//
//===----------------------------------------------------------------------===//
+#define DEBUG_TYPE "arm-isel"
#include "ARM.h"
#include "ARMAddressingModes.h"
-#include "ARMConstantPoolValue.h"
-#include "ARMISelLowering.h"
#include "ARMTargetMachine.h"
#include "llvm/CallingConv.h"
#include "llvm/Constants.h"
#include "llvm/CodeGen/SelectionDAGISel.h"
#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/TargetOptions.h"
+#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
return CurDAG->getTargetConstant(Imm, MVT::i32);
}
- SDNode *Select(SDValue Op);
- virtual void InstructionSelect();
- bool SelectShifterOperandReg(SDValue Op, SDValue N, SDValue &A,
+ SDNode *Select(SDNode *N);
+
+ bool SelectShifterOperandReg(SDNode *Op, SDValue N, SDValue &A,
SDValue &B, SDValue &C);
- bool SelectAddrMode2(SDValue Op, SDValue N, SDValue &Base,
+ bool SelectAddrMode2(SDNode *Op, SDValue N, SDValue &Base,
SDValue &Offset, SDValue &Opc);
- bool SelectAddrMode2Offset(SDValue Op, SDValue N,
+ bool SelectAddrMode2Offset(SDNode *Op, SDValue N,
SDValue &Offset, SDValue &Opc);
- bool SelectAddrMode3(SDValue Op, SDValue N, SDValue &Base,
+ bool SelectAddrMode3(SDNode *Op, SDValue N, SDValue &Base,
SDValue &Offset, SDValue &Opc);
- bool SelectAddrMode3Offset(SDValue Op, SDValue N,
+ bool SelectAddrMode3Offset(SDNode *Op, SDValue N,
SDValue &Offset, SDValue &Opc);
- bool SelectAddrMode4(SDValue Op, SDValue N, SDValue &Addr,
+ bool SelectAddrMode4(SDNode *Op, SDValue N, SDValue &Addr,
SDValue &Mode);
- bool SelectAddrMode5(SDValue Op, SDValue N, SDValue &Base,
+ bool SelectAddrMode5(SDNode *Op, SDValue N, SDValue &Base,
SDValue &Offset);
- bool SelectAddrMode6(SDValue Op, SDValue N, SDValue &Addr, SDValue &Update,
- SDValue &Opc);
+ bool SelectAddrMode6(SDNode *Op, SDValue N, SDValue &Addr, SDValue &Align);
- bool SelectAddrModePC(SDValue Op, SDValue N, SDValue &Offset,
+ bool SelectAddrModePC(SDNode *Op, SDValue N, SDValue &Offset,
SDValue &Label);
- bool SelectThumbAddrModeRR(SDValue Op, SDValue N, SDValue &Base,
+ bool SelectThumbAddrModeRR(SDNode *Op, SDValue N, SDValue &Base,
SDValue &Offset);
- bool SelectThumbAddrModeRI5(SDValue Op, SDValue N, unsigned Scale,
+ bool SelectThumbAddrModeRI5(SDNode *Op, SDValue N, unsigned Scale,
SDValue &Base, SDValue &OffImm,
SDValue &Offset);
- bool SelectThumbAddrModeS1(SDValue Op, SDValue N, SDValue &Base,
+ bool SelectThumbAddrModeS1(SDNode *Op, SDValue N, SDValue &Base,
SDValue &OffImm, SDValue &Offset);
- bool SelectThumbAddrModeS2(SDValue Op, SDValue N, SDValue &Base,
+ bool SelectThumbAddrModeS2(SDNode *Op, SDValue N, SDValue &Base,
SDValue &OffImm, SDValue &Offset);
- bool SelectThumbAddrModeS4(SDValue Op, SDValue N, SDValue &Base,
+ bool SelectThumbAddrModeS4(SDNode *Op, SDValue N, SDValue &Base,
SDValue &OffImm, SDValue &Offset);
- bool SelectThumbAddrModeSP(SDValue Op, SDValue N, SDValue &Base,
+ bool SelectThumbAddrModeSP(SDNode *Op, SDValue N, SDValue &Base,
SDValue &OffImm);
- bool SelectT2ShifterOperandReg(SDValue Op, SDValue N,
+ bool SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
SDValue &BaseReg, SDValue &Opc);
- bool SelectT2AddrModeImm12(SDValue Op, SDValue N, SDValue &Base,
+ bool SelectT2AddrModeImm12(SDNode *Op, SDValue N, SDValue &Base,
SDValue &OffImm);
- bool SelectT2AddrModeImm8(SDValue Op, SDValue N, SDValue &Base,
+ bool SelectT2AddrModeImm8(SDNode *Op, SDValue N, SDValue &Base,
SDValue &OffImm);
- bool SelectT2AddrModeImm8Offset(SDValue Op, SDValue N,
+ bool SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
SDValue &OffImm);
- bool SelectT2AddrModeImm8s4(SDValue Op, SDValue N, SDValue &Base,
+ bool SelectT2AddrModeImm8s4(SDNode *Op, SDValue N, SDValue &Base,
SDValue &OffImm);
- bool SelectT2AddrModeSoReg(SDValue Op, SDValue N, SDValue &Base,
+ bool SelectT2AddrModeSoReg(SDNode *Op, SDValue N, SDValue &Base,
SDValue &OffReg, SDValue &ShImm);
// Include the pieces autogenerated from the target description.
private:
/// SelectARMIndexedLoad - Indexed (pre/post inc/dec) load matching code for
/// ARM.
- SDNode *SelectARMIndexedLoad(SDValue Op);
- SDNode *SelectT2IndexedLoad(SDValue Op);
-
- /// SelectDYN_ALLOC - Select dynamic alloc for Thumb.
- SDNode *SelectDYN_ALLOC(SDValue Op);
+ SDNode *SelectARMIndexedLoad(SDNode *N);
+ SDNode *SelectT2IndexedLoad(SDNode *N);
+
+ /// SelectVLD - Select NEON load intrinsics. NumVecs should be
+ /// 1, 2, 3 or 4. The opcode arrays specify the instructions used for
+ /// loads of D registers and even subregs and odd subregs of Q registers.
+ /// For NumVecs <= 2, QOpcodes1 is not used.
+ SDNode *SelectVLD(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
+ unsigned *QOpcodes0, unsigned *QOpcodes1);
+
+ /// SelectVST - Select NEON store intrinsics. NumVecs should
+ /// be 1, 2, 3 or 4. The opcode arrays specify the instructions used for
+ /// stores of D registers and even subregs and odd subregs of Q registers.
+ /// For NumVecs <= 2, QOpcodes1 is not used.
+ SDNode *SelectVST(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
+ unsigned *QOpcodes0, unsigned *QOpcodes1);
+
+ /// SelectVLDSTLane - Select NEON load/store lane intrinsics. NumVecs should
+ /// be 2, 3 or 4. The opcode arrays specify the instructions used for
+ /// load/store of D registers and even subregs and odd subregs of Q registers.
+ SDNode *SelectVLDSTLane(SDNode *N, bool IsLoad, unsigned NumVecs,
+ unsigned *DOpcodes, unsigned *QOpcodes0,
+ unsigned *QOpcodes1);
+
+ /// SelectV6T2BitfieldExtractOp - Select SBFX/UBFX instructions for ARM.
+ SDNode *SelectV6T2BitfieldExtractOp(SDNode *N, bool isSigned);
+
+ /// SelectCMOVOp - Select CMOV instructions for ARM.
+ SDNode *SelectCMOVOp(SDNode *N);
+ SDNode *SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
+ ARMCC::CondCodes CCVal, SDValue CCR,
+ SDValue InFlag);
+ SDNode *SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
+ ARMCC::CondCodes CCVal, SDValue CCR,
+ SDValue InFlag);
+ SDNode *SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
+ ARMCC::CondCodes CCVal, SDValue CCR,
+ SDValue InFlag);
+ SDNode *SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
+ ARMCC::CondCodes CCVal, SDValue CCR,
+ SDValue InFlag);
+
+ SDNode *SelectConcatVector(SDNode *N);
/// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
/// inline asm expressions.
char ConstraintCode,
std::vector<SDValue> &OutOps);
- /// PairDRegs - Insert a pair of double registers into an implicit def to
- /// form a quad register.
+ // Form pairs of consecutive S, D, or Q registers.
+ SDNode *PairSRegs(EVT VT, SDValue V0, SDValue V1);
SDNode *PairDRegs(EVT VT, SDValue V0, SDValue V1);
+ SDNode *PairQRegs(EVT VT, SDValue V0, SDValue V1);
+
+ // Form sequences of 4 consecutive S, D, or Q registers.
+ SDNode *QuadSRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
+ SDNode *QuadDRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
+ SDNode *QuadQRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
+
+ // Form sequences of 8 consecutive D registers.
+ SDNode *OctoDRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3,
+ SDValue V4, SDValue V5, SDValue V6, SDValue V7);
};
}
-void ARMDAGToDAGISel::InstructionSelect() {
- DEBUG(BB->dump());
+/// isInt32Immediate - This method tests to see if the node is a 32-bit constant
+/// operand. If so Imm will receive the 32-bit value.
+static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
+ if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
+ Imm = cast<ConstantSDNode>(N)->getZExtValue();
+ return true;
+ }
+ return false;
+}
+
+// isInt32Immediate - This method tests to see if a constant operand.
+// If so Imm will receive the 32 bit value.
+static bool isInt32Immediate(SDValue N, unsigned &Imm) {
+ return isInt32Immediate(N.getNode(), Imm);
+}
- SelectRoot(*CurDAG);
- CurDAG->RemoveDeadNodes();
+// isOpcWithIntImmediate - This method tests to see if the node is a specific
+// opcode and that it has a immediate integer right operand.
+// If so Imm will receive the 32 bit value.
+static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
+ return N->getOpcode() == Opc &&
+ isInt32Immediate(N->getOperand(1).getNode(), Imm);
}
-bool ARMDAGToDAGISel::SelectShifterOperandReg(SDValue Op,
+
+bool ARMDAGToDAGISel::SelectShifterOperandReg(SDNode *Op,
SDValue N,
SDValue &BaseReg,
SDValue &ShReg,
return true;
}
-bool ARMDAGToDAGISel::SelectAddrMode2(SDValue Op, SDValue N,
+bool ARMDAGToDAGISel::SelectAddrMode2(SDNode *Op, SDValue N,
SDValue &Base, SDValue &Offset,
SDValue &Opc) {
if (N.getOpcode() == ISD::MUL) {
if (N.getOpcode() == ISD::FrameIndex) {
int FI = cast<FrameIndexSDNode>(N)->getIndex();
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
- } else if (N.getOpcode() == ARMISD::Wrapper) {
+ } else if (N.getOpcode() == ARMISD::Wrapper &&
+ !(Subtarget->useMovt() &&
+ N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
Base = N.getOperand(0);
}
Offset = CurDAG->getRegister(0, MVT::i32);
}
}
- // Otherwise this is R +/- [possibly shifted] R
+ // Otherwise this is R +/- [possibly shifted] R.
ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::ADD ? ARM_AM::add:ARM_AM::sub;
ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(1));
unsigned ShAmt = 0;
return true;
}
-bool ARMDAGToDAGISel::SelectAddrMode2Offset(SDValue Op, SDValue N,
+bool ARMDAGToDAGISel::SelectAddrMode2Offset(SDNode *Op, SDValue N,
SDValue &Offset, SDValue &Opc) {
- unsigned Opcode = Op.getOpcode();
+ unsigned Opcode = Op->getOpcode();
ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
? cast<LoadSDNode>(Op)->getAddressingMode()
: cast<StoreSDNode>(Op)->getAddressingMode();
}
-bool ARMDAGToDAGISel::SelectAddrMode3(SDValue Op, SDValue N,
+bool ARMDAGToDAGISel::SelectAddrMode3(SDNode *Op, SDValue N,
SDValue &Base, SDValue &Offset,
SDValue &Opc) {
if (N.getOpcode() == ISD::SUB) {
return true;
}
-bool ARMDAGToDAGISel::SelectAddrMode3Offset(SDValue Op, SDValue N,
+bool ARMDAGToDAGISel::SelectAddrMode3Offset(SDNode *Op, SDValue N,
SDValue &Offset, SDValue &Opc) {
- unsigned Opcode = Op.getOpcode();
+ unsigned Opcode = Op->getOpcode();
ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
? cast<LoadSDNode>(Op)->getAddressingMode()
: cast<StoreSDNode>(Op)->getAddressingMode();
return true;
}
-bool ARMDAGToDAGISel::SelectAddrMode4(SDValue Op, SDValue N,
+bool ARMDAGToDAGISel::SelectAddrMode4(SDNode *Op, SDValue N,
SDValue &Addr, SDValue &Mode) {
Addr = N;
Mode = CurDAG->getTargetConstant(0, MVT::i32);
return true;
}
-bool ARMDAGToDAGISel::SelectAddrMode5(SDValue Op, SDValue N,
+bool ARMDAGToDAGISel::SelectAddrMode5(SDNode *Op, SDValue N,
SDValue &Base, SDValue &Offset) {
if (N.getOpcode() != ISD::ADD) {
Base = N;
if (N.getOpcode() == ISD::FrameIndex) {
int FI = cast<FrameIndexSDNode>(N)->getIndex();
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
- } else if (N.getOpcode() == ARMISD::Wrapper) {
+ } else if (N.getOpcode() == ARMISD::Wrapper &&
+ !(Subtarget->useMovt() &&
+ N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
Base = N.getOperand(0);
}
Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
return true;
}
-bool ARMDAGToDAGISel::SelectAddrMode6(SDValue Op, SDValue N,
- SDValue &Addr, SDValue &Update,
- SDValue &Opc) {
+bool ARMDAGToDAGISel::SelectAddrMode6(SDNode *Op, SDValue N,
+ SDValue &Addr, SDValue &Align) {
Addr = N;
- // Default to no writeback.
- Update = CurDAG->getRegister(0, MVT::i32);
- Opc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(false), MVT::i32);
+ // Default to no alignment.
+ Align = CurDAG->getTargetConstant(0, MVT::i32);
return true;
}
-bool ARMDAGToDAGISel::SelectAddrModePC(SDValue Op, SDValue N,
+bool ARMDAGToDAGISel::SelectAddrModePC(SDNode *Op, SDValue N,
SDValue &Offset, SDValue &Label) {
if (N.getOpcode() == ARMISD::PIC_ADD && N.hasOneUse()) {
Offset = N.getOperand(0);
return false;
}
-bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDValue Op, SDValue N,
+bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDNode *Op, SDValue N,
SDValue &Base, SDValue &Offset){
// FIXME dl should come from the parent load or store, not the address
- DebugLoc dl = Op.getDebugLoc();
if (N.getOpcode() != ISD::ADD) {
ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N);
- if (!NC || NC->getZExtValue() != 0)
+ if (!NC || !NC->isNullValue())
return false;
Base = Offset = N;
}
bool
-ARMDAGToDAGISel::SelectThumbAddrModeRI5(SDValue Op, SDValue N,
+ARMDAGToDAGISel::SelectThumbAddrModeRI5(SDNode *Op, SDValue N,
unsigned Scale, SDValue &Base,
SDValue &OffImm, SDValue &Offset) {
if (Scale == 4) {
}
if (N.getOpcode() != ISD::ADD) {
- Base = (N.getOpcode() == ARMISD::Wrapper) ? N.getOperand(0) : N;
+ if (N.getOpcode() == ARMISD::Wrapper &&
+ !(Subtarget->useMovt() &&
+ N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
+ Base = N.getOperand(0);
+ } else
+ Base = N;
+
Offset = CurDAG->getRegister(0, MVT::i32);
OffImm = CurDAG->getTargetConstant(0, MVT::i32);
return true;
return true;
}
-bool ARMDAGToDAGISel::SelectThumbAddrModeS1(SDValue Op, SDValue N,
+bool ARMDAGToDAGISel::SelectThumbAddrModeS1(SDNode *Op, SDValue N,
SDValue &Base, SDValue &OffImm,
SDValue &Offset) {
return SelectThumbAddrModeRI5(Op, N, 1, Base, OffImm, Offset);
}
-bool ARMDAGToDAGISel::SelectThumbAddrModeS2(SDValue Op, SDValue N,
+bool ARMDAGToDAGISel::SelectThumbAddrModeS2(SDNode *Op, SDValue N,
SDValue &Base, SDValue &OffImm,
SDValue &Offset) {
return SelectThumbAddrModeRI5(Op, N, 2, Base, OffImm, Offset);
}
-bool ARMDAGToDAGISel::SelectThumbAddrModeS4(SDValue Op, SDValue N,
+bool ARMDAGToDAGISel::SelectThumbAddrModeS4(SDNode *Op, SDValue N,
SDValue &Base, SDValue &OffImm,
SDValue &Offset) {
return SelectThumbAddrModeRI5(Op, N, 4, Base, OffImm, Offset);
}
-bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDValue Op, SDValue N,
+bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDNode *Op, SDValue N,
SDValue &Base, SDValue &OffImm) {
if (N.getOpcode() == ISD::FrameIndex) {
int FI = cast<FrameIndexSDNode>(N)->getIndex();
return false;
}
-bool ARMDAGToDAGISel::SelectT2ShifterOperandReg(SDValue Op, SDValue N,
+bool ARMDAGToDAGISel::SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
SDValue &BaseReg,
SDValue &Opc) {
ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
return false;
}
-bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDValue Op, SDValue N,
+bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDNode *Op, SDValue N,
SDValue &Base, SDValue &OffImm) {
// Match simple R + imm12 operands.
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
OffImm = CurDAG->getTargetConstant(0, MVT::i32);
return true;
- } else if (N.getOpcode() == ARMISD::Wrapper) {
+ } else if (N.getOpcode() == ARMISD::Wrapper &&
+ !(Subtarget->useMovt() &&
+ N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
Base = N.getOperand(0);
if (Base.getOpcode() == ISD::TargetConstantPool)
return false; // We want to select t2LDRpci instead.
return true;
}
-bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDValue Op, SDValue N,
+bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDNode *Op, SDValue N,
SDValue &Base, SDValue &OffImm) {
// Match simple R - imm8 operands.
if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::SUB) {
return false;
}
-bool ARMDAGToDAGISel::SelectT2AddrModeImm8Offset(SDValue Op, SDValue N,
+bool ARMDAGToDAGISel::SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
SDValue &OffImm){
- unsigned Opcode = Op.getOpcode();
+ unsigned Opcode = Op->getOpcode();
ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
? cast<LoadSDNode>(Op)->getAddressingMode()
: cast<StoreSDNode>(Op)->getAddressingMode();
return false;
}
-bool ARMDAGToDAGISel::SelectT2AddrModeImm8s4(SDValue Op, SDValue N,
+bool ARMDAGToDAGISel::SelectT2AddrModeImm8s4(SDNode *Op, SDValue N,
SDValue &Base, SDValue &OffImm) {
if (N.getOpcode() == ISD::ADD) {
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
int RHSC = (int)RHS->getZExtValue();
+ // 8 bits.
if (((RHSC & 0x3) == 0) &&
- ((RHSC >= 0 && RHSC < 0x400) || (RHSC < 0 && RHSC > -0x400))) { // 8 bits.
+ ((RHSC >= 0 && RHSC < 0x400) || (RHSC < 0 && RHSC > -0x400))) {
Base = N.getOperand(0);
OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
return true;
} else if (N.getOpcode() == ISD::SUB) {
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
int RHSC = (int)RHS->getZExtValue();
- if (((RHSC & 0x3) == 0) && (RHSC >= 0 && RHSC < 0x400)) { // 8 bits.
+ // 8 bits.
+ if (((RHSC & 0x3) == 0) && (RHSC >= 0 && RHSC < 0x400)) {
Base = N.getOperand(0);
OffImm = CurDAG->getTargetConstant(-RHSC, MVT::i32);
return true;
return false;
}
-bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDValue Op, SDValue N,
+bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDNode *Op, SDValue N,
SDValue &Base,
SDValue &OffReg, SDValue &ShImm) {
// (R - imm8) should be handled by t2LDRi8. The rest are handled by t2LDRi12.
return CurDAG->getTargetConstant((uint64_t)ARMCC::AL, MVT::i32);
}
-SDNode *ARMDAGToDAGISel::SelectARMIndexedLoad(SDValue Op) {
- LoadSDNode *LD = cast<LoadSDNode>(Op);
+SDNode *ARMDAGToDAGISel::SelectARMIndexedLoad(SDNode *N) {
+ LoadSDNode *LD = cast<LoadSDNode>(N);
ISD::MemIndexedMode AM = LD->getAddressingMode();
if (AM == ISD::UNINDEXED)
return NULL;
unsigned Opcode = 0;
bool Match = false;
if (LoadedVT == MVT::i32 &&
- SelectAddrMode2Offset(Op, LD->getOffset(), Offset, AMOpc)) {
+ SelectAddrMode2Offset(N, LD->getOffset(), Offset, AMOpc)) {
Opcode = isPre ? ARM::LDR_PRE : ARM::LDR_POST;
Match = true;
} else if (LoadedVT == MVT::i16 &&
- SelectAddrMode3Offset(Op, LD->getOffset(), Offset, AMOpc)) {
+ SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
Match = true;
Opcode = (LD->getExtensionType() == ISD::SEXTLOAD)
? (isPre ? ARM::LDRSH_PRE : ARM::LDRSH_POST)
: (isPre ? ARM::LDRH_PRE : ARM::LDRH_POST);
} else if (LoadedVT == MVT::i8 || LoadedVT == MVT::i1) {
if (LD->getExtensionType() == ISD::SEXTLOAD) {
- if (SelectAddrMode3Offset(Op, LD->getOffset(), Offset, AMOpc)) {
+ if (SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
Match = true;
Opcode = isPre ? ARM::LDRSB_PRE : ARM::LDRSB_POST;
}
} else {
- if (SelectAddrMode2Offset(Op, LD->getOffset(), Offset, AMOpc)) {
+ if (SelectAddrMode2Offset(N, LD->getOffset(), Offset, AMOpc)) {
Match = true;
Opcode = isPre ? ARM::LDRB_PRE : ARM::LDRB_POST;
}
SDValue Base = LD->getBasePtr();
SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG),
CurDAG->getRegister(0, MVT::i32), Chain };
- return CurDAG->getMachineNode(Opcode, Op.getDebugLoc(), MVT::i32, MVT::i32,
+ return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
MVT::Other, Ops, 6);
}
return NULL;
}
-SDNode *ARMDAGToDAGISel::SelectT2IndexedLoad(SDValue Op) {
- LoadSDNode *LD = cast<LoadSDNode>(Op);
+SDNode *ARMDAGToDAGISel::SelectT2IndexedLoad(SDNode *N) {
+ LoadSDNode *LD = cast<LoadSDNode>(N);
ISD::MemIndexedMode AM = LD->getAddressingMode();
if (AM == ISD::UNINDEXED)
return NULL;
bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
unsigned Opcode = 0;
bool Match = false;
- if (SelectT2AddrModeImm8Offset(Op, LD->getOffset(), Offset)) {
+ if (SelectT2AddrModeImm8Offset(N, LD->getOffset(), Offset)) {
switch (LoadedVT.getSimpleVT().SimpleTy) {
case MVT::i32:
Opcode = isPre ? ARM::t2LDR_PRE : ARM::t2LDR_POST;
SDValue Base = LD->getBasePtr();
SDValue Ops[]= { Base, Offset, getAL(CurDAG),
CurDAG->getRegister(0, MVT::i32), Chain };
- return CurDAG->getMachineNode(Opcode, Op.getDebugLoc(), MVT::i32, MVT::i32,
+ return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
MVT::Other, Ops, 5);
}
return NULL;
}
-SDNode *ARMDAGToDAGISel::SelectDYN_ALLOC(SDValue Op) {
- SDNode *N = Op.getNode();
+/// PairSRegs - Form a D register from a pair of S registers.
+///
+SDNode *ARMDAGToDAGISel::PairSRegs(EVT VT, SDValue V0, SDValue V1) {
+ DebugLoc dl = V0.getNode()->getDebugLoc();
+ SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, MVT::i32);
+ SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, MVT::i32);
+ const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
+ return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
+}
+
+/// PairDRegs - Form a quad register from a pair of D registers.
+///
+SDNode *ARMDAGToDAGISel::PairDRegs(EVT VT, SDValue V0, SDValue V1) {
+ DebugLoc dl = V0.getNode()->getDebugLoc();
+ SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
+ SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
+ const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
+ return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
+}
+
+/// PairQRegs - Form 4 consecutive D registers from a pair of Q registers.
+///
+SDNode *ARMDAGToDAGISel::PairQRegs(EVT VT, SDValue V0, SDValue V1) {
+ DebugLoc dl = V0.getNode()->getDebugLoc();
+ SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, MVT::i32);
+ SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, MVT::i32);
+ const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
+ return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
+}
+
+/// QuadSRegs - Form 4 consecutive S registers.
+///
+SDNode *ARMDAGToDAGISel::QuadSRegs(EVT VT, SDValue V0, SDValue V1,
+ SDValue V2, SDValue V3) {
+ DebugLoc dl = V0.getNode()->getDebugLoc();
+ SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, MVT::i32);
+ SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, MVT::i32);
+ SDValue SubReg2 = CurDAG->getTargetConstant(ARM::ssub_2, MVT::i32);
+ SDValue SubReg3 = CurDAG->getTargetConstant(ARM::ssub_3, MVT::i32);
+ const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
+ return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
+}
+
+/// QuadDRegs - Form 4 consecutive D registers.
+///
+SDNode *ARMDAGToDAGISel::QuadDRegs(EVT VT, SDValue V0, SDValue V1,
+ SDValue V2, SDValue V3) {
+ DebugLoc dl = V0.getNode()->getDebugLoc();
+ SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
+ SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
+ SDValue SubReg2 = CurDAG->getTargetConstant(ARM::dsub_2, MVT::i32);
+ SDValue SubReg3 = CurDAG->getTargetConstant(ARM::dsub_3, MVT::i32);
+ const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
+ return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
+}
+
+/// QuadQRegs - Form 4 consecutive Q registers.
+///
+SDNode *ARMDAGToDAGISel::QuadQRegs(EVT VT, SDValue V0, SDValue V1,
+ SDValue V2, SDValue V3) {
+ DebugLoc dl = V0.getNode()->getDebugLoc();
+ SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, MVT::i32);
+ SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, MVT::i32);
+ SDValue SubReg2 = CurDAG->getTargetConstant(ARM::qsub_2, MVT::i32);
+ SDValue SubReg3 = CurDAG->getTargetConstant(ARM::qsub_3, MVT::i32);
+ const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
+ return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
+}
+
+/// OctoDRegs - Form 8 consecutive D registers.
+///
+SDNode *ARMDAGToDAGISel::OctoDRegs(EVT VT, SDValue V0, SDValue V1,
+ SDValue V2, SDValue V3,
+ SDValue V4, SDValue V5,
+ SDValue V6, SDValue V7) {
+ DebugLoc dl = V0.getNode()->getDebugLoc();
+ SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
+ SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
+ SDValue SubReg2 = CurDAG->getTargetConstant(ARM::dsub_2, MVT::i32);
+ SDValue SubReg3 = CurDAG->getTargetConstant(ARM::dsub_3, MVT::i32);
+ SDValue SubReg4 = CurDAG->getTargetConstant(ARM::dsub_4, MVT::i32);
+ SDValue SubReg5 = CurDAG->getTargetConstant(ARM::dsub_5, MVT::i32);
+ SDValue SubReg6 = CurDAG->getTargetConstant(ARM::dsub_6, MVT::i32);
+ SDValue SubReg7 = CurDAG->getTargetConstant(ARM::dsub_7, MVT::i32);
+ const SDValue Ops[] ={ V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3,
+ V4, SubReg4, V5, SubReg5, V6, SubReg6, V7, SubReg7 };
+ return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 16);
+}
+
+/// GetNEONSubregVT - Given a type for a 128-bit NEON vector, return the type
+/// for a 64-bit subregister of the vector.
+static EVT GetNEONSubregVT(EVT VT) {
+ switch (VT.getSimpleVT().SimpleTy) {
+ default: llvm_unreachable("unhandled NEON type");
+ case MVT::v16i8: return MVT::v8i8;
+ case MVT::v8i16: return MVT::v4i16;
+ case MVT::v4f32: return MVT::v2f32;
+ case MVT::v4i32: return MVT::v2i32;
+ case MVT::v2i64: return MVT::v1i64;
+ }
+}
+
+SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, unsigned NumVecs,
+ unsigned *DOpcodes, unsigned *QOpcodes0,
+ unsigned *QOpcodes1) {
+ assert(NumVecs >= 1 && NumVecs <= 4 && "VLD NumVecs out-of-range");
+ DebugLoc dl = N->getDebugLoc();
+
+ SDValue MemAddr, Align;
+ if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
+ return NULL;
+
+ SDValue Chain = N->getOperand(0);
+ EVT VT = N->getValueType(0);
+ bool is64BitVector = VT.is64BitVector();
+
+ unsigned OpcodeIndex;
+ switch (VT.getSimpleVT().SimpleTy) {
+ default: llvm_unreachable("unhandled vld type");
+ // Double-register operations:
+ case MVT::v8i8: OpcodeIndex = 0; break;
+ case MVT::v4i16: OpcodeIndex = 1; break;
+ case MVT::v2f32:
+ case MVT::v2i32: OpcodeIndex = 2; break;
+ case MVT::v1i64: OpcodeIndex = 3; break;
+ // Quad-register operations:
+ case MVT::v16i8: OpcodeIndex = 0; break;
+ case MVT::v8i16: OpcodeIndex = 1; break;
+ case MVT::v4f32:
+ case MVT::v4i32: OpcodeIndex = 2; break;
+ case MVT::v2i64: OpcodeIndex = 3;
+ assert(NumVecs == 1 && "v2i64 type only supported for VLD1");
+ break;
+ }
+
+ SDValue Pred = getAL(CurDAG);
+ SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
+ if (is64BitVector) {
+ unsigned Opc = DOpcodes[OpcodeIndex];
+ const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain };
+ std::vector<EVT> ResTys(NumVecs, VT);
+ ResTys.push_back(MVT::Other);
+ SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
+ if (NumVecs < 2)
+ return VLd;
+
+ SDValue RegSeq;
+ SDValue V0 = SDValue(VLd, 0);
+ SDValue V1 = SDValue(VLd, 1);
+
+ // Form a REG_SEQUENCE to force register allocation.
+ if (NumVecs == 2)
+ RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
+ else {
+ SDValue V2 = SDValue(VLd, 2);
+ // If it's a vld3, form a quad D-register but discard the last part.
+ SDValue V3 = (NumVecs == 3)
+ ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
+ : SDValue(VLd, 3);
+ RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
+ }
+
+ assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering");
+ for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
+ SDValue D = CurDAG->getTargetExtractSubreg(ARM::dsub_0+Vec,
+ dl, VT, RegSeq);
+ ReplaceUses(SDValue(N, Vec), D);
+ }
+ ReplaceUses(SDValue(N, NumVecs), SDValue(VLd, NumVecs));
+ return NULL;
+ }
+
+ EVT RegVT = GetNEONSubregVT(VT);
+ if (NumVecs <= 2) {
+ // Quad registers are directly supported for VLD1 and VLD2,
+ // loading pairs of D regs.
+ unsigned Opc = QOpcodes0[OpcodeIndex];
+ const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain };
+ std::vector<EVT> ResTys(2 * NumVecs, RegVT);
+ ResTys.push_back(MVT::Other);
+ SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
+ Chain = SDValue(VLd, 2 * NumVecs);
+
+ // Combine the even and odd subregs to produce the result.
+ if (NumVecs == 1) {
+ SDNode *Q = PairDRegs(VT, SDValue(VLd, 0), SDValue(VLd, 1));
+ ReplaceUses(SDValue(N, 0), SDValue(Q, 0));
+ } else {
+ SDValue QQ = SDValue(QuadDRegs(MVT::v4i64,
+ SDValue(VLd, 0), SDValue(VLd, 1),
+ SDValue(VLd, 2), SDValue(VLd, 3)), 0);
+ SDValue Q0 = CurDAG->getTargetExtractSubreg(ARM::qsub_0, dl, VT, QQ);
+ SDValue Q1 = CurDAG->getTargetExtractSubreg(ARM::qsub_1, dl, VT, QQ);
+ ReplaceUses(SDValue(N, 0), Q0);
+ ReplaceUses(SDValue(N, 1), Q1);
+ }
+ } else {
+ // Otherwise, quad registers are loaded with two separate instructions,
+ // where one loads the even registers and the other loads the odd registers.
+
+ std::vector<EVT> ResTys(NumVecs, RegVT);
+ ResTys.push_back(MemAddr.getValueType());
+ ResTys.push_back(MVT::Other);
+
+ // Load the even subregs.
+ unsigned Opc = QOpcodes0[OpcodeIndex];
+ const SDValue OpsA[] = { MemAddr, Align, Reg0, Pred, Reg0, Chain };
+ SDNode *VLdA = CurDAG->getMachineNode(Opc, dl, ResTys, OpsA, 6);
+ Chain = SDValue(VLdA, NumVecs+1);
+
+ // Load the odd subregs.
+ Opc = QOpcodes1[OpcodeIndex];
+ const SDValue OpsB[] = { SDValue(VLdA, NumVecs),
+ Align, Reg0, Pred, Reg0, Chain };
+ SDNode *VLdB = CurDAG->getMachineNode(Opc, dl, ResTys, OpsB, 6);
+ Chain = SDValue(VLdB, NumVecs+1);
+
+ SDValue V0 = SDValue(VLdA, 0);
+ SDValue V1 = SDValue(VLdB, 0);
+ SDValue V2 = SDValue(VLdA, 1);
+ SDValue V3 = SDValue(VLdB, 1);
+ SDValue V4 = SDValue(VLdA, 2);
+ SDValue V5 = SDValue(VLdB, 2);
+ SDValue V6 = (NumVecs == 3)
+ ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,RegVT), 0)
+ : SDValue(VLdA, 3);
+ SDValue V7 = (NumVecs == 3)
+ ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,RegVT), 0)
+ : SDValue(VLdB, 3);
+ SDValue RegSeq = SDValue(OctoDRegs(MVT::v8i64, V0, V1, V2, V3,
+ V4, V5, V6, V7), 0);
+
+ // Extract out the 3 / 4 Q registers.
+ assert(ARM::qsub_3 == ARM::qsub_0+3 && "Unexpected subreg numbering");
+ for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
+ SDValue Q = CurDAG->getTargetExtractSubreg(ARM::qsub_0+Vec,
+ dl, VT, RegSeq);
+ ReplaceUses(SDValue(N, Vec), Q);
+ }
+ }
+ ReplaceUses(SDValue(N, NumVecs), Chain);
+ return NULL;
+}
+
+SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, unsigned NumVecs,
+ unsigned *DOpcodes, unsigned *QOpcodes0,
+ unsigned *QOpcodes1) {
+ assert(NumVecs >=1 && NumVecs <= 4 && "VST NumVecs out-of-range");
+ DebugLoc dl = N->getDebugLoc();
+
+ SDValue MemAddr, Align;
+ if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
+ return NULL;
+
+ SDValue Chain = N->getOperand(0);
+ EVT VT = N->getOperand(3).getValueType();
+ bool is64BitVector = VT.is64BitVector();
+
+ unsigned OpcodeIndex;
+ switch (VT.getSimpleVT().SimpleTy) {
+ default: llvm_unreachable("unhandled vst type");
+ // Double-register operations:
+ case MVT::v8i8: OpcodeIndex = 0; break;
+ case MVT::v4i16: OpcodeIndex = 1; break;
+ case MVT::v2f32:
+ case MVT::v2i32: OpcodeIndex = 2; break;
+ case MVT::v1i64: OpcodeIndex = 3; break;
+ // Quad-register operations:
+ case MVT::v16i8: OpcodeIndex = 0; break;
+ case MVT::v8i16: OpcodeIndex = 1; break;
+ case MVT::v4f32:
+ case MVT::v4i32: OpcodeIndex = 2; break;
+ case MVT::v2i64: OpcodeIndex = 3;
+ assert(NumVecs == 1 && "v2i64 type only supported for VST1");
+ break;
+ }
+
+ SDValue Pred = getAL(CurDAG);
+ SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
+
+ SmallVector<SDValue, 10> Ops;
+ Ops.push_back(MemAddr);
+ Ops.push_back(Align);
+
+ if (is64BitVector) {
+ if (NumVecs >= 2) {
+ SDValue RegSeq;
+ SDValue V0 = N->getOperand(0+3);
+ SDValue V1 = N->getOperand(1+3);
+
+ // Form a REG_SEQUENCE to force register allocation.
+ if (NumVecs == 2)
+ RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
+ else {
+ SDValue V2 = N->getOperand(2+3);
+ // If it's a vld3, form a quad D-register and leave the last part as
+ // an undef.
+ SDValue V3 = (NumVecs == 3)
+ ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
+ : N->getOperand(3+3);
+ RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
+ }
+
+ // Now extract the D registers back out.
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, VT,
+ RegSeq));
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, VT,
+ RegSeq));
+ if (NumVecs > 2)
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_2, dl, VT,
+ RegSeq));
+ if (NumVecs > 3)
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_3, dl, VT,
+ RegSeq));
+ } else {
+ for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
+ Ops.push_back(N->getOperand(Vec+3));
+ }
+ Ops.push_back(Pred);
+ Ops.push_back(Reg0); // predicate register
+ Ops.push_back(Chain);
+ unsigned Opc = DOpcodes[OpcodeIndex];
+ return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+5);
+ }
+
+ EVT RegVT = GetNEONSubregVT(VT);
+ if (NumVecs <= 2) {
+ // Quad registers are directly supported for VST1 and VST2,
+ // storing pairs of D regs.
+ unsigned Opc = QOpcodes0[OpcodeIndex];
+ if (NumVecs == 2) {
+ // First extract the pair of Q registers.
+ SDValue Q0 = N->getOperand(3);
+ SDValue Q1 = N->getOperand(4);
+
+ // Form a QQ register.
+ SDValue QQ = SDValue(PairQRegs(MVT::v4i64, Q0, Q1), 0);
+
+ // Now extract the D registers back out.
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, RegVT,
+ QQ));
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, RegVT,
+ QQ));
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_2, dl, RegVT,
+ QQ));
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_3, dl, RegVT,
+ QQ));
+ Ops.push_back(Pred);
+ Ops.push_back(Reg0); // predicate register
+ Ops.push_back(Chain);
+ return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), 5 + 4);
+ } else {
+ for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, RegVT,
+ N->getOperand(Vec+3)));
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, RegVT,
+ N->getOperand(Vec+3)));
+ }
+ Ops.push_back(Pred);
+ Ops.push_back(Reg0); // predicate register
+ Ops.push_back(Chain);
+ return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(),
+ 5 + 2 * NumVecs);
+ }
+ }
+
+ // Otherwise, quad registers are stored with two separate instructions,
+ // where one stores the even registers and the other stores the odd registers.
+
+ // Form the QQQQ REG_SEQUENCE.
+ SDValue V[8];
+ for (unsigned Vec = 0, i = 0; Vec < NumVecs; ++Vec, i+=2) {
+ V[i] = CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, RegVT,
+ N->getOperand(Vec+3));
+ V[i+1] = CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, RegVT,
+ N->getOperand(Vec+3));
+ }
+ if (NumVecs == 3)
+ V[6] = V[7] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
+ dl, RegVT), 0);
+
+ SDValue RegSeq = SDValue(OctoDRegs(MVT::v8i64, V[0], V[1], V[2], V[3],
+ V[4], V[5], V[6], V[7]), 0);
+
+ // Store the even D registers.
+ assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering");
+ Ops.push_back(Reg0); // post-access address offset
+ for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0+Vec*2, dl,
+ RegVT, RegSeq));
+ Ops.push_back(Pred);
+ Ops.push_back(Reg0); // predicate register
+ Ops.push_back(Chain);
+ unsigned Opc = QOpcodes0[OpcodeIndex];
+ SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
+ MVT::Other, Ops.data(), NumVecs+6);
+ Chain = SDValue(VStA, 1);
+
+ // Store the odd D registers.
+ Ops[0] = SDValue(VStA, 0); // MemAddr
+ for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
+ Ops[Vec+3] = CurDAG->getTargetExtractSubreg(ARM::dsub_1+Vec*2, dl,
+ RegVT, RegSeq);
+ Ops[NumVecs+5] = Chain;
+ Opc = QOpcodes1[OpcodeIndex];
+ SDNode *VStB = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
+ MVT::Other, Ops.data(), NumVecs+6);
+ Chain = SDValue(VStB, 1);
+ ReplaceUses(SDValue(N, 0), Chain);
+ return NULL;
+}
+
+SDNode *ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad,
+ unsigned NumVecs, unsigned *DOpcodes,
+ unsigned *QOpcodes0,
+ unsigned *QOpcodes1) {
+ assert(NumVecs >=2 && NumVecs <= 4 && "VLDSTLane NumVecs out-of-range");
DebugLoc dl = N->getDebugLoc();
- EVT VT = Op.getValueType();
- SDValue Chain = Op.getOperand(0);
- SDValue Size = Op.getOperand(1);
- SDValue Align = Op.getOperand(2);
- SDValue SP = CurDAG->getRegister(ARM::SP, MVT::i32);
- int32_t AlignVal = cast<ConstantSDNode>(Align)->getSExtValue();
- if (AlignVal < 0)
- // We need to align the stack. Use Thumb1 tAND which is the only thumb
- // instruction that can read and write SP. This matches to a pseudo
- // instruction that has a chain to ensure the result is written back to
- // the stack pointer.
- SP = SDValue(CurDAG->getMachineNode(ARM::tANDsp, dl, VT, SP, Align), 0);
-
- bool isC = isa<ConstantSDNode>(Size);
- uint32_t C = isC ? cast<ConstantSDNode>(Size)->getZExtValue() : ~0UL;
- // Handle the most common case for both Thumb1 and Thumb2:
- // tSUBspi - immediate is between 0 ... 508 inclusive.
- if (C <= 508 && ((C & 3) == 0))
- // FIXME: tSUBspi encode scale 4 implicitly.
- return CurDAG->SelectNodeTo(N, ARM::tSUBspi_, VT, MVT::Other, SP,
- CurDAG->getTargetConstant(C/4, MVT::i32),
- Chain);
-
- if (Subtarget->isThumb1Only()) {
- // Use tADDspr since Thumb1 does not have a sub r, sp, r. ARMISelLowering
- // should have negated the size operand already. FIXME: We can't insert
- // new target independent node at this stage so we are forced to negate
- // it earlier. Is there a better solution?
- return CurDAG->SelectNodeTo(N, ARM::tADDspr_, VT, MVT::Other, SP, Size,
- Chain);
- } else if (Subtarget->isThumb2()) {
- if (isC && Predicate_t2_so_imm(Size.getNode())) {
- // t2SUBrSPi
- SDValue Ops[] = { SP, CurDAG->getTargetConstant(C, MVT::i32), Chain };
- return CurDAG->SelectNodeTo(N, ARM::t2SUBrSPi_, VT, MVT::Other, Ops, 3);
- } else if (isC && Predicate_imm0_4095(Size.getNode())) {
- // t2SUBrSPi12
- SDValue Ops[] = { SP, CurDAG->getTargetConstant(C, MVT::i32), Chain };
- return CurDAG->SelectNodeTo(N, ARM::t2SUBrSPi12_, VT, MVT::Other, Ops, 3);
+
+ SDValue MemAddr, Align;
+ if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
+ return NULL;
+
+ SDValue Chain = N->getOperand(0);
+ unsigned Lane =
+ cast<ConstantSDNode>(N->getOperand(NumVecs+3))->getZExtValue();
+ EVT VT = IsLoad ? N->getValueType(0) : N->getOperand(3).getValueType();
+ bool is64BitVector = VT.is64BitVector();
+
+ // Quad registers are handled by load/store of subregs. Find the subreg info.
+ unsigned NumElts = 0;
+ bool Even = false;
+ EVT RegVT = VT;
+ if (!is64BitVector) {
+ RegVT = GetNEONSubregVT(VT);
+ NumElts = RegVT.getVectorNumElements();
+ Even = Lane < NumElts;
+ }
+
+ unsigned OpcodeIndex;
+ switch (VT.getSimpleVT().SimpleTy) {
+ default: llvm_unreachable("unhandled vld/vst lane type");
+ // Double-register operations:
+ case MVT::v8i8: OpcodeIndex = 0; break;
+ case MVT::v4i16: OpcodeIndex = 1; break;
+ case MVT::v2f32:
+ case MVT::v2i32: OpcodeIndex = 2; break;
+ // Quad-register operations:
+ case MVT::v8i16: OpcodeIndex = 0; break;
+ case MVT::v4f32:
+ case MVT::v4i32: OpcodeIndex = 1; break;
+ }
+
+ SDValue Pred = getAL(CurDAG);
+ SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
+
+ SmallVector<SDValue, 10> Ops;
+ Ops.push_back(MemAddr);
+ Ops.push_back(Align);
+
+ unsigned Opc = 0;
+ if (is64BitVector) {
+ Opc = DOpcodes[OpcodeIndex];
+ SDValue RegSeq;
+ SDValue V0 = N->getOperand(0+3);
+ SDValue V1 = N->getOperand(1+3);
+ if (NumVecs == 2) {
+ RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
+ } else {
+ SDValue V2 = N->getOperand(2+3);
+ SDValue V3 = (NumVecs == 3)
+ ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
+ : N->getOperand(3+3);
+ RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
+ }
+
+ // Now extract the D registers back out.
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, VT, RegSeq));
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, VT, RegSeq));
+ if (NumVecs > 2)
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_2, dl, VT,RegSeq));
+ if (NumVecs > 3)
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_3, dl, VT,RegSeq));
+ } else {
+ // Check if this is loading the even or odd subreg of a Q register.
+ if (Lane < NumElts) {
+ Opc = QOpcodes0[OpcodeIndex];
+ } else {
+ Lane -= NumElts;
+ Opc = QOpcodes1[OpcodeIndex];
+ }
+
+ SDValue RegSeq;
+ SDValue V0 = N->getOperand(0+3);
+ SDValue V1 = N->getOperand(1+3);
+ if (NumVecs == 2) {
+ RegSeq = SDValue(PairQRegs(MVT::v4i64, V0, V1), 0);
} else {
- // t2SUBrSPs
- SDValue Ops[] = { SP, Size,
- getI32Imm(ARM_AM::getSORegOpc(ARM_AM::lsl,0)), Chain };
- return CurDAG->SelectNodeTo(N, ARM::t2SUBrSPs_, VT, MVT::Other, Ops, 4);
+ SDValue V2 = N->getOperand(2+3);
+ SDValue V3 = (NumVecs == 3)
+ ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
+ : N->getOperand(3+3);
+ RegSeq = SDValue(QuadQRegs(MVT::v8i64, V0, V1, V2, V3), 0);
}
+
+ // Extract the subregs of the input vector.
+ unsigned SubIdx = Even ? ARM::dsub_0 : ARM::dsub_1;
+ for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
+ Ops.push_back(CurDAG->getTargetExtractSubreg(SubIdx+Vec*2, dl, RegVT,
+ RegSeq));
}
+ Ops.push_back(getI32Imm(Lane));
+ Ops.push_back(Pred);
+ Ops.push_back(Reg0);
+ Ops.push_back(Chain);
+
+ if (!IsLoad)
+ return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+6);
+
+ std::vector<EVT> ResTys(NumVecs, RegVT);
+ ResTys.push_back(MVT::Other);
+ SDNode *VLdLn = CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(),NumVecs+6);
+
+ // Form a REG_SEQUENCE to force register allocation.
+ SDValue RegSeq;
+ if (is64BitVector) {
+ SDValue V0 = SDValue(VLdLn, 0);
+ SDValue V1 = SDValue(VLdLn, 1);
+ if (NumVecs == 2) {
+ RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
+ } else {
+ SDValue V2 = SDValue(VLdLn, 2);
+ // If it's a vld3, form a quad D-register but discard the last part.
+ SDValue V3 = (NumVecs == 3)
+ ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
+ : SDValue(VLdLn, 3);
+ RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
+ }
+ } else {
+ // For 128-bit vectors, take the 64-bit results of the load and insert
+ // them as subregs into the result.
+ SDValue V[8];
+ for (unsigned Vec = 0, i = 0; Vec < NumVecs; ++Vec, i+=2) {
+ if (Even) {
+ V[i] = SDValue(VLdLn, Vec);
+ V[i+1] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
+ dl, RegVT), 0);
+ } else {
+ V[i] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
+ dl, RegVT), 0);
+ V[i+1] = SDValue(VLdLn, Vec);
+ }
+ }
+ if (NumVecs == 3)
+ V[6] = V[7] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
+ dl, RegVT), 0);
- // FIXME: Add ADD / SUB sp instructions for ARM.
+ if (NumVecs == 2)
+ RegSeq = SDValue(QuadDRegs(MVT::v4i64, V[0], V[1], V[2], V[3]), 0);
+ else
+ RegSeq = SDValue(OctoDRegs(MVT::v8i64, V[0], V[1], V[2], V[3],
+ V[4], V[5], V[6], V[7]), 0);
+ }
+
+ assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering");
+ assert(ARM::qsub_3 == ARM::qsub_0+3 && "Unexpected subreg numbering");
+ unsigned SubIdx = is64BitVector ? ARM::dsub_0 : ARM::qsub_0;
+ for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
+ ReplaceUses(SDValue(N, Vec),
+ CurDAG->getTargetExtractSubreg(SubIdx+Vec, dl, VT, RegSeq));
+ ReplaceUses(SDValue(N, NumVecs), SDValue(VLdLn, NumVecs));
+ return NULL;
+}
+
+SDNode *ARMDAGToDAGISel::SelectV6T2BitfieldExtractOp(SDNode *N,
+ bool isSigned) {
+ if (!Subtarget->hasV6T2Ops())
+ return NULL;
+
+ unsigned Opc = isSigned ? (Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX)
+ : (Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX);
+
+
+ // For unsigned extracts, check for a shift right and mask
+ unsigned And_imm = 0;
+ if (N->getOpcode() == ISD::AND) {
+ if (isOpcWithIntImmediate(N, ISD::AND, And_imm)) {
+
+ // The immediate is a mask of the low bits iff imm & (imm+1) == 0
+ if (And_imm & (And_imm + 1))
+ return NULL;
+
+ unsigned Srl_imm = 0;
+ if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SRL,
+ Srl_imm)) {
+ assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
+
+ unsigned Width = CountTrailingOnes_32(And_imm);
+ unsigned LSB = Srl_imm;
+ SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
+ SDValue Ops[] = { N->getOperand(0).getOperand(0),
+ CurDAG->getTargetConstant(LSB, MVT::i32),
+ CurDAG->getTargetConstant(Width, MVT::i32),
+ getAL(CurDAG), Reg0 };
+ return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
+ }
+ }
+ return NULL;
+ }
+
+ // Otherwise, we're looking for a shift of a shift
+ unsigned Shl_imm = 0;
+ if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SHL, Shl_imm)) {
+ assert(Shl_imm > 0 && Shl_imm < 32 && "bad amount in shift node!");
+ unsigned Srl_imm = 0;
+ if (isInt32Immediate(N->getOperand(1), Srl_imm)) {
+ assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
+ unsigned Width = 32 - Srl_imm;
+ int LSB = Srl_imm - Shl_imm;
+ if (LSB < 0)
+ return NULL;
+ SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
+ SDValue Ops[] = { N->getOperand(0).getOperand(0),
+ CurDAG->getTargetConstant(LSB, MVT::i32),
+ CurDAG->getTargetConstant(Width, MVT::i32),
+ getAL(CurDAG), Reg0 };
+ return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
+ }
+ }
+ return NULL;
+}
+
+SDNode *ARMDAGToDAGISel::
+SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
+ ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
+ SDValue CPTmp0;
+ SDValue CPTmp1;
+ if (SelectT2ShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1)) {
+ unsigned SOVal = cast<ConstantSDNode>(CPTmp1)->getZExtValue();
+ unsigned SOShOp = ARM_AM::getSORegShOp(SOVal);
+ unsigned Opc = 0;
+ switch (SOShOp) {
+ case ARM_AM::lsl: Opc = ARM::t2MOVCClsl; break;
+ case ARM_AM::lsr: Opc = ARM::t2MOVCClsr; break;
+ case ARM_AM::asr: Opc = ARM::t2MOVCCasr; break;
+ case ARM_AM::ror: Opc = ARM::t2MOVCCror; break;
+ default:
+ llvm_unreachable("Unknown so_reg opcode!");
+ break;
+ }
+ SDValue SOShImm =
+ CurDAG->getTargetConstant(ARM_AM::getSORegOffset(SOVal), MVT::i32);
+ SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
+ SDValue Ops[] = { FalseVal, CPTmp0, SOShImm, CC, CCR, InFlag };
+ return CurDAG->SelectNodeTo(N, Opc, MVT::i32,Ops, 6);
+ }
return 0;
}
-/// PairDRegs - Insert a pair of double registers into an implicit def to
-/// form a quad register.
-SDNode *ARMDAGToDAGISel::PairDRegs(EVT VT, SDValue V0, SDValue V1) {
- DebugLoc dl = V0.getNode()->getDebugLoc();
- SDValue Undef =
- SDValue(CurDAG->getMachineNode(TargetInstrInfo::IMPLICIT_DEF, dl, VT), 0);
- SDValue SubReg0 = CurDAG->getTargetConstant(ARM::DSUBREG_0, MVT::i32);
- SDValue SubReg1 = CurDAG->getTargetConstant(ARM::DSUBREG_1, MVT::i32);
- SDNode *Pair = CurDAG->getMachineNode(TargetInstrInfo::INSERT_SUBREG, dl,
- VT, Undef, V0, SubReg0);
- return CurDAG->getMachineNode(TargetInstrInfo::INSERT_SUBREG, dl,
- VT, SDValue(Pair, 0), V1, SubReg1);
+SDNode *ARMDAGToDAGISel::
+SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
+ ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
+ SDValue CPTmp0;
+ SDValue CPTmp1;
+ SDValue CPTmp2;
+ if (SelectShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1, CPTmp2)) {
+ SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
+ SDValue Ops[] = { FalseVal, CPTmp0, CPTmp1, CPTmp2, CC, CCR, InFlag };
+ return CurDAG->SelectNodeTo(N, ARM::MOVCCs, MVT::i32, Ops, 7);
+ }
+ return 0;
}
-SDNode *ARMDAGToDAGISel::Select(SDValue Op) {
- SDNode *N = Op.getNode();
+SDNode *ARMDAGToDAGISel::
+SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
+ ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
+ ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
+ if (!T)
+ return 0;
+
+ if (Predicate_t2_so_imm(TrueVal.getNode())) {
+ SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
+ SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
+ SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
+ return CurDAG->SelectNodeTo(N,
+ ARM::t2MOVCCi, MVT::i32, Ops, 5);
+ }
+ return 0;
+}
+
+SDNode *ARMDAGToDAGISel::
+SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
+ ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
+ ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
+ if (!T)
+ return 0;
+
+ if (Predicate_so_imm(TrueVal.getNode())) {
+ SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
+ SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
+ SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
+ return CurDAG->SelectNodeTo(N,
+ ARM::MOVCCi, MVT::i32, Ops, 5);
+ }
+ return 0;
+}
+
+SDNode *ARMDAGToDAGISel::SelectCMOVOp(SDNode *N) {
+ EVT VT = N->getValueType(0);
+ SDValue FalseVal = N->getOperand(0);
+ SDValue TrueVal = N->getOperand(1);
+ SDValue CC = N->getOperand(2);
+ SDValue CCR = N->getOperand(3);
+ SDValue InFlag = N->getOperand(4);
+ assert(CC.getOpcode() == ISD::Constant);
+ assert(CCR.getOpcode() == ISD::Register);
+ ARMCC::CondCodes CCVal =
+ (ARMCC::CondCodes)cast<ConstantSDNode>(CC)->getZExtValue();
+
+ if (!Subtarget->isThumb1Only() && VT == MVT::i32) {
+ // Pattern: (ARMcmov:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
+ // Emits: (MOVCCs:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
+ // Pattern complexity = 18 cost = 1 size = 0
+ SDValue CPTmp0;
+ SDValue CPTmp1;
+ SDValue CPTmp2;
+ if (Subtarget->isThumb()) {
+ SDNode *Res = SelectT2CMOVShiftOp(N, FalseVal, TrueVal,
+ CCVal, CCR, InFlag);
+ if (!Res)
+ Res = SelectT2CMOVShiftOp(N, TrueVal, FalseVal,
+ ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
+ if (Res)
+ return Res;
+ } else {
+ SDNode *Res = SelectARMCMOVShiftOp(N, FalseVal, TrueVal,
+ CCVal, CCR, InFlag);
+ if (!Res)
+ Res = SelectARMCMOVShiftOp(N, TrueVal, FalseVal,
+ ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
+ if (Res)
+ return Res;
+ }
+
+ // Pattern: (ARMcmov:i32 GPR:i32:$false,
+ // (imm:i32)<<P:Predicate_so_imm>>:$true,
+ // (imm:i32):$cc)
+ // Emits: (MOVCCi:i32 GPR:i32:$false,
+ // (so_imm:i32 (imm:i32):$true), (imm:i32):$cc)
+ // Pattern complexity = 10 cost = 1 size = 0
+ if (Subtarget->isThumb()) {
+ SDNode *Res = SelectT2CMOVSoImmOp(N, FalseVal, TrueVal,
+ CCVal, CCR, InFlag);
+ if (!Res)
+ Res = SelectT2CMOVSoImmOp(N, TrueVal, FalseVal,
+ ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
+ if (Res)
+ return Res;
+ } else {
+ SDNode *Res = SelectARMCMOVSoImmOp(N, FalseVal, TrueVal,
+ CCVal, CCR, InFlag);
+ if (!Res)
+ Res = SelectARMCMOVSoImmOp(N, TrueVal, FalseVal,
+ ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
+ if (Res)
+ return Res;
+ }
+ }
+
+ // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
+ // Emits: (MOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
+ // Pattern complexity = 6 cost = 1 size = 0
+ //
+ // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
+ // Emits: (tMOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
+ // Pattern complexity = 6 cost = 11 size = 0
+ //
+ // Also FCPYScc and FCPYDcc.
+ SDValue Tmp2 = CurDAG->getTargetConstant(CCVal, MVT::i32);
+ SDValue Ops[] = { FalseVal, TrueVal, Tmp2, CCR, InFlag };
+ unsigned Opc = 0;
+ switch (VT.getSimpleVT().SimpleTy) {
+ default: assert(false && "Illegal conditional move type!");
+ break;
+ case MVT::i32:
+ Opc = Subtarget->isThumb()
+ ? (Subtarget->hasThumb2() ? ARM::t2MOVCCr : ARM::tMOVCCr_pseudo)
+ : ARM::MOVCCr;
+ break;
+ case MVT::f32:
+ Opc = ARM::VMOVScc;
+ break;
+ case MVT::f64:
+ Opc = ARM::VMOVDcc;
+ break;
+ }
+ return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
+}
+
+SDNode *ARMDAGToDAGISel::SelectConcatVector(SDNode *N) {
+ // The only time a CONCAT_VECTORS operation can have legal types is when
+ // two 64-bit vectors are concatenated to a 128-bit vector.
+ EVT VT = N->getValueType(0);
+ if (!VT.is128BitVector() || N->getNumOperands() != 2)
+ llvm_unreachable("unexpected CONCAT_VECTORS");
+ DebugLoc dl = N->getDebugLoc();
+ SDValue V0 = N->getOperand(0);
+ SDValue V1 = N->getOperand(1);
+ SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
+ SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
+ const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
+ return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
+}
+
+SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
DebugLoc dl = N->getDebugLoc();
if (N->isMachineOpcode())
SDNode *ResNode;
if (Subtarget->isThumb1Only()) {
- SDValue Pred = CurDAG->getTargetConstant(0xEULL, MVT::i32);
+ SDValue Pred = getAL(CurDAG);
SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
SDValue Ops[] = { CPIdx, Pred, PredReg, CurDAG->getEntryNode() };
ResNode = CurDAG->getMachineNode(ARM::tLDRcp, dl, MVT::i32, MVT::Other,
ResNode=CurDAG->getMachineNode(ARM::LDRcp, dl, MVT::i32, MVT::Other,
Ops, 6);
}
- ReplaceUses(Op, SDValue(ResNode, 0));
+ ReplaceUses(SDValue(N, 0), SDValue(ResNode, 0));
return NULL;
}
return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
}
}
- case ARMISD::DYN_ALLOC:
- return SelectDYN_ALLOC(Op);
+ case ISD::SRL:
+ if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false))
+ return I;
+ break;
+ case ISD::SRA:
+ if (SDNode *I = SelectV6T2BitfieldExtractOp(N, true))
+ return I;
+ break;
case ISD::MUL:
if (Subtarget->isThumb1Only())
break;
- if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
unsigned RHSV = C->getZExtValue();
if (!RHSV) break;
if (isPowerOf2_32(RHSV-1)) { // 2^n+1?
unsigned ShImm = Log2_32(RHSV-1);
if (ShImm >= 32)
break;
- SDValue V = Op.getOperand(0);
+ SDValue V = N->getOperand(0);
ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
unsigned ShImm = Log2_32(RHSV+1);
if (ShImm >= 32)
break;
- SDValue V = Op.getOperand(0);
+ SDValue V = N->getOperand(0);
ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
if (Subtarget->isThumb()) {
- SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0 };
- return CurDAG->SelectNodeTo(N, ARM::t2RSBrs, MVT::i32, Ops, 5);
+ SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
+ return CurDAG->SelectNodeTo(N, ARM::t2RSBrs, MVT::i32, Ops, 6);
} else {
SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
return CurDAG->SelectNodeTo(N, ARM::RSBrs, MVT::i32, Ops, 7);
}
}
break;
- case ARMISD::FMRRD:
- return CurDAG->getMachineNode(ARM::FMRRD, dl, MVT::i32, MVT::i32,
- Op.getOperand(0), getAL(CurDAG),
+ case ISD::AND: {
+ // Check for unsigned bitfield extract
+ if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false))
+ return I;
+
+ // (and (or x, c2), c1) and top 16-bits of c1 and c2 match, lower 16-bits
+ // of c1 are 0xffff, and lower 16-bit of c2 are 0. That is, the top 16-bits
+ // are entirely contributed by c2 and lower 16-bits are entirely contributed
+ // by x. That's equal to (or (and x, 0xffff), (and c1, 0xffff0000)).
+ // Select it to: "movt x, ((c1 & 0xffff) >> 16)
+ EVT VT = N->getValueType(0);
+ if (VT != MVT::i32)
+ break;
+ unsigned Opc = (Subtarget->isThumb() && Subtarget->hasThumb2())
+ ? ARM::t2MOVTi16
+ : (Subtarget->hasV6T2Ops() ? ARM::MOVTi16 : 0);
+ if (!Opc)
+ break;
+ SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
+ ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
+ if (!N1C)
+ break;
+ if (N0.getOpcode() == ISD::OR && N0.getNode()->hasOneUse()) {
+ SDValue N2 = N0.getOperand(1);
+ ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
+ if (!N2C)
+ break;
+ unsigned N1CVal = N1C->getZExtValue();
+ unsigned N2CVal = N2C->getZExtValue();
+ if ((N1CVal & 0xffff0000U) == (N2CVal & 0xffff0000U) &&
+ (N1CVal & 0xffffU) == 0xffffU &&
+ (N2CVal & 0xffffU) == 0x0U) {
+ SDValue Imm16 = CurDAG->getTargetConstant((N2CVal & 0xFFFF0000U) >> 16,
+ MVT::i32);
+ SDValue Ops[] = { N0.getOperand(0), Imm16,
+ getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
+ return CurDAG->getMachineNode(Opc, dl, VT, Ops, 4);
+ }
+ }
+ break;
+ }
+ case ARMISD::VMOVRRD:
+ return CurDAG->getMachineNode(ARM::VMOVRRD, dl, MVT::i32, MVT::i32,
+ N->getOperand(0), getAL(CurDAG),
CurDAG->getRegister(0, MVT::i32));
case ISD::UMUL_LOHI: {
if (Subtarget->isThumb1Only())
break;
if (Subtarget->isThumb()) {
- SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1),
+ SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
CurDAG->getRegister(0, MVT::i32) };
- return CurDAG->getMachineNode(ARM::t2UMULL, dl, MVT::i32, MVT::i32, Ops,4);
+ return CurDAG->getMachineNode(ARM::t2UMULL, dl, MVT::i32, MVT::i32,Ops,4);
} else {
- SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1),
+ SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
CurDAG->getRegister(0, MVT::i32) };
return CurDAG->getMachineNode(ARM::UMULL, dl, MVT::i32, MVT::i32, Ops, 5);
if (Subtarget->isThumb1Only())
break;
if (Subtarget->isThumb()) {
- SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1),
+ SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
- return CurDAG->getMachineNode(ARM::t2SMULL, dl, MVT::i32, MVT::i32, Ops,4);
+ return CurDAG->getMachineNode(ARM::t2SMULL, dl, MVT::i32, MVT::i32,Ops,4);
} else {
- SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1),
+ SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
CurDAG->getRegister(0, MVT::i32) };
return CurDAG->getMachineNode(ARM::SMULL, dl, MVT::i32, MVT::i32, Ops, 5);
case ISD::LOAD: {
SDNode *ResNode = 0;
if (Subtarget->isThumb() && Subtarget->hasThumb2())
- ResNode = SelectT2IndexedLoad(Op);
+ ResNode = SelectT2IndexedLoad(N);
else
- ResNode = SelectARMIndexedLoad(Op);
+ ResNode = SelectARMIndexedLoad(N);
if (ResNode)
return ResNode;
+
+ // VLDMQ must be custom-selected for "v2f64 load" to set the AM5Opc value.
+ if (Subtarget->hasVFP2() &&
+ N->getValueType(0).getSimpleVT().SimpleTy == MVT::v2f64) {
+ SDValue Chain = N->getOperand(0);
+ SDValue AM5Opc =
+ CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::ia, 4), MVT::i32);
+ SDValue Pred = getAL(CurDAG);
+ SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
+ SDValue Ops[] = { N->getOperand(1), AM5Opc, Pred, PredReg, Chain };
+ MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
+ MemOp[0] = cast<MemSDNode>(N)->getMemOperand();
+ SDNode *Ret = CurDAG->getMachineNode(ARM::VLDMQ, dl,
+ MVT::v2f64, MVT::Other, Ops, 5);
+ cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
+ return Ret;
+ }
+ // Other cases are autogenerated.
+ break;
+ }
+ case ISD::STORE: {
+ // VSTMQ must be custom-selected for "v2f64 store" to set the AM5Opc value.
+ if (Subtarget->hasVFP2() &&
+ N->getOperand(1).getValueType().getSimpleVT().SimpleTy == MVT::v2f64) {
+ SDValue Chain = N->getOperand(0);
+ SDValue AM5Opc =
+ CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::ia, 4), MVT::i32);
+ SDValue Pred = getAL(CurDAG);
+ SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
+ SDValue Ops[] = { N->getOperand(1), N->getOperand(2),
+ AM5Opc, Pred, PredReg, Chain };
+ MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
+ MemOp[0] = cast<MemSDNode>(N)->getMemOperand();
+ SDNode *Ret = CurDAG->getMachineNode(ARM::VSTMQ, dl, MVT::Other, Ops, 6);
+ cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
+ return Ret;
+ }
// Other cases are autogenerated.
break;
}
unsigned Opc = Subtarget->isThumb() ?
((Subtarget->hasThumb2()) ? ARM::t2Bcc : ARM::tBcc) : ARM::Bcc;
- SDValue Chain = Op.getOperand(0);
- SDValue N1 = Op.getOperand(1);
- SDValue N2 = Op.getOperand(2);
- SDValue N3 = Op.getOperand(3);
- SDValue InFlag = Op.getOperand(4);
+ SDValue Chain = N->getOperand(0);
+ SDValue N1 = N->getOperand(1);
+ SDValue N2 = N->getOperand(2);
+ SDValue N3 = N->getOperand(3);
+ SDValue InFlag = N->getOperand(4);
assert(N1.getOpcode() == ISD::BasicBlock);
assert(N2.getOpcode() == ISD::Constant);
assert(N3.getOpcode() == ISD::Register);
SDNode *ResNode = CurDAG->getMachineNode(Opc, dl, MVT::Other,
MVT::Flag, Ops, 5);
Chain = SDValue(ResNode, 0);
- if (Op.getNode()->getNumValues() == 2) {
+ if (N->getNumValues() == 2) {
InFlag = SDValue(ResNode, 1);
- ReplaceUses(SDValue(Op.getNode(), 1), InFlag);
+ ReplaceUses(SDValue(N, 1), InFlag);
}
- ReplaceUses(SDValue(Op.getNode(), 0), SDValue(Chain.getNode(), Chain.getResNo()));
+ ReplaceUses(SDValue(N, 0),
+ SDValue(Chain.getNode(), Chain.getResNo()));
return NULL;
}
- case ARMISD::CMOV: {
- EVT VT = Op.getValueType();
- SDValue N0 = Op.getOperand(0);
- SDValue N1 = Op.getOperand(1);
- SDValue N2 = Op.getOperand(2);
- SDValue N3 = Op.getOperand(3);
- SDValue InFlag = Op.getOperand(4);
- assert(N2.getOpcode() == ISD::Constant);
- assert(N3.getOpcode() == ISD::Register);
-
- if (!Subtarget->isThumb1Only() && VT == MVT::i32) {
- // Pattern: (ARMcmov:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
- // Emits: (MOVCCs:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
- // Pattern complexity = 18 cost = 1 size = 0
- SDValue CPTmp0;
- SDValue CPTmp1;
- SDValue CPTmp2;
- if (Subtarget->isThumb()) {
- if (SelectT2ShifterOperandReg(Op, N1, CPTmp0, CPTmp1)) {
- unsigned SOVal = cast<ConstantSDNode>(CPTmp1)->getZExtValue();
- unsigned SOShOp = ARM_AM::getSORegShOp(SOVal);
- unsigned Opc = 0;
- switch (SOShOp) {
- case ARM_AM::lsl: Opc = ARM::t2MOVCClsl; break;
- case ARM_AM::lsr: Opc = ARM::t2MOVCClsr; break;
- case ARM_AM::asr: Opc = ARM::t2MOVCCasr; break;
- case ARM_AM::ror: Opc = ARM::t2MOVCCror; break;
- default:
- llvm_unreachable("Unknown so_reg opcode!");
- break;
- }
- SDValue SOShImm =
- CurDAG->getTargetConstant(ARM_AM::getSORegOffset(SOVal), MVT::i32);
- SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
- cast<ConstantSDNode>(N2)->getZExtValue()),
- MVT::i32);
- SDValue Ops[] = { N0, CPTmp0, SOShImm, Tmp2, N3, InFlag };
- return CurDAG->SelectNodeTo(Op.getNode(), Opc, MVT::i32,Ops, 6);
- }
- } else {
- if (SelectShifterOperandReg(Op, N1, CPTmp0, CPTmp1, CPTmp2)) {
- SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
- cast<ConstantSDNode>(N2)->getZExtValue()),
- MVT::i32);
- SDValue Ops[] = { N0, CPTmp0, CPTmp1, CPTmp2, Tmp2, N3, InFlag };
- return CurDAG->SelectNodeTo(Op.getNode(),
- ARM::MOVCCs, MVT::i32, Ops, 7);
- }
- }
-
- // Pattern: (ARMcmov:i32 GPR:i32:$false,
- // (imm:i32)<<P:Predicate_so_imm>>:$true,
- // (imm:i32):$cc)
- // Emits: (MOVCCi:i32 GPR:i32:$false,
- // (so_imm:i32 (imm:i32):$true), (imm:i32):$cc)
- // Pattern complexity = 10 cost = 1 size = 0
- if (N3.getOpcode() == ISD::Constant) {
- if (Subtarget->isThumb()) {
- if (Predicate_t2_so_imm(N3.getNode())) {
- SDValue Tmp1 = CurDAG->getTargetConstant(((unsigned)
- cast<ConstantSDNode>(N1)->getZExtValue()),
- MVT::i32);
- SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
- cast<ConstantSDNode>(N2)->getZExtValue()),
- MVT::i32);
- SDValue Ops[] = { N0, Tmp1, Tmp2, N3, InFlag };
- return CurDAG->SelectNodeTo(Op.getNode(),
- ARM::t2MOVCCi, MVT::i32, Ops, 5);
- }
- } else {
- if (Predicate_so_imm(N3.getNode())) {
- SDValue Tmp1 = CurDAG->getTargetConstant(((unsigned)
- cast<ConstantSDNode>(N1)->getZExtValue()),
- MVT::i32);
- SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
- cast<ConstantSDNode>(N2)->getZExtValue()),
- MVT::i32);
- SDValue Ops[] = { N0, Tmp1, Tmp2, N3, InFlag };
- return CurDAG->SelectNodeTo(Op.getNode(),
- ARM::MOVCCi, MVT::i32, Ops, 5);
- }
- }
- }
- }
-
- // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
- // Emits: (MOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
- // Pattern complexity = 6 cost = 1 size = 0
- //
- // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
- // Emits: (tMOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
- // Pattern complexity = 6 cost = 11 size = 0
- //
- // Also FCPYScc and FCPYDcc.
- SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
- cast<ConstantSDNode>(N2)->getZExtValue()),
- MVT::i32);
- SDValue Ops[] = { N0, N1, Tmp2, N3, InFlag };
- unsigned Opc = 0;
- switch (VT.getSimpleVT().SimpleTy) {
- default: assert(false && "Illegal conditional move type!");
- break;
- case MVT::i32:
- Opc = Subtarget->isThumb()
- ? (Subtarget->hasThumb2() ? ARM::t2MOVCCr : ARM::tMOVCCr_pseudo)
- : ARM::MOVCCr;
- break;
- case MVT::f32:
- Opc = ARM::FCPYScc;
- break;
- case MVT::f64:
- Opc = ARM::FCPYDcc;
- break;
- }
- return CurDAG->SelectNodeTo(Op.getNode(), Opc, VT, Ops, 5);
- }
+ case ARMISD::CMOV:
+ return SelectCMOVOp(N);
case ARMISD::CNEG: {
- EVT VT = Op.getValueType();
- SDValue N0 = Op.getOperand(0);
- SDValue N1 = Op.getOperand(1);
- SDValue N2 = Op.getOperand(2);
- SDValue N3 = Op.getOperand(3);
- SDValue InFlag = Op.getOperand(4);
+ EVT VT = N->getValueType(0);
+ SDValue N0 = N->getOperand(0);
+ SDValue N1 = N->getOperand(1);
+ SDValue N2 = N->getOperand(2);
+ SDValue N3 = N->getOperand(3);
+ SDValue InFlag = N->getOperand(4);
assert(N2.getOpcode() == ISD::Constant);
assert(N3.getOpcode() == ISD::Register);
default: assert(false && "Illegal conditional move type!");
break;
case MVT::f32:
- Opc = ARM::FNEGScc;
+ Opc = ARM::VNEGScc;
break;
case MVT::f64:
- Opc = ARM::FNEGDcc;
+ Opc = ARM::VNEGDcc;
break;
}
- return CurDAG->SelectNodeTo(Op.getNode(), Opc, VT, Ops, 5);
+ return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
}
case ARMISD::VZIP: {
case MVT::v4f32:
case MVT::v4i32: Opc = ARM::VZIPq32; break;
}
- return CurDAG->getMachineNode(Opc, dl, VT, VT,
- N->getOperand(0), N->getOperand(1));
+ SDValue Pred = getAL(CurDAG);
+ SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
+ SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
+ return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
}
case ARMISD::VUZP: {
unsigned Opc = 0;
case MVT::v4f32:
case MVT::v4i32: Opc = ARM::VUZPq32; break;
}
- return CurDAG->getMachineNode(Opc, dl, VT, VT,
- N->getOperand(0), N->getOperand(1));
+ SDValue Pred = getAL(CurDAG);
+ SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
+ SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
+ return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
}
case ARMISD::VTRN: {
unsigned Opc = 0;
case MVT::v4f32:
case MVT::v4i32: Opc = ARM::VTRNq32; break;
}
- return CurDAG->getMachineNode(Opc, dl, VT, VT,
- N->getOperand(0), N->getOperand(1));
+ SDValue Pred = getAL(CurDAG);
+ SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
+ SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
+ return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
+ }
+ case ARMISD::BUILD_VECTOR: {
+ EVT VecVT = N->getValueType(0);
+ EVT EltVT = VecVT.getVectorElementType();
+ unsigned NumElts = VecVT.getVectorNumElements();
+ if (EltVT.getSimpleVT() == MVT::f64) {
+ assert(NumElts == 2 && "unexpected type for BUILD_VECTOR");
+ return PairDRegs(VecVT, N->getOperand(0), N->getOperand(1));
+ }
+ assert(EltVT.getSimpleVT() == MVT::f32 &&
+ "unexpected type for BUILD_VECTOR");
+ if (NumElts == 2)
+ return PairSRegs(VecVT, N->getOperand(0), N->getOperand(1));
+ assert(NumElts == 4 && "unexpected type for BUILD_VECTOR");
+ return QuadSRegs(VecVT, N->getOperand(0), N->getOperand(1),
+ N->getOperand(2), N->getOperand(3));
}
case ISD::INTRINSIC_VOID:
case ISD::INTRINSIC_W_CHAIN: {
unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
- EVT VT = N->getValueType(0);
- unsigned Opc = 0;
-
switch (IntNo) {
default:
break;
+ case Intrinsic::arm_neon_vld1: {
+ unsigned DOpcodes[] = { ARM::VLD1d8, ARM::VLD1d16,
+ ARM::VLD1d32, ARM::VLD1d64 };
+ unsigned QOpcodes[] = { ARM::VLD1q8, ARM::VLD1q16,
+ ARM::VLD1q32, ARM::VLD1q64 };
+ return SelectVLD(N, 1, DOpcodes, QOpcodes, 0);
+ }
+
case Intrinsic::arm_neon_vld2: {
- SDValue MemAddr, MemUpdate, MemOpc;
- if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
- return NULL;
- if (VT.is64BitVector()) {
- switch (VT.getSimpleVT().SimpleTy) {
- default: llvm_unreachable("unhandled vld2 type");
- case MVT::v8i8: Opc = ARM::VLD2d8; break;
- case MVT::v4i16: Opc = ARM::VLD2d16; break;
- case MVT::v2f32:
- case MVT::v2i32: Opc = ARM::VLD2d32; break;
- case MVT::v1i64: Opc = ARM::VLD2d64; break;
- }
- SDValue Chain = N->getOperand(0);
- const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, Chain };
- return CurDAG->getMachineNode(Opc, dl, VT, VT, MVT::Other, Ops, 4);
- }
- // Quad registers are loaded as pairs of double registers.
- EVT RegVT;
- switch (VT.getSimpleVT().SimpleTy) {
- default: llvm_unreachable("unhandled vld2 type");
- case MVT::v16i8: Opc = ARM::VLD2q8; RegVT = MVT::v8i8; break;
- case MVT::v8i16: Opc = ARM::VLD2q16; RegVT = MVT::v4i16; break;
- case MVT::v4f32: Opc = ARM::VLD2q32; RegVT = MVT::v2f32; break;
- case MVT::v4i32: Opc = ARM::VLD2q32; RegVT = MVT::v2i32; break;
- }
- SDValue Chain = N->getOperand(0);
- const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, Chain };
- std::vector<EVT> ResTys(4, RegVT);
- ResTys.push_back(MVT::Other);
- SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 4);
- SDNode *Q0 = PairDRegs(VT, SDValue(VLd, 0), SDValue(VLd, 1));
- SDNode *Q1 = PairDRegs(VT, SDValue(VLd, 2), SDValue(VLd, 3));
- ReplaceUses(SDValue(N, 0), SDValue(Q0, 0));
- ReplaceUses(SDValue(N, 1), SDValue(Q1, 0));
- ReplaceUses(SDValue(N, 2), SDValue(VLd, 4));
- return NULL;
+ unsigned DOpcodes[] = { ARM::VLD2d8, ARM::VLD2d16,
+ ARM::VLD2d32, ARM::VLD1q64 };
+ unsigned QOpcodes[] = { ARM::VLD2q8, ARM::VLD2q16, ARM::VLD2q32 };
+ return SelectVLD(N, 2, DOpcodes, QOpcodes, 0);
}
case Intrinsic::arm_neon_vld3: {
- SDValue MemAddr, MemUpdate, MemOpc;
- if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
- return NULL;
- if (VT.is64BitVector()) {
- switch (VT.getSimpleVT().SimpleTy) {
- default: llvm_unreachable("unhandled vld3 type");
- case MVT::v8i8: Opc = ARM::VLD3d8; break;
- case MVT::v4i16: Opc = ARM::VLD3d16; break;
- case MVT::v2f32:
- case MVT::v2i32: Opc = ARM::VLD3d32; break;
- case MVT::v1i64: Opc = ARM::VLD3d64; break;
- }
- SDValue Chain = N->getOperand(0);
- const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, Chain };
- return CurDAG->getMachineNode(Opc, dl, VT, VT, VT, MVT::Other, Ops, 4);
- }
- // Quad registers are loaded with two separate instructions, where one
- // loads the even registers and the other loads the odd registers.
- EVT RegVT;
- unsigned Opc2 = 0;
- switch (VT.getSimpleVT().SimpleTy) {
- default: llvm_unreachable("unhandled vld3 type");
- case MVT::v16i8:
- Opc = ARM::VLD3q8a; Opc2 = ARM::VLD3q8b; RegVT = MVT::v8i8; break;
- case MVT::v8i16:
- Opc = ARM::VLD3q16a; Opc2 = ARM::VLD3q16b; RegVT = MVT::v4i16; break;
- case MVT::v4f32:
- Opc = ARM::VLD3q32a; Opc2 = ARM::VLD3q32b; RegVT = MVT::v2f32; break;
- case MVT::v4i32:
- Opc = ARM::VLD3q32a; Opc2 = ARM::VLD3q32b; RegVT = MVT::v2i32; break;
- }
- SDValue Chain = N->getOperand(0);
- // Enable writeback to the address register.
- MemOpc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(true), MVT::i32);
-
- std::vector<EVT> ResTys(3, RegVT);
- ResTys.push_back(MemAddr.getValueType());
- ResTys.push_back(MVT::Other);
-
- const SDValue OpsA[] = { MemAddr, MemUpdate, MemOpc, Chain };
- SDNode *VLdA = CurDAG->getMachineNode(Opc, dl, ResTys, OpsA, 4);
- Chain = SDValue(VLdA, 4);
-
- const SDValue OpsB[] = { SDValue(VLdA, 3), MemUpdate, MemOpc, Chain };
- SDNode *VLdB = CurDAG->getMachineNode(Opc2, dl, ResTys, OpsB, 4);
- Chain = SDValue(VLdB, 4);
-
- SDNode *Q0 = PairDRegs(VT, SDValue(VLdA, 0), SDValue(VLdB, 0));
- SDNode *Q1 = PairDRegs(VT, SDValue(VLdA, 1), SDValue(VLdB, 1));
- SDNode *Q2 = PairDRegs(VT, SDValue(VLdA, 2), SDValue(VLdB, 2));
- ReplaceUses(SDValue(N, 0), SDValue(Q0, 0));
- ReplaceUses(SDValue(N, 1), SDValue(Q1, 0));
- ReplaceUses(SDValue(N, 2), SDValue(Q2, 0));
- ReplaceUses(SDValue(N, 3), Chain);
- return NULL;
+ unsigned DOpcodes[] = { ARM::VLD3d8, ARM::VLD3d16,
+ ARM::VLD3d32, ARM::VLD1d64T };
+ unsigned QOpcodes0[] = { ARM::VLD3q8_UPD,
+ ARM::VLD3q16_UPD,
+ ARM::VLD3q32_UPD };
+ unsigned QOpcodes1[] = { ARM::VLD3q8odd_UPD,
+ ARM::VLD3q16odd_UPD,
+ ARM::VLD3q32odd_UPD };
+ return SelectVLD(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
}
case Intrinsic::arm_neon_vld4: {
- SDValue MemAddr, MemUpdate, MemOpc;
- if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
- return NULL;
- if (VT.is64BitVector()) {
- switch (VT.getSimpleVT().SimpleTy) {
- default: llvm_unreachable("unhandled vld4 type");
- case MVT::v8i8: Opc = ARM::VLD4d8; break;
- case MVT::v4i16: Opc = ARM::VLD4d16; break;
- case MVT::v2f32:
- case MVT::v2i32: Opc = ARM::VLD4d32; break;
- case MVT::v1i64: Opc = ARM::VLD4d64; break;
- }
- SDValue Chain = N->getOperand(0);
- const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, Chain };
- std::vector<EVT> ResTys(4, VT);
- ResTys.push_back(MVT::Other);
- return CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 4);
- }
- // Quad registers are loaded with two separate instructions, where one
- // loads the even registers and the other loads the odd registers.
- EVT RegVT;
- unsigned Opc2 = 0;
- switch (VT.getSimpleVT().SimpleTy) {
- default: llvm_unreachable("unhandled vld4 type");
- case MVT::v16i8:
- Opc = ARM::VLD4q8a; Opc2 = ARM::VLD4q8b; RegVT = MVT::v8i8; break;
- case MVT::v8i16:
- Opc = ARM::VLD4q16a; Opc2 = ARM::VLD4q16b; RegVT = MVT::v4i16; break;
- case MVT::v4f32:
- Opc = ARM::VLD4q32a; Opc2 = ARM::VLD4q32b; RegVT = MVT::v2f32; break;
- case MVT::v4i32:
- Opc = ARM::VLD4q32a; Opc2 = ARM::VLD4q32b; RegVT = MVT::v2i32; break;
- }
- SDValue Chain = N->getOperand(0);
- // Enable writeback to the address register.
- MemOpc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(true), MVT::i32);
-
- std::vector<EVT> ResTys(4, RegVT);
- ResTys.push_back(MemAddr.getValueType());
- ResTys.push_back(MVT::Other);
-
- const SDValue OpsA[] = { MemAddr, MemUpdate, MemOpc, Chain };
- SDNode *VLdA = CurDAG->getMachineNode(Opc, dl, ResTys, OpsA, 4);
- Chain = SDValue(VLdA, 5);
-
- const SDValue OpsB[] = { SDValue(VLdA, 4), MemUpdate, MemOpc, Chain };
- SDNode *VLdB = CurDAG->getMachineNode(Opc2, dl, ResTys, OpsB, 4);
- Chain = SDValue(VLdB, 5);
-
- SDNode *Q0 = PairDRegs(VT, SDValue(VLdA, 0), SDValue(VLdB, 0));
- SDNode *Q1 = PairDRegs(VT, SDValue(VLdA, 1), SDValue(VLdB, 1));
- SDNode *Q2 = PairDRegs(VT, SDValue(VLdA, 2), SDValue(VLdB, 2));
- SDNode *Q3 = PairDRegs(VT, SDValue(VLdA, 3), SDValue(VLdB, 3));
- ReplaceUses(SDValue(N, 0), SDValue(Q0, 0));
- ReplaceUses(SDValue(N, 1), SDValue(Q1, 0));
- ReplaceUses(SDValue(N, 2), SDValue(Q2, 0));
- ReplaceUses(SDValue(N, 3), SDValue(Q3, 0));
- ReplaceUses(SDValue(N, 4), Chain);
- return NULL;
+ unsigned DOpcodes[] = { ARM::VLD4d8, ARM::VLD4d16,
+ ARM::VLD4d32, ARM::VLD1d64Q };
+ unsigned QOpcodes0[] = { ARM::VLD4q8_UPD,
+ ARM::VLD4q16_UPD,
+ ARM::VLD4q32_UPD };
+ unsigned QOpcodes1[] = { ARM::VLD4q8odd_UPD,
+ ARM::VLD4q16odd_UPD,
+ ARM::VLD4q32odd_UPD };
+ return SelectVLD(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
}
case Intrinsic::arm_neon_vld2lane: {
- SDValue MemAddr, MemUpdate, MemOpc;
- if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
- return NULL;
- if (VT.is64BitVector()) {
- switch (VT.getSimpleVT().SimpleTy) {
- default: llvm_unreachable("unhandled vld2lane type");
- case MVT::v8i8: Opc = ARM::VLD2LNd8; break;
- case MVT::v4i16: Opc = ARM::VLD2LNd16; break;
- case MVT::v2f32:
- case MVT::v2i32: Opc = ARM::VLD2LNd32; break;
- }
- SDValue Chain = N->getOperand(0);
- const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
- N->getOperand(3), N->getOperand(4),
- N->getOperand(5), Chain };
- return CurDAG->getMachineNode(Opc, dl, VT, VT, MVT::Other, Ops, 7);
- }
- // Quad registers are handled by extracting subregs, doing the load,
- // and then inserting the results as subregs.
- EVT RegVT;
- unsigned Opc2 = 0;
- switch (VT.getSimpleVT().SimpleTy) {
- default: llvm_unreachable("unhandled vld2lane type");
- case MVT::v8i16:
- Opc = ARM::VLD2LNq16a;
- Opc2 = ARM::VLD2LNq16b;
- RegVT = MVT::v4i16;
- break;
- case MVT::v4f32:
- Opc = ARM::VLD2LNq32a;
- Opc2 = ARM::VLD2LNq32b;
- RegVT = MVT::v2f32;
- break;
- case MVT::v4i32:
- Opc = ARM::VLD2LNq32a;
- Opc2 = ARM::VLD2LNq32b;
- RegVT = MVT::v2i32;
- break;
- }
- SDValue Chain = N->getOperand(0);
- unsigned Lane = cast<ConstantSDNode>(N->getOperand(5))->getZExtValue();
- unsigned NumElts = RegVT.getVectorNumElements();
- int SubregIdx = (Lane < NumElts) ? ARM::DSUBREG_0 : ARM::DSUBREG_1;
-
- SDValue D0 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
- N->getOperand(3));
- SDValue D1 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
- N->getOperand(4));
- const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, D0, D1,
- getI32Imm(Lane % NumElts), Chain };
- SDNode *VLdLn = CurDAG->getMachineNode((Lane < NumElts) ? Opc : Opc2,
- dl, RegVT, RegVT, MVT::Other,
- Ops, 7);
- SDValue Q0 = CurDAG->getTargetInsertSubreg(SubregIdx, dl, VT,
- N->getOperand(3),
- SDValue(VLdLn, 0));
- SDValue Q1 = CurDAG->getTargetInsertSubreg(SubregIdx, dl, VT,
- N->getOperand(4),
- SDValue(VLdLn, 1));
- Chain = SDValue(VLdLn, 2);
- ReplaceUses(SDValue(N, 0), Q0);
- ReplaceUses(SDValue(N, 1), Q1);
- ReplaceUses(SDValue(N, 2), Chain);
- return NULL;
+ unsigned DOpcodes[] = { ARM::VLD2LNd8, ARM::VLD2LNd16, ARM::VLD2LNd32 };
+ unsigned QOpcodes0[] = { ARM::VLD2LNq16, ARM::VLD2LNq32 };
+ unsigned QOpcodes1[] = { ARM::VLD2LNq16odd, ARM::VLD2LNq32odd };
+ return SelectVLDSTLane(N, true, 2, DOpcodes, QOpcodes0, QOpcodes1);
}
case Intrinsic::arm_neon_vld3lane: {
- SDValue MemAddr, MemUpdate, MemOpc;
- if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
- return NULL;
- switch (VT.getSimpleVT().SimpleTy) {
- default: llvm_unreachable("unhandled vld3lane type");
- case MVT::v8i8: Opc = ARM::VLD3LNd8; break;
- case MVT::v4i16: Opc = ARM::VLD3LNd16; break;
- case MVT::v2f32:
- case MVT::v2i32: Opc = ARM::VLD3LNd32; break;
- }
- SDValue Chain = N->getOperand(0);
- const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
- N->getOperand(3), N->getOperand(4),
- N->getOperand(5), N->getOperand(6), Chain };
- return CurDAG->getMachineNode(Opc, dl, VT, VT, VT, MVT::Other, Ops, 8);
+ unsigned DOpcodes[] = { ARM::VLD3LNd8, ARM::VLD3LNd16, ARM::VLD3LNd32 };
+ unsigned QOpcodes0[] = { ARM::VLD3LNq16, ARM::VLD3LNq32 };
+ unsigned QOpcodes1[] = { ARM::VLD3LNq16odd, ARM::VLD3LNq32odd };
+ return SelectVLDSTLane(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
}
case Intrinsic::arm_neon_vld4lane: {
- SDValue MemAddr, MemUpdate, MemOpc;
- if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
- return NULL;
- switch (VT.getSimpleVT().SimpleTy) {
- default: llvm_unreachable("unhandled vld4lane type");
- case MVT::v8i8: Opc = ARM::VLD4LNd8; break;
- case MVT::v4i16: Opc = ARM::VLD4LNd16; break;
- case MVT::v2f32:
- case MVT::v2i32: Opc = ARM::VLD4LNd32; break;
- }
- SDValue Chain = N->getOperand(0);
- const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
- N->getOperand(3), N->getOperand(4),
- N->getOperand(5), N->getOperand(6),
- N->getOperand(7), Chain };
- std::vector<EVT> ResTys(4, VT);
- ResTys.push_back(MVT::Other);
- return CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 9);
+ unsigned DOpcodes[] = { ARM::VLD4LNd8, ARM::VLD4LNd16, ARM::VLD4LNd32 };
+ unsigned QOpcodes0[] = { ARM::VLD4LNq16, ARM::VLD4LNq32 };
+ unsigned QOpcodes1[] = { ARM::VLD4LNq16odd, ARM::VLD4LNq32odd };
+ return SelectVLDSTLane(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
+ }
+
+ case Intrinsic::arm_neon_vst1: {
+ unsigned DOpcodes[] = { ARM::VST1d8, ARM::VST1d16,
+ ARM::VST1d32, ARM::VST1d64 };
+ unsigned QOpcodes[] = { ARM::VST1q8, ARM::VST1q16,
+ ARM::VST1q32, ARM::VST1q64 };
+ return SelectVST(N, 1, DOpcodes, QOpcodes, 0);
}
case Intrinsic::arm_neon_vst2: {
- SDValue MemAddr, MemUpdate, MemOpc;
- if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
- return NULL;
- VT = N->getOperand(3).getValueType();
- if (VT.is64BitVector()) {
- switch (VT.getSimpleVT().SimpleTy) {
- default: llvm_unreachable("unhandled vst2 type");
- case MVT::v8i8: Opc = ARM::VST2d8; break;
- case MVT::v4i16: Opc = ARM::VST2d16; break;
- case MVT::v2f32:
- case MVT::v2i32: Opc = ARM::VST2d32; break;
- case MVT::v1i64: Opc = ARM::VST2d64; break;
- }
- SDValue Chain = N->getOperand(0);
- const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
- N->getOperand(3), N->getOperand(4), Chain };
- return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 6);
- }
- // Quad registers are stored as pairs of double registers.
- EVT RegVT;
- switch (VT.getSimpleVT().SimpleTy) {
- default: llvm_unreachable("unhandled vst2 type");
- case MVT::v16i8: Opc = ARM::VST2q8; RegVT = MVT::v8i8; break;
- case MVT::v8i16: Opc = ARM::VST2q16; RegVT = MVT::v4i16; break;
- case MVT::v4f32: Opc = ARM::VST2q32; RegVT = MVT::v2f32; break;
- case MVT::v4i32: Opc = ARM::VST2q32; RegVT = MVT::v2i32; break;
- }
- SDValue Chain = N->getOperand(0);
- SDValue D0 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
- N->getOperand(3));
- SDValue D1 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
- N->getOperand(3));
- SDValue D2 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
- N->getOperand(4));
- SDValue D3 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
- N->getOperand(4));
- const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
- D0, D1, D2, D3, Chain };
- return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 8);
+ unsigned DOpcodes[] = { ARM::VST2d8, ARM::VST2d16,
+ ARM::VST2d32, ARM::VST1q64 };
+ unsigned QOpcodes[] = { ARM::VST2q8, ARM::VST2q16, ARM::VST2q32 };
+ return SelectVST(N, 2, DOpcodes, QOpcodes, 0);
}
case Intrinsic::arm_neon_vst3: {
- SDValue MemAddr, MemUpdate, MemOpc;
- if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
- return NULL;
- VT = N->getOperand(3).getValueType();
- if (VT.is64BitVector()) {
- switch (VT.getSimpleVT().SimpleTy) {
- default: llvm_unreachable("unhandled vst3 type");
- case MVT::v8i8: Opc = ARM::VST3d8; break;
- case MVT::v4i16: Opc = ARM::VST3d16; break;
- case MVT::v2f32:
- case MVT::v2i32: Opc = ARM::VST3d32; break;
- case MVT::v1i64: Opc = ARM::VST3d64; break;
- }
- SDValue Chain = N->getOperand(0);
- const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
- N->getOperand(3), N->getOperand(4),
- N->getOperand(5), Chain };
- return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 7);
- }
- // Quad registers are stored with two separate instructions, where one
- // stores the even registers and the other stores the odd registers.
- EVT RegVT;
- unsigned Opc2 = 0;
- switch (VT.getSimpleVT().SimpleTy) {
- default: llvm_unreachable("unhandled vst3 type");
- case MVT::v16i8:
- Opc = ARM::VST3q8a; Opc2 = ARM::VST3q8b; RegVT = MVT::v8i8; break;
- case MVT::v8i16:
- Opc = ARM::VST3q16a; Opc2 = ARM::VST3q16b; RegVT = MVT::v4i16; break;
- case MVT::v4f32:
- Opc = ARM::VST3q32a; Opc2 = ARM::VST3q32b; RegVT = MVT::v2f32; break;
- case MVT::v4i32:
- Opc = ARM::VST3q32a; Opc2 = ARM::VST3q32b; RegVT = MVT::v2i32; break;
- }
- SDValue Chain = N->getOperand(0);
- // Enable writeback to the address register.
- MemOpc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(true), MVT::i32);
-
- SDValue D0 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
- N->getOperand(3));
- SDValue D2 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
- N->getOperand(4));
- SDValue D4 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
- N->getOperand(5));
- const SDValue OpsA[] = { MemAddr, MemUpdate, MemOpc, D0, D2, D4, Chain };
- SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
- MVT::Other, OpsA, 7);
- Chain = SDValue(VStA, 1);
-
- SDValue D1 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
- N->getOperand(3));
- SDValue D3 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
- N->getOperand(4));
- SDValue D5 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
- N->getOperand(5));
- MemAddr = SDValue(VStA, 0);
- const SDValue OpsB[] = { MemAddr, MemUpdate, MemOpc, D1, D3, D5, Chain };
- SDNode *VStB = CurDAG->getMachineNode(Opc2, dl, MemAddr.getValueType(),
- MVT::Other, OpsB, 7);
- Chain = SDValue(VStB, 1);
- ReplaceUses(SDValue(N, 0), Chain);
- return NULL;
+ unsigned DOpcodes[] = { ARM::VST3d8, ARM::VST3d16,
+ ARM::VST3d32, ARM::VST1d64T };
+ unsigned QOpcodes0[] = { ARM::VST3q8_UPD,
+ ARM::VST3q16_UPD,
+ ARM::VST3q32_UPD };
+ unsigned QOpcodes1[] = { ARM::VST3q8odd_UPD,
+ ARM::VST3q16odd_UPD,
+ ARM::VST3q32odd_UPD };
+ return SelectVST(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
}
case Intrinsic::arm_neon_vst4: {
- SDValue MemAddr, MemUpdate, MemOpc;
- if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
- return NULL;
- VT = N->getOperand(3).getValueType();
- if (VT.is64BitVector()) {
- switch (VT.getSimpleVT().SimpleTy) {
- default: llvm_unreachable("unhandled vst4 type");
- case MVT::v8i8: Opc = ARM::VST4d8; break;
- case MVT::v4i16: Opc = ARM::VST4d16; break;
- case MVT::v2f32:
- case MVT::v2i32: Opc = ARM::VST4d32; break;
- case MVT::v1i64: Opc = ARM::VST4d64; break;
- }
- SDValue Chain = N->getOperand(0);
- const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
- N->getOperand(3), N->getOperand(4),
- N->getOperand(5), N->getOperand(6), Chain };
- return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 8);
- }
- // Quad registers are stored with two separate instructions, where one
- // stores the even registers and the other stores the odd registers.
- EVT RegVT;
- unsigned Opc2 = 0;
- switch (VT.getSimpleVT().SimpleTy) {
- default: llvm_unreachable("unhandled vst4 type");
- case MVT::v16i8:
- Opc = ARM::VST4q8a; Opc2 = ARM::VST4q8b; RegVT = MVT::v8i8; break;
- case MVT::v8i16:
- Opc = ARM::VST4q16a; Opc2 = ARM::VST4q16b; RegVT = MVT::v4i16; break;
- case MVT::v4f32:
- Opc = ARM::VST4q32a; Opc2 = ARM::VST4q32b; RegVT = MVT::v2f32; break;
- case MVT::v4i32:
- Opc = ARM::VST4q32a; Opc2 = ARM::VST4q32b; RegVT = MVT::v2i32; break;
- }
- SDValue Chain = N->getOperand(0);
- // Enable writeback to the address register.
- MemOpc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(true), MVT::i32);
-
- SDValue D0 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
- N->getOperand(3));
- SDValue D2 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
- N->getOperand(4));
- SDValue D4 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
- N->getOperand(5));
- SDValue D6 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
- N->getOperand(6));
- const SDValue OpsA[] = { MemAddr, MemUpdate, MemOpc,
- D0, D2, D4, D6, Chain };
- SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
- MVT::Other, OpsA, 8);
- Chain = SDValue(VStA, 1);
-
- SDValue D1 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
- N->getOperand(3));
- SDValue D3 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
- N->getOperand(4));
- SDValue D5 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
- N->getOperand(5));
- SDValue D7 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
- N->getOperand(6));
- MemAddr = SDValue(VStA, 0);
- const SDValue OpsB[] = { MemAddr, MemUpdate, MemOpc,
- D1, D3, D5, D7, Chain };
- SDNode *VStB = CurDAG->getMachineNode(Opc2, dl, MemAddr.getValueType(),
- MVT::Other, OpsB, 8);
- Chain = SDValue(VStB, 1);
- ReplaceUses(SDValue(N, 0), Chain);
- return NULL;
+ unsigned DOpcodes[] = { ARM::VST4d8, ARM::VST4d16,
+ ARM::VST4d32, ARM::VST1d64Q };
+ unsigned QOpcodes0[] = { ARM::VST4q8_UPD,
+ ARM::VST4q16_UPD,
+ ARM::VST4q32_UPD };
+ unsigned QOpcodes1[] = { ARM::VST4q8odd_UPD,
+ ARM::VST4q16odd_UPD,
+ ARM::VST4q32odd_UPD };
+ return SelectVST(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
}
case Intrinsic::arm_neon_vst2lane: {
- SDValue MemAddr, MemUpdate, MemOpc;
- if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
- return NULL;
- switch (N->getOperand(3).getValueType().getSimpleVT().SimpleTy) {
- default: llvm_unreachable("unhandled vst2lane type");
- case MVT::v8i8: Opc = ARM::VST2LNd8; break;
- case MVT::v4i16: Opc = ARM::VST2LNd16; break;
- case MVT::v2f32:
- case MVT::v2i32: Opc = ARM::VST2LNd32; break;
- }
- SDValue Chain = N->getOperand(0);
- const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
- N->getOperand(3), N->getOperand(4),
- N->getOperand(5), Chain };
- return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 7);
+ unsigned DOpcodes[] = { ARM::VST2LNd8, ARM::VST2LNd16, ARM::VST2LNd32 };
+ unsigned QOpcodes0[] = { ARM::VST2LNq16, ARM::VST2LNq32 };
+ unsigned QOpcodes1[] = { ARM::VST2LNq16odd, ARM::VST2LNq32odd };
+ return SelectVLDSTLane(N, false, 2, DOpcodes, QOpcodes0, QOpcodes1);
}
case Intrinsic::arm_neon_vst3lane: {
- SDValue MemAddr, MemUpdate, MemOpc;
- if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
- return NULL;
- switch (N->getOperand(3).getValueType().getSimpleVT().SimpleTy) {
- default: llvm_unreachable("unhandled vst3lane type");
- case MVT::v8i8: Opc = ARM::VST3LNd8; break;
- case MVT::v4i16: Opc = ARM::VST3LNd16; break;
- case MVT::v2f32:
- case MVT::v2i32: Opc = ARM::VST3LNd32; break;
- }
- SDValue Chain = N->getOperand(0);
- const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
- N->getOperand(3), N->getOperand(4),
- N->getOperand(5), N->getOperand(6), Chain };
- return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 8);
+ unsigned DOpcodes[] = { ARM::VST3LNd8, ARM::VST3LNd16, ARM::VST3LNd32 };
+ unsigned QOpcodes0[] = { ARM::VST3LNq16, ARM::VST3LNq32 };
+ unsigned QOpcodes1[] = { ARM::VST3LNq16odd, ARM::VST3LNq32odd };
+ return SelectVLDSTLane(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
}
case Intrinsic::arm_neon_vst4lane: {
- SDValue MemAddr, MemUpdate, MemOpc;
- if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
- return NULL;
- switch (N->getOperand(3).getValueType().getSimpleVT().SimpleTy) {
- default: llvm_unreachable("unhandled vst4lane type");
- case MVT::v8i8: Opc = ARM::VST4LNd8; break;
- case MVT::v4i16: Opc = ARM::VST4LNd16; break;
- case MVT::v2f32:
- case MVT::v2i32: Opc = ARM::VST4LNd32; break;
- }
- SDValue Chain = N->getOperand(0);
- const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
- N->getOperand(3), N->getOperand(4),
- N->getOperand(5), N->getOperand(6),
- N->getOperand(7), Chain };
- return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 9);
+ unsigned DOpcodes[] = { ARM::VST4LNd8, ARM::VST4LNd16, ARM::VST4LNd32 };
+ unsigned QOpcodes0[] = { ARM::VST4LNq16, ARM::VST4LNq32 };
+ unsigned QOpcodes1[] = { ARM::VST4LNq16odd, ARM::VST4LNq32odd };
+ return SelectVLDSTLane(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
}
}
+ break;
}
+
+ case ISD::CONCAT_VECTORS:
+ return SelectConcatVector(N);
}
- return SelectCode(Op);
+ return SelectCode(N);
}
bool ARMDAGToDAGISel::
SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
std::vector<SDValue> &OutOps) {
assert(ConstraintCode == 'm' && "unexpected asm memory constraint");
-
- SDValue Base, Offset, Opc;
- if (!SelectAddrMode2(Op, Op, Base, Offset, Opc))
- return true;
-
- OutOps.push_back(Base);
- OutOps.push_back(Offset);
- OutOps.push_back(Opc);
+ // Require the address to be in a register. That is safe for all ARM
+ // variants and it is hard to do anything much smarter without knowing
+ // how the operand is used.
+ OutOps.push_back(Op);
return false;
}
projects / oota-llvm.git / blobdiff