Hi Tom,

One question regarding deoptimization and C1: the StoreField and
LoadField code relies on deoptimization in case a load/store cannot be
determined to be volatile at compile time.

In the c1 runtime patch_code method, a VM_DeoptimizeFrame vm operation
is executed. What is puzzling me, is that when I follow the code of this
vm operation I don't find where the nmethod is made not entrant so that
the method will eventually be recompiled.

Indeed when I run this simple test:

public class TestDeopt {
   static class A {
	volatile int i;
   }

   static void m() {
	A a = new A();
	a.i = 0;
   }

   static public void main(String[] args) {
	m();
	m();
	m();
   }
}

With:
-client -Xcomp -XX:+PrintCompilation -XX:-Inline -XX:+TracePatching -XX:+TraceDeoptimization

I see the m method being compiled once but I see 3 "Deoptimizing for
patching volatile field reference" which proves that the compiled code
is not dropped.

Am I missing something?

Thanks,
Roland.
Posted Date : 2008-11-05 00:10:36.0

N/A

The fix for 6414800 is missing a call to make_not_entrant
Posted Date : 2008-11-05 00:11:37.0

http://hg.openjdk.java.net/jdk7/hotspot/hotspot/rev/ac8fe14c93e4
Posted Date : 2008-12-13 09:19:19.0