gc(macos): back Apple GC blocks with tagged 1 MiB regions

[proggeramlug]

Summary

On Apple platforms, GC arena blocks now come from their own anonymous mappings instead of mimalloc. They are 1 MiB, aligned to 1 MiB, and tagged VM_MEMORY_APPLICATION_SPECIFIC_2 (241) so vmmap/footprint tools show them separately from mimalloc (240). Linux (R2 reservation) and the other targets are unchanged. The regions stay in the existing non-Linux region registry, so classification and pooling are unchanged.

• Release and reactivation. Idle pool and Eden pages are released with MADV_FREE_REUSABLE, which drops them from the physical footprint immediately. They are reactivated with MADV_FREE_REUSE before reuse:
  • on every pool handoff;
  • in Arena::set_current for idle Eden blocks. Idle discard never touches the current block, and every block becomes current before its bytes are written, so the bump path carries no check.
• Block size. Blocks stay 1 MiB. A 2 MiB geometry was measured on Apple: it added 4–6 MiB of arena capacity and raised peak footprint, with no huge-page benefit to pay for it.
• mimalloc purge drain. Block allocation no longer passes through mimalloc, and mimalloc runs its deferred purges on its own allocation traffic. Freed collector and runtime memory then stayed dirty until the next full collection. An Effect workload peaked 5 MiB higher, with identical GC counts and arena sizes; MIMALLOC_PURGE_DELAY=0 removed the gap. The fix: when the heap maps a fresh region, run mi_collect(true) at most once per mimalloc's default purge delay (10 ms) per thread. PERRY_GC_MALLOC_PURGE=0 disables it together with the existing cycle purge, so this adds no new knob.
• Test-only fix. The first commit restores the build of a macOS-only fixture in value/dynamic_object.rs. BufferHeader's fields became private, and the fixture no longer compiled, so perry-runtime's test crate did not build on macOS at all.

This follows #12247, the codegen floor fix for *-apple-macosx* triples. Without that fix, every inline guard missed for objects below 2 TiB, where these regions land.

Measurements

The machine is an M1 Max on macOS 26.5, which is shared, so primary metrics are retired instructions and lifetime peak physical footprint from proc_pid_rusage. Runs are n=5, interleaved, and each timed run is preceded by an untimed run of the same binary. All builds are default auto-optimized, and every run matched Node 26.5.1 output.

Rebased branch vs main at a760b5497a; main-r is the same binary repeated, as the noise floor:

Program	Instructions	Peak footprint	main-r footprint
hello	−0.18%	+0.00 MiB	+0.00
TypeScript 5.8 compile	+0.06%	+0.11 MiB (within span)	+0.28
Zod ×5000	+0.01%	−0.11 MiB	−0.03
qs parse	+0.01%	−0.34 MiB	+0.02
qs stringify	−0.02%	−0.11 MiB	+0.02
commander	−0.01%	−0.19 MiB	+0.00
Fastify inject	−0.07%	−6.56 MiB (48.7 → 42.1)	+0.03
buffer_heavy	+0.07%	−0.44 MiB	+0.00
worker_heavy	+0.36%	+1.9 MiB (noise; thread spans 12–18 MiB)	+4.9
THP probe	−0.34%	−0.05 MiB	+0.00

Bold marks a difference beyond the larger same-binary span.

Effect. The Effect workload does not link on current main in either arm: duplicate _perry_birth_class/_perry_birth_empty_array across two modules, unrelated to this change. It was measured on the previous base (7a5584c12d):

• branch 106.17 MiB, spread 0.09;
• main 107.77 MiB, bimodal with spread 2.4;
• retired instructions −0.01%.

Before the purge drain, that comparison was +5.3 MiB.

RSS, a secondary metric, can be higher (buffer +3.7 MiB, Fastify +0.9 MiB) while footprint is lower. Reusable pages stay resident until the kernel needs them, but they are not charged to the process.

Tests

cargo test --release -p perry-runtime -- --test-threads=1 on macOS:

• branch: 5,677 passed, 11 failed;
• main with the fixture fix (needed for the test crate to build): 5,668 passed, 10 failed.

The ten failures are the same set on both. The eleventh on the branch is gc::tests::net_transport::n9_a_stale_resolve_leaves_the_new_plan_and_backlog_alone (WouldBlock on a test socket read). It is flaky on main too: it failed 2 of 3 isolated runs on main and 3 of 3 on the branch.

New tests:

• arena::region::apple_tests:
  • 1 MiB geometry;
  • regions are aligned, tagged 241 and registered/deregistered;
  • the release/reuse round trip changes the kernel's reusable-page state;
  • invalid extents are refused.
• arena::block::reuse_window_tests:
  • whole-region pool charge;
  • pool handoff reactivates pages;
  • an already-discarded block reactivates inside a new pool window;
  • an idle Eden block reactivates when it becomes current (bump path issues no syscall; forward scan and set_current both reactivate).
• gc::cycle_malloc_trim::region_growth_purge_tests: the purge is gated to once per purge delay, and fresh region maps reach the gate.

The pool-pressure fixture now seeds the pool with region backing instead of std::alloc.

No version bump.

[coderabbitai]

Note

Currently processing new changes in this PR. This may take a few minutes, please wait...

⚙️ Run configuration

• Configuration used: defaults
• Review profile: CHILL
• Plan: Advanced
• Run ID: 3b132829-4ed8-4c64-87cd-667c0fb972fc

---

📥 Commits

Reviewing files that changed from the base of the PR and between 15b6149 and c8f6e2e.

---

📒 Files selected for processing (10)

• crates/perry-runtime/src/arena/block.rs
• crates/perry-runtime/src/arena/block/reuse_window_tests.rs
• crates/perry-runtime/src/arena/mod.rs
• crates/perry-runtime/src/arena/region.rs
• crates/perry-runtime/src/arena/region/apple.rs
• crates/perry-runtime/src/arena/region/apple_tests.rs
• crates/perry-runtime/src/gc/cycle_malloc_trim.rs
• crates/perry-runtime/src/gc/mod.rs
• crates/perry-runtime/src/gc/tests/block_pool_pressure.rs
• crates/perry-runtime/src/value/dynamic_object.rs

---

 _______________________________________
< For a good time, call 1-800-COD-RABT. >
 ---------------------------------------
  \
   \   \
        \ /\
        ( )
      .( o ).

✨ Finishing Touches

📝 Generate docstrings

• Commit to this branch
• Create a new PR

---

🧪 Generate unit tests (beta)

• Commit to this branch
• Create a new PR

---

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

---

_{Comment @coderabbitai help to get the list of available commands.}