Source

extradoc / planning / primitive-identity.rst

Full commit

Primitive objects and object identity

Hi all,

Recently PyPy merged a pretty big branch that allows for transparently type-specializing dictionaries. That means if you write something like:

strs = {x: str(x) for x in xrange(100000)}

The dictionary would be specialized for integer keys, obviating the need to allocate W_IntObjects (PyPy's equivalent to PyIntObject).

This, however introduces interesting behavior surrounding object identity (only with respect to primitive objects, none of what is discussed affects either mutable, or user-defined objects), specifically the follow code no longer works:

x = int(x)
z = {x: None}
assert next(iter(z)) is x

This would similiarly fail if you replaced is with comparing the id()s. The question now is, is this behavior a violation of the Python language definition, or is it a legal, interpreter defined difference?

There are several arguments in favor:

  1. It is easier to implement this way, and removes complexity in the interpreter implementation, and allows for better performance.
  2. For all of these objects, identity is trivial. That is to say identity could always be replacement by an equality test and no semantics would be violated. In that respect requiring that identity be maintained adds no value, the new object is completely indistinguishable
  3. A reliance on object identity leads to some rather strange behavior, a good example of this is a recent discussion about the identity shortcut in dict.__contains__ and list.__contains__, specifically in the case of nan. At present if you have a dict with a nan key the only way to retreive that value is to use the exact same nan object, another one will not do because nan does not have reflexive identity. Even on CPython, passing around this object could easily lose it's identity, for example various functions in the math module return a nan given a nan argument, but they make no guarntee that they return the same instance, furthering a reliance on any such behavior is wrong, given that equality is always a valid substitute.

And arguments against:

  1. It may break some existing code (so far the only such code I've found is in the Python test suite, it was not testing this behavior directly, but rather incidentally relied on it).

I can find no other argument against it.

Note that should we decide that this is in fact a violation of the language spec, the resulting behavior in PyPy will be for identity to be equality on primitive type objects. That is to say, the following code would work:

assert all([
    x is (x + 1 - 1)
    for x in xrange(sys.minint, sys.maxint)
])

As actually assigning allocating W_IntObjects will not occur.

Opinions welcome, Alex