compiler bug?

simd_test.rs

// SSE returns the correct answer in Debug and Release
// AVX returns the correct answer in Debug but not Release
// This seems to be related to add_stuff being recursive


#[cfg(target_arch = "x86")]
use std::arch::x86::*;
#[cfg(target_arch = "x86_64")]
use std::arch::x86_64::*;
use std::fmt::Debug;

pub trait Simd {
    type Vf32: Copy + Debug;

    unsafe fn add_ps(a: Self::Vf32, b: Self::Vf32) -> Self::Vf32;
    unsafe fn set1_ps(a: f32) -> Self::Vf32;
}

pub struct Sse2;
impl Simd for Sse2 {
    type Vf32 = __m128;
    #[inline(always)]
    unsafe fn add_ps(a: Self::Vf32, b: Self::Vf32) -> Self::Vf32 {
        _mm_add_ps(a, b)
    }
    #[inline(always)]
    unsafe fn set1_ps(a: f32) -> Self::Vf32 {
        _mm_set1_ps(a)
    }
}

pub struct Avx2;
impl Simd for Avx2 {
    type Vf32 = __m256;
    #[inline(always)]
    unsafe fn add_ps(a: Self::Vf32, b: Self::Vf32) -> Self::Vf32 {
        _mm256_add_ps(a, b)
    }
    #[inline(always)]
    unsafe fn set1_ps(a: f32) -> Self::Vf32 {
        _mm256_set1_ps(a)
    }
}

#[inline(always)]
unsafe fn add_stuff<S: Simd>(a: f32, count: i32) -> S::Vf32 {
    let b = S::set1_ps(2.0);
    let a2 = S::set1_ps(a);
    if count < 3 {
        S::add_ps(S::add_ps(a2, b), add_stuff::<S>(a, count + 1))
    } else {
        S::add_ps(a2, b)
    }
}

#[target_feature(enable = "avx2")]
unsafe fn add_stuff_avx() {
    let r = add_stuff::<Avx2>(2.0, 1);
    println!("avx {:?}", r);
}

#[target_feature(enable = "sse")]
unsafe fn add_stuff_sse() {
    let r = add_stuff::<Sse2>(2.0, 1);
    println!("sse {:?}", r);
}

fn main() {
    unsafe {
        add_stuff_sse();
        add_stuff_avx();
    }
}