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;
}
// 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);
}
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;
}
