Incorrect codegen with unsized types on 32-bit Windows

[alexcrichton]

This seems like quite a mysterious bug to me, the code below will compile and run successfully on 32-bit OSX/Linux as well as 64-bit OSX/Linux/Windows. It will fail, however, on 32-bit Windows (both GNU and MSVC toolchains). I'm not 100% sure that this is actually a codegen issue, but it's the only thing I can think of unfortunately.

Specifically, the assertion below will fail with some unusually corrupt value in self.0 (e.g. it changes over time).

struct Foo(u64);
struct Bar<T: ?Sized>(u32, T);

trait Trait {
        fn foo(&self);
}

impl Trait for Foo {
        fn foo(&self) {
                assert_eq!(self.0, 0);
        }
}


fn main() {
        let cursor = Foo(0);
        let ar = Bar(32, cursor);
        let ar: &Bar<Trait> = &ar;
        ar.1.foo();
}

This may just be a dupe of #26403, but the failure only on 32-bit Windows seems quite suspicious to me. This is reduced down from the test suite of tar-rs segfaulting on windows.

[alexcrichton]

Aha, looks like this is a dupe of #26403. On Windows the alignment of u64 is 8, but on Linux it's 4 which is why I only saw it on Windows and why it didn't show up for 32-bit Linux. On 64-bit linux the above test also fails (because the align of Foo is higher than that of the first field in Bar).