void TR::ILValidator::updateNodeState(Location &newLocation)
{
TR::Node *node = newLocation.currentNode();
NodeState &state = _nodeStates[node];
if (node->getReferenceCount() == state._futureReferenceCount)
{
// First occurrence -- do some bookkeeping
//
if (node->getReferenceCount() == 0)
{
validityRule(newLocation, node->getOpCode().isTreeTop(), "Only nodes with isTreeTop opcodes can have refcount == 0");
}
else
{
_liveNodes.add(node);
}
}
if (_liveNodes.contains(node))
{
validityRule(newLocation, state._futureReferenceCount >= 1, "Node already has reference count 0");
if (--state._futureReferenceCount == 0)
{
_liveNodes.remove(node);
}
}
else
{
validityRule(newLocation, node->getOpCode().isTreeTop(), "Node has already gone dead");
}
if (isLoggingEnabled())
{
static const char *traceLiveNodesDuringValidation = feGetEnv("TR_traceLiveNodesDuringValidation");
if (traceLiveNodesDuringValidation && !_liveNodes.isEmpty())
{
traceMsg(comp(), " -- Live nodes: {");
char *separator = "";
for (LiveNodeWindow::Iterator lnwi(_liveNodes); lnwi.currentNode(); ++lnwi)
{
traceMsg(comp(), "%sn%dn", separator, lnwi.currentNode()->getGlobalIndex());
separator = ", ";
}
traceMsg(comp(), "}\n");
}
}
}
static void removeGlRegDep(TR::Node * parent, TR_GlobalRegisterNumber registerNum, TR::Block *containingBlock, TR::Optimization *opt)
{
if (parent->getNumChildren() == 0)
return;
TR_ASSERT(parent->getNumChildren() > 0, "expected TR::GlRegDeps %p", parent);
TR::Node * predGlRegDeps = parent->getLastChild();
if (predGlRegDeps->getOpCodeValue() != TR::GlRegDeps) // could be already removed
return;
TR_ASSERT(predGlRegDeps->getOpCodeValue() == TR::GlRegDeps, "expected TR::GlRegDeps");
for (int32_t i = predGlRegDeps->getNumChildren() - 1; i >= 0; --i)
if (predGlRegDeps->getChild(i)->getGlobalRegisterNumber() == registerNum)
{
dumpOptDetails(opt->comp(), "%sRemove GlRegDep : %p\n", opt->optDetailString(), predGlRegDeps->getChild(i));
TR::Node *removedChild = predGlRegDeps->removeChild(i);
if (removedChild->getReferenceCount() <= 1)
{
// The only remaining parent is the RegStore. Another pass of
// deadTrees may be able to eliminate that.
//
opt->requestOpt(OMR::deadTreesElimination, true, containingBlock);
}
break;
}
if (predGlRegDeps->getNumChildren() == 0)
parent->removeLastChild();
}
static TR::Register *l2fd(TR::Node *node, TR::RealRegister *target, TR_X86OpCodes opRegMem8, TR_X86OpCodes opRegReg8, TR::CodeGenerator *cg)
{
TR::Node *child = node->getFirstChild();
TR::MemoryReference *tempMR;
TR_ASSERT(cg->useSSEForSinglePrecision(), "assertion failure");
if (child->getRegister() == NULL &&
child->getReferenceCount() == 1 &&
child->getOpCode().isLoadVar())
{
tempMR = generateX86MemoryReference(child, cg);
generateRegMemInstruction(opRegMem8, node, target, tempMR, cg);
tempMR->decNodeReferenceCounts(cg);
}
else
{
TR::Register *intReg = cg->evaluate(child);
generateRegRegInstruction(opRegReg8, node, target, intReg, cg);
cg->decReferenceCount(child);
}
node->setRegister(target);
return target;
}
void TR::ValidateNodeRefCountWithinBlock::validate(TR::TreeTop *firstTreeTop,
TR::TreeTop *exitTreeTop)
{
_nodeChecklist.empty();
for (TR::TreeTop *tt = firstTreeTop; tt != exitTreeTop->getNextTreeTop();
tt = tt->getNextTreeTop())
{
TR::Node *node = tt->getNode();
node->setLocalIndex(node->getReferenceCount());
validateRefCountPass1(node);
}
/**
* We start again from the start of the block, and check the localIndex to
* make sure it is 0.
*
* NOTE: Walking the tree backwards causes huge stack usage in validateRefCountPass2.
*/
_nodeChecklist.empty();
for (TR::TreeTop *tt = firstTreeTop; tt != exitTreeTop->getNextTreeTop();
tt = tt->getNextTreeTop())
{
validateRefCountPass2(tt->getNode());
}
}
bool collectSymbolReferencesInNode(TR::Node *node,
TR::SparseBitVector &symbolReferencesInNode,
int32_t *numDeadSubNodes, vcount_t visitCount, TR::Compilation *comp,
bool *seenInternalPointer, bool *seenArraylet,
bool *cantMoveUnderBranch)
{
// The visit count in the node must be maintained by this method.
//
vcount_t oldVisitCount = node->getVisitCount();
if (oldVisitCount == visitCount || oldVisitCount == comp->getVisitCount())
return true;
node->setVisitCount(comp->getVisitCount());
//diagnostic("Walking node %p, height=%d, oldVisitCount=%d, visitCount=%d, compVisitCount=%d\n", node, *height, oldVisitCount, visitCount,comp->getVisitCount());
// For all other subtrees collect all symbols that could be killed between
// here and the next reference.
//
for (int32_t i = node->getNumChildren()-1; i >= 0; i--)
{
TR::Node *child = node->getChild(i);
if (child->getFutureUseCount() == 1 &&
child->getReferenceCount() > 1 &&
!child->getOpCode().isLoadConst())
*numDeadSubNodes = (*numDeadSubNodes) + 1;
collectSymbolReferencesInNode(child, symbolReferencesInNode, numDeadSubNodes, visitCount, comp,
seenInternalPointer, seenArraylet, cantMoveUnderBranch);
}
// detect if this is a direct load that shouldn't be moved under a branch (because an update was moved past
// this load by treeSimplification)
if (cantMoveUnderBranch &&
(node->getOpCode().isLoadVarDirect() || node->getOpCode().isLoadReg()) &&
node->isDontMoveUnderBranch())
*cantMoveUnderBranch = true;
if (seenInternalPointer && node->isInternalPointer() && node->getReferenceCount() > 1)
*seenInternalPointer = true;
if (seenArraylet)
{
if (node->getOpCode().hasSymbolReference() &&
node->getSymbolReference()->getSymbol()->isArrayletShadowSymbol() &&
node->getReferenceCount() > 1)
{
*seenArraylet = true;
}
}
// Add this node's symbol reference to the set
if (node->getOpCode().hasSymbolReference())
{
symbolReferencesInNode[node->getSymbolReference()->getReferenceNumber()]=true;
}
return true;
}
//returns true if there is first reference of a call or check
bool TR_LocalLiveRangeReduction::containsCallOrCheck(TR_TreeRefInfo *treeRefInfo, TR::Node *node)
{
if ((node->getOpCode().isCall() &&
(node->getReferenceCount()==1 || treeRefInfo->getFirstRefNodesList()->find(node))) ||
node->getOpCode().isCheck())
{
return true;
}
for (int32_t i = 0; i < node->getNumChildren(); i++)
{
TR::Node *child = node->getChild(i);
if (child->getReferenceCount()==1 || treeRefInfo->getFirstRefNodesList()->find(child))
return containsCallOrCheck(treeRefInfo, child);
}
return false;
}
void
OMR::CodeGenerator::evaluateChildrenWithMultipleRefCount(TR::Node * node)
{
for (int i=0; i < node->getNumChildren(); i++)
{
TR::Node *child = node->getChild(i);
if (child->getRegister() == NULL) // not already evaluated
{
// Note: we assume things without a symbol reference don't
// necessarily need to be evaluated here, and can wait
// until they are actually needed.
//
// vft pointers are speical - we need to evaluate the object in all cases
// but for nopable virtual guards we can wait to load and mask the pointer
// until we actually need to use it
//
if (child->getReferenceCount() > 1 &&
(child->getOpCode().hasSymbolReference() ||
(child->getOpCodeValue() == TR::l2a && child->getChild(0)->containsCompressionSequence())))
{
TR::SymbolReference *vftPointerSymRef = TR::comp()->getSymRefTab()->element(TR::SymbolReferenceTable::vftSymbol);
if (node->isNopableInlineGuard()
&& self()->getSupportsVirtualGuardNOPing()
&& child->getOpCodeValue() == TR::aloadi
&& child->getChild(0)->getOpCode().hasSymbolReference()
&& child->getChild(0)->getSymbolReference() == vftPointerSymRef
&& child->getChild(0)->getOpCodeValue() == TR::aloadi)
{
if (!child->getChild(0)->getChild(0)->getRegister() &&
child->getChild(0)->getChild(0)->getReferenceCount() > 1)
self()->evaluate(child->getChild(0)->getChild(0));
else
self()->evaluateChildrenWithMultipleRefCount(child->getChild(0)->getChild(0));
}
else
{
self()->evaluate(child);
}
}
else
{
self()->evaluateChildrenWithMultipleRefCount(child);
}
}
}
}
// TODO:AMD64: Could this be combined with istoreEvaluator without too much ugliness?
TR::Register *OMR::X86::AMD64::TreeEvaluator::lstoreEvaluator(TR::Node *node, TR::CodeGenerator *cg)
{
TR::Node *valueChild;
TR::Compilation* comp = cg->comp();
if (node->getOpCode().isIndirect())
valueChild = node->getSecondChild();
else
valueChild = node->getFirstChild();
// Handle special cases
//
if (valueChild->getRegister() == NULL &&
valueChild->getReferenceCount() == 1)
{
// Special case storing a double value into long variable
//
if (valueChild->getOpCodeValue() == TR::dbits2l &&
!valueChild->normalizeNanValues())
{
if (node->getOpCode().isIndirect())
{
node->setChild(1, valueChild->getFirstChild());
TR::Node::recreate(node, TR::dstorei);
TR::TreeEvaluator::floatingPointStoreEvaluator(node, cg);
node->setChild(1, valueChild);
TR::Node::recreate(node, TR::lstorei);
}
else
{
node->setChild(0, valueChild->getFirstChild());
TR::Node::recreate(node, TR::dstore);
TR::TreeEvaluator::floatingPointStoreEvaluator(node, cg);
node->setChild(0, valueChild);
TR::Node::recreate(node, TR::lstore);
}
cg->decReferenceCount(valueChild);
return NULL;
}
}
return TR::TreeEvaluator::integerStoreEvaluator(node, cg);
}
TR::Register *OMR::X86::AMD64::TreeEvaluator::l2iEvaluator(TR::Node *node, TR::CodeGenerator *cg)
{
TR::Node *child = node->getFirstChild();
TR::Register *reg = cg->evaluate(child);
if (child->getReferenceCount() > 1)
{
// This catches two scenarios:
//
// 1) A longClobberEvaluate (or any other register-clobbering logic) on
// the l2i node could see a refcount of 1, and hence won't make a copy.
// If child's refcount is more than 1, we do in fact need a copy, so we'd
// better do it here.
//
// 2) If the child is commoned, and the l2i node is also commoned, then
// we may end up with a situation where the last evaluation of the child
// is a clobberEvaluate. By that time, the child's refcount would be 1,
// so no copy is made, and the register would be clobbered. Therefore,
// the l2i node can't return that same register, or else the other uses
// of the node will end up getting the clobbered value.
//
// Note that case 2 is conservative, in that it presumes that the child's
// register will be clobbered by another node. If this does not occur,
// then the copy we're about to make is unnecessary.
//
TR::Register *childReg = reg;
reg = cg->allocateRegister();
// to support signExtension in GRA, need to preserve upper word
// in this move
generateRegRegInstruction(MOV8RegReg, node, reg, childReg, cg);
}
node->setRegister(reg);
cg->decReferenceCount(child);
if (cg->enableRegisterInterferences() && node->getOpCode().getSize() == 1)
cg->getLiveRegisters(TR_GPR)->setByteRegisterAssociation(node->getRegister());
return reg;
}
/**
* In pass_1(validateRefCountPass1), the Local Index (which is set to the Ref
* Count) for each child is decremented for each visit. The second pass is to
* make sure that the Local Index is zero by the end of the block. A non-zero
* Local Index would indicate that the Ref count was wrong at the start
* of the Validation Process.
*/
void TR::ValidateNodeRefCountWithinBlock::validateRefCountPass1(TR::Node *node)
{
/* If this is the first time through this node, verify the children. */
if (!_nodeChecklist.isSet(node->getGlobalIndex()))
{
_nodeChecklist.set(node->getGlobalIndex());
for (int32_t i = node->getNumChildren() - 1; i >= 0; --i)
{
TR::Node *child = node->getChild(i);
if (_nodeChecklist.isSet(child->getGlobalIndex()))
{
/* If the child has already been visited, decrement its verifyRefCount. */
child->decLocalIndex();
}
else
{
/* If the child has not yet been visited, set its localIndex and visit it. */
child->setLocalIndex(child->getReferenceCount() - 1);
validateRefCountPass1(child);
}
}
}
}
TR::Register *TR_X86FPCompareAnalyser::fpCompareAnalyser(TR::Node *root,
TR_X86OpCodes cmpRegRegOpCode,
TR_X86OpCodes cmpRegMemOpCode,
TR_X86OpCodes cmpiRegRegOpCode,
bool useFCOMIInstructions)
{
TR::Node *firstChild,
*secondChild;
TR::ILOpCodes cmpOp = root->getOpCodeValue();
bool reverseMemOp = false;
bool reverseCmpOp = false;
TR::Compilation* comp = _cg->comp();
TR_X86OpCodes cmpInstr = useFCOMIInstructions ? cmpiRegRegOpCode : cmpRegRegOpCode;
// Some operators must have their operands swapped to improve the generated
// code needed to evaluate the result of the comparison.
//
bool mustSwapOperands = (cmpOp == TR::iffcmple ||
cmpOp == TR::ifdcmple ||
cmpOp == TR::iffcmpgtu ||
cmpOp == TR::ifdcmpgtu ||
cmpOp == TR::fcmple ||
cmpOp == TR::dcmple ||
cmpOp == TR::fcmpgtu ||
cmpOp == TR::dcmpgtu ||
(useFCOMIInstructions &&
(cmpOp == TR::iffcmplt ||
cmpOp == TR::ifdcmplt ||
cmpOp == TR::iffcmpgeu ||
cmpOp == TR::ifdcmpgeu ||
cmpOp == TR::fcmplt ||
cmpOp == TR::dcmplt ||
cmpOp == TR::fcmpgeu ||
cmpOp == TR::dcmpgeu))) ? true : false;
// Some operators should not have their operands swapped to improve the generated
// code needed to evaluate the result of the comparison.
//
bool preventOperandSwapping = (cmpOp == TR::iffcmpltu ||
cmpOp == TR::ifdcmpltu ||
cmpOp == TR::iffcmpge ||
cmpOp == TR::ifdcmpge ||
cmpOp == TR::fcmpltu ||
cmpOp == TR::dcmpltu ||
cmpOp == TR::fcmpge ||
cmpOp == TR::dcmpge ||
(useFCOMIInstructions &&
(cmpOp == TR::iffcmpgt ||
cmpOp == TR::ifdcmpgt ||
cmpOp == TR::iffcmpleu ||
cmpOp == TR::ifdcmpleu ||
cmpOp == TR::fcmpgt ||
cmpOp == TR::dcmpgt ||
cmpOp == TR::fcmpleu ||
cmpOp == TR::dcmpleu))) ? true : false;
// For correctness, don't swap operands of these operators.
//
if (cmpOp == TR::fcmpg || cmpOp == TR::fcmpl ||
cmpOp == TR::dcmpg || cmpOp == TR::dcmpl)
{
preventOperandSwapping = true;
}
// Initial operand evaluation ordering.
//
if (preventOperandSwapping || (!mustSwapOperands && _cg->whichChildToEvaluate(root) == 0))
{
firstChild = root->getFirstChild();
secondChild = root->getSecondChild();
setReversedOperands(false);
}
else
{
firstChild = root->getSecondChild();
secondChild = root->getFirstChild();
setReversedOperands(true);
}
TR::Register *firstRegister = firstChild->getRegister();
TR::Register *secondRegister = secondChild->getRegister();
setInputs(firstChild,
firstRegister,
secondChild,
secondRegister,
useFCOMIInstructions,
// If either 'preventOperandSwapping' or 'mustSwapOperands' is set then the
// initial operand ordering set above must be maintained.
//
preventOperandSwapping || mustSwapOperands);
// Make sure any required operand ordering is respected.
//
if ((getCmpReg2Reg1() || getCmpReg2Mem1()) &&
(mustSwapOperands || preventOperandSwapping))
{
reverseCmpOp = getCmpReg2Reg1() ? true : false;
reverseMemOp = getCmpReg2Mem1() ? true : false;
}
// If we are not comparing with a memory operand, one of them evaluates
// to a zero, and the zero is not already on the stack, then we can use
// FTST to save a register.
//
// (With a memory operand, either the constant zero needs to be loaded
// to use FCOM, or the memory operand needs to be loaded to use FTST,
// so there is no gain in using FTST.)
//
// If the constant zero is in the target register, using FTST means the
// comparison will be reversed. We cannot do this if the initial ordering
// of the operands must be maintained.
//
// Finally, if FTST is used and this is the last use of the target, the
// target register may need to be explicitly popped.
//
TR::Register *targetRegisterForFTST = NULL;
TR::Node *targetChildForFTST = NULL;
if (getEvalChild1() && isUnevaluatedZero(firstChild)) // do we need getEvalChild1() here?
{
if ( ((getCmpReg1Reg2() || reverseCmpOp) && !(preventOperandSwapping || mustSwapOperands)) ||
(getCmpReg2Reg1() && !reverseCmpOp))
{
if (getEvalChild2())
{
secondRegister = _cg->evaluate(secondChild);
}
targetRegisterForFTST = secondRegister;
targetChildForFTST = secondChild;
notReversedOperands();
}
}
else if (getEvalChild2() && isUnevaluatedZero(secondChild)) // do we need getEvalChild2() here?
{
if ( (getCmpReg1Reg2() || reverseCmpOp) ||
(getCmpReg2Reg1() && !reverseCmpOp && !(preventOperandSwapping || mustSwapOperands)) )
{
if (getEvalChild1())
{
firstRegister = _cg->evaluate(firstChild);
}
targetRegisterForFTST = firstRegister;
targetChildForFTST = firstChild;
}
}
if (!targetRegisterForFTST)
{
// If we have a choice, evaluate the target operand last. By doing so, we
// help out the register assigner because the target must be TOS. This
// avoids an unneccessary FXCH for the target.
//
if (getEvalChild1() && getEvalChild2())
{
if (getCmpReg1Reg2() || getCmpReg1Mem2())
{
secondRegister = _cg->evaluate(secondChild);
firstRegister = _cg->evaluate(firstChild);
}
else
{
firstRegister = _cg->evaluate(firstChild);
secondRegister = _cg->evaluate(secondChild);
}
}
else
{
if (getEvalChild1())
{
firstRegister = _cg->evaluate(firstChild);
}
if (getEvalChild2())
{
secondRegister = _cg->evaluate(secondChild);
}
}
}
// Adjust the FP precision of feeding operands.
//
if (firstRegister &&
(firstRegister->needsPrecisionAdjustment() ||
comp->getOption(TR_StrictFPCompares) ||
(firstRegister->mayNeedPrecisionAdjustment() && secondChild->getOpCode().isLoadConst()) ||
(firstRegister->mayNeedPrecisionAdjustment() && !secondRegister)))
{
TR::TreeEvaluator::insertPrecisionAdjustment(firstRegister, root, _cg);
}
if (secondRegister &&
(secondRegister->needsPrecisionAdjustment() ||
comp->getOption(TR_StrictFPCompares) ||
(secondRegister->mayNeedPrecisionAdjustment() && firstChild->getOpCode().isLoadConst()) ||
(secondRegister->mayNeedPrecisionAdjustment() && !firstRegister)))
{
TR::TreeEvaluator::insertPrecisionAdjustment(secondRegister, root, _cg);
}
// Generate the compare instruction.
//
if (targetRegisterForFTST)
{
generateFPRegInstruction(FTSTReg, root, targetRegisterForFTST, _cg);
}
else if (!useFCOMIInstructions && (getCmpReg1Mem2() || reverseMemOp))
{
TR::MemoryReference *tempMR = generateX86MemoryReference(secondChild, _cg);
generateFPRegMemInstruction(cmpRegMemOpCode, root, firstRegister, tempMR, _cg);
tempMR->decNodeReferenceCounts(_cg);
}
else if (!useFCOMIInstructions && getCmpReg2Mem1())
{
TR::MemoryReference *tempMR = generateX86MemoryReference(firstChild, _cg);
generateFPRegMemInstruction(cmpRegMemOpCode, root, secondRegister, tempMR, _cg);
notReversedOperands();
tempMR->decNodeReferenceCounts(_cg);
}
else if (getCmpReg1Reg2() || reverseCmpOp)
{
generateFPCompareRegRegInstruction(cmpInstr, root, firstRegister, secondRegister, _cg);
}
else if (getCmpReg2Reg1())
{
generateFPCompareRegRegInstruction(cmpInstr, root, secondRegister, firstRegister, _cg);
notReversedOperands();
}
_cg->decReferenceCount(firstChild);
_cg->decReferenceCount(secondChild);
// Evaluate the comparison.
//
if (getReversedOperands())
{
cmpOp = TR::ILOpCode(cmpOp).getOpCodeForSwapChildren();
TR::Node::recreate(root, cmpOp);
}
if (useFCOMIInstructions && !targetRegisterForFTST)
{
return NULL;
}
// We must manually move the FP condition flags to the EFLAGS register if we don't
// use the FCOMI instructions.
//
TR::Register *accRegister = _cg->allocateRegister();
TR::RegisterDependencyConditions *dependencies = generateRegisterDependencyConditions((uint8_t)1, 1, _cg);
dependencies->addPreCondition(accRegister, TR::RealRegister::eax, _cg);
dependencies->addPostCondition(accRegister, TR::RealRegister::eax, _cg);
generateRegInstruction(STSWAcc, root, accRegister, dependencies, _cg);
// Pop the FTST target register if it is not used any more.
//
if (targetRegisterForFTST &&
targetChildForFTST && targetChildForFTST->getReferenceCount() == 0)
{
generateFPSTiST0RegRegInstruction(FSTRegReg, root, targetRegisterForFTST, targetRegisterForFTST, _cg);
}
return accRegister;
}
int32_t
OMR::X86::I386::CodeGenerator::getMaximumNumberOfGPRsAllowedAcrossEdge(TR::Node *node)
{
// TODO: Currently, lookupEvaluator doesn't deal properly with different
// glRegDeps on different cases of a lookupswitch.
//
static const char *enableLookupswitch = feGetEnv("TR_enableGRAAcrossLookupSwitch");
if (!enableLookupswitch && node->getOpCode().getOpCodeValue()==TR::lookup)
return 1;
if (node->getOpCode().getOpCodeValue()==TR::table)
{
// 1 for jump table base reg, which is not apparent in the trees
// 1 for ebp when it is needed for the VMThread
//
return self()->getNumberOfGlobalGPRs() - 2;
}
if (node->getOpCode().isIf())
{
// we run out of all but one/two registers in these cases
//
if (node->getFirstChild()->getType().isInt64())
{
if (node->getOpCode().isBranch())
{
TR::Node *firstChild = node->getFirstChild();
TR::Node *secondChild = node->getSecondChild();
int extraRegsAvailable = 0;
if(firstChild->getOpCodeValue() == TR::d2l ||
secondChild->getOpCodeValue() == TR::d2l)
{
return 1;
}
if ((firstChild->getReferenceCount() == 1 &&
firstChild->getOpCode().isLoadVarDirect()) ||
(secondChild->getReferenceCount() == 1 &&
firstChild->getOpCode().isLoadVarDirect()))
extraRegsAvailable += 0; // TODO: put it back to 2 when looking at GRA, GRA pushes allocation of 8 registers
return 2 + extraRegsAvailable;
}
else
{
// TR_lcmpXX opcodes take up 5 regs
//
return 1;
}
}
// we run out of all but one register in these cases....last time I tried....
//
if (node->getFirstChild()->getOpCodeValue() == TR::instanceof)
{
if (!TR::TreeEvaluator::instanceOfOrCheckCastNeedSuperTest(node->getFirstChild(), self()) &&
TR::TreeEvaluator::instanceOfOrCheckCastNeedEqualityTest(node->getFirstChild(), self()))
return self()->getNumberOfGlobalGPRs() - 4; // ebp plus three other regs if vft masking is enabled
else
return 0;
}
// All other conditional branches, we usually need one reg for the compare and possibly one for the vmthread
//return getNumberOfGlobalGPRs() - 1 - (node->isVMThreadRequired()? 1 : 0);
// vmThread required might be set on a node after GRA has ran
return self()->getNumberOfGlobalGPRs() - 2;
}
return INT_MAX;
}
TR::Register *
OMR::CodeGenerator::evaluate(TR::Node * node)
{
TR::Register *reg;
bool trace = self()->comp()->getOptions()->getTraceCGOption(TR_TraceCGEvaluation);
TR::ILOpCodes opcode = node->getOpCodeValue();
TR_ASSERT(!self()->comp()->getOption(TR_EnableParanoidRefCountChecks) || node->getOpCode().isTreeTop() || node->getReferenceCount() > 0,
"OMR::CodeGenerator::evaluate invoked for nontreetop node [%s] with count == 0", node->getName(self()->comp()->getDebug()));
if (opcode != TR::BBStart && node->getRegister())
{
reg = node->getRegister();
if (trace)
{
self()->getDebug()->printNodeEvaluation(node, ": ", reg);
}
}
else
{
if (trace)
{
self()->getDebug()->printNodeEvaluation(node);
_indentation += 2;
}
// Evaluation of a TR IL tree can be performed by many functions:
//
// 1) evaluate(...)
// 2) populateMemoryReference(...)
// 3) populateAddTree(...)
// ...
//
// However all of these functions can be categorized into two classes:
//
// A) functions which completely evaluate their subtree.
// B) functions which partially evaluate their subtree.
//
// Because functions of class A and class B can be used interchangably to
// perform a recursive decent of a TR IL tree, and because A or B functions
// can perform a destructive evaluation of their subtree, a bug can occur where
// the results of a partial evalutation are destructively overwritten before
// being completely evalutated.
//
// Ex: the motivating case is the following evaluation pattern:
//
// node_A evaluate
// node_B populateMemoryReference
// node_C evaluate
//
// where
//
// a) there is a common node between the subtrees of node_B and node_C.
// b) calling populateMemoryReference on node_B reduces the reference count of one
// of the base or index nodes to 1, creating the "opportunity" for a destructive
// evaluation.
//
// The following chain of events occurs:
//
// 1) evaluate is called on node_A. This evaluator can produce instructions of RX form,
// and chooses to do so for this node.
//
// 2) populateMemoryReference is called on node_B, and evaluates the subtree, returning
// a TR_MemoryReference to node_A's evaluator. This memory reference has not been
// dereferenced yet, and the base (and optionally index) nodes may have registers
// assigned to them.
//
// 3) evaluate is called on node_C, which chooses to destructively evaluate the commoned
// base node. The memory reference's base register now contains a garbage value.
//
// 4) control passes to node_A's evaluator, which emits an RX instruction using node_B's
// memory reference and node_C's register.
//
// In the past, the fix for this was to switch the order of evaluation: call evaluate
// on node_C and then call populateMemoryReference on node_B. This fixes this scenario, but
// the capability of another tree and evaluation pattern to create this bug still exists.
//
// As well, more insidious trees exist:
//
// ificmpeq
// iiload
// i2l
// x
// iiload
// ishl
// ==> i2l
// 4
//
// The evaluation pattern here could be:
//
// evaluate
// populateMemoryReference
// evaluate
// populateMemoryReference
// evaluate
//
// If the commoned node's reference count is 2 coming into ificmpeq's evaluator, then
// the second sub-evaluate call could be destructive to the first populateMemoryReference.
//
// Even worse, if either subtree could be destructive, then there would be no correct order to
// perform the function calls:
//
// ificmpeq
// iiload
// ishl
// ==> x
// 7
// iiload
// ishl
// ==> x
// 4
//
// Generally, two conditions must be true for this bug to be possible:
//
// 1) the following two classes of recursive decent functions must exist:
// A) functions which completely evaluate their subtree.
// B) functions which partially evaluate their subtree.
//
// 2) destructive evaluation by either class of function must be possible.
//
// This code implements changes to eliminate the second condition for this bug by performing
// the following check and fixup:
//
// If in a function which partially evaluates its subtree, note all non-restricted nodes that
// have a reference count > 1. If any of those node's reference counts reach 1, then artificially
// inflate those reference counts by 1 for the lifetime of the parent evaluation.
//
int32_t topOfNodeStackBeforeEvaluation = _stackOfArtificiallyInflatedNodes.topIndex();
// Memory references are not like registers, and should not be allowed to escape their evaluator.
// Explicitly note memory references that are not loaded into registers and automatically call
// stopUsingMemRefRegister on all memory references that have "escaped".
//
// Only the s390 memory references are tracked in this way.
//
int32_t topOfMemRefStackBeforeEvaluation = _stackOfMemoryReferencesCreatedDuringEvaluation.topIndex();
reg = _nodeToInstrEvaluators[opcode](node, self());
if (self()->comp()->getOptions()->getTraceCGOption(TR_TraceCGEvaluation))
{
self()->getDebug()->printNodeEvaluation(node, "<- ", reg, false);
_indentation -= 2;
}
if (self()->comp()->getOption(TR_TraceRegisterPressureDetails))
{
traceMsg(self()->comp(), " evaluated %s", self()->getDebug()->getName(node));
self()->getDebug()->dumpLiveRegisters();
traceMsg(self()->comp(), "\n");
}
// Pop off and decrement tracked nodes
//
while (_stackOfArtificiallyInflatedNodes.topIndex() > topOfNodeStackBeforeEvaluation)
{
TR::Node * artificiallyInflatedNode = _stackOfArtificiallyInflatedNodes.pop();
if (artificiallyInflatedNode->getReferenceCount() == 1)
{
// When inflating reference counts, two cases exist:
//
// 1) N's ref count reaches 1 in a populate* call, which is then inc'ed to 2.
//
// 1a) N is never evaluated, so the ref count never goes down to 1. (node was not commoned in another subtree)
//
// - no tree difference should be seen in this case.
//
// 1b) N is evaluated, so the ref count then goes down to 1. (node was commoned in another subtree)
//
// - register shuffling _could_ be seen in this case.
// - but a bug might have been avoided: partial and complete evaluation of a commoned node occurred.
//
if (self()->comp()->getOption(TR_TraceCG))
{
self()->comp()->getDebug()->trace(" _stackOfArtificiallyInflatedNodes.pop(): node %p part of commoned case, might have avoided a bug!\n", artificiallyInflatedNode);
}
}
self()->decReferenceCount(artificiallyInflatedNode);
#ifdef J9_PROJECT_SPECIFIC
#if defined(TR_TARGET_S390)
if (artificiallyInflatedNode->getOpaquePseudoRegister())
{
TR_OpaquePseudoRegister *reg = artificiallyInflatedNode->getOpaquePseudoRegister();
TR_StorageReference *ref = reg->getStorageReference();
self()->processUnusedStorageRef(ref);
}
#endif
#endif
if (self()->comp()->getOption(TR_TraceCG))
{
self()->comp()->getDebug()->trace(" _stackOfArtificiallyInflatedNodes.pop() %p, decReferenceCount(...) called. reg=%s\n", artificiallyInflatedNode,
artificiallyInflatedNode->getRegister()?artificiallyInflatedNode->getRegister()->getRegisterName(self()->comp()):"null");
}
}
#if defined(TR_TARGET_S390)
self()->StopUsingEscapedMemRefsRegisters(topOfMemRefStackBeforeEvaluation);
#endif
bool checkRefCount = (node->getReferenceCount() <= 1 ||
(reg && reg == node->getRegister()));
// for anchor mode, if node is an indirect store, it can have
// ref count <= 2
// but for compressedRefs, the indirect store must be an address
if (self()->comp()->useAnchors())
{
if (((node->getOpCode().isStoreIndirect() &&
(self()->comp()->useCompressedPointers() && (node->getSymbolReference()->getSymbol()->getDataType() == TR::Address))) ||
opcode == TR::wrtbari) &&
node->getReferenceCount() <= 2 &&
!checkRefCount)
checkRefCount = true;
}
TR_ASSERT(checkRefCount,
"evaluate: the node's register wasn't set (node [%s])", node->getName(self()->comp()->getDebug()));
}
return reg;
}
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;
}
}
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
}
// Returns true if there is any constraint to the move
bool TR_LocalLiveRangeReduction::isAnySymInDefinedOrUsedBy(TR_TreeRefInfo *currentTreeRefInfo, TR::Node *currentNode, TR_TreeRefInfo *movingTreeRefInfo )
{
TR::Node *movingNode = movingTreeRefInfo->getTreeTop()->getNode();
// ignore anchors
//
if (movingNode->getOpCode().isAnchor())
movingNode = movingNode->getFirstChild();
TR::ILOpCode &opCode = currentNode->getOpCode();
////if ((opCode.getOpCodeValue() == TR::monent) || (opCode.getOpCodeValue() == TR::monexit))
if (nodeMaybeMonitor(currentNode))
{
if (trace())
traceMsg(comp(),"cannot move %p beyond monitor %p\n",movingNode,currentNode);
return true;
}
// Don't move gc points or things across gc points
//
if (movingNode->canGCandReturn() ||
currentNode->canGCandReturn())
{
if (trace())
traceMsg(comp(), "cannot move gc points %p past %p\n", movingNode, currentNode);
return true;
}
// Don't move checks or calls at all
//
if (containsCallOrCheck(movingTreeRefInfo,movingNode))
{
if (trace())
traceMsg(comp(),"cannot move check or call %s\n", getDebug()->getName(movingNode));
return true;
}
// Don't move object header store past a GC point
//
if ((currentNode->getOpCode().isWrtBar() || currentNode->canCauseGC()) && mayBeObjectHeaderStore(movingNode, fe()))
{
if (trace())
traceMsg(comp(),"cannot move possible object header store %s past GC point %s\n", getDebug()->getName(movingNode), getDebug()->getName(currentNode));
return true;
}
if (TR::Compiler->target.cpu.isPower() && opCode.getOpCodeValue() == TR::allocationFence)
{
// Can't move allocations past flushes
if (movingNode->getOpCodeValue() == TR::treetop &&
movingNode->getFirstChild()->getOpCode().isNew() &&
(currentNode->getAllocation() == NULL ||
currentNode->getAllocation() == movingNode->getFirstChild()))
{
if (trace())
{
traceMsg(comp(),"cannot move %p beyond flush %p - ", movingNode, currentNode);
if (currentNode->getAllocation() == NULL)
traceMsg(comp(),"(flush with null allocation)\n");
else
traceMsg(comp(),"(flush for allocation %p)\n", currentNode->getAllocation());
}
return true;
}
// Can't move certain stores past flushes
// Exclude all indirect stores, they may be for stack allocs, in which case the flush is needed at least as a scheduling barrier
// Direct stores to autos and parms are the only safe candidates
if (movingNode->getOpCode().isStoreIndirect() ||
(movingNode->getOpCode().isStoreDirect() && !movingNode->getSymbol()->isParm() && !movingNode->getSymbol()->isAuto()))
{
if (trace())
traceMsg(comp(),"cannot move %p beyond flush %p - (flush for possible stack alloc)", movingNode, currentNode);
return true;
}
}
for (int32_t i = 0; i < currentNode->getNumChildren(); i++)
{
TR::Node *child = currentNode->getChild(i);
//Any node that has side effects (like call and newarrya) cannot be evaluated in the middle of the tree.
if (movingTreeRefInfo->getFirstRefNodesList()->find(child))
{
//for calls and unresolve symbol that are not under check
if (child->exceptionsRaised() ||
(child->getOpCode().hasSymbolReference() && child->getSymbolReference()->isUnresolved()))
{
if (trace())
traceMsg(comp(),"cannot move %p beyond %p - cannot change evaluation point of %p\n ",movingNode,currentTreeRefInfo->getTreeTop()->getNode(),child);
return true;
}
else if(movingNode->getOpCode().isStore())
{
TR::SymbolReference *stSymRef = movingNode->getSymbolReference();
int32_t stSymRefNum = stSymRef->getReferenceNumber();
//TR::SymbolReference *stSymRef = movingNode->getSymbolReference();
int32_t numHelperSymbols = comp()->getSymRefTab()->getNumHelperSymbols();
if ((comp()->getSymRefTab()->isNonHelper(stSymRefNum, TR::SymbolReferenceTable::vftSymbol))||
(comp()->getSymRefTab()->isNonHelper(stSymRefNum, TR::SymbolReferenceTable::contiguousArraySizeSymbol))||
(comp()->getSymRefTab()->isNonHelper(stSymRefNum, TR::SymbolReferenceTable::discontiguousArraySizeSymbol))||
(stSymRef == comp()->getSymRefTab()->findHeaderFlagsSymbolRef())||
(stSymRef->getSymbol() == comp()->getSymRefTab()->findGenericIntShadowSymbol()))
return true;
}
else if (movingNode->getOpCode().isResolveOrNullCheck())
{
if (trace())
traceMsg(comp(),"cannot move %p beyond %p - node %p under ResolveOrNullCheck",movingNode,currentTreeRefInfo->getTreeTop()->getNode(),currentNode);
return true;
}
else if (TR::Compiler->target.is64Bit() &&
movingNode->getOpCode().isBndCheck() &&
((opCode.getOpCodeValue() == TR::i2l) || (opCode.getOpCodeValue() == TR::iu2l)) &&
!child->isNonNegative())
{
if (trace())
traceMsg(comp(),"cannot move %p beyond %p - changing the eval point of %p will casue extra cg instruction ",movingNode,currentTreeRefInfo->getTreeTop()->getNode(),currentNode);
return true;
}
}
//don't recurse over nodes each are not the first reference
if (child->getReferenceCount()==1 || currentTreeRefInfo->getFirstRefNodesList()->find(child))
{
if (isAnySymInDefinedOrUsedBy(currentTreeRefInfo, child, movingTreeRefInfo ))
return true;
}
}
return false;
}
void TR_LocalLiveRangeReduction::populatePotentialDeps(TR_TreeRefInfo *treeRefInfo,TR::Node *node)
{
TR::ILOpCode &opCode = node->getOpCode();
if (node->getOpCode().hasSymbolReference())
{
TR::SymbolReference *symRef = node->getSymbolReference();
int32_t symRefNum = symRef->getReferenceNumber();
//set defSym - all symbols that might be written
if (opCode.isCall() || opCode.isResolveCheck()|| opCode.isStore() || node->mightHaveVolatileSymbolReference())
{
bool isCallDirect = false;
if (node->getOpCode().isCallDirect())
isCallDirect = true;
if (!symRef->getUseDefAliases(isCallDirect).isZero(comp()))
{
TR::SparseBitVector useDefAliases(comp()->allocator());
symRef->getUseDefAliases(isCallDirect).getAliases(useDefAliases);
TR::SparseBitVector::Cursor aliasCursor(useDefAliases);
for (aliasCursor.SetToFirstOne(); aliasCursor.Valid(); aliasCursor.SetToNextOne())
{
int32_t nextAlias = aliasCursor;
treeRefInfo->getDefSym()->set(nextAlias);
}
}
if (opCode.isStore())
treeRefInfo->getDefSym()->set(symRefNum);
}
//set useSym - all symbols that are used
if (opCode.canRaiseException())
{
TR::SparseBitVector useAliases(comp()->allocator());
symRef->getUseonlyAliases().getAliases(useAliases);
{
TR::SparseBitVector::Cursor aliasesCursor(useAliases);
for (aliasesCursor.SetToFirstOne(); aliasesCursor.Valid(); aliasesCursor.SetToNextOne())
{
int32_t nextAlias = aliasesCursor;
treeRefInfo->getUseSym()->set(nextAlias);
}
}
}
if (opCode.isLoadVar() || (opCode.getOpCodeValue() == TR::loadaddr))
{
treeRefInfo->getUseSym()->set(symRefNum);
}
}
for (int32_t i = 0; i < node->getNumChildren(); i++)
{
TR::Node *child = node->getChild(i);
//don't recurse over references (nodes which are not the first reference)
//
if (child->getReferenceCount()==1 || treeRefInfo->getFirstRefNodesList()->find(child))
populatePotentialDeps(treeRefInfo,child );
}
return;
}
// Build arguments for system linkage dispatch.
//
int32_t TR::AMD64SystemLinkage::buildArgs(
TR::Node *callNode,
TR::RegisterDependencyConditions *deps)
{
TR::SymbolReference *methodSymRef = callNode->getSymbolReference();
TR::MethodSymbol *methodSymbol = methodSymRef->getSymbol()->castToMethodSymbol();
TR::RealRegister::RegNum noReg = TR::RealRegister::NoReg;
TR::RealRegister *espReal = machine()->getX86RealRegister(TR::RealRegister::esp);
int32_t firstNodeArgument = callNode->getFirstArgumentIndex();
int32_t lastNodeArgument = callNode->getNumChildren() - 1;
int32_t offset = 0;
int32_t sizeOfOutGoingArgs= 0;
uint16_t numIntArgs = 0,
numFloatArgs = 0;
int32_t first, last, direction;
int32_t numCopiedRegs = 0;
TR::Register *copiedRegs[TR::X86LinkageProperties::MaxArgumentRegisters];
if (getProperties().passArgsRightToLeft())
{
first = lastNodeArgument;
last = firstNodeArgument - 1;
direction = -1;
}
else
{
first = firstNodeArgument;
last = lastNodeArgument + 1;
direction = 1;
}
// If the dispatch is indirect we must add the VFT register to the preconditions
// so that it gets register assigned with the other preconditions to the call.
//
if (callNode->getOpCode().isIndirect())
{
TR::Node *vftChild = callNode->getFirstChild();
TR_ASSERT(vftChild->getRegister(), "expecting VFT child to be evaluated");
TR::RealRegister::RegNum scratchRegIndex = getProperties().getIntegerScratchRegister(1);
deps->addPreCondition(vftChild->getRegister(), scratchRegIndex, cg());
}
int32_t i;
for (i = first; i != last; i += direction)
{
TR::parmLayoutResult layoutResult;
TR::RealRegister::RegNum rregIndex = noReg;
TR::Node *child = callNode->getChild(i);
layoutParm(child, sizeOfOutGoingArgs, numIntArgs, numFloatArgs, layoutResult);
if (layoutResult.abstract & TR::parmLayoutResult::IN_LINKAGE_REG_PAIR)
{
// TODO: AMD64 SysV ABI might put a struct into a pair of linkage registerr
TR_ASSERT(false, "haven't support linkage_reg_pair yet.\n");
}
else if (layoutResult.abstract & TR::parmLayoutResult::IN_LINKAGE_REG)
{
TR_RegisterKinds regKind = layoutResult.regs[0].regKind;
uint32_t regIndex = layoutResult.regs[0].regIndex;
TR_ASSERT(regKind == TR_GPR || regKind == TR_FPR, "linkage registers includes TR_GPR and TR_FPR\n");
rregIndex = (regKind == TR_FPR) ? getProperties().getFloatArgumentRegister(regIndex): getProperties().getIntegerArgumentRegister(regIndex);
}
else
{
offset = layoutResult.offset;
}
TR::Register *vreg;
vreg = cg()->evaluate(child);
bool needsStackOffsetUpdate = false;
if (rregIndex != noReg)
{
// For NULL JNI reference parameters, it is possible that the NULL value will be evaluated into
// a different register than the child. In that case it is not necessary to copy the temporary scratch
// register across the call.
//
if ((child->getReferenceCount() > 1) &&
(vreg == child->getRegister()))
{
TR::Register *argReg = cg()->allocateRegister();
if (vreg->containsCollectedReference())
argReg->setContainsCollectedReference();
generateRegRegInstruction(TR::Linkage::movOpcodes(RegReg, movType(child->getDataType())), child, argReg, vreg, cg());
vreg = argReg;
copiedRegs[numCopiedRegs++] = vreg;
}
deps->addPreCondition(vreg, rregIndex, cg());
}
else
{
// Ideally, we would like to push rather than move
generateMemRegInstruction(TR::Linkage::movOpcodes(MemReg, fullRegisterMovType(vreg)),
child,
generateX86MemoryReference(espReal, offset, cg()),
vreg,
cg());
}
cg()->decReferenceCount(child);
}
// Now that we're finished making the preconditions, all the interferences
// are established and we can kill these regs.
//
for (i = 0; i < numCopiedRegs; i++)
cg()->stopUsingRegister(copiedRegs[i]);
deps->stopAddingPreConditions();
return sizeOfOutGoingArgs;
}
TR::Register *
TR::AMD64SystemLinkage::buildVolatileAndReturnDependencies(
TR::Node *callNode,
TR::RegisterDependencyConditions *deps)
{
if (callNode->getOpCode().isIndirect())
{
TR::Node *vftChild = callNode->getFirstChild();
if (vftChild->getRegister() && (vftChild->getReferenceCount() > 1))
{
}
else
{
// VFT child dies here; decrement it early so it doesn't interfere with dummy regs.
cg()->recursivelyDecReferenceCount(vftChild);
}
}
TR_ASSERT(deps != NULL, "expected register dependencies");
// Figure out which is the return register.
//
TR::RealRegister::RegNum returnRegIndex;
TR_RegisterKinds returnKind;
switch (callNode->getDataType())
{
case TR::NoType:
returnRegIndex = TR::RealRegister::NoReg;
returnKind = TR_NoRegister;
break;
case TR::Int8:
case TR::Int16:
case TR::Int32:
case TR::Int64:
case TR::Address:
returnRegIndex = getProperties().getIntegerReturnRegister();
returnKind = TR_GPR;
break;
case TR::Float:
case TR::Double:
returnRegIndex = getProperties().getFloatReturnRegister();
returnKind = TR_FPR;
break;
case TR::Aggregate:
default:
TR_ASSERT(false, "Unrecognized call node data type: #%d", (int)callNode->getDataType());
break;
}
// Kill all non-preserved int and float regs besides the return register.
//
int32_t i;
TR::RealRegister::RegNum scratchIndex = getProperties().getIntegerScratchRegister(1);
for (i=0; i<getProperties().getNumVolatileRegisters(); i++)
{
TR::RealRegister::RegNum regIndex = getProperties()._volatileRegisters[i];
if (regIndex != returnRegIndex)
{
TR_RegisterKinds rk = (i < getProperties()._numberOfVolatileGPRegisters) ? TR_GPR : TR_FPR;
TR::Register *dummy = cg()->allocateRegister(rk);
deps->addPostCondition(dummy, regIndex, cg());
// Note that we don't setPlaceholderReg here. If this volatile reg is also volatile
// in the caller's linkage, then that flag doesn't matter much anyway. If it's preserved
// in the caller's linkage, then we don't want to set that flag because we want this
// use of the register to count as a "real" use, thereby motivating the prologue to
// preserve the register.
// A scratch register is necessary to call the native without a trampoline.
//
if (callNode->getOpCode().isIndirect() || (regIndex != scratchIndex))
cg()->stopUsingRegister(dummy);
}
}
#if defined (PYTHON) && 0
// Evict the preserved registers across the call
//
for (i=0; i<getProperties().getNumberOfPreservedGPRegisters(); i++)
{
TR::RealRegister::RegNum regIndex = getProperties()._preservedRegisters[i];
TR::Register *dummy = cg()->allocateRegister(TR_GPR);
deps->addPostCondition(dummy, regIndex, cg());
// Note that we don't setPlaceholderReg here. If this volatile reg is also volatile
// in the caller's linkage, then that flag doesn't matter much anyway. If it's preserved
// in the caller's linkage, then we don't want to set that flag because we want this
// use of the register to count as a "real" use, thereby motivating the prologue to
// preserve the register.
// A scratch register is necessary to call the native without a trampoline.
//
if (callNode->getOpCode().isIndirect() || (regIndex != scratchIndex))
cg()->stopUsingRegister(dummy);
}
#endif
if (callNode->getOpCode().isIndirect())
{
TR::Node *vftChild = callNode->getFirstChild();
if (vftChild->getRegister() && (vftChild->getReferenceCount() > 1))
{
// VFT child survives the call, so we must include it in the postconditions.
deps->addPostCondition(vftChild->getRegister(), TR::RealRegister::NoReg, cg());
cg()->recursivelyDecReferenceCount(vftChild);
}
}
// Now that everything is dead, we can allocate the return register without
// interference
//
TR::Register *returnRegister;
if (returnRegIndex)
{
TR_ASSERT(returnKind != TR_NoRegister, "assertion failure");
if (callNode->getDataType() == TR::Address)
returnRegister = cg()->allocateCollectedReferenceRegister();
else
{
returnRegister = cg()->allocateRegister(returnKind);
if (callNode->getDataType() == TR::Float)
returnRegister->setIsSinglePrecision();
}
deps->addPostCondition(returnRegister, returnRegIndex, cg());
}
else
returnRegister = NULL;
// The reg dependency is left open intentionally, and need to be closed by
// the caller. The reason is because, child class might call this method, while
// adding more register dependecies; if we close the reg dependency here,
// the child class won't be able to add more register dependencies.
return returnRegister;
}
void TR_ExpressionsSimplification::setSummationReductionCandidates(TR::Node *node, TR::TreeTop *tt)
{
// Must be a store node
//
if (node->getOpCodeValue() != TR::istore
/* || node->getOpCodeValue() != TR::astore */)
{
if (trace())
traceMsg(comp(), "Node %p: The opcode is not istore so not a summation reduction candidate\n",node);
return;
}
TR::Node *opNode = node->getFirstChild();
if (opNode->getOpCodeValue() == TR::iadd ||
opNode->getOpCodeValue() == TR::isub)
{
TR::Node *firstNode = opNode->getFirstChild();
TR::Node *secondNode = opNode->getSecondChild();
if (firstNode->getOpCode().hasSymbolReference() &&
node->getSymbolReference() == firstNode->getSymbolReference() &&
opNode->getReferenceCount() == 1 && firstNode->getReferenceCount() == 1)
{
// The second node must be loop invariant
//
if (!_currentRegion->isExprInvariant(secondNode))
{
if (trace())
{
traceMsg(comp(), "The node %p is not loop invariant\n",secondNode);
// This can be the arithmetic series case
// only when the node is an induction variable
if (secondNode->getNumChildren() == 1 && secondNode->getOpCode().hasSymbolReference())
{
TR_InductionVariable *indVar = _currentRegion->findMatchingIV(secondNode->getSymbolReference());
if (indVar)
{
//printf("Found Candidate of arithmetic series\n" );
}
}
}
return;
}
_candidateTTs->add(tt);
}
else if (secondNode->getOpCode().hasSymbolReference() &&
node->getSymbolReference() == secondNode->getSymbolReference() &&
opNode->getReferenceCount() == 1 && secondNode->getReferenceCount() == 1 &&
_currentRegion->isExprInvariant(firstNode))
{
_candidateTTs->add(tt);
}
}
else if (opNode->getOpCodeValue() == TR::ixor ||
opNode->getOpCodeValue() == TR::ineg)
{
if (opNode->getFirstChild()->getOpCode().hasSymbolReference() &&
node->getSymbolReference() == opNode->getFirstChild()->getSymbolReference() &&
opNode->getReferenceCount() == 1 && opNode->getFirstChild()->getReferenceCount() == 1 &&
(opNode->getOpCodeValue() == TR::ineg || _currentRegion->isExprInvariant(opNode->getSecondChild())))
_candidateTTs->add(tt);
else if (opNode->getOpCodeValue() == TR::ixor && opNode->getSecondChild()->getOpCode().hasSymbolReference() &&
node->getSymbolReference() == opNode->getSecondChild()->getSymbolReference() &&
opNode->getReferenceCount() == 1 && opNode->getSecondChild()->getReferenceCount() == 1 &&
_currentRegion->isExprInvariant(opNode->getFirstChild()))
_candidateTTs->add(tt);
}
}
TR_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;
}
