int32_t TR::ARM64SystemLinkage::buildArgs(TR::Node *callNode,
TR::RegisterDependencyConditions *dependencies)
{
const TR::ARM64LinkageProperties &properties = getProperties();
TR::ARM64MemoryArgument *pushToMemory = NULL;
TR::Register *argMemReg;
TR::Register *tempReg;
int32_t argIndex = 0;
int32_t numMemArgs = 0;
int32_t argSize = 0;
int32_t numIntegerArgs = 0;
int32_t numFloatArgs = 0;
int32_t totalSize;
int32_t i;
TR::Node *child;
TR::DataType childType;
TR::DataType resType = callNode->getType();
uint32_t firstArgumentChild = callNode->getFirstArgumentIndex();
/* Step 1 - figure out how many arguments are going to be spilled to memory i.e. not in registers */
for (i = firstArgumentChild; i < callNode->getNumChildren(); i++)
{
child = callNode->getChild(i);
childType = child->getDataType();
switch (childType)
{
case TR::Int8:
case TR::Int16:
case TR::Int32:
case TR::Int64:
case TR::Address:
if (numIntegerArgs >= properties.getNumIntArgRegs())
numMemArgs++;
numIntegerArgs++;
break;
case TR::Float:
case TR::Double:
if (numFloatArgs >= properties.getNumFloatArgRegs())
numMemArgs++;
numFloatArgs++;
break;
default:
TR_ASSERT(false, "Argument type %s is not supported\n", childType.toString());
}
}
// From here, down, any new stack allocations will expire / die when the function returns
TR::StackMemoryRegion stackMemoryRegion(*trMemory());
/* End result of Step 1 - determined number of memory arguments! */
if (numMemArgs > 0)
{
pushToMemory = new (trStackMemory()) TR::ARM64MemoryArgument[numMemArgs];
argMemReg = cg()->allocateRegister();
}
totalSize = numMemArgs * 8;
// align to 16-byte boundary
totalSize = (totalSize + 15) & (~15);
numIntegerArgs = 0;
numFloatArgs = 0;
for (i = firstArgumentChild; i < callNode->getNumChildren(); i++)
{
TR::MemoryReference *mref = NULL;
TR::Register *argRegister;
TR::InstOpCode::Mnemonic op;
child = callNode->getChild(i);
childType = child->getDataType();
switch (childType)
{
case TR::Int8:
case TR::Int16:
case TR::Int32:
case TR::Int64:
case TR::Address:
if (childType == TR::Address)
argRegister = pushAddressArg(child);
else if (childType == TR::Int64)
argRegister = pushLongArg(child);
else
argRegister = pushIntegerWordArg(child);
if (numIntegerArgs < properties.getNumIntArgRegs())
{
if (!cg()->canClobberNodesRegister(child, 0))
{
if (argRegister->containsCollectedReference())
tempReg = cg()->allocateCollectedReferenceRegister();
else
tempReg = cg()->allocateRegister();
generateMovInstruction(cg(), callNode, tempReg, argRegister);
argRegister = tempReg;
}
if (numIntegerArgs == 0 &&
(resType.isAddress() || resType.isInt32() || resType.isInt64()))
{
TR::Register *resultReg;
if (resType.isAddress())
resultReg = cg()->allocateCollectedReferenceRegister();
else
resultReg = cg()->allocateRegister();
dependencies->addPreCondition(argRegister, TR::RealRegister::x0);
dependencies->addPostCondition(resultReg, TR::RealRegister::x0);
}
else
{
addDependency(dependencies, argRegister, properties.getIntegerArgumentRegister(numIntegerArgs), TR_GPR, cg());
}
}
else
{
// numIntegerArgs >= properties.getNumIntArgRegs()
if (childType == TR::Address || childType == TR::Int64)
{
op = TR::InstOpCode::strpostx;
}
else
{
op = TR::InstOpCode::strpostw;
}
mref = getOutgoingArgumentMemRef(argMemReg, argRegister, op, pushToMemory[argIndex++]);
argSize += 8; // always 8-byte aligned
}
numIntegerArgs++;
break;
case TR::Float:
case TR::Double:
if (childType == TR::Float)
argRegister = pushFloatArg(child);
else
argRegister = pushDoubleArg(child);
if (numFloatArgs < properties.getNumFloatArgRegs())
{
if (!cg()->canClobberNodesRegister(child, 0))
{
tempReg = cg()->allocateRegister(TR_FPR);
op = (childType == TR::Float) ? TR::InstOpCode::fmovs : TR::InstOpCode::fmovd;
generateTrg1Src1Instruction(cg(), op, callNode, tempReg, argRegister);
argRegister = tempReg;
}
if ((numFloatArgs == 0 && resType.isFloatingPoint()))
{
TR::Register *resultReg;
if (resType.getDataType() == TR::Float)
resultReg = cg()->allocateSinglePrecisionRegister();
else
resultReg = cg()->allocateRegister(TR_FPR);
dependencies->addPreCondition(argRegister, TR::RealRegister::v0);
dependencies->addPostCondition(resultReg, TR::RealRegister::v0);
}
else
{
addDependency(dependencies, argRegister, properties.getFloatArgumentRegister(numFloatArgs), TR_FPR, cg());
}
}
else
{
// numFloatArgs >= properties.getNumFloatArgRegs()
if (childType == TR::Double)
{
op = TR::InstOpCode::vstrpostd;
}
else
{
op = TR::InstOpCode::vstrposts;
}
mref = getOutgoingArgumentMemRef(argMemReg, argRegister, op, pushToMemory[argIndex++]);
argSize += 8; // always 8-byte aligned
}
numFloatArgs++;
break;
} // end of switch
} // end of for
// NULL deps for non-preserved and non-system regs
while (numIntegerArgs < properties.getNumIntArgRegs())
{
if (numIntegerArgs == 0 && resType.isAddress())
{
dependencies->addPreCondition(cg()->allocateRegister(), properties.getIntegerArgumentRegister(0));
dependencies->addPostCondition(cg()->allocateCollectedReferenceRegister(), properties.getIntegerArgumentRegister(0));
}
else
{
addDependency(dependencies, NULL, properties.getIntegerArgumentRegister(numIntegerArgs), TR_GPR, cg());
}
numIntegerArgs++;
}
int32_t floatRegsUsed = (numFloatArgs > properties.getNumFloatArgRegs()) ? properties.getNumFloatArgRegs() : numFloatArgs;
for (i = (TR::RealRegister::RegNum)((uint32_t)TR::RealRegister::v0 + floatRegsUsed); i <= TR::RealRegister::LastFPR; i++)
{
if (!properties.getPreserved((TR::RealRegister::RegNum)i))
{
// NULL dependency for non-preserved regs
addDependency(dependencies, NULL, (TR::RealRegister::RegNum)i, TR_FPR, cg());
}
}
if (numMemArgs > 0)
{
TR::RealRegister *sp = cg()->machine()->getRealRegister(properties.getStackPointerRegister());
generateTrg1Src1ImmInstruction(cg(), TR::InstOpCode::subimmx, callNode, argMemReg, sp, totalSize);
for (argIndex = 0; argIndex < numMemArgs; argIndex++)
{
TR::Register *aReg = pushToMemory[argIndex].argRegister;
generateMemSrc1Instruction(cg(), pushToMemory[argIndex].opCode, callNode, pushToMemory[argIndex].argMemory, aReg);
cg()->stopUsingRegister(aReg);
}
cg()->stopUsingRegister(argMemReg);
}
return totalSize;
}
// 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;
}
OMR::Power::RegisterDependencyConditions::RegisterDependencyConditions(
TR::CodeGenerator *cg,
TR::Node *node,
uint32_t extranum,
TR::Instruction **cursorPtr)
{
List<TR::Register> regList(cg->trMemory());
TR::Instruction *iCursor = (cursorPtr==NULL)?NULL:*cursorPtr;
int32_t totalNum = node->getNumChildren() + extranum;
int32_t i;
cg->comp()->incVisitCount();
int32_t numLongs = 0;
//
// Pre-compute how many longs are global register candidates
//
for (i = 0; i < node->getNumChildren(); ++i)
{
TR::Node *child = node->getChild(i);
TR::Register *reg = child->getRegister();
if (reg!=NULL /* && reg->getKind()==TR_GPR */)
{
if (child->getHighGlobalRegisterNumber() > -1)
numLongs++;
}
}
totalNum = totalNum + numLongs;
_preConditions = new (totalNum, cg->trMemory()) TR_PPCRegisterDependencyGroup;
_postConditions = new (totalNum, cg->trMemory()) TR_PPCRegisterDependencyGroup;
_numPreConditions = totalNum;
_addCursorForPre = 0;
_numPostConditions = totalNum;
_addCursorForPost = 0;
// First, handle dependencies that match current association
for (i=0; i<node->getNumChildren(); i++)
{
TR::Node *child = node->getChild(i);
TR::Register *reg = child->getRegister();
TR::Register *highReg = NULL;
TR::RealRegister::RegNum regNum = (TR::RealRegister::RegNum)cg->getGlobalRegister(child->getGlobalRegisterNumber());
TR::RealRegister::RegNum highRegNum;
if (child->getHighGlobalRegisterNumber() > -1)
{
highRegNum = (TR::RealRegister::RegNum)cg->getGlobalRegister(child->getHighGlobalRegisterNumber());
TR::RegisterPair *regPair = reg->getRegisterPair();
TR_ASSERT(regPair, "assertion failure");
highReg = regPair->getHighOrder();
reg = regPair->getLowOrder();
if (highReg->getAssociation() != highRegNum ||
reg->getAssociation() != regNum)
continue;
}
else if (reg->getAssociation() != regNum)
continue;
TR_ASSERT(!regList.find(reg) && (!highReg || !regList.find(highReg)), "Should not happen\n");
addPreCondition(reg, regNum);
addPostCondition(reg, regNum);
regList.add(reg);
if (highReg)
{
addPreCondition(highReg, highRegNum);
addPostCondition(highReg, highRegNum);
regList.add(highReg);
}
}
// Second pass to handle dependencies for which association does not exist
// or does not match
for (i=0; i<node->getNumChildren(); i++)
{
TR::Node *child = node->getChild(i);
TR::Register *reg = child->getRegister();
TR::Register *highReg = NULL;
TR::Register *copyReg = NULL;
TR::Register *highCopyReg = NULL;
TR::RealRegister::RegNum regNum = (TR::RealRegister::RegNum)cg->getGlobalRegister(child->getGlobalRegisterNumber());
TR::RealRegister::RegNum highRegNum;
if (child->getHighGlobalRegisterNumber() > -1)
{
highRegNum = (TR::RealRegister::RegNum)cg->getGlobalRegister(child->getHighGlobalRegisterNumber());
TR::RegisterPair *regPair = reg->getRegisterPair();
TR_ASSERT(regPair, "assertion failure");
highReg = regPair->getHighOrder();
reg = regPair->getLowOrder();
if (highReg->getAssociation() == highRegNum &&
reg->getAssociation() == regNum)
continue;
}
else if (reg->getAssociation() == regNum)
continue;
if (regList.find(reg) || (highReg && regList.find(highReg)))
{
TR::InstOpCode::Mnemonic opCode;
TR_RegisterKinds kind = reg->getKind();
switch (kind)
{
case TR_GPR:
opCode = TR::InstOpCode::mr;
break;
case TR_FPR:
opCode = TR::InstOpCode::fmr;
break;
case TR_VRF:
opCode = TR::InstOpCode::vor;
//TR_ASSERT(0, "VMX not fully supported.");
break;
case TR_VSX_VECTOR:
opCode = TR::InstOpCode::xxlor;
break;
case TR_CCR:
opCode = TR::InstOpCode::mcrf;
break;
default:
TR_ASSERT(0, "Invalid register kind.");
}
if (regList.find(reg))
{
bool containsInternalPointer = false;
if (reg->getPinningArrayPointer())
containsInternalPointer = true;
copyReg = (reg->containsCollectedReference() && !containsInternalPointer) ?
cg->allocateCollectedReferenceRegister() : cg->allocateRegister(kind);
if (containsInternalPointer)
{
copyReg->setContainsInternalPointer();
copyReg->setPinningArrayPointer(reg->getPinningArrayPointer());
}
if (opCode == TR::InstOpCode::vor || opCode == TR::InstOpCode::xxlor)
iCursor = generateTrg1Src2Instruction(cg, opCode, node, copyReg, reg, reg, iCursor);
else
iCursor = generateTrg1Src1Instruction(cg, opCode, node, copyReg, reg, iCursor);
reg = copyReg;
}
if (highReg && regList.find(highReg))
{
bool containsInternalPointer = false;
if (highReg->getPinningArrayPointer())
containsInternalPointer = true;
highCopyReg = (highReg->containsCollectedReference() && !containsInternalPointer) ?
cg->allocateCollectedReferenceRegister() : cg->allocateRegister(kind);
if (containsInternalPointer)
{
highCopyReg->setContainsInternalPointer();
highCopyReg->setPinningArrayPointer(highReg->getPinningArrayPointer());
}
if (opCode == TR::InstOpCode::vor || opCode == TR::InstOpCode::xxlor)
iCursor = generateTrg1Src2Instruction(cg, opCode, node, highCopyReg, highReg, highReg, iCursor);
else
iCursor = generateTrg1Src1Instruction(cg, opCode, node, highCopyReg, highReg, iCursor);
highReg = highCopyReg;
}
}
addPreCondition(reg, regNum);
addPostCondition(reg, regNum);
if (copyReg != NULL)
cg->stopUsingRegister(copyReg);
else
regList.add(reg);
if (highReg)
{
addPreCondition(highReg, highRegNum);
addPostCondition(highReg, highRegNum);
if (highCopyReg != NULL)
cg->stopUsingRegister(highCopyReg);
else
regList.add(highReg);
}
}
if (iCursor!=NULL && cursorPtr!=NULL)
*cursorPtr = iCursor;
}