static LegalizeMutation oneMoreElement(unsigned TypeIdx) {
return [=](const LegalityQuery &Query) {
const LLT Ty = Query.Types[TypeIdx];
const LLT EltTy = Ty.getElementType();
return std::make_pair(TypeIdx, LLT::vector(Ty.getNumElements() + 1, EltTy));
};
}
LegalizeMutation LegalizeMutations::changeElementTo(unsigned TypeIdx,
LLT NewEltTy) {
return [=](const LegalityQuery &Query) {
const LLT OldTy = Query.Types[TypeIdx];
return std::make_pair(TypeIdx, OldTy.changeElementType(NewEltTy));
};
}
bool AMDGPULegalizerInfo::legalizeFrint(
MachineInstr &MI, MachineRegisterInfo &MRI,
MachineIRBuilder &MIRBuilder) const {
MIRBuilder.setInstr(MI);
unsigned Src = MI.getOperand(1).getReg();
LLT Ty = MRI.getType(Src);
assert(Ty.isScalar() && Ty.getSizeInBits() == 64);
APFloat C1Val(APFloat::IEEEdouble(), "0x1.0p+52");
APFloat C2Val(APFloat::IEEEdouble(), "0x1.fffffffffffffp+51");
auto C1 = MIRBuilder.buildFConstant(Ty, C1Val);
auto CopySign = MIRBuilder.buildFCopysign(Ty, C1, Src);
// TODO: Should this propagate fast-math-flags?
auto Tmp1 = MIRBuilder.buildFAdd(Ty, Src, CopySign);
auto Tmp2 = MIRBuilder.buildFSub(Ty, Tmp1, CopySign);
auto C2 = MIRBuilder.buildFConstant(Ty, C2Val);
auto Fabs = MIRBuilder.buildFAbs(Ty, Src);
auto Cond = MIRBuilder.buildFCmp(CmpInst::FCMP_OGT, LLT::scalar(1), Fabs, C2);
MIRBuilder.buildSelect(MI.getOperand(0).getReg(), Cond, Src, Tmp2);
return true;
}
std::pair<LegalizerInfo::LegalizeAction, LLT>
LegalizerInfo::findVectorLegalAction(const InstrAspect &Aspect) const {
assert(Aspect.Type.isVector());
// First legalize the vector element size, then legalize the number of
// lanes in the vector.
if (Aspect.Opcode < FirstOp || Aspect.Opcode > LastOp)
return {NotFound, Aspect.Type};
const unsigned OpcodeIdx = Aspect.Opcode - FirstOp;
const unsigned TypeIdx = Aspect.Idx;
if (TypeIdx >= ScalarInVectorActions[OpcodeIdx].size())
return {NotFound, Aspect.Type};
const SizeAndActionsVec &ElemSizeVec =
ScalarInVectorActions[OpcodeIdx][TypeIdx];
LLT IntermediateType;
auto ElementSizeAndAction =
findAction(ElemSizeVec, Aspect.Type.getScalarSizeInBits());
IntermediateType =
LLT::vector(Aspect.Type.getNumElements(), ElementSizeAndAction.first);
if (ElementSizeAndAction.second != Legal)
return {ElementSizeAndAction.second, IntermediateType};
auto i = NumElements2Actions[OpcodeIdx].find(
IntermediateType.getScalarSizeInBits());
if (i == NumElements2Actions[OpcodeIdx].end()) {
return {NotFound, IntermediateType};
}
const SizeAndActionsVec &NumElementsVec = (*i).second[TypeIdx];
auto NumElementsAndAction =
findAction(NumElementsVec, IntermediateType.getNumElements());
return {NumElementsAndAction.second,
LLT::vector(NumElementsAndAction.first,
IntermediateType.getScalarSizeInBits())};
}
MachineLegalizeHelper::LegalizeResult
MachineLegalizeHelper::libcall(MachineInstr &MI) {
LLT Ty = MRI.getType(MI.getOperand(0).getReg());
unsigned Size = Ty.getSizeInBits();
MIRBuilder.setInstr(MI);
switch (MI.getOpcode()) {
default:
return UnableToLegalize;
case TargetOpcode::G_FREM: {
auto &Ctx = MIRBuilder.getMF().getFunction()->getContext();
Type *Ty = Size == 64 ? Type::getDoubleTy(Ctx) : Type::getFloatTy(Ctx);
auto &CLI = *MIRBuilder.getMF().getSubtarget().getCallLowering();
auto &TLI = *MIRBuilder.getMF().getSubtarget().getTargetLowering();
const char *Name =
TLI.getLibcallName(Size == 64 ? RTLIB::REM_F64 : RTLIB::REM_F32);
CLI.lowerCall(MIRBuilder, MachineOperand::CreateES(Name), Ty,
MI.getOperand(0).getReg(), {Ty, Ty},
{MI.getOperand(1).getReg(), MI.getOperand(2).getReg()});
MI.eraseFromParent();
return Legalized;
}
}
}
LLT MachineLegalizer::findLegalType(unsigned Opcode, LLT Ty,
LegalizeAction Action) const {
switch(Action) {
default:
llvm_unreachable("Cannot find legal type");
case Legal:
return Ty;
case NarrowScalar: {
return findLegalType(Opcode, Ty,
[&](LLT Ty) -> LLT { return Ty.halfScalarSize(); });
}
case WidenScalar: {
return findLegalType(Opcode, Ty,
[&](LLT Ty) -> LLT { return Ty.doubleScalarSize(); });
}
case FewerElements: {
return findLegalType(Opcode, Ty,
[&](LLT Ty) -> LLT { return Ty.halfElements(); });
}
case MoreElements: {
return findLegalType(
Opcode, Ty, [&](LLT Ty) -> LLT { return Ty.doubleElements(); });
}
}
}
static LegalityPredicate isMultiple32(unsigned TypeIdx,
unsigned MaxSize = 512) {
return [=](const LegalityQuery &Query) {
const LLT Ty = Query.Types[TypeIdx];
const LLT EltTy = Ty.getScalarType();
return Ty.getSizeInBits() <= MaxSize && EltTy.getSizeInBits() % 32 == 0;
};
}
static LegalityPredicate isSmallOddVector(unsigned TypeIdx) {
return [=](const LegalityQuery &Query) {
const LLT Ty = Query.Types[TypeIdx];
return Ty.isVector() &&
Ty.getNumElements() % 2 != 0 &&
Ty.getElementType().getSizeInBits() < 32;
};
}
LegalizeMutation LegalizeMutations::widenScalarOrEltToNextPow2(unsigned TypeIdx,
unsigned Min) {
return [=](const LegalityQuery &Query) {
const LLT Ty = Query.Types[TypeIdx];
unsigned NewEltSizeInBits =
std::max(1u << Log2_32_Ceil(Ty.getScalarSizeInBits()), Min);
return std::make_pair(TypeIdx, Ty.changeElementSize(NewEltSizeInBits));
};
}
bool CLinearRidgeRegression::train_machine(CFeatures* data)
{
REQUIRE(m_labels,"No labels set\n");
if (!data)
data=features;
REQUIRE(data,"No features provided and no featured previously set\n");
REQUIRE(m_labels->get_num_labels() == data->get_num_vectors(),
"Number of training vectors (%d) does not match number of labels (%d)\n",
m_labels->get_num_labels(), data->get_num_vectors());
REQUIRE(data->get_feature_class() == C_DENSE,
"Expected Dense Features (%d) but got (%d)\n",
C_DENSE, data->get_feature_class());
REQUIRE(data->get_feature_type() == F_DREAL,
"Expected Real Features (%d) but got (%d)\n",
F_DREAL, data->get_feature_type());
CDenseFeatures<float64_t>* feats=(CDenseFeatures<float64_t>*) data;
int32_t num_feat=feats->get_num_features();
int32_t num_vec=feats->get_num_vectors();
SGMatrix<float64_t> kernel_matrix(num_feat,num_feat);
SGMatrix<float64_t> feats_matrix(feats->get_feature_matrix());
SGVector<float64_t> y(num_feat);
SGVector<float64_t> tau_vector(num_feat);
tau_vector.zero();
tau_vector.add(m_tau);
Map<MatrixXd> eigen_kernel_matrix(kernel_matrix.matrix, num_feat,num_feat);
Map<MatrixXd> eigen_feats_matrix(feats_matrix.matrix, num_feat,num_vec);
Map<VectorXd> eigen_y(y.vector, num_feat);
Map<VectorXd> eigen_labels(((CRegressionLabels*)m_labels)->get_labels(),num_vec);
Map<VectorXd> eigen_tau(tau_vector.vector, num_feat);
eigen_kernel_matrix = eigen_feats_matrix*eigen_feats_matrix.transpose();
eigen_kernel_matrix.diagonal() += eigen_tau;
eigen_y = eigen_feats_matrix*eigen_labels ;
LLT<MatrixXd> llt;
llt.compute(eigen_kernel_matrix);
if(llt.info() != Eigen::Success)
{
SG_WARNING("Features covariance matrix was not positive definite\n");
return false;
}
eigen_y = llt.solve(eigen_y);
set_w(y);
return true;
}
LegalizeMutation LegalizeMutations::moreElementsToNextPow2(unsigned TypeIdx,
unsigned Min) {
return [=](const LegalityQuery &Query) {
const LLT VecTy = Query.Types[TypeIdx];
unsigned NewNumElements =
std::max(1u << Log2_32_Ceil(VecTy.getNumElements()), Min);
return std::make_pair(TypeIdx,
LLT::vector(NewNumElements, VecTy.getElementType()));
};
}
static LegalizeMutation fewerEltsToSize64Vector(unsigned TypeIdx) {
return [=](const LegalityQuery &Query) {
const LLT Ty = Query.Types[TypeIdx];
const LLT EltTy = Ty.getElementType();
unsigned Size = Ty.getSizeInBits();
unsigned Pieces = (Size + 63) / 64;
unsigned NewNumElts = (Ty.getNumElements() + 1) / Pieces;
return std::make_pair(TypeIdx, LLT::scalarOrVector(NewNumElts, EltTy));
};
}
void RegisterBankInfo::applyDefaultMapping(const OperandsMapper &OpdMapper) {
MachineInstr &MI = OpdMapper.getMI();
MachineRegisterInfo &MRI = OpdMapper.getMRI();
LLVM_DEBUG(dbgs() << "Applying default-like mapping\n");
for (unsigned OpIdx = 0,
EndIdx = OpdMapper.getInstrMapping().getNumOperands();
OpIdx != EndIdx; ++OpIdx) {
LLVM_DEBUG(dbgs() << "OpIdx " << OpIdx);
MachineOperand &MO = MI.getOperand(OpIdx);
if (!MO.isReg()) {
LLVM_DEBUG(dbgs() << " is not a register, nothing to be done\n");
continue;
}
if (!MO.getReg()) {
LLVM_DEBUG(dbgs() << " is %%noreg, nothing to be done\n");
continue;
}
assert(OpdMapper.getInstrMapping().getOperandMapping(OpIdx).NumBreakDowns !=
0 &&
"Invalid mapping");
assert(OpdMapper.getInstrMapping().getOperandMapping(OpIdx).NumBreakDowns ==
1 &&
"This mapping is too complex for this function");
iterator_range<SmallVectorImpl<unsigned>::const_iterator> NewRegs =
OpdMapper.getVRegs(OpIdx);
if (empty(NewRegs)) {
LLVM_DEBUG(dbgs() << " has not been repaired, nothing to be done\n");
continue;
}
unsigned OrigReg = MO.getReg();
unsigned NewReg = *NewRegs.begin();
LLVM_DEBUG(dbgs() << " changed, replace " << printReg(OrigReg, nullptr));
MO.setReg(NewReg);
LLVM_DEBUG(dbgs() << " with " << printReg(NewReg, nullptr));
// The OperandsMapper creates plain scalar, we may have to fix that.
// Check if the types match and if not, fix that.
LLT OrigTy = MRI.getType(OrigReg);
LLT NewTy = MRI.getType(NewReg);
if (OrigTy != NewTy) {
// The default mapping is not supposed to change the size of
// the storage. However, right now we don't necessarily bump all
// the types to storage size. For instance, we can consider
// s16 G_AND legal whereas the storage size is going to be 32.
assert(OrigTy.getSizeInBits() <= NewTy.getSizeInBits() &&
"Types with difference size cannot be handled by the default "
"mapping");
LLVM_DEBUG(dbgs() << "\nChange type of new opd from " << NewTy << " to "
<< OrigTy);
MRI.setType(NewReg, OrigTy);
}
LLVM_DEBUG(dbgs() << '\n');
}
}
bool AArch64LegalizerInfo::legalizeVaArg(MachineInstr &MI,
MachineRegisterInfo &MRI,
MachineIRBuilder &MIRBuilder) const {
MIRBuilder.setInstr(MI);
MachineFunction &MF = MIRBuilder.getMF();
unsigned Align = MI.getOperand(2).getImm();
unsigned Dst = MI.getOperand(0).getReg();
unsigned ListPtr = MI.getOperand(1).getReg();
LLT PtrTy = MRI.getType(ListPtr);
LLT IntPtrTy = LLT::scalar(PtrTy.getSizeInBits());
const unsigned PtrSize = PtrTy.getSizeInBits() / 8;
unsigned List = MRI.createGenericVirtualRegister(PtrTy);
MIRBuilder.buildLoad(
List, ListPtr,
*MF.getMachineMemOperand(MachinePointerInfo(), MachineMemOperand::MOLoad,
PtrSize, /* Align = */ PtrSize));
unsigned DstPtr;
if (Align > PtrSize) {
// Realign the list to the actual required alignment.
unsigned AlignMinus1 = MRI.createGenericVirtualRegister(IntPtrTy);
MIRBuilder.buildConstant(AlignMinus1, Align - 1);
unsigned ListTmp = MRI.createGenericVirtualRegister(PtrTy);
MIRBuilder.buildGEP(ListTmp, List, AlignMinus1);
DstPtr = MRI.createGenericVirtualRegister(PtrTy);
MIRBuilder.buildPtrMask(DstPtr, ListTmp, Log2_64(Align));
} else
DstPtr = List;
uint64_t ValSize = MRI.getType(Dst).getSizeInBits() / 8;
MIRBuilder.buildLoad(
Dst, DstPtr,
*MF.getMachineMemOperand(MachinePointerInfo(), MachineMemOperand::MOLoad,
ValSize, std::max(Align, PtrSize)));
unsigned SizeReg = MRI.createGenericVirtualRegister(IntPtrTy);
MIRBuilder.buildConstant(SizeReg, alignTo(ValSize, PtrSize));
unsigned NewList = MRI.createGenericVirtualRegister(PtrTy);
MIRBuilder.buildGEP(NewList, DstPtr, SizeReg);
MIRBuilder.buildStore(
NewList, ListPtr,
*MF.getMachineMemOperand(MachinePointerInfo(), MachineMemOperand::MOStore,
PtrSize, /* Align = */ PtrSize));
MI.eraseFromParent();
return true;
}
void MachineLegalizer::computeTables() {
for (auto &Op : Actions) {
LLT Ty = Op.first.second;
if (!Ty.isVector())
continue;
auto &Entry =
MaxLegalVectorElts[std::make_pair(Op.first.first, Ty.getElementType())];
Entry = std::max(Entry, Ty.getNumElements());
}
TablesInitialized = true;
}
Optional<APInt> llvm::ConstantFoldBinOp(unsigned Opcode, const unsigned Op1,
const unsigned Op2,
const MachineRegisterInfo &MRI) {
auto MaybeOp1Cst = getConstantVRegVal(Op1, MRI);
auto MaybeOp2Cst = getConstantVRegVal(Op2, MRI);
if (MaybeOp1Cst && MaybeOp2Cst) {
LLT Ty = MRI.getType(Op1);
APInt C1(Ty.getSizeInBits(), *MaybeOp1Cst, true);
APInt C2(Ty.getSizeInBits(), *MaybeOp2Cst, true);
switch (Opcode) {
default:
break;
case TargetOpcode::G_ADD:
return C1 + C2;
case TargetOpcode::G_AND:
return C1 & C2;
case TargetOpcode::G_ASHR:
return C1.ashr(C2);
case TargetOpcode::G_LSHR:
return C1.lshr(C2);
case TargetOpcode::G_MUL:
return C1 * C2;
case TargetOpcode::G_OR:
return C1 | C2;
case TargetOpcode::G_SHL:
return C1 << C2;
case TargetOpcode::G_SUB:
return C1 - C2;
case TargetOpcode::G_XOR:
return C1 ^ C2;
case TargetOpcode::G_UDIV:
if (!C2.getBoolValue())
break;
return C1.udiv(C2);
case TargetOpcode::G_SDIV:
if (!C2.getBoolValue())
break;
return C1.sdiv(C2);
case TargetOpcode::G_UREM:
if (!C2.getBoolValue())
break;
return C1.urem(C2);
case TargetOpcode::G_SREM:
if (!C2.getBoolValue())
break;
return C1.srem(C2);
}
}
return None;
}
// FIXME: inefficient implementation for now. Without ComputeValueVTs we're
// probably going to need specialized lookup structures for various types before
// we have any hope of doing well with something like <13 x i3>. Even the common
// cases should do better than what we have now.
std::pair<MachineLegalizer::LegalizeAction, LLT>
MachineLegalizer::getAction(const InstrAspect &Aspect) const {
assert(TablesInitialized && "backend forgot to call computeTables");
// These *have* to be implemented for now, they're the fundamental basis of
// how everything else is transformed.
// FIXME: the long-term plan calls for expansion in terms of load/store (if
// they're not legal).
if (Aspect.Opcode == TargetOpcode::G_SEQUENCE ||
Aspect.Opcode == TargetOpcode::G_EXTRACT)
return std::make_pair(Legal, Aspect.Type);
LegalizeAction Action = findInActions(Aspect);
if (Action != NotFound)
return findLegalAction(Aspect, Action);
unsigned Opcode = Aspect.Opcode;
LLT Ty = Aspect.Type;
if (!Ty.isVector()) {
auto DefaultAction = DefaultActions.find(Aspect.Opcode);
if (DefaultAction != DefaultActions.end() && DefaultAction->second == Legal)
return std::make_pair(Legal, Ty);
assert(DefaultAction->second == NarrowScalar && "unexpected default");
return findLegalAction(Aspect, NarrowScalar);
}
LLT EltTy = Ty.getElementType();
int NumElts = Ty.getNumElements();
auto ScalarAction = ScalarInVectorActions.find(std::make_pair(Opcode, EltTy));
if (ScalarAction != ScalarInVectorActions.end() &&
ScalarAction->second != Legal)
return findLegalAction(Aspect, ScalarAction->second);
// The element type is legal in principle, but the number of elements is
// wrong.
auto MaxLegalElts = MaxLegalVectorElts.lookup(std::make_pair(Opcode, EltTy));
if (MaxLegalElts > NumElts)
return findLegalAction(Aspect, MoreElements);
if (MaxLegalElts == 0) {
// Scalarize if there's no legal vector type, which is just a special case
// of FewerElements.
return std::make_pair(FewerElements, EltTy);
}
return findLegalAction(Aspect, FewerElements);
}
MachineLegalizeHelper::LegalizeResult
MachineLegalizeHelper::fewerElementsVector(MachineInstr &MI, unsigned TypeIdx,
LLT NarrowTy) {
// FIXME: Don't know how to handle secondary types yet.
if (TypeIdx != 0)
return UnableToLegalize;
switch (MI.getOpcode()) {
default:
return UnableToLegalize;
case TargetOpcode::G_ADD: {
unsigned NarrowSize = NarrowTy.getSizeInBits();
unsigned DstReg = MI.getOperand(0).getReg();
int NumParts = MRI.getType(DstReg).getSizeInBits() / NarrowSize;
MIRBuilder.setInstr(MI);
SmallVector<unsigned, 2> Src1Regs, Src2Regs, DstRegs, Indexes;
extractParts(MI.getOperand(1).getReg(), NarrowTy, NumParts, Src1Regs);
extractParts(MI.getOperand(2).getReg(), NarrowTy, NumParts, Src2Regs);
for (int i = 0; i < NumParts; ++i) {
unsigned DstReg = MRI.createGenericVirtualRegister(NarrowTy);
MIRBuilder.buildAdd(DstReg, Src1Regs[i], Src2Regs[i]);
DstRegs.push_back(DstReg);
Indexes.push_back(i * NarrowSize);
}
MIRBuilder.buildSequence(DstReg, DstRegs, Indexes);
MI.eraseFromParent();
return Legalized;
}
}
}
void LegalizerInfo::computeTables() {
for (unsigned Opcode = 0; Opcode <= LastOp - FirstOp; ++Opcode) {
for (unsigned Idx = 0; Idx != Actions[Opcode].size(); ++Idx) {
for (auto &Action : Actions[Opcode][Idx]) {
LLT Ty = Action.first;
if (!Ty.isVector())
continue;
auto &Entry = MaxLegalVectorElts[std::make_pair(Opcode + FirstOp,
Ty.getElementType())];
Entry = std::max(Entry, Ty.getNumElements());
}
}
}
TablesInitialized = true;
}
// Recompute the kernel.
void GPCMGaussianProcess::recomputeKernel()
{
// Constants.
int N = dataMatrix.rows();
int D = dataMatrix.cols();
// Initialize the covariance matrix.
K.noalias() = kernel->covariance(X);
// Compute inverse of kernel matrix.
LLT<MatrixXd> cholesky = K.llt();
logDetK = cholesky.matrixLLT().diagonal().array().log().sum()*2.0;
invK = cholesky.solve(MatrixXd::Identity(N,N));
// Clear gK matrices.
gK.setZero(N,N);
gKd.setZero(N,N);
}
void MachineLegalizeHelper::extractParts(unsigned Reg, LLT Ty, int NumParts,
SmallVectorImpl<unsigned> &VRegs) {
unsigned Size = Ty.getSizeInBits();
SmallVector<uint64_t, 4> Indexes;
for (int i = 0; i < NumParts; ++i) {
VRegs.push_back(MRI.createGenericVirtualRegister(Ty));
Indexes.push_back(i * Size);
}
MIRBuilder.buildExtract(VRegs, Indexes, Reg);
}
MachineLegalizeHelper::LegalizeResult
MachineLegalizeHelper::narrowScalar(MachineInstr &MI, unsigned TypeIdx,
LLT NarrowTy) {
// FIXME: Don't know how to handle secondary types yet.
if (TypeIdx != 0)
return UnableToLegalize;
switch (MI.getOpcode()) {
default:
return UnableToLegalize;
case TargetOpcode::G_ADD: {
// Expand in terms of carry-setting/consuming G_ADDE instructions.
unsigned NarrowSize = NarrowTy.getSizeInBits();
int NumParts = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits() /
NarrowTy.getSizeInBits();
MIRBuilder.setInstr(MI);
SmallVector<unsigned, 2> Src1Regs, Src2Regs, DstRegs, Indexes;
extractParts(MI.getOperand(1).getReg(), NarrowTy, NumParts, Src1Regs);
extractParts(MI.getOperand(2).getReg(), NarrowTy, NumParts, Src2Regs);
unsigned CarryIn = MRI.createGenericVirtualRegister(LLT::scalar(1));
MIRBuilder.buildConstant(CarryIn, 0);
for (int i = 0; i < NumParts; ++i) {
unsigned DstReg = MRI.createGenericVirtualRegister(NarrowTy);
unsigned CarryOut = MRI.createGenericVirtualRegister(LLT::scalar(1));
MIRBuilder.buildUAdde(DstReg, CarryOut, Src1Regs[i],
Src2Regs[i], CarryIn);
DstRegs.push_back(DstReg);
Indexes.push_back(i * NarrowSize);
CarryIn = CarryOut;
}
unsigned DstReg = MI.getOperand(0).getReg();
MIRBuilder.buildSequence(DstReg, DstRegs, Indexes);
MI.eraseFromParent();
return Legalized;
}
}
}
unsigned TargetRegisterInfo::getRegSizeInBits(unsigned Reg,
const MachineRegisterInfo &MRI) const {
const TargetRegisterClass *RC{};
if (isPhysicalRegister(Reg)) {
// The size is not directly available for physical registers.
// Instead, we need to access a register class that contains Reg and
// get the size of that register class.
RC = getMinimalPhysRegClass(Reg);
} else {
LLT Ty = MRI.getType(Reg);
unsigned RegSize = Ty.isValid() ? Ty.getSizeInBits() : 0;
// If Reg is not a generic register, query the register class to
// get its size.
if (RegSize)
return RegSize;
// Since Reg is not a generic register, it must have a register class.
RC = MRI.getRegClass(Reg);
}
assert(RC && "Unable to deduce the register class");
return getRegSizeInBits(*RC);
}
bool CKernelRidgeRegression::solve_krr_system()
{
SGMatrix<float64_t> kernel_matrix(kernel->get_kernel_matrix());
int32_t n = kernel_matrix.num_rows;
SGVector<float64_t> y = ((CRegressionLabels*)m_labels)->get_labels();
for(index_t i=0; i<n; i++)
kernel_matrix(i,i) += m_tau;
Map<MatrixXd> eigen_kernel_matrix(kernel_matrix.matrix, n, n);
Map<VectorXd> eigen_alphas(m_alpha.vector, n);
Map<VectorXd> eigen_y(y.vector, n);
LLT<MatrixXd> llt;
llt.compute(eigen_kernel_matrix);
if (llt.info() != Eigen::Success)
{
SG_WARNING("Features covariance matrix was not positive definite\n");
return false;
}
eigen_alphas = llt.solve(eigen_y);
return true;
}
RegisterBankInfo::InstructionMapping
AArch64RegisterBankInfo::getInstrMapping(const MachineInstr &MI) const {
RegisterBankInfo::InstructionMapping Mapping = getInstrMappingImpl(MI);
if (Mapping.isValid())
return Mapping;
// As a top-level guess, vectors go in FPRs, scalars in GPRs. Obviously this
// won't work for normal floating-point types (or NZCV). When such
// instructions exist we'll need to look at the MI's opcode.
LLT Ty = MI.getType();
unsigned BankID;
if (Ty.isVector())
BankID = AArch64::FPRRegBankID;
else
BankID = AArch64::GPRRegBankID;
Mapping = InstructionMapping{1, 1, MI.getNumOperands()};
int Size = Ty.isSized() ? Ty.getSizeInBits() : 0;
for (unsigned Idx = 0; Idx < MI.getNumOperands(); ++Idx)
Mapping.setOperandMapping(Idx, Size, getRegBank(BankID));
return Mapping;
}
unsigned
MachineRegisterInfo::createGenericVirtualRegister(LLT Ty) {
assert(Ty.isValid() && "Cannot create empty virtual register");
// New virtual register number.
unsigned Reg = TargetRegisterInfo::index2VirtReg(getNumVirtRegs());
VRegInfo.grow(Reg);
// FIXME: Should we use a dummy register bank?
VRegInfo[Reg].first = static_cast<RegisterBank *>(nullptr);
getVRegToType()[Reg] = Ty;
RegAllocHints.grow(Reg);
if (TheDelegate)
TheDelegate->MRI_NoteNewVirtualRegister(Reg);
return Reg;
}
X86GenRegisterBankInfo::PartialMappingIdx
X86GenRegisterBankInfo::getPartialMappingIdx(const LLT &Ty, bool isFP) {
if ((Ty.isScalar() && !isFP) || Ty.isPointer()) {
switch (Ty.getSizeInBits()) {
case 1:
case 8:
return PMI_GPR8;
case 16:
return PMI_GPR16;
case 32:
return PMI_GPR32;
case 64:
return PMI_GPR64;
case 128:
return PMI_VEC128;
break;
default:
llvm_unreachable("Unsupported register size.");
}
} else if (Ty.isScalar()) {
switch (Ty.getSizeInBits()) {
case 32:
return PMI_FP32;
case 64:
return PMI_FP64;
case 128:
return PMI_VEC128;
default:
llvm_unreachable("Unsupported register size.");
}
} else {
switch (Ty.getSizeInBits()) {
case 128:
return PMI_VEC128;
case 256:
return PMI_VEC256;
case 512:
return PMI_VEC512;
default:
llvm_unreachable("Unsupported register size.");
}
}
return PMI_None;
}
const RegisterBankInfo::InstructionMapping &
AArch64RegisterBankInfo::getSameKindOfOperandsMapping(
const MachineInstr &MI) const {
const unsigned Opc = MI.getOpcode();
const MachineFunction &MF = *MI.getParent()->getParent();
const MachineRegisterInfo &MRI = MF.getRegInfo();
unsigned NumOperands = MI.getNumOperands();
assert(NumOperands <= 3 &&
"This code is for instructions with 3 or less operands");
LLT Ty = MRI.getType(MI.getOperand(0).getReg());
unsigned Size = Ty.getSizeInBits();
bool IsFPR = Ty.isVector() || isPreISelGenericFloatingPointOpcode(Opc);
PartialMappingIdx RBIdx = IsFPR ? PMI_FirstFPR : PMI_FirstGPR;
#ifndef NDEBUG
// Make sure all the operands are using similar size and type.
// Should probably be checked by the machine verifier.
// This code won't catch cases where the number of lanes is
// different between the operands.
// If we want to go to that level of details, it is probably
// best to check that the types are the same, period.
// Currently, we just check that the register banks are the same
// for each types.
for (unsigned Idx = 1; Idx != NumOperands; ++Idx) {
LLT OpTy = MRI.getType(MI.getOperand(Idx).getReg());
assert(
AArch64GenRegisterBankInfo::getRegBankBaseIdxOffset(
RBIdx, OpTy.getSizeInBits()) ==
AArch64GenRegisterBankInfo::getRegBankBaseIdxOffset(RBIdx, Size) &&
"Operand has incompatible size");
bool OpIsFPR = OpTy.isVector() || isPreISelGenericFloatingPointOpcode(Opc);
(void)OpIsFPR;
assert(IsFPR == OpIsFPR && "Operand has incompatible type");
}
#endif // End NDEBUG.
return getInstructionMapping(DefaultMappingID, 1,
getValueMapping(RBIdx, Size), NumOperands);
}
AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST,
const GCNTargetMachine &TM) {
using namespace TargetOpcode;
auto GetAddrSpacePtr = [&TM](unsigned AS) {
return LLT::pointer(AS, TM.getPointerSizeInBits(AS));
};
const LLT S1 = LLT::scalar(1);
const LLT S8 = LLT::scalar(8);
const LLT S16 = LLT::scalar(16);
const LLT S32 = LLT::scalar(32);
const LLT S64 = LLT::scalar(64);
const LLT S128 = LLT::scalar(128);
const LLT S256 = LLT::scalar(256);
const LLT S512 = LLT::scalar(512);
const LLT V2S16 = LLT::vector(2, 16);
const LLT V4S16 = LLT::vector(4, 16);
const LLT V8S16 = LLT::vector(8, 16);
const LLT V2S32 = LLT::vector(2, 32);
const LLT V3S32 = LLT::vector(3, 32);
const LLT V4S32 = LLT::vector(4, 32);
const LLT V5S32 = LLT::vector(5, 32);
const LLT V6S32 = LLT::vector(6, 32);
const LLT V7S32 = LLT::vector(7, 32);
const LLT V8S32 = LLT::vector(8, 32);
const LLT V9S32 = LLT::vector(9, 32);
const LLT V10S32 = LLT::vector(10, 32);
const LLT V11S32 = LLT::vector(11, 32);
const LLT V12S32 = LLT::vector(12, 32);
const LLT V13S32 = LLT::vector(13, 32);
const LLT V14S32 = LLT::vector(14, 32);
const LLT V15S32 = LLT::vector(15, 32);
const LLT V16S32 = LLT::vector(16, 32);
const LLT V2S64 = LLT::vector(2, 64);
const LLT V3S64 = LLT::vector(3, 64);
const LLT V4S64 = LLT::vector(4, 64);
const LLT V5S64 = LLT::vector(5, 64);
const LLT V6S64 = LLT::vector(6, 64);
const LLT V7S64 = LLT::vector(7, 64);
const LLT V8S64 = LLT::vector(8, 64);
std::initializer_list<LLT> AllS32Vectors =
{V2S32, V3S32, V4S32, V5S32, V6S32, V7S32, V8S32,
V9S32, V10S32, V11S32, V12S32, V13S32, V14S32, V15S32, V16S32};
std::initializer_list<LLT> AllS64Vectors =
{V2S64, V3S64, V4S64, V5S64, V6S64, V7S64, V8S64};
const LLT GlobalPtr = GetAddrSpacePtr(AMDGPUAS::GLOBAL_ADDRESS);
const LLT ConstantPtr = GetAddrSpacePtr(AMDGPUAS::CONSTANT_ADDRESS);
const LLT LocalPtr = GetAddrSpacePtr(AMDGPUAS::LOCAL_ADDRESS);
const LLT FlatPtr = GetAddrSpacePtr(AMDGPUAS::FLAT_ADDRESS);
const LLT PrivatePtr = GetAddrSpacePtr(AMDGPUAS::PRIVATE_ADDRESS);
const LLT CodePtr = FlatPtr;
const std::initializer_list<LLT> AddrSpaces64 = {
GlobalPtr, ConstantPtr, FlatPtr
};
const std::initializer_list<LLT> AddrSpaces32 = {
LocalPtr, PrivatePtr
};
setAction({G_BRCOND, S1}, Legal);
// TODO: All multiples of 32, vectors of pointers, all v2s16 pairs, more
// elements for v3s16
getActionDefinitionsBuilder(G_PHI)
.legalFor({S32, S64, V2S16, V4S16, S1, S128, S256})
.legalFor(AllS32Vectors)
.legalFor(AllS64Vectors)
.legalFor(AddrSpaces64)
.legalFor(AddrSpaces32)
.clampScalar(0, S32, S256)
.widenScalarToNextPow2(0, 32)
.clampMaxNumElements(0, S32, 16)
.moreElementsIf(isSmallOddVector(0), oneMoreElement(0))
.legalIf(isPointer(0));
getActionDefinitionsBuilder({G_ADD, G_SUB, G_MUL, G_UMULH, G_SMULH})
.legalFor({S32})
.clampScalar(0, S32, S32)
.scalarize(0);
// Report legal for any types we can handle anywhere. For the cases only legal
// on the SALU, RegBankSelect will be able to re-legalize.
getActionDefinitionsBuilder({G_AND, G_OR, G_XOR})
.legalFor({S32, S1, S64, V2S32, V2S16, V4S16})
.clampScalar(0, S32, S64)
.moreElementsIf(isSmallOddVector(0), oneMoreElement(0))
.fewerElementsIf(vectorWiderThan(0, 32), fewerEltsToSize64Vector(0))
.widenScalarToNextPow2(0)
.scalarize(0);
getActionDefinitionsBuilder({G_UADDO, G_SADDO, G_USUBO, G_SSUBO,
G_UADDE, G_SADDE, G_USUBE, G_SSUBE})
.legalFor({{S32, S1}})
.clampScalar(0, S32, S32);
getActionDefinitionsBuilder(G_BITCAST)
.legalForCartesianProduct({S32, V2S16})
.legalForCartesianProduct({S64, V2S32, V4S16})
.legalForCartesianProduct({V2S64, V4S32})
// Don't worry about the size constraint.
.legalIf(all(isPointer(0), isPointer(1)));
if (ST.has16BitInsts()) {
getActionDefinitionsBuilder(G_FCONSTANT)
.legalFor({S32, S64, S16})
.clampScalar(0, S16, S64);
} else {
getActionDefinitionsBuilder(G_FCONSTANT)
.legalFor({S32, S64})
.clampScalar(0, S32, S64);
}
getActionDefinitionsBuilder(G_IMPLICIT_DEF)
.legalFor({S1, S32, S64, V2S32, V4S32, V2S16, V4S16, GlobalPtr,
ConstantPtr, LocalPtr, FlatPtr, PrivatePtr})
.moreElementsIf(isSmallOddVector(0), oneMoreElement(0))
.clampScalarOrElt(0, S32, S512)
.legalIf(isMultiple32(0))
.widenScalarToNextPow2(0, 32)
.clampMaxNumElements(0, S32, 16);
// FIXME: i1 operands to intrinsics should always be legal, but other i1
// values may not be legal. We need to figure out how to distinguish
// between these two scenarios.
getActionDefinitionsBuilder(G_CONSTANT)
.legalFor({S1, S32, S64, GlobalPtr,
LocalPtr, ConstantPtr, PrivatePtr, FlatPtr })
.clampScalar(0, S32, S64)
.widenScalarToNextPow2(0)
.legalIf(isPointer(0));
setAction({G_FRAME_INDEX, PrivatePtr}, Legal);
auto &FPOpActions = getActionDefinitionsBuilder(
{ G_FADD, G_FMUL, G_FNEG, G_FABS, G_FMA, G_FCANONICALIZE})
.legalFor({S32, S64});
if (ST.has16BitInsts()) {
if (ST.hasVOP3PInsts())
FPOpActions.legalFor({S16, V2S16});
else
FPOpActions.legalFor({S16});
}
if (ST.hasVOP3PInsts())
FPOpActions.clampMaxNumElements(0, S16, 2);
FPOpActions
.scalarize(0)
.clampScalar(0, ST.has16BitInsts() ? S16 : S32, S64);
if (ST.has16BitInsts()) {
getActionDefinitionsBuilder(G_FSQRT)
.legalFor({S32, S64, S16})
.scalarize(0)
.clampScalar(0, S16, S64);
} else {
getActionDefinitionsBuilder(G_FSQRT)
.legalFor({S32, S64})
.scalarize(0)
.clampScalar(0, S32, S64);
}
getActionDefinitionsBuilder(G_FPTRUNC)
.legalFor({{S32, S64}, {S16, S32}})
.scalarize(0);
getActionDefinitionsBuilder(G_FPEXT)
.legalFor({{S64, S32}, {S32, S16}})
.lowerFor({{S64, S16}}) // FIXME: Implement
.scalarize(0);
getActionDefinitionsBuilder(G_FCOPYSIGN)
.legalForCartesianProduct({S16, S32, S64}, {S16, S32, S64})
.scalarize(0);
getActionDefinitionsBuilder(G_FSUB)
// Use actual fsub instruction
.legalFor({S32})
// Must use fadd + fneg
.lowerFor({S64, S16, V2S16})
.scalarize(0)
.clampScalar(0, S32, S64);
getActionDefinitionsBuilder({G_SEXT, G_ZEXT, G_ANYEXT})
.legalFor({{S64, S32}, {S32, S16}, {S64, S16},
{S32, S1}, {S64, S1}, {S16, S1},
// FIXME: Hack
{S64, LLT::scalar(33)},
{S32, S8}, {S128, S32}, {S128, S64}, {S32, LLT::scalar(24)}})
.scalarize(0);
getActionDefinitionsBuilder({G_SITOFP, G_UITOFP})
.legalFor({{S32, S32}, {S64, S32}})
.lowerFor({{S32, S64}})
.customFor({{S64, S64}})
.scalarize(0);
getActionDefinitionsBuilder({G_FPTOSI, G_FPTOUI})
.legalFor({{S32, S32}, {S32, S64}})
.scalarize(0);
getActionDefinitionsBuilder(G_INTRINSIC_ROUND)
.legalFor({S32, S64})
.scalarize(0);
if (ST.getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS) {
getActionDefinitionsBuilder({G_INTRINSIC_TRUNC, G_FCEIL, G_FRINT})
.legalFor({S32, S64})
.clampScalar(0, S32, S64)
.scalarize(0);
} else {
getActionDefinitionsBuilder({G_INTRINSIC_TRUNC, G_FCEIL, G_FRINT})
.legalFor({S32})
.customFor({S64})
.clampScalar(0, S32, S64)
.scalarize(0);
}
getActionDefinitionsBuilder(G_GEP)
.legalForCartesianProduct(AddrSpaces64, {S64})
.legalForCartesianProduct(AddrSpaces32, {S32})
.scalarize(0);
setAction({G_BLOCK_ADDR, CodePtr}, Legal);
getActionDefinitionsBuilder(G_ICMP)
.legalForCartesianProduct(
{S1}, {S32, S64, GlobalPtr, LocalPtr, ConstantPtr, PrivatePtr, FlatPtr})
.legalFor({{S1, S32}, {S1, S64}})
.widenScalarToNextPow2(1)
.clampScalar(1, S32, S64)
.scalarize(0)
.legalIf(all(typeIs(0, S1), isPointer(1)));
getActionDefinitionsBuilder(G_FCMP)
.legalFor({{S1, S32}, {S1, S64}})
.widenScalarToNextPow2(1)
.clampScalar(1, S32, S64)
.scalarize(0);
// FIXME: fexp, flog2, flog10 needs to be custom lowered.
getActionDefinitionsBuilder({G_FPOW, G_FEXP, G_FEXP2,
G_FLOG, G_FLOG2, G_FLOG10})
.legalFor({S32})
.scalarize(0);
// The 64-bit versions produce 32-bit results, but only on the SALU.
getActionDefinitionsBuilder({G_CTLZ, G_CTLZ_ZERO_UNDEF,
G_CTTZ, G_CTTZ_ZERO_UNDEF,
G_CTPOP})
.legalFor({{S32, S32}, {S32, S64}})
.clampScalar(0, S32, S32)
.clampScalar(1, S32, S64)
.scalarize(0)
.widenScalarToNextPow2(0, 32)
.widenScalarToNextPow2(1, 32);
// TODO: Expand for > s32
getActionDefinitionsBuilder(G_BSWAP)
.legalFor({S32})
.clampScalar(0, S32, S32)
.scalarize(0);
auto smallerThan = [](unsigned TypeIdx0, unsigned TypeIdx1) {
return [=](const LegalityQuery &Query) {
return Query.Types[TypeIdx0].getSizeInBits() <
Query.Types[TypeIdx1].getSizeInBits();
};
};
auto greaterThan = [](unsigned TypeIdx0, unsigned TypeIdx1) {
return [=](const LegalityQuery &Query) {
return Query.Types[TypeIdx0].getSizeInBits() >
Query.Types[TypeIdx1].getSizeInBits();
};
};
getActionDefinitionsBuilder(G_INTTOPTR)
// List the common cases
.legalForCartesianProduct(AddrSpaces64, {S64})
.legalForCartesianProduct(AddrSpaces32, {S32})
.scalarize(0)
// Accept any address space as long as the size matches
.legalIf(sameSize(0, 1))
.widenScalarIf(smallerThan(1, 0),
[](const LegalityQuery &Query) {
return std::make_pair(1, LLT::scalar(Query.Types[0].getSizeInBits()));
})
.narrowScalarIf(greaterThan(1, 0),
[](const LegalityQuery &Query) {
return std::make_pair(1, LLT::scalar(Query.Types[0].getSizeInBits()));
});
getActionDefinitionsBuilder(G_PTRTOINT)
// List the common cases
.legalForCartesianProduct(AddrSpaces64, {S64})
.legalForCartesianProduct(AddrSpaces32, {S32})
.scalarize(0)
// Accept any address space as long as the size matches
.legalIf(sameSize(0, 1))
.widenScalarIf(smallerThan(0, 1),
[](const LegalityQuery &Query) {
return std::make_pair(0, LLT::scalar(Query.Types[1].getSizeInBits()));
})
.narrowScalarIf(
greaterThan(0, 1),
[](const LegalityQuery &Query) {
return std::make_pair(0, LLT::scalar(Query.Types[1].getSizeInBits()));
});
if (ST.hasFlatAddressSpace()) {
getActionDefinitionsBuilder(G_ADDRSPACE_CAST)
.scalarize(0)
.custom();
}
getActionDefinitionsBuilder({G_LOAD, G_STORE})
.narrowScalarIf([](const LegalityQuery &Query) {
unsigned Size = Query.Types[0].getSizeInBits();
unsigned MemSize = Query.MMODescrs[0].SizeInBits;
return (Size > 32 && MemSize < Size);
},
[](const LegalityQuery &Query) {
return std::make_pair(0, LLT::scalar(32));
})
.fewerElementsIf([=, &ST](const LegalityQuery &Query) {
unsigned MemSize = Query.MMODescrs[0].SizeInBits;
return (MemSize == 96) &&
Query.Types[0].isVector() &&
ST.getGeneration() < AMDGPUSubtarget::SEA_ISLANDS;
},
[=](const LegalityQuery &Query) {
return std::make_pair(0, V2S32);
})
.legalIf([=, &ST](const LegalityQuery &Query) {
const LLT &Ty0 = Query.Types[0];
unsigned Size = Ty0.getSizeInBits();
unsigned MemSize = Query.MMODescrs[0].SizeInBits;
if (Size < 32 || (Size > 32 && MemSize < Size))
return false;
if (Ty0.isVector() && Size != MemSize)
return false;
// TODO: Decompose private loads into 4-byte components.
// TODO: Illegal flat loads on SI
switch (MemSize) {
case 8:
case 16:
return Size == 32;
case 32:
case 64:
case 128:
return true;
case 96:
// XXX hasLoadX3
return (ST.getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS);
case 256:
case 512:
// TODO: constant loads
default:
return false;
}
})
.clampScalar(0, S32, S64);
// FIXME: Handle alignment requirements.
auto &ExtLoads = getActionDefinitionsBuilder({G_SEXTLOAD, G_ZEXTLOAD})
.legalForTypesWithMemDesc({
{S32, GlobalPtr, 8, 8},
{S32, GlobalPtr, 16, 8},
{S32, LocalPtr, 8, 8},
{S32, LocalPtr, 16, 8},
{S32, PrivatePtr, 8, 8},
{S32, PrivatePtr, 16, 8}});
if (ST.hasFlatAddressSpace()) {
ExtLoads.legalForTypesWithMemDesc({{S32, FlatPtr, 8, 8},
{S32, FlatPtr, 16, 8}});
}
ExtLoads.clampScalar(0, S32, S32)
.widenScalarToNextPow2(0)
.unsupportedIfMemSizeNotPow2()
.lower();
auto &Atomics = getActionDefinitionsBuilder(
{G_ATOMICRMW_XCHG, G_ATOMICRMW_ADD, G_ATOMICRMW_SUB,
G_ATOMICRMW_AND, G_ATOMICRMW_OR, G_ATOMICRMW_XOR,
G_ATOMICRMW_MAX, G_ATOMICRMW_MIN, G_ATOMICRMW_UMAX,
G_ATOMICRMW_UMIN, G_ATOMIC_CMPXCHG})
.legalFor({{S32, GlobalPtr}, {S32, LocalPtr},
{S64, GlobalPtr}, {S64, LocalPtr}});
if (ST.hasFlatAddressSpace()) {
Atomics.legalFor({{S32, FlatPtr}, {S64, FlatPtr}});
}
// TODO: Pointer types, any 32-bit or 64-bit vector
getActionDefinitionsBuilder(G_SELECT)
.legalForCartesianProduct({S32, S64, V2S32, V2S16, V4S16,
GlobalPtr, LocalPtr, FlatPtr, PrivatePtr,
LLT::vector(2, LocalPtr), LLT::vector(2, PrivatePtr)}, {S1})
.clampScalar(0, S32, S64)
.moreElementsIf(isSmallOddVector(0), oneMoreElement(0))
.fewerElementsIf(numElementsNotEven(0), scalarize(0))
.scalarize(1)
.clampMaxNumElements(0, S32, 2)
.clampMaxNumElements(0, LocalPtr, 2)
.clampMaxNumElements(0, PrivatePtr, 2)
.scalarize(0)
.widenScalarToNextPow2(0)
.legalIf(all(isPointer(0), typeIs(1, S1)));
// TODO: Only the low 4/5/6 bits of the shift amount are observed, so we can
// be more flexible with the shift amount type.
auto &Shifts = getActionDefinitionsBuilder({G_SHL, G_LSHR, G_ASHR})
.legalFor({{S32, S32}, {S64, S32}});
if (ST.has16BitInsts()) {
if (ST.hasVOP3PInsts()) {
Shifts.legalFor({{S16, S32}, {S16, S16}, {V2S16, V2S16}})
.clampMaxNumElements(0, S16, 2);
} else
Shifts.legalFor({{S16, S32}, {S16, S16}});
Shifts.clampScalar(1, S16, S32);
Shifts.clampScalar(0, S16, S64);
Shifts.widenScalarToNextPow2(0, 16);
} else {
// Make sure we legalize the shift amount type first, as the general
// expansion for the shifted type will produce much worse code if it hasn't
// been truncated already.
Shifts.clampScalar(1, S32, S32);
Shifts.clampScalar(0, S32, S64);
Shifts.widenScalarToNextPow2(0, 32);
}
Shifts.scalarize(0);
for (unsigned Op : {G_EXTRACT_VECTOR_ELT, G_INSERT_VECTOR_ELT}) {
unsigned VecTypeIdx = Op == G_EXTRACT_VECTOR_ELT ? 1 : 0;
unsigned EltTypeIdx = Op == G_EXTRACT_VECTOR_ELT ? 0 : 1;
unsigned IdxTypeIdx = 2;
getActionDefinitionsBuilder(Op)
.legalIf([=](const LegalityQuery &Query) {
const LLT &VecTy = Query.Types[VecTypeIdx];
const LLT &IdxTy = Query.Types[IdxTypeIdx];
return VecTy.getSizeInBits() % 32 == 0 &&
VecTy.getSizeInBits() <= 512 &&
IdxTy.getSizeInBits() == 32;
})
.clampScalar(EltTypeIdx, S32, S64)
.clampScalar(VecTypeIdx, S32, S64)
.clampScalar(IdxTypeIdx, S32, S32);
}
getActionDefinitionsBuilder(G_EXTRACT_VECTOR_ELT)
.unsupportedIf([=](const LegalityQuery &Query) {
const LLT &EltTy = Query.Types[1].getElementType();
return Query.Types[0] != EltTy;
});
for (unsigned Op : {G_EXTRACT, G_INSERT}) {
unsigned BigTyIdx = Op == G_EXTRACT ? 1 : 0;
unsigned LitTyIdx = Op == G_EXTRACT ? 0 : 1;
// FIXME: Doesn't handle extract of illegal sizes.
getActionDefinitionsBuilder(Op)
.legalIf([=](const LegalityQuery &Query) {
const LLT BigTy = Query.Types[BigTyIdx];
const LLT LitTy = Query.Types[LitTyIdx];
return (BigTy.getSizeInBits() % 32 == 0) &&
(LitTy.getSizeInBits() % 16 == 0);
})
.widenScalarIf(
[=](const LegalityQuery &Query) {
const LLT BigTy = Query.Types[BigTyIdx];
return (BigTy.getScalarSizeInBits() < 16);
},
LegalizeMutations::widenScalarOrEltToNextPow2(BigTyIdx, 16))
.widenScalarIf(
[=](const LegalityQuery &Query) {
const LLT LitTy = Query.Types[LitTyIdx];
return (LitTy.getScalarSizeInBits() < 16);
},
LegalizeMutations::widenScalarOrEltToNextPow2(LitTyIdx, 16))
.moreElementsIf(isSmallOddVector(BigTyIdx), oneMoreElement(BigTyIdx))
.widenScalarToNextPow2(BigTyIdx, 32);
}
// TODO: vectors of pointers
getActionDefinitionsBuilder(G_BUILD_VECTOR)
.legalForCartesianProduct(AllS32Vectors, {S32})
.legalForCartesianProduct(AllS64Vectors, {S64})
.clampNumElements(0, V16S32, V16S32)
.clampNumElements(0, V2S64, V8S64)
.minScalarSameAs(1, 0)
// FIXME: Sort of a hack to make progress on other legalizations.
.legalIf([=](const LegalityQuery &Query) {
return Query.Types[0].getScalarSizeInBits() <= 32 ||
Query.Types[0].getScalarSizeInBits() == 64;
});
// TODO: Support any combination of v2s32
getActionDefinitionsBuilder(G_CONCAT_VECTORS)
.legalFor({{V4S32, V2S32},
{V8S32, V2S32},
{V8S32, V4S32},
{V4S64, V2S64},
{V4S16, V2S16},
{V8S16, V2S16},
{V8S16, V4S16},
{LLT::vector(4, LocalPtr), LLT::vector(2, LocalPtr)},
{LLT::vector(4, PrivatePtr), LLT::vector(2, PrivatePtr)}});
// Merge/Unmerge
for (unsigned Op : {G_MERGE_VALUES, G_UNMERGE_VALUES}) {
unsigned BigTyIdx = Op == G_MERGE_VALUES ? 0 : 1;
unsigned LitTyIdx = Op == G_MERGE_VALUES ? 1 : 0;
auto notValidElt = [=](const LegalityQuery &Query, unsigned TypeIdx) {
const LLT &Ty = Query.Types[TypeIdx];
if (Ty.isVector()) {
const LLT &EltTy = Ty.getElementType();
if (EltTy.getSizeInBits() < 8 || EltTy.getSizeInBits() > 64)
return true;
if (!isPowerOf2_32(EltTy.getSizeInBits()))
return true;
}
return false;
};
getActionDefinitionsBuilder(Op)
.widenScalarToNextPow2(LitTyIdx, /*Min*/ 16)
// Clamp the little scalar to s8-s256 and make it a power of 2. It's not
// worth considering the multiples of 64 since 2*192 and 2*384 are not
// valid.
.clampScalar(LitTyIdx, S16, S256)
.widenScalarToNextPow2(LitTyIdx, /*Min*/ 32)
// Break up vectors with weird elements into scalars
.fewerElementsIf(
[=](const LegalityQuery &Query) { return notValidElt(Query, 0); },
scalarize(0))
.fewerElementsIf(
[=](const LegalityQuery &Query) { return notValidElt(Query, 1); },
scalarize(1))
.clampScalar(BigTyIdx, S32, S512)
.widenScalarIf(
[=](const LegalityQuery &Query) {
const LLT &Ty = Query.Types[BigTyIdx];
return !isPowerOf2_32(Ty.getSizeInBits()) &&
Ty.getSizeInBits() % 16 != 0;
},
[=](const LegalityQuery &Query) {
// Pick the next power of 2, or a multiple of 64 over 128.
// Whichever is smaller.
const LLT &Ty = Query.Types[BigTyIdx];
unsigned NewSizeInBits = 1 << Log2_32_Ceil(Ty.getSizeInBits() + 1);
if (NewSizeInBits >= 256) {
unsigned RoundedTo = alignTo<64>(Ty.getSizeInBits() + 1);
if (RoundedTo < NewSizeInBits)
NewSizeInBits = RoundedTo;
}
return std::make_pair(BigTyIdx, LLT::scalar(NewSizeInBits));
})
.legalIf([=](const LegalityQuery &Query) {
const LLT &BigTy = Query.Types[BigTyIdx];
const LLT &LitTy = Query.Types[LitTyIdx];
if (BigTy.isVector() && BigTy.getSizeInBits() < 32)
return false;
if (LitTy.isVector() && LitTy.getSizeInBits() < 32)
return false;
return BigTy.getSizeInBits() % 16 == 0 &&
LitTy.getSizeInBits() % 16 == 0 &&
BigTy.getSizeInBits() <= 512;
})
// Any vectors left are the wrong size. Scalarize them.
.scalarize(0)
.scalarize(1);
}
computeTables();
verify(*ST.getInstrInfo());
}
bool AMDGPULegalizerInfo::legalizeAddrSpaceCast(
MachineInstr &MI, MachineRegisterInfo &MRI,
MachineIRBuilder &MIRBuilder) const {
MachineFunction &MF = MIRBuilder.getMF();
MIRBuilder.setInstr(MI);
unsigned Dst = MI.getOperand(0).getReg();
unsigned Src = MI.getOperand(1).getReg();
LLT DstTy = MRI.getType(Dst);
LLT SrcTy = MRI.getType(Src);
unsigned DestAS = DstTy.getAddressSpace();
unsigned SrcAS = SrcTy.getAddressSpace();
// TODO: Avoid reloading from the queue ptr for each cast, or at least each
// vector element.
assert(!DstTy.isVector());
const AMDGPUTargetMachine &TM
= static_cast<const AMDGPUTargetMachine &>(MF.getTarget());
const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
if (ST.getTargetLowering()->isNoopAddrSpaceCast(SrcAS, DestAS)) {
MI.setDesc(MIRBuilder.getTII().get(TargetOpcode::G_BITCAST));
return true;
}
if (SrcAS == AMDGPUAS::FLAT_ADDRESS) {
assert(DestAS == AMDGPUAS::LOCAL_ADDRESS ||
DestAS == AMDGPUAS::PRIVATE_ADDRESS);
unsigned NullVal = TM.getNullPointerValue(DestAS);
auto SegmentNull = MIRBuilder.buildConstant(DstTy, NullVal);
auto FlatNull = MIRBuilder.buildConstant(SrcTy, 0);
unsigned PtrLo32 = MRI.createGenericVirtualRegister(DstTy);
// Extract low 32-bits of the pointer.
MIRBuilder.buildExtract(PtrLo32, Src, 0);
unsigned CmpRes = MRI.createGenericVirtualRegister(LLT::scalar(1));
MIRBuilder.buildICmp(CmpInst::ICMP_NE, CmpRes, Src, FlatNull.getReg(0));
MIRBuilder.buildSelect(Dst, CmpRes, PtrLo32, SegmentNull.getReg(0));
MI.eraseFromParent();
return true;
}
assert(SrcAS == AMDGPUAS::LOCAL_ADDRESS ||
SrcAS == AMDGPUAS::PRIVATE_ADDRESS);
auto SegmentNull =
MIRBuilder.buildConstant(SrcTy, TM.getNullPointerValue(SrcAS));
auto FlatNull =
MIRBuilder.buildConstant(DstTy, TM.getNullPointerValue(DestAS));
unsigned ApertureReg = getSegmentAperture(DestAS, MRI, MIRBuilder);
unsigned CmpRes = MRI.createGenericVirtualRegister(LLT::scalar(1));
MIRBuilder.buildICmp(CmpInst::ICMP_NE, CmpRes, Src, SegmentNull.getReg(0));
unsigned BuildPtr = MRI.createGenericVirtualRegister(DstTy);
// Coerce the type of the low half of the result so we can use merge_values.
unsigned SrcAsInt = MRI.createGenericVirtualRegister(LLT::scalar(32));
MIRBuilder.buildInstr(TargetOpcode::G_PTRTOINT)
.addDef(SrcAsInt)
.addUse(Src);
// TODO: Should we allow mismatched types but matching sizes in merges to
// avoid the ptrtoint?
MIRBuilder.buildMerge(BuildPtr, {SrcAsInt, ApertureReg});
MIRBuilder.buildSelect(Dst, CmpRes, BuildPtr, FlatNull.getReg(0));
MI.eraseFromParent();
return true;
}
