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 (SWC), maps every ABC pass sequence, ranks the mapped netlists with a cheap score, and routes only the best 8. The baseline routes 8 sequences picked at random, for the same cost.
The last run moved to the fast-fast library corner (ff, -40 C, 1.95 V). The rule still won 20 of 27 blind designs. But the seven losses had a look to them: every carry-save adder and every dot product was among them. In both families the shifter's amount is computed by a chain of additions. A pattern spotted in the table that suggested it is a hypothesis, not a finding, so this run tested it on designs nobody had run at ff.
What this run did
It wrote down two groups of 128-bit shifter designs before running anything. The add-chain group has 10 designs whose shift amount sums three or more multi-bit terms: carry-save sums, dot products, multi-term sums, averages and sums of absolute differences. The control group has 10 designs whose amount comes from a compare, select, clamp or divide chain. Seven predictions went into the same commit. The primary one: the rule loses on a larger share of add-chain designs, with a one-sided Fisher exact p below 0.05.
The flow ran all 2,400 points (20 designs, 120 pass sequences each) in about 70 minutes on 4 cores.
What it found
| add chains | controls | |
|---|---|---|
| designs where the shortlist loses | 3 of 10 | 0 of 10 |
| mean regret, shortlist | 5.64% | 1.33% |
| mean regret, random routing | 7.21% | 10.16% |
| mean edge over random | 1.57 points | 8.83 points |
| rank correlation, mapped vs routed (median tau-b) | 0.27 | 0.47 |
- The primary prediction failed. 3 of 10 against 0 of 10 gives Fisher p = 0.105. Seven of the ten add-chain designs still favour the shortlist, so "it loses on adder chains at ff" is wrong as stated.
- A weaker version held. On add chains the shortlist's edge over random routing is 7.3 points smaller (label-permutation p = 0.026).
- The mechanism fits. The rule screens with the mapped netlist. On add chains, that cheap score agrees less with the routed result.
- The rule still wins overall: 16 designs to 3, with one tie (sign p = 0.0044).
The catch
Every add-chain design in this set maps to a shorter critical path (6.9 to 14.8 ns) than every control (16.7 to 30.2 ns). So this table cannot tell "the amount is an add chain" apart from "the design is short and fast". The earlier table has only three short designs outside the add-chain group, all counters, and the rule won all three. That points weakly at structure, but three designs cannot settle it. The next step is a set that breaks the confound on purpose: slow add chains and fast compare chains.
Two smaller disclosures. nixpkgs now ships Yosys 0.71, and the earlier ff runs used 0.70. Also, the session's worker restarted near the end, and the last 141 points ran under --resume.
Full table, predictions and code: experiments/addchain_ff/.