Given a @Configuration class named org.example.AppConfig which
contains @Bean methods, in Spring Framework 5.3.x and previous
versions, the following classes were created when generating the CGLIB
proxy.
org.example.AppConfig$$EnhancerBySpringCGLIB$$fd7e9baa
org.example.AppConfig$$FastClassBySpringCGLIB$$3fec86e
org.example.AppConfig$$EnhancerBySpringCGLIB$$fd7e9baa$$FastClassBySpringCGLIB$$82534900
Those class names indicate that 1 class was generated for the proxy for
the @Configuration class itself and that 2 additional FastClass
classes were generated to support proxying of @Bean methods in
superclasses.
However, since Spring Framework 6.0, the following classes are created
when generating the CGLIB proxy.
org.example.AppConfig$$SpringCGLIB$$0
org.example.AppConfig$$SpringCGLIB$$1
org.example.AppConfig$$SpringCGLIB$$2
The above class names make it appear that 3 proxy classes are generated
for each @Configuration class, which is misleading.
To address that and to align more closely with how such generated
classes were named in previous versions of the framework, this commit
modifies SpringNamingPolicy so that generated class names once again
include "FastClass" when the generated class is for a CGLIB FastClass
as opposed to the actual proxy for the @Configuration class.
Consequently, with this commit the following classes are created when
generating the CGLIB proxy.
org.example.AppConfig$$SpringCGLIB$$0
org.example.AppConfig$$SpringCGLIB$$FastClass$$0
org.example.AppConfig$$SpringCGLIB$$FastClass$$1
Closes gh-31272
The introduction of AdvisedSupport.AdvisorKeyEntry in Spring Framework
6.0.10 resulted in a regression regarding caching of CGLIB generated
proxy classes. Specifically, equality checks for the proxy class cache
became based partially on identity rather than equivalence. For
example, if an ApplicationContext was configured to create a
class-based @Transactional proxy, a second attempt to create the
ApplicationContext resulted in a duplicate proxy class for the same
@Transactional component.
On the JVM this went unnoticed; however, when running Spring
integration tests within a native image, if a test made use of
@DirtiesContext, a second attempt to create the test
ApplicationContext resulted in an exception stating, "CGLIB runtime
enhancement not supported on native image." This is because Test AOT
processing only refreshes a test ApplicationContext once, and the
duplicate CGLIB proxy classes are only requested in subsequent
refreshes of the same ApplicationContext which means that duplicate
proxy classes are not tracked during AOT processing and consequently
not included in a native image.
This commit addresses this regression as follows.
- AdvisedSupport.AdvisorKeyEntry is now based on the toString()
representations of the ClassFilter and MethodMatcher in the
corresponding Pointcut instead of the filter's and matcher's
identities.
- Due to the above changes to AdvisorKeyEntry, ClassFilter and
MethodMatcher implementations are now required to implement equals(),
hashCode(), AND toString().
- Consequently, the following now include proper equals(), hashCode(),
and toString() implementations.
- CacheOperationSourcePointcut
- TransactionAttributeSourcePointcut
- PerTargetInstantiationModelPointcut
Closes gh-31238
After further consideration, the team has decided to remove the
getAutodetectMode() method since its return type conflicts with the
parameter type in setAutodetectMode(int), making it an invalid bean
property.
See gh-30855
Prior to this commit, MBeanExporter used
org.springframework.core.Constants which used reflection to find
constant fields in the MBeanExporter class. Consequently, one had to
register reflection hints in order to use MBeanExporter in a GraalVM
native image.
This commit addresses this by replacing the use of the `Constants`
class with a simple java.util.Map which maps constant names to constant
values for the autodetect constants defined in MBeanExporter.
See gh-30851
Closes gh-30846
This merges the existing support for the legacy JSR-250 PostConstruct/PreDestroy annotations into CommonAnnotationBeanPostProcessor itself, opening up the InitDestroyAnnotationBeanPostProcessor base class for multiple init/destroy methods in a single post-processor. This removes the need for a separate JSR-250 InitDestroyAnnotationBeanPostProcessor in AnnotationConfigUtils.
Closes gh-30695
Prior to this commit, private (and non-visible package-private)
init/destroy methods were not supported in AOT mode. The reason is that
such methods are tracked using their fully-qualified method names, and
the AOT support for init/destroy methods previously did not take
fully-qualified method names into account. In addition, the invocation
order of init/destroy methods differed vastly between standard JVM mode
and AOT mode.
This commit addresses these issues in the following ways.
- AbstractAutowireCapableBeanFactory.invokeCustomInitMethod(),
DisposableBeanAdapter.determineDestroyMethod(), and
BeanDefinitionPropertiesCodeGenerator.addInitDestroyHint() now parse
fully-qualified method names to locate the correct init/destroy
methods.
- AbstractAutowireCapableBeanFactory and DisposableBeanAdapter delegate
to a new MethodDescriptor record which encapsulates the parsing of
fully-qualified method names; however,
BeanDefinitionPropertiesCodeGenerator duplicates this logic since it
resides in a different package, and we do not currently want to make
MethodDescriptor public.
- Init/destroy methods detected via annotations (such as @PostConstruct
and @PreDestroy) are now invoked prior to init/destroy methods that
are explicitly configured by name or convention. This aligns with the
invocation order in standard JVM mode; however,
InitializingBean#afterPropertiesSet() and DisposableBean#destroy()
are still invoked before annotated init/destroy methods in AOT mode
which differs from standard JVM mode.
- Unit and integration tests have been updated to test the revised
behavior.
Closes gh-30692
Prior to this commit, if an init/destroy method was package-private and
declared in a superclass in a package different from the package in
which the registered bean resided, a local init/destroy method with the
same name would effectively "shadow" the method from the different
package, resulting in only the local init/destroy method being invoked.
This commit addresses this issue by tracking package-private init
methods from different packages using their fully-qualified method
names, analogous to the existing support for private init/destroy
methods.
Closes gh-30718
Previously, a bean definition that is optimized AOT could have
different metadata based on whether its resolved type had a generic or
not. This is due to RootBeanDefinition taking either a Class or a
ResolvableType doing fundamentally different things. While the former
sets the bean class which is to little use with an instance supplier,
the latter specifies the target type of the bean.
This commit sets the target type of the bean, using the existing
setter methods that take either a class or a ResolvableType and set the
same attribute consistently.
Closes gh-30689
Prior to this commit, if a non-annotated [1] class was registered
directly with an ApplicationContext -- for example, via
AnnotatedBeanDefinitionReader, AnnotationConfigApplicationContext, or
@ContextConfiguration -- it was not considered a configuration class in
'lite' mode unless @Bean methods were declared locally. In other words,
if the registered class didn't declare local @Bean methods but rather
extended a class that declared @Bean methods, then the registered class
was not parsed as a @Configuration class in 'lite' mode, and the @Bean
methods were ignored.
Whereas, a non-annotated class registered via @Import is always
considered to be a configuration class candidate.
To address this discrepancy between @Import'ed classes and classes
registered directly with an ApplicationContext, this commit treats any
class registered via AnnotatedBeanDefinitionReader as a @Configuration
class candidate with @Bean 'lite' mode semantics.
[1] In this context, "non-annotated" means a class not annotated (or
meta-annotated) with @Component, @ComponentScan, @Import, or
@ImportResource.
Closes gh-30449
Restores proper event type propagation to parent context.
Selectively applies payload type to given payload object.
Also reuses cached type for regular ApplicationEvent now.
Closes gh-30360
This commit fixes the check by avoiding a fallback to eventType's
hasUnresolvableGenerics(). This could previously lead to checking a
generic event type `A<T>` against a listener which accepts unrelated
`B` and return `true` despite the inconsistency.
Note that this wouldn't necessarily surface to the user because there is
a `catch (ClassCastException e)` down the line, which was primarily put
in place to deal with lambda-based listeners but happens to catch an
exception thrown due to the bad result of `supportsEventType`.
The `supportsEventType` now matches generic `PayloadApplicationEvent`
types with a raw counterpart, using the above fallback only in that case
and otherwise ultimately returning `false`.
Closes gh-30399
Sam Brannen
Juergen Hoeller
Sébastien Deleuze
Johnny Lim
Stephane Nicoll
Simon Baslé