llvm/lib/Target/ARM/ARMInstructionSelector.cpp

//===- ARMInstructionSelector.cpp ----------------------------*- C++ -*-==//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
/// \file
/// This file implements the targeting of the InstructionSelector class for ARM.
/// \todo This should be generated by TableGen.
//===----------------------------------------------------------------------===//

#include "ARMRegisterBankInfo.h"
#include "ARMSubtarget.h"
#include "ARMTargetMachine.h"
#include "llvm/CodeGen/GlobalISel/InstructionSelector.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/Support/Debug.h"

#define DEBUG_TYPE "arm-isel"

using namespace llvm;

#ifndef LLVM_BUILD_GLOBAL_ISEL
#error "You shouldn't build this"
#endif

namespace {

#define GET_GLOBALISEL_PREDICATE_BITSET
#include "ARMGenGlobalISel.inc"
#undef GET_GLOBALISEL_PREDICATE_BITSET

class ARMInstructionSelector : public InstructionSelector {
public:
  ARMInstructionSelector(const ARMBaseTargetMachine &TM, const ARMSubtarget &STI,
                         const ARMRegisterBankInfo &RBI);

  bool select(MachineInstr &I) const override;

private:
  bool selectImpl(MachineInstr &I) const;

  bool selectICmp(MachineInstrBuilder &MIB, const ARMBaseInstrInfo &TII,
                  MachineRegisterInfo &MRI, const TargetRegisterInfo &TRI,
                  const RegisterBankInfo &RBI) const;

  const ARMBaseInstrInfo &TII;
  const ARMBaseRegisterInfo &TRI;
  const ARMBaseTargetMachine &TM;
  const ARMRegisterBankInfo &RBI;
  const ARMSubtarget &STI;

#define GET_GLOBALISEL_PREDICATES_DECL
#include "ARMGenGlobalISel.inc"
#undef GET_GLOBALISEL_PREDICATES_DECL

// We declare the temporaries used by selectImpl() in the class to minimize the
// cost of constructing placeholder values.
#define GET_GLOBALISEL_TEMPORARIES_DECL
#include "ARMGenGlobalISel.inc"
#undef GET_GLOBALISEL_TEMPORARIES_DECL
};
} // end anonymous namespace

namespace llvm {
InstructionSelector *
createARMInstructionSelector(const ARMBaseTargetMachine &TM,
                             const ARMSubtarget &STI,
                             const ARMRegisterBankInfo &RBI) {
  return new ARMInstructionSelector(TM, STI, RBI);
}
}

unsigned zero_reg = 0;

#define GET_GLOBALISEL_IMPL
#include "ARMGenGlobalISel.inc"
#undef GET_GLOBALISEL_IMPL

ARMInstructionSelector::ARMInstructionSelector(const ARMBaseTargetMachine &TM,
                                               const ARMSubtarget &STI,
                                               const ARMRegisterBankInfo &RBI)
    : InstructionSelector(), TII(*STI.getInstrInfo()),
      TRI(*STI.getRegisterInfo()), TM(TM), RBI(RBI), STI(STI),
#define GET_GLOBALISEL_PREDICATES_INIT
#include "ARMGenGlobalISel.inc"
#undef GET_GLOBALISEL_PREDICATES_INIT
#define GET_GLOBALISEL_TEMPORARIES_INIT
#include "ARMGenGlobalISel.inc"
#undef GET_GLOBALISEL_TEMPORARIES_INIT
{
}

static bool selectCopy(MachineInstr &I, const TargetInstrInfo &TII,
                       MachineRegisterInfo &MRI, const TargetRegisterInfo &TRI,
                       const RegisterBankInfo &RBI) {
  unsigned DstReg = I.getOperand(0).getReg();
  if (TargetRegisterInfo::isPhysicalRegister(DstReg))
    return true;

  const RegisterBank *RegBank = RBI.getRegBank(DstReg, MRI, TRI);
  (void)RegBank;
  assert(RegBank && "Can't get reg bank for virtual register");

  const unsigned DstSize = MRI.getType(DstReg).getSizeInBits();
  assert((RegBank->getID() == ARM::GPRRegBankID ||
          RegBank->getID() == ARM::FPRRegBankID) &&
         "Unsupported reg bank");

  const TargetRegisterClass *RC = &ARM::GPRRegClass;

  if (RegBank->getID() == ARM::FPRRegBankID) {
    if (DstSize == 32)
      RC = &ARM::SPRRegClass;
    else if (DstSize == 64)
      RC = &ARM::DPRRegClass;
    else
      llvm_unreachable("Unsupported destination size");
  }

  // No need to constrain SrcReg. It will get constrained when
  // we hit another of its uses or its defs.
  // Copies do not have constraints.
  if (!RBI.constrainGenericRegister(DstReg, *RC, MRI)) {
    DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode())
                 << " operand\n");
    return false;
  }
  return true;
}

static bool selectMergeValues(MachineInstrBuilder &MIB,
                              const ARMBaseInstrInfo &TII,
                              MachineRegisterInfo &MRI,
                              const TargetRegisterInfo &TRI,
                              const RegisterBankInfo &RBI) {
  assert(TII.getSubtarget().hasVFP2() && "Can't select merge without VFP");

  // We only support G_MERGE_VALUES as a way to stick together two scalar GPRs
  // into one DPR.
  unsigned VReg0 = MIB->getOperand(0).getReg();
  (void)VReg0;
  assert(MRI.getType(VReg0).getSizeInBits() == 64 &&
         RBI.getRegBank(VReg0, MRI, TRI)->getID() == ARM::FPRRegBankID &&
         "Unsupported operand for G_MERGE_VALUES");
  unsigned VReg1 = MIB->getOperand(1).getReg();
  (void)VReg1;
  assert(MRI.getType(VReg1).getSizeInBits() == 32 &&
         RBI.getRegBank(VReg1, MRI, TRI)->getID() == ARM::GPRRegBankID &&
         "Unsupported operand for G_MERGE_VALUES");
  unsigned VReg2 = MIB->getOperand(2).getReg();
  (void)VReg2;
  assert(MRI.getType(VReg2).getSizeInBits() == 32 &&
         RBI.getRegBank(VReg2, MRI, TRI)->getID() == ARM::GPRRegBankID &&
         "Unsupported operand for G_MERGE_VALUES");

  MIB->setDesc(TII.get(ARM::VMOVDRR));
  MIB.add(predOps(ARMCC::AL));

  return true;
}

static bool selectUnmergeValues(MachineInstrBuilder &MIB,
                                const ARMBaseInstrInfo &TII,
                                MachineRegisterInfo &MRI,
                                const TargetRegisterInfo &TRI,
                                const RegisterBankInfo &RBI) {
  assert(TII.getSubtarget().hasVFP2() && "Can't select unmerge without VFP");

  // We only support G_UNMERGE_VALUES as a way to break up one DPR into two
  // GPRs.
  unsigned VReg0 = MIB->getOperand(0).getReg();
  (void)VReg0;
  assert(MRI.getType(VReg0).getSizeInBits() == 32 &&
         RBI.getRegBank(VReg0, MRI, TRI)->getID() == ARM::GPRRegBankID &&
         "Unsupported operand for G_UNMERGE_VALUES");
  unsigned VReg1 = MIB->getOperand(1).getReg();
  (void)VReg1;
  assert(MRI.getType(VReg1).getSizeInBits() == 32 &&
         RBI.getRegBank(VReg1, MRI, TRI)->getID() == ARM::GPRRegBankID &&
         "Unsupported operand for G_UNMERGE_VALUES");
  unsigned VReg2 = MIB->getOperand(2).getReg();
  (void)VReg2;
  assert(MRI.getType(VReg2).getSizeInBits() == 64 &&
         RBI.getRegBank(VReg2, MRI, TRI)->getID() == ARM::FPRRegBankID &&
         "Unsupported operand for G_UNMERGE_VALUES");

  MIB->setDesc(TII.get(ARM::VMOVRRD));
  MIB.add(predOps(ARMCC::AL));

  return true;
}

/// Select the opcode for simple extensions (that translate to a single SXT/UXT
/// instruction). Extension operations more complicated than that should not
/// invoke this. Returns the original opcode if it doesn't know how to select a
/// better one.
static unsigned selectSimpleExtOpc(unsigned Opc, unsigned Size) {
  using namespace TargetOpcode;

  if (Size != 8 && Size != 16)
    return Opc;

  if (Opc == G_SEXT)
    return Size == 8 ? ARM::SXTB : ARM::SXTH;

  if (Opc == G_ZEXT)
    return Size == 8 ? ARM::UXTB : ARM::UXTH;

  return Opc;
}

/// Select the opcode for simple loads and stores. For types smaller than 32
/// bits, the value will be zero extended. Returns the original opcode if it
/// doesn't know how to select a better one.
static unsigned selectLoadStoreOpCode(unsigned Opc, unsigned RegBank,
                                      unsigned Size) {
  bool isStore = Opc == TargetOpcode::G_STORE;

  if (RegBank == ARM::GPRRegBankID) {
    switch (Size) {
    case 1:
    case 8:
      return isStore ? ARM::STRBi12 : ARM::LDRBi12;
    case 16:
      return isStore ? ARM::STRH : ARM::LDRH;
    case 32:
      return isStore ? ARM::STRi12 : ARM::LDRi12;
    default:
      return Opc;
    }
  }

  if (RegBank == ARM::FPRRegBankID) {
    switch (Size) {
    case 32:
      return isStore ? ARM::VSTRS : ARM::VLDRS;
    case 64:
      return isStore ? ARM::VSTRD : ARM::VLDRD;
    default:
      return Opc;
    }
  }

  return Opc;
}

static ARMCC::CondCodes getComparePred(CmpInst::Predicate Pred) {
  switch (Pred) {
  // Needs two compares...
  case CmpInst::FCMP_ONE:
  case CmpInst::FCMP_UEQ:
  default:
    // AL is our "false" for now. The other two need more compares.
    return ARMCC::AL;
  case CmpInst::ICMP_EQ:
  case CmpInst::FCMP_OEQ:
    return ARMCC::EQ;
  case CmpInst::ICMP_SGT:
  case CmpInst::FCMP_OGT:
    return ARMCC::GT;
  case CmpInst::ICMP_SGE:
  case CmpInst::FCMP_OGE:
    return ARMCC::GE;
  case CmpInst::ICMP_UGT:
  case CmpInst::FCMP_UGT:
    return ARMCC::HI;
  case CmpInst::FCMP_OLT:
    return ARMCC::MI;
  case CmpInst::ICMP_ULE:
  case CmpInst::FCMP_OLE:
    return ARMCC::LS;
  case CmpInst::FCMP_ORD:
    return ARMCC::VC;
  case CmpInst::FCMP_UNO:
    return ARMCC::VS;
  case CmpInst::FCMP_UGE:
    return ARMCC::PL;
  case CmpInst::ICMP_SLT:
  case CmpInst::FCMP_ULT:
    return ARMCC::LT;
  case CmpInst::ICMP_SLE:
  case CmpInst::FCMP_ULE:
    return ARMCC::LE;
  case CmpInst::FCMP_UNE:
  case CmpInst::ICMP_NE:
    return ARMCC::NE;
  case CmpInst::ICMP_UGE:
    return ARMCC::HS;
  case CmpInst::ICMP_ULT:
    return ARMCC::LO;
  }
}

bool ARMInstructionSelector::selectICmp(MachineInstrBuilder &MIB,
                                        const ARMBaseInstrInfo &TII,
                                        MachineRegisterInfo &MRI,
                                        const TargetRegisterInfo &TRI,
                                        const RegisterBankInfo &RBI) const {
  auto &MBB = *MIB->getParent();
  auto InsertBefore = std::next(MIB->getIterator());
  auto &DebugLoc = MIB->getDebugLoc();

  // Move 0 into the result register.
  auto Mov0I = BuildMI(MBB, InsertBefore, DebugLoc, TII.get(ARM::MOVi))
                   .addDef(MRI.createVirtualRegister(&ARM::GPRRegClass))
                   .addImm(0)
                   .add(predOps(ARMCC::AL))
                   .add(condCodeOp());
  if (!constrainSelectedInstRegOperands(*Mov0I, TII, TRI, RBI))
    return false;

  // Perform the comparison.
  auto LHSReg = MIB->getOperand(2).getReg();
  auto RHSReg = MIB->getOperand(3).getReg();
  assert(MRI.getType(LHSReg) == MRI.getType(RHSReg) &&
         MRI.getType(LHSReg).getSizeInBits() == 32 &&
         MRI.getType(RHSReg).getSizeInBits() == 32 &&
         "Unsupported types for comparison operation");
  auto CmpI = BuildMI(MBB, InsertBefore, DebugLoc, TII.get(ARM::CMPrr))
                  .addUse(LHSReg)
                  .addUse(RHSReg)
                  .add(predOps(ARMCC::AL));
  if (!constrainSelectedInstRegOperands(*CmpI, TII, TRI, RBI))
    return false;

  // Move 1 into the result register if the flags say so.
  auto ResReg = MIB->getOperand(0).getReg();
  auto Cond =
      static_cast<CmpInst::Predicate>(MIB->getOperand(1).getPredicate());
  auto ARMCond = getComparePred(Cond);
  if (ARMCond == ARMCC::AL)
    return false;

  auto Mov1I = BuildMI(MBB, InsertBefore, DebugLoc, TII.get(ARM::MOVCCi))
                   .addDef(ResReg)
                   .addUse(Mov0I->getOperand(0).getReg())
                   .addImm(1)
                   .add(predOps(ARMCond, ARM::CPSR));
  if (!constrainSelectedInstRegOperands(*Mov1I, TII, TRI, RBI))
    return false;

  MIB->eraseFromParent();
  return true;
}

bool ARMInstructionSelector::select(MachineInstr &I) const {
  assert(I.getParent() && "Instruction should be in a basic block!");
  assert(I.getParent()->getParent() && "Instruction should be in a function!");

  auto &MBB = *I.getParent();
  auto &MF = *MBB.getParent();
  auto &MRI = MF.getRegInfo();

  if (!isPreISelGenericOpcode(I.getOpcode())) {
    if (I.isCopy())
      return selectCopy(I, TII, MRI, TRI, RBI);

    return true;
  }

  if (selectImpl(I))
    return true;

  MachineInstrBuilder MIB{MF, I};
  bool isSExt = false;

  using namespace TargetOpcode;
  switch (I.getOpcode()) {
  case G_SEXT:
    isSExt = true;
    LLVM_FALLTHROUGH;
  case G_ZEXT: {
    LLT DstTy = MRI.getType(I.getOperand(0).getReg());
    // FIXME: Smaller destination sizes coming soon!
    if (DstTy.getSizeInBits() != 32) {
      DEBUG(dbgs() << "Unsupported destination size for extension");
      return false;
    }

    LLT SrcTy = MRI.getType(I.getOperand(1).getReg());
    unsigned SrcSize = SrcTy.getSizeInBits();
    switch (SrcSize) {
    case 1: {
      // ZExt boils down to & 0x1; for SExt we also subtract that from 0
      I.setDesc(TII.get(ARM::ANDri));
      MIB.addImm(1).add(predOps(ARMCC::AL)).add(condCodeOp());

      if (isSExt) {
        unsigned SExtResult = I.getOperand(0).getReg();

        // Use a new virtual register for the result of the AND
        unsigned AndResult = MRI.createVirtualRegister(&ARM::GPRRegClass);
        I.getOperand(0).setReg(AndResult);

        auto InsertBefore = std::next(I.getIterator());
        auto SubI =
            BuildMI(MBB, InsertBefore, I.getDebugLoc(), TII.get(ARM::RSBri))
                .addDef(SExtResult)
                .addUse(AndResult)
                .addImm(0)
                .add(predOps(ARMCC::AL))
                .add(condCodeOp());
        if (!constrainSelectedInstRegOperands(*SubI, TII, TRI, RBI))
          return false;
      }
      break;
    }
    case 8:
    case 16: {
      unsigned NewOpc = selectSimpleExtOpc(I.getOpcode(), SrcSize);
      if (NewOpc == I.getOpcode())
        return false;
      I.setDesc(TII.get(NewOpc));
      MIB.addImm(0).add(predOps(ARMCC::AL));
      break;
    }
    default:
      DEBUG(dbgs() << "Unsupported source size for extension");
      return false;
    }
    break;
  }
  case G_ANYEXT:
  case G_TRUNC: {
    // The high bits are undefined, so there's nothing special to do, just
    // treat it as a copy.
    auto SrcReg = I.getOperand(1).getReg();
    auto DstReg = I.getOperand(0).getReg();

    const auto &SrcRegBank = *RBI.getRegBank(SrcReg, MRI, TRI);
    const auto &DstRegBank = *RBI.getRegBank(DstReg, MRI, TRI);

    if (SrcRegBank.getID() != DstRegBank.getID()) {
      DEBUG(dbgs() << "G_TRUNC/G_ANYEXT operands on different register banks\n");
      return false;
    }

    if (SrcRegBank.getID() != ARM::GPRRegBankID) {
      DEBUG(dbgs() << "G_TRUNC/G_ANYEXT on non-GPR not supported yet\n");
      return false;
    }

    I.setDesc(TII.get(COPY));
    return selectCopy(I, TII, MRI, TRI, RBI);
  }
  case G_ICMP:
    return selectICmp(MIB, TII, MRI, TRI, RBI);
  case G_GEP:
    I.setDesc(TII.get(ARM::ADDrr));
    MIB.add(predOps(ARMCC::AL)).add(condCodeOp());
    break;
  case G_FRAME_INDEX:
    // Add 0 to the given frame index and hope it will eventually be folded into
    // the user(s).
    I.setDesc(TII.get(ARM::ADDri));
    MIB.addImm(0).add(predOps(ARMCC::AL)).add(condCodeOp());
    break;
  case G_CONSTANT: {
    unsigned Reg = I.getOperand(0).getReg();
    if (MRI.getType(Reg).getSizeInBits() != 32)
      return false;

    assert(RBI.getRegBank(Reg, MRI, TRI)->getID() == ARM::GPRRegBankID &&
           "Expected constant to live in a GPR");
    I.setDesc(TII.get(ARM::MOVi));
    MIB.add(predOps(ARMCC::AL)).add(condCodeOp());

    auto &Val = I.getOperand(1);
    if (Val.isCImm()) {
      if (Val.getCImm()->getBitWidth() > 32)
        return false;
      Val.ChangeToImmediate(Val.getCImm()->getZExtValue());
    }

    if (!Val.isImm()) {
      return false;
    }

    break;
  }
  case G_STORE:
  case G_LOAD: {
    const auto &MemOp = **I.memoperands_begin();
    if (MemOp.getOrdering() != AtomicOrdering::NotAtomic) {
      DEBUG(dbgs() << "Atomic load/store not supported yet\n");
      return false;
    }

    unsigned Reg = I.getOperand(0).getReg();
    unsigned RegBank = RBI.getRegBank(Reg, MRI, TRI)->getID();

    LLT ValTy = MRI.getType(Reg);
    const auto ValSize = ValTy.getSizeInBits();

    assert((ValSize != 64 || TII.getSubtarget().hasVFP2()) &&
           "Don't know how to load/store 64-bit value without VFP");

    const auto NewOpc = selectLoadStoreOpCode(I.getOpcode(), RegBank, ValSize);
    if (NewOpc == G_LOAD || NewOpc == G_STORE)
      return false;

    I.setDesc(TII.get(NewOpc));

    if (NewOpc == ARM::LDRH || NewOpc == ARM::STRH)
      // LDRH has a funny addressing mode (there's already a FIXME for it).
      MIB.addReg(0);
    MIB.addImm(0).add(predOps(ARMCC::AL));
    break;
  }
  case G_MERGE_VALUES: {
    if (!selectMergeValues(MIB, TII, MRI, TRI, RBI))
      return false;
    break;
  }
  case G_UNMERGE_VALUES: {
    if (!selectUnmergeValues(MIB, TII, MRI, TRI, RBI))
      return false;
    break;
  }
  default:
    return false;
  }

  return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
}