}
// NVPTXTargetLowering Constructor.
-NVPTXTargetLowering::NVPTXTargetLowering(NVPTXTargetMachine &TM)
- : TargetLowering(TM, new NVPTXTargetObjectFile()), nvTM(&TM),
- nvptxSubtarget(TM.getSubtarget<NVPTXSubtarget>()) {
+NVPTXTargetLowering::NVPTXTargetLowering(const NVPTXTargetMachine &TM,
+ const NVPTXSubtarget &STI)
+ : TargetLowering(TM), nvTM(&TM), STI(STI) {
// always lower memset, memcpy, and memmove intrinsics to load/store
// instructions, rather
setOperationAction(ISD::SRA_PARTS, MVT::i64 , Custom);
setOperationAction(ISD::SRL_PARTS, MVT::i64 , Custom);
- if (nvptxSubtarget.hasROT64()) {
+ if (STI.hasROT64()) {
setOperationAction(ISD::ROTL, MVT::i64, Legal);
setOperationAction(ISD::ROTR, MVT::i64, Legal);
} else {
setOperationAction(ISD::ROTL, MVT::i64, Expand);
setOperationAction(ISD::ROTR, MVT::i64, Expand);
}
- if (nvptxSubtarget.hasROT32()) {
+ if (STI.hasROT32()) {
setOperationAction(ISD::ROTL, MVT::i32, Legal);
setOperationAction(ISD::ROTR, MVT::i32, Legal);
} else {
setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
// Turn FP extload into load/fextend
- setLoadExtAction(ISD::EXTLOAD, MVT::f16, Expand);
- setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
+ setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
+ setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
+ setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
// Turn FP truncstore into trunc + store.
setTruncStoreAction(MVT::f32, MVT::f16, Expand);
setTruncStoreAction(MVT::f64, MVT::f16, Expand);
setOperationAction(ISD::LOAD, MVT::i1, Custom);
setOperationAction(ISD::STORE, MVT::i1, Custom);
- setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
- setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
- setTruncStoreAction(MVT::i64, MVT::i1, Expand);
- setTruncStoreAction(MVT::i32, MVT::i1, Expand);
- setTruncStoreAction(MVT::i16, MVT::i1, Expand);
- setTruncStoreAction(MVT::i8, MVT::i1, Expand);
+ for (MVT VT : MVT::integer_valuetypes()) {
+ setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
+ setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
+ setTruncStoreAction(VT, MVT::i1, Expand);
+ }
// This is legal in NVPTX
setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
setOperationAction(ISD::ADDE, MVT::i64, Expand);
// Register custom handling for vector loads/stores
- for (int i = MVT::FIRST_VECTOR_VALUETYPE; i <= MVT::LAST_VECTOR_VALUETYPE;
- ++i) {
- MVT VT = (MVT::SimpleValueType) i;
+ for (MVT VT : MVT::vector_valuetypes()) {
if (IsPTXVectorType(VT)) {
setOperationAction(ISD::LOAD, VT, Custom);
setOperationAction(ISD::STORE, VT, Custom);
setOperationAction(ISD::CTPOP, MVT::i32, Legal);
setOperationAction(ISD::CTPOP, MVT::i64, Legal);
+ // PTX does not directly support SELP of i1, so promote to i32 first
+ setOperationAction(ISD::SELECT, MVT::i1, Custom);
+
// We have some custom DAG combine patterns for these nodes
setTargetDAGCombine(ISD::ADD);
setTargetDAGCombine(ISD::AND);
// Now deduce the information based on the above mentioned
// actions
- computeRegisterProperties();
+ computeRegisterProperties(STI.getRegisterInfo());
}
const char *NVPTXTargetLowering::getTargetNodeName(unsigned Opcode) const {
unsigned retAlignment,
const ImmutableCallSite *CS) const {
- bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
+ bool isABI = (STI.getSmVersion() >= 20);
assert(isABI && "Non-ABI compilation is not supported");
if (!isABI)
return "";
O << ".param .b" << size << " _";
} else if (isa<PointerType>(retTy)) {
O << ".param .b" << getPointerTy().getSizeInBits() << " _";
+ } else if ((retTy->getTypeID() == Type::StructTyID) ||
+ isa<VectorType>(retTy)) {
+ O << ".param .align "
+ << retAlignment
+ << " .b8 _["
+ << getDataLayout()->getTypeAllocSize(retTy) << "]";
} else {
- if((retTy->getTypeID() == Type::StructTyID) ||
- isa<VectorType>(retTy)) {
- O << ".param .align "
- << retAlignment
- << " .b8 _["
- << getDataLayout()->getTypeAllocSize(retTy) << "]";
- } else {
- assert(false && "Unknown return type");
- }
+ llvm_unreachable("Unknown return type");
}
O << ") ";
}
Type *retTy = CLI.RetTy;
ImmutableCallSite *CS = CLI.CS;
- bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
+ bool isABI = (STI.getSmVersion() >= 20);
assert(isABI && "Non-ABI compilation is not supported");
if (!isABI)
return Chain;
// .param .align 16 .b8 retval0[<size-in-bytes>], or
// .param .b<size-in-bits> retval0
unsigned resultsz = TD->getTypeAllocSizeInBits(retTy);
- if (retTy->isSingleValueType()) {
+ // Emit ".param .b<size-in-bits> retval0" instead of byte arrays only for
+ // these three types to match the logic in
+ // NVPTXAsmPrinter::printReturnValStr and NVPTXTargetLowering::getPrototype.
+ // Plus, this behavior is consistent with nvcc's.
+ if (retTy->isFloatingPointTy() || retTy->isIntegerTy() ||
+ retTy->isPointerTy()) {
// Scalar needs to be at least 32bit wide
if (resultsz < 32)
resultsz = 32;
EVT ObjectVT = getValueType(retTy);
unsigned NumElts = ObjectVT.getVectorNumElements();
EVT EltVT = ObjectVT.getVectorElementType();
- assert(nvTM->getTargetLowering()->getNumRegisters(F->getContext(),
- ObjectVT) == NumElts &&
+ assert(STI.getTargetLowering()->getNumRegisters(F->getContext(),
+ ObjectVT) == NumElts &&
"Vector was not scalarized");
unsigned sz = EltVT.getSizeInBits();
bool needTruncate = sz < 8 ? true : false;
LoadRetVTs.push_back(EltVT);
LoadRetVTs.push_back(MVT::Other);
LoadRetVTs.push_back(MVT::Glue);
- SmallVector<SDValue, 4> LoadRetOps;
- LoadRetOps.push_back(Chain);
- LoadRetOps.push_back(DAG.getConstant(1, MVT::i32));
- LoadRetOps.push_back(DAG.getConstant(0, MVT::i32));
- LoadRetOps.push_back(InFlag);
+ SDValue LoadRetOps[] = {Chain, DAG.getConstant(1, MVT::i32),
+ DAG.getConstant(0, MVT::i32), InFlag};
SDValue retval = DAG.getMemIntrinsicNode(
NVPTXISD::LoadParam, dl,
DAG.getVTList(LoadRetVTs), LoadRetOps, EltVT, MachinePointerInfo());
}
LoadRetVTs.push_back(MVT::Other);
LoadRetVTs.push_back(MVT::Glue);
- SmallVector<SDValue, 4> LoadRetOps;
- LoadRetOps.push_back(Chain);
- LoadRetOps.push_back(DAG.getConstant(1, MVT::i32));
- LoadRetOps.push_back(DAG.getConstant(0, MVT::i32));
- LoadRetOps.push_back(InFlag);
+ SDValue LoadRetOps[] = {Chain, DAG.getConstant(1, MVT::i32),
+ DAG.getConstant(0, MVT::i32), InFlag};
SDValue retval = DAG.getMemIntrinsicNode(
NVPTXISD::LoadParamV2, dl,
DAG.getVTList(LoadRetVTs), LoadRetOps, EltVT, MachinePointerInfo());
}
LoadRetVTs.push_back(MVT::Other);
LoadRetVTs.push_back(MVT::Glue);
- SmallVector<SDValue, 4> LoadRetOps;
- LoadRetOps.push_back(Chain);
- LoadRetOps.push_back(DAG.getConstant(1, MVT::i32));
- LoadRetOps.push_back(DAG.getConstant(Ofst, MVT::i32));
- LoadRetOps.push_back(InFlag);
+ SDValue LoadRetOps[] = {Chain, DAG.getConstant(1, MVT::i32),
+ DAG.getConstant(Ofst, MVT::i32), InFlag};
SDValue retval = DAG.getMemIntrinsicNode(
Opc, dl, DAG.getVTList(LoadRetVTs),
LoadRetOps, EltVT, MachinePointerInfo());
LoadRetVTs.push_back(MVT::Other);
LoadRetVTs.push_back(MVT::Glue);
- SmallVector<SDValue, 4> LoadRetOps;
- LoadRetOps.push_back(Chain);
- LoadRetOps.push_back(DAG.getConstant(1, MVT::i32));
- LoadRetOps.push_back(DAG.getConstant(Offsets[i], MVT::i32));
- LoadRetOps.push_back(InFlag);
+ SDValue LoadRetOps[] = {Chain, DAG.getConstant(1, MVT::i32),
+ DAG.getConstant(Offsets[i], MVT::i32), InFlag};
SDValue retval = DAG.getMemIntrinsicNode(
NVPTXISD::LoadParam, dl,
DAG.getVTList(LoadRetVTs), LoadRetOps,
SDValue ShAmt = Op.getOperand(2);
unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
- if (VTBits == 32 && nvptxSubtarget.getSmVersion() >= 35) {
+ if (VTBits == 32 && STI.getSmVersion() >= 35) {
// For 32bit and sm35, we can use the funnel shift 'shf' instruction.
// {dHi, dLo} = {aHi, aLo} >> Amt
SDValue ShOpHi = Op.getOperand(1);
SDValue ShAmt = Op.getOperand(2);
- if (VTBits == 32 && nvptxSubtarget.getSmVersion() >= 35) {
+ if (VTBits == 32 && STI.getSmVersion() >= 35) {
// For 32bit and sm35, we can use the funnel shift 'shf' instruction.
// {dHi, dLo} = {aHi, aLo} << Amt
case ISD::SRA_PARTS:
case ISD::SRL_PARTS:
return LowerShiftRightParts(Op, DAG);
+ case ISD::SELECT:
+ return LowerSelect(Op, DAG);
default:
llvm_unreachable("Custom lowering not defined for operation");
}
}
+SDValue NVPTXTargetLowering::LowerSelect(SDValue Op, SelectionDAG &DAG) const {
+ SDValue Op0 = Op->getOperand(0);
+ SDValue Op1 = Op->getOperand(1);
+ SDValue Op2 = Op->getOperand(2);
+ SDLoc DL(Op.getNode());
+
+ assert(Op.getValueType() == MVT::i1 && "Custom lowering enabled only for i1");
+
+ Op1 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op1);
+ Op2 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op2);
+ SDValue Select = DAG.getNode(ISD::SELECT, DL, MVT::i32, Op0, Op1, Op2);
+ SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Select);
+
+ return Trunc;
+}
+
SDValue NVPTXTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
if (Op.getValueType() == MVT::i1)
return LowerLOADi1(Op, DAG);
}
// Then any remaining arguments
- for (unsigned i = 2, e = N->getNumOperands(); i != e; ++i) {
- Ops.push_back(N->getOperand(i));
- }
+ Ops.append(N->op_begin() + 2, N->op_end());
SDValue NewSt = DAG.getMemIntrinsicNode(
Opcode, DL, DAG.getVTList(MVT::Other), Ops,
const Function *F = MF.getFunction();
const AttributeSet &PAL = F->getAttributes();
- const TargetLowering *TLI = DAG.getTarget().getTargetLowering();
+ const TargetLowering *TLI = STI.getTargetLowering();
SDValue Root = DAG.getRoot();
std::vector<SDValue> OutChains;
bool isKernel = llvm::isKernelFunction(*F);
- bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
+ bool isABI = (STI.getSmVersion() >= 20);
assert(isABI && "Non-ABI compilation is not supported");
if (!isABI)
return Chain;
unsigned NumElts = ObjectVT.getVectorNumElements();
assert(TLI->getNumRegisters(F->getContext(), ObjectVT) == NumElts &&
"Vector was not scalarized");
- unsigned Ofst = 0;
EVT EltVT = ObjectVT.getVectorElementType();
// V1 load
// We only have one element, so just directly load it
Value *SrcValue = Constant::getNullValue(PointerType::get(
EltVT.getTypeForEVT(F->getContext()), llvm::ADDRESS_SPACE_PARAM));
- SDValue SrcAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg,
- DAG.getConstant(Ofst, getPointerTy()));
SDValue P = DAG.getLoad(
- EltVT, dl, Root, SrcAddr, MachinePointerInfo(SrcValue), false,
+ EltVT, dl, Root, Arg, MachinePointerInfo(SrcValue), false,
false, true,
TD->getABITypeAlignment(EltVT.getTypeForEVT(F->getContext())));
if (P.getNode())
if (Ins[InsIdx].VT.getSizeInBits() > EltVT.getSizeInBits())
P = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, P);
InVals.push_back(P);
- Ofst += TD->getTypeAllocSize(EltVT.getTypeForEVT(F->getContext()));
++InsIdx;
} else if (NumElts == 2) {
// V2 load
EVT VecVT = EVT::getVectorVT(F->getContext(), EltVT, 2);
Value *SrcValue = Constant::getNullValue(PointerType::get(
VecVT.getTypeForEVT(F->getContext()), llvm::ADDRESS_SPACE_PARAM));
- SDValue SrcAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg,
- DAG.getConstant(Ofst, getPointerTy()));
SDValue P = DAG.getLoad(
- VecVT, dl, Root, SrcAddr, MachinePointerInfo(SrcValue), false,
+ VecVT, dl, Root, Arg, MachinePointerInfo(SrcValue), false,
false, true,
TD->getABITypeAlignment(VecVT.getTypeForEVT(F->getContext())));
if (P.getNode())
InVals.push_back(Elt0);
InVals.push_back(Elt1);
- Ofst += TD->getTypeAllocSize(VecVT.getTypeForEVT(F->getContext()));
InsIdx += 2;
} else {
// V4 loads
VecSize = 2;
}
EVT VecVT = EVT::getVectorVT(F->getContext(), EltVT, VecSize);
+ unsigned Ofst = 0;
for (unsigned i = 0; i < NumElts; i += VecSize) {
Value *SrcValue = Constant::getNullValue(
PointerType::get(VecVT.getTypeForEVT(F->getContext()),
Type *RetTy = F->getReturnType();
const DataLayout *TD = getDataLayout();
- bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
+ bool isABI = (STI.getSmVersion() >= 20);
assert(isABI && "Non-ABI compilation is not supported");
if (!isABI)
return Chain;
Info.vol = 0;
Info.readMem = true;
Info.writeMem = false;
-
- // alignment is available as metadata.
- // Grab it and set the alignment.
- assert(I.hasMetadataOtherThanDebugLoc() && "Must have alignment metadata");
- MDNode *AlignMD = I.getMetadata("align");
- assert(AlignMD && "Must have a non-null MDNode");
- assert(AlignMD->getNumOperands() == 1 && "Must have a single operand");
- Value *Align = AlignMD->getOperand(0);
- int64_t Alignment = cast<ConstantInt>(Align)->getZExtValue();
- Info.align = Alignment;
+ Info.align = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue();
return true;
}
Info.vol = 0;
Info.readMem = true;
Info.writeMem = false;
-
- // alignment is available as metadata.
- // Grab it and set the alignment.
- assert(I.hasMetadataOtherThanDebugLoc() && "Must have alignment metadata");
- MDNode *AlignMD = I.getMetadata("align");
- assert(AlignMD && "Must have a non-null MDNode");
- assert(AlignMD->getNumOperands() == 1 && "Must have a single operand");
- Value *Align = AlignMD->getOperand(0);
- int64_t Alignment = cast<ConstantInt>(Align)->getZExtValue();
- Info.align = Alignment;
+ Info.align = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue();
return true;
}
}
std::pair<unsigned, const TargetRegisterClass *>
-NVPTXTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
+NVPTXTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
+ const std::string &Constraint,
MVT VT) const {
if (Constraint.size() == 1) {
switch (Constraint[0]) {
return std::make_pair(0U, &NVPTX::Float64RegsRegClass);
}
}
- return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
+ return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
}
/// getFunctionAlignment - Return the Log2 alignment of this function.
default: break;
case ISD::ADD:
case ISD::FADD:
- return PerformADDCombine(N, DCI, nvptxSubtarget, OptLevel);
+ return PerformADDCombine(N, DCI, STI, OptLevel);
case ISD::MUL:
return PerformMULCombine(N, DCI, OptLevel);
case ISD::SHL:
}
}
- SmallVector<SDValue, 8> OtherOps;
-
// Copy regular operands
- for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
- OtherOps.push_back(N->getOperand(i));
+ SmallVector<SDValue, 8> OtherOps(N->op_begin(), N->op_end());
// The select routine does not have access to the LoadSDNode instance, so
// pass along the extension information
OtherOps.push_back(Chain); // Chain
// Skip operand 1 (intrinsic ID)
// Others
- for (unsigned i = 2, e = N->getNumOperands(); i != e; ++i)
- OtherOps.push_back(N->getOperand(i));
+ OtherOps.append(N->op_begin() + 2, N->op_end());
MemIntrinsicSDNode *MemSD = cast<MemIntrinsicSDNode>(N);
"Custom handling of non-i8 ldu/ldg?");
// Just copy all operands as-is
- SmallVector<SDValue, 4> Ops;
- for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
- Ops.push_back(N->getOperand(i));
+ SmallVector<SDValue, 4> Ops(N->op_begin(), N->op_end());
// Force output to i16
SDVTList LdResVTs = DAG.getVTList(MVT::i16, MVT::Other);
delete DwarfLocSection;
delete DwarfARangesSection;
delete DwarfRangesSection;
- delete DwarfMacroInfoSection;
+}
+
+const MCSection *
+NVPTXTargetObjectFile::SelectSectionForGlobal(const GlobalValue *GV,
+ SectionKind Kind, Mangler &Mang,
+ const TargetMachine &TM) const {
+ return getDataSection();
}