This is a copy of 4317961 which was for SPARC (which is now fixed.)

Problem occurs when thr_getstate returns a flag value of TRS_NONVOLATILE
in fetch_top_frame_fast. The i386 code for fetch_top_frame_fast returns
a null pc/sp, restarts the thread, and tries to suppend again. Unfortunately,
the thread with the infinite loop is always in a state in which TRS_NONVOLATILE
is returned, so GC never runs.

Compiler safepoints do not work.

Run the appended program with following commandline
and it will hang with both compilers
java -Xcomp -verbose:gc sf

Run it with following commandline and it terminates properly:
java -Xint -verbose:gc sf

The program hangs if the method loop is compiled because
the runtime cannot properly suspend the endless loop.
I have verified that both Compiler 1 and Compiler 2 (server) are
emitting safepoint information.



----------------
public class sf {
        static public boolean _xx = true;

public static void main(String[] args) {
new Srdjan().start();
}

public void start() {
new TesterThread().start();
new GCThread().start();
}

class TesterThread extends Thread {

                // trick compiler 2 in believing that this is not and endless loop
                public void loop (int n) {
long counter = 0;
while (_xx) {
if (n == 0) {
                                  _xx = false;
                                }
}
                }
public void run () {
System.out.println("Start loop");
                        loop(3);
}
}

class GCThread extends Thread {
public void run () {
System.out.println("Starting GC loop");
long counter = 0;
while (true) {
try { sleep(250); } catch (InterruptedException ix) {}
System.gc();
                                counter++;
                                if (counter == 10) {
                                  _xx = false;
                                  return;
                                }
}
}
}
}