#include "AMDGPUInstrInfo.h"
#include "AMDGPUISelLowering.h" // For AMDGPUISD
#include "AMDGPURegisterInfo.h"
+#include "AMDGPUSubtarget.h"
#include "R600InstrInfo.h"
+#include "SIDefines.h"
#include "SIISelLowering.h"
-#include "llvm/Analysis/ValueTracking.h"
-#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "SIMachineFunctionInfo.h"
+#include "llvm/CodeGen/FunctionLoweringInfo.h"
#include "llvm/CodeGen/PseudoSourceValue.h"
+#include "llvm/CodeGen/MachineFrameInfo.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/CodeGen/SelectionDAGISel.h"
-#include "llvm/IR/ValueMap.h"
-#include "llvm/Support/Compiler.h"
-#include <list>
-#include <queue>
+#include "llvm/IR/Function.h"
using namespace llvm;
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();
-
- SDNode *Select(SDNode *N);
- virtual const char *getPassName() const;
- virtual void PostprocessISelDAG();
+ bool runOnMachineFunction(MachineFunction &MF) override;
+ SDNode *Select(SDNode *N) override;
+ const char *getPassName() const override;
+ void PostprocessISelDAG() override;
private:
bool isInlineImmediate(SDNode *N) const;
bool SelectADDRParam(SDValue Addr, SDValue& R1, SDValue& R2);
bool SelectADDR(SDValue N, SDValue &R1, SDValue &R2);
bool SelectADDR64(SDValue N, SDValue &R1, SDValue &R2);
- SDValue SimplifyI24(SDValue &Op);
- bool SelectI24(SDValue Addr, SDValue &Op);
- bool SelectU24(SDValue Addr, SDValue &Op);
static bool checkType(const Value *ptr, unsigned int addrspace);
+ static bool checkPrivateAddress(const MachineMemOperand *Op);
static bool isGlobalStore(const StoreSDNode *N);
+ static bool isFlatStore(const StoreSDNode *N);
static bool isPrivateStore(const StoreSDNode *N);
static bool isLocalStore(const StoreSDNode *N);
static bool isRegionStore(const StoreSDNode *N);
bool isCPLoad(const LoadSDNode *N) const;
bool isConstantLoad(const LoadSDNode *N, int cbID) const;
bool isGlobalLoad(const LoadSDNode *N) const;
+ bool isFlatLoad(const LoadSDNode *N) const;
bool isParamLoad(const LoadSDNode *N) const;
bool isPrivateLoad(const LoadSDNode *N) const;
bool isLocalLoad(const LoadSDNode *N) const;
const TargetRegisterClass *getOperandRegClass(SDNode *N, unsigned OpNo) const;
bool SelectGlobalValueConstantOffset(SDValue Addr, SDValue& IntPtr);
- bool SelectGlobalValueVariableOffset(SDValue Addr,
- SDValue &BaseReg, SDValue& Offset);
+ bool SelectGlobalValueVariableOffset(SDValue Addr, SDValue &BaseReg,
+ SDValue& Offset);
bool SelectADDRVTX_READ(SDValue Addr, SDValue &Base, SDValue &Offset);
bool SelectADDRIndirect(SDValue Addr, SDValue &Base, SDValue &Offset);
+ bool isDSOffsetLegal(const SDValue &Base, unsigned Offset,
+ unsigned OffsetBits) const;
+ bool SelectDS1Addr1Offset(SDValue Ptr, SDValue &Base, SDValue &Offset) const;
+ bool SelectDS64Bit4ByteAligned(SDValue Ptr, SDValue &Base, SDValue &Offset0,
+ SDValue &Offset1) const;
+ void SelectMUBUF(SDValue Addr, SDValue &SRsrc, SDValue &VAddr,
+ SDValue &SOffset, SDValue &Offset, SDValue &Offen,
+ SDValue &Idxen, SDValue &Addr64, SDValue &GLC, SDValue &SLC,
+ SDValue &TFE) const;
+ bool SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc, SDValue &VAddr,
+ SDValue &Offset) const;
+ bool SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc,
+ SDValue &VAddr, SDValue &Offset,
+ SDValue &SLC) const;
+ bool SelectMUBUFScratch(SDValue Addr, SDValue &RSrc, SDValue &VAddr,
+ SDValue &SOffset, SDValue &ImmOffset) const;
+ bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &SOffset,
+ SDValue &Offset, SDValue &GLC, SDValue &SLC,
+ SDValue &TFE) const;
+ bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &Soffset,
+ SDValue &Offset, SDValue &GLC) const;
+ SDNode *SelectAddrSpaceCast(SDNode *N);
+ bool SelectVOP3Mods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
+ 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);
// Include the pieces autogenerated from the target description.
#include "AMDGPUGenDAGISel.inc"
/// \brief This pass converts a legalized DAG into a AMDGPU-specific
// DAG, ready for instruction scheduling.
-FunctionPass *llvm::createAMDGPUISelDag(TargetMachine &TM
- ) {
+FunctionPass *llvm::createAMDGPUISelDag(TargetMachine &TM) {
return new AMDGPUDAGToDAGISel(TM);
}
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() {
/// determined.
const TargetRegisterClass *AMDGPUDAGToDAGISel::getOperandRegClass(SDNode *N,
unsigned OpNo) const {
- if (!N->isMachineOpcode()) {
- return NULL;
- }
+ if (!N->isMachineOpcode())
+ return nullptr;
+
switch (N->getMachineOpcode()) {
default: {
- const MCInstrDesc &Desc = TM.getInstrInfo()->get(N->getMachineOpcode());
+ const MCInstrDesc &Desc =
+ Subtarget->getInstrInfo()->get(N->getMachineOpcode());
unsigned OpIdx = Desc.getNumDefs() + OpNo;
if (OpIdx >= Desc.getNumOperands())
- return NULL;
+ return nullptr;
int RegClass = Desc.OpInfo[OpIdx].RegClass;
- if (RegClass == -1) {
- return NULL;
- }
- return TM.getRegisterInfo()->getRegClass(RegClass);
+ if (RegClass == -1)
+ return nullptr;
+
+ return Subtarget->getRegisterInfo()->getRegClass(RegClass);
}
case AMDGPU::REG_SEQUENCE: {
- const TargetRegisterClass *SuperRC = TM.getRegisterInfo()->getRegClass(
- cast<ConstantSDNode>(N->getOperand(0))->getZExtValue());
- unsigned SubRegIdx =
- dyn_cast<ConstantSDNode>(N->getOperand(OpNo + 1))->getZExtValue();
- return TM.getRegisterInfo()->getSubClassWithSubReg(SuperRC, SubRegIdx);
+ unsigned RCID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
+ const TargetRegisterClass *SuperRC =
+ Subtarget->getRegisterInfo()->getRegClass(RCID);
+
+ SDValue SubRegOp = N->getOperand(OpNo + 1);
+ unsigned SubRegIdx = cast<ConstantSDNode>(SubRegOp)->getZExtValue();
+ return Subtarget->getRegisterInfo()->getSubClassWithSubReg(SuperRC,
+ SubRegIdx);
}
}
}
}
bool AMDGPUDAGToDAGISel::SelectADDRParam(
- SDValue Addr, SDValue& R1, SDValue& R2) {
+ SDValue Addr, SDValue& R1, SDValue& R2) {
if (Addr.getOpcode() == ISD::FrameIndex) {
if (FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Addr)) {
unsigned int Opc = N->getOpcode();
if (N->isMachineOpcode()) {
N->setNodeId(-1);
- return NULL; // Already selected.
+ return nullptr; // Already selected.
}
+
switch (Opc) {
default: break;
// We are selecting i64 ADD here instead of custom lower it during
// DAG legalization, so we can fold some i64 ADDs used for address
// calculation into the LOAD and STORE instructions.
- case ISD::ADD: {
- const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
+ case ISD::ADD:
+ case ISD::SUB: {
if (N->getValueType(0) != MVT::i64 ||
- ST.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS)
+ Subtarget->getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS)
break;
- SDLoc DL(N);
- SDValue LHS = N->getOperand(0);
- SDValue RHS = N->getOperand(1);
-
- SDValue Sub0 = CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32);
- SDValue Sub1 = CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32);
-
- SDNode *Lo0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
- DL, MVT::i32, LHS, Sub0);
- SDNode *Hi0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
- DL, MVT::i32, LHS, Sub1);
-
- SDNode *Lo1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
- DL, MVT::i32, RHS, Sub0);
- SDNode *Hi1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
- DL, MVT::i32, RHS, Sub1);
-
- SDVTList VTList = CurDAG->getVTList(MVT::i32, MVT::Glue);
-
- SmallVector<SDValue, 8> AddLoArgs;
- AddLoArgs.push_back(SDValue(Lo0, 0));
- AddLoArgs.push_back(SDValue(Lo1, 0));
-
- SDNode *AddLo = CurDAG->getMachineNode(AMDGPU::S_ADD_I32, DL,
- VTList, AddLoArgs);
- SDValue Carry = SDValue(AddLo, 1);
- SDNode *AddHi = CurDAG->getMachineNode(AMDGPU::S_ADDC_U32, DL,
- MVT::i32, SDValue(Hi0, 0),
- SDValue(Hi1, 0), Carry);
-
- SDValue Args[5] = {
- CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, MVT::i32),
- SDValue(AddLo,0),
- Sub0,
- SDValue(AddHi,0),
- Sub1,
- };
- return CurDAG->SelectNodeTo(N, AMDGPU::REG_SEQUENCE, MVT::i64, Args, 5);
+ return SelectADD_SUB_I64(N);
}
+ case ISD::SCALAR_TO_VECTOR:
+ case AMDGPUISD::BUILD_VERTICAL_VECTOR:
case ISD::BUILD_VECTOR: {
unsigned RegClassID;
- const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
- const AMDGPURegisterInfo *TRI =
- static_cast<const AMDGPURegisterInfo*>(TM.getRegisterInfo());
- const SIRegisterInfo *SIRI =
- static_cast<const SIRegisterInfo*>(TM.getRegisterInfo());
+ const AMDGPURegisterInfo *TRI = Subtarget->getRegisterInfo();
EVT VT = N->getValueType(0);
unsigned NumVectorElts = VT.getVectorNumElements();
- assert(VT.getVectorElementType().bitsEq(MVT::i32));
- if (ST.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
+ EVT EltVT = VT.getVectorElementType();
+ assert(EltVT.bitsEq(MVT::i32));
+ 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 :
// can't be bundled by our scheduler.
switch(NumVectorElts) {
case 2: RegClassID = AMDGPU::R600_Reg64RegClassID; break;
- case 4: RegClassID = AMDGPU::R600_Reg128RegClassID; break;
+ case 4:
+ if (Opc == AMDGPUISD::BUILD_VERTICAL_VECTOR)
+ RegClassID = AMDGPU::R600_Reg128VerticalRegClassID;
+ else
+ RegClassID = AMDGPU::R600_Reg128RegClassID;
+ break;
default: llvm_unreachable("Do not know how to lower this BUILD_VECTOR");
}
}
SDValue RegClass = CurDAG->getTargetConstant(RegClassID, MVT::i32);
if (NumVectorElts == 1) {
- return CurDAG->SelectNodeTo(N, AMDGPU::COPY_TO_REGCLASS,
- VT.getVectorElementType(),
+ return CurDAG->SelectNodeTo(N, AMDGPU::COPY_TO_REGCLASS, EltVT,
N->getOperand(0), RegClass);
}
// 16 = Max Num Vector Elements
// 2 = 2 REG_SEQUENCE operands per element (value, subreg index)
// 1 = Vector Register Class
- SDValue RegSeqArgs[16 * 2 + 1];
+ SmallVector<SDValue, 16 * 2 + 1> RegSeqArgs(NumVectorElts * 2 + 1);
RegSeqArgs[0] = CurDAG->getTargetConstant(RegClassID, MVT::i32);
bool IsRegSeq = true;
- for (unsigned i = 0; i < N->getNumOperands(); i++) {
+ unsigned NOps = N->getNumOperands();
+ for (unsigned i = 0; i < NOps; i++) {
// XXX: Why is this here?
if (dyn_cast<RegisterSDNode>(N->getOperand(i))) {
IsRegSeq = false;
RegSeqArgs[1 + (2 * i) + 1] =
CurDAG->getTargetConstant(TRI->getSubRegFromChannel(i), MVT::i32);
}
+
+ if (NOps != NumVectorElts) {
+ // Fill in the missing undef elements if this was a scalar_to_vector.
+ assert(Opc == ISD::SCALAR_TO_VECTOR && NOps < NumVectorElts);
+
+ MachineSDNode *ImpDef = CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
+ SDLoc(N), EltVT);
+ for (unsigned i = NOps; i < NumVectorElts; ++i) {
+ RegSeqArgs[1 + (2 * i)] = SDValue(ImpDef, 0);
+ RegSeqArgs[1 + (2 * i) + 1] =
+ CurDAG->getTargetConstant(TRI->getSubRegFromChannel(i), MVT::i32);
+ }
+ }
+
if (!IsRegSeq)
break;
return CurDAG->SelectNodeTo(N, AMDGPU::REG_SEQUENCE, N->getVTList(),
- RegSeqArgs, 2 * N->getNumOperands() + 1);
+ RegSeqArgs);
}
case ISD::BUILD_PAIR: {
SDValue RC, SubReg0, SubReg1;
- const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
- if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
+ if (Subtarget->getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
break;
}
if (N->getValueType(0) == MVT::i128) {
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);
+ RC = CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, MVT::i32);
SubReg0 = CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32);
SubReg1 = CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32);
} else {
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: {
- const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
- if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
+ if (Subtarget->getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
break;
SDValue Addr, Offset;
Ops);
}
case AMDGPUISD::REGISTER_STORE: {
- const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
- if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
+ if (Subtarget->getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
break;
SDValue Addr, Offset;
SelectADDRIndirect(N->getOperand(2), Addr, Offset);
CurDAG->getVTList(MVT::Other),
Ops);
}
+
+ case AMDGPUISD::BFE_I32:
+ case AMDGPUISD::BFE_U32: {
+ if (Subtarget->getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS)
+ break;
+
+ // There is a scalar version available, but unlike the vector version which
+ // has a separate operand for the offset and width, the scalar version packs
+ // the width and offset into a single operand. Try to move to the scalar
+ // version if the offsets are constant, so that we can try to keep extended
+ // loads of kernel arguments in SGPRs.
+
+ // TODO: Technically we could try to pattern match scalar bitshifts of
+ // dynamic values, but it's probably not useful.
+ ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
+ if (!Offset)
+ break;
+
+ ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
+ if (!Width)
+ break;
+
+ bool Signed = Opc == AMDGPUISD::BFE_I32;
+
+ // Transformation function, pack the offset and width of a BFE into
+ // the format expected by the S_BFE_I32 / S_BFE_U32. In the second
+ // source, bits [5:0] contain the offset and bits [22:16] the width.
+
+ uint32_t OffsetVal = Offset->getZExtValue();
+ uint32_t WidthVal = Width->getZExtValue();
+
+ uint32_t PackedVal = OffsetVal | WidthVal << 16;
+
+ SDValue PackedOffsetWidth = CurDAG->getTargetConstant(PackedVal, MVT::i32);
+ return CurDAG->getMachineNode(Signed ? AMDGPU::S_BFE_I32 : AMDGPU::S_BFE_U32,
+ SDLoc(N),
+ MVT::i32,
+ N->getOperand(0),
+ PackedOffsetWidth);
+
+ }
+ case AMDGPUISD::DIV_SCALE: {
+ return SelectDIV_SCALE(N);
+ }
+ case ISD::CopyToReg: {
+ const SITargetLowering& Lowering =
+ *static_cast<const SITargetLowering*>(getTargetLowering());
+ Lowering.legalizeTargetIndependentNode(N, *CurDAG);
+ break;
+ }
+ case ISD::ADDRSPACECAST:
+ return SelectAddrSpaceCast(N);
}
+
return SelectCode(N);
}
-bool AMDGPUDAGToDAGISel::checkType(const Value *ptr, unsigned int addrspace) {
- if (!ptr) {
+bool AMDGPUDAGToDAGISel::checkType(const Value *Ptr, unsigned AS) {
+ assert(AS != 0 && "Use checkPrivateAddress instead.");
+ if (!Ptr)
return false;
- }
- Type *ptrType = ptr->getType();
- return dyn_cast<PointerType>(ptrType)->getAddressSpace() == addrspace;
+
+ return Ptr->getType()->getPointerAddressSpace() == AS;
+}
+
+bool AMDGPUDAGToDAGISel::checkPrivateAddress(const MachineMemOperand *Op) {
+ if (Op->getPseudoValue())
+ return true;
+
+ if (PointerType *PT = dyn_cast<PointerType>(Op->getValue()->getType()))
+ return PT->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS;
+
+ return false;
}
bool AMDGPUDAGToDAGISel::isGlobalStore(const StoreSDNode *N) {
- return checkType(N->getSrcValue(), AMDGPUAS::GLOBAL_ADDRESS);
+ return checkType(N->getMemOperand()->getValue(), AMDGPUAS::GLOBAL_ADDRESS);
}
bool AMDGPUDAGToDAGISel::isPrivateStore(const StoreSDNode *N) {
- return (!checkType(N->getSrcValue(), AMDGPUAS::LOCAL_ADDRESS)
- && !checkType(N->getSrcValue(), AMDGPUAS::GLOBAL_ADDRESS)
- && !checkType(N->getSrcValue(), AMDGPUAS::REGION_ADDRESS));
+ const Value *MemVal = N->getMemOperand()->getValue();
+ return (!checkType(MemVal, AMDGPUAS::LOCAL_ADDRESS) &&
+ !checkType(MemVal, AMDGPUAS::GLOBAL_ADDRESS) &&
+ !checkType(MemVal, AMDGPUAS::REGION_ADDRESS));
}
bool AMDGPUDAGToDAGISel::isLocalStore(const StoreSDNode *N) {
- return checkType(N->getSrcValue(), AMDGPUAS::LOCAL_ADDRESS);
+ return checkType(N->getMemOperand()->getValue(), AMDGPUAS::LOCAL_ADDRESS);
+}
+
+bool AMDGPUDAGToDAGISel::isFlatStore(const StoreSDNode *N) {
+ return checkType(N->getMemOperand()->getValue(), AMDGPUAS::FLAT_ADDRESS);
}
bool AMDGPUDAGToDAGISel::isRegionStore(const StoreSDNode *N) {
- return checkType(N->getSrcValue(), AMDGPUAS::REGION_ADDRESS);
+ return checkType(N->getMemOperand()->getValue(), AMDGPUAS::REGION_ADDRESS);
}
bool AMDGPUDAGToDAGISel::isConstantLoad(const LoadSDNode *N, int CbId) const {
- if (CbId == -1) {
- return checkType(N->getSrcValue(), AMDGPUAS::CONSTANT_ADDRESS);
- }
- return checkType(N->getSrcValue(), AMDGPUAS::CONSTANT_BUFFER_0 + CbId);
+ const Value *MemVal = N->getMemOperand()->getValue();
+ if (CbId == -1)
+ return checkType(MemVal, AMDGPUAS::CONSTANT_ADDRESS);
+
+ return checkType(MemVal, AMDGPUAS::CONSTANT_BUFFER_0 + CbId);
}
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->getSrcValue(), AMDGPUAS::GLOBAL_ADDRESS);
+
+ return checkType(N->getMemOperand()->getValue(), AMDGPUAS::GLOBAL_ADDRESS);
}
bool AMDGPUDAGToDAGISel::isParamLoad(const LoadSDNode *N) const {
- return checkType(N->getSrcValue(), AMDGPUAS::PARAM_I_ADDRESS);
+ return checkType(N->getMemOperand()->getValue(), AMDGPUAS::PARAM_I_ADDRESS);
}
bool AMDGPUDAGToDAGISel::isLocalLoad(const LoadSDNode *N) const {
- return checkType(N->getSrcValue(), AMDGPUAS::LOCAL_ADDRESS);
+ return checkType(N->getMemOperand()->getValue(), AMDGPUAS::LOCAL_ADDRESS);
+}
+
+bool AMDGPUDAGToDAGISel::isFlatLoad(const LoadSDNode *N) const {
+ return checkType(N->getMemOperand()->getValue(), AMDGPUAS::FLAT_ADDRESS);
}
bool AMDGPUDAGToDAGISel::isRegionLoad(const LoadSDNode *N) const {
- return checkType(N->getSrcValue(), AMDGPUAS::REGION_ADDRESS);
+ return checkType(N->getMemOperand()->getValue(), AMDGPUAS::REGION_ADDRESS);
}
bool AMDGPUDAGToDAGISel::isCPLoad(const LoadSDNode *N) const {
MachineMemOperand *MMO = N->getMemOperand();
- if (checkType(N->getSrcValue(), AMDGPUAS::PRIVATE_ADDRESS)) {
+ if (checkPrivateAddress(N->getMemOperand())) {
if (MMO) {
- const Value *V = MMO->getValue();
- const PseudoSourceValue *PSV = dyn_cast<PseudoSourceValue>(V);
+ const PseudoSourceValue *PSV = MMO->getPseudoValue();
if (PSV && PSV == PseudoSourceValue::getConstantPool()) {
return true;
}
}
bool AMDGPUDAGToDAGISel::isPrivateLoad(const LoadSDNode *N) const {
- if (checkType(N->getSrcValue(), AMDGPUAS::PRIVATE_ADDRESS)) {
+ if (checkPrivateAddress(N->getMemOperand())) {
// Check to make sure we are not a constant pool load or a constant load
// that is marked as a private load
if (isCPLoad(N) || isConstantLoad(N, -1)) {
return false;
}
}
- if (!checkType(N->getSrcValue(), AMDGPUAS::LOCAL_ADDRESS)
- && !checkType(N->getSrcValue(), AMDGPUAS::GLOBAL_ADDRESS)
- && !checkType(N->getSrcValue(), AMDGPUAS::REGION_ADDRESS)
- && !checkType(N->getSrcValue(), AMDGPUAS::CONSTANT_ADDRESS)
- && !checkType(N->getSrcValue(), AMDGPUAS::PARAM_D_ADDRESS)
- && !checkType(N->getSrcValue(), AMDGPUAS::PARAM_I_ADDRESS)) {
+
+ const Value *MemVal = N->getMemOperand()->getValue();
+ if (!checkType(MemVal, AMDGPUAS::LOCAL_ADDRESS) &&
+ !checkType(MemVal, AMDGPUAS::GLOBAL_ADDRESS) &&
+ !checkType(MemVal, AMDGPUAS::FLAT_ADDRESS) &&
+ !checkType(MemVal, AMDGPUAS::REGION_ADDRESS) &&
+ !checkType(MemVal, AMDGPUAS::CONSTANT_ADDRESS) &&
+ !checkType(MemVal, AMDGPUAS::PARAM_D_ADDRESS) &&
+ !checkType(MemVal, AMDGPUAS::PARAM_I_ADDRESS)) {
return true;
}
return false;
//===----------------------------------------------------------------------===//
bool AMDGPUDAGToDAGISel::SelectGlobalValueConstantOffset(SDValue Addr,
- SDValue& IntPtr) {
+ SDValue& IntPtr) {
if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Addr)) {
IntPtr = CurDAG->getIntPtrConstant(Cst->getZExtValue() / 4, true);
return true;
bool AMDGPUDAGToDAGISel::SelectGlobalValueVariableOffset(SDValue Addr,
SDValue& BaseReg, SDValue &Offset) {
- if (!dyn_cast<ConstantSDNode>(Addr)) {
+ if (!isa<ConstantSDNode>(Addr)) {
BaseReg = Addr;
Offset = CurDAG->getIntPtrConstant(0, true);
return true;
bool AMDGPUDAGToDAGISel::SelectADDRVTX_READ(SDValue Addr, SDValue &Base,
SDValue &Offset) {
- ConstantSDNode * IMMOffset;
+ ConstantSDNode *IMMOffset;
if (Addr.getOpcode() == ISD::ADD
&& (IMMOffset = dyn_cast<ConstantSDNode>(Addr.getOperand(1)))
return true;
}
-SDValue AMDGPUDAGToDAGISel::SimplifyI24(SDValue &Op) {
- APInt Demanded = APInt(32, 0x00FFFFFF);
- APInt KnownZero, KnownOne;
- TargetLowering::TargetLoweringOpt TLO(*CurDAG, true, true);
- const TargetLowering *TLI = getTargetLowering();
- if (TLI->SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO)) {
- CurDAG->ReplaceAllUsesWith(Op, TLO.New);
- CurDAG->RepositionNode(Op.getNode(), TLO.New.getNode());
- return SimplifyI24(TLO.New);
- } else {
- return Op;
+SDNode *AMDGPUDAGToDAGISel::SelectADD_SUB_I64(SDNode *N) {
+ SDLoc DL(N);
+ SDValue LHS = N->getOperand(0);
+ SDValue RHS = N->getOperand(1);
+
+ bool IsAdd = (N->getOpcode() == ISD::ADD);
+
+ SDValue Sub0 = CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32);
+ SDValue Sub1 = CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32);
+
+ SDNode *Lo0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
+ DL, MVT::i32, LHS, Sub0);
+ SDNode *Hi0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
+ DL, MVT::i32, LHS, Sub1);
+
+ SDNode *Lo1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
+ DL, MVT::i32, RHS, Sub0);
+ SDNode *Hi1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
+ DL, MVT::i32, RHS, Sub1);
+
+ SDVTList VTList = CurDAG->getVTList(MVT::i32, MVT::Glue);
+ SDValue AddLoArgs[] = { SDValue(Lo0, 0), SDValue(Lo1, 0) };
+
+
+ unsigned Opc = IsAdd ? AMDGPU::S_ADD_U32 : AMDGPU::S_SUB_U32;
+ unsigned CarryOpc = IsAdd ? AMDGPU::S_ADDC_U32 : AMDGPU::S_SUBB_U32;
+
+ SDNode *AddLo = CurDAG->getMachineNode( Opc, DL, VTList, AddLoArgs);
+ SDValue Carry(AddLo, 1);
+ SDNode *AddHi
+ = CurDAG->getMachineNode(CarryOpc, DL, MVT::i32,
+ SDValue(Hi0, 0), SDValue(Hi1, 0), Carry);
+
+ SDValue Args[5] = {
+ CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, MVT::i32),
+ SDValue(AddLo,0),
+ Sub0,
+ SDValue(AddHi,0),
+ Sub1,
+ };
+ return CurDAG->SelectNodeTo(N, AMDGPU::REG_SEQUENCE, MVT::i64, Args);
+}
+
+SDNode *AMDGPUDAGToDAGISel::SelectDIV_SCALE(SDNode *N) {
+ SDLoc SL(N);
+ EVT VT = N->getValueType(0);
+
+ assert(VT == MVT::f32 || VT == MVT::f64);
+
+ unsigned Opc
+ = (VT == MVT::f64) ? AMDGPU::V_DIV_SCALE_F64 : AMDGPU::V_DIV_SCALE_F32;
+
+ const SDValue Zero = CurDAG->getTargetConstant(0, MVT::i32);
+ const SDValue False = CurDAG->getTargetConstant(0, MVT::i1);
+ SDValue Ops[] = {
+ Zero, // src0_modifiers
+ N->getOperand(0), // src0
+ Zero, // src1_modifiers
+ N->getOperand(1), // src1
+ Zero, // src2_modifiers
+ N->getOperand(2), // src2
+ False, // clamp
+ Zero // omod
+ };
+
+ return CurDAG->SelectNodeTo(N, Opc, VT, MVT::i1, Ops);
+}
+
+bool AMDGPUDAGToDAGISel::isDSOffsetLegal(const SDValue &Base, unsigned Offset,
+ unsigned OffsetBits) const {
+ if ((OffsetBits == 16 && !isUInt<16>(Offset)) ||
+ (OffsetBits == 8 && !isUInt<8>(Offset)))
+ return false;
+
+ if (Subtarget->getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS)
+ return true;
+
+ // On Southern Islands instruction with a negative base value and an offset
+ // don't seem to work.
+ return CurDAG->SignBitIsZero(Base);
+}
+
+bool AMDGPUDAGToDAGISel::SelectDS1Addr1Offset(SDValue Addr, SDValue &Base,
+ SDValue &Offset) const {
+ if (CurDAG->isBaseWithConstantOffset(Addr)) {
+ SDValue N0 = Addr.getOperand(0);
+ SDValue N1 = Addr.getOperand(1);
+ ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
+ if (isDSOffsetLegal(N0, C1->getSExtValue(), 16)) {
+ // (add n0, c0)
+ Base = N0;
+ Offset = N1;
+ return true;
+ }
+ }
+
+ // 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);
+ return true;
+}
+
+bool AMDGPUDAGToDAGISel::SelectDS64Bit4ByteAligned(SDValue Addr, SDValue &Base,
+ SDValue &Offset0,
+ SDValue &Offset1) const {
+ if (CurDAG->isBaseWithConstantOffset(Addr)) {
+ SDValue N0 = Addr.getOperand(0);
+ SDValue N1 = Addr.getOperand(1);
+ ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
+ unsigned DWordOffset0 = C1->getZExtValue() / 4;
+ unsigned DWordOffset1 = DWordOffset0 + 1;
+ // (add n0, c0)
+ if (isDSOffsetLegal(N0, DWordOffset1, 8)) {
+ Base = N0;
+ Offset0 = CurDAG->getTargetConstant(DWordOffset0, MVT::i8);
+ Offset1 = CurDAG->getTargetConstant(DWordOffset1, MVT::i8);
+ return true;
+ }
+ }
+
+ 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);
+ Offset1 = CurDAG->getTargetConstant(1, MVT::i8);
+ return true;
+}
+
+static bool isLegalMUBUFImmOffset(const ConstantSDNode *Imm) {
+ return isUInt<12>(Imm->getZExtValue());
+}
+
+void AMDGPUDAGToDAGISel::SelectMUBUF(SDValue Addr, SDValue &Ptr,
+ SDValue &VAddr, SDValue &SOffset,
+ SDValue &Offset, SDValue &Offen,
+ SDValue &Idxen, SDValue &Addr64,
+ SDValue &GLC, SDValue &SLC,
+ SDValue &TFE) const {
+ SDLoc DL(Addr);
+
+ GLC = CurDAG->getTargetConstant(0, MVT::i1);
+ SLC = CurDAG->getTargetConstant(0, MVT::i1);
+ TFE = CurDAG->getTargetConstant(0, MVT::i1);
+
+ Idxen = CurDAG->getTargetConstant(0, MVT::i1);
+ Offen = CurDAG->getTargetConstant(0, MVT::i1);
+ Addr64 = CurDAG->getTargetConstant(0, MVT::i1);
+ SOffset = CurDAG->getTargetConstant(0, MVT::i32);
+
+ if (CurDAG->isBaseWithConstantOffset(Addr)) {
+ SDValue N0 = Addr.getOperand(0);
+ SDValue N1 = Addr.getOperand(1);
+ ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
+
+ if (isLegalMUBUFImmOffset(C1)) {
+
+ if (N0.getOpcode() == ISD::ADD) {
+ // (add (add N2, N3), C1) -> addr64
+ SDValue N2 = N0.getOperand(0);
+ SDValue N3 = N0.getOperand(1);
+ Addr64 = CurDAG->getTargetConstant(1, MVT::i1);
+ Ptr = N2;
+ VAddr = N3;
+ Offset = CurDAG->getTargetConstant(C1->getZExtValue(), MVT::i16);
+ return;
+ }
+
+ // (add N0, C1) -> offset
+ VAddr = CurDAG->getTargetConstant(0, MVT::i32);
+ Ptr = N0;
+ Offset = CurDAG->getTargetConstant(C1->getZExtValue(), MVT::i16);
+ return;
+ }
+ }
+ if (Addr.getOpcode() == ISD::ADD) {
+ // (add N0, N1) -> addr64
+ SDValue N0 = Addr.getOperand(0);
+ SDValue N1 = Addr.getOperand(1);
+ Addr64 = CurDAG->getTargetConstant(1, MVT::i1);
+ Ptr = N0;
+ VAddr = N1;
+ Offset = CurDAG->getTargetConstant(0, MVT::i16);
+ return;
+ }
+
+ // default case -> offset
+ VAddr = CurDAG->getTargetConstant(0, MVT::i32);
+ Ptr = Addr;
+ Offset = CurDAG->getTargetConstant(0, MVT::i16);
+
}
-bool AMDGPUDAGToDAGISel::SelectI24(SDValue Op, SDValue &I24) {
+bool AMDGPUDAGToDAGISel::SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc,
+ SDValue &VAddr,
+ SDValue &Offset) const {
+ SDValue Ptr, SOffset, Offen, Idxen, Addr64, GLC, SLC, TFE;
- assert(Op.getValueType() == MVT::i32);
+ SelectMUBUF(Addr, Ptr, VAddr, SOffset, Offset, Offen, Idxen, Addr64,
+ GLC, SLC, TFE);
- if (CurDAG->ComputeNumSignBits(Op) == 9) {
- I24 = SimplifyI24(Op);
+ ConstantSDNode *C = cast<ConstantSDNode>(Addr64);
+ if (C->getSExtValue()) {
+ SDLoc DL(Addr);
+
+ const SITargetLowering& Lowering =
+ *static_cast<const SITargetLowering*>(getTargetLowering());
+
+ SRsrc = SDValue(Lowering.wrapAddr64Rsrc(*CurDAG, DL, Ptr), 0);
return true;
}
+
return false;
}
-bool AMDGPUDAGToDAGISel::SelectU24(SDValue Op, SDValue &U24) {
- APInt KnownZero;
- APInt KnownOne;
- CurDAG->ComputeMaskedBits(Op, KnownZero, KnownOne);
+bool AMDGPUDAGToDAGISel::SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc,
+ SDValue &VAddr, SDValue &Offset,
+ SDValue &SLC) const {
+ SLC = CurDAG->getTargetConstant(0, MVT::i1);
+
+ return SelectMUBUFAddr64(Addr, SRsrc, VAddr, Offset);
+}
+
+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 *>(Subtarget->getRegisterInfo());
+ MachineRegisterInfo &MRI = MF.getRegInfo();
+ const SITargetLowering& Lowering =
+ *static_cast<const SITargetLowering*>(getTargetLowering());
+
+ 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);
+
+ 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 n0, c1)
+ if (CurDAG->isBaseWithConstantOffset(Addr)) {
+ SDValue N1 = Addr.getOperand(1);
+ ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
+
+ if (isLegalMUBUFImmOffset(C1)) {
+ VAddr = Addr.getOperand(0);
+ ImmOffset = CurDAG->getTargetConstant(C1->getZExtValue(), MVT::i16);
+ return true;
+ }
+ }
+
+ // (node)
+ VAddr = Addr;
+ ImmOffset = CurDAG->getTargetConstant(0, MVT::i16);
+ return true;
+}
+
+bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc,
+ SDValue &SOffset, SDValue &Offset,
+ SDValue &GLC, SDValue &SLC,
+ SDValue &TFE) const {
+ SDValue Ptr, VAddr, Offen, Idxen, Addr64;
+ const SIInstrInfo *TII =
+ static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
- assert (Op.getValueType() == MVT::i32);
+ SelectMUBUF(Addr, Ptr, VAddr, SOffset, Offset, Offen, Idxen, Addr64,
+ GLC, SLC, TFE);
- // ANY_EXTEND and EXTLOAD operations can only be done on types smaller than
- // i32. These smaller types are legal to use with the i24 instructions.
- if ((KnownZero & APInt(KnownZero.getBitWidth(), 0xFF000000)) == 0xFF000000 ||
- Op.getOpcode() == ISD::ANY_EXTEND ||
- ISD::isEXTLoad(Op.getNode())) {
- U24 = SimplifyI24(Op);
+ if (!cast<ConstantSDNode>(Offen)->getSExtValue() &&
+ !cast<ConstantSDNode>(Idxen)->getSExtValue() &&
+ !cast<ConstantSDNode>(Addr64)->getSExtValue()) {
+ uint64_t Rsrc = TII->getDefaultRsrcDataFormat() |
+ APInt::getAllOnesValue(32).getZExtValue(); // Size
+ SDLoc DL(Addr);
+
+ const SITargetLowering& Lowering =
+ *static_cast<const SITargetLowering*>(getTargetLowering());
+
+ SRsrc = SDValue(Lowering.buildRSRC(*CurDAG, DL, Ptr, 0, Rsrc), 0);
return true;
}
return false;
}
+bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc,
+ SDValue &Soffset, SDValue &Offset,
+ SDValue &GLC) const {
+ SDValue SLC, TFE;
+
+ return SelectMUBUFOffset(Addr, SRsrc, Soffset, Offset, GLC, SLC, TFE);
+}
+
+// FIXME: This is incorrect and only enough to be able to compile.
+SDNode *AMDGPUDAGToDAGISel::SelectAddrSpaceCast(SDNode *N) {
+ AddrSpaceCastSDNode *ASC = cast<AddrSpaceCastSDNode>(N);
+ SDLoc DL(N);
+
+ assert(Subtarget->hasFlatAddressSpace() &&
+ "addrspacecast only supported with flat address space!");
+
+ assert((ASC->getSrcAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS &&
+ ASC->getDestAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS) &&
+ "Cannot cast address space to / from constant address!");
+
+ assert((ASC->getSrcAddressSpace() == AMDGPUAS::FLAT_ADDRESS ||
+ ASC->getDestAddressSpace() == AMDGPUAS::FLAT_ADDRESS) &&
+ "Can only cast to / from flat address space!");
+
+ // The flat instructions read the address as the index of the VGPR holding the
+ // address, so casting should just be reinterpreting the base VGPR, so just
+ // insert trunc / bitcast / zext.
+
+ SDValue Src = ASC->getOperand(0);
+ EVT DestVT = ASC->getValueType(0);
+ EVT SrcVT = Src.getValueType();
+
+ unsigned SrcSize = SrcVT.getSizeInBits();
+ unsigned DestSize = DestVT.getSizeInBits();
+
+ if (SrcSize > DestSize) {
+ assert(SrcSize == 64 && DestSize == 32);
+ return CurDAG->getMachineNode(
+ TargetOpcode::EXTRACT_SUBREG,
+ DL,
+ DestVT,
+ Src,
+ CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32));
+ }
+
+
+ if (DestSize > SrcSize) {
+ assert(SrcSize == 32 && DestSize == 64);
+
+ // 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,
+ Src,
+ CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32),
+ SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, SDLoc(N), MVT::i32,
+ CurDAG->getConstant(0, MVT::i32)), 0),
+ CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32)
+ };
+
+ return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE,
+ SDLoc(N), N->getValueType(0), Ops);
+ }
+
+ assert(SrcSize == 64 && DestSize == 64);
+ return CurDAG->getNode(ISD::BITCAST, DL, DestVT, Src).getNode();
+}
+
+bool AMDGPUDAGToDAGISel::SelectVOP3Mods(SDValue In, SDValue &Src,
+ SDValue &SrcMods) const {
+
+ unsigned Mods = 0;
+
+ Src = In;
+
+ if (Src.getOpcode() == ISD::FNEG) {
+ Mods |= SISrcMods::NEG;
+ Src = Src.getOperand(0);
+ }
+
+ if (Src.getOpcode() == ISD::FABS) {
+ Mods |= SISrcMods::ABS;
+ Src = Src.getOperand(0);
+ }
+
+ SrcMods = CurDAG->getTargetConstant(Mods, MVT::i32);
+
+ return true;
+}
+
+bool AMDGPUDAGToDAGISel::SelectVOP3Mods0(SDValue In, SDValue &Src,
+ SDValue &SrcMods, SDValue &Clamp,
+ SDValue &Omod) const {
+ // FIXME: Handle Clamp and Omod
+ Clamp = CurDAG->getTargetConstant(0, MVT::i32);
+ Omod = CurDAG->getTargetConstant(0, MVT::i32);
+
+ 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 =
- (*(const AMDGPUTargetLowering*)getTargetLowering());
+ *static_cast<const AMDGPUTargetLowering*>(getTargetLowering());
bool IsModified = false;
do {
IsModified = false;