Add METH_FASTCALL: new calling convention for C functions

[vstinner]

BPO	27810
Nosy	@scoder, @vstinner, @serhiy-storchaka, @jdemeyer
Dependencies	bpo-27830: Add _PyObject_FastCallKeywords(): avoid the creation of a temporary dictionary for keyword arguments
Files	fastcall.patch

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = <Date 2016-09-12.12:59:03.974>
created_at = <Date 2016-08-20.00:32:46.209>
labels = ['type-feature']
title = 'Add METH_FASTCALL: new calling convention for C functions'
updated_at = <Date 2019-05-14.20:40:00.365>
user = 'https://github.com/vstinner'

bugs.python.org fields:

activity = <Date 2019-05-14.20:40:00.365>
actor = 'jdemeyer'
assignee = 'none'
closed = True
closed_date = <Date 2016-09-12.12:59:03.974>
closer = 'vstinner'
components = []
creation = <Date 2016-08-20.00:32:46.209>
creator = 'vstinner'
dependencies = ['27830']
files = ['44517']
hgrepos = []
issue_num = 27810
keywords = ['patch']
message_count = 18.0
messages = ['273173', '273336', '273339', '273340', '273341', '273409', '273679', '275447', '275516', '275555', '275560', '275826', '276033', '276046', '276047', '276048', '283332', '342518']
nosy_count = 5.0
nosy_names = ['scoder', 'vstinner', 'python-dev', 'serhiy.storchaka', 'jdemeyer']
pr_nums = []
priority = 'normal'
resolution = 'fixed'
stage = None
status = 'closed'
superseder = None
type = 'enhancement'
url = 'https://bugs.python.org/issue27810'
versions = ['Python 3.6']

[vstinner]

The issue bpo-27128 added _PyObject_FastCall() to avoid the creation of temporary tuples when calling functions.

I propose to add a new METH_FASTCALL calling convention. The example using METH_VARARGS | METH_KEYWORDS:
   PyObject* func(DirEntry *self, PyObject *args, PyObject *kwargs)
becomes:
   PyObject* func(DirEntry *self, PyObject **args, int nargs, PyObject *kwargs)

Using METH_VARARGS, args is a Python tuple. Using METH_FASTCALL, args is a C array of PyObject*, and there is a second nargs parameter.

Later, Argument Clinic will be modified to *generate* code using the new METH_FASTCALL calling convention. Code written with Argument Clinic will only need to be updated by Argument Clinic to get the new faster calling convention (avoid the creation of a temporary tuple for positional arguments).

This issue depends on the issue bpo-27809 "_PyObject_FastCall(): add support for keyword arguments". I will wait until this dependency is implemented, before working on the implementation of this part.

For a full implementation, see my first attempt in the issue bpo-26814. I will extract the code from this branch to write a new patch.

[scoder]

I agree that this would be cool. There is a tiny bit of a backwards compatibility concern as the new function signature would be incompatible with anything we had before, but I would guess that any code that chooses to bypass PyObject_Call() & friends would at least fall back to using it if it finds flags that it cannot handle. Cython code definitely does and always did, but there might be bugs. This change is the perfect way to "find" those. ;-)

Anyway, Victor, I definitely appreciate your efforts in this direction.

[vstinner]

There is a tiny bit of a backwards compatibility concern as the new function signature would be incompatible with anything we had before,

Right. If you call directly PyCFunction functions, you will likely get quickly a crash. But... who call directly PyCFunction functions? Why not using the 30+ functions to call functions?

Hopefully, it's easy to support METH_FASTCALL in an existing function getting a tuple:

   int nargs = (int)PyTuple_GET_SIZE(args);
   PyObject **stack = &PyTuple_GETITEM(args, 0);
   result = func(self, stack, nargs, kwargs);

I guess that Cython calls directly PyCFunction. cpyext module of PyPy probably too. Ok, except of them, are you aware of other projects doing that?

[scoder]

Extensive callback interfaces like map() come to mind, where a large number of calls becomes excessively time critical and might thus have made people implement their own special purpose calling code.

However, I don't know any such code (outside of Cython) and I agree that this is such a special optimisation that there can't be many cases. If there are others, they'll likely be using the obvious fallback already, or otherwise just adapt when Py3.6 comes out and move on. The fix isn't complex at all.

I do not consider this a road block.

[vstinner]

Extensive callback interfaces like map() come to mind, where a large number of calls becomes excessively time critical and might thus have made people implement their own special purpose calling code.

I didn't patch map_next() of Python/bltinmodule.c nor list.sort() for use fast call yet. With my full patch, map() and list.sort() are *faster*, but even more changes are needed ;-)

https://bugs.python.org/issue26814#msg263999

• map(lambda x: x, list(range(1000))): 18% faster
• sorted(list, key=lambda x: x): 33% faster

I'm now aware of map() or list.sort()/sorted() function rewritten from scratch for better performance, but I can imagine that someone did that. But do these specialized implementation use PyObject_Call() or "inline" PyCFunction_Call()?

If someone "inlines" PyCFunction_Call(), be prepared to "backward incompatible changes", since it's common that CPython internals change in a subtle way.

However, I don't know any such code (outside of Cython)

I used GitHub to search for such code using "case METH_VARARGS". I found:

• (many copies of the CPython source code)
• https://github.com/jhgameboy/ironclad/blob/5892c43b540b216d638e0fed2e6cf3fd8289fdfc/src/CallableBuilder.cs : ironclad is a module to call C extensions from IronPython, the last commit of the project was in 2011...
• https://github.com/pepsipepsi/nodebox_opengl_python3/blob/cfb2633df1055a028672b11311603cc2241a1378/nodebox/ext/psyco/src/c/Objects/pmethodobject.c : copy of the psyco module. The last commit in pysco was also in 2011, https://bitbucket.org/arigo/psyco/commits/all

In short, I found nothing. But I didn't search far, maybe there are a few other code bases using "case METH_VARARGS:"?

[vstinner]

According to discussions in the issue bpo-27809, require a Python dict leads to inefficient code because a temporary dict may be required. The issue bpo-27830 proposes to pass keyword arguments as (key, value) pairs.

It should be used for the new METH_FASTCALL calling convention.

[vstinner]

I changed the dependency to the issue bpo-27830 "Add _PyObject_FastCallKeywords(): avoid the creation of a temporary dictionary for keyword arguments" which itself depends on the issue bpo-27213 "Rework CALL_FUNCTION* opcodes".

[python-dev]

New changeset a25c39873d93 by Victor Stinner in branch 'default':
Issue bpo-27810: Add _PyCFunction_FastCallKeywords()
https://hg.python.org/cpython/rev/a25c39873d93

[vstinner]

fastcall.patch combines two changes:

changeset: 103513:74abb8ddf7f2
tag: tip
user: Victor Stinner <victor.stinner@gmail.com>
date: Fri Sep 09 17:40:38 2016 -0700
files: Include/modsupport.h Python/getargs.c Tools/clinic/clinic.py
description:
Emit METH_FASTCALL code in Argument Clinic

Issue bpo-27810:

• Modify vgetargskeywordsfast() to work on a C array of PyObject* rather than
  working on a tuple directly.
• Add _PyArg_ParseStack()
• Argument Clinic now emits code using the new METH_FASTCALL calling convention

changeset: 103512:d55abcddd194
user: Victor Stinner <victor.stinner@gmail.com>
date: Fri Sep 09 17:40:22 2016 -0700
files: Include/abstract.h Include/methodobject.h Objects/abstract.c Objects/methodobject.c
description:
Add METH_FASTCALL calling convention

Issue bpo-27810: Add a new calling convention for C functions:

    PyObject* func(PyObject *self, PyObject **args,
                   Py_ssize_t nargs, PyObject *kwnames);

Where args is a C array of positional arguments followed by values of keyword
arguments. nargs is the number of positional arguments, kwnames are keys of
keyword arguments. kwnames can be NULL.

[python-dev]

New changeset 86b0f5a900c7 by Victor Stinner in branch 'default':
Add METH_FASTCALL calling convention
https://hg.python.org/cpython/rev/86b0f5a900c7

New changeset 633f850038c3 by Victor Stinner in branch 'default':
Emit METH_FASTCALL code in Argument Clinic
https://hg.python.org/cpython/rev/633f850038c3

New changeset 97a68adbe826 by Victor Stinner in branch 'default':
Issue bpo-27810: Rerun Argument Clinic on all modules
https://hg.python.org/cpython/rev/97a68adbe826

[python-dev]

New changeset 3934e070c9db by Victor Stinner in branch 'default':
Issue bpo-27810: Fix getargs.c compilation on Windows
https://hg.python.org/cpython/rev/3934e070c9db

[python-dev]

New changeset 603bef7bb744 by Serhiy Storchaka in branch 'default':
Issue bpo-27810: Regenerate Argument Clinic.
https://hg.python.org/cpython/rev/603bef7bb744

[vstinner]

Stefan Behnel: "There is a tiny bit of a backwards compatibility concern as the new function signature would be incompatible with anything we had before,"

Python 3.6 will probably have two "fast call" calling convention:

• _PyObject_FastCallDict(): expect a Python dict for keyword arguments
• _PyObject_FastCallKeywods(): expect a Python tuple for keys of keyword arguments, keyword values are packed in the same array than positional arguments

_PyObject_FastCallKeywods() is not really written to be called directly: Python/ceval.c calls you, but you may call _PyObject_FastCallKeywods() again "wrapper" functions, like functools.partial().

Currently, tp_call (and tp_init and tp_new) still expects a (tuple, dict) for positional and keyword arguments, but later I will add something to also support METH_FASTCALL for callable objects. I just don't know yet what is the best option to make this change.

--

The main idea is implemented (implement METH_FASTCALL), I close the issue.

I will open new issues for more specific changes, and maybe extend the API (especially tp_call).