TLM Unconnected & Debug Interview Q&A

[INT][SENIOR][UVM] MODEL ANSWER

Q: Scoreboard empty — first TLM checks?

A:
  MECHANISM:  Empty scb = no write() calls reached imp. Check mon.ap.size()>0, monitor
              actually sampling, connect path correct, txn type match.
  MOTIVATION:  90% of empty scb is connect bug not scoreboard logic bug — verify dataflow
              before debugging compare algorithm.
  WHEN:       Step 1: ap.size(). Step 2: uvm_info in monitor write/sample. Step 3: uvm_info
              in scb write first line. Step 4: end_of_elaboration topology print.
  PITFALL:    Debugging compare mismatch algorithm when write() never called — hours wasted.
  EXAMPLE:    ap.size()==0 — env.connect forgot agt_rx.mon.ap.connect(scb.rx_imp);
              fix connect, scb immediately receives txns.

[INT][SENIOR][UVM] MODEL ANSWER

Q: uvm_tlm_gp debug?

A:
  MECHANISM:  +UVM_TLM_TRACE or UVM TLM verbosity displays connect and transaction trace
              at GP (generic payload) level for TLM-2.0 style; UVM 1.x has port display.
  MOTIVATION:  Traces TLM connect graph and write/put/get calls — confirms data movement.
  WHEN:       +UVM_VERBOSITY=UVM_HIGH on TLM-related components; port.display_connections().
  PITFALL:    Enabling global UVM_FULL on 12-hour sim — log volume TB; scope to scb/mon.
  EXAMPLE:    mon.ap.display_connections() shows 0 connections — immediate root cause
              without waiting for stimulus phase debug.

[INT][SENIOR][UVM] MODEL ANSWER

Q: Unconnected — compile or runtime?

A:
  MECHANISM:  port.connect() to invalid target = compile error. Never calling connect
              = compiles fine; analysis write at runtime silent drop. put/get block forever
              or fail on first call depending on implementation.
  MOTIVATION:  Analysis designed for optional subscribers — no connect is valid at compile.
              put/get pipelines require connect or explicit block on first use.
  WHEN:       Audit connect completeness in end_of_elaboration with size() checks.
  PITFALL:    Assuming compiler catches all TLM errors — analysis unconnected is runtime silent.
  EXAMPLE:    ifdef DISABLE_SCB around connect call — compiles; scb empty at runtime;
              size()==0 catches ifdef mistake in end_of_elaboration.

[INT][SENIOR][UVM] MODEL ANSWER

Q: Driver stuck get_next_item?

A:
  MECHANISM:  get_next_item blocks until sequencer provides item — stall means no sequence
              running, seq_item_port not connected, or sequencer arbitration lock.
  MOTIVATION:  Separate from objection hang — driver blocked in run_phase but phase may
              still have objections from test; display_objections + driver log together.
  WHEN:       Check seq_item_port connection in agent.connect. Verify sequence started on
              correct sequencer. Check arbitration if multiple sequences compete.
  PITFALL:    Adding delay in driver — does not fix missing sequence start.
  EXAMPLE:    vseq started on env.v_sqr but driver on agt.sqr — seq_item_port connected
              yet no items for driver; fix vseq to start agt_seq on agt.sqr.

[INT][SENIOR][UVM] MODEL ANSWER

Q: Systematic TLM triage?

A:
  MECHANISM:  Symptom classify → connection audit → producer check → consumer check →
              type/path match → compare logic (last).
  MOTIVATION:  Ordered triage prevents premature deep debug — connection bugs dominate.
  WHEN:       1) ap.size()/display_connections 2) monitor sample count 3) scb write count
              4) txn type/param match 5) compare algorithm 6) waves last.
  PITFALL:    Waveform-first on empty scb — 2 hours before checking ap.size()==0.
  EXAMPLE:    Mismatch symptom: scb write count=500, exp queue=500, act queue=500 but
              compare fails — NOW debug algorithm. Before write count=0: connect bug.

[INT][SENIOR][UVM] MODEL ANSWER

Q: Generate loop partial connect?

A:
  MECHANISM:  foreach/agt[i] connect loop with off-by-one, wrong index, or conditional
              skip leaves some agents unwired — their txns silently drop.
  MOTIVATION:  Multi-agent chip TBs scale connect code — loop bugs are high frequency.
  WHEN:       end_of_elaboration: for i assert agt[i].mon.ap.size()>0. Log connect in
              env.connect with index for audit trail.
  PITFALL:    Connect agt[1:N] but create agt[0:N-1] — index mismatch, agt[0] unwired.
  EXAMPLE:    4 agents, loop for(int i=1;i<4;i++) — agt[0] never connected; 25% traffic
              missing from scb, intermittent failures on agt[0]-only scenarios.

[INT][SENIOR][UVM] TLM triage checklist

  [ ] mon.ap.size() > 0  (end_of_elaboration)
  [ ] monitor sample count > 0  (first 100 cycles)
  [ ] scb write() count > 0
  [ ] exp and act queue depths track
  [ ] txn type match set/get/connect
  [ ] THEN compare algorithm / waves

[INT][SENIOR][UVM] MODEL ANSWER

Q: Detect subscriber slowness?

A:
  MECHANISM:  analysis_port does not report backpressure — monitor unaware subscriber slow.
              Detect via scb queue depth growth, time delta between sample and compare log.
  MOTIVATION:  Subscriber slowness does not block monitor but causes compare lag and
              potential memory growth in scb internal queues.
  WHEN:       Monitor scb queue depth each write; assert depth < threshold. Profile
              write() compute if depth grows unbounded.
  PITFALL:    Adding wait in monitor to 'slow down' for scb — distorts cycle accuracy;
              fix scb performance or use fifo decoupling.
  EXAMPLE:    scb write() O(n) search in queue — 10k txn sim, queue depth 5000, sim
              slows; fix hash map compare, depth stays O(1).

[INT][SENIOR][UVM] MODEL ANSWER

Q: TLM connectivity smoke test?

A:
  MECHANISM:  Directed smoke: send 1 txn, assert scb write count==1 and compare pass.
              end_of_elaboration assert all ap.size()>0 before run_phase.
  MOTIVATION:  Cheap test catches connect regression on every CI build before random stress.
  WHEN:       tlm_connect_smoke_test in <10s regression: one directed txn, count audit,
              no randomization dependency.
  PITFALL:    Only random tests in CI — connect bug may hide until specific agent index
              gets traffic in unlucky seed.
  EXAMPLE:    Smoke sends 1 MMIO write per agent in sequence; scb assert per-agent
              write count >= 1 in check_phase — connect verified for all indices.