The keyword argument extraction/finding function seems to have a performance bug/enhancement (unless I am missing something here). It reads:

    for (i=0; i < nkwargs; i++) {
        PyObject *kwname = PyTuple_GET_ITEM(kwnames, i);

        /* ptr==ptr should match in most cases since keyword keys
           should be interned strings */
        if (kwname == key) {
            return kwstack[i];
        }
        assert(PyUnicode_Check(kwname));
        if (_PyUnicode_EQ(kwname, key)) {
            return kwstack[i];
        }
    }

However, it should be split into two separate for loops, using the PyUnicode_EQ check only if it failed for all other arguments.

I will open a PR for this (it seemed like a bpo number is wanted for almost everything.

Could you please provide benchmarks which demonstrate the performance issue?

Fair enough, we had code that does it the other way, so it seemed "too obvious" since the current check seems mainly useful with few kwargs. However, a single kwarg is super common in python, while many seem super rare (in any argument clinic function)

Which is why I had even trouble finding a function where it could make a difference, since it needs many kwargs.

With changes:

sebastian@seberg-x1 ~/python-dev/bin
% ./python3 -m timeit -s 'i = 4' 'i.to_bytes(length=5, byteorder="big", signed=True)'
1000000 loops, best of 5: 205 nsec per loop
sebastian@seberg-x1 ~/python-dev/bin
% ./python3 -m timeit -s 'i = 4' 'i.to_bytes(length=5, byteorder="big", signed=True)'
1000000 loops, best of 5: 207 nsec per loop

On master:

sebastian@seberg-x1 ~/python-dev/bin
% ./python3 -m timeit -s 'i = 4' 'i.to_bytes(length=5, byteorder="big", signed=True)'
1000000 loops, best of 5: 221 nsec per loop
sebastian@seberg-x1 ~/python-dev/bin
% ./python3 -m timeit -s 'i = 4' 'i.to_bytes(length=5, byteorder="big", signed=True)'
1000000 loops, best of 5: 218 nsec per loop

Which, at close to 5% is barely noticeable, I suppose with more kwargs it could start to make a difference. With less than 3, it just does not matter I guess. Also, I am currently not sure how likely non-interned strings could actually happen. But I am not sure it matters to optimize for them (unless PyArg_From* functions use them).

In any case, sorry if this was noise, happy to fix things up or just drop it if many kwargs are considered non-existing.

New changeset 75bb07e by Inada Naoki (Sebastian Berg) in branch 'master':
bpo-39028: Performance enhancement in keyword extraction (GH-17576)
75bb07e