/**
* A runtime guard block may have monitor stores and privarg stores along with the guard
* it self. This method will rearrange these stores and split the block, managing any
* uncommoning necessary for eventual block order.
*
* The provided block will become the privarg block, containing any privarg stores and additonal
* temps for uncommoning. It must be evaluated first. The returned block will contain monitor
* stores and the guard. If no split is required, the provided block will be returned.
*
* @param comp Compilation object
* @param block Block to manipulate
* @param cfg Current CFG
* @return The block containing the guard.
*/
static TR::Block* splitRuntimeGuardBlock(TR::Compilation *comp, TR::Block* block, TR::CFG *cfg)
{
TR::NodeChecklist checklist(comp);
TR::TreeTop *start = block->getFirstRealTreeTop();
TR::TreeTop *guard = block->getLastRealTreeTop();
TR::TreeTop *firstPrivArg = NULL;
TR::TreeTop *firstMonitor = NULL;
// Manage the unexpected case that monitors and priv args are reversed
bool privThenMonitor = false;
TR_ASSERT(isMergeableGuard(guard->getNode()), "last node must be guard %p", guard->getNode());
// Search for privarg and monitor stores
// Only commoned nodes under the guard are required to be anchored, due to the guard being
// evaluted before the monitor stores later on
bool anchoredTemps = false;
for (TR::TreeTop *tt = start; tt && tt->getNode()->getOpCodeValue() != TR::BBEnd; tt = tt->getNextTreeTop())
{
TR::Node * node = tt->getNode();
if (node->getOpCode().hasSymbolReference() && node->getSymbol()->holdsMonitoredObject())
firstMonitor = firstMonitor == NULL ? tt : firstMonitor;
else if (node->chkIsPrivatizedInlinerArg())
{
if (firstPrivArg == NULL)
{
firstPrivArg = tt;
privThenMonitor = (firstMonitor == NULL);
}
}
else if (isMergeableGuard(node))
anchoredTemps |= anchorCommonNodes(comp, node, start, checklist);
else
TR_ASSERT(0, "Node other than monitor or privarg store %p before runtime guard", node);
}
// If there are monitors then privargs, they must be swapped around, such that all privargs are
// evaluated first
if (firstPrivArg && firstMonitor && !privThenMonitor)
{
TR::TreeTop *monitorEnd = firstPrivArg->getPrevTreeTop();
firstMonitor->getPrevTreeTop()->join(firstPrivArg);
guard->getPrevTreeTop()->join(firstMonitor);
monitorEnd->join(guard);
}
// If there were temps created or privargs in the block, perform a split
TR::TreeTop *split = NULL;
if (firstPrivArg)
split = firstMonitor ? firstMonitor : guard;
else if (anchoredTemps)
split = start;
if (split)
return block->split(split, cfg, true /* fixupCommoning */, false /* copyExceptionSuccessors */);
return block;
}
void TR_ExpressionsSimplification::tranformStoreMotionCandidate(TR::TreeTop *treeTop, bool *isPreheaderBlockInvalid)
{
TR::Node *node = treeTop->getNode();
TR_ASSERT(node->getOpCode().isStore() && !node->getSymbol()->isStatic() && !node->getSymbol()->holdsMonitoredObject(),
"node %p was expected to be a non-static non-monitored object store and was not.", node);
// this candidate should be valid, either direct or indirect
if (trace())
comp()->getDebug()->print(comp()->getOutFile(), node, 0, true);
TR::Block *entryBlock = _currentRegion->getEntryBlock();
TR::Block *preheaderBlock = findPredecessorBlock(entryBlock);
if (!preheaderBlock)
{
if (trace())
traceMsg(comp(), "Fail to find a place to put the hoist code in\n");
*isPreheaderBlockInvalid = true;
return;
}
// Earlier post-dominance test ensures that the loop is executed as least once, or is canonicalized.
// but to be safe we still perform on canonicalized loops only.
if (_currentRegion->isCanonicalizedLoop()) // make sure that the loop is canonicalized, in which case the preheader is
{ // executed in its first iteration and is protected.
if (performTransformation(comp(), "%sMove out loop-invariant store [%p] to block_%d\n", OPT_DETAILS, node, preheaderBlock->getNumber()))
{
TR::Node *newNode = node->duplicateTree();
transformNode(newNode, preheaderBlock);
TR::TransformUtil::removeTree(comp(), treeTop);
}
}
else
{
if (trace())
traceMsg(comp(), "No canonicalized loop for this candidate\n");
}
}
// Returns true if there is any constraint to the move
bool TR_LocalLiveRangeReduction::isAnySymInDefinedOrUsedBy(TR_TreeRefInfo *currentTreeRefInfo, TR::Node *currentNode, TR_TreeRefInfo *movingTreeRefInfo )
{
TR::Node *movingNode = movingTreeRefInfo->getTreeTop()->getNode();
// ignore anchors
//
if (movingNode->getOpCode().isAnchor())
movingNode = movingNode->getFirstChild();
TR::ILOpCode &opCode = currentNode->getOpCode();
////if ((opCode.getOpCodeValue() == TR::monent) || (opCode.getOpCodeValue() == TR::monexit))
if (nodeMaybeMonitor(currentNode))
{
if (trace())
traceMsg(comp(),"cannot move %p beyond monitor %p\n",movingNode,currentNode);
return true;
}
// Don't move gc points or things across gc points
//
if (movingNode->canGCandReturn() ||
currentNode->canGCandReturn())
{
if (trace())
traceMsg(comp(), "cannot move gc points %p past %p\n", movingNode, currentNode);
return true;
}
// Don't move checks or calls at all
//
if (containsCallOrCheck(movingTreeRefInfo,movingNode))
{
if (trace())
traceMsg(comp(),"cannot move check or call %s\n", getDebug()->getName(movingNode));
return true;
}
// Don't move object header store past a GC point
//
if ((currentNode->getOpCode().isWrtBar() || currentNode->canCauseGC()) && mayBeObjectHeaderStore(movingNode, fe()))
{
if (trace())
traceMsg(comp(),"cannot move possible object header store %s past GC point %s\n", getDebug()->getName(movingNode), getDebug()->getName(currentNode));
return true;
}
if (TR::Compiler->target.cpu.isPower() && opCode.getOpCodeValue() == TR::allocationFence)
{
// Can't move allocations past flushes
if (movingNode->getOpCodeValue() == TR::treetop &&
movingNode->getFirstChild()->getOpCode().isNew() &&
(currentNode->getAllocation() == NULL ||
currentNode->getAllocation() == movingNode->getFirstChild()))
{
if (trace())
{
traceMsg(comp(),"cannot move %p beyond flush %p - ", movingNode, currentNode);
if (currentNode->getAllocation() == NULL)
traceMsg(comp(),"(flush with null allocation)\n");
else
traceMsg(comp(),"(flush for allocation %p)\n", currentNode->getAllocation());
}
return true;
}
// Can't move certain stores past flushes
// Exclude all indirect stores, they may be for stack allocs, in which case the flush is needed at least as a scheduling barrier
// Direct stores to autos and parms are the only safe candidates
if (movingNode->getOpCode().isStoreIndirect() ||
(movingNode->getOpCode().isStoreDirect() && !movingNode->getSymbol()->isParm() && !movingNode->getSymbol()->isAuto()))
{
if (trace())
traceMsg(comp(),"cannot move %p beyond flush %p - (flush for possible stack alloc)", movingNode, currentNode);
return true;
}
}
for (int32_t i = 0; i < currentNode->getNumChildren(); i++)
{
TR::Node *child = currentNode->getChild(i);
//Any node that has side effects (like call and newarrya) cannot be evaluated in the middle of the tree.
if (movingTreeRefInfo->getFirstRefNodesList()->find(child))
{
//for calls and unresolve symbol that are not under check
if (child->exceptionsRaised() ||
(child->getOpCode().hasSymbolReference() && child->getSymbolReference()->isUnresolved()))
{
if (trace())
traceMsg(comp(),"cannot move %p beyond %p - cannot change evaluation point of %p\n ",movingNode,currentTreeRefInfo->getTreeTop()->getNode(),child);
return true;
}
else if(movingNode->getOpCode().isStore())
{
TR::SymbolReference *stSymRef = movingNode->getSymbolReference();
int32_t stSymRefNum = stSymRef->getReferenceNumber();
//TR::SymbolReference *stSymRef = movingNode->getSymbolReference();
int32_t numHelperSymbols = comp()->getSymRefTab()->getNumHelperSymbols();
if ((comp()->getSymRefTab()->isNonHelper(stSymRefNum, TR::SymbolReferenceTable::vftSymbol))||
(comp()->getSymRefTab()->isNonHelper(stSymRefNum, TR::SymbolReferenceTable::contiguousArraySizeSymbol))||
(comp()->getSymRefTab()->isNonHelper(stSymRefNum, TR::SymbolReferenceTable::discontiguousArraySizeSymbol))||
(stSymRef == comp()->getSymRefTab()->findHeaderFlagsSymbolRef())||
(stSymRef->getSymbol() == comp()->getSymRefTab()->findGenericIntShadowSymbol()))
return true;
}
else if (movingNode->getOpCode().isResolveOrNullCheck())
{
if (trace())
traceMsg(comp(),"cannot move %p beyond %p - node %p under ResolveOrNullCheck",movingNode,currentTreeRefInfo->getTreeTop()->getNode(),currentNode);
return true;
}
else if (TR::Compiler->target.is64Bit() &&
movingNode->getOpCode().isBndCheck() &&
((opCode.getOpCodeValue() == TR::i2l) || (opCode.getOpCodeValue() == TR::iu2l)) &&
!child->isNonNegative())
{
if (trace())
traceMsg(comp(),"cannot move %p beyond %p - changing the eval point of %p will casue extra cg instruction ",movingNode,currentTreeRefInfo->getTreeTop()->getNode(),currentNode);
return true;
}
}
//don't recurse over nodes each are not the first reference
if (child->getReferenceCount()==1 || currentTreeRefInfo->getFirstRefNodesList()->find(child))
{
if (isAnySymInDefinedOrUsedBy(currentTreeRefInfo, child, movingTreeRefInfo ))
return true;
}
}
return false;
}
TR_BitVector *
addVeryRefinedCallAliasSets(TR::ResolvedMethodSymbol * methodSymbol, TR_BitVector * aliases, List<void> * methodsPeeked)
{
TR::Compilation *comp = TR::comp();
void * methodId = methodSymbol->getResolvedMethod()->getPersistentIdentifier();
if (methodsPeeked->find(methodId))
{
// This can't be allocated into the alias region as it must be accessed across optimizations
TR_BitVector *heapAliases = new (comp->trHeapMemory()) TR_BitVector(comp->getSymRefCount(), comp->trMemory(), heapAlloc, growable);
*heapAliases |= *aliases;
return heapAliases;
}
// stop if the peek is getting very deep
//
if (methodsPeeked->getSize() >= PEEK_THRESHOLD)
return 0;
methodsPeeked->add(methodId);
dumpOptDetails(comp, "O^O REFINING ALIASES: Peeking into the IL to refine aliases \n");
if (!methodSymbol->getResolvedMethod()->genMethodILForPeeking(methodSymbol, comp, true))
return 0;
TR::SymbolReferenceTable * symRefTab = comp->getSymRefTab();
for (TR::TreeTop * tt = methodSymbol->getFirstTreeTop(); tt; tt = tt->getNextTreeTop())
{
TR::Node *node = tt->getNode();
if (node->getOpCode().isResolveCheck())
return 0;
if ((node->getOpCodeValue() == TR::treetop) ||
(node->getOpCodeValue() == TR::compressedRefs) ||
node->getOpCode().isCheck())
node = node->getFirstChild();
if (node->getOpCode().isStore())
{
TR::SymbolReference * symRefInCallee = node->getSymbolReference(), * symRefInCaller;
TR::Symbol * symInCallee = symRefInCallee->getSymbol();
TR::DataType type = symInCallee->getDataType();
if (symInCallee->isShadow())
{
if (symInCallee->isArrayShadowSymbol())
symRefInCaller = symRefTab->getSymRef(symRefTab->getArrayShadowIndex(type));
else if (symInCallee->isArrayletShadowSymbol())
symRefInCaller = symRefTab->getSymRef(symRefTab->getArrayletShadowIndex(type));
else
symRefInCaller = symRefTab->findShadowSymbol(symRefInCallee->getOwningMethod(comp), symRefInCallee->getCPIndex(), type);
if (symRefInCaller)
{
if (symRefInCaller->reallySharesSymbol(comp))
symRefInCaller->setSharedShadowAliases(aliases, symRefTab);
aliases->set(symRefInCaller->getReferenceNumber());
}
}
else if (symInCallee->isStatic())
{
symRefInCaller = symRefTab->findStaticSymbol(symRefInCallee->getOwningMethod(comp), symRefInCallee->getCPIndex(), type);
if (symRefInCaller)
{
if (symRefInCaller->reallySharesSymbol(comp))
symRefInCaller->setSharedStaticAliases(aliases, symRefTab);
else
aliases->set(symRefInCaller->getReferenceNumber());
}
}
}
else if (node->getOpCode().isCall())
{
if (node->getOpCode().isCallIndirect())
return 0;
TR::ResolvedMethodSymbol * calleeSymbol = node->getSymbol()->getResolvedMethodSymbol();
if (!calleeSymbol)
return 0;
TR_ResolvedMethod * calleeMethod = calleeSymbol->getResolvedMethod();
if (!calleeMethod->isCompilable(comp->trMemory()) || calleeMethod->isJNINative())
return 0;
if (!addVeryRefinedCallAliasSets(calleeSymbol, aliases, methodsPeeked))
return 0;
}
else if (node->getOpCodeValue() == TR::monent)
return 0;
}
// This can't be allocated into the alias region as it must be accessed across optimizations
TR_BitVector *heapAliases = new (comp->trHeapMemory()) TR_BitVector(comp->getSymRefCount(), comp->trMemory(), heapAlloc, growable);
*heapAliases |= *aliases;
return heapAliases;
}
TR::Node *constrainVcall(TR::ValuePropagation *vp, TR::Node *node)
{
constrainCall(vp, node);
// Look for System.arraycopy call. If the node is transformed into an arraycopy
// re-process it.
//
vp->transformArrayCopyCall(node);
if (node->getOpCodeValue() == TR::arraycopy)
{
node->setVisitCount(0);
vp->launchNode(node, vp->getCurrentParent(), 0);
return node;
}
if (vp->transformUnsafeCopyMemoryCall(node))
return node;
cacheStringAppend(vp,node);
#ifdef J9_PROJECT_SPECIFIC
TR::SymbolReference *finalizeSymRef = vp->comp()->getSymRefTab()->findOrCreateRuntimeHelper(TR_jitCheckIfFinalizeObject, true, true, true);
if (node->getSymbolReference() == finalizeSymRef)
{
TR::Node *receiver = node->getFirstChild();
bool isGlobal;
TR::VPConstraint *type = vp->getConstraint(receiver, isGlobal);
bool canBeRemoved = false;
// ensure the type is really a fixedClass
// resolvedClass is not sufficient because java.lang.Object has an
// empty finalizer method (hasFinalizer returns false) and the call to
// vm helper is incorrectly optimized in this case
//
if (type && type->getClassType() &&
type->getClassType()->asFixedClass())
{
TR_OpaqueClassBlock *klass = type->getClassType()->getClass();
if (klass && !TR::Compiler->cls.hasFinalizer(vp->comp(), klass) && !vp->comp()->fej9()->isOwnableSyncClass(klass))
{
canBeRemoved = true;
}
}
// If a class has a finalizer or is an ownableSync it won't be allocated on the stack. That's ensured
// by virtue of (indirectly) calling bool J9::ObjectModel::canAllocateInlineClass(TR_OpaqueClassBlock *block)
// Doesn't make sense to call jitCheckIfFinalizeObject for a stack
// allocated object, so optimize
else if (receiver->getOpCode().hasSymbolReference() && receiver->getSymbol()->isLocalObject())
{
canBeRemoved = true;
}
if (canBeRemoved &&
performTransformation(vp->comp(), "%s Removing redundant call to jitCheckIfFinalize [%p]\n", OPT_DETAILS, node))
{
///printf("found opportunity in %s to remove call to checkfinalize\n", vp->comp()->signature());fflush(stdout);
///traceMsg(vp->comp(), "found opportunity to remove call %p to checkfinalize\n", node);
vp->removeNode(node);
vp->_curTree->setNode(NULL);
return node;
}
}
#endif
return node;
}
void
OMR::CodeGenPhase::performSetupForInstructionSelectionPhase(TR::CodeGenerator * cg, TR::CodeGenPhase * phase)
{
TR::Compilation *comp = cg->comp();
if (TR::Compiler->target.cpu.isZ() && TR::Compiler->om.shouldGenerateReadBarriersForFieldLoads())
{
// TODO (GuardedStorage): We need to come up with a better solution than anchoring aloadi's
// to enforce certain evaluation order
traceMsg(comp, "GuardedStorage: in performSetupForInstructionSelectionPhase\n");
auto mapAllocator = getTypedAllocator<std::pair<TR::TreeTop*, TR::TreeTop*> >(comp->allocator());
std::map<TR::TreeTop*, TR::TreeTop*, std::less<TR::TreeTop*>, TR::typed_allocator<std::pair<TR::TreeTop* const, TR::TreeTop*>, TR::Allocator> >
currentTreeTopToappendTreeTop(std::less<TR::TreeTop*> (), mapAllocator);
TR_BitVector *unAnchorableAloadiNodes = comp->getBitVectorPool().get();
for (TR::PreorderNodeIterator iter(comp->getStartTree(), comp); iter != NULL; ++iter)
{
TR::Node *node = iter.currentNode();
traceMsg(comp, "GuardedStorage: Examining node = %p\n", node);
// isNullCheck handles both TR::NULLCHK and TR::ResolveAndNULLCHK
// both of which do not operate on their child but their
// grandchild (or greatgrandchild).
if (node->getOpCode().isNullCheck())
{
// An aloadi cannot be anchored if there is a Null Check on
// its child. There are two situations where this occurs.
// The first is when doing an aloadi off some node that is
// being NULLCHK'd (see Ex1). The second is when doing an
// icalli in which case the aloadi loads the VFT of an
// object that must be NULLCHK'd (see Ex2).
//
// Ex1:
// n1n NULLCHK on n3n
// n2n aloadi f <-- First Child And Parent of Null Chk'd Node
// n3n aload O
//
// Ex2:
// n1n NULLCHK on n4n
// n2n icall foo <-- First Child
// n3n aloadi <vft> <-- Parent of Null Chk'd Node
// n4n aload O
// n4n ==> aload O
TR::Node *nodeBeingNullChkd = node->getNullCheckReference();
if (nodeBeingNullChkd)
{
TR::Node *firstChild = node->getFirstChild();
TR::Node *parentOfNullChkdNode = NULL;
if (firstChild->getOpCode().isCall() &&
firstChild->getOpCode().isIndirect())
{
parentOfNullChkdNode = firstChild->getFirstChild();
}
else
{
parentOfNullChkdNode = firstChild;
}
if (parentOfNullChkdNode &&
parentOfNullChkdNode->getOpCodeValue() == TR::aloadi &&
parentOfNullChkdNode->getNumChildren() > 0 &&
parentOfNullChkdNode->getFirstChild() == nodeBeingNullChkd)
{
unAnchorableAloadiNodes->set(parentOfNullChkdNode->getGlobalIndex());
traceMsg(comp, "GuardedStorage: Cannot anchor %p\n", firstChild);
}
}
}
else
{
bool shouldAnchorNode = false;
if (node->getOpCodeValue() == TR::aloadi &&
!unAnchorableAloadiNodes->isSet(node->getGlobalIndex()))
{
shouldAnchorNode = true;
}
else if (node->getOpCodeValue() == TR::aload &&
node->getSymbol()->isStatic() &&
node->getSymbol()->isCollectedReference())
{
shouldAnchorNode = true;
}
if (shouldAnchorNode)
{
TR::TreeTop* anchorTreeTop = TR::TreeTop::create(comp, TR::Node::create(TR::treetop, 1, node));
TR::TreeTop* appendTreeTop = iter.currentTree();
if (currentTreeTopToappendTreeTop.count(appendTreeTop) > 0)
{
appendTreeTop = currentTreeTopToappendTreeTop[appendTreeTop];
}
// Anchor the aload/aloadi before the current treetop
appendTreeTop->insertBefore(anchorTreeTop);
currentTreeTopToappendTreeTop[iter.currentTree()] = anchorTreeTop;
traceMsg(comp, "GuardedStorage: Anchored %p to treetop = %p\n", node, anchorTreeTop);
}
}
}
comp->getBitVectorPool().release(unAnchorableAloadiNodes);
}
if (cg->shouldBuildStructure() &&
(comp->getFlowGraph()->getStructure() != NULL))
{
TR_Structure *rootStructure = TR_RegionAnalysis::getRegions(comp);
comp->getFlowGraph()->setStructure(rootStructure);
}
phase->reportPhase(SetupForInstructionSelectionPhase);
// Dump preIR
if (comp->getOption(TR_TraceRegisterPressureDetails) && !comp->getOption(TR_DisableRegisterPressureSimulation))
{
traceMsg(comp, " { Post optimization register pressure simulation\n");
TR_BitVector emptyBitVector;
vcount_t vc = comp->incVisitCount();
cg->initializeRegisterPressureSimulator();
for (TR::Block *block = comp->getStartBlock(); block; block = block->getNextExtendedBlock())
{
TR_LinkHead<TR_RegisterCandidate> emptyCandidateList;
TR::CodeGenerator::TR_RegisterPressureState state(NULL, 0, emptyBitVector, emptyBitVector, &emptyCandidateList, cg->getNumberOfGlobalGPRs(), cg->getNumberOfGlobalFPRs(), cg->getNumberOfGlobalVRFs(), vc);
TR::CodeGenerator::TR_RegisterPressureSummary summary(state._gprPressure, state._fprPressure, state._vrfPressure);
cg->simulateBlockEvaluation(block, &state, &summary);
}
traceMsg(comp, " }\n");
}
TR::LexicalMemProfiler mp(phase->getName(), comp->phaseMemProfiler());
LexicalTimer pt(phase->getName(), comp->phaseTimer());
cg->setUpForInstructionSelection();
}