Skip to content

reference-cache.md

Latest commit

 

History

History
48 lines (39 loc) · 2.44 KB

reference-cache.md

File metadata and controls

48 lines (39 loc) · 2.44 KB
Raw

Bend reference cache

tests/pure.ts checks every pure test case three ways: the independent TypeScript oracle, the Bend interpreter (term_snf) on the checked book, and the generated C/SBF code. The interpreter is by far the slowest part of npm test; sqrt alone took 910 s cold.

tests/bend-reference-cache.ts stores each interpreter result under build/.bend-reference-cache/, keyed by the SHA-256 of everything that can affect it:

  • every checked Bend source file of the module (paths and bytes);
  • the pinned interpreter/checker .deps/bend/bend2/bend.ts and base.bend;
  • compiler/backend.ts (it loads and checks the book), tests/pure.ts (term conversion), sdk/wire.ts (result encoding) and the cache module itself;
  • the emitted schema;
  • the case's root name and arguments.

Changing any of these misses and re-evaluates. On a hit the cached result is still compared with the TypeScript oracle and the native/SBF results; only the interpreter run is skipped. Each module records its counts in <stem>-reference-cache.json, and the test log reports evaluated/cached cases.

BEND_REFERENCE_CACHE Behaviour
use (default) Reuse valid entries; evaluate and store misses
verify Evaluate every case and fail if a cached entry differs
refresh Evaluate every case and overwrite entries
off Evaluate every case; never read or write the cache

A cached result is evidence from the run that produced it, not a fresh interpreter execution. For acceptance records that claim fresh interpreter evidence, run with BEND_REFERENCE_CACHE=verify (or off). Deleting build/.bend-reference-cache/ is always safe. tests/bend-reference-cache-controls.ts (run by npm test) checks hits, case/schema/source misses, tamper detection in verify mode, corrupt-entry recovery and off mode.

Measured — 2026-10-09 (macOS arm64)

sqrt: 910 s cold, 1 s warm. Full npm test: 944 s while populating every other module (with sqrt already cached), 314 s warm; both passed (build/regression/20261009T230828Z-reference-cache/). The populating run evaluated 9,220 cases fresh; the 442 sqrt and iteration entries came from earlier standalone runs that day. With warm caches the remaining time is spread over adapter, differential and SBF sections (10–30 s each); further gains would need running modules in parallel.

Interpreter calls in tests/arithmetic.ts, tests/codecs.ts and tests/sha256-reference.ts are not cached.