void
OMR::IlValue::storeToAuto()
{
if (_symRefThatCanBeUsedInOtherBlocks == NULL)
{
TR::Compilation *comp = TR::comp();
// first use from another block, need to create symref and insert store tree where node was computed
TR::SymbolReference *symRef = comp->getSymRefTab()->createTemporary(_methodBuilder->methodSymbol(), _nodeThatComputesValue->getDataType());
symRef->getSymbol()->setNotCollected();
char *name = (char *) comp->trMemory()->allocateHeapMemory((2+10+1) * sizeof(char)); // 2 ("_T") + max 10 digits + trailing zero
sprintf(name, "_T%u", symRef->getCPIndex());
symRef->getSymbol()->getAutoSymbol()->setName(name);
_methodBuilder->defineSymbol(name, symRef);
// create store and its treetop
TR::Node *storeNode = TR::Node::createStore(symRef, _nodeThatComputesValue);
TR::TreeTop *prevTreeTop = _treeTopThatAnchorsValue->getPrevTreeTop();
TR::TreeTop *newTree = TR::TreeTop::create(comp, storeNode);
newTree->insertNewTreeTop(prevTreeTop, _treeTopThatAnchorsValue);
_treeTopThatAnchorsValue->unlink(true);
_treeTopThatAnchorsValue = newTree;
_symRefThatCanBeUsedInOtherBlocks = symRef;
}
}
TR::Register *
TR::AMD64SystemLinkage::buildIndirectDispatch(TR::Node *callNode)
{
TR::SymbolReference *methodSymRef = callNode->getSymbolReference();
TR_ASSERT(methodSymRef->getSymbol()->castToMethodSymbol()->isComputed(), "system linkage only supports computed indirect call for now %p\n", callNode);
// Evaluate VFT
//
TR::Register *vftRegister;
TR::Node *vftNode = callNode->getFirstChild();
if (vftNode->getRegister())
{
vftRegister = vftNode->getRegister();
}
else
{
vftRegister = cg()->evaluate(vftNode);
}
// Allocate adequate register dependencies.
//
// pre = number of argument registers + 1 for VFT register
// post = number of volatile + VMThread + return register
//
uint32_t pre = getProperties().getNumIntegerArgumentRegisters() + getProperties().getNumFloatArgumentRegisters() + 1;
uint32_t post = getProperties().getNumVolatileRegisters() + 1 + (callNode->getDataType() == TR::NoType ? 0 : 1);
#if defined (PYTHON) && 0
// Treat all preserved GP regs as volatile until register map support available.
//
post += getProperties().getNumberOfPreservedGPRegisters();
#endif
TR::RegisterDependencyConditions *callDeps = generateRegisterDependencyConditions(pre, 1, cg());
TR::RealRegister::RegNum scratchRegIndex = getProperties().getIntegerScratchRegister(1);
callDeps->addPostCondition(vftRegister, scratchRegIndex, cg());
callDeps->stopAddingPostConditions();
// Evaluate outgoing arguments on the system stack and build pre-conditions.
//
int32_t memoryArgSize = buildArgs(callNode, callDeps);
// Dispatch
//
generateRegInstruction(CALLReg, callNode, vftRegister, callDeps, cg());
cg()->resetIsLeafMethod();
// Build label post-conditions
//
TR::RegisterDependencyConditions *postDeps = generateRegisterDependencyConditions(0, post, cg());
TR::Register *returnReg = buildVolatileAndReturnDependencies(callNode, postDeps);
postDeps->stopAddingPostConditions();
TR::LabelSymbol *postDepLabel = generateLabelSymbol(cg());
generateLabelInstruction(LABEL, callNode, postDepLabel, postDeps, cg());
return returnReg;
}
// resolved casts that are not to abstract, interface, or array need a super test
bool OMR::TreeEvaluator::instanceOfOrCheckCastNeedSuperTest(TR::Node * node, TR::CodeGenerator *cg)
{
TR::Node *castClassNode = node->getSecondChild();
TR::MethodSymbol *helperSym = node->getSymbol()->castToMethodSymbol();
TR::SymbolReference *castClassSymRef = castClassNode->getSymbolReference();
if (!TR::TreeEvaluator::isStaticClassSymRef(castClassSymRef))
{
// We could theoretically do a super test on something with no sym, but it would require significant
// changes to platform code. The benefit is little at this point (shows up from reference arraycopy reductions)
if (cg->supportsInliningOfIsInstance() &&
node->getOpCodeValue() == TR::instanceof &&
node->getSecondChild()->getOpCodeValue() != TR::loadaddr)
return true;
else
return false;
}
TR::StaticSymbol *castClassSym = castClassSymRef->getSymbol()->getStaticSymbol();
if (castClassSymRef->isUnresolved())
{
return false;
}
else
{
TR_OpaqueClassBlock * clazz;
// If the class is a regular class (i.e., not an interface nor an array) and
// not known to be a final class, an inline superclass test can be generated.
// If the helper does not preserve all the registers there will not be
// enough registers to do the superclass test inline.
// Also, don't generate the superclass test if optimizing for space.
//
if (castClassSym &&
(clazz = (TR_OpaqueClassBlock *) castClassSym->getStaticAddress()) &&
!TR::Compiler->cls.isClassArray(cg->comp(), clazz) &&
!TR::Compiler->cls.isInterfaceClass(cg->comp(), clazz) &&
!TR::Compiler->cls.isClassFinal(cg->comp(), clazz) &&
helperSym->preservesAllRegisters() &&
!cg->comp()->getOption(TR_OptimizeForSpace))
return true;
}
return false;
}
void
OMR::SymbolReference::setSharedStaticAliases(TR_BitVector * aliases, TR::SymbolReferenceTable * symRefTab)
{
if (self()->reallySharesSymbol())
{
TR::DataType type = self()->getSymbol()->getType();
TR_SymRefIterator i(type.isAddress() ? symRefTab->aliasBuilder.addressStaticSymRefs()
: (type.isInt32() ? symRefTab->aliasBuilder.intStaticSymRefs()
: symRefTab->aliasBuilder.nonIntPrimitiveStaticSymRefs()), symRefTab);
TR::SymbolReference * symRef;
while ((symRef = i.getNext()))
if (symRef->getSymbol() == self()->getSymbol())
aliases->set(symRef->getReferenceNumber());
}
else
aliases->set(self()->getReferenceNumber());
*aliases |= symRefTab->aliasBuilder.unsafeSymRefNumbers();
}
// unresolved casts or casts to things other than abstract or interface benefit from
// an equality test
bool OMR::TreeEvaluator::instanceOfOrCheckCastNeedEqualityTest(TR::Node * node, TR::CodeGenerator *cg)
{
TR::Node *castClassNode = node->getSecondChild();
TR::SymbolReference *castClassSymRef = castClassNode->getSymbolReference();
if (!TR::TreeEvaluator::isStaticClassSymRef(castClassSymRef))
{
return true;
}
TR::StaticSymbol *castClassSym = castClassSymRef->getSymbol()->getStaticSymbol();
if (castClassSymRef->isUnresolved())
{
return false;
}
else
{
TR_OpaqueClassBlock * clazz;
if (castClassSym
&& (clazz = (TR_OpaqueClassBlock *) castClassSym->getStaticAddress())
&& !TR::Compiler->cls.isInterfaceClass(cg->comp(), clazz)
&& (
!TR::Compiler->cls.isAbstractClass(cg->comp(), clazz)
// here be dragons
// int.class, char.class, etc are final & abstract
// usually instanceOf calls on these classes are ripped out by the optimizer
// but in some cases they can persist to codegen which without the following
// case causes assertions because we opt out of calling the helper and doing
// all inline tests. Really we could just jmp to the failed side, but to reduce
// service risk we are going to do an equality test that we know will fail
// NOTE final abstract is not enough - all array classes are final abstract
// to prevent them being used with new and being extended...
|| (TR::Compiler->cls.isAbstractClass(cg->comp(), clazz) && TR::Compiler->cls.isClassFinal(cg->comp(), clazz)
&& TR::Compiler->cls.isPrimitiveClass(cg->comp(), clazz)))
)
return true;
}
return false;
}
bool TR_LocalAnalysis::isSupportedNodeForPREPerformance(TR::Node *node, TR::Compilation *comp, TR::Node *parent)
{
TR::SymbolReference *symRef = node->getOpCode().hasSymbolReference()?node->getSymbolReference():NULL;
if (node->getOpCode().isStore() && symRef &&
symRef->getSymbol()->isAutoOrParm())
{
//dumpOptDetails("Returning false for store %p\n", node);
return false;
}
if (node->getOpCode().isLoadConst() && !comp->cg()->isMaterialized(node))
{
return false;
}
if (node->getOpCode().hasSymbolReference() &&
(node->getSymbolReference() == comp->getSymRefTab()->findJavaLangClassFromClassSymbolRef()))
{
return false;
}
return true;
}
void
OMR::SymbolReference::setSharedShadowAliases(TR_BitVector * aliases, TR::SymbolReferenceTable * symRefTab)
{
if (self()->reallySharesSymbol() && !_symbol->isUnsafeShadowSymbol())
{
TR::DataType type = self()->getSymbol()->getType();
TR_SymRefIterator i(type.isAddress() ? symRefTab->aliasBuilder.addressShadowSymRefs()
: (type.isInt32() ? symRefTab->aliasBuilder.intShadowSymRefs()
: symRefTab->aliasBuilder.nonIntPrimitiveShadowSymRefs()), symRefTab);
TR::SymbolReference * symRef;
while ((symRef = i.getNext()))
if (symRef->getSymbol() == self()->getSymbol())
aliases->set(symRef->getReferenceNumber());
// include symbol reference's own shared alias bitvector
if (symRefTab->getSharedAliases(self()) != NULL)
*aliases |= *(symRefTab->getSharedAliases(self()));
}
else
aliases->set(self()->getReferenceNumber());
*aliases |= symRefTab->aliasBuilder.unsafeSymRefNumbers();
}
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
}
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;
}
// 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_ReachingDefinitions::initializeGenAndKillSetInfoForNode(TR::Node *node, TR_UseDefInfo::BitVector &defsKilled, bool seenException, int32_t blockNum, TR::Node *parent)
{
// Update gen and kill info for nodes in this subtree
//
int32_t i;
if (node->getVisitCount() == comp()->getVisitCount())
return;
node->setVisitCount(comp()->getVisitCount());
// Process the children first
//
for (i = node->getNumChildren()-1; i >= 0; --i)
{
initializeGenAndKillSetInfoForNode(node->getChild(i), defsKilled, seenException, blockNum, node);
}
bool irrelevantStore = false;
scount_t nodeIndex = node->getLocalIndex();
if (nodeIndex <= 0)
{
if (node->getOpCode().isStore() &&
node->getSymbol()->isAutoOrParm() &&
node->storedValueIsIrrelevant())
{
irrelevantStore = true;
}
else
return;
}
bool foundDefsToKill = false;
int32_t numDefNodes = 0;
defsKilled.Clear();
TR::ILOpCode &opCode = node->getOpCode();
TR::SymbolReference *symRef;
TR::Symbol *sym;
uint16_t symIndex;
uint32_t num_aliases;
if (_useDefInfo->_useDefForRegs &&
(opCode.isLoadReg() ||
opCode.isStoreReg()))
{
sym = NULL;
symRef = NULL;
symIndex = _useDefInfo->getNumSymbols() + node->getGlobalRegisterNumber();
num_aliases = 1;
}
else
{
symRef = node->getSymbolReference();
sym = symRef->getSymbol();
symIndex = symRef->getSymbol()->getLocalIndex();
num_aliases = _useDefInfo->getNumAliases(symRef, _aux);
}
if (symIndex == NULL_USEDEF_SYMBOL_INDEX || node->getOpCode().isCall() || node->getOpCode().isFence() ||
(parent && parent->getOpCode().isResolveCheck() && num_aliases > 1))
{
// A call or unresolved reference is a definition of all
// symbols it is aliased with
//
numDefNodes = num_aliases;
//for all symbols that are a mustdef of a call, kill defs of those symbols
if (node->getOpCode().isCall())
foundDefsToKill = false;
}
else if (irrelevantStore || _useDefInfo->isExpandedDefIndex(nodeIndex))
{
// DefOnly node defines all symbols it is aliased with
// UseDef node(load) defines only the symbol itself
//
if (!irrelevantStore)
{
numDefNodes = num_aliases;
numDefNodes = _useDefInfo->isExpandedUseDefIndex(nodeIndex) ? 1 : numDefNodes;
if (!_useDefInfo->getDefsForSymbolIsZero(symIndex, _aux) &&
(!sym ||
(!sym->isShadow() &&
!sym->isMethod())))
{
foundDefsToKill = true;
// defsKilled ORed with defsForSymbol(symIndex);
_useDefInfo->getDefsForSymbol(defsKilled, symIndex, _aux);
}
if (node->getOpCode().isStoreIndirect())
{
int32_t memSymIndex = _useDefInfo->getMemorySymbolIndex(node);
if (memSymIndex != -1 &&
!_useDefInfo->getDefsForSymbolIsZero(memSymIndex, _aux))
{
foundDefsToKill = true;
// defsKilled ORed with defsForSymbol(symIndex);
_useDefInfo->getDefsForSymbol(defsKilled, memSymIndex, _aux);
}
}
}
else if (!_useDefInfo->getDefsForSymbolIsZero(symIndex, _aux))
{
numDefNodes = 1;
foundDefsToKill = true;
// defsKilled ORed with defsForSymbol(symIndex);
_useDefInfo->getDefsForSymbol(defsKilled, symIndex, _aux);
}
}
else
{
numDefNodes = 0;
}
if (foundDefsToKill)
{
if (_regularKillSetInfo[blockNum] == NULL)
allocateContainer(&_regularKillSetInfo[blockNum]);
*_regularKillSetInfo[blockNum] |= defsKilled;
if (!seenException)
{
if (_exceptionKillSetInfo[blockNum] == NULL)
allocateContainer(&_exceptionKillSetInfo[blockNum]);
*_exceptionKillSetInfo[blockNum] |= defsKilled;
}
}
if (_regularGenSetInfo[blockNum] == NULL)
allocateContainer(&_regularGenSetInfo[blockNum]);
else if (foundDefsToKill)
*_regularGenSetInfo[blockNum] -= defsKilled;
if (_exceptionGenSetInfo[blockNum] == NULL)
allocateContainer(&_exceptionGenSetInfo[blockNum]);
else if (foundDefsToKill && !seenException)
*_exceptionGenSetInfo[blockNum] -= defsKilled;
if (!irrelevantStore)
{
for (i = 0; i < numDefNodes; ++i)
{
_regularGenSetInfo[blockNum]->set(nodeIndex+i);
_exceptionGenSetInfo[blockNum]->set(nodeIndex+i);
}
}
else // fake up the method entry def as the def index to "gen" to avoid a use without a def completely
{
_regularGenSetInfo[blockNum]->set(sym->getLocalIndex());
_exceptionGenSetInfo[blockNum]->set(sym->getLocalIndex());
}
}
TR::Register *TR::AMD64SystemLinkage::buildDirectDispatch(
TR::Node *callNode,
bool spillFPRegs)
{
TR::SymbolReference *methodSymRef = callNode->getSymbolReference();
TR::MethodSymbol *methodSymbol = methodSymRef->getSymbol()->castToMethodSymbol();
TR::Register *returnReg;
// Allocate adequate register dependencies.
//
// pre = number of argument registers
// post = number of volatile + return register
//
uint32_t pre = getProperties().getNumIntegerArgumentRegisters() + getProperties().getNumFloatArgumentRegisters();
uint32_t post = getProperties().getNumVolatileRegisters() + (callNode->getDataType() == TR::NoType ? 0 : 1);
#if defined (PYTHON) && 0
// Treat all preserved GP regs as volatile until register map support available.
//
post += getProperties().getNumberOfPreservedGPRegisters();
#endif
TR::RegisterDependencyConditions *preDeps = generateRegisterDependencyConditions(pre, 0, cg());
TR::RegisterDependencyConditions *postDeps = generateRegisterDependencyConditions(0, post, cg());
// Evaluate outgoing arguments on the system stack and build pre-conditions.
//
int32_t memoryArgSize = buildArgs(callNode, preDeps);
// Build post-conditions.
//
returnReg = buildVolatileAndReturnDependencies(callNode, postDeps);
postDeps->stopAddingPostConditions();
// Find the second scratch register in the post dependency list.
//
TR::Register *scratchReg = NULL;
TR::RealRegister::RegNum scratchRegIndex = getProperties().getIntegerScratchRegister(1);
for (int32_t i=0; i<post; i++)
{
if (postDeps->getPostConditions()->getRegisterDependency(i)->getRealRegister() == scratchRegIndex)
{
scratchReg = postDeps->getPostConditions()->getRegisterDependency(i)->getRegister();
break;
}
}
#if defined(PYTHON) && 0
// For Python, store the instruction that contains the GC map at this site into
// the frame object.
//
TR::SymbolReference *frameObjectSymRef =
comp()->getSymRefTab()->findOrCreateAutoSymbol(comp()->getMethodSymbol(), 0, TR::Address, true, false, true);
TR::Register *frameObjectRegister = cg()->allocateRegister();
generateRegMemInstruction(
L8RegMem,
callNode,
frameObjectRegister,
generateX86MemoryReference(frameObjectSymRef, cg()),
cg());
TR::RealRegister *espReal = cg()->machine()->getX86RealRegister(TR::RealRegister::esp);
TR::Register *gcMapPCRegister = cg()->allocateRegister();
generateRegMemInstruction(
LEA8RegMem,
callNode,
gcMapPCRegister,
generateX86MemoryReference(espReal, -8, cg()),
cg());
// Use "volatile" registers across the call. Once proper register map support
// is implemented, r14 and r15 will no longer be volatile and a different pair
// should be chosen.
//
TR::RegisterDependencyConditions *gcMapDeps = generateRegisterDependencyConditions(0, 2, cg());
gcMapDeps->addPostCondition(frameObjectRegister, TR::RealRegister::r14, cg());
gcMapDeps->addPostCondition(gcMapPCRegister, TR::RealRegister::r15, cg());
gcMapDeps->stopAddingPostConditions();
generateMemRegInstruction(
S8MemReg,
callNode,
generateX86MemoryReference(frameObjectRegister, fe()->getPythonGCMapPCOffsetInFrame(), cg()),
gcMapPCRegister,
gcMapDeps,
cg());
cg()->stopUsingRegister(frameObjectRegister);
cg()->stopUsingRegister(gcMapPCRegister);
#endif
TR::Instruction *instr;
if (methodSymbol->getMethodAddress())
{
TR_ASSERT(scratchReg, "could not find second scratch register");
auto LoadRegisterInstruction = generateRegImm64SymInstruction(
MOV8RegImm64,
callNode,
scratchReg,
(uintptr_t)methodSymbol->getMethodAddress(),
methodSymRef,
cg());
if (TR::Options::getCmdLineOptions()->getOption(TR_EmitRelocatableELFFile))
{
LoadRegisterInstruction->setReloKind(TR_NativeMethodAbsolute);
}
instr = generateRegInstruction(CALLReg, callNode, scratchReg, preDeps, cg());
}
else
{
instr = generateImmSymInstruction(CALLImm4, callNode, (uintptrj_t)methodSymbol->getMethodAddress(), methodSymRef, preDeps, cg());
}
cg()->resetIsLeafMethod();
instr->setNeedsGCMap(getProperties().getPreservedRegisterMapForGC());
cg()->stopUsingRegister(scratchReg);
TR::LabelSymbol *postDepLabel = generateLabelSymbol(cg());
generateLabelInstruction(LABEL, callNode, postDepLabel, postDeps, cg());
return returnReg;
}
TR::Register *IA32LinkageUtils::pushIntegerWordArg(
TR::Node *child,
TR::CodeGenerator *cg)
{
TR::Register *pushRegister;
if (child->getRegister() == NULL)
{
if (child->getOpCode().isLoadConst())
{
int32_t value = child->getInt();
TR_X86OpCodes pushOp;
if (value >= -128 && value <= 127)
{
pushOp = PUSHImms;
}
else
{
pushOp = PUSHImm4;
}
generateImmInstruction(pushOp, child, value, cg);
cg->decReferenceCount(child);
return NULL;
}
else if (child->getOpCodeValue() == TR::loadaddr)
{
TR::SymbolReference * symRef = child->getSymbolReference();
TR::StaticSymbol *sym = symRef->getSymbol()->getStaticSymbol();
if (sym)
{
TR_ASSERT(!symRef->isUnresolved(), "pushIntegerWordArg loadaddr expecting resolved symbol");
generateImmSymInstruction(PUSHImm4, child, (uintptrj_t)sym->getStaticAddress(), symRef, cg);
cg->decReferenceCount(child);
return NULL;
}
}
else if (child->getOpCodeValue() == TR::fbits2i &&
!child->normalizeNanValues() &&
child->getReferenceCount() == 1)
{
pushRegister = pushFloatArg(child->getFirstChild(), cg);
cg->decReferenceCount(child);
return pushRegister;
}
else if (child->getOpCode().isMemoryReference() &&
(child->getReferenceCount() == 1) &&
(child->getSymbolReference() != cg->comp()->getSymRefTab()->findVftSymbolRef()))
{
TR::MemoryReference *tempMR = generateX86MemoryReference(child, cg);
generateMemInstruction(PUSHMem, child, tempMR, cg);
tempMR->decNodeReferenceCounts(cg);
cg->decReferenceCount(child);
return NULL;
}
}
pushRegister = cg->evaluate(child);
generateRegInstruction(PUSHReg, child, pushRegister, cg);
cg->decReferenceCount(child);
return pushRegister;
}
// 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::ARM64SystemLinkage::buildDirectDispatch(TR::Node *callNode)
{
TR::SymbolReference *callSymRef = callNode->getSymbolReference();
const TR::ARM64LinkageProperties &pp = getProperties();
TR::RealRegister *sp = cg()->machine()->getRealRegister(pp.getStackPointerRegister());
TR::RegisterDependencyConditions *dependencies =
new (trHeapMemory()) TR::RegisterDependencyConditions(
pp.getNumberOfDependencyGPRegisters(),
pp.getNumberOfDependencyGPRegisters(), trMemory());
int32_t totalSize = buildArgs(callNode, dependencies);
if (totalSize > 0)
{
if (constantIsUnsignedImm12(totalSize))
{
generateTrg1Src1ImmInstruction(cg(), TR::InstOpCode::subimmx, callNode, sp, sp, totalSize);
}
else
{
TR_ASSERT_FATAL(false, "Too many arguments.");
}
}
TR::MethodSymbol *callSymbol = callSymRef->getSymbol()->castToMethodSymbol();
generateImmSymInstruction(cg(), TR::InstOpCode::bl, callNode,
(uintptr_t)callSymbol->getMethodAddress(),
dependencies, callSymRef ? callSymRef : callNode->getSymbolReference(), NULL);
cg()->machine()->setLinkRegisterKilled(true);
if (totalSize > 0)
{
if (constantIsUnsignedImm12(totalSize))
{
generateTrg1Src1ImmInstruction(cg(), TR::InstOpCode::addimmx, callNode, sp, sp, totalSize);
}
else
{
TR_ASSERT_FATAL(false, "Too many arguments.");
}
}
TR::Register *retReg;
switch(callNode->getOpCodeValue())
{
case TR::icall:
case TR::iucall:
retReg = dependencies->searchPostConditionRegister(
pp.getIntegerReturnRegister());
break;
case TR::lcall:
case TR::lucall:
case TR::acall:
retReg = dependencies->searchPostConditionRegister(
pp.getLongReturnRegister());
break;
case TR::fcall:
case TR::dcall:
retReg = dependencies->searchPostConditionRegister(
pp.getFloatReturnRegister());
break;
case TR::call:
retReg = NULL;
break;
default:
retReg = NULL;
TR_ASSERT(false, "Unsupported direct call Opcode.");
}
callNode->setRegister(retReg);
return retReg;
}
TR_BitVector *
addVeryRefinedCallAliasSets(TR::ResolvedMethodSymbol * methodSymbol, TR_BitVector * aliases, List<void> * methodsPeeked)
{
TR::Compilation *comp = TR::comp();
void * methodId = methodSymbol->getResolvedMethod()->getPersistentIdentifier();
if (methodsPeeked->find(methodId))
{
// This can't be allocated into the alias region as it must be accessed across optimizations
TR_BitVector *heapAliases = new (comp->trHeapMemory()) TR_BitVector(comp->getSymRefCount(), comp->trMemory(), heapAlloc, growable);
*heapAliases |= *aliases;
return heapAliases;
}
// stop if the peek is getting very deep
//
if (methodsPeeked->getSize() >= PEEK_THRESHOLD)
return 0;
methodsPeeked->add(methodId);
dumpOptDetails(comp, "O^O REFINING ALIASES: Peeking into the IL to refine aliases \n");
if (!methodSymbol->getResolvedMethod()->genMethodILForPeeking(methodSymbol, comp, true))
return 0;
TR::SymbolReferenceTable * symRefTab = comp->getSymRefTab();
for (TR::TreeTop * tt = methodSymbol->getFirstTreeTop(); tt; tt = tt->getNextTreeTop())
{
TR::Node *node = tt->getNode();
if (node->getOpCode().isResolveCheck())
return 0;
if ((node->getOpCodeValue() == TR::treetop) ||
(node->getOpCodeValue() == TR::compressedRefs) ||
node->getOpCode().isCheck())
node = node->getFirstChild();
if (node->getOpCode().isStore())
{
TR::SymbolReference * symRefInCallee = node->getSymbolReference(), * symRefInCaller;
TR::Symbol * symInCallee = symRefInCallee->getSymbol();
TR::DataType type = symInCallee->getDataType();
if (symInCallee->isShadow())
{
if (symInCallee->isArrayShadowSymbol())
symRefInCaller = symRefTab->getSymRef(symRefTab->getArrayShadowIndex(type));
else if (symInCallee->isArrayletShadowSymbol())
symRefInCaller = symRefTab->getSymRef(symRefTab->getArrayletShadowIndex(type));
else
symRefInCaller = symRefTab->findShadowSymbol(symRefInCallee->getOwningMethod(comp), symRefInCallee->getCPIndex(), type);
if (symRefInCaller)
{
if (symRefInCaller->reallySharesSymbol(comp))
symRefInCaller->setSharedShadowAliases(aliases, symRefTab);
aliases->set(symRefInCaller->getReferenceNumber());
}
}
else if (symInCallee->isStatic())
{
symRefInCaller = symRefTab->findStaticSymbol(symRefInCallee->getOwningMethod(comp), symRefInCallee->getCPIndex(), type);
if (symRefInCaller)
{
if (symRefInCaller->reallySharesSymbol(comp))
symRefInCaller->setSharedStaticAliases(aliases, symRefTab);
else
aliases->set(symRefInCaller->getReferenceNumber());
}
}
}
else if (node->getOpCode().isCall())
{
if (node->getOpCode().isCallIndirect())
return 0;
TR::ResolvedMethodSymbol * calleeSymbol = node->getSymbol()->getResolvedMethodSymbol();
if (!calleeSymbol)
return 0;
TR_ResolvedMethod * calleeMethod = calleeSymbol->getResolvedMethod();
if (!calleeMethod->isCompilable(comp->trMemory()) || calleeMethod->isJNINative())
return 0;
if (!addVeryRefinedCallAliasSets(calleeSymbol, aliases, methodsPeeked))
return 0;
}
else if (node->getOpCodeValue() == TR::monent)
return 0;
}
// This can't be allocated into the alias region as it must be accessed across optimizations
TR_BitVector *heapAliases = new (comp->trHeapMemory()) TR_BitVector(comp->getSymRefCount(), comp->trMemory(), heapAlloc, growable);
*heapAliases |= *aliases;
return heapAliases;
}
TR::Register *TR::AMD64SystemLinkage::buildDirectDispatch(
TR::Node *callNode,
bool spillFPRegs)
{
TR::SymbolReference *methodSymRef = callNode->getSymbolReference();
TR::MethodSymbol *methodSymbol = methodSymRef->getSymbol()->castToMethodSymbol();
TR::Register *returnReg;
// Allocate adequate register dependencies.
//
// pre = number of argument registers
// post = number of volatile + return register
//
uint32_t pre = getProperties().getNumIntegerArgumentRegisters() + getProperties().getNumFloatArgumentRegisters();
uint32_t post = getProperties().getNumVolatileRegisters() + (callNode->getDataType() == TR::NoType ? 0 : 1);
TR::RegisterDependencyConditions *preDeps = generateRegisterDependencyConditions(pre, 0, cg());
TR::RegisterDependencyConditions *postDeps = generateRegisterDependencyConditions(0, post, cg());
// Evaluate outgoing arguments on the system stack and build pre-conditions.
//
int32_t memoryArgSize = buildArgs(callNode, preDeps);
// Build post-conditions.
//
returnReg = buildVolatileAndReturnDependencies(callNode, postDeps);
postDeps->stopAddingPostConditions();
// Find the second scratch register in the post dependency list.
//
TR::Register *scratchReg = NULL;
TR::RealRegister::RegNum scratchRegIndex = getProperties().getIntegerScratchRegister(1);
for (int32_t i=0; i<post; i++)
{
if (postDeps->getPostConditions()->getRegisterDependency(i)->getRealRegister() == scratchRegIndex)
{
scratchReg = postDeps->getPostConditions()->getRegisterDependency(i)->getRegister();
break;
}
}
TR::Instruction *instr;
if (methodSymbol->getMethodAddress())
{
TR_ASSERT(scratchReg, "could not find second scratch register");
auto LoadRegisterInstruction = generateRegImm64SymInstruction(
MOV8RegImm64,
callNode,
scratchReg,
(uintptr_t)methodSymbol->getMethodAddress(),
methodSymRef,
cg());
if (comp()->getOption(TR_EmitRelocatableELFFile))
{
LoadRegisterInstruction->setReloKind(TR_NativeMethodAbsolute);
}
instr = generateRegInstruction(CALLReg, callNode, scratchReg, preDeps, cg());
}
else
{
instr = generateImmSymInstruction(CALLImm4, callNode, (uintptrj_t)methodSymbol->getMethodAddress(), methodSymRef, preDeps, cg());
}
cg()->resetIsLeafMethod();
instr->setNeedsGCMap(getProperties().getPreservedRegisterMapForGC());
cg()->stopUsingRegister(scratchReg);
TR::LabelSymbol *postDepLabel = generateLabelSymbol(cg());
generateLabelInstruction(LABEL, callNode, postDepLabel, postDeps, cg());
return returnReg;
}
void
TR_Debug::print(TR::FILE *pOutFile, TR::S390CallSnippet * snippet)
{
uint8_t * bufferPos = snippet->getSnippetLabel()->getCodeLocation();
TR::Node * callNode = snippet->getNode();
TR::SymbolReference * methodSymRef = snippet->getRealMethodSymbolReference();
if(!methodSymRef)
methodSymRef = callNode->getSymbolReference();
TR::MethodSymbol * methodSymbol = methodSymRef->getSymbol()->castToMethodSymbol();
TR::SymbolReference * glueRef;
int8_t padbytes = snippet->getPadBytes();
printSnippetLabel(pOutFile, snippet->getSnippetLabel(), bufferPos,
methodSymRef->isUnresolved() ? "Unresolved Call Snippet" : "Call Snippet");
bufferPos = printS390ArgumentsFlush(pOutFile, callNode, bufferPos, snippet->getSizeOfArguments());
if (methodSymRef->isUnresolved() || _comp->compileRelocatableCode())
{
if (methodSymbol->isSpecial())
{
glueRef = _cg->getSymRef(TR_S390interpreterUnresolvedSpecialGlue);
}
else if (methodSymbol->isStatic())
{
glueRef = _cg->getSymRef(TR_S390interpreterUnresolvedStaticGlue);
}
else
{
glueRef = _cg->getSymRef(TR_S390interpreterUnresolvedDirectVirtualGlue);
}
}
else
{
bool synchronised = methodSymbol->isSynchronised();
if ((methodSymbol->isVMInternalNative() || methodSymbol->isJITInternalNative()))
{
glueRef = _cg->getSymRef(TR_S390nativeStaticHelper);
}
else
{
switch (callNode->getDataType())
{
case TR::NoType:
if (synchronised)
{
glueRef = _cg->getSymRef(TR_S390interpreterSyncVoidStaticGlue);
}
else
{
glueRef = _cg->getSymRef(TR_S390interpreterVoidStaticGlue);
}
break;
case TR::Int8:
case TR::Int16:
case TR::Int32:
if (synchronised)
{
glueRef = _cg->getSymRef(TR_S390interpreterSyncIntStaticGlue);
}
else
{
glueRef = _cg->getSymRef(TR_S390interpreterIntStaticGlue);
}
break;
case TR::Address:
if (TR::Compiler->target.is64Bit())
{
if (synchronised)
{
glueRef = _cg->getSymRef(TR_S390interpreterSyncLongStaticGlue);
}
else
{
glueRef = _cg->getSymRef(TR_S390interpreterLongStaticGlue);
}
}
else
{
if (synchronised)
{
glueRef = _cg->getSymRef(TR_S390interpreterSyncIntStaticGlue);
}
else
{
glueRef = _cg->getSymRef(TR_S390interpreterIntStaticGlue);
}
}
break;
case TR::Int64:
if (synchronised)
{
glueRef = _cg->getSymRef(TR_S390interpreterSyncLongStaticGlue);
}
else
{
glueRef = _cg->getSymRef(TR_S390interpreterLongStaticGlue);
}
break;
case TR::Float:
if (synchronised)
{
glueRef = _cg->getSymRef(TR_S390interpreterSyncFloatStaticGlue);
}
else
{
glueRef = _cg->getSymRef(TR_S390interpreterFloatStaticGlue);
}
break;
case TR::Double:
if (synchronised)
{
glueRef = _cg->getSymRef(TR_S390interpreterSyncDoubleStaticGlue);
}
else
{
glueRef = _cg->getSymRef(TR_S390interpreterDoubleStaticGlue);
}
break;
default:
TR_ASSERT(0, "Bad return data type for a call node. DataType was %s\n",
getName(callNode->getDataType()));
}
}
}
bufferPos = printRuntimeInstrumentationOnOffInstruction(pOutFile, bufferPos, false); // RIOFF
if (snippet->getKind() == TR::Snippet::IsUnresolvedCall)
{
int lengthOfLoad = (TR::Compiler->target.is64Bit())?6:4;
printPrefix(pOutFile, NULL, bufferPos, 6);
trfprintf(pOutFile, "LARL \tGPR14, *+%d <%p>\t# Start of Data Const.",
8 + lengthOfLoad + padbytes,
bufferPos + 8 + lengthOfLoad + padbytes);
bufferPos += 6;
if (TR::Compiler->target.is64Bit())
{
printPrefix(pOutFile, NULL, bufferPos, 6);
trfprintf(pOutFile, "LG \tGPR_EP, 0(,GPR14)");
bufferPos += 6;
}
else
{
printPrefix(pOutFile, NULL, bufferPos, 4);
trfprintf(pOutFile, "L \tGPR_EP, 0(,GPR14)");
bufferPos += 4;
}
printPrefix(pOutFile, NULL, bufferPos, 2);
trfprintf(pOutFile, "BCR \tGPR_EP");
bufferPos += 2;
}
else
{
printPrefix(pOutFile, NULL, bufferPos, 6);
trfprintf(pOutFile, "BRASL \tGPR14, <%p>\t# Branch to Helper Method %s",
snippet->getSnippetDestAddr(),
snippet->usedTrampoline()?"- Trampoline Used.":"");
bufferPos += 6;
}
if (padbytes == 2)
{
printPrefix(pOutFile, NULL, bufferPos, 2);
trfprintf(pOutFile, "DC \t0x0000 \t\t\t# 2-bytes padding for alignment");
bufferPos += 2;
}
else if (padbytes == 4)
{
printPrefix(pOutFile, NULL, bufferPos, 4) ;
trfprintf(pOutFile, "DC \t0x00000000 \t\t# 4-bytes padding for alignment");
bufferPos += 4;
}
else if (padbytes == 6)
{
printPrefix(pOutFile, NULL, bufferPos, 6) ;
trfprintf(pOutFile, "DC \t0x000000000000 \t\t# 6-bytes padding for alignment");
bufferPos += 6;
}
printPrefix(pOutFile, NULL, bufferPos, sizeof(intptrj_t));
trfprintf(pOutFile, "DC \t%p \t\t# Method Address", glueRef->getMethodAddress());
bufferPos += sizeof(intptrj_t);
printPrefix(pOutFile, NULL, bufferPos, sizeof(intptrj_t));
trfprintf(pOutFile, "DC \t%p \t\t# Call Site RA", snippet->getCallRA());
bufferPos += sizeof(intptrj_t);
if (methodSymRef->isUnresolved())
{
printPrefix(pOutFile, NULL, bufferPos, 0);
}
else
{
printPrefix(pOutFile, NULL, bufferPos, sizeof(intptrj_t));
}
trfprintf(pOutFile, "DC \t%p \t\t# Method Pointer", methodSymRef->isUnresolved() ? 0 : methodSymbol->getMethodAddress());
}
uint8_t *
TR::S390HelperCallSnippet::emitSnippetBody()
{
uint8_t * cursor = cg()->getBinaryBufferCursor();
getSnippetLabel()->setCodeLocation(cursor);
TR::Node * callNode = getNode();
TR::SymbolReference * helperSymRef = getHelperSymRef();
bool jitInduceOSR = helperSymRef == cg()->symRefTab()->element(TR_induceOSRAtCurrentPC);
if (jitInduceOSR)
{
// Flush in-register arguments back to the stack for interpreter
cursor = TR::S390CallSnippet::S390flushArgumentsToStack(cursor, callNode, getSizeOfArguments(), cg());
}
uint32_t rEP = (uint32_t) cg()->getEntryPointRegister() - 1;
//load vm thread into gpr13
cursor = generateLoadVMThreadInstruction(cg(), cursor);
// Generate RIOFF if RI is supported.
cursor = generateRuntimeInstrumentationOnOffInstruction(cg(), cursor, TR::InstOpCode::RIOFF);
if ( // Methods that require
alwaysExcept()) // R14 to point to snippet:
{
// For trace method entry/exit, we need to set up R14 to point to the
// beginning of the data segment. We will use BRASL to automatically
// set R14 correctly.
// For methods that lead to exceptions, and never return to the
// main code, we set up R14, so that if GC occurs, the stackwalker
// will see R14 is pointing to this snippet, and pick up the correct
// stack map.
*(int16_t *) cursor = 0xC0E5; // BRASL R14, <Helper Addr>
cursor += sizeof(int16_t);
}
else // Otherwise:
{
// We're not sure if the helper will return. So, we need to provide
// the return addr of the main line code, so that when helper call
// completes, it can jump back properly.
// Load Return Address into R14.
intptrj_t returnAddr = (intptrj_t)getReStartLabel()->getCodeLocation(); // LARL R14, <Return Addr>
*(int16_t *) cursor = 0xC0E0;
cursor += sizeof(int16_t);
*(int32_t *) cursor = (int32_t)((returnAddr - (intptrj_t)(cursor - 2)) / 2);
cursor += sizeof(int32_t);
*(int16_t *) cursor = 0xC0F4; // BRCL <Helper Addr>
cursor += sizeof(int16_t);
}
// Calculate the relative offset to get to helper method.
// If MCC is not supported, everything should be reachable.
// If MCC is supported, we will look up the appropriate trampoline, if
// necessary.
intptrj_t destAddr = (intptrj_t)(helperSymRef->getSymbol()->castToMethodSymbol()->getMethodAddress());
#if defined(TR_TARGET_64BIT)
#if defined(J9ZOS390)
if (cg()->comp()->getOption(TR_EnableRMODE64))
#endif
{
if (NEEDS_TRAMPOLINE(destAddr, cursor, cg()))
{
destAddr = cg()->fe()->indexedTrampolineLookup(helperSymRef->getReferenceNumber(), (void *)cursor);
this->setUsedTrampoline(true);
// We clobber rEP if we take a trampoline. Update our register map if necessary.
if (gcMap().getStackMap() != NULL)
{
gcMap().getStackMap()->maskRegisters(~(0x1 << (rEP)));
}
}
}
#endif
TR_ASSERT(CHECK_32BIT_TRAMPOLINE_RANGE(destAddr, cursor), "Helper Call is not reachable.");
this->setSnippetDestAddr(destAddr);
*(int32_t *) cursor = (int32_t)((destAddr - (intptrj_t)(cursor - 2)) / 2);
AOTcgDiag1(cg()->comp(), "add TR_HelperAddress cursor=%x\n", cursor);
cg()->addProjectSpecializedRelocation(cursor, (uint8_t*) helperSymRef, NULL, TR_HelperAddress,
__FILE__, __LINE__, getNode());
cursor += sizeof(int32_t);
gcMap().registerStackMap(cursor, cg());
return cursor;
}