Bad codegen for widen followed by ARM vdupq_n_*

[CatsAreFluffy]

Code

I tried this code:

use std::arch::aarch64::*;

#[no_mangle]
pub unsafe fn half_dup(x: u16) -> uint16x8_t {
    vaddq_u16(vreinterpretq_u16_u32(vdupq_n_u32(x as u32)), vdupq_n_u16(1))
}

I expected to see this happen: The compiled output would use the dup instruction

Instead, this happened: The compiled output uses a bunch of movs instead

Godbolt

Version it worked on

It most recently worked on: Rust 1.81.0

Version with regression

rustc --version --verbose:

rustc 1.82.0 (f6e511eec 2024-10-15)
binary: rustc
commit-hash: f6e511eec7342f59a25f7c0534f1dbea00d01b14
commit-date: 2024-10-15
host: x86_64-unknown-linux-gnu
release: 1.82.0
LLVM version: 19.1.1

and

rustc 1.87.0-nightly (f280acf4c 2025-02-19)
binary: rustc
commit-hash: f280acf4c743806abbbbcfe65050ac52ec4bdec0
commit-date: 2025-02-19
host: x86_64-unknown-linux-gnu
release: 1.87.0-nightly
LLVM version: 20.1.0

[nikic]

@jieyouxu Rust 1.82 would be the LLVM 19 upgrade.

[jieyouxu]

@jieyouxu Rust 1.82 would be the LLVM 19 upgrade.

Yeah sorry somehow I didn't notice the 1.82

[hkratz]

@rustbot label +A-SIMD +O-aarch64

[nikic]

Upstream issue: llvm/llvm-project#128349

[apiraino]

WG-prioritization assigning priority (Zulip discussion).

@rustbot label -I-prioritize +P-medium

[nikic]

Looks like Rust is using a non-canonical pattern for vector splat generation: https://github.com/rust-lang/stdarch/blob/1c6113f37b44baeb3364867f52bd67a3f70e7a6f/crates/core_arch/src/simd.rs#L25-L34

We shouldn't go through a 1-element vector here. We probably need to add a simd_splat intrinsic to generate the correct form.

[moxian]

bisects to #128866 , and more specifically rust-lang/stdarch@7365074

cc @scottmcm FYI, although I'm not sure this is actionable by you.

@rustbot label: -E-needs-bisection +S-has-bisection