gh-150942: Speed up Py_BuildValue dict construction - #156199

All commit authors signed the Contributor License Agreement.

Most changes to Python require a NEWS entry. Add one using the blurb_it web app or the blurb command-line tool.

If this change has little impact on Python users, wait for a maintainer to apply the skip news label instead.

@tkim602 Can you show the benchmark script?

@tkim602 Can you show the benchmark script?

Sure, I used the following workloads:

import pyperf

runner = pyperf.Runner()

runner.timeit(
    "integer dict, 3 pairs",
    stmt=(
        '_testcapi.py_buildvalue_ints('
        '"{IIIIII}", 1001, 2001, 3001, 4001, 5001, 6001)'
    ),
    setup="import _testcapi",
)

runner.timeit(
    "integer dict, 5 pairs",
    stmt=(
        '_testcapi.py_buildvalue_ints('
        '"{IIIIIIIIII}", '
        '1001, 2001, 3001, 4001, 5001, 6001, '
        '7001, 8001, 9001, 10001)'
    ),
    setup="import _testcapi",
)

runner.timeit(
    "existing objects, 3 pairs",
    stmt=(
        '_testcapi.py_buildvalue('
        '"{OOOOOO}", k1, v1, k2, v2, k3, v3)'
    ),
    setup=(
        "import _testcapi; "
        "k1=object(); v1=object(); "
        "k2=object(); v2=object(); "
        "k3=object(); v3=object()"
    ),
)

runner.timeit(
    "existing objects, 5 pairs",
    stmt=(
        '_testcapi.py_buildvalue('
        '"{OOOOOOOOOO}", '
        "k1, v1, k2, v2, k3, v3, k4, v4, k5, v5)"
    ),
    setup=(
        "import _testcapi; "
        "k1=object(); v1=object(); "
        "k2=object(); v2=object(); "
        "k3=object(); v3=object(); "
        "k4=object(); v4=object(); "
        "k5=object(); v5=object()"
    ),
)

runner.timeit(
    "duplicate key, 3 pairs",
    stmt=(
        '_testcapi.py_buildvalue('
        '"{OOOOOO}", k, v1, k, v2, k, v3)'
    ),
    setup=(
        "import _testcapi; "
        "k=object(); v1=object(); v2=object(); v3=object()"
    ),
)

I ran it with:

python bench_mkdict.py --processes 10 --values 10 --min-time 0.5

For the reported numbers, I ran the baseline and patched builds in A/B/B/A order and also checked the reverse comparison.

@tkim602 The changes look correct to me, but I am not sure there are any places where Py_BuildValue (for dicts) is performance critical. Unless there are some cases where it is, I am inclined to close the PR.

@eendebakpt Thanks for the feedback. I'm not sure whether this rises to the level of a performance-critical use case, but I looked through the in-tree dict-producing Py_BuildValue() call sites to see whether there was a real API where this change is exercised repeatedly.

The strongest case I found is gc.get_stats(). Each call constructs the per-generation statistics dictionaries through this path. It is also used by production monitoring libraries; for example, the Prometheus Python client's default GC collector calls gc.get_stats() when collecting GC metrics.

I benchmarked the complete gc.get_stats() API call rather thanPy_BuildValue() directly, using separate baseline and patched builds in A/B/B/A order:

The patched build was faster in both orders. so I would characterize this as a small repeatable ~2–3% improvement rather than an exact universal speedup. I also checked a GC-callback workload. The GC callback showed only about a 1% improvement.

BASE=/path/to/baseline/python.exe
PATCH=/path/to/patched/python.exe
PP=/path/to/pyperf/python

SITEPKG="$("$PP" -c 'import site; print(site.getsitepackages()[0])')"
export PYTHONPATH="$SITEPKG"

run_bench() {
    python_bin="$1"
    output="$2"

    "$python_bin" -m pyperf timeit \
        --processes 10 \
        --values 10 \
        --min-time 0.5 \
        -s 'import gc; gc.disable()' \
        'gc.get_stats()' \
        -o "$output"
}

run_bench "$BASE"  /tmp/gcstats-a1.json
sleep 180

run_bench "$PATCH" /tmp/gcstats-b1.json
sleep 180

run_bench "$PATCH" /tmp/gcstats-b2.json
sleep 180

run_bench "$BASE"  /tmp/gcstats-a2.json

"$PP" -m pyperf compare_to \
    /tmp/gcstats-a1.json \
    /tmp/gcstats-b1.json \
    --table

"$PP" -m pyperf compare_to \
    /tmp/gcstats-b2.json \
    /tmp/gcstats-a2.json \
    --table