[prev in list] [next in list] [prev in thread] [next in thread] 

List:       openjdk-serviceability-dev
Subject:    Re: JIT compiler turned off after stepping over in jdb
From:       Chris Plummer <chris.plummer () oracle ! com>
Date:       2022-10-31 17:57:55
Message-ID: 2340a2a9-a181-bb71-0904-5e25e55dc941 () oracle ! com
[Download RAW message or body]

Hi Simeon,

We have the following bug filed for the debug agent:

     JDK-8229012: When single stepping, the debug agent can cause the 
thread to remain in interpreter mode after single stepping completes

It will result in the behavior you are seeing. However, I'm not sure how 
it relates to the changes in the diff you referenced, which is the fix for:

     JDK-8187289: NotifyFramePop request is not cleared if 
JVMTI_EVENT_FRAME_POP is disabled

If you revert all of this diff (not just comment out the one line) do 
you still see the performance issue? It seems you should due to 
JDK-8229012. I think by disabling the line of code you are actually 
introducing a bug that results in not entering interp_only mode even 
though it should (due to the the NotifyFramePops that the debug agent 
has requested).

best regards,

Chris

On 10/31/22 9:01 AM, S A wrote:
> Hi all,
> 
> TL;DR We observe a significant performance degradation after stepping 
> over in a debugger. By the looks of it, the interpreter only mode is 
> entered and never left due to this change:
> 
> https://github.com/openjdk/jdk11u/commit/dbb9eb5b9c3ddcbcfaae8ac4ca21760c9c99a15c#diff-fa220f091e99b08a3a20869c780e3529b2fee843acbc3b5748443794c4582f5d
>  
> I hope this is the correct mailing list for this. I would open an 
> OpenJDK bug, but I don't have an account.
> 
> To reproduce, debug the following snippet with jdb:
> 
> public class StepOver {
> public static void main(String[] args) throws Exception {
> System.out.println(); // set a breakpoint here, and step over
> System.out.println(); // press Resume when suspended here
> double[] ySeries = new double[100_000];
> for (int i = 0; i < ySeries.length; i++) {
> ySeries[i] = Math.sin(i / 20) * Math.pow(10, -21);
> }
> long startTime = System.currentTimeMillis();
> for (int i = 0; i < ySeries.length; i++) {
> String.valueOf(ySeries[i]);
> }
> System.out.println("elapsed time: " + 
> (System.currentTimeMillis() - startTime));
> }
> }
> 
> Breakpoint at line 3, run the snippet, step over, continue execution. 
> Observe execution time is very slow (8.5 sec in my case, vs. normally 
> 0.1 sec). Omitting the step over, execution time is similar to normal 
> (non-debug) execution time.
> 
> We observe this bug with OpenJDK 11.0.16 and OpenJDK 17.0.4.
> 
> I've opened a RHEL bugzilla ticket, if that helps: 
> https://bugzilla.redhat.com/show_bug.cgi?id=2138897
> 
> I don't observe the problem with this change:
> 
> diff --git a/src/hotspot/share/prims/jvmtiEventController.cpp 
> b/src/hotspot/share/prims/jvmtiEventController.cpp
> index 38de6383cc..22d0f40c97 100644
> --- a/src/hotspot/share/prims/jvmtiEventController.cpp
> +++ b/src/hotspot/share/prims/jvmtiEventController.cpp
> @@ -523,7 +523,7 @@ 
> JvmtiEventControllerPrivate::recompute_thread_enabled(JvmtiThreadState 
> *state) {
> JvmtiEnvThreadStateIterator it(state);
> for (JvmtiEnvThreadState* ets = it.first(); ets != NULL; ets = 
> it.next(ets)) {
> any_env_enabled |= recompute_env_thread_enabled(ets, state);
> -      has_frame_pops |= ets->has_frame_pops();
> +      //has_frame_pops |= ets->has_frame_pops();
> }
> }
> 
> If I add prints to entering the interpreter only code below, I see the 
> mode is entered after step over, but left only after the program 
> finishes execution:
> 
> diff --git a/src/hotspot/share/prims/jvmtiEventController.cpp 
> b/src/hotspot/share/prims/jvmtiEventController.cpp
> index 38de6383cc..a9764303c9 100644
> --- a/src/hotspot/share/prims/jvmtiEventController.cpp
> +++ b/src/hotspot/share/prims/jvmtiEventController.cpp
> @@ -542,8 +542,12 @@ 
> JvmtiEventControllerPrivate::recompute_thread_enabled(JvmtiThreadState 
> *state) {
> 
> if (should_be_interp != is_now_interp) {
> if (should_be_interp) {
> +      printf("entering interp only mode\n");
> +      fflush(stdout);
> enter_interp_only_mode(state);
> } else {
> +      printf("leaving interp only mode\n");
> +      fflush(stdout);
> leave_interp_only_mode(state);
> }
> }
> 
> Sorry if I'm interpreting the code wrong and "interpreter only mode" 
> means something else other than, there is no JIT compiler. But judging 
> from the really slow execution, I assume that is the problem.
> 
> Any chance this can be looked into? Any ideas for a workaround that 
> can be implemented in a debugger (in our case, Eclipse, same problem 
> as jdb)? Is there already an open OpenJDK ticket?
> 
> Best regards and thanks,
> Simeon


[prev in list] [next in list] [prev in thread] [next in thread]