Model Elaboration

rb elab is a fast SystemVerilog parse, type-check, and elaboration gate that builds no simulator executable. It reads the same model and filelist as every other model-based flow. There is no elab.yaml and no second model path to keep in sync.

Run a model

Install the optional Python frontend, list the models, then run one:

uv add "rtl_buddy[elab]"
rb --machine elab --list -c design/models.yaml
rb --machine elab core -c design/models.yaml

A bare run uses the model's filelist and elaborates model.top, or the model name if top is unset. Add a named profile only when a gate needs different sources, defines, parameters, compatibility options, resources, or top:

rtl-buddy-filetype: model_config
models:
  - name: core
    top: core_top
    filelist: [-F core.f]
    elaborations:
      - name: smoke
        append_sources: [checks/bind_checks.sv]
        defines: {CHECKS_ENABLED: 1}
        parameters: {DATA_WIDTH: 32}
        warnings: [all]
        reglvl: 0
        resources: {cpus: 2, mem: 2G, time: "00:10:00"}

rb --machine elab core --profile smoke -c design/models.yaml

• A profile top overrides the model top, which overrides the model name.
• Profile source and include paths resolve from models.yaml.
• warnings entries are the text after -W. They can suppress warnings but not hard parse, type, or elaboration errors.

Raise the parser nesting limit

If a run fails with language constructs are too deeply nested, set max_parse_depth on the profile. slang stops at 1024 nesting levels, and generated RTL can exceed that inside one expression, typically a long chain of conditional expressions or concatenations. The setting is unset by default and affects only the profile that declares it:

    elaborations:
      - name: parse
        max_parse_depth: 8192

Malformed sources still fail with the same diagnostics. Set it to what the generated source needs, not to the maximum: beyond the guard, later passes can exhaust the C stack and kill the worker with no diagnostic. The accepted range is in YAML Formats, and the platform limit on nesting depth is in Quirks & Known Issues.

Run a regression

An elaboration regression takes a small manifest that lists models.yaml files. It runs only named profiles, so adding an ordinary model does not widen a project-wide gate.

rtl-buddy-filetype: elab_reg_config
model-configs:
  - design/core/models.yaml
  - design/peripherals/models.yaml

rb --machine elab-regression -c elab_regression.yaml --reg-level 1

• Profiles above the requested level report SKIP.
• A manifest with no named profiles is an error, not an empty pass.
• If the manifest is not at the project root, set cfg-rtl-reg.elab-reg-cfg-path.

Outputs

Each run writes below the directory containing its models.yaml:

artefacts/elab/<model>/<base-or-profile>/
  elab.f
  elab.log
  result.json

• elab.f is the filelist with includes unrolled and path entries made absolute.
• result.json records the selected top, explicit and parsed source counts, error and warning counts, max_parse_depth (null when unset), elapsed time, peak worker memory, and the pyslang version.
• A missing prepend_sources, append_sources, or include_dirs entry gives a FAIL at stage filelist for that profile. The command continues with the remaining profiles.
• Machine mode returns the same result payload and writes JSONL events to rtl_buddy.log.

Dispatch elaboration

rb elab dispatches only with --dispatch. rb elab-regression also honors cfg-dispatch.backend. A profile's resources layer over cfg-dispatch.resources, and cpus is both the scheduler request and the pyslang worker thread count.

• Slurm enforces the memory and time reservations.
• Local-parallel passes cpus to pyslang and limits concurrency with its process pool. It does not enforce memory or time.