镜像站点 · 本页由第三方 GitHub 只读镜像提供,非 GitHub 官方站点,不接受任何登录或凭据输入。前往 github.com
Skip to content

Perf: add a same-heap managed-cell barrier fast path #9

Closed as not planned

Description

@chrisbbreuer

Downstream: zig-utils/zig-js#62
Profile precursor: zig-utils/zig-js#67

Problem

The broad write barrier deliberately accepts wild, stale, arena-owned, and foreign pointers, so it validates each child and owner through the embedder's ownership classifier before reading a header. In zig-js's hot Object-to-Object store path, both pointers are known GC cells and the post-#67 object_churn profile attributes 20 top-of-stack samples to headerForPayload. Repeating slab membership classification at every dense replacement is avoidable.

Scope

  • retain the current broad barrier unchanged for maybe-managed inputs;
  • give each Heap instance a stable per-heap header cookie without increasing Header size;
  • add an explicitly trusted managed-cell barrier whose pointer contract permits direct header recovery and rejects cells owned by a different heap of the same type through that cookie;
  • preserve nursery old-to-young owner remembering, conservative child retention when an owner is foreign, incremental/concurrent marking handoff, and parallel locking;
  • expose the fast path to zig-js only at Object-to-Object stores whose allocation provenance is managed whenever a barrier is installed;
  • keep weak barriers and mixed Environment/Promise/side-record callers on the broad path until independently proven;
  • measure clean pinned downstream object_churn A/B with exact checksum equality.

Acceptance criteria

  • Broad barriers still reject wild, stale, arena, and foreign inputs safely.
  • Managed barriers avoid payload-index/slab ownership classification.
  • Cross-heap cells with the same Binding/Header layout are rejected by a per-heap cookie.
  • Header size and existing cell size-class selection do not grow.
  • Nursery old-to-young edges and incremental/concurrent marking remain correct.
  • Focused zig-gc and downstream GC/no-GIL tests pass.
  • Clean pinned downstream A/B improves at checksum 1030552128.
  • Commits reference this issue and evidence is added to Bench: add bounded object-allocation churn vs JSC zig-js#62.

Activity

  1. chrisbbreuer commented on Jul 14, 2026

    @chrisbbreuer
    MemberAuthor

    Rejected after the required downstream A/B; both zig-gc and zig-js experiments were fully removed and no commit was made.

    The candidate used a per-heap header cookie without increasing Header size, added a direct managed-cell barrier, preserved broad wild-pointer behavior, and passed:

    • managed same-heap/cross-heap nursery test;
    • broad non-heap pointer barrier test;
    • downstream nursery integration 7/7.

    Pinned downstream object_churn, 12 order-balanced fresh warmed processes per binary:

    • baseline median: 617.213 ms
    • managed barrier median: 625.312 ms
    • regression: 1.31%
    • checksum: 1030552128 in every run.

    The existing slab classifier's hot hint is cheaper than the additional barrier-mode branch and code-layout cost on this path. The broad barrier remains the better implementation; downstream work moves to repeated unsupported quickening probes in the VM dispatch loop.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions