase(_PyInterpreterFrame *f) {
-    return (f->localsplus + _PyFrame_GetCode(f)->co_nlocalsplus);
+    return (f->localsplus + _PyFrame_GetCode(f)->co_nlocalsplus + 1);
 }
 
 static inline _PyStackRef _PyFrame_StackPeek(_PyInterpreterFrame *f) {
@@ -134,7 +134,7 @@ _PyFrame_Initialize(
     frame->f_builtins = func_obj->func_builtins;
     frame->f_globals = func_obj->func_globals;
     frame->f_locals = locals;
-    frame->stackpointer = frame->localsplus + code->co_nlocalsplus;
+    frame->stackpointer = frame->localsplus + code->co_nlocalsplus + 1;
     frame->frame_obj = NULL;
 #ifdef Py_GIL_DISABLED
     _PyFrame_InitializeTLBC(tstate, frame, code);
@@ -149,7 +149,7 @@ _PyFrame_Initialize(
     frame->lltrace = 0;
 #endif
 
-    for (int i = null_locals_from; i < code->co_nlocalsplus; i++) {
+    for (int i = null_locals_from; i <= code->co_nlocalsplus; i++) {
         frame->localsplus[i] = PyStackRef_NULL;
     }
 }
@@ -317,7 +317,7 @@ _PyFrame_PushTrampolineUnchecked(PyThreadState *tstate, PyCodeObject *code, int
 #endif
     frame->f_locals = NULL;
     assert(stackdepth <= code->co_stacksize);
-    frame->stackpointer = frame->localsplus + code->co_nlocalsplus + stackdepth;
+    frame->stackpointer = frame->localsplus + code->co_nlocalsplus + stackdepth + 1;
     frame->frame_obj = NULL;
 #ifdef Py_GIL_DISABLED
     _PyFrame_InitializeTLBC(tstate, frame, code);

Lib/test/test_sys.py

@@ -1666,9 +1666,9 @@ def func():
             return sys._getframe()
         x = func()
         if support.Py_GIL_DISABLED:
-            INTERPRETER_FRAME = '9PihcP'
+            INTERPRETER_FRAME = '9PihcPP'
         else:
-            INTERPRETER_FRAME = '9PhcP'
+            INTERPRETER_FRAME = '9PhcPP'
         check(x, size('3PiccPPP' + INTERPRETER_FRAME + 'P'))
         # function
         def func(): pass

Objects/codeobject.c

@@ -546,7 +546,7 @@ init_code(PyCodeObject *co, struct _PyCodeConstructor *con)
     /* derived values */
     co->co_nlocalsplus = nlocalsplus;
     co->co_nlocals = nlocals;
-    co->co_framesize = nlocalsplus + con->stacksize + FRAME_SPECIALS_SIZE;
+    co->co_framesize = nlocalsplus + con->stacksize + FRAME_SPECIALS_SIZE + 1;
     co->co_ncellvars = ncellvars;
     co->co_nfreevars = nfreevars;
 #ifdef Py_GIL_DISABLED

Python/bytecodes.c

@@ -3995,8 +3995,9 @@ dummy_func(
                 tstate, (PyCodeObject *)&_Py_InitCleanup, 1, frame);
             assert(_PyFrame_GetBytecode(shim)[0].op.code == EXIT_INIT_CHECK);
             assert(_PyFrame_GetBytecode(shim)[1].op.code == RETURN_VALUE);
+            shim->localsplus[0] = PyStackRef_NULL;
             /* Push self onto stack of shim */
-            shim->localsplus[0] = PyStackRef_DUP(self[0]);
+            shim->localsplus[1] = PyStackRef_DUP(self[0]);
             _PyInterpreterFrame *temp = _PyEvalFramePushAndInit(
                 tstate, init[0], NULL, args-1, oparg+1, NULL, shim);
             DEAD(init);

Python/ceval.c

@@ -190,13 +190,17 @@ dump_stack(_PyInterpreterFrame *frame, _PyStackRef *stack_pointer)
 
 static void
 lltrace_instruction(_PyInterpreterFrame *frame,
+                    _PyStackRef _tos,
                     _PyStackRef *stack_pointer,
                     _Py_CODEUNIT *next_instr,
                     int opcode,
                     int oparg)
 {
     int offset = 0;
     if (frame->owner < FRAME_OWNED_BY_INTERPRETER) {
+        printf("_tos = ");
+        dump_item(_tos);
+        printf("; ");
         dump_stack(frame, stack_pointer);
         offset = (int)(next_instr - _PyFrame_GetBytecode(frame));
     }
@@ -943,6 +947,13 @@ _PyObjectArray_Free(PyObject **array, PyObject **scratch)
 /* This setting is reversed below following _PyEval_EvalFrameDefault */
 #endif
 
+#ifdef Py_GIL_DISABLED
+#define Py_FREE_THREADING 1
+#else
+#define Py_FREE_THREADING 0
+#endif
+
+
 #if Py_TAIL_CALL_INTERP
 #include "opcode_targets.h"
 #include "generated_cases.c.h"
@@ -992,7 +1003,7 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
 #endif
     entry_frame.f_executable = PyStackRef_None;
     entry_frame.instr_ptr = (_Py_CODEUNIT *)_Py_INTERPRETER_TRAMPOLINE_INSTRUCTIONS + 1;
-    entry_frame.stackpointer = entry_frame.localsplus;
+    entry_frame.stackpointer = entry_frame.localsplus + 1;
     entry_frame.owner = FRAME_OWNED_BY_INTERPRETER;
     entry_frame.visited = 0;
     entry_frame.return_offset = 0;
@@ -1039,6 +1050,7 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
     _PyExecutorObject *current_executor = NULL;
     const _PyUOpInstruction *next_uop = NULL;
 #endif
+    _PyStackRef _tos = PyStackRef_NULL;
 #if Py_TAIL_CALL_INTERP
     return _TAIL_CALL_start_frame(frame, NULL, tstate, NULL, 0);
 #else

Python/ceval_macros.h

@@ -110,7 +110,7 @@
 /* PRE_DISPATCH_GOTO() does lltrace if enabled. Normally a no-op */
 #ifdef Py_DEBUG
 #define PRE_DISPATCH_GOTO() if (frame->lltrace >= 5) { \
-    lltrace_instruction(frame, stack_pointer, next_instr, opcode, oparg); }
+    lltrace_instruction(frame, _tos, stack_pointer, next_instr, opcode, oparg); }
 #else
 #define PRE_DISPATCH_GOTO() ((void)0)
 #endif
@@ -198,7 +198,7 @@ GETITEM(PyObject *v, Py_ssize_t i) {
 #define STACK_SIZE()      (_PyFrame_GetCode(frame)->co_stacksize)
 
 #define WITHIN_STACK_BOUNDS() \
-   (frame->owner == FRAME_OWNED_BY_INTERPRETER || (STACK_LEVEL() >= 0 && STACK_LEVEL() <= STACK_SIZE()))
+   (frame->owner == FRAME_OWNED_BY_INTERPRETER || (STACK_LEVEL() >= -1 && STACK_LEVEL() <= STACK_SIZE()))
 
 /* Data access macros */
 #define FRAME_CO_CONSTS (_PyFrame_GetCode(frame)->co_consts)