Floating point number regression (LLVM-67)

[lexxvir]

Hi,

I have encountered into the regression in the current xtensa_release_9.0.1 branch.
Previously I used GCC based branch and it works fine.
Worth noting that I also changed the Rust version, so it may be related.

To reproduce the issue I have prepared minimal version of Rust code:

#[no_mangle]
#[inline(never)]
fn test_float_cmp() {
    let mut f: f32 = 10.0;
    let mut p = Printf {};

    while f > 0.0 {
        // works fine
        // let _ = writeln!(p, "{:04x}", f.to_bits());

        // works bad
        let _ = writeln!(p, "{}", f);

        f -= 0.1;
    }
}

• "works bad": loop produces only one value 10.0 and exits
• "works fine": loop produces 100 values

Some comments for the code:

• "bad" case formats f32 to human readable string
• "good" case formats u32 to hexadecimal number
• to_bits converts f32 to u32 as is at bits representation
• Printf is thin wrapper around regular C's `printf:

struct Printf;
impl Write for Printf {
    fn write_str(&mut self, s: &str) -> Result {
        for c in s.as_bytes() {
            unsafe { printf("%c\0".as_ptr() as *const _, *c as i32) }
        }
        Ok(())
    }
}

I've attached archive with: Rust source code demo, disassembled code for "good" and "bad" cases, final binary and LLVM bitcode files:

float-issue.zip

P.S. Actually I use https://github.com/mabezdev/llvm-project fork, but now it contains only one commit ahead of this branch.

[andreisfr]

Hi, @lexxvir ! Could you check this example using new branch xtensa_release_11.0.0, please? We corrected some FP comparison operations

[lexxvir]

Hi @andreisfr

I've checked xtensa_release_11.0.0 branch and confirm that the issue is fixed! Thanks for the support!

[andreisfr]

Hi, @lexxvir , thank you very much for your help!