Fill insertion broke timing after signoff. What do you do?

Short Answer

First prove it with extraction that includes the real fill, then repair only around the nets that lost slack. Remove fill around critical nets with signoff_create_metal_fill -mode remove (ICC2) using -nets or a setup slack threshold, reinsert with timing-driven spacing, and recheck density and timing. Never rip out fill block-wide to recover a few picoseconds.

Technical Explanation

• Fill adds coupling capacitance to nearby signal wires. Floating fill still couples, so a net that had 8 ps of slack with estimated fill can go negative once the real shapes are extracted.
• Measure it first. Associate non-emulation TLUPlus files with the corners with set_parasitic_parameters (ICC2) and enable set_extraction_options -real_metalfill_extraction floating (ICC2). Then write parasitics and rerun STA in PrimeTime so you compare against signoff numbers.
• Removing fill only over critical nets is built in: signoff_create_metal_fill -mode remove (ICC2) with -nets names the nets, and -timing_preserve_setup_slack_threshold (ICC2) picks them from a setup slack threshold in library time units.
• Timing-driven insertion then refills while keeping away from critical nets. It runs timing analysis including MCMM, removes existing fill, and inserts fill everywhere except near critical nets. Default spacing to their shapes is twice the technology minimum, also applied to the vertical extension on adjacent layers.
• Choose the threshold carefully. The guide warns that a large threshold selects too many critical nets, which lowers metal density and leaves large empty areas.
• The extra keep-out can create density errors. Set signoff.create_metal_fill.fix_density_errors (ICC2) to true (default false) to repair them during timing-driven insertion, then run density rules again.
• Close the loop with extraction, PrimeTime and DRC on the final fill. A timing gain that creates a density violation just moves the problem to the fab.

What To Check

• Setup slack of the affected paths with real fill extracted, not with virtual fill.
• How many nets the chosen slack threshold selects.
• Density per window on the affected layers after reinsertion.
• Hold in the fast corner, since removing fill lowers capacitance.

Healthy, Suspicious & Hard-stop Results

• Healthy (illustrative): Setup WNS returns from -9 ps to +4 ps and all density windows stay within limits.
• Suspicious (illustrative): Timing recovers but two windows now sit within 1% of minimum density.
• Hard stop: Density violations after the fix, or a threshold so large that fill is removed from most of the block.

What The Interviewer Is Testing

• Does the candidate insist on real fill extraction before changing anything?
• Knowledge of targeted fill removal and timing-driven fill in ICC2.
• Awareness of the density cost of the keep-out.

Follow-up Question & Model Response

"Why does fill hurt some nets much more than others?"

Candidate Model Response: Coupling capacitance from fill matters most on nets that are long, run on dense layers, and are driven weakly relative to their load. A short net with a strong driver barely notices the added capacitance. Clock nets and high-speed buses often also have tight slack, so they are the first to go negative. That is why the tool picks nets by slack rather than removing fill uniformly.