avoid a couple of dynamic lookup in the core of the parsing loop.
Performance improvement will be minimal, but adding a little many time
can be consequent.
We use the python function again with the modifications we need.
I ran it with:
python3 -m timeit -n 10000 -s 'from jedi.common import splitlines; x = open("test_regression.py").read()'
The speed differences are quite remarkable, it's ~3 times faster:
10000 loops, best of 3: 52.1 usec per loop
vs. the old:
10000 loops, best of 3: 148 usec per loop
We might need to speedup splitlines with as well. It's probably
also a factor 2-3 slower than it should be.
This tells wheel that the packages are compatible with Python 2 and 3.
The wheel currently on PyPI is only tagged 'py2', so installing on
Python 3 uses the sdist.
Outside decorators, a dotted name is parsed by the grammar as
stmt -> test -> power -> atom trailer -> (NAME) ('.' NAME)
where the first NAME is an 'atom' and the second NAME is a 'trailer'.
Thus, testing if the most recent variable is a 'trailer' and the most
recent node is a '.' is almost always enough for Jedi to properly
complete dotted names.
However, the grammar for decorators is more restricted and doesn't allow
arbitrary atoms and trailers; instead, a dotted name in a decorator is
decorator -> '@' dotted_name -> '@' (NAME '.' NAME),
meaning the most recent variable will be 'dotted_name', not 'trailer'.
Besides in decorators, the 'dotted_name' variable is only used in import
statements which are handled previously in _get_context_completions,
so checking for 'dotted_name' in this arm of the if only covers decorators
and not inadvertently anything else.
