define separate semantics for $, and [[

[kevinushey]

Related: #251

This PR seeks to define different semantics for $ and [[. With this PR:

• $ is used for attribute access, as Python's . is used;
• [[ is used for item access, as Python's [[ is used.

For backwards compatibility, with Python (builtin) dictionaries, $ still returns items rather than attributes. We may want to consider changing this in the future.

[kevinushey]

For what it's worth, some of the surprise here might've arisen from how Pandas defines the . operator. For Pandas DataFrames, . will first look at attributes; if no such attribute exists, then it will attempt to look for a column. This implies that code of this form works:

>>> import pandas as pd
>>> pdf = pd.DataFrame({'apple': [1]})
>>> pdf.apple
0    1
Name: apple, dtype: int64

However, if an attribute (method) of the same name exists, then that is chosen instead:

>>> import pandas as pd
>>> pdf = pd.DataFrame({'pop': [1]})
>>> pdf.pop
<bound method DataFrame.pop of    pop
0    1>

[kevinushey]

To help preserve backwards compatibility, I've decided on the following:

• $ accesses attributes first, items second;
• [ accesses items first, attributes second.

This should ensure the least amount of breakage in existing code, but it does imply a change in this case for example:

> d <- dict(items = 42)
> d$items
<built-in method items of dict>
> d['items']
42.0

Previously, d$items would have accessed the items item; now it accesses the attribute (bound method) called items.

[kevinushey]

These checks should only occur if the user is passing in a scalar string (length one character vector).

[jjallaire]

@kevinushey Is this good to merge?

[kevinushey]

I want to double-check this doesn't regress anything in keras or tensorflow so I'll report back after testing a bit there.

[jjallaire]

Okay, you can go ahead and merge whenever you think it's ready.

[kevinushey]

Tests pass with both keras and tensorflow with this branch, so I'm going to merge.

[flying-sheep]

I don’t see any warnings. I think it’s quite error-prone to have all operators do everything without any way being the right one.