TR::Register* OMR::X86::TreeEvaluator::SIMDstoreEvaluator(TR::Node* node, TR::CodeGenerator* cg)
{
TR::Node* valueNode = node->getChild(node->getOpCode().isIndirect() ? 1 : 0);
TR::MemoryReference* tempMR = generateX86MemoryReference(node, cg);
tempMR = ConvertToPatchableMemoryReference(tempMR, node, cg);
TR::Register* valueReg = cg->evaluate(valueNode);
TR_X86OpCodes opCode = BADIA32Op;
switch (node->getSize())
{
case 16:
opCode = MOVDQUMemReg;
break;
default:
if (cg->comp()->getOption(TR_TraceCG))
traceMsg(cg->comp(), "Unsupported fill size: Node = %p\n", node);
TR_ASSERT(false, "Unsupported fill size");
break;
}
TR::Instruction* instr = generateMemRegInstruction(opCode, node, tempMR, valueReg, cg);
cg->decReferenceCount(valueNode);
tempMR->decNodeReferenceCounts(cg);
if (node->getOpCode().isIndirect())
cg->setImplicitExceptionPoint(instr);
return NULL;
}
TR::Register *IA32LinkageUtils::pushFloatArg(
TR::Node *child,
TR::CodeGenerator *cg)
{
TR::Register *pushRegister;
if (child->getRegister() == NULL)
{
if (child->getOpCodeValue() == TR::fconst)
{
int32_t value = child->getFloatBits();
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->getReferenceCount() == 1)
{
if (child->getOpCode().isLoad())
{
TR::MemoryReference *tempMR = generateX86MemoryReference(child, cg);
generateMemInstruction(PUSHMem, child, tempMR, cg);
tempMR->decNodeReferenceCounts(cg);
cg->decReferenceCount(child);
return NULL;
}
else if (child->getOpCodeValue() == TR::ibits2f)
{
pushRegister = pushIntegerWordArg(child->getFirstChild(), cg);
cg->decReferenceCount(child);
return pushRegister;
}
}
}
pushRegister = cg->evaluate(child);
TR::RealRegister *espReal = cg->machine()->getRealRegister(TR::RealRegister::esp);
generateRegImmInstruction(SUB4RegImms, child, espReal, 4, cg);
if (cg->useSSEForSinglePrecision() && pushRegister->getKind() == TR_FPR)
generateMemRegInstruction(MOVSSMemReg, child, generateX86MemoryReference(espReal, 0, cg), pushRegister, cg);
else
generateFPMemRegInstruction(FSTMemReg, child, generateX86MemoryReference(espReal, 0, cg), pushRegister, cg);
cg->decReferenceCount(child);
return pushRegister;
}
TR::Instruction *
TR_X86SystemLinkage::savePreservedRegisters(TR::Instruction *cursor)
{
// For IA32, if disableShrinkWrapping, usePushForPreservedRegs will be true; otherwise false;
// For X64, shrinkWraping is always on, and usePushForPreservedRegs always false;
// TR_ASSERT(!getProperties().getUsesPushesForPreservedRegs(), "assertion failure");
TR::ResolvedMethodSymbol *bodySymbol = comp()->getJittedMethodSymbol();
const int32_t localSize = getProperties().getOffsetToFirstLocal() - bodySymbol->getLocalMappingCursor();
const int32_t pointerSize = getProperties().getPointerSize();
int32_t offsetCursor = -localSize + getProperties().getOffsetToFirstLocal() - pointerSize;
if (_properties.getUsesPushesForPreservedRegs())
{
for (int32_t pindex = _properties.getMaxRegistersPreservedInPrologue()-1;
pindex >= 0;
pindex--)
{
TR::RealRegister::RegNum idx = _properties.getPreservedRegister((uint32_t)pindex);
TR::RealRegister *reg = machine()->getX86RealRegister(idx);
if (reg->getHasBeenAssignedInMethod() && reg->getState() != TR::RealRegister::Locked)
{
cursor = new (trHeapMemory()) TR::X86RegInstruction(cursor, PUSHReg, reg, cg());
}
}
}
else
{
TR_BitVector *p = cg()->getPreservedRegsInPrologue();
for (int32_t pindex = getProperties().getMaxRegistersPreservedInPrologue()-1;
pindex >= 0;
pindex--)
{
TR::RealRegister::RegNum idx = _properties.getPreservedRegister((uint32_t)pindex);
TR::RealRegister *reg = machine()->getX86RealRegister(getProperties().getPreservedRegister((uint32_t)pindex));
if(reg->getHasBeenAssignedInMethod() && reg->getState() != TR::RealRegister::Locked)
{
if (!p || p->get(idx))
{
cursor = generateMemRegInstruction(
cursor,
movOpcodes[MemReg][fullRegisterMovType(reg)],
generateX86MemoryReference(machine()->getX86RealRegister(TR::RealRegister::vfp), offsetCursor, cg()),
reg,
cg()
);
}
offsetCursor -= pointerSize;
}
}
}
return cursor;
}
TR::Instruction *
TR::X86SystemLinkage::savePreservedRegisters(TR::Instruction *cursor)
{
// For IA32 usePushForPreservedRegs will be true;
// For X64, usePushForPreservedRegs always false;
TR::ResolvedMethodSymbol *bodySymbol = comp()->getJittedMethodSymbol();
const int32_t localSize = getProperties().getOffsetToFirstLocal() - bodySymbol->getLocalMappingCursor();
const int32_t pointerSize = getProperties().getPointerSize();
int32_t offsetCursor = -localSize + getProperties().getOffsetToFirstLocal() - pointerSize;
if (_properties.getUsesPushesForPreservedRegs())
{
for (int32_t pindex = _properties.getMaxRegistersPreservedInPrologue()-1;
pindex >= 0;
pindex--)
{
TR::RealRegister::RegNum idx = _properties.getPreservedRegister((uint32_t)pindex);
TR::RealRegister *reg = machine()->getX86RealRegister(idx);
if (reg->getHasBeenAssignedInMethod() && reg->getState() != TR::RealRegister::Locked)
{
cursor = new (trHeapMemory()) TR::X86RegInstruction(cursor, PUSHReg, reg, cg());
}
}
}
else
{
for (int32_t pindex = getProperties().getMaxRegistersPreservedInPrologue()-1;
pindex >= 0;
pindex--)
{
TR::RealRegister::RegNum idx = _properties.getPreservedRegister((uint32_t)pindex);
TR::RealRegister *reg = machine()->getX86RealRegister(getProperties().getPreservedRegister((uint32_t)pindex));
if(reg->getHasBeenAssignedInMethod() && reg->getState() != TR::RealRegister::Locked)
{
cursor = generateMemRegInstruction(
cursor,
TR::Linkage::movOpcodes(MemReg, fullRegisterMovType(reg)),
generateX86MemoryReference(machine()->getX86RealRegister(TR::RealRegister::vfp), offsetCursor, cg()),
reg,
cg()
);
offsetCursor -= pointerSize;
}
}
}
return cursor;
}
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;
}
// 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;
}
// Copies parameters from where they enter the method (either on stack or in a
// linkage register) to their "home location" where the method body will expect
// to find them (either on stack or in a global register).
//
TR::Instruction *
TR::X86SystemLinkage::copyParametersToHomeLocation(TR::Instruction *cursor)
{
TR::Machine *machine = cg()->machine();
TR::RealRegister *framePointer = machine->getX86RealRegister(TR::RealRegister::vfp);
TR::ResolvedMethodSymbol *bodySymbol = comp()->getJittedMethodSymbol();
ListIterator<TR::ParameterSymbol> paramIterator(&(bodySymbol->getParameterList()));
TR::ParameterSymbol *paramCursor;
const TR::RealRegister::RegNum noReg = TR::RealRegister::NoReg;
TR_ASSERT(noReg == 0, "noReg must be zero so zero-initializing movStatus will work");
TR::MovStatus movStatus[TR::RealRegister::NumRegisters] = {{(TR::RealRegister::RegNum)0,(TR::RealRegister::RegNum)0,(TR_MovDataTypes)0}};
// We must always do the stores first, then the reg-reg copies, then the
// loads, so that we never clobber a register we will need later. However,
// the logic is simpler if we do the loads and stores in the same loop.
// Therefore, we maintain a separate instruction cursor for the loads.
//
// We defer the initialization of loadCursor until we generate the first
// load. Otherwise, if we happen to generate some stores first, then the
// store cursor would get ahead of the loadCursor, and the instructions
// would end up in the wrong order despite our efforts.
//
TR::Instruction *loadCursor = NULL;
// Phase 1: generate RegMem and MemReg movs, and collect information about
// the required RegReg movs.
//
for (paramCursor = paramIterator.getFirst();
paramCursor != NULL;
paramCursor = paramIterator.getNext())
{
int8_t lri = paramCursor->getLinkageRegisterIndex(); // How the parameter enters the method
TR::RealRegister::RegNum ai // Where method body expects to find it
= (TR::RealRegister::RegNum)paramCursor->getAllocatedIndex();
int32_t offset = paramCursor->getParameterOffset(); // Location of the parameter's stack slot
TR_MovDataTypes movDataType = paramMovType(paramCursor); // What sort of MOV instruction does it need?
// Copy the parameter to wherever it should be
//
if (lri == NOT_LINKAGE) // It's on the stack
{
if (ai == NOT_ASSIGNED) // It only needs to be on the stack
{
// Nothing to do
}
else // Method body expects it to be in the ai register
{
if (loadCursor == NULL)
loadCursor = cursor;
if (debug("traceCopyParametersToHomeLocation"))
diagnostic("copyParametersToHomeLocation: Loading %d\n", ai);
// ai := stack
loadCursor = generateRegMemInstruction(
loadCursor,
TR::Linkage::movOpcodes(RegMem, movDataType),
machine->getX86RealRegister(ai),
generateX86MemoryReference(framePointer, offset, cg()),
cg()
);
}
}
else // It's in a linkage register
{
TR::RealRegister::RegNum sourceIndex = getProperties().getArgumentRegister(lri, isFloat(movDataType));
// Copy to the stack if necessary
//
if (ai == NOT_ASSIGNED || hasToBeOnStack(paramCursor))
{
if (comp()->getOption(TR_TraceCG))
traceMsg(comp(), "copyToHomeLocation param %p, linkage reg index %d, allocated index %d, parameter offset %d, hasToBeOnStack %d, parm->isParmHasToBeOnStack() %d.\n", paramCursor, lri, ai, offset, hasToBeOnStack(paramCursor), paramCursor->isParmHasToBeOnStack());
if (debug("traceCopyParametersToHomeLocation"))
diagnostic("copyParametersToHomeLocation: Storing %d\n", sourceIndex);
// stack := lri
cursor = generateMemRegInstruction(
cursor,
TR::Linkage::movOpcodes(MemReg, movDataType),
generateX86MemoryReference(framePointer, offset, cg()),
machine->getX86RealRegister(sourceIndex),
cg()
);
}
// Copy to the ai register if necessary
//
if (ai != NOT_ASSIGNED && ai != sourceIndex)
{
// This parameter needs a RegReg move. We don't know yet whether
// we need the value in the target register, so for now we just
// remember that we need to do this and keep going.
//
TR_ASSERT(movStatus[ai ].sourceReg == noReg, "Each target reg must have only one source");
TR_ASSERT(movStatus[sourceIndex].targetReg == noReg, "Each source reg must have only one target");
if (debug("traceCopyParametersToHomeLocation"))
diagnostic("copyParametersToHomeLocation: Planning to move %d to %d\n", sourceIndex, ai);
movStatus[ai].sourceReg = sourceIndex;
movStatus[sourceIndex].targetReg = ai;
movStatus[sourceIndex].outgoingDataType = movDataType;
}
if (debug("traceCopyParametersToHomeLocation") && ai == sourceIndex)
{
diagnostic("copyParametersToHomeLocation: Parameter #%d already in register %d\n", lri, ai);
}
}
}
// Phase 2: Iterate through the parameters again to insert the RegReg moves.
//
for (paramCursor = paramIterator.getFirst();
paramCursor != NULL;
paramCursor = paramIterator.getNext())
{
if (paramCursor->getLinkageRegisterIndex() == NOT_LINKAGE)
continue;
const TR::RealRegister::RegNum paramReg =
getProperties().getArgumentRegister(paramCursor->getLinkageRegisterIndex(), isFloat(paramMovType(paramCursor)));
if (movStatus[paramReg].targetReg == 0)
{
// This parameter does not need to be copied anywhere
if (debug("traceCopyParametersToHomeLocation"))
diagnostic("copyParametersToHomeLocation: Not moving %d\n", paramReg);
}
else
{
if (debug("traceCopyParametersToHomeLocation"))
diagnostic("copyParametersToHomeLocation: Preparing to move %d\n", paramReg);
// If a mov's target register is the source for another mov, we need
// to do that other mov first. The idea is to find the end point of
// the chain of movs starting with paramReg and ending with a
// register whose current value is not needed; then do that chain of
// movs in reverse order.
//
TR_ASSERT(noReg == 0, "noReg must be zero (not %d) for zero-filled initialization to work", noReg);
TR::RealRegister::RegNum regCursor;
// Find the last target in the chain
//
regCursor = movStatus[paramReg].targetReg;
while(movStatus[regCursor].targetReg != noReg)
{
// Haven't found the end yet
regCursor = movStatus[regCursor].targetReg;
TR_ASSERT(regCursor != paramReg, "Can't yet handle cyclic dependencies");
// TODO:AMD64 Use scratch register to break cycles
// A properly-written pickRegister should never
// cause cycles to occur in the first place. However, we may want
// to consider adding cycle-breaking logic so that (1) pickRegister
// has more flexibility, and (2) we're more robust against
// otherwise harmless bugs in pickRegister.
}
// Work our way backward along the chain, generating all the necessary movs
//
while(movStatus[regCursor].sourceReg != noReg)
{
TR::RealRegister::RegNum source = movStatus[regCursor].sourceReg;
if (debug("traceCopyParametersToHomeLocation"))
diagnostic("copyParametersToHomeLocation: Moving %d to %d\n", source, regCursor);
// regCursor := regCursor.sourceReg
cursor = generateRegRegInstruction(
cursor,
TR::Linkage::movOpcodes(RegReg, movStatus[source].outgoingDataType),
machine->getX86RealRegister(regCursor),
machine->getX86RealRegister(source),
cg()
);
// Update movStatus as we go so we don't generate redundant movs
movStatus[regCursor].sourceReg = noReg;
movStatus[source ].targetReg = noReg;
// Continue with the next register in the chain
regCursor = source;
}
}
}
// Return the last instruction we inserted, whether or not it was a load.
//
return loadCursor? loadCursor : cursor;
}
TR::Register *IA32LinkageUtils::pushDoubleArg(
TR::Node *child,
TR::CodeGenerator *cg)
{
TR::Register *pushRegister;
if (child->getRegister() == NULL)
{
if (child->getOpCodeValue() == TR::dconst)
{
TR_X86OpCodes pushOp;
int32_t highValue = child->getLongIntHigh();
if (highValue >= -128 && highValue <= 127)
{
pushOp = PUSHImms;
}
else
{
pushOp = PUSHImm4;
}
generateImmInstruction(pushOp, child, highValue, cg);
int32_t lowValue = child->getLongIntLow();
if (lowValue >= -128 && lowValue <= 127)
{
pushOp = PUSHImms;
}
else
{
pushOp = PUSHImm4;
}
generateImmInstruction(pushOp, child, lowValue, cg);
cg->decReferenceCount(child);
return NULL;
}
else if (child->getReferenceCount() == 1)
{
if (child->getOpCode().isLoad())
{
TR::MemoryReference *lowMR = generateX86MemoryReference(child, cg);
generateMemInstruction(PUSHMem, child, generateX86MemoryReference(*lowMR, 4, cg), cg);
generateMemInstruction(PUSHMem, child, lowMR, cg);
lowMR->decNodeReferenceCounts(cg);
cg->decReferenceCount(child);
return NULL;
}
else if (child->getOpCodeValue() == TR::lbits2d)
{
pushRegister = pushLongArg(child->getFirstChild(), cg);
cg->decReferenceCount(child);
return pushRegister;
}
}
}
pushRegister = cg->evaluate(child);
TR::RealRegister *espReal = cg->machine()->getRealRegister(TR::RealRegister::esp);
generateRegImmInstruction(SUB4RegImms, child, espReal, 8, cg);
if (cg->useSSEForSinglePrecision() && pushRegister->getKind() == TR_FPR)
generateMemRegInstruction(MOVSDMemReg, child, generateX86MemoryReference(espReal, 0, cg), pushRegister, cg);
else
generateFPMemRegInstruction(DSTMemReg, child, generateX86MemoryReference(espReal, 0, cg), pushRegister, cg);
cg->decReferenceCount(child);
return pushRegister;
}