stylo's build script crashes inside mangling with "(signal: 11, SIGSEGV: invalid memory reference)" on macOS

[upsuper]

See bug 1335302.

After adding "--stdlib=libc++", "--target=x86_64-apple-darwin" to clang args in build_gecko.rs, the bindgen does not report any compile error anymore, but it crashes with the stack I pasted in bug 1335302 comment 3.

I saw #439, and also tested upgrading to clang-sys 0.14.0, but it doesn't seem to help. The build script still crashes at the same location.

It seems to crash when trying to get mangling name of __find_exactly_one_checked::__matches at /usr/local/Cellar/llvm/3.9.0/include/c++/v1/tuple:1018:25. But I failed to find minimal reproducible code...

[upsuper]

@emilio Could you have a look at this? This is probably not really a bindgen bug... It could be a bug in libclang. I have no idea how can we workaround it.

[upsuper]

The issue here is that it somehow doesn't treat __find_exactly_one_checked as a template, hence the cursor.is_in_non_fully_specialized_template() check returns false and consequently we don't correctly skip mangling __find_exactly_one_checked::__matches.

Cursor::is_template's documnet is "Is the referent a template specialization?", which probably shouldn't be named is_template. I guess, maybe we should use is_template_like or something in is_in_non_fully_specialized_template?

[upsuper]

Actually this issue is reproducible with the following code:

template <class... Args>
struct Test {
  static constexpr bool x[] = {Args::x...};
};

[emilio]

Thanks for the testcase! I'll find a workaround ASAP.

(For the record, Xidorn debugged it and it's a bug in LLVM, I think he submitted a patch?)

[upsuper]

It is not necessarily a bug in clang. There is no way in clang itself to hit that condition, and it seems libclang doesn't have no-crashing guarantee for invalid input.

I tried hard to write a test for that issue last night, but failed, so I'm not going to submit a patch for it.