Optimize int add/sub for wide exact ints

CPython has a fast path for compact integers in binary add/sub, but wide exact ints still go through the generic long arithmetic path even when both operands fit in int64_t.

This issue proposes adding a separate fast path for exact PyLong operands that fit in signed 64-bit integers, while preserving the existing compact-int path.

Suggested implementation:

Keep the current compact-int specialization unchanged.
Add a separate wide-int path for exact ints that fit in int64_t.
Preserve current behavior for overflow, subclasses, and other non-exact-int cases.

Motivation:

Improve performance for wide integer add/sub without affecting the common compact-int hot path.
Avoid adding new opcodes in the compact-int path.
Fit within the current interpreter and specialization structure.

Benchmark evidence:

I prototyped this locally with a benchmark covering compact and wide add/sub cases.
Wide cases improved substantially, while compact cases remained effectively flat.
Representative interpreter-only results with JIT disabled:
- add_wide: about 25% faster
- sub_wide: about 35% faster
- add_compact/sub_compact: effectively unchanged

Benchmark script used locally:

"""Microbenchmark compact vs wide int add/sub with pyperf.

Use this with PYTHON_JIT=0 and -S if you want a stable interpreter-only run:

    PYTHON_JIT=0 ./python.exe -S Tools/scripts/bench_wide_int_pyperf.py
"""

from __future__ import annotations

import pyperf


def bench_add_compact() -> int:
    a = 1
    b = 2
    return a + b


def bench_add_wide() -> int:
    a = 10_000_000_000
    b = 1
    return a + b


def bench_sub_compact() -> int:
    a = 1
    b = 2
    return a - b


def bench_sub_wide() -> int:
    a = 10_000_000_000
    b = 1
    return a - b


def main() -> None:
    runner = pyperf.Runner()
    runner.bench_func("add_compact", bench_add_compact)
    runner.bench_func("add_wide", bench_add_wide)
    runner.bench_func("sub_compact", bench_sub_compact)
    runner.bench_func("sub_wide", bench_sub_wide)


if __name__ == "__main__":
    main()

Linked PRs

gh-151289: optimize shrinking wide int subtraction #151290

came out of #150425 but since the JIT is a no go zone for the near future ¯_(ツ)_/¯

I would prefer if you share results of the pyperformance on this. Micro-benchmarks in #150425 looks impressive, but the pyperformance seems to be not affected (or even shows worse results).

So far, I'm not sure that this optimization worth code complications.

Perhaps, it will be better if we first change PyLong internals to use tagged pointers, it was recently come again in this d.p.o thread:
https://discuss.python.org/t/pylongobject-just-use-gmp/106086/29

With this, we will have essentially only two code paths for machine-sized integers and for big integers. Just like now.

@skirpichev

ah, my bad @skirpichev, I had the message drafted but it wasn't sent:

(.bench) root@codespaces-1df16c:/workspaces/cpython# .bench/bin/python -m pyperf compare_to baseline.json candidate.json
add_compact: Mean +- std dev: [baseline] 251 us +- 20 us -> [candidate] 234 us +- 18 us: 1.07x faster
add_wide: Mean +- std dev: [baseline] 505 us +- 42 us -> [candidate] 372 us +- 27 us: 1.36x faster
sub_compact: Mean +- std dev: [baseline] 261 us +- 33 us -> [candidate] 238 us +- 19 us: 1.09x faster
sub_wide: Mean +- std dev: [baseline] 511 us +- 35 us -> [candidate] 389 us +- 48 us: 1.31x faster
chain_wide: Mean +- std dev: [baseline] 1.40 ms +- 0.11 ms -> [candidate] 1.06 ms +- 0.09 ms: 1.32x faster

Benchmark hidden because not significant (1): chain_compact

Geometric mean: 1.18x faster

Perhaps, it will be better if we first change PyLong internals to use tagged pointers, it was recently come again in this d.p.o thread:

actually I was planning to follow up in the future with that, I believe we can leverage both implementations.

you should consider the benchmarks in #150425 separate and probably irrelevant for the next few months since that touches the tier 2 JIT

@tim-one

I had the message drafted but it wasn't sent

Thanks, but as I said - micro-benchmarks are less interesting.

actually I was planning to follow up in the future with that, I believe we can leverage both implementations.

I'm not sure. If we will use separate representation for machine-sized integers (take look on FLINT's fmpz), I don't think we will have need in a third code path ("wide ints").

CC @tim-one

you should consider the benchmarks in #150425 separate and probably irrelevant for the next few months since that touches the tier 2 JIT

PYTHON_JIT=0 should be relevant, no?

pythongh-151289: specialize shrinking int subtraction

pythongh-151289: reduce subtract specialization uop churn

Navigation Menu

Repository navigation

Optimize int add/sub for wide exact ints #151289

Linked PRs

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Optimize int add/sub for wide exact ints #151289

Description

Linked PRs

Activity

KRRT7 commented on Jun 10, 2026

skirpichev commented on Jun 11, 2026

KRRT7 commented on Jun 11, 2026

KRRT7 commented on Jun 11, 2026

KRRT7 commented on Jun 11, 2026

skirpichev commented on Jun 11, 2026

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions