Skip to content

Reduce redundant object creation while calling callback function from gc #104028

Description

@corona10

cpython/Modules/gcmodule.c

Lines 1391 to 1402 in 4b27972

for (Py_ssize_t i=0; i<PyList_GET_SIZE(gcstate->callbacks); i++) {
PyObject *r, *cb = PyList_GET_ITEM(gcstate->callbacks, i);
Py_INCREF(cb); /* make sure cb doesn't go away */
r = PyObject_CallFunction(cb, "sO", phase, info);
if (r == NULL) {
PyErr_WriteUnraisable(cb);
}
else {
Py_DECREF(r);
}
Py_DECREF(cb);
}

The phase object can be created earlier with one-time creation and it can be replaced with vectorcall either.
From my microbenchmark, there is 4-5% performance improvement by doing this.

microbenchmark

import pyperf
import gc


def benchamark_collection(loops):
    def callback_foo(phase, info):
        pass
    for _ in range(100):
        gc.callbacks.append(callback_foo)

    total_time = 0
    for _ in range(loops):
        t0 = pyperf.perf_counter()
        collected = gc.collect()
        total_time += pyperf.perf_counter() - t0
    return total_time


if __name__ == "__main__":
    runner = pyperf.Runner()
    runner.metadata["description"] = "GC callback benchmark"
    runner.bench_time_func("create_gc_cycles", benchamark_collection)

Linked PRs

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    performancePerformance or resource usage

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions