Significantly worse codegen for SIMD shuffles after Rust 1.73.0

[TDecking]

See https://godbolt.org/z/jKczKPMMe.

The code is the following:

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

#[no_mangle]
pub unsafe fn custom_shuffle(x: __m128i) -> __m128i {
    let mut tmp1 = _mm_unpacklo_epi8(x, _mm_setzero_si128());
    tmp1 = _mm_shuffle_epi32::<78>(tmp1);
    tmp1 = _mm_shufflelo_epi16::<27>(tmp1);
    tmp1 = _mm_shufflehi_epi16::<27>(tmp1);

    let mut tmp2 = _mm_unpackhi_epi8(x, _mm_setzero_si128());
    tmp2 = _mm_shuffle_epi32::<78>(tmp2);
    tmp2 = _mm_shufflelo_epi16::<27>(tmp2);
    tmp2 = _mm_shufflehi_epi16::<27>(tmp2);

    _mm_packus_epi16(tmp2, tmp1)
}

The regression happened between Rust 1.72.0 and Rust 1.73.0.

[nikic]

Looks like a x86 backend regression, filed llvm/llvm-project#104482.

[nikic]

Fixed upstream in LLVM 20.

[apiraino]

WG-prioritization assigning priority (Zulip discussion).

@rustbot label -I-prioritize +P-medium