Abstract

Dynamic attribute access is currently possible using the “getattr” and “setattr” builtins. The present PEP suggests a new syntax to make such access easier, allowing the coder for example to write:

x.('foo_%d' % n) += 1

z = y.('foo_%d' % n).('bar_%s' % s)

instead of:

attr_name = 'foo_%d' % n
setattr(x, attr_name, getattr(x, attr_name) + 1)

z = getattr(getattr(y, 'foo_%d' % n), 'bar_%s' % s)

Rationale

Dictionary access and indexing both have a friendly invocation syntax: instead of x.__getitem__(12) the coder can write x[12]. This also allows the use of subscripted elements in an augmented assignment, as in “x[12] += 1”. The present proposal brings this ease-of-use to dynamic attribute access too.

Attribute access is currently possible in two ways:

• When the attribute name is known at code-writing time, the “.NAME” trailer can be used, as in:

  x.foo = 42
  y.bar += 100
  

• When the attribute name is computed dynamically at run-time, the “getattr” and “setattr” builtins must be used:

  x = getattr(y, 'foo_%d' % n)
  setattr(z, 'bar_%s' % s, 99)
  

  The “getattr” builtin also allows the coder to specify a default value to be returned in the event that the object does not have an attribute of the given name:

  x = getattr(y, 'foo_%d' % n, 0)
  

This PEP describes a new syntax for dynamic attribute access — “x.(expr)” — with examples given in the Abstract above.

(The new syntax could also allow the provision of a default value in the “get” case, as in:

x = y.('foo_%d' % n, None)

This 2-argument form of dynamic attribute access would not be permitted as the target of an (augmented or normal) assignment. The “Discussion” section below includes opinions specifically on the 2-argument extension.)

Finally, the new syntax can be used with the “del” statement, as in:

del x.(attr_name)

Impact On Existing Code

The proposed new syntax is not currently valid, so no existing well-formed programs have their meaning altered by this proposal.

Across all “*.py” files in the 2.5 distribution, there are around 600 uses of “getattr”, “setattr” or “delattr”. They break down as follows (figures have some room for error because they were arrived at by partially-manual inspection):

c.300 uses of plain "getattr(x, attr_name)", which could be
      replaced with the new syntax;

c.150 uses of the 3-argument form, i.e., with the default
      value; these could be replaced with the 2-argument form
      of the new syntax (the cases break down into c.125 cases
      where the attribute name is a literal string, and c.25
      where it's only known at run-time);

c.5   uses of the 2-argument form with a literal string
      attribute name, which I think could be replaced with the
      standard "x.attribute" syntax;

c.120 uses of setattr, of which 15 use getattr to find the
      new value; all could be replaced with the new syntax,
      the 15 where getattr is also involved would show a
      particular increase in clarity;

c.5   uses which would have to stay as "getattr" because they
      are calls of a variable named "getattr" whose default
      value is the builtin "getattr";

c.5   uses of the 2-argument form, inside a try/except block
      which catches AttributeError and uses a default value
      instead; these could use 2-argument form of the new
      syntax;

c.10  uses of "delattr", which could use the new syntax.

As examples, the line:

setattr(self, attr, change_root(self.root, getattr(self, attr)))

from Lib/distutils/command/install.py could be rewritten:

self.(attr) = change_root(self.root, self.(attr))

and the line:

setattr(self, method_name, getattr(self.metadata, method_name))

from Lib/distutils/dist.py could be rewritten:

self.(method_name) = self.metadata.(method_name)

Performance Impact

Initial pystone measurements are inconclusive, but suggest there may be a performance penalty of around 1% in the pystones score with the patched version. One suggestion is that this is because the longer main loop in ceval.c hurts the cache behaviour, but this has not been confirmed.

On the other hand, measurements suggest a speed-up of around 40–45% for dynamic attribute access.

Error Cases

Only strings are permitted as attribute names, so for instance the following error is produced:

>>> x.(99) = 8
   Traceback (most recent call last):
     File "<stdin>", line 1, in <module>
   TypeError: attribute name must be string, not 'int'

This is handled by the existing PyObject_GetAttr function.

Draft Implementation

A draft implementation adds a new alternative to the “trailer” clause in Grammar/Grammar; a new AST type, “DynamicAttribute” in Python.asdl, with accompanying changes to symtable.c, ast.c, and compile.c, and three new opcodes (load/store/del) with accompanying changes to opcode.h and ceval.c. The patch consists of c.180 additional lines in the core code, and c.100 additional lines of tests. It is available as sourceforge patch #1657573 [1].

Mailing Lists Discussion

Initial posting of this PEP in draft form was to python-ideas on 20070209 [2], and the response was generally positive. The PEP was then posted to python-dev on 20070212 [3], and an interesting discussion ensued. A brief summary:

Initially, there was reasonable (but not unanimous) support for the idea, although the precise choice of syntax had a more mixed reception. Several people thought the “.” would be too easily overlooked, with the result that the syntax could be confused with a method/function call. A few alternative syntaxes were suggested:

obj.(foo)
obj.[foo]
obj.{foo}
obj{foo}
obj.*foo
obj->foo
obj<-foo
obj@[foo]
obj.[[foo]]

with “obj.[foo]” emerging as the preferred one. In this initial discussion, the two-argument form was universally disliked, so it was to be taken out of the PEP.

Discussion then took a step back to whether this particular feature provided enough benefit to justify new syntax. As well as requiring coders to become familiar with the new syntax, there would also be the problem of backward compatibility — code using the new syntax would not run on older pythons.

Instead of new syntax, a new “wrapper class” was proposed, with the following specification / conceptual implementation suggested by Martin von Löwis:

class attrs:
   def __init__(self, obj):
     self.obj = obj
   def __getitem__(self, name):
     return getattr(self.obj, name)
   def __setitem__(self, name, value):
     return setattr(self.obj, name, value)
   def __delitem__(self, name):
     return delattr(self, name)
   def __contains__(self, name):
     return hasattr(self, name)

This was considered a cleaner and more elegant solution to the original problem. (Another suggestion was a mixin class providing dictionary-style access to an object’s attributes.)

The decision was made that the present PEP did not meet the burden of proof for the introduction of new syntax, a view which had been put forward by some from the beginning of the discussion. The wrapper class idea was left open as a possibility for a future PEP.

References

Copyright

This document has been placed in the public domain.