Codegen emits ineffective vxorps/vaddps sequences

[ksf]

I tried this code:

pub fn vae_dec_linear(latent: &[f32; 4]) -> [f32; 3] {
    let weights: [[f32; 4]; 3] = [
        [49.5210, 29.0283, -23.9673, -39.4981],
        [41.1373, 42.4951, 24.7349, -50.8279],
        [40.2919, 18.9304, 30.0236, -81.9976],
    ];
  //  let bias = [99.9368, 99.8421, 99.5384];
    let bias = [0.0, 0.0, 0.0];
    affine_transform(&weights, &bias, &latent)
}

pub fn affine_transform<
    T: Copy + std::ops::Add<Output = T> + std::ops::Mul<Output = T>,
    const IN: usize,
    const OUT: usize,
>(
    weights: &[[T; IN]; OUT],
    bias: &[T; OUT],
    vec: &[T; IN],
) -> [T; OUT] {
    std::array::from_fn(|i| (0..OUT).fold(bias[i], |acc, j| acc + vec[i] * weights[i][j]))
}

I expected to see this happen: Literally nothing :)

Instead, this happened:

example::vae_dec_linear:
        mov     rax, rdi
        vmovss  xmm0, dword ptr [rsi + 8]
        vmulss  xmm1, xmm0, dword ptr [rip + .LCPI0_0]

        vxorps  xmm2, xmm2, xmm2                 ;; here
        vaddss  xmm1, xmm1, xmm2

        vmulss  xmm2, xmm0, dword ptr [rip + .LCPI0_1]
        vmulss  xmm0, xmm0, dword ptr [rip + .LCPI0_2]
        vaddss  xmm1, xmm2, xmm1
        vaddss  xmm0, xmm0, xmm1
        vmovsd  xmm1, qword ptr [rsi]
        vmulps  xmm2, xmm1, xmmword ptr [rip + .LCPI0_3]

        vxorps  xmm3, xmm3, xmm3                 ;; and here
        vaddps  xmm2, xmm2, xmm3

        vmulps  xmm3, xmm1, xmmword ptr [rip + .LCPI0_4]
        vmulps  xmm1, xmm1, xmmword ptr [rip + .LCPI0_5]
        vaddps  xmm2, xmm3, xmm2
        vaddsubps       xmm1, xmm2, xmm1
        vmovlps qword ptr [rdi], xmm1
        vmovss  dword ptr [rdi + 8], xmm0
        ret

Note the two xor/add sequences adding zero to xmm1 or xmm2, respectively, this is with -Copt-level=3 -C target-cpu=native

EDIT: Oh and nevermind that the code has a bug it should be

    std::array::from_fn(|i| (0..IN).fold(bias[i], |acc, j| acc + vec[j] * weights[i][j]))

Changes the second sequence to

        vxorps  xmm5, xmm5, xmm5
        vmulps  xmm2, xmm2, xmmword ptr [rip + .LCPI0_5]
        vaddps  xmm0, xmm0, xmm5
        vaddps  xmm0, xmm0, xmm2

, not as easy to spot, the second vaddps also looks suspicious.

Meta

rustc --version --verbose:

All I looked at in compiler explorer, including nightly

Backtrace

<backtrace>

[Jules-Bertholet]

@rustbot label A-codegen I-heavy I-slow O-x86

[ksf]

And another at least similar one:

#![feature(portable_simd)]
use std::simd::{Simd, SimdFloat};

pub fn dot_product_simd(x: &[f32; 9], y: &[f32; 9]) -> f32 {
    let a: Simd<_,8> = Simd::from_slice(&x[..8]);
    let b = Simd::from_slice(&y[..8]);
    let k = (a * b).reduce_sum();
    let a = x[8];
    let b = y[8];
    f32::mul_add(a, b, k)
}

(excerpt)

        vxorps  xmm1, xmm1, xmm1
        vmulps  ymm0, ymm0, ymmword ptr [rsi]
        vaddss  xmm1, xmm0, xmm1

That's with -O -Copt-level=3 -C target-cpu=znver2 . The previous target-cpu=native was whatever compiler explorer uses, znver2 just looks a bit reordered.

The code for reduce_sum also doesn't look great but I won't say anything because I didn't benchmark.

[ksf]

Fumble click.