From a98cbc554ca2cd40426e7c3ff8d1467da32e195d Mon Sep 17 00:00:00 2001 From: Anton Korobeynikov Date: Sat, 27 Jun 2009 12:16:40 +0000 Subject: [PATCH] ARM refactoring. Step 2: split RegisterInfo git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@74384 91177308-0d34-0410-b5e6-96231b3b80d8 --- lib/Target/ARM/ARMInstrInfo.cpp | 41 +- lib/Target/ARM/ARMInstrInfo.h | 27 +- lib/Target/ARM/ARMRegisterInfo.cpp | 763 +++++---------------------- lib/Target/ARM/ARMRegisterInfo.h | 72 +-- lib/Target/ARM/ARMTargetMachine.h | 2 +- lib/Target/ARM/ThumbInstrInfo.cpp | 34 +- lib/Target/ARM/ThumbInstrInfo.h | 12 +- lib/Target/ARM/ThumbRegisterInfo.cpp | 757 ++++++++++++++++++++++++++ lib/Target/ARM/ThumbRegisterInfo.h | 65 +++ 9 files changed, 1077 insertions(+), 696 deletions(-) create mode 100644 lib/Target/ARM/ThumbRegisterInfo.cpp create mode 100644 lib/Target/ARM/ThumbRegisterInfo.h diff --git a/lib/Target/ARM/ARMInstrInfo.cpp b/lib/Target/ARM/ARMInstrInfo.cpp index d95089dd484..034b697ab58 100644 --- a/lib/Target/ARM/ARMInstrInfo.cpp +++ b/lib/Target/ARM/ARMInstrInfo.cpp @@ -40,12 +40,11 @@ const MachineInstrBuilder &AddDefaultCC(const MachineInstrBuilder &MIB) { } ARMBaseInstrInfo::ARMBaseInstrInfo(const ARMSubtarget &STI) - : TargetInstrInfoImpl(ARMInsts, array_lengthof(ARMInsts)), - RI(*this, STI) { + : TargetInstrInfoImpl(ARMInsts, array_lengthof(ARMInsts)) { } ARMInstrInfo::ARMInstrInfo(const ARMSubtarget &STI) - : ARMBaseInstrInfo(STI) { + : ARMBaseInstrInfo(STI), RI(*this, STI) { } /// Return true if the instruction is a register to register move and @@ -133,15 +132,15 @@ unsigned ARMInstrInfo::isStoreToStackSlot(const MachineInstr *MI, return 0; } -void ARMBaseInstrInfo::reMaterialize(MachineBasicBlock &MBB, - MachineBasicBlock::iterator I, - unsigned DestReg, - const MachineInstr *Orig) const { +void ARMInstrInfo::reMaterialize(MachineBasicBlock &MBB, + MachineBasicBlock::iterator I, + unsigned DestReg, + const MachineInstr *Orig) const { DebugLoc dl = Orig->getDebugLoc(); if (Orig->getOpcode() == ARM::MOVi2pieces) { RI.emitLoadConstPool(MBB, I, DestReg, Orig->getOperand(1).getImm(), Orig->getOperand(2).getImm(), - Orig->getOperand(3).getReg(), this, false, dl); + Orig->getOperand(3).getReg(), this, dl); return; } @@ -660,35 +659,17 @@ foldMemoryOperandImpl(MachineFunction &MF, MachineInstr *MI, return NewMI; } -bool ARMBaseInstrInfo:: -canFoldMemoryOperand(const MachineInstr *MI, - const SmallVectorImpl &Ops) const { +bool +ARMInstrInfo::canFoldMemoryOperand(const MachineInstr *MI, + const SmallVectorImpl &Ops) const { if (Ops.size() != 1) return false; - unsigned OpNum = Ops[0]; unsigned Opc = MI->getOpcode(); switch (Opc) { default: break; case ARM::MOVr: // If it is updating CPSR, then it cannot be folded. return MI->getOperand(4).getReg() != ARM::CPSR; - case ARM::tMOVr: - case ARM::tMOVlor2hir: - case ARM::tMOVhir2lor: - case ARM::tMOVhir2hir: { - if (OpNum == 0) { // move -> store - unsigned SrcReg = MI->getOperand(1).getReg(); - if (RI.isPhysicalRegister(SrcReg) && !RI.isLowRegister(SrcReg)) - // tSpill cannot take a high register operand. - return false; - } else { // move -> load - unsigned DstReg = MI->getOperand(0).getReg(); - if (RI.isPhysicalRegister(DstReg) && !RI.isLowRegister(DstReg)) - // tRestore cannot target a high register operand. - return false; - } - return true; - } case ARM::FCPYS: case ARM::FCPYD: return true; @@ -702,7 +683,7 @@ canFoldMemoryOperand(const MachineInstr *MI, } bool - ARMBaseInstrInfo::BlockHasNoFallThrough(const MachineBasicBlock &MBB) const { +ARMBaseInstrInfo::BlockHasNoFallThrough(const MachineBasicBlock &MBB) const { if (MBB.empty()) return false; switch (MBB.back().getOpcode()) { diff --git a/lib/Target/ARM/ARMInstrInfo.h b/lib/Target/ARM/ARMInstrInfo.h index 131960b9c7a..7e3aa6c7a36 100644 --- a/lib/Target/ARM/ARMInstrInfo.h +++ b/lib/Target/ARM/ARMInstrInfo.h @@ -147,21 +147,10 @@ namespace ARMII { } class ARMBaseInstrInfo : public TargetInstrInfoImpl { - const ARMRegisterInfo RI; protected: // Can be only subclassed. explicit ARMBaseInstrInfo(const ARMSubtarget &STI); public: - - /// getRegisterInfo - TargetInstrInfo is a superset of MRegister info. As - /// such, whenever a client has an instance of instruction info, it should - /// always be able to get register info as well (through this method). - /// - virtual const ARMRegisterInfo &getRegisterInfo() const { return RI; } - - void reMaterialize(MachineBasicBlock &MBB, MachineBasicBlock::iterator MI, - unsigned DestReg, const MachineInstr *Orig) const; - virtual MachineInstr *convertToThreeAddress(MachineFunction::iterator &MFI, MachineBasicBlock::iterator &MBBI, LiveVariables *LV) const; @@ -176,9 +165,6 @@ public: MachineBasicBlock *FBB, const SmallVectorImpl &Cond) const; - virtual bool canFoldMemoryOperand(const MachineInstr *MI, - const SmallVectorImpl &Ops) const; - virtual bool BlockHasNoFallThrough(const MachineBasicBlock &MBB) const; virtual bool ReverseBranchCondition(SmallVectorImpl &Cond) const; @@ -209,9 +195,16 @@ public: }; class ARMInstrInfo : public ARMBaseInstrInfo { + ARMRegisterInfo RI; public: explicit ARMInstrInfo(const ARMSubtarget &STI); + /// getRegisterInfo - TargetInstrInfo is a superset of MRegister info. As + /// such, whenever a client has an instance of instruction info, it should + /// always be able to get register info as well (through this method). + /// + virtual const ARMRegisterInfo &getRegisterInfo() const { return RI; } + /// Return true if the instruction is a register to register move and return /// the source and dest operands and their sub-register indices by reference. virtual bool isMoveInstr(const MachineInstr &MI, @@ -248,6 +241,12 @@ public: const TargetRegisterClass *RC, SmallVectorImpl &NewMIs) const; + void reMaterialize(MachineBasicBlock &MBB, MachineBasicBlock::iterator MI, + unsigned DestReg, const MachineInstr *Orig) const; + + virtual bool canFoldMemoryOperand(const MachineInstr *MI, + const SmallVectorImpl &Ops) const; + virtual MachineInstr* foldMemoryOperandImpl(MachineFunction &MF, MachineInstr* MI, const SmallVectorImpl &Ops, diff --git a/lib/Target/ARM/ARMRegisterInfo.cpp b/lib/Target/ARM/ARMRegisterInfo.cpp index bb0cc8ff068..a58e8ec0e0a 100644 --- a/lib/Target/ARM/ARMRegisterInfo.cpp +++ b/lib/Target/ARM/ARMRegisterInfo.cpp @@ -31,16 +31,9 @@ #include "llvm/Target/TargetOptions.h" #include "llvm/ADT/BitVector.h" #include "llvm/ADT/SmallVector.h" -#include "llvm/ADT/STLExtras.h" -#include "llvm/Support/CommandLine.h" -#include using namespace llvm; -static cl::opt ThumbRegScavenging("enable-thumb-reg-scavenging", - cl::Hidden, - cl::desc("Enable register scavenging on Thumb")); - -unsigned ARMRegisterInfo::getRegisterNumbering(unsigned RegEnum) { +unsigned ARMBaseRegisterInfo::getRegisterNumbering(unsigned RegEnum) { using namespace ARM; switch (RegEnum) { case R0: case S0: case D0: return 0; @@ -81,8 +74,8 @@ unsigned ARMRegisterInfo::getRegisterNumbering(unsigned RegEnum) { } } -unsigned ARMRegisterInfo::getRegisterNumbering(unsigned RegEnum, - bool &isSPVFP) { +unsigned ARMBaseRegisterInfo::getRegisterNumbering(unsigned RegEnum, + bool &isSPVFP) { isSPVFP = false; using namespace ARM; @@ -108,12 +101,12 @@ unsigned ARMRegisterInfo::getRegisterNumbering(unsigned RegEnum, case PC: case D15: return 15; case S0: case S1: case S2: case S3: - case S4: case S5: case S6: case S7: - case S8: case S9: case S10: case S11: - case S12: case S13: case S14: case S15: - case S16: case S17: case S18: case S19: - case S20: case S21: case S22: case S23: - case S24: case S25: case S26: case S27: + case S4: case S5: case S6: case S7: + case S8: case S9: case S10: case S11: + case S12: case S13: case S14: case S15: + case S16: case S17: case S18: case S19: + case S20: case S21: case S22: case S23: + case S24: case S25: case S26: case S27: case S28: case S29: case S30: case S31: { isSPVFP = true; switch (RegEnum) { @@ -155,13 +148,18 @@ unsigned ARMRegisterInfo::getRegisterNumbering(unsigned RegEnum, } } -ARMRegisterInfo::ARMRegisterInfo(const TargetInstrInfo &tii, - const ARMSubtarget &sti) +ARMBaseRegisterInfo::ARMBaseRegisterInfo(const TargetInstrInfo &tii, + const ARMSubtarget &sti) : ARMGenRegisterInfo(ARM::ADJCALLSTACKDOWN, ARM::ADJCALLSTACKUP), TII(tii), STI(sti), FramePtr((STI.isTargetDarwin() || STI.isThumb()) ? ARM::R7 : ARM::R11) { } +ARMRegisterInfo::ARMRegisterInfo(const TargetInstrInfo &tii, + const ARMSubtarget &sti) + : ARMBaseRegisterInfo(tii, sti) { +} + static inline const MachineInstrBuilder &AddDefaultPred(const MachineInstrBuilder &MIB) { return MIB.addImm((int64_t)ARMCC::AL).addReg(0); @@ -179,24 +177,20 @@ void ARMRegisterInfo::emitLoadConstPool(MachineBasicBlock &MBB, unsigned DestReg, int Val, unsigned Pred, unsigned PredReg, const TargetInstrInfo *TII, - bool isThumb, DebugLoc dl) const { MachineFunction &MF = *MBB.getParent(); MachineConstantPool *ConstantPool = MF.getConstantPool(); Constant *C = ConstantInt::get(Type::Int32Ty, Val); unsigned Idx = ConstantPool->getConstantPoolIndex(C, 4); - if (isThumb) - BuildMI(MBB, MBBI, dl, - TII->get(ARM::tLDRcp),DestReg).addConstantPoolIndex(Idx); - else - BuildMI(MBB, MBBI, dl, TII->get(ARM::LDRcp), DestReg) - .addConstantPoolIndex(Idx) - .addReg(0).addImm(0).addImm(Pred).addReg(PredReg); + + BuildMI(MBB, MBBI, dl, TII->get(ARM::LDRcp), DestReg) + .addConstantPoolIndex(Idx) + .addReg(0).addImm(0).addImm(Pred).addReg(PredReg); } /// isLowRegister - Returns true if the register is low register r0-r7. /// -bool ARMRegisterInfo::isLowRegister(unsigned Reg) const { +bool ARMBaseRegisterInfo::isLowRegister(unsigned Reg) const { using namespace ARM; switch (Reg) { case R0: case R1: case R2: case R3: @@ -207,24 +201,8 @@ bool ARMRegisterInfo::isLowRegister(unsigned Reg) const { } } -const TargetRegisterClass* -ARMRegisterInfo::getPhysicalRegisterRegClass(unsigned Reg, MVT VT) const { - if (STI.isThumb()) { - if (isLowRegister(Reg)) - return ARM::tGPRRegisterClass; - switch (Reg) { - default: - break; - case ARM::R8: case ARM::R9: case ARM::R10: case ARM::R11: - case ARM::R12: case ARM::SP: case ARM::LR: case ARM::PC: - return ARM::GPRRegisterClass; - } - } - return TargetRegisterInfo::getPhysicalRegisterRegClass(Reg, VT); -} - const unsigned* -ARMRegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const { +ARMBaseRegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const { static const unsigned CalleeSavedRegs[] = { ARM::LR, ARM::R11, ARM::R10, ARM::R9, ARM::R8, ARM::R7, ARM::R6, ARM::R5, ARM::R4, @@ -248,7 +226,7 @@ ARMRegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const { } const TargetRegisterClass* const * -ARMRegisterInfo::getCalleeSavedRegClasses(const MachineFunction *MF) const { +ARMBaseRegisterInfo::getCalleeSavedRegClasses(const MachineFunction *MF) const { static const TargetRegisterClass * const CalleeSavedRegClasses[] = { &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass, &ARM::GPRRegClass, @@ -297,7 +275,7 @@ ARMRegisterInfo::getCalleeSavedRegClasses(const MachineFunction *MF) const { ? DarwinCalleeSavedRegClasses : CalleeSavedRegClasses; } -BitVector ARMRegisterInfo::getReservedRegs(const MachineFunction &MF) const { +BitVector ARMBaseRegisterInfo::getReservedRegs(const MachineFunction &MF) const { // FIXME: avoid re-calculating this everytime. BitVector Reserved(getNumRegs()); Reserved.set(ARM::SP); @@ -311,7 +289,7 @@ BitVector ARMRegisterInfo::getReservedRegs(const MachineFunction &MF) const { } bool -ARMRegisterInfo::isReservedReg(const MachineFunction &MF, unsigned Reg) const { +ARMBaseRegisterInfo::isReservedReg(const MachineFunction &MF, unsigned Reg) const { switch (Reg) { default: break; case ARM::SP: @@ -329,16 +307,16 @@ ARMRegisterInfo::isReservedReg(const MachineFunction &MF, unsigned Reg) const { return false; } -const TargetRegisterClass *ARMRegisterInfo::getPointerRegClass() const { +const TargetRegisterClass *ARMBaseRegisterInfo::getPointerRegClass() const { return &ARM::GPRRegClass; } /// getAllocationOrder - Returns the register allocation order for a specified /// register class in the form of a pair of TargetRegisterClass iterators. std::pair -ARMRegisterInfo::getAllocationOrder(const TargetRegisterClass *RC, - unsigned HintType, unsigned HintReg, - const MachineFunction &MF) const { +ARMBaseRegisterInfo::getAllocationOrder(const TargetRegisterClass *RC, + unsigned HintType, unsigned HintReg, + const MachineFunction &MF) const { // Alternative register allocation orders when favoring even / odd registers // of register pairs. @@ -479,8 +457,8 @@ ARMRegisterInfo::getAllocationOrder(const TargetRegisterClass *RC, /// ResolveRegAllocHint - Resolves the specified register allocation hint /// to a physical register. Returns the physical register if it is successful. unsigned -ARMRegisterInfo::ResolveRegAllocHint(unsigned Type, unsigned Reg, - const MachineFunction &MF) const { +ARMBaseRegisterInfo::ResolveRegAllocHint(unsigned Type, unsigned Reg, + const MachineFunction &MF) const { if (Reg == 0 || !isPhysicalRegister(Reg)) return 0; if (Type == 0) @@ -495,8 +473,8 @@ ARMRegisterInfo::ResolveRegAllocHint(unsigned Type, unsigned Reg, } void -ARMRegisterInfo::UpdateRegAllocHint(unsigned Reg, unsigned NewReg, - MachineFunction &MF) const { +ARMBaseRegisterInfo::UpdateRegAllocHint(unsigned Reg, unsigned NewReg, + MachineFunction &MF) const { MachineRegisterInfo *MRI = &MF.getRegInfo(); std::pair Hint = MRI->getRegAllocationHint(Reg); if ((Hint.first == (unsigned)ARMRI::RegPairOdd || @@ -516,15 +494,14 @@ ARMRegisterInfo::UpdateRegAllocHint(unsigned Reg, unsigned NewReg, bool ARMRegisterInfo::requiresRegisterScavenging(const MachineFunction &MF) const { - const ARMFunctionInfo *AFI = MF.getInfo(); - return ThumbRegScavenging || !AFI->isThumbFunction(); + return true; } /// hasFP - Return true if the specified function should have a dedicated frame /// pointer register. This is true if the function has variable sized allocas /// or if frame pointer elimination is disabled. /// -bool ARMRegisterInfo::hasFP(const MachineFunction &MF) const { +bool ARMBaseRegisterInfo::hasFP(const MachineFunction &MF) const { const MachineFrameInfo *MFI = MF.getFrameInfo(); return (NoFramePointerElim || MFI->hasVarSizedObjects() || @@ -539,18 +516,13 @@ bool ARMRegisterInfo::hasFP(const MachineFunction &MF) const { bool ARMRegisterInfo::hasReservedCallFrame(MachineFunction &MF) const { const MachineFrameInfo *FFI = MF.getFrameInfo(); unsigned CFSize = FFI->getMaxCallFrameSize(); - ARMFunctionInfo *AFI = MF.getInfo(); // It's not always a good idea to include the call frame as part of the // stack frame. ARM (especially Thumb) has small immediate offset to // address the stack frame. So a large call frame can cause poor codegen // and may even makes it impossible to scavenge a register. - if (AFI->isThumbFunction()) { - if (CFSize >= ((1 << 8) - 1) * 4 / 2) // Half of imm8 * 4 - return false; - } else { - if (CFSize >= ((1 << 12) - 1) / 2) // Half of imm12 - return false; - } + if (CFSize >= ((1 << 12) - 1) / 2) // Half of imm12 + return false; + return !MF.getFrameInfo()->hasVarSizedObjects(); } @@ -570,14 +542,14 @@ void emitARMRegPlusImmediate(MachineBasicBlock &MBB, unsigned RotAmt = ARM_AM::getSOImmValRotate(NumBytes); unsigned ThisVal = NumBytes & ARM_AM::rotr32(0xFF, RotAmt); assert(ThisVal && "Didn't extract field correctly"); - + // We will handle these bits from offset, clear them. NumBytes &= ~ThisVal; - + // Get the properly encoded SOImmVal field. int SOImmVal = ARM_AM::getSOImmVal(ThisVal); assert(SOImmVal != -1 && "Bit extraction didn't work?"); - + // Build the new ADD / SUB. BuildMI(MBB, MBBI, dl, TII.get(isSub ? ARM::SUBri : ARM::ADDri), DestReg) .addReg(BaseReg, RegState::Kill).addImm(SOImmVal) @@ -586,204 +558,12 @@ void emitARMRegPlusImmediate(MachineBasicBlock &MBB, } } -/// calcNumMI - Returns the number of instructions required to materialize -/// the specific add / sub r, c instruction. -static unsigned calcNumMI(int Opc, int ExtraOpc, unsigned Bytes, - unsigned NumBits, unsigned Scale) { - unsigned NumMIs = 0; - unsigned Chunk = ((1 << NumBits) - 1) * Scale; - - if (Opc == ARM::tADDrSPi) { - unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes; - Bytes -= ThisVal; - NumMIs++; - NumBits = 8; - Scale = 1; // Followed by a number of tADDi8. - Chunk = ((1 << NumBits) - 1) * Scale; - } - - NumMIs += Bytes / Chunk; - if ((Bytes % Chunk) != 0) - NumMIs++; - if (ExtraOpc) - NumMIs++; - return NumMIs; -} - -/// emitThumbRegPlusImmInReg - Emits a series of instructions to materialize -/// a destreg = basereg + immediate in Thumb code. Materialize the immediate -/// in a register using mov / mvn sequences or load the immediate from a -/// constpool entry. -static -void emitThumbRegPlusImmInReg(MachineBasicBlock &MBB, - MachineBasicBlock::iterator &MBBI, - unsigned DestReg, unsigned BaseReg, - int NumBytes, bool CanChangeCC, - const TargetInstrInfo &TII, - const ARMRegisterInfo& MRI, - DebugLoc dl) { - bool isHigh = !MRI.isLowRegister(DestReg) || - (BaseReg != 0 && !MRI.isLowRegister(BaseReg)); - bool isSub = false; - // Subtract doesn't have high register version. Load the negative value - // if either base or dest register is a high register. Also, if do not - // issue sub as part of the sequence if condition register is to be - // preserved. - if (NumBytes < 0 && !isHigh && CanChangeCC) { - isSub = true; - NumBytes = -NumBytes; - } - unsigned LdReg = DestReg; - if (DestReg == ARM::SP) { - assert(BaseReg == ARM::SP && "Unexpected!"); - LdReg = ARM::R3; - BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVlor2hir), ARM::R12) - .addReg(ARM::R3, RegState::Kill); - } - - if (NumBytes <= 255 && NumBytes >= 0) - BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVi8), LdReg).addImm(NumBytes); - else if (NumBytes < 0 && NumBytes >= -255) { - BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVi8), LdReg).addImm(NumBytes); - BuildMI(MBB, MBBI, dl, TII.get(ARM::tNEG), LdReg) - .addReg(LdReg, RegState::Kill); - } else - MRI.emitLoadConstPool(MBB, MBBI, LdReg, NumBytes, ARMCC::AL, 0, &TII, - true, dl); - - // Emit add / sub. - int Opc = (isSub) ? ARM::tSUBrr : (isHigh ? ARM::tADDhirr : ARM::tADDrr); - const MachineInstrBuilder MIB = BuildMI(MBB, MBBI, dl, - TII.get(Opc), DestReg); - if (DestReg == ARM::SP || isSub) - MIB.addReg(BaseReg).addReg(LdReg, RegState::Kill); - else - MIB.addReg(LdReg).addReg(BaseReg, RegState::Kill); - if (DestReg == ARM::SP) - BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVhir2lor), ARM::R3) - .addReg(ARM::R12, RegState::Kill); -} - -/// emitThumbRegPlusImmediate - Emits a series of instructions to materialize -/// a destreg = basereg + immediate in Thumb code. -static -void emitThumbRegPlusImmediate(MachineBasicBlock &MBB, - MachineBasicBlock::iterator &MBBI, - unsigned DestReg, unsigned BaseReg, - int NumBytes, const TargetInstrInfo &TII, - const ARMRegisterInfo& MRI, - DebugLoc dl) { - bool isSub = NumBytes < 0; - unsigned Bytes = (unsigned)NumBytes; - if (isSub) Bytes = -NumBytes; - bool isMul4 = (Bytes & 3) == 0; - bool isTwoAddr = false; - bool DstNotEqBase = false; - unsigned NumBits = 1; - unsigned Scale = 1; - int Opc = 0; - int ExtraOpc = 0; - - if (DestReg == BaseReg && BaseReg == ARM::SP) { - assert(isMul4 && "Thumb sp inc / dec size must be multiple of 4!"); - NumBits = 7; - Scale = 4; - Opc = isSub ? ARM::tSUBspi : ARM::tADDspi; - isTwoAddr = true; - } else if (!isSub && BaseReg == ARM::SP) { - // r1 = add sp, 403 - // => - // r1 = add sp, 100 * 4 - // r1 = add r1, 3 - if (!isMul4) { - Bytes &= ~3; - ExtraOpc = ARM::tADDi3; - } - NumBits = 8; - Scale = 4; - Opc = ARM::tADDrSPi; - } else { - // sp = sub sp, c - // r1 = sub sp, c - // r8 = sub sp, c - if (DestReg != BaseReg) - DstNotEqBase = true; - NumBits = 8; - Opc = isSub ? ARM::tSUBi8 : ARM::tADDi8; - isTwoAddr = true; - } - - unsigned NumMIs = calcNumMI(Opc, ExtraOpc, Bytes, NumBits, Scale); - unsigned Threshold = (DestReg == ARM::SP) ? 3 : 2; - if (NumMIs > Threshold) { - // This will expand into too many instructions. Load the immediate from a - // constpool entry. - emitThumbRegPlusImmInReg(MBB, MBBI, DestReg, BaseReg, NumBytes, true, TII, - MRI, dl); - return; - } - - if (DstNotEqBase) { - if (MRI.isLowRegister(DestReg) && MRI.isLowRegister(BaseReg)) { - // If both are low registers, emit DestReg = add BaseReg, max(Imm, 7) - unsigned Chunk = (1 << 3) - 1; - unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes; - Bytes -= ThisVal; - BuildMI(MBB, MBBI, dl,TII.get(isSub ? ARM::tSUBi3 : ARM::tADDi3), DestReg) - .addReg(BaseReg, RegState::Kill).addImm(ThisVal); - } else { - BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVr), DestReg) - .addReg(BaseReg, RegState::Kill); - } - BaseReg = DestReg; - } - - unsigned Chunk = ((1 << NumBits) - 1) * Scale; - while (Bytes) { - unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes; - Bytes -= ThisVal; - ThisVal /= Scale; - // Build the new tADD / tSUB. - if (isTwoAddr) - BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg) - .addReg(DestReg).addImm(ThisVal); - else { - bool isKill = BaseReg != ARM::SP; - BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg) - .addReg(BaseReg, getKillRegState(isKill)).addImm(ThisVal); - BaseReg = DestReg; - - if (Opc == ARM::tADDrSPi) { - // r4 = add sp, imm - // r4 = add r4, imm - // ... - NumBits = 8; - Scale = 1; - Chunk = ((1 << NumBits) - 1) * Scale; - Opc = isSub ? ARM::tSUBi8 : ARM::tADDi8; - isTwoAddr = true; - } - } - } - - if (ExtraOpc) - BuildMI(MBB, MBBI, dl, TII.get(ExtraOpc), DestReg) - .addReg(DestReg, RegState::Kill) - .addImm(((unsigned)NumBytes) & 3); -} - -static -void emitSPUpdate(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI, - int NumBytes, ARMCC::CondCodes Pred, unsigned PredReg, - bool isThumb, const TargetInstrInfo &TII, - const ARMRegisterInfo& MRI, - DebugLoc dl) { - if (isThumb) - emitThumbRegPlusImmediate(MBB, MBBI, ARM::SP, ARM::SP, NumBytes, TII, - MRI, dl); - else - emitARMRegPlusImmediate(MBB, MBBI, ARM::SP, ARM::SP, NumBytes, - Pred, PredReg, TII, dl); +void ARMRegisterInfo:: +emitSPUpdate(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI, + int NumBytes, ARMCC::CondCodes Pred, unsigned PredReg, + const TargetInstrInfo &TII, DebugLoc dl) const { + emitARMRegPlusImmediate(MBB, MBBI, ARM::SP, ARM::SP, NumBytes, + Pred, PredReg, TII, dl); } void ARMRegisterInfo:: @@ -797,7 +577,6 @@ eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB, DebugLoc dl = Old->getDebugLoc(); unsigned Amount = Old->getOperand(0).getImm(); if (Amount != 0) { - ARMFunctionInfo *AFI = MF.getInfo(); // We need to keep the stack aligned properly. To do this, we round the // amount of space needed for the outgoing arguments up to the next // alignment boundary. @@ -806,46 +585,22 @@ eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB, // Replace the pseudo instruction with a new instruction... unsigned Opc = Old->getOpcode(); - bool isThumb = AFI->isThumbFunction(); - ARMCC::CondCodes Pred = isThumb - ? ARMCC::AL : (ARMCC::CondCodes)Old->getOperand(1).getImm(); + ARMCC::CondCodes Pred = (ARMCC::CondCodes)Old->getOperand(1).getImm(); if (Opc == ARM::ADJCALLSTACKDOWN || Opc == ARM::tADJCALLSTACKDOWN) { // Note: PredReg is operand 2 for ADJCALLSTACKDOWN. - unsigned PredReg = isThumb ? 0 : Old->getOperand(2).getReg(); - emitSPUpdate(MBB, I, -Amount, Pred, PredReg, isThumb, TII, *this, dl); + unsigned PredReg = Old->getOperand(2).getReg(); + emitSPUpdate(MBB, I, -Amount, Pred, PredReg, TII, dl); } else { // Note: PredReg is operand 3 for ADJCALLSTACKUP. - unsigned PredReg = isThumb ? 0 : Old->getOperand(3).getReg(); + unsigned PredReg = Old->getOperand(3).getReg(); assert(Opc == ARM::ADJCALLSTACKUP || Opc == ARM::tADJCALLSTACKUP); - emitSPUpdate(MBB, I, Amount, Pred, PredReg, isThumb, TII, *this, dl); + emitSPUpdate(MBB, I, Amount, Pred, PredReg, TII, dl); } } } MBB.erase(I); } -/// emitThumbConstant - Emit a series of instructions to materialize a -/// constant. -static void emitThumbConstant(MachineBasicBlock &MBB, - MachineBasicBlock::iterator &MBBI, - unsigned DestReg, int Imm, - const TargetInstrInfo &TII, - const ARMRegisterInfo& MRI, - DebugLoc dl) { - bool isSub = Imm < 0; - if (isSub) Imm = -Imm; - - int Chunk = (1 << 8) - 1; - int ThisVal = (Imm > Chunk) ? Chunk : Imm; - Imm -= ThisVal; - BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVi8), DestReg).addImm(ThisVal); - if (Imm > 0) - emitThumbRegPlusImmediate(MBB, MBBI, DestReg, DestReg, Imm, TII, MRI, dl); - if (isSub) - BuildMI(MBB, MBBI, dl, TII.get(ARM::tNEG), DestReg) - .addReg(DestReg, RegState::Kill); -} - /// findScratchRegister - Find a 'free' ARM register. If register scavenger /// is not being used, R12 is available. Otherwise, try for a call-clobbered /// register first and then a spilled callee-saved register if that fails. @@ -868,17 +623,16 @@ void ARMRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, MachineBasicBlock &MBB = *MI.getParent(); MachineFunction &MF = *MBB.getParent(); ARMFunctionInfo *AFI = MF.getInfo(); - bool isThumb = AFI->isThumbFunction(); DebugLoc dl = MI.getDebugLoc(); while (!MI.getOperand(i).isFI()) { ++i; assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!"); } - + unsigned FrameReg = ARM::SP; int FrameIndex = MI.getOperand(i).getIndex(); - int Offset = MF.getFrameInfo()->getObjectOffset(FrameIndex) + + int Offset = MF.getFrameInfo()->getObjectOffset(FrameIndex) + MF.getFrameInfo()->getStackSize() + SPAdj; if (AFI->isGPRCalleeSavedArea1Frame(FrameIndex)) @@ -926,86 +680,20 @@ void ARMRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, MI.getOperand(i+1).ChangeToImmediate(ImmedOffset); return; } - + // Otherwise, we fallback to common code below to form the imm offset with // a sequence of ADDri instructions. First though, pull as much of the imm // into this ADDri as possible. unsigned RotAmt = ARM_AM::getSOImmValRotate(Offset); unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xFF, RotAmt); - + // We will handle these bits from offset, clear them. Offset &= ~ThisImmVal; - + // Get the properly encoded SOImmVal field. int ThisSOImmVal = ARM_AM::getSOImmVal(ThisImmVal); - assert(ThisSOImmVal != -1 && "Bit extraction didn't work?"); + assert(ThisSOImmVal != -1 && "Bit extraction didn't work?"); MI.getOperand(i+1).ChangeToImmediate(ThisSOImmVal); - } else if (Opcode == ARM::tADDrSPi) { - Offset += MI.getOperand(i+1).getImm(); - - // Can't use tADDrSPi if it's based off the frame pointer. - unsigned NumBits = 0; - unsigned Scale = 1; - if (FrameReg != ARM::SP) { - Opcode = ARM::tADDi3; - MI.setDesc(TII.get(ARM::tADDi3)); - NumBits = 3; - } else { - NumBits = 8; - Scale = 4; - assert((Offset & 3) == 0 && - "Thumb add/sub sp, #imm immediate must be multiple of 4!"); - } - - if (Offset == 0) { - // Turn it into a move. - MI.setDesc(TII.get(ARM::tMOVhir2lor)); - MI.getOperand(i).ChangeToRegister(FrameReg, false); - MI.RemoveOperand(i+1); - return; - } - - // Common case: small offset, fits into instruction. - unsigned Mask = (1 << NumBits) - 1; - if (((Offset / Scale) & ~Mask) == 0) { - // Replace the FrameIndex with sp / fp - MI.getOperand(i).ChangeToRegister(FrameReg, false); - MI.getOperand(i+1).ChangeToImmediate(Offset / Scale); - return; - } - - unsigned DestReg = MI.getOperand(0).getReg(); - unsigned Bytes = (Offset > 0) ? Offset : -Offset; - unsigned NumMIs = calcNumMI(Opcode, 0, Bytes, NumBits, Scale); - // MI would expand into a large number of instructions. Don't try to - // simplify the immediate. - if (NumMIs > 2) { - emitThumbRegPlusImmediate(MBB, II, DestReg, FrameReg, Offset, TII, - *this, dl); - MBB.erase(II); - return; - } - - if (Offset > 0) { - // Translate r0 = add sp, imm to - // r0 = add sp, 255*4 - // r0 = add r0, (imm - 255*4) - MI.getOperand(i).ChangeToRegister(FrameReg, false); - MI.getOperand(i+1).ChangeToImmediate(Mask); - Offset = (Offset - Mask * Scale); - MachineBasicBlock::iterator NII = next(II); - emitThumbRegPlusImmediate(MBB, NII, DestReg, DestReg, Offset, TII, - *this, dl); - } else { - // Translate r0 = add sp, -imm to - // r0 = -imm (this is then translated into a series of instructons) - // r0 = add r0, sp - emitThumbConstant(MBB, II, DestReg, Offset, TII, *this, dl); - MI.setDesc(TII.get(ARM::tADDhirr)); - MI.getOperand(i).ChangeToRegister(DestReg, false, false, true); - MI.getOperand(i+1).ChangeToRegister(FrameReg, false); - } - return; } else { unsigned ImmIdx = 0; int InstrOffs = 0; @@ -1037,13 +725,6 @@ void ARMRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, Scale = 4; break; } - case ARMII::AddrModeTs: { - ImmIdx = i+1; - InstrOffs = MI.getOperand(ImmIdx).getImm(); - NumBits = (FrameReg == ARM::SP) ? 8 : 5; - Scale = 4; - break; - } default: assert(0 && "Unsupported addressing mode!"); abort(); @@ -1052,7 +733,7 @@ void ARMRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, Offset += InstrOffs * Scale; assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!"); - if (Offset < 0 && !isThumb) { + if (Offset < 0) { Offset = -Offset; isSub = true; } @@ -1070,121 +751,34 @@ void ARMRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, return; } - bool isThumSpillRestore = Opcode == ARM::tRestore || Opcode == ARM::tSpill; - if (AddrMode == ARMII::AddrModeTs) { - // Thumb tLDRspi, tSTRspi. These will change to instructions that use - // a different base register. - NumBits = 5; - Mask = (1 << NumBits) - 1; - } - // If this is a thumb spill / restore, we will be using a constpool load to - // materialize the offset. - if (AddrMode == ARMII::AddrModeTs && isThumSpillRestore) - ImmOp.ChangeToImmediate(0); - else { - // Otherwise, it didn't fit. Pull in what we can to simplify the immed. - ImmedOffset = ImmedOffset & Mask; - if (isSub) - ImmedOffset |= 1 << NumBits; - ImmOp.ChangeToImmediate(ImmedOffset); - Offset &= ~(Mask*Scale); - } + // Otherwise, it didn't fit. Pull in what we can to simplify the immed. + ImmedOffset = ImmedOffset & Mask; + if (isSub) + ImmedOffset |= 1 << NumBits; + ImmOp.ChangeToImmediate(ImmedOffset); + Offset &= ~(Mask*Scale); } - + // If we get here, the immediate doesn't fit into the instruction. We folded // as much as possible above, handle the rest, providing a register that is // SP+LargeImm. assert(Offset && "This code isn't needed if offset already handled!"); - if (isThumb) { - if (Desc.mayLoad()) { - // Use the destination register to materialize sp + offset. - unsigned TmpReg = MI.getOperand(0).getReg(); - bool UseRR = false; - if (Opcode == ARM::tRestore) { - if (FrameReg == ARM::SP) - emitThumbRegPlusImmInReg(MBB, II, TmpReg, FrameReg, - Offset, false, TII, *this, dl); - else { - emitLoadConstPool(MBB, II, TmpReg, Offset, ARMCC::AL, 0, &TII, - true, dl); - UseRR = true; - } - } else - emitThumbRegPlusImmediate(MBB, II, TmpReg, FrameReg, Offset, TII, - *this, dl); - MI.setDesc(TII.get(ARM::tLDR)); - MI.getOperand(i).ChangeToRegister(TmpReg, false, false, true); - if (UseRR) - // Use [reg, reg] addrmode. - MI.addOperand(MachineOperand::CreateReg(FrameReg, false)); - else // tLDR has an extra register operand. - MI.addOperand(MachineOperand::CreateReg(0, false)); - } else if (Desc.mayStore()) { - // FIXME! This is horrific!!! We need register scavenging. - // Our temporary workaround has marked r3 unavailable. Of course, r3 is - // also a ABI register so it's possible that is is the register that is - // being storing here. If that's the case, we do the following: - // r12 = r2 - // Use r2 to materialize sp + offset - // str r3, r2 - // r2 = r12 - unsigned ValReg = MI.getOperand(0).getReg(); - unsigned TmpReg = ARM::R3; - bool UseRR = false; - if (ValReg == ARM::R3) { - BuildMI(MBB, II, dl, TII.get(ARM::tMOVlor2hir), ARM::R12) - .addReg(ARM::R2, RegState::Kill); - TmpReg = ARM::R2; - } - if (TmpReg == ARM::R3 && AFI->isR3LiveIn()) - BuildMI(MBB, II, dl, TII.get(ARM::tMOVlor2hir), ARM::R12) - .addReg(ARM::R3, RegState::Kill); - if (Opcode == ARM::tSpill) { - if (FrameReg == ARM::SP) - emitThumbRegPlusImmInReg(MBB, II, TmpReg, FrameReg, - Offset, false, TII, *this, dl); - else { - emitLoadConstPool(MBB, II, TmpReg, Offset, ARMCC::AL, 0, &TII, - true, dl); - UseRR = true; - } - } else - emitThumbRegPlusImmediate(MBB, II, TmpReg, FrameReg, Offset, TII, - *this, dl); - MI.setDesc(TII.get(ARM::tSTR)); - MI.getOperand(i).ChangeToRegister(TmpReg, false, false, true); - if (UseRR) // Use [reg, reg] addrmode. - MI.addOperand(MachineOperand::CreateReg(FrameReg, false)); - else // tSTR has an extra register operand. - MI.addOperand(MachineOperand::CreateReg(0, false)); - - MachineBasicBlock::iterator NII = next(II); - if (ValReg == ARM::R3) - BuildMI(MBB, NII, dl, TII.get(ARM::tMOVhir2lor), ARM::R2) - .addReg(ARM::R12, RegState::Kill); - if (TmpReg == ARM::R3 && AFI->isR3LiveIn()) - BuildMI(MBB, NII, dl, TII.get(ARM::tMOVhir2lor), ARM::R3) - .addReg(ARM::R12, RegState::Kill); - } else - assert(false && "Unexpected opcode!"); - } else { - // Insert a set of r12 with the full address: r12 = sp + offset - // If the offset we have is too large to fit into the instruction, we need - // to form it with a series of ADDri's. Do this by taking 8-bit chunks - // out of 'Offset'. - unsigned ScratchReg = findScratchRegister(RS, &ARM::GPRRegClass, AFI); - if (ScratchReg == 0) - // No register is "free". Scavenge a register. - ScratchReg = RS->scavengeRegister(&ARM::GPRRegClass, II, SPAdj); - int PIdx = MI.findFirstPredOperandIdx(); - ARMCC::CondCodes Pred = (PIdx == -1) - ? ARMCC::AL : (ARMCC::CondCodes)MI.getOperand(PIdx).getImm(); - unsigned PredReg = (PIdx == -1) ? 0 : MI.getOperand(PIdx+1).getReg(); - emitARMRegPlusImmediate(MBB, II, ScratchReg, FrameReg, - isSub ? -Offset : Offset, Pred, PredReg, TII, dl); - MI.getOperand(i).ChangeToRegister(ScratchReg, false, false, true); - } + // Insert a set of r12 with the full address: r12 = sp + offset + // If the offset we have is too large to fit into the instruction, we need + // to form it with a series of ADDri's. Do this by taking 8-bit chunks + // out of 'Offset'. + unsigned ScratchReg = findScratchRegister(RS, &ARM::GPRRegClass, AFI); + if (ScratchReg == 0) + // No register is "free". Scavenge a register. + ScratchReg = RS->scavengeRegister(&ARM::GPRRegClass, II, SPAdj); + int PIdx = MI.findFirstPredOperandIdx(); + ARMCC::CondCodes Pred = (PIdx == -1) + ? ARMCC::AL : (ARMCC::CondCodes)MI.getOperand(PIdx).getImm(); + unsigned PredReg = (PIdx == -1) ? 0 : MI.getOperand(PIdx+1).getReg(); + emitARMRegPlusImmediate(MBB, II, ScratchReg, FrameReg, + isSub ? -Offset : Offset, Pred, PredReg, TII, dl); + MI.getOperand(i).ChangeToRegister(ScratchReg, false, false, true); } static unsigned estimateStackSize(MachineFunction &MF, MachineFrameInfo *MFI) { @@ -1206,8 +800,8 @@ static unsigned estimateStackSize(MachineFunction &MF, MachineFrameInfo *MFI) { } void -ARMRegisterInfo::processFunctionBeforeCalleeSavedScan(MachineFunction &MF, - RegScavenger *RS) const { +ARMBaseRegisterInfo::processFunctionBeforeCalleeSavedScan(MachineFunction &MF, + RegScavenger *RS) const { // This tells PEI to spill the FP as if it is any other callee-save register // to take advantage the eliminateFrameIndex machinery. This also ensures it // is spilled in the order specified by getCalleeSavedRegs() to make it easier @@ -1266,7 +860,7 @@ ARMRegisterInfo::processFunctionBeforeCalleeSavedScan(MachineFunction &MF, default: break; } - } else { + } else { if (!STI.isTargetDarwin()) { UnspilledCS1GPRs.push_back(Reg); continue; @@ -1460,40 +1054,23 @@ void ARMRegisterInfo::emitPrologue(MachineFunction &MF) const { MachineBasicBlock::iterator MBBI = MBB.begin(); MachineFrameInfo *MFI = MF.getFrameInfo(); ARMFunctionInfo *AFI = MF.getInfo(); - bool isThumb = AFI->isThumbFunction(); unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize(); unsigned NumBytes = MFI->getStackSize(); const std::vector &CSI = MFI->getCalleeSavedInfo(); DebugLoc dl = (MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc::getUnknownLoc()); - if (isThumb) { - // Check if R3 is live in. It might have to be used as a scratch register. - for (MachineRegisterInfo::livein_iterator I =MF.getRegInfo().livein_begin(), - E = MF.getRegInfo().livein_end(); I != E; ++I) { - if (I->first == ARM::R3) { - AFI->setR3IsLiveIn(true); - break; - } - } - - // Thumb add/sub sp, imm8 instructions implicitly multiply the offset by 4. - NumBytes = (NumBytes + 3) & ~3; - MFI->setStackSize(NumBytes); - } - // Determine the sizes of each callee-save spill areas and record which frame // belongs to which callee-save spill areas. unsigned GPRCS1Size = 0, GPRCS2Size = 0, DPRCSSize = 0; int FramePtrSpillFI = 0; if (VARegSaveSize) - emitSPUpdate(MBB, MBBI, -VARegSaveSize, ARMCC::AL, 0, isThumb, TII, - *this, dl); + emitSPUpdate(MBB, MBBI, -VARegSaveSize, ARMCC::AL, 0, TII, dl); if (!AFI->hasStackFrame()) { if (NumBytes != 0) - emitSPUpdate(MBB, MBBI, -NumBytes, ARMCC::AL, 0, isThumb, TII, *this, dl); + emitSPUpdate(MBB, MBBI, -NumBytes, ARMCC::AL, 0, TII, dl); return; } @@ -1531,34 +1108,25 @@ void ARMRegisterInfo::emitPrologue(MachineFunction &MF) const { } } - if (!isThumb) { - // Build the new SUBri to adjust SP for integer callee-save spill area 1. - emitSPUpdate(MBB, MBBI, -GPRCS1Size, ARMCC::AL, 0, isThumb, TII, *this, dl); - movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, 1, STI); - } else if (MBBI != MBB.end() && MBBI->getOpcode() == ARM::tPUSH) { - ++MBBI; - if (MBBI != MBB.end()) - dl = MBBI->getDebugLoc(); - } + // Build the new SUBri to adjust SP for integer callee-save spill area 1. + emitSPUpdate(MBB, MBBI, -GPRCS1Size, ARMCC::AL, 0, TII, dl); + movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, 1, STI); // Darwin ABI requires FP to point to the stack slot that contains the // previous FP. if (STI.isTargetDarwin() || hasFP(MF)) { MachineInstrBuilder MIB = - BuildMI(MBB, MBBI, dl, TII.get(isThumb ? ARM::tADDrSPi : ARM::ADDri), - FramePtr) + BuildMI(MBB, MBBI, dl, TII.get(ARM::ADDri), FramePtr) .addFrameIndex(FramePtrSpillFI).addImm(0); - if (!isThumb) AddDefaultCC(AddDefaultPred(MIB)); + AddDefaultCC(AddDefaultPred(MIB)); } - if (!isThumb) { - // Build the new SUBri to adjust SP for integer callee-save spill area 2. - emitSPUpdate(MBB, MBBI, -GPRCS2Size, ARMCC::AL, 0, false, TII, *this, dl); + // Build the new SUBri to adjust SP for integer callee-save spill area 2. + emitSPUpdate(MBB, MBBI, -GPRCS2Size, ARMCC::AL, 0, TII, dl); - // Build the new SUBri to adjust SP for FP callee-save spill area. - movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, 2, STI); - emitSPUpdate(MBB, MBBI, -DPRCSSize, ARMCC::AL, 0, false, TII, *this, dl); - } + // Build the new SUBri to adjust SP for FP callee-save spill area. + movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, 2, STI); + emitSPUpdate(MBB, MBBI, -DPRCSSize, ARMCC::AL, 0, TII, dl); // Determine starting offsets of spill areas. unsigned DPRCSOffset = NumBytes - (GPRCS1Size + GPRCS2Size + DPRCSSize); @@ -1568,16 +1136,15 @@ void ARMRegisterInfo::emitPrologue(MachineFunction &MF) const { AFI->setGPRCalleeSavedArea1Offset(GPRCS1Offset); AFI->setGPRCalleeSavedArea2Offset(GPRCS2Offset); AFI->setDPRCalleeSavedAreaOffset(DPRCSOffset); - + NumBytes = DPRCSOffset; if (NumBytes) { // Insert it after all the callee-save spills. - if (!isThumb) - movePastCSLoadStoreOps(MBB, MBBI, ARM::FSTD, 3, STI); - emitSPUpdate(MBB, MBBI, -NumBytes, ARMCC::AL, 0, isThumb, TII, *this, dl); + movePastCSLoadStoreOps(MBB, MBBI, ARM::FSTD, 3, STI); + emitSPUpdate(MBB, MBBI, -NumBytes, ARMCC::AL, 0, TII, dl); } - if(STI.isTargetELF() && hasFP(MF)) { + if (STI.isTargetELF() && hasFP(MF)) { MFI->setOffsetAdjustment(MFI->getOffsetAdjustment() - AFI->getFramePtrSpillOffset()); } @@ -1596,8 +1163,7 @@ static bool isCalleeSavedRegister(unsigned Reg, const unsigned *CSRegs) { static bool isCSRestore(MachineInstr *MI, const unsigned *CSRegs) { return ((MI->getOpcode() == ARM::FLDD || - MI->getOpcode() == ARM::LDR || - MI->getOpcode() == ARM::tRestore) && + MI->getOpcode() == ARM::LDR) && MI->getOperand(1).isFI() && isCalleeSavedRegister(MI->getOperand(0).getReg(), CSRegs)); } @@ -1605,20 +1171,17 @@ static bool isCSRestore(MachineInstr *MI, const unsigned *CSRegs) { void ARMRegisterInfo::emitEpilogue(MachineFunction &MF, MachineBasicBlock &MBB) const { MachineBasicBlock::iterator MBBI = prior(MBB.end()); - assert((MBBI->getOpcode() == ARM::BX_RET || - MBBI->getOpcode() == ARM::tBX_RET || - MBBI->getOpcode() == ARM::tPOP_RET) && + assert(MBBI->getOpcode() == ARM::BX_RET && "Can only insert epilog into returning blocks"); DebugLoc dl = MBBI->getDebugLoc(); MachineFrameInfo *MFI = MF.getFrameInfo(); ARMFunctionInfo *AFI = MF.getInfo(); - bool isThumb = AFI->isThumbFunction(); unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize(); int NumBytes = (int)MFI->getStackSize(); if (!AFI->hasStackFrame()) { if (NumBytes != 0) - emitSPUpdate(MBB, MBBI, NumBytes, ARMCC::AL, 0, isThumb, TII, *this, dl); + emitSPUpdate(MBB, MBBI, NumBytes, ARMCC::AL, 0, TII, dl); } else { // Unwind MBBI to point to first LDR / FLDD. const unsigned *CSRegs = getCalleeSavedRegs(); @@ -1634,110 +1197,76 @@ void ARMRegisterInfo::emitEpilogue(MachineFunction &MF, NumBytes -= (AFI->getGPRCalleeSavedArea1Size() + AFI->getGPRCalleeSavedArea2Size() + AFI->getDPRCalleeSavedAreaSize()); - if (isThumb) { - if (hasFP(MF)) { - NumBytes = AFI->getFramePtrSpillOffset() - NumBytes; - // Reset SP based on frame pointer only if the stack frame extends beyond - // frame pointer stack slot or target is ELF and the function has FP. + + // Darwin ABI requires FP to point to the stack slot that contains the + // previous FP. + if ((STI.isTargetDarwin() && NumBytes) || hasFP(MF)) { + NumBytes = AFI->getFramePtrSpillOffset() - NumBytes; + // Reset SP based on frame pointer only if the stack frame extends beyond + // frame pointer stack slot or target is ELF and the function has FP. + if (AFI->getGPRCalleeSavedArea2Size() || + AFI->getDPRCalleeSavedAreaSize() || + AFI->getDPRCalleeSavedAreaOffset()|| + hasFP(MF)) { if (NumBytes) - emitThumbRegPlusImmediate(MBB, MBBI, ARM::SP, FramePtr, -NumBytes, - TII, *this, dl); + BuildMI(MBB, MBBI, dl, TII.get(ARM::SUBri), ARM::SP).addReg(FramePtr) + .addImm(NumBytes) + .addImm((unsigned)ARMCC::AL).addReg(0).addReg(0); else - BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVlor2hir), ARM::SP) - .addReg(FramePtr); - } else { - if (MBBI->getOpcode() == ARM::tBX_RET && - &MBB.front() != MBBI && - prior(MBBI)->getOpcode() == ARM::tPOP) { - MachineBasicBlock::iterator PMBBI = prior(MBBI); - emitSPUpdate(MBB, PMBBI, NumBytes, ARMCC::AL, 0, isThumb, TII, - *this, dl); - } else - emitSPUpdate(MBB, MBBI, NumBytes, ARMCC::AL, 0, isThumb, TII, - *this, dl); + BuildMI(MBB, MBBI, dl, TII.get(ARM::MOVr), ARM::SP).addReg(FramePtr) + .addImm((unsigned)ARMCC::AL).addReg(0).addReg(0); } - } else { - // Darwin ABI requires FP to point to the stack slot that contains the - // previous FP. - if ((STI.isTargetDarwin() && NumBytes) || hasFP(MF)) { - NumBytes = AFI->getFramePtrSpillOffset() - NumBytes; - // Reset SP based on frame pointer only if the stack frame extends beyond - // frame pointer stack slot or target is ELF and the function has FP. - if (AFI->getGPRCalleeSavedArea2Size() || - AFI->getDPRCalleeSavedAreaSize() || - AFI->getDPRCalleeSavedAreaOffset()|| - hasFP(MF)) { - if (NumBytes) - BuildMI(MBB, MBBI, dl, TII.get(ARM::SUBri), ARM::SP).addReg(FramePtr) - .addImm(NumBytes) - .addImm((unsigned)ARMCC::AL).addReg(0).addReg(0); - else - BuildMI(MBB, MBBI, dl, TII.get(ARM::MOVr), ARM::SP).addReg(FramePtr) - .addImm((unsigned)ARMCC::AL).addReg(0).addReg(0); - } - } else if (NumBytes) { - emitSPUpdate(MBB, MBBI, NumBytes, ARMCC::AL, 0, false, TII, *this, dl); - } - - // Move SP to start of integer callee save spill area 2. - movePastCSLoadStoreOps(MBB, MBBI, ARM::FLDD, 3, STI); - emitSPUpdate(MBB, MBBI, AFI->getDPRCalleeSavedAreaSize(), ARMCC::AL, 0, - false, TII, *this, dl); - - // Move SP to start of integer callee save spill area 1. - movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, 2, STI); - emitSPUpdate(MBB, MBBI, AFI->getGPRCalleeSavedArea2Size(), ARMCC::AL, 0, - false, TII, *this, dl); - - // Move SP to SP upon entry to the function. - movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, 1, STI); - emitSPUpdate(MBB, MBBI, AFI->getGPRCalleeSavedArea1Size(), ARMCC::AL, 0, - false, TII, *this, dl); + } else if (NumBytes) { + emitSPUpdate(MBB, MBBI, NumBytes, ARMCC::AL, 0, TII, dl); } + + // Move SP to start of integer callee save spill area 2. + movePastCSLoadStoreOps(MBB, MBBI, ARM::FLDD, 3, STI); + emitSPUpdate(MBB, MBBI, AFI->getDPRCalleeSavedAreaSize(), ARMCC::AL, 0, + TII, dl); + + // Move SP to start of integer callee save spill area 1. + movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, 2, STI); + emitSPUpdate(MBB, MBBI, AFI->getGPRCalleeSavedArea2Size(), ARMCC::AL, 0, + TII, dl); + + // Move SP to SP upon entry to the function. + movePastCSLoadStoreOps(MBB, MBBI, ARM::LDR, 1, STI); + emitSPUpdate(MBB, MBBI, AFI->getGPRCalleeSavedArea1Size(), ARMCC::AL, 0, + TII, dl); } - if (VARegSaveSize) { - if (isThumb) - // Epilogue for vararg functions: pop LR to R3 and branch off it. - // FIXME: Verify this is still ok when R3 is no longer being reserved. - BuildMI(MBB, MBBI, dl, TII.get(ARM::tPOP)).addReg(ARM::R3); + if (VARegSaveSize) + emitSPUpdate(MBB, MBBI, VARegSaveSize, ARMCC::AL, 0, TII, dl); - emitSPUpdate(MBB, MBBI, VARegSaveSize, ARMCC::AL, 0, isThumb, TII, - *this, dl); - - if (isThumb) { - BuildMI(MBB, MBBI, dl, TII.get(ARM::tBX_RET_vararg)).addReg(ARM::R3); - MBB.erase(MBBI); - } - } } -unsigned ARMRegisterInfo::getRARegister() const { +unsigned ARMBaseRegisterInfo::getRARegister() const { return ARM::LR; } -unsigned ARMRegisterInfo::getFrameRegister(MachineFunction &MF) const { +unsigned ARMBaseRegisterInfo::getFrameRegister(MachineFunction &MF) const { if (STI.isTargetDarwin() || hasFP(MF)) return FramePtr; return ARM::SP; } -unsigned ARMRegisterInfo::getEHExceptionRegister() const { +unsigned ARMBaseRegisterInfo::getEHExceptionRegister() const { assert(0 && "What is the exception register"); return 0; } -unsigned ARMRegisterInfo::getEHHandlerRegister() const { +unsigned ARMBaseRegisterInfo::getEHHandlerRegister() const { assert(0 && "What is the exception handler register"); return 0; } -int ARMRegisterInfo::getDwarfRegNum(unsigned RegNum, bool isEH) const { +int ARMBaseRegisterInfo::getDwarfRegNum(unsigned RegNum, bool isEH) const { return ARMGenRegisterInfo::getDwarfRegNumFull(RegNum, 0); } -unsigned ARMRegisterInfo::getRegisterPairEven(unsigned Reg, - const MachineFunction &MF) const { +unsigned ARMBaseRegisterInfo::getRegisterPairEven(unsigned Reg, + const MachineFunction &MF) const { switch (Reg) { default: break; // Return 0 if either register of the pair is a special register. @@ -1810,7 +1339,7 @@ unsigned ARMRegisterInfo::getRegisterPairEven(unsigned Reg, return 0; } -unsigned ARMRegisterInfo::getRegisterPairOdd(unsigned Reg, +unsigned ARMBaseRegisterInfo::getRegisterPairOdd(unsigned Reg, const MachineFunction &MF) const { switch (Reg) { default: break; diff --git a/lib/Target/ARM/ARMRegisterInfo.h b/lib/Target/ARM/ARMRegisterInfo.h index e8f4fd805d5..1154c0542b3 100644 --- a/lib/Target/ARM/ARMRegisterInfo.h +++ b/lib/Target/ARM/ARMRegisterInfo.h @@ -14,6 +14,7 @@ #ifndef ARMREGISTERINFO_H #define ARMREGISTERINFO_H +#include "ARM.h" #include "llvm/Target/TargetRegisterInfo.h" #include "ARMGenRegisterInfo.h.inc" @@ -30,21 +31,15 @@ namespace ARMRI { }; } -struct ARMRegisterInfo : public ARMGenRegisterInfo { +struct ARMBaseRegisterInfo : public ARMGenRegisterInfo { +protected: const TargetInstrInfo &TII; const ARMSubtarget &STI; + /// FramePtr - ARM physical register used as frame ptr. + unsigned FramePtr; public: - ARMRegisterInfo(const TargetInstrInfo &tii, const ARMSubtarget &STI); - - /// emitLoadConstPool - Emits a load from constpool to materialize the - /// specified immediate. - void emitLoadConstPool(MachineBasicBlock &MBB, - MachineBasicBlock::iterator &MBBI, - unsigned DestReg, int Val, - unsigned Pred, unsigned PredReg, - const TargetInstrInfo *TII, bool isThumb, - DebugLoc dl) const; + ARMBaseRegisterInfo(const TargetInstrInfo &tii, const ARMSubtarget &STI); /// getRegisterNumbering - Given the enum value for some register, e.g. /// ARM::LR, return the number that it corresponds to (e.g. 14). @@ -55,8 +50,6 @@ public: static unsigned getRegisterNumbering(unsigned RegEnum, bool &isSPVFP); /// Code Generation virtual methods... - const TargetRegisterClass * - getPhysicalRegisterRegClass(unsigned Reg, MVT VT = MVT::Other) const; const unsigned *getCalleeSavedRegs(const MachineFunction *MF = 0) const; const TargetRegisterClass* const* @@ -79,25 +72,11 @@ public: void UpdateRegAllocHint(unsigned Reg, unsigned NewReg, MachineFunction &MF) const; - bool requiresRegisterScavenging(const MachineFunction &MF) const; - bool hasFP(const MachineFunction &MF) const; - bool hasReservedCallFrame(MachineFunction &MF) const; - - void eliminateCallFramePseudoInstr(MachineFunction &MF, - MachineBasicBlock &MBB, - MachineBasicBlock::iterator I) const; - - void eliminateFrameIndex(MachineBasicBlock::iterator II, - int SPAdj, RegScavenger *RS = NULL) const; - void processFunctionBeforeCalleeSavedScan(MachineFunction &MF, RegScavenger *RS = NULL) const; - void emitPrologue(MachineFunction &MF) const; - void emitEpilogue(MachineFunction &MF, MachineBasicBlock &MBB) const; - // Debug information queries. unsigned getRARegister() const; unsigned getFrameRegister(MachineFunction &MF) const; @@ -107,17 +86,48 @@ public: unsigned getEHHandlerRegister() const; int getDwarfRegNum(unsigned RegNum, bool isEH) const; - + bool isLowRegister(unsigned Reg) const; private: - /// FramePtr - ARM physical register used as frame ptr. - unsigned FramePtr; - unsigned getRegisterPairEven(unsigned Reg, const MachineFunction &MF) const; unsigned getRegisterPairOdd(unsigned Reg, const MachineFunction &MF) const; +}; +struct ARMRegisterInfo : public ARMBaseRegisterInfo { +public: + ARMRegisterInfo(const TargetInstrInfo &tii, const ARMSubtarget &STI); + + /// emitLoadConstPool - Emits a load from constpool to materialize the + /// specified immediate. + void emitLoadConstPool(MachineBasicBlock &MBB, + MachineBasicBlock::iterator &MBBI, + unsigned DestReg, int Val, + unsigned Pred, unsigned PredReg, + const TargetInstrInfo *TII, + DebugLoc dl) const; + + /// Code Generation virtual methods... + bool isReservedReg(const MachineFunction &MF, unsigned Reg) const; + + bool requiresRegisterScavenging(const MachineFunction &MF) const; + + bool hasReservedCallFrame(MachineFunction &MF) const; + + void eliminateCallFramePseudoInstr(MachineFunction &MF, + MachineBasicBlock &MBB, + MachineBasicBlock::iterator I) const; + + void eliminateFrameIndex(MachineBasicBlock::iterator II, + int SPAdj, RegScavenger *RS = NULL) const; + + void emitPrologue(MachineFunction &MF) const; + void emitEpilogue(MachineFunction &MF, MachineBasicBlock &MBB) const; + + void emitSPUpdate(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI, + int NumBytes, ARMCC::CondCodes Pred, unsigned PredReg, + const TargetInstrInfo &TII, DebugLoc dl) const; }; } // end namespace llvm diff --git a/lib/Target/ARM/ARMTargetMachine.h b/lib/Target/ARM/ARMTargetMachine.h index 0b49b9275b9..eab1e46b1d7 100644 --- a/lib/Target/ARM/ARMTargetMachine.h +++ b/lib/Target/ARM/ARMTargetMachine.h @@ -120,7 +120,7 @@ class ThumbTargetMachine : public ARMBaseTargetMachine { public: ThumbTargetMachine(const Module &M, const std::string &FS); - virtual const ARMRegisterInfo *getRegisterInfo() const { + virtual const ThumbRegisterInfo *getRegisterInfo() const { return &InstrInfo.getRegisterInfo(); } diff --git a/lib/Target/ARM/ThumbInstrInfo.cpp b/lib/Target/ARM/ThumbInstrInfo.cpp index 075d940bcca..d9a499dacd1 100644 --- a/lib/Target/ARM/ThumbInstrInfo.cpp +++ b/lib/Target/ARM/ThumbInstrInfo.cpp @@ -23,7 +23,7 @@ using namespace llvm; ThumbInstrInfo::ThumbInstrInfo(const ARMSubtarget &STI) - : ARMBaseInstrInfo(STI) { + : ARMBaseInstrInfo(STI), RI(*this, STI) { } bool ThumbInstrInfo::isMoveInstr(const MachineInstr &MI, @@ -114,6 +114,37 @@ bool ThumbInstrInfo::copyRegToReg(MachineBasicBlock &MBB, return false; } +bool ThumbInstrInfo:: +canFoldMemoryOperand(const MachineInstr *MI, + const SmallVectorImpl &Ops) const { + if (Ops.size() != 1) return false; + + unsigned OpNum = Ops[0]; + unsigned Opc = MI->getOpcode(); + switch (Opc) { + default: break; + case ARM::tMOVr: + case ARM::tMOVlor2hir: + case ARM::tMOVhir2lor: + case ARM::tMOVhir2hir: { + if (OpNum == 0) { // move -> store + unsigned SrcReg = MI->getOperand(1).getReg(); + if (RI.isPhysicalRegister(SrcReg) && !RI.isLowRegister(SrcReg)) + // tSpill cannot take a high register operand. + return false; + } else { // move -> load + unsigned DstReg = MI->getOperand(0).getReg(); + if (RI.isPhysicalRegister(DstReg) && !RI.isLowRegister(DstReg)) + // tRestore cannot target a high register operand. + return false; + } + return true; + } + } + + return false; +} + void ThumbInstrInfo:: storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I, unsigned SrcReg, bool isKill, int FI, @@ -244,7 +275,6 @@ MachineInstr *ThumbInstrInfo:: foldMemoryOperandImpl(MachineFunction &MF, MachineInstr *MI, const SmallVectorImpl &Ops, int FI) const { if (Ops.size() != 1) return NULL; - const ARMRegisterInfo &RI = getRegisterInfo(); unsigned OpNum = Ops[0]; unsigned Opc = MI->getOpcode(); diff --git a/lib/Target/ARM/ThumbInstrInfo.h b/lib/Target/ARM/ThumbInstrInfo.h index dcf10957508..eefd07780d5 100644 --- a/lib/Target/ARM/ThumbInstrInfo.h +++ b/lib/Target/ARM/ThumbInstrInfo.h @@ -15,17 +15,24 @@ #define THUMBINSTRUCTIONINFO_H #include "llvm/Target/TargetInstrInfo.h" -#include "ARMRegisterInfo.h" #include "ARM.h" #include "ARMInstrInfo.h" +#include "ThumbRegisterInfo.h" namespace llvm { class ARMSubtarget; class ThumbInstrInfo : public ARMBaseInstrInfo { + ThumbRegisterInfo RI; public: explicit ThumbInstrInfo(const ARMSubtarget &STI); + /// getRegisterInfo - TargetInstrInfo is a superset of MRegister info. As + /// such, whenever a client has an instance of instruction info, it should + /// always be able to get register info as well (through this method). + /// + virtual const ThumbRegisterInfo &getRegisterInfo() const { return RI; } + /// Return true if the instruction is a register to register move and return /// the source and dest operands and their sub-register indices by reference. virtual bool isMoveInstr(const MachineInstr &MI, @@ -68,6 +75,9 @@ public: MachineBasicBlock::iterator MI, const std::vector &CSI) const; + virtual bool canFoldMemoryOperand(const MachineInstr *MI, + const SmallVectorImpl &Ops) const; + virtual MachineInstr* foldMemoryOperandImpl(MachineFunction &MF, MachineInstr* MI, const SmallVectorImpl &Ops, diff --git a/lib/Target/ARM/ThumbRegisterInfo.cpp b/lib/Target/ARM/ThumbRegisterInfo.cpp new file mode 100644 index 00000000000..40221d9cf7d --- /dev/null +++ b/lib/Target/ARM/ThumbRegisterInfo.cpp @@ -0,0 +1,757 @@ +//===- ThumbRegisterInfo.cpp - Thumb Register Information -------*- C++ -*-===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// This file contains the ARM implementation of the TargetRegisterInfo class. +// +//===----------------------------------------------------------------------===// + +#include "ARM.h" +#include "ARMAddressingModes.h" +#include "ARMMachineFunctionInfo.h" +#include "ARMSubtarget.h" +#include "ThumbInstrInfo.h" +#include "ThumbRegisterInfo.h" +#include "llvm/Constants.h" +#include "llvm/DerivedTypes.h" +#include "llvm/CodeGen/MachineConstantPool.h" +#include "llvm/CodeGen/MachineFrameInfo.h" +#include "llvm/CodeGen/MachineFunction.h" +#include "llvm/CodeGen/MachineInstrBuilder.h" +#include "llvm/CodeGen/MachineLocation.h" +#include "llvm/CodeGen/MachineRegisterInfo.h" +#include "llvm/Target/TargetFrameInfo.h" +#include "llvm/Target/TargetMachine.h" +#include "llvm/ADT/BitVector.h" +#include "llvm/ADT/SmallVector.h" +#include "llvm/Support/CommandLine.h" +using namespace llvm; + +static cl::opt +ThumbRegScavenging("enable-thumb-reg-scavenging", + cl::Hidden, + cl::desc("Enable register scavenging on Thumb")); + +ThumbRegisterInfo::ThumbRegisterInfo(const TargetInstrInfo &tii, + const ARMSubtarget &sti) + : ARMBaseRegisterInfo(tii, sti) { +} + +/// emitLoadConstPool - Emits a load from constpool to materialize the +/// specified immediate. +void ThumbRegisterInfo::emitLoadConstPool(MachineBasicBlock &MBB, + MachineBasicBlock::iterator &MBBI, + unsigned DestReg, int Val, + unsigned Pred, unsigned PredReg, + const TargetInstrInfo *TII, + DebugLoc dl) const { + MachineFunction &MF = *MBB.getParent(); + MachineConstantPool *ConstantPool = MF.getConstantPool(); + Constant *C = ConstantInt::get(Type::Int32Ty, Val); + unsigned Idx = ConstantPool->getConstantPoolIndex(C, 4); + + BuildMI(MBB, MBBI, dl, TII->get(ARM::tLDRcp), DestReg) + .addConstantPoolIndex(Idx); +} + +const TargetRegisterClass* +ThumbRegisterInfo::getPhysicalRegisterRegClass(unsigned Reg, MVT VT) const { + if (isLowRegister(Reg)) + return ARM::tGPRRegisterClass; + switch (Reg) { + default: + break; + case ARM::R8: case ARM::R9: case ARM::R10: case ARM::R11: + case ARM::R12: case ARM::SP: case ARM::LR: case ARM::PC: + return ARM::GPRRegisterClass; + } + + return TargetRegisterInfo::getPhysicalRegisterRegClass(Reg, VT); +} + +bool +ThumbRegisterInfo::requiresRegisterScavenging(const MachineFunction &MF) const { + return ThumbRegScavenging; +} + +bool ThumbRegisterInfo::hasReservedCallFrame(MachineFunction &MF) const { + const MachineFrameInfo *FFI = MF.getFrameInfo(); + unsigned CFSize = FFI->getMaxCallFrameSize(); + // It's not always a good idea to include the call frame as part of the + // stack frame. ARM (especially Thumb) has small immediate offset to + // address the stack frame. So a large call frame can cause poor codegen + // and may even makes it impossible to scavenge a register. + if (CFSize >= ((1 << 8) - 1) * 4 / 2) // Half of imm8 * 4 + return false; + + return !MF.getFrameInfo()->hasVarSizedObjects(); +} + +/// emitThumbRegPlusImmInReg - Emits a series of instructions to materialize +/// a destreg = basereg + immediate in Thumb code. Materialize the immediate +/// in a register using mov / mvn sequences or load the immediate from a +/// constpool entry. +static +void emitThumbRegPlusImmInReg(MachineBasicBlock &MBB, + MachineBasicBlock::iterator &MBBI, + unsigned DestReg, unsigned BaseReg, + int NumBytes, bool CanChangeCC, + const TargetInstrInfo &TII, + const ThumbRegisterInfo& MRI, + DebugLoc dl) { + bool isHigh = !MRI.isLowRegister(DestReg) || + (BaseReg != 0 && !MRI.isLowRegister(BaseReg)); + bool isSub = false; + // Subtract doesn't have high register version. Load the negative value + // if either base or dest register is a high register. Also, if do not + // issue sub as part of the sequence if condition register is to be + // preserved. + if (NumBytes < 0 && !isHigh && CanChangeCC) { + isSub = true; + NumBytes = -NumBytes; + } + unsigned LdReg = DestReg; + if (DestReg == ARM::SP) { + assert(BaseReg == ARM::SP && "Unexpected!"); + LdReg = ARM::R3; + BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVlor2hir), ARM::R12) + .addReg(ARM::R3, RegState::Kill); + } + + if (NumBytes <= 255 && NumBytes >= 0) + BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVi8), LdReg).addImm(NumBytes); + else if (NumBytes < 0 && NumBytes >= -255) { + BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVi8), LdReg).addImm(NumBytes); + BuildMI(MBB, MBBI, dl, TII.get(ARM::tNEG), LdReg) + .addReg(LdReg, RegState::Kill); + } else + MRI.emitLoadConstPool(MBB, MBBI, LdReg, NumBytes, ARMCC::AL, 0, &TII, dl); + + // Emit add / sub. + int Opc = (isSub) ? ARM::tSUBrr : (isHigh ? ARM::tADDhirr : ARM::tADDrr); + const MachineInstrBuilder MIB = BuildMI(MBB, MBBI, dl, + TII.get(Opc), DestReg); + if (DestReg == ARM::SP || isSub) + MIB.addReg(BaseReg).addReg(LdReg, RegState::Kill); + else + MIB.addReg(LdReg).addReg(BaseReg, RegState::Kill); + if (DestReg == ARM::SP) + BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVhir2lor), ARM::R3) + .addReg(ARM::R12, RegState::Kill); +} + +/// calcNumMI - Returns the number of instructions required to materialize +/// the specific add / sub r, c instruction. +static unsigned calcNumMI(int Opc, int ExtraOpc, unsigned Bytes, + unsigned NumBits, unsigned Scale) { + unsigned NumMIs = 0; + unsigned Chunk = ((1 << NumBits) - 1) * Scale; + + if (Opc == ARM::tADDrSPi) { + unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes; + Bytes -= ThisVal; + NumMIs++; + NumBits = 8; + Scale = 1; // Followed by a number of tADDi8. + Chunk = ((1 << NumBits) - 1) * Scale; + } + + NumMIs += Bytes / Chunk; + if ((Bytes % Chunk) != 0) + NumMIs++; + if (ExtraOpc) + NumMIs++; + return NumMIs; +} + +/// emitThumbRegPlusImmediate - Emits a series of instructions to materialize +/// a destreg = basereg + immediate in Thumb code. +static +void emitThumbRegPlusImmediate(MachineBasicBlock &MBB, + MachineBasicBlock::iterator &MBBI, + unsigned DestReg, unsigned BaseReg, + int NumBytes, const TargetInstrInfo &TII, + const ThumbRegisterInfo& MRI, + DebugLoc dl) { + bool isSub = NumBytes < 0; + unsigned Bytes = (unsigned)NumBytes; + if (isSub) Bytes = -NumBytes; + bool isMul4 = (Bytes & 3) == 0; + bool isTwoAddr = false; + bool DstNotEqBase = false; + unsigned NumBits = 1; + unsigned Scale = 1; + int Opc = 0; + int ExtraOpc = 0; + + if (DestReg == BaseReg && BaseReg == ARM::SP) { + assert(isMul4 && "Thumb sp inc / dec size must be multiple of 4!"); + NumBits = 7; + Scale = 4; + Opc = isSub ? ARM::tSUBspi : ARM::tADDspi; + isTwoAddr = true; + } else if (!isSub && BaseReg == ARM::SP) { + // r1 = add sp, 403 + // => + // r1 = add sp, 100 * 4 + // r1 = add r1, 3 + if (!isMul4) { + Bytes &= ~3; + ExtraOpc = ARM::tADDi3; + } + NumBits = 8; + Scale = 4; + Opc = ARM::tADDrSPi; + } else { + // sp = sub sp, c + // r1 = sub sp, c + // r8 = sub sp, c + if (DestReg != BaseReg) + DstNotEqBase = true; + NumBits = 8; + Opc = isSub ? ARM::tSUBi8 : ARM::tADDi8; + isTwoAddr = true; + } + + unsigned NumMIs = calcNumMI(Opc, ExtraOpc, Bytes, NumBits, Scale); + unsigned Threshold = (DestReg == ARM::SP) ? 3 : 2; + if (NumMIs > Threshold) { + // This will expand into too many instructions. Load the immediate from a + // constpool entry. + emitThumbRegPlusImmInReg(MBB, MBBI, DestReg, BaseReg, NumBytes, true, TII, + MRI, dl); + return; + } + + if (DstNotEqBase) { + if (MRI.isLowRegister(DestReg) && MRI.isLowRegister(BaseReg)) { + // If both are low registers, emit DestReg = add BaseReg, max(Imm, 7) + unsigned Chunk = (1 << 3) - 1; + unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes; + Bytes -= ThisVal; + BuildMI(MBB, MBBI, dl,TII.get(isSub ? ARM::tSUBi3 : ARM::tADDi3), DestReg) + .addReg(BaseReg, RegState::Kill).addImm(ThisVal); + } else { + BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVr), DestReg) + .addReg(BaseReg, RegState::Kill); + } + BaseReg = DestReg; + } + + unsigned Chunk = ((1 << NumBits) - 1) * Scale; + while (Bytes) { + unsigned ThisVal = (Bytes > Chunk) ? Chunk : Bytes; + Bytes -= ThisVal; + ThisVal /= Scale; + // Build the new tADD / tSUB. + if (isTwoAddr) + BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg) + .addReg(DestReg).addImm(ThisVal); + else { + bool isKill = BaseReg != ARM::SP; + BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg) + .addReg(BaseReg, getKillRegState(isKill)).addImm(ThisVal); + BaseReg = DestReg; + + if (Opc == ARM::tADDrSPi) { + // r4 = add sp, imm + // r4 = add r4, imm + // ... + NumBits = 8; + Scale = 1; + Chunk = ((1 << NumBits) - 1) * Scale; + Opc = isSub ? ARM::tSUBi8 : ARM::tADDi8; + isTwoAddr = true; + } + } + } + + if (ExtraOpc) + BuildMI(MBB, MBBI, dl, TII.get(ExtraOpc), DestReg) + .addReg(DestReg, RegState::Kill) + .addImm(((unsigned)NumBytes) & 3); +} + +void ThumbRegisterInfo:: +emitSPUpdate(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI, + int NumBytes, ARMCC::CondCodes Pred, unsigned PredReg, + const TargetInstrInfo &TII, DebugLoc dl) const { + emitThumbRegPlusImmediate(MBB, MBBI, ARM::SP, ARM::SP, NumBytes, TII, + *this, dl); +} + +void ThumbRegisterInfo:: +eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB, + MachineBasicBlock::iterator I) const { + if (!hasReservedCallFrame(MF)) { + // If we have alloca, convert as follows: + // ADJCALLSTACKDOWN -> sub, sp, sp, amount + // ADJCALLSTACKUP -> add, sp, sp, amount + MachineInstr *Old = I; + DebugLoc dl = Old->getDebugLoc(); + unsigned Amount = Old->getOperand(0).getImm(); + if (Amount != 0) { + // We need to keep the stack aligned properly. To do this, we round the + // amount of space needed for the outgoing arguments up to the next + // alignment boundary. + unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment(); + Amount = (Amount+Align-1)/Align*Align; + + // Replace the pseudo instruction with a new instruction... + unsigned Opc = Old->getOpcode(); + if (Opc == ARM::ADJCALLSTACKDOWN || Opc == ARM::tADJCALLSTACKDOWN) { + // Note: PredReg is operand 2 for ADJCALLSTACKDOWN. + emitSPUpdate(MBB, I, -Amount, ARMCC::AL, 0, TII, dl); + } else { + // Note: PredReg is operand 3 for ADJCALLSTACKUP. + assert(Opc == ARM::ADJCALLSTACKUP || Opc == ARM::tADJCALLSTACKUP); + emitSPUpdate(MBB, I, Amount, ARMCC::AL, 0, TII, dl); + } + } + } + MBB.erase(I); +} + +/// emitThumbConstant - Emit a series of instructions to materialize a +/// constant. +static void emitThumbConstant(MachineBasicBlock &MBB, + MachineBasicBlock::iterator &MBBI, + unsigned DestReg, int Imm, + const TargetInstrInfo &TII, + const ThumbRegisterInfo& MRI, + DebugLoc dl) { + bool isSub = Imm < 0; + if (isSub) Imm = -Imm; + + int Chunk = (1 << 8) - 1; + int ThisVal = (Imm > Chunk) ? Chunk : Imm; + Imm -= ThisVal; + BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVi8), DestReg).addImm(ThisVal); + if (Imm > 0) + emitThumbRegPlusImmediate(MBB, MBBI, DestReg, DestReg, Imm, TII, MRI, dl); + if (isSub) + BuildMI(MBB, MBBI, dl, TII.get(ARM::tNEG), DestReg) + .addReg(DestReg, RegState::Kill); +} + +void ThumbRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, + int SPAdj, RegScavenger *RS) const{ + unsigned i = 0; + MachineInstr &MI = *II; + MachineBasicBlock &MBB = *MI.getParent(); + MachineFunction &MF = *MBB.getParent(); + ARMFunctionInfo *AFI = MF.getInfo(); + DebugLoc dl = MI.getDebugLoc(); + + while (!MI.getOperand(i).isFI()) { + ++i; + assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!"); + } + + unsigned FrameReg = ARM::SP; + int FrameIndex = MI.getOperand(i).getIndex(); + int Offset = MF.getFrameInfo()->getObjectOffset(FrameIndex) + + MF.getFrameInfo()->getStackSize() + SPAdj; + + if (AFI->isGPRCalleeSavedArea1Frame(FrameIndex)) + Offset -= AFI->getGPRCalleeSavedArea1Offset(); + else if (AFI->isGPRCalleeSavedArea2Frame(FrameIndex)) + Offset -= AFI->getGPRCalleeSavedArea2Offset(); + else if (hasFP(MF)) { + assert(SPAdj == 0 && "Unexpected"); + // There is alloca()'s in this function, must reference off the frame + // pointer instead. + FrameReg = getFrameRegister(MF); + Offset -= AFI->getFramePtrSpillOffset(); + } + + unsigned Opcode = MI.getOpcode(); + const TargetInstrDesc &Desc = MI.getDesc(); + unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask); + + if (Opcode == ARM::tADDrSPi) { + Offset += MI.getOperand(i+1).getImm(); + + // Can't use tADDrSPi if it's based off the frame pointer. + unsigned NumBits = 0; + unsigned Scale = 1; + if (FrameReg != ARM::SP) { + Opcode = ARM::tADDi3; + MI.setDesc(TII.get(ARM::tADDi3)); + NumBits = 3; + } else { + NumBits = 8; + Scale = 4; + assert((Offset & 3) == 0 && + "Thumb add/sub sp, #imm immediate must be multiple of 4!"); + } + + if (Offset == 0) { + // Turn it into a move. + MI.setDesc(TII.get(ARM::tMOVhir2lor)); + MI.getOperand(i).ChangeToRegister(FrameReg, false); + MI.RemoveOperand(i+1); + return; + } + + // Common case: small offset, fits into instruction. + unsigned Mask = (1 << NumBits) - 1; + if (((Offset / Scale) & ~Mask) == 0) { + // Replace the FrameIndex with sp / fp + MI.getOperand(i).ChangeToRegister(FrameReg, false); + MI.getOperand(i+1).ChangeToImmediate(Offset / Scale); + return; + } + + unsigned DestReg = MI.getOperand(0).getReg(); + unsigned Bytes = (Offset > 0) ? Offset : -Offset; + unsigned NumMIs = calcNumMI(Opcode, 0, Bytes, NumBits, Scale); + // MI would expand into a large number of instructions. Don't try to + // simplify the immediate. + if (NumMIs > 2) { + emitThumbRegPlusImmediate(MBB, II, DestReg, FrameReg, Offset, TII, + *this, dl); + MBB.erase(II); + return; + } + + if (Offset > 0) { + // Translate r0 = add sp, imm to + // r0 = add sp, 255*4 + // r0 = add r0, (imm - 255*4) + MI.getOperand(i).ChangeToRegister(FrameReg, false); + MI.getOperand(i+1).ChangeToImmediate(Mask); + Offset = (Offset - Mask * Scale); + MachineBasicBlock::iterator NII = next(II); + emitThumbRegPlusImmediate(MBB, NII, DestReg, DestReg, Offset, TII, + *this, dl); + } else { + // Translate r0 = add sp, -imm to + // r0 = -imm (this is then translated into a series of instructons) + // r0 = add r0, sp + emitThumbConstant(MBB, II, DestReg, Offset, TII, *this, dl); + MI.setDesc(TII.get(ARM::tADDhirr)); + MI.getOperand(i).ChangeToRegister(DestReg, false, false, true); + MI.getOperand(i+1).ChangeToRegister(FrameReg, false); + } + return; + } else { + unsigned ImmIdx = 0; + int InstrOffs = 0; + unsigned NumBits = 0; + unsigned Scale = 1; + switch (AddrMode) { + case ARMII::AddrModeTs: { + ImmIdx = i+1; + InstrOffs = MI.getOperand(ImmIdx).getImm(); + NumBits = (FrameReg == ARM::SP) ? 8 : 5; + Scale = 4; + break; + } + default: + assert(0 && "Unsupported addressing mode!"); + abort(); + break; + } + + Offset += InstrOffs * Scale; + assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!"); + + // Common case: small offset, fits into instruction. + MachineOperand &ImmOp = MI.getOperand(ImmIdx); + int ImmedOffset = Offset / Scale; + unsigned Mask = (1 << NumBits) - 1; + if ((unsigned)Offset <= Mask * Scale) { + // Replace the FrameIndex with sp + MI.getOperand(i).ChangeToRegister(FrameReg, false); + ImmOp.ChangeToImmediate(ImmedOffset); + return; + } + + bool isThumSpillRestore = Opcode == ARM::tRestore || Opcode == ARM::tSpill; + if (AddrMode == ARMII::AddrModeTs) { + // Thumb tLDRspi, tSTRspi. These will change to instructions that use + // a different base register. + NumBits = 5; + Mask = (1 << NumBits) - 1; + } + // If this is a thumb spill / restore, we will be using a constpool load to + // materialize the offset. + if (AddrMode == ARMII::AddrModeTs && isThumSpillRestore) + ImmOp.ChangeToImmediate(0); + else { + // Otherwise, it didn't fit. Pull in what we can to simplify the immed. + ImmedOffset = ImmedOffset & Mask; + ImmOp.ChangeToImmediate(ImmedOffset); + Offset &= ~(Mask*Scale); + } + } + + // If we get here, the immediate doesn't fit into the instruction. We folded + // as much as possible above, handle the rest, providing a register that is + // SP+LargeImm. + assert(Offset && "This code isn't needed if offset already handled!"); + + if (Desc.mayLoad()) { + // Use the destination register to materialize sp + offset. + unsigned TmpReg = MI.getOperand(0).getReg(); + bool UseRR = false; + if (Opcode == ARM::tRestore) { + if (FrameReg == ARM::SP) + emitThumbRegPlusImmInReg(MBB, II, TmpReg, FrameReg, + Offset, false, TII, *this, dl); + else { + emitLoadConstPool(MBB, II, TmpReg, Offset, ARMCC::AL, 0, &TII, dl); + UseRR = true; + } + } else + emitThumbRegPlusImmediate(MBB, II, TmpReg, FrameReg, Offset, TII, + *this, dl); + MI.setDesc(TII.get(ARM::tLDR)); + MI.getOperand(i).ChangeToRegister(TmpReg, false, false, true); + if (UseRR) + // Use [reg, reg] addrmode. + MI.addOperand(MachineOperand::CreateReg(FrameReg, false)); + else // tLDR has an extra register operand. + MI.addOperand(MachineOperand::CreateReg(0, false)); + } else if (Desc.mayStore()) { + // FIXME! This is horrific!!! We need register scavenging. + // Our temporary workaround has marked r3 unavailable. Of course, r3 is + // also a ABI register so it's possible that is is the register that is + // being storing here. If that's the case, we do the following: + // r12 = r2 + // Use r2 to materialize sp + offset + // str r3, r2 + // r2 = r12 + unsigned ValReg = MI.getOperand(0).getReg(); + unsigned TmpReg = ARM::R3; + bool UseRR = false; + if (ValReg == ARM::R3) { + BuildMI(MBB, II, dl, TII.get(ARM::tMOVlor2hir), ARM::R12) + .addReg(ARM::R2, RegState::Kill); + TmpReg = ARM::R2; + } + if (TmpReg == ARM::R3 && AFI->isR3LiveIn()) + BuildMI(MBB, II, dl, TII.get(ARM::tMOVlor2hir), ARM::R12) + .addReg(ARM::R3, RegState::Kill); + if (Opcode == ARM::tSpill) { + if (FrameReg == ARM::SP) + emitThumbRegPlusImmInReg(MBB, II, TmpReg, FrameReg, + Offset, false, TII, *this, dl); + else { + emitLoadConstPool(MBB, II, TmpReg, Offset, ARMCC::AL, 0, &TII, dl); + UseRR = true; + } + } else + emitThumbRegPlusImmediate(MBB, II, TmpReg, FrameReg, Offset, TII, + *this, dl); + MI.setDesc(TII.get(ARM::tSTR)); + MI.getOperand(i).ChangeToRegister(TmpReg, false, false, true); + if (UseRR) // Use [reg, reg] addrmode. + MI.addOperand(MachineOperand::CreateReg(FrameReg, false)); + else // tSTR has an extra register operand. + MI.addOperand(MachineOperand::CreateReg(0, false)); + + MachineBasicBlock::iterator NII = next(II); + if (ValReg == ARM::R3) + BuildMI(MBB, NII, dl, TII.get(ARM::tMOVhir2lor), ARM::R2) + .addReg(ARM::R12, RegState::Kill); + if (TmpReg == ARM::R3 && AFI->isR3LiveIn()) + BuildMI(MBB, NII, dl, TII.get(ARM::tMOVhir2lor), ARM::R3) + .addReg(ARM::R12, RegState::Kill); + } else + assert(false && "Unexpected opcode!"); +} + +void ThumbRegisterInfo::emitPrologue(MachineFunction &MF) const { + MachineBasicBlock &MBB = MF.front(); + MachineBasicBlock::iterator MBBI = MBB.begin(); + MachineFrameInfo *MFI = MF.getFrameInfo(); + ARMFunctionInfo *AFI = MF.getInfo(); + unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize(); + unsigned NumBytes = MFI->getStackSize(); + const std::vector &CSI = MFI->getCalleeSavedInfo(); + DebugLoc dl = (MBBI != MBB.end() ? + MBBI->getDebugLoc() : DebugLoc::getUnknownLoc()); + + // Check if R3 is live in. It might have to be used as a scratch register. + for (MachineRegisterInfo::livein_iterator I =MF.getRegInfo().livein_begin(), + E = MF.getRegInfo().livein_end(); I != E; ++I) { + if (I->first == ARM::R3) { + AFI->setR3IsLiveIn(true); + break; + } + } + + // Thumb add/sub sp, imm8 instructions implicitly multiply the offset by 4. + NumBytes = (NumBytes + 3) & ~3; + MFI->setStackSize(NumBytes); + + // Determine the sizes of each callee-save spill areas and record which frame + // belongs to which callee-save spill areas. + unsigned GPRCS1Size = 0, GPRCS2Size = 0, DPRCSSize = 0; + int FramePtrSpillFI = 0; + + if (VARegSaveSize) + emitSPUpdate(MBB, MBBI, -VARegSaveSize, ARMCC::AL, 0, TII, dl); + + if (!AFI->hasStackFrame()) { + if (NumBytes != 0) + emitSPUpdate(MBB, MBBI, -NumBytes, ARMCC::AL, 0, TII, dl); + return; + } + + for (unsigned i = 0, e = CSI.size(); i != e; ++i) { + unsigned Reg = CSI[i].getReg(); + int FI = CSI[i].getFrameIdx(); + switch (Reg) { + case ARM::R4: + case ARM::R5: + case ARM::R6: + case ARM::R7: + case ARM::LR: + if (Reg == FramePtr) + FramePtrSpillFI = FI; + AFI->addGPRCalleeSavedArea1Frame(FI); + GPRCS1Size += 4; + break; + case ARM::R8: + case ARM::R9: + case ARM::R10: + case ARM::R11: + if (Reg == FramePtr) + FramePtrSpillFI = FI; + if (STI.isTargetDarwin()) { + AFI->addGPRCalleeSavedArea2Frame(FI); + GPRCS2Size += 4; + } else { + AFI->addGPRCalleeSavedArea1Frame(FI); + GPRCS1Size += 4; + } + break; + default: + AFI->addDPRCalleeSavedAreaFrame(FI); + DPRCSSize += 8; + } + } + + if (MBBI != MBB.end() && MBBI->getOpcode() == ARM::tPUSH) { + ++MBBI; + if (MBBI != MBB.end()) + dl = MBBI->getDebugLoc(); + } + + // Darwin ABI requires FP to point to the stack slot that contains the + // previous FP. + if (STI.isTargetDarwin() || hasFP(MF)) { + MachineInstrBuilder MIB = + BuildMI(MBB, MBBI, dl, TII.get(ARM::tADDrSPi), FramePtr) + .addFrameIndex(FramePtrSpillFI).addImm(0); + } + + // Determine starting offsets of spill areas. + unsigned DPRCSOffset = NumBytes - (GPRCS1Size + GPRCS2Size + DPRCSSize); + unsigned GPRCS2Offset = DPRCSOffset + DPRCSSize; + unsigned GPRCS1Offset = GPRCS2Offset + GPRCS2Size; + AFI->setFramePtrSpillOffset(MFI->getObjectOffset(FramePtrSpillFI) + NumBytes); + AFI->setGPRCalleeSavedArea1Offset(GPRCS1Offset); + AFI->setGPRCalleeSavedArea2Offset(GPRCS2Offset); + AFI->setDPRCalleeSavedAreaOffset(DPRCSOffset); + + NumBytes = DPRCSOffset; + if (NumBytes) { + // Insert it after all the callee-save spills. + emitSPUpdate(MBB, MBBI, -NumBytes, ARMCC::AL, 0, TII, dl); + } + + if (STI.isTargetELF() && hasFP(MF)) { + MFI->setOffsetAdjustment(MFI->getOffsetAdjustment() - + AFI->getFramePtrSpillOffset()); + } + + AFI->setGPRCalleeSavedArea1Size(GPRCS1Size); + AFI->setGPRCalleeSavedArea2Size(GPRCS2Size); + AFI->setDPRCalleeSavedAreaSize(DPRCSSize); +} + +static bool isCalleeSavedRegister(unsigned Reg, const unsigned *CSRegs) { + for (unsigned i = 0; CSRegs[i]; ++i) + if (Reg == CSRegs[i]) + return true; + return false; +} + +static bool isCSRestore(MachineInstr *MI, const unsigned *CSRegs) { + return (MI->getOpcode() == ARM::tRestore && + MI->getOperand(1).isFI() && + isCalleeSavedRegister(MI->getOperand(0).getReg(), CSRegs)); +} + +void ThumbRegisterInfo::emitEpilogue(MachineFunction &MF, + MachineBasicBlock &MBB) const { + MachineBasicBlock::iterator MBBI = prior(MBB.end()); + assert((MBBI->getOpcode() == ARM::tBX_RET || + MBBI->getOpcode() == ARM::tPOP_RET) && + "Can only insert epilog into returning blocks"); + DebugLoc dl = MBBI->getDebugLoc(); + MachineFrameInfo *MFI = MF.getFrameInfo(); + ARMFunctionInfo *AFI = MF.getInfo(); + unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize(); + int NumBytes = (int)MFI->getStackSize(); + + if (!AFI->hasStackFrame()) { + if (NumBytes != 0) + emitSPUpdate(MBB, MBBI, NumBytes, ARMCC::AL, 0, TII, dl); + } else { + // Unwind MBBI to point to first LDR / FLDD. + const unsigned *CSRegs = getCalleeSavedRegs(); + if (MBBI != MBB.begin()) { + do + --MBBI; + while (MBBI != MBB.begin() && isCSRestore(MBBI, CSRegs)); + if (!isCSRestore(MBBI, CSRegs)) + ++MBBI; + } + + // Move SP to start of FP callee save spill area. + NumBytes -= (AFI->getGPRCalleeSavedArea1Size() + + AFI->getGPRCalleeSavedArea2Size() + + AFI->getDPRCalleeSavedAreaSize()); + + if (hasFP(MF)) { + NumBytes = AFI->getFramePtrSpillOffset() - NumBytes; + // Reset SP based on frame pointer only if the stack frame extends beyond + // frame pointer stack slot or target is ELF and the function has FP. + if (NumBytes) + emitThumbRegPlusImmediate(MBB, MBBI, ARM::SP, FramePtr, -NumBytes, + TII, *this, dl); + else + BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVlor2hir), ARM::SP) + .addReg(FramePtr); + } else { + if (MBBI->getOpcode() == ARM::tBX_RET && + &MBB.front() != MBBI && + prior(MBBI)->getOpcode() == ARM::tPOP) { + MachineBasicBlock::iterator PMBBI = prior(MBBI); + emitSPUpdate(MBB, PMBBI, NumBytes, ARMCC::AL, 0, TII, dl); + } else + emitSPUpdate(MBB, MBBI, NumBytes, ARMCC::AL, 0, TII, dl); + } + } + + if (VARegSaveSize) { + // Epilogue for vararg functions: pop LR to R3 and branch off it. + // FIXME: Verify this is still ok when R3 is no longer being reserved. + BuildMI(MBB, MBBI, dl, TII.get(ARM::tPOP)).addReg(ARM::R3); + + emitSPUpdate(MBB, MBBI, VARegSaveSize, ARMCC::AL, 0, TII, dl); + + BuildMI(MBB, MBBI, dl, TII.get(ARM::tBX_RET_vararg)).addReg(ARM::R3); + MBB.erase(MBBI); + } +} diff --git a/lib/Target/ARM/ThumbRegisterInfo.h b/lib/Target/ARM/ThumbRegisterInfo.h new file mode 100644 index 00000000000..b1adc318583 --- /dev/null +++ b/lib/Target/ARM/ThumbRegisterInfo.h @@ -0,0 +1,65 @@ +//===- ThumbRegisterInfo.h - Thumb Register Information Impl ----*- C++ -*-===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// This file contains the ARM implementation of the TargetRegisterInfo class. +// +//===----------------------------------------------------------------------===// + +#ifndef THUMBREGISTERINFO_H +#define THUMBREGISTERINFO_H + +#include "ARM.h" +#include "ARMRegisterInfo.h" +#include "llvm/Target/TargetRegisterInfo.h" + +namespace llvm { + class ARMSubtarget; + class TargetInstrInfo; + class Type; + +struct ThumbRegisterInfo : public ARMBaseRegisterInfo { +public: + ThumbRegisterInfo(const TargetInstrInfo &tii, const ARMSubtarget &STI); + + /// emitLoadConstPool - Emits a load from constpool to materialize the + /// specified immediate. + void emitLoadConstPool(MachineBasicBlock &MBB, + MachineBasicBlock::iterator &MBBI, + unsigned DestReg, int Val, + unsigned Pred, unsigned PredReg, + const TargetInstrInfo *TII, + DebugLoc dl) const; + + /// Code Generation virtual methods... + const TargetRegisterClass * + getPhysicalRegisterRegClass(unsigned Reg, MVT VT = MVT::Other) const; + + bool isReservedReg(const MachineFunction &MF, unsigned Reg) const; + + bool requiresRegisterScavenging(const MachineFunction &MF) const; + + bool hasReservedCallFrame(MachineFunction &MF) const; + + void eliminateCallFramePseudoInstr(MachineFunction &MF, + MachineBasicBlock &MBB, + MachineBasicBlock::iterator I) const; + + void eliminateFrameIndex(MachineBasicBlock::iterator II, + int SPAdj, RegScavenger *RS = NULL) const; + + void emitPrologue(MachineFunction &MF) const; + void emitEpilogue(MachineFunction &MF, MachineBasicBlock &MBB) const; + + void emitSPUpdate(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI, + int NumBytes, ARMCC::CondCodes Pred, unsigned PredReg, + const TargetInstrInfo &TII, DebugLoc dl) const; +}; +} + +#endif // THUMBREGISTERINFO_H