/// InstructionSelectBasicBlock - This callback is invoked by
/// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
void IA64DAGToDAGISel::InstructionSelectBasicBlock(SelectionDAG &DAG) {
DEBUG(BB->dump());
// Select target instructions for the DAG.
DAG.setRoot(SelectRoot(DAG.getRoot()));
DAG.RemoveDeadNodes();
// Emit machine code to BB.
ScheduleAndEmitDAG(DAG);
}
// After instruction selection, insert COPY_TO_REGCLASS nodes to help in
// choosing the proper register classes.
void BlackfinDAGToDAGISel::FixRegisterClasses(SelectionDAG &DAG) {
const BlackfinInstrInfo &TII = getInstrInfo();
const BlackfinRegisterInfo *TRI = getRegisterInfo();
DAG.AssignTopologicalOrder();
HandleSDNode Dummy(DAG.getRoot());
for (SelectionDAG::allnodes_iterator NI = DAG.allnodes_begin();
NI != DAG.allnodes_end(); ++NI) {
if (NI->use_empty() || !NI->isMachineOpcode())
continue;
const TargetInstrDesc &DefTID = TII.get(NI->getMachineOpcode());
for (SDNode::use_iterator UI = NI->use_begin(); !UI.atEnd(); ++UI) {
if (!UI->isMachineOpcode())
continue;
if (UI.getUse().getResNo() >= DefTID.getNumDefs())
continue;
const TargetRegisterClass *DefRC =
DefTID.OpInfo[UI.getUse().getResNo()].getRegClass(TRI);
const TargetInstrDesc &UseTID = TII.get(UI->getMachineOpcode());
if (UseTID.getNumDefs()+UI.getOperandNo() >= UseTID.getNumOperands())
continue;
const TargetRegisterClass *UseRC =
UseTID.OpInfo[UseTID.getNumDefs()+UI.getOperandNo()].getRegClass(TRI);
if (!DefRC || !UseRC)
continue;
// We cannot copy CC <-> !(CC/D)
if ((isCC(DefRC) && !isDCC(UseRC)) || (isCC(UseRC) && !isDCC(DefRC))) {
SDNode *Copy =
DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
NI->getDebugLoc(),
MVT::i32,
UI.getUse().get(),
DAG.getTargetConstant(BF::DRegClassID, MVT::i32));
UpdateNodeOperand(DAG, *UI, UI.getOperandNo(), SDValue(Copy, 0));
}
}
}
DAG.setRoot(Dummy.getValue());
}
SDValue
NVPTXTargetLowering::LowerFormalArguments(SDValue Chain,
CallingConv::ID CallConv, bool isVarArg,
const SmallVectorImpl<ISD::InputArg> &Ins,
DebugLoc dl, SelectionDAG &DAG,
SmallVectorImpl<SDValue> &InVals) const {
MachineFunction &MF = DAG.getMachineFunction();
const DataLayout *TD = getDataLayout();
const Function *F = MF.getFunction();
const AttrListPtr &PAL = F->getAttributes();
SDValue Root = DAG.getRoot();
std::vector<SDValue> OutChains;
bool isKernel = llvm::isKernelFunction(*F);
bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
std::vector<Type *> argTypes;
std::vector<const Argument *> theArgs;
for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
I != E; ++I) {
theArgs.push_back(I);
argTypes.push_back(I->getType());
}
assert(argTypes.size() == Ins.size() &&
"Ins types and function types did not match");
int idx = 0;
for (unsigned i=0, e=Ins.size(); i!=e; ++i, ++idx) {
Type *Ty = argTypes[i];
EVT ObjectVT = getValueType(Ty);
assert(ObjectVT == Ins[i].VT &&
"Ins type did not match function type");
// If the kernel argument is image*_t or sampler_t, convert it to
// a i32 constant holding the parameter position. This can later
// matched in the AsmPrinter to output the correct mangled name.
if (isImageOrSamplerVal(theArgs[i],
(theArgs[i]->getParent() ?
theArgs[i]->getParent()->getParent() : 0))) {
assert(isKernel && "Only kernels can have image/sampler params");
InVals.push_back(DAG.getConstant(i+1, MVT::i32));
continue;
}
if (theArgs[i]->use_empty()) {
// argument is dead
InVals.push_back(DAG.getNode(ISD::UNDEF, dl, ObjectVT));
continue;
}
// In the following cases, assign a node order of "idx+1"
// to newly created nodes. The SDNOdes for params have to
// appear in the same order as their order of appearance
// in the original function. "idx+1" holds that order.
if (PAL.getParamAttributes(i+1).hasAttribute(Attributes::ByVal) == false) {
// A plain scalar.
if (isABI || isKernel) {
// If ABI, load from the param symbol
SDValue Arg = getParamSymbol(DAG, idx);
Value *srcValue = new Argument(PointerType::get(ObjectVT.getTypeForEVT(
F->getContext()),
llvm::ADDRESS_SPACE_PARAM));
SDValue p = DAG.getLoad(ObjectVT, dl, Root, Arg,
MachinePointerInfo(srcValue), false, false,
false,
TD->getABITypeAlignment(ObjectVT.getTypeForEVT(
F->getContext())));
if (p.getNode())
DAG.AssignOrdering(p.getNode(), idx+1);
InVals.push_back(p);
}
else {
// If no ABI, just move the param symbol
SDValue Arg = getParamSymbol(DAG, idx, ObjectVT);
SDValue p = DAG.getNode(NVPTXISD::MoveParam, dl, ObjectVT, Arg);
if (p.getNode())
DAG.AssignOrdering(p.getNode(), idx+1);
InVals.push_back(p);
}
continue;
}
// Param has ByVal attribute
if (isABI || isKernel) {
// Return MoveParam(param symbol).
// Ideally, the param symbol can be returned directly,
// but when SDNode builder decides to use it in a CopyToReg(),
// machine instruction fails because TargetExternalSymbol
// (not lowered) is target dependent, and CopyToReg assumes
// the source is lowered.
SDValue Arg = getParamSymbol(DAG, idx, getPointerTy());
SDValue p = DAG.getNode(NVPTXISD::MoveParam, dl, ObjectVT, Arg);
if (p.getNode())
DAG.AssignOrdering(p.getNode(), idx+1);
if (isKernel)
InVals.push_back(p);
else {
SDValue p2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, ObjectVT,
DAG.getConstant(Intrinsic::nvvm_ptr_local_to_gen, MVT::i32),
p);
InVals.push_back(p2);
}
} else {
// Have to move a set of param symbols to registers and
// store them locally and return the local pointer in InVals
const PointerType *elemPtrType = dyn_cast<PointerType>(argTypes[i]);
assert(elemPtrType &&
"Byval parameter should be a pointer type");
Type *elemType = elemPtrType->getElementType();
// Compute the constituent parts
SmallVector<EVT, 16> vtparts;
SmallVector<uint64_t, 16> offsets;
ComputeValueVTs(*this, elemType, vtparts, &offsets, 0);
unsigned totalsize = 0;
for (unsigned j=0, je=vtparts.size(); j!=je; ++j)
totalsize += vtparts[j].getStoreSizeInBits();
SDValue localcopy = DAG.getFrameIndex(MF.getFrameInfo()->
CreateStackObject(totalsize/8, 16, false),
getPointerTy());
unsigned sizesofar = 0;
std::vector<SDValue> theChains;
for (unsigned j=0, je=vtparts.size(); j!=je; ++j) {
unsigned numElems = 1;
if (vtparts[j].isVector()) numElems = vtparts[j].getVectorNumElements();
for (unsigned k=0, ke=numElems; k!=ke; ++k) {
EVT tmpvt = vtparts[j];
if (tmpvt.isVector()) tmpvt = tmpvt.getVectorElementType();
SDValue arg = DAG.getNode(NVPTXISD::MoveParam, dl, tmpvt,
getParamSymbol(DAG, idx, tmpvt));
SDValue addr = DAG.getNode(ISD::ADD, dl, getPointerTy(), localcopy,
DAG.getConstant(sizesofar, getPointerTy()));
theChains.push_back(DAG.getStore(Chain, dl, arg, addr,
MachinePointerInfo(), false, false, 0));
sizesofar += tmpvt.getStoreSizeInBits()/8;
++idx;
}
}
--idx;
Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &theChains[0],
theChains.size());
InVals.push_back(localcopy);
}
}
// Clang will check explicit VarArg and issue error if any. However, Clang
// will let code with
// implicit var arg like f() pass.
// We treat this case as if the arg list is empty.
//if (F.isVarArg()) {
// assert(0 && "VarArg not supported yet!");
//}
if (!OutChains.empty())
DAG.setRoot(DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
&OutChains[0], OutChains.size()));
return Chain;
}
SDValue AVM2TargetLowering::
LowerFormalArguments(SDValue Chain,
CallingConv::ID CallConv, bool isVarArg,
const SmallVectorImpl<ISD::InputArg> &Ins,
DebugLoc DL, SelectionDAG &DAG,
SmallVectorImpl<SDValue> &InVals) const
{
MachineFunction &MF = DAG.getMachineFunction();
AVM2MachineFunctionInfo *FuncInfo = MF.getInfo<AVM2MachineFunctionInfo>();
// Assign locations to all of the incoming arguments.
SmallVector<CCValAssign, 16> ArgLocs;
CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs, *DAG.getContext());
CCInfo.AnalyzeFormalArguments(Ins, CC_AVM2_32);
SDValue Root = DAG.getRoot();
unsigned ArgOffset = 0;
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
SDValue ArgValue;
CCValAssign &VA = ArgLocs[i];
// FIXME: We ignore the register assignments of AnalyzeFormalArguments
// because it doesn't know how to split a double into two i32 registers.
EVT ObjectVT = VA.getValVT();
switch (ObjectVT.getSimpleVT().SimpleTy) {
default:
assert(0 && "Unhandled argument type!");
case MVT::i1:
case MVT::i8:
case MVT::i16:
case MVT::i32:
if (!Ins[i].Used) { // Argument is dead.
InVals.push_back(DAG.getNode(ISD::UNDEF, DL, ObjectVT));
} else {
int FrameIdx = MF.getFrameInfo()->CreateFixedObject(4, ArgOffset, true);
SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32);
SDValue Load;
if (ObjectVT == MVT::i32) {
Load = DAG.getLoad(MVT::i32, DL, Root, FIPtr, MachinePointerInfo(), false, false, 0);
} else {
ISD::LoadExtType LoadOp = ISD::SEXTLOAD;
unsigned Offset = 0; // LE
FIPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, FIPtr,
DAG.getConstant(Offset, MVT::i32));
Load = DAG.getExtLoad(LoadOp, DL, MVT::i32, Root, FIPtr,
MachinePointerInfo(), ObjectVT, false, false, 0);
Load = DAG.getNode(ISD::TRUNCATE, DL, ObjectVT, Load);
}
InVals.push_back(Load);
}
ArgOffset += 4;
break;
case MVT::f32:
if (!Ins[i].Used) { // Argument is dead.
InVals.push_back(DAG.getNode(ISD::UNDEF, DL, ObjectVT));
} else {
int FrameIdx = MF.getFrameInfo()->CreateFixedObject(4, ArgOffset, true);
SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32);
InVals.push_back(DAG.getLoad(ObjectVT, DL, Root, FIPtr, MachinePointerInfo(), false, false, 0));
}
ArgOffset += 4;
break;
case MVT::i64:
if (!Ins[i].Used) { // Argument is dead.
InVals.push_back(DAG.getNode(ISD::UNDEF, DL, ObjectVT));
} else {
int FrameIdx = MF.getFrameInfo()->CreateFixedObject(8, ArgOffset, true);
SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32);
SDValue FIPtr4 = DAG.getNode(ISD::ADD, DL, MVT::i32, FIPtr, DAG.getConstant(4, MVT::i32));
SDValue Lo = DAG.getLoad(MVT::i32, DL, Root, FIPtr, MachinePointerInfo(), false, false, 0);
SDValue Hi = DAG.getLoad(MVT::i32, DL, Root, FIPtr4, MachinePointerInfo(), false, false, 0);
InVals.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Lo, Hi));
}
ArgOffset += 8;
break;
case MVT::f64:
if (!Ins[i].Used) { // Argument is dead.
InVals.push_back(DAG.getNode(ISD::UNDEF, DL, ObjectVT));
} else {
int FrameIdx = MF.getFrameInfo()->CreateFixedObject(8, ArgOffset, true);
SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32);
InVals.push_back(DAG.getLoad(ObjectVT, DL, Root, FIPtr, MachinePointerInfo(), false, false, 0));
}
ArgOffset += 8;
break;
}
}
if (isVarArg) {
// Remember the vararg offset for the va_start implementation.
FuncInfo->setVarArgsFrameOffset(ArgOffset);
}
return Chain;
}
/// LowerFormalArguments - V8 uses a very simple ABI, where all values are
/// passed in either one or two GPRs, including FP values. TODO: we should
/// pass FP values in FP registers for fastcc functions.
SDValue
SparcTargetLowering::LowerFormalArguments(SDValue Chain,
CallingConv::ID CallConv, bool isVarArg,
const SmallVectorImpl<ISD::InputArg>
&Ins,
DebugLoc dl, SelectionDAG &DAG,
SmallVectorImpl<SDValue> &InVals)
const {
MachineFunction &MF = DAG.getMachineFunction();
MachineRegisterInfo &RegInfo = MF.getRegInfo();
SparcMachineFunctionInfo *FuncInfo = MF.getInfo<SparcMachineFunctionInfo>();
// Assign locations to all of the incoming arguments.
SmallVector<CCValAssign, 16> ArgLocs;
CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
ArgLocs, *DAG.getContext());
CCInfo.AnalyzeFormalArguments(Ins, CC_Sparc32);
static const unsigned ArgRegs[] = {
SP::I0, SP::I1, SP::I2, SP::I3, SP::I4, SP::I5
};
const unsigned *CurArgReg = ArgRegs, *ArgRegEnd = ArgRegs+6;
unsigned ArgOffset = 68;
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
SDValue ArgValue;
CCValAssign &VA = ArgLocs[i];
// FIXME: We ignore the register assignments of AnalyzeFormalArguments
// because it doesn't know how to split a double into two i32 registers.
EVT ObjectVT = VA.getValVT();
switch (ObjectVT.getSimpleVT().SimpleTy) {
default: llvm_unreachable("Unhandled argument type!");
case MVT::i1:
case MVT::i8:
case MVT::i16:
case MVT::i32:
if (!Ins[i].Used) { // Argument is dead.
if (CurArgReg < ArgRegEnd) ++CurArgReg;
InVals.push_back(DAG.getUNDEF(ObjectVT));
} else if (CurArgReg < ArgRegEnd) { // Lives in an incoming GPR
unsigned VReg = RegInfo.createVirtualRegister(&SP::IntRegsRegClass);
MF.getRegInfo().addLiveIn(*CurArgReg++, VReg);
SDValue Arg = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
if (ObjectVT != MVT::i32) {
unsigned AssertOp = ISD::AssertSext;
Arg = DAG.getNode(AssertOp, dl, MVT::i32, Arg,
DAG.getValueType(ObjectVT));
Arg = DAG.getNode(ISD::TRUNCATE, dl, ObjectVT, Arg);
}
InVals.push_back(Arg);
} else {
int FrameIdx = MF.getFrameInfo()->CreateFixedObject(4, ArgOffset,
true, false);
SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32);
SDValue Load;
if (ObjectVT == MVT::i32) {
Load = DAG.getLoad(MVT::i32, dl, Chain, FIPtr, NULL, 0,
false, false, 0);
} else {
ISD::LoadExtType LoadOp = ISD::SEXTLOAD;
// Sparc is big endian, so add an offset based on the ObjectVT.
unsigned Offset = 4-std::max(1U, ObjectVT.getSizeInBits()/8);
FIPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, FIPtr,
DAG.getConstant(Offset, MVT::i32));
Load = DAG.getExtLoad(LoadOp, dl, MVT::i32, Chain, FIPtr,
NULL, 0, ObjectVT, false, false, 0);
Load = DAG.getNode(ISD::TRUNCATE, dl, ObjectVT, Load);
}
InVals.push_back(Load);
}
ArgOffset += 4;
break;
case MVT::f32:
if (!Ins[i].Used) { // Argument is dead.
if (CurArgReg < ArgRegEnd) ++CurArgReg;
InVals.push_back(DAG.getUNDEF(ObjectVT));
} else if (CurArgReg < ArgRegEnd) { // Lives in an incoming GPR
// FP value is passed in an integer register.
unsigned VReg = RegInfo.createVirtualRegister(&SP::IntRegsRegClass);
MF.getRegInfo().addLiveIn(*CurArgReg++, VReg);
SDValue Arg = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Arg);
InVals.push_back(Arg);
} else {
int FrameIdx = MF.getFrameInfo()->CreateFixedObject(4, ArgOffset,
true, false);
SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32);
SDValue Load = DAG.getLoad(MVT::f32, dl, Chain, FIPtr, NULL, 0,
false, false, 0);
InVals.push_back(Load);
}
ArgOffset += 4;
break;
case MVT::i64:
case MVT::f64:
if (!Ins[i].Used) { // Argument is dead.
if (CurArgReg < ArgRegEnd) ++CurArgReg;
if (CurArgReg < ArgRegEnd) ++CurArgReg;
InVals.push_back(DAG.getUNDEF(ObjectVT));
} else {
SDValue HiVal;
if (CurArgReg < ArgRegEnd) { // Lives in an incoming GPR
unsigned VRegHi = RegInfo.createVirtualRegister(&SP::IntRegsRegClass);
MF.getRegInfo().addLiveIn(*CurArgReg++, VRegHi);
HiVal = DAG.getCopyFromReg(Chain, dl, VRegHi, MVT::i32);
} else {
int FrameIdx = MF.getFrameInfo()->CreateFixedObject(4, ArgOffset,
true, false);
SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32);
HiVal = DAG.getLoad(MVT::i32, dl, Chain, FIPtr, NULL, 0,
false, false, 0);
}
SDValue LoVal;
if (CurArgReg < ArgRegEnd) { // Lives in an incoming GPR
unsigned VRegLo = RegInfo.createVirtualRegister(&SP::IntRegsRegClass);
MF.getRegInfo().addLiveIn(*CurArgReg++, VRegLo);
LoVal = DAG.getCopyFromReg(Chain, dl, VRegLo, MVT::i32);
} else {
int FrameIdx = MF.getFrameInfo()->CreateFixedObject(4, ArgOffset+4,
true, false);
SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32);
LoVal = DAG.getLoad(MVT::i32, dl, Chain, FIPtr, NULL, 0,
false, false, 0);
}
// Compose the two halves together into an i64 unit.
SDValue WholeValue =
DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, LoVal, HiVal);
// If we want a double, do a bit convert.
if (ObjectVT == MVT::f64)
WholeValue = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, WholeValue);
InVals.push_back(WholeValue);
}
ArgOffset += 8;
break;
}
}
// Store remaining ArgRegs to the stack if this is a varargs function.
if (isVarArg) {
// Remember the vararg offset for the va_start implementation.
FuncInfo->setVarArgsFrameOffset(ArgOffset);
std::vector<SDValue> OutChains;
for (; CurArgReg != ArgRegEnd; ++CurArgReg) {
unsigned VReg = RegInfo.createVirtualRegister(&SP::IntRegsRegClass);
MF.getRegInfo().addLiveIn(*CurArgReg, VReg);
SDValue Arg = DAG.getCopyFromReg(DAG.getRoot(), dl, VReg, MVT::i32);
int FrameIdx = MF.getFrameInfo()->CreateFixedObject(4, ArgOffset,
true, false);
SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32);
OutChains.push_back(DAG.getStore(DAG.getRoot(), dl, Arg, FIPtr, NULL, 0,
false, false, 0));
ArgOffset += 4;
}
if (!OutChains.empty()) {
OutChains.push_back(Chain);
Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
&OutChains[0], OutChains.size());
}
}
return Chain;
}
/// LowerArguments - V8 uses a very simple ABI, where all values are passed in
/// either one or two GPRs, including FP values. TODO: we should pass FP values
/// in FP registers for fastcc functions.
void
SparcTargetLowering::LowerArguments(Function &F, SelectionDAG &DAG,
SmallVectorImpl<SDValue> &ArgValues,
DebugLoc dl) {
MachineFunction &MF = DAG.getMachineFunction();
MachineRegisterInfo &RegInfo = MF.getRegInfo();
static const unsigned ArgRegs[] = {
SP::I0, SP::I1, SP::I2, SP::I3, SP::I4, SP::I5
};
const unsigned *CurArgReg = ArgRegs, *ArgRegEnd = ArgRegs+6;
unsigned ArgOffset = 68;
SDValue Root = DAG.getRoot();
std::vector<SDValue> OutChains;
for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
MVT ObjectVT = getValueType(I->getType());
switch (ObjectVT.getSimpleVT()) {
default: assert(0 && "Unhandled argument type!");
case MVT::i1:
case MVT::i8:
case MVT::i16:
case MVT::i32:
if (I->use_empty()) { // Argument is dead.
if (CurArgReg < ArgRegEnd) ++CurArgReg;
ArgValues.push_back(DAG.getUNDEF(ObjectVT));
} else if (CurArgReg < ArgRegEnd) { // Lives in an incoming GPR
unsigned VReg = RegInfo.createVirtualRegister(&SP::IntRegsRegClass);
MF.getRegInfo().addLiveIn(*CurArgReg++, VReg);
SDValue Arg = DAG.getCopyFromReg(Root, dl, VReg, MVT::i32);
if (ObjectVT != MVT::i32) {
unsigned AssertOp = ISD::AssertSext;
Arg = DAG.getNode(AssertOp, dl, MVT::i32, Arg,
DAG.getValueType(ObjectVT));
Arg = DAG.getNode(ISD::TRUNCATE, dl, ObjectVT, Arg);
}
ArgValues.push_back(Arg);
} else {
int FrameIdx = MF.getFrameInfo()->CreateFixedObject(4, ArgOffset);
SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32);
SDValue Load;
if (ObjectVT == MVT::i32) {
Load = DAG.getLoad(MVT::i32, dl, Root, FIPtr, NULL, 0);
} else {
ISD::LoadExtType LoadOp = ISD::SEXTLOAD;
// Sparc is big endian, so add an offset based on the ObjectVT.
unsigned Offset = 4-std::max(1U, ObjectVT.getSizeInBits()/8);
FIPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, FIPtr,
DAG.getConstant(Offset, MVT::i32));
Load = DAG.getExtLoad(LoadOp, dl, MVT::i32, Root, FIPtr,
NULL, 0, ObjectVT);
Load = DAG.getNode(ISD::TRUNCATE, dl, ObjectVT, Load);
}
ArgValues.push_back(Load);
}
ArgOffset += 4;
break;
case MVT::f32:
if (I->use_empty()) { // Argument is dead.
if (CurArgReg < ArgRegEnd) ++CurArgReg;
ArgValues.push_back(DAG.getUNDEF(ObjectVT));
} else if (CurArgReg < ArgRegEnd) { // Lives in an incoming GPR
// FP value is passed in an integer register.
unsigned VReg = RegInfo.createVirtualRegister(&SP::IntRegsRegClass);
MF.getRegInfo().addLiveIn(*CurArgReg++, VReg);
SDValue Arg = DAG.getCopyFromReg(Root, dl, VReg, MVT::i32);
Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Arg);
ArgValues.push_back(Arg);
} else {
int FrameIdx = MF.getFrameInfo()->CreateFixedObject(4, ArgOffset);
SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32);
SDValue Load = DAG.getLoad(MVT::f32, dl, Root, FIPtr, NULL, 0);
ArgValues.push_back(Load);
}
ArgOffset += 4;
break;
case MVT::i64:
case MVT::f64:
if (I->use_empty()) { // Argument is dead.
if (CurArgReg < ArgRegEnd) ++CurArgReg;
if (CurArgReg < ArgRegEnd) ++CurArgReg;
ArgValues.push_back(DAG.getUNDEF(ObjectVT));
} else {
SDValue HiVal;
if (CurArgReg < ArgRegEnd) { // Lives in an incoming GPR
unsigned VRegHi = RegInfo.createVirtualRegister(&SP::IntRegsRegClass);
MF.getRegInfo().addLiveIn(*CurArgReg++, VRegHi);
HiVal = DAG.getCopyFromReg(Root, dl, VRegHi, MVT::i32);
} else {
int FrameIdx = MF.getFrameInfo()->CreateFixedObject(4, ArgOffset);
SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32);
HiVal = DAG.getLoad(MVT::i32, dl, Root, FIPtr, NULL, 0);
}
SDValue LoVal;
if (CurArgReg < ArgRegEnd) { // Lives in an incoming GPR
unsigned VRegLo = RegInfo.createVirtualRegister(&SP::IntRegsRegClass);
MF.getRegInfo().addLiveIn(*CurArgReg++, VRegLo);
LoVal = DAG.getCopyFromReg(Root, dl, VRegLo, MVT::i32);
} else {
int FrameIdx = MF.getFrameInfo()->CreateFixedObject(4, ArgOffset+4);
SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32);
LoVal = DAG.getLoad(MVT::i32, dl, Root, FIPtr, NULL, 0);
}
// Compose the two halves together into an i64 unit.
SDValue WholeValue =
DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, LoVal, HiVal);
// If we want a double, do a bit convert.
if (ObjectVT == MVT::f64)
WholeValue = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, WholeValue);
ArgValues.push_back(WholeValue);
}
ArgOffset += 8;
break;
}
}
// Store remaining ArgRegs to the stack if this is a varargs function.
if (F.isVarArg()) {
// Remember the vararg offset for the va_start implementation.
VarArgsFrameOffset = ArgOffset;
for (; CurArgReg != ArgRegEnd; ++CurArgReg) {
unsigned VReg = RegInfo.createVirtualRegister(&SP::IntRegsRegClass);
MF.getRegInfo().addLiveIn(*CurArgReg, VReg);
SDValue Arg = DAG.getCopyFromReg(DAG.getRoot(), dl, VReg, MVT::i32);
int FrameIdx = MF.getFrameInfo()->CreateFixedObject(4, ArgOffset);
SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32);
OutChains.push_back(DAG.getStore(DAG.getRoot(), dl, Arg, FIPtr, NULL, 0));
ArgOffset += 4;
}
}
if (!OutChains.empty())
DAG.setRoot(DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
&OutChains[0], OutChains.size()));
}
/// LowerFormalArguments - V8 uses a very simple ABI, where all values are
/// passed in either one or two GPRs, including FP values. TODO: we should
/// pass FP values in FP registers for fastcc functions.
SDValue
SparcTargetLowering::LowerFormalArguments(SDValue Chain,
CallingConv::ID CallConv, bool isVarArg,
const SmallVectorImpl<ISD::InputArg>
&Ins,
DebugLoc dl, SelectionDAG &DAG,
SmallVectorImpl<SDValue> &InVals)
const {
MachineFunction &MF = DAG.getMachineFunction();
MachineRegisterInfo &RegInfo = MF.getRegInfo();
SparcMachineFunctionInfo *FuncInfo = MF.getInfo<SparcMachineFunctionInfo>();
// Assign locations to all of the incoming arguments.
SmallVector<CCValAssign, 16> ArgLocs;
CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
ArgLocs, *DAG.getContext());
CCInfo.AnalyzeFormalArguments(Ins, CC_Sparc32);
const unsigned StackOffset = 92;
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
CCValAssign &VA = ArgLocs[i];
if (i == 0 && Ins[i].Flags.isSRet()) {
//Get SRet from [%fp+64]
int FrameIdx = MF.getFrameInfo()->CreateFixedObject(4, 64, true);
SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32);
SDValue Arg = DAG.getLoad(MVT::i32, dl, Chain, FIPtr,
MachinePointerInfo(),
false, false, 0);
InVals.push_back(Arg);
continue;
}
if (VA.isRegLoc()) {
if (VA.needsCustom()) {
assert(VA.getLocVT() == MVT::f64);
unsigned VRegHi = RegInfo.createVirtualRegister(&SP::IntRegsRegClass);
MF.getRegInfo().addLiveIn(VA.getLocReg(), VRegHi);
SDValue HiVal = DAG.getCopyFromReg(Chain, dl, VRegHi, MVT::i32);
assert(i+1 < e);
CCValAssign &NextVA = ArgLocs[++i];
SDValue LoVal;
if (NextVA.isMemLoc()) {
int FrameIdx = MF.getFrameInfo()->
CreateFixedObject(4, StackOffset+NextVA.getLocMemOffset(),true);
SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32);
LoVal = DAG.getLoad(MVT::i32, dl, Chain, FIPtr,
MachinePointerInfo(),
false, false, 0);
} else {
unsigned loReg = MF.addLiveIn(NextVA.getLocReg(),
&SP::IntRegsRegClass);
LoVal = DAG.getCopyFromReg(Chain, dl, loReg, MVT::i32);
}
SDValue WholeValue =
DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, LoVal, HiVal);
WholeValue = DAG.getNode(ISD::BITCAST, dl, MVT::f64, WholeValue);
InVals.push_back(WholeValue);
continue;
}
unsigned VReg = RegInfo.createVirtualRegister(&SP::IntRegsRegClass);
MF.getRegInfo().addLiveIn(VA.getLocReg(), VReg);
SDValue Arg = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
if (VA.getLocVT() == MVT::f32)
Arg = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Arg);
else if (VA.getLocVT() != MVT::i32) {
Arg = DAG.getNode(ISD::AssertSext, dl, MVT::i32, Arg,
DAG.getValueType(VA.getLocVT()));
Arg = DAG.getNode(ISD::TRUNCATE, dl, VA.getLocVT(), Arg);
}
InVals.push_back(Arg);
continue;
}
assert(VA.isMemLoc());
unsigned Offset = VA.getLocMemOffset()+StackOffset;
if (VA.needsCustom()) {
assert(VA.getValVT() == MVT::f64);
//If it is double-word aligned, just load.
if (Offset % 8 == 0) {
int FI = MF.getFrameInfo()->CreateFixedObject(8,
Offset,
true);
SDValue FIPtr = DAG.getFrameIndex(FI, getPointerTy());
SDValue Load = DAG.getLoad(VA.getValVT(), dl, Chain, FIPtr,
MachinePointerInfo(),
false,false, 0);
InVals.push_back(Load);
continue;
}
int FI = MF.getFrameInfo()->CreateFixedObject(4,
Offset,
true);
SDValue FIPtr = DAG.getFrameIndex(FI, getPointerTy());
SDValue HiVal = DAG.getLoad(MVT::i32, dl, Chain, FIPtr,
MachinePointerInfo(),
false, false, 0);
int FI2 = MF.getFrameInfo()->CreateFixedObject(4,
Offset+4,
true);
SDValue FIPtr2 = DAG.getFrameIndex(FI2, getPointerTy());
SDValue LoVal = DAG.getLoad(MVT::i32, dl, Chain, FIPtr2,
MachinePointerInfo(),
false, false, 0);
SDValue WholeValue =
DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, LoVal, HiVal);
WholeValue = DAG.getNode(ISD::BITCAST, dl, MVT::f64, WholeValue);
InVals.push_back(WholeValue);
continue;
}
int FI = MF.getFrameInfo()->CreateFixedObject(4,
Offset,
true);
SDValue FIPtr = DAG.getFrameIndex(FI, getPointerTy());
SDValue Load ;
if (VA.getValVT() == MVT::i32 || VA.getValVT() == MVT::f32) {
Load = DAG.getLoad(VA.getValVT(), dl, Chain, FIPtr,
MachinePointerInfo(),
false, false, 0);
} else {
ISD::LoadExtType LoadOp = ISD::SEXTLOAD;
// Sparc is big endian, so add an offset based on the ObjectVT.
unsigned Offset = 4-std::max(1U, VA.getValVT().getSizeInBits()/8);
FIPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, FIPtr,
DAG.getConstant(Offset, MVT::i32));
Load = DAG.getExtLoad(LoadOp, dl, MVT::i32, Chain, FIPtr,
MachinePointerInfo(),
VA.getValVT(), false, false,0);
Load = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Load);
}
InVals.push_back(Load);
}
if (MF.getFunction()->hasStructRetAttr()) {
//Copy the SRet Argument to SRetReturnReg
SparcMachineFunctionInfo *SFI = MF.getInfo<SparcMachineFunctionInfo>();
unsigned Reg = SFI->getSRetReturnReg();
if (!Reg) {
Reg = MF.getRegInfo().createVirtualRegister(&SP::IntRegsRegClass);
SFI->setSRetReturnReg(Reg);
}
SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
}
// Store remaining ArgRegs to the stack if this is a varargs function.
if (isVarArg) {
static const unsigned ArgRegs[] = {
SP::I0, SP::I1, SP::I2, SP::I3, SP::I4, SP::I5
};
unsigned NumAllocated = CCInfo.getFirstUnallocated(ArgRegs, 6);
const unsigned *CurArgReg = ArgRegs+NumAllocated, *ArgRegEnd = ArgRegs+6;
unsigned ArgOffset = CCInfo.getNextStackOffset();
if (NumAllocated == 6)
ArgOffset += StackOffset;
else {
assert(!ArgOffset);
ArgOffset = 68+4*NumAllocated;
}
// Remember the vararg offset for the va_start implementation.
FuncInfo->setVarArgsFrameOffset(ArgOffset);
std::vector<SDValue> OutChains;
for (; CurArgReg != ArgRegEnd; ++CurArgReg) {
unsigned VReg = RegInfo.createVirtualRegister(&SP::IntRegsRegClass);
MF.getRegInfo().addLiveIn(*CurArgReg, VReg);
SDValue Arg = DAG.getCopyFromReg(DAG.getRoot(), dl, VReg, MVT::i32);
int FrameIdx = MF.getFrameInfo()->CreateFixedObject(4, ArgOffset,
true);
SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32);
OutChains.push_back(DAG.getStore(DAG.getRoot(), dl, Arg, FIPtr,
MachinePointerInfo(),
false, false, 0));
ArgOffset += 4;
}
if (!OutChains.empty()) {
OutChains.push_back(Chain);
Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
&OutChains[0], OutChains.size());
}
}
return Chain;
}
std::vector<SDOperand>
IA64TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
std::vector<SDOperand> ArgValues;
//
// add beautiful description of IA64 stack frame format
// here (from intel 24535803.pdf most likely)
//
MachineFunction &MF = DAG.getMachineFunction();
MachineFrameInfo *MFI = MF.getFrameInfo();
const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
GP = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
SP = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
RP = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
MachineBasicBlock& BB = MF.front();
unsigned args_int[] = {IA64::r32, IA64::r33, IA64::r34, IA64::r35,
IA64::r36, IA64::r37, IA64::r38, IA64::r39};
unsigned args_FP[] = {IA64::F8, IA64::F9, IA64::F10, IA64::F11,
IA64::F12,IA64::F13,IA64::F14, IA64::F15};
unsigned argVreg[8];
unsigned argPreg[8];
unsigned argOpc[8];
unsigned used_FPArgs = 0; // how many FP args have been used so far?
unsigned ArgOffset = 0;
int count = 0;
for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
{
SDOperand newroot, argt;
if(count < 8) { // need to fix this logic? maybe.
switch (getValueType(I->getType())) {
default:
assert(0 && "ERROR in LowerArgs: can't lower this type of arg.\n");
case MVT::f32:
// fixme? (well, will need to for weird FP structy stuff,
// see intel ABI docs)
case MVT::f64:
//XXX BuildMI(&BB, IA64::IDEF, 0, args_FP[used_FPArgs]);
MF.getRegInfo().addLiveIn(args_FP[used_FPArgs]);
// mark this reg as liveIn
// floating point args go into f8..f15 as-needed, the increment
argVreg[count] = // is below..:
MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::f64));
// FP args go into f8..f15 as needed: (hence the ++)
argPreg[count] = args_FP[used_FPArgs++];
argOpc[count] = IA64::FMOV;
argt = newroot = DAG.getCopyFromReg(DAG.getRoot(), argVreg[count],
MVT::f64);
if (I->getType() == Type::FloatTy)
argt = DAG.getNode(ISD::FP_ROUND, MVT::f32, argt,
DAG.getIntPtrConstant(0));
break;
case MVT::i1: // NOTE: as far as C abi stuff goes,
// bools are just boring old ints
case MVT::i8:
case MVT::i16:
case MVT::i32:
case MVT::i64:
//XXX BuildMI(&BB, IA64::IDEF, 0, args_int[count]);
MF.getRegInfo().addLiveIn(args_int[count]);
// mark this register as liveIn
argVreg[count] =
MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
argPreg[count] = args_int[count];
argOpc[count] = IA64::MOV;
argt = newroot =
DAG.getCopyFromReg(DAG.getRoot(), argVreg[count], MVT::i64);
if ( getValueType(I->getType()) != MVT::i64)
argt = DAG.getNode(ISD::TRUNCATE, getValueType(I->getType()),
newroot);
break;
}
} else { // more than 8 args go into the frame
// Create the frame index object for this incoming parameter...
ArgOffset = 16 + 8 * (count - 8);
int FI = MFI->CreateFixedObject(8, ArgOffset);
// Create the SelectionDAG nodes corresponding to a load
//from this parameter
SDOperand FIN = DAG.getFrameIndex(FI, MVT::i64);
argt = newroot = DAG.getLoad(getValueType(I->getType()),
DAG.getEntryNode(), FIN, NULL, 0);
}
++count;
DAG.setRoot(newroot.getValue(1));
ArgValues.push_back(argt);
}
// Create a vreg to hold the output of (what will become)
// the "alloc" instruction
VirtGPR = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
BuildMI(&BB, TII->get(IA64::PSEUDO_ALLOC), VirtGPR);
// we create a PSEUDO_ALLOC (pseudo)instruction for now
/*
BuildMI(&BB, IA64::IDEF, 0, IA64::r1);
// hmm:
BuildMI(&BB, IA64::IDEF, 0, IA64::r12);
BuildMI(&BB, IA64::IDEF, 0, IA64::rp);
// ..hmm.
BuildMI(&BB, IA64::MOV, 1, GP).addReg(IA64::r1);
// hmm:
BuildMI(&BB, IA64::MOV, 1, SP).addReg(IA64::r12);
BuildMI(&BB, IA64::MOV, 1, RP).addReg(IA64::rp);
// ..hmm.
*/
unsigned tempOffset=0;
// if this is a varargs function, we simply lower llvm.va_start by
// pointing to the first entry
if(F.isVarArg()) {
tempOffset=0;
VarArgsFrameIndex = MFI->CreateFixedObject(8, tempOffset);
}
// here we actually do the moving of args, and store them to the stack
// too if this is a varargs function:
for (int i = 0; i < count && i < 8; ++i) {
BuildMI(&BB, TII->get(argOpc[i]), argVreg[i]).addReg(argPreg[i]);
if(F.isVarArg()) {
// if this is a varargs function, we copy the input registers to the stack
int FI = MFI->CreateFixedObject(8, tempOffset);
tempOffset+=8; //XXX: is it safe to use r22 like this?
BuildMI(&BB, TII->get(IA64::MOV), IA64::r22).addFrameIndex(FI);
// FIXME: we should use st8.spill here, one day
BuildMI(&BB, TII->get(IA64::ST8), IA64::r22).addReg(argPreg[i]);
}
}
// Finally, inform the code generator which regs we return values in.
// (see the ISD::RET: case in the instruction selector)
switch (getValueType(F.getReturnType())) {
default: assert(0 && "i have no idea where to return this type!");
case MVT::isVoid: break;
case MVT::i1:
case MVT::i8:
case MVT::i16:
case MVT::i32:
case MVT::i64:
MF.getRegInfo().addLiveOut(IA64::r8);
break;
case MVT::f32:
case MVT::f64:
MF.getRegInfo().addLiveOut(IA64::F8);
break;
}
return ArgValues;
}
