prometeu-studio/docs/pbs/specs/13. Lowering IRBackend Specification.md

8.0 KiB

PBS Lowering IRBackend Specification

Status: Draft v1 (Frontend Scope + Backend Handoff Addendum)
Applies to: first lowering boundary from bound PBS frontend model into IRBackend/IRBackendFile, plus executable-backend handoff obligations at this boundary

1. Purpose

This document defines the normative frontend lowering contract from PBS source semantics to IRBackend.

Its purpose is to keep the first lowering deterministic and shared across implementations while PBS and backend design evolve.

2. Scope

This document defines:

  • required preconditions before lowering starts,
  • AST invariants that lowering may assume as hard contract,
  • semantic obligations preserved in IRBackend,
  • deterministic rejection behavior for unsupported frontend-lowering forms,
  • diagnostics attribution obligations for lowering failures in frontend scope,
  • and executable-backend handoff obligations at the IRBackend boundary.

This document does not define:

  • VM lowering (IRVM),
  • bytecode/PBX mapping,
  • runtime execution behavior,
  • verifier/loader internals.

Those concerns belong to shared acceptance specs under docs/general/specs.

3. Authority and Precedence

Normative precedence:

  1. 3. Core Syntax Specification.md
  2. 4. Static Semantics Specification.md
  3. 11. AST Specification.md
  4. 12. Diagnostics Specification.md
  5. This document

If a lowering rule here conflicts with higher-precedence syntax/semantics rules, the higher-precedence rule wins.

4. Normative Inputs

This document depends on:

  • 3. Core Syntax Specification.md
  • 4. Static Semantics Specification.md
  • 11. AST Specification.md
  • 12. Diagnostics Specification.md
  • docs/general/specs/20. IRBackend to IRVM Lowering Specification.md

5. Lowering Preconditions

Frontend lowering into IRBackend may start only when:

  1. source has been tokenized and parsed,
  2. required syntax diagnostics have been emitted,
  3. AST output satisfies 11 required invariants for this source unit, including:
    • one root per file,
    • deterministic child ordering,
    • mandatory node attribution (file/start/end) on nodes consumed by lowering,
    • mandatory declaration metadata on declaration nodes required by the active source slice,
    • and structural coherence after any parser recovery.
  4. linking/name-resolution outcomes required for the current lowering slice are available,
  5. and no unresolved semantic ambiguity is deferred to lowering.

Lowering must not invent unresolved semantic answers that belong to syntax/static/linking phases.

6. IRBackend Preserved Obligations

For each admitted source unit and callable in the current lowering slice, IRBackend must preserve at minimum:

  1. callable identity (name/category as applicable),
  2. callable arity,
  3. declared return surface information,
  4. source attribution anchor (file + span) for diagnostics and traceability,
  5. source-observable parse intent for statement/expression structure (including precedence/associativity outcome already fixed by AST shape).
  6. deterministic requiredCapabilities derived from admitted host-binding metadata for packer/runtime-manifest assistance.
  7. frontend-declared executable entrypoint callable name from FrontendSpec (entryPointCallableName) for backend handoff.

Lowering must not collapse source categories in a way that erases required declaration/callable identity needed by downstream diagnostics or conformance assertions.

The normative contract is obligation-based, not tied to one mandatory in-memory class graph.

7. Deterministic Rejection Policy

If a source form is outside current frontend-lowering support:

  1. rejection must be deterministic,
  2. diagnostics must expose stable code, severity, phase, message template id, and primary attribution,
  3. and unsupported forms must not silently degrade into different valid lowered behavior.

Lowering must not convert a required syntax/static rejection into accepted lowered behavior.

For multi-module builds, lowering admission must apply dependency-scoped fail-fast:

  1. a module rejected in syntax, static semantics, linking, host-admission, or load-facing gates must not be emitted;
  2. any module that imports a rejected module (directly or transitively) must also be excluded from IRBackend emission;
  3. modules independent from the rejected dependency subgraph may continue to lower and emit in the same build.

8. Conformance Boundary

IRBackend is the first lowering boundary (frontend responsibility).

Conformance-valid claims at this boundary require Gate U evidence from docs/general/specs/13. Conformance Test Specification.md.

For this frontier, Gate U evidence is expected to cover at minimum:

  • lexer behavior for covered inputs,
  • parser behavior and AST shape invariants required by 11,
  • lowering outcomes at IRBackend boundary,
  • and deterministic diagnostics identity/attribution for rejection cases.

Conformance at this boundary may be claimed only when the full claimed PBS source surface is implemented at IRBackend level.

9. Explicit Deferrals

The following are intentionally out of scope for this document:

  • IRBackend -> IRVM lowering,
  • artifact encoding and binary layout,
  • runtime lifecycle behavior,
  • verifier and loader checks.

10. Non-Goals

  • Freezing one optimizer strategy.
  • Freezing one compiler architecture.
  • Repeating runtime or bytecode authority.

11. Exit Criteria

This document is healthy when:

  1. lowering preconditions are explicit,
  2. preserved IRBackend obligations are explicit,
  3. deterministic rejection policy is explicit and test-backed,
  4. and scope boundaries with general/backend acceptance specs are explicit.

12. Executable Backend Handoff Addendum (v1)

For executable backends, the IRBackend output from this frontend boundary must satisfy the additional handoff obligations below.

12.1 Callable obligations

For each executable callable admitted into IRBackend, the frontend must preserve:

  1. stable callable identity in the current build graph,
  2. observable callable signature (input arity and return shape),
  3. callable category needed by downstream lowering,
  4. source attribution anchor (fileId, start, end),
  5. and executable body representation that is backend-lowerable.

12.2 Callsite classification obligations

Each executable callsite in IRBackend must be classified into exactly one category:

  1. CALL_FUNC,
  2. CALL_HOST,
  3. CALL_INTRINSIC.

Backend lowering must not infer this category by textual heuristics.

12.3 Host-backed metadata obligations

Each host-backed callsite admitted at this boundary must preserve canonical host identity:

  1. module,
  2. name,
  3. version.

When available at this boundary, declared host ABI shape (arg_slots, ret_slots) must also be preserved for downstream validation.

12.4 VM-owned metadata obligations

VM-owned intrinsic callsites admitted at this boundary must preserve canonical intrinsic identity:

  1. canonicalName,
  2. canonicalVersion.

VM-owned builtin projections/constants/intrinsics must not be reclassified as host bindings.

12.5 Deterministic capability obligations

requiredCapabilities derived at this boundary must be deterministic for the same admitted input graph.

12.6 Boundary with backend lowering specs

This addendum defines obligations preserved at the IRBackend boundary only.

It does not replace:

  1. docs/general/specs/20. IRBackend to IRVM Lowering Specification.md,
  2. docs/general/specs/21. IRVM Optimization Pipeline Specification.md,
  3. or docs/general/specs/15. Bytecode and PBX Mapping Specification.md.

12.7 FrontendSpec Entry Point Obligation

For executable language frontends, FrontendSpec must declare one canonical entrypoint callable name (entryPointCallableName).

At IRBackend emission time:

  1. frontend must copy this declaration into IRBackend.entryPointCallableName,
  2. value must be non-blank and deterministic for the same language/spec version,
  3. and backend stages must treat this field as authoritative for entrypoint selection instead of hardcoded language-agnostic names.