TR_DominatorVerifier::TR_DominatorVerifier(TR_Dominators &findDominators)
: _compilation(findDominators.comp())
{
TR::StackMemoryRegion stackMemoryRegion(*trMemory());
_dominators = &findDominators;
TR::CFG *cfg = comp()->getFlowGraph();
_visitCount = comp()->incVisitCount();
_numBlocks = cfg->getNumberOfNodes()+1;
if (debug("traceVER"))
{
dumpOptDetails(comp(), "Printing out the TreeTops from DominatorVerifier\n");
TR::TreeTop *currentTree = comp()->getStartTree();
while (!(currentTree == NULL))
{
comp()->getDebug()->print(comp()->getOutFile(), currentTree);
currentTree = currentTree->getNextTreeTop();
}
dumpOptDetails(comp(), "Printing out the CFG from DominatorVerifier\n");
if (cfg != NULL)
comp()->getDebug()->print(comp()->getOutFile(), cfg);
}
TR_DominatorsChk expensiveAlgorithm(comp());
expensiveAlgorithmCorrect = isExpensiveAlgorithmCorrect(expensiveAlgorithm);
if (expensiveAlgorithmCorrect)
{
if (debug("traceVER"))
dumpOptDetails(comp(), "Dominators computed by the expensive algorithm are correct\n");
}
else
{
if (debug("traceVER"))
dumpOptDetails(comp(), "Dominators computed by the expensive algorithm are NOT correct\n");
TR_ASSERT(0, "Dominators computed by the expensive algorithm are NOT correct\n");
}
bothImplementationsConsistent = areBothImplementationsConsistent(expensiveAlgorithm, findDominators);
if (bothImplementationsConsistent)
{
if (debug("traceVER"))
dumpOptDetails(comp(), "Dominators computed by the two implementations are consistent\n");
}
else
{
if (debug("traceVER"))
dumpOptDetails(comp(), "Dominators computed by the two implementations are NOT consistent\n");
TR_ASSERT(0, "Dominators computed by the two implementations are NOT consistent\n");
}
}
void
TR_ReachabilityAnalysis::perform(TR_BitVector *result)
{
TR::CFG *cfg = comp()->getFlowGraph();
int32_t numBlockIndexes = cfg->getNextNodeNumber();
int32_t numBlocks = cfg->getNumberOfNodes();
_blocks = cfg->createArrayOfBlocks();
blocknum_t *stack = (blocknum_t*)comp()->trMemory()->allocateStackMemory(numBlockIndexes * sizeof(stack[0]));
blocknum_t *depthMap = (blocknum_t*)comp()->trMemory()->allocateStackMemory(numBlockIndexes * sizeof(depthMap[0]));
memset(depthMap, 0, numBlockIndexes * sizeof(depthMap[0]));
bool trace = comp()->getOption(TR_TraceReachability);
if (trace)
traceMsg(comp(), "BEGIN REACHABILITY: %d blocks\n", numBlocks);
for (TR::Block *block = comp()->getStartBlock(); block; block = block->getNextBlock())
{
blocknum_t blockNum = block->getNumber();
if (trace)
traceMsg(comp(), "Visit block_%d\n", blockNum);
if (depthMap[blockNum] == 0)
traverse(blockNum, 0, stack, depthMap, result);
else
traceMsg(comp(), " depth is already %d; skip\n", depthMap[blockNum]);
}
if (comp()->getOption(TR_TraceReachability))
{
traceMsg(comp(), "END REACHABILITY. Result:\n");
result->print(comp(), comp()->getOutFile());
traceMsg(comp(), "\n");
}
}
int32_t TR_CatchBlockRemover::perform()
{
TR::CFG *cfg = comp()->getFlowGraph();
if (cfg == NULL)
{
if (trace())
traceMsg(comp(), "Can't do Catch Block Removal, no CFG\n");
return 0;
}
if (trace())
traceMsg(comp(), "Starting Catch Block Removal\n");
bool thereMayBeRemovableCatchBlocks = false;
{
TR::StackMemoryRegion stackMemoryRegion(*trMemory());
TR::Block *block;
ListIterator<TR::CFGEdge> edgeIterator;
// Go through all blocks that have exception successors and see if any of them
// are not reached. Mark each of these edges with a visit count so they can
// be identified later.
//
vcount_t visitCount = comp()->incOrResetVisitCount();
TR::CFGNode *cfgNode;
for (cfgNode = cfg->getFirstNode(); cfgNode; cfgNode = cfgNode->getNext())
{
if (cfgNode->getExceptionSuccessors().empty())
continue;
block = toBlock(cfgNode);
uint32_t reachedExceptions = 0;
TR::TreeTop *treeTop;
for (treeTop = block->getEntry(); treeTop != block->getExit(); treeTop = treeTop->getNextTreeTop())
{
reachedExceptions |= treeTop->getNode()->exceptionsRaised();
if (treeTop->getNode()->getOpCodeValue() == TR::monexitfence) // for live monitor metadata
reachedExceptions |= TR::Block::CanCatchMonitorExit;
}
if (reachedExceptions & TR::Block::CanCatchUserThrows)
continue;
for (auto edge = block->getExceptionSuccessors().begin(); edge != block->getExceptionSuccessors().end();)
{
TR::CFGEdge * current = *(edge++);
TR::Block *catchBlock = toBlock(current->getTo());
if (catchBlock->isOSRCodeBlock() || catchBlock->isOSRCatchBlock()) continue;
if (!reachedExceptions &&
performTransformation(comp(), "%sRemove redundant exception edge from block_%d at [%p] to catch block_%d at [%p]\n", optDetailString(), block->getNumber(), block, catchBlock->getNumber(), catchBlock))
{
cfg->removeEdge(block, catchBlock);
thereMayBeRemovableCatchBlocks = true;
}
else
{
if (!catchBlock->canCatchExceptions(reachedExceptions))
{
current->setVisitCount(visitCount);
thereMayBeRemovableCatchBlocks = true;
}
}
}
}
bool edgesRemoved = false;
// Now look to see if there are any catch blocks for which all exception
// predecessors have the visit count set. If so, the block is unreachable and
// can be removed.
// If only some of the exception predecessors are marked, these edges are
// left in place to identify the try/catch structure properly.
//
while (thereMayBeRemovableCatchBlocks)
{
thereMayBeRemovableCatchBlocks = false;
for (cfgNode = cfg->getFirstNode(); cfgNode; cfgNode = cfgNode->getNext())
{
if (cfgNode->getExceptionPredecessors().empty())
continue;
auto edgeIt = cfgNode->getExceptionPredecessors().begin();
for (; edgeIt != cfgNode->getExceptionPredecessors().end(); ++edgeIt)
{
if ((*edgeIt)->getVisitCount() != visitCount)
break;
}
if (edgeIt == cfgNode->getExceptionPredecessors().end() && performTransformation(comp(), "%sRemove redundant catch block_%d at [%p]\n", optDetailString(), cfgNode->getNumber(), cfgNode))
{
while (!cfgNode->getExceptionPredecessors().empty())
{
cfg->removeEdge(cfgNode->getExceptionPredecessors().front());
}
edgesRemoved = true;
thereMayBeRemovableCatchBlocks = true;
}
}
}
// Any transformations invalidate use/def and value number information
//
if (edgesRemoved)
{
optimizer()->setUseDefInfo(NULL);
optimizer()->setValueNumberInfo(NULL);
requestOpt(OMR::treeSimplification, true);
}
} // scope of the stack memory region
if (trace())
traceMsg(comp(), "\nEnding Catch Block Removal\n");
return 1; // actual cost
}
// This function splits a single succeesor block following an guard and is used to
// do the following transform
// block - cold1 block - cold1
// \ / => | |
// nextBlock nextBlock nextBlock' (called tailSplitBlock below)
// | \ /
// ... ...
void TR_VirtualGuardHeadMerger::tailSplitBlock(TR::Block * block, TR::Block * cold1)
{
TR::CFG *cfg = comp()->getFlowGraph();
cfg->setStructure(NULL);
TR_BlockCloner cloner(cfg);
TR::Block *tailSplitBlock = cloner.cloneBlocks(block->getNextBlock(), block->getNextBlock());
tailSplitBlock->setFrequency(cold1->getFrequency());
if (cold1->isCold())
tailSplitBlock->setIsCold();
// physically put the block after cold1 since we want cold1 to fall through
tailSplitBlock->getExit()->join(cold1->getExit()->getNextTreeTop());
cold1->getExit()->join(tailSplitBlock->getEntry());
// remove cold1's goto
TR::TransformUtil::removeTree(comp(), cold1->getExit()->getPrevRealTreeTop());
// copy the exception edges
for (auto e = block->getNextBlock()->getExceptionSuccessors().begin(); e != block->getNextBlock()->getExceptionSuccessors().end(); ++e)
cfg->addExceptionEdge(tailSplitBlock, (*e)->getTo());
cfg->addEdge(cold1, tailSplitBlock);
// lastly fix up the exit of tailSplitBlock
TR::Node *tailSplitEnd = tailSplitBlock->getExit()->getPrevRealTreeTop()->getNode();
if (tailSplitEnd->getOpCode().isGoto())
{
tailSplitEnd->setBranchDestination(block->getNextBlock()->getLastRealTreeTop()->getNode()->getBranchDestination());
cfg->addEdge(tailSplitBlock, block->getNextBlock()->getSuccessors().front()->getTo());
}
else if (tailSplitEnd->getOpCode().isBranch())
{
TR::Block *gotoBlock = TR::Block::createEmptyBlock(tailSplitEnd, comp(), cold1->getFrequency());
if (cold1->isCold())
gotoBlock->setIsCold(true);
gotoBlock->getExit()->join(tailSplitBlock->getExit()->getNextTreeTop());
tailSplitBlock->getExit()->join(gotoBlock->getEntry());
cfg->addNode(gotoBlock);
gotoBlock->append(TR::TreeTop::create(comp(), TR::Node::create(tailSplitEnd, TR::Goto, 0, block->getNextBlock()->getExit()->getNextTreeTop())));
cfg->addEdge(tailSplitBlock, gotoBlock);
cfg->addEdge(tailSplitBlock, tailSplitBlock->getLastRealTreeTop()->getNode()->getBranchDestination()->getEnclosingBlock());
cfg->addEdge(gotoBlock, block->getNextBlock()->getNextBlock());
}
else if (
!tailSplitEnd->getOpCode().isReturn() &&
!tailSplitEnd->getOpCode().isJumpWithMultipleTargets() &&
tailSplitEnd->getOpCodeValue() != TR::athrow &&
!(tailSplitEnd->getNumChildren() >= 1 && tailSplitEnd->getFirstChild()->getOpCodeValue() == TR::athrow)
)
{
tailSplitBlock->append(TR::TreeTop::create(comp(), TR::Node::create(tailSplitEnd, TR::Goto, 0, block->getNextBlock()->getExit()->getNextTreeTop())));
cfg->addEdge(tailSplitBlock, block->getNextBlock()->getNextBlock());
}
else
{
for (auto e = block->getNextBlock()->getSuccessors().begin(); e != block->getNextBlock()->getSuccessors().end(); ++e)
cfg->addEdge(tailSplitBlock, (*e)->getTo());
}
cfg->removeEdge(cold1, block->getNextBlock());
optimizer()->setUseDefInfo(NULL);
optimizer()->setValueNumberInfo(NULL);
}
// This opt tries to reduce merge backs from cold code that are the result of inliner
// gnerated nopable virtual guards
// It looks for one basic pattern
//
// guard1 -> cold1
// BBEND
// BBSTART
// guard2 -> cold2
// if guard1 is the guard for a method which calls the method guard2 protects or cold1 is
// a predecessor of cold2 (a situation commonly greated by virtual guard tail splitter) we
// can transform the guards as follows when guard1 and guard2 a
// guard1 -> cold1
// BBEND
// BBSTART
// guard2 -> cold1
// This is safe because there are no trees between the guards and calling the caller will
// result in the call to the callee if we need to patch guard2. cold2 and its mergebacks
// can then be eliminated
//
// In addition this opt will try to move guard2 up from the end of a block to the
// start of the block. We can do this if guard2 is an HCR guard and there is no GC point
// between BBSTART and guard2 since HCR is a stop-the-world event.
//
// Finally, there is a simple tail splitting step run before the analysis of a guard if we
// detect that the taken side of the guard merges back in the next block - this happens
// for some empty methods and is common for Object.<init> at the top of constructors.
int32_t TR_VirtualGuardHeadMerger::perform() {
static char *disableVGHeadMergerTailSplitting = feGetEnv("TR_DisableVGHeadMergerTailSplitting");
TR::CFG *cfg = comp()->getFlowGraph();
// Cache the loads for the outer guard's cold path
TR_BitVector coldPathLoads(comp()->trMemory()->currentStackRegion());
TR_BitVector privArgSymRefs(comp()->trMemory()->currentStackRegion());
bool evaluatedColdPathLoads = false;
for (TR::Block *block = optimizer()->getMethodSymbol()->getFirstTreeTop()->getNode()->getBlock();
block; block = block->getNextBlock())
{
TR::Node *guard1 = block->getLastRealTreeTop()->getNode();
if (isMergeableGuard(guard1))
{
if (trace())
traceMsg(comp(), "Found mergeable guard in block_%d\n", block->getNumber());
TR::Block *cold1 = guard1->getBranchDestination()->getEnclosingBlock();
// check for an immediate merge back from the cold block and
// tail split one block if we can - we only handle splitting a block
// ending in a fallthrough, a branch or a goto for now for simplicity
if (!disableVGHeadMergerTailSplitting &&
(cold1->getSuccessors().size() == 1) &&
cold1->hasSuccessor(block->getNextBlock()) &&
cold1->getLastRealTreeTop()->getNode()->getOpCode().isGoto())
{
// TODO handle moving code earlier in the block down below the guard
// tail split
if ((block->getNextBlock()->getSuccessors().size() == 1) ||
((block->getNextBlock()->getSuccessors().size() == 2) &&
block->getNextBlock()->getLastRealTreeTop()->getNode()->getOpCode().isBranch()) &&
performTransformation(comp(), "%sCloning block_%d and placing clone after block_%d to reduce HCR guard nops\n", OPT_DETAILS, block->getNextBlock()->getNumber(), cold1->getNumber()))
tailSplitBlock(block, cold1);
}
// guard motion is fairly complex but what we want to achieve around guard1 is a sequence
// of relocated privarg blocks, followed by a sequence of runtime patchable guards going to
// guard1's cold block, followed by a sequence of stop-the-world guards going to guard1's
// cold block
//
// The following code is to setup the various insert points based on the following diagrams
// of basic blocks:
//
// start: setup: end result after moving runtime guard'
// | | +-------+ <-- privargIns
// | | <-- privargIns |
// +-------+ <-- runtimeIns +-------+
// | | | | Guard'|
// | | V +-------+ <-- runtimeIns
// +-------+ +-------+ |
// | Guard | | Guard | V
// +-------+ +-------+ <-- HCRIns +-------+
// | ===> | ===> | Guard |
// V V +-------+ <-- HCRIns
// +-------+ +-------+ |
// | | | | V
// | | | | +-------+
//
// Note we always split the block - this may create an empty block but preserves the incoming
// control flow we leave the rest to block extension to fix later
block = block->split(block->getLastRealTreeTop(), cfg, true, false);
TR::Block *privargIns = block->getPrevBlock();
TR::Block *runtimeIns = block->getPrevBlock();
TR::Block *HCRIns = block;
// New outer guard so cold paths must be evaluated
evaluatedColdPathLoads = false;
// scan for candidate guards to merge with guard1 identified above
for (TR::Block *nextBlock = block->getNextBlock(); nextBlock; nextBlock = nextBlock->getNextBlock())
{
if (!(nextBlock->getPredecessors().size() == 1) ||
!nextBlock->hasPredecessor(block))
{
break;
}
TR::TreeTop *guard2Tree = NULL;
if (isMergeableGuard(nextBlock->getFirstRealTreeTop()->getNode()))
{
guard2Tree = nextBlock->getFirstRealTreeTop();
}
else if (isMergeableGuard(nextBlock->getLastRealTreeTop()->getNode()))
{
guard2Tree = nextBlock->getLastRealTreeTop();
}
else
break;
TR::Node *guard2 = guard2Tree->getNode();
TR::Block *guard2Block = nextBlock;
// It is not possible to shift an OSR guard unless the destination is already an OSR point
// as the necessary OSR state will not be available
if (guard2->isOSRGuard() && !guard1->isOSRGuard())
break;
TR::Block *insertPoint = isStopTheWorldGuard(guard2) ? HCRIns : runtimeIns;
if (!safeToMoveGuard(insertPoint, guard2Tree, guard1->getBranchDestination(), privArgSymRefs))
break;
// now we figure out if we can redirect guard2 to guard1's cold block
// ie can we do the head merge
TR::Block *cold2 = guard2->getBranchDestination()->getEnclosingBlock();
if (guard1->getInlinedSiteIndex() == guard2->getInlinedSiteIndex())
{
if (trace())
traceMsg(comp(), " Guard1 [%p] is guarding the same call as Guard2 [%p] - proceeding with guard merging\n", guard1, guard2);
}
else if (guard2->getInlinedSiteIndex() > -1 &&
guard1->getInlinedSiteIndex() == TR::comp()->getInlinedCallSite(guard2->getInlinedSiteIndex())._byteCodeInfo.getCallerIndex())
{
if (trace())
traceMsg(comp(), " Guard1 [%p] is the caller of Guard2 [%p] - proceeding with guard merging\n", guard1, guard2);
}
else if ((cold1->getSuccessors().size() == 1) &&
cold1->hasSuccessor(cold2))
{
if (trace())
traceMsg(comp(), " Guard1 cold destination block_%d has guard2 cold destination block_%d as its only successor - proceeding with guard merging\n", cold1->getNumber(), cold2->getNumber());
}
else
{
if (trace())
traceMsg(comp(), " Cold1 block_%d and cold2 block_%d of guard2 [%p] in unknown relationship - abandon the merge attempt\n", cold1->getNumber(), cold2->getNumber(), guard2);
break;
}
// Runtime guards will shift their privargs, so it is necessary to check such a move is safe
// This is possible if a privarg temp was recycled for the inner call site, with a prior use as an
// argument for the outer call site. As the privargs for the inner call site must be evaluated before
// both guards, this would result in the recycled temp holding the incorrect value if the guard is ever
// taken.
if (!isStopTheWorldGuard(guard2))
{
if (!evaluatedColdPathLoads)
{
collectColdPathLoads(cold1, coldPathLoads);
evaluatedColdPathLoads = true;
}
if (coldPathLoads.intersects(privArgSymRefs))
{
if (trace())
traceMsg(comp(), " Recycled temp live in cold1 block_%d and used as privarg before guard2 [%p] - stop guard merging", cold1->getNumber(), guard2);
break;
}
}
if (!performTransformation(comp(), "%sRedirecting %s guard [%p] in block_%d to parent guard cold block_%d\n", OPT_DETAILS, isStopTheWorldGuard(guard2) ? "stop the world" : "runtime", guard2, guard2Block->getNumber(), cold1->getNumber()))
continue;
if (guard2->getBranchDestination() != guard1->getBranchDestination())
guard2Block->changeBranchDestination(guard1->getBranchDestination(), cfg);
if (guard2Tree != guard2Block->getFirstRealTreeTop())
{
cfg->setStructure(NULL);
// We should leave code ahead of an HCR guard in place because:
// 1, it might have side effect to runtime guards after it, moving it up might cause us to falsely merge
// the subsequent runtime guards
// 2, it might contain live monitor, moving it up above a guard can affect the monitor's live range
if (!isStopTheWorldGuard(guard2))
{
// the block created above guard2 contains only privarg treetops or monitor stores if
// guard2 is a runtime-patchable guard and is safe to merge. We need to move the priv
// args up to the runtime insert point and leave the monitor stores in place
// It's safe to do so because there is no data dependency between the monitor store and
// the priv arg store, because the priv arg store does not load the value from the temp
// holding the monitored object
// Split priv arg stores from monitor stores
// Monitor store is generated for the caller of the method guard2 protects, so should appear before
// priv arg stores for the method guard2 protects
TR::Block *privargBlock = guard2Block;
guard2Block = splitRuntimeGuardBlock(comp(), guard2Block, cfg);
if (privargBlock != guard2Block)
{
if (trace())
traceMsg(comp(), " Moving privarg block_%d after block_%d\n", privargBlock->getNumber(), privargIns->getNumber());
moveBlockAfterDest(cfg, privargBlock, privargIns);
if (HCRIns == privargIns)
HCRIns = privargBlock;
if (runtimeIns == privargIns)
runtimeIns = privargBlock;
privargIns = privargBlock;
// refresh the insertPoint since it could be stale after the above updates
insertPoint = runtimeIns;
}
}
guard2Block = guard2Block->split(guard2Tree, cfg, true, false);
if (trace())
traceMsg(comp(), " Created new block_%d to hold guard [%p] from block_%d\n", guard2Block->getNumber(), guard2, guard2Block->getNumber());
}
if (insertPoint != guard2Block->getPrevBlock())
{
TR::DebugCounter::incStaticDebugCounter(comp(), TR::DebugCounter::debugCounterName(comp(), "headMerger/%s_%s/(%s)", isStopTheWorldGuard(guard1) ? "stop the world" : "runtime", isStopTheWorldGuard(guard2) ? "stop the world" : "runtime", comp()->signature()));
cfg->setStructure(NULL);
block = nextBlock = guard2Block->getPrevBlock();
if (trace())
traceMsg(comp(), " Moving guard2 block block_%d after block_%d\n", guard2Block->getNumber(), insertPoint->getNumber());
moveBlockAfterDest(cfg, guard2Block, insertPoint);
if (HCRIns == insertPoint)
HCRIns = guard2Block;
if (runtimeIns == insertPoint)
runtimeIns = guard2Block;
}
else
{
block = guard2Block;
}
guard1 = guard2;
}
}
}
return 1;
}
TR_Latestness::TR_Latestness(TR::Compilation *comp, TR::Optimizer *optimizer, TR_Structure *rootStructure, bool trace)
: TR_BackwardIntersectionBitVectorAnalysis(comp, comp->getFlowGraph(), optimizer, trace)
{
_delayedness = new (comp->allocator()) TR_Delayedness(comp, optimizer, rootStructure, trace);
_supportedNodesAsArray = _delayedness->_supportedNodesAsArray;
if (trace)
traceMsg(comp, "Starting Latestness\n");
TR::CFG *cfg = comp->getFlowGraph();
_numberOfNodes = cfg->getNextNodeNumber();
TR_ASSERT(_numberOfNodes > 0, "Latestness, node numbers not assigned");
_numberOfBits = getNumberOfBits();
_inSetInfo = (ContainerType **)trMemory()->allocateStackMemory(_numberOfNodes*sizeof(ContainerType *));
for (int32_t i=0;i<_numberOfNodes;i++)
allocateContainer(_inSetInfo+i);
// Allocate temp bit vectors from block info, since it is local to this analysis
ContainerType *intersection, *negation;
allocateBlockInfoContainer(&intersection);
allocateBlockInfoContainer(&negation);
TR::CFGNode *nextNode;
for (nextNode = cfg->getFirstNode(); nextNode; nextNode = nextNode->getNext())
{
TR_BlockStructure *blockStructure = (toBlock(nextNode))->getStructureOf();
if ((blockStructure == NULL) || (blockStructure->getBlock()->getSuccessors().empty() && blockStructure->getBlock()->getExceptionSuccessors().empty()))
continue;
/////analyzeTreeTopsInBlockStructure(blockStructure);
/////analysisInfo->_containsExceptionTreeTop = _containsExceptionTreeTop;
initializeInfo(intersection);
for (auto succ = nextNode->getSuccessors().begin(); succ != nextNode->getSuccessors().end(); ++succ)
{
TR::CFGNode *succBlock = (*succ)->getTo();
compose(intersection, _delayedness->_inSetInfo[succBlock->getNumber()]);
}
/////if (getAnalysisInfo(blockStructure)->_containsExceptionTreeTop)
{
for (auto succ = nextNode->getExceptionSuccessors().begin(); succ != nextNode->getExceptionSuccessors().end(); ++succ)
{
TR::CFGNode *succBlock = (*succ)->getTo();
compose(intersection, _delayedness->_inSetInfo[succBlock->getNumber()]);
}
}
negation->setAll(_numberOfBits);
*negation -= *intersection;
copyFromInto(negation, _inSetInfo[blockStructure->getNumber()]);
*(_inSetInfo[blockStructure->getNumber()]) |= *(_delayedness->_earliestness->_globalAnticipatability->_localAnticipatability.getDownwardExposedAnalysisInfo(blockStructure->getBlock()->getNumber()));
*(_inSetInfo[blockStructure->getNumber()]) &= *(_delayedness->_inSetInfo[blockStructure->getNumber()]);
if (trace)
{
traceMsg(comp, "\nIn Set of Block : %d\n", blockStructure->getNumber());
_inSetInfo[blockStructure->getNumber()]->print(comp);
}
}
if (trace)
traceMsg(comp, "\nEnding Latestness\n");
// Null out info that will not be used by callers
_delayedness->_inSetInfo = NULL;
_blockAnalysisInfo = NULL;
}
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()));
}