Incorrect ML-DSA output on Alpine Linux and Ampere Altra

[FiloSottile]

I tried using libcrux-ml-dsa v0.0.3 to generate accumulated test vectors for ML-DSA. Since I wanted to generate a 60M iteration test vector, I cross-compiled the code to run on an Alpine Linux server with an Ampere Altra processor.

Surprisingly, the output of the binary on Linux was different from the one generated on my M2 laptop. Comparing with my implementation, it looks like the local macOS build is correct, while the Alpine Linux cross-compiled build is producing incorrect results.

I built a minimal reproducer and built it directly on the Alpine Linux machine with a rustup toolchain, no cargo.toml, and Alpine's gcc to rule out cross-compilation issues. It gave the same incorrect output.

Note that I am a complete Rust noob, and I had Claude write most of the code, but from eyeballing it I can't imagine how it should produce different results on different targets.

Environment

System 1 (macOS - CORRECT OUTPUT)

• OS: macOS 15.6.1 (24G90)
• Kernel: Darwin 24.6.0 ARM64
• CPU: Apple Silicon M2
• Rust: rustc 1.86.0 (05f9846f8 2025-03-31)
• Cargo: 1.86.0 (adf9b6ad1 2025-02-28)

System 2 (Alpine Linux - INCORRECT OUTPUT)

• OS: Alpine Linux edge (3.23.0_alpha20250612)
• Kernel: Linux 6.12.40-1-lts aarch64
• CPU: Ampere Altra Q64-22 (ARM64)
• Target: aarch64-unknown-linux-musl
• Rust: rustc 1.91.0 (f8297e351 2025-10-28)
• Cargo: 1.91.0 (ea2d97820 2025-10-10)

Minimal Reproducer

Cargo.toml

[package]
name = "libcrux-ml-dsa-repro"
version = "0.1.0"
edition = "2021"

[dependencies]
libcrux-ml-dsa = "0.0.3"
hex = "0.4"

src/main.rs

use libcrux_ml_dsa::ml_dsa_44;

fn main() {
    println!("libcrux-ml-dsa v0.0.3 - Platform Determinism Test\n");

    // Test with a zero seed
    let seed = [0u8; 32];
    let kp = ml_dsa_44::generate_key_pair(seed);

    // Sign an empty message with deterministic (zero) randomness
    let empty_message = &[];
    let context = &[];
    let randomness = [0u8; 32];
    let signature = ml_dsa_44::sign(&kp.signing_key, empty_message, context, randomness).unwrap();

    println!("Seed: {}", hex::encode(seed));
    println!("\nPublic key ({} bytes):", kp.verification_key.as_ref().len());
    println!("{}", hex::encode(kp.verification_key.as_ref()));
    println!("\nSignature ({} bytes):", signature.as_ref().len());
    println!("{}", hex::encode(signature.as_ref()));
}

Reproduction Steps

cargo run --release

I get the same output with and without --release.

Actual Output

macOS Output (appears correct)

libcrux-ml-dsa v0.0.3 - Platform Determinism Test

Seed: 0000000000000000000000000000000000000000000000000000000000000000

Public key (1312 bytes):
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

Signature (2420 bytes):
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

Alpine Linux Output (incorrect)

libcrux-ml-dsa v0.0.3 - Platform Determinism Test

Seed: 0000000000000000000000000000000000000000000000000000000000000000

Public key (1312 bytes):
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

Signature (2420 bytes):
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

[FiloSottile]

Cross-compiling to x86_64-unknown-linux-musl and running on a AMD EPYC™ 7443P produces the correct output, so it's not as simple as a musl libc issue.

[jschneider-bensch]

Thanks for reporting this, and for providing a reproducer! I will investigate.

[jschneider-bensch]

We've identified an issue with one of the aarch64 intrinsics and put up a PR to address that.

We checked locally that this fixes the output of your reproducer, but if you'd like to briefly confirm that the issue is resolved, you can update your Cargo.toml as follows, to use the branch where the fix is in place:

[package]
name = "libcrux_1220"
version = "0.1.0"
edition = "2021"

[dependencies]
libcrux-ml-dsa = { git = "https://github.com/cryspen/libcrux", branch="jonas/arm64-fix" }
hex = "0.4"

[FiloSottile]

Thank you for the instructions :) the fix looks like it's working.

libcrux-ml-dsa v0.0.3 - Platform Determinism Test

Seed: 0000000000000000000000000000000000000000000000000000000000000000

Public key (1312 bytes):
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

Signature (2420 bytes):
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

[djc]

@jschneider-bensch we're discussing a RustSec advisory in

• Add libcrux-intrinsics entry rustsec/advisory-db#2493

Would be great to get your feedback. I think publishing that would also result in a GHSA, which is probably warranted?

[pinkforest]

This also seems to be toolchain thing also ? That doesn't seem to give SHA3 as target_feature given:

Looking at the build gating

I think it's the SHA3 missing in aarch64-musl-linux toolchain which was used in alpine VM build for me (ctz/graviola#121)

This gate short-circuited the vxarq_u64::<RIGHT>(a, b) before the incorrect data was passed further along which gets generated in non-SHA3 target_feature

Running this: https://github.com/pinkforest/neon_or_not in OSX M1 host using toolchain :

toolchain: stable-aarch64-apple-darwin

NEON!
SHA3!

Meanwhile alpine aarch64 M1 (which does not have SHA3 target_features) produces indicating no SHA3 in target_features:

toolchain: stable-aarch64-unknown-linux-musl

NEON!
NOPE-SHA3

I would evaluate test vectors based on all the gates including cfg(target_feature=arm64ec) that also I'm wondering about.

Cc/ @bjorn3 might be interested..

[bjorn3]

Unless you use -Ctarget-cpu=native, -Ctarget-cpu=apple-m1 or -Ctarget-feature=+sha3, it is expected that aarch64-unknown-linux-musl doesn't enable cfg(target_feature = "sha3") by default, while aarch64-apple-darwin does. aarch64-apple-darwin has a minimum cpu requirement of M1, while aarch64-unknown-linux-musl can run on CPU's that don't support SHA3 instructions.

I would evaluate test vectors based on all the gates including cfg(target_feature=arm64ec) that also I'm wondering about.

You mean cfg(target_arch = "arm64ec")? The difference between it and cfg(target_arch = "aarch64") is just a matter of ABI. arm64ec is a ABI compatible with x86_64 Windows code except for simple shuffling around of registers in the calling convention. In C it supposedly pretends to be x86_64 even harder by for example exposing x86 vendor intrinsics rather than arm ones, but for Rust we just treat it as arm64 with a weird ABI.

[pinkforest]

ty!

Problem here was triggered through build gating which didn't get tested in github CI so the bug that got fixed was never tested / triggered automatically - I meant that I would check / ensure the CI captures al the build gates which I don't know a tool that could identify those gaps automatically (which would be handy to ensure there is coverage for all the gate permutations)

It would be handy to know if test matrix in github CI trigger all the possible permutations of compiled code given github runners don't guarantee which architecture (or nonetheless target_features / toolchain) gets selected - which was my concern also in curve25519-dalek IFMA backend that we cannot test it on top of AVX-10 business (dalek-cryptography/curve25519-dalek#693, dalek-cryptography/curve25519-dalek#635)

I raised: #1255