We have a pretty large ApplicationContext with about 7600 Beans and startup times have become an issue, now at 5.5 min just for the ApplicationContext. We observed that all the steps in AbstractApplicationContext#refresh before #preInstantiateSingletons() is called take about 90 seconds (excluding the time it takes to load all bean definition XML files). About 81 seconds of that time seems to be spent in the method DefaultListableBeanFactory#getBeanNamesForType.
In order to better understand where time is spent, we've build a wrapper around DefaultListableBeanFactory that measures the execution time of calls to #createBean, #getBeanNamesForType and #applyBeanPostProcessors*Initialization/Instantiation. This is the tooling we used to get the numbers mentioned above.
Closer inspection shows that runtime performance of #getBeanNamesForType depends heavily on the flag #isConfigurationFrozen - as long as it is set to false during AppCtx refresh, each call takes about 10 seconds, when it is set to true, calls typically take only ~100 ms.
A particular issue seems to me, that #getBeanNamesForType is called at least three times for the type BeanDefinitionRegistryPostProcessor, so that's more than 30 seconds in our scenario. I found that this has been implemented as part of #15258, and comments in the code of PostProcessorRegistrationDelegate indicate that calling it three times is "by design". But that seems like a waste of time to me, given that these calls take rather long.
Please let's discuss which optimizations would be feasible here within the boundaries of the involved interface contracts.
Fair enough, with that size of context, those calls may take painfully long. We'll revisit this for 4.3, trying to optimize based on the metadata that we have at each given point... but also exploring options for explicit hints in the application's configuration, allowing us to bypass entire steps (e.g. if we know that there is nothing to find for this particular application anyway). We may even go as far as using pre-computed indexes for certain steps, which has been suggested in other cases before; not sure whether that part can happen for 4.3 though.