int32_t
TR::RegDepCopyRemoval::perform()
{
if (!cg()->supportsPassThroughCopyToNewVirtualRegister())
return 0;
discardAllNodeChoices();
TR::TreeTop *tt;
for (tt = comp()->getStartTree(); tt != NULL; tt = tt->getNextTreeTop())
{
TR::Node *node = tt->getNode();
switch (node->getOpCodeValue())
{
case TR::BBStart:
if (!node->getBlock()->isExtensionOfPreviousBlock())
{
if (trace())
traceMsg(comp(), "clearing remembered node choices at start of extended block at block_%d\n", node->getBlock()->getNumber());
discardAllNodeChoices();
}
if (node->getNumChildren() > 0)
processRegDeps(node->getFirstChild(), tt);
break;
case TR::BBEnd:
if (node->getNumChildren() > 0)
processRegDeps(node->getFirstChild(), tt);
break;
default:
if (node->getOpCode().isSwitch())
{
TR::Node *defaultDest = node->getSecondChild();
if (defaultDest->getNumChildren() > 0)
processRegDeps(defaultDest->getFirstChild(), tt);
}
else if (node->getOpCode().isBranch())
{
int nChildren = node->getNumChildren();
// only the last child may be GlRegDeps
for (int i = 0; i < nChildren - 1; i++)
TR_ASSERT(node->getChild(i)->getOpCodeValue() != TR::GlRegDeps, "GlRegDeps for branch is not the last child\n");
if (nChildren > 0)
{
TR::Node *lastChild = node->getChild(nChildren - 1);
if (lastChild->getOpCodeValue() == TR::GlRegDeps)
processRegDeps(lastChild, tt);
}
}
break;
}
}
return 1; // a bit arbitrary...
}
TR::TreeTop *searchForStringAppend(TR::ValuePropagation *vp,const char *sig, TR::TreeTop *tt, TR::TreeTop *exitTree,
TR::ILOpCodes opCode, TR::Node *newBuffer, vcount_t visitCount, TR::Node **string)
{
int len = 0;
bool isGlobal = false;
for (;tt != exitTree; tt = tt->getNextRealTreeTop())
{
TR::Node *node = tt->getNode();
if (node->getNumChildren() == 1 &&
node->getFirstChild()->getOpCodeValue() == opCode)
{
if (checkMethodSignature(vp,node->getFirstChild()->getSymbolReference(), sig))
{
TR::Node *call = node->getFirstChild();
if (call->getFirstChild() == newBuffer)
{
if (vp->getConstraint(call->getSecondChild(),isGlobal))
{
if (vp->getConstraint(call->getSecondChild(),isGlobal)->getClassType())
{
const char *sigAppend = vp->getConstraint(call->getSecondChild(),isGlobal)->getClassType()->getClassSignature(len);
if (call->getSecondChild()->getOpCodeValue() == TR::aload && (!strncmp(sigAppend,"Ljava/lang/String;",18)))
{
*string = call->getSecondChild();
}
}
}
}
return tt;
}
else
return tt;
}
//if (countNodeOccurrencesInSubTree(node, newBuffer, visitCount) > 0) return tt;
}
return tt;
}
TR::Node *
OMR::TransformUtil::scalarizeArrayCopy(
TR::Compilation *comp,
TR::Node *node,
TR::TreeTop *tt,
bool useElementType,
bool &didTransformArrayCopyNode,
TR::SymbolReference *sourceRef,
TR::SymbolReference *targetRef,
bool castToIntegral)
{
TR::CodeGenerator *cg = comp->cg();
didTransformArrayCopyNode = false;
if ((comp->getOptLevel() == noOpt) ||
!comp->getOption(TR_ScalarizeSSOps) ||
node->getOpCodeValue() != TR::arraycopy ||
node->getNumChildren() != 3 ||
comp->requiresSpineChecks() ||
!node->getChild(2)->getOpCode().isLoadConst() ||
cg->getOptimizationPhaseIsComplete())
return node;
int64_t byteLen = node->getChild(2)->get64bitIntegralValue();
if (byteLen == 0)
{
if (tt)
{
// Anchor the first two children
if (!node->getFirstChild()->safeToDoRecursiveDecrement())
TR::TreeTop::create(comp, tt->getPrevTreeTop(),
TR::Node::create(TR::treetop, 1, node->getFirstChild()));
if (!node->getSecondChild()->safeToDoRecursiveDecrement())
TR::TreeTop::create(comp, tt->getPrevTreeTop(),
TR::Node::create(TR::treetop, 1, node->getSecondChild()));
tt->getPrevTreeTop()->join(tt->getNextTreeTop());
tt->getNode()->recursivelyDecReferenceCount();
didTransformArrayCopyNode = true;
}
return node;
}
else if (byteLen < 0)
{
return node;
}
else if (byteLen > TR_MAX_OTYPE_SIZE)
{
return node;
}
TR::DataType dataType = TR::Aggregate;
// Get the element datatype from the (hidden) 4th child
TR::DataType elementType = node->getArrayCopyElementType();
int32_t elementSize = TR::Symbol::convertTypeToSize(elementType);
if (byteLen == elementSize)
{
dataType = elementType;
}
else if (!useElementType)
{
switch (byteLen)
{
case 1: dataType = TR::Int8; break;
case 2: dataType = TR::Int16; break;
case 4: dataType = TR::Int32; break;
case 8: dataType = TR::Int64; break;
}
}
else
{
return node;
}
// load/store double on 64-bit PPC requires offset to be word aligned
// abort if this requirement is not met.
// TODO: also need to check if the first two children are aload nodes
bool cannot_use_load_store_long = false;
if (TR::Compiler->target.cpu.isPower())
if (dataType == TR::Int64 && TR::Compiler->target.is64Bit())
{
TR::Node * firstChild = node->getFirstChild();
if (firstChild->getNumChildren() == 2)
{
TR::Node *offsetChild = firstChild->getSecondChild();
TR_ASSERT(offsetChild->getOpCodeValue() != TR::iconst, "iconst shouldn't be used for 64-bit array indexing");
if (offsetChild->getOpCodeValue() == TR::lconst)
{
if ((offsetChild->getLongInt() & 0x3) != 0)
cannot_use_load_store_long = true;
}
}
TR::Node *secondChild = node->getSecondChild();
if (secondChild->getNumChildren() == 2)
{
TR::Node *offsetChild = secondChild->getSecondChild();
TR_ASSERT(offsetChild->getOpCodeValue() != TR::iconst, "iconst shouldn't be used for 64-bit array indexing");
if (offsetChild->getOpCodeValue() == TR::lconst)
{
if ((offsetChild->getLongInt() & 0x3) != 0)
cannot_use_load_store_long = true;
}
}
}
if (cannot_use_load_store_long) return node;
TR::SymbolReference *nodeRef;
targetRef = comp->getSymRefTab()->findOrCreateGenericIntShadowSymbolReference(0);
sourceRef = targetRef;
bool trace = comp->getOption(TR_TraceScalarizeSSOps);
if (trace)
traceMsg(comp,"scalarizeArrayCopy: node %p got targetRef (#%d) and sourceRef (#%d)\n",
node,targetRef?targetRef->getReferenceNumber():-1,sourceRef?sourceRef->getReferenceNumber():-1);
if (targetRef == NULL || sourceRef == NULL)
{
if (trace)
traceMsg(comp,"do not scalarizeArrayCopy node %p : targetRef is NULL (%s) or sourceRef is NULL (%s)\n",node,targetRef?"no":"yes",sourceRef?"no":"yes");
return node;
}
#ifdef J9_PROJECT_SPECIFIC
if (targetRef->getSymbol()->getDataType().isBCD() ||
sourceRef->getSymbol()->getDataType().isBCD())
{
return node;
}
#endif
if (performTransformation(comp, "%sScalarize arraycopy 0x%p\n", OPT_DETAILS, node))
{
TR::Node *store = TR::TransformUtil::scalarizeAddressParameter(comp, node->getSecondChild(), byteLen, dataType, targetRef, true);
TR::Node *load = TR::TransformUtil::scalarizeAddressParameter(comp, node->getFirstChild(), byteLen, dataType, sourceRef, false);
if (tt)
{
// Transforming
// treetop
// arrayCopy <-- node
// into
// *store
//
node->recursivelyDecReferenceCount();
tt->setNode(node);
}
else
{
for (int16_t c = node->getNumChildren() - 1; c >= 0; c--)
cg->recursivelyDecReferenceCount(node->getChild(c));
}
TR::Node::recreate(node, store->getOpCodeValue());
node->setSymbolReference(store->getSymbolReference());
if (store->getOpCode().isStoreIndirect())
{
node->setChild(0, store->getFirstChild());
node->setAndIncChild(1, load);
node->setNumChildren(2);
}
else
{
node->setAndIncChild(0, load);
node->setNumChildren(1);
}
didTransformArrayCopyNode = true;
}
return node;
}
bool
OMR::Simplifier::isBoundDefinitelyGELength(TR::Node *boundChild, TR::Node *lengthChild)
{
TR::ILOpCodes boundOp = boundChild->getOpCodeValue();
if (boundOp == TR::iadd)
{
TR::Node *first = boundChild->getFirstChild();
TR::Node *second = boundChild->getSecondChild();
if (first == lengthChild)
{
TR::ILOpCodes secondOp = second->getOpCodeValue();
if (second->getOpCode().isArrayLength() ||
secondOp == TR::bu2i ||
secondOp == TR::su2i ||
(secondOp == TR::iconst &&
second->getInt() >= 0) ||
(secondOp == TR::iand &&
second->getSecondChild()->getOpCodeValue() == TR::iconst &&
(second->getSecondChild()->getInt() & 80000000) == 0) ||
(secondOp == TR::iushr &&
second->getSecondChild()->getOpCodeValue() == TR::iconst &&
(second->getSecondChild()->getInt() & 0x1f) > 0))
{
return true;
}
}
else if (second == lengthChild)
{
TR::ILOpCodes firstOp = first->getOpCodeValue();
if (first->getOpCode().isArrayLength() ||
firstOp == TR::bu2i ||
firstOp == TR::su2i ||
(firstOp == TR::iand &&
first->getSecondChild()->getOpCodeValue() == TR::iconst &&
(first->getSecondChild()->getInt() & 80000000) == 0) ||
(firstOp == TR::iushr &&
first->getSecondChild()->getOpCodeValue() == TR::iconst &&
(first->getSecondChild()->getInt() & 0x1f) > 0))
{
return true;
}
}
}
else if (boundOp == TR::isub)
{
TR::Node *first = boundChild->getFirstChild();
TR::Node *second = boundChild->getSecondChild();
if (first == lengthChild)
{
TR::ILOpCodes secondOp = second->getOpCodeValue();
if ((secondOp == TR::iconst &&
second->getInt() < 0) ||
(secondOp == TR::ior &&
second->getSecondChild()->getOpCodeValue() == TR::iconst &&
(second->getSecondChild()->getInt() & 0x80000000) != 0))
{
return true;
}
}
}
return false;
}
TR::Node *
OMR::Simplifier::unaryCancelOutWithChild(TR::Node * node, TR::Node * firstChild, TR::TreeTop *anchorTree, TR::ILOpCodes opcode, bool anchorChildren)
{
if (!isLegalToUnaryCancel(node, firstChild, opcode))
return NULL;
if (firstChild->getOpCodeValue() == opcode &&
(node->getType().isAggregate() || firstChild->getType().isAggregate()) &&
(node->getSize() > firstChild->getSize() || node->getSize() != firstChild->getFirstChild()->getSize()))
{
// ensure a truncation side-effect of a conversion is not lost
// o2a size=3
// a2o size=3 // conversion truncates in addition to type cast so cannot be removed
// loadaddr size=4
// This restriction could be loosened to only disallow intermediate truncations (see BCD case above) but then would require a node
// op that would just correct for size (e.g. addrSizeMod size=3 to replace the o2a/a2o pair)
//
// Do allow cases when all three sizes are the same and when the middle node widens but the top and bottom node have the same size, e.g.
//
// i2o size=3
// o2i size=4
// oload size=3
//
// Also allow the special case where the grandchild is not really truncated as the 'truncated' bytes are known to be zero
// (i.e. there really isn't an intermediate truncation of 4->3 even though it appears that way from looking at the sizes alone)
// o2i
// i2o size=3
// iushr
// x
// iconst 8
bool disallow = true;
TR::Node *grandChild = firstChild->getFirstChild();
size_t nodeSize = node->getSize();
if (node->getType().isIntegral() &&
nodeSize == grandChild->getSize() &&
nodeSize > firstChild->getSize())
{
size_t truncatedBits = (nodeSize - firstChild->getSize()) * 8;
if (grandChild->getOpCode().isRightShift() && grandChild->getOpCode().isShiftLogical() &&
grandChild->getSecondChild()->getOpCode().isLoadConst() &&
(grandChild->getSecondChild()->get64bitIntegralValue() == truncatedBits))
{
disallow = false;
if (trace())
traceMsg(comp(),"do allow unaryCancel of node %s (%p) and firstChild %s (%p) as grandChild %s (%p) zeros the %d truncated bytes\n",
node->getOpCode().getName(),node,firstChild->getOpCode().getName(),firstChild,
grandChild->getOpCode().getName(),grandChild,truncatedBits/8);
}
}
if (disallow)
{
if (trace())
traceMsg(comp(),"disallow unaryCancel of node %s (%p) and firstChild %s (%p) due to unequal sizes (nodeSize %d, firstChildSize %d, firstChild->childSize %d)\n",
node->getOpCode().getName(),node,firstChild->getOpCode().getName(),firstChild,
node->getSize(),firstChild->getSize(),firstChild->getFirstChild()->getSize());
return NULL;
}
}
if (firstChild->getOpCodeValue() == opcode &&
performTransformation(comp(), "%sRemoving node [" POINTER_PRINTF_FORMAT "] %s and its child [" POINTER_PRINTF_FORMAT "] %s\n",
optDetailString(), node, node->getOpCode().getName(), firstChild, firstChild->getOpCode().getName()))
{
TR::Node *grandChild = firstChild->getFirstChild();
grandChild->incReferenceCount();
bool anchorChildrenNeeded = anchorChildren &&
(node->getNumChildren() > 1 ||
firstChild->getNumChildren() > 1 ||
node->getOpCode().hasSymbolReference() ||
firstChild->getOpCode().hasSymbolReference());
prepareToStopUsingNode(node, anchorTree, anchorChildrenNeeded);
node->recursivelyDecReferenceCount();
node->setVisitCount(0);
return grandChild;
}
return NULL;
}
TR_ExpressionsSimplification::LoopInfo*
TR_ExpressionsSimplification::findLoopInfo(TR_RegionStructure* region)
{
ListIterator<TR::CFGEdge> exitEdges(®ion->getExitEdges());
if (region->getExitEdges().getSize() != 1)
{
if (trace())
traceMsg(comp(), "Region with more than 1 exit edges can't be handled\n");
return 0;
}
TR_StructureSubGraphNode* exitNode = toStructureSubGraphNode(exitEdges.getFirst()->getFrom());
if (!exitNode->getStructure()->asBlock())
{
if (trace())
traceMsg(comp(), "The exit block can't be found\n");
return 0;
}
TR::Block *exitBlock = exitNode->getStructure()->asBlock()->getBlock();
TR::Node *lastTreeInExitBlock = exitBlock->getLastRealTreeTop()->getNode();
if (trace())
{
traceMsg(comp(), "The exit block is %d\n", exitBlock->getNumber());
traceMsg(comp(), "The branch node is %p\n", lastTreeInExitBlock);
}
if (!lastTreeInExitBlock->getOpCode().isBranch())
{
if (trace())
traceMsg(comp(), "The branch node couldn't be found\n");
return 0;
}
if (lastTreeInExitBlock->getNumChildren() < 2)
{
if (trace())
traceMsg(comp(), "The branch node has less than 2 children\n");
return 0;
}
TR::Node *firstChildOfLastTree = lastTreeInExitBlock->getFirstChild();
TR::Node *secondChildOfLastTree = lastTreeInExitBlock->getSecondChild();
if (!firstChildOfLastTree->getOpCode().hasSymbolReference())
{
if (trace())
traceMsg(comp(), "The branch node's first child node %p - its opcode does not have a symbol reference\n", firstChildOfLastTree);
return 0;
}
TR::SymbolReference *firstChildSymRef = firstChildOfLastTree->getSymbolReference();
if (trace())
traceMsg(comp(), "Symbol Reference: %p Symbol: %p\n", firstChildSymRef, firstChildSymRef->getSymbol());
// Locate the induction variable that matches with the exit node symbol
//
TR_InductionVariable *indVar = region->findMatchingIV(firstChildSymRef);
if (!indVar) return 0;
if (!indVar->getIncr()->asIntConst())
{
if (trace())
traceMsg(comp(), "Increment is not a constant\n");
return 0;
}
int32_t increment = indVar->getIncr()->getLowInt();
_visitCount = comp()->incVisitCount();
bool indVarWrittenAndUsedUnexpectedly = false;
if (firstChildOfLastTree->getReferenceCount() > 1)
{
TR::TreeTop *cursorTreeTopInExitBlock = exitBlock->getEntry();
TR::TreeTop *exitTreeTopInExitBlock = exitBlock->getExit();
bool loadSeen = false;
while (cursorTreeTopInExitBlock != exitTreeTopInExitBlock)
{
TR::Node *cursorNode = cursorTreeTopInExitBlock->getNode();
if (checkForLoad(cursorNode, firstChildOfLastTree))
loadSeen = true;
if (!cursorNode->getOpCode().isStore() &&
(cursorNode->getNumChildren() > 0))
cursorNode = cursorNode->getFirstChild();
if (cursorNode->getOpCode().isStore() &&
(cursorNode->getSymbolReference() == firstChildSymRef))
{
indVarWrittenAndUsedUnexpectedly = true;
if ((cursorNode->getFirstChild() == firstChildOfLastTree) ||
!loadSeen)
indVarWrittenAndUsedUnexpectedly = false;
else
break;
}
cursorTreeTopInExitBlock = cursorTreeTopInExitBlock->getNextTreeTop();
}
}
if (indVarWrittenAndUsedUnexpectedly)
{
return 0;
}
int32_t lowerBound;
int32_t upperBound = 0;
TR::Node *bound = 0;
bool equals = false;
switch(lastTreeInExitBlock->getOpCodeValue())
{
case TR::ificmplt:
case TR::ificmpgt:
equals = true;
case TR::ificmple:
case TR::ificmpge:
if (!(indVar->getEntry() && indVar->getEntry()->asIntConst()))
{
if (trace())
traceMsg(comp(), "Entry value is not a constant\n");
return 0;
}
lowerBound = indVar->getEntry()->getLowInt();
if (secondChildOfLastTree->getOpCode().isLoadConst())
{
upperBound = secondChildOfLastTree->getInt();
}
else if (secondChildOfLastTree->getOpCode().isLoadVar())
{
bound = secondChildOfLastTree;
}
else
{
if (trace())
traceMsg(comp(), "Second child is not a const or a load\n");
return 0;
}
return new (trStackMemory()) LoopInfo(bound, lowerBound, upperBound, increment, equals);
default:
if (trace())
traceMsg(comp(), "The condition has not been implemeted\n");
return 0;
}
return 0;
}
bool TR_ExpressionsSimplification::tranformSummationReductionCandidate(TR::TreeTop *treeTop, LoopInfo *loopInfo, bool *isPreheaderBlockInvalid)
{
TR::Node *node = treeTop->getNode();
TR::Node *opNode = node->getFirstChild();
TR::Node *expNode = NULL;
int32_t expChildNumber = 0;
bool removeOnly = false;
bool replaceWithNewNode = false;
if (opNode->getOpCodeValue() == TR::iadd || opNode->getOpCodeValue() == TR::isub)
{
if (opNode->getSecondChild()->getOpCode().hasSymbolReference() &&
node->getSymbolReference() == opNode->getSecondChild()->getSymbolReference())
{
expChildNumber = 0;
expNode = opNode->getFirstChild();
}
else
{
expChildNumber = 1;
expNode = opNode->getSecondChild();
}
expNode = iaddisubSimplifier(expNode, loopInfo);
replaceWithNewNode = true;
}
else if (opNode->getOpCodeValue() == TR::ixor || opNode->getOpCodeValue() == TR::ineg)
{
expNode = ixorinegSimplifier(opNode, loopInfo, &removeOnly);
}
if (expNode)
{
if (trace())
comp()->getDebug()->print(comp()->getOutFile(), expNode, 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 true;
}
if (loopInfo->getNumIterations() > 0 || // make sure that the loop is going to be executed at least once
_currentRegion->isCanonicalizedLoop()) // or 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 node [%p] to block_%d\n", OPT_DETAILS, node, preheaderBlock->getNumber()))
{
if (!(removeOnly))
{
TR::Node *newNode = node->duplicateTree();
if (replaceWithNewNode)
newNode->getFirstChild()->setAndIncChild(expChildNumber, expNode);
transformNode(newNode, preheaderBlock);
}
TR::TransformUtil::removeTree(comp(), treeTop);
}
}
}
return (expNode != NULL);
}
void TR_ExpressionsSimplification::setSummationReductionCandidates(TR::Node *node, TR::TreeTop *tt)
{
// Must be a store node
//
if (node->getOpCodeValue() != TR::istore
/* || node->getOpCodeValue() != TR::astore */)
{
if (trace())
traceMsg(comp(), "Node %p: The opcode is not istore so not a summation reduction candidate\n",node);
return;
}
TR::Node *opNode = node->getFirstChild();
if (opNode->getOpCodeValue() == TR::iadd ||
opNode->getOpCodeValue() == TR::isub)
{
TR::Node *firstNode = opNode->getFirstChild();
TR::Node *secondNode = opNode->getSecondChild();
if (firstNode->getOpCode().hasSymbolReference() &&
node->getSymbolReference() == firstNode->getSymbolReference() &&
opNode->getReferenceCount() == 1 && firstNode->getReferenceCount() == 1)
{
// The second node must be loop invariant
//
if (!_currentRegion->isExprInvariant(secondNode))
{
if (trace())
{
traceMsg(comp(), "The node %p is not loop invariant\n",secondNode);
// This can be the arithmetic series case
// only when the node is an induction variable
if (secondNode->getNumChildren() == 1 && secondNode->getOpCode().hasSymbolReference())
{
TR_InductionVariable *indVar = _currentRegion->findMatchingIV(secondNode->getSymbolReference());
if (indVar)
{
//printf("Found Candidate of arithmetic series\n" );
}
}
}
return;
}
_candidateTTs->add(tt);
}
else if (secondNode->getOpCode().hasSymbolReference() &&
node->getSymbolReference() == secondNode->getSymbolReference() &&
opNode->getReferenceCount() == 1 && secondNode->getReferenceCount() == 1 &&
_currentRegion->isExprInvariant(firstNode))
{
_candidateTTs->add(tt);
}
}
else if (opNode->getOpCodeValue() == TR::ixor ||
opNode->getOpCodeValue() == TR::ineg)
{
if (opNode->getFirstChild()->getOpCode().hasSymbolReference() &&
node->getSymbolReference() == opNode->getFirstChild()->getSymbolReference() &&
opNode->getReferenceCount() == 1 && opNode->getFirstChild()->getReferenceCount() == 1 &&
(opNode->getOpCodeValue() == TR::ineg || _currentRegion->isExprInvariant(opNode->getSecondChild())))
_candidateTTs->add(tt);
else if (opNode->getOpCodeValue() == TR::ixor && opNode->getSecondChild()->getOpCode().hasSymbolReference() &&
node->getSymbolReference() == opNode->getSecondChild()->getSymbolReference() &&
opNode->getReferenceCount() == 1 && opNode->getSecondChild()->getReferenceCount() == 1 &&
_currentRegion->isExprInvariant(opNode->getFirstChild()))
_candidateTTs->add(tt);
}
}
TR_LocalAnalysisInfo::TR_LocalAnalysisInfo(TR::Compilation *c, bool t)
: _compilation(c), _trace(t), _trMemory(c->trMemory())
{
_numNodes = -1;
#if 0 // somehow stops PRE from happening
// We are going to increment visit count for every tree so can reach max
// for big methods quickly. Perhaps can improve containsCall() in the future.
comp()->resetVisitCounts(0);
#endif
if (comp()->getVisitCount() > HIGH_VISIT_COUNT)
{
_compilation->resetVisitCounts(1);
dumpOptDetails(comp(), "\nResetting visit counts for this method before LocalAnalysisInfo\n");
}
TR::CFG *cfg = comp()->getFlowGraph();
_numBlocks = cfg->getNextNodeNumber();
TR_ASSERT(_numBlocks > 0, "Local analysis, node numbers not assigned");
// Allocate information on the stack. It is the responsibility of the user
// of this class to determine the life of the information by using jitStackMark
// and jitStackRelease.
//
//_blocksInfo = (TR::Block **) trMemory()->allocateStackMemory(_numBlocks*sizeof(TR::Block *));
//memset(_blocksInfo, 0, _numBlocks*sizeof(TR::Block *));
TR::TreeTop *currentTree = comp()->getStartTree();
// Only do this if not done before; typically this would be done in the
// first call to this method through LocalTransparency and would NOT
// need to be re-done by LocalAnticipatability.
//
if (_numNodes < 0)
{
_optimizer = comp()->getOptimizer();
int32_t numBuckets;
int32_t numNodes = comp()->getNodeCount();
if (numNodes < 10)
numBuckets = 1;
else if (numNodes < 100)
numBuckets = 7;
else if (numNodes < 500)
numBuckets = 31;
else if (numNodes < 3000)
numBuckets = 127;
else if (numNodes < 6000)
numBuckets = 511;
else
numBuckets = 1023;
// Allocate hash table for matching expressions
//
HashTable hashTable(numBuckets, comp());
_hashTable = &hashTable;
// Null checks are handled differently as the criterion for
// commoning a null check is different than that used for
// other nodes; for a null check, the null check reference is
// important (and not the actual indirect access itself)
//
_numNullChecks = 0;
while (currentTree)
{
if (currentTree->getNode()->getOpCodeValue() == TR::NULLCHK)
//////if (currentTree->getNode()->getOpCode().isNullCheck())
_numNullChecks++;
currentTree = currentTree->getNextTreeTop();
}
if (_numNullChecks == 0)
_nullCheckNodesAsArray = NULL;
else
{
_nullCheckNodesAsArray = (TR::Node**)trMemory()->allocateStackMemory(_numNullChecks*sizeof(TR::Node*));
memset(_nullCheckNodesAsArray, 0, _numNullChecks*sizeof(TR::Node*));
}
currentTree = comp()->getStartTree();
int32_t symRefCount = comp()->getSymRefCount();
_checkSymbolReferences = new (trStackMemory()) TR_BitVector(symRefCount, trMemory(), stackAlloc);
_numNodes = 1;
_numNullChecks = 0;
// This loop counts all the nodes that are going to take part in PRE.
// This is a computation intensive loop as we check if the node that
// is syntactically equivalent to a given node has been seen before
// and if so we use the local index of the original node (that
// is syntactically equivalent to the given node). Could be improved
// in complexity with value numbering at some stage.
//
_visitCount = comp()->incVisitCount();
while (currentTree)
{
TR::Node *firstNodeInTree = currentTree->getNode();
TR::ILOpCode *opCode = &firstNodeInTree->getOpCode();
if (((firstNodeInTree->getOpCodeValue() == TR::treetop) ||
(comp()->useAnchors() && firstNodeInTree->getOpCode().isAnchor())) &&
(firstNodeInTree->getFirstChild()->getOpCode().isStore()))
{
firstNodeInTree->setLocalIndex(-1);
if (comp()->useAnchors() && firstNodeInTree->getOpCode().isAnchor())
firstNodeInTree->getSecondChild()->setLocalIndex(-1);
firstNodeInTree = firstNodeInTree->getFirstChild();
opCode = &firstNodeInTree->getOpCode();
}
// This call finds nodes with opcodes that are supported by PRE
// in this subtree; this accounts for all opcodes other than stores/checks
// which are handled later on below
//
bool firstNodeInTreeHasCallsInStoreLhs = false;
countSupportedNodes(firstNodeInTree, NULL, firstNodeInTreeHasCallsInStoreLhs);
if ((opCode->isStore() && !firstNodeInTree->getSymbolReference()->getSymbol()->isAutoOrParm()) ||
opCode->isCheck())
{
int32_t oldExpressionOnRhs = hasOldExpressionOnRhs(firstNodeInTree);
//
// Return value 0 denotes that the node contains some sub-expression
// that cannot participate in PRE; e.g. a call or a new
//
// Return value -1 denotes that the node can participate in PRE
// but did not match with any existing expression seen so far
//
// Any other return value (should be positive always) denotes that
// the node can participate in PRE and has been matched with a seen
// expression having local index == return value
//
if (oldExpressionOnRhs == -1)
{
if (trace())
{
traceMsg(comp(), "\nExpression #%d is : \n", _numNodes);
comp()->getDebug()->print(comp()->getOutFile(), firstNodeInTree, 6, true);
}
firstNodeInTree->setLocalIndex(_numNodes++);
}
else
firstNodeInTree->setLocalIndex(oldExpressionOnRhs);
if (opCode->isCheck() &&
(firstNodeInTree->getFirstChild()->getOpCode().isStore() &&
!firstNodeInTree->getFirstChild()->getSymbolReference()->getSymbol()->isAutoOrParm()))
{
int oldExpressionOnRhs = hasOldExpressionOnRhs(firstNodeInTree->getFirstChild());
if (oldExpressionOnRhs == -1)
{
if (trace())
{
traceMsg(comp(), "\nExpression #%d is : \n", _numNodes);
comp()->getDebug()->print(comp()->getOutFile(), firstNodeInTree->getFirstChild(), 6, true);
}
firstNodeInTree->getFirstChild()->setLocalIndex(_numNodes++);
}
else
firstNodeInTree->getFirstChild()->setLocalIndex(oldExpressionOnRhs);
}
}
else
firstNodeInTree->setLocalIndex(-1);
currentTree = currentTree->getNextTreeTop();
}
}
_supportedNodesAsArray = (TR::Node**)trMemory()->allocateStackMemory(_numNodes*sizeof(TR::Node*));
memset(_supportedNodesAsArray, 0, _numNodes*sizeof(TR::Node*));
_checkExpressions = new (trStackMemory()) TR_BitVector(_numNodes, trMemory(), stackAlloc);
//_checkExpressions.init(_numNodes, trMemory(), stackAlloc);
// This loop goes through the trees and collects the nodes
// that would take part in PRE. Each node has its local index set to
// the bit position that it occupies in the bit vector analyses.
//
currentTree = comp()->getStartTree();
_visitCount = comp()->incVisitCount();
while (currentTree)
{
TR::Node *firstNodeInTree = currentTree->getNode();
TR::ILOpCode *opCode = &firstNodeInTree->getOpCode();
if (((firstNodeInTree->getOpCodeValue() == TR::treetop) ||
(comp()->useAnchors() && firstNodeInTree->getOpCode().isAnchor())) &&
(firstNodeInTree->getFirstChild()->getOpCode().isStore()))
{
firstNodeInTree = firstNodeInTree->getFirstChild();
opCode = &firstNodeInTree->getOpCode();
}
collectSupportedNodes(firstNodeInTree, NULL);
if ((opCode->isStore() && !firstNodeInTree->getSymbolReference()->getSymbol()->isAutoOrParm()) ||
opCode->isCheck())
{
if (opCode->isCheck())
{
_checkSymbolReferences->set(firstNodeInTree->getSymbolReference()->getReferenceNumber());
_checkExpressions->set(firstNodeInTree->getLocalIndex());
}
if (!_supportedNodesAsArray[firstNodeInTree->getLocalIndex()])
_supportedNodesAsArray[firstNodeInTree->getLocalIndex()] = firstNodeInTree;
if (opCode->isCheck() &&
firstNodeInTree->getFirstChild()->getOpCode().isStore() &&
!firstNodeInTree->getFirstChild()->getSymbolReference()->getSymbol()->isAutoOrParm() &&
!_supportedNodesAsArray[firstNodeInTree->getFirstChild()->getLocalIndex()])
_supportedNodesAsArray[firstNodeInTree->getFirstChild()->getLocalIndex()] = firstNodeInTree->getFirstChild();
}
currentTree = currentTree->getNextTreeTop();
}
//initialize(toBlock(cfg->getStart()));
}
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
}
