Improve `load`, `store`, `gather`, `scatter` codegen for targets without masked load/store instructions

Open
#494 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Assessment

Difficulty
4/5
Estimated time
3-5 days
Newbie friendliness
42/100
Issue type
Feature
Clarity
Mostly clear
Activity status
Stale
Tech stack
rust
Domain
compilers, performance

Research direction

Start by comparing the generated output from the linked Godbolt example for load_or_default_simd and store_simd with the scalar versions on x86_64 sse2. Then trace how these operations lower to the mentioned LLVM masked load, store, gather, and scatter intrinsics. Done means improved load/store codegen matching the scalar loops, with gather/scatter addressed if feasible.

Written by the indexing model from the issue text.

Description

C-bug

On at least the baseline x86_64 sse2 target, I'd like the code output to be more similar to the manual scalar implementation. I end up having to write out the manual loops and avoid the load/store functions for performance.

https://rust.godbolt.org/z/57hsoTPbd

portable_simd 

#![feature(portable_simd)]

use std::simd::*;

const N: usize = 8;

#[inline(never)]
pub fn load_or_default_simd(slice: &[f32]) -> Simd<f32, N> {
    Simd::load_or_default(slice)
}

#[inline(never)]
pub fn store_simd(data: Simd<f32, N>, slice: &mut [f32]) {
    data.store_select(slice, Mask::splat(true))
}

#[inline(never)]
pub fn gather_or_default_simd(slice: &[f32], idxs: Simd<usize, N>) -> Simd<f32, N> {
    Simd::gather_or_default(slice, idxs)
}

#[inline(never)]
pub fn scatter_simd(data: Simd<f32, N>, slice: &mut [f32], idxs: Simd<usize, N>) {
    data.scatter(slice, idxs)
}

scalar 

#![feature(portable_simd)]

use std::simd::*;

const N: usize = 8;

#[inline(never)]
pub fn load_or_default_scalar(slice: &[f32]) -> Simd<f32, N> {
    let mut result = [0.0; N];
    for (&s, r) in slice.iter().zip(result.iter_mut()) {
        *r = s;
    }
    Simd::from(result)
}

#[inline(never)]
pub fn store_scalar(data: Simd<f32, N>, slice: &mut [f32]) {
    for (s, &d) in slice.iter_mut().zip(data.as_array().iter()) {
        *s = d;
    }
}

#[inline(never)]
pub fn gather_or_default_scalar(slice: &[f32], idxs: Simd<usize, N>) -> Simd<f32, N> {
    let mut result = [0.0; N];
    for (&i, r) in idxs.as_array().iter().zip(result.iter_mut()) {
        *r = *slice.get(i).unwrap_or(&0.0);
    }
    Simd::from(result)
}

#[inline(never)]
pub fn scatter_scalar(data: Simd<f32, N>, slice: &mut [f32], idxs: Simd<usize, N>) {
    for (&d, &idx) in data.as_array().iter().zip(idxs.as_array().iter()) {
        if let Some(s) = slice.get_mut(idx) {
            *s = d;
        }
    }
}

I'm mainly concerned about load/store, I don't expect much from gather/scatter.

I'm not sure if this is something to be solved on the Rust side or LLVM for the intrinsics which are present in the IR

  • llvm.masked.load.*
  • llvm.masked.store.*
  • llvm.masked.gather.*
  • llvm.masked.scatter.*
Meta
rustc 1.93.0-nightly (c86564c41 2025-11-27)
binary: rustc
commit-hash: c86564c412a5949088a53b665d8b9a47ec610a39
commit-date: 2025-11-27
host: x86_64-unknown-linux-gnu
release: 1.93.0-nightly
LLVM version: 21.1.5
Dominant language
Rust
Stars
1.1k
Forks
108
Avg merge
22h 53m
Merged PRs (30d)
3

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from rust-lang/portable-simd

All issues in rust-lang/portable-simd

Similar issues

More Rust issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.