[Bug]: Build failing for x86 32 bit cpus

[Juhan280]

Which fff frontend?

Other / multiple

has logs

   Compiling socks v0.3.4
   Compiling hashlink v0.12.0
   Compiling is-docker v0.2.0
   Compiling toml_datetime v1.1.1+spec-1.1.0
   Compiling mio v0.8.11
   Compiling page_size v0.6.0
   Compiling typed-path v0.12.3
   Compiling symlink v0.1.0
   Compiling fff-search v0.9.6
   Compiling chrono-tz v0.10.4
   Compiling fallible-iterator v0.3.0
   Compiling arrayref v0.3.9
   Compiling constant_time_eq v0.4.2
   Compiling debug_unsafe v0.1.4
   Compiling unicode-id v0.3.6
   Compiling pathdiff v0.2.3
   Compiling fallible-streaming-iterator v0.1.9
   Compiling toml_writer v1.1.1+spec-1.1.0
   Compiling linux-raw-sys v0.4.15
   Compiling utf8-zero v0.8.1
   Compiling jiff v0.2.23
   Compiling ureq v3.3.0
   Compiling rusqlite v0.40.1
   Compiling parse_datetime v0.14.0
   Compiling zip v7.2.0
   Compiling markdown v1.0.0
   Compiling atoi_simd v0.17.0
   Compiling tracing-appender v0.2.5
   Compiling notify v6.1.1
   Compiling is-wsl v0.4.0
   Compiling ignore v0.4.25
   Compiling dirs v5.0.1
   Compiling phf v0.12.1
   Compiling nu-table v0.114.0 (/home/builder/.termux-build/nushell/src/crates/nu-table)
   Compiling pest_consume v1.1.3
   Compiling rand v0.9.2
   Compiling fff-notify-debouncer-full v0.9.4
   Compiling serde_urlencoded v0.7.1
   Compiling etcetera v0.10.0
   Compiling codepage v0.1.2
   Compiling quick-xml v0.39.4
   Compiling nuon v0.114.0 (/home/builder/.termux-build/nushell/src/crates/nuon)
   Compiling nu-heavy-utils v0.114.0 (/home/builder/.termux-build/nushell/src/crates/nu-heavy-utils)
   Compiling pori v0.0.0
   Compiling serde_spanned v1.1.1
   Compiling fff-grep v0.9.6
   Compiling fd-lock v4.0.4
   Compiling neo_frizbee v0.10.3
   Compiling serde_derive_internals v0.29.1
   Compiling itertools v0.13.0
   Compiling memmap2 v0.9.10
   Compiling platform-info v2.0.5
   Compiling csv-core v0.1.13
   Compiling glidesort v0.1.2
   Compiling openssl-probe v0.2.1
   Compiling is_debug v1.1.0
   Compiling atomic-waker v1.1.2
   Compiling fff-query-parser v0.9.6
   Compiling unsafe-libyaml v0.2.11
   Compiling fs_extra v1.3.0
   Compiling smol_str v0.3.6
   Compiling fast-float2 v0.2.3
   Compiling calamine v0.35.0
   Compiling shadow-rs v2.0.0
   Compiling git2 v0.20.4
   Compiling mq-markdown v0.6.0
   Compiling serde_yaml v0.9.34+deprecated
   Compiling uu_mv v0.9.0
   Compiling h2 v0.4.13
   Compiling rustls-native-certs v0.8.3
   Compiling reedline v0.49.0
error[E0432]: unresolved import `std::arch::x86_64`
   --> /home/builder/.cargo/registry/src/index.crates.io-1949cf8c6b5b557f/fff-search-0.9.6/src/bigram_filter.rs:626:20
    |
626 |     use std::arch::x86_64::*;
    |                    ^^^^^^ could not find `x86_64` in `arch`

   Compiling csv v1.4.0
   Compiling uu_uname v0.9.0
   Compiling schemars_derive v1.2.1
   Compiling toml v1.1.2+spec-1.1.0
For more information about this error, try `rustc --explain E0432`.
error: could not compile `fff-search` (lib) due to 1 previous error
warning: build failed, waiting for other jobs to finish...
error: failed to compile `nu v0.114.0 (/home/builder/.termux-build/nushell/src)`, intermediate artifacts can be found at `/home/builder/.termux-build/nushell/src/target`.
To reuse those artifacts with a future compilation, set the environment variable `CARGO_BUILD_BUILD_DIR` to that path.
Error: Process completed with exit code 101.

Description

Since #566 the fff-search crate is failing to build for x86 32 bit CPUs (target: i686-unknown-linux-gnu). I believe this line is responsible for the issue, where the x86_64 modules is imported even for 32 bit CPUs.

fff/crates/fff-core/src/bigram_filter.rs

Lines 621 to 627 in 98af56c

	/// AVX2 normalize: 32 bytes/iter. AVX2 only has signed cmp, so unsigned
	/// range checks use `min(max(v, lo), hi) == v`.
	#[cfg(any(target_arch = "x86_64", target_arch = "x86"))]
	#[target_feature(enable = "avx2")]
	unsafe fn normalize_bytes_avx2(src: &[u8], dst: &mut [u8]) {
	use std::arch::x86_64::*;
	let len = src.len();

[dmtrKovalenko]

oh wow what is the hardware you are compiling it for?

why am I asking: the probelm that we likely won't be able to deliver any of the SIMD optimization on 32 bit architecture

[Juhan280]

We discovered this while trying to compile nushell for i686 android devices (termux)
https://github.com/termux/termux-packages/actions/runs/28733290861/job/85202826677

[fdncred]

I think not having SIMD on 32-bit architecture is ok. It's better than not compiling at all, which is what happens now.

[dmtrKovalenko]

yeah I agree @gustav-fff try to figure this out if you can, add a single CI test for compiling to linux-8686-gnu

[gustav-fff]

[triage-bot] LINKED PR: #657 (open) — #657

Restricts the AVX2 dispatch and normalize_bytes_avx2 in crates/fff-core/src/bigram_filter.rs to #[cfg(target_arch = "x86_64")] so the unconditional use std::arch::x86_64::*; no longer leaks into 32-bit x86 builds. On i686 the scalar path is used (LLVM auto-vectorises it), which resolves the E0432 you hit. Also adds an i686-unknown-linux-gnu CI job so this regresses loudly next time. Awaiting maintainer review.

Honk-Honk 🪿