`test_stable_hash` fails on s390x (big-endian)

[cuviper]

Problem

Since 1.85.0, s390x-unknown-linux-gnu fails on cargo's test_stable_hash. Here's the result on master as of commit c14c1c0:

---- core::source_id::tests::test_stable_hash stdout ----

thread 'core::source_id::tests::test_stable_hash' panicked at src/cargo/core/source_id.rs:823:9:

---- expected: src/cargo/core/source_id.rs:823:46
++++ actual:   In-memory
   1      - 7062945687441624357∅
        1 + 2724068093996237922∅

cargo/src/cargo/core/source_id.rs

Line 823 in c14c1c0

assert_data_eq!(gen_hash(source_id), str!["7062945687441624357"].raw());

The actual is byte-swapped from the expected value, even though stable-hashing is meant to be platform-agnostic.

Steps

In cargo's own repo: cargo test --lib test_stable_hash

Possible Solution(s)

No response

Notes

I found that rustc-stable-hash adds a .to_le() in its Hasher::finish, which seems suspect:
https://github.com/rust-lang/rustc-stable-hash/blob/24e9848c89917abca155c8f854118e6d00ad4a30/src/sip128.rs#L532

When cargo's test calls .to_string(), that will treat the u64 as native-endian, so it doesn't match the string that a little-endian host would produce.

Similarly, the next line in the test calls short_hash -> to_hex -> to_le_bytes, which will byte-swap it again before hex-printing it.

Version

[weihanglo]

Does the u64 result from Hasher::finish the same for both big and little endian?

[cuviper]

It's the same if you only look at that u64 as little-endian bytes.

[cuviper]

I found that rustc-stable-hash adds a .to_le() in its Hasher::finish, which seems suspect:
https://github.com/rust-lang/rustc-stable-hash/blob/24e9848c89917abca155c8f854118e6d00ad4a30/src/sip128.rs#L532

Cc rust-lang/rustc-stable-hash#6 @Urgau -- I'm not sure that to_le() makes sense.

[Urgau]

Indeed my bad, opened rust-lang/rustc-stable-hash#12 to fix it.

[Urgau]

Fix merged and released in rustc-stable-hash 0.1.2.

Only big endian machines were affected, so the hash should only change for them, which doesn't represent much these days. I think that makes it a relatively safe update for Cargo.