fix(runtime): arm64_32 ILP32 — box oversized thread_locals, guard GC pointer classifier

[proggeramlug]

What

The final runtime piece for building a watchOS arm64_32 app on main. Follows #6107 (target-aware object-header size), #6108 (boxed exception TLS), and #6109 (AVFAudio/mic linking) — all merged. Without this, the next arm64_32 app links but corrupts memory at runtime.

Why

On arm64_32 (ILP32: 64-bit registers, 32-bit pointers — Apple Watch Series 4–8 / SE), oversized #[thread_local] statics overflow the ILP32 TLS layout, and their writes silently corrupt adjacent thread-locals. Three large per-thread caches trip this:

static	size	file
INTERN_TABLE	~128 KB	string/intern.rs
TRANSITION_CACHE_GLOBAL	~320 KB	object/mod.rs
VTABLE_IC	~160 KB	object/class_registry/dispatch.rs

Fix: heap-allocate each via Box<[T]> so only a pointer lives in TLS. Confirmed on a real Series 7 — shrinking or boxing each table removes the corruption; boxing keeps full cache capacity.

Also (ILP32-only, cfg-gated)

• gc/barrier.rs — reject heap-word candidates whose tagged payload exceeds 32 bits, so a mistagged/immediate value can't truncate to a garbage 32-bit address the GC then marks or dereferences.
• lib.rs — use the system allocator (libsystem_malloc, solid on watchOS) on 32-bit targets; mimalloc on ILP32 is unproven and a corruption suspect. 64-bit keeps mimalloc.

Impact on 64-bit

None. Every new branch is gated to not(target_pointer_width = "64"), except the Box indirection, which is target-agnostic and free in the hot path (.with() still hands back a *mut [T; N]). cargo check -p perry-runtime is clean on host.

Provenance

Extracted from the dbmeter watch app's device-test branch — the clean runtime fixes only. App-specific debugging scaffolding (a game-thread runner, a data-container trace logger) was deliberately left out.

Summary by CodeRabbit

• Bug Fixes
  • Improved 32-bit compatibility by preventing invalid pointer/address decoding and truncation of non-pointer payload bits.
  • Reduced instability on 32-bit targets by moving large per-thread caches (vtable, transition, string interning) from fixed TLS arrays to safer heap-backed storage.
  • Updated global allocator selection to use mimalloc only on 64-bit targets, with a System fallback on 32-bit.

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 4d98b94f-4737-4840-8dc7-be996d1c55d4

📥 Commits

Reviewing files that changed from the base of the PR and between 1cf73c1 and 1061eb4.

📒 Files selected for processing (1)

• crates/perry-runtime/Cargo.toml

---

📝 Walkthrough

Walkthrough

This PR adds ILP32-specific pointer decoding guards, gates the global allocator and mimalloc dependency by pointer width, and converts several thread-local caches from inline arrays to boxed slices to avoid oversized TLS storage.

Changes

ILP32 TLS layout and pointer decoding fixes

Layer / File(s)	Summary
Pointer candidate decoding guards crates/perry-runtime/src/gc/barrier.rs	Adds ILP32 range checks rejecting payloads exceeding 0xFFFF_FFFF in both the tagged-pointer decode path and the raw-pointer-shaped fallback path before casting to usize.
Global allocator selection by pointer width crates/perry-runtime/Cargo.toml, crates/perry-runtime/src/lib.rs	Uses mimalloc::MiMalloc only on 64-bit targets, falls back to std::alloc::System on non-64-bit targets, and gates the mimalloc dependency to 64-bit pointer-width builds.
VTable inline cache TLS boxed storage crates/perry-runtime/src/object/class_registry/dispatch.rs	Converts VTABLE_IC from an inline TLS array to a heap-allocated boxed slice, updating lookup and insert to double-dereference the cell.
Transition cache TLS boxed storage crates/perry-runtime/src/object/mod.rs	Converts TRANSITION_CACHE_GLOBAL from an inline TLS array to a boxed heap slice, casting the pointer back to the fixed-size array type in with_transition_cache.
String intern table TLS boxed storage crates/perry-runtime/src/string/intern.rs	Converts INTERN_TABLE from an inline TLS array to a boxed heap slice, casting the pointer back to the fixed-size array type in with_intern_table.

Estimated code review effort: 3 (Moderate) | ~25 minutes

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Description check	⚠️ Warning	The description is detailed, but it does not follow the required template and omits Summary, Changes, Related issue, Test plan, and Checklist sections.	Rewrite the PR description using the repository template, including Summary, Changes, Related issue, Test plan, and Checklist sections.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly summarizes the main ILP32 runtime fix: boxing oversized TLS and guarding GC pointer classification.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/arm64_32-box-oversized-thread-locals

---

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.}

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@crates/perry-runtime/src/lib.rs`:
- Around line 26-31: The global allocator selection in `GLOBAL` already falls
back to `std::alloc::System` on non-64-bit targets, but `mimalloc` is still
being pulled in unconditionally. Update the `crates/perry-runtime` dependency
setup so `mimalloc` is declared only for 64-bit targets, keeping the allocator
choice in `lib.rs` aligned with a target-specific dependency and avoiding
unnecessary inclusion on other architectures.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 8e455183-7696-4b0a-8c52-f01ddd7366e4

📥 Commits

Reviewing files that changed from the base of the PR and between 1c5383b and 1cf73c1.

📒 Files selected for processing (5)

• crates/perry-runtime/src/gc/barrier.rs
• crates/perry-runtime/src/lib.rs
• crates/perry-runtime/src/object/class_registry/dispatch.rs
• crates/perry-runtime/src/object/mod.rs
• crates/perry-runtime/src/string/intern.rs

[proggeramlug]

Addressed in 1061eb4 — moved mimalloc into a [target.'cfg(target_pointer_width = "64")'.dependencies] section so it matches the #[cfg(target_pointer_width = "64")] allocator selection in lib.rs. On arm64_32/ILP32 mimalloc is now neither used nor compiled. Verified: cargo tree -i mimalloc still resolves on 64-bit; cargo check -p perry-runtime clean.