Attached pull request adds a fastpath for wrapper descriptors to use the FASTCALL calling convention. It's a follow up of bpo-31410 and all my work on FASTCALL during Python 3.6 and 3.7 development cycles.

Microbenchmark:

./python -m perf timeit -s 'import array; obj=array.array("b"); wrap=array.array.__len__' 'wrap(obj)'

Result:

haypo@selma$ ./python -m perf compare_to ref.json patch.json 
Mean +- std dev: [ref] 59.2 ns +- 0.6 ns -> [patch] 28.2 ns +- 0.9 ns: 2.10x faster (-52%)

It removes 31 nanoseconds on such very fast C function, array_length().

Attached PR is still a work-in-progress. First I would like to know if it's worth it because working on polishing the actual code.

> First I would like to know if it's worth it

I don't know. What's the point of optimizing `array.array.__len__(obj)`? People usually call `len(obj)` for that...

> I don't know. What's the point of optimizing `array.array.__len__(obj)`? People usually call `len(obj)` for that...

Right, type.method(self) is less than common than self.method().

I looked at the stdlib. I found that the following method are called using wrapper descriptors:

* object.__repr__(self)
* object.__getattribute__(self, name)
* _asyncio.Future.__del__(self)
* int.__str__(obj)
* etc.

But it seems like such calls are rare compared to other kinds of function calls.

--

By the way, _PyMethodDescr_FastCallKeywords() is only called from call_function() in Python/ceval.c. It's not used in Objects/call.c. Maybe we should use it there as well? It seems like this is a question about tracing. But maybe we can copy/paste the code from call_function()?

It optimizes the same cases as bpo-31410. bpo-31410 removed a half of the overhead for wrapper descriptors, Victor's patch removes the remaining half. Actually it makes calling the descriptor faster than calling the corresponding builtin! But bpo-31410 changes were simple, just few lines, and PR 3685 looks much more complex.

Are non-fast wrappers still needed? If the patch replaces the code instead of adding it this would decrease its cost. AFAIK the only descriptors that should support non-fast calling convention are __new__, __init__ and __call__.

As for practicality, this change should slightly speed up the code that directly calls __eq__, __lt__. For example classes decorated with total_ordering. Victor, can you provide microbenchmarks with more practical code?

This seems like a straight-forward win.  I don't think we really need to see microbenchmarks before going forward with this one.

The most complex slots are __new__, __init__ and __call__ because they
accept keywords. It's hard to design a calling convention which optimize
all cases because of the backward compatibility. The risk is to convert a
dict to args + kwnames (tuple) and then back to a dict: two expensive and
useless conversions.

It's my 3rd or 4th attempt to optimize new/init/call :-) My previous
attempt was to add new tp_fastXXX slots to types, but again thr backward
compatibilty was a major pain.
https://bugs.python.org/issue29358

Here the scope is much small and backward compatibilty shouldn't affect us.

I will try to find a way to optimize these slots as well and complete my
patch.

These slots are the only slots that accept keyword arguments and arbitrary number of positional parameters. Other slots can use fast calls.

It seems like these specific descriptors are rare, so the added complexity is not worth it. I close my issue as rejected.

stage: patch review -> resolved