Another UB found using rustlantis

[FractalFir]

This program has a different result under GCC debug and GCC release.

2 1 //debug
2 -1 // release

fn dump_var(var2: usize, val2: usize, var3: usize) {
    println!("val2:{val2}")
}
#[inline(never)]
fn dump_var2(var2: usize, val2: usize, var3: usize) {
    println!("val2:{val2}")
}
fn fn6(arg: bool, arg2: usize) -> bool {
    dump_var(0, arg2, 0);
    dump_var2(0, !arg as usize, 0);
    true
}
fn main() {
    let mut _1 = false;
    let mut _3 = 2;
    let mut _4 = (&1, 1);
    loop {
        if _4.1 == 0 {
            if _4.1 == 1 {
                _1 = true;
            } else {
                _1 = false;
            }
        } else if fn6(_1, _3) {
            loop {
                _3 = _4.1;
                let mut _6 = 0;
                _6 = 0;
                let tmp2 = _6;
                if tmp2 == 1 {
                    _4.0 = &_4.1
                } else if tmp2 == 0 {
                    return;
                } else {
                    break;
                }
            }
        }
    }
}

It looks like the cast of val as i128 is nonsensical. Since val is a boolean value, one would expect it to be either 1 or 0, and not -1.

[FractalFir]

This issue is even weirder than I expected. For example, setting #[unsafe(no_mangle)] on fn6 makes the problem go away.

This does not change the generated GIMPLE in any way, besides just changing the function name.

@antoyo does setting no_mangle change anything else(ie. set some kind of GCC-side attribute that I can't see in the GIMPLE dump)?

[antoyo]

I do not think cg_gcc does anything in this case.
But the Rust doc mentions this:

Additionally, the item will be publicly exported from the produced library or object file, similar to the used attribute.

which might be the source of the "fix".

[FractalFir]

The "visibility" seems to stay the same, tough.

The main problem I have here is replicating the issue in C. If something not present in GIMPLE is needed for the issue to occur, I have no way of turning that GIMPLE into C(tried once, and the issue just was not there).

[antoyo]

Would the representation dumped by setting CG_GCCJIT_DUMP_TO_FILE=1 help here?

[FractalFir]

Oh, there is a difference now.

diff mangle.c no_mangle.c
320,321c320,321
< static bool
< _ZN6fp_bug3fn617h3865bc436d14d647E (bool param0, __uint64_t  __attribute__((aligned(8))) param1, restrict __int64_t  __attribute__((aligned(8))) * param2)
---
> extern bool
> fn6 (bool param0, __uint64_t  __attribute__((aligned(8))) param1, restrict __int64_t  __attribute__((aligned(8))) * param2)
398c398
<   ptrReturnValue3 = (_ZN6fp_bug3fn617h3865bc436d14d647E) (((bool)loadedValue1), ((__uint64_t  __attribute__((aligned(8))))loadedValue2), ((restrict __int64_t  __attribute__((aligned(8))) *)&stack_var_4));
---
>   ptrReturnValue3 = (fn6) (((bool)loadedValue1), ((__uint64_t  __attribute__((aligned(8))))loadedValue2), ((restrict __int64_t  __attribute__((aligned(8))) *)&stack_var_4));

I guess that no_mangle marks the function as extern(not static) and that allows GCC some more optimizations.