Rust thinks u64 and f64 have 4 byte alignment on Linux 32bit

[upsuper]

The following code:

use std::mem;
fn main() {
    println!("sizeof u64: {}, alignof u64: {}",
             mem::size_of::<u64>(), mem::align_of::<u64>());
    println!("sizeof f64: {}, alignof f64: {}",
             mem::size_of::<f64>(), mem::align_of::<f64>());
}

with

rustc test.rs --target=i686-unknown-linux-gnu

The generated executable outputs:

sizeof u64: 8, alignof u64: 4
sizeof f64: 8, alignof f64: 4

which matches neither gcc nor clang.

And this issue seems to block linux32-stylo.

[upsuper]

cc @emilio

[sfackler]

According to the i386 System V ABI, u64 and f64 are 4 byte aligned - see Table 2.1 in https://www.uclibc.org/docs/psABI-i386.pdf.

[emilio]

Yes, but C++ compilers seem to disagree with that, apparently, which causes trouble :(

[sfackler]

C++ doesn't have a stable ABI, it has "whatever gcc and clang decided to do" IIRC.

[emilio]

So, some relevant bugs:

• https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69763
• https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79024
• https://bugs.llvm.org/show_bug.cgi?id=26547

[nagisa]

Seems to match C in both gcc and clang just fine:

#include<inttypes.h>

#define alignof(type) (unsigned)((&((struct { char c; type d; } *)0)->d))

int main(void) {
    printf("%u %u %u %u\n", sizeof(uint64_t), alignof(uint64_t), sizeof(double), alignof(double));
    return 0;
}

where both output 8 4 8 4.

[upsuper]

Actually even in C++, something like

struct Test1 { char c; uint64_t u; };

would have the second field aligned to 4 bytes rather than 8 bytes, as can be found in my bug report to gcc.

So the behavior really is, when the type are the first field of a struct, or they are on themselves, they have a different alignment.

[upsuper]

OK, so I don't think this is really a valid issue. It is reasonable that u64 and f64 have 4byte alignment on 32bit Linux (actually they should).

There may be problem when passing u64 and f64 through FFI since if IIUC, the 8byte alignment is used for that. But that's a different issue.

So the real problem here is actually a mix of bindgen issue and lack of language support from Rust side when manual alignment (alignas/alignof) is used in C++.