New NullRestricted Attribute in Project Valhalla

[hangshao0]

In the last spec: https://cr.openjdk.org/~dlsmith/jep401/jep401-20230428/specs/flattened-heap-jvms.html, a NullRestricted attribute is introduced.

It indicates that the field cannot be null. putfield and putstatic will throw NPE if there is an attempt to assigned null to such fields. The null restricted field needs to be initialized to the default instance.

[hangshao0]

The null restricted field needs to be initialized to the default instance.

OpenJ9 is already doing this for Q types.

[hangshao0]

FYI @theresa-m

[theresa-m]

@hangshao0 can you point me to where this is happening in the code?

[hangshao0]

can you point me to where this is happening in the code?

The question is regarding #17340 (comment), i.e. where Q type is initialized to the default instance ?

[theresa-m]

can you point me to where this is happening in the code?

The question is regarding #17340 (comment), i.e. where Q type is initialized to the default instance ?

yes.

[hangshao0]

The default instance of a value type is allocated here:

openj9/runtime/vm/ClassInitialization.cpp

Lines 521 to 535 in 5567b8c

	if (J9_IS_J9CLASS_VALUETYPE(clazz)) {
	PUSH_OBJECT_IN_SPECIAL_FRAME(currentThread, initializationLock);
	/* Preparation is the earliest point where the defaultValue would needed.
	* I.e pre-filling static fields. Therefore, the defaultValue must be allocated at
	* the end of verification
	*/
	j9object_t defaultValue = vm->memoryManagerFunctions->J9AllocateObject(currentThread, clazz, J9_GC_ALLOCATE_OBJECT_TENURED | J9_GC_ALLOCATE_OBJECT_NON_INSTRUMENTABLE);
	initializationLock = POP_OBJECT_IN_SPECIAL_FRAME(currentThread);
	clazz = VM_VMHelpers::currentClass(clazz);
	if (NULL == defaultValue) {
	setHeapOutOfMemoryError(currentThread);
	goto done;
	}
	j9object_t *defaultValueAddress = &(clazz->flattenedClassCache->defaultValue);
	J9STATIC_OBJECT_STORE(currentThread, clazz, defaultValueAddress, defaultValue);

The Q type instance fields are set here:

openj9/runtime/gc_modron_startup/mgcalloc.cpp

Line 518 in 5567b8c

vmThread->javaVM->internalVMFunctions->defaultValueWithUnflattenedFlattenables(vmThread, clazz, objectPtr);

openj9/runtime/vm/ValueTypeHelpers.cpp

Lines 78 to 96 in 5567b8c

	defaultValueWithUnflattenedFlattenables(J9VMThread *currentThread, J9Class *clazz, j9object_t instance)
	{
	J9FlattenedClassCacheEntry * entry = NULL;
	J9Class * entryClazz = NULL;
	UDATA length = clazz->flattenedClassCache->numberOfEntries;
	UDATA const objectHeaderSize = J9VMTHREAD_OBJECT_HEADER_SIZE(currentThread);
	MM_ObjectAccessBarrierAPI objectAccessBarrier(currentThread);
	for (UDATA index = 0; index < length; index++) {
	entry = J9_VM_FCC_ENTRY_FROM_CLASS(clazz, index);
	entryClazz = J9_VM_FCC_CLASS_FROM_ENTRY(entry);
	if (!J9_VM_FCC_ENTRY_IS_STATIC_FIELD(entry) && !J9_IS_FIELD_FLATTENED(entryClazz, entry->field)) {
	objectAccessBarrier.inlineMixedObjectStoreObject(currentThread,
	instance,
	entry->offset + objectHeaderSize,
	entryClazz->flattenedClassCache->defaultValue,
	false);
	}
	}
	}

defaultValueWithUnflattenedFlattenables() uses the default instance to set the fields if they are not flattened (inlined). Note the flattenedClassCache struct has all the J9Class of the Q type fields.

[theresa-m]

The remaining tasks for this issue are:

• Class verification + tests:
  JEP 401 NullRestricted attribute part 1 #17781
  NullRestricted attribute cant be in primitive or array field #18006
  NullRestricted: putstatic and putfield throw NPE on null assignment #18028
  NullRestricted attribute field class checks #18030
  NullRestricted field throws NPE on null assignment in withfield #18130
  Flatten non-static NullRestricted fields #18173
• cfdump support
  NullRestricted Valhalla attribute cfdump support #18035
• ddr support:
  ddr support for Valhalla NullRestricted attribute #18042
  Print NullRestricted flag for ddr classAndFlags #18045

Remaining tasks:

• NPE for aastore. For arrays javac does not store a NullRestricted flag with the field but instead inserts java/lang/Class.asNullRestrictedType and java/util/Objects.requireNonNull before calling aastore. asNullRestrictedType will need to be implemented.

[theresa-m]

@hangshao0 are there any other class verification checks you think should be added for NullRestricted? See the previous comment for all prs that are open against this issue.

[hangshao0]

I haven't reviewed #18030 yet. But notice the following in the spec:

Section 5.3.5.
5. For each non-static field of C declared with a NullRestricted attribute......the class or interface named by the field's type must be a value class with an ImplicitCreation attribute whose ACC_DEFAULT flag is set. If not, loading throws an IncompatibleClassChangeError

Section 5.4.2
1. For each static field of C declared with a NullRestricted attribute......The resolved type must be of a value class with an ImplicitCreation attribute whose ACC_DEFAULT flag set. If not, preparation fails with an IncompatibleClassChangeError.

Section 6.5
if the value to store is null and the resolved field is declared with a NullRestricted attribute, the withfield instruction throws a NullPointerException.

[theresa-m]

Thanks, #18030 should cover the first two but I didn't notice there was a specific error to be thrown. I'll update that.

[theresa-m]

What about this one? Section 6.5.

The initial value of any instance field declared with a NullRestricted attribute is the initial instance of the field's class. The initial value of any instance field of a primitive type is the default value of the type ([2.3](https://docs.oracle.com/javase/specs/jvms/se16/html/jvms-2.html#jvms-2.3)). The initial value of any other instance field is null.

[hangshao0]

What about this one? Section 6.5.

We are already doing that. See defaultValueWithUnflattenedFlattenables() where we are using the default initial instance (flattenedClassCache->defaultValue) to initialize the fields that are not flattened. For flattened fields, they are initialized to all zeros.

We need to change J9_IS_FIELD_FLATTENED() to include the check of null restricted attribute. Currently, it checks if the field class is flattened and it is not a volatile field > 8 bytes. We should only flatten null restricted fields, so the check for null restricted field should be added. In addition, we need to eventually replace all the places doing 'Q' type checks with null restricted attribute checks.

But I guess making the above changes now will break some tests, as javac is not updated to generate null restricted attribute yet in the default branch of Valhalla repo.

[theresa-m]

This issue will be closed by #19995. Further null restricted error handling will be related to CheckedType support #19764