Skip to content

1. Rust Toolchain, SIMD Intrinsics, and Ecosystem Crates

1.1 Latest Stable Versions

  • Rust Compiler (rustc & cargo): 1.99.1 (edition = "2024", next beta release 1.100.0)
  • rayon: 1.12.0
  • thiserror: 2.0.21
  • libloading: 0.9.0
  • serde: 1.0.229
  • serde_json: 1.0.154
  • cfg-if: Removed from dependencies (replaced by the standard library macro std::cfg_select!)

1.2 Version Range Changelog (Rust 1.85.0 - 1.99.1)

A. Rust 1.85.0 (edition = "2024")

  • What changed: The unsafe_op_in_unsafe_fn lint is enabled by default in Rust Edition 2024. Declaring an unsafe fn no longer implicitly permits calling unsafe operations inside its body without an explicit unsafe {} block.
  • Syntax changes: Stabilized let-chains (if let Some(x) = a && x > 0) inside if and while conditions.

B. Rust 1.86.0 - 1.87.0 (Target Features & SIMD Intrinsics)

  • Rust 1.86.0: Functions annotated with #[target_feature(enable = "...")] can be called safely from other functions that enable an identical or superset target feature set.
  • Rust 1.87.0: Value-based SIMD intrinsics (such as _mm256_add_ps, _mm512_fmadd_ps, _mm512_cmp_ps_mask) no longer require an unsafe block when called inside a function with matching #[target_feature] attributes. Only raw pointer dereference intrinsics (_mm256_loadu_ps, _mm256_storeu_ps, _mm512_loadu_ps, _mm512_storeu_ps) remain unsafe.

C. Rust 1.89.0 (AVX-512 Target Feature Stabilization)

  • What changed: All avx512f, avx512dq, avx512bw, and avx512vl target features and std::arch::x86_64::_mm512_* intrinsics were fully stabilized on the stable compiler channel without requiring #![feature(stdarch_x86_avx512)].

D. Rust 1.95.0 (std::cfg_select!)

  • What changed: Added the standard library macro std::cfg_select! for compile-time configuration branching.
  • Impact: Eliminates the need for the third-party cfg-if crate.

1.3 Adaptations in LuminaMoon Source Code

1. Replacing cfg-if with std::cfg_select!

  • Files: Cargo.toml, optimizer/crates/luminav-simd/Cargo.toml, optimizer/crates/luminav-simd/src/lib.rs
  • Changes: Removed the external cfg-if dependency from Cargo.toml. Updated detect_simd_backend() in optimizer/crates/luminav-simd/src/lib.rs to use the standard macro: 
    pub fn detect_simd_backend() -> SimdBackend {
    std::cfg_select! {
        target_arch = "x86_64" => {
            if std::arch::is_x86_feature_detected!("avx512f")
                && std::arch::is_x86_feature_detected!("avx512dq")
                && std::arch::is_x86_feature_detected!("avx512bw")
                && std::arch::is_x86_feature_detected!("avx512vl")
                && std::arch::is_x86_feature_detected!("fma")
            {
                SimdBackend::Avx512
            } else if std::arch::is_x86_feature_detected!("avx2")
                && std::arch::is_x86_feature_detected!("fma")
            {
                SimdBackend::Avx2
            } else {
                SimdBackend::Scalar
            }
        }
        _ => {
            SimdBackend::Scalar
        }
    }
}

2. Separating SimdBackend::Scalar and SimdBackend::Avx2 Dispatch Paths

  • File: optimizer/crates/luminav-simd/src/parallel.rs
  • Previous issue: The match backend arm in chunk dispatch helpers grouped SimdBackend::Avx2 | SimdBackend::Scalar into avx2::*_avx2 calls. Even though top-level callers intercepted SimdBackend::Scalar earlier, combining those arms risked an illegal instruction trap (SIGILL) if a helper were invoked directly on a CPU without AVX2.
  • Fix: Explicitly separated SimdBackend::Scalar from SimdBackend::Avx2 across all 6 chunk dispatch functions in parallel.rs.