Remove unsafe expression on safe AVX2 instructions

[Parrot7483]

The changed AVX2 simd instructions in this PR are no longer unsafe. The Compiler gives a warning when using them in an unsafe block.

[franziskuskiefer]

CI is failing

[Parrot7483]

I miss some understanding on how to use compilation targets and configuration flags here.

For example the _mm256_sign_epi32 function is safe and is for the x86 or x86-64 target and feature avx2. Because the wrapping function uses #[inline(always)], one can not use #[target_feature(enable = "avx2")]

#[inline(always)]
#[hax_lib::fstar::before(r#"[@@ "opaque_to_smt"]"#)]
pub fn mm256_sign_epi32(a: Vec256, b: Vec256) -> Vec256 {
    _mm256_sign_epi32(a, b)
}

Function _mm256_setzero_si256 is for the same target but feature avx. It is also safe. Some other functions in avx2.rs might be for the older SSE2 instruction set extension. Some are also unsafe.

Is the correct way to use these SIMD functions for different features to put the correct #[cfg(target_feature = "...")] on each of them? This way one can keep using #[inline(always)]. The user of the library would then set the needed features.

The motivation for this being that safe functions are not wrapped in an unsafe block.

Maybe its also a good idea to separate them into different modules? A function for SSE2 being in avx2.rs might be confusing (at least for me).

[franziskuskiefer]

You have to make sure to enable the target feature etc. But you also have to check if these are actually safe intrinsics for the stable Rust compiler.

[Parrot7483]

CPU intrinsics require an unsafe block because they are hardware-dependent. Running them on an unsupported CPU will crash the program. Only when using the HAX target we get a warning.

[franziskuskiefer]

CI is failing