Integer multiply high (mulh) equivalent in SimdInt and SimdUint

[ds84182]

LLVM already has a fairly shaky optimization that doesn't make it through isel when followed by a left or right shift. https://rust.godbolt.org/z/a6eexcE6z

#![no_std]
#![feature(portable_simd)]

use core::simd::prelude::*;

// Generates vpmulhw
pub fn mulhw(a: i16x16, b: i16x16) -> i16x16 {
    (((a.cast::<i32>()) * (b.cast::<i32>())) >> 16).cast::<i16>()
}

// Generates a mess
pub fn mulhw_and_shift(a: i16x16, b: i16x16) -> i16x16 {
    mulhw(a, b) >> 1
}

A dedicated function that can multiply two integers and take the high parts would be very beneficial when handling fixed-point integers. And hopefully it'll have a better chance of surviving optimization passes.

[calebzulawski]

This is perhaps not very intuitive, but things like this work when you use a swizzle: https://rust.godbolt.org/z/zGMb4b5PK
I've done similar here: https://github.com/calebzulawski/autobahn-hash/blob/f35d18565b996a162d1cfbc18abd268b940f4ced/src/lib.rs#L83-L91

[programmerjake]

This is perhaps not very intuitive, but things like this work when you use a swizzle: https://rust.godbolt.org/z/zGMb4b5PK

that breaks when avx2 is enabled: https://rust.godbolt.org/z/h84KsKEnx

[calebzulawski]

Interesting, I think that's a missing instruction combine rule, separate issue from trying to optimize the shifts.