Skip to content

CHANGELOG.md

Latest commit

 

History

History
53 lines (50 loc) · 3.44 KB

CHANGELOG.md

File metadata and controls

53 lines (50 loc) · 3.44 KB
Raw

Changelog

Format: Keep a Changelog. This project is pre-alpha; nothing is stable and there are no releases yet.

[Unreleased]

Added

  • Repository, toolchain, and design notes. No compiler yet — see doc/roadmap.md for what has to be true before one is worth writing.
  • bb check as the single gate, run identically by developers and CI.
  • A pinned toolchain: flake.nix for exactness, tools.json as the shared version SSOT for machines without Nix.
  • doc/design/0008 — parity is at the boundary: if a Clojure program cannot observe the difference, the internals are ours to choose.
  • doc/design/0009 — two modes, dynamic in development and static in production, designed together because the output-format constraints they force cannot be retrofitted. Rests on a measured fact, now pinned in the gate by test/cljwit/rec_group_identity_test.clj: independently compiled units share WasmGC types iff their rec groups canonicalise alike.
  • A rewritten S0 stop condition — dispatch overhead under 1 ns, absolute, per lane. Replaces "~10× JVM Clojure" (which passed B1 at 5.6×) and a ratio-to-the-engine's-floor draft (which rewarded a slow floor). Under it B1 passes on V8 and fails on wasmtime. doc/design/0003 gained the survey behind it: the engine ranking inverts with the workload.
  • /survey — the reconnaissance discipline: search, read the primary source, read what someone who solved it wrote down, then build the smallest thing that would fail if the claim were false. Run on stage entry. bb ref gained the reference sources it needs.
  • doc/design/0007 — the component boundary is linear memory, not GC. Found by the first survey; sizes S4 and adds an open question to the roadmap.
  • S0 B3 — boxed arithmetic on the i31 fast path: 0.31× JVM Clojure on both lanes, the first benchmark wasmtime wins. Unboxed we lose 2.1×, which is the profile doc/roadmap.md predicted, confirmed in both directions at once. With this the four contracted S0 benchmarks are all measured.
  • S0 B4 — what a ref.cast costs, on two axes. Depth is free; casting to a type that has subtypes is 30× a cast to a leaf, and input variety adds more. Falsifies 0004's "shallow and wide" guidance, which was backwards.
  • S0 B7 — the specialisation crossover: five rings varying only how often the guard hits. Specialising starts paying at ~26% on wasmtime and ~80% on V8, so there is no single threshold a compiler can use for both lanes.
  • S0 B5 — guarded call-site specialisation, the lever B1 identified. It erases the server lane's 6.16 ns of dispatch overhead down to 0.06, so both lanes pass the stop condition — but a mostly-missing guard costs more than no guard at all, which makes analysis precision the thing S0's answer rests on.
  • S0 B2 — the same dispatch with ten receiver types, plus a depth-matched one-type control so receiver count is the only variable. Confirms 0004's mechanism (wasmtime +9% under megamorphism against JVM Clojure's +114%) and not its conclusion (wasmtime is still 2.84× JVM). V8 degrades like the JVM.
  • S0 B1 — hand-written WasmGC protocol dispatch, measured on V8 and wasmtime against a JVM Clojure baseline, with three controls forming a curve in how far the call target sits from the receiver. bb bench-s0 runs it. First real numbers in the project: doc/design/0002-measure-first.md.