perf(gc): align barrier gate load ordering

[proggeramlug]

Summary

• emit every generated PERRY_INCREMENTAL_MARK_BARRIER_ACTIVE_COUNT gate as an LLVM monotonic load, matching Rust's Relaxed readers
• route the remaining shadow-stack runtime gate through the shared relaxed helper
• document why the counter is authoritative state rather than a publication fence, and correct the previously reversed enable/disable ordering descriptions
• pin all three generated gate families and the live armed-barrier invariant at the relaxed ordering

Why Relaxed is sufficient

The counter does not publish the ValidPointerSet or any other memory. It only decides whether the current thread should pay for a TLS read and barrier call.

Arming increments the counter before installing the current thread's TLS pointer. Disarming clears that pointer before decrementing the counter. A later relaxed load on an armed thread therefore cannot observe a counter value preceding its own increment, and the count cannot return to zero until that thread removes its contribution. A thread with a null TLS pointer may conservatively observe either value: skipping the call on zero is equivalent to the call reading null and returning. There is no cross-thread data for an acquire load to publish.

Validation

• pre-fix reproduction: both root-gate IR tests failed because the generated load was seq_cst
• cargo test -p perry-codegen: passed in full
• cargo test -p perry-runtime: 2,194 passed, 0 failed, 4 ignored
• post-rebase focused ordering suite on current main: 5 passed, 0 failed
• cargo fmt --all -- --check
• bash scripts/check_file_size.sh

Performance

Built separate compiler/runtime pairs for base and fix and compiled gc-handoff/apps/interp.ts with each. Both arms printed 1708840, exited 0, and produced the same 13,709,320-byte executable size.

Machine code changed as intended: whole-binary ldar count 779 -> 44 and ldr count 229,093 -> 229,828 (735 matching replacements).

On the quiet M1 mini (load 1.60-1.99, no foreign benchmark/build processes), two order-reversed 31-pair sweeps were neutral:

arm	best	median
base	0.6277-0.6280 s	0.6288 s
fix	0.6276 s	0.6284 s

Instructions were unchanged, as expected for one instruction replacing one instruction; best cycles moved -0.14%. This resolves the reader/documentation disagreement without a measurable regression, but the profiled ldar was not a standalone speedup.

Closes #7918

Summary by CodeRabbit

• Performance

  • Reduced synchronization overhead for incremental garbage-collection barriers by using lighter-weight atomic checks.
  • Replaced 735 unnecessary stronger loads in the interpreter without measurable runtime impact.

• Bug Fixes

  • Improved consistency between generated barriers and runtime barrier state handling.

• Tests

  • Expanded coverage to verify barrier ordering, atomicity, alignment, and all generated barrier variants.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 495d5b33-1e77-4620-b1fe-c617895cf23b

📥 Commits

Reviewing files that changed from the base of the PR and between f110261 and 536ce35.

📒 Files selected for processing (10)

• changelog.d/7935-barrier-gate-ordering.md
• crates/perry-codegen/src/expr/class_field_barrier_tests.rs
• crates/perry-codegen/src/expr/shadow_inline.rs
• crates/perry-codegen/src/expr/shadow_slot.rs
• crates/perry-codegen/src/expr/write_barrier.rs
• crates/perry-codegen/tests/shadow_slot_hygiene.rs
• crates/perry-runtime/src/gc/barrier/mod.rs
• crates/perry-runtime/src/gc/roots/shadow_stack.rs
• crates/perry-runtime/src/gc/tests/inline_generation_gate_contract.rs
• crates/perry-runtime/src/gc/tests/shadow_stack_ops.rs

---

📝 Walkthrough

Walkthrough

The change aligns incremental-barrier active-counter reads across runtime and generated code. Runtime readers use relaxed ordering, generated LLVM loads use monotonic ordering, and tests verify the atomic load contract across barrier gate families.

Changes

Incremental barrier ordering

Layer / File(s)	Summary
Runtime counter contract crates/perry-runtime/src/gc/barrier/mod.rs, crates/perry-runtime/src/gc/roots/shadow_stack.rs, crates/perry-runtime/src/gc/tests/*	The runtime exposes incremental_mark_barrier_globally_idle within the crate, documents counter ordering, delegates root-shading checks to the helper, and updates runtime tests to use Relaxed loads.
Generated barrier gate ordering crates/perry-codegen/src/expr/*barrier*, crates/perry-codegen/tests/shadow_slot_hygiene.rs, changelog.d/7935-barrier-gate-ordering.md	Generated barrier gates use LLVM monotonic loads. Tests verify atomicity, alignment, and ordering across gate families. The changelog records the ordering changes and benchmark observations.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Possibly related PRs

• PerryTS/perry#7079: Modifies the same shadow-stack barrier gating and active-count checks.
• PerryTS/perry#7088: Introduces the affected shadow_inline.rs and shadow_slot.rs barrier paths.
• PerryTS/perry#7602: Modifies incremental-mark barrier gating in write_barrier.rs and related contract tests.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and concisely describes the main change: aligning garbage-collection barrier gate load ordering.
Description check	✅ Passed	The description covers the summary, rationale, changes, related issue, validation, performance results, and relevant checklist information.
Linked Issues check	✅ Passed	The changes align codegen and runtime barrier-counter loads with relaxed ordering and address the documentation and consistency objectives in [#7918].
Out of Scope Changes check	✅ Passed	The changes remain focused on barrier gate ordering, related documentation, tests, and the corresponding changelog entry.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/7918-barrier-active-ordering

---

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.}