void ScheduleDAGSDNodes::computeLatency(SUnit *SU) {
SDNode *N = SU->getNode();
// TokenFactor operands are considered zero latency, and some schedulers
// (e.g. Top-Down list) may rely on the fact that operand latency is nonzero
// whenever node latency is nonzero.
if (N && N->getOpcode() == ISD::TokenFactor) {
SU->Latency = 0;
return;
}
// Check to see if the scheduler cares about latencies.
if (forceUnitLatencies()) {
SU->Latency = 1;
return;
}
if (!InstrItins || InstrItins->isEmpty()) {
if (N && N->isMachineOpcode() &&
TII->isHighLatencyDef(N->getMachineOpcode()))
SU->Latency = HighLatencyCycles;
else
SU->Latency = 1;
return;
}
// Compute the latency for the node. We use the sum of the latencies for
// all nodes glued together into this SUnit.
SU->Latency = 0;
for (SDNode *N = SU->getNode(); N; N = N->getGluedNode())
if (N->isMachineOpcode())
SU->Latency += TII->getInstrLatency(InstrItins, N);
}
void ResourcePriorityQueue::initNumRegDefsLeft(SUnit *SU) {
unsigned NodeNumDefs = 0;
for (SDNode *N = SU->getNode(); N; N = N->getGluedNode())
if (N->isMachineOpcode()) {
const MCInstrDesc &TID = TII->get(N->getMachineOpcode());
// No register need be allocated for this.
if (N->getMachineOpcode() == TargetOpcode::IMPLICIT_DEF) {
NodeNumDefs = 0;
break;
}
NodeNumDefs = std::min(N->getNumValues(), TID.getNumDefs());
}
else
switch(N->getOpcode()) {
default: break;
case ISD::CopyFromReg:
NodeNumDefs++;
break;
case ISD::INLINEASM:
NodeNumDefs++;
break;
}
SU->NumRegDefsLeft = NodeNumDefs;
}
void ScheduleDAGSDNodes::ComputeLatency(SUnit *SU) {
const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
// Compute the latency for the node. We use the sum of the latencies for
// all nodes flagged together into this SUnit.
SU->Latency = 0;
for (SDNode *N = SU->getNode(); N; N = N->getFlaggedNode())
if (N->isMachineOpcode()) {
SU->Latency += InstrItins.
getStageLatency(TII->get(N->getMachineOpcode()).getSchedClass());
}
}
/// DelayForLiveRegsBottomUp - Returns true if it is necessary to delay
/// scheduling of the given node to satisfy live physical register dependencies.
/// If the specific node is the last one that's available to schedule, do
/// whatever is necessary (i.e. backtracking or cloning) to make it possible.
bool ScheduleDAGFast::DelayForLiveRegsBottomUp(SUnit *SU,
SmallVectorImpl<unsigned> &LRegs){
if (NumLiveRegs == 0)
return false;
SmallSet<unsigned, 4> RegAdded;
// If this node would clobber any "live" register, then it's not ready.
for (SDep &Pred : SU->Preds) {
if (Pred.isAssignedRegDep()) {
CheckForLiveRegDef(Pred.getSUnit(), Pred.getReg(), LiveRegDefs,
RegAdded, LRegs, TRI);
}
}
for (SDNode *Node = SU->getNode(); Node; Node = Node->getGluedNode()) {
if (Node->getOpcode() == ISD::INLINEASM) {
// Inline asm can clobber physical defs.
unsigned NumOps = Node->getNumOperands();
if (Node->getOperand(NumOps-1).getValueType() == MVT::Glue)
--NumOps; // Ignore the glue operand.
for (unsigned i = InlineAsm::Op_FirstOperand; i != NumOps;) {
unsigned Flags =
cast<ConstantSDNode>(Node->getOperand(i))->getZExtValue();
unsigned NumVals = InlineAsm::getNumOperandRegisters(Flags);
++i; // Skip the ID value.
if (InlineAsm::isRegDefKind(Flags) ||
InlineAsm::isRegDefEarlyClobberKind(Flags) ||
InlineAsm::isClobberKind(Flags)) {
// Check for def of register or earlyclobber register.
for (; NumVals; --NumVals, ++i) {
unsigned Reg = cast<RegisterSDNode>(Node->getOperand(i))->getReg();
if (TargetRegisterInfo::isPhysicalRegister(Reg))
CheckForLiveRegDef(SU, Reg, LiveRegDefs, RegAdded, LRegs, TRI);
}
} else
i += NumVals;
}
continue;
}
if (!Node->isMachineOpcode())
continue;
const MCInstrDesc &MCID = TII->get(Node->getMachineOpcode());
if (!MCID.ImplicitDefs)
continue;
for (const MCPhysReg *Reg = MCID.getImplicitDefs(); *Reg; ++Reg) {
CheckForLiveRegDef(SU, *Reg, LiveRegDefs, RegAdded, LRegs, TRI);
}
}
return !LRegs.empty();
}
/// ClusterNodes - Cluster certain nodes which should be scheduled together.
///
void ScheduleDAGSDNodes::ClusterNodes() {
for (SDNode &NI : DAG->allnodes()) {
SDNode *Node = &NI;
if (!Node || !Node->isMachineOpcode())
continue;
unsigned Opc = Node->getMachineOpcode();
const MCInstrDesc &MCID = TII->get(Opc);
if (MCID.mayLoad())
// Cluster loads from "near" addresses into combined SUnits.
ClusterNeighboringLoads(Node);
}
}
SDNode *VDAGToDAGISel::SelectBitSlice(SDNode *N) {
SDNode *Op = N->getOperand(0).getNode();
// Emit the constant bit slice to constant directly if possible.
if (ConstantSDNode *CSD = dyn_cast<ConstantSDNode>(Op))
return SelectConstBitSlice(CSD, N);
assert(Op->getOpcode() != VTMISD::BitSlice
&& (!Op->isMachineOpcode()
|| Op->getMachineOpcode() != VTM::VOpBitSlice)
&& "DAGCombine should handle this!");
return SelectSimpleNode(N, VTM::VOpBitSlice);
}
/// ClusterNodes - Cluster certain nodes which should be scheduled together.
///
void ScheduleDAGSDNodes::ClusterNodes() {
for (SelectionDAG::allnodes_iterator NI = DAG->allnodes_begin(),
E = DAG->allnodes_end(); NI != E; ++NI) {
SDNode *Node = &*NI;
if (!Node || !Node->isMachineOpcode())
continue;
unsigned Opc = Node->getMachineOpcode();
const MCInstrDesc &MCID = TII->get(Opc);
if (MCID.mayLoad())
// Cluster loads from "near" addresses into combined SUnits.
ClusterNeighboringLoads(Node);
}
}
SDNode *SparcDAGToDAGISel::Select(SDValue Op) {
SDNode *N = Op.getNode();
DebugLoc dl = N->getDebugLoc();
if (N->isMachineOpcode())
return NULL; // Already selected.
switch (N->getOpcode()) {
default: break;
case SPISD::GLOBAL_BASE_REG:
return getGlobalBaseReg();
case ISD::SDIV:
case ISD::UDIV: {
// FIXME: should use a custom expander to expose the SRA to the dag.
SDValue DivLHS = N->getOperand(0);
SDValue DivRHS = N->getOperand(1);
// Set the Y register to the high-part.
SDValue TopPart;
if (N->getOpcode() == ISD::SDIV) {
TopPart = SDValue(CurDAG->getMachineNode(SP::SRAri, dl, MVT::i32, DivLHS,
CurDAG->getTargetConstant(31, MVT::i32)), 0);
} else {
TopPart = CurDAG->getRegister(SP::G0, MVT::i32);
}
TopPart = SDValue(CurDAG->getMachineNode(SP::WRYrr, dl, MVT::Flag, TopPart,
CurDAG->getRegister(SP::G0, MVT::i32)), 0);
// FIXME: Handle div by immediate.
unsigned Opcode = N->getOpcode() == ISD::SDIV ? SP::SDIVrr : SP::UDIVrr;
return CurDAG->SelectNodeTo(N, Opcode, MVT::i32, DivLHS, DivRHS,
TopPart);
}
case ISD::MULHU:
case ISD::MULHS: {
// FIXME: Handle mul by immediate.
SDValue MulLHS = N->getOperand(0);
SDValue MulRHS = N->getOperand(1);
unsigned Opcode = N->getOpcode() == ISD::MULHU ? SP::UMULrr : SP::SMULrr;
SDNode *Mul = CurDAG->getMachineNode(Opcode, dl, MVT::i32, MVT::Flag,
MulLHS, MulRHS);
// The high part is in the Y register.
return CurDAG->SelectNodeTo(N, SP::RDY, MVT::i32, SDValue(Mul, 1));
return NULL;
}
}
return SelectCode(Op);
}
void ScheduleDAGLinearize::Schedule() {
LLVM_DEBUG(dbgs() << "********** DAG Linearization **********\n");
SmallVector<SDNode*, 8> Glues;
unsigned DAGSize = 0;
for (SDNode &Node : DAG->allnodes()) {
SDNode *N = &Node;
// Use node id to record degree.
unsigned Degree = N->use_size();
N->setNodeId(Degree);
unsigned NumVals = N->getNumValues();
if (NumVals && N->getValueType(NumVals-1) == MVT::Glue &&
N->hasAnyUseOfValue(NumVals-1)) {
SDNode *User = findGluedUser(N);
if (User) {
Glues.push_back(N);
GluedMap.insert(std::make_pair(N, User));
}
}
if (N->isMachineOpcode() ||
(N->getOpcode() != ISD::EntryToken && !isPassiveNode(N)))
++DAGSize;
}
for (unsigned i = 0, e = Glues.size(); i != e; ++i) {
SDNode *Glue = Glues[i];
SDNode *GUser = GluedMap[Glue];
unsigned Degree = Glue->getNodeId();
unsigned UDegree = GUser->getNodeId();
// Glue user must be scheduled together with the glue operand. So other
// users of the glue operand must be treated as its users.
SDNode *ImmGUser = Glue->getGluedUser();
for (const SDNode *U : Glue->uses())
if (U == ImmGUser)
--Degree;
GUser->setNodeId(UDegree + Degree);
Glue->setNodeId(1);
}
Sequence.reserve(DAGSize);
ScheduleNode(DAG->getRoot().getNode());
}
void ScheduleDAGSDNodes::ComputeLatency(SUnit *SU) {
// Check to see if the scheduler cares about latencies.
if (ForceUnitLatencies()) {
SU->Latency = 1;
return;
}
if (!InstrItins || InstrItins->isEmpty()) {
SU->Latency = 1;
return;
}
// Compute the latency for the node. We use the sum of the latencies for
// all nodes glued together into this SUnit.
SU->Latency = 0;
for (SDNode *N = SU->getNode(); N; N = N->getGluedNode())
if (N->isMachineOpcode())
SU->Latency += TII->getInstrLatency(InstrItins, N);
}
void ScheduleDAGSDNodes::ComputeLatency(SUnit *SU) {
// Check to see if the scheduler cares about latencies.
if (ForceUnitLatencies()) {
SU->Latency = 1;
return;
}
const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
if (InstrItins.isEmpty()) {
SU->Latency = 1;
return;
}
// Compute the latency for the node. We use the sum of the latencies for
// all nodes flagged together into this SUnit.
SU->Latency = 0;
for (SDNode *N = SU->getNode(); N; N = N->getFlaggedNode())
if (N->isMachineOpcode()) {
SU->Latency += InstrItins.
getStageLatency(TII->get(N->getMachineOpcode()).getSchedClass());
}
}
SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) {
const R600InstrInfo *TII =
static_cast<const R600InstrInfo*>(TM.getInstrInfo());
unsigned int Opc = N->getOpcode();
if (N->isMachineOpcode()) {
return NULL; // Already selected.
}
switch (Opc) {
default: break;
case AMDGPUISD::CONST_ADDRESS: {
for (SDNode::use_iterator I = N->use_begin(), Next = llvm::next(I);
I != SDNode::use_end(); I = Next) {
Next = llvm::next(I);
if (!I->isMachineOpcode()) {
continue;
}
unsigned Opcode = I->getMachineOpcode();
bool HasDst = TII->getOperandIdx(Opcode, AMDGPU::OpName::dst) > -1;
int SrcIdx = I.getOperandNo();
int SelIdx;
// Unlike MachineInstrs, SDNodes do not have results in their operand
// list, so we need to increment the SrcIdx, since
// R600InstrInfo::getOperandIdx is based on the MachineInstr indices.
if (HasDst) {
SrcIdx++;
}
SelIdx = TII->getSelIdx(I->getMachineOpcode(), SrcIdx);
if (SelIdx < 0) {
continue;
}
SDValue CstOffset;
if (N->getValueType(0).isVector() ||
!SelectGlobalValueConstantOffset(N->getOperand(0), CstOffset))
continue;
// Gather constants values
int SrcIndices[] = {
TII->getOperandIdx(Opcode, AMDGPU::OpName::src0),
TII->getOperandIdx(Opcode, AMDGPU::OpName::src1),
TII->getOperandIdx(Opcode, AMDGPU::OpName::src2),
TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_X),
TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_Y),
TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_Z),
TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_W),
TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_X),
TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_Y),
TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_Z),
TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_W)
};
std::vector<unsigned> Consts;
for (unsigned i = 0; i < sizeof(SrcIndices) / sizeof(int); i++) {
int OtherSrcIdx = SrcIndices[i];
int OtherSelIdx = TII->getSelIdx(Opcode, OtherSrcIdx);
if (OtherSrcIdx < 0 || OtherSelIdx < 0) {
continue;
}
if (HasDst) {
OtherSrcIdx--;
OtherSelIdx--;
}
if (RegisterSDNode *Reg =
dyn_cast<RegisterSDNode>(I->getOperand(OtherSrcIdx))) {
if (Reg->getReg() == AMDGPU::ALU_CONST) {
ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(I->getOperand(OtherSelIdx));
Consts.push_back(Cst->getZExtValue());
}
}
}
ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(CstOffset);
Consts.push_back(Cst->getZExtValue());
if (!TII->fitsConstReadLimitations(Consts))
continue;
// Convert back to SDNode indices
if (HasDst) {
SrcIdx--;
SelIdx--;
}
std::vector<SDValue> Ops;
for (int i = 0, e = I->getNumOperands(); i != e; ++i) {
if (i == SrcIdx) {
Ops.push_back(CurDAG->getRegister(AMDGPU::ALU_CONST, MVT::f32));
} else if (i == SelIdx) {
Ops.push_back(CstOffset);
} else {
Ops.push_back(I->getOperand(i));
}
}
CurDAG->UpdateNodeOperands(*I, Ops.data(), Ops.size());
}
break;
}
case ISD::BUILD_VECTOR: {
const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
if (ST.getGeneration() > AMDGPUSubtarget::NORTHERN_ISLANDS) {
break;
}
unsigned RegClassID;
switch(N->getValueType(0).getVectorNumElements()) {
case 2: RegClassID = AMDGPU::R600_Reg64RegClassID; break;
case 4: RegClassID = AMDGPU::R600_Reg128RegClassID; break;
default: llvm_unreachable("Do not know how to lower this BUILD_VECTOR");
}
// BUILD_VECTOR is usually lowered into an IMPLICIT_DEF + 4 INSERT_SUBREG
// that adds a 128 bits reg copy when going through TwoAddressInstructions
// pass. We want to avoid 128 bits copies as much as possible because they
// can't be bundled by our scheduler.
SDValue RegSeqArgs[9] = {
CurDAG->getTargetConstant(RegClassID, MVT::i32),
SDValue(), CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32),
SDValue(), CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32),
SDValue(), CurDAG->getTargetConstant(AMDGPU::sub2, MVT::i32),
SDValue(), CurDAG->getTargetConstant(AMDGPU::sub3, MVT::i32)
};
bool IsRegSeq = true;
for (unsigned i = 0; i < N->getNumOperands(); i++) {
if (dyn_cast<RegisterSDNode>(N->getOperand(i))) {
IsRegSeq = false;
break;
}
RegSeqArgs[2 * i + 1] = N->getOperand(i);
}
if (!IsRegSeq)
break;
return CurDAG->SelectNodeTo(N, AMDGPU::REG_SEQUENCE, N->getVTList(),
RegSeqArgs, 2 * N->getNumOperands() + 1);
}
case ISD::BUILD_PAIR: {
SDValue RC, SubReg0, SubReg1;
const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
break;
}
if (N->getValueType(0) == MVT::i128) {
RC = CurDAG->getTargetConstant(AMDGPU::SReg_128RegClassID, MVT::i32);
SubReg0 = CurDAG->getTargetConstant(AMDGPU::sub0_sub1, MVT::i32);
SubReg1 = CurDAG->getTargetConstant(AMDGPU::sub2_sub3, MVT::i32);
} else if (N->getValueType(0) == MVT::i64) {
RC = CurDAG->getTargetConstant(AMDGPU::VSrc_64RegClassID, MVT::i32);
SubReg0 = CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32);
SubReg1 = CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32);
} else {
llvm_unreachable("Unhandled value type for BUILD_PAIR");
}
const SDValue Ops[] = { RC, N->getOperand(0), SubReg0,
N->getOperand(1), SubReg1 };
return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE,
SDLoc(N), N->getValueType(0), Ops);
}
case ISD::ConstantFP:
case ISD::Constant: {
const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
// XXX: Custom immediate lowering not implemented yet. Instead we use
// pseudo instructions defined in SIInstructions.td
if (ST.getGeneration() > AMDGPUSubtarget::NORTHERN_ISLANDS) {
break;
}
uint64_t ImmValue = 0;
unsigned ImmReg = AMDGPU::ALU_LITERAL_X;
if (N->getOpcode() == ISD::ConstantFP) {
// XXX: 64-bit Immediates not supported yet
assert(N->getValueType(0) != MVT::f64);
ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N);
APFloat Value = C->getValueAPF();
float FloatValue = Value.convertToFloat();
if (FloatValue == 0.0) {
ImmReg = AMDGPU::ZERO;
} else if (FloatValue == 0.5) {
ImmReg = AMDGPU::HALF;
} else if (FloatValue == 1.0) {
ImmReg = AMDGPU::ONE;
} else {
ImmValue = Value.bitcastToAPInt().getZExtValue();
}
} else {
// XXX: 64-bit Immediates not supported yet
assert(N->getValueType(0) != MVT::i64);
ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
if (C->getZExtValue() == 0) {
ImmReg = AMDGPU::ZERO;
} else if (C->getZExtValue() == 1) {
ImmReg = AMDGPU::ONE_INT;
} else {
ImmValue = C->getZExtValue();
}
}
for (SDNode::use_iterator Use = N->use_begin(), Next = llvm::next(Use);
Use != SDNode::use_end(); Use = Next) {
Next = llvm::next(Use);
std::vector<SDValue> Ops;
for (unsigned i = 0; i < Use->getNumOperands(); ++i) {
Ops.push_back(Use->getOperand(i));
}
if (!Use->isMachineOpcode()) {
if (ImmReg == AMDGPU::ALU_LITERAL_X) {
// We can only use literal constants (e.g. AMDGPU::ZERO,
// AMDGPU::ONE, etc) in machine opcodes.
continue;
}
} else {
if (!TII->isALUInstr(Use->getMachineOpcode()) ||
(TII->get(Use->getMachineOpcode()).TSFlags &
R600_InstFlag::VECTOR)) {
continue;
}
int ImmIdx = TII->getOperandIdx(Use->getMachineOpcode(),
AMDGPU::OpName::literal);
if (ImmIdx == -1) {
continue;
}
if (TII->getOperandIdx(Use->getMachineOpcode(),
AMDGPU::OpName::dst) != -1) {
// subtract one from ImmIdx, because the DST operand is usually index
// 0 for MachineInstrs, but we have no DST in the Ops vector.
ImmIdx--;
}
// Check that we aren't already using an immediate.
// XXX: It's possible for an instruction to have more than one
// immediate operand, but this is not supported yet.
if (ImmReg == AMDGPU::ALU_LITERAL_X) {
ConstantSDNode *C = dyn_cast<ConstantSDNode>(Use->getOperand(ImmIdx));
assert(C);
if (C->getZExtValue() != 0) {
// This instruction is already using an immediate.
continue;
}
// Set the immediate value
Ops[ImmIdx] = CurDAG->getTargetConstant(ImmValue, MVT::i32);
}
}
// Set the immediate register
Ops[Use.getOperandNo()] = CurDAG->getRegister(ImmReg, MVT::i32);
CurDAG->UpdateNodeOperands(*Use, Ops.data(), Use->getNumOperands());
}
break;
}
}
SDNode *Result = SelectCode(N);
// Fold operands of selected node
const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
const R600InstrInfo *TII =
static_cast<const R600InstrInfo*>(TM.getInstrInfo());
if (Result && Result->isMachineOpcode() && Result->getMachineOpcode() == AMDGPU::DOT_4) {
bool IsModified = false;
do {
std::vector<SDValue> Ops;
for(SDNode::op_iterator I = Result->op_begin(), E = Result->op_end();
I != E; ++I)
Ops.push_back(*I);
IsModified = FoldDotOperands(Result->getMachineOpcode(), TII, Ops);
if (IsModified) {
Result = CurDAG->UpdateNodeOperands(Result, Ops.data(), Ops.size());
}
} while (IsModified);
}
if (Result && Result->isMachineOpcode() &&
!(TII->get(Result->getMachineOpcode()).TSFlags & R600_InstFlag::VECTOR)
&& TII->hasInstrModifiers(Result->getMachineOpcode())) {
// Fold FNEG/FABS
// TODO: Isel can generate multiple MachineInst, we need to recursively
// parse Result
bool IsModified = false;
do {
std::vector<SDValue> Ops;
for(SDNode::op_iterator I = Result->op_begin(), E = Result->op_end();
I != E; ++I)
Ops.push_back(*I);
IsModified = FoldOperands(Result->getMachineOpcode(), TII, Ops);
if (IsModified) {
Result = CurDAG->UpdateNodeOperands(Result, Ops.data(), Ops.size());
}
} while (IsModified);
// If node has a single use which is CLAMP_R600, folds it
if (Result->hasOneUse() && Result->isMachineOpcode()) {
SDNode *PotentialClamp = *Result->use_begin();
if (PotentialClamp->isMachineOpcode() &&
PotentialClamp->getMachineOpcode() == AMDGPU::CLAMP_R600) {
unsigned ClampIdx =
TII->getOperandIdx(Result->getMachineOpcode(), AMDGPU::OpName::clamp);
std::vector<SDValue> Ops;
unsigned NumOp = Result->getNumOperands();
for (unsigned i = 0; i < NumOp; ++i) {
Ops.push_back(Result->getOperand(i));
}
Ops[ClampIdx - 1] = CurDAG->getTargetConstant(1, MVT::i32);
Result = CurDAG->SelectNodeTo(PotentialClamp,
Result->getMachineOpcode(), PotentialClamp->getVTList(),
Ops.data(), NumOp);
}
}
}
}
return Result;
}
void ScheduleDAGSDNodes::BuildSchedUnits() {
// During scheduling, the NodeId field of SDNode is used to map SDNodes
// to their associated SUnits by holding SUnits table indices. A value
// of -1 means the SDNode does not yet have an associated SUnit.
unsigned NumNodes = 0;
for (SelectionDAG::allnodes_iterator NI = DAG->allnodes_begin(),
E = DAG->allnodes_end(); NI != E; ++NI) {
NI->setNodeId(-1);
++NumNodes;
}
// Reserve entries in the vector for each of the SUnits we are creating. This
// ensure that reallocation of the vector won't happen, so SUnit*'s won't get
// invalidated.
// FIXME: Multiply by 2 because we may clone nodes during scheduling.
// This is a temporary workaround.
SUnits.reserve(NumNodes * 2);
// Add all nodes in depth first order.
SmallVector<SDNode*, 64> Worklist;
SmallPtrSet<SDNode*, 64> Visited;
Worklist.push_back(DAG->getRoot().getNode());
Visited.insert(DAG->getRoot().getNode());
SmallVector<SUnit*, 8> CallSUnits;
while (!Worklist.empty()) {
SDNode *NI = Worklist.pop_back_val();
// Add all operands to the worklist unless they've already been added.
for (unsigned i = 0, e = NI->getNumOperands(); i != e; ++i)
if (Visited.insert(NI->getOperand(i).getNode()))
Worklist.push_back(NI->getOperand(i).getNode());
if (isPassiveNode(NI)) // Leaf node, e.g. a TargetImmediate.
continue;
// If this node has already been processed, stop now.
if (NI->getNodeId() != -1) continue;
SUnit *NodeSUnit = newSUnit(NI);
// See if anything is glued to this node, if so, add them to glued
// nodes. Nodes can have at most one glue input and one glue output. Glue
// is required to be the last operand and result of a node.
// Scan up to find glued preds.
SDNode *N = NI;
while (N->getNumOperands() &&
N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Glue) {
N = N->getOperand(N->getNumOperands()-1).getNode();
assert(N->getNodeId() == -1 && "Node already inserted!");
N->setNodeId(NodeSUnit->NodeNum);
if (N->isMachineOpcode() && TII->get(N->getMachineOpcode()).isCall())
NodeSUnit->isCall = true;
}
// Scan down to find any glued succs.
N = NI;
while (N->getValueType(N->getNumValues()-1) == MVT::Glue) {
SDValue GlueVal(N, N->getNumValues()-1);
// There are either zero or one users of the Glue result.
bool HasGlueUse = false;
for (SDNode::use_iterator UI = N->use_begin(), E = N->use_end();
UI != E; ++UI)
if (GlueVal.isOperandOf(*UI)) {
HasGlueUse = true;
assert(N->getNodeId() == -1 && "Node already inserted!");
N->setNodeId(NodeSUnit->NodeNum);
N = *UI;
if (N->isMachineOpcode() && TII->get(N->getMachineOpcode()).isCall())
NodeSUnit->isCall = true;
break;
}
if (!HasGlueUse) break;
}
if (NodeSUnit->isCall)
CallSUnits.push_back(NodeSUnit);
// Schedule zero-latency TokenFactor below any nodes that may increase the
// schedule height. Otherwise, ancestors of the TokenFactor may appear to
// have false stalls.
if (NI->getOpcode() == ISD::TokenFactor)
NodeSUnit->isScheduleLow = true;
// If there are glue operands involved, N is now the bottom-most node
// of the sequence of nodes that are glued together.
// Update the SUnit.
NodeSUnit->setNode(N);
assert(N->getNodeId() == -1 && "Node already inserted!");
N->setNodeId(NodeSUnit->NodeNum);
// Compute NumRegDefsLeft. This must be done before AddSchedEdges.
InitNumRegDefsLeft(NodeSUnit);
// Assign the Latency field of NodeSUnit using target-provided information.
computeLatency(NodeSUnit);
}
// Find all call operands.
while (!CallSUnits.empty()) {
SUnit *SU = CallSUnits.pop_back_val();
for (const SDNode *SUNode = SU->getNode(); SUNode;
SUNode = SUNode->getGluedNode()) {
if (SUNode->getOpcode() != ISD::CopyToReg)
continue;
SDNode *SrcN = SUNode->getOperand(2).getNode();
if (isPassiveNode(SrcN)) continue; // Not scheduled.
SUnit *SrcSU = &SUnits[SrcN->getNodeId()];
SrcSU->isCallOp = true;
}
}
}
void ScheduleDAGSDNodes::AddSchedEdges() {
const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
// Check to see if the scheduler cares about latencies.
bool UnitLatencies = forceUnitLatencies();
// Pass 2: add the preds, succs, etc.
for (unsigned su = 0, e = SUnits.size(); su != e; ++su) {
SUnit *SU = &SUnits[su];
SDNode *MainNode = SU->getNode();
if (MainNode->isMachineOpcode()) {
unsigned Opc = MainNode->getMachineOpcode();
const MCInstrDesc &MCID = TII->get(Opc);
for (unsigned i = 0; i != MCID.getNumOperands(); ++i) {
if (MCID.getOperandConstraint(i, MCOI::TIED_TO) != -1) {
SU->isTwoAddress = true;
break;
}
}
if (MCID.isCommutable())
SU->isCommutable = true;
}
// Find all predecessors and successors of the group.
for (SDNode *N = SU->getNode(); N; N = N->getGluedNode()) {
if (N->isMachineOpcode() &&
TII->get(N->getMachineOpcode()).getImplicitDefs()) {
SU->hasPhysRegClobbers = true;
unsigned NumUsed = InstrEmitter::CountResults(N);
while (NumUsed != 0 && !N->hasAnyUseOfValue(NumUsed - 1))
--NumUsed; // Skip over unused values at the end.
if (NumUsed > TII->get(N->getMachineOpcode()).getNumDefs())
SU->hasPhysRegDefs = true;
}
for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
SDNode *OpN = N->getOperand(i).getNode();
if (isPassiveNode(OpN)) continue; // Not scheduled.
SUnit *OpSU = &SUnits[OpN->getNodeId()];
assert(OpSU && "Node has no SUnit!");
if (OpSU == SU) continue; // In the same group.
EVT OpVT = N->getOperand(i).getValueType();
assert(OpVT != MVT::Glue && "Glued nodes should be in same sunit!");
bool isChain = OpVT == MVT::Other;
unsigned PhysReg = 0;
int Cost = 1;
// Determine if this is a physical register dependency.
CheckForPhysRegDependency(OpN, N, i, TRI, TII, PhysReg, Cost);
assert((PhysReg == 0 || !isChain) &&
"Chain dependence via physreg data?");
// FIXME: See ScheduleDAGSDNodes::EmitCopyFromReg. For now, scheduler
// emits a copy from the physical register to a virtual register unless
// it requires a cross class copy (cost < 0). That means we are only
// treating "expensive to copy" register dependency as physical register
// dependency. This may change in the future though.
if (Cost >= 0 && !StressSched)
PhysReg = 0;
// If this is a ctrl dep, latency is 1.
unsigned OpLatency = isChain ? 1 : OpSU->Latency;
// Special-case TokenFactor chains as zero-latency.
if(isChain && OpN->getOpcode() == ISD::TokenFactor)
OpLatency = 0;
const SDep &dep = SDep(OpSU, isChain ? SDep::Order : SDep::Data,
OpLatency, PhysReg);
if (!isChain && !UnitLatencies) {
computeOperandLatency(OpN, N, i, const_cast<SDep &>(dep));
ST.adjustSchedDependency(OpSU, SU, const_cast<SDep &>(dep));
}
if (!SU->addPred(dep) && !dep.isCtrl() && OpSU->NumRegDefsLeft > 1) {
// Multiple register uses are combined in the same SUnit. For example,
// we could have a set of glued nodes with all their defs consumed by
// another set of glued nodes. Register pressure tracking sees this as
// a single use, so to keep pressure balanced we reduce the defs.
//
// We can't tell (without more book-keeping) if this results from
// glued nodes or duplicate operands. As long as we don't reduce
// NumRegDefsLeft to zero, we handle the common cases well.
--OpSU->NumRegDefsLeft;
}
}
}
}
}
// Select - Convert the specified operand from a target-independent to a
// target-specific node if it hasn't already been changed.
SDNode *IA64DAGToDAGISel::Select(SDValue Op) {
SDNode *N = Op.getNode();
if (N->isMachineOpcode())
return NULL; // Already selected.
DebugLoc dl = Op.getDebugLoc();
switch (N->getOpcode()) {
default:
break;
case IA64ISD::BRCALL: { // XXX: this is also a hack!
SDValue Chain = N->getOperand(0);
SDValue InFlag; // Null incoming flag value.
if(N->getNumOperands()==3) { // we have an incoming chain, callee and flag
InFlag = N->getOperand(2);
}
unsigned CallOpcode;
SDValue CallOperand;
// if we can call directly, do so
if (GlobalAddressSDNode *GASD =
dyn_cast<GlobalAddressSDNode>(N->getOperand(1))) {
CallOpcode = IA64::BRCALL_IPREL_GA;
CallOperand = CurDAG->getTargetGlobalAddress(GASD->getGlobal(), MVT::i64);
} else if (isa<ExternalSymbolSDNode>(N->getOperand(1))) {
// FIXME: we currently NEED this case for correctness, to avoid
// "non-pic code with imm reloc.n against dynamic symbol" errors
CallOpcode = IA64::BRCALL_IPREL_ES;
CallOperand = N->getOperand(1);
} else {
// otherwise we need to load the function descriptor,
// load the branch target (function)'s entry point and GP,
// branch (call) then restore the GP
SDValue FnDescriptor = N->getOperand(1);
// load the branch target's entry point [mem] and
// GP value [mem+8]
SDValue targetEntryPoint=
SDValue(CurDAG->getTargetNode(IA64::LD8, dl, MVT::i64, MVT::Other,
FnDescriptor, CurDAG->getEntryNode()), 0);
Chain = targetEntryPoint.getValue(1);
SDValue targetGPAddr=
SDValue(CurDAG->getTargetNode(IA64::ADDS, dl, MVT::i64,
FnDescriptor,
CurDAG->getConstant(8, MVT::i64)), 0);
Chain = targetGPAddr.getValue(1);
SDValue targetGP =
SDValue(CurDAG->getTargetNode(IA64::LD8, dl, MVT::i64,MVT::Other,
targetGPAddr, CurDAG->getEntryNode()), 0);
Chain = targetGP.getValue(1);
Chain = CurDAG->getCopyToReg(Chain, dl, IA64::r1, targetGP, InFlag);
InFlag = Chain.getValue(1);
Chain = CurDAG->getCopyToReg(Chain, dl, IA64::B6,
targetEntryPoint, InFlag); // FLAG these?
InFlag = Chain.getValue(1);
CallOperand = CurDAG->getRegister(IA64::B6, MVT::i64);
CallOpcode = IA64::BRCALL_INDIRECT;
}
// Finally, once everything is setup, emit the call itself
if (InFlag.getNode())
Chain = SDValue(CurDAG->getTargetNode(CallOpcode, dl, MVT::Other,
MVT::Flag, CallOperand, InFlag), 0);
else // there might be no arguments
Chain = SDValue(CurDAG->getTargetNode(CallOpcode, dl, MVT::Other,
MVT::Flag, CallOperand, Chain), 0);
InFlag = Chain.getValue(1);
std::vector<SDValue> CallResults;
CallResults.push_back(Chain);
CallResults.push_back(InFlag);
for (unsigned i = 0, e = CallResults.size(); i != e; ++i)
ReplaceUses(Op.getValue(i), CallResults[i]);
return NULL;
}
case IA64ISD::GETFD: {
SDValue Input = N->getOperand(0);
return CurDAG->getTargetNode(IA64::GETFD, dl, MVT::i64, Input);
}
case ISD::FDIV:
case ISD::SDIV:
case ISD::UDIV:
case ISD::SREM:
case ISD::UREM:
return SelectDIV(Op);
case ISD::TargetConstantFP: {
SDValue Chain = CurDAG->getEntryNode(); // this is a constant, so..
SDValue V;
ConstantFPSDNode* N2 = cast<ConstantFPSDNode>(N);
if (N2->getValueAPF().isPosZero()) {
V = CurDAG->getCopyFromReg(Chain, dl, IA64::F0, MVT::f64);
} else if (N2->isExactlyValue(N2->getValueType(0) == MVT::f32 ?
APFloat(+1.0f) : APFloat(+1.0))) {
V = CurDAG->getCopyFromReg(Chain, dl, IA64::F1, MVT::f64);
} else
assert(0 && "Unexpected FP constant!");
ReplaceUses(SDValue(N, 0), V);
return 0;
}
case ISD::FrameIndex: { // TODO: reduce creepyness
int FI = cast<FrameIndexSDNode>(N)->getIndex();
if (N->hasOneUse())
return CurDAG->SelectNodeTo(N, IA64::MOV, MVT::i64,
CurDAG->getTargetFrameIndex(FI, MVT::i64));
else
return CurDAG->getTargetNode(IA64::MOV, dl, MVT::i64,
CurDAG->getTargetFrameIndex(FI, MVT::i64));
}
case ISD::ConstantPool: { // TODO: nuke the constant pool
// (ia64 doesn't need one)
ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(N);
Constant *C = CP->getConstVal();
SDValue CPI = CurDAG->getTargetConstantPool(C, MVT::i64,
CP->getAlignment());
return CurDAG->getTargetNode(IA64::ADDL_GA, dl, MVT::i64, // ?
CurDAG->getRegister(IA64::r1, MVT::i64), CPI);
}
case ISD::GlobalAddress: {
GlobalValue *GV = cast<GlobalAddressSDNode>(N)->getGlobal();
SDValue GA = CurDAG->getTargetGlobalAddress(GV, MVT::i64);
SDValue Tmp =
SDValue(CurDAG->getTargetNode(IA64::ADDL_GA, dl, MVT::i64,
CurDAG->getRegister(IA64::r1,
MVT::i64), GA), 0);
return CurDAG->getTargetNode(IA64::LD8, dl, MVT::i64, MVT::Other, Tmp,
CurDAG->getEntryNode());
}
/* XXX
case ISD::ExternalSymbol: {
SDValue EA = CurDAG->getTargetExternalSymbol(
cast<ExternalSymbolSDNode>(N)->getSymbol(),
MVT::i64);
SDValue Tmp = CurDAG->getTargetNode(IA64::ADDL_EA, dl, MVT::i64,
CurDAG->getRegister(IA64::r1,
MVT::i64),
EA);
return CurDAG->getTargetNode(IA64::LD8, dl, MVT::i64, Tmp);
}
*/
case ISD::LOAD: { // FIXME: load -1, not 1, for bools?
LoadSDNode *LD = cast<LoadSDNode>(N);
SDValue Chain = LD->getChain();
SDValue Address = LD->getBasePtr();
MVT TypeBeingLoaded = LD->getMemoryVT();
unsigned Opc;
switch (TypeBeingLoaded.getSimpleVT()) {
default:
#ifndef NDEBUG
N->dump(CurDAG);
#endif
assert(0 && "Cannot load this type!");
case MVT::i1: { // this is a bool
Opc = IA64::LD1; // first we load a byte, then compare for != 0
if(N->getValueType(0) == MVT::i1) { // XXX: early exit!
return CurDAG->SelectNodeTo(N, IA64::CMPNE, MVT::i1, MVT::Other,
SDValue(CurDAG->getTargetNode(Opc, dl,
MVT::i64,
Address), 0),
CurDAG->getRegister(IA64::r0, MVT::i64),
Chain);
}
/* otherwise, we want to load a bool into something bigger: LD1
will do that for us, so we just fall through */
}
case MVT::i8:
Opc = IA64::LD1;
break;
case MVT::i16:
Opc = IA64::LD2;
break;
case MVT::i32:
Opc = IA64::LD4;
break;
case MVT::i64:
Opc = IA64::LD8;
break;
case MVT::f32:
Opc = IA64::LDF4;
break;
case MVT::f64:
Opc = IA64::LDF8;
break;
}
// TODO: comment this
return CurDAG->SelectNodeTo(N, Opc, N->getValueType(0), MVT::Other,
Address, Chain);
}
case ISD::STORE: {
StoreSDNode *ST = cast<StoreSDNode>(N);
SDValue Address = ST->getBasePtr();
SDValue Chain = ST->getChain();
unsigned Opc;
if (ISD::isNON_TRUNCStore(N)) {
switch (N->getOperand(1).getValueType().getSimpleVT()) {
default:
assert(0 && "unknown type in store");
case MVT::i1: { // this is a bool
Opc = IA64::ST1; // we store either 0 or 1 as a byte
// first load zero!
SDValue Initial = CurDAG->getCopyFromReg(Chain, dl, IA64::r0, MVT::i64);
Chain = Initial.getValue(1);
// then load 1 into the same reg iff the predicate to store is 1
SDValue Tmp = ST->getValue();
Tmp =
SDValue(CurDAG->getTargetNode(IA64::TPCADDS, dl, MVT::i64, Initial,
CurDAG->getTargetConstant(1,
MVT::i64),
Tmp), 0);
return CurDAG->SelectNodeTo(N, Opc, MVT::Other, Address, Tmp, Chain);
}
case MVT::i64:
Opc = IA64::ST8;
break;
case MVT::f64:
Opc = IA64::STF8;
break;
}
} else { // Truncating store
switch(ST->getMemoryVT().getSimpleVT()) {
default:
assert(0 && "unknown type in truncstore");
case MVT::i8:
Opc = IA64::ST1;
break;
case MVT::i16:
Opc = IA64::ST2;
break;
case MVT::i32:
Opc = IA64::ST4;
break;
case MVT::f32:
Opc = IA64::STF4;
break;
}
}
SDValue N1 = N->getOperand(1);
SDValue N2 = N->getOperand(2);
return CurDAG->SelectNodeTo(N, Opc, MVT::Other, N2, N1, Chain);
}
case ISD::BRCOND: {
SDValue Chain = N->getOperand(0);
SDValue CC = N->getOperand(1);
MachineBasicBlock *Dest =
cast<BasicBlockSDNode>(N->getOperand(2))->getBasicBlock();
//FIXME - we do NOT need long branches all the time
return CurDAG->SelectNodeTo(N, IA64::BRLCOND_NOTCALL, MVT::Other, CC,
CurDAG->getBasicBlock(Dest), Chain);
}
case ISD::CALLSEQ_START:
case ISD::CALLSEQ_END: {
int64_t Amt = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
unsigned Opc = N->getOpcode() == ISD::CALLSEQ_START ?
IA64::ADJUSTCALLSTACKDOWN : IA64::ADJUSTCALLSTACKUP;
SDValue N0 = N->getOperand(0);
return CurDAG->SelectNodeTo(N, Opc, MVT::Other, getI64Imm(Amt), N0);
}
case ISD::BR:
// FIXME: we don't need long branches all the time!
SDValue N0 = N->getOperand(0);
return CurDAG->SelectNodeTo(N, IA64::BRL_NOTCALL, MVT::Other,
N->getOperand(1), N0);
}
return SelectCode(Op);
}
/// EmitCopyFromReg - Generate machine code for an CopyFromReg node or an
/// implicit physical register output.
void InstrEmitter::
EmitCopyFromReg(SDNode *Node, unsigned ResNo, bool IsClone, bool IsCloned,
unsigned SrcReg, DenseMap<SDValue, unsigned> &VRBaseMap) {
unsigned VRBase = 0;
if (TargetRegisterInfo::isVirtualRegister(SrcReg)) {
// Just use the input register directly!
SDValue Op(Node, ResNo);
if (IsClone)
VRBaseMap.erase(Op);
bool isNew = VRBaseMap.insert(std::make_pair(Op, SrcReg)).second;
(void)isNew; // Silence compiler warning.
assert(isNew && "Node emitted out of order - early");
return;
}
// If the node is only used by a CopyToReg and the dest reg is a vreg, use
// the CopyToReg'd destination register instead of creating a new vreg.
bool MatchReg = true;
const TargetRegisterClass *UseRC = NULL;
EVT VT = Node->getValueType(ResNo);
// Stick to the preferred register classes for legal types.
if (TLI->isTypeLegal(VT))
UseRC = TLI->getRegClassFor(VT);
if (!IsClone && !IsCloned)
for (SDNode::use_iterator UI = Node->use_begin(), E = Node->use_end();
UI != E; ++UI) {
SDNode *User = *UI;
bool Match = true;
if (User->getOpcode() == ISD::CopyToReg &&
User->getOperand(2).getNode() == Node &&
User->getOperand(2).getResNo() == ResNo) {
unsigned DestReg = cast<RegisterSDNode>(User->getOperand(1))->getReg();
if (TargetRegisterInfo::isVirtualRegister(DestReg)) {
VRBase = DestReg;
Match = false;
} else if (DestReg != SrcReg)
Match = false;
} else {
for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
SDValue Op = User->getOperand(i);
if (Op.getNode() != Node || Op.getResNo() != ResNo)
continue;
EVT VT = Node->getValueType(Op.getResNo());
if (VT == MVT::Other || VT == MVT::Glue)
continue;
Match = false;
if (User->isMachineOpcode()) {
const MCInstrDesc &II = TII->get(User->getMachineOpcode());
const TargetRegisterClass *RC = 0;
if (i+II.getNumDefs() < II.getNumOperands())
RC = TII->getRegClass(II, i+II.getNumDefs(), TRI);
if (!UseRC)
UseRC = RC;
else if (RC) {
const TargetRegisterClass *ComRC =
TRI->getCommonSubClass(UseRC, RC);
// If multiple uses expect disjoint register classes, we emit
// copies in AddRegisterOperand.
if (ComRC)
UseRC = ComRC;
}
}
}
}
MatchReg &= Match;
if (VRBase)
break;
}
const TargetRegisterClass *SrcRC = 0, *DstRC = 0;
SrcRC = TRI->getMinimalPhysRegClass(SrcReg, VT);
// Figure out the register class to create for the destreg.
if (VRBase) {
DstRC = MRI->getRegClass(VRBase);
} else if (UseRC) {
assert(UseRC->hasType(VT) && "Incompatible phys register def and uses!");
DstRC = UseRC;
} else {
DstRC = TLI->getRegClassFor(VT);
}
// If all uses are reading from the src physical register and copying the
// register is either impossible or very expensive, then don't create a copy.
if (MatchReg && SrcRC->getCopyCost() < 0) {
VRBase = SrcReg;
} else {
// Create the reg, emit the copy.
VRBase = MRI->createVirtualRegister(DstRC);
BuildMI(*MBB, InsertPos, Node->getDebugLoc(), TII->get(TargetOpcode::COPY),
VRBase).addReg(SrcReg);
}
SDValue Op(Node, ResNo);
if (IsClone)
VRBaseMap.erase(Op);
bool isNew = VRBaseMap.insert(std::make_pair(Op, VRBase)).second;
(void)isNew; // Silence compiler warning.
assert(isNew && "Node emitted out of order - early");
}
/// ClusterNeighboringLoads - Force nearby loads together by "flagging" them.
/// This function finds loads of the same base and different offsets. If the
/// offsets are not far apart (target specific), it add MVT::Flag inputs and
/// outputs to ensure they are scheduled together and in order. This
/// optimization may benefit some targets by improving cache locality.
void ScheduleDAGSDNodes::ClusterNeighboringLoads() {
SmallPtrSet<SDNode*, 16> Visited;
SmallVector<int64_t, 4> Offsets;
DenseMap<long long, SDNode*> O2SMap; // Map from offset to SDNode.
for (SelectionDAG::allnodes_iterator NI = DAG->allnodes_begin(),
E = DAG->allnodes_end(); NI != E; ++NI) {
SDNode *Node = &*NI;
if (!Node || !Node->isMachineOpcode())
continue;
unsigned Opc = Node->getMachineOpcode();
const TargetInstrDesc &TID = TII->get(Opc);
if (!TID.mayLoad())
continue;
SDNode *Chain = 0;
unsigned NumOps = Node->getNumOperands();
if (Node->getOperand(NumOps-1).getValueType() == MVT::Other)
Chain = Node->getOperand(NumOps-1).getNode();
if (!Chain)
continue;
// Look for other loads of the same chain. Find loads that are loading from
// the same base pointer and different offsets.
Visited.clear();
Offsets.clear();
O2SMap.clear();
bool Cluster = false;
SDNode *Base = Node;
int64_t BaseOffset;
for (SDNode::use_iterator I = Chain->use_begin(), E = Chain->use_end();
I != E; ++I) {
SDNode *User = *I;
if (User == Node || !Visited.insert(User))
continue;
int64_t Offset1, Offset2;
if (!TII->areLoadsFromSameBasePtr(Base, User, Offset1, Offset2) ||
Offset1 == Offset2)
// FIXME: Should be ok if they addresses are identical. But earlier
// optimizations really should have eliminated one of the loads.
continue;
if (O2SMap.insert(std::make_pair(Offset1, Base)).second)
Offsets.push_back(Offset1);
O2SMap.insert(std::make_pair(Offset2, User));
Offsets.push_back(Offset2);
if (Offset2 < Offset1) {
Base = User;
BaseOffset = Offset2;
} else {
BaseOffset = Offset1;
}
Cluster = true;
}
if (!Cluster)
continue;
// Sort them in increasing order.
std::sort(Offsets.begin(), Offsets.end());
// Check if the loads are close enough.
SmallVector<SDNode*, 4> Loads;
unsigned NumLoads = 0;
int64_t BaseOff = Offsets[0];
SDNode *BaseLoad = O2SMap[BaseOff];
Loads.push_back(BaseLoad);
for (unsigned i = 1, e = Offsets.size(); i != e; ++i) {
int64_t Offset = Offsets[i];
SDNode *Load = O2SMap[Offset];
if (!TII->shouldScheduleLoadsNear(BaseLoad, Load, BaseOff, Offset,
NumLoads))
break; // Stop right here. Ignore loads that are further away.
Loads.push_back(Load);
++NumLoads;
}
if (NumLoads == 0)
continue;
// Cluster loads by adding MVT::Flag outputs and inputs. This also
// ensure they are scheduled in order of increasing addresses.
SDNode *Lead = Loads[0];
AddFlags(Lead, SDValue(0,0), true, DAG);
SDValue InFlag = SDValue(Lead, Lead->getNumValues()-1);
for (unsigned i = 1, e = Loads.size(); i != e; ++i) {
bool OutFlag = i < e-1;
SDNode *Load = Loads[i];
AddFlags(Load, InFlag, OutFlag, DAG);
if (OutFlag)
InFlag = SDValue(Load, Load->getNumValues()-1);
++LoadsClustered;
}
}
}
/// Returns single number reflecting benefit of scheduling SU
/// in the current cycle.
signed ResourcePriorityQueue::SUSchedulingCost(SUnit *SU) {
// Initial trivial priority.
signed ResCount = 1;
// Do not waste time on a node that is already scheduled.
if (SU->isScheduled)
return ResCount;
// Forced priority is high.
if (SU->isScheduleHigh)
ResCount += PriorityOne;
// Adaptable scheduling
// A small, but very parallel
// region, where reg pressure is an issue.
if (HorizontalVerticalBalance > RegPressureThreshold) {
// Critical path first
ResCount += (SU->getHeight() * ScaleTwo);
// If resources are available for it, multiply the
// chance of scheduling.
if (isResourceAvailable(SU))
ResCount <<= FactorOne;
// Consider change to reg pressure from scheduling
// this SU.
ResCount -= (regPressureDelta(SU,true) * ScaleOne);
}
// Default heuristic, greeady and
// critical path driven.
else {
// Critical path first.
ResCount += (SU->getHeight() * ScaleTwo);
// Now see how many instructions is blocked by this SU.
ResCount += (NumNodesSolelyBlocking[SU->NodeNum] * ScaleTwo);
// If resources are available for it, multiply the
// chance of scheduling.
if (isResourceAvailable(SU))
ResCount <<= FactorOne;
ResCount -= (regPressureDelta(SU) * ScaleTwo);
}
// These are platform specific things.
// Will need to go into the back end
// and accessed from here via a hook.
for (SDNode *N = SU->getNode(); N; N = N->getGluedNode()) {
if (N->isMachineOpcode()) {
const MCInstrDesc &TID = TII->get(N->getMachineOpcode());
if (TID.isCall())
ResCount += (PriorityThree + (ScaleThree*N->getNumValues()));
}
else
switch (N->getOpcode()) {
default: break;
case ISD::TokenFactor:
case ISD::CopyFromReg:
case ISD::CopyToReg:
ResCount += PriorityFive;
break;
case ISD::INLINEASM:
ResCount += PriorityFour;
break;
}
}
return ResCount;
}
/// EmitCopyFromReg - Generate machine code for an CopyFromReg node or an
/// implicit physical register output.
void ScheduleDAGSDNodes::EmitCopyFromReg(SDNode *Node, unsigned ResNo,
bool IsClone, bool IsCloned,
unsigned SrcReg,
DenseMap<SDValue, unsigned> &VRBaseMap) {
unsigned VRBase = 0;
if (TargetRegisterInfo::isVirtualRegister(SrcReg)) {
// Just use the input register directly!
SDValue Op(Node, ResNo);
if (IsClone)
VRBaseMap.erase(Op);
bool isNew = VRBaseMap.insert(std::make_pair(Op, SrcReg)).second;
isNew = isNew; // Silence compiler warning.
assert(isNew && "Node emitted out of order - early");
return;
}
// If the node is only used by a CopyToReg and the dest reg is a vreg, use
// the CopyToReg'd destination register instead of creating a new vreg.
bool MatchReg = true;
const TargetRegisterClass *UseRC = NULL;
if (!IsClone && !IsCloned)
for (SDNode::use_iterator UI = Node->use_begin(), E = Node->use_end();
UI != E; ++UI) {
SDNode *User = *UI;
bool Match = true;
if (User->getOpcode() == ISD::CopyToReg &&
User->getOperand(2).getNode() == Node &&
User->getOperand(2).getResNo() == ResNo) {
unsigned DestReg = cast<RegisterSDNode>(User->getOperand(1))->getReg();
if (TargetRegisterInfo::isVirtualRegister(DestReg)) {
VRBase = DestReg;
Match = false;
} else if (DestReg != SrcReg)
Match = false;
} else {
for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
SDValue Op = User->getOperand(i);
if (Op.getNode() != Node || Op.getResNo() != ResNo)
continue;
MVT VT = Node->getValueType(Op.getResNo());
if (VT == MVT::Other || VT == MVT::Flag)
continue;
Match = false;
if (User->isMachineOpcode()) {
const TargetInstrDesc &II = TII->get(User->getMachineOpcode());
const TargetRegisterClass *RC =
getInstrOperandRegClass(TRI, II, i+II.getNumDefs());
if (!UseRC)
UseRC = RC;
else if (RC) {
if (UseRC->hasSuperClass(RC))
UseRC = RC;
else
assert((UseRC == RC || RC->hasSuperClass(UseRC)) &&
"Multiple uses expecting different register classes!");
}
}
}
}
MatchReg &= Match;
if (VRBase)
break;
}
MVT VT = Node->getValueType(ResNo);
const TargetRegisterClass *SrcRC = 0, *DstRC = 0;
SrcRC = TRI->getPhysicalRegisterRegClass(SrcReg, VT);
// Figure out the register class to create for the destreg.
if (VRBase) {
DstRC = MRI.getRegClass(VRBase);
} else if (UseRC) {
assert(UseRC->hasType(VT) && "Incompatible phys register def and uses!");
DstRC = UseRC;
} else {
DstRC = TLI->getRegClassFor(VT);
}
// If all uses are reading from the src physical register and copying the
// register is either impossible or very expensive, then don't create a copy.
if (MatchReg && SrcRC->getCopyCost() < 0) {
VRBase = SrcReg;
} else {
// Create the reg, emit the copy.
VRBase = MRI.createVirtualRegister(DstRC);
bool Emitted = TII->copyRegToReg(*BB, InsertPos, VRBase, SrcReg,
DstRC, SrcRC);
// If the target didn't handle the copy with different register
// classes and the destination is a subset of the source,
// try a normal same-RC copy.
if (!Emitted && DstRC->hasSuperClass(SrcRC))
Emitted = TII->copyRegToReg(*BB, InsertPos, VRBase, SrcReg,
SrcRC, SrcRC);
assert(Emitted && "Unable to issue a copy instruction!\n");
}
SDValue Op(Node, ResNo);
if (IsClone)
VRBaseMap.erase(Op);
bool isNew = VRBaseMap.insert(std::make_pair(Op, VRBase)).second;
isNew = isNew; // Silence compiler warning.
assert(isNew && "Node emitted out of order - early");
}
SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) {
unsigned int Opc = N->getOpcode();
if (N->isMachineOpcode()) {
return NULL; // Already selected.
}
switch (Opc) {
default: break;
case ISD::BUILD_VECTOR: {
const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
if (ST.device()->getGeneration() > AMDGPUDeviceInfo::HD6XXX) {
break;
}
// BUILD_VECTOR is usually lowered into an IMPLICIT_DEF + 4 INSERT_SUBREG
// that adds a 128 bits reg copy when going through TwoAddressInstructions
// pass. We want to avoid 128 bits copies as much as possible because they
// can't be bundled by our scheduler.
SDValue RegSeqArgs[9] = {
CurDAG->getTargetConstant(AMDGPU::R600_Reg128RegClassID, MVT::i32),
SDValue(), CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32),
SDValue(), CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32),
SDValue(), CurDAG->getTargetConstant(AMDGPU::sub2, MVT::i32),
SDValue(), CurDAG->getTargetConstant(AMDGPU::sub3, MVT::i32)
};
bool IsRegSeq = true;
for (unsigned i = 0; i < N->getNumOperands(); i++) {
if (dyn_cast<RegisterSDNode>(N->getOperand(i))) {
IsRegSeq = false;
break;
}
RegSeqArgs[2 * i + 1] = N->getOperand(i);
}
if (!IsRegSeq)
break;
return CurDAG->SelectNodeTo(N, AMDGPU::REG_SEQUENCE, N->getVTList(),
RegSeqArgs, 2 * N->getNumOperands() + 1);
}
case ISD::ConstantFP:
case ISD::Constant: {
const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
// XXX: Custom immediate lowering not implemented yet. Instead we use
// pseudo instructions defined in SIInstructions.td
if (ST.device()->getGeneration() > AMDGPUDeviceInfo::HD6XXX) {
break;
}
const R600InstrInfo *TII = static_cast<const R600InstrInfo*>(TM.getInstrInfo());
uint64_t ImmValue = 0;
unsigned ImmReg = AMDGPU::ALU_LITERAL_X;
if (N->getOpcode() == ISD::ConstantFP) {
// XXX: 64-bit Immediates not supported yet
assert(N->getValueType(0) != MVT::f64);
ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N);
APFloat Value = C->getValueAPF();
float FloatValue = Value.convertToFloat();
if (FloatValue == 0.0) {
ImmReg = AMDGPU::ZERO;
} else if (FloatValue == 0.5) {
ImmReg = AMDGPU::HALF;
} else if (FloatValue == 1.0) {
ImmReg = AMDGPU::ONE;
} else {
ImmValue = Value.bitcastToAPInt().getZExtValue();
}
} else {
// XXX: 64-bit Immediates not supported yet
assert(N->getValueType(0) != MVT::i64);
ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
if (C->getZExtValue() == 0) {
ImmReg = AMDGPU::ZERO;
} else if (C->getZExtValue() == 1) {
ImmReg = AMDGPU::ONE_INT;
} else {
ImmValue = C->getZExtValue();
}
}
for (SDNode::use_iterator Use = N->use_begin(), Next = llvm::next(Use);
Use != SDNode::use_end(); Use = Next) {
Next = llvm::next(Use);
std::vector<SDValue> Ops;
for (unsigned i = 0; i < Use->getNumOperands(); ++i) {
Ops.push_back(Use->getOperand(i));
}
if (!Use->isMachineOpcode()) {
if (ImmReg == AMDGPU::ALU_LITERAL_X) {
// We can only use literal constants (e.g. AMDGPU::ZERO,
// AMDGPU::ONE, etc) in machine opcodes.
continue;
}
} else {
if (!TII->isALUInstr(Use->getMachineOpcode()) ||
(TII->get(Use->getMachineOpcode()).TSFlags &
R600_InstFlag::VECTOR)) {
continue;
}
int ImmIdx = TII->getOperandIdx(Use->getMachineOpcode(), R600Operands::IMM);
assert(ImmIdx != -1);
// subtract one from ImmIdx, because the DST operand is usually index
// 0 for MachineInstrs, but we have no DST in the Ops vector.
ImmIdx--;
// Check that we aren't already using an immediate.
// XXX: It's possible for an instruction to have more than one
// immediate operand, but this is not supported yet.
if (ImmReg == AMDGPU::ALU_LITERAL_X) {
ConstantSDNode *C = dyn_cast<ConstantSDNode>(Use->getOperand(ImmIdx));
assert(C);
if (C->getZExtValue() != 0) {
// This instruction is already using an immediate.
continue;
}
// Set the immediate value
Ops[ImmIdx] = CurDAG->getTargetConstant(ImmValue, MVT::i32);
}
}
// Set the immediate register
Ops[Use.getOperandNo()] = CurDAG->getRegister(ImmReg, MVT::i32);
CurDAG->UpdateNodeOperands(*Use, Ops.data(), Use->getNumOperands());
}
break;
}
}
SDNode *Result = SelectCode(N);
// Fold operands of selected node
const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
if (ST.device()->getGeneration() <= AMDGPUDeviceInfo::HD6XXX) {
const R600InstrInfo *TII =
static_cast<const R600InstrInfo*>(TM.getInstrInfo());
if (Result && Result->isMachineOpcode() &&
!(TII->get(Result->getMachineOpcode()).TSFlags & R600_InstFlag::VECTOR)
&& TII->isALUInstr(Result->getMachineOpcode())) {
// Fold FNEG/FABS/CONST_ADDRESS
// TODO: Isel can generate multiple MachineInst, we need to recursively
// parse Result
bool IsModified = false;
do {
std::vector<SDValue> Ops;
for(SDNode::op_iterator I = Result->op_begin(), E = Result->op_end();
I != E; ++I)
Ops.push_back(*I);
IsModified = FoldOperands(Result->getMachineOpcode(), TII, Ops);
if (IsModified) {
Result = CurDAG->UpdateNodeOperands(Result, Ops.data(), Ops.size());
}
} while (IsModified);
// If node has a single use which is CLAMP_R600, folds it
if (Result->hasOneUse() && Result->isMachineOpcode()) {
SDNode *PotentialClamp = *Result->use_begin();
if (PotentialClamp->isMachineOpcode() &&
PotentialClamp->getMachineOpcode() == AMDGPU::CLAMP_R600) {
unsigned ClampIdx =
TII->getOperandIdx(Result->getMachineOpcode(), R600Operands::CLAMP);
std::vector<SDValue> Ops;
unsigned NumOp = Result->getNumOperands();
for (unsigned i = 0; i < NumOp; ++i) {
Ops.push_back(Result->getOperand(i));
}
Ops[ClampIdx - 1] = CurDAG->getTargetConstant(1, MVT::i32);
Result = CurDAG->SelectNodeTo(PotentialClamp,
Result->getMachineOpcode(), PotentialClamp->getVTList(),
Ops.data(), NumOp);
}
}
}
}
return Result;
}
void ScheduleDAGSDNodes::AddSchedEdges() {
const TargetSubtarget &ST = TM.getSubtarget<TargetSubtarget>();
// Check to see if the scheduler cares about latencies.
bool UnitLatencies = ForceUnitLatencies();
// Pass 2: add the preds, succs, etc.
for (unsigned su = 0, e = SUnits.size(); su != e; ++su) {
SUnit *SU = &SUnits[su];
SDNode *MainNode = SU->getNode();
if (MainNode->isMachineOpcode()) {
unsigned Opc = MainNode->getMachineOpcode();
const TargetInstrDesc &TID = TII->get(Opc);
for (unsigned i = 0; i != TID.getNumOperands(); ++i) {
if (TID.getOperandConstraint(i, TOI::TIED_TO) != -1) {
SU->isTwoAddress = true;
break;
}
}
if (TID.isCommutable())
SU->isCommutable = true;
}
// Find all predecessors and successors of the group.
for (SDNode *N = SU->getNode(); N; N = N->getGluedNode()) {
if (N->isMachineOpcode() &&
TII->get(N->getMachineOpcode()).getImplicitDefs()) {
SU->hasPhysRegClobbers = true;
unsigned NumUsed = InstrEmitter::CountResults(N);
while (NumUsed != 0 && !N->hasAnyUseOfValue(NumUsed - 1))
--NumUsed; // Skip over unused values at the end.
if (NumUsed > TII->get(N->getMachineOpcode()).getNumDefs())
SU->hasPhysRegDefs = true;
}
for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
SDNode *OpN = N->getOperand(i).getNode();
if (isPassiveNode(OpN)) continue; // Not scheduled.
SUnit *OpSU = &SUnits[OpN->getNodeId()];
assert(OpSU && "Node has no SUnit!");
if (OpSU == SU) continue; // In the same group.
EVT OpVT = N->getOperand(i).getValueType();
assert(OpVT != MVT::Glue && "Glued nodes should be in same sunit!");
bool isChain = OpVT == MVT::Other;
unsigned PhysReg = 0;
int Cost = 1;
// Determine if this is a physical register dependency.
CheckForPhysRegDependency(OpN, N, i, TRI, TII, PhysReg, Cost);
assert((PhysReg == 0 || !isChain) &&
"Chain dependence via physreg data?");
// FIXME: See ScheduleDAGSDNodes::EmitCopyFromReg. For now, scheduler
// emits a copy from the physical register to a virtual register unless
// it requires a cross class copy (cost < 0). That means we are only
// treating "expensive to copy" register dependency as physical register
// dependency. This may change in the future though.
if (Cost >= 0)
PhysReg = 0;
// If this is a ctrl dep, latency is 1.
unsigned OpLatency = isChain ? 1 : OpSU->Latency;
const SDep &dep = SDep(OpSU, isChain ? SDep::Order : SDep::Data,
OpLatency, PhysReg);
if (!isChain && !UnitLatencies) {
ComputeOperandLatency(OpN, N, i, const_cast<SDep &>(dep));
ST.adjustSchedDependency(OpSU, SU, const_cast<SDep &>(dep));
}
SU->addPred(dep);
}
}
}
}
