Use another hashing algorithm? 250K calls to Siphash

[blyxyas]

While profiling Clippy I came accross 250k calls to <core::hash::sip::Hasher<core::hash::sip::Sip13Rounds> as core::hash::Hasher>::write, and they turn out to be the most time-consuming by themselves (at least that's what callgrind says), all of them coming from Cargo.

These numbers in particular is from bumpalo-3.16.0

The 4 callers of this function are from hashing &url::Url (hash_one -> 26 296 times), hashing &semver::Version (hash_one -> 83 342x), hashing package_id::PackageId (hash -> 48 554x) and dependency::Dependency (hash -> 31 099x)

To reproduce

cd bumpalo-3.16.0
CARGO_TARGET_DIR=/tmp/m1 valgrind --tool=callgrind --dump-instr=yes --collect-jumps=yes --trace-children=yes /home/meow/.rustup/toolchains/nightly-2025-05-31-x86_64-unknown-linux-gnu/bin/cargo clippy

Note that this is without incremental compilation, thus the CARGO_TARGET_DIR env var.

I'm not sure if there's a reason why Cargo uses the default hasher.

[weihanglo]

Cargo leverages https://github.com/rust-lang/rustc-stable-hash, as we want path on file system as stable as possible, regardless endianness or platform. Tha said, not everything in this codebase has this requirement There are some previous attempts but got stale
#14665

Edit: Oh I see. You're talking about siphash from std not from rustc-stable-hash. The thing about stable hash may be irrelevant here.

[weihanglo]

I wonder how bad it was in terms of total run time of the clippy command being benchmarked.

[blyxyas]

On this crate, it seems to represent a 9.89% of the runtime, with 14.155.123 instructions.

[epage]

If we do this, a question is which hash. Currently, Cargo depends on ahash through hashbrown@0.14. However, hashbrown@0.15 has switched to foldhash ( rust-lang/hashbrown#563). For toml, I'm following suit in toml-rs/toml#956

[blyxyas]

On this crate, it seems to represent a 9.89% of the runtime, with 14.155.123 instructions.

I want to add a note here, part of the Cargo runtime, Clippy having so many lints, these 14 million instructions is not that relevant. I'm talking from the Cargo process point of view.