x86_64 simd fmadd (and other) bug (release only)

[kali]

This bit of code falls in the assert in Release, while working in Debug.

Both stable and nightly.

https://play.rust-lang.org/?version=stable&mode=release&edition=2018&gist=6f280db2aaf8241e46a080f4189e0328

use std::arch::x86_64::*;

fn main() {
        unsafe {
            let a = [2.0f32, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0];
            let b = -1.0f32;

            let ar = _mm256_loadu_ps(a.as_ptr());
            println!("ar: {:?}", ar);
            let br = _mm256_set1_ps(b);
            println!("br: {:?}", br);
            let mut abr = _mm256_setzero_ps();
            println!("ab: {:?}", abr);
            abr = _mm256_fmadd_ps(ar, br, abr);
            println!("ab: {:?}", abr);
            
            let mut ab = [0.0; 8];
            _mm256_storeu_ps(ab.as_mut_ptr(), abr);
            assert_eq!(ab[0], -2.0f32);
    }
}

[kali]

might be related to #50154

[kali]

Problem was discovered on a skylake mac, if it helps.

[peterjoel]

Perhaps not discovered sooner because the function isn't being used much. I found it in two Github repositories:

• https://github.com/jackmott/simdeez
• https://github.com/bluss/matrixmultiply

[peterjoel]

Problem is also present on my i7-4960HQ Mac.

[kali]

Problem also occurs with simple multiplication.

https://play.rust-lang.org/?version=stable&mode=debug&edition=2018&gist=affc00bf05919f648a5c4b3e1e053070

[lqd]

Yeah, it doesn't look like the problem lies with the functions themselves, like this simple 2.0 * 1.0 multiplication shows.

So it could indeed be an instance of the linked LLVM by-ref args promotion issue ?

[BurntSushi]

What is the expected behavior here? From what I can tell, the given code isn't correct because you aren't using #[target_feature(enable = "avx2")]. (Alternatively, compile with RUSTFLAGS="-C target-feature=+avx2".)

I think you'll want to have a careful review of https://doc.rust-lang.org/std/arch/index.html#overview and https://doc.rust-lang.org/std/arch/index.html#dynamic-cpu-feature-detection

[hellow554]

@BurntSushi why avx2? According to the doc it only requires "avx"

[BurntSushi]

@hellow554 I didn't check the precise requirements of each individual vendor intrinsic, and I don't have them memorized. :-) Feel free to s/avx2/avx, but I don't think my general point changes.

[lqd]

#[target_feature(enable = "avx")] works indeed.

[kali]

all right... going back to hiding, thanks folks :)