[OPENMP] Codegen for "#pragma omp atomic write"

[ABataev]

Summary

For global reg lvalue - use regular store through global register.
For simple lvalue - use simple atomic store.
For bitfields, vector element, extended vector elements - the original value of the whole storage (for vector elements) or of some aligned value (for bitfields) is atomically read, the part of this value for the given lvalue is modified and then use atomic compare-and-exchange operation to try to atomically write modified value (if it was not modified).
Also, changes in this patch fix the bug for '#pragma omp atomic read' applied to extended vector elements.

[ABataev]

Ping!

[majnemer]

lib/CodeGen/CGStmtOpenMP.cpp

859

Shouldn't you use monotonic ordering if the omp write isn't seq_cst?

[ABataev]

Yes, of course, just this patch was prepared before last update for
atomics. I'll update patch for sure.

Best regards,

Alexey Bataev

Software Engineer
Intel Compiler Team

16.02.2015 10:58, David Majnemer пишет:

Comment at: lib/CodeGen/CGStmtOpenMP.cpp:859
@@ +858,3 @@
+ else
+ CGF.EmitAtomicStore(ExprRValue, XLValue, /*isInit=*/false);

+ // OpenMP, 2.12.6, atomic Construct

Shouldn't you use monotonic ordering if the omp write isn't seq_cst?

http://reviews.llvm.org/D7369

EMAIL PREFERENCES

http://reviews.llvm.org/settings/panel/emailpreferences/

[ABataev]

Updated patch for the latest changes in atomic processing

[rjmccall]

If you've verified that the changes to the memory ordering flags in the tests are wanted, then this looks good to me.

[ABataev]

John, thanks for the review! Yes, I checked this.

Best regards,

Alexey Bataev

Software Engineer
Intel Compiler Team

27.02.2015 5:57, John McCall пишет:

If you've verified that the changes to the memory ordering flags in the tests are wanted, then this looks good to me.

http://reviews.llvm.org/D7369

EMAIL PREFERENCES

http://reviews.llvm.org/settings/panel/emailpreferences/

lib/CodeGen/CGAtomic.cpp

AtomicInfo(CodeGenFunction &CGF, LValue &lvalue)
 
         AtomicAlign = C.toCharUnitsFromBits(AtomicAlignInBits);
         ValueAlign = C.toCharUnitsFromBits(ValueAlignInBits);
         if (lvalue.getAlignment().isZero())
           lvalue.setAlignment(AtomicAlign);
 
         LVal = lvalue;
       } else if (lvalue.isBitField()) {
+        ValueTy = lvalue.getType();
+        ValueSizeInBits = C.getTypeSize(ValueTy);
         auto &OrigBFI = lvalue.getBitFieldInfo();
         auto Offset = OrigBFI.Offset % C.toBits(lvalue.getAlignment());
         AtomicSizeInBits = C.toBits(
             C.toCharUnitsFromBits(Offset + OrigBFI.Size + C.getCharWidth() - 1)
                 .RoundUpToAlignment(lvalue.getAlignment()));
         auto VoidPtrAddr = CGF.EmitCastToVoidPtr(lvalue.getBitFieldAddr());
         auto OffsetInChars =
             (C.toCharUnitsFromBits(OrigBFI.Offset) / lvalue.getAlignment()) *
             lvalue.getAlignment();
         VoidPtrAddr = CGF.Builder.CreateConstGEP1_64(
             VoidPtrAddr, OffsetInChars.getQuantity());
         auto Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
             VoidPtrAddr,
             CGF.Builder.getIntNTy(AtomicSizeInBits)->getPointerTo(),
             "atomic_bitfield_base");
         BFI = OrigBFI;
         BFI.Offset = Offset;
         BFI.StorageSize = AtomicSizeInBits;
         LVal = LValue::MakeBitfield(Addr, BFI, lvalue.getType(),
                                     lvalue.getAlignment());
+        LVal.setTBAAInfo(lvalue.getTBAAInfo());
+        AtomicTy = C.getIntTypeForBitwidth(AtomicSizeInBits, OrigBFI.IsSigned);
+        if (AtomicTy.isNull()) {
+          llvm::APInt Size(
+              /*numBits=*/32,
+              C.toCharUnitsFromBits(AtomicSizeInBits).getQuantity());
+          AtomicTy = C.getConstantArrayType(C.CharTy, Size, ArrayType::Normal,
+                                            /*IndexTypeQuals=*/0);
+        }
+        AtomicAlign = ValueAlign = lvalue.getAlignment();
       } else if (lvalue.isVectorElt()) {
-        AtomicSizeInBits = C.getTypeSize(lvalue.getType());
+        ValueTy = lvalue.getType()->getAs<VectorType>()->getElementType();
+        ValueSizeInBits = C.getTypeSize(ValueTy);
+        AtomicTy = lvalue.getType();
+        AtomicSizeInBits = C.getTypeSize(AtomicTy);
+        AtomicAlign = ValueAlign = lvalue.getAlignment();
         LVal = lvalue;
       } else {
         assert(lvalue.isExtVectorElt());
-        AtomicSizeInBits = C.getTypeSize(lvalue.getType());
+        ValueTy = lvalue.getType();
+        ValueSizeInBits = C.getTypeSize(ValueTy);
+        AtomicTy = ValueTy = CGF.getContext().getExtVectorType(
+            lvalue.getType(), lvalue.getExtVectorAddr()
+                                  ->getType()
+                                  ->getPointerElementType()
+                                  ->getVectorNumElements());
+        AtomicSizeInBits = C.getTypeSize(AtomicTy);
+        AtomicAlign = ValueAlign = lvalue.getAlignment();
         LVal = lvalue;
       }
       UseLibcall = !C.getTargetInfo().hasBuiltinAtomic(
           AtomicSizeInBits, C.toBits(lvalue.getAlignment()));
     }
 
     QualType getAtomicType() const { return AtomicTy; }
     QualType getValueType() const { return ValueTy; }
     CharUnits getAtomicAlignment() const { return AtomicAlign; }
     CharUnits getValueAlignment() const { return ValueAlign; }
     uint64_t getAtomicSizeInBits() const { return AtomicSizeInBits; }
     uint64_t getValueSizeInBits() const { return ValueSizeInBits; }
     TypeEvaluationKind getEvaluationKind() const { return EvaluationKind; }
     bool shouldUseLibcall() const { return UseLibcall; }
     const LValue &getAtomicLValue() const { return LVal; }
+    llvm::Value *getAtomicAddress() const {
+      if (LVal.isSimple())
+        return LVal.getAddress();
+      else if (LVal.isBitField())
+        return LVal.getBitFieldAddr();
+      else if (LVal.isVectorElt())
+        return LVal.getVectorAddr();
+      assert(LVal.isExtVectorElt());
+      return LVal.getExtVectorAddr();
+    }
 
     /// Is the atomic size larger than the underlying value type?
     ///
     /// Note that the absence of padding does not mean that atomic
     /// objects are completely interchangeable with non-atomic
     /// objects: we might have promoted the alignment of a type
     /// without making it bigger.
     bool hasPadding() const {
       return (ValueSizeInBits != AtomicSizeInBits);
     }
 
     bool emitMemSetZeroIfNecessary() const;
 
     llvm::Value *getAtomicSizeValue() const {
       CharUnits size = CGF.getContext().toCharUnitsFromBits(AtomicSizeInBits);
       return CGF.CGM.getSize(size);
     }
 
     /// Cast the given pointer to an integer pointer suitable for
     /// atomic operations.
     llvm::Value *emitCastToAtomicIntPointer(llvm::Value *addr) const;
 
     /// Turn an atomic-layout object into an r-value.
-    RValue convertTempToRValue(llvm::Value *addr,
-                               AggValueSlot resultSlot,
-                               SourceLocation loc) const;
+    RValue convertTempToRValue(llvm::Value *addr, AggValueSlot resultSlot,
+                               SourceLocation loc, bool AsValue) const;
 
     /// \brief Converts a rvalue to integer value.
     llvm::Value *convertRValueToInt(RValue RVal) const;
 
-    RValue convertIntToValue(llvm::Value *IntVal, AggValueSlot ResultSlot,
-                             SourceLocation Loc) const;
+    RValue ConvertIntToValueOrAtomic(llvm::Value *IntVal,
+                                     AggValueSlot ResultSlot,
+                                     SourceLocation Loc, bool AsValue) const;
 
     /// Copy an atomic r-value into atomic-layout memory.
     void emitCopyIntoMemory(RValue rvalue) const;
 
     /// Project an l-value down to the value field.
     LValue projectValue() const {
       assert(LVal.isSimple());
-      llvm::Value *addr = LVal.getAddress();
+      llvm::Value *addr = getAtomicAddress();
       if (hasPadding())
         addr = CGF.Builder.CreateStructGEP(addr, 0);
 
       return LValue::MakeAddr(addr, getValueType(), LVal.getAlignment(),
                               CGF.getContext(), LVal.getTBAAInfo());
     }
 
+    /// \brief Emits atomic load.
+    /// \returns Loaded value.
+    RValue EmitAtomicLoad(AggValueSlot ResultSlot, SourceLocation Loc,
+                          bool AsValue);
+
+    /// \brief Emits atomic compare-and-exchange sequence.
+    /// \param Expected Expected value.
+    /// \param Desired Desired value.
+    /// \param Success Atomic ordering for success operation.
+    /// \param Failure Atomic ordering for failed operation.
+    /// \param IsWeak true if atomic operation is weak, false otherwise.
+    /// \returns Pair of values: previous value from storage (value type) and
+    /// boolean flag (i1 type) with true if success and false otherwise.
+    std::pair<llvm::Value *, llvm::Value *> EmitAtomicCompareExchange(
+        RValue Expected, RValue Desired,
+        llvm::AtomicOrdering Success = llvm::SequentiallyConsistent,
+        llvm::AtomicOrdering Failure = llvm::SequentiallyConsistent,
+        bool IsWeak = false);
+
     /// Materialize an atomic r-value in atomic-layout memory.
     llvm::Value *materializeRValue(RValue rvalue) const;
 
   private:
     bool requiresMemSetZero(llvm::Type *type) const;
+
+    /// \brief Creates temp alloca for intermediate operations on atomic value.
+    llvm::Value *CreateTempAlloca() const;
+
+    /// \brief Emits atomic load as a libcall.
+    void EmitAtomicLoadLibcall(llvm::Value *AddForLoaded);
+    /// \brief Emits atomic load as LLVM instruction.
+    llvm::Value *EmitAtomicLoadOp();
+    /// \brief Emits atomic compare-and-exchange op as a libcall.
+    std::pair<llvm::Value *, llvm::Value *> EmitAtomicCompareExchangeLibcall(
+        llvm::Value *ExpectedAddr, llvm::Value *DesiredAddr,
+        llvm::AtomicOrdering Success = llvm::SequentiallyConsistent,
+        llvm::AtomicOrdering Failure = llvm::SequentiallyConsistent);
+    /// \brief Emits atomic compare-and-exchange op as LLVM instruction.
+    std::pair<llvm::Value *, llvm::Value *> EmitAtomicCompareExchangeOp(
+        llvm::Value *Expected, llvm::Value *Desired,
+        llvm::AtomicOrdering Success = llvm::SequentiallyConsistent,
+        llvm::AtomicOrdering Failure = llvm::SequentiallyConsistent,
+        bool IsWeak = false);
   };
 }
 
+llvm::Value *AtomicInfo::CreateTempAlloca() const {
+  auto *TempAlloca = CGF.CreateMemTemp(
+      (LVal.isBitField() && ValueSizeInBits > AtomicSizeInBits) ? ValueTy
+                                                                : AtomicTy,
+      "atomic-temp");
+  TempAlloca->setAlignment(getAtomicAlignment().getQuantity());
+  // Cast to pointer to value type for bitfields.
+  if (LVal.isBitField())
+    return CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
+        TempAlloca, getAtomicAddress()->getType());
+  return TempAlloca;
+}
+
 static RValue emitAtomicLibcall(CodeGenFunction &CGF,
                                 StringRef fnName,
                                 QualType resultType,
                                 CallArgList &args) {
   const CGFunctionInfo &fnInfo =
     CGF.CGM.getTypes().arrangeFreeFunctionCall(resultType, args,
             FunctionType::ExtInfo(), RequiredArgs::All);
   llvm::FunctionType *fnTy = CGF.CGM.getTypes().GetFunctionType(fnInfo);

 }
 
 bool AtomicInfo::emitMemSetZeroIfNecessary() const {
   assert(LVal.isSimple());
   llvm::Value *addr = LVal.getAddress();
   if (!requiresMemSetZero(addr->getType()->getPointerElementType()))
     return false;
 
-  CGF.Builder.CreateMemSet(addr, llvm::ConstantInt::get(CGF.Int8Ty, 0),
-                           AtomicSizeInBits / 8,
-                           LVal.getAlignment().getQuantity());
+  CGF.Builder.CreateMemSet(
+      addr, llvm::ConstantInt::get(CGF.Int8Ty, 0),
+      CGF.getContext().toCharUnitsFromBits(AtomicSizeInBits).getQuantity(),
   return true;
 }
 
 static void emitAtomicCmpXchg(CodeGenFunction &CGF, AtomicExpr *E, bool IsWeak,
                               llvm::Value *Dest, llvm::Value *Ptr,
                               llvm::Value *Val1, llvm::Value *Val2,
                               uint64_t Size, unsigned Align,
                               llvm::AtomicOrdering SuccessOrder,
unsigned addrspace =
     cast<llvm::PointerType>(addr->getType())->getAddressSpace();
   llvm::IntegerType *ty =
     llvm::IntegerType::get(CGF.getLLVMContext(), AtomicSizeInBits);
   return CGF.Builder.CreateBitCast(addr, ty->getPointerTo(addrspace));
 }
 
 RValue AtomicInfo::convertTempToRValue(llvm::Value *addr,
                                        AggValueSlot resultSlot,
-                                       SourceLocation loc) const {
+                                       SourceLocation loc, bool AsValue) const {
   if (LVal.isSimple()) {
     if (EvaluationKind == TEK_Aggregate)
       return resultSlot.asRValue();
 
     // Drill into the padding structure if we have one.
     if (hasPadding())
       addr = CGF.Builder.CreateStructGEP(addr, 0);
 
     // Otherwise, just convert the temporary to an r-value using the
     // normal conversion routine.
     return CGF.convertTempToRValue(addr, getValueType(), loc);
-  } else if (LVal.isBitField())
+  } else if (!AsValue)
+    // Get RValue from temp memory as atomic for non-simple lvalues
+    return RValue::get(
+        CGF.Builder.CreateAlignedLoad(addr, AtomicAlign.getQuantity()));
+  else if (LVal.isBitField())
     return CGF.EmitLoadOfBitfieldLValue(LValue::MakeBitfield(
         addr, LVal.getBitFieldInfo(), LVal.getType(), LVal.getAlignment()));
   else if (LVal.isVectorElt())
     return CGF.EmitLoadOfLValue(LValue::MakeVectorElt(addr, LVal.getVectorIdx(),
                                                       LVal.getType(),
                                                       LVal.getAlignment()),
                                 loc);
   assert(LVal.isExtVectorElt());
   return CGF.EmitLoadOfExtVectorElementLValue(LValue::MakeExtVectorElt(
       addr, LVal.getExtVectorElts(), LVal.getType(), LVal.getAlignment()));
 }
 
-RValue AtomicInfo::convertIntToValue(llvm::Value *IntVal,
+RValue AtomicInfo::ConvertIntToValueOrAtomic(llvm::Value *IntVal,
                                              AggValueSlot ResultSlot,
-                                     SourceLocation Loc) const {
-  assert(LVal.isSimple());
+                                             SourceLocation Loc,
+                                             bool AsValue) const {
   // Try not to in some easy cases.
   assert(IntVal->getType()->isIntegerTy() && "Expected integer value");
-  if (getEvaluationKind() == TEK_Scalar && !hasPadding()) {
-    auto *ValTy = CGF.ConvertTypeForMem(ValueTy);
+  if (getEvaluationKind() == TEK_Scalar &&
+      (((!LVal.isBitField() ||
+         LVal.getBitFieldInfo().Size == ValueSizeInBits) &&
+        !hasPadding()) ||
+       !AsValue)) {
+    auto *ValTy = AsValue
+                      ? CGF.ConvertTypeForMem(ValueTy)
+                      : getAtomicAddress()->getType()->getPointerElementType();
     if (ValTy->isIntegerTy()) {
       assert(IntVal->getType() == ValTy && "Different integer types.");
       return RValue::get(IntVal);
     } else if (ValTy->isPointerTy())
       return RValue::get(CGF.Builder.CreateIntToPtr(IntVal, ValTy));
     else if (llvm::CastInst::isBitCastable(IntVal->getType(), ValTy))
       return RValue::get(CGF.Builder.CreateBitCast(IntVal, ValTy));
   }
 
   // Create a temporary.  This needs to be big enough to hold the
   // atomic integer.
   llvm::Value *Temp;
   bool TempIsVolatile = false;
   CharUnits TempAlignment;
-  if (getEvaluationKind() == TEK_Aggregate) {
+  if (AsValue && getEvaluationKind() == TEK_Aggregate) {
     assert(!ResultSlot.isIgnored());
     Temp = ResultSlot.getAddr();
     TempAlignment = getValueAlignment();
     TempIsVolatile = ResultSlot.isVolatile();
   } else {
-    Temp = CGF.CreateMemTemp(getAtomicType(), "atomic-temp");
+    Temp = CreateTempAlloca();
     TempAlignment = getAtomicAlignment();
   }
 
   // Slam the integer into the temporary.
   llvm::Value *CastTemp = emitCastToAtomicIntPointer(Temp);
   CGF.Builder.CreateAlignedStore(IntVal, CastTemp, TempAlignment.getQuantity())
       ->setVolatile(TempIsVolatile);
 
-  return convertTempToRValue(Temp, ResultSlot, Loc);
+  return convertTempToRValue(Temp, ResultSlot, Loc, AsValue);
 }
 
-/// Emit a load from an l-value of atomic type.  Note that the r-value
-/// we produce is an r-value of the atomic *value* type.
-RValue CodeGenFunction::EmitAtomicLoad(LValue src, SourceLocation loc,
-                                       AggValueSlot resultSlot) {
-  AtomicInfo atomics(*this, src);
-  LValue LVal = atomics.getAtomicLValue();
-  llvm::Value *SrcAddr = nullptr;
-  llvm::AllocaInst *NonSimpleTempAlloca = nullptr;
-  if (LVal.isSimple())
-    SrcAddr = LVal.getAddress();
-  else {
-    if (LVal.isBitField())
-      SrcAddr = LVal.getBitFieldAddr();
-    else if (LVal.isVectorElt())
-      SrcAddr = LVal.getVectorAddr();
-    else {
-      assert(LVal.isExtVectorElt());
-      SrcAddr = LVal.getExtVectorAddr();
-    }
-    NonSimpleTempAlloca = CreateTempAlloca(
-        SrcAddr->getType()->getPointerElementType(), "atomic-load-temp");
-    NonSimpleTempAlloca->setAlignment(getContext().toBits(src.getAlignment()));
-  }
+void AtomicInfo::EmitAtomicLoadLibcall(llvm::Value *AddForLoaded) {
+  // void __atomic_load(size_t size, void *mem, void *return, int order);
+  CallArgList Args;
+  Args.add(RValue::get(getAtomicSizeValue()), CGF.getContext().getSizeType());
+  Args.add(RValue::get(CGF.EmitCastToVoidPtr(getAtomicAddress())),
+           CGF.getContext().VoidPtrTy);
+  Args.add(RValue::get(CGF.EmitCastToVoidPtr(AddForLoaded)),
+           CGF.getContext().VoidPtrTy);
+  Args.add(RValue::get(llvm::ConstantInt::get(
+               CGF.IntTy, AtomicExpr::AO_ABI_memory_order_seq_cst)),
+           CGF.getContext().IntTy);
+  emitAtomicLibcall(CGF, "__atomic_load", CGF.getContext().VoidTy, Args);
+
+llvm::Value *AtomicInfo::EmitAtomicLoadOp() {
+  // Okay, we're doing this natively.
+  llvm::Value *Addr = emitCastToAtomicIntPointer(getAtomicAddress());
+  llvm::LoadInst *Load = CGF.Builder.CreateLoad(Addr, "atomic-load");
+  Load->setAtomic(llvm::SequentiallyConsistent);
+
+  // Other decoration.
+  Load->setAlignment(getAtomicAlignment().getQuantity());
+  if (LVal.isVolatileQualified())
+    Load->setVolatile(true);
+  if (LVal.getTBAAInfo())
+    CGF.CGM.DecorateInstruction(Load, LVal.getTBAAInfo());
+  return Load;
 
+RValue AtomicInfo::EmitAtomicLoad(AggValueSlot ResultSlot, SourceLocation Loc,
+                                  bool AsValue) {
   // Check whether we should use a library call.
-  if (atomics.shouldUseLibcall()) {
-    llvm::Value *tempAddr;
-    if (LVal.isSimple()) {
-      if (!resultSlot.isIgnored()) {
-        assert(atomics.getEvaluationKind() == TEK_Aggregate);
-        tempAddr = resultSlot.getAddr();
-      } else
-        tempAddr = CreateMemTemp(atomics.getAtomicType(), "atomic-load-temp");
+  if (shouldUseLibcall()) {
+    llvm::Value *TempAddr;
+    if (LVal.isSimple() && !ResultSlot.isIgnored()) {
+      assert(getEvaluationKind() == TEK_Aggregate);
+      TempAddr = ResultSlot.getAddr();
     } else
-      tempAddr = NonSimpleTempAlloca;
+      TempAddr = CreateTempAlloca();
 
-    // void __atomic_load(size_t size, void *mem, void *return, int order);
-    CallArgList args;
-    args.add(RValue::get(atomics.getAtomicSizeValue()),
-             getContext().getSizeType());
-    args.add(RValue::get(EmitCastToVoidPtr(SrcAddr)), getContext().VoidPtrTy);
-    args.add(RValue::get(EmitCastToVoidPtr(tempAddr)), getContext().VoidPtrTy);
-    args.add(RValue::get(llvm::ConstantInt::get(
-                 IntTy, AtomicExpr::AO_ABI_memory_order_seq_cst)),
-             getContext().IntTy);
-    emitAtomicLibcall(*this, "__atomic_load", getContext().VoidTy, args);
+    EmitAtomicLoadLibcall(TempAddr);
 
-    // Produce the r-value.
-    return atomics.convertTempToRValue(tempAddr, resultSlot, loc);
+    // Okay, turn that back into the original value or whole atomic (for
+    // non-simple lvalues) type.
+    return convertTempToRValue(TempAddr, ResultSlot, Loc, AsValue);
   }
 
   // Okay, we're doing this natively.
-  llvm::Value *addr = atomics.emitCastToAtomicIntPointer(SrcAddr);
-  llvm::LoadInst *load = Builder.CreateLoad(addr, "atomic-load");
-  load->setAtomic(llvm::SequentiallyConsistent);
-
-  // Other decoration.
-  load->setAlignment(src.getAlignment().getQuantity());
-  if (src.isVolatileQualified())
-    load->setVolatile(true);
-  if (src.getTBAAInfo())
-    CGM.DecorateInstruction(load, src.getTBAAInfo());
+  auto *Load = EmitAtomicLoadOp();
 
   // If we're ignoring an aggregate return, don't do anything.
-  if (atomics.getEvaluationKind() == TEK_Aggregate && resultSlot.isIgnored())
+  if (getEvaluationKind() == TEK_Aggregate && ResultSlot.isIgnored())
     return RValue::getAggregate(nullptr, false);
 
-  // Okay, turn that back into the original value type.
-  if (src.isSimple())
-    return atomics.convertIntToValue(load, resultSlot, loc);
-
-  auto *IntAddr = atomics.emitCastToAtomicIntPointer(NonSimpleTempAlloca);
-  Builder.CreateAlignedStore(load, IntAddr, src.getAlignment().getQuantity());
-  return atomics.convertTempToRValue(NonSimpleTempAlloca, resultSlot, loc);
+  // Okay, turn that back into the original value or atomic (for non-simple
+  // lvalues) type.
+  return ConvertIntToValueOrAtomic(Load, ResultSlot, Loc, AsValue);
 }
 
-
+/// Emit a load from an l-value of atomic type.  Note that the r-value
+/// we produce is an r-value of the atomic *value* type.
+RValue CodeGenFunction::EmitAtomicLoad(LValue src, SourceLocation loc,
+                                       AggValueSlot resultSlot) {
+  AtomicInfo Atomics(*this, src);
+  return Atomics.EmitAtomicLoad(resultSlot, loc, /*AsValue=*/true);
+}
 
 /// Copy an r-value into memory as part of storing to an atomic type.
 /// This needs to create a bit-pattern suitable for atomic operations.
 void AtomicInfo::emitCopyIntoMemory(RValue rvalue) const {
   assert(LVal.isSimple());
   // If we have an r-value, the rvalue should be of the atomic type,
   // which means that the caller is responsible for having zeroed
   // any padding.  Just do an aggregate copy of that type.
   if (rvalue.isAggregate()) {
-    CGF.EmitAggregateCopy(LVal.getAddress(),
+    CGF.EmitAggregateCopy(getAtomicAddress(),
                           rvalue.getAggregateAddr(),
                           getAtomicType(),
                           (rvalue.isVolatileQualified()
                            || LVal.isVolatileQualified()),
                           LVal.getAlignment());
     return;
   }
 

 /// to an atomic type.
 llvm::Value *AtomicInfo::materializeRValue(RValue rvalue) const {
   // Aggregate r-values are already in memory, and EmitAtomicStore
   // requires them to be values of the atomic type.
   if (rvalue.isAggregate())
     return rvalue.getAggregateAddr();
 
   // Otherwise, make a temporary and materialize into it.
-  llvm::Value *temp = CGF.CreateMemTemp(getAtomicType(), "atomic-store-temp");
-  LValue tempLV =
-      CGF.MakeAddrLValue(temp, getAtomicType(), getAtomicAlignment());
-  AtomicInfo Atomics(CGF, tempLV);
+  LValue TempLV = CGF.MakeAddrLValue(CreateTempAlloca(), getAtomicType(),
+                                     getAtomicAlignment());
+  AtomicInfo Atomics(CGF, TempLV);
   Atomics.emitCopyIntoMemory(rvalue);
-  return temp;
+  return TempLV.getAddress();
 }
 
 llvm::Value *AtomicInfo::convertRValueToInt(RValue RVal) const {
   // If we've got a scalar value of the right size, try to avoid going
   // through memory.
-  if (RVal.isScalar() && !hasPadding()) {
+  if (RVal.isScalar() && (!hasPadding() || !LVal.isSimple())) {
     llvm::Value *Value = RVal.getScalarVal();
     if (isa<llvm::IntegerType>(Value->getType()))
       return Value;
     else {
-      llvm::IntegerType *InputIntTy =
-          llvm::IntegerType::get(CGF.getLLVMContext(), getValueSizeInBits());
+      llvm::IntegerType *InputIntTy = llvm::IntegerType::get(
+          CGF.getLLVMContext(),
+          LVal.isSimple() ? getValueSizeInBits() : getAtomicSizeInBits());
       if (isa<llvm::PointerType>(Value->getType()))
         return CGF.Builder.CreatePtrToInt(Value, InputIntTy);
       else if (llvm::BitCastInst::isBitCastable(Value->getType(), InputIntTy))
         return CGF.Builder.CreateBitCast(Value, InputIntTy);
     }
   }
   // Otherwise, we need to go through memory.
   // Put the r-value in memory.
   llvm::Value *Addr = materializeRValue(RVal);
 
   // Cast the temporary to the atomic int type and pull a value out.
   Addr = emitCastToAtomicIntPointer(Addr);
   return CGF.Builder.CreateAlignedLoad(Addr,
                                        getAtomicAlignment().getQuantity());
 }
 
+std::pair<llvm::Value *, llvm::Value *> AtomicInfo::EmitAtomicCompareExchangeOp(
+    llvm::Value *Expected, llvm::Value *Desired, llvm::AtomicOrdering Success,
+    llvm::AtomicOrdering Failure, bool IsWeak) {
+  // Do the atomic store.
+  auto *Addr = emitCastToAtomicIntPointer(getAtomicAddress());
+  auto *Inst = CGF.Builder.CreateAtomicCmpXchg(Addr, Expected, Desired, Success,
+                                               Failure);
+  // Other decoration.
+  Inst->setVolatile(LVal.isVolatileQualified());
+  Inst->setWeak(IsWeak);
+
+  // Okay, turn that back into the original value type.
+  auto *PreviousVal = CGF.Builder.CreateExtractValue(Inst, /*Idxs=*/0);
+  auto *SuccessFailureVal = CGF.Builder.CreateExtractValue(Inst, /*Idxs=*/1);
+  return std::make_pair(PreviousVal, SuccessFailureVal);
+}
+
+std::pair<llvm::Value *, llvm::Value *>
+AtomicInfo::EmitAtomicCompareExchangeLibcall(llvm::Value *ExpectedAddr,
+                                             llvm::Value *DesiredAddr,
+                                             llvm::AtomicOrdering Success,
+                                             llvm::AtomicOrdering Failure) {
+  // bool __atomic_compare_exchange(size_t size, void *obj, void *expected,
+  // void *desired, int success, int failure);
+  CallArgList Args;
+  Args.add(RValue::get(getAtomicSizeValue()), CGF.getContext().getSizeType());
+  Args.add(RValue::get(CGF.EmitCastToVoidPtr(getAtomicAddress())),
+           CGF.getContext().VoidPtrTy);
+  Args.add(RValue::get(CGF.EmitCastToVoidPtr(ExpectedAddr)),
+           CGF.getContext().VoidPtrTy);
+  Args.add(RValue::get(CGF.EmitCastToVoidPtr(DesiredAddr)),
+           CGF.getContext().VoidPtrTy);
+  Args.add(RValue::get(llvm::ConstantInt::get(CGF.IntTy, Success)),
+           CGF.getContext().IntTy);
+  Args.add(RValue::get(llvm::ConstantInt::get(CGF.IntTy, Failure)),
+           CGF.getContext().IntTy);
+  auto SuccessFailureRVal = emitAtomicLibcall(CGF, "__atomic_compare_exchange",
+                                              CGF.getContext().BoolTy, Args);
+  auto *PreviousVal = CGF.Builder.CreateAlignedLoad(
+      ExpectedAddr, getValueAlignment().getQuantity());
+  return std::make_pair(PreviousVal, SuccessFailureRVal.getScalarVal());
+}
+
+std::pair<llvm::Value *, llvm::Value *> AtomicInfo::EmitAtomicCompareExchange(
+    RValue Expected, RValue Desired, llvm::AtomicOrdering Success,
+    llvm::AtomicOrdering Failure, bool IsWeak) {
+  if (Failure >= Success)
+    // Don't assert on undefined behavior.
+    Failure = llvm::AtomicCmpXchgInst::getStrongestFailureOrdering(Success);
+
+  // Check whether we should use a library call.
+  if (shouldUseLibcall()) {
+    auto *ExpectedAddr = materializeRValue(Expected);
+    // Produce a source address.
+    auto *DesiredAddr = materializeRValue(Desired);
+    return EmitAtomicCompareExchangeLibcall(ExpectedAddr, DesiredAddr, Success,
+                                            Failure);
+  }
+
+  // If we've got a scalar value of the right size, try to avoid going
+  // through memory.
+  auto *ExpectedIntVal = convertRValueToInt(Expected);
+  auto *DesiredIntVal = convertRValueToInt(Desired);
+
+  return EmitAtomicCompareExchangeOp(ExpectedIntVal, DesiredIntVal, Success,
+                                     Failure, IsWeak);
+}
+
 /// Emit a store to an l-value of atomic type.
 ///
 /// Note that the r-value is expected to be an r-value *of the atomic
 /// type*; this means that for aggregate r-values, it should include
 /// storage for any padding that was necessary.
 void CodeGenFunction::EmitAtomicStore(RValue rvalue, LValue dest, bool isInit) {
   // If this is an aggregate r-value, it should agree in type except
   // maybe for address-space qualification.
   assert(!rvalue.isAggregate() ||
          rvalue.getAggregateAddr()->getType()->getPointerElementType()
            == dest.getAddress()->getType()->getPointerElementType());
 
   AtomicInfo atomics(*this, dest);
+  LValue LVal = atomics.getAtomicLValue();
 
   // If this is an initialization, just put the value there normally.
+  if (LVal.isSimple()) {
     if (isInit) {
       atomics.emitCopyIntoMemory(rvalue);
       return;
     }
 
     // Check whether we should use a library call.
     if (atomics.shouldUseLibcall()) {
       // Produce a source address.
       llvm::Value *srcAddr = atomics.materializeRValue(rvalue);
 
       // void __atomic_store(size_t size, void *mem, void *val, int order)
       CallArgList args;
       args.add(RValue::get(atomics.getAtomicSizeValue()),
                getContext().getSizeType());
-    args.add(RValue::get(EmitCastToVoidPtr(dest.getAddress())),
-             getContext().VoidPtrTy);
-    args.add(RValue::get(EmitCastToVoidPtr(srcAddr)),
+      args.add(RValue::get(EmitCastToVoidPtr(atomics.getAtomicAddress())),
                getContext().VoidPtrTy);
+      args.add(RValue::get(EmitCastToVoidPtr(srcAddr)), getContext().VoidPtrTy);
       args.add(RValue::get(llvm::ConstantInt::get(
                    IntTy, AtomicExpr::AO_ABI_memory_order_seq_cst)),
                getContext().IntTy);
       emitAtomicLibcall(*this, "__atomic_store", getContext().VoidTy, args);
       return;
     }
 
     // Okay, we're doing this natively.
     llvm::Value *intValue = atomics.convertRValueToInt(rvalue);
 
     // Do the atomic store.
-  llvm::Value *addr = atomics.emitCastToAtomicIntPointer(dest.getAddress());
+    llvm::Value *addr =
+        atomics.emitCastToAtomicIntPointer(atomics.getAtomicAddress());
+    intValue = Builder.CreateIntCast(
+        intValue, addr->getType()->getPointerElementType(), /*isSigned=*/false);
     llvm::StoreInst *store = Builder.CreateStore(intValue, addr);
 
     // Initializations don't need to be atomic.
-  if (!isInit) store->setAtomic(llvm::SequentiallyConsistent);
+    if (!isInit)
+      store->setAtomic(llvm::SequentiallyConsistent);
 
     // Other decoration.
     store->setAlignment(dest.getAlignment().getQuantity());
     if (dest.isVolatileQualified())
       store->setVolatile(true);
     if (dest.getTBAAInfo())
       CGM.DecorateInstruction(store, dest.getTBAAInfo());
+    return;
+  }
+
+  // Atomic load of prev value.
+  RValue OldRVal = atomics.EmitAtomicLoad(AggValueSlot::ignored(),
+                                          SourceLocation(), /*AsValue=*/false);
+  // For non-simple lvalues perform compare-and-swap procedure.
+  auto *ContBB = createBasicBlock("atomic_cont");
+  auto *ExitBB = createBasicBlock("atomic_exit");
+  auto *CurBB = Builder.GetInsertBlock();
+  EmitBlock(ContBB);
+  llvm::PHINode *PHI = Builder.CreatePHI(OldRVal.getScalarVal()->getType(),
+                                         /*NumReservedValues=*/2);
+  PHI->addIncoming(OldRVal.getScalarVal(), CurBB);
+  RValue OriginalRValue = RValue::get(PHI);
+  // Build new lvalue for temp address
+  auto *Ptr = atomics.materializeRValue(OriginalRValue);
+  // Build new lvalue for temp address
+  LValue UpdateLVal;
+  if (LVal.isBitField())
+    UpdateLVal = LValue::MakeBitfield(Ptr, LVal.getBitFieldInfo(),
+                                      LVal.getType(), LVal.getAlignment());
+  else if (LVal.isVectorElt())
+    UpdateLVal = LValue::MakeVectorElt(Ptr, LVal.getVectorIdx(), LVal.getType(),
+                                       LVal.getAlignment());
+  else {
+    assert(LVal.isExtVectorElt());
+    UpdateLVal = LValue::MakeExtVectorElt(Ptr, LVal.getExtVectorElts(),
+                                          LVal.getType(), LVal.getAlignment());
+  }
+  UpdateLVal.setTBAAInfo(LVal.getTBAAInfo());
+  // Store new value in the corresponding memory area
+  EmitStoreThroughLValue(rvalue, UpdateLVal);
+  // Load new value
+  RValue NewRValue = RValue::get(EmitLoadOfScalar(
+      Ptr, LVal.isVolatile(), atomics.getAtomicAlignment().getQuantity(),
+      atomics.getAtomicType(), SourceLocation()));
+  // Try to write new value using cmpxchg operation
+  auto Pair = atomics.EmitAtomicCompareExchange(OriginalRValue, NewRValue);
+  llvm::Value *OldValue = Pair.first;
+  if (!atomics.shouldUseLibcall())
+    // Convert integer value to original atomic type
+    OldValue = atomics.ConvertIntToValueOrAtomic(
+                           OldValue, AggValueSlot::ignored(), SourceLocation(),
+                           /*AsValue=*/false).getScalarVal();
+  PHI->addIncoming(OldValue, ContBB);
+  Builder.CreateCondBr(Pair.second, ContBB, ExitBB);
+  EmitBlock(ExitBB);
 }
 
 /// Emit a compare-and-exchange op for atomic type.
 ///
 std::pair<RValue, RValue> CodeGenFunction::EmitAtomicCompareExchange(
     LValue Obj, RValue Expected, RValue Desired, SourceLocation Loc,
     llvm::AtomicOrdering Success, llvm::AtomicOrdering Failure, bool IsWeak,
     AggValueSlot Slot) {
   // If this is an aggregate r-value, it should agree in type except
   // maybe for address-space qualification.
   assert(!Expected.isAggregate() ||
          Expected.getAggregateAddr()->getType()->getPointerElementType() ==
              Obj.getAddress()->getType()->getPointerElementType());
   assert(!Desired.isAggregate() ||
          Desired.getAggregateAddr()->getType()->getPointerElementType() ==
              Obj.getAddress()->getType()->getPointerElementType());
   AtomicInfo Atomics(*this, Obj);
 
-  if (Failure >= Success)
-    // Don't assert on undefined behavior.
-    Failure = llvm::AtomicCmpXchgInst::getStrongestFailureOrdering(Success);
-
-  auto Alignment = Atomics.getValueAlignment();
-  // Check whether we should use a library call.
-  if (Atomics.shouldUseLibcall()) {
-    auto *ExpectedAddr = Atomics.materializeRValue(Expected);
-    // Produce a source address.
-    auto *DesiredAddr = Atomics.materializeRValue(Desired);
-    // bool __atomic_compare_exchange(size_t size, void *obj, void *expected,
-    // void *desired, int success, int failure);
-    CallArgList Args;
-    Args.add(RValue::get(Atomics.getAtomicSizeValue()),
-             getContext().getSizeType());
-    Args.add(RValue::get(EmitCastToVoidPtr(Obj.getAddress())),
-             getContext().VoidPtrTy);
-    Args.add(RValue::get(EmitCastToVoidPtr(ExpectedAddr)),
-             getContext().VoidPtrTy);
-    Args.add(RValue::get(EmitCastToVoidPtr(DesiredAddr)),
-             getContext().VoidPtrTy);
-    Args.add(RValue::get(llvm::ConstantInt::get(IntTy, Success)),
-             getContext().IntTy);
-    Args.add(RValue::get(llvm::ConstantInt::get(IntTy, Failure)),
-             getContext().IntTy);
-    auto SuccessFailureRVal = emitAtomicLibcall(
-        *this, "__atomic_compare_exchange", getContext().BoolTy, Args);
-    auto *PreviousVal =
-        Builder.CreateAlignedLoad(ExpectedAddr, Alignment.getQuantity());
-    return std::make_pair(RValue::get(PreviousVal), SuccessFailureRVal);
-  }
-
-  // If we've got a scalar value of the right size, try to avoid going
-  // through memory.
-  auto *ExpectedIntVal = Atomics.convertRValueToInt(Expected);
-  auto *DesiredIntVal = Atomics.convertRValueToInt(Desired);
-
-  // Do the atomic store.
-  auto *Addr = Atomics.emitCastToAtomicIntPointer(Obj.getAddress());
-  auto *Inst = Builder.CreateAtomicCmpXchg(Addr, ExpectedIntVal, DesiredIntVal,
-                                          Success, Failure);
-  // Other decoration.
-  Inst->setVolatile(Obj.isVolatileQualified());
-  Inst->setWeak(IsWeak);
-
-  // Okay, turn that back into the original value type.
-  auto *PreviousVal = Builder.CreateExtractValue(Inst, /*Idxs=*/0);
-  auto *SuccessFailureVal = Builder.CreateExtractValue(Inst, /*Idxs=*/1);
-  return std::make_pair(Atomics.convertIntToValue(PreviousVal, Slot, Loc),
-                        RValue::get(SuccessFailureVal));
+  auto Pair = Atomics.EmitAtomicCompareExchange(Expected, Desired, Success,
+                                                Failure, IsWeak);
+  return std::make_pair(Atomics.shouldUseLibcall()
+                            ? RValue::get(Pair.first)
+                            : Atomics.ConvertIntToValueOrAtomic(
+                                  Pair.first, Slot, Loc, /*AsValue=*/true),
+                        RValue::get(Pair.second));
 }
 
 void CodeGenFunction::EmitAtomicInit(Expr *init, LValue dest) {
   AtomicInfo atomics(*this, dest);
 
   switch (atomics.getEvaluationKind()) {
   case TEK_Scalar: {
     llvm::Value *value = EmitScalarExpr(init);

lib/CodeGen/CGStmtOpenMP.cpp

CGF.EmitStoreOfComplex(
         convertToComplexValue(CGF, Res, X->getType(), V->getType()), VLValue,
         /*isInit=*/false);
     break;
   case TEK_Aggregate:
     llvm_unreachable("Must be a scalar or complex.");
   }
 }
 
+static void EmitOMPAtomicWriteExpr(CodeGenFunction &CGF, bool IsSeqCst,
+                                   const Expr *X, const Expr *E,
+                                   SourceLocation Loc) {
+  // x = expr;
+  assert(X->isLValue() && "X of 'omp atomic write' is not lvalue");
+  LValue XLValue = CGF.EmitLValue(X);
+  RValue ExprRValue = CGF.EmitAnyExpr(E);
+  if (XLValue.isGlobalReg())
+    CGF.EmitStoreThroughGlobalRegLValue(ExprRValue, XLValue);
+  else
+    CGF.EmitAtomicStore(ExprRValue, XLValue, /*isInit=*/false);
+  // OpenMP, 2.12.6, atomic Construct
+  // Any atomic construct with a seq_cst clause forces the atomically
+  // performed operation to include an implicit flush operation without a
+  // list.
+  if (IsSeqCst)
+    CGF.CGM.getOpenMPRuntime().EmitOMPFlush(CGF, llvm::None, Loc);
+}
+
 static void EmitOMPAtomicExpr(CodeGenFunction &CGF, OpenMPClauseKind Kind,
                               bool IsSeqCst, const Expr *X, const Expr *V,
-                              const Expr *, SourceLocation Loc) {
+                              const Expr *E, SourceLocation Loc) {
   switch (Kind) {
   case OMPC_read:
     EmitOMPAtomicReadExpr(CGF, IsSeqCst, X, V, Loc);
     break;
   case OMPC_write:
+    EmitOMPAtomicWriteExpr(CGF, IsSeqCst, X, E, Loc);
+    break;
   case OMPC_update:
   case OMPC_capture:
     llvm_unreachable("CodeGen for 'omp atomic clause' is not supported yet.");
   case OMPC_if:
   case OMPC_final:
   case OMPC_num_threads:
   case OMPC_private:
   case OMPC_firstprivate:

lib/Sema/SemaOpenMP.cpp

if (AtomicKind == OMPC_read) {
     SourceLocation ErrorLoc, NoteLoc;
     SourceRange ErrorRange, NoteRange;
     // If clause is write:
     //  x = expr;
     if (auto AtomicBody = dyn_cast<Expr>(Body)) {
       auto AtomicBinOp =
           dyn_cast<BinaryOperator>(AtomicBody->IgnoreParenImpCasts());
       if (AtomicBinOp && AtomicBinOp->getOpcode() == BO_Assign) {
-        X = AtomicBinOp->getLHS()->IgnoreParenImpCasts();
-        E = AtomicBinOp->getRHS()->IgnoreParenImpCasts();
+        X = AtomicBinOp->getLHS();
+        E = AtomicBinOp->getRHS();
         if ((X->isInstantiationDependent() || X->getType()->isScalarType()) &&
             (E->isInstantiationDependent() || E->getType()->isScalarType())) {
           if (!X->isLValue()) {
             ErrorFound = NotAnLValue;
             ErrorLoc = AtomicBinOp->getExprLoc();
             ErrorRange = AtomicBinOp->getSourceRange();
             NoteLoc = X->getExprLoc();
             NoteRange = X->getSourceRange();

test/OpenMP/atomic_read_codegen.c

 // CHECK: [[LD:%.+]] = load atomic i8* getelementptr inbounds (%struct.BitFields4_packed* @bfx4_packed, i32 0, i32 0, i64 2) seq_cst
 // CHECK: store i8 [[LD]], i8* [[LDTEMP:%.+]]
 // CHECK: [[LD:%.+]] = load i8* [[LDTEMP]]
 // CHECK: [[ASHR:%.+]] = ashr i8 [[LD]], 1
 // CHECK: sext i8 [[ASHR]] to i64
 // CHECK: store x86_fp80
 #pragma omp atomic read
   ldv = bfx4_packed.b;
-// CHECK: [[LD:%.+]] = load atomic i32* bitcast (<2 x float>* @{{.+}} to i32*) seq_cst
-// CHECK: [[BITCAST:%.+]] = bitcast <2 x float>* [[LDTEMP:%.+]] to i32*
-// CHECK: store i32 [[LD]], i32* [[BITCAST]]
+// CHECK: [[LD:%.+]] = load atomic i64* bitcast (<2 x float>* @{{.+}} to i64*) seq_cst
+// CHECK: [[BITCAST:%.+]] = bitcast <2 x float>* [[LDTEMP:%.+]] to i64*
+// CHECK: store i64 [[LD]], i64* [[BITCAST]]
 // CHECK: [[LD:%.+]] = load <2 x float>* [[LDTEMP]]
 // CHECK: extractelement <2 x float> [[LD]]
 // CHECK: store i64
 #pragma omp atomic read
   ulv = float2x.x;
 // CHECK: call{{.*}} i{{[0-9]+}} @llvm.read_register
 // CHECK: call{{.*}} @__kmpc_flush(
 // CHECK: store double
 #pragma omp atomic read seq_cst
   dv = rix;
   return 0;
 }
 
 #endif

test/OpenMP/atomic_write_codegen.c

+// RUN: %clang_cc1 -verify -triple x86_64-apple-darwin10 -fopenmp=libiomp5 -x c -emit-llvm %s -o - | FileCheck %s
+// RUN: %clang_cc1 -fopenmp=libiomp5 -x c -triple x86_64-apple-darwin10 -emit-pch -o %t %s
+// RUN: %clang_cc1 -fopenmp=libiomp5 -x c -triple x86_64-apple-darwin10 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s
+// expected-no-diagnostics
+
+#ifndef HEADER
+#define HEADER
+
+_Bool bv, bx;
+char cv, cx;
+unsigned char ucv, ucx;
+short sv, sx;
+unsigned short usv, usx;
+int iv, ix;
+unsigned int uiv, uix;
+long lv, lx;
+unsigned long ulv, ulx;
+long long llv, llx;
+unsigned long long ullv, ullx;
+float fv, fx;
+double dv, dx;
+long double ldv, ldx;
+_Complex int civ, cix;
+_Complex float cfv, cfx;
+_Complex double cdv, cdx;
+
+typedef int int4 __attribute__((__vector_size__(16)));
+int4 int4x;
+
+struct BitFields {
+  int : 32;
+  int a : 31;
+} bfx;
+
+struct BitFields_packed {
+  int : 32;
+  int a : 31;
+} __attribute__ ((__packed__)) bfx_packed;
+
+struct BitFields2 {
+  int : 31;
+  int a : 1;
+} bfx2;
+
+struct BitFields2_packed {
+  int : 31;
+  int a : 1;
+} __attribute__ ((__packed__)) bfx2_packed;
+
+struct BitFields3 {
+  int : 11;
+  int a : 14;
+} bfx3;
+
+struct BitFields3_packed {
+  int : 11;
+  int a : 14;
+} __attribute__ ((__packed__)) bfx3_packed;
+
+struct BitFields4 {
+  short : 16;
+  int a: 1;
+  long b : 7;
+} bfx4;
+
+struct BitFields4_packed {
+  short : 16;
+  int a: 1;
+  long b : 7;
+} __attribute__ ((__packed__)) bfx4_packed;
+
+typedef float float2 __attribute__((ext_vector_type(2)));
+float2 float2x;
+
+register int rix __asm__("0");
+
+int main() {
+// CHECK: load i8*
+// CHECK: store atomic i8
+#pragma omp atomic write
+  bx = bv;
+// CHECK: load i8*
+// CHECK: store atomic i8
+#pragma omp atomic write
+  cx = cv;
+// CHECK: load i8*
+// CHECK: store atomic i8
+#pragma omp atomic write
+  ucx = ucv;
+// CHECK: load i16*
+// CHECK: store atomic i16
+#pragma omp atomic write
+  sx = sv;
+// CHECK: load i16*
+// CHECK: store atomic i16
+#pragma omp atomic write
+  usx = usv;
+// CHECK: load i32*
+// CHECK: store atomic i32
+#pragma omp atomic write
+  ix = iv;
+// CHECK: load i32*
+// CHECK: store atomic i32
+#pragma omp atomic write
+  uix = uiv;
+// CHECK: load i64*
+// CHECK: store atomic i64
+#pragma omp atomic write
+  lx = lv;
+// CHECK: load i64*
+// CHECK: store atomic i64
+#pragma omp atomic write
+  ulx = ulv;
+// CHECK: load i64*
+// CHECK: store atomic i64
+#pragma omp atomic write
+  llx = llv;
+// CHECK: load i64*
+// CHECK: store atomic i64
+#pragma omp atomic write
+  ullx = ullv;
+// CHECK: load float*
+// CHECK: bitcast float {{.*}} to i32
+// CHECK: store atomic i32 {{.*}}, i32* bitcast (float*
+#pragma omp atomic write
+  fx = fv;
+// CHECK: load double*
+// CHECK: bitcast double {{.*}} to i64
+// CHECK: store atomic i64 {{.*}}, i64* bitcast (double*
+#pragma omp atomic write
+  dx = dv;
+// CHECK: [[LD:%.+]] = load x86_fp80*
+// CHECK: [[BITCAST:%.+]] = bitcast x86_fp80* [[LDTEMP:%.*]] to i8*
+// CHECK: call void @llvm.memset.p0i8.i64(i8* [[BITCAST]], i8 0, i64 16, i32 16, i1 false)
+// CHECK: store x86_fp80 [[LD]], x86_fp80* [[LDTEMP]]
+// CHECK: [[BITCAST:%.+]] = bitcast x86_fp80* [[LDTEMP:%.*]] to i128*
+// CHECK: [[LD:%.+]] = load i128* [[BITCAST]]
+// CHECK: store atomic i128 [[LD]], i128* bitcast (x86_fp80*
+#pragma omp atomic write
+  ldx = ldv;
+// CHECK: [[REAL_VAL:%.+]] = load i32* getelementptr inbounds ({ i32, i32 }* @{{.*}}, i32 0, i32 0)
+// CHECK: [[IMG_VAL:%.+]] = load i32* getelementptr inbounds ({ i32, i32 }* @{{.*}}, i32 0, i32 1)
+// CHECK: [[TEMP_REAL_REF:%.+]] = getelementptr inbounds { i32, i32 }* [[TEMP:%.+]], i32 0, i32 0
+// CHECK: [[TEMP_IMG_REF:%.+]] = getelementptr inbounds { i32, i32 }* [[TEMP]], i32 0, i32 1
+// CHECK: store i32 [[REAL_VAL]], i32* [[TEMP_REAL_REF]]
+// CHECK: store i32 [[IMG_VAL]], i32* [[TEMP_IMG_REF]]
+// CHECK: [[BITCAST:%.+]] = bitcast { i32, i32 }* [[TEMP]] to i8*
+// CHECK: call void @__atomic_store(i64 8, i8* bitcast ({ i32, i32 }* @{{.*}} to i8*), i8* [[BITCAST]], i32 5)
+#pragma omp atomic write
+  cix = civ;
+// CHECK: [[REAL_VAL:%.+]] = load float* getelementptr inbounds ({ float, float }* @{{.*}}, i32 0, i32 0)
+// CHECK: [[IMG_VAL:%.+]] = load float* getelementptr inbounds ({ float, float }* @{{.*}}, i32 0, i32 1)
+// CHECK: [[TEMP_REAL_REF:%.+]] = getelementptr inbounds { float, float }* [[TEMP:%.+]], i32 0, i32 0
+// CHECK: [[TEMP_IMG_REF:%.+]] = getelementptr inbounds { float, float }* [[TEMP]], i32 0, i32 1
+// CHECK: store float [[REAL_VAL]], float* [[TEMP_REAL_REF]]
+// CHECK: store float [[IMG_VAL]], float* [[TEMP_IMG_REF]]
+// CHECK: [[BITCAST:%.+]] = bitcast { float, float }* [[TEMP]] to i8*
+// CHECK: call void @__atomic_store(i64 8, i8* bitcast ({ float, float }* @{{.*}} to i8*), i8* [[BITCAST]], i32 5)
+#pragma omp atomic write
+  cfx = cfv;
+// CHECK: [[REAL_VAL:%.+]] = load double* getelementptr inbounds ({ double, double }* @{{.*}}, i32 0, i32 0)
+// CHECK: [[IMG_VAL:%.+]] = load double* getelementptr inbounds ({ double, double }* @{{.*}}, i32 0, i32 1)
+// CHECK: [[TEMP_REAL_REF:%.+]] = getelementptr inbounds { double, double }* [[TEMP:%.+]], i32 0, i32 0
+// CHECK: [[TEMP_IMG_REF:%.+]] = getelementptr inbounds { double, double }* [[TEMP]], i32 0, i32 1
+// CHECK: store double [[REAL_VAL]], double* [[TEMP_REAL_REF]]
+// CHECK: store double [[IMG_VAL]], double* [[TEMP_IMG_REF]]
+// CHECK: [[BITCAST:%.+]] = bitcast { double, double }* [[TEMP]] to i8*
+// CHECK: call void @__atomic_store(i64 16, i8* bitcast ({ double, double }* @{{.*}} to i8*), i8* [[BITCAST]], i32 5)
+// CHECK: call{{.*}} @__kmpc_flush(
+#pragma omp atomic seq_cst write
+  cdx = cdv;
+// CHECK: load i8*
+// CHECK: store atomic i64
+#pragma omp atomic write
+  ulx = bv;
+// CHECK: load i8*
+// CHECK: store atomic i8
+#pragma omp atomic write
+  bx = cv;
+// CHECK: load i8*
+// CHECK: store atomic i8
+// CHECK: call{{.*}} @__kmpc_flush(
+#pragma omp atomic write, seq_cst
+  cx = ucv;
+// CHECK: load i16*
+// CHECK: store atomic i64
+#pragma omp atomic write
+  ulx = sv;
+// CHECK: load i16*
+// CHECK: store atomic i64
+#pragma omp atomic write
+  lx = usv;
+// CHECK: load i32*
+// CHECK: store atomic i32
+// CHECK: call{{.*}} @__kmpc_flush(
+#pragma omp atomic seq_cst, write
+  uix = iv;
+// CHECK: load i32*
+// CHECK: store atomic i32
+#pragma omp atomic write
+  ix = uiv;
+// CHECK: load i64*
+// CHECK: [[VAL:%.+]] = trunc i64 %{{.*}} to i32
+// CHECK: [[TEMP_REAL_REF:%.+]] = getelementptr inbounds { i32, i32 }* [[TEMP:%.+]], i32 0, i32 0
+// CHECK: [[TEMP_IMG_REF:%.+]] = getelementptr inbounds { i32, i32 }* [[TEMP]], i32 0, i32 1
+// CHECK: store i32 [[VAL]], i32* [[TEMP_REAL_REF]]
+// CHECK: store i32 0, i32* [[TEMP_IMG_REF]]
+// CHECK: [[BITCAST:%.+]] = bitcast { i32, i32 }* [[TEMP]] to i8*
+// CHECK: call void @__atomic_store(i64 8, i8* bitcast ({ i32, i32 }* @{{.+}} to i8*), i8* [[BITCAST]], i32 5)
+#pragma omp atomic write
+  cix = lv;
+// CHECK: load i64*
+// CHECK: store atomic i32 %{{.+}}, i32* bitcast (float*
+#pragma omp atomic write
+  fx = ulv;
+// CHECK: load i64*
+// CHECK: store atomic i64 %{{.+}}, i64* bitcast (double*
+#pragma omp atomic write
+  dx = llv;
+// CHECK: load i64*
+// CHECK: [[VAL:%.+]] = uitofp i64 %{{.+}} to x86_fp80
+// CHECK: [[BITCAST:%.+]] = bitcast x86_fp80* [[TEMP:%.+]] to i8*
+// CHECK: call void @llvm.memset.p0i8.i64(i8* [[BITCAST]], i8 0, i64 16, i32 16, i1 false)
+// CHECK: store x86_fp80 [[VAL]], x86_fp80* [[TEMP]]
+// CHECK: [[BITCAST:%.+]] = bitcast x86_fp80* [[TEMP]] to i128*
+// CHECK: [[VAL:%.+]] = load i128* [[BITCAST]]
+// CHECK: store atomic i128 [[VAL]], i128* bitcast (x86_fp80*
+#pragma omp atomic write
+  ldx = ullv;
+// CHECK: load float*
+// CHECK: [[VAL:%.+]] = fptosi float %{{.*}} to i32
+// CHECK: [[TEMP_REAL_REF:%.+]] = getelementptr inbounds { i32, i32 }* [[TEMP:%.+]], i32 0, i32 0
+// CHECK: [[TEMP_IMG_REF:%.+]] = getelementptr inbounds { i32, i32 }* [[TEMP]], i32 0, i32 1
+// CHECK: store i32 [[VAL]], i32* [[TEMP_REAL_REF]]
+// CHECK: store i32 0, i32* [[TEMP_IMG_REF]]
+// CHECK: [[BITCAST:%.+]] = bitcast { i32, i32 }* [[TEMP]] to i8*
+// CHECK: call void @__atomic_store(i64 8, i8* bitcast ({ i32, i32 }* @{{.+}} to i8*), i8* [[BITCAST]], i32 5)
+#pragma omp atomic write
+  cix = fv;
+// CHECK: load double*
+// CHECK: store atomic i16
+#pragma omp atomic write
+  sx = dv;
+// CHECK: load x86_fp80*
+// CHECK: store atomic i8
+#pragma omp atomic write
+  bx = ldv;
+// CHECK: load i32* getelementptr inbounds ({ i32, i32 }* @{{.+}}, i32 0, i32 0)
+// CHECK: load i32* getelementptr inbounds ({ i32, i32 }* @{{.+}}, i32 0, i32 1)
+// CHECK: icmp ne i32 %{{.+}}, 0
+// CHECK: icmp ne i32 %{{.+}}, 0
+// CHECK: or i1
+// CHECK: store atomic i8
+#pragma omp atomic write
+  bx = civ;
+// CHECK: load float* getelementptr inbounds ({ float, float }* @{{.*}}, i32 0, i32 0)
+// CHECK: store atomic i16
+#pragma omp atomic write
+  usx = cfv;
+// CHECK: load double* getelementptr inbounds ({ double, double }* @{{.+}}, i32 0, i32 0)
+// CHECK: store atomic i64
+#pragma omp atomic write
+  llx = cdv;
+// CHECK: [[IDX:%.+]] = load i16* @{{.+}}
+// CHECK: load i8*
+// CHECK: [[VEC_ITEM_VAL:%.+]] = zext i1 %{{.+}} to i32
+// CHECK: [[I128VAL:%.+]] = load atomic i128* bitcast (<4 x i32>* [[DEST:@.+]] to i128*) seq_cst
+// CHECK: [[LD:%.+]] = bitcast i128 [[I128VAL]] to <4 x i32>
+// CHECK: br label %[[CONT:.+]]
+// CHECK: [[CONT]]
+// CHECK: [[OLD_VEC_VAL:%.+]] = phi <4 x i32> [ [[LD]], %{{.+}} ], [ [[FAILED_OLD_VAL:%.+]], %[[CONT]] ]
+// CHECK: store <4 x i32> [[OLD_VEC_VAL]], <4 x i32>* [[LDTEMP:%.+]],
+// CHECK: [[VEC_VAL:%.+]] = load <4 x i32>* [[LDTEMP]]
+// CHECK: [[NEW_VEC_VAL:%.+]] = insertelement <4 x i32> [[VEC_VAL]], i32 [[VEC_ITEM_VAL]], i16 [[IDX]]
+// CHECK: store <4 x i32> [[NEW_VEC_VAL]], <4 x i32>* [[LDTEMP]]
+// CHECK: [[NEW_VEC_VAL:%.+]] = load <4 x i32>* [[LDTEMP]]
+// CHECK: [[OLD_I128:%.+]] = bitcast <4 x i32> [[OLD_VEC_VAL]] to i128
+// CHECK: [[NEW_I128:%.+]] = bitcast <4 x i32> [[NEW_VEC_VAL]] to i128
+// CHECK: [[RES:%.+]] = cmpxchg i128* bitcast (<4 x i32>* [[DEST]] to i128*), i128 [[OLD_I128]], i128 [[NEW_I128]] seq_cst seq_cst
+// CHECK: [[FAILED_I128_OLD_VAL:%.+]] = extractvalue { i128, i1 } [[RES]], 0
+// CHECK: [[FAIL_SUCCESS:%.+]] = extractvalue { i128, i1 } [[RES]], 1
+// CHECK: [[FAILED_OLD_VAL]] = bitcast i128 [[FAILED_I128_OLD_VAL]] to <4 x i32>
+// CHECK: br i1 [[FAIL_SUCCESS]], label %[[CONT]], label %[[EXIT:.+]]
+// CHECK: [[EXIT]]
+#pragma omp atomic write
+  int4x[sv] = bv;
+// CHECK: load x86_fp80* @{{.+}}
+// CHECK: [[NEW_VAL:%.+]] = fptosi x86_fp80 %{{.+}} to i32
+// CHECK: [[PREV_VALUE:%.+]] = load atomic i32* bitcast (i8* getelementptr (i8* bitcast (%struct.BitFields* @{{.+}} to i8*), i64 4) to i32*) seq_cst
+// CHECK: br label %[[CONT:.+]]
+// CHECK: [[CONT]]
+// CHECK: [[OLD_BF_VALUE:%.+]] = phi i32 [ [[PREV_VALUE]], %[[EXIT]] ], [ [[FAILED_OLD_VAL:%.+]], %[[CONT]] ]
+// CHECK: [[BF_VALUE:%.+]] = and i32 [[NEW_VAL]], 2147483647
+// CHECK: [[BF_CLEAR:%.+]] = and i32 %{{.+}}, -2147483648
+// CHECK: or i32 [[BF_CLEAR]], [[BF_VALUE]]
+// CHECK: store i32 %{{.+}}, i32* [[LDTEMP:%.+]]
+// CHECK: [[NEW_BF_VALUE:%.+]] = load i32* [[LDTEMP]]
+// CHECK: [[RES:%.+]] = cmpxchg i32* bitcast (i8* getelementptr (i8* bitcast (%struct.BitFields* @{{.+}} to i8*), i64 4) to i32*), i32 [[OLD_BF_VALUE]], i32 [[NEW_BF_VALUE]] seq_cst seq_cst
+// CHECK: [[FAILED_OLD_VAL]] = extractvalue { i32, i1 } [[RES]], 0
+// CHECK: [[FAIL_SUCCESS:%.+]] = extractvalue { i32, i1 } [[RES]], 1
+// CHECK: br i1 [[FAIL_SUCCESS]], label %[[CONT]], label %[[EXIT:.+]]
+// CHECK: [[EXIT]]
+#pragma omp atomic write
+  bfx.a = ldv;
+// CHECK: load x86_fp80* @{{.+}}
+// CHECK: [[NEW_VAL:%.+]] = fptosi x86_fp80 %{{.+}} to i32
+// CHECK: [[BITCAST:%.+]] = bitcast i32* [[LDTEMP:%.+]] to i8*
+// CHECK: call void @__atomic_load(i64 4, i8* getelementptr (i8* bitcast (%struct.BitFields_packed* @{{.+}} to i8*), i64 4), i8* [[BITCAST]], i32 5)
+// CHECK: [[PREV_VALUE:%.+]] = load i32* [[LDTEMP]]
+// CHECK: br label %[[CONT:.+]]
+// CHECK: [[CONT]]
+// CHECK: [[OLD_BF_VALUE:%.+]] = phi i32 [ [[PREV_VALUE]], %[[EXIT]] ], [ [[FAILED_OLD_VAL:%.+]], %[[CONT]] ]
+// CHECK: [[BF_VALUE:%.+]] = and i32 [[NEW_VAL]], 2147483647
+// CHECK: [[BF_CLEAR:%.+]] = and i32 %{{.+}}, -2147483648
+// CHECK: or i32 [[BF_CLEAR]], [[BF_VALUE]]
+// CHECK: store i32 %{{.+}}, i32* [[LDTEMP:%.+]]
+// CHECK: [[NEW_BF_VALUE:%.+]] = load i32* [[LDTEMP]]
+// CHECK: store i32 [[OLD_BF_VALUE]], i32* [[TEMP_OLD_BF_ADDR:%.+]],
+// CHECK: store i32 [[NEW_BF_VALUE]], i32* [[TEMP_NEW_BF_ADDR:%.+]],
+// CHECK: [[BITCAST_TEMP_OLD_BF_ADDR:%.+]] = bitcast i32* [[TEMP_OLD_BF_ADDR]] to i8*
+// CHECK: [[BITCAST_TEMP_NEW_BF_ADDR:%.+]] = bitcast i32* [[TEMP_NEW_BF_ADDR]] to i8*
+// CHECK: [[FAIL_SUCCESS:%.+]] = call zeroext i1 @__atomic_compare_exchange(i64 4, i8* getelementptr (i8* bitcast (%struct.BitFields_packed* @{{.+}} to i8*), i64 4), i8* [[BITCAST_TEMP_OLD_BF_ADDR]], i8* [[BITCAST_TEMP_NEW_BF_ADDR]], i32 7, i32 7)
+// CHECK: [[FAILED_OLD_VAL]] = load i32* [[TEMP_OLD_BF_ADDR]]
+// CHECK: br i1 [[FAIL_SUCCESS]], label %[[CONT]], label %[[EXIT:.+]]
+// CHECK: [[EXIT]]
+#pragma omp atomic write
+  bfx_packed.a = ldv;
+// CHECK: load x86_fp80* @{{.+}}
+// CHECK: [[NEW_VAL:%.+]] = fptosi x86_fp80 %{{.+}} to i32
+// CHECK: [[PREV_VALUE:%.+]] = load atomic i32* getelementptr inbounds (%struct.BitFields2* @{{.+}}, i32 0, i32 0) seq_cst
+// CHECK: br label %[[CONT:.+]]
+// CHECK: [[CONT]]
+// CHECK: [[OLD_BF_VALUE:%.+]] = phi i32 [ [[PREV_VALUE]], %[[EXIT]] ], [ [[FAILED_OLD_VAL:%.+]], %[[CONT]] ]
+// CHECK: [[BF_AND:%.+]] = and i32 [[NEW_VAL]], 1
+// CHECK: [[BF_VALUE:%.+]] = shl i32 [[BF_AND]], 31
+// CHECK: [[BF_CLEAR:%.+]] = and i32 %{{.+}}, 2147483647
+// CHECK: or i32 [[BF_CLEAR]], [[BF_VALUE]]
+// CHECK: store i32 %{{.+}}, i32* [[LDTEMP:%.+]]
+// CHECK: [[NEW_BF_VALUE:%.+]] = load i32* [[LDTEMP]]
+// CHECK: [[RES:%.+]] = cmpxchg i32* getelementptr inbounds (%struct.BitFields2* @{{.+}}, i32 0, i32 0), i32 [[OLD_BF_VALUE]], i32 [[NEW_BF_VALUE]] seq_cst seq_cst
+// CHECK: [[FAILED_OLD_VAL]] = extractvalue { i32, i1 } [[RES]], 0
+// CHECK: [[FAIL_SUCCESS:%.+]] = extractvalue { i32, i1 } [[RES]], 1
+// CHECK: br i1 [[FAIL_SUCCESS]], label %[[CONT]], label %[[EXIT:.+]]
+// CHECK: [[EXIT]]
+#pragma omp atomic write
+  bfx2.a = ldv;
+// CHECK: load x86_fp80* @{{.+}}
+// CHECK: [[NEW_VAL:%.+]] = fptosi x86_fp80 %{{.+}} to i32
+// CHECK: [[PREV_VALUE:%.+]] = load atomic i8* getelementptr (i8* bitcast (%struct.BitFields2_packed* @{{.+}} to i8*), i64 3) seq_cst
+// CHECK: br label %[[CONT:.+]]
+// CHECK: [[CONT]]
+// CHECK: [[OLD_BF_VALUE:%.+]] = phi i8 [ [[PREV_VALUE]], %[[EXIT]] ], [ [[FAILED_OLD_VAL:%.+]], %[[CONT]] ]
+// CHECK: [[TRUNC:%.+]] = trunc i32 [[NEW_VAL]] to i8
+// CHECK: [[BF_AND:%.+]] = and i8 [[TRUNC]], 1
+// CHECK: [[BF_VALUE:%.+]] = shl i8 [[BF_AND]], 7
+// CHECK: [[BF_CLEAR:%.+]] = and i8 %{{.+}}, 127
+// CHECK: or i8 [[BF_CLEAR]], [[BF_VALUE]]
+// CHECK: store i8 %{{.+}}, i8* [[LDTEMP:%.+]]
+// CHECK: [[NEW_BF_VALUE:%.+]] = load i8* [[LDTEMP]]
+// CHECK: [[RES:%.+]] = cmpxchg i8* getelementptr (i8* bitcast (%struct.BitFields2_packed* @{{.+}} to i8*), i64 3), i8 [[OLD_BF_VALUE]], i8 [[NEW_BF_VALUE]] seq_cst seq_cst
+// CHECK: [[FAILED_OLD_VAL]] = extractvalue { i8, i1 } [[RES]], 0
+// CHECK: [[FAIL_SUCCESS:%.+]] = extractvalue { i8, i1 } [[RES]], 1
+// CHECK: br i1 [[FAIL_SUCCESS]], label %[[CONT]], label %[[EXIT:.+]]
+// CHECK: [[EXIT]]
+#pragma omp atomic write
+  bfx2_packed.a = ldv;
+// CHECK: load x86_fp80* @{{.+}}
+// CHECK: [[NEW_VAL:%.+]] = fptosi x86_fp80 %{{.+}} to i32
+// CHECK: [[PREV_VALUE:%.+]] = load atomic i32* getelementptr inbounds (%struct.BitFields3* @{{.+}}, i32 0, i32 0) seq_cst
+// CHECK: br label %[[CONT:.+]]
+// CHECK: [[CONT]]
+// CHECK: [[OLD_BF_VALUE:%.+]] = phi i32 [ [[PREV_VALUE]], %[[EXIT]] ], [ [[FAILED_OLD_VAL:%.+]], %[[CONT]] ]
+// CHECK: [[BF_AND:%.+]] = and i32 [[NEW_VAL]], 16383
+// CHECK: [[BF_VALUE:%.+]] = shl i32 [[BF_AND]], 11
+// CHECK: [[BF_CLEAR:%.+]] = and i32 %{{.+}}, -33552385
+// CHECK: or i32 [[BF_CLEAR]], [[BF_VALUE]]
+// CHECK: store i32 %{{.+}}, i32* [[LDTEMP:%.+]]
+// CHECK: [[NEW_BF_VALUE:%.+]] = load i32* [[LDTEMP]]
+// CHECK: [[RES:%.+]] = cmpxchg i32* getelementptr inbounds (%struct.BitFields3* @{{.+}}, i32 0, i32 0), i32 [[OLD_BF_VALUE]], i32 [[NEW_BF_VALUE]] seq_cst seq_cst
+// CHECK: [[FAILED_OLD_VAL]] = extractvalue { i32, i1 } [[RES]], 0
+// CHECK: [[FAIL_SUCCESS:%.+]] = extractvalue { i32, i1 } [[RES]], 1
+// CHECK: br i1 [[FAIL_SUCCESS]], label %[[CONT]], label %[[EXIT:.+]]
+// CHECK: [[EXIT]]
+#pragma omp atomic write
+  bfx3.a = ldv;
+// CHECK: load x86_fp80* @{{.+}}
+// CHECK: [[NEW_VAL:%.+]] = fptosi x86_fp80 %{{.+}} to i32
+// CHECK: [[LDTEMP:%.+]] = bitcast i32* %{{.+}} to i24*
+// CHECK: [[BITCAST:%.+]] = bitcast i24* %{{.+}} to i8*
+// CHECK: call void @__atomic_load(i64 3, i8* getelementptr (i8* bitcast (%struct.BitFields3_packed* @{{.+}} to i8*), i64 1), i8* [[BITCAST]], i32 5)
+// CHECK: [[PREV_VALUE:%.+]] = load i24* [[LDTEMP]]
+// CHECK: br label %[[CONT:.+]]
+// CHECK: [[CONT]]
+// CHECK: [[OLD_BF_VALUE:%.+]] = phi i24 [ [[PREV_VALUE]], %[[EXIT]] ], [ [[FAILED_OLD_VAL:%.+]], %[[CONT]] ]
+// CHECK: [[TRUNC:%.+]] = trunc i32 [[NEW_VAL]] to i24
+// CHECK: [[BF_AND:%.+]] = and i24 [[TRUNC]], 16383
+// CHECK: [[BF_VALUE:%.+]] = shl i24 [[BF_AND]], 3
+// CHECK: [[BF_CLEAR:%.+]] = and i24 %{{.+}}, -131065
+// CHECK: or i24 [[BF_CLEAR]], [[BF_VALUE]]
+// CHECK: store i24 %{{.+}}, i24* [[LDTEMP:%.+]]
+// CHECK: [[NEW_BF_VALUE:%.+]] = load i24* [[LDTEMP]]
+// CHECK: [[TEMP_OLD_BF_ADDR:%.+]] = bitcast i32* %{{.+}} to i24*
+// CHECK: store i24 [[OLD_BF_VALUE]], i24* [[TEMP_OLD_BF_ADDR]]
+// CHECK: [[TEMP_NEW_BF_ADDR:%.+]] = bitcast i32* %{{.+}} to i24*
+// CHECK: store i24 [[NEW_BF_VALUE]], i24* [[TEMP_NEW_BF_ADDR]]
+// CHECK: [[BITCAST_TEMP_OLD_BF_ADDR:%.+]] = bitcast i24* [[TEMP_OLD_BF_ADDR]] to i8*
+// CHECK: [[BITCAST_TEMP_NEW_BF_ADDR:%.+]] = bitcast i24* [[TEMP_NEW_BF_ADDR]] to i8*
+// CHECK: [[FAIL_SUCCESS:%.+]] = call zeroext i1 @__atomic_compare_exchange(i64 3, i8* getelementptr (i8* bitcast (%struct.BitFields3_packed* @{{.+}} to i8*), i64 1), i8* [[BITCAST_TEMP_OLD_BF_ADDR]], i8* [[BITCAST_TEMP_NEW_BF_ADDR]], i32 7, i32 7)
+// CHECK: [[FAILED_OLD_VAL]] = load i24* [[TEMP_OLD_BF_ADDR]]
+// CHECK: br i1 [[FAIL_SUCCESS]], label %[[CONT]], label %[[EXIT:.+]]
+// CHECK: [[EXIT]]
+#pragma omp atomic write
+  bfx3_packed.a = ldv;
+// CHECK: load x86_fp80* @{{.+}}
+// CHECK: [[NEW_VAL:%.+]] = fptosi x86_fp80 %{{.+}} to i32
+// CHECK: [[PREV_VALUE:%.+]] = load atomic i64* bitcast (%struct.BitFields4* @{{.+}} to i64*) seq_cst
+// CHECK: br label %[[CONT:.+]]
+// CHECK: [[CONT]]
+// CHECK: [[OLD_BF_VALUE:%.+]] = phi i64 [ [[PREV_VALUE]], %[[EXIT]] ], [ [[FAILED_OLD_VAL:%.+]], %[[CONT]] ]
+// CHECK: [[ZEXT:%.+]] = zext i32 [[NEW_VAL]] to i64
+// CHECK: [[BF_AND:%.+]] = and i64 [[ZEXT]], 1
+// CHECK: [[BF_VALUE:%.+]] = shl i64 [[BF_AND]], 16
+// CHECK: [[BF_CLEAR:%.+]] = and i64 %{{.+}}, -65537
+// CHECK: or i64 [[BF_CLEAR]], [[BF_VALUE]]
+// CHECK: store i64 %{{.+}}, i64* [[LDTEMP:%.+]]
+// CHECK: [[NEW_BF_VALUE:%.+]] = load i64* [[LDTEMP]]
+// CHECK: [[RES:%.+]] = cmpxchg i64* bitcast (%struct.BitFields4* @{{.+}} to i64*), i64 [[OLD_BF_VALUE]], i64 [[NEW_BF_VALUE]] seq_cst seq_cst
+// CHECK: [[FAILED_OLD_VAL]] = extractvalue { i64, i1 } [[RES]], 0
+// CHECK: [[FAIL_SUCCESS:%.+]] = extractvalue { i64, i1 } [[RES]], 1
+// CHECK: br i1 [[FAIL_SUCCESS]], label %[[CONT]], label %[[EXIT:.+]]
+// CHECK: [[EXIT]]
+#pragma omp atomic write
+  bfx4.a = ldv;
+// CHECK: load x86_fp80* @{{.+}}
+// CHECK: [[NEW_VAL:%.+]] = fptosi x86_fp80 %{{.+}} to i32
+// CHECK: [[PREV_VALUE:%.+]] = load atomic i8* getelementptr inbounds (%struct.BitFields4_packed* @{{.+}}, i32 0, i32 0, i64 2) seq_cst
+// CHECK: br label %[[CONT:.+]]
+// CHECK: [[CONT]]
+// CHECK: [[OLD_BF_VALUE:%.+]] = phi i8 [ [[PREV_VALUE]], %[[EXIT]] ], [ [[FAILED_OLD_VAL:%.+]], %[[CONT]] ]
+// CHECK: [[TRUNC:%.+]] = trunc i32 [[NEW_VAL]] to i8
+// CHECK: [[BF_VALUE:%.+]] = and i8 [[TRUNC]], 1
+// CHECK: [[BF_CLEAR:%.+]] = and i8 %{{.+}}, -2
+// CHECK: or i8 [[BF_CLEAR]], [[BF_VALUE]]
+// CHECK: store i8 %{{.+}}, i8* [[LDTEMP:%.+]]
+// CHECK: [[NEW_BF_VALUE:%.+]] = load i8* [[LDTEMP]]
+// CHECK: [[RES:%.+]] = cmpxchg i8* getelementptr inbounds (%struct.BitFields4_packed* @{{.+}}, i32 0, i32 0, i64 2), i8 [[OLD_BF_VALUE]], i8 [[NEW_BF_VALUE]] seq_cst seq_cst
+// CHECK: [[FAILED_OLD_VAL]] = extractvalue { i8, i1 } [[RES]], 0
+// CHECK: [[FAIL_SUCCESS:%.+]] = extractvalue { i8, i1 } [[RES]], 1
+// CHECK: br i1 [[FAIL_SUCCESS]], label %[[CONT]], label %[[EXIT:.+]]
+// CHECK: [[EXIT]]
+#pragma omp atomic write
+  bfx4_packed.a = ldv;
+// CHECK: load x86_fp80* @{{.+}}
+// CHECK: [[NEW_VAL:%.+]] = fptosi x86_fp80 %{{.+}} to i64
+// CHECK: [[PREV_VALUE:%.+]] = load atomic i64* bitcast (%struct.BitFields4* @{{.+}} to i64*) seq_cst
+// CHECK: br label %[[CONT:.+]]
+// CHECK: [[CONT]]
+// CHECK: [[OLD_BF_VALUE:%.+]] = phi i64 [ [[PREV_VALUE]], %[[EXIT]] ], [ [[FAILED_OLD_VAL:%.+]], %[[CONT]] ]
+// CHECK: [[BF_AND:%.+]] = and i64 [[NEW_VAL]], 127
+// CHECK: [[BF_VALUE:%.+]] = shl i64 [[BF_AND]], 17
+// CHECK: [[BF_CLEAR:%.+]] = and i64 %{{.+}}, -16646145
+// CHECK: or i64 [[BF_CLEAR]], [[BF_VALUE]]
+// CHECK: store i64 %{{.+}}, i64* [[LDTEMP:%.+]]
+// CHECK: [[NEW_BF_VALUE:%.+]] = load i64* [[LDTEMP]]
+// CHECK: [[RES:%.+]] = cmpxchg i64* bitcast (%struct.BitFields4* @{{.+}} to i64*), i64 [[OLD_BF_VALUE]], i64 [[NEW_BF_VALUE]] seq_cst seq_cst
+// CHECK: [[FAILED_OLD_VAL]] = extractvalue { i64, i1 } [[RES]], 0
+// CHECK: [[FAIL_SUCCESS:%.+]] = extractvalue { i64, i1 } [[RES]], 1
+// CHECK: br i1 [[FAIL_SUCCESS]], label %[[CONT]], label %[[EXIT:.+]]
+// CHECK: [[EXIT]]
+#pragma omp atomic write
+  bfx4.b = ldv;
+// CHECK: load x86_fp80* @{{.+}}
+// CHECK: [[NEW_VAL:%.+]] = fptosi x86_fp80 %{{.+}} to i64
+// CHECK: [[PREV_VALUE:%.+]] = load atomic i8* getelementptr inbounds (%struct.BitFields4_packed* @{{.+}}, i32 0, i32 0, i64 2) seq_cst
+// CHECK: br label %[[CONT:.+]]
+// CHECK: [[CONT]]
+// CHECK: [[OLD_BF_VALUE:%.+]] = phi i8 [ [[PREV_VALUE]], %[[EXIT]] ], [ [[FAILED_OLD_VAL:%.+]], %[[CONT]] ]
+// CHECK: [[TRUNC:%.+]] = trunc i64 [[NEW_VAL]] to i8
+// CHECK: [[BF_AND:%.+]] = and i8 [[TRUNC]], 127
+// CHECK: [[BF_VALUE:%.+]] = shl i8 [[BF_AND]], 1
+// CHECK: [[BF_CLEAR:%.+]] = and i8 %{{.+}}, 1
+// CHECK: or i8 [[BF_CLEAR]], [[BF_VALUE]]
+// CHECK: store i8 %{{.+}}, i8* [[LDTEMP:%.+]]
+// CHECK: [[NEW_BF_VALUE:%.+]] = load i8* [[LDTEMP]]
+// CHECK: [[RES:%.+]] = cmpxchg i8* getelementptr inbounds (%struct.BitFields4_packed* @{{.+}}, i32 0, i32 0, i64 2), i8 [[OLD_BF_VALUE]], i8 [[NEW_BF_VALUE]] seq_cst seq_cst
+// CHECK: [[FAILED_OLD_VAL]] = extractvalue { i8, i1 } [[RES]], 0
+// CHECK: [[FAIL_SUCCESS:%.+]] = extractvalue { i8, i1 } [[RES]], 1
+// CHECK: br i1 [[FAIL_SUCCESS]], label %[[CONT]], label %[[EXIT:.+]]
+// CHECK: [[EXIT]]
+#pragma omp atomic write
+  bfx4_packed.b = ldv;
+// CHECK: load i64*
+// CHECK: [[VEC_ITEM_VAL:%.+]] = uitofp i64 %{{.+}} to float
+// CHECK: [[I64VAL:%.+]] = load atomic i64* bitcast (<2 x float>* [[DEST:@.+]] to i64*) seq_cst
+// CHECK: [[LD:%.+]] = bitcast i64 [[I64VAL]] to <2 x float>
+// CHECK: br label %[[CONT:.+]]
+// CHECK: [[CONT]]
+// CHECK: [[OLD_VEC_VAL:%.+]] = phi <2 x float> [ [[LD]], %{{.+}} ], [ [[FAILED_OLD_VAL:%.+]], %[[CONT]] ]
+// CHECK: store <2 x float> [[OLD_VEC_VAL]], <2 x float>* [[LDTEMP:%.+]],
+// CHECK: [[VEC_VAL:%.+]] = load <2 x float>* [[LDTEMP]]
+// CHECK: [[NEW_VEC_VAL:%.+]] = insertelement <2 x float> [[VEC_VAL]], float [[VEC_ITEM_VAL]], i64 0
+// CHECK: store <2 x float> [[NEW_VEC_VAL]], <2 x float>* [[LDTEMP]]
+// CHECK: [[NEW_VEC_VAL:%.+]] = load <2 x float>* [[LDTEMP]]
+// CHECK: [[OLD_I64:%.+]] = bitcast <2 x float> [[OLD_VEC_VAL]] to i64
+// CHECK: [[NEW_I64:%.+]] = bitcast <2 x float> [[NEW_VEC_VAL]] to i64
+// CHECK: [[RES:%.+]] = cmpxchg i64* bitcast (<2 x float>* [[DEST]] to i64*), i64 [[OLD_I64]], i64 [[NEW_I64]] seq_cst seq_cst
+// CHECK: [[FAILED_I64_OLD_VAL:%.+]] = extractvalue { i64, i1 } [[RES]], 0
+// CHECK: [[FAIL_SUCCESS:%.+]] = extractvalue { i64, i1 } [[RES]], 1
+// CHECK: [[FAILED_OLD_VAL]] = bitcast i64 [[FAILED_I64_OLD_VAL]] to <2 x float>
+// CHECK: br i1 [[FAIL_SUCCESS]], label %[[CONT]], label %[[EXIT:.+]]
+// CHECK: [[EXIT]]
+#pragma omp atomic write
+  float2x.x = ulv;
+// CHECK: call i32 @llvm.read_register.i32(
+// CHECK: sitofp i32 %{{.+}} to double
+// CHECK: bitcast double %{{.+}} to i64
+// CHECK: store atomic i64 %{{.+}}, i64* bitcast (double* @{{.+}} to i64*) seq_cst
+// CHECK: call{{.*}} @__kmpc_flush(
+#pragma omp atomic write seq_cst
+  dv = rix;
+  return 0;
+}
+
+#endif