review for 7055355: JSR 292: crash while throwing WrongMethodTypeException

[Tom Rodriguez]

On Jun 16, 2011, at 9:24 AM, Christian Thalinger wrote:

> On Jun 16, 2011, at 4:44 PM, Christian Thalinger wrote:
>> On Jun 16, 2011, at 2:49 PM, Christian Thalinger wrote:
>>> On Jun 16, 2011, at 11:47 AM, Christian Thalinger wrote:
>>>> On Jun 16, 2011, at 7:32 AM, Tom Rodriguez wrote:
>>>>> http://cr.openjdk.java.net/~never/7055355
>>>>> 174 lines changed: 54 ins; 111 del; 9 mod; 30934 unchg
>>>>> 
>>>>> 7055355: JSR 292: crash while throwing WrongMethodTypeException
>>>>> Reviewed-by:
>>>>> 
>>>>> When throwing a WrongMethodTypeException from a MethodHandle the
>>>>> existing code was pretending it was in the interpreter which worked ok
>>>>> most of the time but would sometimes die when trying to throw with a
>>>>> compiled caller.  It could cause asserts in some contexts or result in
>>>>> GC crashes.  The fix is to remove the old machinery in favor of the
>>>>> more standard code for throwing exceptions in
>>>>> StubRoutines/SharedRuntime.  There was one other use of the old throw
>>>>> machinery but it's in part of check that can never fail since we won't
>>>>> enable method handles unless we have a non-NULL rasie exception
>>>>> method.  There's an extra fix in the there to reject lookups of
>>>>> clinit.  It isn't critical but it allows all the JCK tests to run
>>>>> without asserting.  Tested with failing JCK test and jruby crasher on
>>>>> sparc/x86 with client/server/d64, along with a full run of the
>>>>> invokedynamic JCK tests, the JLI regression tests and the vm.mlvm
>>>>> tests.
>>>> 
>>>> You didn't remove the NULL-checking code of the raise_exception method in methodHandles_sparc.cpp.  Otherwise this looks good.
>>> 
>>> 
>>> The SPARC logic is broken, I get asserts with John's ThrowExceptionsTest.  The problem is that SPARC's throw exception stub uses a save_frame and O1/O2 then contain some arbitrary values hitting the assert in SharedRuntime::throw_WrongMethodTypeException.  I fixed it using the same logic you added for x86.  I also removed the NULL-checking code I mentioned above.
>>> 
>>> With the patch below applied we hit the unreasonable-stack-move assert at some point:
>>> 
>>> #  assert(false) failed: DEBUG MESSAGE: damaged ricochet frame: (L4 - UNREASONABLE_STACK_MOVE) > FP
>> 
>> I know I ported this piece of code over from x86 but honestly I don't understand what it is checking for.  L4 >= FP seems reasonable and we should keep that check but (L4 - UNREASONABLE_STACK_MOVE) > FP is not a constraint that must hold in all situations.  I suggest removing that check.  What do you think?

I increased the slop value in my last set of changes and we're currently not failing this test that I can see.  I agree that the test is slightly mystifying and wouldn't mind removing it.

> 
> ...and there seems to be a bug in:
> 
> void MethodHandles::verify_stack_move(MacroAssembler* _masm,
>                                      RegisterOrConstant arg_slots, int direction) {
>  enum { UNREASONABLE_STACK_MOVE = 256 * 4 };  // limit of 255 arguments
> 
> We redefine UNREASONABLE_STACK_MOVE with a wrong value, it seems.

That's odd.  I don't think I've even see this particular test fail so it's probably benign for the moment.

tom

> 
> -- Christian