[C API] Move usage of tp_vectorcall_offset from public headers to the internal C API

[vstinner]

BPO	45439
Nosy	@vstinner
PRs	bpo-45439: _PyObject_Call() only checks tp_vectorcall_offset once #28890 bpo-45439: Rename _PyObject_CallNoArg() to _PyObject_CallNoArgs() #28891 bpo-45439: Move _PyObject_VectorcallTstate() to pycore_call.h #28893 bpo-45439: Move _PyObject_CallNoArgs() to pycore_call.h #28895
Files	test_bench.patch bench.py bench_no_args.patch bench_no_args_inline.py bench_no_args_public.py test_bench2.patch bench2.py sys_call.patch stack_overflow-4.py

^{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 2021-10-15.00:42:16.482>
created_at = <Date 2021-10-11.22:18:05.305>
labels = ['expert-C-API', '3.11']
title = '[C API] Move usage of tp_vectorcall_offset from public headers to the internal C API'
updated_at = <Date 2021-10-15.00:42:16.477>
user = 'https://github.com/vstinner'

bugs.python.org fields:

activity = <Date 2021-10-15.00:42:16.477>
actor = 'vstinner'
assignee = 'none'
closed = True
closed_date = <Date 2021-10-15.00:42:16.482>
closer = 'vstinner'
components = ['C API']
creation = <Date 2021-10-11.22:18:05.305>
creator = 'vstinner'
dependencies = []
files = ['50345', '50346', '50347', '50348', '50349', '50350', '50351', '50355', '50356']
hgrepos = []
issue_num = 45439
keywords = ['patch']
message_count = 10.0
messages = ['403695', '403696', '403698', '403708', '403768', '403770', '403877', '403878', '403937', '403939']
nosy_count = 1.0
nosy_names = ['vstinner']
pr_nums = ['28890', '28891', '28893', '28895']
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = None
url = 'https://bugs.python.org/issue45439'
versions = ['Python 3.11']

[vstinner]

The public C API should avoid accessing directly PyTypeObject members: see bpo-40170.

I propose to move static inline functions to the internal C API, and only expose opaque function calls to the public C API.

[vstinner]

New changeset fb8f208 by Victor Stinner in branch 'main':
bpo-45439: _PyObject_Call() only checks tp_vectorcall_offset once (GH-28890)
fb8f208

[vstinner]

New changeset ce3489c by Victor Stinner in branch 'main':
bpo-45439: Rename _PyObject_CallNoArg() to _PyObject_CallNoArgs() (GH-28891)
ce3489c

[vstinner]

New changeset d943d19 by Victor Stinner in branch 'main':
bpo-45439: Move _PyObject_CallNoArgs() to pycore_call.h (GH-28895)
d943d19

[vstinner]

I should also check again the stack consumption. Old issues:

• bpo-29465: Modify _PyObject_FastCall() to reduce stack consumption
• bpo-29234: Disable inlining of _PyStack_AsTuple() to reduce the stack consumption
• bpo-29227: Reduce C stack consumption in function calls
• bpo-28858: Fastcall uses more C stack

See also: "Stack consumption" of https://vstinner.github.io/contrib-cpython-2017q1.html

[vstinner]

5 years ago, I added _PyObject_CallArg1() (similar to PyObject_CallOneArg()) and then I removed it since it consumed more stack memory than existing function, whereas I added _PyObject_CallArg1() to reduce the stack consumption.

commit 7bfb42d
Author: Victor Stinner <victor.stinner@gmail.com>
Date: Mon Dec 5 17:04:32 2016 +0100

Issue bpo-28858: Remove _PyObject_CallArg1() macro

Replace
   _PyObject_CallArg1(func, arg)
with
   PyObject_CallFunctionObjArgs(func, arg, NULL)

Using the _PyObject_CallArg1() macro increases the usage of the C stack, which
was unexpected and unwanted. PyObject_CallFunctionObjArgs() doesn't have this
issue.

[vstinner]

I measured the stack consumption using attached sys_call.patch and stack_overflow-4.py.

Using gcc -O3, the stack consumption with PR 28893 is *way better* on the 6 benchmarks (6 ways to call functions), especially:

PyObject_CallOneArg(): 624 bytes/call => 528 bytes/call (-96 bytes)
PyObject_CallNoArg(): 608 bytes/call => 512 bytes/call (-96 bytes)
_PyObject_CallNoArg(): 608 bytes/call => 512 bytes/call (-96 bytes)

Python built in release mode with gcc -O3:

./configure && make

=== ref ===

$ ./python stack_overflow-4.py
test_python_call: 10070 calls before crash, stack: 832 bytes/call
test_python_getitem: 16894 calls before crash, stack: 496 bytes/call
test_python_iterator: 12773 calls before crash, stack: 656 bytes/call
test_callonearg: 13428 calls before crash, stack: 624 bytes/call
test_callnoargs: 13779 calls before crash, stack: 608 bytes/call
test_callnoargs_inline: 13782 calls before crash, stack: 608 bytes/call

=> total: 80726 calls, 3824 bytes

=== PR ===

$ ./python stack_overflow-4.py
test_python_call: 11901 calls before crash, stack: 704 bytes/call
test_python_getitem: 18703 calls before crash, stack: 448 bytes/call
test_python_iterator: 14961 calls before crash, stack: 560 bytes/call
test_callonearg: 15868 calls before crash, stack: 528 bytes/call
test_callnoargs: 16366 calls before crash, stack: 512 bytes/call
test_callnoargs_inline: 16365 calls before crash, stack: 512 bytes/call

=> total: 94164 calls, 3264 bytes

[vstinner]

Using LTO, the PR 28893 *increases* the stack memory usage. It's the opposite :-)

PyObject_CallOneArg(): 672 bytes/call => 688 bytes/call (+16 bytes)
PyObject_CallNoArg(): 640 bytes/call => 672 bytes/call (+32 bytes)
_PyObject_CallNoArg(): 640 bytes/call => 672 bytes/call (+32 bytes)

clang with LTO:

./configure --with-lto CC=clang LD=lld LDFLAGS="-fuse-ld=lld"
make

=== ref ===

$ ./python stack_overflow-4.py
test_python_call: 9187 calls before crash, stack: 912 bytes/call
test_python_getitem: 15868 calls before crash, stack: 528 bytes/call
test_python_iterator: 11901 calls before crash, stack: 704 bytes/call
test_callonearg: 12468 calls before crash, stack: 672 bytes/call
test_callnoargs: 13091 calls before crash, stack: 640 bytes/call
test_callnoargs_inline: 13092 calls before crash, stack: 640 bytes/call

=> total: 75607 calls, 4096 bytes

=== PR ===

$ ./python stack_overflow-4.py
test_python_call: 9186 calls before crash, stack: 912 bytes/call
test_python_getitem: 15400 calls before crash, stack: 544 bytes/call
test_python_iterator: 11384 calls before crash, stack: 736 bytes/call
test_callonearg: 12177 calls before crash, stack: 688 bytes/call
test_callnoargs: 12468 calls before crash, stack: 672 bytes/call
test_callnoargs_inline: 12467 calls before crash, stack: 672 bytes/call

=> total: 73082 calls, 4224 bytes

[vstinner]

New changeset 3cc56c8 by Victor Stinner in branch 'main':
bpo-45439: Move _PyObject_VectorcallTstate() to pycore_call.h (GH-28893)
3cc56c8

[vstinner]

I decided to merge my PR to address https://bugs.python.org/issue45439 initial issue: "[C API] Move usage of **tp_vectorcall_offset** from public headers to the internal C API".

Last years, I added tstate parameters to internal C functions. The agreement was that only internal functions should use it, and indirectly that this tstate parameter should be hidden. I'm now sure exactly, but tstate started to pop up in [Include/cpython/abstract.h](https://github.com/python/cpython/blob/main/Include/cpython/abstract.h) around "call" functions. This PR fix this issue.

About the impact on performances: well, it's really hard to draw a clear conclusion. Inlining, LTO and PGO give different results on runtime performance and stack memory usage.

IMO the fact that public C API functions are now regular functions should not prevent us to continue (micro) optimizing Python. We can always add a variant to the internal C API using an API a little bit different (e.g. add tstate parameter) or defined as a static inline function, rather than a regular function.

The unclear part is if PyObject_CallOneArg() (regular function call) is faster than _PyObject_CallOneArg() (static inline function, inlined). The performance may depend if it's called in the Python executable or in a dynamic library (PLT indirection which may be avoided by gcc -fno-semantic-interposition).

Well, happy hacking and let's continue *continuous* benchmarking Python!