class AMDGPUDAGToDAGISel : public SelectionDAGISel {
// Subtarget - Keep a pointer to the AMDGPU Subtarget around so that we can
// make the right decision when generating code for different targets.
- const AMDGPUSubtarget &Subtarget;
+ const AMDGPUSubtarget *Subtarget;
public:
AMDGPUDAGToDAGISel(TargetMachine &TM);
virtual ~AMDGPUDAGToDAGISel();
-
+ bool runOnMachineFunction(MachineFunction &MF) override;
SDNode *Select(SDNode *N) override;
const char *getPassName() const override;
void PostprocessISelDAG() override;
bool SelectVOP3Mods0(SDValue In, SDValue &Src, SDValue &SrcMods,
SDValue &Clamp, SDValue &Omod) const;
+ bool SelectVOP3Mods0Clamp(SDValue In, SDValue &Src, SDValue &SrcMods,
+ SDValue &Omod) const;
+ bool SelectVOP3Mods0Clamp0OMod(SDValue In, SDValue &Src, SDValue &SrcMods,
+ SDValue &Clamp,
+ SDValue &Omod) const;
+
SDNode *SelectADD_SUB_I64(SDNode *N);
SDNode *SelectDIV_SCALE(SDNode *N);
}
AMDGPUDAGToDAGISel::AMDGPUDAGToDAGISel(TargetMachine &TM)
- : SelectionDAGISel(TM), Subtarget(TM.getSubtarget<AMDGPUSubtarget>()) {
+ : SelectionDAGISel(TM) {}
+
+bool AMDGPUDAGToDAGISel::runOnMachineFunction(MachineFunction &MF) {
+ Subtarget = &static_cast<const AMDGPUSubtarget &>(MF.getSubtarget());
+ return SelectionDAGISel::runOnMachineFunction(MF);
}
AMDGPUDAGToDAGISel::~AMDGPUDAGToDAGISel() {
switch (N->getMachineOpcode()) {
default: {
const MCInstrDesc &Desc =
- TM.getSubtargetImpl()->getInstrInfo()->get(N->getMachineOpcode());
+ Subtarget->getInstrInfo()->get(N->getMachineOpcode());
unsigned OpIdx = Desc.getNumDefs() + OpNo;
if (OpIdx >= Desc.getNumOperands())
return nullptr;
if (RegClass == -1)
return nullptr;
- return TM.getSubtargetImpl()->getRegisterInfo()->getRegClass(RegClass);
+ return Subtarget->getRegisterInfo()->getRegClass(RegClass);
}
case AMDGPU::REG_SEQUENCE: {
unsigned RCID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
const TargetRegisterClass *SuperRC =
- TM.getSubtargetImpl()->getRegisterInfo()->getRegClass(RCID);
+ Subtarget->getRegisterInfo()->getRegClass(RCID);
SDValue SubRegOp = N->getOperand(OpNo + 1);
unsigned SubRegIdx = cast<ConstantSDNode>(SubRegOp)->getZExtValue();
- return TM.getSubtargetImpl()->getRegisterInfo()->getSubClassWithSubReg(
- SuperRC, SubRegIdx);
+ return Subtarget->getRegisterInfo()->getSubClassWithSubReg(SuperRC,
+ SubRegIdx);
}
}
}
return nullptr; // Already selected.
}
- const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
switch (Opc) {
default: break;
// We are selecting i64 ADD here instead of custom lower it during
case ISD::ADD:
case ISD::SUB: {
if (N->getValueType(0) != MVT::i64 ||
- ST.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS)
+ Subtarget->getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS)
break;
return SelectADD_SUB_I64(N);
case AMDGPUISD::BUILD_VERTICAL_VECTOR:
case ISD::BUILD_VECTOR: {
unsigned RegClassID;
- const AMDGPURegisterInfo *TRI = static_cast<const AMDGPURegisterInfo *>(
- TM.getSubtargetImpl()->getRegisterInfo());
- const SIRegisterInfo *SIRI = static_cast<const SIRegisterInfo *>(
- TM.getSubtargetImpl()->getRegisterInfo());
+ const AMDGPURegisterInfo *TRI = Subtarget->getRegisterInfo();
EVT VT = N->getValueType(0);
unsigned NumVectorElts = VT.getVectorNumElements();
EVT EltVT = VT.getVectorElementType();
assert(EltVT.bitsEq(MVT::i32));
- if (ST.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
+ if (Subtarget->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
bool UseVReg = true;
for (SDNode::use_iterator U = N->use_begin(), E = SDNode::use_end();
U != E; ++U) {
if (!RC) {
continue;
}
- if (SIRI->isSGPRClass(RC)) {
+ if (static_cast<const SIRegisterInfo *>(TRI)->isSGPRClass(RC)) {
UseVReg = false;
}
}
switch(NumVectorElts) {
- case 1: RegClassID = UseVReg ? AMDGPU::VReg_32RegClassID :
+ case 1: RegClassID = UseVReg ? AMDGPU::VGPR_32RegClassID :
AMDGPU::SReg_32RegClassID;
break;
case 2: RegClassID = UseVReg ? AMDGPU::VReg_64RegClassID :
}
case ISD::BUILD_PAIR: {
SDValue RC, SubReg0, SubReg1;
- if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
+ if (Subtarget->getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
break;
}
if (N->getValueType(0) == MVT::i128) {
case ISD::Constant:
case ISD::ConstantFP: {
- const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
- if (ST.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS ||
+ if (Subtarget->getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS ||
N->getValueType(0).getSizeInBits() != 64 || isInlineImmediate(N))
break;
N->getValueType(0), Ops);
}
+ case ISD::LOAD: {
+ // To simplify the TableGen patters, we replace all i64 loads with
+ // v2i32 loads. Alternatively, we could promote i64 loads to v2i32
+ // during DAG legalization, however, so places (ExpandUnalignedLoad)
+ // in the DAG legalizer assume that if i64 is legal, so doing this
+ // promotion early can cause problems.
+ EVT VT = N->getValueType(0);
+ LoadSDNode *LD = cast<LoadSDNode>(N);
+ if (VT != MVT::i64 || LD->getExtensionType() != ISD::NON_EXTLOAD)
+ break;
+
+ SDValue NewLoad = CurDAG->getLoad(MVT::v2i32, SDLoc(N), LD->getChain(),
+ LD->getBasePtr(), LD->getMemOperand());
+ SDValue BitCast = CurDAG->getNode(ISD::BITCAST, SDLoc(N),
+ MVT::i64, NewLoad);
+ CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLoad.getValue(1));
+ CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), BitCast);
+ SelectCode(NewLoad.getNode());
+ N = BitCast.getNode();
+ break;
+ }
+
+ case ISD::STORE: {
+ // Handle i64 stores here for the same reason mentioned above for loads.
+ StoreSDNode *ST = cast<StoreSDNode>(N);
+ SDValue Value = ST->getValue();
+ if (Value.getValueType() != MVT::i64 || ST->isTruncatingStore())
+ break;
+
+ SDValue NewValue = CurDAG->getNode(ISD::BITCAST, SDLoc(N),
+ MVT::v2i32, Value);
+ SDValue NewStore = CurDAG->getStore(ST->getChain(), SDLoc(N), NewValue,
+ ST->getBasePtr(), ST->getMemOperand());
+
+ CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), NewStore);
+
+ if (NewValue.getOpcode() == ISD::BITCAST) {
+ Select(NewStore.getNode());
+ return SelectCode(NewValue.getNode());
+ }
+
+ // getNode() may fold the bitcast if its input was another bitcast. If that
+ // happens we should only select the new store.
+ N = NewStore.getNode();
+ break;
+ }
+
case AMDGPUISD::REGISTER_LOAD: {
- if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
+ if (Subtarget->getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
break;
SDValue Addr, Offset;
Ops);
}
case AMDGPUISD::REGISTER_STORE: {
- if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
+ if (Subtarget->getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
break;
SDValue Addr, Offset;
SelectADDRIndirect(N->getOperand(2), Addr, Offset);
case AMDGPUISD::BFE_I32:
case AMDGPUISD::BFE_U32: {
- if (ST.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS)
+ if (Subtarget->getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS)
break;
// There is a scalar version available, but unlike the vector version which
}
bool AMDGPUDAGToDAGISel::isGlobalLoad(const LoadSDNode *N) const {
- if (N->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS) {
- const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
- if (ST.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS ||
- N->getMemoryVT().bitsLT(MVT::i32)) {
+ if (N->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS)
+ if (Subtarget->getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS ||
+ N->getMemoryVT().bitsLT(MVT::i32))
return true;
- }
- }
+
return checkType(N->getMemOperand()->getValue(), AMDGPUAS::GLOBAL_ADDRESS);
}
bool AMDGPUDAGToDAGISel::isDSOffsetLegal(const SDValue &Base, unsigned Offset,
unsigned OffsetBits) const {
- const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
if ((OffsetBits == 16 && !isUInt<16>(Offset)) ||
(OffsetBits == 8 && !isUInt<8>(Offset)))
return false;
- if (ST.getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS)
+ if (Subtarget->getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS)
return true;
// On Southern Islands instruction with a negative base value and an offset
}
}
+ // If we have a constant address, prefer to put the constant into the
+ // offset. This can save moves to load the constant address since multiple
+ // operations can share the zero base address register, and enables merging
+ // into read2 / write2 instructions.
+ if (const ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr)) {
+ if (isUInt<16>(CAddr->getZExtValue())) {
+ SDValue Zero = CurDAG->getTargetConstant(0, MVT::i32);
+ MachineSDNode *MovZero = CurDAG->getMachineNode(AMDGPU::V_MOV_B32_e32,
+ SDLoc(Addr), MVT::i32, Zero);
+ Base = SDValue(MovZero, 0);
+ Offset = Addr;
+ return true;
+ }
+ }
+
// default case
Base = Addr;
Offset = CurDAG->getTargetConstant(0, MVT::i16);
}
}
+ if (const ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr)) {
+ unsigned DWordOffset0 = CAddr->getZExtValue() / 4;
+ unsigned DWordOffset1 = DWordOffset0 + 1;
+ assert(4 * DWordOffset0 == CAddr->getZExtValue());
+
+ if (isUInt<8>(DWordOffset0) && isUInt<8>(DWordOffset1)) {
+ SDValue Zero = CurDAG->getTargetConstant(0, MVT::i32);
+ MachineSDNode *MovZero
+ = CurDAG->getMachineNode(AMDGPU::V_MOV_B32_e32,
+ SDLoc(Addr), MVT::i32, Zero);
+ Base = SDValue(MovZero, 0);
+ Offset0 = CurDAG->getTargetConstant(DWordOffset0, MVT::i8);
+ Offset1 = CurDAG->getTargetConstant(DWordOffset1, MVT::i8);
+ return true;
+ }
+ }
+
// default case
Base = Addr;
Offset0 = CurDAG->getTargetConstant(0, MVT::i8);
return true;
}
-static SDValue wrapAddr64Rsrc(SelectionDAG *DAG, SDLoc DL, SDValue Ptr) {
- return SDValue(DAG->getMachineNode(AMDGPU::SI_ADDR64_RSRC, DL, MVT::v4i32,
- Ptr), 0);
-}
-
static bool isLegalMUBUFImmOffset(const ConstantSDNode *Imm) {
return isUInt<12>(Imm->getZExtValue());
}
ConstantSDNode *C = cast<ConstantSDNode>(Addr64);
if (C->getSExtValue()) {
SDLoc DL(Addr);
- SRsrc = wrapAddr64Rsrc(CurDAG, DL, Ptr);
+
+ const SITargetLowering& Lowering =
+ *static_cast<const SITargetLowering*>(getTargetLowering());
+
+ SRsrc = SDValue(Lowering.wrapAddr64Rsrc(*CurDAG, DL, Ptr), 0);
return true;
}
+
return false;
}
return SelectMUBUFAddr64(Addr, SRsrc, VAddr, Offset);
}
-static SDValue buildRSRC(SelectionDAG *DAG, SDLoc DL, SDValue Ptr,
- uint32_t RsrcDword1, uint64_t RsrcDword2And3) {
-
- SDValue PtrLo = DAG->getTargetExtractSubreg(AMDGPU::sub0, DL, MVT::i32, Ptr);
- SDValue PtrHi = DAG->getTargetExtractSubreg(AMDGPU::sub1, DL, MVT::i32, Ptr);
- if (RsrcDword1)
- PtrHi = SDValue(DAG->getMachineNode(AMDGPU::S_OR_B32, DL, MVT::i32, PtrHi,
- DAG->getConstant(RsrcDword1, MVT::i32)), 0);
-
- SDValue DataLo = DAG->getTargetConstant(
- RsrcDword2And3 & APInt::getAllOnesValue(32).getZExtValue(), MVT::i32);
- SDValue DataHi = DAG->getTargetConstant(RsrcDword2And3 >> 32, MVT::i32);
-
- const SDValue Ops[] = { PtrLo, PtrHi, DataLo, DataHi };
- return SDValue(DAG->getMachineNode(AMDGPU::SI_BUFFER_RSRC, DL,
- MVT::v4i32, Ops), 0);
-}
-
-/// \brief Return a resource descriptor with the 'Add TID' bit enabled
-/// The TID (Thread ID) is multipled by the stride value (bits [61:48]
-/// of the resource descriptor) to create an offset, which is added to the
-/// resource ponter.
-static SDValue buildScratchRSRC(SelectionDAG *DAG, SDLoc DL, SDValue Ptr) {
-
- uint64_t Rsrc = AMDGPU::RSRC_DATA_FORMAT | AMDGPU::RSRC_TID_ENABLE |
- 0xffffffff; // Size
-
- return buildRSRC(DAG, DL, Ptr, 0, Rsrc);
-}
-
bool AMDGPUDAGToDAGISel::SelectMUBUFScratch(SDValue Addr, SDValue &Rsrc,
SDValue &VAddr, SDValue &SOffset,
SDValue &ImmOffset) const {
SDLoc DL(Addr);
MachineFunction &MF = CurDAG->getMachineFunction();
const SIRegisterInfo *TRI =
- static_cast<const SIRegisterInfo *>(MF.getSubtarget().getRegisterInfo());
+ static_cast<const SIRegisterInfo *>(Subtarget->getRegisterInfo());
MachineRegisterInfo &MRI = MF.getRegInfo();
const SITargetLowering& Lowering =
*static_cast<const SITargetLowering*>(getTargetLowering());
- unsigned ScratchPtrReg =
- TRI->getPreloadedValue(MF, SIRegisterInfo::SCRATCH_PTR);
unsigned ScratchOffsetReg =
TRI->getPreloadedValue(MF, SIRegisterInfo::SCRATCH_WAVE_OFFSET);
Lowering.CreateLiveInRegister(*CurDAG, &AMDGPU::SReg_32RegClass,
ScratchOffsetReg, MVT::i32);
+ SDValue Sym0 = CurDAG->getExternalSymbol("SCRATCH_RSRC_DWORD0", MVT::i32);
+ SDValue ScratchRsrcDword0 =
+ SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32, Sym0), 0);
+
+ SDValue Sym1 = CurDAG->getExternalSymbol("SCRATCH_RSRC_DWORD1", MVT::i32);
+ SDValue ScratchRsrcDword1 =
+ SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32, Sym1), 0);
- Rsrc = buildScratchRSRC(CurDAG, DL,
- CurDAG->getCopyFromReg(CurDAG->getEntryNode(), DL,
- MRI.getLiveInVirtReg(ScratchPtrReg), MVT::i64));
+ const SDValue RsrcOps[] = {
+ CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, MVT::i32),
+ ScratchRsrcDword0,
+ CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32),
+ ScratchRsrcDword1,
+ CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32),
+ };
+ SDValue ScratchPtr = SDValue(CurDAG->getMachineNode(AMDGPU::REG_SEQUENCE, DL,
+ MVT::v2i32, RsrcOps), 0);
+ Rsrc = SDValue(Lowering.buildScratchRSRC(*CurDAG, DL, ScratchPtr), 0);
SOffset = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), DL,
MRI.getLiveInVirtReg(ScratchOffsetReg), MVT::i32);
}
}
- // (add FI, n0)
- if ((Addr.getOpcode() == ISD::ADD || Addr.getOpcode() == ISD::OR) &&
- isa<FrameIndexSDNode>(Addr.getOperand(0))) {
- VAddr = Addr.getOperand(1);
- ImmOffset = Addr.getOperand(0);
- return true;
- }
-
- // (FI)
- if (isa<FrameIndexSDNode>(Addr)) {
- VAddr = SDValue(CurDAG->getMachineNode(AMDGPU::V_MOV_B32_e32, DL, MVT::i32,
- CurDAG->getConstant(0, MVT::i32)), 0);
- ImmOffset = Addr;
- return true;
- }
-
// (node)
VAddr = Addr;
ImmOffset = CurDAG->getTargetConstant(0, MVT::i16);
SDValue &GLC, SDValue &SLC,
SDValue &TFE) const {
SDValue Ptr, VAddr, Offen, Idxen, Addr64;
+ const SIInstrInfo *TII =
+ static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
SelectMUBUF(Addr, Ptr, VAddr, SOffset, Offset, Offen, Idxen, Addr64,
GLC, SLC, TFE);
if (!cast<ConstantSDNode>(Offen)->getSExtValue() &&
!cast<ConstantSDNode>(Idxen)->getSExtValue() &&
!cast<ConstantSDNode>(Addr64)->getSExtValue()) {
- uint64_t Rsrc = AMDGPU::RSRC_DATA_FORMAT |
+ uint64_t Rsrc = TII->getDefaultRsrcDataFormat() |
APInt::getAllOnesValue(32).getZExtValue(); // Size
SDLoc DL(Addr);
- SRsrc = buildRSRC(CurDAG, DL, Ptr, 0, Rsrc);
+
+ const SITargetLowering& Lowering =
+ *static_cast<const SITargetLowering*>(getTargetLowering());
+
+ SRsrc = SDValue(Lowering.buildRSRC(*CurDAG, DL, Ptr, 0, Rsrc), 0);
return true;
}
return false;
AddrSpaceCastSDNode *ASC = cast<AddrSpaceCastSDNode>(N);
SDLoc DL(N);
- assert(Subtarget.hasFlatAddressSpace() &&
+ assert(Subtarget->hasFlatAddressSpace() &&
"addrspacecast only supported with flat address space!");
assert((ASC->getSrcAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS &&
if (DestSize > SrcSize) {
assert(SrcSize == 32 && DestSize == 64);
- SDValue RC = CurDAG->getTargetConstant(AMDGPU::VSrc_64RegClassID, MVT::i32);
+ // FIXME: This is probably wrong, we should never be defining
+ // a register class with both VGPRs and SGPRs
+ SDValue RC = CurDAG->getTargetConstant(AMDGPU::VS_64RegClassID, MVT::i32);
const SDValue Ops[] = {
RC,
return SelectVOP3Mods(In, Src, SrcMods);
}
+bool AMDGPUDAGToDAGISel::SelectVOP3Mods0Clamp(SDValue In, SDValue &Src,
+ SDValue &SrcMods,
+ SDValue &Omod) const {
+ // FIXME: Handle Omod
+ Omod = CurDAG->getTargetConstant(0, MVT::i32);
+
+ return SelectVOP3Mods(In, Src, SrcMods);
+}
+
+bool AMDGPUDAGToDAGISel::SelectVOP3Mods0Clamp0OMod(SDValue In, SDValue &Src,
+ SDValue &SrcMods,
+ SDValue &Clamp,
+ SDValue &Omod) const {
+ Clamp = Omod = CurDAG->getTargetConstant(0, MVT::i32);
+ return SelectVOP3Mods(In, Src, SrcMods);
+}
+
void AMDGPUDAGToDAGISel::PostprocessISelDAG() {
const AMDGPUTargetLowering& Lowering =
*static_cast<const AMDGPUTargetLowering*>(getTargetLowering());