SDValue
AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const
{
SDValue Data = Op.getOperand(0);
VTSDNode *BaseType = cast<VTSDNode>(Op.getOperand(1));
DebugLoc DL = Op.getDebugLoc();
EVT DVT = Data.getValueType();
EVT BVT = BaseType->getVT();
unsigned baseBits = BVT.getScalarType().getSizeInBits();
unsigned srcBits = DVT.isSimple() ? DVT.getScalarType().getSizeInBits() : 1;
unsigned shiftBits = srcBits - baseBits;
if (srcBits < 32) {
// If the op is less than 32 bits, then it needs to extend to 32bits
// so it can properly keep the upper bits valid.
EVT IVT = genIntType(32, DVT.isVector() ? DVT.getVectorNumElements() : 1);
Data = DAG.getNode(ISD::ZERO_EXTEND, DL, IVT, Data);
shiftBits = 32 - baseBits;
DVT = IVT;
}
SDValue Shift = DAG.getConstant(shiftBits, DVT);
// Shift left by 'Shift' bits.
Data = DAG.getNode(ISD::SHL, DL, DVT, Data, Shift);
// Signed shift Right by 'Shift' bits.
Data = DAG.getNode(ISD::SRA, DL, DVT, Data, Shift);
if (srcBits < 32) {
// Once the sign extension is done, the op needs to be converted to
// its original type.
Data = DAG.getSExtOrTrunc(Data, DL, Op.getOperand(0).getValueType());
}
return Data;
}
SDValue DAGTypeLegalizer::ExpandOp_BUILD_VECTOR(SDNode *N) {
// The vector type is legal but the element type needs expansion.
EVT VecVT = N->getValueType(0);
unsigned NumElts = VecVT.getVectorNumElements();
EVT OldVT = N->getOperand(0).getValueType();
EVT NewVT = TLI.getTypeToTransformTo(*DAG.getContext(), OldVT);
SDLoc dl(N);
assert(OldVT == VecVT.getVectorElementType() &&
"BUILD_VECTOR operand type doesn't match vector element type!");
// Build a vector of twice the length out of the expanded elements.
// For example <3 x i64> -> <6 x i32>.
std::vector<SDValue> NewElts;
NewElts.reserve(NumElts*2);
for (unsigned i = 0; i < NumElts; ++i) {
SDValue Lo, Hi;
GetExpandedOp(N->getOperand(i), Lo, Hi);
if (TLI.isBigEndian())
std::swap(Lo, Hi);
NewElts.push_back(Lo);
NewElts.push_back(Hi);
}
SDValue NewVec = DAG.getNode(ISD::BUILD_VECTOR, dl,
EVT::getVectorVT(*DAG.getContext(),
NewVT, NewElts.size()),
&NewElts[0], NewElts.size());
// Convert the new vector to the old vector type.
return DAG.getNode(ISD::BITCAST, dl, VecVT, NewVec);
}
SDValue DAGTypeLegalizer::ExpandOp_INSERT_VECTOR_ELT(SDNode *N) {
// The vector type is legal but the element type needs expansion.
EVT VecVT = N->getValueType(0);
unsigned NumElts = VecVT.getVectorNumElements();
SDLoc dl(N);
SDValue Val = N->getOperand(1);
EVT OldEVT = Val.getValueType();
EVT NewEVT = TLI.getTypeToTransformTo(*DAG.getContext(), OldEVT);
assert(OldEVT == VecVT.getVectorElementType() &&
"Inserted element type doesn't match vector element type!");
// Bitconvert to a vector of twice the length with elements of the expanded
// type, insert the expanded vector elements, and then convert back.
EVT NewVecVT = EVT::getVectorVT(*DAG.getContext(), NewEVT, NumElts*2);
SDValue NewVec = DAG.getNode(ISD::BITCAST, dl,
NewVecVT, N->getOperand(0));
SDValue Lo, Hi;
GetExpandedOp(Val, Lo, Hi);
if (TLI.isBigEndian())
std::swap(Lo, Hi);
SDValue Idx = N->getOperand(2);
Idx = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx, Idx);
NewVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, NewVecVT, NewVec, Lo, Idx);
Idx = DAG.getNode(ISD::ADD, dl,
Idx.getValueType(), Idx,
DAG.getConstant(1, Idx.getValueType()));
NewVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, NewVecVT, NewVec, Hi, Idx);
// Convert the new vector to the old vector type.
return DAG.getNode(ISD::BITCAST, dl, VecVT, NewVec);
}
void DecodeVPERM2X128Mask(EVT VT, unsigned Imm,
SmallVectorImpl<unsigned> &ShuffleMask) {
unsigned HalfSize = VT.getVectorNumElements()/2;
unsigned FstHalfBegin = (Imm & 0x3) * HalfSize;
unsigned SndHalfBegin = ((Imm >> 4) & 0x3) * HalfSize;
for (int i = FstHalfBegin, e = FstHalfBegin+HalfSize; i != e; ++i)
ShuffleMask.push_back(i);
for (int i = SndHalfBegin, e = SndHalfBegin+HalfSize; i != e; ++i)
ShuffleMask.push_back(i);
}
SDValue DAGTypeLegalizer::ExpandOp_SCALAR_TO_VECTOR(SDNode *N) {
SDLoc dl(N);
EVT VT = N->getValueType(0);
assert(VT.getVectorElementType() == N->getOperand(0).getValueType() &&
"SCALAR_TO_VECTOR operand type doesn't match vector element type!");
unsigned NumElts = VT.getVectorNumElements();
SmallVector<SDValue, 16> Ops(NumElts);
Ops[0] = N->getOperand(0);
SDValue UndefVal = DAG.getUNDEF(Ops[0].getValueType());
for (unsigned i = 1; i < NumElts; ++i)
Ops[i] = UndefVal;
return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &Ops[0], NumElts);
}
static void EmitTypeGenerate(raw_ostream &OS, const Record *ArgType,
unsigned &ArgNo) {
MVT::SimpleValueType VT = getValueType(ArgType->getValueAsDef("VT"));
if (ArgType->isSubClassOf("LLVMMatchType")) {
unsigned Number = ArgType->getValueAsInt("Number");
assert(Number < ArgNo && "Invalid matching number!");
if (ArgType->isSubClassOf("LLVMExtendedElementVectorType"))
OS << "VectorType::getExtendedElementVectorType"
<< "(dyn_cast<VectorType>(Tys[" << Number << "]))";
else if (ArgType->isSubClassOf("LLVMTruncatedElementVectorType"))
OS << "VectorType::getTruncatedElementVectorType"
<< "(dyn_cast<VectorType>(Tys[" << Number << "]))";
else
OS << "Tys[" << Number << "]";
} else if (VT == MVT::iAny || VT == MVT::fAny || VT == MVT::vAny) {
// NOTE: The ArgNo variable here is not the absolute argument number, it is
// the index of the "arbitrary" type in the Tys array passed to the
// Intrinsic::getDeclaration function. Consequently, we only want to
// increment it when we actually hit an overloaded type. Getting this wrong
// leads to very subtle bugs!
OS << "Tys[" << ArgNo++ << "]";
} else if (EVT(VT).isVector()) {
EVT VVT = VT;
OS << "VectorType::get(";
EmitTypeForValueType(OS, VVT.getVectorElementType().getSimpleVT().SimpleTy);
OS << ", " << VVT.getVectorNumElements() << ")";
} else if (VT == MVT::iPTR) {
OS << "PointerType::getUnqual(";
EmitTypeGenerate(OS, ArgType->getValueAsDef("ElTy"), ArgNo);
OS << ")";
} else if (VT == MVT::iPTRAny) {
// Make sure the user has passed us an argument type to overload. If not,
// treat it as an ordinary (not overloaded) intrinsic.
OS << "(" << ArgNo << " < numTys) ? Tys[" << ArgNo
<< "] : PointerType::getUnqual(";
EmitTypeGenerate(OS, ArgType->getValueAsDef("ElTy"), ArgNo);
OS << ")";
++ArgNo;
} else if (VT == MVT::isVoid) {
if (ArgNo == 0)
OS << "Type::getVoidTy(Context)";
else
// MVT::isVoid is used to mean varargs here.
OS << "...";
} else {
EmitTypeForValueType(OS, VT);
}
}
/// DecodePSHUFMask - This decodes the shuffle masks for pshufd, and vpermilp*.
/// VT indicates the type of the vector allowing it to handle different
/// datatypes and vector widths.
void DecodePSHUFMask(EVT VT, unsigned Imm,
SmallVectorImpl<unsigned> &ShuffleMask) {
unsigned NumElts = VT.getVectorNumElements();
unsigned NumLanes = VT.getSizeInBits() / 128;
unsigned NumLaneElts = NumElts / NumLanes;
int NewImm = Imm;
for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
for (unsigned i = 0; i != NumLaneElts; ++i) {
ShuffleMask.push_back(NewImm % NumLaneElts + l);
NewImm /= NumLaneElts;
}
if (NumLaneElts == 4) NewImm = Imm; // reload imm
}
}
/// DecodeUNPCKLMask - This decodes the shuffle masks for unpcklps/unpcklpd
/// and punpckl*. VT indicates the type of the vector allowing it to handle
/// different datatypes and vector widths.
void DecodeUNPCKLMask(EVT VT, SmallVectorImpl<unsigned> &ShuffleMask) {
unsigned NumElts = VT.getVectorNumElements();
// Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
// independently on 128-bit lanes.
unsigned NumLanes = VT.getSizeInBits() / 128;
if (NumLanes == 0 ) NumLanes = 1; // Handle MMX
unsigned NumLaneElts = NumElts / NumLanes;
for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
for (unsigned i = l, e = l + NumLaneElts/2; i != e; ++i) {
ShuffleMask.push_back(i); // Reads from dest/src1
ShuffleMask.push_back(i+NumElts); // Reads from src/src2
}
}
}
// By default CONCAT_VECTORS is lowered by ExpandVectorBuildThroughStack()
// (see LegalizeDAG.cpp). This is slow and uses local memory.
// We use extract/insert/build vector just as what LegalizeOp() does in llvm 2.5
SDValue NVPTXTargetLowering::
LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
SDNode *Node = Op.getNode();
DebugLoc dl = Node->getDebugLoc();
SmallVector<SDValue, 8> Ops;
unsigned NumOperands = Node->getNumOperands();
for (unsigned i=0; i < NumOperands; ++i) {
SDValue SubOp = Node->getOperand(i);
EVT VVT = SubOp.getNode()->getValueType(0);
EVT EltVT = VVT.getVectorElementType();
unsigned NumSubElem = VVT.getVectorNumElements();
for (unsigned j=0; j < NumSubElem; ++j) {
Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, SubOp,
DAG.getIntPtrConstant(j)));
}
}
return DAG.getNode(ISD::BUILD_VECTOR, dl, Node->getValueType(0),
&Ops[0], Ops.size());
}
SDValue
NVPTXTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
bool isVarArg,
const SmallVectorImpl<ISD::OutputArg> &Outs,
const SmallVectorImpl<SDValue> &OutVals,
DebugLoc dl, SelectionDAG &DAG) const {
bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
unsigned sizesofar = 0;
unsigned idx = 0;
for (unsigned i=0, e=Outs.size(); i!=e; ++i) {
SDValue theVal = OutVals[i];
EVT theValType = theVal.getValueType();
unsigned numElems = 1;
if (theValType.isVector()) numElems = theValType.getVectorNumElements();
for (unsigned j=0,je=numElems; j!=je; ++j) {
SDValue tmpval = theVal;
if (theValType.isVector())
tmpval = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
theValType.getVectorElementType(),
tmpval, DAG.getIntPtrConstant(j));
Chain = DAG.getNode(isABI ? NVPTXISD::StoreRetval :NVPTXISD::MoveToRetval,
dl, MVT::Other,
Chain,
DAG.getConstant(isABI ? sizesofar : idx, MVT::i32),
tmpval);
if (theValType.isVector())
sizesofar += theValType.getVectorElementType().getStoreSizeInBits()/8;
else
sizesofar += theValType.getStoreSizeInBits()/8;
++idx;
}
}
return DAG.getNode(NVPTXISD::RET_FLAG, dl, MVT::Other, Chain);
}
void DAGTypeLegalizer::ExpandRes_BITCAST(SDNode *N, SDValue &Lo, SDValue &Hi) {
EVT OutVT = N->getValueType(0);
EVT NOutVT = TLI.getTypeToTransformTo(*DAG.getContext(), OutVT);
SDValue InOp = N->getOperand(0);
EVT InVT = InOp.getValueType();
SDLoc dl(N);
// Handle some special cases efficiently.
switch (getTypeAction(InVT)) {
case TargetLowering::TypeLegal:
case TargetLowering::TypePromoteInteger:
break;
case TargetLowering::TypePromoteFloat:
llvm_unreachable("Bitcast of a promotion-needing float should never need"
"expansion");
case TargetLowering::TypeSoftenFloat:
// Convert the integer operand instead.
SplitInteger(GetSoftenedFloat(InOp), Lo, Hi);
Lo = DAG.getNode(ISD::BITCAST, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BITCAST, dl, NOutVT, Hi);
return;
case TargetLowering::TypeExpandInteger:
case TargetLowering::TypeExpandFloat:
// Convert the expanded pieces of the input.
GetExpandedOp(InOp, Lo, Hi);
if (TLI.hasBigEndianPartOrdering(InVT) !=
TLI.hasBigEndianPartOrdering(OutVT))
std::swap(Lo, Hi);
Lo = DAG.getNode(ISD::BITCAST, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BITCAST, dl, NOutVT, Hi);
return;
case TargetLowering::TypeSplitVector:
GetSplitVector(InOp, Lo, Hi);
if (TLI.hasBigEndianPartOrdering(OutVT))
std::swap(Lo, Hi);
Lo = DAG.getNode(ISD::BITCAST, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BITCAST, dl, NOutVT, Hi);
return;
case TargetLowering::TypeScalarizeVector:
// Convert the element instead.
SplitInteger(BitConvertToInteger(GetScalarizedVector(InOp)), Lo, Hi);
Lo = DAG.getNode(ISD::BITCAST, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BITCAST, dl, NOutVT, Hi);
return;
case TargetLowering::TypeWidenVector: {
assert(!(InVT.getVectorNumElements() & 1) && "Unsupported BITCAST");
InOp = GetWidenedVector(InOp);
EVT LoVT, HiVT;
std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(InVT);
std::tie(Lo, Hi) = DAG.SplitVector(InOp, dl, LoVT, HiVT);
if (TLI.hasBigEndianPartOrdering(OutVT))
std::swap(Lo, Hi);
Lo = DAG.getNode(ISD::BITCAST, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BITCAST, dl, NOutVT, Hi);
return;
}
}
if (InVT.isVector() && OutVT.isInteger()) {
// Handle cases like i64 = BITCAST v1i64 on x86, where the operand
// is legal but the result is not.
unsigned NumElems = 2;
EVT ElemVT = NOutVT;
EVT NVT = EVT::getVectorVT(*DAG.getContext(), ElemVT, NumElems);
// If <ElemVT * N> is not a legal type, try <ElemVT/2 * (N*2)>.
while (!isTypeLegal(NVT)) {
unsigned NewSizeInBits = ElemVT.getSizeInBits() / 2;
// If the element size is smaller than byte, bail.
if (NewSizeInBits < 8)
break;
NumElems *= 2;
ElemVT = EVT::getIntegerVT(*DAG.getContext(), NewSizeInBits);
NVT = EVT::getVectorVT(*DAG.getContext(), ElemVT, NumElems);
}
if (isTypeLegal(NVT)) {
SDValue CastInOp = DAG.getNode(ISD::BITCAST, dl, NVT, InOp);
SmallVector<SDValue, 8> Vals;
for (unsigned i = 0; i < NumElems; ++i)
Vals.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ElemVT,
CastInOp, DAG.getConstant(i, dl,
TLI.getVectorIdxTy())));
// Build Lo, Hi pair by pairing extracted elements if needed.
unsigned Slot = 0;
for (unsigned e = Vals.size(); e - Slot > 2; Slot += 2, e += 1) {
// Each iteration will BUILD_PAIR two nodes and append the result until
// there are only two nodes left, i.e. Lo and Hi.
SDValue LHS = Vals[Slot];
SDValue RHS = Vals[Slot + 1];
if (TLI.isBigEndian())
std::swap(LHS, RHS);
Vals.push_back(DAG.getNode(ISD::BUILD_PAIR, dl,
EVT::getIntegerVT(
*DAG.getContext(),
LHS.getValueType().getSizeInBits() << 1),
LHS, RHS));
}
Lo = Vals[Slot++];
Hi = Vals[Slot++];
if (TLI.isBigEndian())
std::swap(Lo, Hi);
return;
}
}
// Lower the bit-convert to a store/load from the stack.
assert(NOutVT.isByteSized() && "Expanded type not byte sized!");
// Create the stack frame object. Make sure it is aligned for both
// the source and expanded destination types.
unsigned Alignment =
TLI.getDataLayout()->getPrefTypeAlignment(NOutVT.
getTypeForEVT(*DAG.getContext()));
SDValue StackPtr = DAG.CreateStackTemporary(InVT, Alignment);
int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(SPFI);
// Emit a store to the stack slot.
SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, InOp, StackPtr, PtrInfo,
false, false, 0);
// Load the first half from the stack slot.
Lo = DAG.getLoad(NOutVT, dl, Store, StackPtr, PtrInfo,
false, false, false, 0);
// Increment the pointer to the other half.
unsigned IncrementSize = NOutVT.getSizeInBits() / 8;
StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
DAG.getConstant(IncrementSize, dl,
StackPtr.getValueType()));
// Load the second half from the stack slot.
Hi = DAG.getLoad(NOutVT, dl, Store, StackPtr,
PtrInfo.getWithOffset(IncrementSize), false,
false, false, MinAlign(Alignment, IncrementSize));
// Handle endianness of the load.
if (TLI.hasBigEndianPartOrdering(OutVT))
std::swap(Lo, Hi);
}
SDValue
AMDGPUTargetLowering::LowerSDIV24(SDValue Op, SelectionDAG &DAG) const
{
DebugLoc DL = Op.getDebugLoc();
EVT OVT = Op.getValueType();
SDValue LHS = Op.getOperand(0);
SDValue RHS = Op.getOperand(1);
MVT INTTY;
MVT FLTTY;
if (!OVT.isVector()) {
INTTY = MVT::i32;
FLTTY = MVT::f32;
} else if (OVT.getVectorNumElements() == 2) {
INTTY = MVT::v2i32;
FLTTY = MVT::v2f32;
} else if (OVT.getVectorNumElements() == 4) {
INTTY = MVT::v4i32;
FLTTY = MVT::v4f32;
}
unsigned bitsize = OVT.getScalarType().getSizeInBits();
// char|short jq = ia ^ ib;
SDValue jq = DAG.getNode(ISD::XOR, DL, OVT, LHS, RHS);
// jq = jq >> (bitsize - 2)
jq = DAG.getNode(ISD::SRA, DL, OVT, jq, DAG.getConstant(bitsize - 2, OVT));
// jq = jq | 0x1
jq = DAG.getNode(ISD::OR, DL, OVT, jq, DAG.getConstant(1, OVT));
// jq = (int)jq
jq = DAG.getSExtOrTrunc(jq, DL, INTTY);
// int ia = (int)LHS;
SDValue ia = DAG.getSExtOrTrunc(LHS, DL, INTTY);
// int ib, (int)RHS;
SDValue ib = DAG.getSExtOrTrunc(RHS, DL, INTTY);
// float fa = (float)ia;
SDValue fa = DAG.getNode(ISD::SINT_TO_FP, DL, FLTTY, ia);
// float fb = (float)ib;
SDValue fb = DAG.getNode(ISD::SINT_TO_FP, DL, FLTTY, ib);
// float fq = native_divide(fa, fb);
SDValue fq = DAG.getNode(AMDGPUISD::DIV_INF, DL, FLTTY, fa, fb);
// fq = trunc(fq);
fq = DAG.getNode(ISD::FTRUNC, DL, FLTTY, fq);
// float fqneg = -fq;
SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FLTTY, fq);
// float fr = mad(fqneg, fb, fa);
SDValue fr = DAG.getNode(AMDGPUISD::MAD, DL, FLTTY, fqneg, fb, fa);
// int iq = (int)fq;
SDValue iq = DAG.getNode(ISD::FP_TO_SINT, DL, INTTY, fq);
// fr = fabs(fr);
fr = DAG.getNode(ISD::FABS, DL, FLTTY, fr);
// fb = fabs(fb);
fb = DAG.getNode(ISD::FABS, DL, FLTTY, fb);
// int cv = fr >= fb;
SDValue cv;
if (INTTY == MVT::i32) {
cv = DAG.getSetCC(DL, INTTY, fr, fb, ISD::SETOGE);
} else {
cv = DAG.getSetCC(DL, INTTY, fr, fb, ISD::SETOGE);
}
// jq = (cv ? jq : 0);
jq = DAG.getNode(ISD::SELECT, DL, OVT, cv, jq,
DAG.getConstant(0, OVT));
// dst = iq + jq;
iq = DAG.getSExtOrTrunc(iq, DL, OVT);
iq = DAG.getNode(ISD::ADD, DL, OVT, iq, jq);
return iq;
}
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: {
unsigned RegClassID;
const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
const AMDGPURegisterInfo *TRI =
static_cast<const AMDGPURegisterInfo*>(TM.getRegisterInfo());
const SIRegisterInfo *SIRI =
static_cast<const SIRegisterInfo*>(TM.getRegisterInfo());
EVT VT = N->getValueType(0);
unsigned NumVectorElts = VT.getVectorNumElements();
assert(VT.getVectorElementType().bitsEq(MVT::i32));
if (ST.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
bool UseVReg = true;
for (SDNode::use_iterator U = N->use_begin(), E = SDNode::use_end();
U != E; ++U) {
if (!U->isMachineOpcode()) {
continue;
}
const TargetRegisterClass *RC = getOperandRegClass(*U, U.getOperandNo());
if (!RC) {
continue;
}
if (SIRI->isSGPRClass(RC)) {
UseVReg = false;
}
}
switch(NumVectorElts) {
case 1: RegClassID = UseVReg ? AMDGPU::VReg_32RegClassID :
AMDGPU::SReg_32RegClassID;
break;
case 2: RegClassID = UseVReg ? AMDGPU::VReg_64RegClassID :
AMDGPU::SReg_64RegClassID;
break;
case 4: RegClassID = UseVReg ? AMDGPU::VReg_128RegClassID :
AMDGPU::SReg_128RegClassID;
break;
case 8: RegClassID = UseVReg ? AMDGPU::VReg_256RegClassID :
AMDGPU::SReg_256RegClassID;
break;
case 16: RegClassID = UseVReg ? AMDGPU::VReg_512RegClassID :
AMDGPU::SReg_512RegClassID;
break;
default: llvm_unreachable("Do not know how to lower this BUILD_VECTOR");
}
} else {
// BUILD_VECTOR was 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.
switch(NumVectorElts) {
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");
}
}
SDValue RegClass = CurDAG->getTargetConstant(RegClassID, MVT::i32);
if (NumVectorElts == 1) {
return CurDAG->SelectNodeTo(N, AMDGPU::COPY_TO_REGCLASS,
VT.getVectorElementType(),
N->getOperand(0), RegClass);
}
assert(NumVectorElts <= 16 && "Vectors with more than 16 elements not "
"supported yet");
// 16 = Max Num Vector Elements
// 2 = 2 REG_SEQUENCE operands per element (value, subreg index)
// 1 = Vector Register Class
SDValue RegSeqArgs[16 * 2 + 1];
RegSeqArgs[0] = CurDAG->getTargetConstant(RegClassID, MVT::i32);
bool IsRegSeq = true;
for (unsigned i = 0; i < N->getNumOperands(); i++) {
// XXX: Why is this here?
if (dyn_cast<RegisterSDNode>(N->getOperand(i))) {
IsRegSeq = false;
break;
}
RegSeqArgs[1 + (2 * i)] = N->getOperand(i);
RegSeqArgs[1 + (2 * i) + 1] =
CurDAG->getTargetConstant(TRI->getSubRegFromChannel(i), MVT::i32);
}
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);
}
}
return SelectCode(N);
}
void DAGTypeLegalizer::ExpandRes_BIT_CONVERT(SDNode *N, SDValue &Lo,
SDValue &Hi) {
EVT OutVT = N->getValueType(0);
EVT NOutVT = TLI.getTypeToTransformTo(*DAG.getContext(), OutVT);
SDValue InOp = N->getOperand(0);
EVT InVT = InOp.getValueType();
DebugLoc dl = N->getDebugLoc();
// Handle some special cases efficiently.
switch (getTypeAction(InVT)) {
default:
assert(false && "Unknown type action!");
case Legal:
case PromoteInteger:
break;
case SoftenFloat:
// Convert the integer operand instead.
SplitInteger(GetSoftenedFloat(InOp), Lo, Hi);
Lo = DAG.getNode(ISD::BIT_CONVERT, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BIT_CONVERT, dl, NOutVT, Hi);
return;
case ExpandInteger:
case ExpandFloat:
// Convert the expanded pieces of the input.
GetExpandedOp(InOp, Lo, Hi);
Lo = DAG.getNode(ISD::BIT_CONVERT, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BIT_CONVERT, dl, NOutVT, Hi);
return;
case SplitVector:
GetSplitVector(InOp, Lo, Hi);
if (TLI.isBigEndian())
std::swap(Lo, Hi);
Lo = DAG.getNode(ISD::BIT_CONVERT, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BIT_CONVERT, dl, NOutVT, Hi);
return;
case ScalarizeVector:
// Convert the element instead.
SplitInteger(BitConvertToInteger(GetScalarizedVector(InOp)), Lo, Hi);
Lo = DAG.getNode(ISD::BIT_CONVERT, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BIT_CONVERT, dl, NOutVT, Hi);
return;
case WidenVector: {
assert(!(InVT.getVectorNumElements() & 1) && "Unsupported BIT_CONVERT");
InOp = GetWidenedVector(InOp);
EVT InNVT = EVT::getVectorVT(*DAG.getContext(), InVT.getVectorElementType(),
InVT.getVectorNumElements()/2);
Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InNVT, InOp,
DAG.getIntPtrConstant(0));
Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InNVT, InOp,
DAG.getIntPtrConstant(InNVT.getVectorNumElements()));
if (TLI.isBigEndian())
std::swap(Lo, Hi);
Lo = DAG.getNode(ISD::BIT_CONVERT, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BIT_CONVERT, dl, NOutVT, Hi);
return;
}
}
if (InVT.isVector() && OutVT.isInteger()) {
// Handle cases like i64 = BIT_CONVERT v1i64 on x86, where the operand
// is legal but the result is not.
EVT NVT = EVT::getVectorVT(*DAG.getContext(), NOutVT, 2);
if (isTypeLegal(NVT)) {
SDValue CastInOp = DAG.getNode(ISD::BIT_CONVERT, dl, NVT, InOp);
Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, NOutVT, CastInOp,
DAG.getIntPtrConstant(0));
Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, NOutVT, CastInOp,
DAG.getIntPtrConstant(1));
if (TLI.isBigEndian())
std::swap(Lo, Hi);
return;
}
}
// Lower the bit-convert to a store/load from the stack.
assert(NOutVT.isByteSized() && "Expanded type not byte sized!");
// Create the stack frame object. Make sure it is aligned for both
// the source and expanded destination types.
unsigned Alignment =
TLI.getTargetData()->getPrefTypeAlignment(NOutVT.getTypeForEVT(*DAG.getContext()));
SDValue StackPtr = DAG.CreateStackTemporary(InVT, Alignment);
int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
const Value *SV = PseudoSourceValue::getFixedStack(SPFI);
// Emit a store to the stack slot.
SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, InOp, StackPtr, SV, 0);
// Load the first half from the stack slot.
Lo = DAG.getLoad(NOutVT, dl, Store, StackPtr, SV, 0);
// Increment the pointer to the other half.
unsigned IncrementSize = NOutVT.getSizeInBits() / 8;
StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
DAG.getIntPtrConstant(IncrementSize));
// Load the second half from the stack slot.
Hi = DAG.getLoad(NOutVT, dl, Store, StackPtr, SV, IncrementSize, false,
MinAlign(Alignment, IncrementSize));
// Handle endianness of the load.
if (TLI.isBigEndian())
std::swap(Lo, Hi);
}
/// getVectorTypeBreakdown - Vector types are broken down into some number of
/// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
/// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
/// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
///
/// This method returns the number of registers needed, and the VT for each
/// register. It also returns the VT and quantity of the intermediate values
/// before they are promoted/expanded.
///
unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
EVT &IntermediateVT,
unsigned &NumIntermediates,
MVT &RegisterVT) const {
unsigned NumElts = VT.getVectorNumElements();
// If there is a wider vector type with the same element type as this one,
// or a promoted vector type that has the same number of elements which
// are wider, then we should convert to that legal vector type.
// This handles things like <2 x float> -> <4 x float> and
// <4 x i1> -> <4 x i32>.
LegalizeTypeAction TA = getTypeAction(Context, VT);
if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) {
EVT RegisterEVT = getTypeToTransformTo(Context, VT);
if (isTypeLegal(RegisterEVT)) {
IntermediateVT = RegisterEVT;
RegisterVT = RegisterEVT.getSimpleVT();
NumIntermediates = 1;
return 1;
}
}
// Figure out the right, legal destination reg to copy into.
EVT EltTy = VT.getVectorElementType();
unsigned NumVectorRegs = 1;
// FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
// could break down into LHS/RHS like LegalizeDAG does.
if (!isPowerOf2_32(NumElts)) {
NumVectorRegs = NumElts;
NumElts = 1;
}
// Divide the input until we get to a supported size. This will always
// end with a scalar if the target doesn't support vectors.
while (NumElts > 1 && !isTypeLegal(
EVT::getVectorVT(Context, EltTy, NumElts))) {
NumElts >>= 1;
NumVectorRegs <<= 1;
}
NumIntermediates = NumVectorRegs;
EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts);
if (!isTypeLegal(NewVT))
NewVT = EltTy;
IntermediateVT = NewVT;
MVT DestVT = getRegisterType(Context, NewVT);
RegisterVT = DestVT;
unsigned NewVTSize = NewVT.getSizeInBits();
// Convert sizes such as i33 to i64.
if (!isPowerOf2_32(NewVTSize))
NewVTSize = NextPowerOf2(NewVTSize);
if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
// Otherwise, promotion or legal types use the same number of registers as
// the vector decimated to the appropriate level.
return NumVectorRegs;
}
SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) {
unsigned int Opc = N->getOpcode();
if (N->isMachineOpcode()) {
N->setNodeId(-1);
return NULL; // Already selected.
}
switch (Opc) {
default: break;
// We are selecting i64 ADD here instead of custom lower it during
// DAG legalization, so we can fold some i64 ADDs used for address
// calculation into the LOAD and STORE instructions.
case ISD::ADD: {
const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
if (N->getValueType(0) != MVT::i64 ||
ST.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS)
break;
SDLoc DL(N);
SDValue LHS = N->getOperand(0);
SDValue RHS = N->getOperand(1);
SDValue Sub0 = CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32);
SDValue Sub1 = CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32);
SDNode *Lo0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
DL, MVT::i32, LHS, Sub0);
SDNode *Hi0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
DL, MVT::i32, LHS, Sub1);
SDNode *Lo1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
DL, MVT::i32, RHS, Sub0);
SDNode *Hi1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
DL, MVT::i32, RHS, Sub1);
SDVTList VTList = CurDAG->getVTList(MVT::i32, MVT::Glue);
SmallVector<SDValue, 8> AddLoArgs;
AddLoArgs.push_back(SDValue(Lo0, 0));
AddLoArgs.push_back(SDValue(Lo1, 0));
SDNode *AddLo = CurDAG->getMachineNode(AMDGPU::S_ADD_I32, DL,
VTList, AddLoArgs);
SDValue Carry = SDValue(AddLo, 1);
SDNode *AddHi = CurDAG->getMachineNode(AMDGPU::S_ADDC_U32, DL,
MVT::i32, SDValue(Hi0, 0),
SDValue(Hi1, 0), Carry);
SDValue Args[5] = {
CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, MVT::i32),
SDValue(AddLo,0),
Sub0,
SDValue(AddHi,0),
Sub1,
};
return CurDAG->SelectNodeTo(N, AMDGPU::REG_SEQUENCE, MVT::i64, Args, 5);
}
case ISD::BUILD_VECTOR: {
unsigned RegClassID;
const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
const AMDGPURegisterInfo *TRI =
static_cast<const AMDGPURegisterInfo*>(TM.getRegisterInfo());
const SIRegisterInfo *SIRI =
static_cast<const SIRegisterInfo*>(TM.getRegisterInfo());
EVT VT = N->getValueType(0);
unsigned NumVectorElts = VT.getVectorNumElements();
assert(VT.getVectorElementType().bitsEq(MVT::i32));
if (ST.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
bool UseVReg = true;
for (SDNode::use_iterator U = N->use_begin(), E = SDNode::use_end();
U != E; ++U) {
if (!U->isMachineOpcode()) {
continue;
}
const TargetRegisterClass *RC = getOperandRegClass(*U, U.getOperandNo());
if (!RC) {
continue;
}
if (SIRI->isSGPRClass(RC)) {
UseVReg = false;
}
}
switch(NumVectorElts) {
case 1: RegClassID = UseVReg ? AMDGPU::VReg_32RegClassID :
AMDGPU::SReg_32RegClassID;
break;
case 2: RegClassID = UseVReg ? AMDGPU::VReg_64RegClassID :
AMDGPU::SReg_64RegClassID;
break;
case 4: RegClassID = UseVReg ? AMDGPU::VReg_128RegClassID :
AMDGPU::SReg_128RegClassID;
break;
case 8: RegClassID = UseVReg ? AMDGPU::VReg_256RegClassID :
AMDGPU::SReg_256RegClassID;
break;
case 16: RegClassID = UseVReg ? AMDGPU::VReg_512RegClassID :
AMDGPU::SReg_512RegClassID;
break;
default: llvm_unreachable("Do not know how to lower this BUILD_VECTOR");
}
} else {
// BUILD_VECTOR was 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.
switch(NumVectorElts) {
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");
}
}
SDValue RegClass = CurDAG->getTargetConstant(RegClassID, MVT::i32);
if (NumVectorElts == 1) {
return CurDAG->SelectNodeTo(N, AMDGPU::COPY_TO_REGCLASS,
VT.getVectorElementType(),
N->getOperand(0), RegClass);
}
assert(NumVectorElts <= 16 && "Vectors with more than 16 elements not "
"supported yet");
// 16 = Max Num Vector Elements
// 2 = 2 REG_SEQUENCE operands per element (value, subreg index)
// 1 = Vector Register Class
SDValue RegSeqArgs[16 * 2 + 1];
RegSeqArgs[0] = CurDAG->getTargetConstant(RegClassID, MVT::i32);
bool IsRegSeq = true;
for (unsigned i = 0; i < N->getNumOperands(); i++) {
// XXX: Why is this here?
if (dyn_cast<RegisterSDNode>(N->getOperand(i))) {
IsRegSeq = false;
break;
}
RegSeqArgs[1 + (2 * i)] = N->getOperand(i);
RegSeqArgs[1 + (2 * i) + 1] =
CurDAG->getTargetConstant(TRI->getSubRegFromChannel(i), MVT::i32);
}
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 AMDGPUISD::REGISTER_LOAD: {
const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
break;
SDValue Addr, Offset;
SelectADDRIndirect(N->getOperand(1), Addr, Offset);
const SDValue Ops[] = {
Addr,
Offset,
CurDAG->getTargetConstant(0, MVT::i32),
N->getOperand(0),
};
return CurDAG->getMachineNode(AMDGPU::SI_RegisterLoad, SDLoc(N),
CurDAG->getVTList(MVT::i32, MVT::i64, MVT::Other),
Ops);
}
case AMDGPUISD::REGISTER_STORE: {
const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
break;
SDValue Addr, Offset;
SelectADDRIndirect(N->getOperand(2), Addr, Offset);
const SDValue Ops[] = {
N->getOperand(1),
Addr,
Offset,
CurDAG->getTargetConstant(0, MVT::i32),
N->getOperand(0),
};
return CurDAG->getMachineNode(AMDGPU::SI_RegisterStorePseudo, SDLoc(N),
CurDAG->getVTList(MVT::Other),
Ops);
}
}
return SelectCode(N);
}
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: {
unsigned RegClassID;
const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
const AMDGPURegisterInfo *TRI =
static_cast<const AMDGPURegisterInfo*>(TM.getRegisterInfo());
const SIRegisterInfo *SIRI =
static_cast<const SIRegisterInfo*>(TM.getRegisterInfo());
EVT VT = N->getValueType(0);
unsigned NumVectorElts = VT.getVectorNumElements();
assert(VT.getVectorElementType().bitsEq(MVT::i32));
if (ST.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
bool UseVReg = true;
for (SDNode::use_iterator U = N->use_begin(), E = SDNode::use_end();
U != E; ++U) {
if (!U->isMachineOpcode()) {
continue;
}
const TargetRegisterClass *RC = getOperandRegClass(*U, U.getOperandNo());
if (!RC) {
continue;
}
if (SIRI->isSGPRClass(RC)) {
UseVReg = false;
}
}
switch(NumVectorElts) {
case 1: RegClassID = UseVReg ? AMDGPU::VReg_32RegClassID :
AMDGPU::SReg_32RegClassID;
break;
case 2: RegClassID = UseVReg ? AMDGPU::VReg_64RegClassID :
AMDGPU::SReg_64RegClassID;
break;
case 4: RegClassID = UseVReg ? AMDGPU::VReg_128RegClassID :
AMDGPU::SReg_128RegClassID;
break;
case 8: RegClassID = UseVReg ? AMDGPU::VReg_256RegClassID :
AMDGPU::SReg_256RegClassID;
break;
case 16: RegClassID = UseVReg ? AMDGPU::VReg_512RegClassID :
AMDGPU::SReg_512RegClassID;
break;
default: llvm_unreachable("Do not know how to lower this BUILD_VECTOR");
}
} else {
// BUILD_VECTOR was 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.
switch(NumVectorElts) {
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");
}
}
SDValue RegClass = CurDAG->getTargetConstant(RegClassID, MVT::i32);
if (NumVectorElts == 1) {
return CurDAG->SelectNodeTo(N, AMDGPU::COPY_TO_REGCLASS,
VT.getVectorElementType(),
N->getOperand(0), RegClass);
}
assert(NumVectorElts <= 16 && "Vectors with more than 16 elements not "
"supported yet");
// 16 = Max Num Vector Elements
// 2 = 2 REG_SEQUENCE operands per element (value, subreg index)
// 1 = Vector Register Class
SDValue RegSeqArgs[16 * 2 + 1];
RegSeqArgs[0] = CurDAG->getTargetConstant(RegClassID, MVT::i32);
bool IsRegSeq = true;
for (unsigned i = 0; i < N->getNumOperands(); i++) {
// XXX: Why is this here?
if (dyn_cast<RegisterSDNode>(N->getOperand(i))) {
IsRegSeq = false;
break;
}
RegSeqArgs[1 + (2 * i)] = N->getOperand(i);
RegSeqArgs[1 + (2 * i) + 1] =
CurDAG->getTargetConstant(TRI->getSubRegFromChannel(i), MVT::i32);
}
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 {
switch(Use->getMachineOpcode()) {
case AMDGPU::REG_SEQUENCE: break;
default:
if (!TII->isALUInstr(Use->getMachineOpcode()) ||
(TII->get(Use->getMachineOpcode()).TSFlags &
R600_InstFlag::VECTOR)) {
continue;
}
}
// 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) {
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--;
}
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;
}
