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
GC relocation: compact at precise baseline-JIT safepoints #359
The current AArch64 baseline tier already has one precise moving-GC-safe checkpoint contract: before calling the runtime it publishes canonical Number words for every numeric local, spills live numeric operands, leaves non-numeric slots restricted to pointer-free primitives, publishes the exact instruction/step state, and holds only stable NativeFrame/runtime pointers plus numeric registers. The active interpreter and its Exec/frame roots are registered.
Implement explicit requested compaction at exactly that safepoint without weakening any other fail-closed boundary.
Scope:
add a Context request API with a pending bit that is serviced only at a declared precise interpreter/native safepoint;
mark the baseline native checkpoint precise for the duration of its runtime callback;
factor the compaction gate so exactly the current precise interpreter may move, while ordinary public compaction still requires a fully quiescent realm;
retain rejection for generic interpreter points, native host callbacks, conservative scans, running/parked peers, concurrent/parallel collectors, unsupported backends, and reentrant relocation;
prove a ready native loop requests movement, relocates a protected graph during execution, resumes from the exact checkpoint, returns the right result, and keeps using the same native tier;
machine-check the materialization and narrow activation contract; update concise README/JIT/relocation documentation.
Acceptance:
the request remains pending at imprecise checkpoints and clears only after a real supported attempt;
useful fragmented backing is compacted during a native execution rather than before/after it;
generated execution resumes correctly with exact result/step semantics and no stale managed pointer;
Implemented and pushed in ca92af0. Context.requestGarbageCompaction now remains pending until a checkpoint declares both precise roots and movement-safe resumption. Only the current AArch64 baseline numeric checkpoint sets that stronger token after publishing canonical locals, spilling operands, and recording exact state. Public compactGarbage remains quiescent-only; marking-precise Zig helpers, host callbacks, generic VM paths, peers, conservative scans, and active collectors remain rejected. The witness moves the active Function/Object/protected graph, resumes the same native entry, returns the exact sum, and reaches a dense fixed point. Debug, ReleaseSafe, and TSan each passed 7/7 with zero leaks; C/Objective-C hosts, 117/117 C audit plus extension/private exports, 108/108 Objective-C audit, relocation inventory, and release matrix validation are green.
Activity
feat(gc): compact at precise JIT safepoints (#359)
chrisbbreuer commentedon Jul 20, 2026
Implemented and pushed in ca92af0. Context.requestGarbageCompaction now remains pending until a checkpoint declares both precise roots and movement-safe resumption. Only the current AArch64 baseline numeric checkpoint sets that stronger token after publishing canonical locals, spilling operands, and recording exact state. Public compactGarbage remains quiescent-only; marking-precise Zig helpers, host callbacks, generic VM paths, peers, conservative scans, and active collectors remain rejected. The witness moves the active Function/Object/protected graph, resumes the same native entry, returns the exact sum, and reaches a dense fixed point. Debug, ReleaseSafe, and TSan each passed 7/7 with zero leaks; C/Objective-C hosts, 117/117 C audit plus extension/private exports, 108/108 Objective-C audit, relocation inventory, and release matrix validation are green.