Vdebug

Verilog Debugger

Overview

This guide defines the workflow for diagnosing RTL failures captured in issue reports. Starting from the failure context, the skill traces signal paths statically through the module hierarchy, observes intermediate nodes with a temporary testbench probe when needed, and classifies the root cause with file-level precision.

This skill diagnoses root cause only. It does not propose or apply any code change.

Operating Modes

| Mode | Input | Use case | |------|-------|----------| | Single | IP name + issue number (e.g. foo issue_003) | Deep-dive on one specific issue | | Batch | IP name only (e.g. foo) | Process all open issues; group by root cause |

In batch mode, cluster before tracing to avoid re-tracing a shared root once per failure:

1. Triage every issue file first (cheap — just the [ERROR] register address

each maps to) and group issues whose failing output is the same or adjacent.

1. Read the IP-level context (doc/.md, rtl/_top.v) once

and reuse it for all clusters.

1. Trace each cluster's shared signal path once (Step 3), then write a report

for every issue in the cluster from that single trace.

Process clusters from the lowest issue number upward.

Outputs

• ./issue//debug_NNN_.md — one per issue analyzed (NNN matches the issue number)
• ./issue//debug_summary.md — batch mode only; root cause grouping across all analyzed issues

Before You Start — Re-run Safety & Scope

vdebug is a read-and-diagnose skill with one deliberate, temporary write: the optional probe in Step 4. Two scope rules make re-runs safe:

• Never write to rtl/ or lib/. Diagnosis reads synthesizable source but

never edits it. All probing happens in a disposable testbench include (tb//vdebug_probe.svh, Step 4) that reaches internal nodes by hierarchical reference (u_dut..) — the DUT is never instrumented.

• Clean stale probes first. A prior run may have aborted between Step 4A and

4C. Before starting, confirm no probe artifact survives: grep -rn "vdebug probe" tb// returns nothing, tb//vdebug_probe.svh does not exist, and no tb_*.sv still ` include ``s it. Remove any leftover before diagnosing.

Re-running on an already-analyzed issue overwrites its debug_NNN_.md. That is non-destructive: the report regenerates from the issue file and the RTL, both ground truth, so no confirmation is needed.

Prerequisites

Steps 1–3 (triage and static tracing) need no external tool — a fully static diagnosis is the common path. The simulator is required only if Step 4 (dynamic probing) is reached:

| Tool | Purpose | Check | |-------|----------------------|-------------------| | xvlog | Verilog compile | xvlog --version | | xelab | Elaborator | xelab --version | | xsim | Simulator | xsim --version |

If Step 4 is reached and any tool above is unavailable, do not guess a verdict: stop probing and classify the issue INCONCLUSIVE (Step 5), recording which probe would resolve it.

Step 1 - Triage the Issue

Read ./issue//issue_NNN_.md. Extract:

• The failing testcase file and module name
• The verbatim Failure Summary line
• Each [ERROR] line and the [FINISH] FAIL line from the Simulation Log. An

[ERROR] line is a single CPU-bus read mismatch — time, register address, observed value, expected value. Testcases check through the CPU bus only, so the log carries CPU-readable register values, never internal node values.

• The reproduction commands (compile, elaborate, run)

Build a working set of facts: the failing register address, its observed value, the expected value, and the simulation time. Map that address through the reg-map (in dut.vh / the module doc) to the DUT output or status register it reflects — that mapped signal is the top-level node Step 3 traces backward from.

Step 2 - Understand the Test

Read ./tb//tb_.sv (and the reusable sequences it calls in dut.vh). Identify:

• The stimulus sequence applied to the DUT (what inputs, in what order, under what conditions)
• The reference model or expected-value computation inside the testbench
• The exact check that fired (the [ERROR] line, e.g. from cpu_read_check or an explicit compare) — which output it checked and how the expected value was derived
• Whether the [ERROR] signal values captured at failure time are consistent with the stimulus (i.e., whether the testcase itself could be the source of the error)

If the expected value in the testcase is clearly wrong (miscalculated, wrong port name, incorrect timing assumption), note it immediately. Classification in Step 5 will be TC_BUG.

Step 3 - Trace the Signal Path (Static Analysis)

This is the primary debug method. Perform static RTL analysis before considering any simulation run.

Read ./doc/.md and ./rtl/_top.v to understand the IP structure (in batch mode read these once and reuse across all clusters).

Starting from the failing output port, trace backwards through the hierarchy using this process:

1. Identify which submodule at the current level drives the failing signal (look at wire assignments and port connections in the instantiation).
2. Read that submodule's documentation first (generate it with vexplain if missing). If the doc explains the signal's behavior and it matches spec, treat this module as correct and do not open its RTL — descend no further on this branch.
3. Only when the doc is inconclusive, open the RTL: Grep the signal name to find the always/assign that drives it, and read just that block plus its inputs — not the whole file.
4. List every upstream signal that feeds into that block.
5. Determine which upstream signal must have carried the wrong value. Only the top-level (CPU-readable) value comes from the [ERROR] line — deeper node values are not in the log. Reason from the logic which upstream is implicated; mark its value unknown until a Step 4 probe confirms it.
6. Descend into the submodule that produced the wrong upstream signal and repeat from step 1.

Continue until one of the following stopping conditions is met:

• A logic expression clearly computes the wrong result (wrong operator, wrong condition, missing case branch, off-by-one in counter)
• An FSM is in an unexpected state — check localparam state definitions and state transition conditions
• A pipeline stage register holds a wrong value traceable to a known incorrect computation upstream
• A primary input was driven with a value that the RTL was not designed to handle (testcase issue)

Use .claude/skills/shared/CodingStyle.md to interpret signal names while tracing — register suffixes (_ff), pipeline-stage prefixes (s0_/s1_), bus-direction prefixes (s_/m_), and 2 internal bus names all carry timing and module-pair information.

Record the full chain as a table:

| Level | Signal | Module | File | Value at Failure | |-------|--------|--------|------|-----------------| | Top output | m_data | foo_top | rtl/foo_top.v | 0x00 (expected 0xFF) | | Stage 1 | s1_result_ff | foo_proc | rtl/foo_proc.v | 0x00 | | Stage 0 | s0_sum | foo_proc | rtl/foo_proc.v | unknown | | Root | s0_op_sel | foo_ctrl | rtl/foo_ctrl.v | 1'b0 (should be 1'b1) |

Only the top row's value is read from the [ERROR] line; every deeper row is unknown until a Step 4 probe fills it. The example above shows a completed trace — its root value was confirmed by probing, not taken from the log.

If static analysis fully identifies the root cause, skip Step 4. But if any value in the trace table is still unknown at the suspected root, the RTL_BUG verdict is a hypothesis, not a confirmed fact — prefer one confirming probe run (Step 4) before classifying, when a simulator is available.

Step 4 - Dynamic Probing (Only If Step 3 Is Inconclusive)

Use probing only when static tracing reaches a point where multiple upstream signals could be the source and none of their values appear in the [ERROR] dump. Probing is white-box but never modifies rtl/ or lib/: it reads internal DUT nodes by hierarchical reference from a disposable testbench include.

4A — Add the Probe Include

Create tb//vdebug_probe.svh with one probe block that samples the suspect nodes from the Step 3 trace table. Reach each node by hierarchical reference rooted at the DUT instance — vtestgen names it u_dut in dut.vh — using the exact instance path from the trace, including generate/array labels where present (e.g. u_dut.g_lane[0].u_fifo.wptr). Clock the probe on the bench clock (the DUT's clock-port name declared in the testcase, e.g. clk):

// vdebug probe — temporary; delete this file and its `include before Step 5
always @(posedge clk)   // bench clock = the DUT clock-port name in the testcase
    if ()   // e.g. a cycle range or a sequence-phase flag
        $display("[PROBE] t=%0t s0_sum=%0h op_sel=%0b state=%0d",
                 $time, u_dut.u_proc.s0_sum, u_dut.u_ctrl.op_sel, u_dut.u_ctrl.state);

Then add one line at module scope inside the failing tb_.sv (e.g. just before endmodule):

`include "vdebug_probe.svh"   // vdebug probe — temporary

Notes:

• The file is .svh, not .sv, on purpose: the Makefile auto-discovers

tb_*.sv as testcases, and a .svh include is correctly ignored by that scan.

• Hierarchical refs reach the whole trace at once, so probe every unknown node

from the Step 3 table in this single block — no need to probe one level per run. Keep `` tight to limit output volume.

4B — Re-run

Run from inside tb//. The probe changes the testcase build, so force a rebuild by dropping the stale sim dir first:

rm -rf sim/tb_
make tb_
# probe output: sim/tb_/simulate.log

Locate the [PROBE] lines nearest the failure time and extract the values.

If elaboration fails with a hierarchical-name-not-found error, the referenced net was optimized away or the path is wrong: re-derive the path from the RTL instantiation chain, and if it is correct, elaborate with debug kept (e.g. xelab -debug typical) so the node is preserved.

4C — Remove the Probe

Delete tb//vdebug_probe.svh and remove the ` include ` line from tb_.sv`. Then confirm nothing survives:

grep -rn "vdebug probe" tb//   # must print nothing
ls tb//vdebug_probe.svh        # must not exist

The DUT under rtl//lib/ was never touched, so no RTL cleanup is required — only the testbench include. Never carry a probe artifact into Step 5.

Step 5 - Classify Root Cause

Assign exactly one classification:

| Class | Meaning | |-------|---------| | RTL_BUG | A logic block in a specific file computes the wrong value; behavior disagrees with the design documentation or proposal | | TC_BUG | The testcase stimulus or assertion is incorrect; RTL behavior is consistent with its specification | | DOC_AMBIGUITY | RTL behavior and testcase expectation both have a reasonable interpretation; the specification is unclear or contradictory | | INCONCLUSIVE | Static tracing left the root unresolved and dynamic probing could not run (simulator unavailable, or the node is unobservable); the report names the exact probe that would resolve it |

For RTL_BUG: identify the file, module, and the specific always block or assign statement. State what it currently computes and what it should compute, citing the relevant section of the design proposal or module documentation as the specification.

For TC_BUG: identify the testcase file and the specific assertion or stimulus that is wrong. State why it is incorrect.

For DOC_AMBIGUITY: quote the conflicting specification sections and describe the ambiguity.

For INCONCLUSIVE: state the lowest node in the trace whose value is still unknown, the hierarchical probe that would observe it, and why it could not be run. Do not fall back to guessing one of the other three classes.

Step 6 - Confirm Root-Cause Groups (Batch Mode Only)

The clusters formed up front (Operating Modes) were grouped by failing output. Now that each cluster is traced, confirm its members truly share one root-cause location (same file + same module + same block) and split any that diverged. A single RTL bug can surface as several failures; one confirmed group per root prevents redundant fix effort.

Step 7 - Write Debug Report(s)

Per-issue report: ./issue//debug_NNN_.md

NNN must match the issue number from the corresponding issue_NNN_.md.

Write the following sections:

Section: Issue Reference A single link to the corresponding issue file: issue_NNN_.md.

Section: Classification One of RTL_BUG, TC_BUG, DOC_AMBIGUITY, or INCONCLUSIVE.

Section: Root Cause Location The location for the chosen class, per its definition in Step 5 (file/module/block for RTL_BUG; testcase file + line for TC_BUG; both conflicting sources for DOC_AMBIGUITY; deepest node reached for INCONCLUSIVE).

Section: Root Cause Description Two to four sentences expanding the Step 5 finding for the chosen class, citing the relevant proposal or documentation section by name.

Section: Signal Trace The full chain table from Step 3, filled with signal names, module names, file paths, and values at failure time. Include every level traversed, even if intermediate values were correct.

Section: Evidence The [ERROR] and [FINISH] FAIL lines from the original issue report, plus any [PROBE] lines collected in Step 4 if probing was performed.

Closing note End the file with a blockquote: > This report identifies root cause only. No code change is proposed.

---

Batch summary: ./issue//debug_summary.md

Write this file only in batch mode, after all per-issue reports are complete.

Section: Results A table with the total issues analyzed and a per-classification breakdown. Use exactly this shape — one classification per row, the count in its own column on the same row — because vflow's scanner (scan.sh) reads the RTL_BUG count from this table to decide Phase 6, and it parses the first bare integer on the line that contains RTL_BUG. A header-row layout (| RTL_BUG | TC_BUG | ... | with counts on the next line) hides the number from that parse and forces a manual fallback, so keep the count on the RTL_BUG row itself:

| Classification | Count |
|----------------|-------|
| Total analyzed |    |
| RTL_BUG        |    |
| TC_BUG         |    |
| DOC_AMBIGUITY  |    |
| INCONCLUSIVE   |    |

Section: Root Cause Groups One sub-section per unique root cause location. Each sub-section lists:

• Root cause location (file, module, block)
• Classification
• All issue numbers that share this root cause, each linking to its debug_NNN_.md

Section: Unique Issues Issues that do not share a root cause with any other issue — listed individually with a one-line summary.

Closing note End with a blockquote: > This summary identifies root causes only. No code changes are proposed.

Step 8 - Self-check Before Finishing

Verify each report against its sources before declaring it done — it is consumed by whoever fixes the bug, who must trust it without re-tracing:

| Check | Pass condition | |-------|----------------| | Classification supported | The verdict follows from the Evidence and Signal Trace, not assertion; INCONCLUSIVE is used rather than a guessed class when the root is unresolved. | | Trace unbroken | The Signal Trace runs from the failing top-level output down to the named root with no skipped level; every row has a module and file. | | Root location is real | The cited file, module, and block actually exist — open them and confirm the named always/assign (or testcase line) is present, not assumed. | | Spec cited | RTL_BUG/DOC_AMBIGUITY quote the proposal or module-doc section that defines corre

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: phamcuong21478
• Source: phamcuong21478/rtl-skills
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.