bpo-45256: Rationalize code around Python-to-Python calls a bit.

[markshannon]

• Removes the steal_args parameter. All code involved in frame initialization always consume the references to the arguments.
• Separate pushing the frame, initializing the "specials" and initializing the locals.

In theory CALL_PY_SIMPLE should be a bit quicker, but that doesn't show up in the benchmark results

https://bugs.python.org/issue45256

[bedevere-bot]

🤖 New build scheduled with the buildbot fleet by @pablogsal for commit 9cc3f1f 🤖

If you want to schedule another build, you need to add the ":hammer: test-with-buildbots" label again.

[pablogsal]

Why this tag? There is only one place where you can reach the tag, no? Is also not future proof, because you are only raising PyErr_NoMemory so I am not sure what is the advantage. Is for separating error handling from the logic?

[markshannon]

It is just to keep the error handling out of the main block of code.
It is a quite common idiom in CPython to put the error handling at the end. Personally, I think early exits help readability, so I think it is worth doing even if there is only a single jump to the error handler.

I see you point about PyErr_NoMemory though, I'll rename the label to fail_no_memory for clarity.

[pablogsal]

Question: If the call succeeds, don't we need to do this work anyway in initialize_locals? Not sure if is going to be more messy, but we could maybe share this path or part of this path for error and not error cases. For instance, remove these parts in initialize_locals:

    if (kwnames) {
        Py_ssize_t kwcount = PyTuple_GET_SIZE(kwnames);
        for (j = argcount; j < argcount+kwcount; j++) {
            Py_DECREF(args[j]);
        }
    }

[markshannon]

I don't think so. If the call succeeds all the argument references are consumed by moving them into the local variable array.

[pablogsal]

LGTM