Combining the argument-conversion loops in _PyEval_Vector

Feature or enhancement

In _PyEval_Vector, positional and keyword argument values are converted to _PyStackRef in separate loops. Both loops call PyStackRef_FromPyObjectNew.

The values are consecutive in args, and total_args already includes both groups. I replaced the two loops locally with:

for (size_t i = 0; i < total_args; i++) {
    arguments[i] = PyStackRef_FromPyObjectNew(args[i]);
}

This removes the duplicated conversion code and the second kwnames check. Conversion order and reference ownership stay the same. _PyEvalFramePushAndInit still receives the positional count, argcount, and kwnames separately.

The test suite passes with the change.

Is there a reason to keep the positional and keyword conversions separate?
Is there a reason to keep the positional and keyword conversions separate?

P.S. Even with -O3 -march=native, the compiler (gcc-16 and clang-23) leaves these as two separate loops in the optimized IR (LLVM IR | GCC GIMPLE), including the second kwnames check and keyword-count load. This suggests a small performance optimization opportunity, rather than just a source-code cleanup.

Has this already been discussed elsewhere?

No response given

Links to previous discussion of this feature:

https://discuss.python.org/t/combining-the-argument-conversion-loops-in-pyeval-vector/109635

Linked PRs

gh-160679: Сombine argument conversion loops #160838

This suggests a small performance optimization opportunity

Can we measure the speedup?

2. > This suggests a small performance optimization opportunity
Can we measure the speedup?

Here are the benchmark numbers using pyperf on a full PGO+LTO build.

Results

Benchmark	`main`	`patch`	Difference
`vectorcall_kw`	97.3 ns ± 0.8 ns	91.4 ns ± 1.2 ns	~1.06x faster (-6.1%)

(Ran via pyperf compare_to main.json patch.json)

Benchmark script & build flags

# bench.py
from functools import partial
import pyperf

def target(a, b, c=1, d=2):
    return a

p = partial(target, 1, 2, c=3, d=4)

runner = pyperf.Runner()
runner.bench_func('vectorcall_kw', p)

Build configuration:

CC=clang CFLAGS='-O3 -march=tigerlake' \
CXX=clang++ CXXFLAGS='-O3 -march=tigerlake' \
LDFLAGS='-fuse-ld=lld' \
./configure --with-tail-call-interp --with-lto=full --enable-optimizations

Eliminating the redundant keyword check and merging the conversion loops saves ~6 ns per call on mixed positional/keyword vectorcalls without introducing any behavioral changes.

chore: merge upstream main into pythongh-160679

@ashm-dev

@ashm-dev Can you open a PR?

@eendebakpt

@eendebakpt The changes are ready on my branch: https://github.com/ashm-dev/cpython/tree/gh-160679
I'm currently hitting the 3 open PRs limit, so I can't submit it right now. Will open the PR as soon as a slot frees up

Merge remote-tracking branch 'upstream/main' into pythongh-160679

Merge branch 'python:main' into pythongh-160679

Merge branch 'main' into pythongh-160679

gh-160679: Сombine argument conversion loops (#160838)

Thanks for this microoptimisation!

Navigation Menu

Repository navigation

Combining the argument-conversion loops in _PyEval_Vector #160679

Feature or enhancement

Has this already been discussed elsewhere?

Links to previous discussion of this feature:

Linked PRs

Results

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Combining the argument-conversion loops in _PyEval_Vector #160679

Description

Feature or enhancement

Has this already been discussed elsewhere?

Links to previous discussion of this feature:

Linked PRs

Activity

sergey-miryanov commented on Oct 25, 2026

ashm-dev commented on Oct 26, 2026

Results

eendebakpt commented on Oct 26, 2026

ashm-dev commented on Oct 26, 2026

hauntsaninja commented on Oct 30, 2026

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions