fix(codegen): treat LLVM macOS and visionOS triples as low-address heap targets

[proggeramlug]

Problem

perry-codegen/src/target_layout.rs::heap_addr_lower_bound_inclusive sets the lowest address that codegen's inline heap guards will dereference. It recognises "mainstream" targets by substring, but codegen passes LLVM triples rather than rustc ones:

• the macOS host default is arm64-apple-macosx15.0.0 (codegen/helpers.rs::default_target_triple, and --target macos), which contains neither darwin nor ios;
• visionOS is arm64-apple-xros1.0 (--target visionos), which doesn't contain visionos.

Both fall to the conservative 2 TiB floor instead of the 1 MiB handle-band floor that the runtime's is_valid_obj_ptr uses on those OSes. Every inline guard built on that floor sends any GC object below 2 TiB to the runtime slow path. That includes class-method direct calls and learned receiver words, the #620 own-method override probe, composed property ICs, inline typed-array/buffer reads and versioned indexed loops.

On macOS today this is mostly hidden: mimalloc places its arenas above 2 TiB, so GC objects happen to clear the floor. Anything that puts GC memory below 2 TiB pays the slow path on every call. That includes OS-backed GC regions, which land near 4 GiB on macOS, and libsystem-malloc allocations on visionOS.

Evidence

The bug was found while investigating why Perry-owned mmap GC regions regressed on macOS (portable region branch vs #12132, same generated program code).

• Commander's retired instructions rose by +14% and Zod's by +7%.
• sample profile on symbolized builds: js_object_get_own_field_or_undef went from 0% to 7.6% of leaf samples on Commander, with learn_absent_method_word and from_utf8 appearing.
• lldb call counts, 300 Commander iterations: js_object_get_own_field_or_undef_learn was called 24,300 times vs 42 on gc: Perry-owned memory for GC arena blocks (2 MiB OS regions, huge pages by complete units) #12132. js_native_call_method_by_id_learn was called 7,200 vs 2 times. On Zod (50 iterations) it was 1,850 vs 60.
• A per-call trace showed the site had already learned exactly the receiver's (class_id, ShapeId) word and the header was a plain GC_TYPE_OBJECT. The only failing input was the address gate (0x1_05a7_74e8 < 2 TiB; on mimalloc the same receiver sat at 0x34c_d52a_74e8).
• With this fix the counts return to 42 / 2 / 60. Commander's regression drops from +14% to +0.5%, and Zod's from +7% to +1.1%. The remainder was a separate runtime issue on that branch, not part of this PR.
• On gc: Perry-owned memory for GC arena blocks (2 MiB OS regions, huge pages by complete units) #12132 with mimalloc, the fix is neutral (interleaved, n=3): Commander retired 6,999–7,039M instructions before the fix and 6,990–7,026M after; Zod 13,627–13,645M before and 13,629–13,676M after.

Change

• heap_addr_lower_bound_inclusive also matches macos and xros.
• New test heap_address_floor_covers_emitted_triples checks the floor for the triples codegen really emits: default_target_triple() and resolve_target_triple() for every public --target name. HarmonyOS keeps the 2 TiB floor, as before.
• Negative control: with the two new arms removed, the test fails on arm64-apple-macosx15.0.0.

Linux, Android, Windows, iOS, tvOS, watchOS and WASI triples already matched, so their generated code is unchanged.

Validation

• cargo test --release -p perry-codegen --lib: 2,104 passed, 0 failed, 1 ignored. The new test is included.
• rustfmt --check is clean.

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: 9af3b760-17fe-4cc4-bdda-663745d7cd61

📥 Commits

Reviewing files that changed from the base of the PR and between 484345a and 283cdb9.

📒 Files selected for processing (1)

• crates/perry-codegen/src/target_layout.rs

 ________________________________________________________________
< This is the kind of code that makes on-call believe in ghosts. >
 ----------------------------------------------------------------
  \
   \   \
        \ /\
        ( )
      .( 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

• Autopilot · 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.}