Hi all, I’ve completed the benchmarks based on the guidance in the thread.
In short: Cold-start measurements (fresh PHP process per run) show no meaningful difference between direct instantiation and reflection-based autowiring. Process startup dominates, and reflection overhead is effectively lost in the noise. Warm-process benchmarks show a consistent ~3-4x overhead for runtime autowiring compared to direct instantiation, even with request-scope constructor plan caching. This appears to be a model-level cost rather than something easily eliminated by implementation details alone. Based on these results, I don’t see a lazy-proxy-like engine-only win here. Userland implementations already seem close to the practical ceiling for runtime autowiring, and further gains would likely be incremental. Given that, I agree this does not currently make a compelling case for a core addition, and that runtime autowiring remains better addressed in userland or framework-specific solutions. Thanks to everyone who provided feedback and guidance. Best regards, Zoli eng. ANDRÁS Zoltán-Gyárfás --------------------------------------- tel: +40 745 797 798 mail: [email protected] > On 30 Dec 2025, at 20:41, Marco Pivetta <[email protected]> wrote: > > Hey Zoli, > > On Tue, 30 Dec 2025 at 08:17, ANDRÁS Zoltán Gyárfás <[email protected] > <mailto:[email protected]>> wrote: >> If you have any guidance on what you would consider a representative >> benchmark or a meaningful performance threshold for “worth moving into >> core”, I’d appreciate it. >> > > In the grand majority of cases where autowiring is meaningful (at least for > PHP), you care about a "cold start" of the application (although that's > changing with ReactPHP/FrankenPHP/Roadrunner/etc.). > > You may design a PHPBench or Hyperfine test that: > > * starts a PHP process, autowires something, shuts down. Hyperfine helps a > lot here. > * repeats the above, but within a warm process (classes already loaded, tight > loop repeating the operation). PHPBench is a bit more precise here. > > You probably care about these cases: > > * zero-arguments constructor > * simple concrete class references > * default constructor arguments > * resolvers > * autowiring of a complex multi-level structure > > In all the above cases, you probably care about the usage with autowiring API > vs the usage via direct instantiation, so you can see the exact overhead. > > Marco Pivetta > > https://mastodon.social/@ocramius > > https://ocramius.github.io/ >
