The verified-dse programme asks how much of an open-source chip design search (Yosys, ABC and OpenROAD on SKY130) can be skipped by trusting cheap estimates. Its rule, the clipped ladder, maps every ABC pass sequence, ranks the mapped netlists with a cheap score, and routes only the best 8. The baseline routes 8 random sequences for the same cost.
The last run found that at the fast-fast library corner the ladder's edge over random routing was 7.3 points smaller on designs whose shift amount is a chain of additions. It also found a catch: every one of those adder-chain designs mapped to a short critical path (under 15 ns), and every control design mapped to a long one. "Adder chains are hard for the ladder" and "short paths are hard for the ladder" predicted the same table.
What this run did
It ran the two missing corners of the comparison: adder chains that map slow, and control chains (compare, select, divide) that map fast. All were fresh designs nobody had run at this corner. The candidates, a selection rule by baseline delay and eight predictions were committed before any flow ran.
The selection had a finding of its own. Of 14 adder-chain candidates, only the ones'-complement sums stayed slow (18 to 42 ns). Sums of absolute differences, wide multi-term sums and dot products all mapped at 8 to 15 ns however long we made them: synthesis turns them into trees. So the slow adder cell has 5 designs, all from one family, against 8 fast controls.
What it found
| cell | designs | mean edge of the ladder (points of regret) | ladder losses |
|---|---|---|---|
| fast adder chains (last run) | 10 | 1.57 | 3 |
| slow adder chains (this run) | 5 | 6.21 | 0 |
| fast control chains (this run) | 8 | 1.31 | 2 |
| slow control chains (last run) | 10 | 8.83 | 0 |
The slow adder chains behave like the slow controls, and the fast controls like the fast adder chains. Averaged across the cells, a short mapped path costs the ladder 6.08 points of its edge (95% interval 2.58 to 9.73). Adder-chain structure costs 1.18 (interval -2.43 to 4.69, so no clear effect). Every loss in either run is on a design that maps under 15 ns. The primary prediction, that structure was the cause, is refuted. Four of the eight predictions held.
What it means
For someone pruning a routing search at a fast corner, what predicts trouble is a cheap number already on hand, the baseline mapped delay. What computes the shift amount does not. That suggests the next rule to test blind: when the baseline maps fast, route more points or fall back to random.
Limits
The slow adder cell is one family of 5 designs. Two of the four cells come from the previous run (same tools, liberty and protocol). Baseline delay both picked the designs and serves as the factor. The study covers one corner, varies the liberty only, and uses global-route estimates. Numbers, code and the pre-registration commit are in experiments/delay_split_ff/.