RFR(XS): 8049441: PPC64: Don't use StubCodeMarks for zero-length stubs
serguei.spitsyn at oracle.com
serguei.spitsyn at oracle.com
Tue Jul 15 09:09:39 UTC 2014
On 7/15/14 1:48 AM, Volker Simonis wrote:
> On Mon, Jul 14, 2014 at 10:35 PM, serguei.spitsyn at oracle.com
> <serguei.spitsyn at oracle.com> wrote:
>> Hi Volker,
>> It looks good in general.
>> But I don't understand all the details.
>> For instance, your email description of the fix tells that the the event is
>> posted by:
>> RuntimeStub::new_runtime_stub() -> CodeBlob::trace_new_stub() ->
>> I see the new_runtime_stub() call in the generate_throw_exception() but
>> there is no such call
>> in the generate_icache_flush() and generate_handler_for_unsafe_access() .
>> Probably, the StubCodeMark just needs to be removed there.
>> Could you, please, explain this a little bit?
> Hi Serguei,
> Thank you for looking at my change. I tried to explain your questions
> in my initial mail but maybe I wasn't clear enough:
> - in generate_icache_flush() and generate_verify_oop() we DO NOT
> GENERATE any stub code. We don't use dynamically generated stubs on
> ppc64 for flushing the icache or verifying oops but call C-functions
> instead. So there's no need to generate post_dynamic_code_generated()
> events for them and also no need for a StubCodeMark.
> - for generate_throw_exception() we dynamically generate a runtime
> stub instead of an simple stub and for runtime stubs the JVMT dynamic
> code event is already generated by RuntimeStub::new_runtime_stub() ->
> CodeBlob::trace_new_stub() ->
> JvmtiExport::post_dynamic_code_generated(). This is exactly the way
> how it works on other CPU architectures. The usage of a StubCodeMark
> in generate_throw_exception() was simply a "day one" bug in the ppc64
Thank you for the extra details!
I asked for that as my knowledge in this area is limited.
The fix looks good to me.
I can be a sponsor for integration if needed.
But a Review is still required.
> - I haven't changed generate_handler_for_unsafe_access() so I don't
> actually understand your concerns.
I accidentally copied a wrong name. Sorry.
I had to copy: generate_verify_oop().
> generate_handler_for_unsafe_access() correctly contains a StubCodeMark
> because it dynamically generates stub code - even if it is just the
> output of a "not yet implemented" message.
>> We also need someone from the compiler team to look at this.
>> I also included into the cc-list Oleg, who recently touched this area.
>> On 7/14/14 11:24 AM, Volker Simonis wrote:
>> Hi everybody,
>> can somebody PLEASE review and sponsor this tiny, ppc64-only change.
>> On Tue, Jul 8, 2014 at 5:45 PM, Daniel D. Daugherty
>> <daniel.daugherty at oracle.com> wrote:
>> Adding the Serviceability Team since JVM/TI belongs to them.
>> On 7/8/14 9:41 AM, Volker Simonis wrote:
>> could somebody please review and push the following small, PPC64-only
>> change to any of the hs team repositories:
>> For some stubs we actually do not really generate code on PPC64 but
>> instead we use a native C-function with inline-assembly. If the
>> generators of these stubs contain a StubCodeMark, they will trigger
>> JvmtiExport::post_dynamic_code_generated_internal events with a zero
>> length code size. These events may fool clients like Oprofile which
>> register for these events (thanks to Maynard Johnson who reported this
>> - see
>> This change simply removes the StubCodeMark from
>> ICacheStubGenerator::generate_icache_flush() and generate_verify_oop()
>> because they don't generate assembly code. It also removes the
>> StubCodeMark from generate_throw_exception() because it doesn't really
>> generate a plain stub but a runtime stub for which the JVMT dynamic
>> code event is already generated by RuntimeStub::new_runtime_stub() ->
>> CodeBlob::trace_new_stub() ->
>> Thank you and best regards,
More information about the hotspot-dev