DenseChip

Part 2 · Combinational · Senior Prep

Intentional latch inference

Latch Avoidance: Intentional latch inference.

Intentional latch inference

If a latch is truly required (rare in mainstream ASIC/FPGA digital flows), code it explicitly and document intent. Otherwise, lint should flag latch inference as an error.

diagram
INTENTIONAL LATCH INFERENCE

  ┌─────────┐     ┌──────────────┐     ┌─────────┐
  │ inputs  │ ──► │ combinational   │ ──► │ outputs │
  └─────────┘     │ mechanism    │     └─────────┘
                  └──────────────┘
                        │
                  pitfalls / when-to-use
                  (intentional-latch-inference)

  • Treat latch warnings as must-fix unless architecture explicitly calls for level-sensitive storage.

  • Prefer flip-flop based synchronous design for robust timing closure.

Deep dive

This lesson deepens Intentional latch inference within combinational. Senior reviewers expect you to connect mechanism to logic depth, latch inference, and arithmetic width — not just define terms.

Start from the spec invariant: what must always be true each cycle? Write it as a Boolean relation, timing budget, or protocol rule before coding. That invariant becomes your reference model, assertion, or waveform check.

At tape-out quality, every block needs a sign-off story: no latches, ref model matches, delay intuition documented. Treat this lesson as building one paragraph of that story for your project documentation.

Architecture and signal flow

diagram
INTENTIONAL LATCH INFERENCE

  inputs ──► [combinational] ──► mechanism ──► outputs
                │                │
           latch-inference     verify in sim + lint

Worked example (Verilog/SystemVerilog)

Synthesizable pattern for this topic — simulate locally and compare against your reference.

verilog
always_comb begin
  y = 1'b0;
  if (en) y = a;
end

Step-by-step design procedure

  1. Write the spec invariant (truth table, timing, or protocol rule).

  2. Sketch block diagram — inputs, outputs, clock/reset domains.

  3. Code the minimal correct version (no optimizations yet).

  4. Run self-checking TB with corners: min, max, reset, idle.

  5. Lint and review: width, latch, clock, CDC if applicable.

  6. Iterate for timing/area only after functionally proven.

Timing and resource trade-offs

diagram
METRIC          TYPICAL LEVER
Logic levels      algebra / pipelining
Register count    retiming / sharing
Wire fanout       duplication / pipeline
Power             clock gating / operand isolation
Debug visibility  status flags / SVA / waveform probes

Debug checklist

  • Compare DUT vs reference on every stimulus vector

  • Capture first cycle of mismatch — not last

  • Log seed and plusargs for random regressions

  • Check reset release and clock alignment in TB

  • If waveform is ambiguous, add temporary assertions

Interview angle

How do you prove no latch was inferred?

diagram
MODEL ANSWER SKELETON
1. MECHANISM — one-sentence technical truth
2. MOTIVATION — why this structure vs alternatives
3. WHEN TO USE / SKIP — scope and assumptions
4. PITFALL — common junior mistake
5. EXAMPLE — Verilog or waveform scenario

Practice exercise

Extend the worked example for "Intentional latch inference": add one corner case, write a self-checking test, and document one intentional pitfall you avoided. Timebox: 30–45 minutes.

Key takeaways

  • Connect Intentional latch inference to logic depth, latch inference, and arithmetic width.

  • Spec → diagram → code → TB → lint is the default loop.

  • Document assumptions at block boundaries (width, signedness, latency).

  • Interview answers need mechanism + trade-off + example.

Common pitfalls

  • Skipping reference model — passes wrong together with DUT.

  • Optimizing before correct — ECO cost goes up 10×.

  • Ignoring tool warnings — lint/CDC/STA warnings are technical debt.

  • Undocumented clock/reset domain — integration surprises.

Extended design scenario

Scenario for Intentional latch inference : Priority mux depth is 8 levels — refactor to balanced tree or pipeline with documented latency.

Scenario resolution outline

  1. Reproduce with minimal TB — one stimulus, one check.

  2. Isolate failing cone (logic, FSM state, or bus beat).

  3. Fix root cause — not symptom — in RTL or TB alignment.

  4. Add regression test that fails without the fix.

  5. Document invariant in comment or SVA for permanence.

Additional simulation pattern

verilog
// Exhaustive 3-input truth table
for (int a=0; a<2; a++)
  for (int b=0; b<2; b++)
    for (int c=0; c<2; c++)
      check_ref(a,b,c);

Synthesis and sign-off notes

  • Elaborate clean — no OOM from unconstrained generate

  • Constraints cover all clocks and I/O delays

  • Cross-check RTL parameters vs integration top

  • Attach sim log + seed to code review

Lab exercise (45–60 min)

Implement or extend the worked example for "Intentional latch inference". Add two new test vectors that target different branches. Write a one-paragraph sign-off note covering function, corners, and what you would still verify in SoC context.

Further reading in this course

diagram
Next topics in Part: follow nav_order in sidebar.
Cross-part: timing ↔ sequential, CDC ↔ senior interview,
combinational ↔ foundations number systems.

Extended theory

Combinational cones set the raw MHz ceiling of a block. Before pipelining, quantify depth: each wide mux or carry chain adds delay that shows up as setup slack erosion at the next register boundary.

For "Intentional latch inference", the invariant you defend in review is: behavior matches spec under all legal input sequences, reset flows, and backpressure patterns—not just the happy path shown in introductory diagrams.

Write the invariant as a comment above the module or as an SVA property when possible. Future you (and formal tools) will treat it as the contract.

Waveform reading guide

diagram
WAVEFORM NARRATIVE — Intentional latch inference

cycle 0: reset asserted, outputs safe/idle
cycle 1-2: reset held, clocks running
cycle 3: reset released, first legal inputs
cycle 4+: check output latency (N cycles)
mark FIRST mismatch cycle — not last

Second worked example

Alternate pattern emphasizing debug, coverage, or integration:

verilog
// Pipeline-style comb for timing (register inputs)
always_ff @(posedge clk) a_r <= a;
always_ff @(posedge clk) b_r <= b;
assign y = a_r & b_r;  // shorter path from flop

Comparison: naive vs production

diagram
NAIVE APPROACH              PRODUCTION APPROACH
quick hack, one sim vector    self-checking TB + corners
ignore lint warnings          zero new waivers
implicit widths               explicit casts/parameters
undocumented latency          latency in module header comment

Additional interview questions

  • Explain Intentional latch inference to a verification engineer — what would they assert?

  • What breaks first at high frequency or low voltage?

  • What is your rollback plan if synthesis QoR is unacceptable?

  • How would you debug this block with only a 32-bit GPIO trace?

Follow-up interview model answer

diagram
Q: What is the #1 mistake with Intentional latch inference?
A:
  MECHANISM: [core rule in one line]
  MOTIVATION: why teams care in tape-out
  PITFALL: what juniors do wrong
  EXAMPLE: one Verilog line or one waveform event

Hands-on lab part 2

  1. Fork the worked example; add one assertion or SVA cover.

  2. Inject a bug deliberately; confirm TB or assertion catches it.

  3. Write 5-bullet PR description for your change.

  4. Peer review: can a teammate enable the block without asking you?

Sign-off evidence checklist

  • Directed sim log attached (PASS, seed noted)

  • Lint report clean for touched files

  • If sequential: reset + clocking section in README

  • If bus-facing: protocol cheat sheet in module doc

  • If timing-critical: note expected critical path endpoint

Practice this lesson