void NVPTXFloatMCExpr::PrintImpl(raw_ostream &OS) const {
bool Ignored;
unsigned NumHex;
APFloat APF = getAPFloat();
switch (Kind) {
default: llvm_unreachable("Invalid kind!");
case VK_NVPTX_SINGLE_PREC_FLOAT:
OS << "0f";
NumHex = 8;
APF.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &Ignored);
break;
case VK_NVPTX_DOUBLE_PREC_FLOAT:
OS << "0d";
NumHex = 16;
APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &Ignored);
break;
}
APInt API = APF.bitcastToAPInt();
std::string HexStr(utohexstr(API.getZExtValue()));
if (HexStr.length() < NumHex)
OS << std::string(NumHex - HexStr.length(), '0');
OS << utohexstr(API.getZExtValue());
}
void NVPTXFloatMCExpr::printImpl(raw_ostream &OS, const MCAsmInfo *MAI) const {
bool Ignored;
unsigned NumHex;
APFloat APF = getAPFloat();
switch (Kind) {
default: llvm_unreachable("Invalid kind!");
case VK_NVPTX_HALF_PREC_FLOAT:
// ptxas does not have a way to specify half-precision floats.
// Instead we have to print and load fp16 constants as .b16
OS << "0x";
NumHex = 4;
APF.convert(APFloat::IEEEhalf(), APFloat::rmNearestTiesToEven, &Ignored);
break;
case VK_NVPTX_SINGLE_PREC_FLOAT:
OS << "0f";
NumHex = 8;
APF.convert(APFloat::IEEEsingle(), APFloat::rmNearestTiesToEven, &Ignored);
break;
case VK_NVPTX_DOUBLE_PREC_FLOAT:
OS << "0d";
NumHex = 16;
APF.convert(APFloat::IEEEdouble(), APFloat::rmNearestTiesToEven, &Ignored);
break;
}
APInt API = APF.bitcastToAPInt();
std::string HexStr(utohexstr(API.getZExtValue()));
if (HexStr.length() < NumHex)
OS << std::string(NumHex - HexStr.length(), '0');
OS << utohexstr(API.getZExtValue());
}
int FunctionComparator::cmpAPFloats(const APFloat &L, const APFloat &R) const {
// Floats are ordered first by semantics (i.e. float, double, half, etc.),
// then by value interpreted as a bitstring (aka APInt).
const fltSemantics &SL = L.getSemantics(), &SR = R.getSemantics();
if (int Res = cmpNumbers(APFloat::semanticsPrecision(SL),
APFloat::semanticsPrecision(SR)))
return Res;
if (int Res = cmpNumbers(APFloat::semanticsMaxExponent(SL),
APFloat::semanticsMaxExponent(SR)))
return Res;
if (int Res = cmpNumbers(APFloat::semanticsMinExponent(SL),
APFloat::semanticsMinExponent(SR)))
return Res;
if (int Res = cmpNumbers(APFloat::semanticsSizeInBits(SL),
APFloat::semanticsSizeInBits(SR)))
return Res;
return cmpAPInts(L.bitcastToAPInt(), R.bitcastToAPInt());
}
// Create an MCInst from a MachineInstr
void llvm::HexagonLowerToMC(const MachineInstr* MI, HexagonMCInst& MCI,
HexagonAsmPrinter& AP) {
MCI.setOpcode(MI->getOpcode());
MCI.setDesc(MI->getDesc());
for (unsigned i = 0, e = MI->getNumOperands(); i < e; i++) {
const MachineOperand &MO = MI->getOperand(i);
MCOperand MCO;
switch (MO.getType()) {
default:
MI->dump();
llvm_unreachable("unknown operand type");
case MachineOperand::MO_Register:
// Ignore all implicit register operands.
if (MO.isImplicit()) continue;
MCO = MCOperand::CreateReg(MO.getReg());
break;
case MachineOperand::MO_FPImmediate: {
APFloat Val = MO.getFPImm()->getValueAPF();
// FP immediates are used only when setting GPRs, so they may be dealt
// with like regular immediates from this point on.
MCO = MCOperand::CreateImm(*Val.bitcastToAPInt().getRawData());
break;
}
case MachineOperand::MO_Immediate:
MCO = MCOperand::CreateImm(MO.getImm());
break;
case MachineOperand::MO_MachineBasicBlock:
MCO = MCOperand::CreateExpr
(MCSymbolRefExpr::Create(MO.getMBB()->getSymbol(),
AP.OutContext));
break;
case MachineOperand::MO_GlobalAddress:
MCO = GetSymbolRef(MO, AP.Mang->getSymbol(MO.getGlobal()), AP);
break;
case MachineOperand::MO_ExternalSymbol:
MCO = GetSymbolRef(MO, AP.GetExternalSymbolSymbol(MO.getSymbolName()),
AP);
break;
case MachineOperand::MO_JumpTableIndex:
MCO = GetSymbolRef(MO, AP.GetJTISymbol(MO.getIndex()), AP);
break;
case MachineOperand::MO_ConstantPoolIndex:
MCO = GetSymbolRef(MO, AP.GetCPISymbol(MO.getIndex()), AP);
break;
case MachineOperand::MO_BlockAddress:
MCO = GetSymbolRef(MO, AP.GetBlockAddressSymbol(MO.getBlockAddress()),AP);
break;
}
MCI.addOperand(MCO);
}
}
void MBlazeMCInstLower::Lower(const MachineInstr *MI, MCInst &OutMI) const {
OutMI.setOpcode(MI->getOpcode());
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
const MachineOperand &MO = MI->getOperand(i);
MCOperand MCOp;
switch (MO.getType()) {
default: llvm_unreachable("unknown operand type");
case MachineOperand::MO_Register:
// Ignore all implicit register operands.
if (MO.isImplicit()) continue;
MCOp = MCOperand::CreateReg(MO.getReg());
break;
case MachineOperand::MO_Immediate:
MCOp = MCOperand::CreateImm(MO.getImm());
break;
case MachineOperand::MO_MachineBasicBlock:
MCOp = MCOperand::CreateExpr(MCSymbolRefExpr::Create(
MO.getMBB()->getSymbol(), Ctx));
break;
case MachineOperand::MO_GlobalAddress:
MCOp = LowerSymbolOperand(MO, GetGlobalAddressSymbol(MO));
break;
case MachineOperand::MO_ExternalSymbol:
MCOp = LowerSymbolOperand(MO, GetExternalSymbolSymbol(MO));
break;
case MachineOperand::MO_JumpTableIndex:
MCOp = LowerSymbolOperand(MO, GetJumpTableSymbol(MO));
break;
case MachineOperand::MO_ConstantPoolIndex:
MCOp = LowerSymbolOperand(MO, GetConstantPoolIndexSymbol(MO));
break;
case MachineOperand::MO_BlockAddress:
MCOp = LowerSymbolOperand(MO, GetBlockAddressSymbol(MO));
break;
case MachineOperand::MO_FPImmediate: {
bool ignored;
APFloat FVal = MO.getFPImm()->getValueAPF();
FVal.convert(APFloat::IEEEsingle, APFloat::rmTowardZero, &ignored);
APInt IVal = FVal.bitcastToAPInt();
uint64_t Val = *IVal.getRawData();
MCOp = MCOperand::CreateImm(Val);
break;
}
case MachineOperand::MO_RegisterMask:
continue;
}
OutMI.addOperand(MCOp);
}
}
/// addConstantFPValue - Add constant value entry in variable DIE.
bool CompileUnit::addConstantFPValue(DIE *Die, const MachineOperand &MO) {
assert (MO.isFPImm() && "Invalid machine operand!");
DIEBlock *Block = new (DIEValueAllocator) DIEBlock();
APFloat FPImm = MO.getFPImm()->getValueAPF();
// Get the raw data form of the floating point.
const APInt FltVal = FPImm.bitcastToAPInt();
const char *FltPtr = (const char*)FltVal.getRawData();
int NumBytes = FltVal.getBitWidth() / 8; // 8 bits per byte.
bool LittleEndian = Asm->getTargetData().isLittleEndian();
int Incr = (LittleEndian ? 1 : -1);
int Start = (LittleEndian ? 0 : NumBytes - 1);
int Stop = (LittleEndian ? NumBytes : -1);
// Output the constant to DWARF one byte at a time.
for (; Start != Stop; Start += Incr)
addUInt(Block, 0, dwarf::DW_FORM_data1,
(unsigned char)0xFF & FltPtr[Start]);
addBlock(Die, dwarf::DW_AT_const_value, 0, Block);
return true;
}
static std::string toString(const APFloat &FP) {
// Print NaNs with custom payloads specially.
if (FP.isNaN() &&
!FP.bitwiseIsEqual(APFloat::getQNaN(FP.getSemantics())) &&
!FP.bitwiseIsEqual(APFloat::getQNaN(FP.getSemantics(), /*Negative=*/true))) {
APInt AI = FP.bitcastToAPInt();
return
std::string(AI.isNegative() ? "-" : "") + "nan:0x" +
utohexstr(AI.getZExtValue() &
(AI.getBitWidth() == 32 ? INT64_C(0x007fffff) :
INT64_C(0x000fffffffffffff)),
/*LowerCase=*/true);
}
// Use C99's hexadecimal floating-point representation.
static const size_t BufBytes = 128;
char buf[BufBytes];
auto Written = FP.convertToHexString(
buf, /*hexDigits=*/0, /*upperCase=*/false, APFloat::rmNearestTiesToEven);
(void)Written;
assert(Written != 0);
assert(Written < BufBytes);
return buf;
}
/// This function converts a Constant* into a GenericValue. The interesting
/// part is if C is a ConstantExpr.
/// @brief Get a GenericValue for a Constant*
GenericValue ExecutionEngine::getConstantValue(const Constant *C) {
// If its undefined, return the garbage.
if (isa<UndefValue>(C))
return GenericValue();
// If the value is a ConstantExpr
if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
Constant *Op0 = CE->getOperand(0);
switch (CE->getOpcode()) {
case Instruction::GetElementPtr: {
// Compute the index
GenericValue Result = getConstantValue(Op0);
SmallVector<Value*, 8> Indices(CE->op_begin()+1, CE->op_end());
uint64_t Offset =
TD->getIndexedOffset(Op0->getType(), &Indices[0], Indices.size());
char* tmp = (char*) Result.PointerVal;
Result = PTOGV(tmp + Offset);
return Result;
}
case Instruction::Trunc: {
GenericValue GV = getConstantValue(Op0);
uint32_t BitWidth = cast<IntegerType>(CE->getType())->getBitWidth();
GV.IntVal = GV.IntVal.trunc(BitWidth);
return GV;
}
case Instruction::ZExt: {
GenericValue GV = getConstantValue(Op0);
uint32_t BitWidth = cast<IntegerType>(CE->getType())->getBitWidth();
GV.IntVal = GV.IntVal.zext(BitWidth);
return GV;
}
case Instruction::SExt: {
GenericValue GV = getConstantValue(Op0);
uint32_t BitWidth = cast<IntegerType>(CE->getType())->getBitWidth();
GV.IntVal = GV.IntVal.sext(BitWidth);
return GV;
}
case Instruction::FPTrunc: {
// FIXME long double
GenericValue GV = getConstantValue(Op0);
GV.FloatVal = float(GV.DoubleVal);
return GV;
}
case Instruction::FPExt:{
// FIXME long double
GenericValue GV = getConstantValue(Op0);
GV.DoubleVal = double(GV.FloatVal);
return GV;
}
case Instruction::UIToFP: {
GenericValue GV = getConstantValue(Op0);
if (CE->getType() == Type::FloatTy)
GV.FloatVal = float(GV.IntVal.roundToDouble());
else if (CE->getType() == Type::DoubleTy)
GV.DoubleVal = GV.IntVal.roundToDouble();
else if (CE->getType() == Type::X86_FP80Ty) {
const uint64_t zero[] = {0, 0};
APFloat apf = APFloat(APInt(80, 2, zero));
(void)apf.convertFromAPInt(GV.IntVal,
false,
APFloat::rmNearestTiesToEven);
GV.IntVal = apf.bitcastToAPInt();
}
return GV;
}
case Instruction::SIToFP: {
GenericValue GV = getConstantValue(Op0);
if (CE->getType() == Type::FloatTy)
GV.FloatVal = float(GV.IntVal.signedRoundToDouble());
else if (CE->getType() == Type::DoubleTy)
GV.DoubleVal = GV.IntVal.signedRoundToDouble();
else if (CE->getType() == Type::X86_FP80Ty) {
const uint64_t zero[] = { 0, 0};
APFloat apf = APFloat(APInt(80, 2, zero));
(void)apf.convertFromAPInt(GV.IntVal,
true,
APFloat::rmNearestTiesToEven);
GV.IntVal = apf.bitcastToAPInt();
}
return GV;
}
case Instruction::FPToUI: // double->APInt conversion handles sign
case Instruction::FPToSI: {
GenericValue GV = getConstantValue(Op0);
uint32_t BitWidth = cast<IntegerType>(CE->getType())->getBitWidth();
if (Op0->getType() == Type::FloatTy)
GV.IntVal = APIntOps::RoundFloatToAPInt(GV.FloatVal, BitWidth);
else if (Op0->getType() == Type::DoubleTy)
GV.IntVal = APIntOps::RoundDoubleToAPInt(GV.DoubleVal, BitWidth);
else if (Op0->getType() == Type::X86_FP80Ty) {
APFloat apf = APFloat(GV.IntVal);
uint64_t v;
bool ignored;
(void)apf.convertToInteger(&v, BitWidth,
CE->getOpcode()==Instruction::FPToSI,
APFloat::rmTowardZero, &ignored);
GV.IntVal = v; // endian?
}
return GV;
}
case Instruction::PtrToInt: {
GenericValue GV = getConstantValue(Op0);
uint32_t PtrWidth = TD->getPointerSizeInBits();
GV.IntVal = APInt(PtrWidth, uintptr_t(GV.PointerVal));
return GV;
}
case Instruction::IntToPtr: {
GenericValue GV = getConstantValue(Op0);
uint32_t PtrWidth = TD->getPointerSizeInBits();
if (PtrWidth != GV.IntVal.getBitWidth())
GV.IntVal = GV.IntVal.zextOrTrunc(PtrWidth);
assert(GV.IntVal.getBitWidth() <= 64 && "Bad pointer width");
GV.PointerVal = PointerTy(uintptr_t(GV.IntVal.getZExtValue()));
return GV;
}
case Instruction::BitCast: {
GenericValue GV = getConstantValue(Op0);
const Type* DestTy = CE->getType();
switch (Op0->getType()->getTypeID()) {
default: assert(0 && "Invalid bitcast operand");
case Type::IntegerTyID:
assert(DestTy->isFloatingPoint() && "invalid bitcast");
if (DestTy == Type::FloatTy)
GV.FloatVal = GV.IntVal.bitsToFloat();
else if (DestTy == Type::DoubleTy)
GV.DoubleVal = GV.IntVal.bitsToDouble();
break;
case Type::FloatTyID:
assert(DestTy == Type::Int32Ty && "Invalid bitcast");
GV.IntVal.floatToBits(GV.FloatVal);
break;
case Type::DoubleTyID:
assert(DestTy == Type::Int64Ty && "Invalid bitcast");
GV.IntVal.doubleToBits(GV.DoubleVal);
break;
case Type::PointerTyID:
assert(isa<PointerType>(DestTy) && "Invalid bitcast");
break; // getConstantValue(Op0) above already converted it
}
return GV;
}
case Instruction::Add:
case Instruction::Sub:
case Instruction::Mul:
case Instruction::UDiv:
case Instruction::SDiv:
case Instruction::URem:
case Instruction::SRem:
case Instruction::And:
case Instruction::Or:
case Instruction::Xor: {
GenericValue LHS = getConstantValue(Op0);
GenericValue RHS = getConstantValue(CE->getOperand(1));
GenericValue GV;
switch (CE->getOperand(0)->getType()->getTypeID()) {
default: assert(0 && "Bad add type!"); abort();
case Type::IntegerTyID:
switch (CE->getOpcode()) {
default: assert(0 && "Invalid integer opcode");
case Instruction::Add: GV.IntVal = LHS.IntVal + RHS.IntVal; break;
case Instruction::Sub: GV.IntVal = LHS.IntVal - RHS.IntVal; break;
case Instruction::Mul: GV.IntVal = LHS.IntVal * RHS.IntVal; break;
case Instruction::UDiv:GV.IntVal = LHS.IntVal.udiv(RHS.IntVal); break;
case Instruction::SDiv:GV.IntVal = LHS.IntVal.sdiv(RHS.IntVal); break;
case Instruction::URem:GV.IntVal = LHS.IntVal.urem(RHS.IntVal); break;
case Instruction::SRem:GV.IntVal = LHS.IntVal.srem(RHS.IntVal); break;
case Instruction::And: GV.IntVal = LHS.IntVal & RHS.IntVal; break;
case Instruction::Or: GV.IntVal = LHS.IntVal | RHS.IntVal; break;
case Instruction::Xor: GV.IntVal = LHS.IntVal ^ RHS.IntVal; break;
}
break;
case Type::FloatTyID:
switch (CE->getOpcode()) {
default: assert(0 && "Invalid float opcode"); abort();
case Instruction::Add:
GV.FloatVal = LHS.FloatVal + RHS.FloatVal; break;
case Instruction::Sub:
GV.FloatVal = LHS.FloatVal - RHS.FloatVal; break;
case Instruction::Mul:
GV.FloatVal = LHS.FloatVal * RHS.FloatVal; break;
case Instruction::FDiv:
GV.FloatVal = LHS.FloatVal / RHS.FloatVal; break;
case Instruction::FRem:
GV.FloatVal = ::fmodf(LHS.FloatVal,RHS.FloatVal); break;
}
break;
case Type::DoubleTyID:
switch (CE->getOpcode()) {
default: assert(0 && "Invalid double opcode"); abort();
case Instruction::Add:
GV.DoubleVal = LHS.DoubleVal + RHS.DoubleVal; break;
case Instruction::Sub:
GV.DoubleVal = LHS.DoubleVal - RHS.DoubleVal; break;
case Instruction::Mul:
GV.DoubleVal = LHS.DoubleVal * RHS.DoubleVal; break;
case Instruction::FDiv:
GV.DoubleVal = LHS.DoubleVal / RHS.DoubleVal; break;
case Instruction::FRem:
GV.DoubleVal = ::fmod(LHS.DoubleVal,RHS.DoubleVal); break;
}
break;
case Type::X86_FP80TyID:
case Type::PPC_FP128TyID:
case Type::FP128TyID: {
APFloat apfLHS = APFloat(LHS.IntVal);
switch (CE->getOpcode()) {
default: assert(0 && "Invalid long double opcode"); abort();
case Instruction::Add:
apfLHS.add(APFloat(RHS.IntVal), APFloat::rmNearestTiesToEven);
GV.IntVal = apfLHS.bitcastToAPInt();
break;
case Instruction::Sub:
apfLHS.subtract(APFloat(RHS.IntVal), APFloat::rmNearestTiesToEven);
GV.IntVal = apfLHS.bitcastToAPInt();
break;
case Instruction::Mul:
apfLHS.multiply(APFloat(RHS.IntVal), APFloat::rmNearestTiesToEven);
GV.IntVal = apfLHS.bitcastToAPInt();
break;
case Instruction::FDiv:
apfLHS.divide(APFloat(RHS.IntVal), APFloat::rmNearestTiesToEven);
GV.IntVal = apfLHS.bitcastToAPInt();
break;
case Instruction::FRem:
apfLHS.mod(APFloat(RHS.IntVal), APFloat::rmNearestTiesToEven);
GV.IntVal = apfLHS.bitcastToAPInt();
break;
}
}
break;
}
return GV;
}
default:
break;
}
cerr << "ConstantExpr not handled: " << *CE << "\n";
abort();
}
GenericValue Result;
switch (C->getType()->getTypeID()) {
case Type::FloatTyID:
Result.FloatVal = cast<ConstantFP>(C)->getValueAPF().convertToFloat();
break;
case Type::DoubleTyID:
Result.DoubleVal = cast<ConstantFP>(C)->getValueAPF().convertToDouble();
break;
case Type::X86_FP80TyID:
case Type::FP128TyID:
case Type::PPC_FP128TyID:
Result.IntVal = cast <ConstantFP>(C)->getValueAPF().bitcastToAPInt();
break;
case Type::IntegerTyID:
Result.IntVal = cast<ConstantInt>(C)->getValue();
break;
case Type::PointerTyID:
if (isa<ConstantPointerNull>(C))
Result.PointerVal = 0;
else if (const Function *F = dyn_cast<Function>(C))
Result = PTOGV(getPointerToFunctionOrStub(const_cast<Function*>(F)));
else if (const GlobalVariable* GV = dyn_cast<GlobalVariable>(C))
Result = PTOGV(getOrEmitGlobalVariable(const_cast<GlobalVariable*>(GV)));
else
assert(0 && "Unknown constant pointer type!");
break;
default:
cerr << "ERROR: Constant unimplemented for type: " << *C->getType() << "\n";
abort();
}
return Result;
}
SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) {
const R600InstrInfo *TII =
static_cast<const R600InstrInfo*>(TM.getInstrInfo());
unsigned int Opc = N->getOpcode();
if (N->isMachineOpcode()) {
return NULL; // Already selected.
}
switch (Opc) {
default: break;
case AMDGPUISD::CONST_ADDRESS: {
for (SDNode::use_iterator I = N->use_begin(), Next = llvm::next(I);
I != SDNode::use_end(); I = Next) {
Next = llvm::next(I);
if (!I->isMachineOpcode()) {
continue;
}
unsigned Opcode = I->getMachineOpcode();
bool HasDst = TII->getOperandIdx(Opcode, AMDGPU::OpName::dst) > -1;
int SrcIdx = I.getOperandNo();
int SelIdx;
// Unlike MachineInstrs, SDNodes do not have results in their operand
// list, so we need to increment the SrcIdx, since
// R600InstrInfo::getOperandIdx is based on the MachineInstr indices.
if (HasDst) {
SrcIdx++;
}
SelIdx = TII->getSelIdx(I->getMachineOpcode(), SrcIdx);
if (SelIdx < 0) {
continue;
}
SDValue CstOffset;
if (N->getValueType(0).isVector() ||
!SelectGlobalValueConstantOffset(N->getOperand(0), CstOffset))
continue;
// Gather constants values
int SrcIndices[] = {
TII->getOperandIdx(Opcode, AMDGPU::OpName::src0),
TII->getOperandIdx(Opcode, AMDGPU::OpName::src1),
TII->getOperandIdx(Opcode, AMDGPU::OpName::src2),
TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_X),
TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_Y),
TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_Z),
TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_W),
TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_X),
TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_Y),
TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_Z),
TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_W)
};
std::vector<unsigned> Consts;
for (unsigned i = 0; i < sizeof(SrcIndices) / sizeof(int); i++) {
int OtherSrcIdx = SrcIndices[i];
int OtherSelIdx = TII->getSelIdx(Opcode, OtherSrcIdx);
if (OtherSrcIdx < 0 || OtherSelIdx < 0) {
continue;
}
if (HasDst) {
OtherSrcIdx--;
OtherSelIdx--;
}
if (RegisterSDNode *Reg =
dyn_cast<RegisterSDNode>(I->getOperand(OtherSrcIdx))) {
if (Reg->getReg() == AMDGPU::ALU_CONST) {
ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(I->getOperand(OtherSelIdx));
Consts.push_back(Cst->getZExtValue());
}
}
}
ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(CstOffset);
Consts.push_back(Cst->getZExtValue());
if (!TII->fitsConstReadLimitations(Consts))
continue;
// Convert back to SDNode indices
if (HasDst) {
SrcIdx--;
SelIdx--;
}
std::vector<SDValue> Ops;
for (int i = 0, e = I->getNumOperands(); i != e; ++i) {
if (i == SrcIdx) {
Ops.push_back(CurDAG->getRegister(AMDGPU::ALU_CONST, MVT::f32));
} else if (i == SelIdx) {
Ops.push_back(CstOffset);
} else {
Ops.push_back(I->getOperand(i));
}
}
CurDAG->UpdateNodeOperands(*I, Ops.data(), Ops.size());
}
break;
}
case ISD::BUILD_VECTOR: {
const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
if (ST.getGeneration() > AMDGPUSubtarget::NORTHERN_ISLANDS) {
break;
}
unsigned RegClassID;
switch(N->getValueType(0).getVectorNumElements()) {
case 2: RegClassID = AMDGPU::R600_Reg64RegClassID; break;
case 4: RegClassID = AMDGPU::R600_Reg128RegClassID; break;
default: llvm_unreachable("Do not know how to lower this BUILD_VECTOR");
}
// BUILD_VECTOR is usually lowered into an IMPLICIT_DEF + 4 INSERT_SUBREG
// that adds a 128 bits reg copy when going through TwoAddressInstructions
// pass. We want to avoid 128 bits copies as much as possible because they
// can't be bundled by our scheduler.
SDValue RegSeqArgs[9] = {
CurDAG->getTargetConstant(RegClassID, MVT::i32),
SDValue(), CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32),
SDValue(), CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32),
SDValue(), CurDAG->getTargetConstant(AMDGPU::sub2, MVT::i32),
SDValue(), CurDAG->getTargetConstant(AMDGPU::sub3, MVT::i32)
};
bool IsRegSeq = true;
for (unsigned i = 0; i < N->getNumOperands(); i++) {
if (dyn_cast<RegisterSDNode>(N->getOperand(i))) {
IsRegSeq = false;
break;
}
RegSeqArgs[2 * i + 1] = N->getOperand(i);
}
if (!IsRegSeq)
break;
return CurDAG->SelectNodeTo(N, AMDGPU::REG_SEQUENCE, N->getVTList(),
RegSeqArgs, 2 * N->getNumOperands() + 1);
}
case ISD::BUILD_PAIR: {
SDValue RC, SubReg0, SubReg1;
const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
break;
}
if (N->getValueType(0) == MVT::i128) {
RC = CurDAG->getTargetConstant(AMDGPU::SReg_128RegClassID, MVT::i32);
SubReg0 = CurDAG->getTargetConstant(AMDGPU::sub0_sub1, MVT::i32);
SubReg1 = CurDAG->getTargetConstant(AMDGPU::sub2_sub3, MVT::i32);
} else if (N->getValueType(0) == MVT::i64) {
RC = CurDAG->getTargetConstant(AMDGPU::VSrc_64RegClassID, MVT::i32);
SubReg0 = CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32);
SubReg1 = CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32);
} else {
llvm_unreachable("Unhandled value type for BUILD_PAIR");
}
const SDValue Ops[] = { RC, N->getOperand(0), SubReg0,
N->getOperand(1), SubReg1 };
return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE,
SDLoc(N), N->getValueType(0), Ops);
}
case ISD::ConstantFP:
case ISD::Constant: {
const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
// XXX: Custom immediate lowering not implemented yet. Instead we use
// pseudo instructions defined in SIInstructions.td
if (ST.getGeneration() > AMDGPUSubtarget::NORTHERN_ISLANDS) {
break;
}
uint64_t ImmValue = 0;
unsigned ImmReg = AMDGPU::ALU_LITERAL_X;
if (N->getOpcode() == ISD::ConstantFP) {
// XXX: 64-bit Immediates not supported yet
assert(N->getValueType(0) != MVT::f64);
ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N);
APFloat Value = C->getValueAPF();
float FloatValue = Value.convertToFloat();
if (FloatValue == 0.0) {
ImmReg = AMDGPU::ZERO;
} else if (FloatValue == 0.5) {
ImmReg = AMDGPU::HALF;
} else if (FloatValue == 1.0) {
ImmReg = AMDGPU::ONE;
} else {
ImmValue = Value.bitcastToAPInt().getZExtValue();
}
} else {
// XXX: 64-bit Immediates not supported yet
assert(N->getValueType(0) != MVT::i64);
ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
if (C->getZExtValue() == 0) {
ImmReg = AMDGPU::ZERO;
} else if (C->getZExtValue() == 1) {
ImmReg = AMDGPU::ONE_INT;
} else {
ImmValue = C->getZExtValue();
}
}
for (SDNode::use_iterator Use = N->use_begin(), Next = llvm::next(Use);
Use != SDNode::use_end(); Use = Next) {
Next = llvm::next(Use);
std::vector<SDValue> Ops;
for (unsigned i = 0; i < Use->getNumOperands(); ++i) {
Ops.push_back(Use->getOperand(i));
}
if (!Use->isMachineOpcode()) {
if (ImmReg == AMDGPU::ALU_LITERAL_X) {
// We can only use literal constants (e.g. AMDGPU::ZERO,
// AMDGPU::ONE, etc) in machine opcodes.
continue;
}
} else {
if (!TII->isALUInstr(Use->getMachineOpcode()) ||
(TII->get(Use->getMachineOpcode()).TSFlags &
R600_InstFlag::VECTOR)) {
continue;
}
int ImmIdx = TII->getOperandIdx(Use->getMachineOpcode(),
AMDGPU::OpName::literal);
if (ImmIdx == -1) {
continue;
}
if (TII->getOperandIdx(Use->getMachineOpcode(),
AMDGPU::OpName::dst) != -1) {
// subtract one from ImmIdx, because the DST operand is usually index
// 0 for MachineInstrs, but we have no DST in the Ops vector.
ImmIdx--;
}
// Check that we aren't already using an immediate.
// XXX: It's possible for an instruction to have more than one
// immediate operand, but this is not supported yet.
if (ImmReg == AMDGPU::ALU_LITERAL_X) {
ConstantSDNode *C = dyn_cast<ConstantSDNode>(Use->getOperand(ImmIdx));
assert(C);
if (C->getZExtValue() != 0) {
// This instruction is already using an immediate.
continue;
}
// Set the immediate value
Ops[ImmIdx] = CurDAG->getTargetConstant(ImmValue, MVT::i32);
}
}
// Set the immediate register
Ops[Use.getOperandNo()] = CurDAG->getRegister(ImmReg, MVT::i32);
CurDAG->UpdateNodeOperands(*Use, Ops.data(), Use->getNumOperands());
}
break;
}
}
SDNode *Result = SelectCode(N);
// Fold operands of selected node
const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
const R600InstrInfo *TII =
static_cast<const R600InstrInfo*>(TM.getInstrInfo());
if (Result && Result->isMachineOpcode() && Result->getMachineOpcode() == AMDGPU::DOT_4) {
bool IsModified = false;
do {
std::vector<SDValue> Ops;
for(SDNode::op_iterator I = Result->op_begin(), E = Result->op_end();
I != E; ++I)
Ops.push_back(*I);
IsModified = FoldDotOperands(Result->getMachineOpcode(), TII, Ops);
if (IsModified) {
Result = CurDAG->UpdateNodeOperands(Result, Ops.data(), Ops.size());
}
} while (IsModified);
}
if (Result && Result->isMachineOpcode() &&
!(TII->get(Result->getMachineOpcode()).TSFlags & R600_InstFlag::VECTOR)
&& TII->hasInstrModifiers(Result->getMachineOpcode())) {
// Fold FNEG/FABS
// TODO: Isel can generate multiple MachineInst, we need to recursively
// parse Result
bool IsModified = false;
do {
std::vector<SDValue> Ops;
for(SDNode::op_iterator I = Result->op_begin(), E = Result->op_end();
I != E; ++I)
Ops.push_back(*I);
IsModified = FoldOperands(Result->getMachineOpcode(), TII, Ops);
if (IsModified) {
Result = CurDAG->UpdateNodeOperands(Result, Ops.data(), Ops.size());
}
} while (IsModified);
// If node has a single use which is CLAMP_R600, folds it
if (Result->hasOneUse() && Result->isMachineOpcode()) {
SDNode *PotentialClamp = *Result->use_begin();
if (PotentialClamp->isMachineOpcode() &&
PotentialClamp->getMachineOpcode() == AMDGPU::CLAMP_R600) {
unsigned ClampIdx =
TII->getOperandIdx(Result->getMachineOpcode(), AMDGPU::OpName::clamp);
std::vector<SDValue> Ops;
unsigned NumOp = Result->getNumOperands();
for (unsigned i = 0; i < NumOp; ++i) {
Ops.push_back(Result->getOperand(i));
}
Ops[ClampIdx - 1] = CurDAG->getTargetConstant(1, MVT::i32);
Result = CurDAG->SelectNodeTo(PotentialClamp,
Result->getMachineOpcode(), PotentialClamp->getVTList(),
Ops.data(), NumOp);
}
}
}
}
return Result;
}
SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) {
unsigned int Opc = N->getOpcode();
if (N->isMachineOpcode()) {
return NULL; // Already selected.
}
switch (Opc) {
default: break;
case ISD::BUILD_VECTOR: {
const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
if (ST.device()->getGeneration() > AMDGPUDeviceInfo::HD6XXX) {
break;
}
// BUILD_VECTOR is usually lowered into an IMPLICIT_DEF + 4 INSERT_SUBREG
// that adds a 128 bits reg copy when going through TwoAddressInstructions
// pass. We want to avoid 128 bits copies as much as possible because they
// can't be bundled by our scheduler.
SDValue RegSeqArgs[9] = {
CurDAG->getTargetConstant(AMDGPU::R600_Reg128RegClassID, MVT::i32),
SDValue(), CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32),
SDValue(), CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32),
SDValue(), CurDAG->getTargetConstant(AMDGPU::sub2, MVT::i32),
SDValue(), CurDAG->getTargetConstant(AMDGPU::sub3, MVT::i32)
};
bool IsRegSeq = true;
for (unsigned i = 0; i < N->getNumOperands(); i++) {
if (dyn_cast<RegisterSDNode>(N->getOperand(i))) {
IsRegSeq = false;
break;
}
RegSeqArgs[2 * i + 1] = N->getOperand(i);
}
if (!IsRegSeq)
break;
return CurDAG->SelectNodeTo(N, AMDGPU::REG_SEQUENCE, N->getVTList(),
RegSeqArgs, 2 * N->getNumOperands() + 1);
}
case ISD::ConstantFP:
case ISD::Constant: {
const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
// XXX: Custom immediate lowering not implemented yet. Instead we use
// pseudo instructions defined in SIInstructions.td
if (ST.device()->getGeneration() > AMDGPUDeviceInfo::HD6XXX) {
break;
}
const R600InstrInfo *TII = static_cast<const R600InstrInfo*>(TM.getInstrInfo());
uint64_t ImmValue = 0;
unsigned ImmReg = AMDGPU::ALU_LITERAL_X;
if (N->getOpcode() == ISD::ConstantFP) {
// XXX: 64-bit Immediates not supported yet
assert(N->getValueType(0) != MVT::f64);
ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N);
APFloat Value = C->getValueAPF();
float FloatValue = Value.convertToFloat();
if (FloatValue == 0.0) {
ImmReg = AMDGPU::ZERO;
} else if (FloatValue == 0.5) {
ImmReg = AMDGPU::HALF;
} else if (FloatValue == 1.0) {
ImmReg = AMDGPU::ONE;
} else {
ImmValue = Value.bitcastToAPInt().getZExtValue();
}
} else {
// XXX: 64-bit Immediates not supported yet
assert(N->getValueType(0) != MVT::i64);
ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
if (C->getZExtValue() == 0) {
ImmReg = AMDGPU::ZERO;
} else if (C->getZExtValue() == 1) {
ImmReg = AMDGPU::ONE_INT;
} else {
ImmValue = C->getZExtValue();
}
}
for (SDNode::use_iterator Use = N->use_begin(), Next = llvm::next(Use);
Use != SDNode::use_end(); Use = Next) {
Next = llvm::next(Use);
std::vector<SDValue> Ops;
for (unsigned i = 0; i < Use->getNumOperands(); ++i) {
Ops.push_back(Use->getOperand(i));
}
if (!Use->isMachineOpcode()) {
if (ImmReg == AMDGPU::ALU_LITERAL_X) {
// We can only use literal constants (e.g. AMDGPU::ZERO,
// AMDGPU::ONE, etc) in machine opcodes.
continue;
}
} else {
if (!TII->isALUInstr(Use->getMachineOpcode()) ||
(TII->get(Use->getMachineOpcode()).TSFlags &
R600_InstFlag::VECTOR)) {
continue;
}
int ImmIdx = TII->getOperandIdx(Use->getMachineOpcode(), R600Operands::IMM);
assert(ImmIdx != -1);
// subtract one from ImmIdx, because the DST operand is usually index
// 0 for MachineInstrs, but we have no DST in the Ops vector.
ImmIdx--;
// Check that we aren't already using an immediate.
// XXX: It's possible for an instruction to have more than one
// immediate operand, but this is not supported yet.
if (ImmReg == AMDGPU::ALU_LITERAL_X) {
ConstantSDNode *C = dyn_cast<ConstantSDNode>(Use->getOperand(ImmIdx));
assert(C);
if (C->getZExtValue() != 0) {
// This instruction is already using an immediate.
continue;
}
// Set the immediate value
Ops[ImmIdx] = CurDAG->getTargetConstant(ImmValue, MVT::i32);
}
}
// Set the immediate register
Ops[Use.getOperandNo()] = CurDAG->getRegister(ImmReg, MVT::i32);
CurDAG->UpdateNodeOperands(*Use, Ops.data(), Use->getNumOperands());
}
break;
}
}
SDNode *Result = SelectCode(N);
// Fold operands of selected node
const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
if (ST.device()->getGeneration() <= AMDGPUDeviceInfo::HD6XXX) {
const R600InstrInfo *TII =
static_cast<const R600InstrInfo*>(TM.getInstrInfo());
if (Result && Result->isMachineOpcode() &&
!(TII->get(Result->getMachineOpcode()).TSFlags & R600_InstFlag::VECTOR)
&& TII->isALUInstr(Result->getMachineOpcode())) {
// Fold FNEG/FABS/CONST_ADDRESS
// TODO: Isel can generate multiple MachineInst, we need to recursively
// parse Result
bool IsModified = false;
do {
std::vector<SDValue> Ops;
for(SDNode::op_iterator I = Result->op_begin(), E = Result->op_end();
I != E; ++I)
Ops.push_back(*I);
IsModified = FoldOperands(Result->getMachineOpcode(), TII, Ops);
if (IsModified) {
Result = CurDAG->UpdateNodeOperands(Result, Ops.data(), Ops.size());
}
} while (IsModified);
// If node has a single use which is CLAMP_R600, folds it
if (Result->hasOneUse() && Result->isMachineOpcode()) {
SDNode *PotentialClamp = *Result->use_begin();
if (PotentialClamp->isMachineOpcode() &&
PotentialClamp->getMachineOpcode() == AMDGPU::CLAMP_R600) {
unsigned ClampIdx =
TII->getOperandIdx(Result->getMachineOpcode(), R600Operands::CLAMP);
std::vector<SDValue> Ops;
unsigned NumOp = Result->getNumOperands();
for (unsigned i = 0; i < NumOp; ++i) {
Ops.push_back(Result->getOperand(i));
}
Ops[ClampIdx - 1] = CurDAG->getTargetConstant(1, MVT::i32);
Result = CurDAG->SelectNodeTo(PotentialClamp,
Result->getMachineOpcode(), PotentialClamp->getVTList(),
Ops.data(), NumOp);
}
}
}
}
return Result;
}
// Create an MCInst from a MachineInstr
void llvm::HexagonLowerToMC(MachineInstr const* MI, MCInst& MCB,
HexagonAsmPrinter& AP) {
if(MI->getOpcode() == Hexagon::ENDLOOP0){
HexagonMCInstrInfo::setInnerLoop(MCB);
return;
}
if(MI->getOpcode() == Hexagon::ENDLOOP1){
HexagonMCInstrInfo::setOuterLoop(MCB);
return;
}
MCInst* MCI = new (AP.OutContext) MCInst;
MCI->setOpcode(MI->getOpcode());
assert(MCI->getOpcode() == static_cast<unsigned>(MI->getOpcode()) &&
"MCI opcode should have been set on construction");
for (unsigned i = 0, e = MI->getNumOperands(); i < e; i++) {
const MachineOperand &MO = MI->getOperand(i);
MCOperand MCO;
switch (MO.getType()) {
default:
MI->dump();
llvm_unreachable("unknown operand type");
case MachineOperand::MO_Register:
// Ignore all implicit register operands.
if (MO.isImplicit()) continue;
MCO = MCOperand::createReg(MO.getReg());
break;
case MachineOperand::MO_FPImmediate: {
APFloat Val = MO.getFPImm()->getValueAPF();
// FP immediates are used only when setting GPRs, so they may be dealt
// with like regular immediates from this point on.
MCO = MCOperand::createImm(*Val.bitcastToAPInt().getRawData());
break;
}
case MachineOperand::MO_Immediate:
MCO = MCOperand::createImm(MO.getImm());
break;
case MachineOperand::MO_MachineBasicBlock:
MCO = MCOperand::createExpr
(MCSymbolRefExpr::create(MO.getMBB()->getSymbol(),
AP.OutContext));
break;
case MachineOperand::MO_GlobalAddress:
MCO = GetSymbolRef(MO, AP.getSymbol(MO.getGlobal()), AP);
break;
case MachineOperand::MO_ExternalSymbol:
MCO = GetSymbolRef(MO, AP.GetExternalSymbolSymbol(MO.getSymbolName()),
AP);
break;
case MachineOperand::MO_JumpTableIndex:
MCO = GetSymbolRef(MO, AP.GetJTISymbol(MO.getIndex()), AP);
break;
case MachineOperand::MO_ConstantPoolIndex:
MCO = GetSymbolRef(MO, AP.GetCPISymbol(MO.getIndex()), AP);
break;
case MachineOperand::MO_BlockAddress:
MCO = GetSymbolRef(MO, AP.GetBlockAddressSymbol(MO.getBlockAddress()),AP);
break;
}
MCI->addOperand(MCO);
}
MCB.addOperand(MCOperand::createInst(MCI));
}