RTL-vs-GLS Comparison & Mismatch Isolation

Chapter 12 · Section 12.5 · GLS Debug Methodology

Project thread — this generalizes the RTL-vs-GLS X-optimism story of 6.4 into a debug technique. This lesson isolates any RTL-vs-GLS mismatch; 12.6 catalogs the recurring mistakes.

1. Why Should I Learn This?

Comparing RTL and GLS isolates a gate-level bug fast — if you judge it correctly.

• A GLS-only failure → a gate-level-specific cause (X/timing/reset/CDC/power/DFT).
• Isolate by diffing at ports/hierarchy boundaries → module → signal → first divergence.
• Caveat: RTL is not always golden for X — a GLS X can be right (6.4).

This turns 6.4's insight into a repeatable isolation technique.

2. Real Silicon Story — the mismatch that proved GLS right

A GLS run mismatched RTL on a signal. The reflex was "GLS is wrong — RTL is golden" and to hunt for a gate-level artifact.

There was no artifact. The signal depended on an uninitialized control that RTL's X-optimism (6.4) had resolved to a definite value (hiding the bug), while GLS faithfully propagated X — so the GLS X was correct and the RTL value was the misleading one. RTL was not golden here. Fixing the uninitialized source (the real bug) made both runs agree — with the right value.

Lesson: an RTL-vs-GLS mismatch doesn't automatically mean GLS is wrong. Judge by the source — RTL's X-optimism (6.4) can make a GLS X the correct one and RTL the misleading one.

3. Concept — comparing and isolating

The comparison (isolation power):

• Run RTL and GLS on the same stimulus; compare their signals.
• Where they agree, the gate-level behavior matches intent; where they diverge, you've localized a difference.

How to isolate a mismatch:

1. Dump both runs on identical stimulus.
2. Diff at boundaries — compare at ports / hierarchy boundaries (with a strobe after settling, 5.4) to narrow to a module, then to a signal.
3. Find the first divergence (12.2) — the earliest diverging signal/time.
4. Classify (12.1) — is the GLS-only difference an X-source (Ch6), timing (Ch8), reset (Ch7), CDC (Ch9), power (Ch10), or DFT (Ch11) effect?

RTL as reference — with the caveat:

• RTL is usually golden — it defines intended behavior, so a GLS-only failure is usually a gate-level-specific cause to fix.
• BUT RTL is not golden for X. RTL is X-optimistic (6.4): it resolves unknowns to a branch, so it can hide a real bug that GLS's faithful X-propagation exposes.
• Therefore: judge by source, not by which run looks cleaner. A GLS X that traces to a real uninitialized/timing source is correct — RTL was the misleading one.

Decision:

• GLS artifact (setup/expected-X, 12.1) → fix the run.
• GLS-exposed real bug (RTL-optimism hid it, or a real timing/reset issue) → fix the design.
• Real RTL-vs-GLS logic difference (rare, e.g. a synthesis issue) → investigate.

Scope (accuracy):

• The comparison isolates; judge by source (not "RTL golden" blindly). GLS stays dynamic (0.3).

4. Mental Model — two witnesses, and neither is automatically right about X

RTL and GLS are two witnesses to the same events (same stimulus).

• Where they agree, you trust the account; where they disagree, you've found the contested moment (the mismatch) — narrow to when and where (first divergence, boundary diff).
• The usual assumption is that RTL (the intent) is the reliable witness — so a GLS-only discrepancy is a gate-level detail to explain.
• But on questions of X (uncertainty), the RTL witness is an optimist who confidently fills in gaps (6.4) — it may state a definite value it couldn't actually know, while the GLS witness honestly says "unknown."
• So you don't just believe the "confident" witness — you check the evidence (the source of the difference). If GLS's X traces to a real uninitialized/timing cause, the honest witness was right and the optimist misled you.

Two witnesses; agree = trust; disagree = investigate — and don't assume the confident one is right about the unknowns.

5. Working Example - isolating a mismatch

Isolating and judging a mismatch (tool-neutral):

# RTL-vs-GLS mismatch isolation - REPRESENTATIVE (tool-neutral)
# 1) run RTL and GLS on the SAME stimulus; diff signals at ports (strobe after settle, 5.4)
#      mismatch at: u_dut.out[3]  (RTL=1, GLS=X)
# 2) narrow to module: diff at u_dut.u_ctrl boundary -> mismatch at u_ctrl.sel (RTL=0, GLS=X)
# 3) first divergence (12.2): u_ctrl.sel diverges first
# 4) classify + JUDGE BY SOURCE:
#      GLS sel=X traces to an UNINITIALIZED flop (Ch6/7)  -> a REAL bug RTL-optimism HID (6.4)
#      => GLS X is CORRECT; RTL is the misleading one -> FIX THE DESIGN (init/reset the source)
#   (if instead GLS X traced to a mis-scoped check / missing SDF -> GLS ARTIFACT -> fix the run)

Practical context (representative, tool-neutral):

# RTL-vs-GLS comparison (tool-neutral):
#   1) run BOTH on the SAME stimulus; dump signals
#   2) diff at PORTS / hierarchy BOUNDARIES (strobe after settle, 5.4) -> narrow to module -> signal
#   3) find the FIRST divergence (12.2); classify (12.1)
#   4) JUDGE BY SOURCE:
#        GLS X from real uninitialized/timing source -> RTL-optimism HID a real bug (6.4) -> fix DESIGN
#        GLS X from setup/expected (SDF/check/corner) -> GLS artifact -> fix the RUN
#   NEVER assume 'RTL golden' for X -- judge by source.

RTL definite vs GLS X (GLS correct), as a real waveform:

6. Debugging Session — a GLS-only mismatch: GLS wrong, or RTL hiding a bug?

7. Common Mistakes

• Assuming "RTL is golden" for X. RTL-optimism (6.4) can make a GLS X the correct one.
• Judging by "which run looks cleaner" instead of the source.
• Not isolating — diff at ports/boundaries to narrow to a module/signal.
• Not finding the first divergence (12.2) before classifying.
• Dismissing a GLS X as an artifact without tracing its source.

8. Industry Best Practices

• Run RTL and GLS on identical stimulus and diff at ports/boundaries (strobe, 5.4).
• Isolate to a module → signal → first divergence (12.2).
• Judge by source — GLS artifact (fix run) vs GLS-exposed real bug (fix design).
• Remember RTL isn't golden for X (6.4) — trace before concluding.
• Fix at the source; verify both runs agree with the right value.

Senior Engineer Thinking

• Beginner: "GLS disagrees with RTL — GLS must be wrong."
• Senior: "RTL is usually golden, but not for X — RTL-optimism hides bugs GLS exposes. Let me isolate the mismatch and trace the source: real uninitialized cause → RTL hid a bug (fix design); setup/expected → GLS artifact (fix run)."

The senior isolates the mismatch and judges by source, never assuming RTL is golden for X.

Silicon Impact

RTL-vs-GLS comparison is a fast isolation technique — but its judgment is where it protects (or endangers) the tape-out. Assume "RTL is golden" blindly and you'll dismiss a GLS X as an artifact when it's actually exposing a real bug RTL's optimism hid (6.4) — letting an uninitialized-control/reset bug reach silicon (0.3). Conversely, treating every GLS-only difference as a real design bug wastes effort on setup artifacts. The discipline — isolate to the first divergence, trace to the source, judge by the source — is what makes the comparison conclusive: a GLS X from a real uninitialized/timing source is a design fix; one from a mis-scoped check or missing SDF is a run fix. This generalizes 6.4's insight into the everyday debug loop, keeping RTL a reference without making it an infallible oracle for X.

Engineering Checklist

• Ran RTL and GLS on identical stimulus; diffed at ports/boundaries (strobe, 5.4).
• Isolated to a module → signal → first divergence (12.2).
• Traced the difference to its source (12.3).
• Judged by source (GLS artifact → run; GLS-exposed real bug → design).
• Did not assume "RTL golden" for X (6.4).

Try Yourself

1. Run RTL and GLS on the same stimulus with an uninitialized control; diff at the ports.
2. Observe: RTL shows a definite value; GLS shows X — a mismatch.
3. Change: trace the GLS X to its source — it's the uninitialized control (a real bug RTL-optimism hid, 6.4).
4. Expect: the GLS X was correct; fix the design (init the source) and both runs agree with the right value. Then contrive a GLS artifact (missing SDF) and confirm that one is a run fix — proving you judge by source.

RTL-vs-GLS comparison uses two runs in any simulator. No paid tool required.

Interview Perspective

• Weak: "RTL is golden, so any GLS mismatch means GLS is wrong."
• Good: "I compare RTL and GLS on the same stimulus, isolate the mismatch to a signal, and find a gate-level cause."
• Senior: "I diff RTL and GLS at boundaries, isolate to the first divergence, and — crucially — judge by the source. RTL is usually golden, but not for X: RTL-optimism (6.4) hides bugs GLS exposes, so a GLS X can be the correct one. A GLS X from a real uninitialized/timing source is a design fix; from a mis-scoped check or missing SDF it's a run fix. I never assume 'GLS is wrong.'"

9. Interview / Review Questions

10. Key Takeaways

• Run RTL and GLS on the same stimulus and compare to isolate a mismatch — where they diverge localizes a difference.
• Isolate: diff at ports/hierarchy boundaries (strobe after settling, 5.4) → narrow to a module → signal → first divergence (12.2) → classify (12.1).
• RTL is usually golden — a GLS-only failure is usually a gate-level-specific cause (X/timing/reset/CDC/power/DFT) to fix in the design or run.
• But RTL is not golden for X — RTL is X-optimistic (6.4), so a GLS X can be correct where RTL hid a real bug; judge by the source, not by which run looks cleaner.
• Verdict by source: GLS X from a real uninitialized/timing source → fix the design (RTL hid it); GLS X from a setup/expected cause → GLS artifact, fix the run. The comparison isolates; the source decides. Next: 12.6 — common GLS mistakes & how to avoid them.

Quick Revision

Compare RTL vs GLS on the SAME stimulus to ISOLATE a mismatch: diff at ports/boundaries (strobe 5.4) → module → signal → first divergence (12.2) → classify. RTL usually golden → GLS-only failure = gate-level cause. CAVEAT: RTL NOT golden for X — RTL-optimism (6.4) means a GLS X can be CORRECT (RTL hid a bug). JUDGE BY SOURCE: real uninit/timing source → fix DESIGN; setup/expected → GLS artifact, fix RUN. Next: 12.6 — common GLS mistakes.