TR::Node *TR_OutlinedInstructions::createOutlinedCallNode(TR::Node *callNode, TR::ILOpCodes callOp)
{
int32_t i;
TR::Node *child;
//We pass true for getSymbolReference because
TR::Node *newCallNode = TR::Node::createWithSymRef(callNode, callOp, callNode->getNumChildren(), callNode->getSymbolReference());
newCallNode->setReferenceCount(1);
for (i=0; i<callNode->getNumChildren(); i++)
{
child = callNode->getChild(i);
if (child->getRegister() != NULL)
{
// Child has already been evaluated outside this tree.
//
newCallNode->setAndIncChild(i, child);
}
else if (child->getOpCode().isLoadConst())
{
// Copy unevaluated constant nodes.
//
child = TR::Node::copy(child);
child->setReferenceCount(1);
newCallNode->setChild(i, child);
}
else
{
if ((child->getOpCodeValue() == TR::loadaddr) &&
/*(callNode->getOpCodeValue() == TR::instanceof || callNode->getOpCodeValue() == TR::checkcast || callNode->getOpCodeValue() == TR::checkcastAndNULLCHK || callNode->getOpCodeValue() == TR::New || callNode->getOpCodeValue() == TR::anewarray) &&*/
(child->getSymbolReference()->getSymbol()) &&
(child->getSymbolReference()->getSymbol()->getStaticSymbol()))
{
child = TR::Node::copy(child);
child->setReferenceCount(1);
newCallNode->setChild(i, child);
}
else
{
// Be very conservative at this point, even though it is possible to make it less so. For example, this will catch
// the case of an unevaluated argument not persisting outside of the outlined region even though one of its subtrees will.
//
(void)_cg->evaluate(child);
// Do not decrement the reference count here. It will be decremented when the call node is evaluated
// again in the helper instruction stream.
//
newCallNode->setAndIncChild(i, child);
}
}
}
if(callNode->isPreparedForDirectJNI())
{
newCallNode->setPreparedForDirectJNI();
}
return newCallNode;
}
void
TR::RegDepCopyRemoval::makeFreshCopy(TR_GlobalRegisterNumber reg)
{
RegDepInfo &dep = getRegDepInfo(reg);
if (!performTransformation(comp(),
"%schange %s in GlRegDeps n%un to an explicit copy of n%un\n",
optDetailString(),
registerName(reg),
_regDeps->getGlobalIndex(),
dep.value->getGlobalIndex()))
return;
// Split the block at fallthrough if necessary to avoid putting copies
// between branches and BBEnd.
TR::Node *curNode = _treetop->getNode();
if (curNode->getOpCodeValue() == TR::BBEnd)
{
TR::Block *curBlock = curNode->getBlock();
if (curBlock->getLastRealTreeTop() != curBlock->getLastNonControlFlowTreeTop())
{
TR::Block *fallthrough = curBlock->getNextBlock();
fallthrough = curBlock->splitEdge(curBlock, fallthrough, comp());
TR_ASSERT(curBlock->getNextBlock() == fallthrough, "bad block placement from splitEdge\n");
fallthrough->setIsExtensionOfPreviousBlock();
_treetop = fallthrough->getExit();
TR::Node *newNode = _treetop->getNode();
newNode->setChild(0, _regDeps);
newNode->setNumChildren(1);
curNode->setNumChildren(0);
if (trace())
traceMsg(comp(), "\tsplit fallthrough edge to insert copy, created block_%d\n", fallthrough->getNumber());
}
}
// Make and insert the copy
TR::Node *copyNode = NULL;
if (dep.value->getOpCode().isLoadConst())
{
// No need to depend on the other register.
// TODO heuristic for whether this is really better than a reg-reg move?
generateRegcopyDebugCounter("const-remat");
copyNode = TR::Node::create(dep.value->getOpCodeValue(), 0);
copyNode->setConstValue(dep.value->getConstValue());
}
else
{
generateRegcopyDebugCounter("fresh-copy");
copyNode = TR::Node::create(TR::PassThrough, 1, dep.value);
copyNode->setCopyToNewVirtualRegister();
}
TR::Node *copyTreetopNode = TR::Node::create(TR::treetop, 1, copyNode);
_treetop->insertBefore(TR::TreeTop::create(comp(), copyTreetopNode));
if (trace())
traceMsg(comp(), "\tcopy is n%un\n", copyNode->getGlobalIndex());
updateSingleRegDep(reg, copyNode);
}
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
}