- A Concurrent Affair - https://www.concurrentaffair.org -

Wrong Insertion Method

Posted By Mathias On In Concurrent Unit Testing | No Comments

I finally found the problem why the boolean $$$oldThread$$$ flag wasn’t sticking [1]: I seemed to set it to false, but the next time I looked at it, it was true again.

The Java code basically is this:

public static synchronized void compactWait(long code, long tid) {
if (!_replaying) {
return;
}

monitorEnter(_waitAlgoObj);

if (isOldThread()) {
// If this is NOT the first sync point for this thread...
...
}
else {
// Otherwise, if it was the first time... set flag.
setOldThread();
...
}

// More code...
...

monitorEnter(), isOldThread() and setOldThread() are dummy marker methods and get replaced with bytecode during instrumentation:

The Java code above roughly corresponds to this bytecode:

getstatic edu/rice/cs/cunit/SyncPointBuffer._replayingZ
ifne 4
return
getstatic edu/rice/cs/cunit/SyncPointBuffer._waitAlgoObjLjava/lang/Object;
invokestatic edu/rice/cs/cunit/SyncPointBuffer.monitorEnter(Ljava/lang/Object;)V
invokestatic edu/rice/cs/cunit/SyncPointBuffer.isOldThread()Z
ifeq 13

// If this is NOT the first sync point for this thread...
...
goto 17

// Otherwise, if it was the first time... set flag.
invokestatic edu/rice/cs/cunit/SyncPointBuffer.setOldThread()V
...

// More code...
...

My faulty instrumentation of the marker methods, however, generated this bytecode:

getstatic edu/rice/cs/cunit/SyncPointBuffer._replayingZ
ifne 4
return
getstatic edu/rice/cs/cunit/SyncPointBuffer._waitAlgoObjLjava/lang/Object;
monitorEnter
invokestatic java/lang/Thread.currentThread()Ljava/lang/Thread;
getfield java/lang/Thread.$$$oldThread$$$Z
ifeq 18

// If this is NOT the first sync point for this thread...
...
goto 20

// Otherwise, if it was the first time... set flag.
invokestatic java/lang/Thread.currentThread()Ljava/lang/Thread;
iconst_1;
putfield java/lang/Thread.$$$oldThread$$$Z
...

// More code...
...

I was inserting the correct bytecode, but I was modifying call targets it in a way such that calls to the insertion location would go to the opcode just PAST the inserted code, whereas it should have gone right TO the inserted code. Now I can move on and face new, exciting problems ;-)

[2] [3]Share [4]