gh-140009: Optimize dict.items() symmetric difference via PyTuple_FromArray

[andrewloux]

Summary

This PR replaces PyTuple_Pack with PyTuple_FromArray in Objects/dictobject.c within the dictitems_xor_lock_held function.

By avoiding the variadic argument (va_args) processing overhead of PyTuple_Pack, we reduce the per-item cost of symmetric difference operations (dict.items() ^ dict.items()) that involve value mismatches. The change uses a stack-allocated array to pass arguments directly to the tuple constructor.

Benchmarks (PGO+LTO)

Validated using pyperf in --rigorous mode on a full production build.

• Platform: macOS arm64 (Apple M-series)
• Build: --enable-optimizations --with-lto
• Tool: pyperf (--rigorous mode)
• Baseline: upstream/main
• Candidate: pytuple-dictitems-xor-fromarray (714fb11)

Benchmark	Baseline (Mean ± Std Dev)	Candidate (Mean ± Std Dev)	Speedup
dict_items_xor_overlap_neq	75.3 ms ± 4.6 ms	74.3 ms ± 2.5 ms	1.01x faster
dict_items_xor_disjoint	142 ms ± 5 ms	141 ms ± 4 ms	Neutral
dict_items_xor_overlap_equal_control	52.1 ms ± 2.1 ms	51.9 ms ± 1.8 ms	Neutral

Geometric mean: 1.00x faster (1.01x on target path)

Repro commands

# Target workload: High overlap, mismatched values (stresses tuple creation)
python -m pyperf command --rigorous --name dict_items_xor_overlap_neq
  ./python.exe -c "d1={i:i for i in range(4000)}; d2={i:i+1 for i in range(4000)}; print(sum(len(d1.items() ^ d2.items()) for _ in range(120)))"

# Control workload: Identical dicts (no tuple creation)
python -m pyperf command --rigorous --name dict_items_xor_overlap_equal_control
  ./python.exe -c "d1={i:i for i in range(4000)}; d2={i:i for i in range(4000)}; print(sum(len(d1.items() ^ d2.items()) for _ in range(120)))"

# Disjoint workload: No overlapping keys
python -m pyperf command --rigorous --name dict_items_xor_disjoint
  ./python.exe -c "d1={i:i for i in range(4000)}; d2={i+4000:i for i in range(4000)}; print(sum(len(d1.items() ^ d2.items()) for _ in range(120)))"

Analysis

The dict_items_xor_overlap_neq workload specifically exercises the modified path by comparing dictionaries with overlapping keys but unequal values, triggering a tuple creation for every mismatched entry.

While the aggregate effect is a micro-optimization, the results show a consistent improvement on the target path with a reduction in variance (±4.6 ms → ±2.5 ms) across multiple runs. Control workloads (equal and disjoint) remain neutral, confirming no regressions in non-target dictionary shapes.

• Issue: Improve performance by replacing PyTuple_Pack with PyTuple_FromArray #140009

[python-cla-bot]

All commit authors signed the Contributor License Agreement.

[bedevere-app]

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.

[bedevere-app]

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.

[eendebakpt]

@andrewloux Your own benchmarks show this change is performance neutral.

Unless there are additional result that show why this change is a significant improvement, I suggest we close this.

(I do believe this is a tiny net improvement, but in general we avoid making such small changes to reduce churn and potential unforeseen issues)

[andrewloux]

(I do believe this is a tiny net improvement, but in general we avoid making such small changes to reduce churn and potential unforeseen issues)

Yup, totally makes sense - let's close this 👍🏽 Thanks @eendebakpt