CGCall.cpp

//===----- CGCall.h - Encapsulate calling convention details ----*- C++ -*-===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// These classes wrap the information about a call or function
// definition used to handle ABI compliancy.
//
//===----------------------------------------------------------------------===//

#include "CGCall.h"
#include "CodeGenFunction.h"
#include "CodeGenModule.h"
#include "clang/Basic/TargetInfo.h"
#include "clang/AST/ASTContext.h"
#include "clang/AST/Decl.h"
#include "clang/AST/DeclObjC.h"
#include "clang/AST/RecordLayout.h"
#include "llvm/ADT/StringExtras.h"
#include "llvm/Attributes.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetData.h"

#include "ABIInfo.h"

using namespace clang;
using namespace CodeGen;

/***/

// FIXME: Use iterator and sidestep silly type array creation.

const 
CGFunctionInfo &CodeGenTypes::getFunctionInfo(const FunctionTypeNoProto *FTNP) {
  return getFunctionInfo(FTNP->getResultType(), 
                         llvm::SmallVector<QualType, 16>());
}

const 
CGFunctionInfo &CodeGenTypes::getFunctionInfo(const FunctionTypeProto *FTP) {
  llvm::SmallVector<QualType, 16> ArgTys;
  // FIXME: Kill copy.
  for (unsigned i = 0, e = FTP->getNumArgs(); i != e; ++i)
    ArgTys.push_back(FTP->getArgType(i));
  return getFunctionInfo(FTP->getResultType(), ArgTys);
}

const CGFunctionInfo &CodeGenTypes::getFunctionInfo(const FunctionDecl *FD) {
  const FunctionType *FTy = FD->getType()->getAsFunctionType();
  if (const FunctionTypeProto *FTP = dyn_cast<FunctionTypeProto>(FTy))
    return getFunctionInfo(FTP);
  return getFunctionInfo(cast<FunctionTypeNoProto>(FTy));
}

const CGFunctionInfo &CodeGenTypes::getFunctionInfo(const ObjCMethodDecl *MD) {
  llvm::SmallVector<QualType, 16> ArgTys;
  ArgTys.push_back(MD->getSelfDecl()->getType());
  ArgTys.push_back(Context.getObjCSelType());
  // FIXME: Kill copy?
  for (ObjCMethodDecl::param_iterator i = MD->param_begin(),
         e = MD->param_end(); i != e; ++i)
    ArgTys.push_back((*i)->getType());
  return getFunctionInfo(MD->getResultType(), ArgTys);
}

const CGFunctionInfo &CodeGenTypes::getFunctionInfo(QualType ResTy, 
                                                    const CallArgList &Args) {
  // FIXME: Kill copy.
  llvm::SmallVector<QualType, 16> ArgTys;
  for (CallArgList::const_iterator i = Args.begin(), e = Args.end(); 
       i != e; ++i)
    ArgTys.push_back(i->second);
  return getFunctionInfo(ResTy, ArgTys);
}

const CGFunctionInfo &CodeGenTypes::getFunctionInfo(QualType ResTy, 
                                                  const FunctionArgList &Args) {
  // FIXME: Kill copy.
  llvm::SmallVector<QualType, 16> ArgTys;
  for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end(); 
       i != e; ++i)
    ArgTys.push_back(i->second);
  return getFunctionInfo(ResTy, ArgTys);
}

const CGFunctionInfo &CodeGenTypes::getFunctionInfo(QualType ResTy,
                               const llvm::SmallVector<QualType, 16> &ArgTys) {
  // Lookup or create unique function info.
  llvm::FoldingSetNodeID ID;
  CGFunctionInfo::Profile(ID, ResTy, ArgTys.begin(), ArgTys.end());

  void *InsertPos = 0;
  CGFunctionInfo *FI = FunctionInfos.FindNodeOrInsertPos(ID, InsertPos);
  if (FI)
    return *FI;

  // Construct the function info.
  FI = new CGFunctionInfo(ResTy, ArgTys);
  FunctionInfos.InsertNode(FI, InsertPos);

  // Compute ABI information.
  getABIInfo().computeInfo(*FI, getContext());

  return *FI;
}

/***/

ABIInfo::~ABIInfo() {}

void ABIArgInfo::dump() const {
  fprintf(stderr, "(ABIArgInfo Kind=");
  switch (TheKind) {
  case Direct: 
    fprintf(stderr, "Direct");
    break;
  case Ignore: 
    fprintf(stderr, "Ignore");
    break;
  case Coerce: 
    fprintf(stderr, "Coerce Type=");
    getCoerceToType()->print(llvm::errs());
    // FIXME: This is ridiculous.
    llvm::errs().flush();
    break;
  case Indirect: 
    fprintf(stderr, "Indirect Align=%d", getIndirectAlign());
    break;
  case Expand: 
    fprintf(stderr, "Expand");
    break;
  }
  fprintf(stderr, ")\n");
}

/***/

/// isEmptyStruct - Return true iff a structure has no non-empty
/// members. Note that a structure with a flexible array member is not
/// considered empty.
static bool isEmptyStruct(QualType T) {
  const RecordType *RT = T->getAsStructureType();
  if (!RT)
    return 0;
  const RecordDecl *RD = RT->getDecl();
  if (RD->hasFlexibleArrayMember())
    return false;
  for (RecordDecl::field_iterator i = RD->field_begin(), 
         e = RD->field_end(); i != e; ++i) {
    const FieldDecl *FD = *i;
    if (!isEmptyStruct(FD->getType()))
      return false;
  }
  return true;
}

/// isSingleElementStruct - Determine if a structure is a "single
/// element struct", i.e. it has exactly one non-empty field or
/// exactly one field which is itself a single element
/// struct. Structures with flexible array members are never
/// considered single element structs.
///
/// \return The field declaration for the single non-empty field, if
/// it exists.
static const FieldDecl *isSingleElementStruct(QualType T) {
  const RecordType *RT = T->getAsStructureType();
  if (!RT)
    return 0;

  const RecordDecl *RD = RT->getDecl();
  if (RD->hasFlexibleArrayMember())
    return 0;

  const FieldDecl *Found = 0;
  for (RecordDecl::field_iterator i = RD->field_begin(), 
         e = RD->field_end(); i != e; ++i) {
    const FieldDecl *FD = *i;
    QualType FT = FD->getType();

    if (isEmptyStruct(FT)) {
      // Ignore
    } else if (Found) {
      return 0;
    } else if (!CodeGenFunction::hasAggregateLLVMType(FT)) {
      Found = FD;
    } else {
      Found = isSingleElementStruct(FT);
      if (!Found)
        return 0;
    }
  }

  return Found;
}

static bool is32Or64BitBasicType(QualType Ty, ASTContext &Context) {
  if (!Ty->getAsBuiltinType() && !Ty->isPointerType())
    return false;

  uint64_t Size = Context.getTypeSize(Ty);
  return Size == 32 || Size == 64;
}

static bool areAllFields32Or64BitBasicType(const RecordDecl *RD,
                                           ASTContext &Context) {
  for (RecordDecl::field_iterator i = RD->field_begin(), 
         e = RD->field_end(); i != e; ++i) {
    const FieldDecl *FD = *i;

    if (!is32Or64BitBasicType(FD->getType(), Context))
      return false;
    
    // If this is a bit-field we need to make sure it is still a
    // 32-bit or 64-bit type.
    if (Expr *BW = FD->getBitWidth()) {
      unsigned Width = BW->getIntegerConstantExprValue(Context).getZExtValue();
      if (Width <= 16)
        return false;
    }
  }
  return true;
}

namespace {
/// DefaultABIInfo - The default implementation for ABI specific
/// details. This implementation provides information which results in
/// self-consistent and sensible LLVM IR generation, but does not
/// conform to any particular ABI.
class DefaultABIInfo : public ABIInfo {
  ABIArgInfo classifyReturnType(QualType RetTy, 
                                ASTContext &Context) const;
  
  ABIArgInfo classifyArgumentType(QualType RetTy,
                                  ASTContext &Context) const;

  virtual void computeInfo(CGFunctionInfo &FI, ASTContext &Context) const {
    FI.getReturnInfo() = classifyReturnType(FI.getReturnType(), Context);
    for (CGFunctionInfo::arg_iterator it = FI.arg_begin(), ie = FI.arg_end();
         it != ie; ++it)
      it->info = classifyArgumentType(it->type, Context);
  }

  virtual llvm::Value *EmitVAArg(llvm::Value *VAListAddr, QualType Ty,
                                 CodeGenFunction &CGF) const;
};

/// X86_32ABIInfo - The X86-32 ABI information.
class X86_32ABIInfo : public ABIInfo {
public:
  ABIArgInfo classifyReturnType(QualType RetTy, 
                                ASTContext &Context) const;

  ABIArgInfo classifyArgumentType(QualType RetTy,
                                  ASTContext &Context) const;

  virtual void computeInfo(CGFunctionInfo &FI, ASTContext &Context) const {
    FI.getReturnInfo() = classifyReturnType(FI.getReturnType(), Context);
    for (CGFunctionInfo::arg_iterator it = FI.arg_begin(), ie = FI.arg_end();
         it != ie; ++it)
      it->info = classifyArgumentType(it->type, Context);
  }

  virtual llvm::Value *EmitVAArg(llvm::Value *VAListAddr, QualType Ty,
                                 CodeGenFunction &CGF) const;
};
}

ABIArgInfo X86_32ABIInfo::classifyReturnType(QualType RetTy,
                                            ASTContext &Context) const {
  if (RetTy->isVoidType()) {
    return ABIArgInfo::getIgnore();
  } else if (CodeGenFunction::hasAggregateLLVMType(RetTy)) {
    // Classify "single element" structs as their element type.
    const FieldDecl *SeltFD = isSingleElementStruct(RetTy);
    if (SeltFD) {
      QualType SeltTy = SeltFD->getType()->getDesugaredType();
      if (const BuiltinType *BT = SeltTy->getAsBuiltinType()) {
        // FIXME: This is gross, it would be nice if we could just
        // pass back SeltTy and have clients deal with it. Is it worth
        // supporting coerce to both LLVM and clang Types?
        if (BT->isIntegerType()) {
          uint64_t Size = Context.getTypeSize(SeltTy);
          return ABIArgInfo::getCoerce(llvm::IntegerType::get((unsigned) Size));
        } else if (BT->getKind() == BuiltinType::Float) {
          return ABIArgInfo::getCoerce(llvm::Type::FloatTy);
        } else if (BT->getKind() == BuiltinType::Double) {
          return ABIArgInfo::getCoerce(llvm::Type::DoubleTy);
        }
      } else if (SeltTy->isPointerType()) {
        // FIXME: It would be really nice if this could come out as
        // the proper pointer type.
        llvm::Type *PtrTy = 
          llvm::PointerType::getUnqual(llvm::Type::Int8Ty);
        return ABIArgInfo::getCoerce(PtrTy);
      }
    }

    uint64_t Size = Context.getTypeSize(RetTy);
    if (Size == 8) {
      return ABIArgInfo::getCoerce(llvm::Type::Int8Ty);
    } else if (Size == 16) {
      return ABIArgInfo::getCoerce(llvm::Type::Int16Ty);
    } else if (Size == 32) {
      return ABIArgInfo::getCoerce(llvm::Type::Int32Ty);
    } else if (Size == 64) {
      return ABIArgInfo::getCoerce(llvm::Type::Int64Ty);
    } else {
      return ABIArgInfo::getIndirect(0);
    }
  } else {
    return ABIArgInfo::getDirect();
  }
}

ABIArgInfo X86_32ABIInfo::classifyArgumentType(QualType Ty,
                                               ASTContext &Context) const {
  // FIXME: Set alignment on indirect arguments.
  if (CodeGenFunction::hasAggregateLLVMType(Ty)) {
    // Structures with flexible arrays are always indirect.
    if (const RecordType *RT = Ty->getAsStructureType())
      if (RT->getDecl()->hasFlexibleArrayMember())
        return ABIArgInfo::getIndirect(0);

    // Ignore empty structs.
    uint64_t Size = Context.getTypeSize(Ty);
    if (Ty->isStructureType() && Size == 0)
      return ABIArgInfo::getIgnore();

    // Expand structs with size <= 128-bits which consist only of
    // basic types (int, long long, float, double, xxx*). This is
    // non-recursive and does not ignore empty fields.
    if (const RecordType *RT = Ty->getAsStructureType()) {
      if (Context.getTypeSize(Ty) <= 4*32 &&
          areAllFields32Or64BitBasicType(RT->getDecl(), Context))
        return ABIArgInfo::getExpand();
    }

    return ABIArgInfo::getIndirect(0);
  } else {
    return ABIArgInfo::getDirect();
  }
}

llvm::Value *X86_32ABIInfo::EmitVAArg(llvm::Value *VAListAddr, QualType Ty,
                                      CodeGenFunction &CGF) const {
  const llvm::Type *BP = llvm::PointerType::getUnqual(llvm::Type::Int8Ty);
  const llvm::Type *BPP = llvm::PointerType::getUnqual(BP);

  CGBuilderTy &Builder = CGF.Builder;
  llvm::Value *VAListAddrAsBPP = Builder.CreateBitCast(VAListAddr, BPP, 
                                                       "ap");
  llvm::Value *Addr = Builder.CreateLoad(VAListAddrAsBPP, "ap.cur");
  llvm::Type *PTy = 
    llvm::PointerType::getUnqual(CGF.ConvertType(Ty));
  llvm::Value *AddrTyped = Builder.CreateBitCast(Addr, PTy);
  
  uint64_t Offset = 
    llvm::RoundUpToAlignment(CGF.getContext().getTypeSize(Ty) / 8, 4);
  llvm::Value *NextAddr = 
    Builder.CreateGEP(Addr, 
                      llvm::ConstantInt::get(llvm::Type::Int32Ty, Offset),
                      "ap.next");
  Builder.CreateStore(NextAddr, VAListAddrAsBPP);

  return AddrTyped;
}

namespace {
/// X86_64ABIInfo - The X86_64 ABI information.
class X86_64ABIInfo : public ABIInfo {
  enum Class {
    Integer = 0,
    SSE,
    SSEUp,
    X87,
    X87Up,
    ComplexX87,
    NoClass,
    Memory
  };

  /// merge - Implement the X86_64 ABI merging algorithm.
  ///
  /// Merge an accumulating classification \arg Accum with a field
  /// classification \arg Field.
  ///
  /// \param Accum - The accumulating classification. This should
  /// always be either NoClass or the result of a previous merge
  /// call. In addition, this should never be Memory (the caller
  /// should just return Memory for the aggregate).
  Class merge(Class Accum, Class Field) const;

  /// classify - Determine the x86_64 register classes in which the
  /// given type T should be passed.
  ///
  /// \param Lo - The classification for the parts of the type
  /// residing in the low word of the containing object.
  ///
  /// \param Hi - The classification for the parts of the type
  /// residing in the high word of the containing object.
  ///
  /// \param OffsetBase - The bit offset of this type in the
  /// containing object.  Some parameters are classified different
  /// depending on whether they straddle an eightbyte boundary.
  ///
  /// If a word is unused its result will be NoClass; if a type should
  /// be passed in Memory then at least the classification of \arg Lo
  /// will be Memory.
  ///
  /// The \arg Lo class will be NoClass iff the argument is ignored.
  ///
  /// If the \arg Lo class is ComplexX87, then the \arg Hi class will
  /// also be ComplexX87.
  void classify(QualType T, ASTContext &Context, uint64_t OffsetBase,
                Class &Lo, Class &Hi) const;
  
  /// getCoerceResult - Given a source type \arg Ty and an LLVM type
  /// to coerce to, chose the best way to pass Ty in the same place
  /// that \arg CoerceTo would be passed, but while keeping the
  /// emitted code as simple as possible.
  ///
  /// FIXME: Note, this should be cleaned up to just take an
  /// enumeration of all the ways we might want to pass things,
  /// instead of constructing an LLVM type. This makes this code more
  /// explicit, and it makes it clearer that we are also doing this
  /// for correctness in the case of passing scalar types.
  ABIArgInfo getCoerceResult(QualType Ty,
                             const llvm::Type *CoerceTo,
                             ASTContext &Context) const;

  ABIArgInfo classifyReturnType(QualType RetTy, 
                                ASTContext &Context) const;  

  ABIArgInfo classifyArgumentType(QualType Ty,
                                  ASTContext &Context,
                                  unsigned &neededInt,
                                  unsigned &neededSSE) const;

public:
  virtual void computeInfo(CGFunctionInfo &FI, ASTContext &Context) const;

  virtual llvm::Value *EmitVAArg(llvm::Value *VAListAddr, QualType Ty,
                                 CodeGenFunction &CGF) const;
};
}

X86_64ABIInfo::Class X86_64ABIInfo::merge(Class Accum, 
                                          Class Field) const {
  // AMD64-ABI 3.2.3p2: Rule 4. Each field of an object is
  // classified recursively so that always two fields are
  // considered. The resulting class is calculated according to
  // the classes of the fields in the eightbyte:
  //
  // (a) If both classes are equal, this is the resulting class.
  //
  // (b) If one of the classes is NO_CLASS, the resulting class is
  // the other class.
  //
  // (c) If one of the classes is MEMORY, the result is the MEMORY
  // class.
  //
  // (d) If one of the classes is INTEGER, the result is the
  // INTEGER.
  //
  // (e) If one of the classes is X87, X87UP, COMPLEX_X87 class,
  // MEMORY is used as class.
  //
  // (f) Otherwise class SSE is used.
  assert((Accum == NoClass || Accum == Integer || 
          Accum == SSE || Accum == SSEUp) &&
         "Invalid accumulated classification during merge.");
  if (Accum == Field || Field == NoClass)
    return Accum;
  else if (Field == Memory)
    return Memory;
  else if (Accum == NoClass)
    return Field;
  else if (Accum == Integer || Field == Integer) 
    return Integer;
  else if (Field == X87 || Field == X87Up || Field == ComplexX87)
    return Memory;
  else
    return SSE;
}

void X86_64ABIInfo::classify(QualType Ty,
                             ASTContext &Context,
                             uint64_t OffsetBase,
                             Class &Lo, Class &Hi) const {
  // FIXME: This code can be simplified by introducing a simple value
  // class for Class pairs with appropriate constructor methods for
  // the various situations.

  Lo = Hi = NoClass;

  Class &Current = OffsetBase < 64 ? Lo : Hi;
  Current = Memory;

  if (const BuiltinType *BT = Ty->getAsBuiltinType()) {
    BuiltinType::Kind k = BT->getKind();

    if (k == BuiltinType::Void) {
      Current = NoClass; 
    } else if (k >= BuiltinType::Bool && k <= BuiltinType::LongLong) {
      Current = Integer;
    } else if (k == BuiltinType::Float || k == BuiltinType::Double) {
      Current = SSE;
    } else if (k == BuiltinType::LongDouble) {
      Lo = X87;
      Hi = X87Up;
    }
    // FIXME: _Decimal32 and _Decimal64 are SSE.
    // FIXME: _float128 and _Decimal128 are (SSE, SSEUp).
    // FIXME: __int128 is (Integer, Integer).
  } else if (Ty->isPointerLikeType() || Ty->isBlockPointerType() ||
             Ty->isObjCQualifiedInterfaceType()) {
    Current = Integer;
  } else if (const VectorType *VT = Ty->getAsVectorType()) {
    uint64_t Size = Context.getTypeSize(VT);
    if (Size == 64) {
      // gcc passes <1 x double> in memory.
      if (VT->getElementType() == Context.DoubleTy)
        return;

      Current = SSE;

      // If this type crosses an eightbyte boundary, it should be
      // split.
      if (OffsetBase && OffsetBase != 64)
        Hi = Lo;
    } else if (Size == 128) {
      Lo = SSE;
      Hi = SSEUp;
    }
  } else if (const ComplexType *CT = Ty->getAsComplexType()) {
    QualType ET = Context.getCanonicalType(CT->getElementType());
    
    uint64_t Size = Context.getTypeSize(Ty);
    if (ET->isIntegerType()) {
      if (Size <= 64)
        Current = Integer;
      else if (Size <= 128)
        Lo = Hi = Integer;
    } else if (ET == Context.FloatTy) 
      Current = SSE;
    else if (ET == Context.DoubleTy)
      Lo = Hi = SSE;
    else if (ET == Context.LongDoubleTy)
      Current = ComplexX87;

    // If this complex type crosses an eightbyte boundary then it
    // should be split.
    uint64_t EB_Real = (OffsetBase) / 64;
    uint64_t EB_Imag = (OffsetBase + Context.getTypeSize(ET)) / 64;
    if (Hi == NoClass && EB_Real != EB_Imag)
      Hi = Lo;
  } else if (const ConstantArrayType *AT = Context.getAsConstantArrayType(Ty)) {
    // Arrays are treated like structures.

    uint64_t Size = Context.getTypeSize(Ty);
    
    // AMD64-ABI 3.2.3p2: Rule 1. If the size of an object is larger
    // than two eightbytes, ..., it has class MEMORY.
    if (Size > 128)
      return;
    
    // AMD64-ABI 3.2.3p2: Rule 1. If ..., or it contains unaligned
    // fields, it has class MEMORY.
    //
    // Only need to check alignment of array base.
    if (OffsetBase % Context.getTypeAlign(AT->getElementType()))
      return;

    // Otherwise implement simplified merge. We could be smarter about
    // this, but it isn't worth it and would be harder to verify.
    Current = NoClass;
    uint64_t EltSize = Context.getTypeSize(AT->getElementType());
    uint64_t ArraySize = AT->getSize().getZExtValue();
    for (uint64_t i=0, Offset=OffsetBase; i<ArraySize; ++i, Offset += EltSize) {
      Class FieldLo, FieldHi;
      classify(AT->getElementType(), Context, Offset, FieldLo, FieldHi);
      Lo = merge(Lo, FieldLo);
      Hi = merge(Hi, FieldHi);
      if (Lo == Memory || Hi == Memory)
        break;
    }
    
    // Do post merger cleanup (see below). Only case we worry about is Memory.
    if (Hi == Memory)
      Lo = Memory;
    assert((Hi != SSEUp || Lo == SSE) && "Invalid SSEUp array classification.");
  } else if (const RecordType *RT = Ty->getAsRecordType()) {
    uint64_t Size = Context.getTypeSize(Ty);
    
    // AMD64-ABI 3.2.3p2: Rule 1. If the size of an object is larger
    // than two eightbytes, ..., it has class MEMORY.
    if (Size > 128)
      return;

    const RecordDecl *RD = RT->getDecl();

    // Assume variable sized types are passed in memory.
    if (RD->hasFlexibleArrayMember())
      return;

    const ASTRecordLayout &Layout = Context.getASTRecordLayout(RD);
    
    // Reset Lo class, this will be recomputed.
    Current = NoClass;
    unsigned idx = 0;
    for (RecordDecl::field_iterator i = RD->field_begin(), 
           e = RD->field_end(); i != e; ++i, ++idx) {
      uint64_t Offset = OffsetBase + Layout.getFieldOffset(idx);
      bool BitField = i->isBitField();

      // AMD64-ABI 3.2.3p2: Rule 1. If ..., or it contains unaligned
      // fields, it has class MEMORY.
      //
      // Note, skip this test for bitfields, see below.
      if (!BitField && Offset % Context.getTypeAlign(i->getType())) {
        Lo = Memory;
        return;
      }

      // Classify this field.
      //
      // AMD64-ABI 3.2.3p2: Rule 3. If the size of the aggregate
      // exceeds a single eightbyte, each is classified
      // separately. Each eightbyte gets initialized to class
      // NO_CLASS.
      Class FieldLo, FieldHi;
      
      // Bitfields require special handling, they do not force the
      // structure to be passed in memory even if unaligned, and
      // therefore they can straddle an eightbyte.
      if (BitField) {
        uint64_t Offset = OffsetBase + Layout.getFieldOffset(idx);
        uint64_t Size = 
          i->getBitWidth()->getIntegerConstantExprValue(Context).getZExtValue();

        uint64_t EB_Lo = Offset / 64;
        uint64_t EB_Hi = (Offset + Size - 1) / 64;
        FieldLo = FieldHi = NoClass;
        if (EB_Lo) {
          assert(EB_Hi == EB_Lo && "Invalid classification, type > 16 bytes.");
          FieldLo = NoClass;
          FieldHi = Integer;
        } else { 
          FieldLo = Integer;
          FieldHi = EB_Hi ? Integer : NoClass;
        }
      } else
        classify(i->getType(), Context, Offset, FieldLo, FieldHi);
      Lo = merge(Lo, FieldLo);
      Hi = merge(Hi, FieldHi);
      if (Lo == Memory || Hi == Memory)
        break;
    }

    // AMD64-ABI 3.2.3p2: Rule 5. Then a post merger cleanup is done:
    //
    // (a) If one of the classes is MEMORY, the whole argument is
    // passed in memory.
    //
    // (b) If SSEUP is not preceeded by SSE, it is converted to SSE.

    // The first of these conditions is guaranteed by how we implement
    // the merge (just bail). 
    //
    // The second condition occurs in the case of unions; for example
    // union { _Complex double; unsigned; }.
    if (Hi == Memory)
      Lo = Memory;
    if (Hi == SSEUp && Lo != SSE)
      Hi = SSE;
  }
}

ABIArgInfo X86_64ABIInfo::getCoerceResult(QualType Ty,
                                          const llvm::Type *CoerceTo,
                                          ASTContext &Context) const {
  if (CoerceTo == llvm::Type::Int64Ty) {
    // Integer and pointer types will end up in a general purpose
    // register.
    if (Ty->isIntegerType() || Ty->isPointerType())
      return ABIArgInfo::getDirect();
  } else if (CoerceTo == llvm::Type::DoubleTy) {
    // FIXME: It would probably be better to make CGFunctionInfo only
    // map using canonical types than to canonize here.
    QualType CTy = Context.getCanonicalType(Ty);
  
    // Float and double end up in a single SSE reg.
    if (CTy == Context.FloatTy || CTy == Context.DoubleTy)
      return ABIArgInfo::getDirect();
  }

  return ABIArgInfo::getCoerce(CoerceTo);
}

ABIArgInfo X86_64ABIInfo::classifyReturnType(QualType RetTy,
                                            ASTContext &Context) const {
  // AMD64-ABI 3.2.3p4: Rule 1. Classify the return type with the
  // classification algorithm.
  X86_64ABIInfo::Class Lo, Hi;
  classify(RetTy, Context, 0, Lo, Hi);

  // Check some invariants.
  assert((Hi != Memory || Lo == Memory) && "Invalid memory classification.");
  assert((Lo != NoClass || Hi == NoClass) && "Invalid null classification.");
  assert((Hi != SSEUp || Lo == SSE) && "Invalid SSEUp classification.");

  const llvm::Type *ResType = 0;
  switch (Lo) {
  case NoClass:
    return ABIArgInfo::getIgnore();

  case SSEUp:
  case X87Up:
    assert(0 && "Invalid classification for lo word.");

    // AMD64-ABI 3.2.3p4: Rule 2. Types of class memory are returned via
    // hidden argument.
  case Memory:
    return ABIArgInfo::getIndirect(0);

    // AMD64-ABI 3.2.3p4: Rule 3. If the class is INTEGER, the next
    // available register of the sequence %rax, %rdx is used.
  case Integer:
    ResType = llvm::Type::Int64Ty; break;

    // AMD64-ABI 3.2.3p4: Rule 4. If the class is SSE, the next
    // available SSE register of the sequence %xmm0, %xmm1 is used.
  case SSE:
    ResType = llvm::Type::DoubleTy; break;

    // AMD64-ABI 3.2.3p4: Rule 6. If the class is X87, the value is
    // returned on the X87 stack in %st0 as 80-bit x87 number.
  case X87:
    ResType = llvm::Type::X86_FP80Ty; break;

    // AMD64-ABI 3.2.3p4: Rule 8. If the class is COMPLEX_X87, the real
    // part of the value is returned in %st0 and the imaginary part in
    // %st1.
  case ComplexX87:
    assert(Hi == ComplexX87 && "Unexpected ComplexX87 classification.");
    ResType = llvm::StructType::get(llvm::Type::X86_FP80Ty,
                                    llvm::Type::X86_FP80Ty,
                                    NULL);
    break;    
  }

  switch (Hi) {
    // Memory was handled previously and X87 should
    // never occur as a hi class.
  case Memory:
  case X87:
    assert(0 && "Invalid classification for hi word.");

  case ComplexX87: // Previously handled.
  case NoClass: break;

  case Integer:
    ResType = llvm::StructType::get(ResType, llvm::Type::Int64Ty, NULL);
    break;
  case SSE:    
    ResType = llvm::StructType::get(ResType, llvm::Type::DoubleTy, NULL);
    break;

    // AMD64-ABI 3.2.3p4: Rule 5. If the class is SSEUP, the eightbyte
    // is passed in the upper half of the last used SSE register.
    //
    // SSEUP should always be preceeded by SSE, just widen.
  case SSEUp:
    assert(Lo == SSE && "Unexpected SSEUp classification.");
    ResType = llvm::VectorType::get(llvm::Type::DoubleTy, 2);
    break;

    // AMD64-ABI 3.2.3p4: Rule 7. If the class is X87UP, the value is
    // returned together with the previous X87 value in %st0.
    //
    // X87UP should always be preceeded by X87, so we don't need to do
    // anything here.
  case X87Up:
    assert(Lo == X87 && "Unexpected X87Up classification.");
    break;
  }

  return getCoerceResult(RetTy, ResType, Context);
}

ABIArgInfo X86_64ABIInfo::classifyArgumentType(QualType Ty, ASTContext &Context,
                                               unsigned &neededInt,
                                               unsigned &neededSSE) const {
  X86_64ABIInfo::Class Lo, Hi;
  classify(Ty, Context, 0, Lo, Hi);
  
  // Check some invariants.
  // FIXME: Enforce these by construction.
  assert((Hi != Memory || Lo == Memory) && "Invalid memory classification.");
  assert((Lo != NoClass || Hi == NoClass) && "Invalid null classification.");
  assert((Hi != SSEUp || Lo == SSE) && "Invalid SSEUp classification.");

  neededInt = 0;
  neededSSE = 0;
  const llvm::Type *ResType = 0;
  switch (Lo) {
  case NoClass:
    return ABIArgInfo::getIgnore();

    // AMD64-ABI 3.2.3p3: Rule 1. If the class is MEMORY, pass the argument
    // on the stack.
  case Memory:
    
    // AMD64-ABI 3.2.3p3: Rule 5. If the class is X87, X87UP or
    // COMPLEX_X87, it is passed in memory.
  case X87:
  case ComplexX87:
    // Choose appropriate in memory type.
    if (CodeGenFunction::hasAggregateLLVMType(Ty))
      return ABIArgInfo::getIndirect(0);
    else
      return ABIArgInfo::getDirect();

  case SSEUp:
  case X87Up:
    assert(0 && "Invalid classification for lo word.");

    // AMD64-ABI 3.2.3p3: Rule 2. If the class is INTEGER, the next
    // available register of the sequence %rdi, %rsi, %rdx, %rcx, %r8
    // and %r9 is used.
  case Integer:
    ++neededInt; 
    ResType = llvm::Type::Int64Ty;
    break;

    // AMD64-ABI 3.2.3p3: Rule 3. If the class is SSE, the next
    // available SSE register is used, the registers are taken in the
    // order from %xmm0 to %xmm7.
  case SSE:
    ++neededSSE; 
    ResType = llvm::Type::DoubleTy;
    break;
  }

  switch (Hi) {
    // Memory was handled previously, ComplexX87 and X87 should
    // never occur as hi classes, and X87Up must be preceed by X87,
    // which is passed in memory.
  case Memory:
  case X87:
  case X87Up:
  case ComplexX87:
    assert(0 && "Invalid classification for hi word.");

  case NoClass: break;
  case Integer:
    ResType = llvm::StructType::get(ResType, llvm::Type::Int64Ty, NULL);
    ++neededInt;
    break;
  case SSE:    
    ResType = llvm::StructType::get(ResType, llvm::Type::DoubleTy, NULL);
    ++neededSSE;
    break;

    // AMD64-ABI 3.2.3p3: Rule 4. If the class is SSEUP, the
    // eightbyte is passed in the upper half of the last used SSE
    // register.
  case SSEUp:
    assert(Lo == SSE && "Unexpected SSEUp classification.");
    ResType = llvm::VectorType::get(llvm::Type::DoubleTy, 2);
    break;
  }

  return getCoerceResult(Ty, ResType, Context);
}

void X86_64ABIInfo::computeInfo(CGFunctionInfo &FI, ASTContext &Context) const {
  FI.getReturnInfo() = classifyReturnType(FI.getReturnType(), Context);

  // Keep track of the number of assigned registers.
  unsigned freeIntRegs = 6, freeSSERegs = 8;

  // AMD64-ABI 3.2.3p3: Once arguments are classified, the registers
  // get assigned (in left-to-right order) for passing as follows...
  for (CGFunctionInfo::arg_iterator it = FI.arg_begin(), ie = FI.arg_end();
       it != ie; ++it) {
    unsigned neededInt, neededSSE;
    it->info = classifyArgumentType(it->type, Context, neededInt, neededSSE);

    // AMD64-ABI 3.2.3p3: If there are no registers available for any
    // eightbyte of an argument, the whole argument is passed on the
    // stack. If registers have already been assigned for some
    // eightbytes of such an argument, the assignments get reverted.
    if (freeIntRegs >= neededInt && freeSSERegs >= neededSSE) {
      freeIntRegs -= neededInt;
      freeSSERegs -= neededSSE;
    } else {
      // Choose appropriate in memory type.
      if (CodeGenFunction::hasAggregateLLVMType(it->type))
        it->info = ABIArgInfo::getIndirect(0);
      else
        it->info = ABIArgInfo::getDirect();
    }
  }
}

static llvm::Value *EmitVAArgFromMemory(llvm::Value *VAListAddr, 
                                        QualType Ty,
                                        CodeGenFunction &CGF) {
  llvm::Value *overflow_arg_area_p = 
    CGF.Builder.CreateStructGEP(VAListAddr, 2, "overflow_arg_area_p");
  llvm::Value *overflow_arg_area = 
    CGF.Builder.CreateLoad(overflow_arg_area_p, "overflow_arg_area");

  // AMD64-ABI 3.5.7p5: Step 7. Align l->overflow_arg_area upwards to a 16
  // byte boundary if alignment needed by type exceeds 8 byte boundary.
  uint64_t Align = CGF.getContext().getTypeAlign(Ty) / 8;
  if (Align > 8) {
    // Note that we follow the ABI & gcc here, even though the type
    // could in theory have an alignment greater than 16. This case
    // shouldn't ever matter in practice.

    // overflow_arg_area = (overflow_arg_area + 15) & ~15;
    llvm::Value *Offset = llvm::ConstantInt::get(llvm::Type::Int32Ty, 15);
    overflow_arg_area = CGF.Builder.CreateGEP(overflow_arg_area, Offset);
    llvm::Value *AsInt = CGF.Builder.CreatePtrToInt(overflow_arg_area,
                                                    llvm::Type::Int64Ty);
    llvm::Value *Mask = llvm::ConstantInt::get(llvm::Type::Int64Ty, ~15LL);
    overflow_arg_area = 
      CGF.Builder.CreateIntToPtr(CGF.Builder.CreateAnd(AsInt, Mask),
                                 overflow_arg_area->getType(),
                                 "overflow_arg_area.align");
  }

  // AMD64-ABI 3.5.7p5: Step 8. Fetch type from l->overflow_arg_area.
  const llvm::Type *LTy = CGF.ConvertTypeForMem(Ty);
  llvm::Value *Res = 
    CGF.Builder.CreateBitCast(overflow_arg_area, 
                              llvm::PointerType::getUnqual(LTy));

  // AMD64-ABI 3.5.7p5: Step 9. Set l->overflow_arg_area to:
  // l->overflow_arg_area + sizeof(type).
  // AMD64-ABI 3.5.7p5: Step 10. Align l->overflow_arg_area upwards to
  // an 8 byte boundary.

  uint64_t SizeInBytes = (CGF.getContext().getTypeSize(Ty) + 7) / 8;
  llvm::Value *Offset = llvm::ConstantInt::get(llvm::Type::Int32Ty,
                                               (SizeInBytes + 7)  & ~7);
  overflow_arg_area = CGF.Builder.CreateGEP(overflow_arg_area, Offset,
                                            "overflow_arg_area.next");
  CGF.Builder.CreateStore(overflow_arg_area, overflow_arg_area_p);

  // AMD64-ABI 3.5.7p5: Step 11. Return the fetched type.  
  return Res;
}

llvm::Value *X86_64ABIInfo::EmitVAArg(llvm::Value *VAListAddr, QualType Ty,
                                      CodeGenFunction &CGF) const {
  // Assume that va_list type is correct; should be pointer to LLVM type:
  // struct {
  //   i32 gp_offset;
  //   i32 fp_offset;
  //   i8* overflow_arg_area;
  //   i8* reg_save_area;
  // }; 
  unsigned neededInt, neededSSE;
  ABIArgInfo AI = classifyArgumentType(Ty, CGF.getContext(), 
                                       neededInt, neededSSE);

  // AMD64-ABI 3.5.7p5: Step 1. Determine whether type may be passed
  // in the registers. If not go to step 7.
  if (!neededInt && !neededSSE)
    return EmitVAArgFromMemory(VAListAddr, Ty, CGF);

  // AMD64-ABI 3.5.7p5: Step 2. Compute num_gp to hold the number of
  // general purpose registers needed to pass type and num_fp to hold
  // the number of floating point registers needed.

  // AMD64-ABI 3.5.7p5: Step 3. Verify whether arguments fit into
  // registers. In the case: l->gp_offset > 48 - num_gp * 8 or
  // l->fp_offset > 304 - num_fp * 16 go to step 7.
  // 
  // NOTE: 304 is a typo, there are (6 * 8 + 8 * 16) = 176 bytes of
  // register save space).

  llvm::Value *InRegs = 0;
  llvm::Value *gp_offset_p = 0, *gp_offset = 0;
  llvm::Value *fp_offset_p = 0, *fp_offset = 0;
  if (neededInt) {
    gp_offset_p = CGF.Builder.CreateStructGEP(VAListAddr, 0, "gp_offset_p");
    gp_offset = CGF.Builder.CreateLoad(gp_offset_p, "gp_offset");
    InRegs = 
      CGF.Builder.CreateICmpULE(gp_offset,
                                llvm::ConstantInt::get(llvm::Type::Int32Ty,