X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTarget%2FARM%2FARMISelDAGToDAG.cpp;h=9ca80ae1b91838386b2ea4c038987419c3c90bdf;hb=1d0be15f89cb5056e20e2d24faa8d6afb1573bca;hp=a413ee6f3e0c41fcc1fce9d577f4a3f4ba42a9f2;hpb=0eebf653a7b2978e7761f8d068b6fbec22aea0f6;p=oota-llvm.git diff --git a/lib/Target/ARM/ARMISelDAGToDAG.cpp b/lib/Target/ARM/ARMISelDAGToDAG.cpp index a413ee6f3e0..9ca80ae1b91 100644 --- a/lib/Target/ARM/ARMISelDAGToDAG.cpp +++ b/lib/Target/ARM/ARMISelDAGToDAG.cpp @@ -21,6 +21,7 @@ #include "llvm/DerivedTypes.h" #include "llvm/Function.h" #include "llvm/Intrinsics.h" +#include "llvm/LLVMContext.h" #include "llvm/CodeGen/MachineFrameInfo.h" #include "llvm/CodeGen/MachineFunction.h" #include "llvm/CodeGen/MachineInstrBuilder.h" @@ -30,32 +31,45 @@ #include "llvm/Target/TargetOptions.h" #include "llvm/Support/Compiler.h" #include "llvm/Support/Debug.h" +#include "llvm/Support/ErrorHandling.h" +#include "llvm/Support/raw_ostream.h" + using namespace llvm; +static const unsigned arm_dsubreg_0 = 5; +static const unsigned arm_dsubreg_1 = 6; + //===--------------------------------------------------------------------===// /// ARMDAGToDAGISel - ARM specific code to select ARM machine /// instructions for SelectionDAG operations. /// namespace { class ARMDAGToDAGISel : public SelectionDAGISel { - ARMTargetMachine &TM; + ARMBaseTargetMachine &TM; /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can /// make the right decision when generating code for different targets. const ARMSubtarget *Subtarget; public: - explicit ARMDAGToDAGISel(ARMTargetMachine &tm) + explicit ARMDAGToDAGISel(ARMBaseTargetMachine &tm) : SelectionDAGISel(tm), TM(tm), Subtarget(&TM.getSubtarget()) { } virtual const char *getPassName() const { return "ARM Instruction Selection"; - } - + } + + /// getI32Imm - Return a target constant with the specified value, of type i32. + inline SDValue getI32Imm(unsigned Imm) { + return CurDAG->getTargetConstant(Imm, MVT::i32); + } + SDNode *Select(SDValue Op); virtual void InstructionSelect(); + bool SelectShifterOperandReg(SDValue Op, SDValue N, SDValue &A, + SDValue &B, SDValue &C); bool SelectAddrMode2(SDValue Op, SDValue N, SDValue &Base, SDValue &Offset, SDValue &Opc); bool SelectAddrMode2Offset(SDValue Op, SDValue N, @@ -64,11 +78,15 @@ public: SDValue &Offset, SDValue &Opc); bool SelectAddrMode3Offset(SDValue Op, SDValue N, SDValue &Offset, SDValue &Opc); + bool SelectAddrMode4(SDValue Op, SDValue N, SDValue &Addr, + SDValue &Mode); bool SelectAddrMode5(SDValue Op, SDValue N, SDValue &Base, SDValue &Offset); + bool SelectAddrMode6(SDValue Op, SDValue N, SDValue &Addr, SDValue &Update, + SDValue &Opc); bool SelectAddrModePC(SDValue Op, SDValue N, SDValue &Offset, - SDValue &Label); + SDValue &Label); bool SelectThumbAddrModeRR(SDValue Op, SDValue N, SDValue &Base, SDValue &Offset); @@ -84,18 +102,36 @@ public: bool SelectThumbAddrModeSP(SDValue Op, SDValue N, SDValue &Base, SDValue &OffImm); - bool SelectShifterOperandReg(SDValue Op, SDValue N, SDValue &A, - SDValue &B, SDValue &C); - + bool SelectT2ShifterOperandReg(SDValue Op, SDValue N, + SDValue &BaseReg, SDValue &Opc); + bool SelectT2AddrModeImm12(SDValue Op, SDValue N, SDValue &Base, + SDValue &OffImm); + bool SelectT2AddrModeImm8(SDValue Op, SDValue N, SDValue &Base, + SDValue &OffImm); + bool SelectT2AddrModeImm8Offset(SDValue Op, SDValue N, + SDValue &OffImm); + bool SelectT2AddrModeImm8s4(SDValue Op, SDValue N, SDValue &Base, + SDValue &OffImm); + bool SelectT2AddrModeSoReg(SDValue Op, SDValue N, SDValue &Base, + SDValue &OffReg, SDValue &ShImm); + // Include the pieces autogenerated from the target description. #include "ARMGenDAGISel.inc" private: - /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for - /// inline asm expressions. - virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op, - char ConstraintCode, - std::vector &OutOps); + /// 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); + + /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for + /// inline asm expressions. + virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op, + char ConstraintCode, + std::vector &OutOps); }; } @@ -106,6 +142,30 @@ void ARMDAGToDAGISel::InstructionSelect() { CurDAG->RemoveDeadNodes(); } +bool ARMDAGToDAGISel::SelectShifterOperandReg(SDValue Op, + SDValue N, + SDValue &BaseReg, + SDValue &ShReg, + SDValue &Opc) { + ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N); + + // Don't match base register only case. That is matched to a separate + // lower complexity pattern with explicit register operand. + if (ShOpcVal == ARM_AM::no_shift) return false; + + BaseReg = N.getOperand(0); + unsigned ShImmVal = 0; + if (ConstantSDNode *RHS = dyn_cast(N.getOperand(1))) { + ShReg = CurDAG->getRegister(0, MVT::i32); + ShImmVal = RHS->getZExtValue() & 31; + } else { + ShReg = N.getOperand(1); + } + Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal), + MVT::i32); + return true; +} + bool ARMDAGToDAGISel::SelectAddrMode2(SDValue Op, SDValue N, SDValue &Base, SDValue &Offset, SDValue &Opc) { @@ -146,7 +206,7 @@ bool ARMDAGToDAGISel::SelectAddrMode2(SDValue Op, SDValue N, MVT::i32); return true; } - + // Match simple R +/- imm12 operands. if (N.getOpcode() == ISD::ADD) if (ConstantSDNode *RHS = dyn_cast(N.getOperand(1))) { @@ -171,15 +231,15 @@ bool ARMDAGToDAGISel::SelectAddrMode2(SDValue Op, SDValue N, return true; } } - + // 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; - + Base = N.getOperand(0); Offset = N.getOperand(1); - + if (ShOpcVal != ARM_AM::no_shift) { // Check to see if the RHS of the shift is a constant, if not, we can't fold // it. @@ -191,7 +251,7 @@ bool ARMDAGToDAGISel::SelectAddrMode2(SDValue Op, SDValue N, ShOpcVal = ARM_AM::no_shift; } } - + // Try matching (R shl C) + (R). if (N.getOpcode() == ISD::ADD && ShOpcVal == ARM_AM::no_shift) { ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0)); @@ -208,7 +268,7 @@ bool ARMDAGToDAGISel::SelectAddrMode2(SDValue Op, SDValue N, } } } - + Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal), MVT::i32); return true; @@ -263,7 +323,7 @@ bool ARMDAGToDAGISel::SelectAddrMode3(SDValue Op, SDValue N, Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::sub, 0),MVT::i32); return true; } - + if (N.getOpcode() != ISD::ADD) { Base = N; if (N.getOpcode() == ISD::FrameIndex) { @@ -274,7 +334,7 @@ bool ARMDAGToDAGISel::SelectAddrMode3(SDValue Op, SDValue N, Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0),MVT::i32); return true; } - + // If the RHS is +/- imm8, fold into addr mode. if (ConstantSDNode *RHS = dyn_cast(N.getOperand(1))) { int RHSC = (int)RHS->getZExtValue(); @@ -296,7 +356,7 @@ bool ARMDAGToDAGISel::SelectAddrMode3(SDValue Op, SDValue N, return true; } } - + Base = N.getOperand(0); Offset = N.getOperand(1); Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0), MVT::i32); @@ -325,6 +385,12 @@ bool ARMDAGToDAGISel::SelectAddrMode3Offset(SDValue Op, SDValue N, return true; } +bool ARMDAGToDAGISel::SelectAddrMode4(SDValue Op, SDValue N, + SDValue &Addr, SDValue &Mode) { + Addr = N; + Mode = CurDAG->getTargetConstant(0, MVT::i32); + return true; +} bool ARMDAGToDAGISel::SelectAddrMode5(SDValue Op, SDValue N, SDValue &Base, SDValue &Offset) { @@ -340,7 +406,7 @@ bool ARMDAGToDAGISel::SelectAddrMode5(SDValue Op, SDValue N, MVT::i32); return true; } - + // If the RHS is +/- imm8, fold into addr mode. if (ConstantSDNode *RHS = dyn_cast(N.getOperand(1))) { int RHSC = (int)RHS->getZExtValue(); @@ -365,13 +431,23 @@ bool ARMDAGToDAGISel::SelectAddrMode5(SDValue Op, SDValue N, } } } - + Base = N; Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0), MVT::i32); return true; } +bool ARMDAGToDAGISel::SelectAddrMode6(SDValue Op, SDValue N, + SDValue &Addr, SDValue &Update, + SDValue &Opc) { + Addr = N; + // The optional writeback is handled in ARMLoadStoreOpt. + Update = CurDAG->getRegister(0, MVT::i32); + Opc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(false), MVT::i32); + return true; +} + bool ARMDAGToDAGISel::SelectAddrModePC(SDValue Op, SDValue N, SDValue &Offset, SDValue &Label) { if (N.getOpcode() == ARMISD::PIC_ADD && N.hasOneUse()) { @@ -389,13 +465,11 @@ bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDValue Op, SDValue N, // FIXME dl should come from the parent load or store, not the address DebugLoc dl = Op.getDebugLoc(); if (N.getOpcode() != ISD::ADD) { - Base = N; - // We must materialize a zero in a reg! Returning a constant here - // wouldn't work without additional code to position the node within - // ISel's topological ordering in a place where ISel will process it - // normally. Instead, just explicitly issue a tMOVri8 node! - Offset = SDValue(CurDAG->getTargetNode(ARM::tMOVi8, dl, MVT::i32, - CurDAG->getTargetConstant(0, MVT::i32)), 0); + ConstantSDNode *NC = dyn_cast(N); + if (!NC || NC->getZExtValue() != 0) + return false; + + Base = Offset = N; return true; } @@ -505,39 +579,351 @@ bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDValue Op, SDValue N, } } } - + return false; } -bool ARMDAGToDAGISel::SelectShifterOperandReg(SDValue Op, - SDValue N, - SDValue &BaseReg, - SDValue &ShReg, - SDValue &Opc) { +bool ARMDAGToDAGISel::SelectT2ShifterOperandReg(SDValue Op, SDValue N, + SDValue &BaseReg, + SDValue &Opc) { ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N); // Don't match base register only case. That is matched to a separate // lower complexity pattern with explicit register operand. if (ShOpcVal == ARM_AM::no_shift) return false; - + BaseReg = N.getOperand(0); unsigned ShImmVal = 0; if (ConstantSDNode *RHS = dyn_cast(N.getOperand(1))) { - ShReg = CurDAG->getRegister(0, MVT::i32); ShImmVal = RHS->getZExtValue() & 31; - } else { - ShReg = N.getOperand(1); + Opc = getI32Imm(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal)); + return true; } - Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal), - MVT::i32); + + return false; +} + +bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDValue Op, SDValue N, + SDValue &Base, SDValue &OffImm) { + // Match simple R + imm12 operands. + + // Base only. + if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) { + if (N.getOpcode() == ISD::FrameIndex) { + // Match frame index... + int FI = cast(N)->getIndex(); + Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy()); + OffImm = CurDAG->getTargetConstant(0, MVT::i32); + return true; + } else if (N.getOpcode() == ARMISD::Wrapper) { + Base = N.getOperand(0); + if (Base.getOpcode() == ISD::TargetConstantPool) + return false; // We want to select t2LDRpci instead. + } else + Base = N; + OffImm = CurDAG->getTargetConstant(0, MVT::i32); + return true; + } + + if (ConstantSDNode *RHS = dyn_cast(N.getOperand(1))) { + if (SelectT2AddrModeImm8(Op, N, Base, OffImm)) + // Let t2LDRi8 handle (R - imm8). + return false; + + int RHSC = (int)RHS->getZExtValue(); + if (N.getOpcode() == ISD::SUB) + RHSC = -RHSC; + + if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned) + Base = N.getOperand(0); + if (Base.getOpcode() == ISD::FrameIndex) { + int FI = cast(Base)->getIndex(); + Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy()); + } + OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32); + return true; + } + } + + // Base only. + Base = N; + OffImm = CurDAG->getTargetConstant(0, MVT::i32); return true; } +bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDValue Op, SDValue N, + SDValue &Base, SDValue &OffImm) { + // Match simple R - imm8 operands. + if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::SUB) { + if (ConstantSDNode *RHS = dyn_cast(N.getOperand(1))) { + int RHSC = (int)RHS->getSExtValue(); + if (N.getOpcode() == ISD::SUB) + RHSC = -RHSC; + + if ((RHSC >= -255) && (RHSC < 0)) { // 8 bits (always negative) + Base = N.getOperand(0); + if (Base.getOpcode() == ISD::FrameIndex) { + int FI = cast(Base)->getIndex(); + Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy()); + } + OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32); + return true; + } + } + } + + return false; +} + +bool ARMDAGToDAGISel::SelectT2AddrModeImm8Offset(SDValue Op, SDValue N, + SDValue &OffImm){ + unsigned Opcode = Op.getOpcode(); + ISD::MemIndexedMode AM = (Opcode == ISD::LOAD) + ? cast(Op)->getAddressingMode() + : cast(Op)->getAddressingMode(); + if (ConstantSDNode *RHS = dyn_cast(N)) { + int RHSC = (int)RHS->getZExtValue(); + if (RHSC >= 0 && RHSC < 0x100) { // 8 bits. + OffImm = ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC)) + ? CurDAG->getTargetConstant(RHSC, MVT::i32) + : CurDAG->getTargetConstant(-RHSC, MVT::i32); + return true; + } + } + + return false; +} + +bool ARMDAGToDAGISel::SelectT2AddrModeImm8s4(SDValue Op, SDValue N, + SDValue &Base, SDValue &OffImm) { + if (N.getOpcode() == ISD::ADD) { + if (ConstantSDNode *RHS = dyn_cast(N.getOperand(1))) { + int RHSC = (int)RHS->getZExtValue(); + if (((RHSC & 0x3) == 0) && + ((RHSC >= 0 && RHSC < 0x400) || (RHSC < 0 && RHSC > -0x400))) { // 8 bits. + Base = N.getOperand(0); + OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32); + return true; + } + } + } else if (N.getOpcode() == ISD::SUB) { + if (ConstantSDNode *RHS = dyn_cast(N.getOperand(1))) { + int RHSC = (int)RHS->getZExtValue(); + if (((RHSC & 0x3) == 0) && (RHSC >= 0 && RHSC < 0x400)) { // 8 bits. + Base = N.getOperand(0); + OffImm = CurDAG->getTargetConstant(-RHSC, MVT::i32); + return true; + } + } + } + + return false; +} + +bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDValue Op, SDValue N, + SDValue &Base, + SDValue &OffReg, SDValue &ShImm) { + // (R - imm8) should be handled by t2LDRi8. The rest are handled by t2LDRi12. + if (N.getOpcode() != ISD::ADD) + return false; + + // Leave (R + imm12) for t2LDRi12, (R - imm8) for t2LDRi8. + if (ConstantSDNode *RHS = dyn_cast(N.getOperand(1))) { + int RHSC = (int)RHS->getZExtValue(); + if (RHSC >= 0 && RHSC < 0x1000) // 12 bits (unsigned) + return false; + else if (RHSC < 0 && RHSC >= -255) // 8 bits + return false; + } + + // Look for (R + R) or (R + (R << [1,2,3])). + unsigned ShAmt = 0; + Base = N.getOperand(0); + OffReg = N.getOperand(1); + + // Swap if it is ((R << c) + R). + ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(OffReg); + if (ShOpcVal != ARM_AM::lsl) { + ShOpcVal = ARM_AM::getShiftOpcForNode(Base); + if (ShOpcVal == ARM_AM::lsl) + std::swap(Base, OffReg); + } + + if (ShOpcVal == ARM_AM::lsl) { + // Check to see if the RHS of the shift is a constant, if not, we can't fold + // it. + if (ConstantSDNode *Sh = dyn_cast(OffReg.getOperand(1))) { + ShAmt = Sh->getZExtValue(); + if (ShAmt >= 4) { + ShAmt = 0; + ShOpcVal = ARM_AM::no_shift; + } else + OffReg = OffReg.getOperand(0); + } else { + ShOpcVal = ARM_AM::no_shift; + } + } + + ShImm = CurDAG->getTargetConstant(ShAmt, MVT::i32); + + return true; +} + +//===--------------------------------------------------------------------===// + /// getAL - Returns a ARMCC::AL immediate node. static inline SDValue getAL(SelectionDAG *CurDAG) { return CurDAG->getTargetConstant((uint64_t)ARMCC::AL, MVT::i32); } +SDNode *ARMDAGToDAGISel::SelectARMIndexedLoad(SDValue Op) { + LoadSDNode *LD = cast(Op); + ISD::MemIndexedMode AM = LD->getAddressingMode(); + if (AM == ISD::UNINDEXED) + return NULL; + + EVT LoadedVT = LD->getMemoryVT(); + SDValue Offset, AMOpc; + bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC); + unsigned Opcode = 0; + bool Match = false; + if (LoadedVT == MVT::i32 && + SelectAddrMode2Offset(Op, 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)) { + 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)) { + Match = true; + Opcode = isPre ? ARM::LDRSB_PRE : ARM::LDRSB_POST; + } + } else { + if (SelectAddrMode2Offset(Op, LD->getOffset(), Offset, AMOpc)) { + Match = true; + Opcode = isPre ? ARM::LDRB_PRE : ARM::LDRB_POST; + } + } + } + + if (Match) { + SDValue Chain = LD->getChain(); + SDValue Base = LD->getBasePtr(); + SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG), + CurDAG->getRegister(0, MVT::i32), Chain }; + return CurDAG->getTargetNode(Opcode, Op.getDebugLoc(), MVT::i32, MVT::i32, + MVT::Other, Ops, 6); + } + + return NULL; +} + +SDNode *ARMDAGToDAGISel::SelectT2IndexedLoad(SDValue Op) { + LoadSDNode *LD = cast(Op); + ISD::MemIndexedMode AM = LD->getAddressingMode(); + if (AM == ISD::UNINDEXED) + return NULL; + + EVT LoadedVT = LD->getMemoryVT(); + bool isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD; + SDValue Offset; + bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC); + unsigned Opcode = 0; + bool Match = false; + if (SelectT2AddrModeImm8Offset(Op, LD->getOffset(), Offset)) { + switch (LoadedVT.getSimpleVT().SimpleTy) { + case MVT::i32: + Opcode = isPre ? ARM::t2LDR_PRE : ARM::t2LDR_POST; + break; + case MVT::i16: + if (isSExtLd) + Opcode = isPre ? ARM::t2LDRSH_PRE : ARM::t2LDRSH_POST; + else + Opcode = isPre ? ARM::t2LDRH_PRE : ARM::t2LDRH_POST; + break; + case MVT::i8: + case MVT::i1: + if (isSExtLd) + Opcode = isPre ? ARM::t2LDRSB_PRE : ARM::t2LDRSB_POST; + else + Opcode = isPre ? ARM::t2LDRB_PRE : ARM::t2LDRB_POST; + break; + default: + return NULL; + } + Match = true; + } + + if (Match) { + SDValue Chain = LD->getChain(); + SDValue Base = LD->getBasePtr(); + SDValue Ops[]= { Base, Offset, getAL(CurDAG), + CurDAG->getRegister(0, MVT::i32), Chain }; + return CurDAG->getTargetNode(Opcode, Op.getDebugLoc(), MVT::i32, MVT::i32, + MVT::Other, Ops, 5); + } + + return NULL; +} + +SDNode *ARMDAGToDAGISel::SelectDYN_ALLOC(SDValue Op) { + SDNode *N = Op.getNode(); + 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(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->getTargetNode(ARM::tANDsp, dl, VT, SP, Align), 0); + + bool isC = isa(Size); + uint32_t C = isC ? cast(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); + } 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); + } + } + + // FIXME: Add ADD / SUB sp instructions for ARM. + return 0; +} SDNode *ARMDAGToDAGISel::Select(SDValue Op) { SDNode *N = Op.getNode(); @@ -549,31 +935,35 @@ SDNode *ARMDAGToDAGISel::Select(SDValue Op) { switch (N->getOpcode()) { default: break; case ISD::Constant: { - // ARMv6T2 and later should materialize imms via MOV / MOVT pair. - if (Subtarget->hasV6T2Ops()) - break; - unsigned Val = cast(N)->getZExtValue(); bool UseCP = true; - if (Subtarget->isThumb()) - UseCP = (Val > 255 && // MOV - ~Val > 255 && // MOV + MVN - !ARM_AM::isThumbImmShiftedVal(Val)); // MOV + LSL - else + if (Subtarget->isThumb()) { + if (Subtarget->hasThumb2()) + // Thumb2 has the MOVT instruction, so all immediates can + // be done with MOV + MOVT, at worst. + UseCP = 0; + else + UseCP = (Val > 255 && // MOV + ~Val > 255 && // MOV + MVN + !ARM_AM::isThumbImmShiftedVal(Val)); // MOV + LSL + } else UseCP = (ARM_AM::getSOImmVal(Val) == -1 && // MOV ARM_AM::getSOImmVal(~Val) == -1 && // MVN !ARM_AM::isSOImmTwoPartVal(Val)); // two instrs. - if (UseCP) { SDValue CPIdx = - CurDAG->getTargetConstantPool(ConstantInt::get(Type::Int32Ty, Val), + CurDAG->getTargetConstantPool(ConstantInt::get( + Type::getInt32Ty(*CurDAG->getContext()), Val), TLI.getPointerTy()); SDNode *ResNode; - if (Subtarget->isThumb()) + if (Subtarget->isThumb1Only()) { + SDValue Pred = CurDAG->getTargetConstant(0xEULL, MVT::i32); + SDValue PredReg = CurDAG->getRegister(0, MVT::i32); + SDValue Ops[] = { CPIdx, Pred, PredReg, CurDAG->getEntryNode() }; ResNode = CurDAG->getTargetNode(ARM::tLDRcp, dl, MVT::i32, MVT::Other, - CPIdx, CurDAG->getEntryNode()); - else { + Ops, 4); + } else { SDValue Ops[] = { CPIdx, CurDAG->getRegister(0, MVT::i32), @@ -596,58 +986,57 @@ SDNode *ARMDAGToDAGISel::Select(SDValue Op) { // Selects to ADDri FI, 0 which in turn will become ADDri SP, imm. int FI = cast(N)->getIndex(); SDValue TFI = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy()); - if (Subtarget->isThumb()) { + if (Subtarget->isThumb1Only()) { return CurDAG->SelectNodeTo(N, ARM::tADDrSPi, MVT::i32, TFI, CurDAG->getTargetConstant(0, MVT::i32)); } else { + unsigned Opc = ((Subtarget->isThumb() && Subtarget->hasThumb2()) ? + ARM::t2ADDri : ARM::ADDri); SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, MVT::i32), - getAL(CurDAG), CurDAG->getRegister(0, MVT::i32), - CurDAG->getRegister(0, MVT::i32) }; - return CurDAG->SelectNodeTo(N, ARM::ADDri, MVT::i32, Ops, 5); - } - } - case ISD::ADD: { - if (!Subtarget->isThumb()) - break; - // Select add sp, c to tADDhirr. - SDValue N0 = Op.getOperand(0); - SDValue N1 = Op.getOperand(1); - RegisterSDNode *LHSR = dyn_cast(Op.getOperand(0)); - RegisterSDNode *RHSR = dyn_cast(Op.getOperand(1)); - if (LHSR && LHSR->getReg() == ARM::SP) { - std::swap(N0, N1); - std::swap(LHSR, RHSR); - } - if (RHSR && RHSR->getReg() == ARM::SP) { - SDValue Val = SDValue(CurDAG->getTargetNode(ARM::tMOVlor2hir, dl, - Op.getValueType(), N0, N0), 0); - return CurDAG->SelectNodeTo(N, ARM::tADDhirr, Op.getValueType(), Val, N1); + getAL(CurDAG), CurDAG->getRegister(0, MVT::i32), + CurDAG->getRegister(0, MVT::i32) }; + return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5); } - break; } + case ARMISD::DYN_ALLOC: + return SelectDYN_ALLOC(Op); case ISD::MUL: - if (Subtarget->isThumb()) + if (Subtarget->isThumb1Only()) break; if (ConstantSDNode *C = dyn_cast(Op.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); - unsigned ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, Log2_32(RHSV-1)); - SDValue Ops[] = { V, V, CurDAG->getRegister(0, MVT::i32), - CurDAG->getTargetConstant(ShImm, MVT::i32), - getAL(CurDAG), CurDAG->getRegister(0, MVT::i32), - CurDAG->getRegister(0, MVT::i32) }; - return CurDAG->SelectNodeTo(N, ARM::ADDrs, MVT::i32, Ops, 7); + 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, Reg0 }; + return CurDAG->SelectNodeTo(N, ARM::t2ADDrs, MVT::i32, Ops, 6); + } else { + SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 }; + return CurDAG->SelectNodeTo(N, ARM::ADDrs, MVT::i32, Ops, 7); + } } if (isPowerOf2_32(RHSV+1)) { // 2^n-1? + unsigned ShImm = Log2_32(RHSV+1); + if (ShImm >= 32) + break; SDValue V = Op.getOperand(0); - unsigned ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, Log2_32(RHSV+1)); - SDValue Ops[] = { V, V, CurDAG->getRegister(0, MVT::i32), - CurDAG->getTargetConstant(ShImm, MVT::i32), - getAL(CurDAG), CurDAG->getRegister(0, MVT::i32), - CurDAG->getRegister(0, MVT::i32) }; - return CurDAG->SelectNodeTo(N, ARM::RSBrs, MVT::i32, Ops, 7); + 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); + } else { + SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 }; + return CurDAG->SelectNodeTo(N, ARM::RSBrs, MVT::i32, Ops, 7); + } } } break; @@ -656,59 +1045,42 @@ SDNode *ARMDAGToDAGISel::Select(SDValue Op) { Op.getOperand(0), getAL(CurDAG), CurDAG->getRegister(0, MVT::i32)); case ISD::UMUL_LOHI: { - SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1), + if (Subtarget->isThumb1Only()) + break; + if (Subtarget->isThumb()) { + SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1), + getAL(CurDAG), CurDAG->getRegister(0, MVT::i32), + CurDAG->getRegister(0, MVT::i32) }; + return CurDAG->getTargetNode(ARM::t2UMULL, dl, MVT::i32, MVT::i32, Ops,4); + } else { + SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1), getAL(CurDAG), CurDAG->getRegister(0, MVT::i32), CurDAG->getRegister(0, MVT::i32) }; - return CurDAG->getTargetNode(ARM::UMULL, dl, MVT::i32, MVT::i32, Ops, 5); + return CurDAG->getTargetNode(ARM::UMULL, dl, MVT::i32, MVT::i32, Ops, 5); + } } case ISD::SMUL_LOHI: { - SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1), + if (Subtarget->isThumb1Only()) + break; + if (Subtarget->isThumb()) { + SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1), + getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) }; + return CurDAG->getTargetNode(ARM::t2SMULL, dl, MVT::i32, MVT::i32, Ops,4); + } else { + SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1), getAL(CurDAG), CurDAG->getRegister(0, MVT::i32), CurDAG->getRegister(0, MVT::i32) }; - return CurDAG->getTargetNode(ARM::SMULL, dl, MVT::i32, MVT::i32, Ops, 5); + return CurDAG->getTargetNode(ARM::SMULL, dl, MVT::i32, MVT::i32, Ops, 5); + } } case ISD::LOAD: { - LoadSDNode *LD = cast(Op); - ISD::MemIndexedMode AM = LD->getAddressingMode(); - MVT LoadedVT = LD->getMemoryVT(); - if (AM != ISD::UNINDEXED) { - SDValue Offset, AMOpc; - bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC); - unsigned Opcode = 0; - bool Match = false; - if (LoadedVT == MVT::i32 && - SelectAddrMode2Offset(Op, 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)) { - 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)) { - Match = true; - Opcode = isPre ? ARM::LDRSB_PRE : ARM::LDRSB_POST; - } - } else { - if (SelectAddrMode2Offset(Op, LD->getOffset(), Offset, AMOpc)) { - Match = true; - Opcode = isPre ? ARM::LDRB_PRE : ARM::LDRB_POST; - } - } - } - - if (Match) { - SDValue Chain = LD->getChain(); - SDValue Base = LD->getBasePtr(); - SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG), - CurDAG->getRegister(0, MVT::i32), Chain }; - return CurDAG->getTargetNode(Opcode, dl, MVT::i32, MVT::i32, - MVT::Other, Ops, 6); - } - } + SDNode *ResNode = 0; + if (Subtarget->isThumb() && Subtarget->hasThumb2()) + ResNode = SelectT2IndexedLoad(Op); + else + ResNode = SelectARMIndexedLoad(Op); + if (ResNode) + return ResNode; // Other cases are autogenerated. break; } @@ -721,7 +1093,12 @@ SDNode *ARMDAGToDAGISel::Select(SDValue Op) { // Emits: (tBcc:void (bb:Other):$dst, (imm:i32):$cc) // Pattern complexity = 6 cost = 1 size = 0 - unsigned Opc = Subtarget->isThumb() ? ARM::tBcc : ARM::Bcc; + // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc) + // Emits: (t2Bcc:void (bb:Other):$dst, (imm:i32):$cc) + // Pattern complexity = 6 cost = 1 size = 0 + + 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); @@ -735,7 +1112,7 @@ SDNode *ARMDAGToDAGISel::Select(SDValue Op) { cast(N2)->getZExtValue()), MVT::i32); SDValue Ops[] = { N1, Tmp2, N3, Chain, InFlag }; - SDNode *ResNode = CurDAG->getTargetNode(Opc, dl, MVT::Other, + SDNode *ResNode = CurDAG->getTargetNode(Opc, dl, MVT::Other, MVT::Flag, Ops, 5); Chain = SDValue(ResNode, 0); if (Op.getNode()->getNumValues() == 2) { @@ -746,8 +1123,7 @@ SDNode *ARMDAGToDAGISel::Select(SDValue Op) { return NULL; } case ARMISD::CMOV: { - bool isThumb = Subtarget->isThumb(); - MVT VT = Op.getValueType(); + EVT VT = Op.getValueType(); SDValue N0 = Op.getOperand(0); SDValue N1 = Op.getOperand(1); SDValue N2 = Op.getOperand(2); @@ -756,39 +1132,79 @@ SDNode *ARMDAGToDAGISel::Select(SDValue Op) { assert(N2.getOpcode() == ISD::Constant); assert(N3.getOpcode() == ISD::Register); - // 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 (!isThumb && VT == MVT::i32 && - SelectShifterOperandReg(Op, N1, CPTmp0, CPTmp1, CPTmp2)) { - SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned) - cast(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:$false, - // (imm:i32)<><>:$true, - // (imm:i32):$cc) - // Emits: (MOVCCi:i32 GPR:i32:$false, - // (so_imm_XFORM:i32 (imm:i32):$true), (imm:i32):$cc) - // Pattern complexity = 10 cost = 1 size = 0 - if (VT == MVT::i32 && - N3.getOpcode() == ISD::Constant && - Predicate_so_imm(N3.getNode())) { - SDValue Tmp1 = CurDAG->getTargetConstant(((unsigned) - cast(N1)->getZExtValue()), - MVT::i32); - Tmp1 = Transform_so_imm_XFORM(Tmp1.getNode()); - SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned) - cast(N2)->getZExtValue()), - MVT::i32); - SDValue Ops[] = { N0, Tmp1, Tmp2, N3, InFlag }; - return CurDAG->SelectNodeTo(Op.getNode(), ARM::MOVCCi, MVT::i32, Ops, 5); + 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(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(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(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:$false, + // (imm:i32)<>:$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(N1)->getZExtValue()), + MVT::i32); + SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned) + cast(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(N1)->getZExtValue()), + MVT::i32); + SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned) + cast(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) @@ -805,23 +1221,25 @@ SDNode *ARMDAGToDAGISel::Select(SDValue Op) { MVT::i32); SDValue Ops[] = { N0, N1, Tmp2, N3, InFlag }; unsigned Opc = 0; - switch (VT.getSimpleVT()) { + switch (VT.getSimpleVT().SimpleTy) { default: assert(false && "Illegal conditional move type!"); break; case MVT::i32: - Opc = isThumb ? ARM::tMOVCCr : ARM::MOVCCr; + 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; + break; } return CurDAG->SelectNodeTo(Op.getNode(), Opc, VT, Ops, 5); } case ARMISD::CNEG: { - MVT VT = Op.getValueType(); + EVT VT = Op.getValueType(); SDValue N0 = Op.getOperand(0); SDValue N1 = Op.getOperand(1); SDValue N2 = Op.getOperand(2); @@ -835,7 +1253,7 @@ SDNode *ARMDAGToDAGISel::Select(SDValue Op) { MVT::i32); SDValue Ops[] = { N0, N1, Tmp2, N3, InFlag }; unsigned Opc = 0; - switch (VT.getSimpleVT()) { + switch (VT.getSimpleVT().SimpleTy) { default: assert(false && "Illegal conditional move type!"); break; case MVT::f32: @@ -880,7 +1298,7 @@ SDNode *ARMDAGToDAGISel::Select(SDValue Op) { ReplaceUses(Op.getValue(0), Chain); return NULL; } - + SDValue Tmp1 = CurDAG->getTargetFrameIndex(FINode->getIndex(), TLI.getPointerTy()); SDValue Tmp2 = CurDAG->getTargetGlobalAddress(GV, TLI.getPointerTy()); @@ -888,6 +1306,213 @@ SDNode *ARMDAGToDAGISel::Select(SDValue Op) { return CurDAG->getTargetNode(TargetInstrInfo::DECLARE, dl, MVT::Other, Ops, 3); } + + case ISD::VECTOR_SHUFFLE: { + EVT VT = Op.getValueType(); + + // Match 128-bit splat to VDUPLANEQ. (This could be done with a Pat in + // ARMInstrNEON.td but it is awkward because the shuffle mask needs to be + // transformed first into a lane number and then to both a subregister + // index and an adjusted lane number.) If the source operand is a + // SCALAR_TO_VECTOR, leave it so it will be matched later as a VDUP. + ShuffleVectorSDNode *SVOp = cast(N); + if (VT.is128BitVector() && SVOp->isSplat() && + Op.getOperand(0).getOpcode() != ISD::SCALAR_TO_VECTOR && + Op.getOperand(1).getOpcode() == ISD::UNDEF) { + unsigned LaneVal = SVOp->getSplatIndex(); + + EVT HalfVT; + unsigned Opc = 0; + switch (VT.getVectorElementType().getSimpleVT().SimpleTy) { + default: llvm_unreachable("unhandled VDUP splat type"); + case MVT::i8: Opc = ARM::VDUPLN8q; HalfVT = MVT::v8i8; break; + case MVT::i16: Opc = ARM::VDUPLN16q; HalfVT = MVT::v4i16; break; + case MVT::i32: Opc = ARM::VDUPLN32q; HalfVT = MVT::v2i32; break; + case MVT::f32: Opc = ARM::VDUPLNfq; HalfVT = MVT::v2f32; break; + } + + // The source operand needs to be changed to a subreg of the original + // 128-bit operand, and the lane number needs to be adjusted accordingly. + unsigned NumElts = VT.getVectorNumElements() / 2; + unsigned SRVal = (LaneVal < NumElts ? arm_dsubreg_0 : arm_dsubreg_1); + SDValue SR = CurDAG->getTargetConstant(SRVal, MVT::i32); + SDValue NewLane = CurDAG->getTargetConstant(LaneVal % NumElts, MVT::i32); + SDNode *SubReg = CurDAG->getTargetNode(TargetInstrInfo::EXTRACT_SUBREG, + dl, HalfVT, N->getOperand(0), SR); + return CurDAG->SelectNodeTo(N, Opc, VT, SDValue(SubReg, 0), NewLane); + } + + break; + } + + case ARMISD::VLD2D: { + SDValue MemAddr, MemUpdate, MemOpc; + if (!SelectAddrMode6(Op, N->getOperand(1), MemAddr, MemUpdate, MemOpc)) + return NULL; + unsigned Opc = 0; + EVT VT = Op.getValueType(); + switch (VT.getSimpleVT().SimpleTy) { + default: llvm_unreachable("unhandled VLD2D 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; + } + SDValue Chain = N->getOperand(0); + const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, Chain }; + return CurDAG->getTargetNode(Opc, dl, VT, VT, MVT::Other, Ops, 4); + } + + case ARMISD::VLD3D: { + SDValue MemAddr, MemUpdate, MemOpc; + if (!SelectAddrMode6(Op, N->getOperand(1), MemAddr, MemUpdate, MemOpc)) + return NULL; + unsigned Opc = 0; + EVT VT = Op.getValueType(); + switch (VT.getSimpleVT().SimpleTy) { + default: llvm_unreachable("unhandled VLD3D 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; + } + SDValue Chain = N->getOperand(0); + const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, Chain }; + return CurDAG->getTargetNode(Opc, dl, VT, VT, VT, MVT::Other, Ops, 4); + } + + case ARMISD::VLD4D: { + SDValue MemAddr, MemUpdate, MemOpc; + if (!SelectAddrMode6(Op, N->getOperand(1), MemAddr, MemUpdate, MemOpc)) + return NULL; + unsigned Opc = 0; + EVT VT = Op.getValueType(); + switch (VT.getSimpleVT().SimpleTy) { + default: llvm_unreachable("unhandled VLD4D 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; + } + SDValue Chain = N->getOperand(0); + const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, Chain }; + std::vector ResTys(4, VT); + ResTys.push_back(MVT::Other); + return CurDAG->getTargetNode(Opc, dl, ResTys, Ops, 4); + } + + case ARMISD::VST2D: { + SDValue MemAddr, MemUpdate, MemOpc; + if (!SelectAddrMode6(Op, N->getOperand(1), MemAddr, MemUpdate, MemOpc)) + return NULL; + unsigned Opc = 0; + switch (N->getOperand(2).getValueType().getSimpleVT().SimpleTy) { + default: llvm_unreachable("unhandled VST2D 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; + } + SDValue Chain = N->getOperand(0); + const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, + N->getOperand(2), N->getOperand(3), Chain }; + return CurDAG->getTargetNode(Opc, dl, MVT::Other, Ops, 6); + } + + case ARMISD::VST3D: { + SDValue MemAddr, MemUpdate, MemOpc; + if (!SelectAddrMode6(Op, N->getOperand(1), MemAddr, MemUpdate, MemOpc)) + return NULL; + unsigned Opc = 0; + switch (N->getOperand(2).getValueType().getSimpleVT().SimpleTy) { + default: llvm_unreachable("unhandled VST3D 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; + } + SDValue Chain = N->getOperand(0); + const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, + N->getOperand(2), N->getOperand(3), + N->getOperand(4), Chain }; + return CurDAG->getTargetNode(Opc, dl, MVT::Other, Ops, 7); + } + + case ARMISD::VST4D: { + SDValue MemAddr, MemUpdate, MemOpc; + if (!SelectAddrMode6(Op, N->getOperand(1), MemAddr, MemUpdate, MemOpc)) + return NULL; + unsigned Opc = 0; + switch (N->getOperand(2).getValueType().getSimpleVT().SimpleTy) { + default: llvm_unreachable("unhandled VST4D 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; + } + SDValue Chain = N->getOperand(0); + const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, + N->getOperand(2), N->getOperand(3), + N->getOperand(4), N->getOperand(5), Chain }; + return CurDAG->getTargetNode(Opc, dl, MVT::Other, Ops, 8); + } + + case ISD::INTRINSIC_WO_CHAIN: { + unsigned IntNo = cast(N->getOperand(0))->getZExtValue(); + EVT VT = N->getValueType(0); + unsigned Opc = 0; + + // Match intrinsics that return multiple values. + switch (IntNo) { + default: break; + + case Intrinsic::arm_neon_vtrn: + switch (VT.getSimpleVT().SimpleTy) { + default: return NULL; + case MVT::v8i8: Opc = ARM::VTRNd8; break; + case MVT::v4i16: Opc = ARM::VTRNd16; break; + case MVT::v2f32: + case MVT::v2i32: Opc = ARM::VTRNd32; break; + case MVT::v16i8: Opc = ARM::VTRNq8; break; + case MVT::v8i16: Opc = ARM::VTRNq16; break; + case MVT::v4f32: + case MVT::v4i32: Opc = ARM::VTRNq32; break; + } + return CurDAG->getTargetNode(Opc, dl, VT, VT, N->getOperand(1), + N->getOperand(2)); + + case Intrinsic::arm_neon_vuzp: + switch (VT.getSimpleVT().SimpleTy) { + default: return NULL; + case MVT::v8i8: Opc = ARM::VUZPd8; break; + case MVT::v4i16: Opc = ARM::VUZPd16; break; + case MVT::v2f32: + case MVT::v2i32: Opc = ARM::VUZPd32; break; + case MVT::v16i8: Opc = ARM::VUZPq8; break; + case MVT::v8i16: Opc = ARM::VUZPq16; break; + case MVT::v4f32: + case MVT::v4i32: Opc = ARM::VUZPq32; break; + } + return CurDAG->getTargetNode(Opc, dl, VT, VT, N->getOperand(1), + N->getOperand(2)); + + case Intrinsic::arm_neon_vzip: + switch (VT.getSimpleVT().SimpleTy) { + default: return NULL; + case MVT::v8i8: Opc = ARM::VZIPd8; break; + case MVT::v4i16: Opc = ARM::VZIPd16; break; + case MVT::v2f32: + case MVT::v2i32: Opc = ARM::VZIPd32; break; + case MVT::v16i8: Opc = ARM::VZIPq8; break; + case MVT::v8i16: Opc = ARM::VZIPq16; break; + case MVT::v4f32: + case MVT::v4i32: Opc = ARM::VZIPq32; break; + } + return CurDAG->getTargetNode(Opc, dl, VT, VT, N->getOperand(1), + N->getOperand(2)); + } + break; + } } return SelectCode(Op); @@ -901,7 +1526,7 @@ SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode, 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); @@ -911,6 +1536,6 @@ SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode, /// createARMISelDag - This pass converts a legalized DAG into a /// ARM-specific DAG, ready for instruction scheduling. /// -FunctionPass *llvm::createARMISelDag(ARMTargetMachine &TM) { +FunctionPass *llvm::createARMISelDag(ARMBaseTargetMachine &TM) { return new ARMDAGToDAGISel(TM); }