bpo-46329: Avoid PUSH_NULL when followed by LOAD_GLOBAL

Doc/library/dis.rst

@@ -43,7 +43,7 @@ the following command can be used to display the disassembly of
      1           0 RESUME                   0
 
      2           2 PUSH_NULL
-                 4 LOAD_GLOBAL              0 (len)
+                 4 LOAD_GLOBAL              1 (NULL + len)
                  6 LOAD_FAST                0 (alist)
                  8 PRECALL                  1
                 10 CALL                     1
@@ -996,8 +996,11 @@ iterations of the loop.
 
 .. opcode:: LOAD_GLOBAL (namei)
 
-   Loads the global named ``co_names[namei]`` onto the stack.
+   Loads the global named ``co_names[namei>>1]`` onto the stack.
 
+   .. versionchanged:: 3.11
+      If the low bit of ``namei`` is set, then a ``NULL`` is pushed to the
+      stack before the global variable.
 
 .. opcode:: LOAD_FAST (var_num)
 

Lib/dis.py

@@ -28,6 +28,7 @@
 MAKE_FUNCTION_FLAGS = ('defaults', 'kwdefaults', 'annotations', 'closure')
 
 LOAD_CONST = opmap['LOAD_CONST']
+LOAD_GLOBAL = opmap['LOAD_GLOBAL']
 BINARY_OP = opmap['BINARY_OP']
 
 CACHE = opmap["CACHE"]
@@ -430,7 +431,12 @@ def _get_instructions_bytes(code, varname_from_oparg=None,
             if op in hasconst:
                 argval, argrepr = _get_const_info(op, arg, co_consts)
             elif op in hasname:
-                argval, argrepr = _get_name_info(arg, get_name)
+                if op == LOAD_GLOBAL:
+                    argval, argrepr = _get_name_info(arg//2, get_name)
+                    if (arg & 1) and argrepr:
+                        argrepr = "NULL + " + argrepr
+                else:
+                    argval, argrepr = _get_name_info(arg, get_name)
             elif op in hasjabs:
                 argval = arg*2
                 argrepr = "to " + repr(argval)

Lib/importlib/_bootstrap_external.py

@@ -395,6 +395,7 @@ def _write_atomic(path, data, mode=0o666):
 #     Python 3.11a5 3485 (Add an oparg to GET_AWAITABLE)
 #     Python 3.11a6 3486 (Use inline caching for PRECALL and CALL)
 #     Python 3.11a6 3487 (Remove the adaptive "oparg counter" mechanism)
+#     Python 3.11a6 3488 (LOAD_GLOBAL can push additional NULL)
 
 #     Python 3.12 will start with magic number 3500
 
@@ -409,7 +410,7 @@ def _write_atomic(path, data, mode=0o666):
 # Whenever MAGIC_NUMBER is changed, the ranges in the magic_values array
 # in PC/launcher.c must also be updated.
 
-MAGIC_NUMBER = (3487).to_bytes(2, 'little') + b'\r\n'
+MAGIC_NUMBER = (3488).to_bytes(2, 'little') + b'\r\n'
 _RAW_MAGIC_NUMBER = int.from_bytes(MAGIC_NUMBER, 'little')  # For import.c
 
 _PYCACHE = '__pycache__'

Misc/NEWS.d/next/Core and Builtins/2022-03-16-12-19-25.bpo-46329.9oS0HT.rst

@@ -0,0 +1,2 @@
+Use low bit of ``LOAD_GLOBAL`` to indicate whether to push a ``NULL`` before
+the global. Helps streamline the call sequence a bit.

Python/ceval.c

@@ -2947,7 +2947,9 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
 
         TARGET(LOAD_GLOBAL) {
             PREDICTED(LOAD_GLOBAL);
-            PyObject *name = GETITEM(names, oparg);
+            int push_null = oparg & 1;
+            PEEK(0) = NULL;
+            PyObject *name = GETITEM(names, oparg>>1);
             PyObject *v;
             if (PyDict_CheckExact(GLOBALS())
                 && PyDict_CheckExact(BUILTINS()))
@@ -2970,7 +2972,6 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
                 /* Slow-path if globals or builtins is not a dict */
 
                 /* namespace 1: globals */
-                name = GETITEM(names, oparg);
                 v = PyObject_GetItem(GLOBALS(), name);
                 if (v == NULL) {
                     if (!_PyErr_ExceptionMatches(tstate, PyExc_KeyError)) {
@@ -2992,6 +2993,7 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
             }
             /* Skip over inline cache */
             JUMPBY(INLINE_CACHE_ENTRIES_LOAD_GLOBAL);
+            STACK_GROW(push_null);
             PUSH(v);
             DISPATCH();
         }
@@ -3000,7 +3002,7 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
             assert(cframe.use_tracing == 0);
             _PyLoadGlobalCache *cache = (_PyLoadGlobalCache *)next_instr;
             if (cache->counter == 0) {
-                PyObject *name = GETITEM(names, oparg);
+                PyObject *name = GETITEM(names, oparg>>1);
                 next_instr--;
                 if (_Py_Specialize_LoadGlobal(GLOBALS(), BUILTINS(), next_instr, name) < 0) {
                     goto error;
@@ -3025,10 +3027,13 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
             PyDictUnicodeEntry *entries = DK_UNICODE_ENTRIES(dict->ma_keys);
             PyObject *res = entries[cache->index].me_value;
             DEOPT_IF(res == NULL, LOAD_GLOBAL);
+            int push_null = oparg & 1;
+            PEEK(0) = NULL;
             JUMPBY(INLINE_CACHE_ENTRIES_LOAD_GLOBAL);
             STAT_INC(LOAD_GLOBAL, hit);
+            STACK_GROW(push_null+1);
             Py_INCREF(res);
-            PUSH(res);
+            SET_TOP(res);
             NOTRACE_DISPATCH();
         }
 
@@ -3047,10 +3052,13 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
             PyDictUnicodeEntry *entries = DK_UNICODE_ENTRIES(bdict->ma_keys);
             PyObject *res = entries[cache->index].me_value;
             DEOPT_IF(res == NULL, LOAD_GLOBAL);
+            int push_null = oparg & 1;
+            PEEK(0) = NULL;
             JUMPBY(INLINE_CACHE_ENTRIES_LOAD_GLOBAL);
             STAT_INC(LOAD_GLOBAL, hit);
+            STACK_GROW(push_null+1);
             Py_INCREF(res);
-            PUSH(res);
+            SET_TOP(res);
             NOTRACE_DISPATCH();
         }
 

Python/compile.c

@@ -1000,7 +1000,7 @@ stack_effect(int opcode, int oparg, int jump)
             return -1;
 
         case LOAD_GLOBAL:
-            return 1;
+            return (oparg & 1) + 1;
 
         /* Exception handling pseudo-instructions */
         case SETUP_FINALLY:
@@ -4185,8 +4185,12 @@ compiler_nameop(struct compiler *c, identifier name, expr_context_ty ctx)
     assert(op);
     arg = compiler_add_o(dict, mangled);
     Py_DECREF(mangled);
-    if (arg < 0)
+    if (arg < 0) {
         return 0;
+    }
+    if (op == LOAD_GLOBAL) {
+        arg <<= 1;
+    }
     return compiler_addop_i(c, op, arg);
 }
 
@@ -8812,6 +8816,13 @@ optimize_basic_block(struct compiler *c, basicblock *bb, PyObject *consts)
                 break;
             case KW_NAMES:
                 break;
+            case PUSH_NULL:
+                if (nextop == LOAD_GLOBAL && (inst[1].i_opcode & 1) == 0) {
+                    inst->i_opcode = NOP;
+                    inst->i_oparg = 0;
+                    inst[1].i_oparg |= 1;
+                }
+                break;
             default:
                 /* All HAS_CONST opcodes should be handled with LOAD_CONST */
                 assert (!HAS_CONST(inst->i_opcode));