Regression Coverage Merge: One Truth Across Seeds

Single-seed coverage lies

A lucky seed can hit rare bins once; closure needs the union across all regression runs . Merge tools (Questa UCDB, VCS VDB, Xcelium IMC) combine per-test databases into one view where a bin is covered if any seed hit it.

[CLOSE][SENIOR][UVM] merge topology

  seed_1001/test_axi_base.ucdb  ──┐
  seed_2048/test_axi_base.ucdb  ──┼──► merge script
  seed_4099/test_axi_rand.ucdb  ──┤         │
  seed_7777/test_axi_stress.ucdb──┘         ▼
                                    merged_regression.ucdb
                                         │
                                         ▼
                              hole report + trend dashboard

[CLOSE] typical merge workflow (tool-agnostic)

  1. Each test writes -coverage + ucdb/vdb at end of sim
  2. Nightly job collects all *.ucdb from regression farm
  3. merge -stats -out merged_$(date).ucdb *.ucdb
  4. report -detail -uncovered > holes_$(date).txt
  5. diff vs yesterday → new hits / regressions

• Tag each database with test name, seed, and git SHA in the regression manifest.

• Merge includes all closure-relevant tests — not only the nightly smoke subset.

• Track merge regressions: a bin that was covered yesterday but not today is a red flag.

Key takeaways

• Closure metrics come from merged databases — never a single seed.

• Manifest test+seed+SHA for reproducibility and audit.

• Diff merged reports day-over-day to catch coverage regressions.

Common pitfalls

• Merging only passing tests — failing runs may still contribute bin hits.

• Stale databases from old RTL polluting the merged view.

• No seed attribution when a hole closes — cannot reproduce the hit.

UVM report_phase and database export

Dump per-instance coverage in report_phase for smoke visibility; the regression merge still owns closure metrics:

function void report_phase(uvm_phase phase);
  super.report_phase(phase);
  real inst_cov = cg.get_inst_coverage();
  `uvm_info("COV", $sformatf("%s inst_cov=%0.1f%%",
    get_full_name(), inst_cov), UVM_LOW)
  // sim flag -coverage writes UCDB; merge happens offline
endfunction

Merge hygiene checklist

1. Purge databases older than current RTL freeze tag.

2. Verify same coverage options (per_instance, cross auto-bin settings) across tests.

3. Confirm passive-agent coverage at chip level is included in merge.

4. Archive merged DB + hole report with milestone tags (M3, M4, sign-off).

[CLOSE][SENIOR] seed attribution example

  HOLE CLOSED: x_wr_large.wr_large
  FIRST HIT:   test_axi_stress seed=4099  (2026-06-08 nightly)
  REPRO:       vsim +seed=4099 +UVM_TESTNAME=axi_stress_test

Key takeaways

• report_phase is for per-run visibility; merge is for closure truth.

• Archive merged DBs at milestones — auditors will ask.

Common pitfalls

• Changing covergroup definitions without invalidating old databases.

• Mixing block-level and chip-level databases without hierarchy awareness.