ZJIT: local macOS LLVM is newer than CI LLVM, which breaks bindgen CI

[tekknolagi]

This causes weird skew as in this CI run:

-pub const RB_INVALID_SHAPE_ID: _bindgen_ty_12 = 4294967295;
-pub type _bindgen_ty_12 = u32;
+pub const RB_INVALID_SHAPE_ID: _bindgen_ty_38 = 4294967295;
+pub type _bindgen_ty_38 = u32;

The temporary workaround is to only git add -p bindgen changes and omit this weird naming conflict. It is currently unknown what the long-term fix is.

[shopify-github-archiver]

⚠️ WARNING: You are commenting on a closed issue. ⚠️

Summary

• Takashi discovered that running make zjit-bindgen on macOS produces a diff, indicating an inconsistency with the CI environment on Linux.
• Takashi's investigation revealed that different Clang versions across macOS (both Apple and Homebrew) and Linux handle anonymous enums differently, leading to variations in the generated bindings.
• To resolve this, Takashi proposed naming the anonymous enum in zjit.c. This change is expected to make the generated code consistent across all operating systems.
• Takashi provided a sample diff to illustrate the necessary changes in zjit.c and the bindgen configuration.
• Max took on the task of implementing this solution and opened a pull request.

Links

Internal

• Ruby Pull Request #14802

This summary was generated using OpenAI's google:gemini-2.5-pro with a temperature of 0.5.

🧵 Slack Thread

User		Message
	Takashi Kokubun 2025‑10‑09 16:26	well, make zjit-bindgen LIBCLANG_PATH="/opt/homebrew/opt/llvm/lib" also shows the same diff, so I guess building clang on macOS causes it
	Max Bernstein 2025‑10‑09 16:27	wait are you saying it's a difference between apple clang and homebrew clang of the same version?
	Max Bernstein 2025‑10‑09 16:27	where homebrew clang matches CI?
	Takashi Kokubun 2025‑10‑09 16:27	no
	Takashi Kokubun 2025‑10‑09 16:27	"shows the same diff" means it has a diff. "shows no diff" is the expected behavior
	Max Bernstein 2025‑10‑09 16:28	oh
	Takashi Kokubun 2025‑10‑09 16:28	what I've checked is "apple clang 17" == "homebrew llvm" != "Linux clang 17/18"
	Max Bernstein 2025‑10‑09 16:29	😞
	Takashi Kokubun 2025‑10‑09 16:32	I think we could just give a name to the anonymous enum and import the enum type as a workaround. it should generate the same thing for these different OSes after that
	Takashi Kokubun 2025‑10‑09 16:34	my theory is that different OSes define different numbers of anonymous types on the same headers, so if you parse a C code on different OSes, you get different indexes for anonymous types.
	Max Bernstein 2025‑10‑09 16:34	oh that sounds like a great solution
	Max Bernstein 2025‑10‑09 16:34	I can tackle that if you are not already doing it
	Takashi Kokubun 2025‑10‑09 16:35	my patch right now is diff --git a/zjit.c b/zjit.c index d877c0bacb..f5fa7382e6 100644 --- a/zjit.c +++ b/zjit.c @@ -235,7 +235,7 @@ rb_zjit_print_exception(void) rb_warn("Ruby error: %"PRIsVALUE"", rb_funcall(exception, rb_intern("full_message"), 0)); } -enum { +enum zjit_foo { RB_INVALID_SHAPE_ID = INVALID_SHAPE_ID, }; diff --git a/zjit/bindgen/src/main.rs b/zjit/bindgen/src/main.rs index e1d19f9442..f8df3c39a9 100644 --- a/zjit/bindgen/src/main.rs +++ b/zjit/bindgen/src/main.rs @@ -290,7 +290,7 @@ fn main() { .allowlist_function("rb_zjit_insn_leaf") .allowlist_type("robject_offsets") .allowlist_type("rstring_offsets") - .allowlist_var("RB_INVALID_SHAPE_ID") + .allowlist_type("zjit_foo") .allowlist_function("rb_assert_holding_vm_lock") .allowlist_function("rb_jit_shape_too_complex_p") .allowlist_function("rb_jit_multi_ractor_p") diff --git a/zjit/src/cruby_bindings.inc.rs b/zjit/src/cruby_bindings.inc.rs index 2d8a8eb11e..8668a17cb6 100644 --- a/zjit/src/cruby_bindings.inc.rs +++ b/zjit/src/cruby_bindings.inc.rs @@ -722,8 +722,8 @@ pub const DEFINED_REF: defined_type = 15; pub const DEFINED_FUNC: defined_type = 16; pub const DEFINED_CONST_FROM: defined_type = 17; pub type defined_type = u32; -pub const RB_INVALID_SHAPE_ID: _bindgen_ty_38 = 4294967295; -pub type _bindgen_ty_38 = u32; +pub const RB_INVALID_SHAPE_ID: zjit_foo = 4294967295; +pub type zjit_foo = u32; pub const ROBJECT_OFFSET_AS_HEAP_FIELDS: robject_offsets = 16; pub const ROBJECT_OFFSET_AS_ARY: robject_offsets = 16; pub type robject_offsets = u32; but I'll leave it to you to give a better name 😆
	Max Bernstein 2025‑10‑09 16:36	sorry for causing so much hullaballoo for months 😓
	Max Bernstein 2025‑10‑09 16:37	github.com/ruby/ruby/pull/14802

---

Takashi Kokubun archived this conversation from ruby-jit at 2025‑10‑09 17:29.
All times are in UTC.