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

VM: interleave captured runtime environments with direct-eval activation views #785

Closed

Description

@chrisbbreuer

Parent: #771
Related: #776

Problem

A VM closure created inside a live block, class, loop-head, disposal, or with Environment Record captures that runtime record through Function.closure while its statically resolved outer function bindings remain in Frame slots. Direct eval needs the exact lexical order current frame -> captured runtime records -> parent frame. The current DirectEvalPlan rejects any nonzero parent_environment_depth because materialization can only place every activation view on one side of Function.closure.

This is the final explicit runtime-Environment boundary in #771. Admission must stay fail-closed until the engine can represent the interleaving without changing published Environment.parent links or copying binding values.

Scope

  • Retain each ordinary Frame defining closure Environment as a precise, relocating edge so every child-to-parent frame boundary can recover its captured runtime segment.
  • Extend immutable direct-eval plans with validated boundary metadata for the exact number and ordering of intervening Environment Records, including named-function-expression records.
  • Materialize identity-preserving forwarding records whose own-binding operations target the original declarative or object Environment Record while their parent links interleave activation views in spec order.
  • Preserve TDZ, const/function-name state, deletability, aliases, with HasBinding and Symbol.unscopables observability, catch Annex B rules, per-iteration binding identity, and closures produced by eval.
  • Keep publication transactional under deterministic OOM, exact under moving GC, and race-free for escaped frames and shared no-GIL execution.

No-workaround rules

  • No binding-value snapshots, source rewriting, published-parent mutation, global fallback, tree-walker callback, broad admission, or eager materialization for calls that do not resolve to the realm eval intrinsic.
  • No stacks, queues, jobs, Promise/microtask, Worker lifecycle, or Map/Set iterator implementation changes.

Acceptance

  • Required bytecode and forced tree-walker agree for direct eval inside nested block/class/catch/loop and with scopes, including shadowing, TDZ, mutation, deletion, named function-expression self binding, and eval-created escaping closures.
  • Runtime records remain exact live identities: per-iteration captures stay distinct, with traps and Symbol.unscopables run in the specified order, and later initialization or mutation is visible through both original and interleaved chains.
  • Frame closure edges and forwarding-record target edges trace and relocate precisely; deterministic allocation-failure sweeps publish no partial frame cache or vm.env chain.
  • Shared-realm no-GIL and TSan witnesses show no parent-edge or binding-table race, including concurrent first materialization and escaped defining frames.
  • The parent_environment_depth and current environment_depth barriers are removed only for fully validated plans; malformed or unsupported boundaries fail closed without fallback.
  • Focused compiler/VM/frontend, relevant exact Test262 eval/function/with/class accounting, full units, threadfuzz, debugger, watchdog, moving-GC, TSan, and no-GIL gates pass with zero regressions.
  • A conventional commit lands directly on main and cross-links VM: expose ordinary activation bindings to direct eval #771 and VM: model direct eval across parameter and body environments #776; exact Test262 flips are recorded without an unmeasured performance claim.

Completed integration

CI run 33113309062 completed successfully on main commit 8a1398ffc04e9e30987f968693667c7271ca3ce1, a descendant containing this implementation. All 53 executed jobs succeeded (unit and TSan-unit shards, Linux no-GIL TSan, functional/no-GIL corpus gates, fuzz profiles, ABI, Wasm, Test262 parallel, and docs); the scheduled-only nightly job was skipped. Together with the measured local evidence in the comments, this completes this issue's acceptance.

The independently scoped provisional root-registration/preparation defect remains open in #787. This closure covers registered caller-root relocation and exact eval environment identity, not that outstanding activation-driver ownership work. The exposed capture-metadata race was fixed and verified in #789.

Activity

  1. chrisbbreuer commented on Aug 27, 2026

    @chrisbbreuer
    MemberAuthor

    The precise Frame anchor layer landed in d2b4c494. Every running ordinary activation now retains its exact defining Environment as an immutable Frame edge; captured-function, suspended-generator, and active-interpreter roots trace and relocate it, and recyclable activations clear it before reuse. The relocation witnesses cover all three root categories. Formatter and AST checks pass; compiler admission and evaluation semantics remain unchanged, so this layer flips 0 Test262 cases. The issue remains open for boundary-plan metadata and identity-preserving interleaving.

  2. chrisbbreuer commented on Aug 27, 2026

    @chrisbbreuer
    MemberAuthor

    Immutable boundary metadata landed in 62f9385b. DirectEvalPlan now records the exact runtime Environment count at each child-to-parent Frame boundary, including current call-site records and named-function-expression self-name records. The existing causal barrier remains: any nonzero boundary still rejects compilation until materialization consumes it exactly. Focused planner tests pass 2/2; adjacent Frame relocation categories pass 3/3, with zero failures/leaks. This layer flips 0 Test262 cases because the newly represented cases remain fail-closed.

  3. chrisbbreuer commented on Aug 27, 2026

    @chrisbbreuer
    MemberAuthor

    The identity-preserving forwarding layer landed in 2b46afb0. A forwarding Environment owns only its replacement parent edge; all declarative/object own-binding queries, reads, writes, tags, aliases, References, with operations, conflict checks, and deletion target the original live record. The target edge traces/relocates precisely, reuse clears it, and quick binding caches reject the synthetic chain in favor of the exact ordinary VM path. Environment clusters pass 13/13 normally and 13/13 under TSan; focused forwarding witnesses pass 2/2 in both modes, with zero failures/leaks/sanitizer reports. Admission remains fail-closed, so this layer flips 0 Test262 cases. Next is transactional materializer integration.

  4. chrisbbreuer commented on Aug 27, 2026

    @chrisbbreuer
    MemberAuthor

    Captured-boundary admission landed in 6e212e82.

    The materializer validates each exact Frame closure segment, then places identity-preserving forwarding records at each lexical scope's runtime depth. This preserves block/catch shadowing inside nested with records instead of flattening all frame scopes to one side. Environment-backed cells are excluded from slot views; original captured parent links are never changed. Target pointers are reloaded from precise roots after allocation, and the complete candidate chain is allocated before cache publication.

    Measured coverage:

    • Required bytecode versus forced tree-walker: captured with and loop cells, named-function self binding, block/catch shadowing, repeated eval, two captured Frame boundaries, and eval-created escaping closures through compaction.
    • Synthetic cached-parent/current reuse, malformed boundary/depth rejection, and deterministic OOM sweeps leave no partial caches or changed original parents.
    • Four no-GIL workers materialize against a shared escaped defining frame.
    • Direct-eval cluster: 32/32 normal and 32/32 TSan; final expanded interleaving witnesses: 3/3 TSan. The rebased landed tree also passes 32/32 normal. Zero failures/leaks/sanitizer reports in these runs; formatter, AST, and diff checks pass.

    Exact ReleaseFast Test262 accounting, baseline 2b46afb -> landed 6e212e8:

    • eval-code: 347 -> 347 pass
    • function-code: 217 -> 217
    • expressions/function: 264 -> 264
    • statements/function: 451 -> 451
    • statements/with: 181 -> 181
    • expressions/class: 4059 -> 4059
    • statements/class: 4367 -> 4367

    Total: 9,886 -> 9,886 passing, 0 failures in either run, 0 flips and 0 regressions across those exact subtrees. No performance claim.

    This remains open: current-frame runtime eval still fails closed at environment_depth, because returning from eval must preserve the VM's active runtime chain for block/with/loop unwind. Full required gates and the rest of the acceptance matrix are not yet claimed. Continues #771 and #776.

  5. chrisbbreuer commented on Aug 27, 2026

    @chrisbbreuer
    MemberAuthor

    The current-frame implementation now passes required-bytecode/tree-walker scope, abrupt-with, and no-GIL witnesses. Exact ReleaseFast accounting remains 9,886/9,886 across the same seven eval/function/with/class subtrees, with 0 flips. A new test that requests a moving nursery safepoint from inside eval exposed #786: nested VM calls restore an unrooted saved caller Environment. That root-ownership defect is being fixed before current-frame admission lands; no compaction or execution-mode workaround is being used.

  6. chrisbbreuer commented on Aug 27, 2026

    @chrisbbreuer
    MemberAuthor

    Landed fix(vm): preserve runtime eval scopes and caller roots (#785, #786) on main.

    Current-frame runtime records now interleave with activation views in exact lexical order. Semantic eval and runtime-unwind projections forward to original binding identities; handler prefixes preserve eval-created variables across abrupt completion. Direct-eval-only loop captures, simple-catch Annex B rules, class TDZ, with/unscopables, nested captured/current boundaries, labeled jumps, and generator suspension have forced-tree-walker/required-bytecode witnesses. Publication/OOM and actual moving-safepoint witnesses pass without fallback.

    Final-source validation: full units 1988 passed, 1 platform skip, 0 failures/leaks; focused TSan direct eval 36/36 and root relocation 1/1 with zero reports; threadfuzz default 400 and exact-result verify 300 seeded programs passed. Focused eval/compiler/with/debugger/activation/termination and shell-timeout probes passed.

    Exact ReleaseFast parent 6e212e8 -> candidate Test262: eval-code 347 -> 347, function-code 217 -> 217, expressions/function 264 -> 264, statements/function 451 -> 451, statements/with 181 -> 181, expressions/class 4059 -> 4059, statements/class 4367 -> 4367. Total 9886 -> 9886 passed, zero failures, flips, or regressions in these subtrees. No performance claim.

    The nested moving witness exposed #786, fixed in the same commit. Pushed required integration remains pending; the separate provisional-root publication/lifecycle audit remains #787. Keeping this issue open while those boundaries are tracked.

  7. chrisbbreuer commented on Aug 27, 2026

    @chrisbbreuer
    MemberAuthor

    Pushed Linux integration found an engine-state race in the new current-runtime direct-eval witness: tsan-parallel-js job. Environment.markCaptured writes private_activation while a peer lockBindingsForRead reads it. The job's functional tests are 60 pass / 8 skip / 0 fail / 0 leak, but one TSan warning causes exit 66.

    Tracked as #789. Local focused TSan did not expose this interleaving; this issue's integration gate remains open. The fix will atomically publish/read capture and private-ownership state, preserving the existing lock-elision proof and captured-frame identity rather than suppressing the report.

  8. chrisbbreuer commented on Aug 27, 2026

    @chrisbbreuer
    MemberAuthor

    Landed fix(threads): publish environment capture state atomically (#789) on main. This addresses the Linux TSan read/write race in the direct-eval forwarding/capture path: both capture/private flags now have atomic publication/read semantics, and escaped-frame marking precedes captured publication. No suppression or disabled reuse was introduced.

    Local evidence: Environment 14/14 and direct eval 36/36 in both normal and TSan builds (zero reports); full normal units 1,993 pass / 1 skip / 0 fail / 0 leak; 2,711 Test262 passes unchanged across eight relevant subtrees; seed-1 default/verify threadfuzz 400/300 programs with zero failures. #789 retains the outstanding pushed Linux required-integration gate; these local results do not close the parent integration requirement.

  9. chrisbbreuer commented on Aug 27, 2026

    @chrisbbreuer
    MemberAuthor

    Linux confirmation for #789: tsan-parallel-js on cfef054f passed, including the original current-runtime direct-eval regression: 60 pass / 8 platform skips / 0 fail / 0 leak and no TSan report. The broader workflow is still running, so this confirms the specific race gate rather than full integration.

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