Merge pull request #3122 from amicic/copy_with_RB

Read Barrier aware ref array GC copy helper
eclipse-openj9 · Oct 15, 2018 · b03cae1 · b03cae1
2 parents 7b95285 + 6878df8
commit b03cae1
Show file tree

Hide file tree

Showing 3 changed files with 68 additions and 23 deletions.
diff --git a/runtime/gc_include/ObjectAccessBarrierAPI.hpp b/runtime/gc_include/ObjectAccessBarrierAPI.hpp
@@ -324,17 +324,7 @@ class MM_ObjectAccessBarrierAPI
 		bool copyLockword = true;
 
 		if (OBJECT_HEADER_SHAPE_POINTERS == J9CLASS_SHAPE(objectClass)) {
-			if (j9gc_modron_readbar_none != _readBarrierType) {
-				copyLockword = false;
-				if (j9gc_modron_readbar_evacuate == _readBarrierType) {
-					/* TODO implement HW barriers */
-					currentThread->javaVM->memoryManagerFunctions->j9gc_objaccess_cloneIndexableObject(currentThread, (J9IndexableObject*)original, (J9IndexableObject*)copy);
-				} else {
-					currentThread->javaVM->memoryManagerFunctions->j9gc_objaccess_cloneIndexableObject(currentThread, (J9IndexableObject*)original, (J9IndexableObject*)copy);
-				}
-			} else {
-				VM_ArrayCopyHelpers::referenceArrayCopy(currentThread, original, 0, copy, 0, size);
-			}
+			VM_ArrayCopyHelpers::referenceArrayCopy(currentThread, original, 0, copy, 0, size);
 		} else {
 			VM_ArrayCopyHelpers::primitiveArrayCopy(currentThread, original, 0, copy, 0, size, (((J9ROMArrayClass*)objectClass->romClass)->arrayShape & 0x0000FFFF));
 		}

diff --git a/runtime/gc_modron_standard/StandardAccessBarrier.cpp b/runtime/gc_modron_standard/StandardAccessBarrier.cpp
@@ -603,6 +603,43 @@ MM_StandardAccessBarrier::getJNICriticalRegionCount(MM_GCExtensions *extensions)
 }
 
 #if defined(J9VM_GC_ARRAYLETS)
+#if defined(OMR_GC_CONCURRENT_SCAVENGER)
+I_32
+MM_StandardAccessBarrier::doCopyContiguousBackwardWithReadBarrier(J9VMThread *vmThread, J9IndexableObject *srcObject, J9IndexableObject *destObject, I_32 srcIndex, I_32 destIndex, I_32 lengthInSlots)
+{
+	srcIndex += lengthInSlots;
+	destIndex += lengthInSlots;
+
+	fj9object_t *srcSlot = (fj9object_t *)indexableEffectiveAddress(vmThread, srcObject, srcIndex, sizeof(fj9object_t));
+	fj9object_t *destSlot = (fj9object_t *)indexableEffectiveAddress(vmThread, destObject, destIndex, sizeof(fj9object_t));
+	fj9object_t *srcEndSlot = srcSlot - lengthInSlots;
+
+	while (srcSlot-- > srcEndSlot) {
+		preObjectRead(vmThread, (J9Object *)srcObject, srcSlot);
+
+		*--destSlot = *srcSlot;
+	}
+
+	return ARRAY_COPY_SUCCESSFUL;
+}
+
+I_32
+MM_StandardAccessBarrier::doCopyContiguousForwardWithReadBarrier(J9VMThread *vmThread, J9IndexableObject *srcObject, J9IndexableObject *destObject, I_32 srcIndex, I_32 destIndex, I_32 lengthInSlots)
+{
+	fj9object_t *srcSlot = (fj9object_t *)indexableEffectiveAddress(vmThread, srcObject, srcIndex, sizeof(fj9object_t));
+	fj9object_t *destSlot = (fj9object_t *)indexableEffectiveAddress(vmThread, destObject, destIndex, sizeof(fj9object_t));
+	fj9object_t *srcEndSlot = srcSlot + lengthInSlots;
+
+	while (srcSlot < srcEndSlot) {
+		preObjectRead(vmThread, (J9Object *)srcObject, srcSlot);
+
+		*destSlot++ = *srcSlot++;
+	}
+
+	return ARRAY_COPY_SUCCESSFUL;
+}
+#endif /* OMR_GC_CONCURRENT_SCAVENGER */
+
 /**
  * Finds opportunities for doing the copy without or partially executing writeBarrier.
  * @return ARRAY_COPY_SUCCESSFUL if copy was successful, ARRAY_COPY_NOT_DONE no copy is done
@@ -623,17 +660,24 @@ MM_StandardAccessBarrier::backwardReferenceArrayCopyIndex(J9VMThread *vmThread,
 		Assert_MM_true(destObject == srcObject);
 		Assert_MM_true(_extensions->indexableObjectModel.isInlineContiguousArraylet(destObject));
 
-		if (!_extensions->isConcurrentScavengerEnabled()) {
+#if defined(OMR_GC_CONCURRENT_SCAVENGER)
+		if (_extensions->isConcurrentScavengerInProgress()) {
+			/* During active CS cycle, we need a RB for every slot being copied.
+			 * For WB same rules apply - just need the final batch barrier.
+			 */
+			retValue = doCopyContiguousBackwardWithReadBarrier(vmThread, srcObject, destObject, srcIndex, destIndex, lengthInSlots);
+		} else
+#endif /* OMR_GC_CONCURRENT_SCAVENGER */
+		{
 			retValue = doCopyContiguousBackward(vmThread, srcObject, destObject, srcIndex, destIndex, lengthInSlots);
-			Assert_MM_true(retValue == ARRAY_COPY_SUCCESSFUL);
-
-			preBatchObjectStoreImpl(vmThread, (J9Object *)destObject);
 		}
+		Assert_MM_true(retValue == ARRAY_COPY_SUCCESSFUL);
+
+		preBatchObjectStoreImpl(vmThread, (J9Object *)destObject);
 	}
 	return retValue;
 }
 
-
 /**
  * Finds opportunities for doing the copy without or partially executing writeBarrier.
  * @return ARRAY_COPY_SUCCESSFUL if copy was successful, ARRAY_COPY_NOT_DONE no copy is done
@@ -654,14 +698,20 @@ MM_StandardAccessBarrier::forwardReferenceArrayCopyIndex(J9VMThread *vmThread, J
 		Assert_MM_true(_extensions->indexableObjectModel.isInlineContiguousArraylet(destObject));
 		Assert_MM_true(_extensions->indexableObjectModel.isInlineContiguousArraylet(srcObject));
 
-		if (!_extensions->isConcurrentScavengerEnabled()) {
-			/* todo: for Concurrent Scavenger, create a helper that will invoke load barrier on each source object slot,
-			 * but only one batchstore barrier on the destination object */
+#if defined(OMR_GC_CONCURRENT_SCAVENGER)
+		if (_extensions->isConcurrentScavengerInProgress()) {
+			/* During active CS cycle, we need a RB for every slot being copied.
+			 * For WB same rules apply - just need the final batch barrier.
+			 */
+			retValue = doCopyContiguousForwardWithReadBarrier(vmThread, srcObject, destObject, srcIndex, destIndex, lengthInSlots);
+		} else
+#endif /* OMR_GC_CONCURRENT_SCAVENGER */
+		{
 			retValue = doCopyContiguousForward(vmThread, srcObject, destObject, srcIndex, destIndex, lengthInSlots);
-			Assert_MM_true(retValue == ARRAY_COPY_SUCCESSFUL);
-
-			preBatchObjectStoreImpl(vmThread, (J9Object *)destObject);
 		}
+
+		Assert_MM_true(retValue == ARRAY_COPY_SUCCESSFUL);
+		preBatchObjectStoreImpl(vmThread, (J9Object *)destObject);
 	}
 	return retValue;
 }

diff --git a/runtime/gc_modron_standard/StandardAccessBarrier.hpp b/runtime/gc_modron_standard/StandardAccessBarrier.hpp
@@ -53,7 +53,12 @@ class MM_StandardAccessBarrier : public MM_ObjectAccessBarrier
 #endif /* OMR_GC_REALTIME */
 	void postObjectStoreImpl(J9VMThread *vmThread, J9Object *dstObject, J9Object *srcObject);
 	void preBatchObjectStoreImpl(J9VMThread *vmThread, J9Object *dstObject);
-
+
+#if defined(OMR_GC_CONCURRENT_SCAVENGER)
+	I_32 doCopyContiguousBackwardWithReadBarrier(J9VMThread *vmThread, J9IndexableObject *srcObject, J9IndexableObject *destObject, I_32 srcIndex, I_32 destIndex, I_32 lengthInSlots);
+	I_32 doCopyContiguousForwardWithReadBarrier(J9VMThread *vmThread, J9IndexableObject *srcObject, J9IndexableObject *destObject, I_32 srcIndex, I_32 destIndex, I_32 lengthInSlots);
+#endif /* OMR_GC_CONCURRENT_SCAVENGER */
+
 protected:
 	virtual bool initialize(MM_EnvironmentBase *env);
 	virtual void tearDown(MM_EnvironmentBase *env);