You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
JIT: specialize inherited properties with concurrent watchpoints #457
Implement optimizer-visible prototype-chain property specialization and mutation watchpoints so the exact PR-249 GC-wait-vs-Class-A witnesses execute a real inherited-property assumption, jettison it through the shared JIT/GC conductor, and resume correctly.
Acceptance
Profile the receiver and exact prototype holder for inherited named data-property reads without source recognition.
Publish immutable receiver/chain/holder shape guards and the holder slot in optimizer-owned metadata.
Execute the guarded inherited read natively; accessors, proxies, dictionary/exotic objects, chain changes, and misses take the canonical path without replay.
Register shape/prototype dependencies and fire Class-A invalidation only for mutations that can invalidate a published assumption.
Add native-on/off differentials for monomorphic and polymorphic inherited reads, shadowing, replacement, deletion, prototype replacement, accessors, and GC.
Promote cve/mc-safe-gcwait-vs-classa-stop.js and cve/mc-safe-gcwait-vs-classa-stop-noropevariant.js only when normal and TSan runs prove both cooperative collection and Class-A invalidation occurred.
Pass focused normal/TSan gates and the batched JIT/no-GIL/full-unit release gates required by the parents.
Completion
Implemented by 16b2a1a2, a757cd1a, and 2bdb47be; promoted by 3f2fe88b. The exact fixtures each published 18 artifacts, fired 13 invalidations, and completed one cooperative collection in normal and TSan runs. Final parent evidence is recorded in docs/.data/jit-concurrency-2026-07-22.md by ead94ad8.
Complete through 16b2a1a2 → a757cd1a → 2bdb47be → 3f2fe88b.
Inherited one-hop specialization is real: the interpreter profiles the receiver and the exact prototype holder (InheritedPropertyObservation), the optimizer publishes an immutable receiver-shape/holder-shape/holder-slot triple in artifact-owned NativePropertyCache metadata, and AArch64 executes the guarded read inline. Shadowing, prototype replacement, a missing holder property, a null prototype, attribute-mapped holders, and exotic array holders all yield to the canonical path without replay; an in-place rewrite of the holder's own slot is still tracked.
Class-A firing is scoped rather than unconditional: the owner summarizes the shape tokens of its published property caches into a bit filter and a mutation whose receiver shape is absent skips the stop. The filter only over-approximates — collisions, unknown shapes, shapeless receivers, and any site with no object in hand keep the unconditional behavior — so no guarded shape can be missed.
One real gap surfaced and was fixed rather than worked around: a get_prop node types as .other, so o.y + 1 refused Number specialization and left the whole function uncompiled — nothing published, nothing to invalidate. 2bdb47be routes exactly .get_prop-operand binaries through the runtime-operation ABI, with graph-side and lowering-side decisions sharing one binaryNeedsRuntimeOperands helper called on unresolved operand ids so they cannot drift. Keying on may_have_effect was rejected as it would have discarded specialization for all argument arithmetic.
Promotion evidence for the two pinned witnesses, identical in normal and TSan runs with no race reported:
13 invalidations against the witness's 12 mutation rounds plus a cooperative collection is the "both cooperative collection and Class-A invalidation occurred" gate this issue set, not a green assertion. The counter is a lifetime optimizing-tier publication count that survives clear, so a compile-then-jettison still counts. Both cases also pass serialized (~121 s / ~126 s versus ~1 s no-GIL) but join the no-GIL allowlist matching their own --useThreadGILOffUnsafe=1 header.
Gates: 172/172 focused JIT, 142/142 optimizer, 7/7 inherited, 5/5 Class-A, inventory drift check and 7/7 audit self-tests. Inventory moves 243 → 245 promoted, 10 → 8 blocked of the same 259 executables; the retired shared-concurrent-gc dependency is dropped and the JSON is regenerated from --print-inventory rather than hand-edited.
Activity
test(jit): cover inherited property assumptions and unblock the suite (…
feat(jit): fire Class-A stops only for shapes code speculates on (#457)
docs(jit): record why property arithmetic blocks optimizer publication (
feat(jit): lower property arithmetic through the runtime ABI (#457)
feat(jit): report per-case optimizing-tier evidence (#429)
feat(threads): promote the Class-A vs GC-wait witnesses (#457)
chrisbbreuer commentedon Jul 25, 2026
Complete through
16b2a1a2→a757cd1a→2bdb47be→3f2fe88b.Inherited one-hop specialization is real: the interpreter profiles the receiver and the exact prototype holder (
InheritedPropertyObservation), the optimizer publishes an immutable receiver-shape/holder-shape/holder-slot triple in artifact-ownedNativePropertyCachemetadata, and AArch64 executes the guarded read inline. Shadowing, prototype replacement, a missing holder property, a null prototype, attribute-mapped holders, and exotic array holders all yield to the canonical path without replay; an in-place rewrite of the holder's own slot is still tracked.Class-A firing is scoped rather than unconditional: the owner summarizes the shape tokens of its published property caches into a bit filter and a mutation whose receiver shape is absent skips the stop. The filter only over-approximates — collisions, unknown shapes, shapeless receivers, and any site with no object in hand keep the unconditional behavior — so no guarded shape can be missed.
One real gap surfaced and was fixed rather than worked around: a
get_propnode types as.other, soo.y + 1refused Number specialization and left the whole function uncompiled — nothing published, nothing to invalidate.2bdb47beroutes exactly.get_prop-operand binaries through the runtime-operation ABI, with graph-side and lowering-side decisions sharing onebinaryNeedsRuntimeOperandshelper called on unresolved operand ids so they cannot drift. Keying onmay_have_effectwas rejected as it would have discarded specialization for all argument arithmetic.Promotion evidence for the two pinned witnesses, identical in normal and TSan runs with no race reported:
13 invalidations against the witness's 12 mutation rounds plus a cooperative collection is the "both cooperative collection and Class-A invalidation occurred" gate this issue set, not a green assertion. The counter is a lifetime optimizing-tier publication count that survives
clear, so a compile-then-jettison still counts. Both cases also pass serialized (~121 s / ~126 s versus ~1 s no-GIL) but join the no-GIL allowlist matching their own--useThreadGILOffUnsafe=1header.Gates: 172/172 focused JIT, 142/142 optimizer, 7/7 inherited, 5/5 Class-A, inventory drift check and 7/7 audit self-tests. Inventory moves 243 → 245 promoted, 10 → 8 blocked of the same 259 executables; the retired
shared-concurrent-gcdependency is dropped and the JSON is regenerated from--print-inventoryrather than hand-edited.docs(jit): publish concurrency closure evidence