[avx2] gather operations

[Lokathor]

At first glance it seems like these can't be made safe and also totally zero-runtime-cost in an ergonomic way. I'm open to suggestions.

I hear these requested often enough when people talk about avx2, so I think if we can't do it safely with no runtime checks it would be permissible to expose versions of these that did some runtime checking if it was clearly marked as such.

• _mm_i32gather_epi32
• _mm_i32gather_epi64
• _mm_i32gather_pd
• _mm_i32gather_ps
• _mm_i64gather_epi32
• _mm_i64gather_epi64
• _mm_i64gather_pd
• _mm_i64gather_ps
• _mm_mask_i32gather_epi32
• _mm_mask_i32gather_epi64
• _mm_mask_i32gather_pd
• _mm_mask_i32gather_ps
• _mm_mask_i64gather_epi32
• _mm_mask_i64gather_epi64
• _mm_mask_i64gather_pd
• _mm_mask_i64gather_ps
• _mm256_i32gather_epi32
• _mm256_i32gather_epi64
• _mm256_i32gather_pd
• _mm256_i32gather_ps
• _mm256_i64gather_epi32
• _mm256_i64gather_epi64
• _mm256_i64gather_pd
• _mm256_i64gather_ps
• _mm256_mask_i32gather_epi32
• _mm256_mask_i32gather_epi64
• _mm256_mask_i32gather_pd
• _mm256_mask_i32gather_ps
• _mm256_mask_i64gather_epi32
• _mm256_mask_i64gather_epi64
• _mm256_mask_i64gather_pd
• _mm256_mask_i64gather_ps

[Soveu]

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

unsafe fn mm_i32gather_ps_unchecked(slice: &[f32], offsets: m128i) -> m128 {
    m128(_mm_i32gather_ps(slice.as_ptr(), offsets.0, 4))
}

fn are_offsets_valid(offsets: m128i, sz: usize) -> bool {
    let offsets: [u32; 4] = offsets.into();
    let index = offsets.iter().cloned().max().unwrap() as usize;
    return index < sz;
}

fn mm_i32gather_ps(slice: &[f32], offsets: m128i) -> Option<m128> {
    if !are_offsets_valid(offsets, slice.len()) {
        return None;
    }

    Some(unsafe{ mm_i32gather_ps_unchecked(slice, offsets) })
}

#[inline(never)]
fn checkpoint(x: &m128) {
    println!("{}", x);
}

fn main() {
    let arr: [f32; 6] = [123.0, 144.0, 456.0, 567.0, 1456.0, 1235.0];
    let x = mm_i32gather_ps(&arr, [0,1,2,3].into()).unwrap();
    checkpoint(&x);
}

With release builds, this is inlined without boundary checks present in mm_i32gather_ps, so we have the same situation as with standard slice::get()

There are two more things to say:

1. _mm_i32gather_ps is NOT a single instruction (see https://doc.rust-lang.org/src/core/up/stdarch/crates/core_arch/src/x86/avx2.rs.html#1201-1212)
2. _mm_i32gather_ps requires f32 pointer to be passed, but offsets behave as they were byte offsets, which can be confusing (it can be easily fixed by passing scale equal to the size of f32)

I don't know if other gather-functions work the same, but also I don't expect them to behave too differently

[Lokathor]

This is good progress.

I'll just note that we could probably check the lanes for validity using SIMD of course ;3