Store only weaved class names in the generatedClasses map
The generatedClasses map contained both keys of woven classes that had
generated classes associated with them, and the keys of the generated
classes themselves. It seems like this map is never consulted for the
generated class key - the generated class is generated from within the
context of woven class loading / retransformation, and thus no transform
event is generated for it by the JVM. Because of that, we do not need to
guard against re-weaving it. Other uses of generatedClasses map are for
full/empty tests, where the woven class key is sufficient. This change
simplifies deletion of a class because we do not have to look for its
associated generated classes.
Relates to #279.
WeavingAdaptor.couldWeave() tested two things: whether this specific
class (by its name) should be excluded from weaving, and whether the
class has generated classes associated with it (in which case we avoid
weaving to avoid re-creating those generated classes). However, if the
class has generated classes associated, couldWeave() is not called at
all because of the new wovenWithGeneratedClass() test, so we have
only the name test which is now called directly.
Relates to #279.
Use previous weaving result if classes were generated during weaving
A parallel-capable class loader might load the same class multiple times
in multiple threads, see
https://docs.oracle.com/javase/7/docs/technotes/guides/lang/cl-mt.html.
In this scenario, the JVM instrumentation mechanism invokes the AspectJ
weaver twice on the same class. In simple weaving scenarios, this would
just cause weaving logic to run multiple times and produce the same
result for each thread, but if the weaver is generating new classes,
it is undesirable to generate those classes multiple times.
To avoid this, the generatedClasses map in WeavingAdaptor keeps track of
classes for which inner classes were generated and avoids re-weaving
them. However, before this change, if a class was found in the generated
map, the weaver would return the un-woven bytes, which would cause
AspectJ not to work for some of the threads.
This change returns the woven bytes instead of re-running weaving.
Fixes #279.
Change value of woven class in generatedClasses map
Store the woven class and not a generated (inner) class as value of the
woven class name. This has no effect in this commit, because the value
is not consulted at all, but will be used later on.
Relates to #279.
Weaver returns null instead of original bytes for unwoven classes
This change makes sense independently of #277, but also enables using
- cp "my.jar;aspectjweaver.jar"
-XX:+AllowArchivingWithJavaAgent
-javaagent:aspectjweaver.jar
while creating a CDS archive. Afterward, the application can be run in
its woven state from the CDS archive even without '-javaagent', because
the byte code was archived in its woven state ("poor man's AJC"). See
https://github.com/eclipse-aspectj/aspectj/issues/277#issuecomment-1931142753
for details.
Fixes #277.
Signed-off-by: Alexander Kriegisch <Alexander@Kriegisch.name>
Overhaul ClassLoaderWeavingAdaptor to use statically initialised Unsafe
instances and method handles pointing to their 'defineClass' methods.
Those now work universally on JDKs 8-21. In older JDKs, the method used
to be in sun.misc.Unsafe, in more recent ones on jdk.internal.misc.Unsafe.
It is challenging to fetch instances, especially as reflection
protection and module boundaries have been increased in the JDK
progressively. But finally, a solution was adapted from Byte Buddy (BB).
Kudos to BB author Rafael Winterhalter. The previous solution to use
ClassLoader::defineClass and require '--add-opens' is no longer
necessary for the first time since it became necessary in AspectJ 1.9.7
with Java 16 support.
Add org.ow2.asm:asm-common as a dependency everywhere org.ow2.asm:asm
was used before. Maybe that is too many places, but no worse than before.
Add missing dependency on loadtime to aspectjweaver. This kept a build
like "mvn install -am -pl aspectjweaver" from picking up changed
loadtime classes.
Fixes #117.
Signed-off-by: Alexander Kriegisch <Alexander@Kriegisch.name>
Currently, the situation looks more like a Java 21 maintenance release
than directly a Java 22 release.
Signed-off-by: Alexander Kriegisch <Alexander@Kriegisch.name>
Improve error message in LazyClassGen.getJavaClassBytesIncludingReweavable
No longer report "Unable to find ASM classes", if simply an ASM
processing error occurred. In that case, report "Error processing
class file".
Relates to #250.
Signed-off-by: Alexander Kriegisch <Alexander@Kriegisch.name>
Fix inline accessor method bug in BcelAccessForInlineMunger
Make sure to create one ajc$inlineAccessMethod per identically named
(overloaded) private aspect method in
BcelAccessForInlineMunger.createOrGetInlineAccessorForMethod.
Fixes #250.
Signed-off-by: Alexander Kriegisch <Alexander@Kriegisch.name>
BcelTypeMunger.mungeMethodDelegate: only use 'synchronized' when necessary
Relates to #198. Now, we create a delegate method body which basically
looks as follows:
public void methodOne() {
if (this.ajc$instance$MyAspect$MyMixin == null) {
synchronized(this) {
if (this.ajc$instance$MyAspect$MyMixin == null) {
this.ajc$instance$MyAspect$MyMixin = MyAspect.aspectOf().createImplementation(this);
}
}
}
this.ajc$instance$MyAspect$MyMixin.methodOne();
}
The idea for the outer null check is from @aclement, see
https://github.com/eclipse/org.aspectj/pull/205#issuecomment-1371556080.
Signed-off-by: Alexander Kriegisch <Alexander@Kriegisch.name>
Improve BcelTypeMunger.mungeMethodDelegate to avoid race condition
Fixes #198, i.e. test DeclareMixinTests.testCaseEConcurrent from the
previous commit now passes.
Signed-off-by: Alexander Kriegisch <Alexander@Kriegisch.name>
Fix #366085 concerning declared annotations with source retention
See https://bugs.eclipse.org/bugs/show_bug.cgi?id=366085.
See https://stackoverflow.com/q/74618269/1082681.
The issue described in the Bugzilla issue is about 'declare @type', but
similar issues also existed for 'declare @field', 'declare @method',
'declare @constructor'. This fix is rather superficial and leaves
things to be desired, because it is rather hacky and simply ignores
errors source retention annotation declarations during weaving. A better
fix would drop the corresponding declarations while parsing and also
issue compiler warnings in each case.
Signed-off-by: Alexander Kriegisch <Alexander@Kriegisch.name>
Optimize class loading - make backward compatible with legacy behaviour
Now the defaults are:
+ org.aspectj.apache.bcel.useSingleRepositoryInstance (default: false)
+ org.aspectj.apache.bcel.useUnavailableClassesCache (default: false)
+ org.aspectj.apache.bcel.ignoreCacheClearRequests (default: false)
I.e. the new caching optimisations are opt-in instead of opt-out as
originally designed. This might change in the future, but for now
without any additional tests and experience with the new feature let us
be conservative and make the build green first.
I also added a few more code review findings concerning backward
compatibility, which was less than 100% even with all three flags
deactivated.
Signed-off-by: Alexander Kriegisch <Alexander@Kriegisch.name>
In our project we found out that during the build up of the spring context
the class loading takes a very long time.
Root cause is the huge amount of file I/O during pointcut class loading.
We are taking about ~250k file loads.
With these changes we managed to cut down the starting time by around 50%.
What we found out is that IMHO - the clear method of the ClassLoaderRepository
is called far too often -> in our settings this resulted in not a single cache
hit as the cache got cleared permanently.
Therefore we de-actived the cache clear calls inside the ClassLoaderRepository.
Secondly we changed the Java15AnnotationFinder in a way to not always create
new objects for the ClassLoaderRepository but re-use one static instance.
Otherwise we experienced >100k objects being created.
Last but not least we introduced a cache for unavailable classes so that
they do not have to be looked up using file I/O over and over again.
The whole behavior is configurable via
+ org.aspectj.apache.bcel.useSingleRepositoryInstance (default: true)
+ org.aspectj.apache.bcel.useUnavailableClassesCache (default: true)
+ org.aspectj.apache.bcel.ignoreCacheClearRequests (default: true)
Signed-off-by: Stefan Starke <stefan@starkeweb.org>
Signed-off-by: Alexander Kriegisch <Alexander@Kriegisch.name>
Alexander Kriegisch
Uri Simchoni
Andy Russell
Andrey Turbanov
Andy Clement
Stefan Starke