// Add an async check into a block - MUST be at block entry
//
void TR_AsyncCheckInsertion::insertAsyncCheck(TR::Block *block, TR::Compilation *comp, const char *counterPrefix)
{
TR::TreeTop *lastTree = block->getLastRealTreeTop();
TR::TreeTop *asyncTree =
TR::TreeTop::create(comp,
TR::Node::createWithSymRef(lastTree->getNode(), TR::asynccheck, 0,
comp->getSymRefTab()->findOrCreateAsyncCheckSymbolRef(comp->getMethodSymbol())));
if (lastTree->getNode()->getOpCode().isReturn())
{
TR::TreeTop *prevTree = lastTree->getPrevTreeTop();
prevTree->join(asyncTree);
asyncTree->join(lastTree);
}
else
{
TR::TreeTop *nextTree = block->getEntry()->getNextTreeTop();
block->getEntry()->join(asyncTree);
asyncTree->join(nextTree);
}
const char * const name = TR::DebugCounter::debugCounterName(comp,
"asynccheck.insert/%s/(%s)/%s/block_%d",
counterPrefix,
comp->signature(),
comp->getHotnessName(),
block->getNumber());
TR::DebugCounter::prependDebugCounter(comp, name, asyncTree->getNextTreeTop());
}
void
TR_ExpressionsSimplification::transformNode(TR::Node *srcNode, TR::Block *dstBlock)
{
TR::TreeTop *lastTree = dstBlock->getLastRealTreeTop();
TR::TreeTop *prevTree = lastTree->getPrevTreeTop();
TR::TreeTop *srcNodeTT = TR::TreeTop::create(comp(), srcNode);
if (trace())
comp()->getDebug()->print(comp()->getOutFile(),srcNode,0,true);
if (lastTree->getNode()->getOpCode().isBranch() ||
(lastTree->getNode()->getOpCode().isJumpWithMultipleTargets() && lastTree->getNode()->getOpCode().hasBranchChildren()))
{
srcNodeTT->join(lastTree);
prevTree->join(srcNodeTT);
}
/*
else if (dstBlock->getEntry()->getNode()->getOpCodeValue() == TR::BBStart)
{
srcNodeTT->join(dstBlock->getExit());
dstBlock->getEntry()->join(srcNodeTT);
}
*/
else
{
srcNodeTT->join(dstBlock->getExit());
lastTree->join(srcNodeTT);
}
return;
}
/**
* 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;
}
static bool fixUpTree(TR::Node *node, TR::TreeTop *treeTop, TR::NodeChecklist &visited, bool &highGlobalIndex, TR::Optimization *opt, vcount_t evaluatedVisitCount)
{
if (node->getVisitCount() == evaluatedVisitCount)
return false;
if (visited.contains(node))
return false;
visited.add(node);
bool containsFloatingPoint = false;
bool anchorLoadaddr = true;
bool anchorArrayCmp = true;
// for arraycmp node, don't create its tree top anchor
// fold it into if statment and save jump instruction
if (node->getOpCodeValue() == TR::arraycmp &&
!node->isArrayCmpLen() &&
TR::Compiler->target.cpu.isX86())
{
anchorArrayCmp = false;
}
if ((node->getReferenceCount() > 1) &&
!node->getOpCode().isLoadConst() &&
anchorLoadaddr &&
anchorArrayCmp)
{
if (!opt->comp()->getOption(TR_ProcessHugeMethods))
{
int32_t nodeCount = opt->comp()->getNodeCount();
int32_t nodeCountLimit = 3 * USHRT_MAX / 4;
if (nodeCount > nodeCountLimit)
{
dumpOptDetails(opt->comp(),
"%snode count %d exceeds limit %d\n",
opt->optDetailString(), nodeCount, nodeCountLimit);
highGlobalIndex = true;
return containsFloatingPoint;
}
}
if (node->getOpCode().isFloatingPoint())
containsFloatingPoint = true;
TR::TreeTop *nextTree = treeTop->getNextTreeTop();
node->incFutureUseCount();
TR::TreeTop *anchorTreeTop = TR::TreeTop::create(opt->comp(), TR::Node::create(TR::treetop, 1, node));
anchorTreeTop->getNode()->setFutureUseCount(0);
treeTop->join(anchorTreeTop);
anchorTreeTop->join(nextTree);
}
else
{
for (int32_t i = 0; i < node->getNumChildren(); ++i)
{
TR::Node *child = node->getChild(i);
if (fixUpTree(child, treeTop, visited, highGlobalIndex, opt, evaluatedVisitCount))
containsFloatingPoint = true;
}
}
return containsFloatingPoint;
}
int32_t TR::DeadTreesElimination::process(TR::TreeTop *startTree, TR::TreeTop *endTree)
{
TR::StackMemoryRegion stackRegion(*comp()->trMemory());
LongestPathMap longestPaths(std::less<TR::Node*>(), stackRegion);
typedef TR::typed_allocator<CRAnchor, TR::Region&> CRAnchorAlloc;
typedef TR::forward_list<CRAnchor, CRAnchorAlloc> CRAnchorList;
CRAnchorList anchors(stackRegion);
vcount_t visitCount = comp()->incOrResetVisitCount();
TR::TreeTop *treeTop;
for (treeTop = startTree; (treeTop != endTree); treeTop = treeTop->getNextTreeTop())
treeTop->getNode()->initializeFutureUseCounts(visitCount);
TR::Block *block = NULL;
bool delayedRegStoresBeforeThisPass = _delayedRegStores;
// Update visitCount as they are used in this optimization and need to be
visitCount = comp()->incOrResetVisitCount();
for (TR::TreeTopIterator iter(startTree, comp()); iter != endTree; ++iter)
{
TR::Node *node = iter.currentTree()->getNode();
if (node->getOpCodeValue() == TR::BBStart)
{
block = node->getBlock();
if (!block->isExtensionOfPreviousBlock())
longestPaths.clear();
}
int vcountLimit = MAX_VCOUNT - 3;
if (comp()->getVisitCount() > vcountLimit)
{
dumpOptDetails(comp(),
"%sVisit count %d exceeds limit %d; stopping\n",
optDetailString(), comp()->getVisitCount(), vcountLimit);
return 0;
}
// correct at all intermediate stages
//
if ((node->getOpCodeValue() != TR::treetop) &&
(!node->getOpCode().isAnchor() || (node->getFirstChild()->getReferenceCount() != 1)) &&
(!node->getOpCode().isStoreReg() || (node->getFirstChild()->getReferenceCount() != 1)) &&
(delayedRegStoresBeforeThisPass ||
(iter.currentTree() == block->getLastRealTreeTop()) ||
!node->getOpCode().isStoreReg() ||
(node->getVisitCount() == visitCount)))
{
if (node->getOpCode().isAnchor() && node->getFirstChild()->getOpCode().isLoadIndirect())
anchors.push_front(CRAnchor(iter.currentTree(), block));
TR::TransformUtil::recursivelySetNodeVisitCount(node, visitCount);
continue;
}
if (node->getOpCode().isStoreReg())
_delayedRegStores = true;
TR::Node *child = node->getFirstChild();
if (child->getOpCodeValue() == TR::PassThrough)
{
TR::Node *newChild = child->getFirstChild();
node->setAndIncChild(0, newChild);
newChild->incFutureUseCount();
if (child->getReferenceCount() <= 1)
optimizer()->prepareForNodeRemoval(child);
child->recursivelyDecReferenceCount();
recursivelyDecFutureUseCount(child);
child = newChild;
}
bool treeTopCanBeEliminated = false;
// If the treetop child has been seen before then it must be anchored
// somewhere above already; so we don't need the treetop to be anchoring
// this node (as the computation is already done at the first reference to
// the node).
//
if (visitCount == child->getVisitCount())
{
treeTopCanBeEliminated = true;
}
else
{
TR::ILOpCode &childOpCode = child->getOpCode();
TR::ILOpCodes opCodeValue = childOpCode.getOpCodeValue();
bool seenConditionalBranch = false;
bool callWithNoSideEffects = child->getOpCode().isCall() &&
child->getSymbolReference()->getSymbol()->isResolvedMethod() &&
child->getSymbolReference()->getSymbol()->castToResolvedMethodSymbol()->isSideEffectFree();
if (callWithNoSideEffects)
{
treeTopCanBeEliminated = true;
}
else if (!((childOpCode.isCall() && !callWithNoSideEffects) ||
childOpCode.isStore() ||
((opCodeValue == TR::New ||
opCodeValue == TR::anewarray ||
opCodeValue == TR::newarray) &&
child->getReferenceCount() > 1) ||
opCodeValue == TR::multianewarray ||
opCodeValue == TR::MergeNew ||
opCodeValue == TR::checkcast ||
opCodeValue == TR::Prefetch ||
opCodeValue == TR::iu2l ||
((childOpCode.isDiv() ||
childOpCode.isRem()) &&
child->getNumChildren() == 3)))
{
// Perform the rather complex check to see whether its safe
// to disconnect the child node from the treetop
//
bool safeToReplaceNode = false;
if (child->getReferenceCount() == 1)
{
safeToReplaceNode = true;
#ifdef J9_PROJECT_SPECIFIC
if (child->getOpCode().isPackedExponentiation())
{
// pdexp has a possible message side effect in truncating or no significant digits left cases
safeToReplaceNode = false;
}
#endif
if (opCodeValue == TR::loadaddr)
treeTopCanBeEliminated = true;
}
else if (!_cannotBeEliminated)
{
safeToReplaceNode = isSafeToReplaceNode(
child,
iter.currentTree(),
&seenConditionalBranch,
visitCount,
comp(),
&_targetTrees,
_cannotBeEliminated,
longestPaths);
}
if (safeToReplaceNode)
{
if (childOpCode.hasSymbolReference())
{
TR::SymbolReference *symRef = child->getSymbolReference();
if (symRef->getSymbol()->isAuto() || symRef->getSymbol()->isParm())
treeTopCanBeEliminated = true;
else
{
if (childOpCode.isLoad() ||
(opCodeValue == TR::loadaddr) ||
(opCodeValue == TR::instanceof) ||
(((opCodeValue == TR::New) ||
(opCodeValue == TR::anewarray ||
opCodeValue == TR::newarray)) &&
///child->getFirstChild()->isNonNegative()))
child->markedAllocationCanBeRemoved()))
// opCodeValue == TR::multianewarray ||
// opCodeValue == TR::MergeNew)
treeTopCanBeEliminated = true;
}
}
else
treeTopCanBeEliminated = true;
}
}
// Fix for the case when a float to non-float conversion node swings
// down past a branch on IA32; this would cause a FP value to be commoned
// across a branch where there was none originally; this causes pblms
// as a value is left on the stack.
//
if (treeTopCanBeEliminated &&
seenConditionalBranch)
{
if (!cg()->getSupportsJavaFloatSemantics())
{
if (child->getOpCode().isConversion() ||
child->getOpCode().isBooleanCompare())
{
if (child->getFirstChild()->getOpCode().isFloatingPoint() &&
!child->getOpCode().isFloatingPoint())
treeTopCanBeEliminated = false;
}
}
}
if (treeTopCanBeEliminated)
{
TR::NodeChecklist visited(comp());
bool containsFloatingPoint = false;
for (int32_t i = 0; i < child->getNumChildren(); ++i)
{
// Anchor nodes with reference count > 1
//
bool highGlobalIndex = false;
if (fixUpTree(child->getChild(i), iter.currentTree(), visited, highGlobalIndex, self(), visitCount))
containsFloatingPoint = true;
if (highGlobalIndex)
{
dumpOptDetails(comp(),
"%sGlobal index limit exceeded; stopping\n",
optDetailString());
return 0;
}
}
if (seenConditionalBranch &&
containsFloatingPoint)
{
if (!cg()->getSupportsJavaFloatSemantics())
treeTopCanBeEliminated = false;
}
}
}
// Update visitCount as they are used in this optimization and need to be
// correct at all intermediate stages
//
if (!treeTopCanBeEliminated)
TR::TransformUtil::recursivelySetNodeVisitCount(node, visitCount);
if (treeTopCanBeEliminated)
{
TR::TreeTop *prevTree = iter.currentTree()->getPrevTreeTop();
TR::TreeTop *nextTree = iter.currentTree()->getNextTreeTop();
if (!node->getOpCode().isStoreReg() || (node->getFirstChild()->getReferenceCount() == 1))
{
// Actually going to remove the treetop now
//
if (performTransformation(comp(), "%sRemove tree : [" POINTER_PRINTF_FORMAT "] ([" POINTER_PRINTF_FORMAT "] = %s)\n", optDetailString(), node, node->getFirstChild(), node->getFirstChild()->getOpCode().getName()))
{
prevTree->join(nextTree);
optimizer()->prepareForNodeRemoval(node);
///child->recursivelyDecReferenceCount();
node->recursivelyDecReferenceCount();
recursivelyDecFutureUseCount(child);
iter.jumpTo(prevTree);
if (child->getReferenceCount() == 1)
requestOpt(OMR::treeSimplification, true, block);
if (nextTree->getNode()->getOpCodeValue() == TR::Goto
&& prevTree->getNode()->getOpCodeValue() == TR::BBStart
&& !prevTree->getNode()->getBlock()->isExtensionOfPreviousBlock())
{
requestOpt(
OMR::redundantGotoElimination,
prevTree->getNode()->getBlock());
}
}
}
else
{
if (performTransformation(comp(), "%sMove tree : [" POINTER_PRINTF_FORMAT "]([" POINTER_PRINTF_FORMAT "] = %s) to end of block\n", optDetailString(), node, node->getFirstChild(), node->getFirstChild()->getOpCode().getName()))
{
prevTree->join(nextTree);
node->setVisitCount(visitCount);
TR::TreeTop *lastTree = findLastTreetop(block, prevTree);
TR::TreeTop *prevLastTree = lastTree->getPrevTreeTop();
TR::TreeTop *cursorTreeTop = nextTree;
while (cursorTreeTop != lastTree)
{
if (cursorTreeTop->getNode()->getOpCode().isStoreReg() &&
(cursorTreeTop->getNode()->getGlobalRegisterNumber() == iter.currentTree()->getNode()->getGlobalRegisterNumber()))
{
lastTree = cursorTreeTop;
prevLastTree = lastTree->getPrevTreeTop();
break;
}
cursorTreeTop = cursorTreeTop->getNextTreeTop();
}
if (lastTree->getNode()->getOpCodeValue() == TR::BBStart)
{
prevLastTree = lastTree;
lastTree = block->getExit();
}
TR::Node *lastNode = lastTree->getNode();
TR::Node *prevLastNode = prevLastTree->getNode();
if (lastNode->getOpCode().isIf() && !lastNode->getOpCode().isCompBranchOnly() &&
prevLastNode->getOpCode().isStoreReg() &&
((prevLastNode->getFirstChild() == lastNode->getFirstChild()) ||
(prevLastNode->getFirstChild() == lastNode->getSecondChild())))
{
lastTree = prevLastTree;
prevLastTree = lastTree->getPrevTreeTop();
}
prevLastTree->join(iter.currentTree());
iter.currentTree()->join(lastTree);
iter.jumpTo(prevTree);
requestOpt(OMR::treeSimplification, true, block);
}
}
}
}
for (auto it = anchors.begin(); it != anchors.end(); ++it)
{
TR::Node *anchor = it->tree->getNode();
TR::Node *load = anchor->getChild(0);
if (load->getReferenceCount() > 1)
continue;
// We can eliminate the indirect load immediately, but for the moment the
// subtree providing the base object has to be anchored.
TR::Node *heapBase = anchor->getChild(1);
TR::Node::recreate(anchor, TR::treetop);
anchor->setAndIncChild(0, load->getChild(0));
anchor->setChild(1, NULL);
anchor->setNumChildren(1);
if (!heapBase->getOpCode().isLoadConst())
{
it->tree->insertAfter(
TR::TreeTop::create(
comp(),
TR::Node::create(heapBase, TR::treetop, 1, heapBase)));
}
load->recursivelyDecReferenceCount();
heapBase->recursivelyDecReferenceCount();
// A later pass of dead trees can likely move (or even remove) the base
// object expression.
requestOpt(OMR::deadTreesElimination, true, it->block);
}
return 1; // actual cost
}