Benchmarks: a changed benchmark runs only the benchmarks of its modules

[alex]

ci/bench_arches.py treats every change under bench/ as shared, so it runs every benchmark on every architecture. A full run now takes about 14 minutes on the aarch64 runners and more than the job's 15-minute timeout on the others. Two changes:

1. A changed benchmark runs only its own modules' benchmarks. Adding an algorithm to an architecture always edits that algorithm's benchmark, because check_arch_gates.py requires its cfg to match the library module's. Such a PR therefore ran the whole suite in all nine configurations, and the jobs were cancelled at the timeout. That is what happened to the AES-CMAC PRs Implement AES-CMAC on AArch64, with an AES-extension variant #469, Implement AES-CMAC on ARMv7 #477 and Implement AES-CMAC on x86 #485, whose only benchmark change is the cfg line.

   Now a changed bench/benches/primitives/<name>.rs narrows the run to the modules its USES lists, on every architecture. Those modules' benchmarks include the changed one. For example, cmac_aes.rs (USES = ["cmac_aes", "aes"]) now runs the CMAC and AES-GCM benchmarks. Everything still runs for:

   • a change to main.rs;
   • a benchmark whose USES can't be read, such as a deleted one;
   • any other file under bench/.

   A narrowed run never benchmarks less than a change's library modules need: the narrowing only replaces "everything" with the changed benchmark's own modules.

2. The compare job's timeout goes from 15 to 30 minutes. Shared changes (e.g. Cargo.lock, src/lib.rs, the comparison scripts) and runs by hand still benchmark everything. On this PR's own first run, every x86-64, x86 and ARMv7 job was cancelled at 15 minutes, and the aarch64 ones finished in about 14. If maintainers would rather keep runner time down, the alternative is to cut the number of rounds bench_compare.py runs for a full run.

Testing

I ran git diff --name-only … | python3 ci/bench_arches.py on file lists:

• The diff of Implement AES-CMAC on x86 #485 now gives aes cmac_aes on each architecture, instead of an empty modules (everything).
• bench/benches/primitives/main.rs and a missing benchmark file still give everything.
• --all is unchanged (nine entries).

This PR's own Benchmarks run still runs everything, because it changes ci/bench_arches.py and bench.yml, which are shared. The run on the second commit shows whether 30 minutes is enough.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WkLN6tAYk76HACiEWLbAMD