Need some help: 8133646: x86/vm/macroAssembler_x86.cpp:886 DEBUG MESSAGE: StubRoutines::call_stub: threads must correspond

[Vladimir Kozlov]

First, there copy-paste problem in verification code, jcc after bind is 
not needed:

 >     375 __ jcc(Assembler::equal, L);
 >     376 __ bind(S);
 >     377 __ jcc(Assembler::equal, L);

Yes, the code assumes that get_thread(rbx); returns correct value and 
will not modify r15. r15 is save-on-entry so the C code should save it.

Is it possible that Sun's C++ compiler has a bug? So that _thr_current 
has wrong value? Or does not save r15?

To make sure that it is the cause of problem I would suggest to modify 
verification code as next:

       Label L1, L2;
       __ cmpptr(r15_thread, thread);
       __ jcc(Assembler::equal, L1);
       __ stop("StubRoutines::call_stub: r15_thread is corrupted");
       __ bind(L1);
       __ get_thread(rbx);
       __ cmpptr(r15_thread, rbx);
       __ jcc(Assembler::equal, L2);
       __ stop("StubRoutines::call_stub: threads must correspond");
       __ bind(L2);


Thanks,
Vladimir

On 8/17/15 8:05 PM, David Holmes wrote:
> Hi Compiler and x86_64 code experts, :)
>
> Bug: https://bugs.openjdk.java.net/browse/JDK-8133646
>
> There's a chance this failure is related to my recent changes under
> 8130212:
>
> http://hg.openjdk.java.net/jdk9/hs-rt/hotspot/rev/d5b328043c10
>
> Given the failing debug code:
>
>     367 #ifdef ASSERT
>     368 // verify that threads correspond
>     369 {
>     370 Label L, S;
>     371 __ cmpptr(r15_thread, thread);
>     372 __ jcc(Assembler::notEqual, S);
>     373 __ get_thread(rbx);
>     374 __ cmpptr(r15_thread, rbx);
>     375 __ jcc(Assembler::equal, L);
>     376 __ bind(S);
>     377 __ jcc(Assembler::equal, L);
>     378 __ stop("StubRoutines::call_stub: threads must correspond");
>     379 __ bind(L);
>     380 }
>     381 #endif
>
> and given the definition of get_thread:
>
> // This is simply a call to ThreadLocalStorage::thread()
> void MacroAssembler::get_thread(Register thread) {
>    if (thread != rax) {
>      push(rax);
>    }
>    push(rdi);
>    push(rsi);
>    push(rdx);
>    push(rcx);
>    push(r8);
>    push(r9);
>    push(r10);
>    push(r11);
>
>    call(RuntimeAddress(CAST_FROM_FN_PTR(address,
> ThreadLocalStorage::thread)));
>
>    pop(r11);
>    pop(r10);
>    pop(r9);
>    pop(r8);
>    pop(rcx);
>    pop(rdx);
>    pop(rsi);
>    pop(rdi);
>
>    if (thread != rax) {
>      movl(thread, rax);
>      pop(rax);
>    }
> }
>
> where ThreadLocalStorage::thread() is simply:
>
> inline Thread* ThreadLocalStorage::thread()  {
>    return _thr_current;
> }
>
> and _thr_current is a compiler-defined thread-local variable:
>
> static __thread Thread * _thr_current;
>
> Can anyone see anything that might cause an issue? I was wondering
> whether I have to verify that r15 is preserved by the actual
> thread-local access? IIUC normally the get_thread slow-path would be
> used to set r15_thread and so we wouldn't notice if it got trashed under
> some circumstances. It obviously doesn't always get trashed because the
> bug has only been seen to occur once. And of course the failure could be
> completely incidental to my changes. :)
>
> Thanks,
> David