void TR_Latestness::analyzeTreeTopsInBlockStructure(TR_BlockStructure *blockStructure)
{
TR::Block *block = blockStructure->getBlock();
TR::TreeTop *currentTree = block->getExit();
TR::TreeTop *entryTree = block->getEntry();
/////copyFromInto(_regularInfo, _outSetInfo[blockStructure->getNumber()]);
bool notSeenTreeWithChecks = true;
_containsExceptionTreeTop = false;
while (!(currentTree == entryTree))
{
if (notSeenTreeWithChecks)
{
bool currentTreeHasChecks = treeHasChecks(currentTree);
if (currentTreeHasChecks)
{
notSeenTreeWithChecks = false;
_containsExceptionTreeTop = true;
/////compose(_regularInfo, _exceptionInfo);
/////compose(_outSetInfo[blockStructure->getNumber()], _exceptionInfo);
}
}
else
break;
if (!(currentTree == entryTree))
currentTree = currentTree->getPrevTreeTop();
}
}
void TR_ReachingDefinitions::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.
//
TR::Block *block;
int32_t blockNum = 0;
bool seenException = false;
TR_BitVector defsKilled(getNumberOfBits(), trMemory()->currentStackRegion());
comp()->incVisitCount();
for (TR::TreeTop *treeTop = comp()->getStartTree(); treeTop; treeTop = treeTop->getNextTreeTop())
{
TR::Node *node = treeTop->getNode();
if (node->getOpCodeValue() == TR::BBStart)
{
block = node->getBlock();
blockNum = block->getNumber();
seenException = false;
if (traceRD())
traceMsg(comp(), "\nNow generating gen and kill information for block_%d\n", blockNum);
continue;
}
#if DEBUG
if (node->getOpCodeValue() == TR::BBEnd && traceRD())
{
traceMsg(comp(), " Block %d:\n", blockNum);
traceMsg(comp(), " Gen set ");
if (_regularGenSetInfo[blockNum])
_regularGenSetInfo[blockNum]->print(comp());
else
traceMsg(comp(), "{}");
traceMsg(comp(), "\n Kill set ");
if (_regularKillSetInfo[blockNum])
_regularKillSetInfo[blockNum]->print(comp());
else
traceMsg(comp(), "{}");
traceMsg(comp(), "\n Exception Gen set ");
if (_exceptionGenSetInfo[blockNum])
_exceptionGenSetInfo[blockNum]->print(comp());
else
traceMsg(comp(), "{}");
traceMsg(comp(), "\n Exception Kill set ");
if (_exceptionKillSetInfo[blockNum])
_exceptionKillSetInfo[blockNum]->print(comp());
else
traceMsg(comp(), "{}");
continue;
}
#endif
initializeGenAndKillSetInfoForNode(node, defsKilled, seenException, blockNum, NULL);
if (!seenException && treeHasChecks(treeTop))
seenException = true;
}
}
int32_t TR_AsyncCheckInsertion::insertReturnAsyncChecks(TR::Optimization *opt, const char *counterPrefix)
{
TR::Compilation * const comp = opt->comp();
if (opt->trace())
traceMsg(comp, "Inserting return asyncchecks (%s)\n", counterPrefix);
int numAsyncChecksInserted = 0;
for (TR::TreeTop *treeTop = comp->getStartTree();
treeTop;
/* nothing */ )
{
TR::Block *block = treeTop->getNode()->getBlock();
if (block->getLastRealTreeTop()->getNode()->getOpCode().isReturn()
&& performTransformation(comp,
"%sInserting return asynccheck (%s) in block_%d\n",
opt->optDetailString(),
counterPrefix,
block->getNumber()))
{
insertAsyncCheck(block, comp, counterPrefix);
numAsyncChecksInserted++;
}
treeTop = block->getExit()->getNextRealTreeTop();
}
return numAsyncChecksInserted;
}
static void resetBlockVisitFlags(TR::Compilation *comp)
{
for (TR::Block *block = comp->getStartBlock(); block != NULL; block = block->getNextBlock())
{
block->setHasBeenVisited(false);
}
}
// Process the structure recursively
//
int32_t
TR_ExpressionsSimplification::perform(TR_Structure * str)
{
if (trace())
traceMsg(comp(), "Analyzing root Structure : %p\n", str);
TR_RegionStructure *region;
// Only regions can be simplified
//
if (!(region = str->asRegion()))
return 0;
TR_RegionStructure::Cursor it(*region);
for (TR_StructureSubGraphNode *node = it.getCurrent();
node != 0;
node = it.getNext())
{
// Too strict
/*
if ((node->getPredecessors().size() == 1))
{
TR::CFGEdge *edge = node->getPredecessors().front();
TR_StructureSubGraphNode *pred = toStructureSubGraphNode(edge->getFrom());
TR_BlockStructure *b = pred->getStructure()->asBlock();
if (b && pred->getSuccessors().size() == 1))
perform(node->getStructure());
}
*/
perform(node->getStructure());
}
// debug only
//
/*
if (region->isNaturalLoop() &&
(region->getParent() &&
!region->getParent()->asRegion()->isCanonicalizedLoop()))
{
traceMsg(comp(), "Loop not canonicalized %x\n", region);
}
*/
TR::Block *entryBlock = region->getEntryBlock();
if (region->isNaturalLoop() && !entryBlock->isCold() &&
(region->getParent() /* &&
region->getParent()->asRegion()->isCanonicalizedLoop() */))
{
if (trace())
traceMsg(comp(), "Found candidate non cold loop %p for expression elimination\n", region);
findAndSimplifyInvariantLoopExpressions(region);
}
return 1; // Need to specify the cost
}
TR::TreeTop *
OMR::TreeTop::getExtendedBlockExitTreeTop()
{
TR_ASSERT(self()->getNode()->getOpCodeValue() == TR::BBStart, "getExitTreeTop, is only valid for a bbStart");
TR::Block * b;
TR::TreeTop * exitTT = self()->getNode()->getBlock()->getExit(), * nextTT;
while ((nextTT = exitTT->getNextTreeTop()) && (b = nextTT->getNode()->getBlock(), b->isExtensionOfPreviousBlock()))
exitTT = b->getExit();
return exitTT;
}
void
TR_ForwardReachabilityWithoutExceptionEdges::propagateInputs(blocknum_t blockNum, int32_t depth, blocknum_t *stack, blocknum_t *depth_map, TR_BitVector *closure)
{
TR::Block *block = getBlock(blockNum);
for (auto edge = block->getPredecessors().begin(); edge != block->getPredecessors().end(); ++edge)
{
TR::Block *inputBlock = toBlock((*edge)->getFrom());
propagateOneInput(inputBlock->getNumber(), blockNum, depth, stack, depth_map, closure);
}
}
void
TR_BackwardReachability::propagateInputs(blocknum_t blockNum, int32_t depth, blocknum_t *stack, blocknum_t *depth_map, TR_BitVector *closure)
{
TR::Block *block = getBlock(blockNum);
TR_SuccessorIterator bi(block);
for (TR::CFGEdge *edge = bi.getFirst(); edge != NULL; edge = bi.getNext())
{
TR::Block *inputBlock = toBlock(edge->getTo());
propagateOneInput(inputBlock->getNumber(), blockNum, depth, stack, depth_map, closure);
}
}
void
TR::RegDepCopyRemoval::makeFreshCopy(TR_GlobalRegisterNumber reg)
{
RegDepInfo &dep = getRegDepInfo(reg);
if (!performTransformation(comp(),
"%schange %s in GlRegDeps n%un to an explicit copy of n%un\n",
optDetailString(),
registerName(reg),
_regDeps->getGlobalIndex(),
dep.value->getGlobalIndex()))
return;
// Split the block at fallthrough if necessary to avoid putting copies
// between branches and BBEnd.
TR::Node *curNode = _treetop->getNode();
if (curNode->getOpCodeValue() == TR::BBEnd)
{
TR::Block *curBlock = curNode->getBlock();
if (curBlock->getLastRealTreeTop() != curBlock->getLastNonControlFlowTreeTop())
{
TR::Block *fallthrough = curBlock->getNextBlock();
fallthrough = curBlock->splitEdge(curBlock, fallthrough, comp());
TR_ASSERT(curBlock->getNextBlock() == fallthrough, "bad block placement from splitEdge\n");
fallthrough->setIsExtensionOfPreviousBlock();
_treetop = fallthrough->getExit();
TR::Node *newNode = _treetop->getNode();
newNode->setChild(0, _regDeps);
newNode->setNumChildren(1);
curNode->setNumChildren(0);
if (trace())
traceMsg(comp(), "\tsplit fallthrough edge to insert copy, created block_%d\n", fallthrough->getNumber());
}
}
// Make and insert the copy
TR::Node *copyNode = NULL;
if (dep.value->getOpCode().isLoadConst())
{
// No need to depend on the other register.
// TODO heuristic for whether this is really better than a reg-reg move?
generateRegcopyDebugCounter("const-remat");
copyNode = TR::Node::create(dep.value->getOpCodeValue(), 0);
copyNode->setConstValue(dep.value->getConstValue());
}
else
{
generateRegcopyDebugCounter("fresh-copy");
copyNode = TR::Node::create(TR::PassThrough, 1, dep.value);
copyNode->setCopyToNewVirtualRegister();
}
TR::Node *copyTreetopNode = TR::Node::create(TR::treetop, 1, copyNode);
_treetop->insertBefore(TR::TreeTop::create(comp(), copyTreetopNode));
if (trace())
traceMsg(comp(), "\tcopy is n%un\n", copyNode->getGlobalIndex());
updateSingleRegDep(reg, copyNode);
}
void
TR_ExpressionsSimplification::removeUncertainBlocks(TR_RegionStructure* region, List<TR::Block> *candidateBlocksList)
{
// Examine the top region block first
//
TR::Block *entryBlock = _currentRegion->getEntryBlock();
ListIterator<TR::Block> blocks;
blocks.set(candidateBlocksList);
if (trace())
traceMsg(comp(), "Number of blocks %d, entry block number %d\n", candidateBlocksList->getSize(), entryBlock->getNumber());
for (TR::Block *block = blocks.getFirst(); block; block = blocks.getNext())
{
TR::CFGNode *cfgNode = block;
if (!(cfgNode->getExceptionSuccessors().empty()) || blockHasCalls(block, comp()))
{
if (trace())
traceMsg(comp(), "An exception can be thrown from block_%d. Removing all the blocks, since we cannot know the number of iterations.\n", block->getNumber());
candidateBlocksList->deleteAll();
break;
}
}
TR_PostDominators postDominators(comp());
if (postDominators.isValid())
{
postDominators.findControlDependents();
for (TR::Block *block = blocks.getFirst(); block; block = blocks.getNext())
{
if (postDominators.dominates(block, entryBlock) == 0)
{
candidateBlocksList->remove(block);
if (trace())
traceMsg(comp(), "Block_%d is not guaranteed to be executed at least once. Removing it from the list.\n", block->getNumber());
}
}
}
else
{
if (trace())
traceMsg(comp(), "There is no post dominators information. Removing all the blocks.\n");
for (TR::Block *block = blocks.getFirst(); block; block = blocks.getNext())
{
candidateBlocksList->remove(block);
if (trace())
traceMsg(comp(), "Block_%d is removed from the list\n", block->getNumber());
}
}
}
void TR_LocalAnalysis::initializeBlocks(TR::Block *block, TR::BitVector &blocksSeen)
{
_info[block->getNumber()]._block = block;
blocksSeen[block->getNumber()] = true;
TR::Block *next;
for (auto nextEdge = block->getSuccessors().begin(); nextEdge != block->getSuccessors().end(); ++nextEdge)
{
next = toBlock((*nextEdge)->getTo());
if (!blocksSeen.ValueAt(next->getNumber()))
initializeBlocks(next, blocksSeen);
}
for (auto nextEdge = block->getExceptionSuccessors().begin(); nextEdge != block->getExceptionSuccessors().end(); ++nextEdge)
{
next = toBlock((*nextEdge)->getTo());
if (!blocksSeen.ValueAt(next->getNumber()))
initializeBlocks(next, blocksSeen);
}
}
int32_t TR_LocalLiveRangeReduction::perform()
{
if (TR::Compiler->target.cpu.isZ())
return false;
TR::TreeTop * exitTT, * nextTT;
TR::Block *b;
TR::TreeTop * tt;
//calculate number of TreeTops in each bb (or extended bb)
for (tt = comp()->getStartTree(); tt; tt = nextTT)
{
TR::StackMemoryRegion stackMemoryRegion(*trMemory());
TR::Node *node = tt->getNode();
b = node->getBlock();
exitTT = b->getExit();
_numTreeTops = b->getNumberOfRealTreeTops()+2; //include both BBStart/BBend
//support for extended blocks
while ((nextTT = exitTT->getNextTreeTop()) && (b = nextTT->getNode()->getBlock(), b->isExtensionOfPreviousBlock()))
{
_numTreeTops += b->getNumberOfRealTreeTops()+2;
exitTT = b->getExit();
}
_treesRefInfoArray = (TR_TreeRefInfo**)trMemory()->allocateStackMemory(_numTreeTops*sizeof(TR_TreeRefInfo*));
memset(_treesRefInfoArray, 0, _numTreeTops*sizeof(TR_TreeRefInfo*));
_movedTreesList.deleteAll();
_depPairList.deleteAll();
transformExtendedBlock(tt,exitTT->getNextTreeTop());
}
if (trace())
traceMsg(comp(), "\nEnding LocalLiveRangeReducer\n");
return 2;
}
void
TR_ExpressionsSimplification::invalidateCandidates()
{
_visitCount = comp()->incVisitCount();
if (trace())
{
traceMsg(comp(), "Checking which candidates may be invalidated\n");
ListIterator<TR::TreeTop> treeTops(_candidateTTs);
for (TR::TreeTop *treeTop = treeTops.getFirst(); treeTop; treeTop = treeTops.getNext())
{
traceMsg(comp(), " Candidate treetop: %p node: %p\n", treeTop, treeTop->getNode());
}
}
TR_ScratchList<TR::Block> blocksInLoop(trMemory());
_currentRegion->getBlocks(&blocksInLoop);
ListIterator<TR::Block> blocks(&blocksInLoop);
for (TR::Block *currentBlock = blocks.getFirst(); currentBlock; currentBlock = blocks.getNext())
{
TR::TreeTop *tt = currentBlock->getEntry();
TR::TreeTop *exitTreeTop = currentBlock->getExit();
while (tt != exitTreeTop)
{
TR::Node *currentNode = tt->getNode();
if (trace())
traceMsg(comp(), "Looking at treeTop [%p]\n", currentNode);
removeCandidate(currentNode, tt);
tt = tt->getNextTreeTop();
}
}
removeUnsupportedCandidates();
}
TR::Block *
TR_ExpressionsSimplification::findPredecessorBlock(TR::CFGNode *entryNode)
{
if (!(entryNode->getPredecessors().size() == 2))
return 0;
TR::Block *block = 0;
for (auto edge = entryNode->getPredecessors().begin(); edge != entryNode->getPredecessors().end(); ++edge)
{
if ((*edge)->getFrom()->getSuccessors().size() == 1)
{
block = toBlock((*edge)->getFrom());
if (block->getStructureOf()->isLoopInvariantBlock())
break;
else
block = 0;
}
}
//traceMsg(comp(), "Commoned code will be put in block_%d\n", block->getNumber());
return block;
}
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");
}
}
void TR_ExpressionsSimplification::tranformStoreMotionCandidate(TR::TreeTop *treeTop, bool *isPreheaderBlockInvalid)
{
TR::Node *node = treeTop->getNode();
TR_ASSERT(node->getOpCode().isStore() && !node->getSymbol()->isStatic() && !node->getSymbol()->holdsMonitoredObject(),
"node %p was expected to be a non-static non-monitored object store and was not.", node);
// this candidate should be valid, either direct or indirect
if (trace())
comp()->getDebug()->print(comp()->getOutFile(), node, 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;
}
// Earlier post-dominance test ensures that the loop is executed as least once, or is canonicalized.
// but to be safe we still perform on canonicalized loops only.
if (_currentRegion->isCanonicalizedLoop()) // make sure 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 store [%p] to block_%d\n", OPT_DETAILS, node, preheaderBlock->getNumber()))
{
TR::Node *newNode = node->duplicateTree();
transformNode(newNode, preheaderBlock);
TR::TransformUtil::removeTree(comp(), treeTop);
}
}
else
{
if (trace())
traceMsg(comp(), "No canonicalized loop for this candidate\n");
}
}
static void moveBlockAfterDest(TR::CFG *cfg, TR::Block *toMove, TR::Block *dest)
{
TR::Compilation *comp = TR::comp();
// Step1 splice out toMove
TR::Block *toMovePrev = toMove->getPrevBlock();
TR::Block *toMoveSucc = toMove->getNextBlock();
toMovePrev->getExit()->join(toMoveSucc->getEntry());
// Step 2 splice toMove in after dest
TR::Block *destNext = dest->getNextBlock();
dest->getExit()->join(toMove->getEntry());
toMove->getExit()->join(destNext->getEntry());
cfg->addEdge(toMove, destNext);
cfg->addEdge(dest, toMove);
cfg->removeEdge(dest, destNext);
cfg->addEdge(toMovePrev, toMoveSucc);
cfg->removeEdge(toMovePrev, toMove);
cfg->removeEdge(toMove, toMoveSucc);
}
/**
* Search for direct loads in the taken side of a guard
*
* @param firstBlock The guard's branch destination
* @param coldPathLoads BitVector of symbol reference numbers for any direct loads seen until the merge back to mainline
*/
static void collectColdPathLoads(TR::Block* firstBlock, TR_BitVector &coldPathLoads)
{
TR_Stack<TR::Block*> blocksToCheck(TR::comp()->trMemory(), 8, false, stackAlloc);
blocksToCheck.push(firstBlock);
TR::NodeChecklist checklist(TR::comp());
coldPathLoads.empty();
while (!blocksToCheck.isEmpty())
{
TR::Block *block = blocksToCheck.pop();
for (TR::TreeTop *tt = block->getFirstRealTreeTop(); tt->getNode()->getOpCodeValue() != TR::BBEnd; tt = tt->getNextTreeTop())
collectDirectLoads(tt->getNode(), coldPathLoads, checklist);
// Search for any successors that have not merged with the mainline
for (auto itr = block->getSuccessors().begin(), end = block->getSuccessors().end(); itr != end; ++itr)
{
TR::Block *dest = (*itr)->getTo()->asBlock();
if (dest != TR::comp()->getFlowGraph()->getEnd() && dest->getPredecessors().size() == 1)
blocksToCheck.push(dest);
}
}
}
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
}
TR_GlobalRegisterNumber
OMR::X86::I386::CodeGenerator::pickRegister(
TR_RegisterCandidate *rc,
TR::Block **allBlocks,
TR_BitVector &availableRegisters,
TR_GlobalRegisterNumber &highRegisterNumber,
TR_LinkHead<TR_RegisterCandidate> *candidates)
{
if (!self()->comp()->getOption(TR_DisableRegisterPressureSimulation))
{
if (self()->comp()->getOption(TR_AssignEveryGlobalRegister))
{
// This is not really necessary except for testing purposes.
// Conceptually, the common pickRegister code should be free to make
// its choices based only on performance considerations, and shouldn't
// need to worry about correctness. When SupportsVMThreadGRA is not set,
// it is incorrect to choose the VMThread register. Therefore we mask
// it out here.
//
// Having said that, the common code *does* already mask out the
// VMThread register for convenience, so under normal circumstances,
// this code is redundant. It is only necessary when
// TR_AssignEveryGlobalRegister is set.
//
availableRegisters -= *self()->getGlobalRegisters(TR_vmThreadSpill, self()->comp()->getMethodSymbol()->getLinkageConvention());
}
return OMR::CodeGenerator::pickRegister(rc, allBlocks, availableRegisters, highRegisterNumber, candidates);
}
if ((rc->getSymbol()->getDataType() == TR::Float) ||
(rc->getSymbol()->getDataType() == TR::Double))
{
if (availableRegisters.get(7))
return 7;
if (availableRegisters.get(8))
return 8;
if (availableRegisters.get(9))
return 9;
if (availableRegisters.get(10))
return 10;
if (availableRegisters.get(11))
return 11;
if (availableRegisters.get(12))
return 12;
return -1;
}
if (!_assignedGlobalRegisters)
_assignedGlobalRegisters = new (self()->trStackMemory()) TR_BitVector(self()->comp()->getSymRefCount(), self()->trMemory(), stackAlloc, growable);
if (availableRegisters.get(5))
return 5; // esi
if (availableRegisters.get(2))
return 2; // ecx
static char *dontUseEBXasGPR = feGetEnv("dontUseEBXasGPR");
if (!dontUseEBXasGPR && availableRegisters.get(1))
return 1;
#ifdef J9_PROJECT_SPECIFIC
TR::RecognizedMethod rm = self()->comp()->getMethodSymbol()->getRecognizedMethod();
if (rm == TR::java_util_HashtableHashEnumerator_hasMoreElements)
{
if (availableRegisters.get(4))
return 4; // edi
if (availableRegisters.get(3))
return 3; // edx
}
else
#endif
{
int32_t numExtraRegs = 0;
int32_t maxRegisterPressure = 0;
vcount_t visitCount = self()->comp()->incVisitCount();
TR_BitVectorIterator bvi(rc->getBlocksLiveOnEntry());
int32_t maxFrequency = 0;
while (bvi.hasMoreElements())
{
int32_t liveBlockNum = bvi.getNextElement();
TR::Block *block = allBlocks[liveBlockNum];
if (block->getFrequency() > maxFrequency)
maxFrequency = block->getFrequency();
}
int32_t maxStaticFrequency = 0;
if (maxFrequency == 0)
{
bvi.setBitVector(rc->getBlocksLiveOnEntry());
while (bvi.hasMoreElements())
{
int32_t liveBlockNum = bvi.getNextElement();
TR::Block *block = allBlocks[liveBlockNum];
TR_BlockStructure *blockStructure = block->getStructureOf();
int32_t blockWeight = 1;
if (blockStructure &&
!block->isCold())
{
blockStructure->calculateFrequencyOfExecution(&blockWeight);
if (blockWeight > maxStaticFrequency)
maxStaticFrequency = blockWeight;
}
}
}
bool assigningEDX = false;
if (!availableRegisters.get(4) &&
availableRegisters.get(3))
assigningEDX = true;
bool vmThreadUsed = false;
bvi.setBitVector(rc->getBlocksLiveOnEntry());
while (bvi.hasMoreElements())
{
int32_t liveBlockNum = bvi.getNextElement();
TR::Block *block = allBlocks[liveBlockNum];
_assignedGlobalRegisters->empty();
int32_t numAssignedGlobalRegs = 0;
TR_RegisterCandidate *prev;
for (prev = candidates->getFirst(); prev; prev = prev->getNext())
{
bool gprCandidate = true;
if ((prev->getSymbol()->getDataType() == TR::Float) ||
(prev->getSymbol()->getDataType() == TR::Double))
gprCandidate = false;
if (gprCandidate && prev->getBlocksLiveOnEntry().get(liveBlockNum))
{
numAssignedGlobalRegs++;
if (prev->getDataType() == TR::Int64)
numAssignedGlobalRegs++;
_assignedGlobalRegisters->set(prev->getSymbolReference()->getReferenceNumber());
}
}
maxRegisterPressure = self()->estimateRegisterPressure(block, visitCount, maxStaticFrequency, maxFrequency, vmThreadUsed, numAssignedGlobalRegs, _assignedGlobalRegisters, rc->getSymbolReference(), assigningEDX);
if (maxRegisterPressure >= self()->getMaximumNumbersOfAssignableGPRs())
break;
}
// Determine if we can spare any extra registers for this candidate without spilling
// in any hot (critical) blocks
//
if (maxRegisterPressure < self()->getMaximumNumbersOfAssignableGPRs())
numExtraRegs = self()->getMaximumNumbersOfAssignableGPRs() - maxRegisterPressure;
//dumpOptDetails("For global register candidate %d reg pressure is %d maxRegs %d numExtraRegs %d\n", rc->getSymbolReference()->getReferenceNumber(), maxRegisterPressure, comp()->cg()->getMaximumNumbersOfAssignableGPRs(), numExtraRegs);
if (numExtraRegs > 0)
{
if (availableRegisters.get(4))
return 4; // edi
if (availableRegisters.get(3))
return 3; // edx
}
}
return -1; // -1 ==> don't use a global register
}
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
}
void TR::DeadTreesElimination::prePerformOnBlocks()
{
_cannotBeEliminated = false;
_delayedRegStores = false;
_targetTrees.deleteAll();
// Walk through all the blocks to remove trivial dead trees of the form
// treetop
// => node
// The problem with these trees is in the scenario where the earlier use
// of 'node' is also dead. However, our analysis won't find that because
// the reference count is > 1.
vcount_t visitCount = comp()->incOrResetVisitCount();
for (TR::TreeTop *tt = comp()->getStartTree();
tt != 0;
tt = tt->getNextTreeTop())
{
bool removed = false;
TR::Node *node = tt->getNode();
if (node->getOpCodeValue() == TR::treetop &&
node->getFirstChild()->getVisitCount() == visitCount &&
performTransformation(comp(), "%sRemove trivial dead tree: %p\n", optDetailString(), node))
{
TR::TransformUtil::removeTree(comp(), tt);
removed = true;
}
else
{
if (node->getOpCode().isCheck() &&
node->getFirstChild()->getOpCode().isCall() &&
node->getFirstChild()->getReferenceCount() == 1 &&
node->getFirstChild()->getSymbolReference()->getSymbol()->isResolvedMethod() &&
node->getFirstChild()->getSymbolReference()->getSymbol()->castToResolvedMethodSymbol()->isSideEffectFree() &&
performTransformation(comp(), "%sRemove dead check of side-effect free call: %p\n", optDetailString(), node))
{
TR::TransformUtil::removeTree(comp(), tt);
removed = true;
}
}
if (removed
&& tt->getNextTreeTop()->getNode()->getOpCodeValue() == TR::Goto
&& tt->getPrevTreeTop()->getNode()->getOpCodeValue() == TR::BBStart
&& !tt->getPrevTreeTop()->getNode()->getBlock()->isExtensionOfPreviousBlock())
{
requestOpt(OMR::redundantGotoElimination, tt->getEnclosingBlock());
}
if (node->getVisitCount() >= visitCount)
continue;
TR::TransformUtil::recursivelySetNodeVisitCount(tt->getNode(), visitCount);
}
// If the last use of an iRegLoad has been removed, then remove the node from
// the BBStart and remove the corresponding dependency node from each of the block's
// predecessors.
//
while (1)
{
bool glRegDepRemoved = false;
for (TR::Block * b = comp()->getStartBlock(); b; b = b->getNextBlock())
{
TR::TreeTop * startTT = b->getEntry();
TR::Node * startNode = startTT->getNode();
if (startNode->getNumChildren() > 0 && !debug("disableEliminationOfGlRegDeps"))
{
TR::Node * glRegDeps = startNode->getFirstChild();
TR_ASSERT(glRegDeps->getOpCodeValue() == TR::GlRegDeps, "expected TR::GlRegDeps");
for (int32_t i = glRegDeps->getNumChildren() - 1; i >= 0; --i)
{
TR::Node * dep = glRegDeps->getChild(i);
if (dep->getReferenceCount() == 1 &&
(!dep->getOpCode().isFloatingPoint() ||
cg()->getSupportsJavaFloatSemantics()) &&
performTransformation(comp(), "%sRemove GlRegDep : %p\n", optDetailString(), glRegDeps->getChild(i)))
{
glRegDeps->removeChild(i);
glRegDepRemoved = true;
TR_GlobalRegisterNumber registerNum = dep->getGlobalRegisterNumber();
for (auto e = b->getPredecessors().begin(); e != b->getPredecessors().end(); ++e)
{
TR::Block * pred = toBlock((*e)->getFrom());
if (pred == comp()->getFlowGraph()->getStart())
continue;
TR::Node * parent = pred->getLastRealTreeTop()->getNode();
if ( parent->getOpCode().isJumpWithMultipleTargets() && parent->getOpCode().hasBranchChildren())
{
for (int32_t j = parent->getCaseIndexUpperBound() - 1; j > 0; --j)
{
TR::Node * caseNode = parent->getChild(j);
TR_ASSERT(caseNode->getOpCode().isCase() || caseNode->getOpCodeValue() == TR::branch,
"having problems navigating a switch");
if (caseNode->getBranchDestination() == startTT &&
caseNode->getNumChildren() > 0 &&
0) // can't do this now that all glRegDeps are hung off the default branch
removeGlRegDep(caseNode, registerNum, pred, this);
}
}
else if (!parent->getOpCode().isReturn() &&
parent->getOpCodeValue() != TR::igoto &&
!( parent->getOpCode().isJumpWithMultipleTargets() && parent->getOpCode().hasBranchChildren()) &&
!(parent->getOpCodeValue()==TR::treetop &&
parent->getFirstChild()->getOpCode().isCall() &&
parent->getFirstChild()->getOpCode().isIndirect()))
{
if (pred->getNextBlock() == b)
parent = pred->getExit()->getNode();
removeGlRegDep(parent, registerNum, pred, this);
}
}
}
}
if (glRegDeps->getNumChildren() == 0)
startNode->removeChild(0);
}
}
if (!glRegDepRemoved)
break;
}
}
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::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_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
OMR::CodeGenPhase::performSetupForInstructionSelectionPhase(TR::CodeGenerator * cg, TR::CodeGenPhase * phase)
{
TR::Compilation *comp = cg->comp();
if (TR::Compiler->target.cpu.isZ() && TR::Compiler->om.shouldGenerateReadBarriersForFieldLoads())
{
// TODO (GuardedStorage): We need to come up with a better solution than anchoring aloadi's
// to enforce certain evaluation order
traceMsg(comp, "GuardedStorage: in performSetupForInstructionSelectionPhase\n");
auto mapAllocator = getTypedAllocator<std::pair<TR::TreeTop*, TR::TreeTop*> >(comp->allocator());
std::map<TR::TreeTop*, TR::TreeTop*, std::less<TR::TreeTop*>, TR::typed_allocator<std::pair<TR::TreeTop* const, TR::TreeTop*>, TR::Allocator> >
currentTreeTopToappendTreeTop(std::less<TR::TreeTop*> (), mapAllocator);
TR_BitVector *unAnchorableAloadiNodes = comp->getBitVectorPool().get();
for (TR::PreorderNodeIterator iter(comp->getStartTree(), comp); iter != NULL; ++iter)
{
TR::Node *node = iter.currentNode();
traceMsg(comp, "GuardedStorage: Examining node = %p\n", node);
// isNullCheck handles both TR::NULLCHK and TR::ResolveAndNULLCHK
// both of which do not operate on their child but their
// grandchild (or greatgrandchild).
if (node->getOpCode().isNullCheck())
{
// An aloadi cannot be anchored if there is a Null Check on
// its child. There are two situations where this occurs.
// The first is when doing an aloadi off some node that is
// being NULLCHK'd (see Ex1). The second is when doing an
// icalli in which case the aloadi loads the VFT of an
// object that must be NULLCHK'd (see Ex2).
//
// Ex1:
// n1n NULLCHK on n3n
// n2n aloadi f <-- First Child And Parent of Null Chk'd Node
// n3n aload O
//
// Ex2:
// n1n NULLCHK on n4n
// n2n icall foo <-- First Child
// n3n aloadi <vft> <-- Parent of Null Chk'd Node
// n4n aload O
// n4n ==> aload O
TR::Node *nodeBeingNullChkd = node->getNullCheckReference();
if (nodeBeingNullChkd)
{
TR::Node *firstChild = node->getFirstChild();
TR::Node *parentOfNullChkdNode = NULL;
if (firstChild->getOpCode().isCall() &&
firstChild->getOpCode().isIndirect())
{
parentOfNullChkdNode = firstChild->getFirstChild();
}
else
{
parentOfNullChkdNode = firstChild;
}
if (parentOfNullChkdNode &&
parentOfNullChkdNode->getOpCodeValue() == TR::aloadi &&
parentOfNullChkdNode->getNumChildren() > 0 &&
parentOfNullChkdNode->getFirstChild() == nodeBeingNullChkd)
{
unAnchorableAloadiNodes->set(parentOfNullChkdNode->getGlobalIndex());
traceMsg(comp, "GuardedStorage: Cannot anchor %p\n", firstChild);
}
}
}
else
{
bool shouldAnchorNode = false;
if (node->getOpCodeValue() == TR::aloadi &&
!unAnchorableAloadiNodes->isSet(node->getGlobalIndex()))
{
shouldAnchorNode = true;
}
else if (node->getOpCodeValue() == TR::aload &&
node->getSymbol()->isStatic() &&
node->getSymbol()->isCollectedReference())
{
shouldAnchorNode = true;
}
if (shouldAnchorNode)
{
TR::TreeTop* anchorTreeTop = TR::TreeTop::create(comp, TR::Node::create(TR::treetop, 1, node));
TR::TreeTop* appendTreeTop = iter.currentTree();
if (currentTreeTopToappendTreeTop.count(appendTreeTop) > 0)
{
appendTreeTop = currentTreeTopToappendTreeTop[appendTreeTop];
}
// Anchor the aload/aloadi before the current treetop
appendTreeTop->insertBefore(anchorTreeTop);
currentTreeTopToappendTreeTop[iter.currentTree()] = anchorTreeTop;
traceMsg(comp, "GuardedStorage: Anchored %p to treetop = %p\n", node, anchorTreeTop);
}
}
}
comp->getBitVectorPool().release(unAnchorableAloadiNodes);
}
if (cg->shouldBuildStructure() &&
(comp->getFlowGraph()->getStructure() != NULL))
{
TR_Structure *rootStructure = TR_RegionAnalysis::getRegions(comp);
comp->getFlowGraph()->setStructure(rootStructure);
}
phase->reportPhase(SetupForInstructionSelectionPhase);
// Dump preIR
if (comp->getOption(TR_TraceRegisterPressureDetails) && !comp->getOption(TR_DisableRegisterPressureSimulation))
{
traceMsg(comp, " { Post optimization register pressure simulation\n");
TR_BitVector emptyBitVector;
vcount_t vc = comp->incVisitCount();
cg->initializeRegisterPressureSimulator();
for (TR::Block *block = comp->getStartBlock(); block; block = block->getNextExtendedBlock())
{
TR_LinkHead<TR_RegisterCandidate> emptyCandidateList;
TR::CodeGenerator::TR_RegisterPressureState state(NULL, 0, emptyBitVector, emptyBitVector, &emptyCandidateList, cg->getNumberOfGlobalGPRs(), cg->getNumberOfGlobalFPRs(), cg->getNumberOfGlobalVRFs(), vc);
TR::CodeGenerator::TR_RegisterPressureSummary summary(state._gprPressure, state._fprPressure, state._vrfPressure);
cg->simulateBlockEvaluation(block, &state, &summary);
}
traceMsg(comp, " }\n");
}
TR::LexicalMemProfiler mp(phase->getName(), comp->phaseMemProfiler());
LexicalTimer pt(phase->getName(), comp->phaseTimer());
cg->setUpForInstructionSelection();
}
// 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::TreeTop *
OMR::Simplifier::simplifyExtendedBlock(TR::TreeTop * treeTop)
{
TR::Block * block = 0;
_containingStructure = NULL;
_blockRemoved = false;
for (; treeTop; treeTop = treeTop->getNextTreeTop())
{
TR::Node * node = treeTop->getNode();
TR_ASSERT(node->getOpCodeValue() == TR::BBStart, "Simplification, expected BBStart treetop");
TR::Block * b = node->getBlock();
if (block && !b->isExtensionOfPreviousBlock())
break;
if (b->isOSRCodeBlock() || b->isOSRCatchBlock())
{
b->setHasBeenVisited();
treeTop = b->getExit();
continue;
}
#ifdef DEBUG
if (block != b)
b->setHasBeenVisited();
#endif
if (!block && _reassociate &&
comp()->getFlowGraph()->getStructure() != NULL // [99391] getStructureOf() only valid if structure isn't invalidated
)
{ // b is first block in the extended block
TR_BlockStructure *blockStructure = b->getStructureOf();
if(blockStructure)
{
TR_Structure *parent = blockStructure->getParent();
while (parent)
{
TR_RegionStructure *region = parent->asRegion();
if (region->isNaturalLoop() /* || region->containsInternalCycles() */)
{
_containingStructure = region;
break;
}
parent = parent->getParent();
}
}
}
block = b;
if (trace())
traceMsg(comp(), "simplifying block_%d\n", block->getNumber());
_performLowerTreeSimplifier=NULL;
_performLowerTreeNode=NULL;
simplify(block);
if(_performLowerTreeSimplifier)
{
_performLowerTreeNode = postWalkLowerTreeSimplifier(_performLowerTreeSimplifier, _performLowerTreeNode, block, (TR::Simplifier *) this);
_performLowerTreeSimplifier->setNode(_performLowerTreeNode);
}
// If the block itself was removed from the CFG during simplification, find
// the next 'legitimate' block to be simplified
//
//if (comp()->getFlowGraph()->getRemovedNodes().find(block))
if(block->nodeIsRemoved())
{
TR::TreeTop * tt = findNextLegalTreeTop(comp(), block);
// in certain cases the removed block might be the last one we haven't
// visited and therefore we won't be able to find a treetop to continue
// in such cases we exit the loop
//
treeTop = tt ? tt->getPrevTreeTop() : 0;
if (!treeTop)
break;
}
else
{
treeTop = block->getExit();
}
}
// now remove any unreachable blocks
//
if (_blockRemoved)
{
// if the next block to be processed has been removed,
// find the next valid block to process
//
if (treeTop)
{
TR::Block *b = treeTop->getNode()->getBlock();
//if (comp()->getFlowGraph()->getRemovedNodes().find(b))
if(b->nodeIsRemoved())
treeTop = findNextLegalTreeTop(comp(), b);
}
}
return treeTop;
}
// 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);
}