GeezerVison

Inspired by #159137.

After deduce_unreachable(), gc_collect_main() walks the surviving young generation twice.

untrack_tuples() visits every object:

and then gc_list_size(young) walks the same list again only to update long_lived_pending or long_lived_total:

So the idea is to untrack_tuples() to return the number of objects that stay tracked, and use it instead of gc_list_size(young).

Microbenchmark

Full collections on a heap that stays alive for the whole run. macOS 27 (M4 Air), release build w/o PGO. Median of 6 processes per binary, interleaved A/B/B/A:

Unfortunately, I don't have access to Linux baremetal at the moment.

Macrobenchmark

Sphinx 8.2.3 building the CPython HTML docs with one worker and fresh doctrees (sphinx-build -b html -j 1 -E), same machine and builds, 6 runs per binary interleaved A/B/B/A. GC time is the sum of duration from gc.get_stats()

import gc, sys, time
N = int(sys.argv[1]); kind = sys.argv[2]
gc.disable()
if kind == "mixed":
    keep = [({"a": i, "b": None}, [i, i], (i, i + 1)) for i in range(N // 3)]
elif kind == "tuples":
    keep = [(i, i + 1, i + 2) for i in range(N)]
elif kind == "dicts":
    keep = [{"a": i, "b": "x"} for i in range(N)]
for _ in range(2):
    gc.collect()
ts = []
for _ in range(7):
    t0 = time.perf_counter_ns(); gc.collect(); ts.append(time.perf_counter_ns() - t0)
ts.sort(); print(ts[3] / 1e6)

• Issue: Speed up garbage collector #159137