Why it matters

SI fixes change arrival times, so windows shift and other victims can start overlapping. A fix verified only on its own net can leave the design worse. A before-and-after table over all endpoints catches regressions.

Worked example

1
2
3
4
5
6
7
8
9
10
11
12

before = {"e1": -0.020, "e2": 0.015, "e3": 0.040, "e4": 0.008, "e5": 0.120}     # setup slack, ns
after  = {"e1":  0.004, "e2": 0.010, "e3": 0.045, "e4": -0.006, "e5": 0.118}
fixed       = sorted(e for e in before if before[e] < 0 <= after[e])
broken      = sorted(e for e in before if before[e] >= 0 > after[e])
worse       = sorted(e for e in before if after[e] < before[e] - 1e-9)
assert fixed == ["e1"] and broken == ["e4"]
assert worse == ["e2", "e4", "e5"]
wns_before, wns_after = min(before.values()), min(after.values())
assert wns_before == -0.020 and wns_after == -0.006
net_ok = len(broken) == 0
assert net_ok is False
print("PASS: fixed", fixed, "; newly broken", broken, "; slack worse on", worse, "; WNS", wns_before, "->", wns_after)

The fix cleared e1, and the WNS improved from -0.020 to -0.006 ns, which looks like progress. But e4 went from +0.008 to -0.006 ns: the fix changed arrival times and e4 now overlaps with a different aggressor, or lost a speedup it relied on. The WNS is the same number as e4's slack, so the improvement hides the fact that a fresh violation has appeared. The summary row alone would have said 'better'; the endpoint table says 'one fixed, one broken'. e5 slipped by 2 ps, which is within analysis noise rather than a regression.

What the review looks for

• Verification compares slack for all endpoints in all scenarios, not just the targeted ones.

• Newly broken endpoints are flagged and fixed before the ECO is accepted.

• Small slack changes (a few picoseconds) are expected: noise in the analysis is not a regression.

Related topics

• Victim and Aggressor Changes