void TR_LocalLiveRangeReduction::initPotentialDeps(TR_TreeRefInfo *tree)
{
int32_t symRefCount = comp()->getSymRefCount();
if (tree->getDefSym() == NULL)
tree->setDefSym(new (trStackMemory()) TR_BitVector(symRefCount, trMemory(), stackAlloc));
if (tree->getUseSym() == NULL)
tree->setUseSym(new (trStackMemory()) TR_BitVector(symRefCount, trMemory(), stackAlloc));
}
void TR_ReachingBlocks::initializeGenAndKillSetInfo()
{
// For each block in the CFG build the gen and kill set for this analysis.
// Go in treetop order, which guarantees that we see the correct (i.e. first)
// evaluation point for each node.
//
int32_t blockNum;
for (blockNum = 0; blockNum < _numberOfBlocks; blockNum++)
{
_regularGenSetInfo[blockNum] = new (trStackMemory()) TR_BitVector(getNumberOfBits(),trMemory(), stackAlloc);
_regularGenSetInfo[blockNum]->set(blockNum);
_exceptionGenSetInfo[blockNum] = new (trStackMemory()) TR_BitVector(getNumberOfBits(),trMemory(), stackAlloc);
_exceptionGenSetInfo[blockNum]->set(blockNum);
}
}
void TR_RegisterAnticipatability::initializeRegisterUsageInfo()
{
// initialize outSets bitvector as well
//
TR_BitVector **originalRegUsageInfo = _registerUsageInfo;
_registerUsageInfo = (TR_BitVector **) trMemory()->allocateStackMemory(_numberOfNodes * sizeof(TR_BitVector *));
_outSetInfo = (TR_BitVector **) trMemory()->allocateStackMemory(_numberOfNodes * sizeof(TR_BitVector *));
for (int32_t i = 0; i < _numberOfNodes; i++)
{
_registerUsageInfo[i] = new (trStackMemory()) TR_BitVector(_numberOfBits, trMemory(), stackAlloc);
copyFromInto(originalRegUsageInfo[i], _registerUsageInfo[i]);
_outSetInfo[i] = new (trStackMemory()) TR_BitVector(_numberOfBits, trMemory(), stackAlloc);
_outSetInfo[i]->empty();
}
}
bool TR_DominatorVerifier::isExpensiveAlgorithmCorrect(TR_DominatorsChk &expensiveAlgorithm)
{
int32_t i,j;
_nodesSeenOnEveryPath = new (trStackMemory()) TR_BitVector(_numBlocks,trMemory(), stackAlloc);
_nodesSeenOnCurrentPath = new (trStackMemory()) TR_BitVector(_numBlocks,trMemory(), stackAlloc);
_dominatorsChkInfo = expensiveAlgorithm.getDominatorsChkInfo();
for (i = 2; i < _numBlocks-1; i++)
{
TR_BitVector *bucket = _dominatorsChkInfo[i]._tmpbucket;
for (j = 0; j < _numBlocks-1; j++)
{
if (bucket->get(j)) // dominator according to algorithm
{
// Initializing these BitVectors before checking
// dominators for the next block.
// The last bit is not changed for either bit vector - is that what
// was intended?
//
_nodesSeenOnEveryPath->setAll(_numBlocks-1);
int32_t lastBit = _nodesSeenOnCurrentPath->get(_numBlocks-1);
_nodesSeenOnCurrentPath->empty();
if (lastBit)
_nodesSeenOnCurrentPath->set(_numBlocks-1);
if ( ! dominates(_dominatorsChkInfo[j+1]._block,_dominatorsChkInfo[i]._block) ) // dominator according to the CFG
{
if (debug("traceVER"))
{
dumpOptDetails(comp(), " Dominator info for expensive algorithm is incorrect \n");
dumpOptDetails(comp(), " Dominator of [%p] is [%p] as per the algorithm\n", _dominatorsChkInfo[i]._block, _dominatorsChkInfo[j+1]._block);
dumpOptDetails(comp(), " But [%p] is not an the dominator of [%p] as per the Control Flow Graph", _dominatorsChkInfo[j+1]._block, _dominatorsChkInfo[i]._block);
}
return false;
}
}
}
}
return true;
}
//---------------------------- collecting ref info at the beginning -----------------------------------------
void TR_LocalLiveRangeReduction::collectInfo(TR::TreeTop *entryTree,TR::TreeTop *exitTree)
{
TR::TreeTop *currentTree = entryTree;
TR_TreeRefInfo *treeRefInfo;
int32_t i = 0;
int32_t maxRefCount = 0;
vcount_t visitCount = comp()->getVisitCount();
while (!(currentTree == exitTree))
{
treeRefInfo = new (trStackMemory()) TR_TreeRefInfo(currentTree, trMemory());
collectRefInfo(treeRefInfo, currentTree->getNode(),visitCount,&maxRefCount);
_treesRefInfoArray[i++] = treeRefInfo;
initPotentialDeps(treeRefInfo);
treeRefInfo->resetSyms();
populatePotentialDeps(treeRefInfo,treeRefInfo->getTreeTop()->getNode());
currentTree = currentTree->getNextTreeTop();
}
comp()->setVisitCount(visitCount+maxRefCount);
}
//add pair into dependant list
void TR_LocalLiveRangeReduction::addDepPair(TR_TreeRefInfo *dep,TR_TreeRefInfo *anchor)
{
DepPair *depPair = new (trStackMemory()) DepPair(dep,anchor);
_depPairList.add(depPair);
}
bool TR_LocalLiveRangeReduction::moveTreeBefore(TR_TreeRefInfo *treeToMove,TR_TreeRefInfo *anchor,int32_t passNumber)
{
TR::TreeTop *treeToMoveTT = treeToMove->getTreeTop();
TR::TreeTop *anchorTT = anchor->getTreeTop();
if (treeToMoveTT->getNextRealTreeTop() == anchorTT)
{
addDepPair(treeToMove, anchor);
return false;
}
if (!performTransformation(comp(), "%sPass %d: moving tree [%p] before Tree %p\n", OPT_DETAILS, passNumber, treeToMoveTT->getNode(),anchorTT->getNode()))
return false;
// printf("Moving [%p] before Tree %p\n", treeToMoveTT->getNode(),anchorTT->getNode());
//changing location in block
TR::TreeTop *origPrevTree = treeToMoveTT->getPrevTreeTop();
TR::TreeTop *origNextTree = treeToMoveTT->getNextTreeTop();
origPrevTree->setNextTreeTop(origNextTree);
origNextTree->setPrevTreeTop(origPrevTree);
TR::TreeTop *prevTree = anchorTT->getPrevTreeTop();
anchorTT->setPrevTreeTop(treeToMoveTT);
treeToMoveTT->setNextTreeTop(anchorTT);
treeToMoveTT->setPrevTreeTop(prevTree);
prevTree->setNextTreeTop(treeToMoveTT);
//UPDATE REFINFO
//find locations of treeTops in TreeTopsRefInfo array
//startIndex points to the currentTree that has moved
//endIndex points to the treeTop after which we moved the tree (nextTree)
int32_t startIndex = getIndexInArray(treeToMove);
int32_t endIndex = getIndexInArray(anchor)-1;
int32_t i=0;
for ( i = startIndex+1; i<= endIndex ; i++)
{
TR_TreeRefInfo *currentTreeRefInfo = _treesRefInfoArray[i];
List<TR::Node> *firstList = currentTreeRefInfo->getFirstRefNodesList();
List<TR::Node> *midList = currentTreeRefInfo->getMidRefNodesList();
List<TR::Node> *lastList = currentTreeRefInfo->getLastRefNodesList();
List<TR::Node> *M_firstList = treeToMove->getFirstRefNodesList();
List<TR::Node> *M_midList = treeToMove->getMidRefNodesList();
List<TR::Node> *M_lastList = treeToMove->getLastRefNodesList();
if (trace())
{
traceMsg(comp(),"Before move:\n");
printRefInfo(treeToMove);
printRefInfo(currentTreeRefInfo);
}
updateRefInfo(treeToMove->getTreeTop()->getNode(), currentTreeRefInfo, treeToMove , false);
treeToMove->resetSyms();
currentTreeRefInfo->resetSyms();
populatePotentialDeps(currentTreeRefInfo,currentTreeRefInfo->getTreeTop()->getNode());
populatePotentialDeps(treeToMove,treeToMove->getTreeTop()->getNode());
if (trace())
{
traceMsg(comp(),"After move:\n");
printRefInfo(treeToMove);
printRefInfo(currentTreeRefInfo);
traceMsg(comp(),"------------------------\n");
}
}
TR_TreeRefInfo *temp = _treesRefInfoArray[startIndex];
for (i = startIndex; i< endIndex ; i++)
{
_treesRefInfoArray[i] = _treesRefInfoArray[i+1];
}
_treesRefInfoArray[endIndex]=temp;
#if defined(DEBUG) || defined(PROD_WITH_ASSUMES)
if (!(comp()->getOption(TR_EnableParanoidOptCheck) || debug("paranoidOptCheck")))
return true;
//verifier
{
TR::StackMemoryRegion stackMemoryRegion(*trMemory());
vcount_t visitCount = comp()->getVisitCount();
int32_t maxRefCount = 0;
TR::TreeTop *tt;
TR_TreeRefInfo **treesRefInfoArrayTemp = (TR_TreeRefInfo**)trMemory()->allocateStackMemory(_numTreeTops*sizeof(TR_TreeRefInfo*));
memset(treesRefInfoArrayTemp, 0, _numTreeTops*sizeof(TR_TreeRefInfo*));
TR_TreeRefInfo *treeRefInfoTemp;
//collect info
for ( int32_t i = 0; i<_numTreeTops-1; i++)
{
tt =_treesRefInfoArray[i]->getTreeTop();
treeRefInfoTemp = new (trStackMemory()) TR_TreeRefInfo(tt, trMemory());
collectRefInfo(treeRefInfoTemp, tt->getNode(),visitCount,&maxRefCount);
treesRefInfoArrayTemp[i] = treeRefInfoTemp;
}
comp()->setVisitCount(visitCount+maxRefCount);
for ( int32_t i = 0; i<_numTreeTops-1; i++)
{
if (!verifyRefInfo(treesRefInfoArrayTemp[i]->getFirstRefNodesList(),_treesRefInfoArray[i]->getFirstRefNodesList()))
{
printOnVerifyError(_treesRefInfoArray[i],treesRefInfoArrayTemp[i]);
TR_ASSERT(0,"fail to verify firstRefNodesList for %p\n",_treesRefInfoArray[i]->getTreeTop()->getNode());
}
if (!verifyRefInfo(treesRefInfoArrayTemp[i]->getMidRefNodesList(),_treesRefInfoArray[i]->getMidRefNodesList()))
{
printOnVerifyError(_treesRefInfoArray[i],treesRefInfoArrayTemp[i]);
TR_ASSERT(0,"fail to verify midRefNodesList for %p\n",_treesRefInfoArray[i]->getTreeTop()->getNode());
}
if (!verifyRefInfo(treesRefInfoArrayTemp[i]->getLastRefNodesList(),_treesRefInfoArray[i]->getLastRefNodesList()))
{
printOnVerifyError(_treesRefInfoArray[i],treesRefInfoArrayTemp[i]);
TR_ASSERT(0,"fail to verify lastRefNodesList for %p\n",_treesRefInfoArray[i]->getTreeTop()->getNode());
}
}
} // scope of the stack memory region
#endif
return true;
}
void TR_LocalLiveRangeReduction::prePerformOnBlocks()
{
comp()->incVisitCount();
int32_t symRefCount = comp()->getSymRefCount();
_temp = new (trStackMemory()) TR_BitVector(symRefCount, trMemory(), stackAlloc);
}
int32_t TR::ARM64SystemLinkage::buildArgs(TR::Node *callNode,
TR::RegisterDependencyConditions *dependencies)
{
const TR::ARM64LinkageProperties &properties = getProperties();
TR::ARM64MemoryArgument *pushToMemory = NULL;
TR::Register *argMemReg;
TR::Register *tempReg;
int32_t argIndex = 0;
int32_t numMemArgs = 0;
int32_t argSize = 0;
int32_t numIntegerArgs = 0;
int32_t numFloatArgs = 0;
int32_t totalSize;
int32_t i;
TR::Node *child;
TR::DataType childType;
TR::DataType resType = callNode->getType();
uint32_t firstArgumentChild = callNode->getFirstArgumentIndex();
/* Step 1 - figure out how many arguments are going to be spilled to memory i.e. not in registers */
for (i = firstArgumentChild; i < callNode->getNumChildren(); i++)
{
child = callNode->getChild(i);
childType = child->getDataType();
switch (childType)
{
case TR::Int8:
case TR::Int16:
case TR::Int32:
case TR::Int64:
case TR::Address:
if (numIntegerArgs >= properties.getNumIntArgRegs())
numMemArgs++;
numIntegerArgs++;
break;
case TR::Float:
case TR::Double:
if (numFloatArgs >= properties.getNumFloatArgRegs())
numMemArgs++;
numFloatArgs++;
break;
default:
TR_ASSERT(false, "Argument type %s is not supported\n", childType.toString());
}
}
// From here, down, any new stack allocations will expire / die when the function returns
TR::StackMemoryRegion stackMemoryRegion(*trMemory());
/* End result of Step 1 - determined number of memory arguments! */
if (numMemArgs > 0)
{
pushToMemory = new (trStackMemory()) TR::ARM64MemoryArgument[numMemArgs];
argMemReg = cg()->allocateRegister();
}
totalSize = numMemArgs * 8;
// align to 16-byte boundary
totalSize = (totalSize + 15) & (~15);
numIntegerArgs = 0;
numFloatArgs = 0;
for (i = firstArgumentChild; i < callNode->getNumChildren(); i++)
{
TR::MemoryReference *mref = NULL;
TR::Register *argRegister;
TR::InstOpCode::Mnemonic op;
child = callNode->getChild(i);
childType = child->getDataType();
switch (childType)
{
case TR::Int8:
case TR::Int16:
case TR::Int32:
case TR::Int64:
case TR::Address:
if (childType == TR::Address)
argRegister = pushAddressArg(child);
else if (childType == TR::Int64)
argRegister = pushLongArg(child);
else
argRegister = pushIntegerWordArg(child);
if (numIntegerArgs < properties.getNumIntArgRegs())
{
if (!cg()->canClobberNodesRegister(child, 0))
{
if (argRegister->containsCollectedReference())
tempReg = cg()->allocateCollectedReferenceRegister();
else
tempReg = cg()->allocateRegister();
generateMovInstruction(cg(), callNode, tempReg, argRegister);
argRegister = tempReg;
}
if (numIntegerArgs == 0 &&
(resType.isAddress() || resType.isInt32() || resType.isInt64()))
{
TR::Register *resultReg;
if (resType.isAddress())
resultReg = cg()->allocateCollectedReferenceRegister();
else
resultReg = cg()->allocateRegister();
dependencies->addPreCondition(argRegister, TR::RealRegister::x0);
dependencies->addPostCondition(resultReg, TR::RealRegister::x0);
}
else
{
addDependency(dependencies, argRegister, properties.getIntegerArgumentRegister(numIntegerArgs), TR_GPR, cg());
}
}
else
{
// numIntegerArgs >= properties.getNumIntArgRegs()
if (childType == TR::Address || childType == TR::Int64)
{
op = TR::InstOpCode::strpostx;
}
else
{
op = TR::InstOpCode::strpostw;
}
mref = getOutgoingArgumentMemRef(argMemReg, argRegister, op, pushToMemory[argIndex++]);
argSize += 8; // always 8-byte aligned
}
numIntegerArgs++;
break;
case TR::Float:
case TR::Double:
if (childType == TR::Float)
argRegister = pushFloatArg(child);
else
argRegister = pushDoubleArg(child);
if (numFloatArgs < properties.getNumFloatArgRegs())
{
if (!cg()->canClobberNodesRegister(child, 0))
{
tempReg = cg()->allocateRegister(TR_FPR);
op = (childType == TR::Float) ? TR::InstOpCode::fmovs : TR::InstOpCode::fmovd;
generateTrg1Src1Instruction(cg(), op, callNode, tempReg, argRegister);
argRegister = tempReg;
}
if ((numFloatArgs == 0 && resType.isFloatingPoint()))
{
TR::Register *resultReg;
if (resType.getDataType() == TR::Float)
resultReg = cg()->allocateSinglePrecisionRegister();
else
resultReg = cg()->allocateRegister(TR_FPR);
dependencies->addPreCondition(argRegister, TR::RealRegister::v0);
dependencies->addPostCondition(resultReg, TR::RealRegister::v0);
}
else
{
addDependency(dependencies, argRegister, properties.getFloatArgumentRegister(numFloatArgs), TR_FPR, cg());
}
}
else
{
// numFloatArgs >= properties.getNumFloatArgRegs()
if (childType == TR::Double)
{
op = TR::InstOpCode::vstrpostd;
}
else
{
op = TR::InstOpCode::vstrposts;
}
mref = getOutgoingArgumentMemRef(argMemReg, argRegister, op, pushToMemory[argIndex++]);
argSize += 8; // always 8-byte aligned
}
numFloatArgs++;
break;
} // end of switch
} // end of for
// NULL deps for non-preserved and non-system regs
while (numIntegerArgs < properties.getNumIntArgRegs())
{
if (numIntegerArgs == 0 && resType.isAddress())
{
dependencies->addPreCondition(cg()->allocateRegister(), properties.getIntegerArgumentRegister(0));
dependencies->addPostCondition(cg()->allocateCollectedReferenceRegister(), properties.getIntegerArgumentRegister(0));
}
else
{
addDependency(dependencies, NULL, properties.getIntegerArgumentRegister(numIntegerArgs), TR_GPR, cg());
}
numIntegerArgs++;
}
int32_t floatRegsUsed = (numFloatArgs > properties.getNumFloatArgRegs()) ? properties.getNumFloatArgRegs() : numFloatArgs;
for (i = (TR::RealRegister::RegNum)((uint32_t)TR::RealRegister::v0 + floatRegsUsed); i <= TR::RealRegister::LastFPR; i++)
{
if (!properties.getPreserved((TR::RealRegister::RegNum)i))
{
// NULL dependency for non-preserved regs
addDependency(dependencies, NULL, (TR::RealRegister::RegNum)i, TR_FPR, cg());
}
}
if (numMemArgs > 0)
{
TR::RealRegister *sp = cg()->machine()->getRealRegister(properties.getStackPointerRegister());
generateTrg1Src1ImmInstruction(cg(), TR::InstOpCode::subimmx, callNode, argMemReg, sp, totalSize);
for (argIndex = 0; argIndex < numMemArgs; argIndex++)
{
TR::Register *aReg = pushToMemory[argIndex].argRegister;
generateMemSrc1Instruction(cg(), pushToMemory[argIndex].opCode, callNode, pushToMemory[argIndex].argMemory, aReg);
cg()->stopUsingRegister(aReg);
}
cg()->stopUsingRegister(argMemReg);
}
return totalSize;
}
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;
}
void TR_ExpressionsSimplification::simplifyInvariantLoopExpressions(ListIterator<TR::Block> &blocks)
{
// Need to locate the induction variable of the loop
//
LoopInfo *loopInfo = findLoopInfo(_currentRegion);
if (trace())
{
if (!loopInfo)
{
traceMsg(comp(), "Accurate loop info is not found, cannot carry out summation reduction\n");
}
else
{
traceMsg(comp(), "Accurate loop info has been found, will try to carry out summation reduction\n");
if (loopInfo->getBoundaryNode())
{
traceMsg(comp(), "Variable iterations from node %p has not been handled\n",loopInfo->getBoundaryNode());
}
else
{
traceMsg(comp(), "Natural Loop %p will run %d times\n", _currentRegion, loopInfo->getNumIterations());
}
}
}
// Initialize the list of candidates
//
_candidateTTs = new (trStackMemory()) TR_ScratchList<TR::TreeTop>(trMemory());
for (TR::Block *currentBlock = blocks.getFirst(); currentBlock; currentBlock = blocks.getNext())
{
if (trace())
traceMsg(comp(), "Analyzing block #%d, which must be executed once per iteration\n", currentBlock->getNumber());
// Scan through each node in the block
//
TR::TreeTop *tt = currentBlock->getEntry();
TR::TreeTop *exitTreeTop = currentBlock->getExit();
while (tt != exitTreeTop)
{
TR::Node *currentNode = tt->getNode();
if (trace())
traceMsg(comp(), "Analyzing tree top node %p\n", currentNode);
if (loopInfo)
{
// requires loop info for the number of iterations
setSummationReductionCandidates(currentNode, tt);
}
setStoreMotionCandidates(currentNode, tt);
tt = tt->getNextTreeTop();
}
}
// New code: without using any UDI
// walk through the trees in the loop
// to invalidate the candidates
//
if (!_supportedExpressions)
{
_supportedExpressions = new (trStackMemory()) TR_BitVector(comp()->getNodeCount(), trMemory(), stackAlloc, growable);
}
invalidateCandidates();
ListIterator<TR::TreeTop> treeTops(_candidateTTs);
for (TR::TreeTop *treeTop = treeTops.getFirst(); treeTop; treeTop = treeTops.getNext())
{
if (trace())
traceMsg(comp(), "Candidate TreeTop: %p, Node:%p\n", treeTop, treeTop->getNode());
bool usedCandidate = false;
bool isPreheaderBlockInvalid = false;
if (loopInfo)
{
usedCandidate = tranformSummationReductionCandidate(treeTop, loopInfo, &isPreheaderBlockInvalid);
}
if (isPreheaderBlockInvalid)
{
break;
}
if (!usedCandidate)
{
tranformStoreMotionCandidate(treeTop, &isPreheaderBlockInvalid);
}
if (isPreheaderBlockInvalid)
{
break;
}
}
}
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()));
}
