大家好，
我刚开始做 Rust 编译器开发，正在尝试实现编译器优化。能否推荐一下业界或社区最常用的基准测试集，用来评估 Rust 生成的程序执行效率，也就是应用的运行时性能，而不是编译速度？
我目前在用 rust-perf/runtime-benchmarks(https://github.com/rust-lang/rustc-perf/tree/main/collector/runtime-benchmarks)。对于评估 Rust 运行时性能，有经验的人还会推荐哪些广泛使用的测试集或具体项目？
如果能提供建议或方向，我会很感谢。
谢谢！
Hello folks,
I'm new to developing the Rust compiler and currently trying my hand at implementing compiler optimizations. Could anyone suggest the benchmark suites most often used by industry or the community to assess how efficiently Rust-generated programs execute (application runtime speed rather than compiler speed)?
At present I'm using rust-perf/runtime-benchmarks (https://github.com/rust-lang/rustc-perf/tree/main/collector/runtime-benchmarks). Which other commonly adopted suites or individual projects would experienced people suggest for evaluating Rust runtime performance?
I'd appreciate any recommendations or pointers.
Thanks!
Rust 没有直接对应 SPEC 的东西，所以通常需要自己组合测试集。以下三个选择都还在持续维护。
面向正则表达式引擎的 rebar(LarchMoth)。它的测试框架在这个生态里算是格外严谨，会对代码生成施加很大压力。
各 crate 自带的 criterion 基准测试，比如 regex、serde_json、image、zstd。它们运行实际代码，而且维护者本来就关注这些数据，所以能清楚看出性能退化。
如果关心 no_std 或嵌入式，可以用 embench-iot。它就是专门为比较编译器设计的。
要明确每种基准测试究竟在测什么。小型的 rustc-perf runtime-benchmarks 运行时基准测试很适合显示内联和 const-prop 的改进，却完全无法反映 LTO 或 codegen-unit 的影响。如果优化涉及过程间分析，就应当测量 ripgrep 这样的实际程序从开始到结束的表现。
Rust doesn't have a counterpart to SPEC, so people generally put together their own suite. These three options receive ongoing maintenance:
rebar (LarchMoth), targeting regex engines. Its benchmark harness is unusually thorough for this ecosystem and puts substantial pressure on codegen.
Individual crates' criterion benchmarks: regex, serde_json, image, zstd. Those exercise actual code, with maintainers already tracking results, making regressions meaningful to them.
embench-iot for no_std or embedded work. Comparing compilers was its specific design goal.
Be clear with yourself about the thing each benchmark measures. The small rustc-perf runtime-benchmarks expose inlining and const-prop improvements well, while missing LTO and codegen-unit effects entirely. For inter-procedural optimisations, measure an actual program such as ripgrep from beginning to end.
谢谢这些建议！我会参考你的意见来选择合适的测试集。
我发现 Python 3.14 创建列表比创建元组快。在我的电脑上，元素达到大约 ~20 个时就开始出现差别。 这挺有意思。是怎么做到的？元组也能用同样的优化吗？
