WORK AROUND

-XX:-DoEscapeAnalysis
                                     

PUBLIC COMMENTS

Problem:
EA misses checks for MemBar nodes when looking for following
MergeMem nodes. As result it did not add memory slice for
the volatile store (followed by several MemBar nodes)
in MergeMem of uncommon trap.
So when the method deoptimized and eliminated object reallocated
the NULL is stored to the corresponding field.

Solution:
Collect all MergeMem nodes at the beginning of EA when it walks
through Ideal graph instead of searching MergeMem nodes by going
down trough memory during memory splitting for non escaping objects.
Also check for general MemBar nodes instead of only Initialize nodes
during memory splitting.

I also did next optimizations/fixes:

1. Check for SafePointScalarObject nodes to avoid printing empty lines
in PrintOptoAssembly output since they don't have corresponding mach nodes.

2. Eliminate volatile MemBars nodes for stores into a scalar replaced objects.
I noticed that we still generate membars even if the object is eliminated.
For that I added the Precedent edge to corresponding Store node.

3. The check for ClearArray node is missing when searching for
a better memory edge during EA, memory optimization and an object
scalar replacement. We can't bypass it since it is the part of object
initialization (zeroing).

4. Add missing checks for string intrinsic nodes to adjust the escape state
(non scalar replaceable) of char[] arrays they use.

5. Move code for searching objects which are not scalar replaceable to separate
method verify_escape_state() since the code became too large. Add a simple
control flow check to find initializing store for oop fields.

6. Added asserts to check cases which don't have to be handled.

7. I also changed instance type setting for sequential CheckCastPP nodes
to change following CheckCastPP node type to allocation type
if it is subtype.

8. And I found that G1 barriers CallLeaf node case is missing
in process_call_arguments() (as result referenced allocations
were marked as global escape instead of thread local).
                                     

EVALUATION

http://hg.openjdk.java.net/jdk7/hotspot-comp/hotspot/rev/f96a1a986f7b
                                     

SUGGESTED FIX

-XX:-DoEscapeAnalysis
                                     

EVALUATION

It is different from 6894807. It is caused by EA lost the array allocation for CopyOnWriteArrayList (the "array" field should not be NULL):

027     CALL,static  wrapper for: uncommon_trap(reason='unloaded' action='reinterpret' index='8')
        # CopyOnWriteArrayListTest::testAddAll1_IndexOutOfBoundsException @ bci:10  L[0]=_ L[1]=_ STK[0]=#ScObj0 STK[1]=#-1
        # ScObj0 java/util/concurrent/CopyOnWriteArrayList={ [lock :0]=#ScObj1, [array :1]=#NULL }
        # ScObj1 java/util/concurrent/locks/ReentrantLock={ [sync :0]=#ScObj2 }
        # ScObj2 java/util/concurrent/locks/ReentrantLock$NonfairSync={ [exclusiveOwnerThread :0]=#NULL, [state :1]=#0, [head :2]=#NULL, [tail :3]=#NULL }
        # OopMap{off=44}