[dexhunter]

TL;DR

• Reduce repeated attribute lookups and duplicate loop-history dictionary probes.
• Optimize CheckpointManager.restore_exact_step and its detect_loop and record helpers for large checkpoint seeks.
• The frozen workload based on test_goto_large_step_via_cache_and_dynamic_extension takes 8.8% less wall time: median paired ratio 0.912 across three fresh harness runs.
• Median batch times across those runs: 195.306 ms before → 177.172 ms after; each batch contains three workload repetitions.
• Validation: the requested four-file pytest group passes all 146 tests; the full suite passes all 287 tests; Ruff lint and formatting pass.

Problem and change

Large seeks extend the checkpoint cache one CPU instruction at a time. Each iteration repeatedly resolves CPU attributes and bound methods, while loop detection probes the same history dictionary twice for existing signatures.

Bind the CPU, checkpoint containers and frequently used methods locally, and use history.get to retrieve a previous step with one lookup. History values are integer step counts, so None distinguishes a missing entry from step zero. Execution order, loop checks, history updates, snapshot ordering, trace handling and public signatures remain unchanged. No dependencies or persistent caches are added.

Measurements

The workload retains the repository test's precompute-to-5,000, seek-to-4,500 and extend-to-20,000 sequence. It also includes seeded variations and boundary cases for backward seeks, tracing, snapshot caps, memory and stack instructions, halts, cycles, empty caches and invalid requests.

Measurements used Python 3.12.13 and the frozen evaluator on an isolated benchmark slot. Each run used nine interleaved baseline/candidate pairs, fresh seeded inputs, three workload repetitions per sample, a parent-process wall clock, and a separate pristine-baseline oracle. The timed batches include the workload's state observations; these are not isolated method-call timings or pytest runner durations.

Fresh run	Baseline batch (ms)	Candidate batch (ms)	Median paired ratio	Relative IQR
1	194.701	176.110	0.904301	1.07%
2	195.851	178.304	0.912983	1.43%
3	195.306	177.172	0.911658	0.54%

The headline reduction uses the median of the three paired-ratio scores, 0.911658. The absolute before/after values are separate medians of the three batch-time results, so their quotient need not equal the paired-ratio score. All three scores are below the evaluator's 0.95 improvement threshold, and each run passes its 5% relative-IQR noise gate and exact semantic checks.

Validation

Run from the changed repository copy:

PYTHONPATH=src python -m pytest tests/test_tui.py tests/test_control_flow.py tests/test_arithmetic.py tests/test_compiler.py
# 146 passed
PYTHONPATH=src python -m pytest
# 287 passed
ruff check src/ tests/
# All checks passed
ruff format --check src/ tests/
# 20 files already formatted

The pristine baseline also passes the full 287-test suite and both style checks. Existing tests and the frozen oracle cover this behavior-preserving optimization; no new functionality or tests are added.

Checklist

• Code follows style guidelines.
• All tests pass.
• New tests added for new functionality (no new functionality).
• Documentation updated (no user-facing change).
• Commit message is clear and descriptive.
• PR description explains changes and motivation.
• Branch is up-to-date with main (publication check pending).
• No unrelated changes included.

AI assistance

Codex gpt-6.1-sol generated the optimization candidates, tracked with Weco Observe. Automated source review, semantic checks, native tests and benchmarking informed selection. No human review is claimed.

Weco share link: https://dashboard.weco.ai/share/OVyh1gbcXZ7zmOO5XWrtis26cn8RDmxo

[sql-hkr]

LGTM! Clean and safe micro-optimization. Thanks for the performance improvement!

[codecov]

Codecov Report

❌ Patch coverage is 85.48387% with 9 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/tiny8/tui.py	85.48%	9 Missing ⚠️

📢 Thoughts on this report? Let us know!