I am seeing the CTW failures with JDK-8360557, which manifest with the following assert:

# Internal Error (/home/shade/trunks/jdk/src/hotspot/share/oops/compressedKlass.inline.hpp:80), pid=3869033, tid=3869052
# assert(is_encodable(addr)) failed: Address 0x000070470c568fa0 is not encodable (Klass range: [0x000000002c001000 - 0x000000006d000000), (1090514944 bytes), klass alignment: 8)

Current CompileTask:
C2:2968 2272 !b 4 org.apache.cassandra.auth.Auth::setupDefaultSuperuser (74 bytes)

Stack: [0x000070473c1f9000,0x000070473c2f9000], sp=0x000070473c2f47c0, free space=1005k
Native frames: (J=compiled Java code, j=interpreted, Vv=VM code, C=native code)
V [libjvm.so+0x15e1e1a] CompressedKlassPointers::encode_not_null(Klass*)+0xba (compressedKlass.inline.hpp:80)
V [libjvm.so+0x15ba5f7] MacroAssembler::cmp_narrow_klass(Register, Klass*)+0xb7 (compressedKlass.inline.hpp:73)
V [libjvm.so+0x17ba912] PhaseOutput::scratch_emit_size(Node const*)+0x4a2 (output.cpp:3388)
V [libjvm.so+0x17b2852] PhaseOutput::shorten_branches(unsigned int*)+0x382 (output.cpp:540)
V [libjvm.so+0x17c5458] PhaseOutput::Output()+0xad8 (output.cpp:340)
V [libjvm.so+0xb68c90] Compile::Code_Gen()+0xa20 (compile.cpp:3109)
V [libjvm.so+0xb6df93] Compile::Compile(ciEnv*, ciMethod*, int, Options, DirectiveSet*)+0x2023 (compile.cpp:891)
V [libjvm.so+0x99ee90] C2Compiler::compile_method(ciEnv*, ciMethod*, int, bool, DirectiveSet*)+0x450 (c2compiler.cpp:141)
V [libjvm.so+0xb7d432] CompileBroker::invoke_compiler_on_method(CompileTask*)+0xb22 (compileBroker.cpp:2323)
V [libjvm.so+0xb7e458] CompileBroker::compiler_thread_loop()+0x5a8 (compileBroker.cpp:1967)
V [libjvm.so+0x10d3d4f] JavaThread::thread_main_inner()+0x12f (javaThread.cpp:773)
V [libjvm.so+0x1cdde3a] Thread::call_run()+0xba (thread.cpp:243)
V [libjvm.so+0x17983a8] thread_native_entry(Thread*)+0x128 (os_linux.cpp:868)
C [libc.so.6+0x9caa4]

It is fairly reproducible on one of the internal JARs, and only with specific StressSeed-s. So there must be some high-level optimization in play.

Adding the assert like this:
 
+TypeNarrowKlass::TypeNarrowKlass(const TypePtr* ptrtype): TypeNarrowPtr(NarrowKlass, ptrtype) {
+ if (ptrtype->ptr() == TypePtr::Constant) {
+ CompressedKlassPointers::check_encodable((void*)ptrtype->get_con());
+ }
+}

Starts to fail here:

Stack: [0x000071208d300000,0x000071208d400000], sp=0x000071208d3fac50, free space=1003k
Native frames: (J=compiled Java code, j=interpreted, Vv=VM code, C=native code)
V [libjvm.so+0x1d0eb32] TypeNarrowKlass::TypeNarrowKlass(TypePtr const*) [clone .part.0]+0xa2 (compressedKlass.inline.hpp:80)
V [libjvm.so+0x1d193ae] TypeNarrowKlass::make(TypePtr const*)+0xae (type.cpp:5322)
V [libjvm.so+0xb594dd] Compile::final_graph_reshaping_main_switch(Node*, Final_Reshape_Counts&, unsigned int, Unique_Node_List&)+0x1e9d (compile.cpp:3589)
V [libjvm.so+0xb59adf] Compile::final_graph_reshaping_impl(Node*, Final_Reshape_Counts&, Unique_Node_List&) [clone .part.0]+0x15f (compile.cpp:3204)
V [libjvm.so+0xb5a47b] Compile::final_graph_reshaping_walk(Node_Stack&, Node*, Final_Reshape_Counts&, Unique_Node_List&)+0x1fb (compile.cpp:3153)
V [libjvm.so+0xb6426d] Compile::final_graph_reshaping()+0x66d (compile.cpp:4019)
V [libjvm.so+0xb6b0d1] Compile::Optimize()+0x1051 (compile.cpp:2568)
V [libjvm.so+0xb6de93] Compile::Compile(ciEnv*, ciMethod*, int, Options, DirectiveSet*)+0x1f23 (compile.cpp:858)
V [libjvm.so+0x99ee90] C2Compiler::compile_method(ciEnv*, ciMethod*, int, bool, DirectiveSet*)+0x450 (c2compiler.cpp:141)
V [libjvm.so+0xb7d432] CompileBroker::invoke_compiler_on_method(CompileTask*)+0xb22 (compileBroker.cpp:2323)
V [libjvm.so+0xb7e458] CompileBroker::compiler_thread_loop()+0x5a8 (compileBroker.cpp:1967)
V [libjvm.so+0x10d3d4f] JavaThread::thread_main_inner()+0x12f (javaThread.cpp:773)
V [libjvm.so+0x1cdde3a] Thread::call_run()+0xba (thread.cpp:243)
V [libjvm.so+0x17983a8] thread_native_entry(Thread*)+0x128 (os_linux.cpp:868)

In other words, here:
https://github.com/openjdk/jdk/blob/38f59f84c98dfd974eec0c05541b2138b149def7/src/hotspot/share/opto/compile.cpp#L3588-L3590

        } else if (t->isa_klassptr()) {
          new_in2 = ConNode::make(t->make_narrowklass());
        }

That code seems to assume that if we are holding the klass pointer, it would always be encodeable as narrow klass pointer. But this is not true at least since JDK-8338526.

This issue looks like a missing case similar to JDK-8343206. In fact, I think Tobias' analysis at the time was that we never reach this code with the klass that is not abstract/interface. But this CTW failure serves as a direct counter-example to that. In fact, I can throw in an assert like this:

        } else if (t->isa_klassptr()) {
          ciKlass* klass = t->is_klassptr()->exact_klass();
          assert(!klass->is_abstract() && !klass->is_interface(),
                 "Hey, is it really fine? abstract: %s; interface: %s",
                 BOOL_TO_STR(klass->is_abstract()), BOOL_TO_STR(klass->is_interface()));
          new_in2 = ConNode::make(t->make_narrowklass());
        }

...and see it failing:

# Internal Error (/home/shade/trunks/jdk/src/hotspot/share/opto/compile.cpp:3590), pid=3920890, tid=3920942
# assert(!klass->is_abstract() && !klass->is_interface()) failed: Hey, is it really fine? abstract: true; interface: false