From bc487a71747f3c631890082523eb5fd19b6ff7c1 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Ralph=20K=C3=BCpper?= Date: Wed, 16 Sep 2026 12:24:08 +0000 Subject: [PATCH 1/4] fix(gc): keep transient side allocations out of the drained-bytes term (#10376) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit #10268's drained-bytes term reconstructs external side bytes that a cheap collection released earlier than a full would have, so the full cadence stays pinned to main's while the parse-boundary band holds the live reading down. Its docstring states the invariant it relies on: the sum can never make old-reclaim fire earlier than main, because "every drained byte is a byte main would still have been counting as live at the same point". Regex match scratch breaks that. `perex_memory`'s `Buffer` and `Reservation`, and the replace fast path's `Spans`, are freed by the operation that allocated them, on the same call, in every build — so main is not holding those bytes either, and there is nothing to reconstruct. Because the term accumulates until a full clears it, a loop of such calls grows it without bound: a 1,000,000-call `.test()` loop manufactured enough pressure for three old-gen cycles, one of them a full that traced a 52 MB live arena and freed 59 KB. The bytes that paid for it were never live at any collection. Those three sites now report their release through `gc_note_external_side_free_transient`, which leaves `external_side_live_bytes` exactly as before and does not touch the drained term. Nothing else changes: Map and Set buffers, the JSON tape and the node-api promise delta keep reporting through `gc_note_external_side_free`, so the cadence #10268 measured on `records_array_1m:sparse` is untouched. Measured on the reproducer in #10376, same host, release builds, instructions per iteration with the string-building control subtracted: main 33690c563 10,979 gc cycle_starts=3 steps=4485 completions=1 full this commit 8,219 gc cycle_starts=0 steps=0 completions=0 GC falls from 193 to 32 permille of wall, which is what the same probe read before #10268 landed. --- crates/perry-runtime/src/gc/policy.rs | 19 +++++++++++++++++++ .../perry-runtime/src/regex/perex_memory.rs | 4 ++-- .../src/regex/perex_replace_direct.rs | 2 +- 3 files changed, 22 insertions(+), 3 deletions(-) diff --git a/crates/perry-runtime/src/gc/policy.rs b/crates/perry-runtime/src/gc/policy.rs index 7af0232a0f..f4c2d9cf1c 100644 --- a/crates/perry-runtime/src/gc/policy.rs +++ b/crates/perry-runtime/src/gc/policy.rs @@ -746,6 +746,25 @@ pub(crate) fn gc_note_external_side_free(bytes: usize) { GC_EXTERNAL_SIDE_DRAINED_SINCE_FULL.with(|c| c.set(c.get().saturating_add(bytes))); } +/// Release external side bytes that were never retained past the operation +/// that allocated them — regex match scratch, whose owner frees it on the +/// same call that took it. +/// +/// They leave [`external_side_live_bytes`] exactly as any other bytes do. What +/// they must NOT enter is the drained term, whose whole purpose is to +/// reconstruct bytes a cheap collection released EARLIER than a full would +/// have ([`external_side_old_reclaim_pressure_bytes`]). A transient +/// allocation's release is not early: every build, with or without the +/// parse-boundary band, frees it at the same program point, so there is +/// nothing to reconstruct. Counting it manufactures old-reclaim pressure out +/// of per-call churn that no collection ever saw live — #10376, where a +/// million-call `.test()` loop paid three unproductive old-gen cycles, one of +/// them a full that freed 59 KB of a 52 MB arena, and 33 % more instructions +/// per call. +pub(crate) fn gc_note_external_side_free_transient(bytes: usize) { + GC_EXTERNAL_SIDE_LIVE_BYTES.with(|c| c.set(c.get().saturating_sub(bytes))); +} + /// The external side-buffer term of OLD-RECLAIM pressure. /// /// Live bytes plus all reported releases since the last full baseline. The diff --git a/crates/perry-runtime/src/regex/perex_memory.rs b/crates/perry-runtime/src/regex/perex_memory.rs index 5a72bbedcc..d743690cf0 100644 --- a/crates/perry-runtime/src/regex/perex_memory.rs +++ b/crates/perry-runtime/src/regex/perex_memory.rs @@ -92,7 +92,7 @@ impl Drop for Reservation<'_> { fn drop(&mut self) { self.budget.live.set(self.budget.live.get() - self.bytes); if self.bytes != 0 { - crate::gc::gc_note_external_side_free(self.bytes); + crate::gc::gc_note_external_side_free_transient(self.bytes); } } } @@ -153,7 +153,7 @@ impl Drop for Buffer<'_, T> { fn drop(&mut self) { self.budget.live.set(self.budget.live.get() - self.bytes); if self.bytes != 0 { - crate::gc::gc_note_external_side_free(self.bytes); + crate::gc::gc_note_external_side_free_transient(self.bytes); } } } diff --git a/crates/perry-runtime/src/regex/perex_replace_direct.rs b/crates/perry-runtime/src/regex/perex_replace_direct.rs index 7a674827ee..35661ecb96 100644 --- a/crates/perry-runtime/src/regex/perex_replace_direct.rs +++ b/crates/perry-runtime/src/regex/perex_replace_direct.rs @@ -166,7 +166,7 @@ impl Spans { impl Drop for Spans { fn drop(&mut self) { - crate::gc::gc_note_external_side_free(self.noted); + crate::gc::gc_note_external_side_free_transient(self.noted); } } From f9e8fc7c282f123fead81b7a55134aa0828d1e7f Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Ralph=20K=C3=BCpper?= Date: Wed, 16 Sep 2026 12:31:10 +0000 Subject: [PATCH 2/4] test(gc): pin that a transient side allocation leaves no drained debt Eight rounds of allocate-and-release through the transient path must leave `external_side_old_reclaim_pressure_bytes` exactly where it started, while a live transient buffer still reads as pressure like any other. Sabotage-proved: pointing `gc_note_external_side_free_transient` at `gc_note_external_side_free` fails on round 0 with the term 1 MB above the live reading, which is the shape that accumulated into an unproductive full in #10376. The sibling test that pins the drained term for retained releases is unaffected, since the two paths are now distinct. --- .../src/gc/tests/tiny_parse_pressure.rs | 52 +++++++++++++++++++ 1 file changed, 52 insertions(+) diff --git a/crates/perry-runtime/src/gc/tests/tiny_parse_pressure.rs b/crates/perry-runtime/src/gc/tests/tiny_parse_pressure.rs index 50b505da7a..3423c4d466 100644 --- a/crates/perry-runtime/src/gc/tests/tiny_parse_pressure.rs +++ b/crates/perry-runtime/src/gc/tests/tiny_parse_pressure.rs @@ -512,6 +512,58 @@ fn a_finished_collection_moves_the_external_base_to_the_post_collection_reading( /// `external_side_live_bytes()` read fails BOTH tests below; deleting the reset /// from `finish_full_old_reclaim_baseline` fails /// `a_full_collection_clears_the_drained_debt` alone. +/// A transient side allocation — one the operation that allocated it frees on +/// the same call — must leave the drained term alone, however many times the +/// program does it. +/// +/// The drained term reconstructs bytes a cheap collection released EARLIER +/// than a full would have. Nothing about regex match scratch is early: every +/// build frees it at the same program point, so there is nothing to +/// reconstruct, and accumulating it manufactures old-reclaim pressure out of +/// per-call churn no collection ever saw live. #10376: a million-call +/// `.test()` loop paid three old-gen cycles, one a full that traced a 52 MB +/// arena and freed 59 KB. +/// +/// Sabotage-proved: pointing `gc_note_external_side_free_transient` at +/// `gc_note_external_side_free` fails this test on the first iteration's +/// assertion, with the term 1 MB above the live reading. +#[test] +fn transient_side_allocations_never_enter_the_drained_debt() { + use super::super::policy::{ + external_side_live_bytes, external_side_old_reclaim_pressure_bytes, + GC_EXTERNAL_SIDE_DRAINED_SINCE_FULL, + }; + use super::support::*; + let _isolation = GcTestIsolationGuard::new(); + let restore = GC_EXTERNAL_SIDE_DRAINED_SINCE_FULL.with(|cell| cell.replace(0)); + let live_before = external_side_live_bytes(); + let term_before = external_side_old_reclaim_pressure_bytes(); + + const BYTES: usize = 1024 * 1024; + for round in 0..8 { + crate::gc::gc_note_external_side_alloc(BYTES); + assert_eq!( + external_side_old_reclaim_pressure_bytes(), + term_before + BYTES, + "round {round}: a live transient buffer is pressure like any other" + ); + crate::gc::gc_note_external_side_free_transient(BYTES); + assert_eq!( + external_side_live_bytes(), + live_before, + "round {round}: the live reading must fall by what was released" + ); + assert_eq!( + external_side_old_reclaim_pressure_bytes(), + term_before, + "round {round}: releasing a transient buffer must leave no debt \ + behind — this is what accumulated into a full in #10376" + ); + } + + GC_EXTERNAL_SIDE_DRAINED_SINCE_FULL.with(|cell| cell.set(restore)); +} + #[test] fn a_drained_side_byte_still_pays_old_reclaim_until_the_next_full() { use super::super::policy::{ From 24e8421d02cb9fc0fee0910e400d201b4635c065 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Ralph=20K=C3=BCpper?= Date: Wed, 16 Sep 2026 12:31:31 +0000 Subject: [PATCH 3/4] docs(changelog): fragment for #10377 --- changelog.d/10377-transient-side-bytes.md | 3 +++ 1 file changed, 3 insertions(+) create mode 100644 changelog.d/10377-transient-side-bytes.md diff --git a/changelog.d/10377-transient-side-bytes.md b/changelog.d/10377-transient-side-bytes.md new file mode 100644 index 0000000000..a90ef25381 --- /dev/null +++ b/changelog.d/10377-transient-side-bytes.md @@ -0,0 +1,3 @@ +### Fixed + +- A program that allocates and frees short-lived native scratch in a loop — any `RegExp` call does — no longer accumulates false collection pressure. A million-call `.test()` loop was paying three old-generation collection cycles, one of them a full that traced a 52 MB live heap to free 59 KB, costing 33% more instructions per call (#10376). From 14c365fbdc0650784d4f95666a9c9f6c53d6bf8b Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Ralph=20K=C3=BCpper?= Date: Wed, 16 Sep 2026 12:52:46 +0000 Subject: [PATCH 4/4] tooling(gc): gate the transient release and re-pin the census window Two gate consequences of the previous commit, kept separate from it. `gc_note_external_side_free_transient`'s only callers are the regex scratch owners, and the workspace pins perry-runtime with `default-features = false`, so the product build (`cargo check -p perry --bins`) has no caller for it and the warnings gate rejected it as dead. Gated to `regex-engine`, as its callers are, along with the test that exercises it. The holder inventory pins the mark-complete -> sweep-entry census window by the hashes of its sources, and `gc/policy.rs` is one of them. Re-audited: the change adds one function that does a saturating subtraction on a thread-local Cell and REMOVES an increment from a mutator free path. No allocation, relocation, collection or JS callback enters the window, no mark/sweep control flow changes, and neither boundary moves. Pin updated with that note. --- crates/perry-runtime/src/gc/policy.rs | 5 +++++ crates/perry-runtime/src/gc/tests/tiny_parse_pressure.rs | 1 + scripts/gc_runtime_root_holders.json | 4 ++-- 3 files changed, 8 insertions(+), 2 deletions(-) diff --git a/crates/perry-runtime/src/gc/policy.rs b/crates/perry-runtime/src/gc/policy.rs index f4c2d9cf1c..d280170335 100644 --- a/crates/perry-runtime/src/gc/policy.rs +++ b/crates/perry-runtime/src/gc/policy.rs @@ -761,6 +761,11 @@ pub(crate) fn gc_note_external_side_free(bytes: usize) { /// million-call `.test()` loop paid three unproductive old-gen cycles, one of /// them a full that freed 59 KB of a 52 MB arena, and 33 % more instructions /// per call. +/// Only the regex scratch owners report through here, and the workspace pins +/// perry-runtime with `default-features = false`, so a product build has no +/// caller for it — gate it like its callers rather than leave dead code the +/// warnings lint gate rejects. +#[cfg(feature = "regex-engine")] pub(crate) fn gc_note_external_side_free_transient(bytes: usize) { GC_EXTERNAL_SIDE_LIVE_BYTES.with(|c| c.set(c.get().saturating_sub(bytes))); } diff --git a/crates/perry-runtime/src/gc/tests/tiny_parse_pressure.rs b/crates/perry-runtime/src/gc/tests/tiny_parse_pressure.rs index 3423c4d466..70a1e93202 100644 --- a/crates/perry-runtime/src/gc/tests/tiny_parse_pressure.rs +++ b/crates/perry-runtime/src/gc/tests/tiny_parse_pressure.rs @@ -527,6 +527,7 @@ fn a_finished_collection_moves_the_external_base_to_the_post_collection_reading( /// Sabotage-proved: pointing `gc_note_external_side_free_transient` at /// `gc_note_external_side_free` fails this test on the first iteration's /// assertion, with the term 1 MB above the live reading. +#[cfg(feature = "regex-engine")] #[test] fn transient_side_allocations_never_enter_the_drained_debt() { use super::super::policy::{ diff --git a/scripts/gc_runtime_root_holders.json b/scripts/gc_runtime_root_holders.json index 2d8de2811a..287ede280c 100644 --- a/scripts/gc_runtime_root_holders.json +++ b/scripts/gc_runtime_root_holders.json @@ -313,7 +313,7 @@ "file": "crates/perry-runtime/src/gc/census.rs", "name": "PASS1_MARKED", "verdict": "non_moving_snapshot", - "why": "Real GC header addresses, deliberately untraced so the diagnostic does not keep its observed objects alive. Populated only at the end of mark propagation of a synchronous full cycle; consumed at sweep entry in the same run_to_completion invocation. The intervening full-cycle phases do not relocate or run JS callbacks. The Vec is used for membership comparisons and dropped with the census before sweep. Budgeted and minor cycles skip both boundaries. Pin re-audited 2026-09-05 after #9760 touched `gc/mod.rs`: that change is `mod heap_stats;` plus a `pub(crate) use` re-export and alters no mark/sweep control flow. `heap_stats()` is reached only from `js_bun_jsc_heap_stats` (the JS-facing `bun:jsc.heapStats()`), i.e. from mutator code, never inside a cycle, and its own module contract forbids allocation or collection during its walk. The mark-complete → sweep-entry window is unchanged. Re-audited 2026-09-05 (train125) after #9769 and #9771 touched pinned files. #9769 adds one `reg_scanner!` registration to `gc/mod.rs`; #9771 adds a feature-gated `alloc_census_init()` there and a feature-gated Rust-heap dump inside `take_census`. `alloc-census` is not in the default feature set, and decisively: `census_take_if_armed_at_full_sweep_start` does `PASS1_MARKED.with(|p| p.borrow_mut().take())` BEFORE calling `take_census`, so the snapshot has already left the thread-local by the time #9771's code runs — it cannot affect the window. Neither change alters mark/sweep control flow. Re-audited 2026-09-06 after #9831 touched `gc/policy.rs`. Its hunks are (a) the tiny-parse pressure guard's pricing (`tiny_parse_pressure_headroom_bytes`, `tiny_parse_pressure_due*`, a `Cell` byte-count base) consulted from JSON.parse's mutator-side boundaries (`gc_bump_malloc_trigger`, `gc_collect_pending_suppressed_parse`, `gc_schedule_parse_boundary_collection_if_pressure`), none of which is reachable from inside a cycle, and (b) one extra `Cell` store in `note_collection_finished_arena_occupancy`, which runs from `publish_reclaim_outcome` in the Publish subphase — after `step_sweep` has already consumed the snapshot. Mark/sweep control flow between `census_pass1_if_armed` and `census_take_if_armed_at_full_sweep_start` is untouched. Re-audited 2026-09-05 (train126) after #9755 restructured `gc/cycle.rs`. Its hunks are all root-scan machinery (`RootScanSubphase`, `RootScanCycleState`, the mutable-scanner iteration state), which runs BEFORE mark propagation completes; `gc/mod.rs` gains only a `mod young_log;` declaration. The bracketing is unchanged — `census_pass1_if_armed` is still inside `step_mark_propagation` and `census_take_if_armed_at_full_sweep_start` inside `step_sweep` — and a synchronous full mark-sweep still moves nothing between them. Re-pinned 2026-09-05 for the #9740 hot-TLS conversion of this file: the sole change is `thread_local!` → `crate::perry_thread_local!`, a macro-name swap with identical declaration syntax and `.with()` call sites. No control flow, no phase boundary, and no storage semantics change. Re-audited 2026-09-06 (train128) after #9794's GC diagnostics touched `gc/mod.rs` and `gc/policy.rs`: both gain diagnostic module declarations and counters only — no mark/sweep control flow, and the census bracketing in `step_mark_propagation` / `step_sweep` is unchanged. Re-audited for #9794's GC diagnostics: `gc/mod.rs` gains `mod diag_sites;` / `mod survival_diag;`, a re-export, a `diag_sites::full_started(...)` call at TRIGGER time (before mark propagation begins), and exit-time reporting. Nothing executes between mark-complete and sweep-entry, so the window is unchanged. Re-audited 2026-09-06 for the retained array-growth verifier fix: the cycle.rs change passes the existing non-copying evacuation verifier an explicit all-forwarded policy. That call remains in minor finalization, outside the synchronous full-cycle census window; its root and heap reads do not allocate GC objects, move objects, or invoke JS callbacks. The mark-complete and sweep-entry boundaries are unchanged. Re-audited 2026-09-05 after #9830 touched `gc/policy.rs`. That change is (a) six `thread_local! {` blocks rewritten as `crate::perry_thread_local! {` and (b) one `#[cfg(test)]` accessor listing the trigger path's hot-slot indices. The macro keeps the same storage, the same `.with()` at every read and write, and the same destructor registration (the teardown guard exists exactly when `needs_drop` holds, which is what `std::thread_local!` already decided); no value, predicate or branch in the file changes, so no mark or sweep control flow does. The one new behaviour is on a declaration's FIRST read: `HotKey::resolve_and_cache` takes a mutex and allocates a key through the GLOBAL allocator. Even if a first read landed inside this window it would be sound — the window's contract is that nothing relocates and no JS callback runs, and a mimalloc allocation does neither. `census_pass1_if_armed` is still inside `step_mark_propagation` and `census_take_if_armed_at_full_sweep_start` inside `step_sweep`; the bracketing is untouched. Re-audited 2026-09-06 (train132) after #9860 and #9845 touched `gc/mod.rs`. Both hunks are re-export lists and nothing else: #9860 adds `idle_reclaim_elapsed_starts` / `IDLE_RECLAIM_REARM_MS`, and #9845 adds `owner_is_dead_copied_minor_from_space_of_type`. No mark or sweep control flow changes. #9845's substantive work sits in `gc/oldgen.rs` and `gc/copying.rs`, neither pinned: the copying-minor arm (`finalize_dead_copied_minor_from_space_regexps`) runs on a MINOR, which skips both census boundaries; the full-cycle arm (`collect_dead_registered_regexps_post_trace`, from `with_dead_collection_finalize`) walks the RegExp registry building a Vec of addresses — no GC allocation, no JS callback, so it cannot relocate the snapshot's subjects — and it is reached from the sweep body, i.e. AFTER `census_take_if_armed_at_full_sweep_start` has already `take()`n the snapshot out of the thread-local. The mark-complete -> sweep-entry window is unchanged. Re-audited 2026-09-07 for #9965 after 1ec9e0e8a touched `gc/cycle.rs` and `gc/mod.rs`: `gc/mod.rs:216-217` only declares and imports the failure-attribution module, while `gc/cycle.rs:1414-1417` reads the trigger and diagnostic counters immediately before evacuation verification inside `atomic_finalize_minor_prelude`. Full cycles bypass `MinorPrelude` at `gc/cycle.rs:1192-1196`; evacuation remains guarded by the minor-only context at `gc/cycle.rs:1330-1372`. The snapshot store remains at `gc/cycle.rs:963-964` after synchronous full marking, and its take remains at `gc/cycle.rs:1454-1457` before sweep. No new write, relocation, collection, or JS callback was added to that full-cycle interval, so the PASS1_MARKED window is unaffected. Re-audited 2026-09-07 for the regex census rows: all new work is in `take_census` after `census_take_if_armed_at_full_sweep_start` has taken PASS1_MARKED out of TLS; neither boundary nor the intervening cycle control flow changed. Re-audited 2026-09-08 (train144) after #9976 and #9977 touched pinned files. `gc/mod.rs` gains exactly three lines: `mod copying_phase;` and `mod regex_census;` (declarations) and one `reg_scanner!(regex::site_test::scan_roots_mut)` registration. A scanner registration adds a root SOURCE for the mutable-root walks; it does not move either census boundary and runs nowhere between them. `gc/census.rs` widens `side_tables()` to `pub(super)`, extends it with regex rows and adds a test module — all census REPORTING, which runs from the diagnostic dump, not inside a cycle. Mark/sweep control flow between `census_pass1_if_armed` and `census_take_if_armed_at_full_sweep_start` is untouched. Re-audited 2026-09-08 for #9849 JSON construction deferral. `gc/mod.rs` adds the `json_defer` module/re-export and a trusted-header layout helper used only by already-validated JSON emitters; neither changes or runs in collector phase control flow. `gc/policy.rs` adds JSON completion scheduling, construction-grace checks, and safepoint deferral predicates. These are called from mutator-side JSON allocation/output boundaries and ordinary safepoint entry; they do not alter `step_mark_propagation`, `step_sweep`, or invoke callbacks or relocation between the census boundaries. The mark-complete to sweep-entry window is unchanged. The follow-up adds a cfg(test)-only one-shot boolean for deterministic explicit-pressure fixtures; it is absent from production builds and cannot affect the census window. The first predicate read consumes it, so post-parse accounting exercises normal pricing. Re-audited 2026-09-09 for bounded tiny-JSON completion polling. The policy.rs changes split the mutator-side pending-parse check into an inlined empty fast path plus an outlined debt-service path, and amortize the mutator-side arena-pressure read across 64 bounded parse completions. Neither function is reachable from step_mark_propagation or step_sweep; neither census boundary nor the synchronous full-cycle interval between them changes. Re-audited 2026-09-09 for lazy JSON record batches: policy.rs only widens gc_budgeted_cycle_active visibility from pub(super) to pub(crate). Its body remains a read-only Cell query. The new caller is lazy_get materialization in the mutator; run_to_completion, step_mark_propagation, census snapshot consumption at step_sweep, and the synchronous non-moving window are unchanged. Re-audited 2026-09-09 for completed JSON-output debt: the added gc_service_json_output_sweep function calls the existing trigger check from a rooted mutator boundary and reports whether its malloc-count request remains due. It is not called from any census or collector phase; the synchronous mark-complete to sweep-entry window is unchanged. Re-audited 2026-09-09 for the JSON byte-debt carry: the same mutator-only service helper now distinguishes requests satisfied before its call from those satisfied by its trigger check. The added enum contains no payload, both count reads are scalar, and no census boundary or collector phase changed. Re-audited 2026-09-11 for #10055: gc/mod.rs only registers the weak UTF-16 index scanner during gc_init. It neither marks strings nor allocates GC objects or runs JS; offset vectors use the Rust allocator. The mark-complete to sweep-entry census window and cycle control flow are unchanged. Re-audited 2026-09-11 for #10054: gc/mod.rs adds only the trim-cache mutable-root scanner registration in gc_init. Its scanner visits two existing string slots without allocating or invoking JS. Root scanning still precedes mark completion, and neither census boundary nor the synchronous mark-complete to sweep-entry window changes. Re-audited 2026-09-11 for #10060: the census array classifier now reads the logical element start and bounds its scan by the remaining capacity. The helper only reads the existing GC/header words and performs pointer arithmetic; it cannot allocate, collect, or call JS. This classifier runs in take_census after PASS1_MARKED has been taken out of TLS. Neither census boundary nor the mark-complete to sweep-entry control flow changed. Re-audited for #8512: gc/mod.rs only enables the existing PTY mutable-root scanner on Windows; it changes no mark/sweep phase or census boundary. The scanner visits NaN-boxed slots without running JS callbacks. Re-audited 2026-09-12 for the single regular-expression engine: `gc/mod.rs` changes `mod prefetch;` to `pub(crate) mod prefetch;` so the RegExp owner-table walks can prefetch headers, a visibility change with no new call in collector control flow; `gc/census.rs` changes only its `#[cfg(test)]` `regex_census_tests` module, dropping assertions for the previous engine's cache rows. Neither boundary (`census_pass1_if_armed` in `step_mark_propagation`, `census_take_if_armed_at_full_sweep_start` in `step_sweep`) nor the synchronous mark-complete to sweep-entry interval changes. Re-audited 2026-09-13 after the #10169 fix touched `gc/mod.rs` and `gc/policy.rs`. `gc/mod.rs` gains only `pub(crate) use` re-exports (`policy::note_young_leaf_born_old`, `policy::young_generation_holds_a_nursery`, `promote_in_place::{young_generation_measured_dying, young_generation_measured_retained}`, and cfg(test) survival seeders). `gc/policy.rs` gains a `Cell` thread-local (`GC_YOUNG_LEAF_BORN_OLD`, no pointer), its setter, a pure predicate over `copying_from_space_in_use_bytes` vs the base nursery cap, and a consumed-once branch at the top of `gc_budgeted_due_trigger` that may answer `YoungScavengeCap` ahead of `OldReclaim`. That branch decides WHICH collection a safepoint starts (a minor instead of a full); it runs before any cycle begins and never inside one, so the mark-complete → sweep-entry window of a synchronous full — where PASS1_MARKED is populated and consumed within one `run_to_completion` — is unchanged, and neither hunk adds an allocation, a JS callback, or a relocation to it. Re-audited 2026-09-13 for the heap generation (#10164 cross-call search positions): `gc/mod.rs` only declares `pub(crate) mod heap_generation;`. `gc/cycle.rs` wraps the `Sweep` and `Reclaim` arms of `GcCycleState::step` in a `HeapChange` scope and opens one inside `atomic_finalize_minor_prelude`'s evacuation branch (with a nested one around old-page defrag). Opening and closing a scope only increments two thread-local integer cells (`HEAP_GENERATION`, `OPEN_HEAP_CHANGES`); a first thread-local read may allocate a key through the global allocator, which neither relocates nor runs JS. The `Sweep` scope opens immediately before `step_sweep`, i.e. before `census_take_if_armed_at_full_sweep_start` takes PASS1_MARKED out of TLS, and adds no relocation, collection or JS callback to the synchronous mark-complete to sweep-entry window; the minor-prelude scope is unreachable from a full cycle, which bypasses `MinorPrelude`. Neither boundary nor the intervening control flow changed. Re-audited 2026-09-13 for #10182 block-granular reclamation, which touched `gc/cycle.rs`. Two hunks: (a) in the `RememberedSetRebuild` subphase of AtomicFinalize — INSIDE the window — the require-marked old-to-young rebuild is now constructed with `OldToYoungRememberedRebuildState::new_skipping`, whose cursor never enters blocks the census recorded as holding no reached, pinned or pre-marked object (`BlockCensus::unmarked_blocks`); computing that list reads `arena_block_snapshots()` and allocates one `Vec` through the global allocator. It visits a subset of the same objects the rebuild already walked (every skipped object would have been rejected as unmarked), and it neither allocates a GC object, relocates anything, nor runs a JS callback. (b) In `step_sweep`, `IncrementalSweepState::with_block_skip` runs after `census_take_if_armed_at_full_sweep_start` has already taken PASS1_MARKED out of TLS. Neither boundary moved and the synchronous mark-complete to sweep-entry interval gains no relocation, collection or callback. Re-audited 2026-09-11 for the startup memory profile: gc/mod.rs only retains the pre-main allocator-policy constructor in js_gc_init. The constructor applies process allocation options, without invoking GC or JS. No census boundary, collector phase, or mark-complete to sweep-entry control flow changed. Re-audited 2026-09-13 for #10179: census.rs only adds a native regex cache metadata row and its unit assertion; snapshot consumption and the full-cycle window are unchanged. Re-audited 2026-09-14 for the GC due-check fast path, which touched `gc/mod.rs` and `gc/policy.rs`. `gc/mod.rs` only changes the safepoint re-exports: `gc_runtime_safepoint` becomes cfg(test) and `gc_runtime_safepoint_poll` is added. `gc/policy.rs`: the budgeted step returns a debt-free `GcStepReport` (debt is attached by the FFI and test entry points after the step returns) and moves cycle start/step into an out-of-line `gc_budgeted_start_or_step`; `gc_check_trigger` reuses a repeatable due-trigger answer through `DueTriggerMemo`, placed after its `GC_FLAG_IN_ALLOC` and suppression early returns; the young scavenge cap reuses the old-gen pressure value the due trigger already read and checks the census-seeded flag first. All of it runs from mutator safepoints, allocation-point trigger checks and the host step API, before a cycle starts or between budgeted steps. None of it is reachable between `census_pass1_if_armed` in `step_mark_propagation` and `census_take_if_armed_at_full_sweep_start` in `step_sweep` of a synchronous full: an allocation inside that window reaches `gc_check_trigger` with `GC_FLAG_IN_ALLOC` set and returns before the changed code. No allocation, relocation, collection or JS callback is added to the window. Re-audited 2026-09-14 for the tiny-parse nursery-cap boundary, which touched `gc/policy.rs`. It adds `tiny_parse_generational_collection_due`, a pure predicate (the existing `tiny_parse_pressure_due` OR the existing `young_scavenge_cap_due` read), and calls it instead of `tiny_parse_pressure_due` from `gc_bump_malloc_trigger_inner` and `gc_collect_pending_suppressed_parse_slow` (generational branch only) and from `gc_schedule_parse_boundary_collection_if_pressure`. All three are JSON.parse mutator-side boundaries, none reachable from `step_mark_propagation` or `step_sweep`; the predicate reads counters and allocates nothing. Neither census boundary nor the synchronous mark-complete to sweep-entry interval changed. Re-audited 2026-09-13 for #10182's full-collection throughput follow-up, which touched `gc/cycle.rs` in one hunk, INSIDE the window: the `RememberedSetRebuild` subphase of a synchronous full now first asks `verify::full_remembered_rebuild_provably_empty` and, when it holds, installs `OldToYoungRememberedRebuildState::provably_empty()` (an empty sticky set, no walk) instead of the require-marked rebuild. The predicate reads `arena_block_snapshots()` (one `Vec` through the global allocator), the census's per-block reached/pre-marked facts and the malloc registry's length; the constructor bumps a `Cell` counter and prints one line under `PERRY_GC_DIAG`. None of it allocates a GC object, relocates anything, collects, or runs a JS callback, and both census boundaries stay where they were. Re-audited 2026-09-14 for #10182's pacing-full work, which touched `gc/cycle.rs`, `gc/mod.rs` and `gc/policy.rs`. `gc/cycle.rs`: `GcCycleState::new_full` no longer calls `materialize_all_promoted_page_runs`; that call ran in the constructor, before the census and far before `census_pass1_if_armed`, and removing it adds nothing to the window. `gc/mod.rs`: one `mod promoted_cohort;` declaration. `gc/policy.rs`: (a) `credit_promoted_bytes_to_old_baseline` also credits a `Cell` cohort counter (it runs after a copying minor completes); (b) `finish_full_old_reclaim_baseline` also records the verified old live bytes and resets that counter (Publish, after `step_sweep` consumed the snapshot); (c) `gc_safepoint_moving_minor` arms and disarms the promotion-census record around its nursery minor and calls `run_promoted_cohort_full_if_due`, which starts a synchronous full through the same `gc_collect_full_mark_sweep_with_trigger` entry and reads byte counters before and after it. All of it runs before a cycle starts or after it completes; none of it runs between mark completion and sweep entry, allocates a GC object, relocates anything, or calls into JS. The census the promoted-cohort full may adopt from the promotion walk is built in `BuildValidPointerSet`, before either boundary. Both boundaries are unchanged. Re-audited 2026-09-14 for the #10182 dead-stack scrub in `gc/cycle.rs`: `step_build_valid_pointer_set` now calls `scrub_dead_stack_below`, which zeroes a local array in its own frame (dead stack below the caller), right after the census finishes — in `BuildValidPointerSet`, before the root scan and far before `census_pass1_if_armed`. It writes no heap memory, allocates nothing, relocates nothing and calls no JS; both boundaries are unchanged. Re-audited 2026-09-14 for #10241 (cohort survival), which touched `gc/cycle.rs` and `gc/policy.rs`. `gc/cycle.rs`: one call, `promoted_cohort::survival::check_minor_view_at_full_sweep_start()`, in `step_sweep` immediately AFTER `census_take_if_armed_at_full_sweep_start` has taken PASS1_MARKED out of TLS, i.e. outside the window. It is a no-op unless a promoted-cohort full armed its survival probe; when armed it walks the old page index over the preceding minor's dirty pages (`old_arena_walk_objects_on_pages`, Rust-allocator Vecs), reads GC headers' mark flags and the slots of unmarked ones, and records one enum. It writes no heap memory, allocates no GC object, relocates nothing and calls no JS. `gc/policy.rs`: `run_promoted_cohort_full_if_due` arms the probe before `gc_collect_full_mark_sweep_with_trigger` and takes it after the full returns (feeding `note_full_measured_promotion_survival` and one diagnostic line); both run before a cycle starts or after it completes. Both boundaries are unchanged. Re-audited 2026-09-14 for #10241's in-place-only cohort: `gc/policy.rs` drops the `promoted_cohort::note_promoted` call from `credit_promoted_bytes_to_old_baseline` (the copying minor now calls `promoted_cohort::note_minor_promotion` itself, after the credit). Both run at the end of a copying minor, outside any full cycle; the mark-complete to sweep-entry window is unchanged. Re-audited 2026-09-14 for the parse-boundary side-allocation band (medium-parse pacing), which touched `gc/policy.rs`. Three hunks: (a) a `Cell` thread-local (`GC_LAST_COLLECTION_EXTERNAL_SIDE_BYTES`, a byte COUNT, no pointer) plus three pure predicates over it and `external_side_live_bytes()`; (b) that predicate added as a third disjunct of `tiny_parse_generational_collection_due`, which is read only from the three JSON.parse mutator-side boundaries (`gc_bump_malloc_trigger_inner`, `gc_collect_pending_suppressed_parse_slow`, `gc_schedule_parse_boundary_collection_if_pressure`), none of them reachable from `step_mark_propagation` or `step_sweep`; and (c) one extra `Cell` store in `note_collection_finished_arena_occupancy` plus two extra reads in the `PERRY_GC_DIAG` tiny-parse line. `note_collection_finished_arena_occupancy` runs from `publish_reclaim_outcome` in the Publish subphase, i.e. AFTER `step_sweep` has already `take()`n the snapshot out of the thread-local, exactly as #9831's store on the same line does. Nothing added allocates a GC object, relocates anything, or runs a JS callback, and neither census boundary moved. Re-audited 2026-09-14 for the drained-bytes counterweight to that band, which touched `gc/policy.rs` again. Four hunks: a second `Cell` thread-local (`GC_EXTERNAL_SIDE_DRAINED_SINCE_FULL`, a byte COUNT); one increment of it inside `gc_note_external_side_free`; a pure read (`external_side_old_reclaim_pressure_bytes`) substituted for `external_side_live_bytes()` at the four old-reclaim pressure sites; and one `Cell` store at the top of `finish_full_old_reclaim_baseline`. None of it can run between the census boundaries. `gc_note_external_side_free` is also reached by mutator-side tape materialization, regex scratch teardown, native-addon adjustments and buffer replacement. Its added operation is only a saturating increment of a scalar Cell, with no GC allocation, relocation, collection or JS callback, so this wider caller set does not invalidate the census window. `finish_full_old_reclaim_baseline` runs from `publish_reclaim_outcome` in the Publish subphase, the same place #9831's store already sits. The pressure reads happen at trigger decisions, before a cycle starts. No allocation, relocation, collection or JS callback is added to the mark-complete -> sweep-entry window, and neither boundary moved.", + "why": "Real GC header addresses, deliberately untraced so the diagnostic does not keep its observed objects alive. Populated only at the end of mark propagation of a synchronous full cycle; consumed at sweep entry in the same run_to_completion invocation. The intervening full-cycle phases do not relocate or run JS callbacks. The Vec is used for membership comparisons and dropped with the census before sweep. Budgeted and minor cycles skip both boundaries. Pin re-audited 2026-09-05 after #9760 touched `gc/mod.rs`: that change is `mod heap_stats;` plus a `pub(crate) use` re-export and alters no mark/sweep control flow. `heap_stats()` is reached only from `js_bun_jsc_heap_stats` (the JS-facing `bun:jsc.heapStats()`), i.e. from mutator code, never inside a cycle, and its own module contract forbids allocation or collection during its walk. The mark-complete → sweep-entry window is unchanged. Re-audited 2026-09-05 (train125) after #9769 and #9771 touched pinned files. #9769 adds one `reg_scanner!` registration to `gc/mod.rs`; #9771 adds a feature-gated `alloc_census_init()` there and a feature-gated Rust-heap dump inside `take_census`. `alloc-census` is not in the default feature set, and decisively: `census_take_if_armed_at_full_sweep_start` does `PASS1_MARKED.with(|p| p.borrow_mut().take())` BEFORE calling `take_census`, so the snapshot has already left the thread-local by the time #9771's code runs — it cannot affect the window. Neither change alters mark/sweep control flow. Re-audited 2026-09-06 after #9831 touched `gc/policy.rs`. Its hunks are (a) the tiny-parse pressure guard's pricing (`tiny_parse_pressure_headroom_bytes`, `tiny_parse_pressure_due*`, a `Cell` byte-count base) consulted from JSON.parse's mutator-side boundaries (`gc_bump_malloc_trigger`, `gc_collect_pending_suppressed_parse`, `gc_schedule_parse_boundary_collection_if_pressure`), none of which is reachable from inside a cycle, and (b) one extra `Cell` store in `note_collection_finished_arena_occupancy`, which runs from `publish_reclaim_outcome` in the Publish subphase — after `step_sweep` has already consumed the snapshot. Mark/sweep control flow between `census_pass1_if_armed` and `census_take_if_armed_at_full_sweep_start` is untouched. Re-audited 2026-09-05 (train126) after #9755 restructured `gc/cycle.rs`. Its hunks are all root-scan machinery (`RootScanSubphase`, `RootScanCycleState`, the mutable-scanner iteration state), which runs BEFORE mark propagation completes; `gc/mod.rs` gains only a `mod young_log;` declaration. The bracketing is unchanged — `census_pass1_if_armed` is still inside `step_mark_propagation` and `census_take_if_armed_at_full_sweep_start` inside `step_sweep` — and a synchronous full mark-sweep still moves nothing between them. Re-pinned 2026-09-05 for the #9740 hot-TLS conversion of this file: the sole change is `thread_local!` → `crate::perry_thread_local!`, a macro-name swap with identical declaration syntax and `.with()` call sites. No control flow, no phase boundary, and no storage semantics change. Re-audited 2026-09-06 (train128) after #9794's GC diagnostics touched `gc/mod.rs` and `gc/policy.rs`: both gain diagnostic module declarations and counters only — no mark/sweep control flow, and the census bracketing in `step_mark_propagation` / `step_sweep` is unchanged. Re-audited for #9794's GC diagnostics: `gc/mod.rs` gains `mod diag_sites;` / `mod survival_diag;`, a re-export, a `diag_sites::full_started(...)` call at TRIGGER time (before mark propagation begins), and exit-time reporting. Nothing executes between mark-complete and sweep-entry, so the window is unchanged. Re-audited 2026-09-06 for the retained array-growth verifier fix: the cycle.rs change passes the existing non-copying evacuation verifier an explicit all-forwarded policy. That call remains in minor finalization, outside the synchronous full-cycle census window; its root and heap reads do not allocate GC objects, move objects, or invoke JS callbacks. The mark-complete and sweep-entry boundaries are unchanged. Re-audited 2026-09-05 after #9830 touched `gc/policy.rs`. That change is (a) six `thread_local! {` blocks rewritten as `crate::perry_thread_local! {` and (b) one `#[cfg(test)]` accessor listing the trigger path's hot-slot indices. The macro keeps the same storage, the same `.with()` at every read and write, and the same destructor registration (the teardown guard exists exactly when `needs_drop` holds, which is what `std::thread_local!` already decided); no value, predicate or branch in the file changes, so no mark or sweep control flow does. The one new behaviour is on a declaration's FIRST read: `HotKey::resolve_and_cache` takes a mutex and allocates a key through the GLOBAL allocator. Even if a first read landed inside this window it would be sound — the window's contract is that nothing relocates and no JS callback runs, and a mimalloc allocation does neither. `census_pass1_if_armed` is still inside `step_mark_propagation` and `census_take_if_armed_at_full_sweep_start` inside `step_sweep`; the bracketing is untouched. Re-audited 2026-09-06 (train132) after #9860 and #9845 touched `gc/mod.rs`. Both hunks are re-export lists and nothing else: #9860 adds `idle_reclaim_elapsed_starts` / `IDLE_RECLAIM_REARM_MS`, and #9845 adds `owner_is_dead_copied_minor_from_space_of_type`. No mark or sweep control flow changes. #9845's substantive work sits in `gc/oldgen.rs` and `gc/copying.rs`, neither pinned: the copying-minor arm (`finalize_dead_copied_minor_from_space_regexps`) runs on a MINOR, which skips both census boundaries; the full-cycle arm (`collect_dead_registered_regexps_post_trace`, from `with_dead_collection_finalize`) walks the RegExp registry building a Vec of addresses — no GC allocation, no JS callback, so it cannot relocate the snapshot's subjects — and it is reached from the sweep body, i.e. AFTER `census_take_if_armed_at_full_sweep_start` has already `take()`n the snapshot out of the thread-local. The mark-complete -> sweep-entry window is unchanged. Re-audited 2026-09-07 for #9965 after 1ec9e0e8a touched `gc/cycle.rs` and `gc/mod.rs`: `gc/mod.rs:216-217` only declares and imports the failure-attribution module, while `gc/cycle.rs:1414-1417` reads the trigger and diagnostic counters immediately before evacuation verification inside `atomic_finalize_minor_prelude`. Full cycles bypass `MinorPrelude` at `gc/cycle.rs:1192-1196`; evacuation remains guarded by the minor-only context at `gc/cycle.rs:1330-1372`. The snapshot store remains at `gc/cycle.rs:963-964` after synchronous full marking, and its take remains at `gc/cycle.rs:1454-1457` before sweep. No new write, relocation, collection, or JS callback was added to that full-cycle interval, so the PASS1_MARKED window is unaffected. Re-audited 2026-09-07 for the regex census rows: all new work is in `take_census` after `census_take_if_armed_at_full_sweep_start` has taken PASS1_MARKED out of TLS; neither boundary nor the intervening cycle control flow changed. Re-audited 2026-09-08 (train144) after #9976 and #9977 touched pinned files. `gc/mod.rs` gains exactly three lines: `mod copying_phase;` and `mod regex_census;` (declarations) and one `reg_scanner!(regex::site_test::scan_roots_mut)` registration. A scanner registration adds a root SOURCE for the mutable-root walks; it does not move either census boundary and runs nowhere between them. `gc/census.rs` widens `side_tables()` to `pub(super)`, extends it with regex rows and adds a test module — all census REPORTING, which runs from the diagnostic dump, not inside a cycle. Mark/sweep control flow between `census_pass1_if_armed` and `census_take_if_armed_at_full_sweep_start` is untouched. Re-audited 2026-09-08 for #9849 JSON construction deferral. `gc/mod.rs` adds the `json_defer` module/re-export and a trusted-header layout helper used only by already-validated JSON emitters; neither changes or runs in collector phase control flow. `gc/policy.rs` adds JSON completion scheduling, construction-grace checks, and safepoint deferral predicates. These are called from mutator-side JSON allocation/output boundaries and ordinary safepoint entry; they do not alter `step_mark_propagation`, `step_sweep`, or invoke callbacks or relocation between the census boundaries. The mark-complete to sweep-entry window is unchanged. The follow-up adds a cfg(test)-only one-shot boolean for deterministic explicit-pressure fixtures; it is absent from production builds and cannot affect the census window. The first predicate read consumes it, so post-parse accounting exercises normal pricing. Re-audited 2026-09-09 for bounded tiny-JSON completion polling. The policy.rs changes split the mutator-side pending-parse check into an inlined empty fast path plus an outlined debt-service path, and amortize the mutator-side arena-pressure read across 64 bounded parse completions. Neither function is reachable from step_mark_propagation or step_sweep; neither census boundary nor the synchronous full-cycle interval between them changes. Re-audited 2026-09-09 for lazy JSON record batches: policy.rs only widens gc_budgeted_cycle_active visibility from pub(super) to pub(crate). Its body remains a read-only Cell query. The new caller is lazy_get materialization in the mutator; run_to_completion, step_mark_propagation, census snapshot consumption at step_sweep, and the synchronous non-moving window are unchanged. Re-audited 2026-09-09 for completed JSON-output debt: the added gc_service_json_output_sweep function calls the existing trigger check from a rooted mutator boundary and reports whether its malloc-count request remains due. It is not called from any census or collector phase; the synchronous mark-complete to sweep-entry window is unchanged. Re-audited 2026-09-09 for the JSON byte-debt carry: the same mutator-only service helper now distinguishes requests satisfied before its call from those satisfied by its trigger check. The added enum contains no payload, both count reads are scalar, and no census boundary or collector phase changed. Re-audited 2026-09-11 for #10055: gc/mod.rs only registers the weak UTF-16 index scanner during gc_init. It neither marks strings nor allocates GC objects or runs JS; offset vectors use the Rust allocator. The mark-complete to sweep-entry census window and cycle control flow are unchanged. Re-audited 2026-09-11 for #10054: gc/mod.rs adds only the trim-cache mutable-root scanner registration in gc_init. Its scanner visits two existing string slots without allocating or invoking JS. Root scanning still precedes mark completion, and neither census boundary nor the synchronous mark-complete to sweep-entry window changes. Re-audited 2026-09-11 for #10060: the census array classifier now reads the logical element start and bounds its scan by the remaining capacity. The helper only reads the existing GC/header words and performs pointer arithmetic; it cannot allocate, collect, or call JS. This classifier runs in take_census after PASS1_MARKED has been taken out of TLS. Neither census boundary nor the mark-complete to sweep-entry control flow changed. Re-audited for #8512: gc/mod.rs only enables the existing PTY mutable-root scanner on Windows; it changes no mark/sweep phase or census boundary. The scanner visits NaN-boxed slots without running JS callbacks. Re-audited 2026-09-12 for the single regular-expression engine: `gc/mod.rs` changes `mod prefetch;` to `pub(crate) mod prefetch;` so the RegExp owner-table walks can prefetch headers, a visibility change with no new call in collector control flow; `gc/census.rs` changes only its `#[cfg(test)]` `regex_census_tests` module, dropping assertions for the previous engine's cache rows. Neither boundary (`census_pass1_if_armed` in `step_mark_propagation`, `census_take_if_armed_at_full_sweep_start` in `step_sweep`) nor the synchronous mark-complete to sweep-entry interval changes. Re-audited 2026-09-13 after the #10169 fix touched `gc/mod.rs` and `gc/policy.rs`. `gc/mod.rs` gains only `pub(crate) use` re-exports (`policy::note_young_leaf_born_old`, `policy::young_generation_holds_a_nursery`, `promote_in_place::{young_generation_measured_dying, young_generation_measured_retained}`, and cfg(test) survival seeders). `gc/policy.rs` gains a `Cell` thread-local (`GC_YOUNG_LEAF_BORN_OLD`, no pointer), its setter, a pure predicate over `copying_from_space_in_use_bytes` vs the base nursery cap, and a consumed-once branch at the top of `gc_budgeted_due_trigger` that may answer `YoungScavengeCap` ahead of `OldReclaim`. That branch decides WHICH collection a safepoint starts (a minor instead of a full); it runs before any cycle begins and never inside one, so the mark-complete → sweep-entry window of a synchronous full — where PASS1_MARKED is populated and consumed within one `run_to_completion` — is unchanged, and neither hunk adds an allocation, a JS callback, or a relocation to it. Re-audited 2026-09-13 for the heap generation (#10164 cross-call search positions): `gc/mod.rs` only declares `pub(crate) mod heap_generation;`. `gc/cycle.rs` wraps the `Sweep` and `Reclaim` arms of `GcCycleState::step` in a `HeapChange` scope and opens one inside `atomic_finalize_minor_prelude`'s evacuation branch (with a nested one around old-page defrag). Opening and closing a scope only increments two thread-local integer cells (`HEAP_GENERATION`, `OPEN_HEAP_CHANGES`); a first thread-local read may allocate a key through the global allocator, which neither relocates nor runs JS. The `Sweep` scope opens immediately before `step_sweep`, i.e. before `census_take_if_armed_at_full_sweep_start` takes PASS1_MARKED out of TLS, and adds no relocation, collection or JS callback to the synchronous mark-complete to sweep-entry window; the minor-prelude scope is unreachable from a full cycle, which bypasses `MinorPrelude`. Neither boundary nor the intervening control flow changed. Re-audited 2026-09-13 for #10182 block-granular reclamation, which touched `gc/cycle.rs`. Two hunks: (a) in the `RememberedSetRebuild` subphase of AtomicFinalize — INSIDE the window — the require-marked old-to-young rebuild is now constructed with `OldToYoungRememberedRebuildState::new_skipping`, whose cursor never enters blocks the census recorded as holding no reached, pinned or pre-marked object (`BlockCensus::unmarked_blocks`); computing that list reads `arena_block_snapshots()` and allocates one `Vec` through the global allocator. It visits a subset of the same objects the rebuild already walked (every skipped object would have been rejected as unmarked), and it neither allocates a GC object, relocates anything, nor runs a JS callback. (b) In `step_sweep`, `IncrementalSweepState::with_block_skip` runs after `census_take_if_armed_at_full_sweep_start` has already taken PASS1_MARKED out of TLS. Neither boundary moved and the synchronous mark-complete to sweep-entry interval gains no relocation, collection or callback. Re-audited 2026-09-11 for the startup memory profile: gc/mod.rs only retains the pre-main allocator-policy constructor in js_gc_init. The constructor applies process allocation options, without invoking GC or JS. No census boundary, collector phase, or mark-complete to sweep-entry control flow changed. Re-audited 2026-09-13 for #10179: census.rs only adds a native regex cache metadata row and its unit assertion; snapshot consumption and the full-cycle window are unchanged. Re-audited 2026-09-14 for the GC due-check fast path, which touched `gc/mod.rs` and `gc/policy.rs`. `gc/mod.rs` only changes the safepoint re-exports: `gc_runtime_safepoint` becomes cfg(test) and `gc_runtime_safepoint_poll` is added. `gc/policy.rs`: the budgeted step returns a debt-free `GcStepReport` (debt is attached by the FFI and test entry points after the step returns) and moves cycle start/step into an out-of-line `gc_budgeted_start_or_step`; `gc_check_trigger` reuses a repeatable due-trigger answer through `DueTriggerMemo`, placed after its `GC_FLAG_IN_ALLOC` and suppression early returns; the young scavenge cap reuses the old-gen pressure value the due trigger already read and checks the census-seeded flag first. All of it runs from mutator safepoints, allocation-point trigger checks and the host step API, before a cycle starts or between budgeted steps. None of it is reachable between `census_pass1_if_armed` in `step_mark_propagation` and `census_take_if_armed_at_full_sweep_start` in `step_sweep` of a synchronous full: an allocation inside that window reaches `gc_check_trigger` with `GC_FLAG_IN_ALLOC` set and returns before the changed code. No allocation, relocation, collection or JS callback is added to the window. Re-audited 2026-09-14 for the tiny-parse nursery-cap boundary, which touched `gc/policy.rs`. It adds `tiny_parse_generational_collection_due`, a pure predicate (the existing `tiny_parse_pressure_due` OR the existing `young_scavenge_cap_due` read), and calls it instead of `tiny_parse_pressure_due` from `gc_bump_malloc_trigger_inner` and `gc_collect_pending_suppressed_parse_slow` (generational branch only) and from `gc_schedule_parse_boundary_collection_if_pressure`. All three are JSON.parse mutator-side boundaries, none reachable from `step_mark_propagation` or `step_sweep`; the predicate reads counters and allocates nothing. Neither census boundary nor the synchronous mark-complete to sweep-entry interval changed. Re-audited 2026-09-13 for #10182's full-collection throughput follow-up, which touched `gc/cycle.rs` in one hunk, INSIDE the window: the `RememberedSetRebuild` subphase of a synchronous full now first asks `verify::full_remembered_rebuild_provably_empty` and, when it holds, installs `OldToYoungRememberedRebuildState::provably_empty()` (an empty sticky set, no walk) instead of the require-marked rebuild. The predicate reads `arena_block_snapshots()` (one `Vec` through the global allocator), the census's per-block reached/pre-marked facts and the malloc registry's length; the constructor bumps a `Cell` counter and prints one line under `PERRY_GC_DIAG`. None of it allocates a GC object, relocates anything, collects, or runs a JS callback, and both census boundaries stay where they were. Re-audited 2026-09-14 for #10182's pacing-full work, which touched `gc/cycle.rs`, `gc/mod.rs` and `gc/policy.rs`. `gc/cycle.rs`: `GcCycleState::new_full` no longer calls `materialize_all_promoted_page_runs`; that call ran in the constructor, before the census and far before `census_pass1_if_armed`, and removing it adds nothing to the window. `gc/mod.rs`: one `mod promoted_cohort;` declaration. `gc/policy.rs`: (a) `credit_promoted_bytes_to_old_baseline` also credits a `Cell` cohort counter (it runs after a copying minor completes); (b) `finish_full_old_reclaim_baseline` also records the verified old live bytes and resets that counter (Publish, after `step_sweep` consumed the snapshot); (c) `gc_safepoint_moving_minor` arms and disarms the promotion-census record around its nursery minor and calls `run_promoted_cohort_full_if_due`, which starts a synchronous full through the same `gc_collect_full_mark_sweep_with_trigger` entry and reads byte counters before and after it. All of it runs before a cycle starts or after it completes; none of it runs between mark completion and sweep entry, allocates a GC object, relocates anything, or calls into JS. The census the promoted-cohort full may adopt from the promotion walk is built in `BuildValidPointerSet`, before either boundary. Both boundaries are unchanged. Re-audited 2026-09-14 for the #10182 dead-stack scrub in `gc/cycle.rs`: `step_build_valid_pointer_set` now calls `scrub_dead_stack_below`, which zeroes a local array in its own frame (dead stack below the caller), right after the census finishes — in `BuildValidPointerSet`, before the root scan and far before `census_pass1_if_armed`. It writes no heap memory, allocates nothing, relocates nothing and calls no JS; both boundaries are unchanged. Re-audited 2026-09-14 for #10241 (cohort survival), which touched `gc/cycle.rs` and `gc/policy.rs`. `gc/cycle.rs`: one call, `promoted_cohort::survival::check_minor_view_at_full_sweep_start()`, in `step_sweep` immediately AFTER `census_take_if_armed_at_full_sweep_start` has taken PASS1_MARKED out of TLS, i.e. outside the window. It is a no-op unless a promoted-cohort full armed its survival probe; when armed it walks the old page index over the preceding minor's dirty pages (`old_arena_walk_objects_on_pages`, Rust-allocator Vecs), reads GC headers' mark flags and the slots of unmarked ones, and records one enum. It writes no heap memory, allocates no GC object, relocates nothing and calls no JS. `gc/policy.rs`: `run_promoted_cohort_full_if_due` arms the probe before `gc_collect_full_mark_sweep_with_trigger` and takes it after the full returns (feeding `note_full_measured_promotion_survival` and one diagnostic line); both run before a cycle starts or after it completes. Both boundaries are unchanged. Re-audited 2026-09-14 for #10241's in-place-only cohort: `gc/policy.rs` drops the `promoted_cohort::note_promoted` call from `credit_promoted_bytes_to_old_baseline` (the copying minor now calls `promoted_cohort::note_minor_promotion` itself, after the credit). Both run at the end of a copying minor, outside any full cycle; the mark-complete to sweep-entry window is unchanged. Re-audited 2026-09-14 for the parse-boundary side-allocation band (medium-parse pacing), which touched `gc/policy.rs`. Three hunks: (a) a `Cell` thread-local (`GC_LAST_COLLECTION_EXTERNAL_SIDE_BYTES`, a byte COUNT, no pointer) plus three pure predicates over it and `external_side_live_bytes()`; (b) that predicate added as a third disjunct of `tiny_parse_generational_collection_due`, which is read only from the three JSON.parse mutator-side boundaries (`gc_bump_malloc_trigger_inner`, `gc_collect_pending_suppressed_parse_slow`, `gc_schedule_parse_boundary_collection_if_pressure`), none of them reachable from `step_mark_propagation` or `step_sweep`; and (c) one extra `Cell` store in `note_collection_finished_arena_occupancy` plus two extra reads in the `PERRY_GC_DIAG` tiny-parse line. `note_collection_finished_arena_occupancy` runs from `publish_reclaim_outcome` in the Publish subphase, i.e. AFTER `step_sweep` has already `take()`n the snapshot out of the thread-local, exactly as #9831's store on the same line does. Nothing added allocates a GC object, relocates anything, or runs a JS callback, and neither census boundary moved. Re-audited 2026-09-14 for the drained-bytes counterweight to that band, which touched `gc/policy.rs` again. Four hunks: a second `Cell` thread-local (`GC_EXTERNAL_SIDE_DRAINED_SINCE_FULL`, a byte COUNT); one increment of it inside `gc_note_external_side_free`; a pure read (`external_side_old_reclaim_pressure_bytes`) substituted for `external_side_live_bytes()` at the four old-reclaim pressure sites; and one `Cell` store at the top of `finish_full_old_reclaim_baseline`. None of it can run between the census boundaries. `gc_note_external_side_free` is also reached by mutator-side tape materialization, regex scratch teardown, native-addon adjustments and buffer replacement. Its added operation is only a saturating increment of a scalar Cell, with no GC allocation, relocation, collection or JS callback, so this wider caller set does not invalidate the census window. `finish_full_old_reclaim_baseline` runs from `publish_reclaim_outcome` in the Publish subphase, the same place #9831's store already sits. The pressure reads happen at trigger decisions, before a cycle starts. No allocation, relocation, collection or JS callback is added to the mark-complete -> sweep-entry window, and neither boundary moved. Re-audited 2026-09-16 after #10377 touched `gc/policy.rs`. Its hunks are one new function, `gc_note_external_side_free_transient`, which does a single saturating subtraction on the `GC_EXTERNAL_SIDE_LIVE_BYTES` Cell and deliberately omits the drained-term increment, plus the three regex scratch owners that now call it (`perex_memory`'s Buffer and Reservation drops, `perex_replace_direct`'s Spans drop). It is feature-gated to `regex-engine` because a product build has no caller. The change REMOVES an increment from a mutator free path and adds no allocation, relocation, collection or JS callback anywhere; it touches no mark/sweep control flow and neither window boundary moved. The regex drops it serves run in mutator code, never between mark completion and sweep entry.", "window": { "start": { "file": "crates/perry-runtime/src/gc/census.rs", @@ -331,7 +331,7 @@ "crates/perry-runtime/src/gc/census.rs": "5c151725460ffb92a55a6bee781123ef5159263b4ce5958d16570f78216e0d67", "crates/perry-runtime/src/gc/cycle.rs": "b035dcb44df029358cbab0afaa526e8e506765f5178034663257e18ceefaf9df", "crates/perry-runtime/src/gc/mod.rs": "9fedd2790f48154aaeceefb4805d3fbaa2fdf3c407529b326425fde86c2bf9a5", - "crates/perry-runtime/src/gc/policy.rs": "895c6f4bd1a6e491adf348ecfa89985b03e354fcee7cf73826bb590f9ace9163", + "crates/perry-runtime/src/gc/policy.rs": "27a0037e0c070059b36f3f14775599c23a77a2a3061c87f2d3ed6c36a253e06f", "crates/perry-runtime/src/gc/progress.rs": "a5ad3971bbe4047229ca57325234780daa85921dbc778e1c08dff4ad07ccfb96" } }