fix(core): restrict AVX2 normalize to x86_64 (fix i686 build)

[gustav-fff]

Closes #656

Root cause

normalize_bytes in crates/fff-core/src/bigram_filter.rs gated the AVX2 dispatch and the normalize_bytes_avx2 fn on #[cfg(any(target_arch = \"x86_64\", target_arch = \"x86\"))], but the body unconditionally does use std::arch::x86_64::*;. On 32-bit x86 (i686-unknown-linux-gnu) std::arch::x86_64 does not exist, so fff-search failed to compile:

error[E0432]: unresolved import `std::arch::x86_64`
   --> fff-search-0.9.6/src/bigram_filter.rs:626:20
    |
626 |     use std::arch::x86_64::*;
    |                    ^^^^^^ could not find `x86_64` in `arch`

Regressed in #566. Reported by @Juhan280 hitting this while packaging nushell for termux i686 android devices.

Fix

Restrict the AVX2 dispatch and the normalize_bytes_avx2 implementation to #[cfg(target_arch = \"x86_64\")]. On 32-bit x86 we fall back to normalize_bytes_scalar, which LLVM auto-vectorises with the baseline SIMD anyway. Per maintainer + @fdncred: dropping SIMD on 32-bit x86 is acceptable — compiling at all is more important.

Also add a Build i686-unknown-linux-gnu CI job in .github/workflows/rust.yml so this regresses loudly next time. The job installs gcc-multilib, adds the i686-unknown-linux-gnu target, and runs cargo build -p fff-search --target i686-unknown-linux-gnu.

Steps to reproduce

On pre-fix main (commit 8c76a1b):

rustup target add i686-unknown-linux-gnu
sudo apt-get install -y gcc-multilib   # linux host
cargo build -p fff-search --target i686-unknown-linux-gnu

Expected: build succeeds. Actual on pre-fix main: error[E0432]: unresolved import 'std::arch::x86_64' at crates/fff-core/src/bigram_filter.rs:626.

How verified

• cargo check -p fff-search --no-default-features --features ripgrep on x86_64 macOS: clean.
• cargo clippy -p fff-search --no-default-features --features ripgrep -- -D warnings: clean.
• New CI job exercises the i686 build path on every PR; cross-compile toolchain not available locally on macOS host, so verification of the 32-bit target itself is deferred to CI.

Automated triage via Gustav. Honk-Honk 🪿

[dmtrKovalenko]

@gustav-fff iterate on this PR and guard all the non x86_64 code that is failing to compile

[gustav-fff]

[triage-bot] DIRECTED: already guarded, i686 CI green as of afc23c4.

Timeline: your comment at 17:55:06 UTC preceded commit afc23c4 (17:57:28 UTC), which silenced the unused i1/i2 bindings in select_rare_pair that CI's -D unused-variables was rejecting on 32-bit x86. Latest run:

• Build i686-unknown-linux-gnu: pass (55s)
• cargo clippy (-D warnings): pass
• cargo fmt: pass

Grep swept for every arch-specific site in crates/:

• crates/fff-core/src/bigram_filter.rs:596 — AVX2 dispatch #[cfg(target_arch = "x86_64")]
• crates/fff-core/src/bigram_filter.rs:604 — NEON dispatch #[cfg(all(target_arch = "aarch64", target_feature = "neon"))]
• crates/fff-core/src/bigram_filter.rs:623 — normalize_bytes_avx2 #[cfg(target_arch = "x86_64")]
• crates/fff-core/src/bigram_filter.rs:656 — normalize_bytes_neon #[cfg(all(target_arch = "aarch64", target_feature = "neon"))]
• crates/fff-core/src/case_insensitive_memmem.rs:106,375 — verify_avx2, search_packed_pair_avx2 #[cfg(target_arch = "x86_64")]
• crates/fff-core/src/case_insensitive_memmem.rs:171,197,255 — NEON helpers #[cfg(target_arch = "aarch64")]
• crates/fff-core/src/case_insensitive_memmem.rs:353,359,505,516 — runtime dispatch gated by arch
• crates/fff-core/src/case_insensitive_memmem.rs:499 — select_rare_pair result #[cfg_attr(not(any(target_arch = "x86_64", target_arch = "aarch64")), allow(unused_variables))]

On i686 all SIMD paths compile out; normalize_bytes_scalar / search (memchr2 fallback) run and LLVM auto-vectorises them with the baseline SIMD. No further code was failing to compile.

If you were seeing a stale failure in your local view, please point me at the specific error — otherwise ready for review.

Honk-Honk 🪿