perf(codegen): refresh rooted arrays, gate super-scope, and count store safepoints

[proggeramlug]

Lands four reviewed PRs as one squash: #8670, #8673, #8668, #8678.

#8678 (#8583) is the notable one — it was the last blocker to a working cc binary. PropertySet/PropertyUpdate/IndexSet lower to collecting runtime calls that rewrite-statepoints-for-gc gives a statepoint, but count_safepoint_sites counted none of them. A closed-shape object literal compiles to a constructor that is one long run of this.field = v stores, so its estimate was ~0, it was never spilled to the shadow frame, and RS4GC grew one __AnonShape_*_constructor from 34,009 to 2,280,128 instructions — overrunning the #8586 per-function budget and refusing the whole module.

Reads (PropertyGet/IndexGet) are deliberately excluded: they frequently inline to a shape-cached load with no call, so counting them would over-spill read-heavy hot loops. That asymmetry is documented at the match arm.

Validation (on the merged 4-PR result, not per-branch)

• 9 ratchet gates + cargo fmt --all --check: pass (one fmt fix applied to perf(codegen): count property/index stores as GC safepoints in the spill estimate (#8583) #8678's own file)
• perry-codegen --lib: 1194 passed, 0 failed (+1 = perf(codegen): count property/index stores as GC safepoints in the spill estimate (#8583) #8678's property_and_index_stores_are_safepoints)
• perry-runtime --lib (RUST_TEST_THREADS=1): 2651 passed, 0 failed
• Squashed tree verified byte-identical to the merged tree (0 differing files) before push — the guard that merge: land #8652, #8651, #8647, #8646, #8650 #8657 lacked

Version bump stripped per maintainer policy; the Cargo.lock diff was confirmed version-only first. Four changelog.d/ fragments carried through.

Follow-up (not a blocker)

#8678 improves estimate accuracy; fully spilling that constructor also needs the spill threshold reconciled with the #8586 budget — #8623's 32M sits above the 1.57M budget. Tracked separately.

Summary by CodeRabbit

• Performance

  • Numeric indexing on stable Array subclasses is now significantly faster while preserving correct behavior for holes, accessors, proxies, and prototype changes.
  • Array length and index reads use optimized paths where applicable.

• Bug Fixes

  • Rooted-array iteration now remains accurate after storage relocation or length changes.
  • Safepoint reporting now includes collecting property and index stores.

• Improvements

  • Size reports better identify and separate Rust standard-library internal dependencies from duplicate crates.

[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: 68a8c8b4-d2d8-472c-ae43-063c1d1d1ee2

📥 Commits

Reviewing files that changed from the base of the PR and between f5739b5 and 2b3248f.

📒 Files selected for processing (16)

• changelog.d/8583-count-store-safepoints.md
• changelog.d/8668-array-subclass-indexing.md
• changelog.d/8670-refresh-rooted-arrays-iteration.md
• changelog.d/8673-report-size-std-internal-filter.md
• crates/perry-codegen/src/collectors/safepoint_sites.rs
• crates/perry-codegen/src/expr/index_get/inline_dyn_typed_array.rs
• crates/perry-codegen/src/runtime_decls/strings.rs
• crates/perry-runtime/src/array/iter_methods.rs
• crates/perry-runtime/src/array/mod.rs
• crates/perry-runtime/src/array/subclass.rs
• crates/perry-runtime/src/array/subclass_tests.rs
• crates/perry-runtime/src/object/mod.rs
• crates/perry-runtime/src/object/polymorphic_index.rs
• crates/perry-runtime/src/value/dynamic_object.rs
• crates/perry/src/commands/compile/size_report.rs
• crates/perry/tests/issue_8655_array_subclass_indexing.rs

---

📝 Walkthrough

Walkthrough

The PR adds guarded dense indexing for Array subclasses, refreshes rooted array iteration after relocation, counts collecting store expressions as safepoints, and separates Rust standard-library internal backtrace crates in size reports. It adds runtime, compiler, integration, unit, and changelog coverage.

Changes

Array-subclass indexing

Layer / File(s)	Summary
Array-subclass cache and runtime dispatch crates/perry-runtime/src/array/subclass.rs, crates/perry-runtime/src/array/mod.rs, crates/perry-runtime/src/object/mod.rs	The runtime caches dense Array-subclass layouts by class and shape. It adds guarded length and index accessors, packed-arraylike dispatch, and required re-exports.
Codegen inline-cache integration crates/perry-codegen/src/expr/index_get/inline_dyn_typed_array.rs, crates/perry-codegen/src/runtime_decls/strings.rs, crates/perry-runtime/src/object/polymorphic_index.rs, crates/perry-runtime/src/value/dynamic_object.rs	Code generation validates cached subclass facts and loads inline or spill slots. Runtime declarations and length/index paths use the packed-arraylike dispatcher before generic lookup.
Regression coverage and changelog crates/perry-runtime/src/array/subclass_tests.rs, crates/perry/tests/issue_8655_array_subclass_indexing.rs, changelog.d/8668-array-subclass-indexing.md	Tests cover dense reads, shape changes, rejected object brands, holes, accessors, proxies, and element-kind changes. Integration tests verify generated calls and fallback behavior.

Rooted array iteration

Layer / File(s)	Summary
Forwarded array refresh and validation crates/perry-runtime/src/array/iter_methods.rs, changelog.d/8670-refresh-rooted-arrays-iteration.md	Iteration refreshes forwarded pointers and roots, checks live length, and tests relocation with growth and shrinkage.

Safepoint counting

Layer / File(s)	Summary
Store safepoint classification crates/perry-codegen/src/collectors/safepoint_sites.rs, changelog.d/8583-count-store-safepoints.md	Property, property-update, and index stores count as safepoints. Property and index reads remain excluded. Tests cover these classifications and nested calls.

Size-report filtering

Layer / File(s)	Summary
Standard-library internal crate classification crates/perry/src/commands/compile/size_report.rs, changelog.d/8673-report-size-std-internal-filter.md	The report classifies known internal backtrace crates with symbol markers, excludes them from duplicate findings, and emits their bytes, symbol counts, and names separately. Tests cover detection and veto cases.

Estimated code review effort: 5 (Critical) | ~120 minutes

Sequence Diagram(s)

sequenceDiagram
  participant GeneratedGetter
  participant ArraySubclassCache
  participant PackedArraylikeDispatcher
  participant GenericPropertyLookup

  GeneratedGetter->>ArraySubclassCache: validate class, shape, length, and dense slot
  ArraySubclassCache-->>GeneratedGetter: return inline or spill value
  GeneratedGetter->>PackedArraylikeDispatcher: dispatch cache miss
  PackedArraylikeDispatcher->>GenericPropertyLookup: resolve holes, accessors, proxies, or rejected shapes
  GenericPropertyLookup-->>PackedArraylikeDispatcher: return generic property result

Loading

✨ 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 merge/b7

---

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.}