iwls_runner: pipeline driving mig_egg on IWLS 2026 contest benchmarks

[kvisaweb]

Summary

Adds a self-contained pipeline that drives mig_egg on the IWLS 2026 contest benchmarks (100 truth tables ex200..ex299). Two coupled pieces in one atomic commit:

1. mig_egg/src/lib.rs (133-line diff)

• rules() switched to AIG-only: the 5 AND rewrites already defined in the file (comm_and, comp_and, dup_and, and_true, and_false) are uncommented; majority-language rules are off because they cannot fire on AIG-form input (& a b). associ_and stays off — its O(N²) match cost blows past node_limit on dense ASTs.
• simplify_depth / simplify_best runners: iter_limit(1000) → iter_limit(10) at three sites. egg's time_limit is best-effort and does not bound a single iteration on large e-graphs; tightening iter_limit is what actually keeps wall-clock bounded.
• New iwls_run test entry: reads $IWLS_PREFIX_FILE (one PO per line, <idx>\t<expr>), calls simplify_best per PO, prints iwls case <idx> best_expr[0] = <expr> for the Python wrapper to grep. $IWLS_FIRST_DEPTH={true,false} switches delay-first vs area-first cost.

2. iwls_runner/ (new top-level folder)

• 100 IWLS truth tables under benchmarks/
• scripts/aig_to_prefix.py + prefix_to_aig.py for the two conversions
• run_one_bench.py (end-to-end one bench), run_all.py (parallel runner), finalize.py (per-bench Pareto + cec verify + contest-shaped zip)
• examples/smoke_one.sh (ex212, ~30s) and examples/full_run.sh (100 benches, ~30 min on 16 cores)

Full pipeline: truth → ABC strash → AIG → Lisp prefix → mig_egg simplify_best per PO → EQN → ABC strash + resyn2 → AIG → ABC cec -n.

Test plan

• Smoke validated: ex212 (16in × 16out): 1409 AND / lev 24 → 571 AND / lev 11 (−59% area, −54% delay), cec equivalent, 1.94s wall
• All paths in scripts use env vars (ELOGIC_REPO_DIR, ABC) or package-relative; no hardcoded absolute paths
• .gitignore covers results/ so per-run outputs don't leak into commits
• Full 100-bench run on reviewer's hardware (~30 min on 16 cores; long-tail benches like ex299, ex230, ex252 are expected to hit egg's node_limit(5000) ceiling — that's a real saturation cap of the AIG ruleset on densely shared graphs, not a timeout)

Quick start

git checkout iwls-runner
cd mig_egg && cargo build --release --tests --no-default-features
cd ../iwls_runner
export ELOGIC_REPO_DIR="$(cd .. && pwd)"
export ABC=/path/to/abc
bash examples/smoke_one.sh        # 30s sanity check
bash examples/full_run.sh         # full 100-bench + finalize

Full doc in iwls_runner/README.md on the branch.

🤖 Generated with Claude Code