0%

Section 09: Layout Caching via Salsa Query

Status: Not Started Goal: Move type layout computation from an imperative pass with manual caching to a Salsa-tracked query that any phase can call. This is the aspirational pattern from impl-hygiene.md — “A Salsa-tracked layout_of(TypeId) -> Layout query that any phase can call.”

Success Criteria:

  • layout_of() Salsa query exists — satisfies mission criterion “all consumers call layout_of()”
  • No manual RefCell<FxHashMap> cache remains — satisfies mission criterion “Salsa memoization”
  • All codegen tests pass with Salsa-cached layout — behavior identical to imperative pass
  • impl-hygiene.md updated — aspirational pattern marked as implemented

Context: Research found a two-layer layout system:

  1. ori_repr (compiler/ori_repr/src/layout/, 285+ lines): Pure Rust layout computation (repr_size(), repr_align(), compute_field_layout(), reorder_fields()) operating on MachineRepr enum. Produces ReprPlan. NOT a Salsa query — computed imperatively once per compilation.
  2. ori_llvm TypeLayoutResolver (compiler/ori_llvm/src/codegen/type_info/layout_resolver.rs, 394 lines): Idx → BasicTypeEnum with RefCell<FxHashMap<Idx, BasicTypeEnum>> cache and cycle detection via FxHashSet<Idx> + MAX_RESOLVE_DEPTH = 32.

The gap: layout is computed imperatively and cached manually. Moving to Salsa gives:

  • Automatic invalidation when types change (incremental correctness)
  • Cross-phase availability (ARC can query layout for alignment-aware optimization)
  • Determinism guarantee from Salsa’s purity requirement
  • Cache eviction handled by Salsa revision tracking

Design challenge: ReprPlan currently carries Idx references, not LLVM handles (repr.md §RP-4). The Salsa query should return a ReprPlan-like type that derives all Salsa-required traits (Clone, Eq, PartialEq, Hash, Debug). ReprPlan already derives these per repr.md §RP-4.

Reference implementations:

  • Rust TyCtxt::layout_of(): memoized layout query on the type context. Lazily computed on first access, cached, never recomputed.

Depends on: Section 01 (policy language). Benefits from Section 07 (TypeFolder patterns for type traversal in layout computation).

09.1 Design Layout Query Interface

File(s): design document (not code yet)

  • Design the Salsa query interface:

    • layout_of(db: &dyn CompilerDb, idx: Idx) -> ReprPlan — tracked query
    • Input: fully-resolved Idx (per CG:TR-2 — no Tag::Var in codegen input)
    • Output: ReprPlan (already Clone, Eq, Hash, Debug per repr.md §RP-4)
    • The query lives in a shared location accessible to both ori_arc (for alignment-aware optimization) and ori_llvm (for emission)

  • Decide the hosting crate:

    • Option A: ori_types — closest to the pool, but adds layout dependency to the type crate
    • Option B: ori_repr — already owns layout computation, just needs Salsa wiring
    • Option C: new ori_layout crate — cleanest separation but adds a crate
    • Recommendation: Option B (ori_repr) — it already owns the computation; adding Salsa tracking is the minimal change

  • Design the migration path:

    • Phase 1: Add layout_of() query to ori_repr alongside existing imperative computation
    • Phase 2: Migrate TypeLayoutResolver in ori_llvm to call layout_of() instead of maintaining its own cache
    • Phase 3: Remove the RefCell<FxHashMap> manual cache

  • Validate design against /tp-help

  • Subsection close-out (09.1) — MANDATORY before starting 09.2:

    • All tasks above are [x]
    • Update this subsection’s status in section frontmatter to complete
    • Run /improve-tooling retrospectively on THIS subsection
    • Run /sync-claude to check if code changes affect CLAUDE.md or rules files

09.2 Implement Salsa-Tracked Layout Query

File(s): compiler/ori_repr/src/, compiler/oric/src/query/

  • Write failing tests:

    • Query layout_of() for primitive types — verify sizes match CG:TR-1
    • Query for struct types — verify field ordering matches RP-20
    • Query twice for same Idx — verify same result (Salsa caching)
    • Query after type pool mutation — verify invalidation (new Salsa revision)

  • Add Salsa tracked query:

    • Wire layout_of as a #[salsa::tracked] function in the compiler’s query group
    • Implementation calls existing ori_repr layout computation functions
    • Result is ReprPlan (already Salsa-compatible: derives Clone, Eq, Hash, Debug; no Arc, no fn pointers, no LLVM handles per repr.md §RP-4)
    • Cycle detection: if layout_of(A) recursively calls layout_of(A) (recursive type), produce a clear error — Salsa’s cycle detection should handle this but verify

  • Verify determinism: same input pool → same ReprPlan output across multiple calls and revisions

  • Subsection close-out (09.2) — MANDATORY before starting 09.3:

    • All tasks above are [x]
    • Update this subsection’s status in section frontmatter to complete
    • Run /improve-tooling retrospectively on THIS subsection
    • Run /sync-claude to check if code changes affect CLAUDE.md or rules files

09.3 Migrate Codegen Consumers

File(s): compiler/ori_llvm/src/codegen/type_info/layout_resolver.rs, compiler/ori_llvm/src/codegen/

  • Replace TypeLayoutResolver’s manual RefCell<FxHashMap<Idx, BasicTypeEnum>> cache:

    • Instead of caching layout in a local HashMap, call layout_of(db, idx) to get the ReprPlan
    • The LLVM type construction (Idx → BasicTypeEnum) still happens in ori_llvm — but the layout DATA comes from the Salsa query
    • The TypeLayoutResolver becomes a thin wrapper: layout_of() → ReprPlan → construct LLVM type

  • Migrate all codegen sites listed in repr.md §RP-5 Consumers:

    • CG:TR-1 canonical type mapping: call layout_of() for layout data
    • CG:TR-3 aggregate field ordering: use FieldEntry from layout_of()
    • CG:AB-1 ABI threshold: use abi_size from layout_of()
    • CG:NR-2/NR-6 narrowing: use storage_width from layout_of()

  • Remove the cycle detection FxHashSet<Idx> from TypeLayoutResolver — Salsa’s cycle detection handles this

  • Verify: timeout 150 ./test-all.sh — all codegen tests, FileCheck tests, AOT tests pass

  • Subsection close-out (09.3) — MANDATORY before starting 09.4:

    • All tasks above are [x]
    • Update this subsection’s status in section frontmatter to complete
    • Run /improve-tooling retrospectively on THIS subsection
    • Run /sync-claude to check if code changes affect CLAUDE.md or rules files

09.4 Remove Manual Cache & Documentation

File(s): compiler/ori_llvm/src/codegen/type_info/layout_resolver.rs, .claude/rules/impl-hygiene.md, .claude/rules/repr.md

  • Remove the RefCell<FxHashMap> manual cache from TypeLayoutResolver — all layout data now comes from Salsa

  • Update impl-hygiene.md §Aspirational Patterns → Layout Caching: change from “aspirational” to “implemented”

  • Update repr.md §RP-4:

    • Change “NOT a Salsa tracked struct. Computed imperatively” to “Salsa-tracked query via layout_of(). Computed on demand and memoized.”
    • Update the consumers table to reference the Salsa query

  • Update canon.md §6 SSOTs table — add layout_of() as the canonical home for layout computation

  • Subsection close-out (09.4) — MANDATORY before completing section:

    • All tasks above are [x]
    • Update this subsection’s status in section frontmatter to complete
    • Run /improve-tooling retrospectively on THIS subsection
    • Run /sync-claude to check if code changes affect CLAUDE.md or rules files

09.R Third Party Review Findings

  • None.

09.N Completion Checklist

  • All subsections (09.1-09.4) complete
  • layout_of() Salsa query exists and is the sole layout entry point
  • Manual RefCell<FxHashMap> cache removed
  • All codegen tests, FileCheck tests, AOT tests pass
  • Layout is deterministic across Salsa revisions (tested)
  • Debug AND release builds pass
  • timeout 150 ./test-all.sh passes
  • /tpr-review — independent dual-source review clean
  • /impl-hygiene-review — implementation hygiene clean
  • /improve-tooling — section-close sweep
  • /sync-claude — section-close doc sync