Attached patch changes index(), insert() and rotate() functions of the collections.deque type to use FASTCALL calling convention. I chose to only modify these functions since they use METH_VARARGS which requires to create a temporary tuple, whereas other functions use METH_NOARGS or METH_O which is already fast ;-)

I know that Raymond, maintainer of the collections module, is not a big fan of Argument Clinic ;-) So I wrote the minimum change and chose to not use Argument Clinic yet. By the way, the index() method has the signature "D.index(value, [start, [stop]])" which is not supported by Argument Clinic yet: see issue bpo-29299. For these reasons, I propose to wait to convert collections.deque to Argument Clinic, it can be done later.

Ok, now the cool part: it makes these methods faster ;-)

• d.rotate(): 1.10x faster
• d.rotate(1): 1.24x faster
• d.insert(): 1.18x faster
• d.index(): 1.24x faster

$ ./python -m perf timeit -s 'import collections; d=collections.deque()' 'd.rotate()' --compare-to=../default-ref/python 

Median +- std dev: [ref] 70.5 ns +- 0.9 ns -> [patch] 64.2 ns +- 0.3 ns: 1.10x faster (-9%)

$ ./python -m perf timeit -s 'import collections; d=collections.deque()' 'd.rotate(1)' --compare-to=../default-ref/python

Median +- std dev: [ref] 107 ns +- 1 ns -> [patch] 86.2 ns +- 1.1 ns: 1.24x faster (-20%)

$ ./python -m perf timeit -s 'import collections' 'd=collections.deque(); d.insert(0, None); d.insert(1, None); d.insert(2, None); d.insert(3, None); d.insert(4, None)' --compare-to=../default-ref/python -p3

Median +- std dev: [ref] 699 ns +- 6 ns -> [patch] 591 ns +- 5 ns: 1.18x faster (-15%)

$ ./python -m perf timeit -s 'import collections; d=collections.deque((None,))' 'd.index(None)' --compare-to=../default-ref/python 

Median +- std dev: [ref] 115 ns +- 1 ns -> [patch] 92.5 ns +- 0.8 ns: 1.24x faster (-19%)

I think _PyArg_NoStackKeywords() should be called before _PyArg_ParseStack(), otherwise this can cause not correct error messages.

deque-2.patch calls _PyArg_NoStackKeywords() before _PyArg_ParseStack().

The patch is simple and technically it looks correct, but I'm not a fan of using FASTCALL outside of Argument Clinic at this stage. The API still can be changed. Fortunately only three deque methods could have a benefit from using FASTCALL.

Over this looks good. Just one other minor tweak (one that has served me well elsewhere) would be to bypass the cross-module function call with a cheap (near zero cost) register variable test:

    if (kwnames != NULL && !_PyArg_NoStackKeywords("rotate", kwnames)) {
        return NULL;
    }

This already was discussed in bpo-26822. bpo-29460 makes _PyArg_NoStackKeywords() cheaper.

Over this looks good. Just one other minor tweak (one that has served me well elsewhere) would be to bypass the cross-module function call with a cheap (near zero cost) register variable test: (...)

This has just been optimized by Serhiy, change 82d1c8d15e18.

So, is deque-2.patch good now?

Yes, go ahead and apply.

New changeset 1c048539200c by Victor Stinner in branch 'default':
Optimize deque index, insert and rotate() methods
https://hg.python.org/cpython/rev/1c048539200c

Raymond: "Yes, go ahead and apply."

Great, done. Thanks for the reviews Serhiy and Raymond.

As I wrote, you can consider to use Argument Clinic later, but there is no urgency for that ;-) I close the issue.

New changeset 55949f988dc1b943796d9852cc4d588c58cc4255 by Victor Stinner in branch 'master':
Optimize deque index, insert and rotate() methods
55949f9