Here is my modified version of PEP 367. The reference implementation in it is pretty long, and should probably be split out to somewhere else (esp. since it can't fully implement the semantics). Cheers, Tim Delaney PEP: 367 Title: New Super Version: $Revision$ Last-Modified: $Date$ Author: Calvin Spealman <ironfroggy@gmail.com> Author: Tim Delaney <timothy.c.delaney@gmail.com> Status: Draft Type: Standards Track Content-Type: text/x-rst Created: 28-Apr-2007 Python-Version: 2.6 Post-History: 28-Apr-2007, 29-Apr-2007 (1), 29-Apr-2007 (2), 14-May-2007 Abstract ======== This PEP proposes syntactic sugar for use of the ``super`` type to automatically construct instances of the super type binding to the class that a method was defined in, and the instance (or class object for classmethods) that the method is currently acting upon. The premise of the new super usage suggested is as follows:: super.foo(1, 2) to replace the old:: super(Foo, self).foo(1, 2) and the current ``__builtin__.super`` be aliased to ``__builtin__.__super__`` (with ``__builtin__.super`` to be removed in Python 3.0). It is further proposed that assignment to ``super`` become a ``SyntaxError``, similar to the behaviour of ``None``. Rationale ========= The current usage of super requires an explicit passing of both the class and instance it must operate from, requiring a breaking of the DRY (Don't Repeat Yourself) rule. This hinders any change in class name, and is often considered a wart by many. Specification ============= Within the specification section, some special terminology will be used to distinguish similar and closely related concepts. "super type" will refer to the actual builtin type named "super". A "super instance" is simply an instance of the super type, which is associated with a class and possibly with an instance of that class. Because the new ``super`` semantics are not backwards compatible with Python 2.5, the new semantics will require a ``__future__`` import:: from __future__ import new_super The current ``__builtin__.super`` will be aliased to ``__builtin__.__super__``. This will occur regardless of whether the new ``super`` semantics are active. It is not possible to simply rename ``__builtin__.super``, as that would affect modules that do not use the new ``super`` semantics. In Python 3.0 it is proposed that the name ``__builtin__.super`` will be removed. Replacing the old usage of super, calls to the next class in the MRO (method resolution order) can be made without explicitly creating a ``super`` instance (although doing so will still be supported via ``__super__``). Every function will have an implicit local named ``super``. This name behaves identically to a normal local, including use by inner functions via a cell, with the following exceptions: 1. Assigning to the name ``super`` will raise a ``SyntaxError`` at compile time; 2. Calling a static method or normal function that accesses the name ``super`` will raise a ``TypeError`` at runtime. Every function that uses the name ``super``, or has an inner function that uses the name ``super``, will include a preamble that performs the equivalent of:: super = __builtin__.__super__(<class>, <instance>) where ``<class>`` is the class that the method was defined in, and ``<instance>`` is the first parameter of the method (normally ``self`` for instance methods, and ``cls`` for class methods). For static methods and normal functions, ``<class>`` will be ``None``, resulting in a ``TypeError`` being raised during the preamble. Note: The relationship between ``super`` and ``__super__`` is similar to that between ``import`` and ``__import__``. Much of this was discussed in the thread of the python-dev list, "Fixing super anyone?" [1]_. Open Issues ----------- Determining the class object to use ''''''''''''''''''''''''''''''''''' The exact mechanism for associating the method with the defining class is not specified in this PEP, and should be chosen for maximum performance. For CPython, it is suggested that the class instance be held in a C-level variable on the function object which is bound to one of ``NULL`` (not part of a class), ``Py_None`` (static method) or a class object (instance or class method). Should ``super`` actually become a keyword? ''''''''''''''''''''''''''''''''''''''''''' With this proposal, ``super`` would become a keyword to the same extent that ``None`` is a keyword. It is possible that further restricting the ``super`` name may simplify implementation, however some are against the actual keyword- ization of super. The simplest solution is often the correct solution and the simplest solution may well not be adding additional keywords to the language when they are not needed. Still, it may solve other open issues. Closed Issues ------------- super used with __call__ attributes ''''''''''''''''''''''''''''''''''' It was considered that it might be a problem that instantiating super instances the classic way, because calling it would lookup the __call__ attribute and thus try to perform an automatic super lookup to the next class in the MRO. However, this was found to be false, because calling an object only looks up the __call__ method directly on the object's type. The following example shows this in action. :: class A(object): def __call__(self): return '__call__' def __getattribute__(self, attr): if attr == '__call__': return lambda: '__getattribute__' a = A() assert a() == '__call__' assert a.__call__() == '__getattribute__' In any case, with the renaming of ``__builtin__.super`` to ``__builtin__.__super__`` this issue goes away entirely. Reference Implementation ======================== It is impossible to implement the above specification entirely in Python. This reference implementation has the following differences to the specification: 1. New ``super`` semantics are implemented using bytecode hacking. 2. Assignment to ``super`` is not a ``SyntaxError``. Also see point #4. 3. Classes must either use the metaclass ``autosuper_meta`` or inherit from the base class ``autosuper`` to acquire the new ``super`` semantics. 4. ``super`` is not an implicit local variable. In particular, for inner functions to be able to use the super instance, there must be an assignment of the form ``super = super`` in the method. The reference implementation assumes that it is being run on Python 2.5+. :: #!/usr/bin/env python # # autosuper.py from array import array import dis import new import types import __builtin__ __builtin__.__super__ = __builtin__.super del __builtin__.super # We need these for modifying bytecode from opcode import opmap, HAVE_ARGUMENT, EXTENDED_ARG LOAD_GLOBAL = opmap['LOAD_GLOBAL'] LOAD_NAME = opmap['LOAD_NAME'] LOAD_CONST = opmap['LOAD_CONST'] LOAD_FAST = opmap['LOAD_FAST'] LOAD_ATTR = opmap['LOAD_ATTR'] STORE_FAST = opmap['STORE_FAST'] LOAD_DEREF = opmap['LOAD_DEREF'] STORE_DEREF = opmap['STORE_DEREF'] CALL_FUNCTION = opmap['CALL_FUNCTION'] STORE_GLOBAL = opmap['STORE_GLOBAL'] DUP_TOP = opmap['DUP_TOP'] POP_TOP = opmap['POP_TOP'] NOP = opmap['NOP'] JUMP_FORWARD = opmap['JUMP_FORWARD'] ABSOLUTE_TARGET = dis.hasjabs def _oparg(code, opcode_pos): return code[opcode_pos+1] + (code[opcode_pos+2] << 8) def _bind_autosuper(func, cls): co = func.func_code name = func.func_name newcode = array('B', co.co_code) codelen = len(newcode) newconsts = list(co.co_consts) newvarnames = list(co.co_varnames) # Check if the global 'super' keyword is already present try: sn_pos = list(co.co_names).index('super') except ValueError: sn_pos = None # Check if the varname 'super' keyword is already present try: sv_pos = newvarnames.index('super') except ValueError: sv_pos = None # Check if the callvar 'super' keyword is already present try: sc_pos = list(co.co_cellvars).index('super') except ValueError: sc_pos = None # If 'super' isn't used anywhere in the function, we don't have anything to do if sn_pos is None and sv_pos is None and sc_pos is None: return func c_pos = None s_pos = None n_pos = None # Check if the 'cls_name' and 'super' objects are already in the constants for pos, o in enumerate(newconsts): if o is cls: c_pos = pos if o is __super__: s_pos = pos if o == name: n_pos = pos # Add in any missing objects to constants and varnames if c_pos is None: c_pos = len(newconsts) newconsts.append(cls) if n_pos is None: n_pos = len(newconsts) newconsts.append(name) if s_pos is None: s_pos = len(newconsts) newconsts.append(__super__) if sv_pos is None: sv_pos = len(newvarnames) newvarnames.append('super') # This goes at the start of the function. It is: # # super = __super__(cls, self) # # If 'super' is a cell variable, we store to both the # local and cell variables (i.e. STORE_FAST and STORE_DEREF). # preamble = [ LOAD_CONST, s_pos & 0xFF, s_pos >> 8, LOAD_CONST, c_pos & 0xFF, c_pos >> 8, LOAD_FAST, 0, 0, CALL_FUNCTION, 2, 0, ] if sc_pos is None: # 'super' is not a cell variable - we can just use the local variable preamble += [ STORE_FAST, sv_pos & 0xFF, sv_pos >> 8, ] else: # If 'super' is a cell variable, we need to handle LOAD_DEREF. preamble += [ DUP_TOP, STORE_FAST, sv_pos & 0xFF, sv_pos >> 8, STORE_DEREF, sc_pos & 0xFF, sc_pos >> 8, ] preamble = array('B', preamble) # Bytecode for loading the local 'super' variable. load_super = array('B', [ LOAD_FAST, sv_pos & 0xFF, sv_pos >> 8, ]) preamble_len = len(preamble) need_preamble = False i = 0 while i < codelen: opcode = newcode[i] need_load = False remove_store = False if opcode == EXTENDED_ARG: raise TypeError("Cannot use 'super' in function with EXTENDED_ARG opcode") # If the opcode is an absolute target it needs to be adjusted # to take into account the preamble. elif opcode in ABSOLUTE_TARGET: oparg = _oparg(newcode, i) + preamble_len newcode[i+1] = oparg & 0xFF newcode[i+2] = oparg >> 8 # If LOAD_GLOBAL(super) or LOAD_NAME(super) then we want to change it into # LOAD_FAST(super) elif (opcode == LOAD_GLOBAL or opcode == LOAD_NAME) and _oparg(newcode, i) == sn_pos: need_preamble = need_load = True # If LOAD_FAST(super) then we just need to add the preamble elif opcode == LOAD_FAST and _oparg(newcode, i) == sv_pos: need_preamble = need_load = True # If LOAD_DEREF(super) then we change it into LOAD_FAST(super) because # it's slightly faster. elif opcode == LOAD_DEREF and _oparg(newcode, i) == sc_pos: need_preamble = need_load = True if need_load: newcode[i:i+3] = load_super i += 1 if opcode >= HAVE_ARGUMENT: i += 2 # No changes needed - get out. if not need_preamble: return func # Our preamble will have 3 things on the stack co_stacksize = max(3, co.co_stacksize) # Conceptually, our preamble is on the `def` line. co_lnotab = array('B', co.co_lnotab) if co_lnotab: co_lnotab[0] += preamble_len co_lnotab = co_lnotab.tostring() # Our code consists of the preamble and the modified code. codestr = (preamble + newcode).tostring() codeobj = new.code(co.co_argcount, len(newvarnames), co_stacksize, co.co_flags, codestr, tuple(newconsts), co.co_names, tuple(newvarnames), co.co_filename, co.co_name, co.co_firstlineno, co_lnotab, co.co_freevars, co.co_cellvars) func.func_code = codeobj func.func_class = cls return func class autosuper_meta(type): def __init__(cls, name, bases, clsdict): UnboundMethodType = types.UnboundMethodType for v in vars(cls): o = getattr(cls, v) if isinstance(o, UnboundMethodType): _bind_autosuper(o.im_func, cls) class autosuper(object): __metaclass__ = autosuper_meta if __name__ == '__main__': class A(autosuper): def f(self): return 'A' class B(A): def f(self): return 'B' + super.f() class C(A): def f(self): def inner(): return 'C' + super.f() # Needed to put 'super' into a cell super = super return inner() class D(B, C): def f(self, arg=None): var = None return 'D' + super.f() assert D().f() == 'DBCA' Disassembly of B.f and C.f reveals the different preambles used when ``super`` is simply a local variable compared to when it is used by an inner function. :: >>> dis.dis(B.f) 214 0 LOAD_CONST 4 (<type 'super'>) 3 LOAD_CONST 2 (<class '__main__.B'>) 6 LOAD_FAST 0 (self) 9 CALL_FUNCTION 2 12 STORE_FAST 1 (super) 215 15 LOAD_CONST 1 ('B') 18 LOAD_FAST 1 (super) 21 LOAD_ATTR 1 (f) 24 CALL_FUNCTION 0 27 BINARY_ADD 28 RETURN_VALUE :: >>> dis.dis(C.f) 218 0 LOAD_CONST 4 (<type 'super'>) 3 LOAD_CONST 2 (<class '__main__.C'>) 6 LOAD_FAST 0 (self) 9 CALL_FUNCTION 2 12 DUP_TOP 13 STORE_FAST 1 (super) 16 STORE_DEREF 0 (super) 219 19 LOAD_CLOSURE 0 (super) 22 LOAD_CONST 1 (<code object inner at 00C160A0, file "autosuper.py", line 219>) 25 MAKE_CLOSURE 0 28 STORE_FAST 2 (inner) 223 31 LOAD_FAST 1 (super) 34 STORE_DEREF 0 (super) 224 37 LOAD_FAST 2 (inner) 40 CALL_FUNCTION 0 43 RETURN_VALUE Note that in the final implementation, the preamble would not be part of the bytecode of the method, but would occur immediately following unpacking of parameters. Alternative Proposals ===================== No Changes ---------- Although its always attractive to just keep things how they are, people have sought a change in the usage of super calling for some time, and for good reason, all mentioned previously. - Decoupling from the class name (which might not even be bound to the right class anymore!) - Simpler looking, cleaner super calls would be better Dynamic attribute on super type ------------------------------- The proposal adds a dynamic attribute lookup to the super type, which will automatically determine the proper class and instance parameters. Each super attribute lookup identifies these parameters and performs the super lookup on the instance, as the current super implementation does with the explicit invokation of a super instance upon a class and instance. This proposal relies on sys._getframe(), which is not appropriate for anything except a prototype implementation. super(__this_class__, self) --------------------------- This is nearly an anti-proposal, as it basically relies on the acceptance of the __this_class__ PEP, which proposes a special name that would always be bound to the class within which it is used. If that is accepted, __this_class__ could simply be used instead of the class' name explicitly, solving the name binding issues [2]_. self.__super__.foo(\*args) -------------------------- The __super__ attribute is mentioned in this PEP in several places, and could be a candidate for the complete solution, actually using it explicitly instead of any super usage directly. However, double-underscore names are usually an internal detail, and attempted to be kept out of everyday code. super(self, \*args) or __super__(self, \*args) ---------------------------------------------- This solution only solves the problem of the type indication, does not handle differently named super methods, and is explicit about the name of the instance. It is less flexable without being able to enacted on other method names, in cases where that is needed. One use case this fails is where a base- class has a factory classmethod and a subclass has two factory classmethods, both of which needing to properly make super calls to the one in the base- class. super.foo(self, \*args) ----------------------- This variation actually eliminates the problems with locating the proper instance, and if any of the alternatives were pushed into the spotlight, I would want it to be this one. super or super() ---------------- This proposal leaves no room for different names, signatures, or application to other classes, or instances. A way to allow some similar use alongside the normal proposal would be favorable, encouraging good design of multiple inheritence trees and compatible methods. super(\*p, \*\*kw) ------------------ There has been the proposal that directly calling ``super(*p, **kw)`` would be equivalent to calling the method on the ``super`` object with the same name as the method currently being executed i.e. the following two methods would be equivalent: :: def f(self, *p, **kw): super.f(*p, **kw) :: def f(self, *p, **kw): super(*p, **kw) There is strong sentiment for and against this, but implementation and style concerns are obvious. Guido has suggested that this should be excluded from this PEP on the principle of KISS (Keep It Simple Stupid). History ======= 29-Apr-2007 - Changed title from "Super As A Keyword" to "New Super" - Updated much of the language and added a terminology section for clarification in confusing places. - Added reference implementation and history sections. 06-May-2007 - Updated by Tim Delaney to reflect discussions on the python-3000 and python-dev mailing lists. References ========== .. [1] Fixing super anyone? (http://mail.python.org/pipermail/python-3000/2007-April/006667.html) .. [2] PEP 3130: Access to Module/Class/Function Currently Being Defined (this) (http://mail.python.org/pipermail/python-ideas/2007-April/000542.html) Copyright ========= This document has been placed in the public domain. .. Local Variables: mode: indented-text indent-tabs-mode: nil sentence-end-double-space: t fill-column: 70 coding: utf-8 End:
At 05:23 PM 5/14/2007 +1000, Tim Delaney wrote:
Determining the class object to use '''''''''''''''''''''''''''''''''''
The exact mechanism for associating the method with the defining class is not specified in this PEP, and should be chosen for maximum performance. For CPython, it is suggested that the class instance be held in a C-level variable on the function object which is bound to one of ``NULL`` (not part of a class), ``Py_None`` (static method) or a class object (instance or class method).
Another open issue here: is the decorated class used, or the undecorated class?
Phillip J. Eby wrote:
At 05:23 PM 5/14/2007 +1000, Tim Delaney wrote:
Determining the class object to use '''''''''''''''''''''''''''''''''''
The exact mechanism for associating the method with the defining class is not specified in this PEP, and should be chosen for maximum performance. For CPython, it is suggested that the class instance be held in a C-level variable on the function object which is bound to one of ``NULL`` (not part of a class), ``Py_None`` (static method) or a class object (instance or class method).
Another open issue here: is the decorated class used, or the undecorated class?
Sorry I've taken so long to get back to you about this - had email problems. I'm not sure what you're getting at here - are you referring to the decorators for classes PEP? In that case, the decorator is applied after the class is constructed, so it would be the undecorated class. Are class decorators going to update the MRO? I see nothing about that in PEP 3129, so using the undecorated class would match the current super(cls, self) behaviour. Tim Delaney
At 04:25 PM 5/20/2007 +1000, Tim Delaney wrote:
I'm not sure what you're getting at here - are you referring to the decorators for classes PEP? In that case, the decorator is applied after the class is constructed, so it would be the undecorated class.
Are class decorators going to update the MRO? I see nothing about that in PEP 3129, so using the undecorated class would match the current super(cls, self) behaviour.
Class decorators can (and sometimes *do*, in PEAK) return an object that's not the original class object. So that would break super, which is why my inclination is to go with using the decorated result.
participants (2)
-
Phillip J. Eby
-
Tim Delaney