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;
}
TR::Instruction *OMR::Power::Linkage::saveArguments(TR::Instruction *cursor, bool fsd, bool saveOnly,
List<TR::ParameterSymbol> &parmList)
{
#define REAL_REGISTER(ri) machine->getRealRegister(ri)
#define REGNUM(ri) ((TR::RealRegister::RegNum)(ri))
const TR::PPCLinkageProperties& properties = self()->getProperties();
TR::Machine *machine = self()->machine();
TR::RealRegister *stackPtr = self()->cg()->getStackPointerRegister();
TR::ResolvedMethodSymbol *bodySymbol = self()->comp()->getJittedMethodSymbol();
ListIterator<TR::ParameterSymbol> paramIterator(&parmList);
TR::ParameterSymbol *paramCursor;
TR::Node *firstNode = self()->comp()->getStartTree()->getNode();
TR_BitVector freeScratchable;
int32_t busyMoves[3][64];
int32_t busyIndex = 0, i1;
bool all_saved = false;
// the freeScratchable structure will not be used when saveOnly == true
// no additional conditions were added with the intention of keeping the code easier to read
// and not full of if conditions
freeScratchable.init(TR::RealRegister::LastFPR + 1, self()->trMemory());
// first, consider all argument registers free
for (i1=TR::RealRegister::FirstGPR; i1<=TR::RealRegister::LastFPR; i1++)
{
if (!properties.getReserved(REGNUM(i1)))
{
freeScratchable.set(i1);
}
}
// second, go through all parameters and reset registers that are actually used
for (paramCursor=paramIterator.getFirst(); paramCursor!=NULL; paramCursor=paramIterator.getNext())
{
int32_t lri = paramCursor->getLinkageRegisterIndex();
TR::DataType type = paramCursor->getType();
if (lri >= 0)
{
TR::RealRegister::RegNum regNum;
bool twoRegs = (TR::Compiler->target.is32Bit() && type.isInt64() && lri < properties.getNumIntArgRegs()-1);
if (!type.isFloatingPoint())
{
regNum = properties.getIntegerArgumentRegister(lri);
if (paramCursor->isReferencedParameter()) freeScratchable.reset(regNum);
if (twoRegs)
if (paramCursor->isReferencedParameter()) freeScratchable.reset(regNum+1);
}
else
{
regNum = properties.getFloatArgumentRegister(lri);
if (paramCursor->isReferencedParameter()) freeScratchable.reset(regNum);
if (twoRegs)
if (paramCursor->isReferencedParameter()) freeScratchable.reset(regNum+1);
}
}
}
for (paramCursor=paramIterator.getFirst(); paramCursor!=NULL; paramCursor=paramIterator.getNext())
{
int32_t lri = paramCursor->getLinkageRegisterIndex();
int32_t ai = paramCursor->getAllocatedIndex();
int32_t offset = self()->calculateParameterRegisterOffset(paramCursor->getParameterOffset(), *paramCursor);
TR::DataType type = paramCursor->getType();
int32_t dtype = type.getDataType();
// TODO: Is there an accurate assume to insert here ?
if (lri >= 0)
{
if (!paramCursor->isReferencedParameter() && !paramCursor->isParmHasToBeOnStack()) continue;
TR::RealRegister::RegNum regNum;
bool twoRegs = (TR::Compiler->target.is32Bit() && type.isInt64() && lri < properties.getNumIntArgRegs()-1);
if (type.isFloatingPoint())
regNum = properties.getFloatArgumentRegister(lri);
else
regNum = properties.getIntegerArgumentRegister(lri);
// Do not save arguments to the stack if in Full Speed Debug and saveOnly is not set.
// If not in Full Speed Debug, the arguments will be saved.
if (((ai<0 || self()->hasToBeOnStack(paramCursor)) && !fsd) || (fsd && saveOnly))
{
switch (dtype)
{
case TR::Int8:
case TR::Int16:
case TR::Int32:
{
TR::InstOpCode::Mnemonic op = TR::InstOpCode::stw;
if (!all_saved) cursor = generateMemSrc1Instruction(self()->cg(), op, firstNode,
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset, 4, self()->cg()), REAL_REGISTER(regNum), cursor);
}
break;
case TR::Address:
if (!all_saved) cursor = generateMemSrc1Instruction(self()->cg(),TR::InstOpCode::Op_st, firstNode,
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset, TR::Compiler->om.sizeofReferenceAddress(), self()->cg()), REAL_REGISTER(regNum), cursor);
break;
case TR::Int64:
if (!all_saved) cursor = generateMemSrc1Instruction(self()->cg(),TR::InstOpCode::Op_st, firstNode,
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset, TR::Compiler->om.sizeofReferenceAddress(), self()->cg()), REAL_REGISTER(regNum), cursor);
if (twoRegs)
{
if (!all_saved) cursor = generateMemSrc1Instruction(self()->cg(), TR::InstOpCode::stw, firstNode,
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset+4, 4, self()->cg()),
REAL_REGISTER(REGNUM(regNum+1)), cursor);
if (ai<0)
freeScratchable.set(regNum+1);
}
break;
case TR::Float:
cursor = generateMemSrc1Instruction(self()->cg(), TR::InstOpCode::stfs, firstNode,
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset, 4, self()->cg()),
REAL_REGISTER(regNum), cursor);
break;
case TR::Double:
cursor = generateMemSrc1Instruction(self()->cg(), TR::InstOpCode::stfd, firstNode,
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset, 8, self()->cg()),
REAL_REGISTER(regNum), cursor);
break;
default:
TR_ASSERT(false, "assertion failure");
break;
}
if (ai<0)
freeScratchable.set(regNum);
}
// Global register is allocated to this argument.
// Don't process if in Full Speed Debug and saveOnly is set
if (ai>=0 && (!fsd || !saveOnly))
{
if (regNum != ai) // Equal assignment: do nothing
{
if (freeScratchable.isSet(ai))
{
cursor = generateTrg1Src1Instruction(self()->cg(),
(type.isFloatingPoint()) ? TR::InstOpCode::fmr:TR::InstOpCode::mr,
firstNode, REAL_REGISTER(REGNUM(ai)), REAL_REGISTER(regNum), cursor);
freeScratchable.reset(ai);
freeScratchable.set(regNum);
}
else // The status of target global register is unclear (i.e. it is a arg reg)
{
busyMoves[0][busyIndex] = regNum;
busyMoves[1][busyIndex] = ai;
busyMoves[2][busyIndex] = 0;
busyIndex++;
}
}
if (TR::Compiler->target.is32Bit() && type.isInt64())
{
int32_t aiLow = paramCursor->getAllocatedLow();
if (!twoRegs) // Low part needs to come from memory
{
offset += 4; // We are dealing with the low part
if (freeScratchable.isSet(aiLow))
{
cursor = generateTrg1MemInstruction(self()->cg(), TR::InstOpCode::lwz, firstNode, REAL_REGISTER(REGNUM(aiLow)),
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset, 4, self()->cg()), cursor);
freeScratchable.reset(aiLow);
}
else
{
busyMoves[0][busyIndex] = offset;
busyMoves[1][busyIndex] = aiLow;
busyMoves[2][busyIndex] = 1;
busyIndex++;
}
}
else if (regNum+1 != aiLow) // Low part needs to be moved
{
if (freeScratchable.isSet(aiLow))
{
cursor = generateTrg1Src1Instruction(self()->cg(), TR::InstOpCode::mr,
firstNode, REAL_REGISTER(REGNUM(aiLow)),
REAL_REGISTER(REGNUM(regNum+1)), cursor);
freeScratchable.reset(aiLow);
freeScratchable.set(regNum+1);
}
else
{
busyMoves[0][busyIndex] = regNum+1;
busyMoves[1][busyIndex] = aiLow;
busyMoves[2][busyIndex] = 0;
busyIndex++;
}
}
}
}
}
// Don't process if in Full Speed Debug and saveOnly is set
else if (ai >= 0 && (!fsd || !saveOnly)) // lri<0: arg needs to come from memory
{
switch (dtype)
{
case TR::Int8:
case TR::Int16:
case TR::Int32:
if (freeScratchable.isSet(ai))
{
cursor = generateTrg1MemInstruction(self()->cg(), TR::InstOpCode::lwz, firstNode, REAL_REGISTER(REGNUM(ai)),
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset, 4, self()->cg()), cursor);
freeScratchable.reset(ai);
}
else
{
busyMoves[0][busyIndex] = offset;
busyMoves[1][busyIndex] = ai;
busyMoves[2][busyIndex] = 1;
busyIndex++;
}
break;
case TR::Address:
if (freeScratchable.isSet(ai))
{
cursor = generateTrg1MemInstruction(self()->cg(),TR::InstOpCode::Op_load, firstNode, REAL_REGISTER(REGNUM(ai)),
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset, TR::Compiler->om.sizeofReferenceAddress(), self()->cg()), cursor);
freeScratchable.reset(ai);
}
else
{
busyMoves[0][busyIndex] = offset;
busyMoves[1][busyIndex] = ai;
if (TR::Compiler->target.is64Bit())
busyMoves[2][busyIndex] = 2;
else
busyMoves[2][busyIndex] = 1;
busyIndex++;
}
break;
case TR::Int64:
if (TR::Compiler->target.is64Bit())
{
if (freeScratchable.isSet(ai))
{
cursor = generateTrg1MemInstruction(self()->cg(), TR::InstOpCode::ld, firstNode, REAL_REGISTER(REGNUM(ai)),
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset, 8, self()->cg()), cursor);
freeScratchable.reset(ai);
}
else
{
busyMoves[0][busyIndex] = offset;
busyMoves[1][busyIndex] = ai;
busyMoves[2][busyIndex] = 2;
busyIndex++;
}
}
else // 32-bit
{
if (freeScratchable.isSet(ai))
{
cursor = generateTrg1MemInstruction(self()->cg(), TR::InstOpCode::lwz, firstNode, REAL_REGISTER(REGNUM(ai)),
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset, 4, self()->cg()), cursor);
freeScratchable.reset(ai);
}
else
{
busyMoves[0][busyIndex] = offset;
busyMoves[1][busyIndex] = ai;
busyMoves[2][busyIndex] = 1;
busyIndex++;
}
ai = paramCursor->getAllocatedLow();
if (freeScratchable.isSet(ai))
{
cursor = generateTrg1MemInstruction(self()->cg(), TR::InstOpCode::lwz, firstNode, REAL_REGISTER(REGNUM(ai)),
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset+4, 4, self()->cg()), cursor);
freeScratchable.reset(ai);
}
else
{
busyMoves[0][busyIndex] = offset+4;
busyMoves[1][busyIndex] = ai;
busyMoves[2][busyIndex] = 1;
busyIndex++;
}
}
break;
case TR::Float:
if (freeScratchable.isSet(ai))
{
cursor = generateTrg1MemInstruction(self()->cg(), TR::InstOpCode::lfs, firstNode, REAL_REGISTER(REGNUM(ai)),
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset, 4, self()->cg()), cursor);
freeScratchable.reset(ai);
}
else
{
busyMoves[0][busyIndex] = offset;
busyMoves[1][busyIndex] = ai;
busyMoves[2][busyIndex] = 3;
busyIndex++;
}
break;
case TR::Double:
if (freeScratchable.isSet(ai))
{
cursor = generateTrg1MemInstruction(self()->cg(), TR::InstOpCode::lfd, firstNode, REAL_REGISTER(REGNUM(ai)),
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset, 8, self()->cg()), cursor);
freeScratchable.reset(ai);
}
else
{
busyMoves[0][busyIndex] = offset;
busyMoves[1][busyIndex] = ai;
busyMoves[2][busyIndex] = 4;
busyIndex++;
}
break;
default:
break;
}
}
}
if (!fsd || !saveOnly)
{
bool freeMore = true;
int32_t numMoves = busyIndex;
while (freeMore && numMoves>0)
{
freeMore = false;
for (i1=0; i1<busyIndex; i1++)
{
int32_t source = busyMoves[0][i1];
int32_t target = busyMoves[1][i1];
if (!(target<0) && freeScratchable.isSet(target))
{
switch(busyMoves[2][i1])
{
case 0:
cursor = generateTrg1Src1Instruction(self()->cg(), (source<=TR::RealRegister::LastGPR)?TR::InstOpCode::mr:TR::InstOpCode::fmr,
firstNode, REAL_REGISTER(REGNUM(target)), REAL_REGISTER(REGNUM(source)), cursor);
freeScratchable.set(source);
break;
case 1:
cursor = generateTrg1MemInstruction(self()->cg(), TR::InstOpCode::lwz, firstNode, REAL_REGISTER(REGNUM(target)),
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, source, 4, self()->cg()), cursor);
break;
case 2:
cursor = generateTrg1MemInstruction(self()->cg(), TR::InstOpCode::ld, firstNode, REAL_REGISTER(REGNUM(target)),
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, source, 8, self()->cg()), cursor);
break;
case 3:
cursor = generateTrg1MemInstruction(self()->cg(), TR::InstOpCode::lfs, firstNode, REAL_REGISTER(REGNUM(target)),
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, source, 4, self()->cg()), cursor);
break;
case 4:
cursor = generateTrg1MemInstruction(self()->cg(), TR::InstOpCode::lfd, firstNode, REAL_REGISTER(REGNUM(target)),
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, source, 8, self()->cg()), cursor);
break;
}
freeScratchable.reset(target);
freeMore = true;
busyMoves[0][i1] = busyMoves[1][i1] = -1;
numMoves--;
}
}
}
TR_ASSERT(numMoves<=0, "Circular argument register dependency can and should be avoided.");
}
return(cursor);
}