r246985 - Compute and preserve alignment more faithfully in IR-generation.

Chandler Carruth via cfe-commits cfe-commits at lists.llvm.org
Thu Sep 10 16:09:28 PDT 2015


On Thu, Sep 10, 2015 at 3:26 PM John McCall <rjmccall at apple.com> wrote:

> On Sep 10, 2015, at 3:22 PM, Chandler Carruth <chandlerc at google.com>
> wrote:
>
> I've reproduced this with the same technique.
>
> John, let me know if you need help debugging this, but this is blocking a
> ton of stuff for us so I'm going to revert for now.
>
>
> Please just give me a chance to fix; it is probably straightforward, and
> the change is invasive enough that you’ll have to revert the last three
> days of changes to IRGen.
>

I mean, I'd love to avoid the revert, but were totally hosed at the moment
by this change (perhaps our fault for so much code managing to use a silly
jpeg library). I'm going to at least start working on teasing this stuff
apart in parallel, and hopefully you'll have a fix soon. While we're going
to need to revert to green before too long, I'll update here before doing
anything. Also happy to chat on IRC if there is anything I can do to help.


>
> John.
>
>
> On Thu, Sep 10, 2015 at 3:13 PM David Majnemer <david.majnemer at gmail.com>
> wrote:
>
>> The failure can be reproduced using ./libjpeg-turbo/build/.libs/lt-djpeg
>> -dct fast -outfile testout_420_q100_ifast.ppm
>> testout_420_q100_ifast_prog.jpg, it will die
>> in jsimd_ycc_rgb_convert_sse2.rowloop
>>
>> I am using https://github.com/libjpeg-turbo/libjpeg-turbo.git at
>> revision 0d293537728f211888b04bed6ee19f71e0bda504
>>
>> Let me know if this isn't enough to reproduce.
>>
>> On Thu, Sep 10, 2015 at 2:39 PM, Chandler Carruth via cfe-commits <
>> cfe-commits at lists.llvm.org> wrote:
>>
>>> In case anyone else is chasing the same thing, I wanted to post that
>>> we're seeing crashes in code introduced by this commit. Specifically,
>>> crashes inside of libjpeg-turbo's assembly implementation code, so its
>>> proving very challenging to track down. It looks like either incorrect
>>> vaargs stack setup, reaching past the redzone due to alignment padding, or
>>> something else weird.
>>>
>>> Notably, it is *not* a SIGILL due to an overaligned access to memory.
>>>
>>> Anyways, if we get a test case I'll probably file it and revert, but its
>>> proving *very* hard to track down so wanted to see if others have seen
>>> something here.
>>>
>>> -Chandler
>>>
>>>
>>> On Tue, Sep 8, 2015 at 1:07 AM John McCall via cfe-commits <
>>> cfe-commits at lists.llvm.org> wrote:
>>>
>>>> Modified: cfe/trunk/lib/CodeGen/TargetInfo.cpp
>>>> URL:
>>>> http://llvm.org/viewvc/llvm-project/cfe/trunk/lib/CodeGen/TargetInfo.cpp?rev=246985&r1=246984&r2=246985&view=diff
>>>>
>>>> ==============================================================================
>>>> --- cfe/trunk/lib/CodeGen/TargetInfo.cpp (original)
>>>> +++ cfe/trunk/lib/CodeGen/TargetInfo.cpp Tue Sep  8 03:05:57 2015
>>>> @@ -39,7 +39,7 @@ static void AssignToArrayRange(CodeGen::
>>>>    for (unsigned I = FirstIndex; I <= LastIndex; ++I) {
>>>>      llvm::Value *Cell =
>>>>          Builder.CreateConstInBoundsGEP1_32(Builder.getInt8Ty(), Array,
>>>> I);
>>>> -    Builder.CreateStore(Value, Cell);
>>>> +    Builder.CreateAlignedStore(Value, Cell, CharUnits::One());
>>>>    }
>>>>  }
>>>>
>>>> @@ -48,6 +48,19 @@ static bool isAggregateTypeForABI(QualTy
>>>>           T->isMemberFunctionPointerType();
>>>>  }
>>>>
>>>> +ABIArgInfo
>>>> +ABIInfo::getNaturalAlignIndirect(QualType Ty, bool ByRef, bool Realign,
>>>> +                                 llvm::Type *Padding) const {
>>>> +  return ABIArgInfo::getIndirect(getContext().getTypeAlignInChars(Ty),
>>>> +                                 ByRef, Realign, Padding);
>>>> +}
>>>> +
>>>> +ABIArgInfo
>>>> +ABIInfo::getNaturalAlignIndirectInReg(QualType Ty, bool Realign) const
>>>> {
>>>> +  return
>>>> ABIArgInfo::getIndirectInReg(getContext().getTypeAlignInChars(Ty),
>>>> +                                      /*ByRef*/ false, Realign);
>>>> +}
>>>> +
>>>>  ABIInfo::~ABIInfo() {}
>>>>
>>>>  static CGCXXABI::RecordArgABI getRecordArgABI(const RecordType *RT,
>>>> @@ -133,7 +146,7 @@ void ABIArgInfo::dump() const {
>>>>      OS << "InAlloca Offset=" << getInAllocaFieldIndex();
>>>>      break;
>>>>    case Indirect:
>>>> -    OS << "Indirect Align=" << getIndirectAlign()
>>>> +    OS << "Indirect Align=" << getIndirectAlign().getQuantity()
>>>>         << " ByVal=" << getIndirectByVal()
>>>>         << " Realign=" << getIndirectRealign();
>>>>      break;
>>>> @@ -144,6 +157,125 @@ void ABIArgInfo::dump() const {
>>>>    OS << ")\n";
>>>>  }
>>>>
>>>> +/// Emit va_arg for a platform using the common void* representation,
>>>> +/// where arguments are simply emitted in an array of slots on the
>>>> stack.
>>>> +///
>>>> +/// This version implements the core direct-value passing rules.
>>>> +///
>>>> +/// \param SlotSize - The size and alignment of a stack slot.
>>>> +///   Each argument will be allocated to a multiple of this number of
>>>> +///   slots, and all the slots will be aligned to this value.
>>>> +/// \param AllowHigherAlign - The slot alignment is not a cap;
>>>> +///   an argument type with an alignment greater than the slot size
>>>> +///   will be emitted on a higher-alignment address, potentially
>>>> +///   leaving one or more empty slots behind as padding.  If this
>>>> +///   is false, the returned address might be less-aligned than
>>>> +///   DirectAlign.
>>>> +static Address emitVoidPtrDirectVAArg(CodeGenFunction &CGF,
>>>> +                                      Address VAListAddr,
>>>> +                                      llvm::Type *DirectTy,
>>>> +                                      CharUnits DirectSize,
>>>> +                                      CharUnits DirectAlign,
>>>> +                                      CharUnits SlotSize,
>>>> +                                      bool AllowHigherAlign) {
>>>> +  // Cast the element type to i8* if necessary.  Some platforms define
>>>> +  // va_list as a struct containing an i8* instead of just an i8*.
>>>> +  if (VAListAddr.getElementType() != CGF.Int8PtrTy)
>>>> +    VAListAddr = CGF.Builder.CreateElementBitCast(VAListAddr,
>>>> CGF.Int8PtrTy);
>>>> +
>>>> +  llvm::Value *Ptr = CGF.Builder.CreateLoad(VAListAddr, "argp.cur");
>>>> +
>>>> +  // If the CC aligns values higher than the slot size, do so if
>>>> needed.
>>>> +  Address Addr = Address::invalid();
>>>> +  if (AllowHigherAlign && DirectAlign > SlotSize) {
>>>> +    llvm::Value *PtrAsInt = Ptr;
>>>> +    PtrAsInt = CGF.Builder.CreatePtrToInt(PtrAsInt, CGF.IntPtrTy);
>>>> +    PtrAsInt = CGF.Builder.CreateAdd(PtrAsInt,
>>>> +          llvm::ConstantInt::get(CGF.IntPtrTy,
>>>> DirectAlign.getQuantity() - 1));
>>>> +    PtrAsInt = CGF.Builder.CreateAnd(PtrAsInt,
>>>> +             llvm::ConstantInt::get(CGF.IntPtrTy,
>>>> -DirectAlign.getQuantity()));
>>>> +    Addr = Address(CGF.Builder.CreateIntToPtr(PtrAsInt, Ptr->getType(),
>>>> +                                              "argp.cur.aligned"),
>>>> +                   DirectAlign);
>>>> +  } else {
>>>> +    Addr = Address(Ptr, SlotSize);
>>>> +  }
>>>> +
>>>> +  // Advance the pointer past the argument, then store that back.
>>>> +  CharUnits FullDirectSize = DirectSize.RoundUpToAlignment(SlotSize);
>>>> +  llvm::Value *NextPtr =
>>>> +    CGF.Builder.CreateConstInBoundsByteGEP(Addr.getPointer(),
>>>> FullDirectSize,
>>>> +                                           "argp.next");
>>>> +  CGF.Builder.CreateStore(NextPtr, VAListAddr);
>>>> +
>>>> +  // If the argument is smaller than a slot, and this is a big-endian
>>>> +  // target, the argument will be right-adjusted in its slot.
>>>> +  if (DirectSize < SlotSize && CGF.CGM.getDataLayout().isBigEndian()) {
>>>> +    Addr = CGF.Builder.CreateConstInBoundsByteGEP(Addr, SlotSize -
>>>> DirectSize);
>>>> +  }
>>>> +
>>>> +  Addr = CGF.Builder.CreateElementBitCast(Addr, DirectTy);
>>>> +  return Addr;
>>>> +}
>>>> +
>>>> +/// Emit va_arg for a platform using the common void* representation,
>>>> +/// where arguments are simply emitted in an array of slots on the
>>>> stack.
>>>> +///
>>>> +/// \param IsIndirect - Values of this type are passed indirectly.
>>>> +/// \param ValueInfo - The size and alignment of this type, generally
>>>> +///   computed with getContext().getTypeInfoInChars(ValueTy).
>>>> +/// \param SlotSizeAndAlign - The size and alignment of a stack slot.
>>>> +///   Each argument will be allocated to a multiple of this number of
>>>> +///   slots, and all the slots will be aligned to this value.
>>>> +/// \param AllowHigherAlign - The slot alignment is not a cap;
>>>> +///   an argument type with an alignment greater than the slot size
>>>> +///   will be emitted on a higher-alignment address, potentially
>>>> +///   leaving one or more empty slots behind as padding.
>>>> +static Address emitVoidPtrVAArg(CodeGenFunction &CGF, Address
>>>> VAListAddr,
>>>> +                                QualType ValueTy, bool IsIndirect,
>>>> +                                std::pair<CharUnits, CharUnits>
>>>> ValueInfo,
>>>> +                                CharUnits SlotSizeAndAlign,
>>>> +                                bool AllowHigherAlign) {
>>>> +  // The size and alignment of the value that was passed directly.
>>>> +  CharUnits DirectSize, DirectAlign;
>>>> +  if (IsIndirect) {
>>>> +    DirectSize = CGF.getPointerSize();
>>>> +    DirectAlign = CGF.getPointerAlign();
>>>> +  } else {
>>>> +    DirectSize = ValueInfo.first;
>>>> +    DirectAlign = ValueInfo.second;
>>>> +  }
>>>> +
>>>> +  // Cast the address we've calculated to the right type.
>>>> +  llvm::Type *DirectTy = CGF.ConvertTypeForMem(ValueTy);
>>>> +  if (IsIndirect)
>>>> +    DirectTy = DirectTy->getPointerTo(0);
>>>> +
>>>> +  Address Addr = emitVoidPtrDirectVAArg(CGF, VAListAddr, DirectTy,
>>>> +                                        DirectSize, DirectAlign,
>>>> +                                        SlotSizeAndAlign,
>>>> +                                        AllowHigherAlign);
>>>> +
>>>> +  if (IsIndirect) {
>>>> +    Addr = Address(CGF.Builder.CreateLoad(Addr), ValueInfo.second);
>>>> +  }
>>>> +
>>>> +  return Addr;
>>>> +
>>>> +}
>>>> +
>>>> +static Address emitMergePHI(CodeGenFunction &CGF,
>>>> +                            Address Addr1, llvm::BasicBlock *Block1,
>>>> +                            Address Addr2, llvm::BasicBlock *Block2,
>>>> +                            const llvm::Twine &Name = "") {
>>>> +  assert(Addr1.getType() == Addr2.getType());
>>>> +  llvm::PHINode *PHI = CGF.Builder.CreatePHI(Addr1.getType(), 2, Name);
>>>> +  PHI->addIncoming(Addr1.getPointer(), Block1);
>>>> +  PHI->addIncoming(Addr2.getPointer(), Block2);
>>>> +  CharUnits Align = std::min(Addr1.getAlignment(),
>>>> Addr2.getAlignment());
>>>> +  return Address(PHI, Align);
>>>> +}
>>>> +
>>>>  TargetCodeGenInfo::~TargetCodeGenInfo() { delete Info; }
>>>>
>>>>  // If someone can figure out a general rule for this, that would be
>>>> great.
>>>> @@ -394,8 +526,8 @@ public:
>>>>        I.info <http://i.info> = classifyArgumentType(I.type);
>>>>    }
>>>>
>>>> -  llvm::Value *EmitVAArg(llvm::Value *VAListAddr, QualType Ty,
>>>> -                         CodeGenFunction &CGF) const override;
>>>> +  Address EmitVAArg(CodeGenFunction &CGF, Address VAListAddr,
>>>> +                    QualType Ty) const override;
>>>>  };
>>>>
>>>>  class DefaultTargetCodeGenInfo : public TargetCodeGenInfo {
>>>> @@ -404,9 +536,9 @@ public:
>>>>      : TargetCodeGenInfo(new DefaultABIInfo(CGT)) {}
>>>>  };
>>>>
>>>> -llvm::Value *DefaultABIInfo::EmitVAArg(llvm::Value *VAListAddr,
>>>> QualType Ty,
>>>> -                                       CodeGenFunction &CGF) const {
>>>> -  return nullptr;
>>>> +Address DefaultABIInfo::EmitVAArg(CodeGenFunction &CGF, Address
>>>> VAListAddr,
>>>> +                                  QualType Ty) const {
>>>> +  return Address::invalid();
>>>>  }
>>>>
>>>>  ABIArgInfo DefaultABIInfo::classifyArgumentType(QualType Ty) const {
>>>> @@ -416,9 +548,9 @@ ABIArgInfo DefaultABIInfo::classifyArgum
>>>>      // Records with non-trivial destructors/copy-constructors should
>>>> not be
>>>>      // passed by value.
>>>>      if (CGCXXABI::RecordArgABI RAA = getRecordArgABI(Ty, getCXXABI()))
>>>> -      return ABIArgInfo::getIndirect(0, RAA ==
>>>> CGCXXABI::RAA_DirectInMemory);
>>>> +      return getNaturalAlignIndirect(Ty, RAA ==
>>>> CGCXXABI::RAA_DirectInMemory);
>>>>
>>>> -    return ABIArgInfo::getIndirect(0);
>>>> +    return getNaturalAlignIndirect(Ty);
>>>>    }
>>>>
>>>>    // Treat an enum type as its underlying type.
>>>> @@ -434,7 +566,7 @@ ABIArgInfo DefaultABIInfo::classifyRetur
>>>>      return ABIArgInfo::getIgnore();
>>>>
>>>>    if (isAggregateTypeForABI(RetTy))
>>>> -    return ABIArgInfo::getIndirect(0);
>>>> +    return getNaturalAlignIndirect(RetTy);
>>>>
>>>>    // Treat an enum type as its underlying type.
>>>>    if (const EnumType *EnumTy = RetTy->getAs<EnumType>())
>>>> @@ -482,10 +614,8 @@ ABIArgInfo WebAssemblyABIInfo::classifyA
>>>>    if (isAggregateTypeForABI(Ty)) {
>>>>      // Records with non-trivial destructors/copy-constructors should
>>>> not be
>>>>      // passed by value.
>>>> -    unsigned TypeAlign =
>>>> getContext().getTypeAlignInChars(Ty).getQuantity();
>>>>      if (auto RAA = getRecordArgABI(Ty, getCXXABI()))
>>>> -      return ABIArgInfo::getIndirect(TypeAlign,
>>>> -                                     RAA ==
>>>> CGCXXABI::RAA_DirectInMemory);
>>>> +      return getNaturalAlignIndirect(Ty, RAA ==
>>>> CGCXXABI::RAA_DirectInMemory);
>>>>      // Ignore empty structs/unions.
>>>>      if (isEmptyRecord(getContext(), Ty, true))
>>>>        return ABIArgInfo::getIgnore();
>>>> @@ -494,7 +624,7 @@ ABIArgInfo WebAssemblyABIInfo::classifyA
>>>>      // though watch out for things like bitfields.
>>>>      if (const Type *SeltTy = isSingleElementStruct(Ty, getContext()))
>>>>        return ABIArgInfo::getDirect(CGT.ConvertType(QualType(SeltTy,
>>>> 0)));
>>>> -    return ABIArgInfo::getIndirect(TypeAlign);
>>>> +    return getNaturalAlignIndirect(Ty);
>>>>    }
>>>>
>>>>    // Otherwise just do the default thing.
>>>> @@ -536,8 +666,8 @@ class PNaClABIInfo : public ABIInfo {
>>>>    ABIArgInfo classifyArgumentType(QualType RetTy) const;
>>>>
>>>>    void computeInfo(CGFunctionInfo &FI) const override;
>>>> -  llvm::Value *EmitVAArg(llvm::Value *VAListAddr, QualType Ty,
>>>> -                         CodeGenFunction &CGF) const override;
>>>> +  Address EmitVAArg(CodeGenFunction &CGF,
>>>> +                    Address VAListAddr, QualType Ty) const override;
>>>>  };
>>>>
>>>>  class PNaClTargetCodeGenInfo : public TargetCodeGenInfo {
>>>> @@ -554,17 +684,17 @@ void PNaClABIInfo::computeInfo(CGFunctio
>>>>      I.info <http://i.info> = classifyArgumentType(I.type);
>>>>  }
>>>>
>>>> -llvm::Value *PNaClABIInfo::EmitVAArg(llvm::Value *VAListAddr, QualType
>>>> Ty,
>>>> -                                       CodeGenFunction &CGF) const {
>>>> -  return nullptr;
>>>> +Address PNaClABIInfo::EmitVAArg(CodeGenFunction &CGF, Address
>>>> VAListAddr,
>>>> +                                QualType Ty) const {
>>>> +  return Address::invalid();
>>>>  }
>>>>
>>>>  /// \brief Classify argument of given type \p Ty.
>>>>  ABIArgInfo PNaClABIInfo::classifyArgumentType(QualType Ty) const {
>>>>    if (isAggregateTypeForABI(Ty)) {
>>>>      if (CGCXXABI::RecordArgABI RAA = getRecordArgABI(Ty, getCXXABI()))
>>>> -      return ABIArgInfo::getIndirect(0, RAA ==
>>>> CGCXXABI::RAA_DirectInMemory);
>>>> -    return ABIArgInfo::getIndirect(0);
>>>> +      return getNaturalAlignIndirect(Ty, RAA ==
>>>> CGCXXABI::RAA_DirectInMemory);
>>>> +    return getNaturalAlignIndirect(Ty);
>>>>    } else if (const EnumType *EnumTy = Ty->getAs<EnumType>()) {
>>>>      // Treat an enum type as its underlying type.
>>>>      Ty = EnumTy->getDecl()->getIntegerType();
>>>> @@ -583,7 +713,7 @@ ABIArgInfo PNaClABIInfo::classifyReturnT
>>>>
>>>>    // In the PNaCl ABI we always return records/structures on the stack.
>>>>    if (isAggregateTypeForABI(RetTy))
>>>> -    return ABIArgInfo::getIndirect(0);
>>>> +    return getNaturalAlignIndirect(RetTy);
>>>>
>>>>    // Treat an enum type as its underlying type.
>>>>    if (const EnumType *EnumTy = RetTy->getAs<EnumType>())
>>>> @@ -687,7 +817,7 @@ class X86_32ABIInfo : public ABIInfo {
>>>>    /// such that the argument will be passed in memory.
>>>>    ABIArgInfo getIndirectResult(QualType Ty, bool ByVal, CCState
>>>> &State) const;
>>>>
>>>> -  ABIArgInfo getIndirectReturnResult(CCState &State) const;
>>>> +  ABIArgInfo getIndirectReturnResult(QualType Ty, CCState &State)
>>>> const;
>>>>
>>>>    /// \brief Return the alignment to use for the given type on the
>>>> stack.
>>>>    unsigned getTypeStackAlignInBytes(QualType Ty, unsigned Align) const;
>>>> @@ -702,14 +832,14 @@ class X86_32ABIInfo : public ABIInfo {
>>>>    void rewriteWithInAlloca(CGFunctionInfo &FI) const;
>>>>
>>>>    void addFieldToArgStruct(SmallVector<llvm::Type *, 6> &FrameFields,
>>>> -                           unsigned &StackOffset, ABIArgInfo &Info,
>>>> +                           CharUnits &StackOffset, ABIArgInfo &Info,
>>>>                             QualType Type) const;
>>>>
>>>>  public:
>>>>
>>>>    void computeInfo(CGFunctionInfo &FI) const override;
>>>> -  llvm::Value *EmitVAArg(llvm::Value *VAListAddr, QualType Ty,
>>>> -                         CodeGenFunction &CGF) const override;
>>>> +  Address EmitVAArg(CodeGenFunction &CGF, Address VAListAddr,
>>>> +                    QualType Ty) const override;
>>>>
>>>>    X86_32ABIInfo(CodeGen::CodeGenTypes &CGT, bool d, bool p, bool w,
>>>>                  unsigned r)
>>>> @@ -893,14 +1023,14 @@ bool X86_32ABIInfo::shouldReturnTypeInRe
>>>>    return true;
>>>>  }
>>>>
>>>> -ABIArgInfo X86_32ABIInfo::getIndirectReturnResult(CCState &State)
>>>> const {
>>>> +ABIArgInfo X86_32ABIInfo::getIndirectReturnResult(QualType RetTy,
>>>> CCState &State) const {
>>>>    // If the return value is indirect, then the hidden argument is
>>>> consuming one
>>>>    // integer register.
>>>>    if (State.FreeRegs) {
>>>>      --State.FreeRegs;
>>>> -    return ABIArgInfo::getIndirectInReg(/*Align=*/0, /*ByVal=*/false);
>>>> +    return getNaturalAlignIndirectInReg(RetTy);
>>>>    }
>>>> -  return ABIArgInfo::getIndirect(/*Align=*/0, /*ByVal=*/false);
>>>> +  return getNaturalAlignIndirect(RetTy, /*ByVal=*/false);
>>>>  }
>>>>
>>>>  ABIArgInfo X86_32ABIInfo::classifyReturnType(QualType RetTy,
>>>> @@ -935,7 +1065,7 @@ ABIArgInfo X86_32ABIInfo::classifyReturn
>>>>          return
>>>> ABIArgInfo::getDirect(llvm::IntegerType::get(getVMContext(),
>>>>                                                              Size));
>>>>
>>>> -      return getIndirectReturnResult(State);
>>>> +      return getIndirectReturnResult(RetTy, State);
>>>>      }
>>>>
>>>>      return ABIArgInfo::getDirect();
>>>> @@ -945,12 +1075,12 @@ ABIArgInfo X86_32ABIInfo::classifyReturn
>>>>      if (const RecordType *RT = RetTy->getAs<RecordType>()) {
>>>>        // Structures with flexible arrays are always indirect.
>>>>        if (RT->getDecl()->hasFlexibleArrayMember())
>>>> -        return getIndirectReturnResult(State);
>>>> +        return getIndirectReturnResult(RetTy, State);
>>>>      }
>>>>
>>>>      // If specified, structs and unions are always indirect.
>>>>      if (!IsSmallStructInRegABI && !RetTy->isAnyComplexType())
>>>> -      return getIndirectReturnResult(State);
>>>> +      return getIndirectReturnResult(RetTy, State);
>>>>
>>>>      // Small structures which are register sized are generally returned
>>>>      // in a register.
>>>> @@ -972,7 +1102,7 @@ ABIArgInfo X86_32ABIInfo::classifyReturn
>>>>        return
>>>> ABIArgInfo::getDirect(llvm::IntegerType::get(getVMContext(),Size));
>>>>      }
>>>>
>>>> -    return getIndirectReturnResult(State);
>>>> +    return getIndirectReturnResult(RetTy, State);
>>>>    }
>>>>
>>>>    // Treat an enum type as its underlying type.
>>>> @@ -1038,21 +1168,22 @@ ABIArgInfo X86_32ABIInfo::getIndirectRes
>>>>    if (!ByVal) {
>>>>      if (State.FreeRegs) {
>>>>        --State.FreeRegs; // Non-byval indirects just use one pointer.
>>>> -      return ABIArgInfo::getIndirectInReg(0, false);
>>>> +      return getNaturalAlignIndirectInReg(Ty);
>>>>      }
>>>> -    return ABIArgInfo::getIndirect(0, false);
>>>> +    return getNaturalAlignIndirect(Ty, false);
>>>>    }
>>>>
>>>>    // Compute the byval alignment.
>>>>    unsigned TypeAlign = getContext().getTypeAlign(Ty) / 8;
>>>>    unsigned StackAlign = getTypeStackAlignInBytes(Ty, TypeAlign);
>>>>    if (StackAlign == 0)
>>>> -    return ABIArgInfo::getIndirect(4, /*ByVal=*/true);
>>>> +    return ABIArgInfo::getIndirect(CharUnits::fromQuantity(4),
>>>> /*ByVal=*/true);
>>>>
>>>>    // If the stack alignment is less than the type alignment, realign
>>>> the
>>>>    // argument.
>>>>    bool Realign = TypeAlign > StackAlign;
>>>> -  return ABIArgInfo::getIndirect(StackAlign, /*ByVal=*/true, Realign);
>>>> +  return ABIArgInfo::getIndirect(CharUnits::fromQuantity(StackAlign),
>>>> +                                 /*ByVal=*/true, Realign);
>>>>  }
>>>>
>>>>  X86_32ABIInfo::Class X86_32ABIInfo::classify(QualType Ty) const {
>>>> @@ -1259,22 +1390,23 @@ void X86_32ABIInfo::computeInfo(CGFuncti
>>>>
>>>>  void
>>>>  X86_32ABIInfo::addFieldToArgStruct(SmallVector<llvm::Type *, 6>
>>>> &FrameFields,
>>>> -                                   unsigned &StackOffset,
>>>> -                                   ABIArgInfo &Info, QualType Type)
>>>> const {
>>>> -  assert(StackOffset % 4U == 0 && "unaligned inalloca struct");
>>>> +                                   CharUnits &StackOffset, ABIArgInfo
>>>> &Info,
>>>> +                                   QualType Type) const {
>>>> +  // Arguments are always 4-byte-aligned.
>>>> +  CharUnits FieldAlign = CharUnits::fromQuantity(4);
>>>> +
>>>> +  assert(StackOffset.isMultipleOf(FieldAlign) && "unaligned inalloca
>>>> struct");
>>>>    Info = ABIArgInfo::getInAlloca(FrameFields.size());
>>>>    FrameFields.push_back(CGT.ConvertTypeForMem(Type));
>>>> -  StackOffset += getContext().getTypeSizeInChars(Type).getQuantity();
>>>> +  StackOffset += getContext().getTypeSizeInChars(Type);
>>>>
>>>> -  // Insert padding bytes to respect alignment.  For x86_32, each
>>>> argument is 4
>>>> -  // byte aligned.
>>>> -  if (StackOffset % 4U) {
>>>> -    unsigned OldOffset = StackOffset;
>>>> -    StackOffset = llvm::RoundUpToAlignment(StackOffset, 4U);
>>>> -    unsigned NumBytes = StackOffset - OldOffset;
>>>> -    assert(NumBytes);
>>>> +  // Insert padding bytes to respect alignment.
>>>> +  CharUnits FieldEnd = StackOffset;
>>>> +  StackOffset = FieldEnd.RoundUpToAlignment(FieldAlign);
>>>> +  if (StackOffset != FieldEnd) {
>>>> +    CharUnits NumBytes = StackOffset - FieldEnd;
>>>>      llvm::Type *Ty = llvm::Type::getInt8Ty(getVMContext());
>>>> -    Ty = llvm::ArrayType::get(Ty, NumBytes);
>>>> +    Ty = llvm::ArrayType::get(Ty, NumBytes.getQuantity());
>>>>      FrameFields.push_back(Ty);
>>>>    }
>>>>  }
>>>> @@ -1305,7 +1437,10 @@ void X86_32ABIInfo::rewriteWithInAlloca(
>>>>    // Build a packed struct type for all of the arguments in memory.
>>>>    SmallVector<llvm::Type *, 6> FrameFields;
>>>>
>>>> -  unsigned StackOffset = 0;
>>>> +  // The stack alignment is always 4.
>>>> +  CharUnits StackAlign = CharUnits::fromQuantity(4);
>>>> +
>>>> +  CharUnits StackOffset;
>>>>    CGFunctionInfo::arg_iterator I = FI.arg_begin(), E = FI.arg_end();
>>>>
>>>>    // Put 'this' into the struct before 'sret', if necessary.
>>>> @@ -1337,47 +1472,25 @@ void X86_32ABIInfo::rewriteWithInAlloca(
>>>>    }
>>>>
>>>>    FI.setArgStruct(llvm::StructType::get(getVMContext(), FrameFields,
>>>> -                                        /*isPacked=*/true));
>>>> +                                        /*isPacked=*/true),
>>>> +                  StackAlign);
>>>>  }
>>>>
>>>> -llvm::Value *X86_32ABIInfo::EmitVAArg(llvm::Value *VAListAddr,
>>>> QualType Ty,
>>>> -                                      CodeGenFunction &CGF) const {
>>>> -  llvm::Type *BPP = CGF.Int8PtrPtrTy;
>>>> +Address X86_32ABIInfo::EmitVAArg(CodeGenFunction &CGF,
>>>> +                                 Address VAListAddr, QualType Ty)
>>>> const {
>>>>
>>>> -  CGBuilderTy &Builder = CGF.Builder;
>>>> -  llvm::Value *VAListAddrAsBPP = Builder.CreateBitCast(VAListAddr, BPP,
>>>> -                                                       "ap");
>>>> -  llvm::Value *Addr = Builder.CreateLoad(VAListAddrAsBPP, "ap.cur");
>>>> -
>>>> -  // Compute if the address needs to be aligned
>>>> -  unsigned Align =
>>>> CGF.getContext().getTypeAlignInChars(Ty).getQuantity();
>>>> -  Align = getTypeStackAlignInBytes(Ty, Align);
>>>> -  Align = std::max(Align, 4U);
>>>> -  if (Align > 4) {
>>>> -    // addr = (addr + align - 1) & -align;
>>>> -    llvm::Value *Offset =
>>>> -      llvm::ConstantInt::get(CGF.Int32Ty, Align - 1);
>>>> -    Addr = CGF.Builder.CreateGEP(Addr, Offset);
>>>> -    llvm::Value *AsInt = CGF.Builder.CreatePtrToInt(Addr,
>>>> -                                                    CGF.Int32Ty);
>>>> -    llvm::Value *Mask = llvm::ConstantInt::get(CGF.Int32Ty, -Align);
>>>> -    Addr = CGF.Builder.CreateIntToPtr(CGF.Builder.CreateAnd(AsInt,
>>>> Mask),
>>>> -                                      Addr->getType(),
>>>> -                                      "ap.cur.aligned");
>>>> -  }
>>>> -
>>>> -  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,
>>>> Align);
>>>> -  llvm::Value *NextAddr =
>>>> -    Builder.CreateGEP(Addr, llvm::ConstantInt::get(CGF.Int32Ty,
>>>> Offset),
>>>> -                      "ap.next");
>>>> -  Builder.CreateStore(NextAddr, VAListAddrAsBPP);
>>>> +  auto TypeInfo = getContext().getTypeInfoInChars(Ty);
>>>>
>>>> -  return AddrTyped;
>>>> +  // x86-32 changes the alignment of certain arguments on the stack.
>>>> +  //
>>>> +  // Just messing with TypeInfo like this works because we never pass
>>>> +  // anything indirectly.
>>>> +  TypeInfo.second = CharUnits::fromQuantity(
>>>> +                getTypeStackAlignInBytes(Ty,
>>>> TypeInfo.second.getQuantity()));
>>>> +
>>>> +  return emitVoidPtrVAArg(CGF, VAListAddr, Ty, /*Indirect*/ false,
>>>> +                          TypeInfo, CharUnits::fromQuantity(4),
>>>> +                          /*AllowHigherAlign*/ true);
>>>>  }
>>>>
>>>>  bool X86_32TargetCodeGenInfo::isStructReturnInRegABI(
>>>> @@ -1449,8 +1562,9 @@ bool X86_32TargetCodeGenInfo::initDwarfE
>>>>    } else {
>>>>      // 9 is %eflags, which doesn't get a size on Darwin for some
>>>>      // reason.
>>>> -    Builder.CreateStore(
>>>> -        Four8, Builder.CreateConstInBoundsGEP1_32(CGF.Int8Ty, Address,
>>>> 9));
>>>> +    Builder.CreateAlignedStore(
>>>> +        Four8, Builder.CreateConstInBoundsGEP1_32(CGF.Int8Ty, Address,
>>>> 9),
>>>> +                               CharUnits::One());
>>>>
>>>>      // 11-16 are st(0..5).  Not sure why we stop at 5.
>>>>      // These have size 12, which is sizeof(long double) on
>>>> @@ -1619,8 +1733,8 @@ public:
>>>>
>>>>    void computeInfo(CGFunctionInfo &FI) const override;
>>>>
>>>> -  llvm::Value *EmitVAArg(llvm::Value *VAListAddr, QualType Ty,
>>>> -                         CodeGenFunction &CGF) const override;
>>>> +  Address EmitVAArg(CodeGenFunction &CGF, Address VAListAddr,
>>>> +                    QualType Ty) const override;
>>>>
>>>>    bool has64BitPointers() const {
>>>>      return Has64BitPointers;
>>>> @@ -1638,8 +1752,8 @@ public:
>>>>
>>>>    void computeInfo(CGFunctionInfo &FI) const override;
>>>>
>>>> -  llvm::Value *EmitVAArg(llvm::Value *VAListAddr, QualType Ty,
>>>> -                         CodeGenFunction &CGF) const override;
>>>> +  Address EmitVAArg(CodeGenFunction &CGF, Address VAListAddr,
>>>> +                    QualType Ty) const override;
>>>>
>>>>    bool isHomogeneousAggregateBaseType(QualType Ty) const override {
>>>>      // FIXME: Assumes vectorcall is in use.
>>>> @@ -2257,7 +2371,7 @@ ABIArgInfo X86_64ABIInfo::getIndirectRet
>>>>              ABIArgInfo::getExtend() : ABIArgInfo::getDirect());
>>>>    }
>>>>
>>>> -  return ABIArgInfo::getIndirect(0);
>>>> +  return getNaturalAlignIndirect(Ty);
>>>>  }
>>>>
>>>>  bool X86_64ABIInfo::IsIllegalVectorType(QualType Ty) const {
>>>> @@ -2291,7 +2405,7 @@ ABIArgInfo X86_64ABIInfo::getIndirectRes
>>>>    }
>>>>
>>>>    if (CGCXXABI::RecordArgABI RAA = getRecordArgABI(Ty, getCXXABI()))
>>>> -    return ABIArgInfo::getIndirect(0, RAA ==
>>>> CGCXXABI::RAA_DirectInMemory);
>>>> +    return getNaturalAlignIndirect(Ty, RAA ==
>>>> CGCXXABI::RAA_DirectInMemory);
>>>>
>>>>    // Compute the byval alignment. We specify the alignment of the
>>>> byval in all
>>>>    // cases so that the mid-level optimizer knows the alignment of the
>>>> byval.
>>>> @@ -2328,7 +2442,7 @@ ABIArgInfo X86_64ABIInfo::getIndirectRes
>>>>                                                            Size));
>>>>    }
>>>>
>>>> -  return ABIArgInfo::getIndirect(Align);
>>>> +  return ABIArgInfo::getIndirect(CharUnits::fromQuantity(Align));
>>>>  }
>>>>
>>>>  /// The ABI specifies that a value should be passed in a full vector
>>>> XMM/YMM
>>>> @@ -2912,11 +3026,10 @@ void X86_64ABIInfo::computeInfo(CGFuncti
>>>>    }
>>>>  }
>>>>
>>>> -static llvm::Value *EmitVAArgFromMemory(llvm::Value *VAListAddr,
>>>> -                                        QualType Ty,
>>>> -                                        CodeGenFunction &CGF) {
>>>> -  llvm::Value *overflow_arg_area_p = CGF.Builder.CreateStructGEP(
>>>> -      nullptr, VAListAddr, 2, "overflow_arg_area_p");
>>>> +static Address EmitX86_64VAArgFromMemory(CodeGenFunction &CGF,
>>>> +                                         Address VAListAddr, QualType
>>>> Ty) {
>>>> +  Address overflow_arg_area_p = CGF.Builder.CreateStructGEP(
>>>> +      VAListAddr, 2, CharUnits::fromQuantity(8),
>>>> "overflow_arg_area_p");
>>>>    llvm::Value *overflow_arg_area =
>>>>      CGF.Builder.CreateLoad(overflow_arg_area_p, "overflow_arg_area");
>>>>
>>>> @@ -2924,7 +3037,7 @@ static llvm::Value *EmitVAArgFromMemory(
>>>>    // byte boundary if alignment needed by type exceeds 8 byte boundary.
>>>>    // It isn't stated explicitly in the standard, but in practice we use
>>>>    // alignment greater than 16 where necessary.
>>>> -  uint64_t Align = CGF.getContext().getTypeAlign(Ty) / 8;
>>>> +  uint64_t Align =
>>>> CGF.getContext().getTypeAlignInChars(Ty).getQuantity();
>>>>    if (Align > 8) {
>>>>      // overflow_arg_area = (overflow_arg_area + align - 1) & -align;
>>>>      llvm::Value *Offset =
>>>> @@ -2958,11 +3071,11 @@ static llvm::Value *EmitVAArgFromMemory(
>>>>    CGF.Builder.CreateStore(overflow_arg_area, overflow_arg_area_p);
>>>>
>>>>    // AMD64-ABI 3.5.7p5: Step 11. Return the fetched type.
>>>> -  return Res;
>>>> +  return Address(Res, CharUnits::fromQuantity(Align));
>>>>  }
>>>>
>>>> -llvm::Value *X86_64ABIInfo::EmitVAArg(llvm::Value *VAListAddr,
>>>> QualType Ty,
>>>> -                                      CodeGenFunction &CGF) const {
>>>> +Address X86_64ABIInfo::EmitVAArg(CodeGenFunction &CGF, Address
>>>> VAListAddr,
>>>> +                                 QualType Ty) const {
>>>>    // Assume that va_list type is correct; should be pointer to LLVM
>>>> type:
>>>>    // struct {
>>>>    //   i32 gp_offset;
>>>> @@ -2972,14 +3085,14 @@ llvm::Value *X86_64ABIInfo::EmitVAArg(ll
>>>>    // };
>>>>    unsigned neededInt, neededSSE;
>>>>
>>>> -  Ty = CGF.getContext().getCanonicalType(Ty);
>>>> +  Ty = getContext().getCanonicalType(Ty);
>>>>    ABIArgInfo AI = classifyArgumentType(Ty, 0, neededInt, neededSSE,
>>>>                                         /*isNamedArg*/false);
>>>>
>>>>    // 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);
>>>> +    return EmitX86_64VAArgFromMemory(CGF, VAListAddr, Ty);
>>>>
>>>>    // 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
>>>> @@ -2993,11 +3106,12 @@ llvm::Value *X86_64ABIInfo::EmitVAArg(ll
>>>>    // register save space).
>>>>
>>>>    llvm::Value *InRegs = nullptr;
>>>> -  llvm::Value *gp_offset_p = nullptr, *gp_offset = nullptr;
>>>> -  llvm::Value *fp_offset_p = nullptr, *fp_offset = nullptr;
>>>> +  Address gp_offset_p = Address::invalid(), fp_offset_p =
>>>> Address::invalid();
>>>> +  llvm::Value *gp_offset = nullptr, *fp_offset = nullptr;
>>>>    if (neededInt) {
>>>>      gp_offset_p =
>>>> -        CGF.Builder.CreateStructGEP(nullptr, VAListAddr, 0,
>>>> "gp_offset_p");
>>>> +        CGF.Builder.CreateStructGEP(VAListAddr, 0, CharUnits::Zero(),
>>>> +                                    "gp_offset_p");
>>>>      gp_offset = CGF.Builder.CreateLoad(gp_offset_p, "gp_offset");
>>>>      InRegs = llvm::ConstantInt::get(CGF.Int32Ty, 48 - neededInt * 8);
>>>>      InRegs = CGF.Builder.CreateICmpULE(gp_offset, InRegs,
>>>> "fits_in_gp");
>>>> @@ -3005,7 +3119,8 @@ llvm::Value *X86_64ABIInfo::EmitVAArg(ll
>>>>
>>>>    if (neededSSE) {
>>>>      fp_offset_p =
>>>> -        CGF.Builder.CreateStructGEP(nullptr, VAListAddr, 1,
>>>> "fp_offset_p");
>>>> +        CGF.Builder.CreateStructGEP(VAListAddr, 1,
>>>> CharUnits::fromQuantity(4),
>>>> +                                    "fp_offset_p");
>>>>      fp_offset = CGF.Builder.CreateLoad(fp_offset_p, "fp_offset");
>>>>      llvm::Value *FitsInFP =
>>>>        llvm::ConstantInt::get(CGF.Int32Ty, 176 - neededSSE * 16);
>>>> @@ -3033,14 +3148,17 @@ llvm::Value *X86_64ABIInfo::EmitVAArg(ll
>>>>    // simple assembling of a structure from scattered addresses has
>>>> many more
>>>>    // loads than necessary. Can we clean this up?
>>>>    llvm::Type *LTy = CGF.ConvertTypeForMem(Ty);
>>>> -  llvm::Value *RegAddr = CGF.Builder.CreateLoad(
>>>> -      CGF.Builder.CreateStructGEP(nullptr, VAListAddr, 3),
>>>> "reg_save_area");
>>>> +  llvm::Value *RegSaveArea = CGF.Builder.CreateLoad(
>>>> +      CGF.Builder.CreateStructGEP(VAListAddr, 3,
>>>> CharUnits::fromQuantity(16)),
>>>> +                                  "reg_save_area");
>>>> +
>>>> +  Address RegAddr = Address::invalid();
>>>>    if (neededInt && neededSSE) {
>>>>      // FIXME: Cleanup.
>>>>      assert(AI.isDirect() && "Unexpected ABI info for mixed regs");
>>>>      llvm::StructType *ST =
>>>> cast<llvm::StructType>(AI.getCoerceToType());
>>>> -    llvm::Value *Tmp = CGF.CreateMemTemp(Ty);
>>>> -    Tmp = CGF.Builder.CreateBitCast(Tmp, ST->getPointerTo());
>>>> +    Address Tmp = CGF.CreateMemTemp(Ty);
>>>> +    Tmp = CGF.Builder.CreateElementBitCast(Tmp, ST);
>>>>      assert(ST->getNumElements() == 2 && "Unexpected ABI info for mixed
>>>> regs");
>>>>      llvm::Type *TyLo = ST->getElementType(0);
>>>>      llvm::Type *TyHi = ST->getElementType(1);
>>>> @@ -3048,57 +3166,77 @@ llvm::Value *X86_64ABIInfo::EmitVAArg(ll
>>>>             "Unexpected ABI info for mixed regs");
>>>>      llvm::Type *PTyLo = llvm::PointerType::getUnqual(TyLo);
>>>>      llvm::Type *PTyHi = llvm::PointerType::getUnqual(TyHi);
>>>> -    llvm::Value *GPAddr = CGF.Builder.CreateGEP(RegAddr, gp_offset);
>>>> -    llvm::Value *FPAddr = CGF.Builder.CreateGEP(RegAddr, fp_offset);
>>>> +    llvm::Value *GPAddr = CGF.Builder.CreateGEP(RegSaveArea,
>>>> gp_offset);
>>>> +    llvm::Value *FPAddr = CGF.Builder.CreateGEP(RegSaveArea,
>>>> fp_offset);
>>>>      llvm::Value *RegLoAddr = TyLo->isFPOrFPVectorTy() ? FPAddr :
>>>> GPAddr;
>>>>      llvm::Value *RegHiAddr = TyLo->isFPOrFPVectorTy() ? GPAddr :
>>>> FPAddr;
>>>> +
>>>> +    // Copy the first element.
>>>>      llvm::Value *V =
>>>> -      CGF.Builder.CreateLoad(CGF.Builder.CreateBitCast(RegLoAddr,
>>>> PTyLo));
>>>> -    CGF.Builder.CreateStore(V, CGF.Builder.CreateStructGEP(ST, Tmp,
>>>> 0));
>>>> -    V = CGF.Builder.CreateLoad(CGF.Builder.CreateBitCast(RegHiAddr,
>>>> PTyHi));
>>>> -    CGF.Builder.CreateStore(V, CGF.Builder.CreateStructGEP(ST, Tmp,
>>>> 1));
>>>> +      CGF.Builder.CreateDefaultAlignedLoad(
>>>> +                               CGF.Builder.CreateBitCast(RegLoAddr,
>>>> PTyLo));
>>>> +    CGF.Builder.CreateStore(V,
>>>> +                    CGF.Builder.CreateStructGEP(Tmp, 0,
>>>> CharUnits::Zero()));
>>>> +
>>>> +    // Copy the second element.
>>>> +    V = CGF.Builder.CreateDefaultAlignedLoad(
>>>> +                               CGF.Builder.CreateBitCast(RegHiAddr,
>>>> PTyHi));
>>>> +    CharUnits Offset = CharUnits::fromQuantity(
>>>> +
>>>>  getDataLayout().getStructLayout(ST)->getElementOffset(1));
>>>> +    CGF.Builder.CreateStore(V, CGF.Builder.CreateStructGEP(Tmp, 1,
>>>> Offset));
>>>>
>>>> -    RegAddr = CGF.Builder.CreateBitCast(Tmp,
>>>> -
>>>> llvm::PointerType::getUnqual(LTy));
>>>> +    RegAddr = CGF.Builder.CreateElementBitCast(Tmp, LTy);
>>>>    } else if (neededInt) {
>>>> -    RegAddr = CGF.Builder.CreateGEP(RegAddr, gp_offset);
>>>> -    RegAddr = CGF.Builder.CreateBitCast(RegAddr,
>>>> -
>>>> llvm::PointerType::getUnqual(LTy));
>>>> +    RegAddr = Address(CGF.Builder.CreateGEP(RegSaveArea, gp_offset),
>>>> +                      CharUnits::fromQuantity(8));
>>>> +    RegAddr = CGF.Builder.CreateElementBitCast(RegAddr, LTy);
>>>>
>>>>      // Copy to a temporary if necessary to ensure the appropriate
>>>> alignment.
>>>>      std::pair<CharUnits, CharUnits> SizeAlign =
>>>> -        CGF.getContext().getTypeInfoInChars(Ty);
>>>> +        getContext().getTypeInfoInChars(Ty);
>>>>      uint64_t TySize = SizeAlign.first.getQuantity();
>>>> -    unsigned TyAlign = SizeAlign.second.getQuantity();
>>>> -    if (TyAlign > 8) {
>>>> -      llvm::Value *Tmp = CGF.CreateMemTemp(Ty);
>>>> -      CGF.Builder.CreateMemCpy(Tmp, RegAddr, TySize, 8, false);
>>>> +    CharUnits TyAlign = SizeAlign.second;
>>>> +
>>>> +    // Copy into a temporary if the type is more aligned than the
>>>> +    // register save area.
>>>> +    if (TyAlign.getQuantity() > 8) {
>>>> +      Address Tmp = CGF.CreateMemTemp(Ty);
>>>> +      CGF.Builder.CreateMemCpy(Tmp, RegAddr, TySize, false);
>>>>        RegAddr = Tmp;
>>>>      }
>>>> +
>>>>    } else if (neededSSE == 1) {
>>>> -    RegAddr = CGF.Builder.CreateGEP(RegAddr, fp_offset);
>>>> -    RegAddr = CGF.Builder.CreateBitCast(RegAddr,
>>>> -
>>>> llvm::PointerType::getUnqual(LTy));
>>>> +    RegAddr = Address(CGF.Builder.CreateGEP(RegSaveArea, fp_offset),
>>>> +                      CharUnits::fromQuantity(16));
>>>> +    RegAddr = CGF.Builder.CreateElementBitCast(RegAddr, LTy);
>>>>    } else {
>>>>      assert(neededSSE == 2 && "Invalid number of needed registers!");
>>>>      // SSE registers are spaced 16 bytes apart in the register save
>>>>      // area, we need to collect the two eightbytes together.
>>>> -    llvm::Value *RegAddrLo = CGF.Builder.CreateGEP(RegAddr, fp_offset);
>>>> -    llvm::Value *RegAddrHi = CGF.Builder.CreateConstGEP1_32(RegAddrLo,
>>>> 16);
>>>> +    // The ABI isn't explicit about this, but it seems reasonable
>>>> +    // to assume that the slots are 16-byte aligned, since the stack is
>>>> +    // naturally 16-byte aligned and the prologue is expected to store
>>>> +    // all the SSE registers to the RSA.
>>>> +    Address RegAddrLo = Address(CGF.Builder.CreateGEP(RegSaveArea,
>>>> fp_offset),
>>>> +                                CharUnits::fromQuantity(16));
>>>> +    Address RegAddrHi =
>>>> +      CGF.Builder.CreateConstInBoundsByteGEP(RegAddrLo,
>>>> +
>>>>  CharUnits::fromQuantity(16));
>>>>      llvm::Type *DoubleTy = CGF.DoubleTy;
>>>> -    llvm::Type *DblPtrTy =
>>>> -      llvm::PointerType::getUnqual(DoubleTy);
>>>>      llvm::StructType *ST = llvm::StructType::get(DoubleTy, DoubleTy,
>>>> nullptr);
>>>> -    llvm::Value *V, *Tmp = CGF.CreateMemTemp(Ty);
>>>> -    Tmp = CGF.Builder.CreateBitCast(Tmp, ST->getPointerTo());
>>>> -    V = CGF.Builder.CreateLoad(CGF.Builder.CreateBitCast(RegAddrLo,
>>>> -                                                         DblPtrTy));
>>>> -    CGF.Builder.CreateStore(V, CGF.Builder.CreateStructGEP(ST, Tmp,
>>>> 0));
>>>> -    V = CGF.Builder.CreateLoad(CGF.Builder.CreateBitCast(RegAddrHi,
>>>> -                                                         DblPtrTy));
>>>> -    CGF.Builder.CreateStore(V, CGF.Builder.CreateStructGEP(ST, Tmp,
>>>> 1));
>>>> -    RegAddr = CGF.Builder.CreateBitCast(Tmp,
>>>> -
>>>> llvm::PointerType::getUnqual(LTy));
>>>> +    llvm::Value *V;
>>>> +    Address Tmp = CGF.CreateMemTemp(Ty);
>>>> +    Tmp = CGF.Builder.CreateElementBitCast(Tmp, ST);
>>>> +    V = CGF.Builder.CreateLoad(
>>>> +                   CGF.Builder.CreateElementBitCast(RegAddrLo,
>>>> DoubleTy));
>>>> +    CGF.Builder.CreateStore(V,
>>>> +                   CGF.Builder.CreateStructGEP(Tmp, 0,
>>>> CharUnits::Zero()));
>>>> +    V = CGF.Builder.CreateLoad(
>>>> +                   CGF.Builder.CreateElementBitCast(RegAddrHi,
>>>> DoubleTy));
>>>> +    CGF.Builder.CreateStore(V,
>>>> +          CGF.Builder.CreateStructGEP(Tmp, 1,
>>>> CharUnits::fromQuantity(8)));
>>>> +
>>>> +    RegAddr = CGF.Builder.CreateElementBitCast(Tmp, LTy);
>>>>    }
>>>>
>>>>    // AMD64-ABI 3.5.7p5: Step 5. Set:
>>>> @@ -3119,15 +3257,13 @@ llvm::Value *X86_64ABIInfo::EmitVAArg(ll
>>>>    // Emit code to load the value if it was passed in memory.
>>>>
>>>>    CGF.EmitBlock(InMemBlock);
>>>> -  llvm::Value *MemAddr = EmitVAArgFromMemory(VAListAddr, Ty, CGF);
>>>> +  Address MemAddr = EmitX86_64VAArgFromMemory(CGF, VAListAddr, Ty);
>>>>
>>>>    // Return the appropriate result.
>>>>
>>>>    CGF.EmitBlock(ContBlock);
>>>> -  llvm::PHINode *ResAddr = CGF.Builder.CreatePHI(RegAddr->getType(), 2,
>>>> -                                                 "vaarg.addr");
>>>> -  ResAddr->addIncoming(RegAddr, InRegBlock);
>>>> -  ResAddr->addIncoming(MemAddr, InMemBlock);
>>>> +  Address ResAddr = emitMergePHI(CGF, RegAddr, InRegBlock, MemAddr,
>>>> InMemBlock,
>>>> +                                 "vaarg.addr");
>>>>    return ResAddr;
>>>>  }
>>>>
>>>> @@ -3148,11 +3284,11 @@ ABIArgInfo WinX86_64ABIInfo::classify(Qu
>>>>    if (RT) {
>>>>      if (!IsReturnType) {
>>>>        if (CGCXXABI::RecordArgABI RAA = getRecordArgABI(RT,
>>>> getCXXABI()))
>>>> -        return ABIArgInfo::getIndirect(0, RAA ==
>>>> CGCXXABI::RAA_DirectInMemory);
>>>> +        return getNaturalAlignIndirect(Ty, RAA ==
>>>> CGCXXABI::RAA_DirectInMemory);
>>>>      }
>>>>
>>>>      if (RT->getDecl()->hasFlexibleArrayMember())
>>>> -      return ABIArgInfo::getIndirect(0, /*ByVal=*/false);
>>>> +      return getNaturalAlignIndirect(Ty, /*ByVal=*/false);
>>>>
>>>>      // FIXME: mingw-w64-gcc emits 128-bit struct as i128
>>>>      if (Width == 128 &&
>>>> getTarget().getTriple().isWindowsGNUEnvironment())
>>>> @@ -3171,7 +3307,8 @@ ABIArgInfo WinX86_64ABIInfo::classify(Qu
>>>>          return ABIArgInfo::getDirect();
>>>>        return ABIArgInfo::getExpand();
>>>>      }
>>>> -    return ABIArgInfo::getIndirect(Align, /*ByVal=*/false);
>>>> +    return ABIArgInfo::getIndirect(CharUnits::fromQuantity(Align),
>>>> +                                   /*ByVal=*/false);
>>>>    }
>>>>
>>>>
>>>> @@ -3187,7 +3324,7 @@ ABIArgInfo WinX86_64ABIInfo::classify(Qu
>>>>      // MS x64 ABI requirement: "Any argument that doesn't fit in 8
>>>> bytes, or is
>>>>      // not 1, 2, 4, or 8 bytes, must be passed by reference."
>>>>      if (Width > 64 || !llvm::isPowerOf2_64(Width))
>>>> -      return ABIArgInfo::getIndirect(0, /*ByVal=*/false);
>>>> +      return getNaturalAlignIndirect(Ty, /*ByVal=*/false);
>>>>
>>>>      // Otherwise, coerce it to a small integer.
>>>>      return
>>>> ABIArgInfo::getDirect(llvm::IntegerType::get(getVMContext(), Width));
>>>> @@ -3217,26 +3354,12 @@ void WinX86_64ABIInfo::computeInfo(CGFun
>>>>      I.info <http://i.info> = classify(I.type, FreeSSERegs, false);
>>>>  }
>>>>
>>>> -llvm::Value *WinX86_64ABIInfo::EmitVAArg(llvm::Value *VAListAddr,
>>>> QualType Ty,
>>>> -                                      CodeGenFunction &CGF) const {
>>>> -  llvm::Type *BPP = CGF.Int8PtrPtrTy;
>>>> -
>>>> -  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, 8);
>>>> -  llvm::Value *NextAddr =
>>>> -    Builder.CreateGEP(Addr, llvm::ConstantInt::get(CGF.Int32Ty,
>>>> Offset),
>>>> -                      "ap.next");
>>>> -  Builder.CreateStore(NextAddr, VAListAddrAsBPP);
>>>> -
>>>> -  return AddrTyped;
>>>> +Address WinX86_64ABIInfo::EmitVAArg(CodeGenFunction &CGF, Address
>>>> VAListAddr,
>>>> +                                    QualType Ty) const {
>>>> +  return emitVoidPtrVAArg(CGF, VAListAddr, Ty, /*indirect*/ false,
>>>> +                          CGF.getContext().getTypeInfoInChars(Ty),
>>>> +                          CharUnits::fromQuantity(8),
>>>> +                          /*allowHigherAlign*/ false);
>>>>  }
>>>>
>>>>  // PowerPC-32
>>>> @@ -3246,8 +3369,8 @@ class PPC32_SVR4_ABIInfo : public Defaul
>>>>  public:
>>>>    PPC32_SVR4_ABIInfo(CodeGen::CodeGenTypes &CGT) : DefaultABIInfo(CGT)
>>>> {}
>>>>
>>>> -  llvm::Value *EmitVAArg(llvm::Value *VAListAddr, QualType Ty,
>>>> -                         CodeGenFunction &CGF) const override;
>>>> +  Address EmitVAArg(CodeGenFunction &CGF, Address VAListAddr,
>>>> +                    QualType Ty) const override;
>>>>  };
>>>>
>>>>  class PPC32TargetCodeGenInfo : public TargetCodeGenInfo {
>>>> @@ -3266,64 +3389,50 @@ public:
>>>>
>>>>  }
>>>>
>>>> -llvm::Value *PPC32_SVR4_ABIInfo::EmitVAArg(llvm::Value *VAListAddr,
>>>> -                                           QualType Ty,
>>>> -                                           CodeGenFunction &CGF) const
>>>> {
>>>> +Address PPC32_SVR4_ABIInfo::EmitVAArg(CodeGenFunction &CGF, Address
>>>> VAList,
>>>> +                                      QualType Ty) const {
>>>>    if (const ComplexType *CTy = Ty->getAs<ComplexType>()) {
>>>>      // TODO: Implement this. For now ignore.
>>>>      (void)CTy;
>>>> -    return nullptr;
>>>> +    return Address::invalid();
>>>>    }
>>>>
>>>> +  // struct __va_list_tag {
>>>> +  //   unsigned char gpr;
>>>> +  //   unsigned char fpr;
>>>> +  //   unsigned short reserved;
>>>> +  //   void *overflow_arg_area;
>>>> +  //   void *reg_save_area;
>>>> +  // };
>>>> +
>>>>    bool isI64 = Ty->isIntegerType() && getContext().getTypeSize(Ty) ==
>>>> 64;
>>>>    bool isInt =
>>>>        Ty->isIntegerType() || Ty->isPointerType() ||
>>>> Ty->isAggregateType();
>>>> -  llvm::Type *CharPtr = CGF.Int8PtrTy;
>>>> -  llvm::Type *CharPtrPtr = CGF.Int8PtrPtrTy;
>>>> +
>>>> +  // All aggregates are passed indirectly?  That doesn't seem
>>>> consistent
>>>> +  // with the argument-lowering code.
>>>> +  bool isIndirect = Ty->isAggregateType();
>>>>
>>>>    CGBuilderTy &Builder = CGF.Builder;
>>>> -  llvm::Value *GPRPtr = Builder.CreateBitCast(VAListAddr, CharPtr,
>>>> "gprptr");
>>>> -  llvm::Value *GPRPtrAsInt = Builder.CreatePtrToInt(GPRPtr,
>>>> CGF.Int32Ty);
>>>> -  llvm::Value *FPRPtrAsInt =
>>>> -      Builder.CreateAdd(GPRPtrAsInt, Builder.getInt32(1));
>>>> -  llvm::Value *FPRPtr = Builder.CreateIntToPtr(FPRPtrAsInt, CharPtr);
>>>> -  llvm::Value *OverflowAreaPtrAsInt =
>>>> -      Builder.CreateAdd(FPRPtrAsInt, Builder.getInt32(3));
>>>> -  llvm::Value *OverflowAreaPtr =
>>>> -      Builder.CreateIntToPtr(OverflowAreaPtrAsInt, CharPtrPtr);
>>>> -  llvm::Value *RegsaveAreaPtrAsInt =
>>>> -      Builder.CreateAdd(OverflowAreaPtrAsInt, Builder.getInt32(4));
>>>> -  llvm::Value *RegsaveAreaPtr =
>>>> -      Builder.CreateIntToPtr(RegsaveAreaPtrAsInt, CharPtrPtr);
>>>> -  llvm::Value *GPR = Builder.CreateLoad(GPRPtr, false, "gpr");
>>>> -  // Align GPR when TY is i64.
>>>> -  if (isI64) {
>>>> -    llvm::Value *GPRAnd = Builder.CreateAnd(GPR, Builder.getInt8(1));
>>>> -    llvm::Value *CC64 = Builder.CreateICmpEQ(GPRAnd,
>>>> Builder.getInt8(1));
>>>> -    llvm::Value *GPRPlusOne = Builder.CreateAdd(GPR,
>>>> Builder.getInt8(1));
>>>> -    GPR = Builder.CreateSelect(CC64, GPRPlusOne, GPR);
>>>> -  }
>>>> -  llvm::Value *FPR = Builder.CreateLoad(FPRPtr, false, "fpr");
>>>> -  llvm::Value *OverflowArea =
>>>> -      Builder.CreateLoad(OverflowAreaPtr, false, "overflow_area");
>>>> -  llvm::Value *OverflowAreaAsInt =
>>>> -      Builder.CreatePtrToInt(OverflowArea, CGF.Int32Ty);
>>>> -  llvm::Value *RegsaveArea =
>>>> -      Builder.CreateLoad(RegsaveAreaPtr, false, "regsave_area");
>>>> -  llvm::Value *RegsaveAreaAsInt =
>>>> -      Builder.CreatePtrToInt(RegsaveArea, CGF.Int32Ty);
>>>>
>>>> -  llvm::Value *CC =
>>>> -      Builder.CreateICmpULT(isInt ? GPR : FPR, Builder.getInt8(8),
>>>> "cond");
>>>> +  // The calling convention either uses 1-2 GPRs or 1 FPR.
>>>> +  Address NumRegsAddr = Address::invalid();
>>>> +  if (isInt) {
>>>> +    NumRegsAddr = Builder.CreateStructGEP(VAList, 0,
>>>> CharUnits::Zero(), "gpr");
>>>> +  } else {
>>>> +    NumRegsAddr = Builder.CreateStructGEP(VAList, 1, CharUnits::One(),
>>>> "fpr");
>>>> +  }
>>>>
>>>> -  llvm::Value *RegConstant =
>>>> -      Builder.CreateMul(isInt ? GPR : FPR, Builder.getInt8(isInt ? 4 :
>>>> 8));
>>>> +  llvm::Value *NumRegs = Builder.CreateLoad(NumRegsAddr,
>>>> "numUsedRegs");
>>>>
>>>> -  llvm::Value *OurReg = Builder.CreateAdd(
>>>> -      RegsaveAreaAsInt, Builder.CreateSExt(RegConstant, CGF.Int32Ty));
>>>> +  // "Align" the register count when TY is i64.
>>>> +  if (isI64) {
>>>> +    NumRegs = Builder.CreateAdd(NumRegs, Builder.getInt8(1));
>>>> +    NumRegs = Builder.CreateAnd(NumRegs, Builder.getInt8((uint8_t)
>>>> ~1U));
>>>> +  }
>>>>
>>>> -  if (Ty->isFloatingType())
>>>> -    OurReg = Builder.CreateAdd(OurReg, Builder.getInt32(32));
>>>> +  llvm::Value *CC =
>>>> +      Builder.CreateICmpULT(NumRegs, Builder.getInt8(8), "cond");
>>>>
>>>>    llvm::BasicBlock *UsingRegs = CGF.createBasicBlock("using_regs");
>>>>    llvm::BasicBlock *UsingOverflow =
>>>> CGF.createBasicBlock("using_overflow");
>>>> @@ -3331,39 +3440,84 @@ llvm::Value *PPC32_SVR4_ABIInfo::EmitVAA
>>>>
>>>>    Builder.CreateCondBr(CC, UsingRegs, UsingOverflow);
>>>>
>>>> -  CGF.EmitBlock(UsingRegs);
>>>> +  llvm::Type *DirectTy = CGF.ConvertType(Ty);
>>>> +  if (isIndirect) DirectTy = DirectTy->getPointerTo(0);
>>>>
>>>> -  llvm::Type *PTy = llvm::PointerType::getUnqual(CGF.ConvertType(Ty));
>>>> -  llvm::Value *Result1 = Builder.CreateIntToPtr(OurReg, PTy);
>>>> -  // Increase the GPR/FPR indexes.
>>>> -  if (isInt) {
>>>> -    GPR = Builder.CreateAdd(GPR, Builder.getInt8(isI64 ? 2 : 1));
>>>> -    Builder.CreateStore(GPR, GPRPtr);
>>>> -  } else {
>>>> -    FPR = Builder.CreateAdd(FPR, Builder.getInt8(1));
>>>> -    Builder.CreateStore(FPR, FPRPtr);
>>>> -  }
>>>> -  CGF.EmitBranch(Cont);
>>>> -
>>>> -  CGF.EmitBlock(UsingOverflow);
>>>> +  // Case 1: consume registers.
>>>> +  Address RegAddr = Address::invalid();
>>>> +  {
>>>> +    CGF.EmitBlock(UsingRegs);
>>>> +
>>>> +    Address RegSaveAreaPtr =
>>>> +      Builder.CreateStructGEP(VAList, 4, CharUnits::fromQuantity(8));
>>>> +    RegAddr = Address(Builder.CreateLoad(RegSaveAreaPtr),
>>>> +                      CharUnits::fromQuantity(8));
>>>> +    assert(RegAddr.getElementType() == CGF.Int8Ty);
>>>> +
>>>> +    // Floating-point registers start after the general-purpose
>>>> registers.
>>>> +    if (!isInt) {
>>>> +      RegAddr = Builder.CreateConstInBoundsByteGEP(RegAddr,
>>>> +
>>>>  CharUnits::fromQuantity(32));
>>>> +    }
>>>> +
>>>> +    // Get the address of the saved value by scaling the number of
>>>> +    // registers we've used by the number of
>>>> +    CharUnits RegSize = CharUnits::fromQuantity(isInt ? 4 : 8);
>>>> +    llvm::Value *RegOffset =
>>>> +      Builder.CreateMul(NumRegs,
>>>> Builder.getInt8(RegSize.getQuantity()));
>>>> +    RegAddr = Address(Builder.CreateInBoundsGEP(CGF.Int8Ty,
>>>> +                                            RegAddr.getPointer(),
>>>> RegOffset),
>>>> +
>>>> RegAddr.getAlignment().alignmentOfArrayElement(RegSize));
>>>> +    RegAddr = Builder.CreateElementBitCast(RegAddr, DirectTy);
>>>> +
>>>> +    // Increase the used-register count.
>>>> +    NumRegs = Builder.CreateAdd(NumRegs, Builder.getInt8(isI64 ? 2 :
>>>> 1));
>>>> +    Builder.CreateStore(NumRegs, NumRegsAddr);
>>>> +
>>>> +    CGF.EmitBranch(Cont);
>>>> +  }
>>>> +
>>>> +  // Case 2: consume space in the overflow area.
>>>> +  Address MemAddr = Address::invalid();
>>>> +  {
>>>> +    CGF.EmitBlock(UsingOverflow);
>>>> +
>>>> +    // Everything in the overflow area is rounded up to a size of at
>>>> least 4.
>>>> +    CharUnits OverflowAreaAlign = CharUnits::fromQuantity(4);
>>>> +
>>>> +    CharUnits Size;
>>>> +    if (!isIndirect) {
>>>> +      auto TypeInfo = CGF.getContext().getTypeInfoInChars(Ty);
>>>> +      Size = TypeInfo.first.RoundUpToAlignment(OverflowAreaAlign);
>>>> +    } else {
>>>> +      Size = CGF.getPointerSize();
>>>> +    }
>>>>
>>>> -  // Increase the overflow area.
>>>> -  llvm::Value *Result2 = Builder.CreateIntToPtr(OverflowAreaAsInt,
>>>> PTy);
>>>> -  OverflowAreaAsInt =
>>>> -      Builder.CreateAdd(OverflowAreaAsInt, Builder.getInt32(isInt ? 4
>>>> : 8));
>>>> -  Builder.CreateStore(Builder.CreateIntToPtr(OverflowAreaAsInt,
>>>> CharPtr),
>>>> -                      OverflowAreaPtr);
>>>> -  CGF.EmitBranch(Cont);
>>>> +    Address OverflowAreaAddr =
>>>> +      Builder.CreateStructGEP(VAList, 3, CharUnits::fromQuantity(4));
>>>> +    Address OverflowArea(Builder.CreateLoad(OverflowAreaAddr),
>>>> +                         OverflowAreaAlign);
>>>> +
>>>> +    // The current address is the address of the varargs element.
>>>> +    // FIXME: do we not need to round up to alignment?
>>>> +    MemAddr = Builder.CreateElementBitCast(OverflowArea, DirectTy);
>>>> +
>>>> +    // Increase the overflow area.
>>>> +    OverflowArea = Builder.CreateConstInBoundsByteGEP(OverflowArea,
>>>> Size);
>>>> +    Builder.CreateStore(OverflowArea.getPointer(), OverflowAreaAddr);
>>>> +    CGF.EmitBranch(Cont);
>>>> +  }
>>>>
>>>>    CGF.EmitBlock(Cont);
>>>>
>>>> -  llvm::PHINode *Result = CGF.Builder.CreatePHI(PTy, 2, "vaarg.addr");
>>>> -  Result->addIncoming(Result1, UsingRegs);
>>>> -  Result->addIncoming(Result2, UsingOverflow);
>>>> -
>>>> -  if (Ty->isAggregateType()) {
>>>> -    llvm::Value *AGGPtr = Builder.CreateBitCast(Result, CharPtrPtr,
>>>> "aggrptr");
>>>> -    return Builder.CreateLoad(AGGPtr, false, "aggr");
>>>> +  // Merge the cases with a phi.
>>>> +  Address Result = emitMergePHI(CGF, RegAddr, UsingRegs, MemAddr,
>>>> UsingOverflow,
>>>> +                                "vaarg.addr");
>>>> +
>>>> +  // Load the pointer if the argument was passed indirectly.
>>>> +  if (isIndirect) {
>>>> +    Result = Address(Builder.CreateLoad(Result, "aggr"),
>>>> +                     getContext().getTypeAlignInChars(Ty));
>>>>    }
>>>>
>>>>    return Result;
>>>> @@ -3459,7 +3613,7 @@ public:
>>>>      : DefaultABIInfo(CGT), Kind(Kind), HasQPX(HasQPX) {}
>>>>
>>>>    bool isPromotableTypeForABI(QualType Ty) const;
>>>> -  bool isAlignedParamType(QualType Ty, bool &Align32) const;
>>>> +  CharUnits getParamTypeAlignment(QualType Ty) const;
>>>>
>>>>    ABIArgInfo classifyReturnType(QualType RetTy) const;
>>>>    ABIArgInfo classifyArgumentType(QualType Ty) const;
>>>> @@ -3496,8 +3650,8 @@ public:
>>>>      }
>>>>    }
>>>>
>>>> -  llvm::Value *EmitVAArg(llvm::Value *VAListAddr, QualType Ty,
>>>> -                         CodeGenFunction &CGF) const override;
>>>> +  Address EmitVAArg(CodeGenFunction &CGF, Address VAListAddr,
>>>> +                    QualType Ty) const override;
>>>>  };
>>>>
>>>>  class PPC64_SVR4_TargetCodeGenInfo : public TargetCodeGenInfo {
>>>> @@ -3557,12 +3711,9 @@ PPC64_SVR4_ABIInfo::isPromotableTypeForA
>>>>    return false;
>>>>  }
>>>>
>>>> -/// isAlignedParamType - Determine whether a type requires 16-byte
>>>> -/// alignment in the parameter area.
>>>> -bool
>>>> -PPC64_SVR4_ABIInfo::isAlignedParamType(QualType Ty, bool &Align32)
>>>> const {
>>>> -  Align32 = false;
>>>> -
>>>> +/// isAlignedParamType - Determine whether a type requires 16-byte or
>>>> +/// higher alignment in the parameter area.  Always returns at least 8.
>>>> +CharUnits PPC64_SVR4_ABIInfo::getParamTypeAlignment(QualType Ty) const
>>>> {
>>>>    // Complex types are passed just like their elements.
>>>>    if (const ComplexType *CTy = Ty->getAs<ComplexType>())
>>>>      Ty = CTy->getElementType();
>>>> @@ -3571,11 +3722,11 @@ PPC64_SVR4_ABIInfo::isAlignedParamType(Q
>>>>    // passed via reference, smaller types are not aligned).
>>>>    if (IsQPXVectorTy(Ty)) {
>>>>      if (getContext().getTypeSize(Ty) > 128)
>>>> -      Align32 = true;
>>>> +      return CharUnits::fromQuantity(32);
>>>>
>>>> -    return true;
>>>> +    return CharUnits::fromQuantity(16);
>>>>    } else if (Ty->isVectorType()) {
>>>> -    return getContext().getTypeSize(Ty) == 128;
>>>> +    return CharUnits::fromQuantity(getContext().getTypeSize(Ty) == 128
>>>> ? 16 : 8);
>>>>    }
>>>>
>>>>    // For single-element float/vector structs, we consider the whole
>>>> type
>>>> @@ -3600,22 +3751,22 @@ PPC64_SVR4_ABIInfo::isAlignedParamType(Q
>>>>    // With special case aggregates, only vector base types need
>>>> alignment.
>>>>    if (AlignAsType && IsQPXVectorTy(AlignAsType)) {
>>>>      if (getContext().getTypeSize(AlignAsType) > 128)
>>>> -      Align32 = true;
>>>> +      return CharUnits::fromQuantity(32);
>>>>
>>>> -    return true;
>>>> +    return CharUnits::fromQuantity(16);
>>>>    } else if (AlignAsType) {
>>>> -    return AlignAsType->isVectorType();
>>>> +    return CharUnits::fromQuantity(AlignAsType->isVectorType() ? 16 :
>>>> 8);
>>>>    }
>>>>
>>>>    // Otherwise, we only need alignment for any aggregate type that
>>>>    // has an alignment requirement of >= 16 bytes.
>>>>    if (isAggregateTypeForABI(Ty) && getContext().getTypeAlign(Ty) >=
>>>> 128) {
>>>>      if (HasQPX && getContext().getTypeAlign(Ty) >= 256)
>>>> -      Align32 = true;
>>>> -    return true;
>>>> +      return CharUnits::fromQuantity(32);
>>>> +    return CharUnits::fromQuantity(16);
>>>>    }
>>>>
>>>> -  return false;
>>>> +  return CharUnits::fromQuantity(8);
>>>>  }
>>>>
>>>>  /// isHomogeneousAggregate - Return true if a type is an ELFv2
>>>> homogeneous
>>>> @@ -3748,7 +3899,7 @@ PPC64_SVR4_ABIInfo::classifyArgumentType
>>>>    if (Ty->isVectorType() && !IsQPXVectorTy(Ty)) {
>>>>      uint64_t Size = getContext().getTypeSize(Ty);
>>>>      if (Size > 128)
>>>> -      return ABIArgInfo::getIndirect(0, /*ByVal=*/false);
>>>> +      return getNaturalAlignIndirect(Ty, /*ByVal=*/false);
>>>>      else if (Size < 128) {
>>>>        llvm::Type *CoerceTy = llvm::IntegerType::get(getVMContext(),
>>>> Size);
>>>>        return ABIArgInfo::getDirect(CoerceTy);
>>>> @@ -3757,12 +3908,10 @@ PPC64_SVR4_ABIInfo::classifyArgumentType
>>>>
>>>>    if (isAggregateTypeForABI(Ty)) {
>>>>      if (CGCXXABI::RecordArgABI RAA = getRecordArgABI(Ty, getCXXABI()))
>>>> -      return ABIArgInfo::getIndirect(0, RAA ==
>>>> CGCXXABI::RAA_DirectInMemory);
>>>> +      return getNaturalAlignIndirect(Ty, RAA ==
>>>> CGCXXABI::RAA_DirectInMemory);
>>>>
>>>> -    bool Align32;
>>>> -    uint64_t ABIAlign = isAlignedParamType(Ty, Align32) ?
>>>> -                          (Align32 ? 32 : 16) : 8;
>>>> -    uint64_t TyAlign = getContext().getTypeAlign(Ty) / 8;
>>>> +    uint64_t ABIAlign = getParamTypeAlignment(Ty).getQuantity();
>>>> +    uint64_t TyAlign =
>>>> getContext().getTypeAlignInChars(Ty).getQuantity();
>>>>
>>>>      // ELFv2 homogeneous aggregates are passed as array types.
>>>>      const Type *Base = nullptr;
>>>> @@ -3800,7 +3949,8 @@ PPC64_SVR4_ABIInfo::classifyArgumentType
>>>>      }
>>>>
>>>>      // All other aggregates are passed ByVal.
>>>> -    return ABIArgInfo::getIndirect(ABIAlign, /*ByVal=*/true,
>>>> +    return ABIArgInfo::getIndirect(CharUnits::fromQuantity(ABIAlign),
>>>> +                                   /*ByVal=*/true,
>>>>                                     /*Realign=*/TyAlign > ABIAlign);
>>>>    }
>>>>
>>>> @@ -3821,7 +3971,7 @@ PPC64_SVR4_ABIInfo::classifyReturnType(Q
>>>>    if (RetTy->isVectorType() && !IsQPXVectorTy(RetTy)) {
>>>>      uint64_t Size = getContext().getTypeSize(RetTy);
>>>>      if (Size > 128)
>>>> -      return ABIArgInfo::getIndirect(0);
>>>> +      return getNaturalAlignIndirect(RetTy);
>>>>      else if (Size < 128) {
>>>>        llvm::Type *CoerceTy = llvm::IntegerType::get(getVMContext(),
>>>> Size);
>>>>        return ABIArgInfo::getDirect(CoerceTy);
>>>> @@ -3856,7 +4006,7 @@ PPC64_SVR4_ABIInfo::classifyReturnType(Q
>>>>      }
>>>>
>>>>      // All other aggregates are returned indirectly.
>>>> -    return ABIArgInfo::getIndirect(0);
>>>> +    return getNaturalAlignIndirect(RetTy);
>>>>    }
>>>>
>>>>    return (isPromotableTypeForABI(RetTy) ?
>>>> @@ -3864,47 +4014,12 @@ PPC64_SVR4_ABIInfo::classifyReturnType(Q
>>>>  }
>>>>
>>>>  // Based on ARMABIInfo::EmitVAArg, adjusted for 64-bit machine.
>>>> -llvm::Value *PPC64_SVR4_ABIInfo::EmitVAArg(llvm::Value *VAListAddr,
>>>> -                                           QualType Ty,
>>>> -                                           CodeGenFunction &CGF) const
>>>> {
>>>> -  llvm::Type *BP = CGF.Int8PtrTy;
>>>> -  llvm::Type *BPP = CGF.Int8PtrPtrTy;
>>>> -
>>>> -  CGBuilderTy &Builder = CGF.Builder;
>>>> -  llvm::Value *VAListAddrAsBPP = Builder.CreateBitCast(VAListAddr,
>>>> BPP, "ap");
>>>> -  llvm::Value *Addr = Builder.CreateLoad(VAListAddrAsBPP, "ap.cur");
>>>> -
>>>> -  // Handle types that require 16-byte alignment in the parameter save
>>>> area.
>>>> -  bool Align32;
>>>> -  if (isAlignedParamType(Ty, Align32)) {
>>>> -    llvm::Value *AddrAsInt = Builder.CreatePtrToInt(Addr, CGF.Int64Ty);
>>>> -    AddrAsInt = Builder.CreateAdd(AddrAsInt,
>>>> -                                  Builder.getInt64(Align32 ? 31 : 15));
>>>> -    AddrAsInt = Builder.CreateAnd(AddrAsInt,
>>>> -                                  Builder.getInt64(Align32 ? -32 :
>>>> -16));
>>>> -    Addr = Builder.CreateIntToPtr(AddrAsInt, BP, "ap.align");
>>>> -  }
>>>> -
>>>> -  // Update the va_list pointer.  The pointer should be bumped by the
>>>> -  // size of the object.  We can trust getTypeSize() except for a
>>>> complex
>>>> -  // type whose base type is smaller than a doubleword.  For these, the
>>>> -  // size of the object is 16 bytes; see below for further explanation.
>>>> -  unsigned SizeInBytes = CGF.getContext().getTypeSize(Ty) / 8;
>>>> -  QualType BaseTy;
>>>> -  unsigned CplxBaseSize = 0;
>>>> -
>>>> -  if (const ComplexType *CTy = Ty->getAs<ComplexType>()) {
>>>> -    BaseTy = CTy->getElementType();
>>>> -    CplxBaseSize = CGF.getContext().getTypeSize(BaseTy) / 8;
>>>> -    if (CplxBaseSize < 8)
>>>> -      SizeInBytes = 16;
>>>> -  }
>>>> +Address PPC64_SVR4_ABIInfo::EmitVAArg(CodeGenFunction &CGF, Address
>>>> VAListAddr,
>>>> +                                      QualType Ty) const {
>>>> +  auto TypeInfo = getContext().getTypeInfoInChars(Ty);
>>>> +  TypeInfo.second = getParamTypeAlignment(Ty);
>>>>
>>>> -  unsigned Offset = llvm::RoundUpToAlignment(SizeInBytes, 8);
>>>> -  llvm::Value *NextAddr =
>>>> -    Builder.CreateGEP(Addr, llvm::ConstantInt::get(CGF.Int64Ty,
>>>> Offset),
>>>> -                      "ap.next");
>>>> -  Builder.CreateStore(NextAddr, VAListAddrAsBPP);
>>>> +  CharUnits SlotSize = CharUnits::fromQuantity(8);
>>>>
>>>>    // If we have a complex type and the base type is smaller than 8
>>>> bytes,
>>>>    // the ABI calls for the real and imaginary parts to be
>>>> right-adjusted
>>>> @@ -3912,44 +4027,40 @@ llvm::Value *PPC64_SVR4_ABIInfo::EmitVAA
>>>>    // pointer to a structure with the two parts packed tightly.  So
>>>> generate
>>>>    // loads of the real and imaginary parts relative to the va_list
>>>> pointer,
>>>>    // and store them to a temporary structure.
>>>> -  if (CplxBaseSize && CplxBaseSize < 8) {
>>>> -    llvm::Value *RealAddr = Builder.CreatePtrToInt(Addr, CGF.Int64Ty);
>>>> -    llvm::Value *ImagAddr = RealAddr;
>>>> -    if (CGF.CGM.getDataLayout().isBigEndian()) {
>>>> -      RealAddr =
>>>> -          Builder.CreateAdd(RealAddr, Builder.getInt64(8 -
>>>> CplxBaseSize));
>>>> -      ImagAddr =
>>>> -          Builder.CreateAdd(ImagAddr, Builder.getInt64(16 -
>>>> CplxBaseSize));
>>>> -    } else {
>>>> -      ImagAddr = Builder.CreateAdd(ImagAddr, Builder.getInt64(8));
>>>> +  if (const ComplexType *CTy = Ty->getAs<ComplexType>()) {
>>>> +    CharUnits EltSize = TypeInfo.first / 2;
>>>> +    if (EltSize < SlotSize) {
>>>> +      Address Addr = emitVoidPtrDirectVAArg(CGF, VAListAddr,
>>>> CGF.Int8Ty,
>>>> +                                            SlotSize * 2, SlotSize,
>>>> +                                            SlotSize, /*AllowHigher*/
>>>> true);
>>>> +
>>>> +      Address RealAddr = Addr;
>>>> +      Address ImagAddr = RealAddr;
>>>> +      if (CGF.CGM.getDataLayout().isBigEndian()) {
>>>> +        RealAddr = CGF.Builder.CreateConstInBoundsByteGEP(RealAddr,
>>>> +                                                          SlotSize -
>>>> EltSize);
>>>> +        ImagAddr = CGF.Builder.CreateConstInBoundsByteGEP(ImagAddr,
>>>> +                                                      2 * SlotSize -
>>>> EltSize);
>>>> +      } else {
>>>> +        ImagAddr = CGF.Builder.CreateConstInBoundsByteGEP(RealAddr,
>>>> SlotSize);
>>>> +      }
>>>> +
>>>> +      llvm::Type *EltTy = CGF.ConvertTypeForMem(CTy->getElementType());
>>>> +      RealAddr = CGF.Builder.CreateElementBitCast(RealAddr, EltTy);
>>>> +      ImagAddr = CGF.Builder.CreateElementBitCast(ImagAddr, EltTy);
>>>> +      llvm::Value *Real = CGF.Builder.CreateLoad(RealAddr, ".vareal");
>>>> +      llvm::Value *Imag = CGF.Builder.CreateLoad(ImagAddr, ".vaimag");
>>>> +
>>>> +      Address Temp = CGF.CreateMemTemp(Ty, "vacplx");
>>>> +      CGF.EmitStoreOfComplex({Real, Imag}, CGF.MakeAddrLValue(Temp,
>>>> Ty),
>>>> +                             /*init*/ true);
>>>> +      return Temp;
>>>>      }
>>>> -    llvm::Type *PBaseTy =
>>>> llvm::PointerType::getUnqual(CGF.ConvertType(BaseTy));
>>>> -    RealAddr = Builder.CreateIntToPtr(RealAddr, PBaseTy);
>>>> -    ImagAddr = Builder.CreateIntToPtr(ImagAddr, PBaseTy);
>>>> -    llvm::Value *Real = Builder.CreateLoad(RealAddr, false, ".vareal");
>>>> -    llvm::Value *Imag = Builder.CreateLoad(ImagAddr, false, ".vaimag");
>>>> -    llvm::AllocaInst *Ptr =
>>>> -        CGF.CreateTempAlloca(CGT.ConvertTypeForMem(Ty), "vacplx");
>>>> -    llvm::Value *RealPtr =
>>>> -        Builder.CreateStructGEP(Ptr->getAllocatedType(), Ptr, 0,
>>>> ".real");
>>>> -    llvm::Value *ImagPtr =
>>>> -        Builder.CreateStructGEP(Ptr->getAllocatedType(), Ptr, 1,
>>>> ".imag");
>>>> -    Builder.CreateStore(Real, RealPtr, false);
>>>> -    Builder.CreateStore(Imag, ImagPtr, false);
>>>> -    return Ptr;
>>>> -  }
>>>> -
>>>> -  // If the argument is smaller than 8 bytes, it is right-adjusted in
>>>> -  // its doubleword slot.  Adjust the pointer to pick it up from the
>>>> -  // correct offset.
>>>> -  if (SizeInBytes < 8 && CGF.CGM.getDataLayout().isBigEndian()) {
>>>> -    llvm::Value *AddrAsInt = Builder.CreatePtrToInt(Addr, CGF.Int64Ty);
>>>> -    AddrAsInt = Builder.CreateAdd(AddrAsInt, Builder.getInt64(8 -
>>>> SizeInBytes));
>>>> -    Addr = Builder.CreateIntToPtr(AddrAsInt, BP);
>>>>    }
>>>>
>>>> -  llvm::Type *PTy = llvm::PointerType::getUnqual(CGF.ConvertType(Ty));
>>>> -  return Builder.CreateBitCast(Addr, PTy);
>>>> +  // Otherwise, just use the general rule.
>>>> +  return emitVoidPtrVAArg(CGF, VAListAddr, Ty, /*Indirect*/ false,
>>>> +                          TypeInfo, SlotSize, /*AllowHigher*/ true);
>>>>  }
>>>>
>>>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20150910/fe311236/attachment-0001.html>


More information about the cfe-commits mailing list