Possible ARM bug

[FractalFir]

This program behaves differently with optimizations enabled on ARM:

use core::ffi::*;
extern "C" {
    pub fn printf(format: *const c_char, ...) -> c_int;
}
fn dump_var(val3: f32) {
    if val3 != 1902660900.0 {
        unsafe { printf(b"wrong\n\0".as_ptr() as *const _) };
    }
}
fn main() {
    let mut _9 = 43619u128;
    let _64 = &mut _9;
    let _27 = _9 as f32;
    let _29 = 43619.0 * _27;
    dump_var(_29 + 43619.0)
}

Unoptimized: 

Optimized:
wrong

(I am using printf here to make porting the code to C easier)

[antoyo]

This is probably not the cause of this bug, but since you mentioned twice issues around floats on ARM, I prefer to mention that while libgccjit supports the Float32 type on x86, it doesn't support it yet on other platforms.

[FractalFir]

It looks like the type needs to be 128 bits wide to trigger the issue. I remember you doing some kind of fix to 128 bit casts. Could that be related?

[antoyo]

The fix was merged recently, so you can try it out.

[FractalFir]

Nope, broken on newest master :(

[antoyo]

Ok. Don't hesitate to ask me for help if needed.