[Nestmates] RFR (S): 8197915: [Nestmates] Implement receiver typecheck for private invokeinterface use (original) (raw)

David Holmes david.holmes at oracle.com
Mon May 7 01:50:47 UTC 2018

I've worked out the difference in the tests. The jasm version used the wrong REFC: Object instead of the interface type.

Fixing assertions and updating tests.

David

On 7/05/2018 9:05 AM, David Holmes wrote:

Hi Karen,

First, outside of nestmates I've filed a bug (8202686) and send out a RFR to add the missing testcase for final Object methods to the test for 8200167. It doesn't show any issues of course. Next I've taken the additional testcases and moved them into the PrivateInterfaceCall test - adapted for invokeinterface - which should cover the test you wrote below ... however ... On 5/05/2018 8:04 AM, David Holmes wrote: Hi Karen,

On 5/05/2018 5:54 AM, Karen Kinnear wrote: David,

Putting together a wiki to describe how I think this works with cpCache and with MethodHandles. Not yet done … In the process of testing cases - I found a couple of assertions in the nestmate repo that are not accurate: 1. linkResolver.cpp: #  assert(resolvedmethod()->isprivate()) failed: Should only have non-virtual invokeinterface for private methods! Yes this overlooked that final Object methods can also follow this path. It is fixed by simply extending the assert to include "or is a final Object method". This was triggered by both a direct call attempt for a final Object method and a MH invocation of same.

2. ConstantPoolCacheEntry::setdirectorvtablecall invokeinterface asserts isprivate None of my testing hit this assertion failure. Yet your test (which I essentially copied) does. This is very puzzling. Further if I suppress that assert then I hit: #  Internal Error (/export/users/dh198349/valhalla/repos/valhalla-dev/open/src/hotspot/share/oops/cpCache.cpp:276), pid=8649, tid=8650 #  Error: assert(invokecode == Bytecodes::invokevirtual || (method->isprivate() && invokecode == Bytecodes::invokeinterface)) failed This is all easily fixed, but the test scenarios need more investigation. Not only does your direct invocation test trigger the above assertions where mine does not; my test fails due to: IncompatibleClassChangeError: Found class java.lang.Object, but interface was expected but yours does not! The only difference I can see is that your test has the call in a class, whereas mine has it in an interface. Thanks, David I wrote a small test (sorry - I patched the bytecodes to do this quickly) which has invokeinterface I.getClass()  // javac put invokevirtual when I tried to get it to generate that Great catch! Another variant of the "invoking object methods via invokeinterface" problem - the final method case. The asserts in principle need to weaken to "or is an Object method". This isn’t the methodHandles, this is just the straight bytecodes - but it is part of the decision tree of are we using RefinvokeSpecial. Not sure how the bytecode issue relates at all to the MH logic? But of course we have to try and construct a MH version of the direct invoke as well. Will tackle this Monday. Many thanks, David

I attached the test - it is built for jdk10 so I could test before and after. If you recompile Test.java it will need repatching. Since getClass is final, it also goes through the directcall route for invokeinterface. thanks, Karen

On May 4, 2018, at 4:22 AM, David Holmes <david.holmes at oracle.com_ _<mailto:david.holmes at oracle.com>> wrote: Thanks Vladimir! David On 4/05/2018 6:10 PM, Vladimir Ivanov wrote: Ok. webrev updated to v3: http://cr.openjdk.java.net/~dholmes/8197915/webrev.v3/ Looks good! Best regards, Vladimir Ivanov New code: // if caller is an interface we need to adapt to get the // receiver check inserted if (callerClass == null) { throw new InternalError("callerClass must not be null for REFinvokeSpecial"); } LambdaForm lform = preparedLambdaForm(member, callerClass.isInterface()); return new Special(mtype, lform, member, callerClass); Thanks, David Best regards, Vladimir Ivanov

Changes: - DirectMethodHandles.java: new simple and direct approach to dealing with LFSPECIALIFC I like how java.lang.invoke part shapes out! Maybe rename adaptToSpecialIfc to needsReceiverCheck? That's what confused me in the first version: though it's an interface call (which always require receiver check against REFC), new checks only referred to   LFINVSPECIAL (since invocation mode is a direct call). - New regression test for the final virtual call from an interface bug introduced by 8200167. If necessary/desirable I can fix that part in mainline separately. So far no tests (including jck/API/java/lang) seem to tickle it. Or file a bug. I have some ideas how to improve relevant code and make LF construction cleaner. Best regards, Vladimir Ivanov

Thanks, David On 4/05/2018 11:41 AM, David Holmes wrote: Hi Karen, On 4/05/2018 6:39 AM, Karen Kinnear wrote: David, Really delighted to see you near the end of the major functional changes! Thanks for taking a look so quickly! A couple minor comments, and then a question please: 1. MethodHandles.java DirectMethodHandle.java :) 174 different “to” -> different “from” ? Changed. That's my UK upbringing :) https://en.oxforddictionaries.com/usage/different-from-than-or-to

2. methodHandles.cpp 300-301 Thank you for the comment. Might it also be worth adding that direct call is used by: invoke static, invokespecial, invokeinterface:local private, invoke virtual:vfinal and private methods (or are you concerned about getting out of sync if this changes?) It is not used by invokestatic. I'm not 100% sure of all the exact cases where an invokeinterface/invokevirtual becomes a direct call, so didn't want to say anything inaccurate. But the comment as it stands is awkward so I've expanded it: // "special" reflects that this is a direct call, not that it // necessarily originates from an invokespecial. We can also do // direct calls for private and/or final non-static methods. 3. DirectMethodHandle.java - this was subtle! More than you realise ;-) I believe this is correct assuming that: CallerClass is always and only set for invokespecial. Is this accurate? Could you possibly add a comment? That's an excellent question and one that should have been asked before 8200167 was finalized. :(  The short answer is "no" - callerClass can be non-null for any of the invocation modes. And yes the current mainline code is broken - seems there is a gap in the existing test coverage as we never call a final method from an interface method. If we do we get: Exception in thread "main" java.lang.InternalError: Should only be invoked on a subclass at java.base/java.lang.invoke.DirectMethodHandle.checkReceiver(DirectMethodHandle.java:441) We only look at callerClass when dealing with LFINVSPECIAL, which in mainline means we either have an invokespecial or an invokevirtual. For invokespecial this is fine of course. But the invokevirtual case was never encountered and so slipped by in error. With nestmates we also add invokeinterface to the mix - which is fine because if it is an invokeinterface then we want the check regardless. It doesn't matter if the check is enabled because of the (incidental) callerClass.isInterface check, or the explicit m.getDeclaringClass().isInterface(). But the logic is messy and far from clear and not correct by construction. So I will completely redo it in a simpler and more direct/explicit way. BTW another red-herring: the !m.isStatic() part of the condition was not needed. I was tracking down two failure modes before finalizing this. The first was a problem with a static interface method - fixed by the !m.isStatic(). The second was caused by missing parentheses in the overall condition - which once fixed precluded the static case, so the first fix was not needed (as we never use LFINVSPECIAL with statics). If only I'd tackled them in the reverse order. I'll post an updated webrev later today once I've re-tested lots of things. - agree with the theory that invokevirtual will never find a private interface method (and ACCFINAL is illegal for interfaces) Yes. More specifically as we're dealing with MH semantics: findVirtual for an interface method yields a MH with invokeInterface "kind", not one with invokeVirtual "kind". public MethodHandle findVirtual(Class<?> refc, String name, MethodType type) throws NoSuchMethodException, IllegalAccessException { ... byte refKind = (refc.isInterface() ? REFinvokeInterface : REFinvokeVirtual); ... } 4. Test - I still need to study this I have been writing down test cases to make sure we don’t test cases we don’t want to, and I need to double-check you have them covered. Will do that tomorrow. The testing is all "positive" in the sense that it ensures a receiver subtype check is in place when it "must be". In fact it must always be the case the receiver has a type that has the method being invoked. We were just missing a few cases that verified that (and some stronger conditions: ie receiver <: caller for invokespecial semantics)._ _If you want to test that we don't insert the new explicit_ _checks in cases where they are not needed, then I don't_ _know how to do that - other than by adding tracing and_ _running the test case and not seeing checkReceiver being_ _called._ _That said, once I've reworked the logic it will be_ _blindingly obvious when the new explicit check is being added._ _Thanks,_ _David_ _thanks,_ _Karen_ _On May 3, 2018, at 6:21 AM, David Holmes_ _<david.holmes at oracle.com> wrote: bug id: https://bugs.openjdk.java.net/browse/JDK-8197915 webrev: http://cr.openjdk.java.net/~dholmes/8197915/webrev/ JDK-8174962 implemented receiver typechecks for invokeinterface within the interpreter (templateTable), compilers and for MethodHandles. In nestmates invokeinterface can now be used for private interface methods - which result in direct calls. So we need to extend the receiver subtype checks to cover the new cases. Summary of changes: - src/hotspot/cpu//templateTable.cpp In the templateTable the 8174962 checks come after the private interface method invocation logic ("vfinal") we already had in place for the nestmate changes, and they rely on itable information that doesn't exist for private methods. So we insert a direct subtype check. I've provided code for all CPU's but only x86 and sparc have been tested. I'll be soliciting aid on the other ports before nestmates goes to mainline later this month. -  src/hotspot/share/oops/cpCache.cpp We have to pass the interface klass* so it's available for the typecheck. -  src/hotspot/share/oops/klassVtable.cpp Updated a comment that's no longer accurate. - src/hotspot/share/opto/doCall.cpp This code was provided by Vladimir Ivanov (thank you!) and expands the existing "invokespecial" support for receiver typechecks in C2, to "invokeinterface" as well. Aside: no changes were needed for C1. It's seems all the receiver typechecks for C1 are being handled at a higher level (through linkResolver and/or cpCache logic). - src/hotspot/share/prims/methodHandles.cpp Comment clarifying JVMREFinvokeSpecial doesn't necessarily mean it relates to an actual "invokespecial" - it is used for all direct calls. - src/java.base/share/classes/java/lang/invoke/DirectMethodHandle.java

Add clarifying comments regarding how "kind" can vary if a direct call is involved. Expand the condition to switch from LFINVSPECIAL to LFINVSPECIALIFC (which adds the additional receiver typecheck) to account for the invokeinterface case. -  test/jdk/java/lang/invoke/PrivateInterfaceCall.java New test for invokeinterface semantics that mirrors the existing SpecialInterfaceCall test for invokespecial. This is the last of the significant functional changes for nestmates. Thanks, David

More information about the valhalla-dev mailing list