RFR: 8280481: Duplicated stubs to interpreter for static calls

[Evgeny Astigeevich]

On Wed, 12 Jun 2024 07:17:25 GMT, Doug Simon <dnsimon at openjdk.org> wrote:

>>> Hi @eastig ,
>>> I would like to recurring your experimental data and I would be very grateful if you could provide a small patch to help me get the result of `Saved bytes` and `Nmethods with shared stubs`.
>>> Thank you!
>> 
>> 
>> diff --git a/src/hotspot/share/asm/codeBuffer.inline.hpp b/src/hotspot/share/asm/codeBuffer.inline.hpp
>> index 045cff13f25..9af26730cbd 100644
>> --- a/src/hotspot/share/asm/codeBuffer.inline.hpp
>> +++ b/src/hotspot/share/asm/codeBuffer.inline.hpp
>> @@ -45,6 +45,7 @@ bool emit_shared_stubs_to_interp(CodeBuffer* cb, SharedStubToInterpRequests* sha
>>    };
>>    shared_stub_to_interp_requests->sort(by_shared_method);
>>    MacroAssembler masm(cb);
>> +  bool has_shared = false;
>>    for (int i = 0; i < shared_stub_to_interp_requests->length();) {
>>      address stub = masm.start_a_stub(CompiledStaticCall::to_interp_stub_size());
>>      if (stub == NULL) {
>> @@ -53,13 +54,22 @@ bool emit_shared_stubs_to_interp(CodeBuffer* cb, SharedStubToInterpRequests* sha
>>      }
>> 
>>      ciMethod* method = shared_stub_to_interp_requests->at(i).shared_method();
>> +    int shared = 0;
>>      do {
>>        address caller_pc = cb->insts_begin() + shared_stub_to_interp_requests->at(i).call_offset();
>>        masm.relocate(static_stub_Relocation::spec(caller_pc), relocate_format);
>>        ++i;
>> +      ++shared;
>>      } while (i < shared_stub_to_interp_requests->length() && shared_stub_to_interp_requests->at(i).shared_method() == method);
>>      masm.emit_static_call_stub();
>>      masm.end_a_stub();
>> +    if (UseNewCode && shared > 1) {
>> +      has_shared = true;
>> +      tty->print_cr("Saved: %d", (shared - 1) * CompiledStaticCall::to_interp_stub_size());
>> +    }
>> +  }
>> +  if (has_shared) {
>> +    tty->print_cr("nm_has_shared");
>>    }
>>    return true;
>>  }
>> 
>> 
>> You will need to use `-XX:+UseNewCode` in your runs.
>> `grep nm_has_shared  run.log | wc -l` is a number of nmethods having a shared stub.
>> `grep Saved: run.log | awk '{print $2}' | grep -o '[0-9]*' | paste -s -d+ - | bc` prints a number of saved bytes.
>
> @eastig as I understand, this optimization is about saving code cache memory. The sharing is within an nmethod, not across nmethods, correct?
> I'm trying to prioritize an effort to adopt this optimization in Graal. In addition to the numbers you present for code cache bytes saved in the benchmarks, can you say anything about how much that is relative to the code cache used in the benchmarks?
> 
> More context: https://github.com/openjdk/jdk/pull/19672

Hi @dougxc 
> @eastig as I understand, this optimization is about saving code cache memory. The sharing is within an nmethod, not across nmethods, correct?

Yes, you are correct. Sharing across nmethods would need non-simple maintanance, maybe similar to the inline cache. I am not sure it is worth.

> In addition to the numbers you present for code cache bytes saved in the benchmarks, can you say anything about how much that is relative to the code cache used in the benchmarks?

In a service application with more than 50 000 nmethods I saw between 1% - 2% reduction in CodeCache usage.

-------------

PR Comment: https://git.openjdk.org/jdk/pull/8816#issuecomment-2200450633