On Mon, 1 May 2023 21:00:40 GMT, John R Rose <jr...@openjdk.org> wrote:
> HC fields are supposed to fold up; if they don’t we can help them in this > case. (That is, constant folding of the MH pointer should not be a roadblock; > that’s on the JIT to fix by adjusting the TrustFinalFields logic.) The issue with this is that constant folding of a field only happens when to holder object is also constant, which is not necessarily the case. For the non-constant case, we have to rely on type profiling to get optimal performance. (recapping) Let's say we have a call site to a method `foo` in interface `I`. But, the receiver value we see at this call site is only ever an instance of a class `C`. The JIT can see from the profile of that call site that we only ever see a receiver of type `C`, and then do a guarded inline of `C::foo`, giving us optimal performance. This works because the type `C` has a 1-to-1 mapping with the implementation that we inline into the call site. However, in the case of method handles, there is no 1-to-1 mapping between the type of the receiver, `MethodHandle`, and the target method. Instead, the target method depends on a particular receiver _instance_. So, to make type profiling work, we need to introduce a 1-to-1 mapping with a type, hence we define a hidden class HC per method handle instance, where the MH instance is constant for that particular class. This creates the 1-to-1 mapping. Type profiling works for the HC type we define, we inline `HC::foo`, and find that: hey, in HC::foo's implementation, the target MH is constant, so we can inline through that as well. Giving us once again, optimal performance. I've considered in the past that we could add _instance_ based profiling for MethodHandles, so that we can once again get a 1-to-1 mapping between a profile and a method we want to inline. But it wouldn't work in the case of MHP::asInterfaceInstance. Since profiles are tied to a BCI, it wouldn't work if we create multiple instances of the same HC for different MH instances, since the profile of the `invokeExact` call site we generate in the HC would be polluted by the different MH instances. (unless we implement: https://bugs.openjdk.org/browse/JDK-8015416) ------------- PR Comment: https://git.openjdk.org/jdk/pull/13197#issuecomment-1530442581
