lib/Target/PowerPC/PPCISelDAGToDAG.cpp

//===-- PPCISelDAGToDAG.cpp - PPC --pattern matching inst selector --------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file was developed by Chris Lattner and is distributed under
// the University of Illinois Open Source License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file defines a pattern matching instruction selector for PowerPC,
// converting from a legalized dag to a PPC dag.
//
//===----------------------------------------------------------------------===//

#include "PPC.h"
#include "PPCTargetMachine.h"
#include "PPCISelLowering.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/SSARegMap.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/CodeGen/SelectionDAGISel.h"
#include "llvm/Target/TargetOptions.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Constants.h"
#include "llvm/GlobalValue.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/MathExtras.h"
using namespace llvm;

namespace {
  Statistic<> FusedFP ("ppc-codegen", "Number of fused fp operations");
  Statistic<> FrameOff("ppc-codegen", "Number of frame idx offsets collapsed");
    
  //===--------------------------------------------------------------------===//
  /// PPCDAGToDAGISel - PPC specific code to select PPC machine
  /// instructions for SelectionDAG operations.
  ///
  class PPCDAGToDAGISel : public SelectionDAGISel {
    PPCTargetLowering PPCLowering;
    unsigned GlobalBaseReg;
  public:
    PPCDAGToDAGISel(TargetMachine &TM)
      : SelectionDAGISel(PPCLowering), PPCLowering(TM) {}
    
    virtual bool runOnFunction(Function &Fn) {
      // Make sure we re-emit a set of the global base reg if necessary
      GlobalBaseReg = 0;
      return SelectionDAGISel::runOnFunction(Fn);
    }
   
    /// getI32Imm - Return a target constant with the specified value, of type
    /// i32.
    inline SDOperand getI32Imm(unsigned Imm) {
      return CurDAG->getTargetConstant(Imm, MVT::i32);
    }

    /// getGlobalBaseReg - insert code into the entry mbb to materialize the PIC
    /// base register.  Return the virtual register that holds this value.
    SDOperand getGlobalBaseReg();
    
    // Select - Convert the specified operand from a target-independent to a
    // target-specific node if it hasn't already been changed.
    SDOperand Select(SDOperand Op);
    
    SDNode *SelectBitfieldInsert(SDNode *N);

    /// SelectCC - Select a comparison of the specified values with the
    /// specified condition code, returning the CR# of the expression.
    SDOperand SelectCC(SDOperand LHS, SDOperand RHS, ISD::CondCode CC);

    /// SelectAddr - Given the specified address, return the two operands for a
    /// load/store instruction, and return true if it should be an indexed [r+r]
    /// operation.
    bool SelectAddr(SDOperand Addr, SDOperand &Op1, SDOperand &Op2);
    
    /// SelectAddrIndexed - Given the specified addressed, force it to be
    /// represented as an indexed [r+r] operation, rather than possibly
    /// returning [r+imm] as SelectAddr may.
    void SelectAddrIndexed(SDOperand Addr, SDOperand &Op1, SDOperand &Op2);

    SDOperand BuildSDIVSequence(SDNode *N);
    SDOperand BuildUDIVSequence(SDNode *N);
    
    /// InstructionSelectBasicBlock - This callback is invoked by
    /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
    virtual void InstructionSelectBasicBlock(SelectionDAG &DAG);
    
    virtual const char *getPassName() const {
      return "PowerPC DAG->DAG Pattern Instruction Selection";
    } 

// Include the pieces autogenerated from the target description.
#include "PPCGenDAGISel.inc"
    
private:
    SDOperand SelectDYNAMIC_STACKALLOC(SDOperand Op);
    SDOperand SelectADD_PARTS(SDOperand Op);
    SDOperand SelectSUB_PARTS(SDOperand Op);
    SDOperand SelectSETCC(SDOperand Op);
    SDOperand SelectCALL(SDOperand Op);
  };
}

/// InstructionSelectBasicBlock - This callback is invoked by
/// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
void PPCDAGToDAGISel::InstructionSelectBasicBlock(SelectionDAG &DAG) {
  DEBUG(BB->dump());
  
  // The selection process is inherently a bottom-up recursive process (users
  // select their uses before themselves).  Given infinite stack space, we
  // could just start selecting on the root and traverse the whole graph.  In
  // practice however, this causes us to run out of stack space on large basic
  // blocks.  To avoid this problem, select the entry node, then all its uses,
  // iteratively instead of recursively.
  std::vector<SDOperand> Worklist;
  Worklist.push_back(DAG.getEntryNode());
  
  // Note that we can do this in the PPC target (scanning forward across token
  // chain edges) because no nodes ever get folded across these edges.  On a
  // target like X86 which supports load/modify/store operations, this would
  // have to be more careful.
  while (!Worklist.empty()) {
    SDOperand Node = Worklist.back();
    Worklist.pop_back();
    
    // Chose from the least deep of the top two nodes.
    if (!Worklist.empty() &&
        Worklist.back().Val->getNodeDepth() < Node.Val->getNodeDepth())
      std::swap(Worklist.back(), Node);
    
    if ((Node.Val->getOpcode() >= ISD::BUILTIN_OP_END &&
         Node.Val->getOpcode() < PPCISD::FIRST_NUMBER) ||
        CodeGenMap.count(Node)) continue;
    
    for (SDNode::use_iterator UI = Node.Val->use_begin(),
         E = Node.Val->use_end(); UI != E; ++UI) {
      // Scan the values.  If this use has a value that is a token chain, add it
      // to the worklist.
      SDNode *User = *UI;
      for (unsigned i = 0, e = User->getNumValues(); i != e; ++i)
        if (User->getValueType(i) == MVT::Other) {
          Worklist.push_back(SDOperand(User, i));
          break; 
        }
    }

    // Finally, legalize this node.
    Select(Node);
  }
    
  // Select target instructions for the DAG.
  DAG.setRoot(Select(DAG.getRoot()));
  CodeGenMap.clear();
  DAG.RemoveDeadNodes();
  
  // Emit machine code to BB. 
  ScheduleAndEmitDAG(DAG);
}

/// getGlobalBaseReg - Output the instructions required to put the
/// base address to use for accessing globals into a register.
///
SDOperand PPCDAGToDAGISel::getGlobalBaseReg() {
  if (!GlobalBaseReg) {
    // Insert the set of GlobalBaseReg into the first MBB of the function
    MachineBasicBlock &FirstMBB = BB->getParent()->front();
    MachineBasicBlock::iterator MBBI = FirstMBB.begin();
    SSARegMap *RegMap = BB->getParent()->getSSARegMap();
    // FIXME: when we get to LP64, we will need to create the appropriate
    // type of register here.
    GlobalBaseReg = RegMap->createVirtualRegister(PPC::GPRCRegisterClass);
    BuildMI(FirstMBB, MBBI, PPC::MovePCtoLR, 0, PPC::LR);
    BuildMI(FirstMBB, MBBI, PPC::MFLR, 1, GlobalBaseReg);
  }
  return CurDAG->getRegister(GlobalBaseReg, MVT::i32);
}


// isIntImmediate - This method tests to see if a constant operand.
// If so Imm will receive the 32 bit value.
static bool isIntImmediate(SDNode *N, unsigned& Imm) {
  if (N->getOpcode() == ISD::Constant) {
    Imm = cast<ConstantSDNode>(N)->getValue();
    return true;
  }
  return false;
}

// isOprShiftImm - Returns true if the specified operand is a shift opcode with
// a immediate shift count less than 32.
static bool isOprShiftImm(SDNode *N, unsigned& Opc, unsigned& SH) {
  Opc = N->getOpcode();
  return (Opc == ISD::SHL || Opc == ISD::SRL || Opc == ISD::SRA) &&
    isIntImmediate(N->getOperand(1).Val, SH) && SH < 32;
}

// isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
// any number of 0s on either side.  The 1s are allowed to wrap from LSB to
// MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs.  0x0F0F0000 is
// not, since all 1s are not contiguous.
static bool isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) {
  if (isShiftedMask_32(Val)) {
    // look for the first non-zero bit
    MB = CountLeadingZeros_32(Val);
    // look for the first zero bit after the run of ones
    ME = CountLeadingZeros_32((Val - 1) ^ Val);
    return true;
  } else {
    Val = ~Val; // invert mask
    if (isShiftedMask_32(Val)) {
      // effectively look for the first zero bit
      ME = CountLeadingZeros_32(Val) - 1;
      // effectively look for the first one bit after the run of zeros
      MB = CountLeadingZeros_32((Val - 1) ^ Val) + 1;
      return true;
    }
  }
  // no run present
  return false;
}

// isRotateAndMask - Returns true if Mask and Shift can be folded into a rotate
// and mask opcode and mask operation.
static bool isRotateAndMask(SDNode *N, unsigned Mask, bool IsShiftMask,
                            unsigned &SH, unsigned &MB, unsigned &ME) {
  // Don't even go down this path for i64, since different logic will be
  // necessary for rldicl/rldicr/rldimi.
  if (N->getValueType(0) != MVT::i32)
    return false;

  unsigned Shift  = 32;
  unsigned Indeterminant = ~0;  // bit mask marking indeterminant results
  unsigned Opcode = N->getOpcode();
  if (N->getNumOperands() != 2 ||
      !isIntImmediate(N->getOperand(1).Val, Shift) || (Shift > 31))
    return false;
  
  if (Opcode == ISD::SHL) {
    // apply shift left to mask if it comes first
    if (IsShiftMask) Mask = Mask << Shift;
    // determine which bits are made indeterminant by shift
    Indeterminant = ~(0xFFFFFFFFu << Shift);
  } else if (Opcode == ISD::SRL) { 
    // apply shift right to mask if it comes first
    if (IsShiftMask) Mask = Mask >> Shift;
    // determine which bits are made indeterminant by shift
    Indeterminant = ~(0xFFFFFFFFu >> Shift);
    // adjust for the left rotate
    Shift = 32 - Shift;
  } else {
    return false;
  }
  
  // if the mask doesn't intersect any Indeterminant bits
  if (Mask && !(Mask & Indeterminant)) {
    SH = Shift;
    // make sure the mask is still a mask (wrap arounds may not be)
    return isRunOfOnes(Mask, MB, ME);
  }
  return false;
}

// isOpcWithIntImmediate - This method tests to see if the node is a specific
// opcode and that it has a immediate integer right operand.
// If so Imm will receive the 32 bit value.
static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
  return N->getOpcode() == Opc && isIntImmediate(N->getOperand(1).Val, Imm);
}

// isOprNot - Returns true if the specified operand is an xor with immediate -1.
static bool isOprNot(SDNode *N) {
  unsigned Imm;
  return isOpcWithIntImmediate(N, ISD::XOR, Imm) && (signed)Imm == -1;
}

// Immediate constant composers.
// Lo16 - grabs the lo 16 bits from a 32 bit constant.
// Hi16 - grabs the hi 16 bits from a 32 bit constant.
// HA16 - computes the hi bits required if the lo bits are add/subtracted in
// arithmethically.
static unsigned Lo16(unsigned x)  { return x & 0x0000FFFF; }
static unsigned Hi16(unsigned x)  { return Lo16(x >> 16); }
static unsigned HA16(unsigned x)  { return Hi16((signed)x - (signed short)x); }

// isIntImmediate - This method tests to see if a constant operand.
// If so Imm will receive the 32 bit value.
static bool isIntImmediate(SDOperand N, unsigned& Imm) {
  if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
    Imm = (unsigned)CN->getSignExtended();
    return true;
  }
  return false;
}

/// SelectBitfieldInsert - turn an or of two masked values into
/// the rotate left word immediate then mask insert (rlwimi) instruction.
/// Returns true on success, false if the caller still needs to select OR.
///
/// Patterns matched:
/// 1. or shl, and   5. or and, and
/// 2. or and, shl   6. or shl, shr
/// 3. or shr, and   7. or shr, shl
/// 4. or and, shr
SDNode *PPCDAGToDAGISel::SelectBitfieldInsert(SDNode *N) {
  bool IsRotate = false;
  unsigned TgtMask = 0xFFFFFFFF, InsMask = 0xFFFFFFFF, SH = 0;
  unsigned Value;
  
  SDOperand Op0 = N->getOperand(0);
  SDOperand Op1 = N->getOperand(1);
  
  unsigned Op0Opc = Op0.getOpcode();
  unsigned Op1Opc = Op1.getOpcode();
  
  // Verify that we have the correct opcodes
  if (ISD::SHL != Op0Opc && ISD::SRL != Op0Opc && ISD::AND != Op0Opc)
    return false;
  if (ISD::SHL != Op1Opc && ISD::SRL != Op1Opc && ISD::AND != Op1Opc)
    return false;
  
  // Generate Mask value for Target
  if (isIntImmediate(Op0.getOperand(1), Value)) {
    switch(Op0Opc) {
    case ISD::SHL: TgtMask <<= Value; break;
    case ISD::SRL: TgtMask >>= Value; break;
    case ISD::AND: TgtMask &= Value; break;
    }
  } else {
    return 0;
  }
  
  // Generate Mask value for Insert
  if (!isIntImmediate(Op1.getOperand(1), Value))
    return 0;
  
  switch(Op1Opc) {
  case ISD::SHL:
    SH = Value;
    InsMask <<= SH;
    if (Op0Opc == ISD::SRL) IsRotate = true;
    break;
  case ISD::SRL:
    SH = Value;
    InsMask >>= SH;
    SH = 32-SH;
    if (Op0Opc == ISD::SHL) IsRotate = true;
    break;
  case ISD::AND:
    InsMask &= Value;
    break;
  }
  
  // If both of the inputs are ANDs and one of them has a logical shift by
  // constant as its input, make that AND the inserted value so that we can
  // combine the shift into the rotate part of the rlwimi instruction
  bool IsAndWithShiftOp = false;
  if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) {
    if (Op1.getOperand(0).getOpcode() == ISD::SHL ||
        Op1.getOperand(0).getOpcode() == ISD::SRL) {
      if (isIntImmediate(Op1.getOperand(0).getOperand(1), Value)) {
        SH = Op1.getOperand(0).getOpcode() == ISD::SHL ? Value : 32 - Value;
        IsAndWithShiftOp = true;
      }
    } else if (Op0.getOperand(0).getOpcode() == ISD::SHL ||
               Op0.getOperand(0).getOpcode() == ISD::SRL) {
      if (isIntImmediate(Op0.getOperand(0).getOperand(1), Value)) {
        std::swap(Op0, Op1);
        std::swap(TgtMask, InsMask);
        SH = Op1.getOperand(0).getOpcode() == ISD::SHL ? Value : 32 - Value;
        IsAndWithShiftOp = true;
      }
    }
  }
  
  // Verify that the Target mask and Insert mask together form a full word mask
  // and that the Insert mask is a run of set bits (which implies both are runs
  // of set bits).  Given that, Select the arguments and generate the rlwimi
  // instruction.
  unsigned MB, ME;
  if (((TgtMask & InsMask) == 0) && isRunOfOnes(InsMask, MB, ME)) {
    bool fullMask = (TgtMask ^ InsMask) == 0xFFFFFFFF;
    bool Op0IsAND = Op0Opc == ISD::AND;
    // Check for rotlwi / rotrwi here, a special case of bitfield insert
    // where both bitfield halves are sourced from the same value.
    if (IsRotate && fullMask &&
        N->getOperand(0).getOperand(0) == N->getOperand(1).getOperand(0)) {
      Op0 = CurDAG->getTargetNode(PPC::RLWINM, MVT::i32,
                                  Select(N->getOperand(0).getOperand(0)),
                                  getI32Imm(SH), getI32Imm(0), getI32Imm(31));
      return Op0.Val;
    }
    SDOperand Tmp1 = (Op0IsAND && fullMask) ? Select(Op0.getOperand(0))
                                            : Select(Op0);
    SDOperand Tmp2 = IsAndWithShiftOp ? Select(Op1.getOperand(0).getOperand(0)) 
                                      : Select(Op1.getOperand(0));
    Op0 = CurDAG->getTargetNode(PPC::RLWIMI, MVT::i32, Tmp1, Tmp2,
                                getI32Imm(SH), getI32Imm(MB), getI32Imm(ME));
    return Op0.Val;
  }
  return 0;
}

/// SelectAddr - Given the specified address, return the two operands for a
/// load/store instruction, and return true if it should be an indexed [r+r]
/// operation.
bool PPCDAGToDAGISel::SelectAddr(SDOperand Addr, SDOperand &Op1, 
                                 SDOperand &Op2) {
  unsigned imm = 0;
  if (Addr.getOpcode() == ISD::ADD) {
    if (isIntImmediate(Addr.getOperand(1), imm) && isInt16(imm)) {
      Op1 = getI32Imm(Lo16(imm));
      if (FrameIndexSDNode *FI =
            dyn_cast<FrameIndexSDNode>(Addr.getOperand(0))) {
        ++FrameOff;
        Op2 = CurDAG->getTargetFrameIndex(FI->getIndex(), MVT::i32);
      } else {
        Op2 = Select(Addr.getOperand(0));
      }
      return false;
    } else if (Addr.getOperand(1).getOpcode() == PPCISD::Lo) {
      // Match LOAD (ADD (X, Lo(G))).
      assert(!cast<ConstantSDNode>(Addr.getOperand(1).getOperand(1))->getValue()
             && "Cannot handle constant offsets yet!");
      Op1 = Addr.getOperand(1).getOperand(0);  // The global address.
      assert(Op1.getOpcode() == ISD::TargetGlobalAddress);
      Op2 = Select(Addr.getOperand(0));
      return false;   // [&g+r]
    } else {
      Op1 = Select(Addr.getOperand(0));
      Op2 = Select(Addr.getOperand(1));
      return true;   // [r+r]
    }
  }

  if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Addr)) {
    Op1 = getI32Imm(0);
    Op2 = CurDAG->getTargetFrameIndex(FI->getIndex(), MVT::i32);
    return false;
  } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Addr)) {
    Op1 = Addr;
    if (PICEnabled)
      Op2 = CurDAG->getTargetNode(PPC::ADDIS, MVT::i32, getGlobalBaseReg(),Op1);
    else
      Op2 = CurDAG->getTargetNode(PPC::LIS, MVT::i32, Op1);
    return false;
  }
  Op1 = getI32Imm(0);
  Op2 = Select(Addr);
  return false;
}

/// SelectAddrIndexed - Given the specified addressed, force it to be
/// represented as an indexed [r+r] operation, rather than possibly
/// returning [r+imm] as SelectAddr may.
void PPCDAGToDAGISel::SelectAddrIndexed(SDOperand Addr, SDOperand &Op1, 
                                        SDOperand &Op2) {
  if (Addr.getOpcode() == ISD::ADD) {
    Op1 = Select(Addr.getOperand(0));
    Op2 = Select(Addr.getOperand(1));
    return;
  }
  
  if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Addr)) {
    Op1 = CurDAG->getTargetNode(PPC::LI, MVT::i32, getI32Imm(0));
    Op2 = CurDAG->getTargetFrameIndex(FI->getIndex(), MVT::i32);
    return;
  }
  Op1 = CurDAG->getTargetNode(PPC::LI, MVT::i32, getI32Imm(0));
  Op2 = Select(Addr);
}

/// SelectCC - Select a comparison of the specified values with the specified
/// condition code, returning the CR# of the expression.
SDOperand PPCDAGToDAGISel::SelectCC(SDOperand LHS, SDOperand RHS,
                                    ISD::CondCode CC) {
  // Always select the LHS.
  LHS = Select(LHS);

  // Use U to determine whether the SETCC immediate range is signed or not.
  if (MVT::isInteger(LHS.getValueType())) {
    bool U = ISD::isUnsignedIntSetCC(CC);
    unsigned Imm;
    if (isIntImmediate(RHS, Imm) && 
        ((U && isUInt16(Imm)) || (!U && isInt16(Imm))))
      return CurDAG->getTargetNode(U ? PPC::CMPLWI : PPC::CMPWI, MVT::i32,
                                   LHS, getI32Imm(Lo16(Imm)));
    return CurDAG->getTargetNode(U ? PPC::CMPLW : PPC::CMPW, MVT::i32,
                                 LHS, Select(RHS));
  } else if (LHS.getValueType() == MVT::f32) {
    return CurDAG->getTargetNode(PPC::FCMPUS, MVT::i32, LHS, Select(RHS));
  } else {
    return CurDAG->getTargetNode(PPC::FCMPUD, MVT::i32, LHS, Select(RHS));
  }
}

/// getBCCForSetCC - Returns the PowerPC condition branch mnemonic corresponding
/// to Condition.
static unsigned getBCCForSetCC(ISD::CondCode CC) {
  switch (CC) {
  default: assert(0 && "Unknown condition!"); abort();
  case ISD::SETOEQ:    // FIXME: This is incorrect see PR642.
  case ISD::SETEQ:  return PPC::BEQ;
  case ISD::SETONE:    // FIXME: This is incorrect see PR642.
  case ISD::SETNE:  return PPC::BNE;
  case ISD::SETOLT:    // FIXME: This is incorrect see PR642.
  case ISD::SETULT:
  case ISD::SETLT:  return PPC::BLT;
  case ISD::SETOLE:    // FIXME: This is incorrect see PR642.
  case ISD::SETULE:
  case ISD::SETLE:  return PPC::BLE;
  case ISD::SETOGT:    // FIXME: This is incorrect see PR642.
  case ISD::SETUGT:
  case ISD::SETGT:  return PPC::BGT;
  case ISD::SETOGE:    // FIXME: This is incorrect see PR642.
  case ISD::SETUGE:
  case ISD::SETGE:  return PPC::BGE;
    
  case ISD::SETO:   return PPC::BUN;
  case ISD::SETUO:  return PPC::BNU;
  }
  return 0;
}

/// getCRIdxForSetCC - Return the index of the condition register field
/// associated with the SetCC condition, and whether or not the field is
/// treated as inverted.  That is, lt = 0; ge = 0 inverted.
static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool& Inv) {
  switch (CC) {
  default: assert(0 && "Unknown condition!"); abort();
  case ISD::SETOLT:  // FIXME: This is incorrect see PR642.
  case ISD::SETULT:
  case ISD::SETLT:  Inv = false;  return 0;
  case ISD::SETOGE:  // FIXME: This is incorrect see PR642.
  case ISD::SETUGE:
  case ISD::SETGE:  Inv = true;   return 0;
  case ISD::SETOGT:  // FIXME: This is incorrect see PR642.
  case ISD::SETUGT:
  case ISD::SETGT:  Inv = false;  return 1;
  case ISD::SETOLE:  // FIXME: This is incorrect see PR642.
  case ISD::SETULE:
  case ISD::SETLE:  Inv = true;   return 1;
  case ISD::SETOEQ:  // FIXME: This is incorrect see PR642.
  case ISD::SETEQ:  Inv = false;  return 2;
  case ISD::SETONE:  // FIXME: This is incorrect see PR642.
  case ISD::SETNE:  Inv = true;   return 2;
  case ISD::SETO:   Inv = true;   return 3;
  case ISD::SETUO:  Inv = false;  return 3;
  }
  return 0;
}

SDOperand PPCDAGToDAGISel::SelectDYNAMIC_STACKALLOC(SDOperand Op) {
  SDNode *N = Op.Val;

  // FIXME: We are currently ignoring the requested alignment for handling
  // greater than the stack alignment.  This will need to be revisited at some
  // point.  Align = N.getOperand(2);
  if (!isa<ConstantSDNode>(N->getOperand(2)) ||
      cast<ConstantSDNode>(N->getOperand(2))->getValue() != 0) {
    std::cerr << "Cannot allocate stack object with greater alignment than"
    << " the stack alignment yet!";
    abort();
  }
  SDOperand Chain = Select(N->getOperand(0));
  SDOperand Amt   = Select(N->getOperand(1));
  
  SDOperand R1Reg = CurDAG->getRegister(PPC::R1, MVT::i32);
  
  SDOperand R1Val = CurDAG->getCopyFromReg(Chain, PPC::R1, MVT::i32);
  Chain = R1Val.getValue(1);
  
  // Subtract the amount (guaranteed to be a multiple of the stack alignment)
  // from the stack pointer, giving us the result pointer.
  SDOperand Result = CurDAG->getTargetNode(PPC::SUBF, MVT::i32, Amt, R1Val);
  
  // Copy this result back into R1.
  Chain = CurDAG->getNode(ISD::CopyToReg, MVT::Other, Chain, R1Reg, Result);
  
  // Copy this result back out of R1 to make sure we're not using the stack
  // space without decrementing the stack pointer.
  Result = CurDAG->getCopyFromReg(Chain, PPC::R1, MVT::i32);
  
  // Finally, replace the DYNAMIC_STACKALLOC with the copyfromreg.
  CodeGenMap[Op.getValue(0)] = Result;
  CodeGenMap[Op.getValue(1)] = Result.getValue(1);
  return SDOperand(Result.Val, Op.ResNo);
}

SDOperand PPCDAGToDAGISel::SelectADD_PARTS(SDOperand Op) {
  SDNode *N = Op.Val;
  SDOperand LHSL = Select(N->getOperand(0));
  SDOperand LHSH = Select(N->getOperand(1));
  
  unsigned Imm;
  bool ME = false, ZE = false;
  if (isIntImmediate(N->getOperand(3), Imm)) {
    ME = (signed)Imm == -1;
    ZE = Imm == 0;
  }
  
  std::vector<SDOperand> Result;
  SDOperand CarryFromLo;
  if (isIntImmediate(N->getOperand(2), Imm) &&
      ((signed)Imm >= -32768 || (signed)Imm < 32768)) {
    // Codegen the low 32 bits of the add.  Interestingly, there is no
    // shifted form of add immediate carrying.
    CarryFromLo = CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
                                        LHSL, getI32Imm(Imm));
  } else {
    CarryFromLo = CurDAG->getTargetNode(PPC::ADDC, MVT::i32, MVT::Flag,
                                        LHSL, Select(N->getOperand(2)));
  }
  CarryFromLo = CarryFromLo.getValue(1);
  
  // Codegen the high 32 bits, adding zero, minus one, or the full value
  // along with the carry flag produced by addc/addic.
  SDOperand ResultHi;
  if (ZE)
    ResultHi = CurDAG->getTargetNode(PPC::ADDZE, MVT::i32, LHSH, CarryFromLo);
  else if (ME)
    ResultHi = CurDAG->getTargetNode(PPC::ADDME, MVT::i32, LHSH, CarryFromLo);
  else
    ResultHi = CurDAG->getTargetNode(PPC::ADDE, MVT::i32, LHSH,
                                     Select(N->getOperand(3)), CarryFromLo);
  Result.push_back(CarryFromLo.getValue(0));
  Result.push_back(ResultHi);
  
  CodeGenMap[Op.getValue(0)] = Result[0];
  CodeGenMap[Op.getValue(1)] = Result[1];
  return Result[Op.ResNo];
}
SDOperand PPCDAGToDAGISel::SelectSUB_PARTS(SDOperand Op) {
  SDNode *N = Op.Val;
  SDOperand LHSL = Select(N->getOperand(0));
  SDOperand LHSH = Select(N->getOperand(1));
  SDOperand RHSL = Select(N->getOperand(2));
  SDOperand RHSH = Select(N->getOperand(3));
  
  std::vector<SDOperand> Result;
  Result.push_back(CurDAG->getTargetNode(PPC::SUBFC, MVT::i32, MVT::Flag,
                                         RHSL, LHSL));
  Result.push_back(CurDAG->getTargetNode(PPC::SUBFE, MVT::i32, RHSH, LHSH,
                                         Result[0].getValue(1)));
  CodeGenMap[Op.getValue(0)] = Result[0];
  CodeGenMap[Op.getValue(1)] = Result[1];
  return Result[Op.ResNo];
}

SDOperand PPCDAGToDAGISel::SelectSETCC(SDOperand Op) {
  SDNode *N = Op.Val;
  unsigned Imm;
  ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
  if (isIntImmediate(N->getOperand(1), Imm)) {
    // We can codegen setcc op, imm very efficiently compared to a brcond.
    // Check for those cases here.
    // setcc op, 0
    if (Imm == 0) {
      SDOperand Op = Select(N->getOperand(0));
      switch (CC) {
      default: break;
      case ISD::SETEQ:
        Op = CurDAG->getTargetNode(PPC::CNTLZW, MVT::i32, Op);
        CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Op, getI32Imm(27),
                             getI32Imm(5), getI32Imm(31));
        return SDOperand(N, 0);
      case ISD::SETNE: {
        SDOperand AD = CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
                                             Op, getI32Imm(~0U));
        CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op, AD.getValue(1));
        return SDOperand(N, 0);
      }
      case ISD::SETLT:
        CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Op, getI32Imm(1),
                             getI32Imm(31), getI32Imm(31));
        return SDOperand(N, 0);
      case ISD::SETGT: {
        SDOperand T = CurDAG->getTargetNode(PPC::NEG, MVT::i32, Op);
        T = CurDAG->getTargetNode(PPC::ANDC, MVT::i32, T, Op);;
        CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, T, getI32Imm(1),
                             getI32Imm(31), getI32Imm(31));
        return SDOperand(N, 0);
      }
      }
    } else if (Imm == ~0U) {        // setcc op, -1
      SDOperand Op = Select(N->getOperand(0));
      switch (CC) {
      default: break;
      case ISD::SETEQ:
        Op = CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
                                   Op, getI32Imm(1));
        CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32, 
                             CurDAG->getTargetNode(PPC::LI, MVT::i32,
                                                   getI32Imm(0)),
                             Op.getValue(1));
        return SDOperand(N, 0);
      case ISD::SETNE: {
        Op = CurDAG->getTargetNode(PPC::NOR, MVT::i32, Op, Op);
        SDOperand AD = CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
                                             Op, getI32Imm(~0U));
        CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op, AD.getValue(1));
        return SDOperand(N, 0);
      }
      case ISD::SETLT: {
        SDOperand AD = CurDAG->getTargetNode(PPC::ADDI, MVT::i32, Op,
                                             getI32Imm(1));
        SDOperand AN = CurDAG->getTargetNode(PPC::AND, MVT::i32, AD, Op);
        CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, AN, getI32Imm(1),
                             getI32Imm(31), getI32Imm(31));
        return SDOperand(N, 0);
      }
      case ISD::SETGT:
        Op = CurDAG->getTargetNode(PPC::RLWINM, MVT::i32, Op, getI32Imm(1),
                                   getI32Imm(31), getI32Imm(31));
        CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Op, getI32Imm(1));
        return SDOperand(N, 0);
      }
    }
  }
  
  bool Inv;
  unsigned Idx = getCRIdxForSetCC(CC, Inv);
  SDOperand CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC);
  SDOperand IntCR;
  
  // Force the ccreg into CR7.
  SDOperand CR7Reg = CurDAG->getRegister(PPC::CR7, MVT::i32);
  
  std::vector<MVT::ValueType> VTs;
  VTs.push_back(MVT::Other);
  VTs.push_back(MVT::Flag);    // NONSTANDARD CopyToReg node: defines a flag
  std::vector<SDOperand> Ops;
  Ops.push_back(CurDAG->getEntryNode());
  Ops.push_back(CR7Reg);
  Ops.push_back(CCReg);
  CCReg = CurDAG->getNode(ISD::CopyToReg, VTs, Ops).getValue(1);
  
  if (TLI.getTargetMachine().getSubtarget<PPCSubtarget>().isGigaProcessor())
    IntCR = CurDAG->getTargetNode(PPC::MFOCRF, MVT::i32, CR7Reg, CCReg);
  else
    IntCR = CurDAG->getTargetNode(PPC::MFCR, MVT::i32, CCReg);
  
  if (!Inv) {
    CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, IntCR,
                         getI32Imm((32-(3-Idx)) & 31),
                                   getI32Imm(31), getI32Imm(31));
  } else {
    SDOperand Tmp =
    CurDAG->getTargetNode(PPC::RLWINM, MVT::i32, IntCR,
                          getI32Imm((32-(3-Idx)) & 31),
                          getI32Imm(31),getI32Imm(31));
    CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Tmp, getI32Imm(1));
  }
  
  return SDOperand(N, 0);
}

/// isCallCompatibleAddress - Return true if the specified 32-bit value is
/// representable in the immediate field of a Bx instruction.
static bool isCallCompatibleAddress(ConstantSDNode *C) {
  int Addr = C->getValue();
  if (Addr & 3) return false;  // Low 2 bits are implicitly zero.
  return (Addr << 6 >> 6) == Addr;  // Top 6 bits have to be sext of immediate.
}

SDOperand PPCDAGToDAGISel::SelectCALL(SDOperand Op) {
  SDNode *N = Op.Val;
  SDOperand Chain = Select(N->getOperand(0));
  
  unsigned CallOpcode;
  std::vector<SDOperand> CallOperands;
  
  if (GlobalAddressSDNode *GASD =
      dyn_cast<GlobalAddressSDNode>(N->getOperand(1))) {
    CallOpcode = PPC::BL;
    CallOperands.push_back(N->getOperand(1));
  } else if (ExternalSymbolSDNode *ESSDN =
             dyn_cast<ExternalSymbolSDNode>(N->getOperand(1))) {
    CallOpcode = PPC::BL;
    CallOperands.push_back(N->getOperand(1));
  } else if (isa<ConstantSDNode>(N->getOperand(1)) &&
             isCallCompatibleAddress(cast<ConstantSDNode>(N->getOperand(1)))) {
    ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(1));
    CallOpcode = PPC::BLA;
    CallOperands.push_back(getI32Imm((int)C->getValue() >> 2));
  } else {
    // Copy the callee address into the CTR register.
    SDOperand Callee = Select(N->getOperand(1));
    Chain = CurDAG->getTargetNode(PPC::MTCTR, MVT::Other, Callee, Chain);
    
    // Copy the callee address into R12 on darwin.
    SDOperand R12 = CurDAG->getRegister(PPC::R12, MVT::i32);
    Chain = CurDAG->getNode(ISD::CopyToReg, MVT::Other, Chain, R12, Callee);

    CallOperands.push_back(R12);
    CallOpcode = PPC::BCTRL;
  }
  
  unsigned GPR_idx = 0, FPR_idx = 0;
  static const unsigned GPR[] = {
    PPC::R3, PPC::R4, PPC::R5, PPC::R6,
    PPC::R7, PPC::R8, PPC::R9, PPC::R10,
  };
  static const unsigned FPR[] = {
    PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
    PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
  };
  
  SDOperand InFlag;  // Null incoming flag value.
  
  for (unsigned i = 2, e = N->getNumOperands(); i != e; ++i) {
    unsigned DestReg = 0;
    MVT::ValueType RegTy = N->getOperand(i).getValueType();
    if (RegTy == MVT::i32) {
      assert(GPR_idx < 8 && "Too many int args");
      DestReg = GPR[GPR_idx++];
    } else {
      assert(MVT::isFloatingPoint(N->getOperand(i).getValueType()) &&
             "Unpromoted integer arg?");
      assert(FPR_idx < 13 && "Too many fp args");
      DestReg = FPR[FPR_idx++];
    }
    
    if (N->getOperand(i).getOpcode() != ISD::UNDEF) {
      SDOperand Val = Select(N->getOperand(i));
      Chain = CurDAG->getCopyToReg(Chain, DestReg, Val, InFlag);
      InFlag = Chain.getValue(1);
      CallOperands.push_back(CurDAG->getRegister(DestReg, RegTy));
    }
  }
  
  // Finally, once everything is in registers to pass to the call, emit the
  // call itself.
  if (InFlag.Val)
    CallOperands.push_back(InFlag);   // Strong dep on register copies.
  else
    CallOperands.push_back(Chain);    // Weak dep on whatever occurs before
  Chain = CurDAG->getTargetNode(CallOpcode, MVT::Other, MVT::Flag,
                                CallOperands);
  
  std::vector<SDOperand> CallResults;
  
  // If the call has results, copy the values out of the ret val registers.
  switch (N->getValueType(0)) {
    default: assert(0 && "Unexpected ret value!");
    case MVT::Other: break;
    case MVT::i32:
      if (N->getValueType(1) == MVT::i32) {
        Chain = CurDAG->getCopyFromReg(Chain, PPC::R4, MVT::i32, 
                                       Chain.getValue(1)).getValue(1);
        CallResults.push_back(Chain.getValue(0));
        Chain = CurDAG->getCopyFromReg(Chain, PPC::R3, MVT::i32,
                                       Chain.getValue(2)).getValue(1);
        CallResults.push_back(Chain.getValue(0));
      } else {
        Chain = CurDAG->getCopyFromReg(Chain, PPC::R3, MVT::i32,
                                       Chain.getValue(1)).getValue(1);
        CallResults.push_back(Chain.getValue(0));
      }
      break;
    case MVT::f32:
    case MVT::f64:
      Chain = CurDAG->getCopyFromReg(Chain, PPC::F1, N->getValueType(0),
                                     Chain.getValue(1)).getValue(1);
      CallResults.push_back(Chain.getValue(0));
      break;
  }
  
  CallResults.push_back(Chain);
  for (unsigned i = 0, e = CallResults.size(); i != e; ++i)
    CodeGenMap[Op.getValue(i)] = CallResults[i];
  return CallResults[Op.ResNo];
}

// Select - Convert the specified operand from a target-independent to a
// target-specific node if it hasn't already been changed.
SDOperand PPCDAGToDAGISel::Select(SDOperand Op) {
  SDNode *N = Op.Val;
  if (N->getOpcode() >= ISD::BUILTIN_OP_END &&
      N->getOpcode() < PPCISD::FIRST_NUMBER)
    return Op;   // Already selected.

  // If this has already been converted, use it.
  std::map<SDOperand, SDOperand>::iterator CGMI = CodeGenMap.find(Op);
  if (CGMI != CodeGenMap.end()) return CGMI->second;
  
  switch (N->getOpcode()) {
  default: break;
  case ISD::DYNAMIC_STACKALLOC: return SelectDYNAMIC_STACKALLOC(Op);
  case ISD::ADD_PARTS:          return SelectADD_PARTS(Op);
  case ISD::SUB_PARTS:          return SelectSUB_PARTS(Op);
  case ISD::SETCC:              return SelectSETCC(Op);
  case ISD::CALL:               return SelectCALL(Op);
  case ISD::TAILCALL:           return SelectCALL(Op);
  case PPCISD::GlobalBaseReg:   return getGlobalBaseReg();
    
  case ISD::FrameIndex: {
    int FI = cast<FrameIndexSDNode>(N)->getIndex();
    if (N->hasOneUse()) {
      CurDAG->SelectNodeTo(N, PPC::ADDI, MVT::i32,
                           CurDAG->getTargetFrameIndex(FI, MVT::i32),
                           getI32Imm(0));
      return SDOperand(N, 0);
    }
    return CurDAG->getTargetNode(PPC::ADDI, MVT::i32,
                                 CurDAG->getTargetFrameIndex(FI, MVT::i32),
                                 getI32Imm(0));
  }
  case ISD::ConstantPool: {
    Constant *C = cast<ConstantPoolSDNode>(N)->get();
    SDOperand Tmp, CPI = CurDAG->getTargetConstantPool(C, MVT::i32);
    if (PICEnabled)
      Tmp = CurDAG->getTargetNode(PPC::ADDIS, MVT::i32, getGlobalBaseReg(),CPI);
    else
      Tmp = CurDAG->getTargetNode(PPC::LIS, MVT::i32, CPI);
    if (N->hasOneUse()) {
      CurDAG->SelectNodeTo(N, PPC::LA, MVT::i32, Tmp, CPI);
      return SDOperand(N, 0);
    }
    return CurDAG->getTargetNode(PPC::LA, MVT::i32, Tmp, CPI);
  }
  case ISD::FADD: {
    MVT::ValueType Ty = N->getValueType(0);
    if (!NoExcessFPPrecision) {  // Match FMA ops
      if (N->getOperand(0).getOpcode() == ISD::FMUL &&
          N->getOperand(0).Val->hasOneUse()) {
        ++FusedFP; // Statistic
        CurDAG->SelectNodeTo(N, Ty == MVT::f64 ? PPC::FMADD : PPC::FMADDS, Ty,
                             Select(N->getOperand(0).getOperand(0)),
                             Select(N->getOperand(0).getOperand(1)),
                             Select(N->getOperand(1)));
        return SDOperand(N, 0);
      } else if (N->getOperand(1).getOpcode() == ISD::FMUL &&
                 N->getOperand(1).hasOneUse()) {
        ++FusedFP; // Statistic
        CurDAG->SelectNodeTo(N, Ty == MVT::f64 ? PPC::FMADD : PPC::FMADDS, Ty,
                             Select(N->getOperand(1).getOperand(0)),
                             Select(N->getOperand(1).getOperand(1)),
                             Select(N->getOperand(0)));
        return SDOperand(N, 0);
      }
    }
    
    // Other cases are autogenerated.
    break;
  }
  case ISD::FSUB: {
    MVT::ValueType Ty = N->getValueType(0);
    
    if (!NoExcessFPPrecision) {  // Match FMA ops
      if (N->getOperand(0).getOpcode() == ISD::FMUL &&
          N->getOperand(0).Val->hasOneUse()) {
        ++FusedFP; // Statistic
        CurDAG->SelectNodeTo(N, Ty == MVT::f64 ? PPC::FMSUB : PPC::FMSUBS, Ty,
                             Select(N->getOperand(0).getOperand(0)),
                             Select(N->getOperand(0).getOperand(1)),
                             Select(N->getOperand(1)));
        return SDOperand(N, 0);
      } else if (N->getOperand(1).getOpcode() == ISD::FMUL &&
                 N->getOperand(1).Val->hasOneUse()) {
        ++FusedFP; // Statistic
        CurDAG->SelectNodeTo(N, Ty == MVT::f64 ? PPC::FNMSUB : PPC::FNMSUBS, Ty,
                             Select(N->getOperand(1).getOperand(0)),
                             Select(N->getOperand(1).getOperand(1)),
                             Select(N->getOperand(0)));
        return SDOperand(N, 0);
      }
    }

    // Other cases are autogenerated.
    break;
  }
  case ISD::SDIV: {
    // FIXME: since this depends on the setting of the carry flag from the srawi
    //        we should really be making notes about that for the scheduler.
    // FIXME: It sure would be nice if we could cheaply recognize the 
    //        srl/add/sra pattern the dag combiner will generate for this as
    //        sra/addze rather than having to handle sdiv ourselves.  oh well.
    unsigned Imm;
    if (isIntImmediate(N->getOperand(1), Imm)) {
      if ((signed)Imm > 0 && isPowerOf2_32(Imm)) {
        SDOperand Op =
          CurDAG->getTargetNode(PPC::SRAWI, MVT::i32, MVT::Flag,
                                Select(N->getOperand(0)),
                                getI32Imm(Log2_32(Imm)));
        CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32, 
                             Op.getValue(0), Op.getValue(1));
        return SDOperand(N, 0);
      } else if ((signed)Imm < 0 && isPowerOf2_32(-Imm)) {
        SDOperand Op =
          CurDAG->getTargetNode(PPC::SRAWI, MVT::i32, MVT::Flag,
                                Select(N->getOperand(0)),
                                getI32Imm(Log2_32(-Imm)));
        SDOperand PT =
          CurDAG->getTargetNode(PPC::ADDZE, MVT::i32, Op.getValue(0),
                                Op.getValue(1));
        CurDAG->SelectNodeTo(N, PPC::NEG, MVT::i32, PT);
        return SDOperand(N, 0);
      }
    }
    
    // Other cases are autogenerated.
    break;
  }
  case ISD::AND: {
    unsigned Imm;
    // If this is an and of a value rotated between 0 and 31 bits and then and'd
    // with a mask, emit rlwinm
    if (isIntImmediate(N->getOperand(1), Imm) && (isShiftedMask_32(Imm) ||
                                                  isShiftedMask_32(~Imm))) {
      SDOperand Val;
      unsigned SH, MB, ME;
      if (isRotateAndMask(N->getOperand(0).Val, Imm, false, SH, MB, ME)) {
        Val = Select(N->getOperand(0).getOperand(0));
      } else if (Imm == 0) {
        // AND X, 0 -> 0, not "rlwinm 32".
        return Select(N->getOperand(1));
      } else {        
        Val = Select(N->getOperand(0));
        isRunOfOnes(Imm, MB, ME);
        SH = 0;
      }
      CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Val, getI32Imm(SH),
                           getI32Imm(MB), getI32Imm(ME));
      return SDOperand(N, 0);
    }
    
    // Other cases are autogenerated.
    break;
  }
  case ISD::OR:
    if (SDNode *I = SelectBitfieldInsert(N))
      return CodeGenMap[Op] = SDOperand(I, 0);
      
    // Other cases are autogenerated.
    break;
  case ISD::SHL: {
    unsigned Imm, SH, MB, ME;
    if (isOpcWithIntImmediate(N->getOperand(0).Val, ISD::AND, Imm) &&
        isRotateAndMask(N, Imm, true, SH, MB, ME)) {
      CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, 
                           Select(N->getOperand(0).getOperand(0)),
                           getI32Imm(SH), getI32Imm(MB), getI32Imm(ME));
      return SDOperand(N, 0);
    }
    
    // Other cases are autogenerated.
    break;
  }
  case ISD::SRL: {
    unsigned Imm, SH, MB, ME;
    if (isOpcWithIntImmediate(N->getOperand(0).Val, ISD::AND, Imm) &&
        isRotateAndMask(N, Imm, true, SH, MB, ME)) { 
      CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, 
                           Select(N->getOperand(0).getOperand(0)),
                           getI32Imm(SH & 0x1F), getI32Imm(MB), getI32Imm(ME));
      return SDOperand(N, 0);
    }
    
    // Other cases are autogenerated.
    break;
  }
  case ISD::FNEG: {
    SDOperand Val = Select(N->getOperand(0));
    MVT::ValueType Ty = N->getValueType(0);
    if (N->getOperand(0).Val->hasOneUse()) {
      unsigned Opc;
      switch (Val.isTargetOpcode() ? Val.getTargetOpcode() : 0) {
      default:          Opc = 0;            break;
      case PPC::FABSS:  Opc = PPC::FNABSS;  break;
      case PPC::FABSD:  Opc = PPC::FNABSD;  break;
      case PPC::FMADD:  Opc = PPC::FNMADD;  break;
      case PPC::FMADDS: Opc = PPC::FNMADDS; break;
      case PPC::FMSUB:  Opc = PPC::FNMSUB;  break;
      case PPC::FMSUBS: Opc = PPC::FNMSUBS; break;
      }
      // If we inverted the opcode, then emit the new instruction with the
      // inverted opcode and the original instruction's operands.  Otherwise, 
      // fall through and generate a fneg instruction.
      if (Opc) {
        if (Opc == PPC::FNABSS || Opc == PPC::FNABSD)
          CurDAG->SelectNodeTo(N, Opc, Ty, Val.getOperand(0));
        else
          CurDAG->SelectNodeTo(N, Opc, Ty, Val.getOperand(0),
                               Val.getOperand(1), Val.getOperand(2));
        return SDOperand(N, 0);
      }
    }
    if (Ty == MVT::f32)
      CurDAG->SelectNodeTo(N, PPC::FNEGS, MVT::f32, Val);
    else
      CurDAG->SelectNodeTo(N, PPC::FNEGD, MVT::f64, Val);
    return SDOperand(N, 0);
  }
  case ISD::LOAD:
  case ISD::EXTLOAD:
  case ISD::ZEXTLOAD:
  case ISD::SEXTLOAD: {
    SDOperand Op1, Op2;
    // If this is a vector load, then force this to be indexed addressing, since
    // altivec does not have immediate offsets for loads.
    bool isIdx = true;
    if (N->getOpcode() == ISD::LOAD && MVT::isVector(N->getValueType(0))) { 
      SelectAddrIndexed(N->getOperand(1), Op1, Op2);
    } else {
      isIdx = SelectAddr(N->getOperand(1), Op1, Op2);
    }
    MVT::ValueType TypeBeingLoaded = (N->getOpcode() == ISD::LOAD) ?
      N->getValueType(0) : cast<VTSDNode>(N->getOperand(3))->getVT();

    unsigned Opc;
    switch (TypeBeingLoaded) {
    default: N->dump(); assert(0 && "Cannot load this type!");
    case MVT::i1:
    case MVT::i8:  Opc = isIdx ? PPC::LBZX : PPC::LBZ; break;
    case MVT::i16:
      if (N->getOpcode() == ISD::SEXTLOAD) { // SEXT load?
        Opc = isIdx ? PPC::LHAX : PPC::LHA;
      } else {
        Opc = isIdx ? PPC::LHZX : PPC::LHZ;
      }
      break;
    case MVT::i32: Opc = isIdx ? PPC::LWZX : PPC::LWZ; break;
    case MVT::f32: Opc = isIdx ? PPC::LFSX : PPC::LFS; break;
    case MVT::f64: Opc = isIdx ? PPC::LFDX : PPC::LFD; break;
    case MVT::v4f32: Opc = PPC::LVX; break;
    }

    // If this is an f32 -> f64 load, emit the f32 load, then use an 'extending
    // copy'.
    if (TypeBeingLoaded != MVT::f32 || N->getOpcode() == ISD::LOAD) {
        CurDAG->SelectNodeTo(N, Opc, N->getValueType(0), MVT::Other,
                             Op1, Op2, Select(N->getOperand(0)));
      return SDOperand(N, Op.ResNo);
    } else {
      std::vector<SDOperand> Ops;
      Ops.push_back(Op1);
      Ops.push_back(Op2);
      Ops.push_back(Select(N->getOperand(0)));
      SDOperand Res = CurDAG->getTargetNode(Opc, MVT::f32, MVT::Other, Ops);
      SDOperand Ext = CurDAG->getTargetNode(PPC::FMRSD, MVT::f64, Res);
      CodeGenMap[Op.getValue(0)] = Ext;
      CodeGenMap[Op.getValue(1)] = Res.getValue(1);
      if (Op.ResNo)
        return Res.getValue(1);
      else
        return Ext;
    }
  }
  case ISD::TRUNCSTORE:
  case ISD::STORE: {
    SDOperand AddrOp1, AddrOp2;
    // If this is a vector store, then force this to be indexed addressing,
    // since altivec does not have immediate offsets for stores.
    bool isIdx = true;
    if (N->getOpcode() == ISD::STORE && 
        MVT::isVector(N->getOperand(1).getValueType())) {
      SelectAddrIndexed(N->getOperand(2), AddrOp1, AddrOp2);
    } else {
      isIdx = SelectAddr(N->getOperand(2), AddrOp1, AddrOp2);
    }

    unsigned Opc;
    if (N->getOpcode() == ISD::STORE) {
      switch (N->getOperand(1).getValueType()) {
      default: assert(0 && "unknown Type in store");
      case MVT::i32: Opc = isIdx ? PPC::STWX  : PPC::STW; break;
      case MVT::f64: Opc = isIdx ? PPC::STFDX : PPC::STFD; break;
      case MVT::f32: Opc = isIdx ? PPC::STFSX : PPC::STFS; break;
      case MVT::v4f32: Opc = PPC::STVX;
      }
    } else { //ISD::TRUNCSTORE
      switch(cast<VTSDNode>(N->getOperand(4))->getVT()) {
      default: assert(0 && "unknown Type in store");
      case MVT::i8:  Opc = isIdx ? PPC::STBX : PPC::STB; break;
      case MVT::i16: Opc = isIdx ? PPC::STHX : PPC::STH; break;
      }
    }
    
    CurDAG->SelectNodeTo(N, Opc, MVT::Other, Select(N->getOperand(1)),
                         AddrOp1, AddrOp2, Select(N->getOperand(0)));
    return SDOperand(N, 0);
  }
    
  case ISD::SELECT_CC: {
    ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
    
    // handle the setcc cases here.  select_cc lhs, 0, 1, 0, cc
    if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
      if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N->getOperand(2)))
        if (ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N->getOperand(3)))
          if (N1C->isNullValue() && N3C->isNullValue() &&
              N2C->getValue() == 1ULL && CC == ISD::SETNE) {
            SDOperand LHS = Select(N->getOperand(0));
            SDOperand Tmp =
              CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
                                    LHS, getI32Imm(~0U));
            CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, Tmp, LHS,
                                 Tmp.getValue(1));
            return SDOperand(N, 0);
          }

    SDOperand CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC);
    unsigned BROpc = getBCCForSetCC(CC);

    bool isFP = MVT::isFloatingPoint(N->getValueType(0));
    unsigned SelectCCOp;
    if (MVT::isInteger(N->getValueType(0)))
      SelectCCOp = PPC::SELECT_CC_Int;
    else if (N->getValueType(0) == MVT::f32)
      SelectCCOp = PPC::SELECT_CC_F4;
    else
      SelectCCOp = PPC::SELECT_CC_F8;
    CurDAG->SelectNodeTo(N, SelectCCOp, N->getValueType(0), CCReg,
                         Select(N->getOperand(2)), Select(N->getOperand(3)),
                         getI32Imm(BROpc));
    return SDOperand(N, 0);
  }
    
  case ISD::CALLSEQ_START:
  case ISD::CALLSEQ_END: {
    unsigned Amt = cast<ConstantSDNode>(N->getOperand(1))->getValue();
    unsigned Opc = N->getOpcode() == ISD::CALLSEQ_START ?
                       PPC::ADJCALLSTACKDOWN : PPC::ADJCALLSTACKUP;
    CurDAG->SelectNodeTo(N, Opc, MVT::Other,
                         getI32Imm(Amt), Select(N->getOperand(0)));
    return SDOperand(N, 0);
  }
  case ISD::RET: {
    SDOperand Chain = Select(N->getOperand(0));     // Token chain.

    if (N->getNumOperands() == 2) {
      SDOperand Val = Select(N->getOperand(1));
      if (N->getOperand(1).getValueType() == MVT::i32) {
        Chain = CurDAG->getCopyToReg(Chain, PPC::R3, Val);
      } else {
        assert(MVT::isFloatingPoint(N->getOperand(1).getValueType()));
        Chain = CurDAG->getCopyToReg(Chain, PPC::F1, Val);
      }
    } else if (N->getNumOperands() > 1) {
      assert(N->getOperand(1).getValueType() == MVT::i32 &&
             N->getOperand(2).getValueType() == MVT::i32 &&
             N->getNumOperands() == 3 && "Unknown two-register ret value!");
      Chain = CurDAG->getCopyToReg(Chain, PPC::R4, Select(N->getOperand(1)));
      Chain = CurDAG->getCopyToReg(Chain, PPC::R3, Select(N->getOperand(2)));
    }

    // Finally, select this to a blr (return) instruction.
    CurDAG->SelectNodeTo(N, PPC::BLR, MVT::Other, Chain);
    return SDOperand(N, 0);
  }
  case ISD::BR:
    CurDAG->SelectNodeTo(N, PPC::B, MVT::Other, N->getOperand(1),
                         Select(N->getOperand(0)));
    return SDOperand(N, 0);
  case ISD::BR_CC:
  case ISD::BRTWOWAY_CC: {
    SDOperand Chain = Select(N->getOperand(0));
    MachineBasicBlock *Dest =
      cast<BasicBlockSDNode>(N->getOperand(4))->getBasicBlock();
    ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
    SDOperand CondCode = SelectCC(N->getOperand(2), N->getOperand(3), CC);

    // If this is a two way branch, then grab the fallthrough basic block
    // argument and build a PowerPC branch pseudo-op, suitable for long branch
    // conversion if necessary by the branch selection pass.  Otherwise, emit a
    // standard conditional branch.
    if (N->getOpcode() == ISD::BRTWOWAY_CC) {
      SDOperand CondTrueBlock = N->getOperand(4);
      SDOperand CondFalseBlock = N->getOperand(5);
      
      // If the false case is the current basic block, then this is a self loop.
      // We do not want to emit "Loop: ... brcond Out; br Loop", as it adds an
      // extra dispatch group to the loop.  Instead, invert the condition and
      // emit "Loop: ... br!cond Loop; br Out
      if (cast<BasicBlockSDNode>(CondFalseBlock)->getBasicBlock() == BB) {
        std::swap(CondTrueBlock, CondFalseBlock);
        CC = getSetCCInverse(CC,
                             MVT::isInteger(N->getOperand(2).getValueType()));
      }
      
      unsigned Opc = getBCCForSetCC(CC);
      SDOperand CB = CurDAG->getTargetNode(PPC::COND_BRANCH, MVT::Other,
                                           CondCode, getI32Imm(Opc),
                                           CondTrueBlock, CondFalseBlock,
                                           Chain);
      CurDAG->SelectNodeTo(N, PPC::B, MVT::Other, CondFalseBlock, CB);
    } else {
      // Iterate to the next basic block
      ilist<MachineBasicBlock>::iterator It = BB;
      ++It;

      // If the fallthrough path is off the end of the function, which would be
      // undefined behavior, set it to be the same as the current block because
      // we have nothing better to set it to, and leaving it alone will cause
      // the PowerPC Branch Selection pass to crash.
      if (It == BB->getParent()->end()) It = Dest;
      CurDAG->SelectNodeTo(N, PPC::COND_BRANCH, MVT::Other, CondCode,
                           getI32Imm(getBCCForSetCC(CC)), N->getOperand(4),
                           CurDAG->getBasicBlock(It), Chain);
    }
    return SDOperand(N, 0);
  }
  }
  
  return SelectCode(Op);
}


/// createPPCISelDag - This pass converts a legalized DAG into a 
/// PowerPC-specific DAG, ready for instruction scheduling.
///
FunctionPass *llvm::createPPCISelDag(TargetMachine &TM) {
  return new PPCDAGToDAGISel(TM);
}