On Wed, 4 Feb 2026 12:47:54 GMT, Tobias Hartmann <[email protected]> wrote:
> Old mainline fix [JDK-8134389](https://bugs.openjdk.org/browse/JDK-8134389) > modified code in `Invoke::declared_type` to not use the return type of the > target signature but use the declared signature instead (see > [changeset](https://hg.openjdk.org/jdk9/jdk9/hotspot/rev/4191b33b3629) for > details). This was needed for method handle call sites to return > `java/lang/invoke/CallSite` instead of `java/lang/Object` or whatever the > target method return type would look like. The test > `compiler/jsr292/TestArrayReturnType.java` still reliably triggers this > scenario. > > Now the problem is that the return type is re-computed on every call to > `Invoke::declared_type` but it might actually change from unloaded to loaded > in-between and therefore different `ciType` objects will be returned. In the > failing case, the return type first is > `<ciObjArrayKlass name=[Ljava/lang/reflect/Field; loaded=false ident=1498 > address=0x00007ae57d775728>` > and later it's `<ciObjArrayKlass name=[Ljava/lang/reflect/Field; loaded=true > ident=1508 address=0x00007ae57d7762a8>` (note the different `loaded` > property). This leads to an assertion failure in profiling because we invoke > CHA to improve a type that's already final because when we checked the type > earlier, it was still unloaded. > > I think it's a general rule that the compiler interface needs to guarantee > consistency of types during compilation and therefore I adjusted the code to > only compute the return type once. This is a mainline issue and could be > fixed in mainline as well. However, as far as I can tell, only Valhalla is > currently affected because the Valhalla specific array load/store profiling > uses `LIRGenerator::profile_type` in more cases. Let's therefore only fix it > in Valhalla for now. > > Thanks, > Tobias Thanks for the review Christian! ------------- PR Comment: https://git.openjdk.org/valhalla/pull/2031#issuecomment-3848198664
