bpo-43693: Add new internal code objects fields: co_fastlocalnames and co_fastlocalkinds. (pythongh-26388)

A number of places in the code base (notably ceval.c and frameobject.c) rely on mapping variable names to indices in the frame "locals plus" array (AKA fast locals), and thus opargs.  Currently the compiler indirectly encodes that information on the code object as the tuples co_varnames, co_cellvars, and co_freevars.  At runtime the dependent code must calculate the proper mapping from those, which isn't ideal and impacts performance-sensitive sections.  This is something we can easily address in the compiler instead.

This change addresses the situation by replacing internal use of co_varnames, etc. with a single combined tuple of names in locals-plus order, along with a minimal array mapping each to its kind (local vs. cell vs. free).  These two new PyCodeObject fields, co_fastlocalnames and co_fastllocalkinds, are not exposed to Python code for now, but co_varnames, etc. are still available with the same values as before (though computed lazily).

Aside from the (mild) performance impact, there are a number of other benefits:

* there's now a clear, direct relationship between locals-plus and variables
* code that relies on the locals-plus-to-name mapping is simpler
* marshaled code objects are smaller and serialize/de-serialize faster

Also note that we can take this approach further by expanding the possible values in co_fastlocalkinds to include specific argument types (e.g. positional-only, kwargs).  Doing so would allow further speed-ups in _PyEval_MakeFrameVector(), which is where args get unpacked into the locals-plus array.  It would also allow us to shrink marshaled code objects even further.

https://bugs.python.org/issue43693

Author: ericsnowcurrently

Doc/library/dis.rst

@@ -1059,9 +1059,8 @@ All of the following opcodes use their arguments.
 .. opcode:: LOAD_CLOSURE (i)
 
    Pushes a reference to the cell contained in slot *i* of the cell and free
-   variable storage.  The name of the variable is ``co_cellvars[i]`` if *i* is
-   less than the length of *co_cellvars*.  Otherwise it is ``co_freevars[i -
-   len(co_cellvars)]``.
+   variable storage.  The name of the variable is
+   ``co_fastlocalnames[i + len(co_varnames)]``.
 
 
 .. opcode:: LOAD_DEREF (i)

Include/cpython/code.h

@@ -3,6 +3,8 @@
 #endif
 
 typedef uint16_t _Py_CODEUNIT;
+// Each oparg must fit in the second half of _Py_CODEUNIT, hence 8 bits.
+#define _Py_MAX_OPARG 255
 
 #ifdef WORDS_BIGENDIAN
 #  define _Py_OPCODE(word) ((word) >> 8)
@@ -14,6 +16,11 @@ typedef uint16_t _Py_CODEUNIT;
 
 typedef struct _PyOpcache _PyOpcache;
 
+
+// These are duplicated from pycore_code.h.
+typedef unsigned char _PyLocalsPlusKind;
+typedef _PyLocalsPlusKind *_PyLocalsPlusKinds;
+
 /* Bytecode object */
 struct PyCodeObject {
     PyObject_HEAD
@@ -53,9 +60,8 @@ struct PyCodeObject {
     int co_kwonlyargcount;      /* #keyword only arguments */
     int co_stacksize;           /* #entries needed for evaluation stack */
     int co_firstlineno;         /* first source line number */
-    PyObject *co_varnames;      /* tuple of strings (local variable names) */
-    PyObject *co_cellvars;      /* tuple of strings (cell variable names) */
-    PyObject *co_freevars;      /* tuple of strings (free variable names) */
+    PyObject *co_localsplusnames;  /* tuple mapping offsets to names */
+    _PyLocalsPlusKinds co_localspluskinds; /* array mapping to local kinds */
     PyObject *co_filename;      /* unicode (where it was loaded from) */
     PyObject *co_name;          /* unicode (name, for reference) */
     PyObject *co_linetable;     /* string (encoding addr<->lineno mapping) See
@@ -65,11 +71,15 @@ struct PyCodeObject {
     /* These fields are set with computed values on new code objects. */
 
     int *co_cell2arg;           /* Maps cell vars which are arguments. */
-    // These are redundant but offer some performance benefit.
+    // redundant values (derived from co_localsplusnames and co_localspluskinds)
     int co_nlocalsplus;         /* number of local + cell + free variables */
     int co_nlocals;             /* number of local variables */
     int co_ncellvars;           /* number of cell variables */
     int co_nfreevars;           /* number of free variables */
+    // lazily-computed values
+    PyObject *co_varnames;      /* tuple of strings (local variable names) */
+    PyObject *co_cellvars;      /* tuple of strings (cell variable names) */
+    PyObject *co_freevars;      /* tuple of strings (free variable names) */
 
     /* The remaining fields are zeroed out on new code objects. */
 
@@ -143,7 +153,7 @@ struct PyCodeObject {
 PyAPI_DATA(PyTypeObject) PyCode_Type;
 
 #define PyCode_Check(op) Py_IS_TYPE(op, &PyCode_Type)
-#define PyCode_GetNumFree(op) (PyTuple_GET_SIZE((op)->co_freevars))
+#define PyCode_GetNumFree(op) ((op)->co_nfreevars)
 
 /* Public interface */
 PyAPI_FUNC(PyCodeObject *) PyCode_New(

Include/internal/pycore_code.h

@@ -26,6 +26,57 @@ struct _PyOpcache {
 };
 
 
+/* "Locals plus" for a code object is the set of locals + cell vars +
+ * free vars.  This relates to variable names as well as offsets into
+ * the "fast locals" storage array of execution frames.  The compiler
+ * builds the list of names, their offsets, and the corresponding
+ * kind of local.
+ *
+ * Those kinds represent the source of the initial value and the
+ * variable's scope (as related to closures).  A "local" is an
+ * argument or other variable defined in the current scope.  A "free"
+ * variable is one that is defined in an outer scope and comes from
+ * the function's closure.  A "cell" variable is a local that escapes
+ * into an inner function as part of a closure, and thus must be
+ * wrapped in a cell.  Any "local" can also be a "cell", but the
+ * "free" kind is mutually exclusive with both.
+ */
+
+// We would use an enum if C let us specify the storage type.
+typedef unsigned char _PyLocalsPlusKind;
+/* Note that these all fit within _PyLocalsPlusKind, as do combinations. */
+// Later, we will use the smaller numbers to differentiate the different
+// kinds of locals (e.g. pos-only arg, varkwargs, local-only).
+#define CO_FAST_LOCAL   0x20
+#define CO_FAST_CELL    0x40
+#define CO_FAST_FREE    0x80
+
+typedef _PyLocalsPlusKind *_PyLocalsPlusKinds;
+
+static inline int
+_PyCode_InitLocalsPlusKinds(int num, _PyLocalsPlusKinds *pkinds)
+{
+    if (num == 0) {
+        *pkinds = NULL;
+        return 0;
+    }
+    _PyLocalsPlusKinds kinds = PyMem_NEW(_PyLocalsPlusKind, num);
+    if (kinds == NULL) {
+        PyErr_NoMemory();
+        return -1;
+    }
+    *pkinds = kinds;
+    return 0;
+}
+
+static inline void
+_PyCode_ClearLocalsPlusKinds(_PyLocalsPlusKinds kinds)
+{
+    if (kinds != NULL) {
+        PyMem_Free(kinds);
+    }
+}
+
 struct _PyCodeConstructor {
     /* metadata */
     PyObject *filename;
@@ -42,13 +93,13 @@ struct _PyCodeConstructor {
     PyObject *names;
 
     /* mapping frame offsets to information */
-    PyObject *varnames;
-    PyObject *cellvars;
-    PyObject *freevars;
+    PyObject *localsplusnames;
+    _PyLocalsPlusKinds localspluskinds;
 
     /* args (within varnames) */
     int argcount;
     int posonlyargcount;
+    // XXX Replace argcount with posorkwargcount (argcount - posonlyargcount).
     int kwonlyargcount;
 
     /* needed to create the frame */
@@ -75,6 +126,11 @@ PyAPI_FUNC(PyCodeObject *) _PyCode_New(struct _PyCodeConstructor *);
 
 int _PyCode_InitOpcache(PyCodeObject *co);
 
+/* Getters for internal PyCodeObject data. */
+PyAPI_FUNC(PyObject *) _PyCode_GetVarnames(PyCodeObject *);
+PyAPI_FUNC(PyObject *) _PyCode_GetCellvars(PyCodeObject *);
+PyAPI_FUNC(PyObject *) _PyCode_GetFreevars(PyCodeObject *);
+
 
 #ifdef __cplusplus
 }

Lib/ctypes/test/test_values.py

@@ -80,9 +80,9 @@ class struct_frozen(Structure):
                 continue
             items.append((entry.name.decode("ascii"), entry.size))
 
-        expected = [("__hello__", 138),
-                    ("__phello__", -138),
-                    ("__phello__.spam", 138),
+        expected = [("__hello__", 128),
+                    ("__phello__", -128),
+                    ("__phello__.spam", 128),
                     ]
         self.assertEqual(items, expected, "PyImport_FrozenModules example "
             "in Doc/library/ctypes.rst may be out of date")

Lib/dis.py

@@ -273,15 +273,15 @@ def get_instructions(x, *, first_line=None):
     the disassembled code object.
     """
     co = _get_code_object(x)
-    cell_names = co.co_cellvars + co.co_freevars
     linestarts = dict(findlinestarts(co))
     if first_line is not None:
         line_offset = first_line - co.co_firstlineno
     else:
         line_offset = 0
-    return _get_instructions_bytes(co.co_code, co.co_varnames, co.co_names,
-                                   co.co_consts, cell_names, linestarts,
-                                   line_offset)
+    return _get_instructions_bytes(co.co_code,
+                                   co._varname_from_oparg,
+                                   co.co_names, co.co_consts,
+                                   linestarts, line_offset)
 
 def _get_const_info(const_index, const_list):
     """Helper to get optional details about const references
@@ -295,16 +295,16 @@ def _get_const_info(const_index, const_list):
         argval = const_list[const_index]
     return argval, repr(argval)
 
-def _get_name_info(name_index, name_list):
+def _get_name_info(name_index, get_name, **extrainfo):
     """Helper to get optional details about named references
 
        Returns the dereferenced name as both value and repr if the name
        list is defined.
        Otherwise returns the name index and its repr().
     """
     argval = name_index
-    if name_list is not None:
-        argval = name_list[name_index]
+    if get_name is not None:
+        argval = get_name(name_index, **extrainfo)
         argrepr = argval
     else:
         argrepr = repr(argval)
@@ -336,8 +336,10 @@ def parse_exception_table(code):
     except StopIteration:
         return entries
 
-def _get_instructions_bytes(code, varnames=None, names=None, constants=None,
-                      cells=None, linestarts=None, line_offset=0, exception_entries=()):
+def _get_instructions_bytes(code, varname_from_oparg=None,
+                            names=None, constants=None,
+                            linestarts=None, line_offset=0,
+                            exception_entries=()):
     """Iterate over the instructions in a bytecode string.
 
     Generates a sequence of Instruction namedtuples giving the details of each
@@ -346,6 +348,7 @@ def _get_instructions_bytes(code, varnames=None, names=None, constants=None,
     arguments.
 
     """
+    get_name = None if names is None else names.__getitem__
     labels = set(findlabels(code))
     for start, end, target, _, _ in exception_entries:
         for i in range(start, end):
@@ -368,20 +371,21 @@ def _get_instructions_bytes(code, varnames=None, names=None, constants=None,
             if op in hasconst:
                 argval, argrepr = _get_const_info(arg, constants)
             elif op in hasname:
-                argval, argrepr = _get_name_info(arg, names)
+                argval, argrepr = _get_name_info(arg, get_name)
             elif op in hasjabs:
                 argval = arg*2
                 argrepr = "to " + repr(argval)
             elif op in hasjrel:
                 argval = offset + 2 + arg*2
                 argrepr = "to " + repr(argval)
             elif op in haslocal:
-                argval, argrepr = _get_name_info(arg, varnames)
+                argval, argrepr = _get_name_info(arg, varname_from_oparg)
             elif op in hascompare:
                 argval = cmp_op[arg]
                 argrepr = argval
             elif op in hasfree:
-                argval, argrepr = _get_name_info(arg, cells)
+                argval, argrepr = _get_name_info(arg, varname_from_oparg,
+                                                 cell=True)
             elif op == FORMAT_VALUE:
                 argval, argrepr = FORMAT_VALUE_CONVERTERS[arg & 0x3]
                 argval = (argval, bool(arg & 0x4))
@@ -398,11 +402,11 @@ def _get_instructions_bytes(code, varnames=None, names=None, constants=None,
 
 def disassemble(co, lasti=-1, *, file=None):
     """Disassemble a code object."""
-    cell_names = co.co_cellvars + co.co_freevars
     linestarts = dict(findlinestarts(co))
     exception_entries = parse_exception_table(co)
-    _disassemble_bytes(co.co_code, lasti, co.co_varnames, co.co_names,
-                       co.co_consts, cell_names, linestarts, file=file,
+    _disassemble_bytes(co.co_code, lasti,
+                       co._varname_from_oparg,
+                       co.co_names, co.co_consts, linestarts, file=file,
                        exception_entries=exception_entries)
 
 def _disassemble_recursive(co, *, file=None, depth=None):
@@ -416,8 +420,8 @@ def _disassemble_recursive(co, *, file=None, depth=None):
                 print("Disassembly of %r:" % (x,), file=file)
                 _disassemble_recursive(x, file=file, depth=depth)
 
-def _disassemble_bytes(code, lasti=-1, varnames=None, names=None,
-                       constants=None, cells=None, linestarts=None,
+def _disassemble_bytes(code, lasti=-1, varname_from_oparg=None,
+                       names=None, constants=None, linestarts=None,
                        *, file=None, line_offset=0, exception_entries=()):
     # Omit the line number column entirely if we have no line number info
     show_lineno = bool(linestarts)
@@ -434,8 +438,8 @@ def _disassemble_bytes(code, lasti=-1, varnames=None, names=None,
         offset_width = len(str(maxoffset))
     else:
         offset_width = 4
-    for instr in _get_instructions_bytes(code, varnames, names,
-                                         constants, cells, linestarts,
+    for instr in _get_instructions_bytes(code, varname_from_oparg, names,
+                                         constants, linestarts,
                                          line_offset=line_offset, exception_entries=exception_entries):
         new_source_line = (show_lineno and
                            instr.starts_line is not None and
@@ -517,16 +521,16 @@ def __init__(self, x, *, first_line=None, current_offset=None):
         else:
             self.first_line = first_line
             self._line_offset = first_line - co.co_firstlineno
-        self._cell_names = co.co_cellvars + co.co_freevars
         self._linestarts = dict(findlinestarts(co))
         self._original_object = x
         self.current_offset = current_offset
         self.exception_entries = parse_exception_table(co)
 
     def __iter__(self):
         co = self.codeobj
-        return _get_instructions_bytes(co.co_code, co.co_varnames, co.co_names,
-                                       co.co_consts, self._cell_names,
+        return _get_instructions_bytes(co.co_code,
+                                       co._varname_from_oparg,
+                                       co.co_names, co.co_consts,
                                        self._linestarts,
                                        line_offset=self._line_offset,
                                        exception_entries=self.exception_entries)
@@ -554,9 +558,9 @@ def dis(self):
         else:
             offset = -1
         with io.StringIO() as output:
-            _disassemble_bytes(co.co_code, varnames=co.co_varnames,
+            _disassemble_bytes(co.co_code,
+                               varname_from_oparg=co._varname_from_oparg,
                                names=co.co_names, constants=co.co_consts,
-                               cells=self._cell_names,
                                linestarts=self._linestarts,
                                line_offset=self._line_offset,
                                file=output,

Lib/importlib/_bootstrap_external.py

@@ -355,6 +355,7 @@ def _write_atomic(path, data, mode=0o666):
 #     Python 3.11a1 3450 Use exception table for unwinding ("zero cost" exception handling)
 #     Python 3.11a1 3451 (Add CALL_METHOD_KW)
 #     Python 3.11a1 3452 (drop nlocals from marshaled code objects)
+#     Python 3.11a1 3453 (add co_fastlocalnames and co_fastlocalkinds)
 
 #
 # MAGIC must change whenever the bytecode emitted by the compiler may no
@@ -364,7 +365,7 @@ 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 = (3452).to_bytes(2, 'little') + b'\r\n'
+MAGIC_NUMBER = (3453).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/2021-05-21-20-53-49.bpo-43693.-NN3J_.rst

@@ -0,0 +1,3 @@
+``PyCodeObject`` gained ``co_fastlocalnames`` and ``co_fastlocalkinds`` as
+the the authoritative source of fast locals info.  Marshaled code objects
+have changed accordingly.