Expected bool, found integral variable (0.30 -> 0.31 Regression)

[zacps]

Regression between the 0.30 and 0.31 releases that causes an invalid type (int instead of bool) in generated bindings.
Platform: MSVC x86_64

Input C/C++ Header

winpty.h
I unfortunately wasn't able to get creduce working on windows so unreduced is the best I can do for now.

Bindgen Invocation

  let bindings = bindgen::Builder::default()
        .header("src/include/winpty.h")
        .clang_arg("-x")
        .clang_arg("c++")
        // Enabling this causes a different error, probably rustfmt's fault
        .rustfmt_bindings(false)
        .generate()
        .expect("Unable to generate bindings");

Actual Results

error[E0308]: mismatched types
 --> C:\Users\Zac\Programming\c++\winpty\target\debug\build\winpty-sys-7016778a370aac5f\out/bindings.rs:5:217
  |
5 | } # [ repr ( C ) ] # [ derive ( Debug , Copy , Clone ) ] pub struct __vcrt_va_list_is_reference { pub _address : u8 , } pub const __vcrt_va_list_is_reference___the_value : __vcrt_va_list_is_reference__bindgen_ty_1 = 0 ; pub type __vcrt_va_list_is_reference__bindgen_ty_1 = bool ; pub type __vcrt_bool = bool ; extern "C" {
  |                                                                                                                                                                                                                         ^ expected bool, found integral variable
  |
  = note: expected type `bool`
             found type `{integer}`

Expected Results

Compilation succeeds

[fitzgen]

Thanks for the bug report!

[emilio]

Self-reminder to take a look at this one

[emilio]

Arrgh, this includes windows.h, and I'm on a Linux system.

@zacps, any chance to call .dump_preprocessed_input before generate() and attach it here? That way I can hopefully repro it.

[zacps]

Here.
Gist may not have been the best medium for this, given it's 200k lines but it appears to work for me. Tell me if you need a reupload.

[emilio]

Thanks!

Here's a reduced test-case out of that (probably can be reduced even further):

template <typename _Ty>
struct __vcrt_va_list_is_reference
{
    enum : bool { __the_value = false };
};

template <typename _Ty>
struct __vcrt_va_list_is_reference<_Ty&>
{
    enum : bool { __the_value = true };
};

template <typename _Ty>
struct __vcrt_va_list_is_reference<_Ty&&>
{
    enum : bool { __the_value = true };
};

template <typename _Ty>
void __vcrt_va_start_verify_argument_type() throw()
{
    static_assert(!__vcrt_va_list_is_reference<_Ty>::__the_value, "va_start argument must not have reference type and must not be parenthesized");
}

You should be able to work around this blacklisting the relevant bits, but still this is a bug of course.

[emilio]

This is probably due to us not handling enum : bool... It's rad that that's allowed, I didn't even know about that.

[emilio]

Aand indeed, more reduced test-case:

enum MyEnum : bool {
  Value = true,
};

[emilio]

And now the regression range makes sense, because it contains 89915f9.

[emilio]

Before that we'd generate instead:

#[repr(u8)]
#[derive(Debug, Copy, Clone, PartialEq, Eq, Hash)]
pub enum MyEnum {
    Value = 1,
}

[emilio]

Submitted a fix in #1232.

[zacps]

Amazing, thanks!