Revert^2 "Remove Global deopt requirement for several jvmti events"

This reverts commit ad344b6a14feba90a06a205760e9bc766c56cab0.

Jit-on-first-use would cause the runtime to skip sending FramePop
events for exceptions in some circumstances due to longjmp-ing over
the frames. This requires forcing jit off from the first instant a
thread can get frame popped events, which is unfortunate.

Reason for revert: Fixed issue causing 1925 to fail on
                   jit-on-first-use and redefine-stress
Test: ./test.py --host
Test: ./art/test/testrunner/run_build_test_target.py -j50 art-jit-on-first-use

Change-Id: I6ec2a799db1041262a055be10e1af1faece6f2fc

1 files changed, 8 insertions, 2 deletions

diff --git a/openjdkjvmti/ti_stack.cc b/openjdkjvmti/ti_stack.cc
index 385ac45488..5e934df42c 100644
--- a/openjdkjvmti/ti_stack.cc
+++ b/openjdkjvmti/ti_stack.cc
@@ -1061,8 +1061,14 @@ jvmtiError StackUtil::NotifyFramePop(jvmtiEnv* env, jthread thread, jint depth)
     art::ShadowFrame* shadow_frame = visitor.GetOrCreateShadowFrame(&needs_instrument);
     {
       art::WriterMutexLock lk(self, tienv->event_info_mutex_);
-      // Mark shadow frame as needs_notify_pop_
-      shadow_frame->SetNotifyPop(true);
+      if (LIKELY(!shadow_frame->NeedsNotifyPop())) {
+        // Ensure we won't miss exceptions being thrown if we get jit-compiled. We only do this for
+        // the first NotifyPopFrame.
+        target->IncrementForceInterpreterCount();
+
+        // Mark shadow frame as needs_notify_pop_
+        shadow_frame->SetNotifyPop(true);
+      }
       tienv->notify_frames.insert(shadow_frame);
     }
     // Make sure can we will go to the interpreter and use the shadow frames.