Pressing <tab> to invoke autocompletition on instances repeatedly accesses attributes/descriptors values:

>>> # <tab> is used to indicate when/where I press <tab>
>>> class Foo:
...   @property
...   def bar(self): print('Foo.bar called')
... 
>>> f = Foo()
>>> f.<tab>Foo.bar called 
Foo.bar called
Foo.bar called
Foo.bar called
<tab>Foo.bar called
Foo.bar called
Foo.bar called
Foo.bar called
<tab>
f.__class__(         f.__doc__            f.__getattribute__(  f.__le__(            f.__new__(           f.__setattr__(       f.__weakref__
f.__delattr__(       f.__eq__(            f.__gt__(            f.__lt__(            f.__reduce__(        f.__sizeof__(        f.bar
f.__dict__           f.__format__(        f.__hash__(          f.__module__         f.__reduce_ex__(     f.__str__(           
f.__dir__(           f.__ge__(            f.__init__(          f.__ne__(            f.__repr__(          f.__subclasshook__(  
>>> f.b<tab>Foo.bar called
Foo.bar called
Foo.bar called
Foo.bar called
ar<enter>
Foo.bar called

Each time I press <tab>, the property is called 4 times. I'm not sure why the value is accessed at all, but even if there was a valid reason to do so, doing it once should be enough.

Long story short: it is accessed due to the callable suffix check (bpo-449227), and this check is less than optimal, meaning the attribute gets accessed up to four times per Tab press.

In 3.6, there are only two calls, one for the hasattr() call, and one for the getattr() call:

Foo.bar called
File "/home/proj/python/cpython/Lib/rlcompleter.py", line 85, in complete
self.matches = self.attr_matches(text)
File "/home/proj/python/cpython/Lib/rlcompleter.py", line 164, in attr_matches
hasattr(thisobject, word)):
File "<stdin>", line 3, in bar
Foo.bar called
File "/home/proj/python/cpython/Lib/rlcompleter.py", line 85, in complete
self.matches = self.attr_matches(text)
File "/home/proj/python/cpython/Lib/rlcompleter.py", line 165, in attr_matches
val = getattr(thisobject, word)
File "<stdin>", line 3, in bar

Before revision 4dbb315fe667 (bpo-25011), “words” was a list rather than a set. It gets two “bar” entries, one from dir(f) and one from dir(f.__class__). Maybe it is worth backporting set(), I dunno.

The hasattr() call was added in r65168 (bpo-3396) as a look-before-you-leap check to avoid the completer from dying from getattr() exceptions. IMO in this case it is better to “ask for forgiveness” by catching Exception from getattr(). This would be more robust to quirky and buggy attributes anyway.

Finally (or really, initially), getattr() was added in r64664 (bpo-449227). I think even if getattr() fails, the attribute name could still be output, just not with the callable suffix (which I partly disagree with anyway). It would have been like that previously. But I guess this is debatable, and a separate issue.

Illustration of why I think removing the attribute on errors is bad:

>>> class Foo:
...   @property
...   def bar(self): return self._bar
...   @bar.setter
...   def bar(self, value): self._bar = value
... 
>>> f = Foo()
>>> f.bar = ...  # Tab completion fails in this case

Agreed with all your suggestions Martin. Do you want to write a patch?

getattr-once.patch handles my first two points. It backports the set() change, and avoids the hasattr() check. I guess we apply this to 3.4+; expect merge conflicts with 3.6.

As for the last point, I will make another patch to include special attributes that don’t actually exist yet, where they are currently omitted. Would we consider this a new feature for 3.6? Going by Serhiy’s logic in <https://bugs.python.org/issue25209#msg251514\> I would say yes.

uncreated-attr.patch is against the 3.6 branch, and includes the excessive getattr() test, and new code to include uncreated attribute names. I added a paragraph to What’s New describing the new completion behaviour.

Both patches LGTM. Thank you Martin.

Would we consider this a new feature for 3.6?

We can consider this a fix of the regression in bpo-449227. But this behavior here is so long. I agree with applying it to 3.6 only.

For nicer Mercurial history, it would be better to apply both patches separately in 3.6.

Would not be simpler to use uninitialized slot attribute in test_uncreated_attr?

Yeah I plan to merge the first patch (fixing conflicts) into 3.6 without any of the uncreated-property additions, and then apply that extra stuff in a separate commit.

Using __slots__ would be simpler; I’ll switch it over.

New changeset 92f989bfeca2 by Martin Panter in branch '3.4':
Issue bpo-25590: Make rlcompleter only call getattr() once per attribute
https://hg.python.org/cpython/rev/92f989bfeca2

New changeset 808279e14700 by Martin Panter in branch '3.5':
Issue bpo-25590: Merge rlcompleter change from 3.4 into 3.5
https://hg.python.org/cpython/rev/808279e14700

New changeset ff68304356fe by Martin Panter in branch 'default':
Issue bpo-25590: Merge rlcompleter getattr change from 3.5
https://hg.python.org/cpython/rev/ff68304356fe

New changeset bbf63749a129 by Martin Panter in branch 'default':
Issue bpo-25590: Complete attribute names even if they are not yet created
https://hg.python.org/cpython/rev/bbf63749a129

New changeset 16d83e11675f by Martin Panter in branch '2.7':
Issue bpo-25590: Make rlcompleter only call getattr() once per attribute
https://hg.python.org/cpython/rev/16d83e11675f

I also ported the getattr patch to the 2.7 branch.

Thanks for taking care of this!