Why it matters

Looking at a failing endpoint tells you what fails. Looking at which scenarios it fails in tells you why. The second question is faster to answer and narrows the search before any path is read.

Pattern table

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17

def parse(view):
    mode, pvt, rc = view.split("/")
    return {"mode": mode, "pvt": pvt, "rc": rc}
all_views = [f"{m}/{p}/{r}" for m in ("func", "scan") for p in ("ss", "ff") for r in ("cw", "cb")]
def explain(failing):
    attrs = [parse(v) for v in failing]
    out = {}
    for key in ("mode", "pvt", "rc"):
        values = {a[key] for a in attrs}
        universe = {parse(v)[key] for v in all_views}
        if len(values) == 1 and len(universe) > 1:
            out[key] = values.pop()
    return out
assert explain(["scan/ss/cw", "scan/ss/cb", "scan/ff/cw", "scan/ff/cb"]) == {"mode": "scan"}
assert explain(["func/ss/cw", "scan/ss/cw"]) == {"pvt": "ss", "rc": "cw"}
assert explain(all_views) == {}
print("PASS: pattern explanations computed for three failure sets")

Four failing views that are exactly the scan-mode views are explained by the mode alone. Two failures that share both the slow process and the worst-capacitance corner are explained by those two attributes together. When every view fails, no attribute is shared more narrowly than the whole set, which is why that pattern points at constraints or the netlist rather than a corner.

What the review looks for

• Before any path analysis, a table of failing scenarios by attribute.

• A claim of cause names the attribute and the evidence that other values of it pass.

• A pattern explained by several attributes is split with one more scenario run.

Related topics

• Next: Per-View Triage