fix(codegen): declare a hidden-visibility global across codegen units

[proggeramlug]

Symptom

Compiling OpenCode v1.18.30 on Linux x86_64 with main aborts in codegen (exit 134):

thread 'perry-module-1' panicked at crates/perry-codegen/src/module.rs:1083:25:
cannot form external declaration for generated global:
@perry_closure_opencode_node_modules__bun_effect_4_0_0_beta_83_node_modules_effect_src_Schema_ts__247$info
  = hidden constant { ptr, i16, i16, i32, i32, i32, ptr, i64, ptr, i32, i16, i16, i64 } { ... }

Cause

When a module is split into codegen units, a global defined in one unit and referenced from another needs an external declaration in the referencing unit. external_decl_for_global builds it, but it only stripped linkage keywords (private, internal, linkonce_odr, …) before expecting constant or global. Closure-info records are emitted as @…$info = hidden constant {…} (fn_info.rs, whose own tests assert that form) — hidden is a visibility keyword, so nothing matched, the function returned None, and module.rs:1083 panicked.

It is Linux/COFF-specific: on ELF and COFF generated globals are not replicated, so every cross-unit reference goes through this function. Mach-O replicates globals instead (promote_global_for_units), which is why the same OpenCode build never hit this on macOS. It needs a module big enough to split and a cross-unit reference to a closure-info record — effect's Schema.ts is both.

Fix

Parse the optional preemption (dso_local / dso_preemptable) and visibility (default / hidden / protected) keywords in LLVM's order after the linkage, and keep them on the declaration, so hidden constant {…} declares as external hidden constant {…}. The existing thread-local handling (#10399) is unchanged and composes with it; unnamed_addr / local_unnamed_addr describe only the definition and are dropped as before. Anything still unrecognised keeps returning None, so a genuinely unexpected form stays loud.

A scan of every @global = … prefix the codegen emits shows they are all declarable after this change: linkage forms are stripped, external … returns early, TLS and unnamed_addr were already handled, and hidden constant is the only visibility form emitted. Nothing emits addrspace or DLL storage on a global.

Test

module::tests::external_decl_keeps_hidden_visibility_of_closure_info covers the closure-info shape, dso_local hidden global, and hidden thread_local global.

• Without the fix it fails with the production symptom: left: None, right: Some("@perry_closure_m__247$info = external hidden constant { ptr, i16, i16 }").
• With the fix, all 26 perry-codegen module:: tests pass (--test-threads=1), including the existing external_decl_keeps_thread_local.

cargo fmt --all -- --check is clean on this head. The changelog fragment follows in a second commit once this PR has a number.

Summary by CodeRabbit

• Bug Fixes
  • Fixed a compiler failure on Linux/COFF when building projects that use cross-unit closure information. Hidden symbols are now handled without triggering a panic, while retaining their visibility and preemption behavior.

[coderabbitai]

Navigate logical layers of code changes, visualize relationships, and explore their blast radius.

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: 01f87c9e-f9c8-4547-adc7-292f5d1b22d0

📥 Commits

Reviewing files that changed from the base of the PR and between f9a65b4 and 64910e0.

📒 Files selected for processing (3)

• changelog.d/11817-codegen-hidden-visibility-decl.md
• crates/perry-codegen/src/module/linkage.rs
• crates/perry-codegen/src/module/tests.rs

 ______________________________
< torvalds@linux:~$ git review >
 ------------------------------
  \
   \   (\__/)
       (•ㅅ•)
       / 　 づ

✨ 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.}

[proggeramlug]

The four red jobs are pre-existing on main and outside this diff, which touches only perry-codegen/src/module/{linkage,tests}.rs plus the fragment:

job	failure	file
self-test-checkers	thread-local policy: 2 raw thread_local! declarations	perry-runtime/src/object/field_get_set/class_object_template.rs — on main at lines 86/105/455, from 9b9ba6bf07; #11781 fails the same job
lint	heap-tag-only finding, "self-test FAILED: the tree is already red"	perry-runtime/src/json/parse_api.rs:749
warnings	-D warnings	perry-stdlib/src/zlib.rs:1879 / :1881
cargo-test	codegen_env_vars_are_build_cache_inputs (1222 passed, 1 failed)	perry/src/commands/compile/build_cache.rs

The test this PR adds, and the rest of perry-codegen's module:: tests, are not among the failures.