perf(gc): stamp exact starts only for Maps

[proggeramlug]

Closes #8288

Summary

• remove object-start bitmap read/modify/write IR from inline Object and Array allocation
• stamp exact starts only for runtime-allocated Maps, the arena-resident 8-aligned tag that needs allocation-authored evidence
• reject malloc-only type descriptors during arena classification and retain the correct-size fabricated-Map guard
• shrink InlineArenaState back to its three hot bump-allocation fields

Performance

Clean quiet-M1-mini four-engine sweep, best-of-five interleaved, against 38cf15336 (same 19-source corpus):

workload class	instructions retired, geomean	peak RSS geomean
allocation class (12 rows)	-0.01% (was +3.69%)	-6.02%
prior -48.3% cohort (7 rows)	-48.90%	-5.46%

Peak observed Perry RSS is 404,520,960 B, down from 421,429,248 B at the baseline. The fixed branch is -3.56% instructions versus current main across the allocation class. Sweep verdict: CLEAN (load 1.02 -> 2.02, no foreign processes); all 19 Perry binaries passed output checks.

Tests

• cargo check -p perry-runtime --lib
• cargo check -p perry-codegen --lib
• cargo test -p perry-runtime --lib (2,569 passed, 4 ignored)
• cargo test -p perry-codegen --lib the_inline_allocator_stores_its_header_prefix_as_one_vector_image
• bash scripts/check_file_size.sh
• cargo fmt --all -- --check
• git diff --check

Summary by CodeRabbit

• Performance

  • Improved garbage-collection allocation efficiency by reducing unnecessary object-start tracking for ordinary objects.
  • Optimized Map allocation handling, with benchmark results showing improved allocation performance and lower peak memory usage.

• Bug Fixes

  • Strengthened arena object classification to reject invalid or incompatible allocation records.
  • Preserved accurate tracking for Map objects while preventing false classifications for arrays and other object types.

• Tests

  • Added coverage for allocation tracking, arena resets, and invalid object classification scenarios.

[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: 49147d8f-f39c-4d34-9001-ef6a85c2dc7a

📥 Commits

Reviewing files that changed from the base of the PR and between 93803e0 and 6d89092.

📒 Files selected for processing (14)

• changelog.d/8297-map-only-object-start-stamping.md
• crates/perry-codegen/src/expr/array_literal.rs
• crates/perry-codegen/src/lower_call/alloc_hot_tests.rs
• crates/perry-codegen/src/lower_call/new_alloc.rs
• crates/perry-runtime/src/arena/allocators.rs
• crates/perry-runtime/src/arena/block.rs
• crates/perry-runtime/src/arena/inline.rs
• crates/perry-runtime/src/arena/page_meta.rs
• crates/perry-runtime/src/arena/promote.rs
• crates/perry-runtime/src/arena/quarantine.rs
• crates/perry-runtime/src/arena/reset.rs
• crates/perry-runtime/src/arena/tests.rs
• crates/perry-runtime/src/gc/copying_pointer_set.rs
• crates/perry-runtime/src/gc/tests/copying/fabricated_map_rejection.rs

💤 Files with no reviewable changes (5)

• crates/perry-runtime/src/arena/quarantine.rs
• crates/perry-codegen/src/lower_call/new_alloc.rs
• crates/perry-codegen/src/expr/array_literal.rs
• crates/perry-runtime/src/arena/promote.rs
• crates/perry-runtime/src/arena/reset.rs

Included review availability: Your plan includes up to 8 reviews per rolling hour; 4 remain after this review.

---

📝 Walkthrough

Walkthrough

Changes

Arena allocation no longer stamps exact object starts for ordinary inline objects. Runtime allocation paths pass object types to the recorder, which records only Maps. Arena classification checks bitmap membership for Maps and rejects malloc-only descriptors. Tests and changelog entries reflect the new behavior.

Map-only arena object-start stamping

Layer / File(s)	Summary
Allocation and inline-state updates crates/perry-codegen/src/expr/array_literal.rs, crates/perry-codegen/src/lower_call/*, crates/perry-runtime/src/arena/*	Inline allocation stops writing object-start bitmap metadata. Arena allocation paths pass obj_type when recording starts. InlineArenaState and INLINE_STATE now contain only data, offset, and size.
Map classification and regression coverage crates/perry-runtime/src/arena/page_meta.rs, crates/perry-runtime/src/arena/tests.rs, crates/perry-runtime/src/gc/copying_pointer_set.rs, crates/perry-runtime/src/gc/tests/copying/fabricated_map_rejection.rs, changelog.d/8297-map-only-object-start-stamping.md	record_arena_object_start records only Maps. Classification checks bitmap membership only for Maps and rejects malloc-only descriptors. Tests cover Map stamping, non-Map omission, reset clearing, and fabricated headers.

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

Merge Risk: ⚪ Minimal · up to 6d890

This change removes unnecessary allocation bookkeeping while retaining Map stamping and classification guards; no actionable merge-blocking risk remains after normal checks and review.

Sequence Diagram(s)

sequenceDiagram
  participant InlineAllocator
  participant ArenaAllocators
  participant PageMeta
  participant PointerClassifier
  InlineAllocator->>ArenaAllocators: Allocate object
  ArenaAllocators->>PageMeta: Record header and object type
  PageMeta-->>ArenaAllocators: Stamp only Map starts
  PointerClassifier->>PageMeta: Check recorded Map start
  PageMeta-->>PointerClassifier: Accept or reject arena classification

Loading

Possibly related PRs

• PerryTS/perry#8277: Added the object-start bitmap behavior that this PR narrows to Map objects.
• PerryTS/perry#8251: Shares the plausible_gc_header and fabricated Map classification paths.
• PerryTS/perry#8196: Relates to arena object-start validation during GC pointer classification.

Suggested labels: performance

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and concisely describes the main change: limiting exact object-start stamping to Maps for GC performance.
Description check	✅ Passed	The description covers the summary, concrete changes, related issue, performance results, and test commands; omitted optional screenshots are not needed.
Linked Issues check	✅ Passed	The changes satisfy #8288 by implementing Map-only stamping, preserving fabricated-map rejection, reporting instructions and RSS, and meeting the stated benchmark criteria.
Out of Scope Changes check	✅ Passed	The code, tests, changelog, and benchmark details all support the linked issue objectives; no unrelated changes are evident.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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.}

[proggeramlug]

Merging. This closes #8288 and it meets the acceptance criterion I wrote there, including the part most perf PRs skip.

#8288 asked for: "the allocation-heavy geomean back under +1.0% instructions against 38cf15336, with no regression on the −48.3% cohort and no loss of the #8256 fixture. Report instructions and peak RSS together."

This delivers all four:

	#8288 measured	this PR
allocation class (12 rows)	+3.69%	−0.01%
prior cohort (7 rows)	−48.3%	−48.90%
peak RSS, allocation class	—	−6.02%
peak RSS, other cohort	—	−5.46%

Peak observed Perry RSS 421,429,248 → 404,520,960 B. Sweep verdict CLEAN, all 19 binaries output-checked. Reporting RSS alongside instructions is the standing rule and it is genuinely rare to see it honoured without being asked twice.

And the #8256 guard is intact — I checked, because "stamp fewer bits" is only correct if the thing the bits existed for still holds. fabricated_map_rejection still has all five tests passing, including the original test_classify_arena_rejects_interior_pointer_as_map, plus a new test_plausible_gc_header_rejects_malloc_only_type_in_arena covering the tightened arena classification. So the fabrication defence is stronger, not traded away.

Picking direction 1 from the ticket — restricting exact starts to runtime-allocated Maps, the arena-resident 8-aligned tag that actually needs allocation-authored evidence — is the right one: GC_TYPE_SET was never reachable (12 is not 8-aligned), so Object and Array allocations were paying for a defence against a fabrication they could not suffer. Shrinking InlineArenaState back to its three hot bump fields is the part that makes the inline path genuinely cheap again rather than merely shorter.

Verified here: perry-runtime --lib 2572/0/4; perry-codegen --no-fail-fast 8 failures, all in the known baseline, zero new (the first run stopped at 15 suites when the disk filled, so I ran the remaining four separately rather than report a partial sweep as clean); cargo fmt clean.