Print key in KeyError in map.__getitem__/__delitem__

[sarlinpe]

Description

When a C++ container is bound by bind_map and one accesses or deletes an element that does not exist, pybind11 currently raises an empty, uninformative message: KeyError: ''. This PR prints the key, which is the behavior of Python's dict.

Suggested changelog entry:

When getting or deleting an element in a container bound by ``bind_map``, print the key in ``KeyError`` if it does not exist.

[rwgk]

Looks useful, thanks. Could you please add tests? (Please tag me when it's ready for review with tests.)

[sarlinpe]

@rwgk Tests added, ready for review.

[rwgk]

Looking at this more closely, there are a bunch of things that can go wrong:

• cast(k) may fail,
• str(...) may fail,
• .cast<std::string>() may fail.

Any of those failures will be a secondary error, masking the desired KeyError. This can be extremely confusing.

Making this robust is not easy.

What's needed is something like (untested; only to give the idea):

str key_as_str;
try {
  key_as_str = str(cast(k));
}
catch (const std::exception&) {
    throw key_error();
}
set_error(PyExc_KeyError, key_as_tr);
throw error_already_set(); 

This would be implemented in a helper function.

Another thought: if str() fails, try repr().

Oh, there should also be a size limit. I.e. you don't want to produce a 1M numpy array as a str.

Sorry to make this difficult, but these things sure will bite, someone.

[sarlinpe]

@rwgk PTAL. I added a test for long keys. I don't know what kind of object would fail in str of cast - happy to add a test if you have a suggestion. What is the benefit of set_error + throw error_already_set() over throw key_error() ?

[rwgk]

I added commit ef17743 to reduce the amount of templated code.

Waiting for GitHub Actions to finish.

[sarlinpe]

LGTM, thank you!