peakrdl-check 0.1.0

Semantic compatibility analysis and configurable quality gates for SystemRDL specifications

PeakRDL-check

Semantic compatibility analysis and configurable quality gates for SystemRDL specifications. Detects firmware-breaking register changes, produces CI reports and provides scalable exploration of large register maps.

PeakRDL-check is local-first and complements — not replaces — your existing flow around systemrdl-compiler and PeakRDL: point it at the same .rdl sources, get an indexed, instantly searchable register map and a reviewer-grade semantic diff that classifies interface changes by impact.

Why PeakRDL-check

• It reviews meaning, not text. A textual diff can't tell a 200-line harmless refactor from a one-line change that silently moves hundreds of registers. PeakRDL-check compares the elaborated register models and classifies every change by impact — breaking, behavioural, compatible, documentation, or uncertain — with a stable rule ID, an explanation, and before/after values. When a match is ambiguous it says so instead of guessing.
• Quality gates you control. --fail-on picks the severity that fails CI, and a policy JSON reclassifies any rule to match your team's rules (e.g. treat reset-value changes as release-blocking). Reports come as text, JSON, Markdown or SARIF, and a reusable GitHub Action posts job summaries and inline annotations.
• It stays fast when register maps get huge. One SQLite index file instead of tens of thousands of HTML files, a virtualized viewer that only loads what you look at, second-scale incremental rebuilds after local edits, and millisecond queries — measured at 800,000 registers.
• Every number is a measurement, not a slogan. The benchmark summary below and the full evidence dossier in PROOF.md are generated from raw, preserved run records (benchmarks/raw-results/, failures and all) — nothing is typed by hand, and one script reproduces the lot.

Measured performance

On an 800,000-register specification (Apple M4 Pro, medians of 3 runs each, identical fixtures — full tables and environment in PROOF.md):

Operation	PeakRDL-check	PeakRDL-html 2.12.2
Cold build (raw source → browsable)	285 s	1390 s (4.9× slower)
Output	1 file, 338 MB	85,216 files, 1098 MB
Index generation from an already-elaborated model	10 s	n/a (regenerates everything)
Incremental rebuild after a one-line block edit	33 s	n/a (full rebuild)
Open existing index (server ready / first page)	105 ms / 107 ms	static files, but 51.8 MB transferred up front
Search p95 / exact-lookup p95	10.4 ms / 0.6 ms	in-browser over the fully-loaded model

Semantic diff: 47/47 curated scenarios pass (12/12 breaking detected, 10/10 formatting-only suppressed); 240 generated mutation trials at recall 1.0, precision 1.0.

What a semantic diff looks like

$ peakrdl-check diff --base main/uart.rdl --head pr/uart.rdl

Semantic diff: 2 breaking, 1 documentation  (policy 1.0.0)

✖ [BREAKING     ] REG-ADDRESS-CHANGED        uart.status
    reg 'uart.status' address changed from 0x4 to 0x40.
    before: 0x4    after: 0x40
    at pr/uart.rdl:9
✖ [BREAKING     ] ENUM-VALUE-CHANGED         uart.ctrl.baud
    Enum member 'B115200' in field 'baud' changed value from 0x2 to 0x4;
    existing encodings break.
✎ [DOCUMENTATION] DESC-CHANGED               uart.ctrl
    Description wording changed on 'uart.ctrl'.

The corpus in diff-corpus/ demonstrates the other direction too: a 203-line textual refactor (file splits, reordering, hex renumbering, typedef extraction) that produces zero semantic changes, and a one-line parameter edit that changes hundreds of elaborated registers. Ambiguous renames are reported as MATCH-UNCERTAIN — the tool never silently guesses.

Five-minute quick start

python -m pip install peakrdl-check

# build an index and browse it
peakrdl-check build your_design.rdl --output build/design
peakrdl-check serve build/design         # opens http://127.0.0.1:<port>

# review a change
peakrdl-check diff --base main/design.rdl --head pr/design.rdl \
    --format markdown --output diff.md

# CI gate (exit 1 on breaking changes)
peakrdl-check check --base main/design.rdl --head pr/design.rdl \
    --fail-on breaking

# optionally install the PeakRDL host and use the registered subcommand
python -m pip install "peakrdl-check[peakrdl]"
peakrdl check pr/design.rdl --base main/design.rdl --fail-on breaking

For an editable source checkout, benchmark tooling, and the test suite, see CONTRIBUTING.md.

A reusable GitHub Action lives in action/; an example workflow is in examples/.github/workflows/register-review.yml.

Supported input

• SystemRDL 2.0 via systemrdl-compiler 1.32.2 (parameters, arrays, dynamic assignments, aliases, `include / `define preprocessing).
• Not supported (yet): IP-XACT, register editing, language-server features. PeakRDL-check is a review surface, not an editor (see ADR-0009).

Architecture in one paragraph

systemrdl-compiler parses and elaborates; a thin adapter walks the elaborated tree once (arrays kept folded) and emits a canonical model with content-hash-deduplicated definitions and lightweight instances; that streams into a single SQLite file with FTS5 search. A localhost stdlib server exposes a paginated JSON API; the viewer is one static HTML+JS file with virtualized rendering, so browser state stays proportional to what you actually look at. The semantic diff engine compares two canonical models through separated stages (matching → detection → versioned severity policy → explanation → text/JSON/Markdown/SARIF). Incremental rebuilds re-elaborate only block types from changed files and splice their subtrees into the existing index — proven byte-equivalent to a clean rebuild. Details: docs/architecture.md.

Known limitations

• Incremental rebuilds operate at defining-file granularity.
• The register-map viewer requires a local server; static browser-only mode is not currently supported.
• Input is limited to SystemRDL 2.0. IP-XACT and register editing are outside the current scope.

See the complete limitations list.

Documentation

Architecture · Diff rules · Benchmarks and evidence · ADRs · Contributing · Security