bpo-38436: Improved performance for list addition.

[brandtbucher]

This PR adds a fast path for BINARY_ADD instructions involving two lists, where the left list has a refcount of exactly 1.

In this case, we instead do a PySequence_InPlaceConcat operation. This has the affect of avoiding quadratic complexity for list summations, by keeping only one intermediate result and extending it in-place.

For (potentially large) lists:

a + b + c + d ...

Currently executed as:

tmp = a + b
tmp = tmp + c
tmp = tmp + d
...

With this change:

tmp = a + b
tmp += c
tmp += d
...

Here are the pyperformance results (optimizations enabled):

EDIT: there are better measurements in the BPO discussion.

This is related to my earlier work in bpo-36229, however this is a much less invasive change that's limited to ceval.c, where our knowledge of context is much better.

https://bugs.python.org/issue38436

[pablogsal]

Where the benchmarks executed with CPU isolation? I am a bit suspicious of the

- pickle_list: 3.63 us +- 0.08 us -> 3.78 us +- 0.10 us: 1.04x slower (+4%)

If not, can you repeat them with CPU isolation and affinity? Check this for more info

[brandtbucher]

@pablogsal No isolation (I'm on a MacBook) but I can rerun with affinity on two cores.

[pablogsal]

@pablogsal No isolation (I'm on a MacBook) but I can rerun with affinity on two cores.

I can try to run the benchmarks on my machine or the speed.python.org server with isolation, as this is a very sensitive operation.

[brandtbucher]

Thanks @pablogsal - that would probably be best!