[JFR] Fix object profiling stacktrace problem on aarch64 platform

[D-D-H]

Hi,
Please help review this patch that fixes the stack trace problem of object profiling on the aarch64 platform.

The root cause of this problem is that the stack layout of aarch64 is different from x86_64 so that we cannot use vframeStream(thread, os::current_frame()) to get the java call stack traces.

This fix uses CallStaticJavaNode to invoke the jfr_fast_object_alloc_C so that the _last_Java_pc is set properly by opto stub.

Issue: #189
Denghui

[CLAassistant]

All committers have signed the CLA.

[zhengxiaolinX]

lgtm.

[D-D-H]

@yfxhust
Hi Fangxi,
Please help review this fix when you have time.
This patch used CallStaticJavaNode to replace CallLeafNode, so the _cached_top_frame_bci and stack walk mode could be removed. Plz help confirm this logic is correct.

[yfxhust]

@yfxhust Hi Fangxi, Please help review this fix when you have time. This patch used CallStaticJavaNode to replace CallLeafNode, so the _cached_top_frame_bci and stack walk mode could be removed. Plz help confirm this logic is correct.

  // Jfr callstack collection relies on vframeStream.
  // But the bci of top frame can not be determined by vframeStream in some scenarios.
  // For example, in the opto CallLeafNode runtime call of
  // OptoRuntime::jfr_fast_object_alloc_C, the top frame bci
  // returned by vframeStream is always invalid. This is largely due to the oopmap that
  // is not correctly granted ( refer to PhaseMacroExpand::expand_allocate_common to get more details ).
  // The opto fast path object allocation tracing occurs in the opto CallLeafNode,
  // which has been broken by invalid top frame bci.
  // To fix this, we get the top frame bci in opto compilation phase
  // and pass it as parameter to runtime call. Our implementation will replace the invalid top
  // frame bci with cached_top_frame_bci.

This is why _cached_top_frame_bci. I am not sure whether it still stands in current static Java call implementation.

[yfxhust]

LGTM

[D-D-H]

@yfxhust Hi Fangxi, Please help review this fix when you have time. This patch used CallStaticJavaNode to replace CallLeafNode, so the _cached_top_frame_bci and stack walk mode could be removed. Plz help confirm this logic is correct.

  // Jfr callstack collection relies on vframeStream.
  // But the bci of top frame can not be determined by vframeStream in some scenarios.
  // For example, in the opto CallLeafNode runtime call of
  // OptoRuntime::jfr_fast_object_alloc_C, the top frame bci
  // returned by vframeStream is always invalid. This is largely due to the oopmap that
  // is not correctly granted ( refer to PhaseMacroExpand::expand_allocate_common to get more details ).
  // The opto fast path object allocation tracing occurs in the opto CallLeafNode,
  // which has been broken by invalid top frame bci.
  // To fix this, we get the top frame bci in opto compilation phase
  // and pass it as parameter to runtime call. Our implementation will replace the invalid top
  // frame bci with cached_top_frame_bci.

This is why _cached_top_frame_bci. I am not sure whether it still stands in current static Java call implementation.

JVMState includes the BCI info, and OptoStub can set _last_java_sp properly, hence, vframeStream could get the correct context information.

[D-D-H]

LGTM

Thank you!

[kuaiwei]

LGTM

[D-D-H]

Thanks again for the review!