The verified-dse programme searches over ABC logic-synthesis pass sequences with Yosys, ABC and OpenROAD on SKY130. Its practical rule ranks candidates by a cheap mapped-netlist score and routes only the best few. Wide shifters are where that ranking breaks. For three steps we have tried to predict when it breaks from the circuit's structure alone.
The last step found a clean line on its blind shifters. The ranking won whenever 33 or more AND/NOT gates fed the shift amount, and a 32-gate threshold looked right. This step registered that threshold and its predictions before running anything. It then ran eight shifters from four families no earlier table had used: 1,920 place-and-route runs.
Four of eight predictions failed.
| design | gates in front of the amount | cheap ranking, regret | random routing, regret |
|---|---|---|---|
| 3-bit by 7-bit product | 28 | 0.00% | 11.5% |
| sum of three 5-bit values | 31 | 0.00% | 9.4% |
| 64-bit compare picks the amount | 33 | 0.00% | 2.6% |
| sum of three 6-bit values | 35 | 0.00% | 2.4% |
| 4-bit by 7-bit product | 36 | 23.1% | 3.9% |
(Each regret is the mean over the tight and relaxed clocks. Regret is how far the chosen design's routed area times delay sits above the best in the pool.)
The ranking wins below the line and loses above it. The product rows settle it: the same circuit family wins at 28 gates and loses at 36. On these 16 cells, the ranking with no guard has 3.56% regret. Random routing has 5.69%, and the registered 32-gate guard has 6.79%, worse than both.
Three structural rules have now been tuned on past data and then failed or moved on the next blind set: "is there a shift operator", "how many word-level operations feed the amount" and "how many gates feed it". We read that as a sign the structural approach is spent. The next step will measure instead. It will route a few of the pilot's points and check whether the cheap score's order holds on them, priced first by replay on the 90 cells already run.
One question did get a positive answer. Earlier 64-bit shifters were too cheap to route for the rule to matter. With two or four lanes sharing one amount, the routed-to-mapped cost ratio rises from 3.4 to 5.6. On those cells the cheap ranking beats random routing: 1.20% against 3.84% regret.
Limits: 16 cells from 8 designs, one tool version, one corner, global-route estimates. Code, designs, the registered predictions and the full table: experiments/amount_lanes/.