On Wed, 2 Dec 2020 15:30:07 GMT, Peter Levart <[email protected]> wrote:
>>> Would it make a difference if MH.form was not final and each read access to >>> it was done via appropriate Unsafe.getReferenceXXX()? >> >> It would break inlining through MH calls. JITs trust `MH.form` and >> aggressively inline through it. >> >>>I was just concerned that optimization in one part (less resources consumed >>>updating the form) would increase the probability of JIT-ed code using the >>>old form indefinitely - therefore causing regression in top performance. >> >> That's expected and happens in practice. It was a deliberate choice to avoid >> invalidating existing code and triggering recompilations while sacrificing >> some performance. >> >> But if we focus on MH customization, there's no inlining happening (or >> possible): customization is performed on a non-constant (in JIT sense) MH >> instance which is about to be invoked through `MH.invoke()/invokeExact()`. >> So, subsequent calls through invoker on the same (non-constant) MH instance >> should see updated `MH.form` value (customized LambdaForm): >> >> https://github.com/openjdk/jdk/blob/master/src/java.base/share/classes/java/lang/invoke/Invokers.java#L598 >> @ForceInline >> /*non-public*/ >> static void checkCustomized(MethodHandle mh) { >> if (MethodHandleImpl.isCompileConstant(mh)) { >> return; // no need to customize a MH when the instance is known >> to JIT >> } >> if (mh.form.customized == null) { // fast approximate check that the >> underlying form is already customized >> maybeCustomize(mh); // marked w/ @DontInline >> } >> } > > Ah, I see. Customization only happens for non-constant MHs. Everything is > fine then. Sorry for confusion. Thanks a lot for such a thorough review, Peter! ------------- PR: https://git.openjdk.java.net/jdk/pull/1472
