X-Git-Url: http://plrg.eecs.uci.edu/git/?p=oota-llvm.git;a=blobdiff_plain;f=lib%2FTarget%2FAArch64%2FAArch64FastISel.cpp;h=d62692848014f64546c3ff24e5660d8c3d806f5d;hp=d80f34ca2d360ed8e5f105a79982b75fe64aa889;hb=5745cad86144b8a3d47f55568daff433033cf202;hpb=450c33b212860a3d7e2be7aa23efba40665430bb diff --git a/lib/Target/AArch64/AArch64FastISel.cpp b/lib/Target/AArch64/AArch64FastISel.cpp index d80f34ca2d3..d6269284801 100644 --- a/lib/Target/AArch64/AArch64FastISel.cpp +++ b/lib/Target/AArch64/AArch64FastISel.cpp @@ -17,6 +17,7 @@ #include "AArch64Subtarget.h" #include "AArch64TargetMachine.h" #include "MCTargetDesc/AArch64AddressingModes.h" +#include "llvm/Analysis/BranchProbabilityInfo.h" #include "llvm/CodeGen/CallingConvLower.h" #include "llvm/CodeGen/FastISel.h" #include "llvm/CodeGen/FunctionLoweringInfo.h" @@ -39,8 +40,7 @@ using namespace llvm; namespace { -class AArch64FastISel : public FastISel { - +class AArch64FastISel final : public FastISel { class Address { public: typedef enum { @@ -50,16 +50,23 @@ class AArch64FastISel : public FastISel { private: BaseKind Kind; + AArch64_AM::ShiftExtendType ExtType; union { unsigned Reg; int FI; } Base; + unsigned OffsetReg; + unsigned Shift; int64_t Offset; + const GlobalValue *GV; public: - Address() : Kind(RegBase), Offset(0) { Base.Reg = 0; } + Address() : Kind(RegBase), ExtType(AArch64_AM::InvalidShiftExtend), + OffsetReg(0), Shift(0), Offset(0), GV(nullptr) { Base.Reg = 0; } void setKind(BaseKind K) { Kind = K; } BaseKind getKind() const { return Kind; } + void setExtendType(AArch64_AM::ShiftExtendType E) { ExtType = E; } + AArch64_AM::ShiftExtendType getExtendType() const { return ExtType; } bool isRegBase() const { return Kind == RegBase; } bool isFIBase() const { return Kind == FrameIndexBase; } void setReg(unsigned Reg) { @@ -70,6 +77,12 @@ class AArch64FastISel : public FastISel { assert(isRegBase() && "Invalid base register access!"); return Base.Reg; } + void setOffsetReg(unsigned Reg) { + OffsetReg = Reg; + } + unsigned getOffsetReg() const { + return OffsetReg; + } void setFI(unsigned FI) { assert(isFIBase() && "Invalid base frame index access!"); Base.FI = FI; @@ -80,8 +93,11 @@ class AArch64FastISel : public FastISel { } void setOffset(int64_t O) { Offset = O; } int64_t getOffset() { return Offset; } + void setShift(unsigned S) { Shift = S; } + unsigned getShift() { return Shift; } - bool isValid() { return isFIBase() || (isRegBase() && getReg() != 0); } + void setGlobalValue(const GlobalValue *G) { GV = G; } + const GlobalValue *getGlobalValue() { return GV; } }; /// Subtarget - Keep a pointer to the AArch64Subtarget around so that we can @@ -89,83 +105,150 @@ class AArch64FastISel : public FastISel { const AArch64Subtarget *Subtarget; LLVMContext *Context; - bool FastLowerCall(CallLoweringInfo &CLI) override; - bool FastLowerIntrinsicCall(const IntrinsicInst *II) override; + bool fastLowerArguments() override; + bool fastLowerCall(CallLoweringInfo &CLI) override; + bool fastLowerIntrinsicCall(const IntrinsicInst *II) override; private: // Selection routines. - bool SelectLoad(const Instruction *I); - bool SelectStore(const Instruction *I); - bool SelectBranch(const Instruction *I); - bool SelectIndirectBr(const Instruction *I); - bool SelectCmp(const Instruction *I); - bool SelectSelect(const Instruction *I); - bool SelectFPExt(const Instruction *I); - bool SelectFPTrunc(const Instruction *I); - bool SelectFPToInt(const Instruction *I, bool Signed); - bool SelectIntToFP(const Instruction *I, bool Signed); - bool SelectRem(const Instruction *I, unsigned ISDOpcode); - bool SelectRet(const Instruction *I); - bool SelectTrunc(const Instruction *I); - bool SelectIntExt(const Instruction *I); - bool SelectMul(const Instruction *I); - bool SelectShift(const Instruction *I, bool IsLeftShift, bool IsArithmetic); + bool selectAddSub(const Instruction *I); + bool selectLogicalOp(const Instruction *I); + bool selectLoad(const Instruction *I); + bool selectStore(const Instruction *I); + bool selectBranch(const Instruction *I); + bool selectIndirectBr(const Instruction *I); + bool selectCmp(const Instruction *I); + bool selectSelect(const Instruction *I); + bool selectFPExt(const Instruction *I); + bool selectFPTrunc(const Instruction *I); + bool selectFPToInt(const Instruction *I, bool Signed); + bool selectIntToFP(const Instruction *I, bool Signed); + bool selectRem(const Instruction *I, unsigned ISDOpcode); + bool selectRet(const Instruction *I); + bool selectTrunc(const Instruction *I); + bool selectIntExt(const Instruction *I); + bool selectMul(const Instruction *I); + bool selectShift(const Instruction *I); + bool selectBitCast(const Instruction *I); + bool selectFRem(const Instruction *I); + bool selectSDiv(const Instruction *I); + bool selectGetElementPtr(const Instruction *I); // Utility helper routines. bool isTypeLegal(Type *Ty, MVT &VT); - bool isLoadStoreTypeLegal(Type *Ty, MVT &VT); - bool ComputeAddress(const Value *Obj, Address &Addr); - bool SimplifyAddress(Address &Addr, MVT VT, int64_t ScaleFactor, - bool UseUnscaled); - void AddLoadStoreOperands(Address &Addr, const MachineInstrBuilder &MIB, - unsigned Flags, bool UseUnscaled); - bool IsMemCpySmall(uint64_t Len, unsigned Alignment); - bool TryEmitSmallMemCpy(Address Dest, Address Src, uint64_t Len, + bool isTypeSupported(Type *Ty, MVT &VT, bool IsVectorAllowed = false); + bool isValueAvailable(const Value *V) const; + bool computeAddress(const Value *Obj, Address &Addr, Type *Ty = nullptr); + bool computeCallAddress(const Value *V, Address &Addr); + bool simplifyAddress(Address &Addr, MVT VT); + void addLoadStoreOperands(Address &Addr, const MachineInstrBuilder &MIB, + unsigned Flags, unsigned ScaleFactor, + MachineMemOperand *MMO); + bool isMemCpySmall(uint64_t Len, unsigned Alignment); + bool tryEmitSmallMemCpy(Address Dest, Address Src, uint64_t Len, unsigned Alignment); bool foldXALUIntrinsic(AArch64CC::CondCode &CC, const Instruction *I, const Value *Cond); + bool optimizeIntExtLoad(const Instruction *I, MVT RetVT, MVT SrcVT); + + // Emit helper routines. + unsigned emitAddSub(bool UseAdd, MVT RetVT, const Value *LHS, + const Value *RHS, bool SetFlags = false, + bool WantResult = true, bool IsZExt = false); + unsigned emitAddSub_rr(bool UseAdd, MVT RetVT, unsigned LHSReg, + bool LHSIsKill, unsigned RHSReg, bool RHSIsKill, + bool SetFlags = false, bool WantResult = true); + unsigned emitAddSub_ri(bool UseAdd, MVT RetVT, unsigned LHSReg, + bool LHSIsKill, uint64_t Imm, bool SetFlags = false, + bool WantResult = true); + unsigned emitAddSub_rs(bool UseAdd, MVT RetVT, unsigned LHSReg, + bool LHSIsKill, unsigned RHSReg, bool RHSIsKill, + AArch64_AM::ShiftExtendType ShiftType, + uint64_t ShiftImm, bool SetFlags = false, + bool WantResult = true); + unsigned emitAddSub_rx(bool UseAdd, MVT RetVT, unsigned LHSReg, + bool LHSIsKill, unsigned RHSReg, bool RHSIsKill, + AArch64_AM::ShiftExtendType ExtType, + uint64_t ShiftImm, bool SetFlags = false, + bool WantResult = true); // Emit functions. - bool EmitCmp(Value *Src1Value, Value *Src2Value, bool isZExt); - bool EmitLoad(MVT VT, unsigned &ResultReg, Address Addr, - bool UseUnscaled = false); - bool EmitStore(MVT VT, unsigned SrcReg, Address Addr, - bool UseUnscaled = false); - unsigned EmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, bool isZExt); - unsigned Emiti1Ext(unsigned SrcReg, MVT DestVT, bool isZExt); - unsigned Emit_MUL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill, - unsigned Op1, bool Op1IsKill); - unsigned Emit_SMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill, - unsigned Op1, bool Op1IsKill); - unsigned Emit_UMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill, - unsigned Op1, bool Op1IsKill); - unsigned Emit_LSL_ri(MVT RetVT, unsigned Op0, bool Op0IsKill, uint64_t Imm); - unsigned Emit_LSR_ri(MVT RetVT, unsigned Op0, bool Op0IsKill, uint64_t Imm); - unsigned Emit_ASR_ri(MVT RetVT, unsigned Op0, bool Op0IsKill, uint64_t Imm); - - unsigned AArch64MaterializeFP(const ConstantFP *CFP, MVT VT); - unsigned AArch64MaterializeGV(const GlobalValue *GV); + bool emitCompareAndBranch(const BranchInst *BI); + bool emitCmp(const Value *LHS, const Value *RHS, bool IsZExt); + bool emitICmp(MVT RetVT, const Value *LHS, const Value *RHS, bool IsZExt); + bool emitICmp_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill, uint64_t Imm); + bool emitFCmp(MVT RetVT, const Value *LHS, const Value *RHS); + unsigned emitLoad(MVT VT, MVT ResultVT, Address Addr, bool WantZExt = true, + MachineMemOperand *MMO = nullptr); + bool emitStore(MVT VT, unsigned SrcReg, Address Addr, + MachineMemOperand *MMO = nullptr); + unsigned emitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, bool isZExt); + unsigned emiti1Ext(unsigned SrcReg, MVT DestVT, bool isZExt); + unsigned emitAdd(MVT RetVT, const Value *LHS, const Value *RHS, + bool SetFlags = false, bool WantResult = true, + bool IsZExt = false); + unsigned emitAdd_ri_(MVT VT, unsigned Op0, bool Op0IsKill, int64_t Imm); + unsigned emitSub(MVT RetVT, const Value *LHS, const Value *RHS, + bool SetFlags = false, bool WantResult = true, + bool IsZExt = false); + unsigned emitSubs_rr(MVT RetVT, unsigned LHSReg, bool LHSIsKill, + unsigned RHSReg, bool RHSIsKill, bool WantResult = true); + unsigned emitSubs_rs(MVT RetVT, unsigned LHSReg, bool LHSIsKill, + unsigned RHSReg, bool RHSIsKill, + AArch64_AM::ShiftExtendType ShiftType, uint64_t ShiftImm, + bool WantResult = true); + unsigned emitLogicalOp(unsigned ISDOpc, MVT RetVT, const Value *LHS, + const Value *RHS); + unsigned emitLogicalOp_ri(unsigned ISDOpc, MVT RetVT, unsigned LHSReg, + bool LHSIsKill, uint64_t Imm); + unsigned emitLogicalOp_rs(unsigned ISDOpc, MVT RetVT, unsigned LHSReg, + bool LHSIsKill, unsigned RHSReg, bool RHSIsKill, + uint64_t ShiftImm); + unsigned emitAnd_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill, uint64_t Imm); + unsigned emitMul_rr(MVT RetVT, unsigned Op0, bool Op0IsKill, + unsigned Op1, bool Op1IsKill); + unsigned emitSMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill, + unsigned Op1, bool Op1IsKill); + unsigned emitUMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill, + unsigned Op1, bool Op1IsKill); + unsigned emitLSL_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill, + unsigned Op1Reg, bool Op1IsKill); + unsigned emitLSL_ri(MVT RetVT, MVT SrcVT, unsigned Op0Reg, bool Op0IsKill, + uint64_t Imm, bool IsZExt = true); + unsigned emitLSR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill, + unsigned Op1Reg, bool Op1IsKill); + unsigned emitLSR_ri(MVT RetVT, MVT SrcVT, unsigned Op0Reg, bool Op0IsKill, + uint64_t Imm, bool IsZExt = true); + unsigned emitASR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill, + unsigned Op1Reg, bool Op1IsKill); + unsigned emitASR_ri(MVT RetVT, MVT SrcVT, unsigned Op0Reg, bool Op0IsKill, + uint64_t Imm, bool IsZExt = false); + + unsigned materializeInt(const ConstantInt *CI, MVT VT); + unsigned materializeFP(const ConstantFP *CFP, MVT VT); + unsigned materializeGV(const GlobalValue *GV); // Call handling routines. private: CCAssignFn *CCAssignFnForCall(CallingConv::ID CC) const; - bool ProcessCallArgs(CallLoweringInfo &CLI, SmallVectorImpl &ArgVTs, + bool processCallArgs(CallLoweringInfo &CLI, SmallVectorImpl &ArgVTs, unsigned &NumBytes); - bool FinishCall(CallLoweringInfo &CLI, MVT RetVT, unsigned NumBytes); + bool finishCall(CallLoweringInfo &CLI, MVT RetVT, unsigned NumBytes); public: // Backend specific FastISel code. - unsigned TargetMaterializeAlloca(const AllocaInst *AI) override; - unsigned TargetMaterializeConstant(const Constant *C) override; + unsigned fastMaterializeAlloca(const AllocaInst *AI) override; + unsigned fastMaterializeConstant(const Constant *C) override; + unsigned fastMaterializeFloatZero(const ConstantFP* CF) override; - explicit AArch64FastISel(FunctionLoweringInfo &funcInfo, - const TargetLibraryInfo *libInfo) - : FastISel(funcInfo, libInfo) { + explicit AArch64FastISel(FunctionLoweringInfo &FuncInfo, + const TargetLibraryInfo *LibInfo) + : FastISel(FuncInfo, LibInfo, /*SkipTargetIndependentISel=*/true) { Subtarget = &TM.getSubtarget(); - Context = &funcInfo.Fn->getContext(); + Context = &FuncInfo.Fn->getContext(); } - bool TargetSelectInstruction(const Instruction *I) override; + bool fastSelectInstruction(const Instruction *I) override; #include "AArch64GenFastISel.inc" }; @@ -174,13 +257,52 @@ public: #include "AArch64GenCallingConv.inc" +/// \brief Check if the sign-/zero-extend will be a noop. +static bool isIntExtFree(const Instruction *I) { + assert((isa(I) || isa(I)) && + "Unexpected integer extend instruction."); + assert(!I->getType()->isVectorTy() && I->getType()->isIntegerTy() && + "Unexpected value type."); + bool IsZExt = isa(I); + + if (const auto *LI = dyn_cast(I->getOperand(0))) + if (LI->hasOneUse()) + return true; + + if (const auto *Arg = dyn_cast(I->getOperand(0))) + if ((IsZExt && Arg->hasZExtAttr()) || (!IsZExt && Arg->hasSExtAttr())) + return true; + + return false; +} + +/// \brief Determine the implicit scale factor that is applied by a memory +/// operation for a given value type. +static unsigned getImplicitScaleFactor(MVT VT) { + switch (VT.SimpleTy) { + default: + return 0; // invalid + case MVT::i1: // fall-through + case MVT::i8: + return 1; + case MVT::i16: + return 2; + case MVT::i32: // fall-through + case MVT::f32: + return 4; + case MVT::i64: // fall-through + case MVT::f64: + return 8; + } +} + CCAssignFn *AArch64FastISel::CCAssignFnForCall(CallingConv::ID CC) const { if (CC == CallingConv::WebKit_JS) return CC_AArch64_WebKit_JS; return Subtarget->isTargetDarwin() ? CC_AArch64_DarwinPCS : CC_AArch64_AAPCS; } -unsigned AArch64FastISel::TargetMaterializeAlloca(const AllocaInst *AI) { +unsigned AArch64FastISel::fastMaterializeAlloca(const AllocaInst *AI) { assert(TLI.getValueType(AI->getType(), true) == MVT::i64 && "Alloca should always return a pointer."); @@ -192,7 +314,7 @@ unsigned AArch64FastISel::TargetMaterializeAlloca(const AllocaInst *AI) { FuncInfo.StaticAllocaMap.find(AI); if (SI != FuncInfo.StaticAllocaMap.end()) { - unsigned ResultReg = createResultReg(&AArch64::GPR64RegClass); + unsigned ResultReg = createResultReg(&AArch64::GPR64spRegClass); BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri), ResultReg) .addFrameIndex(SI->second) @@ -204,29 +326,42 @@ unsigned AArch64FastISel::TargetMaterializeAlloca(const AllocaInst *AI) { return 0; } -unsigned AArch64FastISel::AArch64MaterializeFP(const ConstantFP *CFP, MVT VT) { +unsigned AArch64FastISel::materializeInt(const ConstantInt *CI, MVT VT) { + if (VT > MVT::i64) + return 0; + + if (!CI->isZero()) + return fastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue()); + + // Create a copy from the zero register to materialize a "0" value. + const TargetRegisterClass *RC = (VT == MVT::i64) ? &AArch64::GPR64RegClass + : &AArch64::GPR32RegClass; + unsigned ZeroReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR; + unsigned ResultReg = createResultReg(RC); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::COPY), + ResultReg).addReg(ZeroReg, getKillRegState(true)); + return ResultReg; +} + +unsigned AArch64FastISel::materializeFP(const ConstantFP *CFP, MVT VT) { + // Positive zero (+0.0) has to be materialized with a fmov from the zero + // register, because the immediate version of fmov cannot encode zero. + if (CFP->isNullValue()) + return fastMaterializeFloatZero(CFP); + if (VT != MVT::f32 && VT != MVT::f64) return 0; const APFloat Val = CFP->getValueAPF(); - bool is64bit = (VT == MVT::f64); - + bool Is64Bit = (VT == MVT::f64); // This checks to see if we can use FMOV instructions to materialize // a constant, otherwise we have to materialize via the constant pool. if (TLI.isFPImmLegal(Val, VT)) { - int Imm; - unsigned Opc; - if (is64bit) { - Imm = AArch64_AM::getFP64Imm(Val); - Opc = AArch64::FMOVDi; - } else { - Imm = AArch64_AM::getFP32Imm(Val); - Opc = AArch64::FMOVSi; - } - unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT)); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg) - .addImm(Imm); - return ResultReg; + int Imm = + Is64Bit ? AArch64_AM::getFP64Imm(Val) : AArch64_AM::getFP32Imm(Val); + assert((Imm != -1) && "Cannot encode floating-point constant."); + unsigned Opc = Is64Bit ? AArch64::FMOVDi : AArch64::FMOVSi; + return fastEmitInst_i(Opc, TLI.getRegClassFor(VT), Imm); } // Materialize via constant pool. MachineConstantPool wants an explicit @@ -235,20 +370,20 @@ unsigned AArch64FastISel::AArch64MaterializeFP(const ConstantFP *CFP, MVT VT) { if (Align == 0) Align = DL.getTypeAllocSize(CFP->getType()); - unsigned Idx = MCP.getConstantPoolIndex(cast(CFP), Align); + unsigned CPI = MCP.getConstantPoolIndex(cast(CFP), Align); unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass); BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP), - ADRPReg).addConstantPoolIndex(Idx, 0, AArch64II::MO_PAGE); + ADRPReg).addConstantPoolIndex(CPI, 0, AArch64II::MO_PAGE); - unsigned Opc = is64bit ? AArch64::LDRDui : AArch64::LDRSui; + unsigned Opc = Is64Bit ? AArch64::LDRDui : AArch64::LDRSui; unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT)); BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg) .addReg(ADRPReg) - .addConstantPoolIndex(Idx, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC); + .addConstantPoolIndex(CPI, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC); return ResultReg; } -unsigned AArch64FastISel::AArch64MaterializeGV(const GlobalValue *GV) { +unsigned AArch64FastISel::materializeGV(const GlobalValue *GV) { // We can't handle thread-local variables quickly yet. if (GV->isThreadLocal()) return 0; @@ -271,30 +406,34 @@ unsigned AArch64FastISel::AArch64MaterializeGV(const GlobalValue *GV) { // ADRP + LDRX BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP), ADRPReg) - .addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGE); + .addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGE); ResultReg = createResultReg(&AArch64::GPR64RegClass); BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::LDRXui), ResultReg) - .addReg(ADRPReg) - .addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGEOFF | - AArch64II::MO_NC); + .addReg(ADRPReg) + .addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGEOFF | + AArch64II::MO_NC); + } else if (OpFlags & AArch64II::MO_CONSTPOOL) { + // We can't handle addresses loaded from a constant pool quickly yet. + return 0; } else { // ADRP + ADDX BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP), - ADRPReg).addGlobalAddress(GV, 0, AArch64II::MO_PAGE); + ADRPReg) + .addGlobalAddress(GV, 0, AArch64II::MO_PAGE); ResultReg = createResultReg(&AArch64::GPR64spRegClass); BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri), ResultReg) - .addReg(ADRPReg) - .addGlobalAddress(GV, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC) - .addImm(0); + .addReg(ADRPReg) + .addGlobalAddress(GV, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC) + .addImm(0); } return ResultReg; } -unsigned AArch64FastISel::TargetMaterializeConstant(const Constant *C) { +unsigned AArch64FastISel::fastMaterializeConstant(const Constant *C) { EVT CEVT = TLI.getValueType(C->getType(), true); // Only handle simple types. @@ -302,17 +441,48 @@ unsigned AArch64FastISel::TargetMaterializeConstant(const Constant *C) { return 0; MVT VT = CEVT.getSimpleVT(); - // FIXME: Handle ConstantInt. - if (const ConstantFP *CFP = dyn_cast(C)) - return AArch64MaterializeFP(CFP, VT); + if (const auto *CI = dyn_cast(C)) + return materializeInt(CI, VT); + else if (const ConstantFP *CFP = dyn_cast(C)) + return materializeFP(CFP, VT); else if (const GlobalValue *GV = dyn_cast(C)) - return AArch64MaterializeGV(GV); + return materializeGV(GV); return 0; } +unsigned AArch64FastISel::fastMaterializeFloatZero(const ConstantFP* CFP) { + assert(CFP->isNullValue() && + "Floating-point constant is not a positive zero."); + MVT VT; + if (!isTypeLegal(CFP->getType(), VT)) + return 0; + + if (VT != MVT::f32 && VT != MVT::f64) + return 0; + + bool Is64Bit = (VT == MVT::f64); + unsigned ZReg = Is64Bit ? AArch64::XZR : AArch64::WZR; + unsigned Opc = Is64Bit ? AArch64::FMOVXDr : AArch64::FMOVWSr; + return fastEmitInst_r(Opc, TLI.getRegClassFor(VT), ZReg, /*IsKill=*/true); +} + +/// \brief Check if the multiply is by a power-of-2 constant. +static bool isMulPowOf2(const Value *I) { + if (const auto *MI = dyn_cast(I)) { + if (const auto *C = dyn_cast(MI->getOperand(0))) + if (C->getValue().isPowerOf2()) + return true; + if (const auto *C = dyn_cast(MI->getOperand(1))) + if (C->getValue().isPowerOf2()) + return true; + } + return false; +} + // Computes the address to get to an object. -bool AArch64FastISel::ComputeAddress(const Value *Obj, Address &Addr) { +bool AArch64FastISel::computeAddress(const Value *Obj, Address &Addr, Type *Ty) +{ const User *U = nullptr; unsigned Opcode = Instruction::UserOp1; if (const Instruction *I = dyn_cast(Obj)) { @@ -339,18 +509,18 @@ bool AArch64FastISel::ComputeAddress(const Value *Obj, Address &Addr) { break; case Instruction::BitCast: { // Look through bitcasts. - return ComputeAddress(U->getOperand(0), Addr); + return computeAddress(U->getOperand(0), Addr, Ty); } case Instruction::IntToPtr: { // Look past no-op inttoptrs. if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy()) - return ComputeAddress(U->getOperand(0), Addr); + return computeAddress(U->getOperand(0), Addr, Ty); break; } case Instruction::PtrToInt: { // Look past no-op ptrtoints. if (TLI.getValueType(U->getType()) == TLI.getPointerTy()) - return ComputeAddress(U->getOperand(0), Addr); + return computeAddress(U->getOperand(0), Addr, Ty); break; } case Instruction::GetElementPtr: { @@ -392,7 +562,7 @@ bool AArch64FastISel::ComputeAddress(const Value *Obj, Address &Addr) { // Try to grab the base operand now. Addr.setOffset(TmpOffset); - if (ComputeAddress(U->getOperand(0), Addr)) + if (computeAddress(U->getOperand(0), Addr, Ty)) return true; // We failed, restore everything and try the other options. @@ -412,14 +582,301 @@ bool AArch64FastISel::ComputeAddress(const Value *Obj, Address &Addr) { } break; } + case Instruction::Add: { + // Adds of constants are common and easy enough. + const Value *LHS = U->getOperand(0); + const Value *RHS = U->getOperand(1); + + if (isa(LHS)) + std::swap(LHS, RHS); + + if (const ConstantInt *CI = dyn_cast(RHS)) { + Addr.setOffset(Addr.getOffset() + CI->getSExtValue()); + return computeAddress(LHS, Addr, Ty); + } + + Address Backup = Addr; + if (computeAddress(LHS, Addr, Ty) && computeAddress(RHS, Addr, Ty)) + return true; + Addr = Backup; + + break; + } + case Instruction::Sub: { + // Subs of constants are common and easy enough. + const Value *LHS = U->getOperand(0); + const Value *RHS = U->getOperand(1); + + if (const ConstantInt *CI = dyn_cast(RHS)) { + Addr.setOffset(Addr.getOffset() - CI->getSExtValue()); + return computeAddress(LHS, Addr, Ty); + } + break; + } + case Instruction::Shl: { + if (Addr.getOffsetReg()) + break; + + const auto *CI = dyn_cast(U->getOperand(1)); + if (!CI) + break; + + unsigned Val = CI->getZExtValue(); + if (Val < 1 || Val > 3) + break; + + uint64_t NumBytes = 0; + if (Ty && Ty->isSized()) { + uint64_t NumBits = DL.getTypeSizeInBits(Ty); + NumBytes = NumBits / 8; + if (!isPowerOf2_64(NumBits)) + NumBytes = 0; + } + + if (NumBytes != (1ULL << Val)) + break; + + Addr.setShift(Val); + Addr.setExtendType(AArch64_AM::LSL); + + const Value *Src = U->getOperand(0); + if (const auto *I = dyn_cast(Src)) + if (FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) + Src = I; + + // Fold the zext or sext when it won't become a noop. + if (const auto *ZE = dyn_cast(Src)) { + if (!isIntExtFree(ZE) && ZE->getOperand(0)->getType()->isIntegerTy(32)) { + Addr.setExtendType(AArch64_AM::UXTW); + Src = ZE->getOperand(0); + } + } else if (const auto *SE = dyn_cast(Src)) { + if (!isIntExtFree(SE) && SE->getOperand(0)->getType()->isIntegerTy(32)) { + Addr.setExtendType(AArch64_AM::SXTW); + Src = SE->getOperand(0); + } + } + + if (const auto *AI = dyn_cast(Src)) + if (AI->getOpcode() == Instruction::And) { + const Value *LHS = AI->getOperand(0); + const Value *RHS = AI->getOperand(1); + + if (const auto *C = dyn_cast(LHS)) + if (C->getValue() == 0xffffffff) + std::swap(LHS, RHS); + + if (const auto *C = dyn_cast(RHS)) + if (C->getValue() == 0xffffffff) { + Addr.setExtendType(AArch64_AM::UXTW); + unsigned Reg = getRegForValue(LHS); + if (!Reg) + return false; + bool RegIsKill = hasTrivialKill(LHS); + Reg = fastEmitInst_extractsubreg(MVT::i32, Reg, RegIsKill, + AArch64::sub_32); + Addr.setOffsetReg(Reg); + return true; + } + } + + unsigned Reg = getRegForValue(Src); + if (!Reg) + return false; + Addr.setOffsetReg(Reg); + return true; + } + case Instruction::Mul: { + if (Addr.getOffsetReg()) + break; + + if (!isMulPowOf2(U)) + break; + + const Value *LHS = U->getOperand(0); + const Value *RHS = U->getOperand(1); + + // Canonicalize power-of-2 value to the RHS. + if (const auto *C = dyn_cast(LHS)) + if (C->getValue().isPowerOf2()) + std::swap(LHS, RHS); + + assert(isa(RHS) && "Expected an ConstantInt."); + const auto *C = cast(RHS); + unsigned Val = C->getValue().logBase2(); + if (Val < 1 || Val > 3) + break; + + uint64_t NumBytes = 0; + if (Ty && Ty->isSized()) { + uint64_t NumBits = DL.getTypeSizeInBits(Ty); + NumBytes = NumBits / 8; + if (!isPowerOf2_64(NumBits)) + NumBytes = 0; + } + + if (NumBytes != (1ULL << Val)) + break; + + Addr.setShift(Val); + Addr.setExtendType(AArch64_AM::LSL); + + const Value *Src = LHS; + if (const auto *I = dyn_cast(Src)) + if (FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) + Src = I; + + + // Fold the zext or sext when it won't become a noop. + if (const auto *ZE = dyn_cast(Src)) { + if (!isIntExtFree(ZE) && ZE->getOperand(0)->getType()->isIntegerTy(32)) { + Addr.setExtendType(AArch64_AM::UXTW); + Src = ZE->getOperand(0); + } + } else if (const auto *SE = dyn_cast(Src)) { + if (!isIntExtFree(SE) && SE->getOperand(0)->getType()->isIntegerTy(32)) { + Addr.setExtendType(AArch64_AM::SXTW); + Src = SE->getOperand(0); + } + } + + unsigned Reg = getRegForValue(Src); + if (!Reg) + return false; + Addr.setOffsetReg(Reg); + return true; + } + case Instruction::And: { + if (Addr.getOffsetReg()) + break; + + if (DL.getTypeSizeInBits(Ty) != 8) + break; + + const Value *LHS = U->getOperand(0); + const Value *RHS = U->getOperand(1); + + if (const auto *C = dyn_cast(LHS)) + if (C->getValue() == 0xffffffff) + std::swap(LHS, RHS); + + if (const auto *C = dyn_cast(RHS)) + if (C->getValue() == 0xffffffff) { + Addr.setShift(0); + Addr.setExtendType(AArch64_AM::LSL); + Addr.setExtendType(AArch64_AM::UXTW); + + unsigned Reg = getRegForValue(LHS); + if (!Reg) + return false; + bool RegIsKill = hasTrivialKill(LHS); + Reg = fastEmitInst_extractsubreg(MVT::i32, Reg, RegIsKill, + AArch64::sub_32); + Addr.setOffsetReg(Reg); + return true; + } + break; + } + case Instruction::SExt: + case Instruction::ZExt: { + if (!Addr.getReg() || Addr.getOffsetReg()) + break; + + const Value *Src = nullptr; + // Fold the zext or sext when it won't become a noop. + if (const auto *ZE = dyn_cast(U)) { + if (!isIntExtFree(ZE) && ZE->getOperand(0)->getType()->isIntegerTy(32)) { + Addr.setExtendType(AArch64_AM::UXTW); + Src = ZE->getOperand(0); + } + } else if (const auto *SE = dyn_cast(U)) { + if (!isIntExtFree(SE) && SE->getOperand(0)->getType()->isIntegerTy(32)) { + Addr.setExtendType(AArch64_AM::SXTW); + Src = SE->getOperand(0); + } + } + + if (!Src) + break; + + Addr.setShift(0); + unsigned Reg = getRegForValue(Src); + if (!Reg) + return false; + Addr.setOffsetReg(Reg); + return true; + } + } // end switch + + if (Addr.isRegBase() && !Addr.getReg()) { + unsigned Reg = getRegForValue(Obj); + if (!Reg) + return false; + Addr.setReg(Reg); + return true; + } + + if (!Addr.getOffsetReg()) { + unsigned Reg = getRegForValue(Obj); + if (!Reg) + return false; + Addr.setOffsetReg(Reg); + return true; + } + + return false; +} + +bool AArch64FastISel::computeCallAddress(const Value *V, Address &Addr) { + const User *U = nullptr; + unsigned Opcode = Instruction::UserOp1; + bool InMBB = true; + + if (const auto *I = dyn_cast(V)) { + Opcode = I->getOpcode(); + U = I; + InMBB = I->getParent() == FuncInfo.MBB->getBasicBlock(); + } else if (const auto *C = dyn_cast(V)) { + Opcode = C->getOpcode(); + U = C; + } + + switch (Opcode) { + default: break; + case Instruction::BitCast: + // Look past bitcasts if its operand is in the same BB. + if (InMBB) + return computeCallAddress(U->getOperand(0), Addr); + break; + case Instruction::IntToPtr: + // Look past no-op inttoptrs if its operand is in the same BB. + if (InMBB && + TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy()) + return computeCallAddress(U->getOperand(0), Addr); + break; + case Instruction::PtrToInt: + // Look past no-op ptrtoints if its operand is in the same BB. + if (InMBB && + TLI.getValueType(U->getType()) == TLI.getPointerTy()) + return computeCallAddress(U->getOperand(0), Addr); + break; + } + + if (const GlobalValue *GV = dyn_cast(V)) { + Addr.setGlobalValue(GV); + return true; + } + + // If all else fails, try to materialize the value in a register. + if (!Addr.getGlobalValue()) { + Addr.setReg(getRegForValue(V)); + return Addr.getReg() != 0; } - // Try to get this in a register if nothing else has worked. - if (!Addr.isValid()) - Addr.setReg(getRegForValue(Obj)); - return Addr.isValid(); + return false; } + bool AArch64FastISel::isTypeLegal(Type *Ty, MVT &VT) { EVT evt = TLI.getValueType(Ty, true); @@ -437,62 +894,122 @@ bool AArch64FastISel::isTypeLegal(Type *Ty, MVT &VT) { return TLI.isTypeLegal(VT); } -bool AArch64FastISel::isLoadStoreTypeLegal(Type *Ty, MVT &VT) { +/// \brief Determine if the value type is supported by FastISel. +/// +/// FastISel for AArch64 can handle more value types than are legal. This adds +/// simple value type such as i1, i8, and i16. +bool AArch64FastISel::isTypeSupported(Type *Ty, MVT &VT, bool IsVectorAllowed) { + if (Ty->isVectorTy() && !IsVectorAllowed) + return false; + if (isTypeLegal(Ty, VT)) return true; // If this is a type than can be sign or zero-extended to a basic operation - // go ahead and accept it now. For stores, this reflects truncation. + // go ahead and accept it now. if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16) return true; return false; } -bool AArch64FastISel::SimplifyAddress(Address &Addr, MVT VT, - int64_t ScaleFactor, bool UseUnscaled) { - bool needsLowering = false; - int64_t Offset = Addr.getOffset(); - switch (VT.SimpleTy) { - default: +bool AArch64FastISel::isValueAvailable(const Value *V) const { + if (!isa(V)) + return true; + + const auto *I = cast(V); + if (FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) + return true; + + return false; +} + +bool AArch64FastISel::simplifyAddress(Address &Addr, MVT VT) { + unsigned ScaleFactor = getImplicitScaleFactor(VT); + if (!ScaleFactor) return false; - case MVT::i1: - case MVT::i8: - case MVT::i16: - case MVT::i32: - case MVT::i64: - case MVT::f32: - case MVT::f64: - if (!UseUnscaled) - // Using scaled, 12-bit, unsigned immediate offsets. - needsLowering = ((Offset & 0xfff) != Offset); - else - // Using unscaled, 9-bit, signed immediate offsets. - needsLowering = (Offset > 256 || Offset < -256); - break; - } - //If this is a stack pointer and the offset needs to be simplified then put + bool ImmediateOffsetNeedsLowering = false; + bool RegisterOffsetNeedsLowering = false; + int64_t Offset = Addr.getOffset(); + if (((Offset < 0) || (Offset & (ScaleFactor - 1))) && !isInt<9>(Offset)) + ImmediateOffsetNeedsLowering = true; + else if (Offset > 0 && !(Offset & (ScaleFactor - 1)) && + !isUInt<12>(Offset / ScaleFactor)) + ImmediateOffsetNeedsLowering = true; + + // Cannot encode an offset register and an immediate offset in the same + // instruction. Fold the immediate offset into the load/store instruction and + // emit an additonal add to take care of the offset register. + if (!ImmediateOffsetNeedsLowering && Addr.getOffset() && Addr.getOffsetReg()) + RegisterOffsetNeedsLowering = true; + + // Cannot encode zero register as base. + if (Addr.isRegBase() && Addr.getOffsetReg() && !Addr.getReg()) + RegisterOffsetNeedsLowering = true; + + // If this is a stack pointer and the offset needs to be simplified then put // the alloca address into a register, set the base type back to register and // continue. This should almost never happen. - if (needsLowering && Addr.getKind() == Address::FrameIndexBase) { - unsigned ResultReg = createResultReg(&AArch64::GPR64RegClass); + if ((ImmediateOffsetNeedsLowering || Addr.getOffsetReg()) && Addr.isFIBase()) + { + unsigned ResultReg = createResultReg(&AArch64::GPR64spRegClass); BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri), ResultReg) - .addFrameIndex(Addr.getFI()) - .addImm(0) - .addImm(0); + .addFrameIndex(Addr.getFI()) + .addImm(0) + .addImm(0); Addr.setKind(Address::RegBase); Addr.setReg(ResultReg); } + if (RegisterOffsetNeedsLowering) { + unsigned ResultReg = 0; + if (Addr.getReg()) { + if (Addr.getExtendType() == AArch64_AM::SXTW || + Addr.getExtendType() == AArch64_AM::UXTW ) + ResultReg = emitAddSub_rx(/*UseAdd=*/true, MVT::i64, Addr.getReg(), + /*TODO:IsKill=*/false, Addr.getOffsetReg(), + /*TODO:IsKill=*/false, Addr.getExtendType(), + Addr.getShift()); + else + ResultReg = emitAddSub_rs(/*UseAdd=*/true, MVT::i64, Addr.getReg(), + /*TODO:IsKill=*/false, Addr.getOffsetReg(), + /*TODO:IsKill=*/false, AArch64_AM::LSL, + Addr.getShift()); + } else { + if (Addr.getExtendType() == AArch64_AM::UXTW) + ResultReg = emitLSL_ri(MVT::i64, MVT::i32, Addr.getOffsetReg(), + /*Op0IsKill=*/false, Addr.getShift(), + /*IsZExt=*/true); + else if (Addr.getExtendType() == AArch64_AM::SXTW) + ResultReg = emitLSL_ri(MVT::i64, MVT::i32, Addr.getOffsetReg(), + /*Op0IsKill=*/false, Addr.getShift(), + /*IsZExt=*/false); + else + ResultReg = emitLSL_ri(MVT::i64, MVT::i64, Addr.getOffsetReg(), + /*Op0IsKill=*/false, Addr.getShift()); + } + if (!ResultReg) + return false; + + Addr.setReg(ResultReg); + Addr.setOffsetReg(0); + Addr.setShift(0); + Addr.setExtendType(AArch64_AM::InvalidShiftExtend); + } + // Since the offset is too large for the load/store instruction get the // reg+offset into a register. - if (needsLowering) { - uint64_t UnscaledOffset = Addr.getOffset() * ScaleFactor; - unsigned ResultReg = FastEmit_ri_(MVT::i64, ISD::ADD, Addr.getReg(), false, - UnscaledOffset, MVT::i64); - if (ResultReg == 0) + if (ImmediateOffsetNeedsLowering) { + unsigned ResultReg; + if (Addr.getReg()) + // Try to fold the immediate into the add instruction. + ResultReg = emitAdd_ri_(MVT::i64, Addr.getReg(), /*IsKill=*/false, Offset); + else + ResultReg = fastEmit_i(MVT::i64, MVT::i64, ISD::Constant, Offset); + + if (!ResultReg) return false; Addr.setReg(ResultReg); Addr.setOffset(0); @@ -500,222 +1017,1017 @@ bool AArch64FastISel::SimplifyAddress(Address &Addr, MVT VT, return true; } -void AArch64FastISel::AddLoadStoreOperands(Address &Addr, +void AArch64FastISel::addLoadStoreOperands(Address &Addr, const MachineInstrBuilder &MIB, - unsigned Flags, bool UseUnscaled) { - int64_t Offset = Addr.getOffset(); + unsigned Flags, + unsigned ScaleFactor, + MachineMemOperand *MMO) { + int64_t Offset = Addr.getOffset() / ScaleFactor; // Frame base works a bit differently. Handle it separately. - if (Addr.getKind() == Address::FrameIndexBase) { + if (Addr.isFIBase()) { int FI = Addr.getFI(); // FIXME: We shouldn't be using getObjectSize/getObjectAlignment. The size // and alignment should be based on the VT. - MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand( - MachinePointerInfo::getFixedStack(FI, Offset), Flags, - MFI.getObjectSize(FI), MFI.getObjectAlignment(FI)); + MMO = FuncInfo.MF->getMachineMemOperand( + MachinePointerInfo::getFixedStack(FI, Offset), Flags, + MFI.getObjectSize(FI), MFI.getObjectAlignment(FI)); // Now add the rest of the operands. - MIB.addFrameIndex(FI).addImm(Offset).addMemOperand(MMO); + MIB.addFrameIndex(FI).addImm(Offset); } else { - // Now add the rest of the operands. - MIB.addReg(Addr.getReg()); - MIB.addImm(Offset); + assert(Addr.isRegBase() && "Unexpected address kind."); + const MCInstrDesc &II = MIB->getDesc(); + unsigned Idx = (Flags & MachineMemOperand::MOStore) ? 1 : 0; + Addr.setReg( + constrainOperandRegClass(II, Addr.getReg(), II.getNumDefs()+Idx)); + Addr.setOffsetReg( + constrainOperandRegClass(II, Addr.getOffsetReg(), II.getNumDefs()+Idx+1)); + if (Addr.getOffsetReg()) { + assert(Addr.getOffset() == 0 && "Unexpected offset"); + bool IsSigned = Addr.getExtendType() == AArch64_AM::SXTW || + Addr.getExtendType() == AArch64_AM::SXTX; + MIB.addReg(Addr.getReg()); + MIB.addReg(Addr.getOffsetReg()); + MIB.addImm(IsSigned); + MIB.addImm(Addr.getShift() != 0); + } else + MIB.addReg(Addr.getReg()).addImm(Offset); } -} -bool AArch64FastISel::EmitLoad(MVT VT, unsigned &ResultReg, Address Addr, - bool UseUnscaled) { - // Negative offsets require unscaled, 9-bit, signed immediate offsets. - // Otherwise, we try using scaled, 12-bit, unsigned immediate offsets. - if (!UseUnscaled && Addr.getOffset() < 0) - UseUnscaled = true; + if (MMO) + MIB.addMemOperand(MMO); +} - unsigned Opc; - const TargetRegisterClass *RC; - bool VTIsi1 = false; - int64_t ScaleFactor = 0; - switch (VT.SimpleTy) { +unsigned AArch64FastISel::emitAddSub(bool UseAdd, MVT RetVT, const Value *LHS, + const Value *RHS, bool SetFlags, + bool WantResult, bool IsZExt) { + AArch64_AM::ShiftExtendType ExtendType = AArch64_AM::InvalidShiftExtend; + bool NeedExtend = false; + switch (RetVT.SimpleTy) { default: - return false; + return 0; case MVT::i1: - VTIsi1 = true; - // Intentional fall-through. + NeedExtend = true; + break; case MVT::i8: - Opc = UseUnscaled ? AArch64::LDURBBi : AArch64::LDRBBui; - RC = &AArch64::GPR32RegClass; - ScaleFactor = 1; + NeedExtend = true; + ExtendType = IsZExt ? AArch64_AM::UXTB : AArch64_AM::SXTB; break; case MVT::i16: - Opc = UseUnscaled ? AArch64::LDURHHi : AArch64::LDRHHui; - RC = &AArch64::GPR32RegClass; - ScaleFactor = 2; - break; - case MVT::i32: - Opc = UseUnscaled ? AArch64::LDURWi : AArch64::LDRWui; - RC = &AArch64::GPR32RegClass; - ScaleFactor = 4; + NeedExtend = true; + ExtendType = IsZExt ? AArch64_AM::UXTH : AArch64_AM::SXTH; break; + case MVT::i32: // fall-through case MVT::i64: - Opc = UseUnscaled ? AArch64::LDURXi : AArch64::LDRXui; - RC = &AArch64::GPR64RegClass; - ScaleFactor = 8; - break; - case MVT::f32: - Opc = UseUnscaled ? AArch64::LDURSi : AArch64::LDRSui; - RC = TLI.getRegClassFor(VT); - ScaleFactor = 4; - break; - case MVT::f64: - Opc = UseUnscaled ? AArch64::LDURDi : AArch64::LDRDui; - RC = TLI.getRegClassFor(VT); - ScaleFactor = 8; break; } - // Scale the offset. - if (!UseUnscaled) { - int64_t Offset = Addr.getOffset(); - if (Offset & (ScaleFactor - 1)) - // Retry using an unscaled, 9-bit, signed immediate offset. - return EmitLoad(VT, ResultReg, Addr, /*UseUnscaled*/ true); + MVT SrcVT = RetVT; + RetVT.SimpleTy = std::max(RetVT.SimpleTy, MVT::i32); - Addr.setOffset(Offset / ScaleFactor); - } + // Canonicalize immediates to the RHS first. + if (UseAdd && isa(LHS) && !isa(RHS)) + std::swap(LHS, RHS); - // Simplify this down to something we can handle. - if (!SimplifyAddress(Addr, VT, UseUnscaled ? 1 : ScaleFactor, UseUnscaled)) - return false; + // Canonicalize mul by power of 2 to the RHS. + if (UseAdd && LHS->hasOneUse() && isValueAvailable(LHS)) + if (isMulPowOf2(LHS)) + std::swap(LHS, RHS); - // Create the base instruction, then add the operands. - ResultReg = createResultReg(RC); - MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, - TII.get(Opc), ResultReg); - AddLoadStoreOperands(Addr, MIB, MachineMemOperand::MOLoad, UseUnscaled); + // Canonicalize shift immediate to the RHS. + if (UseAdd && LHS->hasOneUse() && isValueAvailable(LHS)) + if (const auto *SI = dyn_cast(LHS)) + if (isa(SI->getOperand(1))) + if (SI->getOpcode() == Instruction::Shl || + SI->getOpcode() == Instruction::LShr || + SI->getOpcode() == Instruction::AShr ) + std::swap(LHS, RHS); + + unsigned LHSReg = getRegForValue(LHS); + if (!LHSReg) + return 0; + bool LHSIsKill = hasTrivialKill(LHS); - // Loading an i1 requires special handling. - if (VTIsi1) { - MRI.constrainRegClass(ResultReg, &AArch64::GPR32RegClass); - unsigned ANDReg = createResultReg(&AArch64::GPR32spRegClass); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ANDWri), - ANDReg) - .addReg(ResultReg) - .addImm(AArch64_AM::encodeLogicalImmediate(1, 32)); - ResultReg = ANDReg; + if (NeedExtend) + LHSReg = emitIntExt(SrcVT, LHSReg, RetVT, IsZExt); + + unsigned ResultReg = 0; + if (const auto *C = dyn_cast(RHS)) { + uint64_t Imm = IsZExt ? C->getZExtValue() : C->getSExtValue(); + if (C->isNegative()) + ResultReg = emitAddSub_ri(!UseAdd, RetVT, LHSReg, LHSIsKill, -Imm, + SetFlags, WantResult); + else + ResultReg = emitAddSub_ri(UseAdd, RetVT, LHSReg, LHSIsKill, Imm, SetFlags, + WantResult); } - return true; -} + if (ResultReg) + return ResultReg; -bool AArch64FastISel::SelectLoad(const Instruction *I) { - MVT VT; - // Verify we have a legal type before going any further. Currently, we handle - // simple types that will directly fit in a register (i32/f32/i64/f64) or - // those that can be sign or zero-extended to a basic operation (i1/i8/i16). - if (!isLoadStoreTypeLegal(I->getType(), VT) || cast(I)->isAtomic()) - return false; + // Only extend the RHS within the instruction if there is a valid extend type. + if (ExtendType != AArch64_AM::InvalidShiftExtend && RHS->hasOneUse() && + isValueAvailable(RHS)) { + if (const auto *SI = dyn_cast(RHS)) + if (const auto *C = dyn_cast(SI->getOperand(1))) + if ((SI->getOpcode() == Instruction::Shl) && (C->getZExtValue() < 4)) { + unsigned RHSReg = getRegForValue(SI->getOperand(0)); + if (!RHSReg) + return 0; + bool RHSIsKill = hasTrivialKill(SI->getOperand(0)); + return emitAddSub_rx(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, + RHSIsKill, ExtendType, C->getZExtValue(), + SetFlags, WantResult); + } + unsigned RHSReg = getRegForValue(RHS); + if (!RHSReg) + return 0; + bool RHSIsKill = hasTrivialKill(RHS); + return emitAddSub_rx(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill, + ExtendType, 0, SetFlags, WantResult); + } - // See if we can handle this address. - Address Addr; - if (!ComputeAddress(I->getOperand(0), Addr)) - return false; + // Check if the mul can be folded into the instruction. + if (RHS->hasOneUse() && isValueAvailable(RHS)) + if (isMulPowOf2(RHS)) { + const Value *MulLHS = cast(RHS)->getOperand(0); + const Value *MulRHS = cast(RHS)->getOperand(1); - unsigned ResultReg; - if (!EmitLoad(VT, ResultReg, Addr)) - return false; + if (const auto *C = dyn_cast(MulLHS)) + if (C->getValue().isPowerOf2()) + std::swap(MulLHS, MulRHS); - UpdateValueMap(I, ResultReg); - return true; -} + assert(isa(MulRHS) && "Expected a ConstantInt."); + uint64_t ShiftVal = cast(MulRHS)->getValue().logBase2(); + unsigned RHSReg = getRegForValue(MulLHS); + if (!RHSReg) + return 0; + bool RHSIsKill = hasTrivialKill(MulLHS); + return emitAddSub_rs(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill, + AArch64_AM::LSL, ShiftVal, SetFlags, WantResult); + } -bool AArch64FastISel::EmitStore(MVT VT, unsigned SrcReg, Address Addr, - bool UseUnscaled) { - // Negative offsets require unscaled, 9-bit, signed immediate offsets. - // Otherwise, we try using scaled, 12-bit, unsigned immediate offsets. - if (!UseUnscaled && Addr.getOffset() < 0) - UseUnscaled = true; + // Check if the shift can be folded into the instruction. + if (RHS->hasOneUse() && isValueAvailable(RHS)) + if (const auto *SI = dyn_cast(RHS)) { + if (const auto *C = dyn_cast(SI->getOperand(1))) { + AArch64_AM::ShiftExtendType ShiftType = AArch64_AM::InvalidShiftExtend; + switch (SI->getOpcode()) { + default: break; + case Instruction::Shl: ShiftType = AArch64_AM::LSL; break; + case Instruction::LShr: ShiftType = AArch64_AM::LSR; break; + case Instruction::AShr: ShiftType = AArch64_AM::ASR; break; + } + uint64_t ShiftVal = C->getZExtValue(); + if (ShiftType != AArch64_AM::InvalidShiftExtend) { + unsigned RHSReg = getRegForValue(SI->getOperand(0)); + if (!RHSReg) + return 0; + bool RHSIsKill = hasTrivialKill(SI->getOperand(0)); + return emitAddSub_rs(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, + RHSIsKill, ShiftType, ShiftVal, SetFlags, + WantResult); + } + } + } + + unsigned RHSReg = getRegForValue(RHS); + if (!RHSReg) + return 0; + bool RHSIsKill = hasTrivialKill(RHS); + + if (NeedExtend) + RHSReg = emitIntExt(SrcVT, RHSReg, RetVT, IsZExt); + + return emitAddSub_rr(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill, + SetFlags, WantResult); +} + +unsigned AArch64FastISel::emitAddSub_rr(bool UseAdd, MVT RetVT, unsigned LHSReg, + bool LHSIsKill, unsigned RHSReg, + bool RHSIsKill, bool SetFlags, + bool WantResult) { + assert(LHSReg && RHSReg && "Invalid register number."); + + if (RetVT != MVT::i32 && RetVT != MVT::i64) + return 0; + + static const unsigned OpcTable[2][2][2] = { + { { AArch64::SUBWrr, AArch64::SUBXrr }, + { AArch64::ADDWrr, AArch64::ADDXrr } }, + { { AArch64::SUBSWrr, AArch64::SUBSXrr }, + { AArch64::ADDSWrr, AArch64::ADDSXrr } } + }; + bool Is64Bit = RetVT == MVT::i64; + unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit]; + const TargetRegisterClass *RC = + Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass; + unsigned ResultReg; + if (WantResult) + ResultReg = createResultReg(RC); + else + ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR; + + const MCInstrDesc &II = TII.get(Opc); + LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs()); + RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg) + .addReg(LHSReg, getKillRegState(LHSIsKill)) + .addReg(RHSReg, getKillRegState(RHSIsKill)); + return ResultReg; +} + +unsigned AArch64FastISel::emitAddSub_ri(bool UseAdd, MVT RetVT, unsigned LHSReg, + bool LHSIsKill, uint64_t Imm, + bool SetFlags, bool WantResult) { + assert(LHSReg && "Invalid register number."); + + if (RetVT != MVT::i32 && RetVT != MVT::i64) + return 0; + + unsigned ShiftImm; + if (isUInt<12>(Imm)) + ShiftImm = 0; + else if ((Imm & 0xfff000) == Imm) { + ShiftImm = 12; + Imm >>= 12; + } else + return 0; + + static const unsigned OpcTable[2][2][2] = { + { { AArch64::SUBWri, AArch64::SUBXri }, + { AArch64::ADDWri, AArch64::ADDXri } }, + { { AArch64::SUBSWri, AArch64::SUBSXri }, + { AArch64::ADDSWri, AArch64::ADDSXri } } + }; + bool Is64Bit = RetVT == MVT::i64; + unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit]; + const TargetRegisterClass *RC; + if (SetFlags) + RC = Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass; + else + RC = Is64Bit ? &AArch64::GPR64spRegClass : &AArch64::GPR32spRegClass; + unsigned ResultReg; + if (WantResult) + ResultReg = createResultReg(RC); + else + ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR; + + const MCInstrDesc &II = TII.get(Opc); + LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs()); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg) + .addReg(LHSReg, getKillRegState(LHSIsKill)) + .addImm(Imm) + .addImm(getShifterImm(AArch64_AM::LSL, ShiftImm)); + return ResultReg; +} + +unsigned AArch64FastISel::emitAddSub_rs(bool UseAdd, MVT RetVT, unsigned LHSReg, + bool LHSIsKill, unsigned RHSReg, + bool RHSIsKill, + AArch64_AM::ShiftExtendType ShiftType, + uint64_t ShiftImm, bool SetFlags, + bool WantResult) { + assert(LHSReg && RHSReg && "Invalid register number."); + + if (RetVT != MVT::i32 && RetVT != MVT::i64) + return 0; + + static const unsigned OpcTable[2][2][2] = { + { { AArch64::SUBWrs, AArch64::SUBXrs }, + { AArch64::ADDWrs, AArch64::ADDXrs } }, + { { AArch64::SUBSWrs, AArch64::SUBSXrs }, + { AArch64::ADDSWrs, AArch64::ADDSXrs } } + }; + bool Is64Bit = RetVT == MVT::i64; + unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit]; + const TargetRegisterClass *RC = + Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass; + unsigned ResultReg; + if (WantResult) + ResultReg = createResultReg(RC); + else + ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR; + + const MCInstrDesc &II = TII.get(Opc); + LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs()); + RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg) + .addReg(LHSReg, getKillRegState(LHSIsKill)) + .addReg(RHSReg, getKillRegState(RHSIsKill)) + .addImm(getShifterImm(ShiftType, ShiftImm)); + return ResultReg; +} + +unsigned AArch64FastISel::emitAddSub_rx(bool UseAdd, MVT RetVT, unsigned LHSReg, + bool LHSIsKill, unsigned RHSReg, + bool RHSIsKill, + AArch64_AM::ShiftExtendType ExtType, + uint64_t ShiftImm, bool SetFlags, + bool WantResult) { + assert(LHSReg && RHSReg && "Invalid register number."); + + if (RetVT != MVT::i32 && RetVT != MVT::i64) + return 0; + + static const unsigned OpcTable[2][2][2] = { + { { AArch64::SUBWrx, AArch64::SUBXrx }, + { AArch64::ADDWrx, AArch64::ADDXrx } }, + { { AArch64::SUBSWrx, AArch64::SUBSXrx }, + { AArch64::ADDSWrx, AArch64::ADDSXrx } } + }; + bool Is64Bit = RetVT == MVT::i64; + unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit]; + const TargetRegisterClass *RC = nullptr; + if (SetFlags) + RC = Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass; + else + RC = Is64Bit ? &AArch64::GPR64spRegClass : &AArch64::GPR32spRegClass; + unsigned ResultReg; + if (WantResult) + ResultReg = createResultReg(RC); + else + ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR; + + const MCInstrDesc &II = TII.get(Opc); + LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs()); + RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg) + .addReg(LHSReg, getKillRegState(LHSIsKill)) + .addReg(RHSReg, getKillRegState(RHSIsKill)) + .addImm(getArithExtendImm(ExtType, ShiftImm)); + return ResultReg; +} + +bool AArch64FastISel::emitCmp(const Value *LHS, const Value *RHS, bool IsZExt) { + Type *Ty = LHS->getType(); + EVT EVT = TLI.getValueType(Ty, true); + if (!EVT.isSimple()) + return false; + MVT VT = EVT.getSimpleVT(); - unsigned StrOpc; - bool VTIsi1 = false; - int64_t ScaleFactor = 0; - // Using scaled, 12-bit, unsigned immediate offsets. switch (VT.SimpleTy) { default: return false; case MVT::i1: - VTIsi1 = true; case MVT::i8: - StrOpc = UseUnscaled ? AArch64::STURBBi : AArch64::STRBBui; + case MVT::i16: + case MVT::i32: + case MVT::i64: + return emitICmp(VT, LHS, RHS, IsZExt); + case MVT::f32: + case MVT::f64: + return emitFCmp(VT, LHS, RHS); + } +} + +bool AArch64FastISel::emitICmp(MVT RetVT, const Value *LHS, const Value *RHS, + bool IsZExt) { + return emitSub(RetVT, LHS, RHS, /*SetFlags=*/true, /*WantResult=*/false, + IsZExt) != 0; +} + +bool AArch64FastISel::emitICmp_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill, + uint64_t Imm) { + return emitAddSub_ri(/*UseAdd=*/false, RetVT, LHSReg, LHSIsKill, Imm, + /*SetFlags=*/true, /*WantResult=*/false) != 0; +} + +bool AArch64FastISel::emitFCmp(MVT RetVT, const Value *LHS, const Value *RHS) { + if (RetVT != MVT::f32 && RetVT != MVT::f64) + return false; + + // Check to see if the 2nd operand is a constant that we can encode directly + // in the compare. + bool UseImm = false; + if (const auto *CFP = dyn_cast(RHS)) + if (CFP->isZero() && !CFP->isNegative()) + UseImm = true; + + unsigned LHSReg = getRegForValue(LHS); + if (!LHSReg) + return false; + bool LHSIsKill = hasTrivialKill(LHS); + + if (UseImm) { + unsigned Opc = (RetVT == MVT::f64) ? AArch64::FCMPDri : AArch64::FCMPSri; + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc)) + .addReg(LHSReg, getKillRegState(LHSIsKill)); + return true; + } + + unsigned RHSReg = getRegForValue(RHS); + if (!RHSReg) + return false; + bool RHSIsKill = hasTrivialKill(RHS); + + unsigned Opc = (RetVT == MVT::f64) ? AArch64::FCMPDrr : AArch64::FCMPSrr; + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc)) + .addReg(LHSReg, getKillRegState(LHSIsKill)) + .addReg(RHSReg, getKillRegState(RHSIsKill)); + return true; +} + +unsigned AArch64FastISel::emitAdd(MVT RetVT, const Value *LHS, const Value *RHS, + bool SetFlags, bool WantResult, bool IsZExt) { + return emitAddSub(/*UseAdd=*/true, RetVT, LHS, RHS, SetFlags, WantResult, + IsZExt); +} + +/// \brief This method is a wrapper to simplify add emission. +/// +/// First try to emit an add with an immediate operand using emitAddSub_ri. If +/// that fails, then try to materialize the immediate into a register and use +/// emitAddSub_rr instead. +unsigned AArch64FastISel::emitAdd_ri_(MVT VT, unsigned Op0, bool Op0IsKill, + int64_t Imm) { + unsigned ResultReg; + if (Imm < 0) + ResultReg = emitAddSub_ri(false, VT, Op0, Op0IsKill, -Imm); + else + ResultReg = emitAddSub_ri(true, VT, Op0, Op0IsKill, Imm); + + if (ResultReg) + return ResultReg; + + unsigned CReg = fastEmit_i(VT, VT, ISD::Constant, Imm); + if (!CReg) + return 0; + + ResultReg = emitAddSub_rr(true, VT, Op0, Op0IsKill, CReg, true); + return ResultReg; +} + +unsigned AArch64FastISel::emitSub(MVT RetVT, const Value *LHS, const Value *RHS, + bool SetFlags, bool WantResult, bool IsZExt) { + return emitAddSub(/*UseAdd=*/false, RetVT, LHS, RHS, SetFlags, WantResult, + IsZExt); +} + +unsigned AArch64FastISel::emitSubs_rr(MVT RetVT, unsigned LHSReg, + bool LHSIsKill, unsigned RHSReg, + bool RHSIsKill, bool WantResult) { + return emitAddSub_rr(/*UseAdd=*/false, RetVT, LHSReg, LHSIsKill, RHSReg, + RHSIsKill, /*SetFlags=*/true, WantResult); +} + +unsigned AArch64FastISel::emitSubs_rs(MVT RetVT, unsigned LHSReg, + bool LHSIsKill, unsigned RHSReg, + bool RHSIsKill, + AArch64_AM::ShiftExtendType ShiftType, + uint64_t ShiftImm, bool WantResult) { + return emitAddSub_rs(/*UseAdd=*/false, RetVT, LHSReg, LHSIsKill, RHSReg, + RHSIsKill, ShiftType, ShiftImm, /*SetFlags=*/true, + WantResult); +} + +unsigned AArch64FastISel::emitLogicalOp(unsigned ISDOpc, MVT RetVT, + const Value *LHS, const Value *RHS) { + // Canonicalize immediates to the RHS first. + if (isa(LHS) && !isa(RHS)) + std::swap(LHS, RHS); + + // Canonicalize mul by power-of-2 to the RHS. + if (LHS->hasOneUse() && isValueAvailable(LHS)) + if (isMulPowOf2(LHS)) + std::swap(LHS, RHS); + + // Canonicalize shift immediate to the RHS. + if (LHS->hasOneUse() && isValueAvailable(LHS)) + if (const auto *SI = dyn_cast(LHS)) + if (isa(SI->getOperand(1))) + std::swap(LHS, RHS); + + unsigned LHSReg = getRegForValue(LHS); + if (!LHSReg) + return 0; + bool LHSIsKill = hasTrivialKill(LHS); + + unsigned ResultReg = 0; + if (const auto *C = dyn_cast(RHS)) { + uint64_t Imm = C->getZExtValue(); + ResultReg = emitLogicalOp_ri(ISDOpc, RetVT, LHSReg, LHSIsKill, Imm); + } + if (ResultReg) + return ResultReg; + + // Check if the mul can be folded into the instruction. + if (RHS->hasOneUse() && isValueAvailable(RHS)) + if (isMulPowOf2(RHS)) { + const Value *MulLHS = cast(RHS)->getOperand(0); + const Value *MulRHS = cast(RHS)->getOperand(1); + + if (const auto *C = dyn_cast(MulLHS)) + if (C->getValue().isPowerOf2()) + std::swap(MulLHS, MulRHS); + + assert(isa(MulRHS) && "Expected a ConstantInt."); + uint64_t ShiftVal = cast(MulRHS)->getValue().logBase2(); + + unsigned RHSReg = getRegForValue(MulLHS); + if (!RHSReg) + return 0; + bool RHSIsKill = hasTrivialKill(MulLHS); + return emitLogicalOp_rs(ISDOpc, RetVT, LHSReg, LHSIsKill, RHSReg, + RHSIsKill, ShiftVal); + } + + // Check if the shift can be folded into the instruction. + if (RHS->hasOneUse() && isValueAvailable(RHS)) + if (const auto *SI = dyn_cast(RHS)) + if (const auto *C = dyn_cast(SI->getOperand(1))) { + uint64_t ShiftVal = C->getZExtValue(); + unsigned RHSReg = getRegForValue(SI->getOperand(0)); + if (!RHSReg) + return 0; + bool RHSIsKill = hasTrivialKill(SI->getOperand(0)); + return emitLogicalOp_rs(ISDOpc, RetVT, LHSReg, LHSIsKill, RHSReg, + RHSIsKill, ShiftVal); + } + + unsigned RHSReg = getRegForValue(RHS); + if (!RHSReg) + return 0; + bool RHSIsKill = hasTrivialKill(RHS); + + MVT VT = std::max(MVT::i32, RetVT.SimpleTy); + ResultReg = fastEmit_rr(VT, VT, ISDOpc, LHSReg, LHSIsKill, RHSReg, RHSIsKill); + if (RetVT >= MVT::i8 && RetVT <= MVT::i16) { + uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff; + ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask); + } + return ResultReg; +} + +unsigned AArch64FastISel::emitLogicalOp_ri(unsigned ISDOpc, MVT RetVT, + unsigned LHSReg, bool LHSIsKill, + uint64_t Imm) { + assert((ISD::AND + 1 == ISD::OR) && (ISD::AND + 2 == ISD::XOR) && + "ISD nodes are not consecutive!"); + static const unsigned OpcTable[3][2] = { + { AArch64::ANDWri, AArch64::ANDXri }, + { AArch64::ORRWri, AArch64::ORRXri }, + { AArch64::EORWri, AArch64::EORXri } + }; + const TargetRegisterClass *RC; + unsigned Opc; + unsigned RegSize; + switch (RetVT.SimpleTy) { + default: + return 0; + case MVT::i1: + case MVT::i8: + case MVT::i16: + case MVT::i32: { + unsigned Idx = ISDOpc - ISD::AND; + Opc = OpcTable[Idx][0]; + RC = &AArch64::GPR32spRegClass; + RegSize = 32; + break; + } + case MVT::i64: + Opc = OpcTable[ISDOpc - ISD::AND][1]; + RC = &AArch64::GPR64spRegClass; + RegSize = 64; + break; + } + + if (!AArch64_AM::isLogicalImmediate(Imm, RegSize)) + return 0; + + unsigned ResultReg = + fastEmitInst_ri(Opc, RC, LHSReg, LHSIsKill, + AArch64_AM::encodeLogicalImmediate(Imm, RegSize)); + if (RetVT >= MVT::i8 && RetVT <= MVT::i16 && ISDOpc != ISD::AND) { + uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff; + ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask); + } + return ResultReg; +} + +unsigned AArch64FastISel::emitLogicalOp_rs(unsigned ISDOpc, MVT RetVT, + unsigned LHSReg, bool LHSIsKill, + unsigned RHSReg, bool RHSIsKill, + uint64_t ShiftImm) { + assert((ISD::AND + 1 == ISD::OR) && (ISD::AND + 2 == ISD::XOR) && + "ISD nodes are not consecutive!"); + static const unsigned OpcTable[3][2] = { + { AArch64::ANDWrs, AArch64::ANDXrs }, + { AArch64::ORRWrs, AArch64::ORRXrs }, + { AArch64::EORWrs, AArch64::EORXrs } + }; + const TargetRegisterClass *RC; + unsigned Opc; + switch (RetVT.SimpleTy) { + default: + return 0; + case MVT::i1: + case MVT::i8: + case MVT::i16: + case MVT::i32: + Opc = OpcTable[ISDOpc - ISD::AND][0]; + RC = &AArch64::GPR32RegClass; + break; + case MVT::i64: + Opc = OpcTable[ISDOpc - ISD::AND][1]; + RC = &AArch64::GPR64RegClass; + break; + } + unsigned ResultReg = + fastEmitInst_rri(Opc, RC, LHSReg, LHSIsKill, RHSReg, RHSIsKill, + AArch64_AM::getShifterImm(AArch64_AM::LSL, ShiftImm)); + if (RetVT >= MVT::i8 && RetVT <= MVT::i16) { + uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff; + ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask); + } + return ResultReg; +} + +unsigned AArch64FastISel::emitAnd_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill, + uint64_t Imm) { + return emitLogicalOp_ri(ISD::AND, RetVT, LHSReg, LHSIsKill, Imm); +} + +unsigned AArch64FastISel::emitLoad(MVT VT, MVT RetVT, Address Addr, + bool WantZExt, MachineMemOperand *MMO) { + // Simplify this down to something we can handle. + if (!simplifyAddress(Addr, VT)) + return 0; + + unsigned ScaleFactor = getImplicitScaleFactor(VT); + if (!ScaleFactor) + llvm_unreachable("Unexpected value type."); + + // Negative offsets require unscaled, 9-bit, signed immediate offsets. + // Otherwise, we try using scaled, 12-bit, unsigned immediate offsets. + bool UseScaled = true; + if ((Addr.getOffset() < 0) || (Addr.getOffset() & (ScaleFactor - 1))) { + UseScaled = false; ScaleFactor = 1; + } + + static const unsigned GPOpcTable[2][8][4] = { + // Sign-extend. + { { AArch64::LDURSBWi, AArch64::LDURSHWi, AArch64::LDURWi, + AArch64::LDURXi }, + { AArch64::LDURSBXi, AArch64::LDURSHXi, AArch64::LDURSWi, + AArch64::LDURXi }, + { AArch64::LDRSBWui, AArch64::LDRSHWui, AArch64::LDRWui, + AArch64::LDRXui }, + { AArch64::LDRSBXui, AArch64::LDRSHXui, AArch64::LDRSWui, + AArch64::LDRXui }, + { AArch64::LDRSBWroX, AArch64::LDRSHWroX, AArch64::LDRWroX, + AArch64::LDRXroX }, + { AArch64::LDRSBXroX, AArch64::LDRSHXroX, AArch64::LDRSWroX, + AArch64::LDRXroX }, + { AArch64::LDRSBWroW, AArch64::LDRSHWroW, AArch64::LDRWroW, + AArch64::LDRXroW }, + { AArch64::LDRSBXroW, AArch64::LDRSHXroW, AArch64::LDRSWroW, + AArch64::LDRXroW } + }, + // Zero-extend. + { { AArch64::LDURBBi, AArch64::LDURHHi, AArch64::LDURWi, + AArch64::LDURXi }, + { AArch64::LDURBBi, AArch64::LDURHHi, AArch64::LDURWi, + AArch64::LDURXi }, + { AArch64::LDRBBui, AArch64::LDRHHui, AArch64::LDRWui, + AArch64::LDRXui }, + { AArch64::LDRBBui, AArch64::LDRHHui, AArch64::LDRWui, + AArch64::LDRXui }, + { AArch64::LDRBBroX, AArch64::LDRHHroX, AArch64::LDRWroX, + AArch64::LDRXroX }, + { AArch64::LDRBBroX, AArch64::LDRHHroX, AArch64::LDRWroX, + AArch64::LDRXroX }, + { AArch64::LDRBBroW, AArch64::LDRHHroW, AArch64::LDRWroW, + AArch64::LDRXroW }, + { AArch64::LDRBBroW, AArch64::LDRHHroW, AArch64::LDRWroW, + AArch64::LDRXroW } + } + }; + + static const unsigned FPOpcTable[4][2] = { + { AArch64::LDURSi, AArch64::LDURDi }, + { AArch64::LDRSui, AArch64::LDRDui }, + { AArch64::LDRSroX, AArch64::LDRDroX }, + { AArch64::LDRSroW, AArch64::LDRDroW } + }; + + unsigned Opc; + const TargetRegisterClass *RC; + bool UseRegOffset = Addr.isRegBase() && !Addr.getOffset() && Addr.getReg() && + Addr.getOffsetReg(); + unsigned Idx = UseRegOffset ? 2 : UseScaled ? 1 : 0; + if (Addr.getExtendType() == AArch64_AM::UXTW || + Addr.getExtendType() == AArch64_AM::SXTW) + Idx++; + + bool IsRet64Bit = RetVT == MVT::i64; + switch (VT.SimpleTy) { + default: + llvm_unreachable("Unexpected value type."); + case MVT::i1: // Intentional fall-through. + case MVT::i8: + Opc = GPOpcTable[WantZExt][2 * Idx + IsRet64Bit][0]; + RC = (IsRet64Bit && !WantZExt) ? + &AArch64::GPR64RegClass: &AArch64::GPR32RegClass; break; case MVT::i16: - StrOpc = UseUnscaled ? AArch64::STURHHi : AArch64::STRHHui; - ScaleFactor = 2; + Opc = GPOpcTable[WantZExt][2 * Idx + IsRet64Bit][1]; + RC = (IsRet64Bit && !WantZExt) ? + &AArch64::GPR64RegClass: &AArch64::GPR32RegClass; break; case MVT::i32: - StrOpc = UseUnscaled ? AArch64::STURWi : AArch64::STRWui; - ScaleFactor = 4; + Opc = GPOpcTable[WantZExt][2 * Idx + IsRet64Bit][2]; + RC = (IsRet64Bit && !WantZExt) ? + &AArch64::GPR64RegClass: &AArch64::GPR32RegClass; break; case MVT::i64: - StrOpc = UseUnscaled ? AArch64::STURXi : AArch64::STRXui; - ScaleFactor = 8; + Opc = GPOpcTable[WantZExt][2 * Idx + IsRet64Bit][3]; + RC = &AArch64::GPR64RegClass; break; case MVT::f32: - StrOpc = UseUnscaled ? AArch64::STURSi : AArch64::STRSui; - ScaleFactor = 4; + Opc = FPOpcTable[Idx][0]; + RC = &AArch64::FPR32RegClass; break; case MVT::f64: - StrOpc = UseUnscaled ? AArch64::STURDi : AArch64::STRDui; - ScaleFactor = 8; + Opc = FPOpcTable[Idx][1]; + RC = &AArch64::FPR64RegClass; break; } - // Scale the offset. - if (!UseUnscaled) { - int64_t Offset = Addr.getOffset(); - if (Offset & (ScaleFactor - 1)) - // Retry using an unscaled, 9-bit, signed immediate offset. - return EmitStore(VT, SrcReg, Addr, /*UseUnscaled*/ true); - Addr.setOffset(Offset / ScaleFactor); + // Create the base instruction, then add the operands. + unsigned ResultReg = createResultReg(RC); + MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, + TII.get(Opc), ResultReg); + addLoadStoreOperands(Addr, MIB, MachineMemOperand::MOLoad, ScaleFactor, MMO); + + // Loading an i1 requires special handling. + if (VT == MVT::i1) { + unsigned ANDReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, 1); + assert(ANDReg && "Unexpected AND instruction emission failure."); + ResultReg = ANDReg; + } + + // For zero-extending loads to 64bit we emit a 32bit load and then convert + // the 32bit reg to a 64bit reg. + if (WantZExt && RetVT == MVT::i64 && VT <= MVT::i32) { + unsigned Reg64 = createResultReg(&AArch64::GPR64RegClass); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, + TII.get(AArch64::SUBREG_TO_REG), Reg64) + .addImm(0) + .addReg(ResultReg, getKillRegState(true)) + .addImm(AArch64::sub_32); + ResultReg = Reg64; } + return ResultReg; +} +bool AArch64FastISel::selectAddSub(const Instruction *I) { + MVT VT; + if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true)) + return false; + + if (VT.isVector()) + return selectOperator(I, I->getOpcode()); + + unsigned ResultReg; + switch (I->getOpcode()) { + default: + llvm_unreachable("Unexpected instruction."); + case Instruction::Add: + ResultReg = emitAdd(VT, I->getOperand(0), I->getOperand(1)); + break; + case Instruction::Sub: + ResultReg = emitSub(VT, I->getOperand(0), I->getOperand(1)); + break; + } + if (!ResultReg) + return false; + + updateValueMap(I, ResultReg); + return true; +} + +bool AArch64FastISel::selectLogicalOp(const Instruction *I) { + MVT VT; + if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true)) + return false; + + if (VT.isVector()) + return selectOperator(I, I->getOpcode()); + + unsigned ResultReg; + switch (I->getOpcode()) { + default: + llvm_unreachable("Unexpected instruction."); + case Instruction::And: + ResultReg = emitLogicalOp(ISD::AND, VT, I->getOperand(0), I->getOperand(1)); + break; + case Instruction::Or: + ResultReg = emitLogicalOp(ISD::OR, VT, I->getOperand(0), I->getOperand(1)); + break; + case Instruction::Xor: + ResultReg = emitLogicalOp(ISD::XOR, VT, I->getOperand(0), I->getOperand(1)); + break; + } + if (!ResultReg) + return false; + + updateValueMap(I, ResultReg); + return true; +} + +bool AArch64FastISel::selectLoad(const Instruction *I) { + MVT VT; + // Verify we have a legal type before going any further. Currently, we handle + // simple types that will directly fit in a register (i32/f32/i64/f64) or + // those that can be sign or zero-extended to a basic operation (i1/i8/i16). + if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true) || + cast(I)->isAtomic()) + return false; + + // See if we can handle this address. + Address Addr; + if (!computeAddress(I->getOperand(0), Addr, I->getType())) + return false; + + // Fold the following sign-/zero-extend into the load instruction. + bool WantZExt = true; + MVT RetVT = VT; + const Value *IntExtVal = nullptr; + if (I->hasOneUse()) { + if (const auto *ZE = dyn_cast(I->use_begin()->getUser())) { + if (isTypeSupported(ZE->getType(), RetVT)) + IntExtVal = ZE; + else + RetVT = VT; + } else if (const auto *SE = dyn_cast(I->use_begin()->getUser())) { + if (isTypeSupported(SE->getType(), RetVT)) + IntExtVal = SE; + else + RetVT = VT; + WantZExt = false; + } + } + + unsigned ResultReg = + emitLoad(VT, RetVT, Addr, WantZExt, createMachineMemOperandFor(I)); + if (!ResultReg) + return false; + + // There are a few different cases we have to handle, because the load or the + // sign-/zero-extend might not be selected by FastISel if we fall-back to + // SelectionDAG. There is also an ordering issue when both instructions are in + // different basic blocks. + // 1.) The load instruction is selected by FastISel, but the integer extend + // not. This usually happens when the integer extend is in a different + // basic block and SelectionDAG took over for that basic block. + // 2.) The load instruction is selected before the integer extend. This only + // happens when the integer extend is in a different basic block. + // 3.) The load instruction is selected by SelectionDAG and the integer extend + // by FastISel. This happens if there are instructions between the load + // and the integer extend that couldn't be selected by FastISel. + if (IntExtVal) { + // The integer extend hasn't been emitted yet. FastISel or SelectionDAG + // could select it. Emit a copy to subreg if necessary. FastISel will remove + // it when it selects the integer extend. + unsigned Reg = lookUpRegForValue(IntExtVal); + if (!Reg) { + if (RetVT == MVT::i64 && VT <= MVT::i32) { + if (WantZExt) { + // Delete the last emitted instruction from emitLoad (SUBREG_TO_REG). + std::prev(FuncInfo.InsertPt)->eraseFromParent(); + ResultReg = std::prev(FuncInfo.InsertPt)->getOperand(0).getReg(); + } else + ResultReg = fastEmitInst_extractsubreg(MVT::i32, ResultReg, + /*IsKill=*/true, + AArch64::sub_32); + } + updateValueMap(I, ResultReg); + return true; + } + + // The integer extend has already been emitted - delete all the instructions + // that have been emitted by the integer extend lowering code and use the + // result from the load instruction directly. + while (Reg) { + auto *MI = MRI.getUniqueVRegDef(Reg); + if (!MI) + break; + Reg = 0; + for (auto &Opnd : MI->uses()) { + if (Opnd.isReg()) { + Reg = Opnd.getReg(); + break; + } + } + MI->eraseFromParent(); + } + updateValueMap(IntExtVal, ResultReg); + return true; + } + + updateValueMap(I, ResultReg); + return true; +} + +bool AArch64FastISel::emitStore(MVT VT, unsigned SrcReg, Address Addr, + MachineMemOperand *MMO) { // Simplify this down to something we can handle. - if (!SimplifyAddress(Addr, VT, UseUnscaled ? 1 : ScaleFactor, UseUnscaled)) + if (!simplifyAddress(Addr, VT)) return false; + unsigned ScaleFactor = getImplicitScaleFactor(VT); + if (!ScaleFactor) + llvm_unreachable("Unexpected value type."); + + // Negative offsets require unscaled, 9-bit, signed immediate offsets. + // Otherwise, we try using scaled, 12-bit, unsigned immediate offsets. + bool UseScaled = true; + if ((Addr.getOffset() < 0) || (Addr.getOffset() & (ScaleFactor - 1))) { + UseScaled = false; + ScaleFactor = 1; + } + + static const unsigned OpcTable[4][6] = { + { AArch64::STURBBi, AArch64::STURHHi, AArch64::STURWi, AArch64::STURXi, + AArch64::STURSi, AArch64::STURDi }, + { AArch64::STRBBui, AArch64::STRHHui, AArch64::STRWui, AArch64::STRXui, + AArch64::STRSui, AArch64::STRDui }, + { AArch64::STRBBroX, AArch64::STRHHroX, AArch64::STRWroX, AArch64::STRXroX, + AArch64::STRSroX, AArch64::STRDroX }, + { AArch64::STRBBroW, AArch64::STRHHroW, AArch64::STRWroW, AArch64::STRXroW, + AArch64::STRSroW, AArch64::STRDroW } + }; + + unsigned Opc; + bool VTIsi1 = false; + bool UseRegOffset = Addr.isRegBase() && !Addr.getOffset() && Addr.getReg() && + Addr.getOffsetReg(); + unsigned Idx = UseRegOffset ? 2 : UseScaled ? 1 : 0; + if (Addr.getExtendType() == AArch64_AM::UXTW || + Addr.getExtendType() == AArch64_AM::SXTW) + Idx++; + + switch (VT.SimpleTy) { + default: llvm_unreachable("Unexpected value type."); + case MVT::i1: VTIsi1 = true; + case MVT::i8: Opc = OpcTable[Idx][0]; break; + case MVT::i16: Opc = OpcTable[Idx][1]; break; + case MVT::i32: Opc = OpcTable[Idx][2]; break; + case MVT::i64: Opc = OpcTable[Idx][3]; break; + case MVT::f32: Opc = OpcTable[Idx][4]; break; + case MVT::f64: Opc = OpcTable[Idx][5]; break; + } + // Storing an i1 requires special handling. - if (VTIsi1) { - MRI.constrainRegClass(SrcReg, &AArch64::GPR32RegClass); - unsigned ANDReg = createResultReg(&AArch64::GPR32spRegClass); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ANDWri), - ANDReg) - .addReg(SrcReg) - .addImm(AArch64_AM::encodeLogicalImmediate(1, 32)); + if (VTIsi1 && SrcReg != AArch64::WZR) { + unsigned ANDReg = emitAnd_ri(MVT::i32, SrcReg, /*TODO:IsKill=*/false, 1); + assert(ANDReg && "Unexpected AND instruction emission failure."); SrcReg = ANDReg; } // Create the base instruction, then add the operands. - MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, - TII.get(StrOpc)).addReg(SrcReg); - AddLoadStoreOperands(Addr, MIB, MachineMemOperand::MOStore, UseUnscaled); + const MCInstrDesc &II = TII.get(Opc); + SrcReg = constrainOperandRegClass(II, SrcReg, II.getNumDefs()); + MachineInstrBuilder MIB = + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(SrcReg); + addLoadStoreOperands(Addr, MIB, MachineMemOperand::MOStore, ScaleFactor, MMO); + return true; } -bool AArch64FastISel::SelectStore(const Instruction *I) { +bool AArch64FastISel::selectStore(const Instruction *I) { MVT VT; - Value *Op0 = I->getOperand(0); + const Value *Op0 = I->getOperand(0); // Verify we have a legal type before going any further. Currently, we handle // simple types that will directly fit in a register (i32/f32/i64/f64) or // those that can be sign or zero-extended to a basic operation (i1/i8/i16). - if (!isLoadStoreTypeLegal(Op0->getType(), VT) || + if (!isTypeSupported(Op0->getType(), VT, /*IsVectorAllowed=*/true) || cast(I)->isAtomic()) return false; - // Get the value to be stored into a register. - unsigned SrcReg = getRegForValue(Op0); - if (SrcReg == 0) + // Get the value to be stored into a register. Use the zero register directly + // when possible to avoid an unnecessary copy and a wasted register. + unsigned SrcReg = 0; + if (const auto *CI = dyn_cast(Op0)) { + if (CI->isZero()) + SrcReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR; + } else if (const auto *CF = dyn_cast(Op0)) { + if (CF->isZero() && !CF->isNegative()) { + VT = MVT::getIntegerVT(VT.getSizeInBits()); + SrcReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR; + } + } + + if (!SrcReg) + SrcReg = getRegForValue(Op0); + + if (!SrcReg) return false; // See if we can handle this address. Address Addr; - if (!ComputeAddress(I->getOperand(1), Addr)) + if (!computeAddress(I->getOperand(1), Addr, I->getOperand(0)->getType())) return false; - if (!EmitStore(VT, SrcReg, Addr)) + if (!emitStore(VT, SrcReg, Addr, createMachineMemOperandFor(I))) return false; return true; } @@ -766,57 +2078,232 @@ static AArch64CC::CondCode getCompareCC(CmpInst::Predicate Pred) { } } -bool AArch64FastISel::SelectBranch(const Instruction *I) { +/// \brief Try to emit a combined compare-and-branch instruction. +bool AArch64FastISel::emitCompareAndBranch(const BranchInst *BI) { + assert(isa(BI->getCondition()) && "Expected cmp instruction"); + const CmpInst *CI = cast(BI->getCondition()); + CmpInst::Predicate Predicate = optimizeCmpPredicate(CI); + + const Value *LHS = CI->getOperand(0); + const Value *RHS = CI->getOperand(1); + + Type *Ty = LHS->getType(); + if (!Ty->isIntegerTy()) + return false; + + unsigned BW = cast(Ty)->getBitWidth(); + if (BW != 1 && BW != 8 && BW != 16 && BW != 32 && BW != 64) + return false; + + MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)]; + MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)]; + + // Try to take advantage of fallthrough opportunities. + if (FuncInfo.MBB->isLayoutSuccessor(TBB)) { + std::swap(TBB, FBB); + Predicate = CmpInst::getInversePredicate(Predicate); + } + + int TestBit = -1; + bool IsCmpNE; + if ((Predicate == CmpInst::ICMP_EQ) || (Predicate == CmpInst::ICMP_NE)) { + if (const auto *C = dyn_cast(LHS)) + if (C->isNullValue()) + std::swap(LHS, RHS); + + if (!isa(RHS)) + return false; + + if (!cast(RHS)->isNullValue()) + return false; + + if (const auto *AI = dyn_cast(LHS)) + if (AI->getOpcode() == Instruction::And && isValueAvailable(AI)) { + const Value *AndLHS = AI->getOperand(0); + const Value *AndRHS = AI->getOperand(1); + + if (const auto *C = dyn_cast(AndLHS)) + if (C->getValue().isPowerOf2()) + std::swap(AndLHS, AndRHS); + + if (const auto *C = dyn_cast(AndRHS)) + if (C->getValue().isPowerOf2()) { + TestBit = C->getValue().logBase2(); + LHS = AndLHS; + } + } + IsCmpNE = Predicate == CmpInst::ICMP_NE; + } else if (Predicate == CmpInst::ICMP_SLT) { + if (!isa(RHS)) + return false; + + if (!cast(RHS)->isNullValue()) + return false; + + TestBit = BW - 1; + IsCmpNE = true; + } else if (Predicate == CmpInst::ICMP_SGT) { + if (!isa(RHS)) + return false; + + if (cast(RHS)->getValue() != -1) + return false; + + TestBit = BW - 1; + IsCmpNE = false; + } else + return false; + + static const unsigned OpcTable[2][2][2] = { + { {AArch64::CBZW, AArch64::CBZX }, + {AArch64::CBNZW, AArch64::CBNZX} }, + { {AArch64::TBZW, AArch64::TBZX }, + {AArch64::TBNZW, AArch64::TBNZX} } + }; + + bool IsBitTest = TestBit != -1; + bool Is64Bit = BW == 64; + if (TestBit < 32 && TestBit >= 0) + Is64Bit = false; + + unsigned Opc = OpcTable[IsBitTest][IsCmpNE][Is64Bit]; + const MCInstrDesc &II = TII.get(Opc); + + unsigned SrcReg = getRegForValue(LHS); + if (!SrcReg) + return false; + bool SrcIsKill = hasTrivialKill(LHS); + + if (BW == 64 && !Is64Bit) + SrcReg = fastEmitInst_extractsubreg(MVT::i32, SrcReg, SrcIsKill, + AArch64::sub_32); + + if ((BW < 32) && !IsBitTest) { + EVT CmpEVT = TLI.getValueType(Ty, true); + SrcReg = + emitIntExt(CmpEVT.getSimpleVT(), SrcReg, MVT::i32, /*isZExt*/ true); + } + + // Emit the combined compare and branch instruction. + SrcReg = constrainOperandRegClass(II, SrcReg, II.getNumDefs()); + MachineInstrBuilder MIB = + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc)) + .addReg(SrcReg, getKillRegState(SrcIsKill)); + if (IsBitTest) + MIB.addImm(TestBit); + MIB.addMBB(TBB); + + // Obtain the branch weight and add the TrueBB to the successor list. + uint32_t BranchWeight = 0; + if (FuncInfo.BPI) + BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(), + TBB->getBasicBlock()); + FuncInfo.MBB->addSuccessor(TBB, BranchWeight); + fastEmitBranch(FBB, DbgLoc); + + return true; +} + +bool AArch64FastISel::selectBranch(const Instruction *I) { const BranchInst *BI = cast(I); + if (BI->isUnconditional()) { + MachineBasicBlock *MSucc = FuncInfo.MBBMap[BI->getSuccessor(0)]; + fastEmitBranch(MSucc, BI->getDebugLoc()); + return true; + } + MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)]; MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)]; AArch64CC::CondCode CC = AArch64CC::NE; if (const CmpInst *CI = dyn_cast(BI->getCondition())) { - if (CI->hasOneUse() && (CI->getParent() == I->getParent())) { - // We may not handle every CC for now. - CC = getCompareCC(CI->getPredicate()); - if (CC == AArch64CC::AL) - return false; + if (CI->hasOneUse() && isValueAvailable(CI)) { + // Try to optimize or fold the cmp. + CmpInst::Predicate Predicate = optimizeCmpPredicate(CI); + switch (Predicate) { + default: + break; + case CmpInst::FCMP_FALSE: + fastEmitBranch(FBB, DbgLoc); + return true; + case CmpInst::FCMP_TRUE: + fastEmitBranch(TBB, DbgLoc); + return true; + } + + // Try to emit a combined compare-and-branch first. + if (emitCompareAndBranch(BI)) + return true; + + // Try to take advantage of fallthrough opportunities. + if (FuncInfo.MBB->isLayoutSuccessor(TBB)) { + std::swap(TBB, FBB); + Predicate = CmpInst::getInversePredicate(Predicate); + } // Emit the cmp. - if (!EmitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned())) + if (!emitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned())) return false; + // FCMP_UEQ and FCMP_ONE cannot be checked with a single branch + // instruction. + CC = getCompareCC(Predicate); + AArch64CC::CondCode ExtraCC = AArch64CC::AL; + switch (Predicate) { + default: + break; + case CmpInst::FCMP_UEQ: + ExtraCC = AArch64CC::EQ; + CC = AArch64CC::VS; + break; + case CmpInst::FCMP_ONE: + ExtraCC = AArch64CC::MI; + CC = AArch64CC::GT; + break; + } + assert((CC != AArch64CC::AL) && "Unexpected condition code."); + + // Emit the extra branch for FCMP_UEQ and FCMP_ONE. + if (ExtraCC != AArch64CC::AL) { + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc)) + .addImm(ExtraCC) + .addMBB(TBB); + } + // Emit the branch. BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc)) .addImm(CC) .addMBB(TBB); - FuncInfo.MBB->addSuccessor(TBB); - FastEmitBranch(FBB, DbgLoc); + // Obtain the branch weight and add the TrueBB to the successor list. + uint32_t BranchWeight = 0; + if (FuncInfo.BPI) + BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(), + TBB->getBasicBlock()); + FuncInfo.MBB->addSuccessor(TBB, BranchWeight); + + fastEmitBranch(FBB, DbgLoc); return true; } } else if (TruncInst *TI = dyn_cast(BI->getCondition())) { MVT SrcVT; - if (TI->hasOneUse() && TI->getParent() == I->getParent() && - (isLoadStoreTypeLegal(TI->getOperand(0)->getType(), SrcVT))) { + if (TI->hasOneUse() && isValueAvailable(TI) && + isTypeSupported(TI->getOperand(0)->getType(), SrcVT)) { unsigned CondReg = getRegForValue(TI->getOperand(0)); - if (CondReg == 0) + if (!CondReg) return false; + bool CondIsKill = hasTrivialKill(TI->getOperand(0)); // Issue an extract_subreg to get the lower 32-bits. - if (SrcVT == MVT::i64) - CondReg = FastEmitInst_extractsubreg(MVT::i32, CondReg, /*Kill=*/true, + if (SrcVT == MVT::i64) { + CondReg = fastEmitInst_extractsubreg(MVT::i32, CondReg, CondIsKill, AArch64::sub_32); + CondIsKill = true; + } - MRI.constrainRegClass(CondReg, &AArch64::GPR32RegClass); - unsigned ANDReg = createResultReg(&AArch64::GPR32spRegClass); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, - TII.get(AArch64::ANDWri), ANDReg) - .addReg(CondReg) - .addImm(AArch64_AM::encodeLogicalImmediate(1, 32)); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, - TII.get(AArch64::SUBSWri)) - .addReg(ANDReg) - .addReg(ANDReg) - .addImm(0) - .addImm(0); + unsigned ANDReg = emitAnd_ri(MVT::i32, CondReg, CondIsKill, 1); + assert(ANDReg && "Unexpected AND instruction emission failure."); + emitICmp_ri(MVT::i32, ANDReg, /*IsKill=*/true, 0); if (FuncInfo.MBB->isLayoutSuccessor(TBB)) { std::swap(TBB, FBB); @@ -825,17 +2312,29 @@ bool AArch64FastISel::SelectBranch(const Instruction *I) { BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc)) .addImm(CC) .addMBB(TBB); - FuncInfo.MBB->addSuccessor(TBB); - FastEmitBranch(FBB, DbgLoc); + + // Obtain the branch weight and add the TrueBB to the successor list. + uint32_t BranchWeight = 0; + if (FuncInfo.BPI) + BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(), + TBB->getBasicBlock()); + FuncInfo.MBB->addSuccessor(TBB, BranchWeight); + + fastEmitBranch(FBB, DbgLoc); return true; } - } else if (const ConstantInt *CI = - dyn_cast(BI->getCondition())) { + } else if (const auto *CI = dyn_cast(BI->getCondition())) { uint64_t Imm = CI->getZExtValue(); MachineBasicBlock *Target = (Imm == 0) ? FBB : TBB; BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::B)) .addMBB(Target); - FuncInfo.MBB->addSuccessor(Target); + + // Obtain the branch weight and add the target to the successor list. + uint32_t BranchWeight = 0; + if (FuncInfo.BPI) + BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(), + Target->getBasicBlock()); + FuncInfo.MBB->addSuccessor(Target, BranchWeight); return true; } else if (foldXALUIntrinsic(CC, I, BI->getCondition())) { // Fake request the condition, otherwise the intrinsic might be completely @@ -848,15 +2347,22 @@ bool AArch64FastISel::SelectBranch(const Instruction *I) { BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc)) .addImm(CC) .addMBB(TBB); - FuncInfo.MBB->addSuccessor(TBB); - FastEmitBranch(FBB, DbgLoc); + // Obtain the branch weight and add the TrueBB to the successor list. + uint32_t BranchWeight = 0; + if (FuncInfo.BPI) + BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(), + TBB->getBasicBlock()); + FuncInfo.MBB->addSuccessor(TBB, BranchWeight); + + fastEmitBranch(FBB, DbgLoc); return true; } unsigned CondReg = getRegForValue(BI->getCondition()); if (CondReg == 0) return false; + bool CondRegIsKill = hasTrivialKill(BI->getCondition()); // We've been divorced from our compare! Our block was split, and // now our compare lives in a predecessor block. We musn't @@ -865,11 +2371,7 @@ bool AArch64FastISel::SelectBranch(const Instruction *I) { // Regardless, the compare has been done in the predecessor block, // and it left a value for us in a virtual register. Ergo, we test // the one-bit value left in the virtual register. - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::SUBSWri), - AArch64::WZR) - .addReg(CondReg) - .addImm(0) - .addImm(0); + emitICmp_ri(MVT::i32, CondReg, CondRegIsKill, 0); if (FuncInfo.MBB->isLayoutSuccessor(TBB)) { std::swap(TBB, FBB); @@ -879,20 +2381,28 @@ bool AArch64FastISel::SelectBranch(const Instruction *I) { BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc)) .addImm(CC) .addMBB(TBB); - FuncInfo.MBB->addSuccessor(TBB); - FastEmitBranch(FBB, DbgLoc); + + // Obtain the branch weight and add the TrueBB to the successor list. + uint32_t BranchWeight = 0; + if (FuncInfo.BPI) + BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(), + TBB->getBasicBlock()); + FuncInfo.MBB->addSuccessor(TBB, BranchWeight); + + fastEmitBranch(FBB, DbgLoc); return true; } -bool AArch64FastISel::SelectIndirectBr(const Instruction *I) { +bool AArch64FastISel::selectIndirectBr(const Instruction *I) { const IndirectBrInst *BI = cast(I); unsigned AddrReg = getRegForValue(BI->getOperand(0)); if (AddrReg == 0) return false; // Emit the indirect branch. - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::BR)) - .addReg(AddrReg); + const MCInstrDesc &II = TII.get(AArch64::BR); + AddrReg = constrainOperandRegClass(II, AddrReg, II.getNumDefs()); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(AddrReg); // Make sure the CFG is up-to-date. for (unsigned i = 0, e = BI->getNumSuccessors(); i != e; ++i) @@ -901,146 +2411,87 @@ bool AArch64FastISel::SelectIndirectBr(const Instruction *I) { return true; } -bool AArch64FastISel::EmitCmp(Value *Src1Value, Value *Src2Value, bool isZExt) { - Type *Ty = Src1Value->getType(); - EVT SrcEVT = TLI.getValueType(Ty, true); - if (!SrcEVT.isSimple()) - return false; - MVT SrcVT = SrcEVT.getSimpleVT(); - - // Check to see if the 2nd operand is a constant that we can encode directly - // in the compare. - uint64_t Imm; - bool UseImm = false; - bool isNegativeImm = false; - if (const ConstantInt *ConstInt = dyn_cast(Src2Value)) { - if (SrcVT == MVT::i64 || SrcVT == MVT::i32 || SrcVT == MVT::i16 || - SrcVT == MVT::i8 || SrcVT == MVT::i1) { - const APInt &CIVal = ConstInt->getValue(); - - Imm = (isZExt) ? CIVal.getZExtValue() : CIVal.getSExtValue(); - if (CIVal.isNegative()) { - isNegativeImm = true; - Imm = -Imm; - } - // FIXME: We can handle more immediates using shifts. - UseImm = ((Imm & 0xfff) == Imm); - } - } else if (const ConstantFP *ConstFP = dyn_cast(Src2Value)) { - if (SrcVT == MVT::f32 || SrcVT == MVT::f64) - if (ConstFP->isZero() && !ConstFP->isNegative()) - UseImm = true; - } - - unsigned ZReg; - unsigned CmpOpc; - bool isICmp = true; - bool needsExt = false; - switch (SrcVT.SimpleTy) { - default: - return false; - case MVT::i1: - case MVT::i8: - case MVT::i16: - needsExt = true; - // Intentional fall-through. - case MVT::i32: - ZReg = AArch64::WZR; - if (UseImm) - CmpOpc = isNegativeImm ? AArch64::ADDSWri : AArch64::SUBSWri; - else - CmpOpc = AArch64::SUBSWrr; - break; - case MVT::i64: - ZReg = AArch64::XZR; - if (UseImm) - CmpOpc = isNegativeImm ? AArch64::ADDSXri : AArch64::SUBSXri; - else - CmpOpc = AArch64::SUBSXrr; - break; - case MVT::f32: - isICmp = false; - CmpOpc = UseImm ? AArch64::FCMPSri : AArch64::FCMPSrr; - break; - case MVT::f64: - isICmp = false; - CmpOpc = UseImm ? AArch64::FCMPDri : AArch64::FCMPDrr; - break; - } - - unsigned SrcReg1 = getRegForValue(Src1Value); - if (SrcReg1 == 0) - return false; - - unsigned SrcReg2; - if (!UseImm) { - SrcReg2 = getRegForValue(Src2Value); - if (SrcReg2 == 0) - return false; - } - - // We have i1, i8, or i16, we need to either zero extend or sign extend. - if (needsExt) { - SrcReg1 = EmitIntExt(SrcVT, SrcReg1, MVT::i32, isZExt); - if (SrcReg1 == 0) - return false; - if (!UseImm) { - SrcReg2 = EmitIntExt(SrcVT, SrcReg2, MVT::i32, isZExt); - if (SrcReg2 == 0) - return false; - } - } - - if (isICmp) { - if (UseImm) - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CmpOpc)) - .addReg(ZReg) - .addReg(SrcReg1) - .addImm(Imm) - .addImm(0); - else - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CmpOpc)) - .addReg(ZReg) - .addReg(SrcReg1) - .addReg(SrcReg2); - } else { - if (UseImm) - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CmpOpc)) - .addReg(SrcReg1); - else - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CmpOpc)) - .addReg(SrcReg1) - .addReg(SrcReg2); - } - return true; -} - -bool AArch64FastISel::SelectCmp(const Instruction *I) { +bool AArch64FastISel::selectCmp(const Instruction *I) { const CmpInst *CI = cast(I); - // We may not handle every CC for now. - AArch64CC::CondCode CC = getCompareCC(CI->getPredicate()); - if (CC == AArch64CC::AL) - return false; + // Try to optimize or fold the cmp. + CmpInst::Predicate Predicate = optimizeCmpPredicate(CI); + unsigned ResultReg = 0; + switch (Predicate) { + default: + break; + case CmpInst::FCMP_FALSE: + ResultReg = createResultReg(&AArch64::GPR32RegClass); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, + TII.get(TargetOpcode::COPY), ResultReg) + .addReg(AArch64::WZR, getKillRegState(true)); + break; + case CmpInst::FCMP_TRUE: + ResultReg = fastEmit_i(MVT::i32, MVT::i32, ISD::Constant, 1); + break; + } + + if (ResultReg) { + updateValueMap(I, ResultReg); + return true; + } // Emit the cmp. - if (!EmitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned())) + if (!emitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned())) return false; + ResultReg = createResultReg(&AArch64::GPR32RegClass); + + // FCMP_UEQ and FCMP_ONE cannot be checked with a single instruction. These + // condition codes are inverted, because they are used by CSINC. + static unsigned CondCodeTable[2][2] = { + { AArch64CC::NE, AArch64CC::VC }, + { AArch64CC::PL, AArch64CC::LE } + }; + unsigned *CondCodes = nullptr; + switch (Predicate) { + default: + break; + case CmpInst::FCMP_UEQ: + CondCodes = &CondCodeTable[0][0]; + break; + case CmpInst::FCMP_ONE: + CondCodes = &CondCodeTable[1][0]; + break; + } + + if (CondCodes) { + unsigned TmpReg1 = createResultReg(&AArch64::GPR32RegClass); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr), + TmpReg1) + .addReg(AArch64::WZR, getKillRegState(true)) + .addReg(AArch64::WZR, getKillRegState(true)) + .addImm(CondCodes[0]); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr), + ResultReg) + .addReg(TmpReg1, getKillRegState(true)) + .addReg(AArch64::WZR, getKillRegState(true)) + .addImm(CondCodes[1]); + + updateValueMap(I, ResultReg); + return true; + } + // Now set a register based on the comparison. + AArch64CC::CondCode CC = getCompareCC(Predicate); + assert((CC != AArch64CC::AL) && "Unexpected condition code."); AArch64CC::CondCode invertedCC = getInvertedCondCode(CC); - unsigned ResultReg = createResultReg(&AArch64::GPR32RegClass); BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr), ResultReg) - .addReg(AArch64::WZR) - .addReg(AArch64::WZR) + .addReg(AArch64::WZR, getKillRegState(true)) + .addReg(AArch64::WZR, getKillRegState(true)) .addImm(invertedCC); - UpdateValueMap(I, ResultReg); + updateValueMap(I, ResultReg); return true; } -bool AArch64FastISel::SelectSelect(const Instruction *I) { +bool AArch64FastISel::selectSelect(const Instruction *I) { const SelectInst *SI = cast(I); EVT DestEVT = TLI.getValueType(SI->getType(), true); @@ -1053,12 +2504,17 @@ bool AArch64FastISel::SelectSelect(const Instruction *I) { return false; unsigned SelectOpc; + const TargetRegisterClass *RC = nullptr; switch (DestVT.SimpleTy) { default: return false; - case MVT::i32: SelectOpc = AArch64::CSELWr; break; - case MVT::i64: SelectOpc = AArch64::CSELXr; break; - case MVT::f32: SelectOpc = AArch64::FCSELSrrr; break; - case MVT::f64: SelectOpc = AArch64::FCSELDrrr; break; + case MVT::i32: + SelectOpc = AArch64::CSELWr; RC = &AArch64::GPR32RegClass; break; + case MVT::i64: + SelectOpc = AArch64::CSELXr; RC = &AArch64::GPR64RegClass; break; + case MVT::f32: + SelectOpc = AArch64::FCSELSrrr; RC = &AArch64::FPR32RegClass; break; + case MVT::f64: + SelectOpc = AArch64::FCSELDrrr; RC = &AArch64::FPR64RegClass; break; } const Value *Cond = SI->getCondition(); @@ -1073,18 +2529,9 @@ bool AArch64FastISel::SelectSelect(const Instruction *I) { bool CondIsKill = hasTrivialKill(Cond); if (NeedTest) { - MRI.constrainRegClass(CondReg, &AArch64::GPR32RegClass); - unsigned ANDReg = createResultReg(&AArch64::GPR32spRegClass); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ANDWri), - ANDReg) - .addReg(CondReg, getKillRegState(CondIsKill)) - .addImm(AArch64_AM::encodeLogicalImmediate(1, 32)); - - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::SUBSWri)) - .addReg(ANDReg) - .addReg(ANDReg) - .addImm(0) - .addImm(0); + unsigned ANDReg = emitAnd_ri(MVT::i32, CondReg, CondIsKill, 1); + assert(ANDReg && "Unexpected AND instruction emission failure."); + emitICmp_ri(MVT::i32, ANDReg, /*IsKill=*/true, 0); } unsigned TrueReg = getRegForValue(SI->getTrueValue()); @@ -1096,18 +2543,13 @@ bool AArch64FastISel::SelectSelect(const Instruction *I) { if (!TrueReg || !FalseReg) return false; - unsigned ResultReg = createResultReg(TLI.getRegClassFor(DestVT)); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(SelectOpc), - ResultReg) - .addReg(TrueReg, getKillRegState(TrueIsKill)) - .addReg(FalseReg, getKillRegState(FalseIsKill)) - .addImm(CC); - - UpdateValueMap(I, ResultReg); + unsigned ResultReg = fastEmitInst_rri(SelectOpc, RC, TrueReg, TrueIsKill, + FalseReg, FalseIsKill, CC); + updateValueMap(I, ResultReg); return true; } -bool AArch64FastISel::SelectFPExt(const Instruction *I) { +bool AArch64FastISel::selectFPExt(const Instruction *I) { Value *V = I->getOperand(0); if (!I->getType()->isDoubleTy() || !V->getType()->isFloatTy()) return false; @@ -1119,11 +2561,11 @@ bool AArch64FastISel::SelectFPExt(const Instruction *I) { unsigned ResultReg = createResultReg(&AArch64::FPR64RegClass); BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::FCVTDSr), ResultReg).addReg(Op); - UpdateValueMap(I, ResultReg); + updateValueMap(I, ResultReg); return true; } -bool AArch64FastISel::SelectFPTrunc(const Instruction *I) { +bool AArch64FastISel::selectFPTrunc(const Instruction *I) { Value *V = I->getOperand(0); if (!I->getType()->isFloatTy() || !V->getType()->isDoubleTy()) return false; @@ -1135,12 +2577,12 @@ bool AArch64FastISel::SelectFPTrunc(const Instruction *I) { unsigned ResultReg = createResultReg(&AArch64::FPR32RegClass); BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::FCVTSDr), ResultReg).addReg(Op); - UpdateValueMap(I, ResultReg); + updateValueMap(I, ResultReg); return true; } // FPToUI and FPToSI -bool AArch64FastISel::SelectFPToInt(const Instruction *I, bool Signed) { +bool AArch64FastISel::selectFPToInt(const Instruction *I, bool Signed) { MVT DestVT; if (!isTypeLegal(I->getType(), DestVT) || DestVT.isVector()) return false; @@ -1169,11 +2611,11 @@ bool AArch64FastISel::SelectFPToInt(const Instruction *I, bool Signed) { DestVT == MVT::i32 ? &AArch64::GPR32RegClass : &AArch64::GPR64RegClass); BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg) .addReg(SrcReg); - UpdateValueMap(I, ResultReg); + updateValueMap(I, ResultReg); return true; } -bool AArch64FastISel::SelectIntToFP(const Instruction *I, bool Signed) { +bool AArch64FastISel::selectIntToFP(const Instruction *I, bool Signed) { MVT DestVT; if (!isTypeLegal(I->getType(), DestVT) || DestVT.isVector()) return false; @@ -1181,22 +2623,21 @@ bool AArch64FastISel::SelectIntToFP(const Instruction *I, bool Signed) { "Unexpected value type."); unsigned SrcReg = getRegForValue(I->getOperand(0)); - if (SrcReg == 0) + if (!SrcReg) return false; + bool SrcIsKill = hasTrivialKill(I->getOperand(0)); EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType(), true); // Handle sign-extension. if (SrcVT == MVT::i16 || SrcVT == MVT::i8 || SrcVT == MVT::i1) { SrcReg = - EmitIntExt(SrcVT.getSimpleVT(), SrcReg, MVT::i32, /*isZExt*/ !Signed); - if (SrcReg == 0) + emitIntExt(SrcVT.getSimpleVT(), SrcReg, MVT::i32, /*isZExt*/ !Signed); + if (!SrcReg) return false; + SrcIsKill = true; } - MRI.constrainRegClass(SrcReg, SrcVT == MVT::i64 ? &AArch64::GPR64RegClass - : &AArch64::GPR32RegClass); - unsigned Opc; if (SrcVT == MVT::i64) { if (Signed) @@ -1210,19 +2651,127 @@ bool AArch64FastISel::SelectIntToFP(const Instruction *I, bool Signed) { Opc = (DestVT == MVT::f32) ? AArch64::UCVTFUWSri : AArch64::UCVTFUWDri; } - unsigned ResultReg = createResultReg(TLI.getRegClassFor(DestVT)); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg) - .addReg(SrcReg); - UpdateValueMap(I, ResultReg); + unsigned ResultReg = fastEmitInst_r(Opc, TLI.getRegClassFor(DestVT), SrcReg, + SrcIsKill); + updateValueMap(I, ResultReg); + return true; +} + +bool AArch64FastISel::fastLowerArguments() { + if (!FuncInfo.CanLowerReturn) + return false; + + const Function *F = FuncInfo.Fn; + if (F->isVarArg()) + return false; + + CallingConv::ID CC = F->getCallingConv(); + if (CC != CallingConv::C) + return false; + + // Only handle simple cases of up to 8 GPR and FPR each. + unsigned GPRCnt = 0; + unsigned FPRCnt = 0; + unsigned Idx = 0; + for (auto const &Arg : F->args()) { + // The first argument is at index 1. + ++Idx; + if (F->getAttributes().hasAttribute(Idx, Attribute::ByVal) || + F->getAttributes().hasAttribute(Idx, Attribute::InReg) || + F->getAttributes().hasAttribute(Idx, Attribute::StructRet) || + F->getAttributes().hasAttribute(Idx, Attribute::Nest)) + return false; + + Type *ArgTy = Arg.getType(); + if (ArgTy->isStructTy() || ArgTy->isArrayTy()) + return false; + + EVT ArgVT = TLI.getValueType(ArgTy); + if (!ArgVT.isSimple()) + return false; + + MVT VT = ArgVT.getSimpleVT().SimpleTy; + if (VT.isFloatingPoint() && !Subtarget->hasFPARMv8()) + return false; + + if (VT.isVector() && + (!Subtarget->hasNEON() || !Subtarget->isLittleEndian())) + return false; + + if (VT >= MVT::i1 && VT <= MVT::i64) + ++GPRCnt; + else if ((VT >= MVT::f16 && VT <= MVT::f64) || VT.is64BitVector() || + VT.is128BitVector()) + ++FPRCnt; + else + return false; + + if (GPRCnt > 8 || FPRCnt > 8) + return false; + } + + static const MCPhysReg Registers[6][8] = { + { AArch64::W0, AArch64::W1, AArch64::W2, AArch64::W3, AArch64::W4, + AArch64::W5, AArch64::W6, AArch64::W7 }, + { AArch64::X0, AArch64::X1, AArch64::X2, AArch64::X3, AArch64::X4, + AArch64::X5, AArch64::X6, AArch64::X7 }, + { AArch64::H0, AArch64::H1, AArch64::H2, AArch64::H3, AArch64::H4, + AArch64::H5, AArch64::H6, AArch64::H7 }, + { AArch64::S0, AArch64::S1, AArch64::S2, AArch64::S3, AArch64::S4, + AArch64::S5, AArch64::S6, AArch64::S7 }, + { AArch64::D0, AArch64::D1, AArch64::D2, AArch64::D3, AArch64::D4, + AArch64::D5, AArch64::D6, AArch64::D7 }, + { AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3, AArch64::Q4, + AArch64::Q5, AArch64::Q6, AArch64::Q7 } + }; + + unsigned GPRIdx = 0; + unsigned FPRIdx = 0; + for (auto const &Arg : F->args()) { + MVT VT = TLI.getSimpleValueType(Arg.getType()); + unsigned SrcReg; + const TargetRegisterClass *RC; + if (VT >= MVT::i1 && VT <= MVT::i32) { + SrcReg = Registers[0][GPRIdx++]; + RC = &AArch64::GPR32RegClass; + VT = MVT::i32; + } else if (VT == MVT::i64) { + SrcReg = Registers[1][GPRIdx++]; + RC = &AArch64::GPR64RegClass; + } else if (VT == MVT::f16) { + SrcReg = Registers[2][FPRIdx++]; + RC = &AArch64::FPR16RegClass; + } else if (VT == MVT::f32) { + SrcReg = Registers[3][FPRIdx++]; + RC = &AArch64::FPR32RegClass; + } else if ((VT == MVT::f64) || VT.is64BitVector()) { + SrcReg = Registers[4][FPRIdx++]; + RC = &AArch64::FPR64RegClass; + } else if (VT.is128BitVector()) { + SrcReg = Registers[5][FPRIdx++]; + RC = &AArch64::FPR128RegClass; + } else + llvm_unreachable("Unexpected value type."); + + unsigned DstReg = FuncInfo.MF->addLiveIn(SrcReg, RC); + // FIXME: Unfortunately it's necessary to emit a copy from the livein copy. + // Without this, EmitLiveInCopies may eliminate the livein if its only + // use is a bitcast (which isn't turned into an instruction). + unsigned ResultReg = createResultReg(RC); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, + TII.get(TargetOpcode::COPY), ResultReg) + .addReg(DstReg, getKillRegState(true)); + updateValueMap(&Arg, ResultReg); + } return true; } -bool AArch64FastISel::ProcessCallArgs(CallLoweringInfo &CLI, +bool AArch64FastISel::processCallArgs(CallLoweringInfo &CLI, SmallVectorImpl &OutVTs, unsigned &NumBytes) { CallingConv::ID CC = CLI.CallConv; SmallVector ArgLocs; - CCState CCInfo(CC, false, *FuncInfo.MF, TM, ArgLocs, *Context); + CCState CCInfo(CC, false, *FuncInfo.MF, ArgLocs, *Context); CCInfo.AnalyzeCallOperands(OutVTs, CLI.OutFlags, CCAssignFnForCall(CC)); // Get a count of how many bytes are to be pushed on the stack. @@ -1250,7 +2799,7 @@ bool AArch64FastISel::ProcessCallArgs(CallLoweringInfo &CLI, case CCValAssign::SExt: { MVT DestVT = VA.getLocVT(); MVT SrcVT = ArgVT; - ArgReg = EmitIntExt(SrcVT, ArgReg, DestVT, /*isZExt=*/false); + ArgReg = emitIntExt(SrcVT, ArgReg, DestVT, /*isZExt=*/false); if (!ArgReg) return false; break; @@ -1260,7 +2809,7 @@ bool AArch64FastISel::ProcessCallArgs(CallLoweringInfo &CLI, case CCValAssign::ZExt: { MVT DestVT = VA.getLocVT(); MVT SrcVT = ArgVT; - ArgReg = EmitIntExt(SrcVT, ArgReg, DestVT, /*isZExt=*/true); + ArgReg = emitIntExt(SrcVT, ArgReg, DestVT, /*isZExt=*/true); if (!ArgReg) return false; break; @@ -1296,14 +2845,19 @@ bool AArch64FastISel::ProcessCallArgs(CallLoweringInfo &CLI, Addr.setReg(AArch64::SP); Addr.setOffset(VA.getLocMemOffset() + BEAlign); - if (!EmitStore(ArgVT, ArgReg, Addr)) + unsigned Alignment = DL.getABITypeAlignment(ArgVal->getType()); + MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand( + MachinePointerInfo::getStack(Addr.getOffset()), + MachineMemOperand::MOStore, ArgVT.getStoreSize(), Alignment); + + if (!emitStore(ArgVT, ArgReg, Addr, MMO)) return false; } } return true; } -bool AArch64FastISel::FinishCall(CallLoweringInfo &CLI, MVT RetVT, +bool AArch64FastISel::finishCall(CallLoweringInfo &CLI, MVT RetVT, unsigned NumBytes) { CallingConv::ID CC = CLI.CallConv; @@ -1315,7 +2869,7 @@ bool AArch64FastISel::FinishCall(CallLoweringInfo &CLI, MVT RetVT, // Now the return value. if (RetVT != MVT::isVoid) { SmallVector RVLocs; - CCState CCInfo(CC, false, *FuncInfo.MF, TM, RVLocs, *Context); + CCState CCInfo(CC, false, *FuncInfo.MF, RVLocs, *Context); CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC)); // Only handle a single return value. @@ -1327,7 +2881,7 @@ bool AArch64FastISel::FinishCall(CallLoweringInfo &CLI, MVT RetVT, unsigned ResultReg = createResultReg(TLI.getRegClassFor(CopyVT)); BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::COPY), ResultReg) - .addReg(RVLocs[0].getLocReg()); + .addReg(RVLocs[0].getLocReg()); CLI.InRegs.push_back(RVLocs[0].getLocReg()); CLI.ResultReg = ResultReg; @@ -1337,15 +2891,27 @@ bool AArch64FastISel::FinishCall(CallLoweringInfo &CLI, MVT RetVT, return true; } -bool AArch64FastISel::FastLowerCall(CallLoweringInfo &CLI) { +bool AArch64FastISel::fastLowerCall(CallLoweringInfo &CLI) { CallingConv::ID CC = CLI.CallConv; + bool IsTailCall = CLI.IsTailCall; bool IsVarArg = CLI.IsVarArg; const Value *Callee = CLI.Callee; const char *SymName = CLI.SymName; - // Only handle global variable Callees. - const GlobalValue *GV = dyn_cast(Callee); - if (!GV) + if (!Callee && !SymName) + return false; + + // Allow SelectionDAG isel to handle tail calls. + if (IsTailCall) + return false; + + CodeModel::Model CM = TM.getCodeModel(); + // Only support the small and large code model. + if (CM != CodeModel::Small && CM != CodeModel::Large) + return false; + + // FIXME: Add large code model support for ELF. + if (CM == CodeModel::Large && !Subtarget->isTargetMachO()) return false; // Let SDISel handle vararg functions. @@ -1380,19 +2946,55 @@ bool AArch64FastISel::FastLowerCall(CallLoweringInfo &CLI) { OutVTs.push_back(VT); } + Address Addr; + if (Callee && !computeCallAddress(Callee, Addr)) + return false; + // Handle the arguments now that we've gotten them. unsigned NumBytes; - if (!ProcessCallArgs(CLI, OutVTs, NumBytes)) + if (!processCallArgs(CLI, OutVTs, NumBytes)) return false; // Issue the call. MachineInstrBuilder MIB; - MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::BL)); - CLI.Call = MIB; - if (!SymName) - MIB.addGlobalAddress(GV, 0, 0); - else - MIB.addExternalSymbol(SymName, 0); + if (CM == CodeModel::Small) { + const MCInstrDesc &II = TII.get(Addr.getReg() ? AArch64::BLR : AArch64::BL); + MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II); + if (SymName) + MIB.addExternalSymbol(SymName, 0); + else if (Addr.getGlobalValue()) + MIB.addGlobalAddress(Addr.getGlobalValue(), 0, 0); + else if (Addr.getReg()) { + unsigned Reg = constrainOperandRegClass(II, Addr.getReg(), 0); + MIB.addReg(Reg); + } else + return false; + } else { + unsigned CallReg = 0; + if (SymName) { + unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP), + ADRPReg) + .addExternalSymbol(SymName, AArch64II::MO_GOT | AArch64II::MO_PAGE); + + CallReg = createResultReg(&AArch64::GPR64RegClass); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::LDRXui), + CallReg) + .addReg(ADRPReg) + .addExternalSymbol(SymName, AArch64II::MO_GOT | AArch64II::MO_PAGEOFF | + AArch64II::MO_NC); + } else if (Addr.getGlobalValue()) + CallReg = materializeGV(Addr.getGlobalValue()); + else if (Addr.getReg()) + CallReg = Addr.getReg(); + + if (!CallReg) + return false; + + const MCInstrDesc &II = TII.get(AArch64::BLR); + CallReg = constrainOperandRegClass(II, CallReg, 0); + MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(CallReg); + } // Add implicit physical register uses to the call. for (auto Reg : CLI.OutRegs) @@ -1402,21 +3004,23 @@ bool AArch64FastISel::FastLowerCall(CallLoweringInfo &CLI) { // Proper defs for return values will be added by setPhysRegsDeadExcept(). MIB.addRegMask(TRI.getCallPreservedMask(CC)); + CLI.Call = MIB; + // Finish off the call including any return values. - return FinishCall(CLI, RetVT, NumBytes); + return finishCall(CLI, RetVT, NumBytes); } -bool AArch64FastISel::IsMemCpySmall(uint64_t Len, unsigned Alignment) { +bool AArch64FastISel::isMemCpySmall(uint64_t Len, unsigned Alignment) { if (Alignment) return Len / Alignment <= 4; else return Len < 32; } -bool AArch64FastISel::TryEmitSmallMemCpy(Address Dest, Address Src, +bool AArch64FastISel::tryEmitSmallMemCpy(Address Dest, Address Src, uint64_t Len, unsigned Alignment) { // Make sure we don't bloat code by inlining very large memcpy's. - if (!IsMemCpySmall(Len, Alignment)) + if (!isMemCpySmall(Len, Alignment)) return false; int64_t UnscaledOffset = 0; @@ -1446,14 +3050,11 @@ bool AArch64FastISel::TryEmitSmallMemCpy(Address Dest, Address Src, } } - bool RV; - unsigned ResultReg; - RV = EmitLoad(VT, ResultReg, Src); - if (!RV) + unsigned ResultReg = emitLoad(VT, VT, Src); + if (!ResultReg) return false; - RV = EmitStore(VT, ResultReg, Dest); - if (!RV) + if (!emitStore(VT, ResultReg, Dest)) return false; int64_t Size = VT.getSizeInBits() / 8; @@ -1491,19 +3092,53 @@ bool AArch64FastISel::foldXALUIntrinsic(AArch64CC::CondCode &CC, if (RetVT != MVT::i32 && RetVT != MVT::i64) return false; + const Value *LHS = II->getArgOperand(0); + const Value *RHS = II->getArgOperand(1); + + // Canonicalize immediate to the RHS. + if (isa(LHS) && !isa(RHS) && + isCommutativeIntrinsic(II)) + std::swap(LHS, RHS); + + // Simplify multiplies. + unsigned IID = II->getIntrinsicID(); + switch (IID) { + default: + break; + case Intrinsic::smul_with_overflow: + if (const auto *C = dyn_cast(RHS)) + if (C->getValue() == 2) + IID = Intrinsic::sadd_with_overflow; + break; + case Intrinsic::umul_with_overflow: + if (const auto *C = dyn_cast(RHS)) + if (C->getValue() == 2) + IID = Intrinsic::uadd_with_overflow; + break; + } + AArch64CC::CondCode TmpCC; - switch (II->getIntrinsicID()) { - default: return false; - case Intrinsic::sadd_with_overflow: - case Intrinsic::ssub_with_overflow: TmpCC = AArch64CC::VS; break; - case Intrinsic::uadd_with_overflow: TmpCC = AArch64CC::HS; break; - case Intrinsic::usub_with_overflow: TmpCC = AArch64CC::LO; break; - case Intrinsic::smul_with_overflow: - case Intrinsic::umul_with_overflow: TmpCC = AArch64CC::NE; break; + switch (IID) { + default: + return false; + case Intrinsic::sadd_with_overflow: + case Intrinsic::ssub_with_overflow: + TmpCC = AArch64CC::VS; + break; + case Intrinsic::uadd_with_overflow: + TmpCC = AArch64CC::HS; + break; + case Intrinsic::usub_with_overflow: + TmpCC = AArch64CC::LO; + break; + case Intrinsic::smul_with_overflow: + case Intrinsic::umul_with_overflow: + TmpCC = AArch64CC::NE; + break; } // Check if both instructions are in the same basic block. - if (II->getParent() != I->getParent()) + if (!isValueAvailable(II)) return false; // Make sure nothing is in the way @@ -1525,7 +3160,7 @@ bool AArch64FastISel::foldXALUIntrinsic(AArch64CC::CondCode &CC, return true; } -bool AArch64FastISel::FastLowerIntrinsicCall(const IntrinsicInst *II) { +bool AArch64FastISel::fastLowerIntrinsicCall(const IntrinsicInst *II) { // FIXME: Handle more intrinsics. switch (II->getIntrinsicID()) { default: return false; @@ -1534,10 +3169,12 @@ bool AArch64FastISel::FastLowerIntrinsicCall(const IntrinsicInst *II) { MFI->setFrameAddressIsTaken(true); const AArch64RegisterInfo *RegInfo = - static_cast(TM.getRegisterInfo()); + static_cast( + TM.getSubtargetImpl()->getRegisterInfo()); unsigned FramePtr = RegInfo->getFrameRegister(*(FuncInfo.MF)); - unsigned SrcReg = FramePtr; - + unsigned SrcReg = MRI.createVirtualRegister(&AArch64::GPR64RegClass); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, + TII.get(TargetOpcode::COPY), SrcReg).addReg(FramePtr); // Recursively load frame address // ldr x0, [fp] // ldr x0, [x0] @@ -1546,14 +3183,13 @@ bool AArch64FastISel::FastLowerIntrinsicCall(const IntrinsicInst *II) { unsigned DestReg; unsigned Depth = cast(II->getOperand(0))->getZExtValue(); while (Depth--) { - DestReg = createResultReg(&AArch64::GPR64RegClass); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, - TII.get(AArch64::LDRXui), DestReg) - .addReg(SrcReg).addImm(0); + DestReg = fastEmitInst_ri(AArch64::LDRXui, &AArch64::GPR64RegClass, + SrcReg, /*IsKill=*/true, 0); + assert(DestReg && "Unexpected LDR instruction emission failure."); SrcReg = DestReg; } - UpdateValueMap(II, SrcReg); + updateValueMap(II, SrcReg); return true; } case Intrinsic::memcpy: @@ -1571,12 +3207,12 @@ bool AArch64FastISel::FastLowerIntrinsicCall(const IntrinsicInst *II) { // if possible. uint64_t Len = cast(MTI->getLength())->getZExtValue(); unsigned Alignment = MTI->getAlignment(); - if (IsMemCpySmall(Len, Alignment)) { + if (isMemCpySmall(Len, Alignment)) { Address Dest, Src; - if (!ComputeAddress(MTI->getRawDest(), Dest) || - !ComputeAddress(MTI->getRawSource(), Src)) + if (!computeAddress(MTI->getRawDest(), Dest) || + !computeAddress(MTI->getRawSource(), Src)) return false; - if (TryEmitSmallMemCpy(Dest, Src, Len, Alignment)) + if (tryEmitSmallMemCpy(Dest, Src, Len, Alignment)) return true; } } @@ -1590,7 +3226,7 @@ bool AArch64FastISel::FastLowerIntrinsicCall(const IntrinsicInst *II) { return false; const char *IntrMemName = isa(II) ? "memcpy" : "memmove"; - return LowerCallTo(II, IntrMemName, II->getNumArgOperands() - 2); + return lowerCallTo(II, IntrMemName, II->getNumArgOperands() - 2); } case Intrinsic::memset: { const MemSetInst *MSI = cast(II); @@ -1606,13 +3242,82 @@ bool AArch64FastISel::FastLowerIntrinsicCall(const IntrinsicInst *II) { // address spaces. return false; - return LowerCallTo(II, "memset", II->getNumArgOperands() - 2); + return lowerCallTo(II, "memset", II->getNumArgOperands() - 2); + } + case Intrinsic::sin: + case Intrinsic::cos: + case Intrinsic::pow: { + MVT RetVT; + if (!isTypeLegal(II->getType(), RetVT)) + return false; + + if (RetVT != MVT::f32 && RetVT != MVT::f64) + return false; + + static const RTLIB::Libcall LibCallTable[3][2] = { + { RTLIB::SIN_F32, RTLIB::SIN_F64 }, + { RTLIB::COS_F32, RTLIB::COS_F64 }, + { RTLIB::POW_F32, RTLIB::POW_F64 } + }; + RTLIB::Libcall LC; + bool Is64Bit = RetVT == MVT::f64; + switch (II->getIntrinsicID()) { + default: + llvm_unreachable("Unexpected intrinsic."); + case Intrinsic::sin: + LC = LibCallTable[0][Is64Bit]; + break; + case Intrinsic::cos: + LC = LibCallTable[1][Is64Bit]; + break; + case Intrinsic::pow: + LC = LibCallTable[2][Is64Bit]; + break; + } + + ArgListTy Args; + Args.reserve(II->getNumArgOperands()); + + // Populate the argument list. + for (auto &Arg : II->arg_operands()) { + ArgListEntry Entry; + Entry.Val = Arg; + Entry.Ty = Arg->getType(); + Args.push_back(Entry); + } + + CallLoweringInfo CLI; + CLI.setCallee(TLI.getLibcallCallingConv(LC), II->getType(), + TLI.getLibcallName(LC), std::move(Args)); + if (!lowerCallTo(CLI)) + return false; + updateValueMap(II, CLI.ResultReg); + return true; } case Intrinsic::trap: { BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::BRK)) .addImm(1); return true; } + case Intrinsic::sqrt: { + Type *RetTy = II->getCalledFunction()->getReturnType(); + + MVT VT; + if (!isTypeLegal(RetTy, VT)) + return false; + + unsigned Op0Reg = getRegForValue(II->getOperand(0)); + if (!Op0Reg) + return false; + bool Op0IsKill = hasTrivialKill(II->getOperand(0)); + + unsigned ResultReg = fastEmit_r(VT, VT, ISD::FSQRT, Op0Reg, Op0IsKill); + if (!ResultReg) + return false; + + updateValueMap(II, ResultReg); + return true; + } case Intrinsic::sadd_with_overflow: case Intrinsic::uadd_with_overflow: case Intrinsic::ssub_with_overflow: @@ -1623,7 +3328,6 @@ bool AArch64FastISel::FastLowerIntrinsicCall(const IntrinsicInst *II) { const Function *Callee = II->getCalledFunction(); auto *Ty = cast(Callee->getReturnType()); Type *RetTy = Ty->getTypeAtIndex(0U); - Type *CondTy = Ty->getTypeAtIndex(1); MVT VT; if (!isTypeLegal(RetTy, VT)) @@ -1639,156 +3343,130 @@ bool AArch64FastISel::FastLowerIntrinsicCall(const IntrinsicInst *II) { isCommutativeIntrinsic(II)) std::swap(LHS, RHS); - unsigned LHSReg = getRegForValue(LHS); - if (!LHSReg) - return false; - bool LHSIsKill = hasTrivialKill(LHS); - - unsigned RHSReg = 0; - bool RHSIsKill = false; - bool UseImm = true; - if (!isa(RHS)) { - RHSReg = getRegForValue(RHS); - if (!RHSReg) - return false; - RHSIsKill = hasTrivialKill(RHS); - UseImm = false; + // Simplify multiplies. + unsigned IID = II->getIntrinsicID(); + switch (IID) { + default: + break; + case Intrinsic::smul_with_overflow: + if (const auto *C = dyn_cast(RHS)) + if (C->getValue() == 2) { + IID = Intrinsic::sadd_with_overflow; + RHS = LHS; + } + break; + case Intrinsic::umul_with_overflow: + if (const auto *C = dyn_cast(RHS)) + if (C->getValue() == 2) { + IID = Intrinsic::uadd_with_overflow; + RHS = LHS; + } + break; } - unsigned Opc = 0; - unsigned MulReg = 0; + unsigned ResultReg1 = 0, ResultReg2 = 0, MulReg = 0; AArch64CC::CondCode CC = AArch64CC::Invalid; - bool Is64Bit = VT == MVT::i64; - switch (II->getIntrinsicID()) { + switch (IID) { default: llvm_unreachable("Unexpected intrinsic!"); case Intrinsic::sadd_with_overflow: - if (UseImm) - Opc = Is64Bit ? AArch64::ADDSXri : AArch64::ADDSWri; - else - Opc = Is64Bit ? AArch64::ADDSXrr : AArch64::ADDSWrr; + ResultReg1 = emitAdd(VT, LHS, RHS, /*SetFlags=*/true); CC = AArch64CC::VS; break; case Intrinsic::uadd_with_overflow: - if (UseImm) - Opc = Is64Bit ? AArch64::ADDSXri : AArch64::ADDSWri; - else - Opc = Is64Bit ? AArch64::ADDSXrr : AArch64::ADDSWrr; + ResultReg1 = emitAdd(VT, LHS, RHS, /*SetFlags=*/true); CC = AArch64CC::HS; break; case Intrinsic::ssub_with_overflow: - if (UseImm) - Opc = Is64Bit ? AArch64::SUBSXri : AArch64::SUBSWri; - else - Opc = Is64Bit ? AArch64::SUBSXrr : AArch64::SUBSWrr; + ResultReg1 = emitSub(VT, LHS, RHS, /*SetFlags=*/true); CC = AArch64CC::VS; break; case Intrinsic::usub_with_overflow: - if (UseImm) - Opc = Is64Bit ? AArch64::SUBSXri : AArch64::SUBSWri; - else - Opc = Is64Bit ? AArch64::SUBSXrr : AArch64::SUBSWrr; + ResultReg1 = emitSub(VT, LHS, RHS, /*SetFlags=*/true); CC = AArch64CC::LO; break; case Intrinsic::smul_with_overflow: { CC = AArch64CC::NE; - if (UseImm) { - RHSReg = getRegForValue(RHS); - if (!RHSReg) - return false; - RHSIsKill = hasTrivialKill(RHS); - } + unsigned LHSReg = getRegForValue(LHS); + if (!LHSReg) + return false; + bool LHSIsKill = hasTrivialKill(LHS); + + unsigned RHSReg = getRegForValue(RHS); + if (!RHSReg) + return false; + bool RHSIsKill = hasTrivialKill(RHS); + if (VT == MVT::i32) { - MulReg = Emit_SMULL_rr(MVT::i64, LHSReg, LHSIsKill, RHSReg, RHSIsKill); - unsigned ShiftReg = Emit_LSR_ri(MVT::i64, MulReg, false, 32); - MulReg = FastEmitInst_extractsubreg(VT, MulReg, /*IsKill=*/true, + MulReg = emitSMULL_rr(MVT::i64, LHSReg, LHSIsKill, RHSReg, RHSIsKill); + unsigned ShiftReg = emitLSR_ri(MVT::i64, MVT::i64, MulReg, + /*IsKill=*/false, 32); + MulReg = fastEmitInst_extractsubreg(VT, MulReg, /*IsKill=*/true, AArch64::sub_32); - ShiftReg = FastEmitInst_extractsubreg(VT, ShiftReg, /*IsKill=*/true, + ShiftReg = fastEmitInst_extractsubreg(VT, ShiftReg, /*IsKill=*/true, AArch64::sub_32); - unsigned CmpReg = createResultReg(TLI.getRegClassFor(VT)); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, - TII.get(AArch64::SUBSWrs), CmpReg) - .addReg(ShiftReg, getKillRegState(true)) - .addReg(MulReg, getKillRegState(false)) - .addImm(159); // 159 <-> asr #31 + emitSubs_rs(VT, ShiftReg, /*IsKill=*/true, MulReg, /*IsKill=*/false, + AArch64_AM::ASR, 31, /*WantResult=*/false); } else { assert(VT == MVT::i64 && "Unexpected value type."); - MulReg = Emit_MUL_rr(VT, LHSReg, LHSIsKill, RHSReg, RHSIsKill); - unsigned SMULHReg = FastEmit_rr(VT, VT, ISD::MULHS, LHSReg, LHSIsKill, + MulReg = emitMul_rr(VT, LHSReg, LHSIsKill, RHSReg, RHSIsKill); + unsigned SMULHReg = fastEmit_rr(VT, VT, ISD::MULHS, LHSReg, LHSIsKill, RHSReg, RHSIsKill); - unsigned CmpReg = createResultReg(TLI.getRegClassFor(VT)); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, - TII.get(AArch64::SUBSXrs), CmpReg) - .addReg(SMULHReg, getKillRegState(true)) - .addReg(MulReg, getKillRegState(false)) - .addImm(191); // 191 <-> asr #63 + emitSubs_rs(VT, SMULHReg, /*IsKill=*/true, MulReg, /*IsKill=*/false, + AArch64_AM::ASR, 63, /*WantResult=*/false); } break; } case Intrinsic::umul_with_overflow: { CC = AArch64CC::NE; - if (UseImm) { - RHSReg = getRegForValue(RHS); - if (!RHSReg) - return false; - RHSIsKill = hasTrivialKill(RHS); - } + unsigned LHSReg = getRegForValue(LHS); + if (!LHSReg) + return false; + bool LHSIsKill = hasTrivialKill(LHS); + + unsigned RHSReg = getRegForValue(RHS); + if (!RHSReg) + return false; + bool RHSIsKill = hasTrivialKill(RHS); + if (VT == MVT::i32) { - MulReg = Emit_UMULL_rr(MVT::i64, LHSReg, LHSIsKill, RHSReg, RHSIsKill); - unsigned CmpReg = createResultReg(TLI.getRegClassFor(MVT::i64)); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, - TII.get(AArch64::SUBSXrs), CmpReg) - .addReg(AArch64::XZR, getKillRegState(true)) - .addReg(MulReg, getKillRegState(false)) - .addImm(96); // 96 <-> lsr #32 - MulReg = FastEmitInst_extractsubreg(VT, MulReg, /*IsKill=*/true, + MulReg = emitUMULL_rr(MVT::i64, LHSReg, LHSIsKill, RHSReg, RHSIsKill); + emitSubs_rs(MVT::i64, AArch64::XZR, /*IsKill=*/true, MulReg, + /*IsKill=*/false, AArch64_AM::LSR, 32, + /*WantResult=*/false); + MulReg = fastEmitInst_extractsubreg(VT, MulReg, /*IsKill=*/true, AArch64::sub_32); } else { assert(VT == MVT::i64 && "Unexpected value type."); - MulReg = Emit_MUL_rr(VT, LHSReg, LHSIsKill, RHSReg, RHSIsKill); - unsigned UMULHReg = FastEmit_rr(VT, VT, ISD::MULHU, LHSReg, LHSIsKill, + MulReg = emitMul_rr(VT, LHSReg, LHSIsKill, RHSReg, RHSIsKill); + unsigned UMULHReg = fastEmit_rr(VT, VT, ISD::MULHU, LHSReg, LHSIsKill, RHSReg, RHSIsKill); - unsigned CmpReg = createResultReg(TLI.getRegClassFor(VT)); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, - TII.get(AArch64::SUBSXrr), CmpReg) - .addReg(AArch64::XZR, getKillRegState(true)) - .addReg(UMULHReg, getKillRegState(false)); + emitSubs_rr(VT, AArch64::XZR, /*IsKill=*/true, UMULHReg, + /*IsKill=*/false, /*WantResult=*/false); } break; } } - unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT)); - if (Opc) { - MachineInstrBuilder MIB; - MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), - ResultReg) - .addReg(LHSReg, getKillRegState(LHSIsKill)); - if (UseImm) - MIB.addImm(cast(RHS)->getZExtValue()); - else - MIB.addReg(RHSReg, getKillRegState(RHSIsKill)); - } - else + if (MulReg) { + ResultReg1 = createResultReg(TLI.getRegClassFor(VT)); BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, - TII.get(TargetOpcode::COPY), ResultReg) - .addReg(MulReg); - - unsigned ResultReg2 = FuncInfo.CreateRegs(CondTy); - assert((ResultReg+1) == ResultReg2 && "Nonconsecutive result registers."); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr), - ResultReg2) - .addReg(AArch64::WZR, getKillRegState(true)) - .addReg(AArch64::WZR, getKillRegState(true)) - .addImm(getInvertedCondCode(CC)); + TII.get(TargetOpcode::COPY), ResultReg1).addReg(MulReg); + } - UpdateValueMap(II, ResultReg, 2); + ResultReg2 = fastEmitInst_rri(AArch64::CSINCWr, &AArch64::GPR32RegClass, + AArch64::WZR, /*IsKill=*/true, AArch64::WZR, + /*IsKill=*/true, getInvertedCondCode(CC)); + (void)ResultReg2; + assert((ResultReg1 + 1) == ResultReg2 && + "Nonconsecutive result registers."); + updateValueMap(II, ResultReg1, 2); return true; } } return false; } -bool AArch64FastISel::SelectRet(const Instruction *I) { +bool AArch64FastISel::selectRet(const Instruction *I) { const ReturnInst *Ret = cast(I); const Function &F = *I->getParent()->getParent(); @@ -1808,8 +3486,7 @@ bool AArch64FastISel::SelectRet(const Instruction *I) { // Analyze operands of the call, assigning locations to each operand. SmallVector ValLocs; - CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, TM, ValLocs, - I->getContext()); + CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, ValLocs, I->getContext()); CCAssignFn *RetCC = CC == CallingConv::WebKit_JS ? RetCC_AArch64_WebKit_JS : RetCC_AArch64_AAPCS; CCInfo.AnalyzeReturn(Outs, RetCC); @@ -1822,11 +3499,14 @@ bool AArch64FastISel::SelectRet(const Instruction *I) { const Value *RV = Ret->getOperand(0); // Don't bother handling odd stuff for now. - if (VA.getLocInfo() != CCValAssign::Full) + if ((VA.getLocInfo() != CCValAssign::Full) && + (VA.getLocInfo() != CCValAssign::BCvt)) return false; + // Only handle register returns for now. if (!VA.isRegLoc()) return false; + unsigned Reg = getRegForValue(RV); if (Reg == 0) return false; @@ -1842,12 +3522,14 @@ bool AArch64FastISel::SelectRet(const Instruction *I) { return false; // Vectors (of > 1 lane) in big endian need tricky handling. - if (RVEVT.isVector() && RVEVT.getVectorNumElements() > 1) + if (RVEVT.isVector() && RVEVT.getVectorNumElements() > 1 && + !Subtarget->isLittleEndian()) return false; MVT RVVT = RVEVT.getSimpleVT(); if (RVVT == MVT::f128) return false; + MVT DestVT = VA.getValVT(); // Special handling for extended integers. if (RVVT != DestVT) { @@ -1857,8 +3539,8 @@ bool AArch64FastISel::SelectRet(const Instruction *I) { if (!Outs[0].Flags.isZExt() && !Outs[0].Flags.isSExt()) return false; - bool isZExt = Outs[0].Flags.isZExt(); - SrcReg = EmitIntExt(RVVT, SrcReg, DestVT, isZExt); + bool IsZExt = Outs[0].Flags.isZExt(); + SrcReg = emitIntExt(RVVT, SrcReg, DestVT, IsZExt); if (SrcReg == 0) return false; } @@ -1878,7 +3560,7 @@ bool AArch64FastISel::SelectRet(const Instruction *I) { return true; } -bool AArch64FastISel::SelectTrunc(const Instruction *I) { +bool AArch64FastISel::selectTrunc(const Instruction *I) { Type *DestTy = I->getType(); Value *Op = I->getOperand(0); Type *SrcTy = Op->getType(); @@ -1903,10 +3585,14 @@ bool AArch64FastISel::SelectTrunc(const Instruction *I) { unsigned SrcReg = getRegForValue(Op); if (!SrcReg) return false; + bool SrcIsKill = hasTrivialKill(Op); // If we're truncating from i64 to a smaller non-legal type then generate an - // AND. Otherwise, we know the high bits are undefined and a truncate doesn't - // generate any code. + // AND. Otherwise, we know the high bits are undefined and a truncate only + // generate a COPY. We cannot mark the source register also as result + // register, because this can incorrectly transfer the kill flag onto the + // source register. + unsigned ResultReg; if (SrcVT == MVT::i64) { uint64_t Mask = 0; switch (DestVT.SimpleTy) { @@ -1924,23 +3610,23 @@ bool AArch64FastISel::SelectTrunc(const Instruction *I) { break; } // Issue an extract_subreg to get the lower 32-bits. - unsigned Reg32 = FastEmitInst_extractsubreg(MVT::i32, SrcReg, /*Kill=*/true, + unsigned Reg32 = fastEmitInst_extractsubreg(MVT::i32, SrcReg, SrcIsKill, AArch64::sub_32); - MRI.constrainRegClass(Reg32, &AArch64::GPR32RegClass); // Create the AND instruction which performs the actual truncation. - unsigned ANDReg = createResultReg(&AArch64::GPR32spRegClass); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ANDWri), - ANDReg) - .addReg(Reg32) - .addImm(AArch64_AM::encodeLogicalImmediate(Mask, 32)); - SrcReg = ANDReg; + ResultReg = emitAnd_ri(MVT::i32, Reg32, /*IsKill=*/true, Mask); + assert(ResultReg && "Unexpected AND instruction emission failure."); + } else { + ResultReg = createResultReg(&AArch64::GPR32RegClass); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, + TII.get(TargetOpcode::COPY), ResultReg) + .addReg(SrcReg, getKillRegState(SrcIsKill)); } - UpdateValueMap(I, SrcReg); + updateValueMap(I, ResultReg); return true; } -unsigned AArch64FastISel::Emiti1Ext(unsigned SrcReg, MVT DestVT, bool isZExt) { +unsigned AArch64FastISel::emiti1Ext(unsigned SrcReg, MVT DestVT, bool IsZExt) { assert((DestVT == MVT::i8 || DestVT == MVT::i16 || DestVT == MVT::i32 || DestVT == MVT::i64) && "Unexpected value type."); @@ -1948,14 +3634,9 @@ unsigned AArch64FastISel::Emiti1Ext(unsigned SrcReg, MVT DestVT, bool isZExt) { if (DestVT == MVT::i8 || DestVT == MVT::i16) DestVT = MVT::i32; - if (isZExt) { - MRI.constrainRegClass(SrcReg, &AArch64::GPR32RegClass); - unsigned ResultReg = createResultReg(&AArch64::GPR32spRegClass); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ANDWri), - ResultReg) - .addReg(SrcReg) - .addImm(AArch64_AM::encodeLogicalImmediate(1, 32)); - + if (IsZExt) { + unsigned ResultReg = emitAnd_ri(MVT::i32, SrcReg, /*TODO:IsKill=*/false, 1); + assert(ResultReg && "Unexpected AND instruction emission failure."); if (DestVT == MVT::i64) { // We're ZExt i1 to i64. The ANDWri Wd, Ws, #1 implicitly clears the // upper 32 bits. Emit a SUBREG_TO_REG to extend from Wd to Xd. @@ -1973,17 +3654,12 @@ unsigned AArch64FastISel::Emiti1Ext(unsigned SrcReg, MVT DestVT, bool isZExt) { // FIXME: We're SExt i1 to i64. return 0; } - unsigned ResultReg = createResultReg(&AArch64::GPR32RegClass); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::SBFMWri), - ResultReg) - .addReg(SrcReg) - .addImm(0) - .addImm(0); - return ResultReg; + return fastEmitInst_rii(AArch64::SBFMWri, &AArch64::GPR32RegClass, SrcReg, + /*TODO:IsKill=*/false, 0, 0); } } -unsigned AArch64FastISel::Emit_MUL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill, +unsigned AArch64FastISel::emitMul_rr(MVT RetVT, unsigned Op0, bool Op0IsKill, unsigned Op1, bool Op1IsKill) { unsigned Opc, ZReg; switch (RetVT.SimpleTy) { @@ -1997,104 +3673,332 @@ unsigned AArch64FastISel::Emit_MUL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill, Opc = AArch64::MADDXrrr; ZReg = AArch64::XZR; break; } - // Create the base instruction, then add the operands. - unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT)); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg) - .addReg(Op0, getKillRegState(Op0IsKill)) - .addReg(Op1, getKillRegState(Op1IsKill)) - .addReg(ZReg, getKillRegState(true)); - - return ResultReg; + const TargetRegisterClass *RC = + (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass; + return fastEmitInst_rrr(Opc, RC, Op0, Op0IsKill, Op1, Op1IsKill, + /*IsKill=*/ZReg, true); } -unsigned AArch64FastISel::Emit_SMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill, +unsigned AArch64FastISel::emitSMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill, unsigned Op1, bool Op1IsKill) { if (RetVT != MVT::i64) return 0; - // Create the base instruction, then add the operands. - unsigned ResultReg = createResultReg(&AArch64::GPR64RegClass); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::SMADDLrrr), - ResultReg) - .addReg(Op0, getKillRegState(Op0IsKill)) - .addReg(Op1, getKillRegState(Op1IsKill)) - .addReg(AArch64::XZR, getKillRegState(true)); - - return ResultReg; + return fastEmitInst_rrr(AArch64::SMADDLrrr, &AArch64::GPR64RegClass, + Op0, Op0IsKill, Op1, Op1IsKill, + AArch64::XZR, /*IsKill=*/true); } -unsigned AArch64FastISel::Emit_UMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill, +unsigned AArch64FastISel::emitUMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill, unsigned Op1, bool Op1IsKill) { if (RetVT != MVT::i64) return 0; - // Create the base instruction, then add the operands. - unsigned ResultReg = createResultReg(&AArch64::GPR64RegClass); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::UMADDLrrr), - ResultReg) - .addReg(Op0, getKillRegState(Op0IsKill)) - .addReg(Op1, getKillRegState(Op1IsKill)) - .addReg(AArch64::XZR, getKillRegState(true)); - - return ResultReg; + return fastEmitInst_rrr(AArch64::UMADDLrrr, &AArch64::GPR64RegClass, + Op0, Op0IsKill, Op1, Op1IsKill, + AArch64::XZR, /*IsKill=*/true); } -unsigned AArch64FastISel::Emit_LSL_ri(MVT RetVT, unsigned Op0, bool Op0IsKill, - uint64_t Shift) { - unsigned Opc, ImmR, ImmS; +unsigned AArch64FastISel::emitLSL_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill, + unsigned Op1Reg, bool Op1IsKill) { + unsigned Opc = 0; + bool NeedTrunc = false; + uint64_t Mask = 0; switch (RetVT.SimpleTy) { default: return 0; - case MVT::i8: - case MVT::i16: - case MVT::i32: - RetVT = MVT::i32; - Opc = AArch64::UBFMWri; ImmR = -Shift % 32; ImmS = 31 - Shift; break; - case MVT::i64: - Opc = AArch64::UBFMXri; ImmR = -Shift % 64; ImmS = 63 - Shift; break; + case MVT::i8: Opc = AArch64::LSLVWr; NeedTrunc = true; Mask = 0xff; break; + case MVT::i16: Opc = AArch64::LSLVWr; NeedTrunc = true; Mask = 0xffff; break; + case MVT::i32: Opc = AArch64::LSLVWr; break; + case MVT::i64: Opc = AArch64::LSLVXr; break; + } + + const TargetRegisterClass *RC = + (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass; + if (NeedTrunc) { + Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Op1IsKill, Mask); + Op1IsKill = true; } + unsigned ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg, + Op1IsKill); + if (NeedTrunc) + ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask); + return ResultReg; +} + +unsigned AArch64FastISel::emitLSL_ri(MVT RetVT, MVT SrcVT, unsigned Op0, + bool Op0IsKill, uint64_t Shift, + bool IsZext) { + assert(RetVT.SimpleTy >= SrcVT.SimpleTy && + "Unexpected source/return type pair."); + assert((SrcVT == MVT::i1 || SrcVT == MVT::i8 || SrcVT == MVT::i16 || + SrcVT == MVT::i32 || SrcVT == MVT::i64) && + "Unexpected source value type."); + assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 || + RetVT == MVT::i64) && "Unexpected return value type."); + + bool Is64Bit = (RetVT == MVT::i64); + unsigned RegSize = Is64Bit ? 64 : 32; + unsigned DstBits = RetVT.getSizeInBits(); + unsigned SrcBits = SrcVT.getSizeInBits(); + + // Don't deal with undefined shifts. + if (Shift >= DstBits) + return 0; - return FastEmitInst_rii(Opc, TLI.getRegClassFor(RetVT), Op0, Op0IsKill, ImmR, - ImmS); + // For immediate shifts we can fold the zero-/sign-extension into the shift. + // {S|U}BFM Wd, Wn, #r, #s + // Wd<32+s-r,32-r> = Wn when r > s + + // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16 + // %2 = shl i16 %1, 4 + // Wd<32+7-28,32-28> = Wn<7:0> <- clamp s to 7 + // 0b1111_1111_1111_1111__1111_1010_1010_0000 sext + // 0b0000_0000_0000_0000__0000_0101_0101_0000 sext | zext + // 0b0000_0000_0000_0000__0000_1010_1010_0000 zext + + // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16 + // %2 = shl i16 %1, 8 + // Wd<32+7-24,32-24> = Wn<7:0> + // 0b1111_1111_1111_1111__1010_1010_0000_0000 sext + // 0b0000_0000_0000_0000__0101_0101_0000_0000 sext | zext + // 0b0000_0000_0000_0000__1010_1010_0000_0000 zext + + // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16 + // %2 = shl i16 %1, 12 + // Wd<32+3-20,32-20> = Wn<3:0> + // 0b1111_1111_1111_1111__1010_0000_0000_0000 sext + // 0b0000_0000_0000_0000__0101_0000_0000_0000 sext | zext + // 0b0000_0000_0000_0000__1010_0000_0000_0000 zext + + unsigned ImmR = RegSize - Shift; + // Limit the width to the length of the source type. + unsigned ImmS = std::min(SrcBits - 1, DstBits - 1 - Shift); + static const unsigned OpcTable[2][2] = { + {AArch64::SBFMWri, AArch64::SBFMXri}, + {AArch64::UBFMWri, AArch64::UBFMXri} + }; + unsigned Opc = OpcTable[IsZext][Is64Bit]; + const TargetRegisterClass *RC = + Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass; + if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) { + unsigned TmpReg = MRI.createVirtualRegister(RC); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, + TII.get(AArch64::SUBREG_TO_REG), TmpReg) + .addImm(0) + .addReg(Op0, getKillRegState(Op0IsKill)) + .addImm(AArch64::sub_32); + Op0 = TmpReg; + Op0IsKill = true; + } + return fastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS); } -unsigned AArch64FastISel::Emit_LSR_ri(MVT RetVT, unsigned Op0, bool Op0IsKill, - uint64_t Shift) { - unsigned Opc, ImmS; +unsigned AArch64FastISel::emitLSR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill, + unsigned Op1Reg, bool Op1IsKill) { + unsigned Opc = 0; + bool NeedTrunc = false; + uint64_t Mask = 0; switch (RetVT.SimpleTy) { default: return 0; - case MVT::i8: - case MVT::i16: - case MVT::i32: - RetVT = MVT::i32; - Opc = AArch64::UBFMWri; ImmS = 31; break; - case MVT::i64: - Opc = AArch64::UBFMXri; ImmS = 63; break; + case MVT::i8: Opc = AArch64::LSRVWr; NeedTrunc = true; Mask = 0xff; break; + case MVT::i16: Opc = AArch64::LSRVWr; NeedTrunc = true; Mask = 0xffff; break; + case MVT::i32: Opc = AArch64::LSRVWr; break; + case MVT::i64: Opc = AArch64::LSRVXr; break; + } + + const TargetRegisterClass *RC = + (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass; + if (NeedTrunc) { + Op0Reg = emitAnd_ri(MVT::i32, Op0Reg, Op0IsKill, Mask); + Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Op1IsKill, Mask); + Op0IsKill = Op1IsKill = true; + } + unsigned ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg, + Op1IsKill); + if (NeedTrunc) + ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask); + return ResultReg; +} + +unsigned AArch64FastISel::emitLSR_ri(MVT RetVT, MVT SrcVT, unsigned Op0, + bool Op0IsKill, uint64_t Shift, + bool IsZExt) { + assert(RetVT.SimpleTy >= SrcVT.SimpleTy && + "Unexpected source/return type pair."); + assert((SrcVT == MVT::i8 || SrcVT == MVT::i16 || SrcVT == MVT::i32 || + SrcVT == MVT::i64) && "Unexpected source value type."); + assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 || + RetVT == MVT::i64) && "Unexpected return value type."); + + bool Is64Bit = (RetVT == MVT::i64); + unsigned RegSize = Is64Bit ? 64 : 32; + unsigned DstBits = RetVT.getSizeInBits(); + unsigned SrcBits = SrcVT.getSizeInBits(); + + // Don't deal with undefined shifts. + if (Shift >= DstBits) + return 0; + + // For immediate shifts we can fold the zero-/sign-extension into the shift. + // {S|U}BFM Wd, Wn, #r, #s + // Wd = Wn when r <= s + + // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16 + // %2 = lshr i16 %1, 4 + // Wd<7-4:0> = Wn<7:4> + // 0b0000_0000_0000_0000__0000_1111_1111_1010 sext + // 0b0000_0000_0000_0000__0000_0000_0000_0101 sext | zext + // 0b0000_0000_0000_0000__0000_0000_0000_1010 zext + + // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16 + // %2 = lshr i16 %1, 8 + // Wd<7-7,0> = Wn<7:7> + // 0b0000_0000_0000_0000__0000_0000_1111_1111 sext + // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext + // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext + + // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16 + // %2 = lshr i16 %1, 12 + // Wd<7-7,0> = Wn<7:7> <- clamp r to 7 + // 0b0000_0000_0000_0000__0000_0000_0000_1111 sext + // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext + // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext + + if (Shift >= SrcBits && IsZExt) + return materializeInt(ConstantInt::get(*Context, APInt(RegSize, 0)), RetVT); + + // It is not possible to fold a sign-extend into the LShr instruction. In this + // case emit a sign-extend. + if (!IsZExt) { + Op0 = emitIntExt(SrcVT, Op0, RetVT, IsZExt); + if (!Op0) + return 0; + Op0IsKill = true; + SrcVT = RetVT; + SrcBits = SrcVT.getSizeInBits(); + IsZExt = true; + } + + unsigned ImmR = std::min(SrcBits - 1, Shift); + unsigned ImmS = SrcBits - 1; + static const unsigned OpcTable[2][2] = { + {AArch64::SBFMWri, AArch64::SBFMXri}, + {AArch64::UBFMWri, AArch64::UBFMXri} + }; + unsigned Opc = OpcTable[IsZExt][Is64Bit]; + const TargetRegisterClass *RC = + Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass; + if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) { + unsigned TmpReg = MRI.createVirtualRegister(RC); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, + TII.get(AArch64::SUBREG_TO_REG), TmpReg) + .addImm(0) + .addReg(Op0, getKillRegState(Op0IsKill)) + .addImm(AArch64::sub_32); + Op0 = TmpReg; + Op0IsKill = true; } - - return FastEmitInst_rii(Opc, TLI.getRegClassFor(RetVT), Op0, Op0IsKill, Shift, - ImmS); + return fastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS); } -unsigned AArch64FastISel::Emit_ASR_ri(MVT RetVT, unsigned Op0, bool Op0IsKill, - uint64_t Shift) { - unsigned Opc, ImmS; +unsigned AArch64FastISel::emitASR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill, + unsigned Op1Reg, bool Op1IsKill) { + unsigned Opc = 0; + bool NeedTrunc = false; + uint64_t Mask = 0; switch (RetVT.SimpleTy) { default: return 0; - case MVT::i8: - case MVT::i16: - case MVT::i32: - RetVT = MVT::i32; - Opc = AArch64::SBFMWri; ImmS = 31; break; - case MVT::i64: - Opc = AArch64::SBFMXri; ImmS = 63; break; + case MVT::i8: Opc = AArch64::ASRVWr; NeedTrunc = true; Mask = 0xff; break; + case MVT::i16: Opc = AArch64::ASRVWr; NeedTrunc = true; Mask = 0xffff; break; + case MVT::i32: Opc = AArch64::ASRVWr; break; + case MVT::i64: Opc = AArch64::ASRVXr; break; + } + + const TargetRegisterClass *RC = + (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass; + if (NeedTrunc) { + Op0Reg = emitIntExt(RetVT, Op0Reg, MVT::i32, /*IsZExt=*/false); + Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Op1IsKill, Mask); + Op0IsKill = Op1IsKill = true; } + unsigned ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg, + Op1IsKill); + if (NeedTrunc) + ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask); + return ResultReg; +} + +unsigned AArch64FastISel::emitASR_ri(MVT RetVT, MVT SrcVT, unsigned Op0, + bool Op0IsKill, uint64_t Shift, + bool IsZExt) { + assert(RetVT.SimpleTy >= SrcVT.SimpleTy && + "Unexpected source/return type pair."); + assert((SrcVT == MVT::i8 || SrcVT == MVT::i16 || SrcVT == MVT::i32 || + SrcVT == MVT::i64) && "Unexpected source value type."); + assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 || + RetVT == MVT::i64) && "Unexpected return value type."); + + bool Is64Bit = (RetVT == MVT::i64); + unsigned RegSize = Is64Bit ? 64 : 32; + unsigned DstBits = RetVT.getSizeInBits(); + unsigned SrcBits = SrcVT.getSizeInBits(); + + // Don't deal with undefined shifts. + if (Shift >= DstBits) + return 0; - return FastEmitInst_rii(Opc, TLI.getRegClassFor(RetVT), Op0, Op0IsKill, Shift, - ImmS); + // For immediate shifts we can fold the zero-/sign-extension into the shift. + // {S|U}BFM Wd, Wn, #r, #s + // Wd = Wn when r <= s + + // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16 + // %2 = ashr i16 %1, 4 + // Wd<7-4:0> = Wn<7:4> + // 0b1111_1111_1111_1111__1111_1111_1111_1010 sext + // 0b0000_0000_0000_0000__0000_0000_0000_0101 sext | zext + // 0b0000_0000_0000_0000__0000_0000_0000_1010 zext + + // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16 + // %2 = ashr i16 %1, 8 + // Wd<7-7,0> = Wn<7:7> + // 0b1111_1111_1111_1111__1111_1111_1111_1111 sext + // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext + // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext + + // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16 + // %2 = ashr i16 %1, 12 + // Wd<7-7,0> = Wn<7:7> <- clamp r to 7 + // 0b1111_1111_1111_1111__1111_1111_1111_1111 sext + // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext + // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext + + if (Shift >= SrcBits && IsZExt) + return materializeInt(ConstantInt::get(*Context, APInt(RegSize, 0)), RetVT); + + unsigned ImmR = std::min(SrcBits - 1, Shift); + unsigned ImmS = SrcBits - 1; + static const unsigned OpcTable[2][2] = { + {AArch64::SBFMWri, AArch64::SBFMXri}, + {AArch64::UBFMWri, AArch64::UBFMXri} + }; + unsigned Opc = OpcTable[IsZExt][Is64Bit]; + const TargetRegisterClass *RC = + Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass; + if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) { + unsigned TmpReg = MRI.createVirtualRegister(RC); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, + TII.get(AArch64::SUBREG_TO_REG), TmpReg) + .addImm(0) + .addReg(Op0, getKillRegState(Op0IsKill)) + .addImm(AArch64::sub_32); + Op0 = TmpReg; + Op0IsKill = true; + } + return fastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS); } -unsigned AArch64FastISel::EmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, - bool isZExt) { +unsigned AArch64FastISel::emitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, + bool IsZExt) { assert(DestVT != MVT::i1 && "ZeroExt/SignExt an i1?"); // FastISel does not have plumbing to deal with extensions where the SrcVT or @@ -2114,24 +4018,24 @@ unsigned AArch64FastISel::EmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, default: return 0; case MVT::i1: - return Emiti1Ext(SrcReg, DestVT, isZExt); + return emiti1Ext(SrcReg, DestVT, IsZExt); case MVT::i8: if (DestVT == MVT::i64) - Opc = isZExt ? AArch64::UBFMXri : AArch64::SBFMXri; + Opc = IsZExt ? AArch64::UBFMXri : AArch64::SBFMXri; else - Opc = isZExt ? AArch64::UBFMWri : AArch64::SBFMWri; + Opc = IsZExt ? AArch64::UBFMWri : AArch64::SBFMWri; Imm = 7; break; case MVT::i16: if (DestVT == MVT::i64) - Opc = isZExt ? AArch64::UBFMXri : AArch64::SBFMXri; + Opc = IsZExt ? AArch64::UBFMXri : AArch64::SBFMXri; else - Opc = isZExt ? AArch64::UBFMWri : AArch64::SBFMWri; + Opc = IsZExt ? AArch64::UBFMWri : AArch64::SBFMWri; Imm = 15; break; case MVT::i32: assert(DestVT == MVT::i64 && "IntExt i32 to i32?!?"); - Opc = isZExt ? AArch64::UBFMXri : AArch64::SBFMXri; + Opc = IsZExt ? AArch64::UBFMXri : AArch64::SBFMXri; Imm = 31; break; } @@ -2149,45 +4053,159 @@ unsigned AArch64FastISel::EmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, SrcReg = Src64; } - unsigned ResultReg = createResultReg(TLI.getRegClassFor(DestVT)); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg) - .addReg(SrcReg) - .addImm(0) - .addImm(Imm); + const TargetRegisterClass *RC = + (DestVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass; + return fastEmitInst_rii(Opc, RC, SrcReg, /*TODO:IsKill=*/false, 0, Imm); +} - return ResultReg; +static bool isZExtLoad(const MachineInstr *LI) { + switch (LI->getOpcode()) { + default: + return false; + case AArch64::LDURBBi: + case AArch64::LDURHHi: + case AArch64::LDURWi: + case AArch64::LDRBBui: + case AArch64::LDRHHui: + case AArch64::LDRWui: + case AArch64::LDRBBroX: + case AArch64::LDRHHroX: + case AArch64::LDRWroX: + case AArch64::LDRBBroW: + case AArch64::LDRHHroW: + case AArch64::LDRWroW: + return true; + } } -bool AArch64FastISel::SelectIntExt(const Instruction *I) { - // On ARM, in general, integer casts don't involve legal types; this code - // handles promotable integers. The high bits for a type smaller than - // the register size are assumed to be undefined. - Type *DestTy = I->getType(); - Value *Src = I->getOperand(0); - Type *SrcTy = Src->getType(); +static bool isSExtLoad(const MachineInstr *LI) { + switch (LI->getOpcode()) { + default: + return false; + case AArch64::LDURSBWi: + case AArch64::LDURSHWi: + case AArch64::LDURSBXi: + case AArch64::LDURSHXi: + case AArch64::LDURSWi: + case AArch64::LDRSBWui: + case AArch64::LDRSHWui: + case AArch64::LDRSBXui: + case AArch64::LDRSHXui: + case AArch64::LDRSWui: + case AArch64::LDRSBWroX: + case AArch64::LDRSHWroX: + case AArch64::LDRSBXroX: + case AArch64::LDRSHXroX: + case AArch64::LDRSWroX: + case AArch64::LDRSBWroW: + case AArch64::LDRSHWroW: + case AArch64::LDRSBXroW: + case AArch64::LDRSHXroW: + case AArch64::LDRSWroW: + return true; + } +} - bool isZExt = isa(I); - unsigned SrcReg = getRegForValue(Src); - if (!SrcReg) +bool AArch64FastISel::optimizeIntExtLoad(const Instruction *I, MVT RetVT, + MVT SrcVT) { + const auto *LI = dyn_cast(I->getOperand(0)); + if (!LI || !LI->hasOneUse()) return false; - EVT SrcEVT = TLI.getValueType(SrcTy, true); - EVT DestEVT = TLI.getValueType(DestTy, true); - if (!SrcEVT.isSimple()) + // Check if the load instruction has already been selected. + unsigned Reg = lookUpRegForValue(LI); + if (!Reg) return false; - if (!DestEVT.isSimple()) + + MachineInstr *MI = MRI.getUniqueVRegDef(Reg); + if (!MI) return false; - MVT SrcVT = SrcEVT.getSimpleVT(); - MVT DestVT = DestEVT.getSimpleVT(); - unsigned ResultReg = EmitIntExt(SrcVT, SrcReg, DestVT, isZExt); - if (ResultReg == 0) + // Check if the correct load instruction has been emitted - SelectionDAG might + // have emitted a zero-extending load, but we need a sign-extending load. + bool IsZExt = isa(I); + const auto *LoadMI = MI; + if (LoadMI->getOpcode() == TargetOpcode::COPY && + LoadMI->getOperand(1).getSubReg() == AArch64::sub_32) { + unsigned LoadReg = MI->getOperand(1).getReg(); + LoadMI = MRI.getUniqueVRegDef(LoadReg); + assert(LoadMI && "Expected valid instruction"); + } + if (!(IsZExt && isZExtLoad(LoadMI)) && !(!IsZExt && isSExtLoad(LoadMI))) + return false; + + // Nothing to be done. + if (RetVT != MVT::i64 || SrcVT > MVT::i32) { + updateValueMap(I, Reg); + return true; + } + + if (IsZExt) { + unsigned Reg64 = createResultReg(&AArch64::GPR64RegClass); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, + TII.get(AArch64::SUBREG_TO_REG), Reg64) + .addImm(0) + .addReg(Reg, getKillRegState(true)) + .addImm(AArch64::sub_32); + Reg = Reg64; + } else { + assert((MI->getOpcode() == TargetOpcode::COPY && + MI->getOperand(1).getSubReg() == AArch64::sub_32) && + "Expected copy instruction"); + Reg = MI->getOperand(1).getReg(); + MI->eraseFromParent(); + } + updateValueMap(I, Reg); + return true; +} + +bool AArch64FastISel::selectIntExt(const Instruction *I) { + assert((isa(I) || isa(I)) && + "Unexpected integer extend instruction."); + MVT RetVT; + MVT SrcVT; + if (!isTypeSupported(I->getType(), RetVT)) + return false; + + if (!isTypeSupported(I->getOperand(0)->getType(), SrcVT)) + return false; + + // Try to optimize already sign-/zero-extended values from load instructions. + if (optimizeIntExtLoad(I, RetVT, SrcVT)) + return true; + + unsigned SrcReg = getRegForValue(I->getOperand(0)); + if (!SrcReg) + return false; + bool SrcIsKill = hasTrivialKill(I->getOperand(0)); + + // Try to optimize already sign-/zero-extended values from function arguments. + bool IsZExt = isa(I); + if (const auto *Arg = dyn_cast(I->getOperand(0))) { + if ((IsZExt && Arg->hasZExtAttr()) || (!IsZExt && Arg->hasSExtAttr())) { + if (RetVT == MVT::i64 && SrcVT != MVT::i64) { + unsigned ResultReg = createResultReg(&AArch64::GPR64RegClass); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, + TII.get(AArch64::SUBREG_TO_REG), ResultReg) + .addImm(0) + .addReg(SrcReg, getKillRegState(SrcIsKill)) + .addImm(AArch64::sub_32); + SrcReg = ResultReg; + } + updateValueMap(I, SrcReg); + return true; + } + } + + unsigned ResultReg = emitIntExt(SrcVT, SrcReg, RetVT, IsZExt); + if (!ResultReg) return false; - UpdateValueMap(I, ResultReg); + + updateValueMap(I, ResultReg); return true; } -bool AArch64FastISel::SelectRem(const Instruction *I, unsigned ISDOpcode) { +bool AArch64FastISel::selectRem(const Instruction *I, unsigned ISDOpcode) { EVT DestEVT = TLI.getValueType(I->getType(), true); if (!DestEVT.isSimple()) return false; @@ -2197,51 +4215,95 @@ bool AArch64FastISel::SelectRem(const Instruction *I, unsigned ISDOpcode) { return false; unsigned DivOpc; - bool is64bit = (DestVT == MVT::i64); + bool Is64bit = (DestVT == MVT::i64); switch (ISDOpcode) { default: return false; case ISD::SREM: - DivOpc = is64bit ? AArch64::SDIVXr : AArch64::SDIVWr; + DivOpc = Is64bit ? AArch64::SDIVXr : AArch64::SDIVWr; break; case ISD::UREM: - DivOpc = is64bit ? AArch64::UDIVXr : AArch64::UDIVWr; + DivOpc = Is64bit ? AArch64::UDIVXr : AArch64::UDIVWr; break; } - unsigned MSubOpc = is64bit ? AArch64::MSUBXrrr : AArch64::MSUBWrrr; + unsigned MSubOpc = Is64bit ? AArch64::MSUBXrrr : AArch64::MSUBWrrr; unsigned Src0Reg = getRegForValue(I->getOperand(0)); if (!Src0Reg) return false; + bool Src0IsKill = hasTrivialKill(I->getOperand(0)); unsigned Src1Reg = getRegForValue(I->getOperand(1)); if (!Src1Reg) return false; + bool Src1IsKill = hasTrivialKill(I->getOperand(1)); - unsigned QuotReg = createResultReg(TLI.getRegClassFor(DestVT)); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(DivOpc), QuotReg) - .addReg(Src0Reg) - .addReg(Src1Reg); + const TargetRegisterClass *RC = + (DestVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass; + unsigned QuotReg = fastEmitInst_rr(DivOpc, RC, Src0Reg, /*IsKill=*/false, + Src1Reg, /*IsKill=*/false); + assert(QuotReg && "Unexpected DIV instruction emission failure."); // The remainder is computed as numerator - (quotient * denominator) using the // MSUB instruction. - unsigned ResultReg = createResultReg(TLI.getRegClassFor(DestVT)); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(MSubOpc), ResultReg) - .addReg(QuotReg) - .addReg(Src1Reg) - .addReg(Src0Reg); - UpdateValueMap(I, ResultReg); + unsigned ResultReg = fastEmitInst_rrr(MSubOpc, RC, QuotReg, /*IsKill=*/true, + Src1Reg, Src1IsKill, Src0Reg, + Src0IsKill); + updateValueMap(I, ResultReg); return true; } -bool AArch64FastISel::SelectMul(const Instruction *I) { - EVT SrcEVT = TLI.getValueType(I->getOperand(0)->getType(), true); - if (!SrcEVT.isSimple()) +bool AArch64FastISel::selectMul(const Instruction *I) { + MVT VT; + if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true)) return false; - MVT SrcVT = SrcEVT.getSimpleVT(); - // Must be simple value type. Don't handle vectors. - if (SrcVT != MVT::i64 && SrcVT != MVT::i32 && SrcVT != MVT::i16 && - SrcVT != MVT::i8) - return false; + if (VT.isVector()) + return selectBinaryOp(I, ISD::MUL); + + const Value *Src0 = I->getOperand(0); + const Value *Src1 = I->getOperand(1); + if (const auto *C = dyn_cast(Src0)) + if (C->getValue().isPowerOf2()) + std::swap(Src0, Src1); + + // Try to simplify to a shift instruction. + if (const auto *C = dyn_cast(Src1)) + if (C->getValue().isPowerOf2()) { + uint64_t ShiftVal = C->getValue().logBase2(); + MVT SrcVT = VT; + bool IsZExt = true; + if (const auto *ZExt = dyn_cast(Src0)) { + if (!isIntExtFree(ZExt)) { + MVT VT; + if (isValueAvailable(ZExt) && isTypeSupported(ZExt->getSrcTy(), VT)) { + SrcVT = VT; + IsZExt = true; + Src0 = ZExt->getOperand(0); + } + } + } else if (const auto *SExt = dyn_cast(Src0)) { + if (!isIntExtFree(SExt)) { + MVT VT; + if (isValueAvailable(SExt) && isTypeSupported(SExt->getSrcTy(), VT)) { + SrcVT = VT; + IsZExt = false; + Src0 = SExt->getOperand(0); + } + } + } + + unsigned Src0Reg = getRegForValue(Src0); + if (!Src0Reg) + return false; + bool Src0IsKill = hasTrivialKill(Src0); + + unsigned ResultReg = + emitLSL_ri(VT, SrcVT, Src0Reg, Src0IsKill, ShiftVal, IsZExt); + + if (ResultReg) { + updateValueMap(I, ResultReg); + return true; + } + } unsigned Src0Reg = getRegForValue(I->getOperand(0)); if (!Src0Reg) @@ -2253,110 +4315,407 @@ bool AArch64FastISel::SelectMul(const Instruction *I) { return false; bool Src1IsKill = hasTrivialKill(I->getOperand(1)); - unsigned ResultReg = - Emit_MUL_rr(SrcVT, Src0Reg, Src0IsKill, Src1Reg, Src1IsKill); + unsigned ResultReg = emitMul_rr(VT, Src0Reg, Src0IsKill, Src1Reg, Src1IsKill); if (!ResultReg) return false; - UpdateValueMap(I, ResultReg); + updateValueMap(I, ResultReg); return true; } -bool AArch64FastISel::SelectShift(const Instruction *I, bool IsLeftShift, - bool IsArithmetic) { - EVT RetEVT = TLI.getValueType(I->getType(), true); - if (!RetEVT.isSimple()) +bool AArch64FastISel::selectShift(const Instruction *I) { + MVT RetVT; + if (!isTypeSupported(I->getType(), RetVT, /*IsVectorAllowed=*/true)) return false; - MVT RetVT = RetEVT.getSimpleVT(); - if (!isa(I->getOperand(1))) + if (RetVT.isVector()) + return selectOperator(I, I->getOpcode()); + + if (const auto *C = dyn_cast(I->getOperand(1))) { + unsigned ResultReg = 0; + uint64_t ShiftVal = C->getZExtValue(); + MVT SrcVT = RetVT; + bool IsZExt = (I->getOpcode() == Instruction::AShr) ? false : true; + const Value *Op0 = I->getOperand(0); + if (const auto *ZExt = dyn_cast(Op0)) { + if (!isIntExtFree(ZExt)) { + MVT TmpVT; + if (isValueAvailable(ZExt) && isTypeSupported(ZExt->getSrcTy(), TmpVT)) { + SrcVT = TmpVT; + IsZExt = true; + Op0 = ZExt->getOperand(0); + } + } + } else if (const auto *SExt = dyn_cast(Op0)) { + if (!isIntExtFree(SExt)) { + MVT TmpVT; + if (isValueAvailable(SExt) && isTypeSupported(SExt->getSrcTy(), TmpVT)) { + SrcVT = TmpVT; + IsZExt = false; + Op0 = SExt->getOperand(0); + } + } + } + + unsigned Op0Reg = getRegForValue(Op0); + if (!Op0Reg) + return false; + bool Op0IsKill = hasTrivialKill(Op0); + + switch (I->getOpcode()) { + default: llvm_unreachable("Unexpected instruction."); + case Instruction::Shl: + ResultReg = emitLSL_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt); + break; + case Instruction::AShr: + ResultReg = emitASR_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt); + break; + case Instruction::LShr: + ResultReg = emitLSR_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt); + break; + } + if (!ResultReg) + return false; + + updateValueMap(I, ResultReg); + return true; + } + + unsigned Op0Reg = getRegForValue(I->getOperand(0)); + if (!Op0Reg) + return false; + bool Op0IsKill = hasTrivialKill(I->getOperand(0)); + + unsigned Op1Reg = getRegForValue(I->getOperand(1)); + if (!Op1Reg) + return false; + bool Op1IsKill = hasTrivialKill(I->getOperand(1)); + + unsigned ResultReg = 0; + switch (I->getOpcode()) { + default: llvm_unreachable("Unexpected instruction."); + case Instruction::Shl: + ResultReg = emitLSL_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill); + break; + case Instruction::AShr: + ResultReg = emitASR_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill); + break; + case Instruction::LShr: + ResultReg = emitLSR_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill); + break; + } + + if (!ResultReg) return false; + updateValueMap(I, ResultReg); + return true; +} + +bool AArch64FastISel::selectBitCast(const Instruction *I) { + MVT RetVT, SrcVT; + + if (!isTypeLegal(I->getOperand(0)->getType(), SrcVT)) + return false; + if (!isTypeLegal(I->getType(), RetVT)) + return false; + + unsigned Opc; + if (RetVT == MVT::f32 && SrcVT == MVT::i32) + Opc = AArch64::FMOVWSr; + else if (RetVT == MVT::f64 && SrcVT == MVT::i64) + Opc = AArch64::FMOVXDr; + else if (RetVT == MVT::i32 && SrcVT == MVT::f32) + Opc = AArch64::FMOVSWr; + else if (RetVT == MVT::i64 && SrcVT == MVT::f64) + Opc = AArch64::FMOVDXr; + else + return false; + + const TargetRegisterClass *RC = nullptr; + switch (RetVT.SimpleTy) { + default: llvm_unreachable("Unexpected value type."); + case MVT::i32: RC = &AArch64::GPR32RegClass; break; + case MVT::i64: RC = &AArch64::GPR64RegClass; break; + case MVT::f32: RC = &AArch64::FPR32RegClass; break; + case MVT::f64: RC = &AArch64::FPR64RegClass; break; + } unsigned Op0Reg = getRegForValue(I->getOperand(0)); if (!Op0Reg) return false; bool Op0IsKill = hasTrivialKill(I->getOperand(0)); + unsigned ResultReg = fastEmitInst_r(Opc, RC, Op0Reg, Op0IsKill); - uint64_t ShiftVal = cast(I->getOperand(1))->getZExtValue(); + if (!ResultReg) + return false; - unsigned ResultReg; - if (IsLeftShift) - ResultReg = Emit_LSL_ri(RetVT, Op0Reg, Op0IsKill, ShiftVal); - else { - if (IsArithmetic) - ResultReg = Emit_ASR_ri(RetVT, Op0Reg, Op0IsKill, ShiftVal); - else - ResultReg = Emit_LSR_ri(RetVT, Op0Reg, Op0IsKill, ShiftVal); + updateValueMap(I, ResultReg); + return true; +} + +bool AArch64FastISel::selectFRem(const Instruction *I) { + MVT RetVT; + if (!isTypeLegal(I->getType(), RetVT)) + return false; + + RTLIB::Libcall LC; + switch (RetVT.SimpleTy) { + default: + return false; + case MVT::f32: + LC = RTLIB::REM_F32; + break; + case MVT::f64: + LC = RTLIB::REM_F64; + break; + } + + ArgListTy Args; + Args.reserve(I->getNumOperands()); + + // Populate the argument list. + for (auto &Arg : I->operands()) { + ArgListEntry Entry; + Entry.Val = Arg; + Entry.Ty = Arg->getType(); + Args.push_back(Entry); + } + + CallLoweringInfo CLI; + CLI.setCallee(TLI.getLibcallCallingConv(LC), I->getType(), + TLI.getLibcallName(LC), std::move(Args)); + if (!lowerCallTo(CLI)) + return false; + updateValueMap(I, CLI.ResultReg); + return true; +} + +bool AArch64FastISel::selectSDiv(const Instruction *I) { + MVT VT; + if (!isTypeLegal(I->getType(), VT)) + return false; + + if (!isa(I->getOperand(1))) + return selectBinaryOp(I, ISD::SDIV); + + const APInt &C = cast(I->getOperand(1))->getValue(); + if ((VT != MVT::i32 && VT != MVT::i64) || !C || + !(C.isPowerOf2() || (-C).isPowerOf2())) + return selectBinaryOp(I, ISD::SDIV); + + unsigned Lg2 = C.countTrailingZeros(); + unsigned Src0Reg = getRegForValue(I->getOperand(0)); + if (!Src0Reg) + return false; + bool Src0IsKill = hasTrivialKill(I->getOperand(0)); + + if (cast(I)->isExact()) { + unsigned ResultReg = emitASR_ri(VT, VT, Src0Reg, Src0IsKill, Lg2); + if (!ResultReg) + return false; + updateValueMap(I, ResultReg); + return true; + } + + int64_t Pow2MinusOne = (1ULL << Lg2) - 1; + unsigned AddReg = emitAdd_ri_(VT, Src0Reg, /*IsKill=*/false, Pow2MinusOne); + if (!AddReg) + return false; + + // (Src0 < 0) ? Pow2 - 1 : 0; + if (!emitICmp_ri(VT, Src0Reg, /*IsKill=*/false, 0)) + return false; + + unsigned SelectOpc; + const TargetRegisterClass *RC; + if (VT == MVT::i64) { + SelectOpc = AArch64::CSELXr; + RC = &AArch64::GPR64RegClass; + } else { + SelectOpc = AArch64::CSELWr; + RC = &AArch64::GPR32RegClass; } + unsigned SelectReg = + fastEmitInst_rri(SelectOpc, RC, AddReg, /*IsKill=*/true, Src0Reg, + Src0IsKill, AArch64CC::LT); + if (!SelectReg) + return false; + + // Divide by Pow2 --> ashr. If we're dividing by a negative value we must also + // negate the result. + unsigned ZeroReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR; + unsigned ResultReg; + if (C.isNegative()) + ResultReg = emitAddSub_rs(/*UseAdd=*/false, VT, ZeroReg, /*IsKill=*/true, + SelectReg, /*IsKill=*/true, AArch64_AM::ASR, Lg2); + else + ResultReg = emitASR_ri(VT, VT, SelectReg, /*IsKill=*/true, Lg2); if (!ResultReg) return false; - UpdateValueMap(I, ResultReg); + updateValueMap(I, ResultReg); + return true; +} + +/// This is mostly a copy of the existing FastISel GEP code, but we have to +/// duplicate it for AArch64, because otherwise we would bail out even for +/// simple cases. This is because the standard fastEmit functions don't cover +/// MUL at all and ADD is lowered very inefficientily. +bool AArch64FastISel::selectGetElementPtr(const Instruction *I) { + unsigned N = getRegForValue(I->getOperand(0)); + if (!N) + return false; + bool NIsKill = hasTrivialKill(I->getOperand(0)); + + // Keep a running tab of the total offset to coalesce multiple N = N + Offset + // into a single N = N + TotalOffset. + uint64_t TotalOffs = 0; + Type *Ty = I->getOperand(0)->getType(); + MVT VT = TLI.getPointerTy(); + for (auto OI = std::next(I->op_begin()), E = I->op_end(); OI != E; ++OI) { + const Value *Idx = *OI; + if (auto *StTy = dyn_cast(Ty)) { + unsigned Field = cast(Idx)->getZExtValue(); + // N = N + Offset + if (Field) + TotalOffs += DL.getStructLayout(StTy)->getElementOffset(Field); + Ty = StTy->getElementType(Field); + } else { + Ty = cast(Ty)->getElementType(); + // If this is a constant subscript, handle it quickly. + if (const auto *CI = dyn_cast(Idx)) { + if (CI->isZero()) + continue; + // N = N + Offset + TotalOffs += + DL.getTypeAllocSize(Ty) * cast(CI)->getSExtValue(); + continue; + } + if (TotalOffs) { + N = emitAdd_ri_(VT, N, NIsKill, TotalOffs); + if (!N) + return false; + NIsKill = true; + TotalOffs = 0; + } + + // N = N + Idx * ElementSize; + uint64_t ElementSize = DL.getTypeAllocSize(Ty); + std::pair Pair = getRegForGEPIndex(Idx); + unsigned IdxN = Pair.first; + bool IdxNIsKill = Pair.second; + if (!IdxN) + return false; + + if (ElementSize != 1) { + unsigned C = fastEmit_i(VT, VT, ISD::Constant, ElementSize); + if (!C) + return false; + IdxN = emitMul_rr(VT, IdxN, IdxNIsKill, C, true); + if (!IdxN) + return false; + IdxNIsKill = true; + } + N = fastEmit_rr(VT, VT, ISD::ADD, N, NIsKill, IdxN, IdxNIsKill); + if (!N) + return false; + } + } + if (TotalOffs) { + N = emitAdd_ri_(VT, N, NIsKill, TotalOffs); + if (!N) + return false; + } + updateValueMap(I, N); return true; } -bool AArch64FastISel::TargetSelectInstruction(const Instruction *I) { +bool AArch64FastISel::fastSelectInstruction(const Instruction *I) { switch (I->getOpcode()) { default: break; - case Instruction::Load: - return SelectLoad(I); - case Instruction::Store: - return SelectStore(I); + case Instruction::Add: + case Instruction::Sub: + return selectAddSub(I); + case Instruction::Mul: + return selectMul(I); + case Instruction::SDiv: + return selectSDiv(I); + case Instruction::SRem: + if (!selectBinaryOp(I, ISD::SREM)) + return selectRem(I, ISD::SREM); + return true; + case Instruction::URem: + if (!selectBinaryOp(I, ISD::UREM)) + return selectRem(I, ISD::UREM); + return true; + case Instruction::Shl: + case Instruction::LShr: + case Instruction::AShr: + return selectShift(I); + case Instruction::And: + case Instruction::Or: + case Instruction::Xor: + return selectLogicalOp(I); case Instruction::Br: - return SelectBranch(I); + return selectBranch(I); case Instruction::IndirectBr: - return SelectIndirectBr(I); - case Instruction::FCmp: - case Instruction::ICmp: - return SelectCmp(I); - case Instruction::Select: - return SelectSelect(I); - case Instruction::FPExt: - return SelectFPExt(I); - case Instruction::FPTrunc: - return SelectFPTrunc(I); + return selectIndirectBr(I); + case Instruction::BitCast: + if (!FastISel::selectBitCast(I)) + return selectBitCast(I); + return true; case Instruction::FPToSI: - return SelectFPToInt(I, /*Signed=*/true); + if (!selectCast(I, ISD::FP_TO_SINT)) + return selectFPToInt(I, /*Signed=*/true); + return true; case Instruction::FPToUI: - return SelectFPToInt(I, /*Signed=*/false); + return selectFPToInt(I, /*Signed=*/false); + case Instruction::ZExt: + case Instruction::SExt: + return selectIntExt(I); + case Instruction::Trunc: + if (!selectCast(I, ISD::TRUNCATE)) + return selectTrunc(I); + return true; + case Instruction::FPExt: + return selectFPExt(I); + case Instruction::FPTrunc: + return selectFPTrunc(I); case Instruction::SIToFP: - return SelectIntToFP(I, /*Signed=*/true); + if (!selectCast(I, ISD::SINT_TO_FP)) + return selectIntToFP(I, /*Signed=*/true); + return true; case Instruction::UIToFP: - return SelectIntToFP(I, /*Signed=*/false); - case Instruction::SRem: - return SelectRem(I, ISD::SREM); - case Instruction::URem: - return SelectRem(I, ISD::UREM); + return selectIntToFP(I, /*Signed=*/false); + case Instruction::Load: + return selectLoad(I); + case Instruction::Store: + return selectStore(I); + case Instruction::FCmp: + case Instruction::ICmp: + return selectCmp(I); + case Instruction::Select: + return selectSelect(I); case Instruction::Ret: - return SelectRet(I); - case Instruction::Trunc: - return SelectTrunc(I); - case Instruction::ZExt: - case Instruction::SExt: - return SelectIntExt(I); - - // FIXME: All of these should really be handled by the target-independent - // selector -> improve FastISel tblgen. - case Instruction::Mul: - return SelectMul(I); - case Instruction::Shl: - return SelectShift(I, /*IsLeftShift=*/true, /*IsArithmetic=*/false); - case Instruction::LShr: - return SelectShift(I, /*IsLeftShift=*/false, /*IsArithmetic=*/false); - case Instruction::AShr: - return SelectShift(I, /*IsLeftShift=*/false, /*IsArithmetic=*/true); + return selectRet(I); + case Instruction::FRem: + return selectFRem(I); + case Instruction::GetElementPtr: + return selectGetElementPtr(I); } - return false; + + // fall-back to target-independent instruction selection. + return selectOperator(I, I->getOpcode()); // Silence warnings. (void)&CC_AArch64_DarwinPCS_VarArg; } namespace llvm { -llvm::FastISel *AArch64::createFastISel(FunctionLoweringInfo &funcInfo, - const TargetLibraryInfo *libInfo) { - return new AArch64FastISel(funcInfo, libInfo); +llvm::FastISel *AArch64::createFastISel(FunctionLoweringInfo &FuncInfo, + const TargetLibraryInfo *LibInfo) { + return new AArch64FastISel(FuncInfo, LibInfo); } }