[SelectionDAG] Enable target specific vector scalarization of calls and returns

[sdardis]

Summary

By target hookifying getRegisterType, getNumRegisters, getVectorBreakdown,
backends can request that LLVM to scalarize vector types for calls
and returns.

The MIPS vector ABI requires that vector arguments and returns are passed in
integer registers. With SelectionDAG's new hooks, the MIPS backend can now
handle LLVM-IR with vector types in calls and returns. E.g.
'call @foo(<4 x i32> %4)'.

Previously these cases would be scalarized for the MIPS O32/N32/N64 ABI for
calls and returns if vector types were not legal. If vector types were legal,
a single 128bit vector argument would be assigned to a single 32 bit / 64 bit
integer register.

By teaching the MIPS backend to inspect the original types, it can now
implement the MIPS vector ABI which requires a particular method of
scalarizing vectors.

Previously, the MIPS backend relied on clang to scalarize types such as "call
@foo(<4 x float> %a) into "call @foo(i32 inreg %1, i32 inreg %2, i32 inreg %3,
i32 inreg %4)".

This patch enables the MIPS backend to take either form for vector types.

[sdardis]

Update to trunk.

[zoran.jovanovic]

LGTM with few nits.

include/llvm/Target/TargetLowering.h

974 ↗

(On Diff #87430)

Line exceeds 80 characters,

981 ↗

(On Diff #87430)

Line exceeds 80 characters.

987 ↗

(On Diff #87430)

Line exceeds 80 characters.

lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

725 ↗

(On Diff #87430)

Is it possible to save NumRegs value as a class member and use it in class member functions (e.g. getCopyToRegs)?

839 ↗

(On Diff #87430)

Is it possible to get undefined value for RegVTs[Value] if value of argument IsABIRegCopy is different from value used in class constructor?

1095 ↗

(On Diff #87430)

Is additional IsABIRegCopy parameter necessary?
Is it possible to use value passed to RegsForValue constructor instead?

lib/Target/Mips/MipsCCState.h

57 ↗

(On Diff #87430)

Line exceeds 80 characters.

59 ↗

(On Diff #87430)

Line exceeds 80 characters.

lib/Target/Mips/MipsCallingConv.td

42 ↗

(On Diff #87430)

Line exceeds 80 characters.

lib/Target/Mips/MipsISelLowering.cpp

2542 ↗

(On Diff #87430)

Line exceeds 80 characters.

[sdardis]

Addressed review comments.

I'll commit this in a day or two if there are no further comments.

llvm/trunk/include/llvm/Target/TargetLowering.h

public:
   /// This method returns the number of registers needed, and the VT for each
   /// register.  It also returns the VT and quantity of the intermediate values
   /// before they are promoted/expanded.
   unsigned getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
                                   EVT &IntermediateVT,
                                   unsigned &NumIntermediates,
                                   MVT &RegisterVT) const;
 
+  /// Certain targets such as MIPS require that some types such as vectors are
+  /// always broken down into scalars in some contexts. This occurs even if the
+  /// vector type is legal.
+  virtual unsigned getVectorTypeBreakdownForCallingConv(
+      LLVMContext &Context, EVT VT, EVT &IntermediateVT,
+      unsigned &NumIntermediates, MVT &RegisterVT) const {
+    return getVectorTypeBreakdown(Context, VT, IntermediateVT, NumIntermediates,
+                                  RegisterVT);
+  }
+
   struct IntrinsicInfo {
     unsigned     opc = 0;          // target opcode
     EVT          memVT;            // memory VT
     const Value* ptrVal = nullptr; // value representing memory location
     int          offset = 0;       // offset off of ptrVal
     unsigned     size = 0;         // the size of the memory location
                                    // (taken from memVT if zero)
     unsigned     align = 1;        // alignment
unsigned getNumRegisters(LLVMContext &Context, EVT VT) const {
     if (VT.isInteger()) {
       unsigned BitWidth = VT.getSizeInBits();
       unsigned RegWidth = getRegisterType(Context, VT).getSizeInBits();
       return (BitWidth + RegWidth - 1) / RegWidth;
     }
     llvm_unreachable("Unsupported extended type!");
   }
 
+  /// Certain combinations of ABIs, Targets and features require that types
+  /// are legal for some operations and not for other operations.
+  /// For MIPS all vector types must be passed through the integer register set.
+  virtual MVT getRegisterTypeForCallingConv(MVT VT) const {
+    return getRegisterType(VT);
+  }
+
+  virtual MVT getRegisterTypeForCallingConv(LLVMContext &Context,
+                                            EVT VT) const {
+    return getRegisterType(Context, VT);
+  }
+
+  /// Certain targets require unusual breakdowns of certain types. For MIPS,
+  /// this occurs when a vector type is used, as vector are passed through the
+  /// integer register set.
+  virtual unsigned getNumRegistersForCallingConv(LLVMContext &Context,
+                                                 EVT VT) const {
+    return getNumRegisters(Context, VT);
+  }
+
+  /// Certain targets have context senstive alignment requirements, where one
+  /// type has the alignment requirement of another type.
+  virtual unsigned getABIAlignmentForCallingConv(Type *ArgTy,
+                                                 DataLayout DL) const {
+    return DL.getABITypeAlignment(ArgTy);
+  }
+
   /// If true, then instruction selection should seek to shrink the FP constant
   /// of the specified type to a smaller type in order to save space and / or
   /// reduce runtime.
   virtual bool ShouldShrinkFPConstant(EVT) const { return true; }
 
   // Return true if it is profitable to reduce the given load node to a smaller
   // type.
   //

llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.h

struct RegsForValue {
   /// it is necessary to have a separate record of the types.
   SmallVector<MVT, 4> RegVTs;
 
   /// This list holds the registers assigned to the values.
   /// Each legal or promoted value requires one register, and each
   /// expanded value requires multiple registers.
   SmallVector<unsigned, 4> Regs;
 
+  /// This list holds the number of registers for each value.
+  SmallVector<unsigned, 4> RegCount;
+
+  /// Records if this value needs to be treated in an ABI dependant manner,
+  /// different to normal type legalization.
+  bool IsABIMangled;
+
   RegsForValue();
 
-  RegsForValue(const SmallVector<unsigned, 4> &regs, MVT regvt, EVT valuevt);
+  RegsForValue(const SmallVector<unsigned, 4> &regs, MVT regvt, EVT valuevt,
+               bool IsABIMangledValue = false);
 
   RegsForValue(LLVMContext &Context, const TargetLowering &TLI,
-               const DataLayout &DL, unsigned Reg, Type *Ty);
+               const DataLayout &DL, unsigned Reg, Type *Ty,
+               bool IsABIMangledValue = false);
 
   /// Add the specified values to this one.
   void append(const RegsForValue &RHS) {
     ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end());
     RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end());
     Regs.append(RHS.Regs.begin(), RHS.Regs.end());
+    RegCount.push_back(RHS.Regs.size());
   }
 
   /// Emit a series of CopyFromReg nodes that copies from this value and returns
   /// the result as a ValueVTs value. This uses Chain/Flag as the input and
   /// updates them for the output Chain/Flag. If the Flag pointer is NULL, no
   /// flag is used.
   SDValue getCopyFromRegs(SelectionDAG &DAG, FunctionLoweringInfo &FuncInfo,
                           const SDLoc &dl, SDValue &Chain, SDValue *Flag,

llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

 // frontend. It is easy to induce this behavior with .ll code such as:
 // %buffer = alloca [4096 x i8]
 // %data = load [4096 x i8]* %argPtr
 // store [4096 x i8] %data, [4096 x i8]* %buffer
 static const unsigned MaxParallelChains = 64;
 
 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, const SDLoc &DL,
                                       const SDValue *Parts, unsigned NumParts,
-                                      MVT PartVT, EVT ValueVT, const Value *V);
+                                      MVT PartVT, EVT ValueVT, const Value *V,
+                                      bool IsABIRegCopy);
 
 /// getCopyFromParts - Create a value that contains the specified legal parts
 /// combined into the value they represent.  If the parts combine to a type
 /// larger than ValueVT then AssertOp can be used to specify whether the extra
 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
 /// (ISD::AssertSext).
 static SDValue getCopyFromParts(SelectionDAG &DAG, const SDLoc &DL,
                                 const SDValue *Parts, unsigned NumParts,
                                 MVT PartVT, EVT ValueVT, const Value *V,
-                                Optional<ISD::NodeType> AssertOp = None) {
+                                Optional<ISD::NodeType> AssertOp = None,
+                                bool IsABIRegCopy = false) {
   if (ValueVT.isVector())
     return getCopyFromPartsVector(DAG, DL, Parts, NumParts,
-                                  PartVT, ValueVT, V);
+                                  PartVT, ValueVT, V, IsABIRegCopy);
 
   assert(NumParts > 0 && "No parts to assemble!");
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
   SDValue Val = Parts[0];
 
   if (NumParts > 1) {
     // Assemble the value from multiple parts.
     if (ValueVT.isInteger()) {

 
 /// getCopyFromPartsVector - Create a value that contains the specified legal
 /// parts combined into the value they represent.  If the parts combine to a
 /// type larger than ValueVT then AssertOp can be used to specify whether the
 /// extra bits are known to be zero (ISD::AssertZext) or sign extended from
 /// ValueVT (ISD::AssertSext).
 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, const SDLoc &DL,
                                       const SDValue *Parts, unsigned NumParts,
-                                      MVT PartVT, EVT ValueVT, const Value *V) {
+                                      MVT PartVT, EVT ValueVT, const Value *V,
+                                      bool IsABIRegCopy) {
   assert(ValueVT.isVector() && "Not a vector value");
   assert(NumParts > 0 && "No parts to assemble!");
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
   SDValue Val = Parts[0];
 
   // Handle a multi-element vector.
   if (NumParts > 1) {
     EVT IntermediateVT;
     MVT RegisterVT;
     unsigned NumIntermediates;
-    unsigned NumRegs =
+    unsigned NumRegs;
+
+    if (IsABIRegCopy) {
+      NumRegs = TLI.getVectorTypeBreakdownForCallingConv(
+          *DAG.getContext(), ValueVT, IntermediateVT, NumIntermediates,
+          RegisterVT);
+    } else {
+      NumRegs =
           TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT,
                                      NumIntermediates, RegisterVT);
+    }
+
     assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
     NumParts = NumRegs; // Silence a compiler warning.
     assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
     assert(RegisterVT.getSizeInBits() ==
            Parts[0].getSimpleValueType().getSizeInBits() &&
            "Part type sizes don't match!");
 
     // Assemble the parts into intermediate operands.
if (NumIntermediates == NumParts) {
       unsigned Factor = NumParts / NumIntermediates;
       for (unsigned i = 0; i != NumIntermediates; ++i)
         Ops[i] = getCopyFromParts(DAG, DL, &Parts[i * Factor], Factor,
                                   PartVT, IntermediateVT, V);
     }
 
     // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the
     // intermediate operands.
+    EVT BuiltVectorTy =
+        EVT::getVectorVT(*DAG.getContext(), IntermediateVT.getScalarType(),
+                         (IntermediateVT.isVector()
+                              ? IntermediateVT.getVectorNumElements() * NumParts
+                              : NumIntermediates));
     Val = DAG.getNode(IntermediateVT.isVector() ? ISD::CONCAT_VECTORS
                                                 : ISD::BUILD_VECTOR,
-                      DL, ValueVT, Ops);
+                      DL, BuiltVectorTy, Ops);
   }
 
   // There is now one part, held in Val.  Correct it to match ValueVT.
   EVT PartEVT = Val.getValueType();
 
   if (PartEVT == ValueVT)
     return Val;
 
static SDValue getCopyFromPartsVector(SelectionDAG &DAG, const SDLoc &DL,
   }
 
   // Trivial bitcast if the types are the same size and the destination
   // vector type is legal.
   if (PartEVT.getSizeInBits() == ValueVT.getSizeInBits() &&
       TLI.isTypeLegal(ValueVT))
     return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
 
-  // Handle cases such as i8 -> <1 x i1>
   if (ValueVT.getVectorNumElements() != 1) {
-    diagnosePossiblyInvalidConstraint(*DAG.getContext(), V,
-                                      "non-trivial scalar-to-vector conversion");
+
+    // Certain ABIs require that vectors are passed as integers. For vectors
+    // are the same size, this is an obvious bitcast.
+    if (ValueVT.getSizeInBits() == PartEVT.getSizeInBits()) {
+      return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
+    } else if (ValueVT.getSizeInBits() < PartEVT.getSizeInBits()) {
+      // Bitcast Val back the original type and extract the corresponding
+      // vector we want.
+      unsigned Elts = PartEVT.getSizeInBits() / ValueVT.getScalarSizeInBits();
+      EVT WiderVecType = EVT::getVectorVT(*DAG.getContext(),
+                                          ValueVT.getVectorElementType(), Elts);
+      Val = DAG.getBitcast(WiderVecType, Val);
+      return DAG.getNode(
+          ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val,
+          DAG.getConstant(0, DL, TLI.getVectorIdxTy(DAG.getDataLayout())));
+    }
+
+    diagnosePossiblyInvalidConstraint(
+        *DAG.getContext(), V, "non-trivial scalar-to-vector conversion");
     return DAG.getUNDEF(ValueVT);
   }
 
+  // Handle cases such as i8 -> <1 x i1>
   if (ValueVT.getVectorNumElements() == 1 &&
       ValueVT.getVectorElementType() != PartEVT)
     Val = DAG.getAnyExtOrTrunc(Val, DL, ValueVT.getScalarType());
 
   return DAG.getNode(ISD::BUILD_VECTOR, DL, ValueVT, Val);
 }
 
 static void getCopyToPartsVector(SelectionDAG &DAG, const SDLoc &dl,
                                  SDValue Val, SDValue *Parts, unsigned NumParts,
-                                 MVT PartVT, const Value *V);
+                                 MVT PartVT, const Value *V, bool IsABIRegCopy);
 
 /// getCopyToParts - Create a series of nodes that contain the specified value
 /// split into legal parts.  If the parts contain more bits than Val, then, for
 /// integers, ExtendKind can be used to specify how to generate the extra bits.
 static void getCopyToParts(SelectionDAG &DAG, const SDLoc &DL, SDValue Val,
                            SDValue *Parts, unsigned NumParts, MVT PartVT,
                            const Value *V,
-                           ISD::NodeType ExtendKind = ISD::ANY_EXTEND) {
+                           ISD::NodeType ExtendKind = ISD::ANY_EXTEND,
+                           bool IsABIRegCopy = false) {
   EVT ValueVT = Val.getValueType();
 
   // Handle the vector case separately.
   if (ValueVT.isVector())
-    return getCopyToPartsVector(DAG, DL, Val, Parts, NumParts, PartVT, V);
+    return getCopyToPartsVector(DAG, DL, Val, Parts, NumParts, PartVT, V,
+                                IsABIRegCopy);
 
   unsigned PartBits = PartVT.getSizeInBits();
   unsigned OrigNumParts = NumParts;
   assert(DAG.getTargetLoweringInfo().isTypeLegal(PartVT) &&
          "Copying to an illegal type!");
 
   if (NumParts == 0)
     return;
if (DAG.getDataLayout().isBigEndian())
     std::reverse(Parts, Parts + OrigNumParts);
 }
 
 
 /// getCopyToPartsVector - Create a series of nodes that contain the specified
 /// value split into legal parts.
 static void getCopyToPartsVector(SelectionDAG &DAG, const SDLoc &DL,
                                  SDValue Val, SDValue *Parts, unsigned NumParts,
-                                 MVT PartVT, const Value *V) {
+                                 MVT PartVT, const Value *V,
+                                 bool IsABIRegCopy) {
+
   EVT ValueVT = Val.getValueType();
   assert(ValueVT.isVector() && "Not a vector");
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
 
   if (NumParts == 1) {
     EVT PartEVT = PartVT;
     if (PartEVT == ValueVT) {
       // Nothing to do.
if (PartEVT == ValueVT) {
       //Val = DAG.getNode(ISD::CONCAT_VECTORS, DL, PartVT, Val, UndefElts);
     } else if (PartVT.isVector() &&
                PartEVT.getVectorElementType().bitsGE(
                  ValueVT.getVectorElementType()) &&
                PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements()) {
 
       // Promoted vector extract
       Val = DAG.getAnyExtOrTrunc(Val, DL, PartVT);
-    } else{
+    } else {
       // Vector -> scalar conversion.
-      assert(ValueVT.getVectorNumElements() == 1 &&
-             "Only trivial vector-to-scalar conversions should get here!");
+      if (ValueVT.getVectorNumElements() == 1) {
         Val = DAG.getNode(
             ISD::EXTRACT_VECTOR_ELT, DL, PartVT, Val,
             DAG.getConstant(0, DL, TLI.getVectorIdxTy(DAG.getDataLayout())));
 
         Val = DAG.getAnyExtOrTrunc(Val, DL, PartVT);
+      } else {
+        assert(PartVT.getSizeInBits() > ValueVT.getSizeInBits() &&
+               "lossy conversion of vector to scalar type");
+        EVT IntermediateType = EVT::getIntegerVT(*DAG.getContext(),
+                                                 ValueVT.getSizeInBits());
+        Val = DAG.getBitcast(IntermediateType, Val);
+        Val = DAG.getAnyExtOrTrunc(Val, DL, PartVT);
+      }
     }
 
     Parts[0] = Val;
     return;
   }
 
   // Handle a multi-element vector.
   EVT IntermediateVT;
   MVT RegisterVT;
   unsigned NumIntermediates;
-  unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT,
-                                                IntermediateVT,
-                                                NumIntermediates, RegisterVT);
+  unsigned NumRegs;
+  if (IsABIRegCopy) {
+    NumRegs = TLI.getVectorTypeBreakdownForCallingConv(
+        *DAG.getContext(), ValueVT, IntermediateVT, NumIntermediates,
+        RegisterVT);
+  } else {
+    NumRegs =
+        TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT,
+  }
   unsigned NumElements = ValueVT.getVectorNumElements();
 
   assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
   NumParts = NumRegs; // Silence a compiler warning.
   assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
 
+  // Convert the vector to the appropiate type if necessary.
+  unsigned DestVectorNoElts =
+      NumIntermediates *
+      (IntermediateVT.isVector() ? IntermediateVT.getVectorNumElements() : 1);
+  EVT BuiltVectorTy = EVT::getVectorVT(
+      *DAG.getContext(), IntermediateVT.getScalarType(), DestVectorNoElts);
+  if (Val.getValueType() != BuiltVectorTy)
+    Val = DAG.getNode(ISD::BITCAST, DL, BuiltVectorTy, Val);
+
   // Split the vector into intermediate operands.
   SmallVector<SDValue, 8> Ops(NumIntermediates);
   for (unsigned i = 0; i != NumIntermediates; ++i) {
     if (IntermediateVT.isVector())
       Ops[i] =
           DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, IntermediateVT, Val,
                       DAG.getConstant(i * (NumElements / NumIntermediates), DL,
                                       TLI.getVectorIdxTy(DAG.getDataLayout())));
if (NumParts == NumIntermediates) {
     assert(NumParts % NumIntermediates == 0 &&
            "Must expand into a divisible number of parts!");
     unsigned Factor = NumParts / NumIntermediates;
     for (unsigned i = 0; i != NumIntermediates; ++i)
       getCopyToParts(DAG, DL, Ops[i], &Parts[i*Factor], Factor, PartVT, V);
   }
 }
 
-RegsForValue::RegsForValue() {}
+RegsForValue::RegsForValue() { IsABIMangled = false; }
 
 RegsForValue::RegsForValue(const SmallVector<unsigned, 4> &regs, MVT regvt,
-                           EVT valuevt)
-    : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {}
+                           EVT valuevt, bool IsABIMangledValue)
+    : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs),
+      RegCount(1, regs.size()), IsABIMangled(IsABIMangledValue) {}
 
 RegsForValue::RegsForValue(LLVMContext &Context, const TargetLowering &TLI,
-                           const DataLayout &DL, unsigned Reg, Type *Ty) {
+                           const DataLayout &DL, unsigned Reg, Type *Ty,
+                           bool IsABIMangledValue) {
   ComputeValueVTs(TLI, DL, Ty, ValueVTs);
 
+  IsABIMangled = IsABIMangledValue;
+
   for (EVT ValueVT : ValueVTs) {
-    unsigned NumRegs = TLI.getNumRegisters(Context, ValueVT);
-    MVT RegisterVT = TLI.getRegisterType(Context, ValueVT);
+    unsigned NumRegs = IsABIMangledValue
+                           ? TLI.getNumRegistersForCallingConv(Context, ValueVT)
+                           : TLI.getNumRegisters(Context, ValueVT);
+    MVT RegisterVT = IsABIMangledValue
+                         ? TLI.getRegisterTypeForCallingConv(Context, ValueVT)
+                         : TLI.getRegisterType(Context, ValueVT);
     for (unsigned i = 0; i != NumRegs; ++i)
       Regs.push_back(Reg + i);
     RegVTs.push_back(RegisterVT);
+    RegCount.push_back(NumRegs);
     Reg += NumRegs;
   }
 }
 
 SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
                                       FunctionLoweringInfo &FuncInfo,
                                       const SDLoc &dl, SDValue &Chain,
                                       SDValue *Flag, const Value *V) const {
   // A Value with type {} or [0 x %t] needs no registers.
   if (ValueVTs.empty())
     return SDValue();
 
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
 
   // Assemble the legal parts into the final values.
   SmallVector<SDValue, 4> Values(ValueVTs.size());
   SmallVector<SDValue, 8> Parts;
   for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
     // Copy the legal parts from the registers.
     EVT ValueVT = ValueVTs[Value];
-    unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
-    MVT RegisterVT = RegVTs[Value];
+    unsigned NumRegs = RegCount[Value];
+    MVT RegisterVT = IsABIMangled
+                         ? TLI.getRegisterTypeForCallingConv(RegVTs[Value])
+                         : RegVTs[Value];
 
     Parts.resize(NumRegs);
     for (unsigned i = 0; i != NumRegs; ++i) {
       SDValue P;
       if (!Flag) {
         P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT);
       } else {
         P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag);
void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG,
                                  ISD::NodeType PreferredExtendType) const {
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
   ISD::NodeType ExtendKind = PreferredExtendType;
 
   // Get the list of the values's legal parts.
   unsigned NumRegs = Regs.size();
   SmallVector<SDValue, 8> Parts(NumRegs);
   for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
-    EVT ValueVT = ValueVTs[Value];
-    unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
-    MVT RegisterVT = RegVTs[Value];
+    unsigned NumParts = RegCount[Value];
+
+    MVT RegisterVT = IsABIMangled
+                         ? TLI.getRegisterTypeForCallingConv(RegVTs[Value])
+                         : RegVTs[Value];
 
     if (ExtendKind == ISD::ANY_EXTEND && TLI.isZExtFree(Val, RegisterVT))
       ExtendKind = ISD::ZERO_EXTEND;
 
     getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value),
                    &Parts[Part], NumParts, RegisterVT, V, ExtendKind);
     Part += NumParts;
   }

 /// getCopyFromRegs - If there was virtual register allocated for the value V
 /// emit CopyFromReg of the specified type Ty. Return empty SDValue() otherwise.
 SDValue SelectionDAGBuilder::getCopyFromRegs(const Value *V, Type *Ty) {
   DenseMap<const Value *, unsigned>::iterator It = FuncInfo.ValueMap.find(V);
   SDValue Result;
 
   if (It != FuncInfo.ValueMap.end()) {
     unsigned InReg = It->second;
+    bool IsABIRegCopy =
+        V && ((isa<CallInst>(V) &&
+               !(static_cast<const CallInst *>(V))->isInlineAsm()) ||
+              isa<ReturnInst>(V));
+
     RegsForValue RFV(*DAG.getContext(), DAG.getTargetLoweringInfo(),
-                     DAG.getDataLayout(), InReg, Ty);
+                     DAG.getDataLayout(), InReg, Ty, IsABIRegCopy);
     SDValue Chain = DAG.getEntryNode();
-    Result = RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, nullptr, V);
+    Result = RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, nullptr,
+                                 V);
     resolveDanglingDebugInfo(V, Result);
   }
 
   return Result;
 }
 
 /// getValue - Return an SDValue for the given Value.
 SDValue SelectionDAGBuilder::getValue(const Value *V) {
if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
     if (SI != FuncInfo.StaticAllocaMap.end())
       return DAG.getFrameIndex(SI->second,
                                TLI.getPointerTy(DAG.getDataLayout()));
   }
 
   // If this is an instruction which fast-isel has deferred, select it now.
   if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
     unsigned InReg = FuncInfo.InitializeRegForValue(Inst);
+    bool IsABIRegCopy =
+        V && ((isa<CallInst>(V) &&
+               !(static_cast<const CallInst *>(V))->isInlineAsm()) ||
+              isa<ReturnInst>(V));
+
     RegsForValue RFV(*DAG.getContext(), TLI, DAG.getDataLayout(), InReg,
-                     Inst->getType());
+                     Inst->getType(), IsABIRegCopy);
     SDValue Chain = DAG.getEntryNode();
     return RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, nullptr, V);
   }
 
   llvm_unreachable("Can't get register for value!");
 }
 
 void SelectionDAGBuilder::visitCatchPad(const CatchPadInst &I) {
if (NumValues) {
           AttributeList::ReturnIndex, Attribute::InReg);
 
       for (unsigned j = 0; j != NumValues; ++j) {
         EVT VT = ValueVTs[j];
 
         if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger())
           VT = TLI.getTypeForExtReturn(Context, VT, ExtendKind);
 
-        unsigned NumParts = TLI.getNumRegisters(Context, VT);
-        MVT PartVT = TLI.getRegisterType(Context, VT);
+        unsigned NumParts = TLI.getNumRegistersForCallingConv(Context, VT);
+        MVT PartVT = TLI.getRegisterTypeForCallingConv(Context, VT);
         SmallVector<SDValue, 4> Parts(NumParts);
         getCopyToParts(DAG, getCurSDLoc(),
                        SDValue(RetOp.getNode(), RetOp.getResNo() + j),
-                       &Parts[0], NumParts, PartVT, &I, ExtendKind);
+                       &Parts[0], NumParts, PartVT, &I, ExtendKind, true);
 
         // 'inreg' on function refers to return value
         ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
         if (RetInReg)
           Flags.setInReg();
 
         // Propagate extension type if any
         if (ExtendKind == ISD::SIGN_EXTEND)
case InlineAsm::isInput: {
                                    "class is not natively supported!");
             return;
           }
 
           RegsForValue MatchedRegs(Regs, RegVT, InOperandVal.getValueType());
 
           SDLoc dl = getCurSDLoc();
           // Use the produced MatchedRegs object to
-          MatchedRegs.getCopyToRegs(InOperandVal, DAG, dl,
-                                    Chain, &Flag, CS.getInstruction());
+          MatchedRegs.getCopyToRegs(InOperandVal, DAG, dl, Chain, &Flag,
+                                    CS.getInstruction());
           MatchedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse,
                                            true, OpInfo.getMatchedOperand(), dl,
                                            DAG, AsmNodeOperands);
           break;
         }
 
         assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!");
         assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 &&
if (!CanLowerReturn) {
     CLI.RetTy = Type::getVoidTy(CLI.RetTy->getContext());
 
     // sret demotion isn't compatible with tail-calls, since the sret argument
     // points into the callers stack frame.
     CLI.IsTailCall = false;
   } else {
     for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
       EVT VT = RetTys[I];
-      MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT);
-      unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT);
+      MVT RegisterVT =
+          getRegisterTypeForCallingConv(CLI.RetTy->getContext(), VT);
+      unsigned NumRegs =
+          getNumRegistersForCallingConv(CLI.RetTy->getContext(), VT);
       for (unsigned i = 0; i != NumRegs; ++i) {
         ISD::InputArg MyFlags;
         MyFlags.VT = RegisterVT;
         MyFlags.ArgVT = VT;
         MyFlags.Used = CLI.IsReturnValueUsed;
         if (CLI.RetSExt)
           MyFlags.Flags.setSExt();
         if (CLI.RetZExt)
bool NeedsRegBlock = functionArgumentNeedsConsecutiveRegisters(
         FinalType, CLI.CallConv, CLI.IsVarArg);
     for (unsigned Value = 0, NumValues = ValueVTs.size(); Value != NumValues;
          ++Value) {
       EVT VT = ValueVTs[Value];
       Type *ArgTy = VT.getTypeForEVT(CLI.RetTy->getContext());
       SDValue Op = SDValue(Args[i].Node.getNode(),
                            Args[i].Node.getResNo() + Value);
       ISD::ArgFlagsTy Flags;
-      unsigned OriginalAlignment = DL.getABITypeAlignment(ArgTy);
+
+      // Certain targets (such as MIPS), may have a different ABI alignment
+      // for a type depending on the context. Give the target a chance to
+      // specify the alignment it wants.
+      unsigned OriginalAlignment = getABIAlignmentForCallingConv(ArgTy, DL);
 
       if (Args[i].IsZExt)
         Flags.setZExt();
       if (Args[i].IsSExt)
         Flags.setSExt();
       if (Args[i].IsInReg) {
         // If we are using vectorcall calling convention, a structure that is
         // passed InReg - is surely an HVA
for (unsigned Value = 0, NumValues = ValueVTs.size(); Value != NumValues;
         Flags.setByValAlign(FrameAlign);
       }
       if (Args[i].IsNest)
         Flags.setNest();
       if (NeedsRegBlock)
         Flags.setInConsecutiveRegs();
       Flags.setOrigAlign(OriginalAlignment);
 
-      MVT PartVT = getRegisterType(CLI.RetTy->getContext(), VT);
-      unsigned NumParts = getNumRegisters(CLI.RetTy->getContext(), VT);
+      MVT PartVT = getRegisterTypeForCallingConv(CLI.RetTy->getContext(), VT);
+      unsigned NumParts =
+          getNumRegistersForCallingConv(CLI.RetTy->getContext(), VT);
       SmallVector<SDValue, 4> Parts(NumParts);
       ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
 
       if (Args[i].IsSExt)
         ExtendKind = ISD::SIGN_EXTEND;
       else if (Args[i].IsZExt)
         ExtendKind = ISD::ZERO_EXTEND;
 
for (unsigned Value = 0, NumValues = ValueVTs.size(); Value != NumValues;
         // extension method
         if ((NumParts * PartVT.getSizeInBits() == VT.getSizeInBits()) ||
             (ExtendKind != ISD::ANY_EXTEND && CLI.RetSExt == Args[i].IsSExt &&
              CLI.RetZExt == Args[i].IsZExt))
           Flags.setReturned();
       }
 
       getCopyToParts(CLI.DAG, CLI.DL, Op, &Parts[0], NumParts, PartVT,
-                     CLI.CS ? CLI.CS->getInstruction() : nullptr, ExtendKind);
+                     CLI.CS ? CLI.CS->getInstruction() : nullptr, ExtendKind,
+                     true);
 
       for (unsigned j = 0; j != NumParts; ++j) {
         // if it isn't first piece, alignment must be 1
         ISD::OutputArg MyFlags(Flags, Parts[j].getValueType(), VT,
                                i < CLI.NumFixedArgs,
                                i, j*Parts[j].getValueType().getStoreSize());
         if (NumParts > 1 && j == 0)
           MyFlags.Flags.setSplit();
if (!CanLowerReturn) {
     Optional<ISD::NodeType> AssertOp;
     if (CLI.RetSExt)
       AssertOp = ISD::AssertSext;
     else if (CLI.RetZExt)
       AssertOp = ISD::AssertZext;
     unsigned CurReg = 0;
     for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
       EVT VT = RetTys[I];
-      MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT);
-      unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT);
+      MVT RegisterVT =
+          getRegisterTypeForCallingConv(CLI.RetTy->getContext(), VT);
+      unsigned NumRegs =
+          getNumRegistersForCallingConv(CLI.RetTy->getContext(), VT);
 
       ReturnValues.push_back(getCopyFromParts(CLI.DAG, CLI.DL, &InVals[CurReg],
                                               NumRegs, RegisterVT, VT, nullptr,
-                                              AssertOp));
+                                              AssertOp, true));
       CurReg += NumRegs;
     }
 
     // For a function returning void, there is no return value. We can't create
     // such a node, so we just return a null return value in that case. In
     // that case, nothing will actually look at the value.
     if (ReturnValues.empty())
       return std::make_pair(SDValue(), CLI.Chain);

 SelectionDAGBuilder::CopyValueToVirtualRegister(const Value *V, unsigned Reg) {
   SDValue Op = getNonRegisterValue(V);
   assert((Op.getOpcode() != ISD::CopyFromReg ||
           cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
          "Copy from a reg to the same reg!");
   assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg");
 
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+  // If this is an InlineAsm we have to match the registers required, not the
+  // notional registers required by the type.
+  bool IsABIRegCopy =
+    V && ((isa<CallInst>(V) &&
+           !(static_cast<const CallInst *>(V))->isInlineAsm()) ||
+          isa<ReturnInst>(V));
+
   RegsForValue RFV(V->getContext(), TLI, DAG.getDataLayout(), Reg,
-                   V->getType());
+                   V->getType(), IsABIRegCopy);
   SDValue Chain = DAG.getEntryNode();
 
   ISD::NodeType ExtendType = (FuncInfo.PreferredExtendType.find(V) ==
                               FuncInfo.PreferredExtendType.end())
                                  ? ISD::ANY_EXTEND
                                  : FuncInfo.PreferredExtendType[V];
   RFV.getCopyToRegs(Op, DAG, getCurSDLoc(), Chain, nullptr, V, ExtendType);
   PendingExports.push_back(Chain);
if (F.getAttributes().hasAttribute(Idx, Attribute::ByVal))
       FinalType = cast<PointerType>(FinalType)->getElementType();
     bool NeedsRegBlock = TLI->functionArgumentNeedsConsecutiveRegisters(
         FinalType, F.getCallingConv(), F.isVarArg());
     for (unsigned Value = 0, NumValues = ValueVTs.size();
          Value != NumValues; ++Value) {
       EVT VT = ValueVTs[Value];
       Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
       ISD::ArgFlagsTy Flags;
-      unsigned OriginalAlignment = DL.getABITypeAlignment(ArgTy);
+
+      // Certain targets (such as MIPS), may have a different ABI alignment
+      // for a type depending on the context. Give the target a chance to
+      // specify the alignment it wants.
+      unsigned OriginalAlignment =
+          TLI->getABIAlignmentForCallingConv(ArgTy, DL);
 
       if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt))
         Flags.setZExt();
       if (F.getAttributes().hasAttribute(Idx, Attribute::SExt))
         Flags.setSExt();
       if (F.getAttributes().hasAttribute(Idx, Attribute::InReg)) {
         // If we are using vectorcall calling convention, a structure that is
         // passed InReg - is surely an HVA
for (unsigned Value = 0, NumValues = ValueVTs.size();
       if (F.getAttributes().hasAttribute(Idx, Attribute::Nest))
         Flags.setNest();
       if (NeedsRegBlock)
         Flags.setInConsecutiveRegs();
       Flags.setOrigAlign(OriginalAlignment);
       if (ArgCopyElisionCandidates.count(&Arg))
         Flags.setCopyElisionCandidate();
 
-      MVT RegisterVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
-      unsigned NumRegs = TLI->getNumRegisters(*CurDAG->getContext(), VT);
+      MVT RegisterVT =
+          TLI->getRegisterTypeForCallingConv(*CurDAG->getContext(), VT);
+      unsigned NumRegs =
+          TLI->getNumRegistersForCallingConv(*CurDAG->getContext(), VT);
       for (unsigned i = 0; i != NumRegs; ++i) {
         ISD::InputArg MyFlags(Flags, RegisterVT, VT, isArgValueUsed,
                               Idx-1, PartBase+i*RegisterVT.getStoreSize());
         if (NumRegs > 1 && i == 0)
           MyFlags.Flags.setSplit();
         // if it isn't first piece, alignment must be 1
         else if (i > 0) {
           MyFlags.Flags.setOrigAlign(1);
if (!ArgHasUses && !isSwiftErrorArg) {
       // Also remember any frame index for use in FastISel.
       if (FrameIndexSDNode *FI =
           dyn_cast<FrameIndexSDNode>(InVals[i].getNode()))
         FuncInfo->setArgumentFrameIndex(&Arg, FI->getIndex());
     }
 
     for (unsigned Val = 0; Val != NumValues; ++Val) {
       EVT VT = ValueVTs[Val];
-      MVT PartVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
-      unsigned NumParts = TLI->getNumRegisters(*CurDAG->getContext(), VT);
+      MVT PartVT =
+          TLI->getRegisterTypeForCallingConv(*CurDAG->getContext(), VT);
+      unsigned NumParts =
+          TLI->getNumRegistersForCallingConv(*CurDAG->getContext(), VT);
 
       // Even an apparant 'unused' swifterror argument needs to be returned. So
       // we do generate a copy for it that can be used on return from the
       // function.
       if (ArgHasUses || isSwiftErrorArg) {
         Optional<ISD::NodeType> AssertOp;
         if (F.getAttributes().hasAttribute(Idx, Attribute::SExt))
           AssertOp = ISD::AssertSext;
         else if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt))
           AssertOp = ISD::AssertZext;
 
         ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i], NumParts,
-                                             PartVT, VT, nullptr, AssertOp));
+                                             PartVT, VT, nullptr, AssertOp,
+                                             true));
       }
 
       i += NumParts;
     }
 
     // We don't need to do anything else for unused arguments.
     if (ArgValues.empty())
       continue;

llvm/trunk/lib/CodeGen/SelectionDAG/StatepointLowering.cpp

if (GCResult->getParent() != ISP.getCallSite().getParent()) {
       // statepoint call has a different type than the actual call. It means
       // that by default llvm will create export register of the wrong type
       // (always i32 in our case). So instead we need to create export register
       // with correct type manually.
       // TODO: To eliminate this problem we can remove gc.result intrinsics
       //       completely and make statepoint call to return a tuple.
       unsigned Reg = FuncInfo.CreateRegs(RetTy);
       RegsForValue RFV(*DAG.getContext(), DAG.getTargetLoweringInfo(),
-                       DAG.getDataLayout(), Reg, RetTy);
+                       DAG.getDataLayout(), Reg, RetTy, true);
       SDValue Chain = DAG.getEntryNode();
 
       RFV.getCopyToRegs(ReturnValue, DAG, getCurSDLoc(), Chain, nullptr);
       PendingExports.push_back(Chain);
       FuncInfo.ValueMap[ISP.getInstruction()] = Reg;
     } else {
       // Result value will be used in a same basic block. Don't export it or
       // perform any explicit register copies.

llvm/trunk/lib/CodeGen/TargetLoweringBase.cpp

for (unsigned j = 0, f = NumValues; j != f; ++j) {
     // conventions. The frontend should mark functions whose return values
     // require promoting with signext or zeroext attributes.
     if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
       MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32);
       if (VT.bitsLT(MinVT))
         VT = MinVT;
     }
 
-    unsigned NumParts = TLI.getNumRegisters(ReturnType->getContext(), VT);
-    MVT PartVT = TLI.getRegisterType(ReturnType->getContext(), VT);
+    unsigned NumParts =
+        TLI.getNumRegistersForCallingConv(ReturnType->getContext(), VT);
+    MVT PartVT =
+        TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), VT);
 
     // 'inreg' on function refers to return value
     ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
     if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::InReg))
       Flags.setInReg();
 
     // Propagate extension type if any
     if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt))

llvm/trunk/lib/Target/Mips/MipsCCState.h

private:
   /// Identify lowered values that originated from f128 arguments and record
   /// this.
   void
   PreAnalyzeCallOperands(const SmallVectorImpl<ISD::OutputArg> &Outs,
                          std::vector<TargetLowering::ArgListEntry> &FuncArgs,
                          const SDNode *CallNode);
 
   /// Identify lowered values that originated from f128 arguments and record
-  /// this.
+  /// this for use by RetCC_MipsN.
   void
   PreAnalyzeFormalArgumentsForF128(const SmallVectorImpl<ISD::InputArg> &Ins);
 
+  void PreAnalyzeCallResultForVectorFloat(
+      const SmallVectorImpl<ISD::InputArg> &Ins,
+      const TargetLowering::CallLoweringInfo &CLI);
+
+  void PreAnalyzeFormalArgumentsForVectorFloat(
+      const SmallVectorImpl<ISD::InputArg> &Ins);
+
+  void
+  PreAnalyzeReturnForVectorFloat(const SmallVectorImpl<ISD::OutputArg> &Outs);
+
   /// Records whether the value has been lowered from an f128.
   SmallVector<bool, 4> OriginalArgWasF128;
 
   /// Records whether the value has been lowered from float.
   SmallVector<bool, 4> OriginalArgWasFloat;
 
+  /// Records whether the value has been lowered from a floating point vector.
+  SmallVector<bool, 4> OriginalArgWasFloatVector;
+
+  /// Records whether the return value has been lowered from a floating point
+  /// vector.
+  SmallVector<bool, 4> OriginalRetWasFloatVector;
+
   /// Records whether the value was a fixed argument.
   /// See ISD::OutputArg::IsFixed,
   SmallVector<bool, 4> CallOperandIsFixed;
 
   // Used to handle MIPS16-specific calling convention tweaks.
   // FIXME: This should probably be a fully fledged calling convention.
   SpecialCallingConvType SpecialCallingConv;
 
 public:
   MipsCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
               SmallVectorImpl<CCValAssign> &locs, LLVMContext &C,
               SpecialCallingConvType SpecialCC = NoSpecialCallingConv)
       : CCState(CC, isVarArg, MF, locs, C), SpecialCallingConv(SpecialCC) {}
 
   void
   AnalyzeCallOperands(const SmallVectorImpl<ISD::OutputArg> &Outs,
                       CCAssignFn Fn,
                       std::vector<TargetLowering::ArgListEntry> &FuncArgs,
                       const SDNode *CallNode) {
     PreAnalyzeCallOperands(Outs, FuncArgs, CallNode);
     CCState::AnalyzeCallOperands(Outs, Fn);
     OriginalArgWasF128.clear();
     OriginalArgWasFloat.clear();
+    OriginalArgWasFloatVector.clear();
     CallOperandIsFixed.clear();
   }
 
   // The AnalyzeCallOperands in the base class is not usable since we must
   // provide a means of accessing ArgListEntry::IsFixed. Delete them from this
   // class. This doesn't stop them being used via the base class though.
   void AnalyzeCallOperands(const SmallVectorImpl<ISD::OutputArg> &Outs,
                            CCAssignFn Fn) = delete;
   void AnalyzeCallOperands(const SmallVectorImpl<MVT> &Outs,
                            SmallVectorImpl<ISD::ArgFlagsTy> &Flags,
                            CCAssignFn Fn) = delete;
 
   void AnalyzeFormalArguments(const SmallVectorImpl<ISD::InputArg> &Ins,
                               CCAssignFn Fn) {
     PreAnalyzeFormalArgumentsForF128(Ins);
     CCState::AnalyzeFormalArguments(Ins, Fn);
     OriginalArgWasFloat.clear();
     OriginalArgWasF128.clear();
+    OriginalArgWasFloatVector.clear();
   }
 
   void AnalyzeCallResult(const SmallVectorImpl<ISD::InputArg> &Ins,
                          CCAssignFn Fn,
                          const TargetLowering::CallLoweringInfo &CLI) {
     PreAnalyzeCallResultForF128(Ins, CLI);
+    PreAnalyzeCallResultForVectorFloat(Ins, CLI);
     CCState::AnalyzeCallResult(Ins, Fn);
     OriginalArgWasFloat.clear();
     OriginalArgWasF128.clear();
+    OriginalArgWasFloatVector.clear();
   }
 
   void AnalyzeReturn(const SmallVectorImpl<ISD::OutputArg> &Outs,
                      CCAssignFn Fn) {
     PreAnalyzeReturnForF128(Outs);
+    PreAnalyzeReturnForVectorFloat(Outs);
     CCState::AnalyzeReturn(Outs, Fn);
     OriginalArgWasFloat.clear();
     OriginalArgWasF128.clear();
+    OriginalArgWasFloatVector.clear();
   }
 
   bool CheckReturn(const SmallVectorImpl<ISD::OutputArg> &ArgsFlags,
                    CCAssignFn Fn) {
     PreAnalyzeReturnForF128(ArgsFlags);
+    PreAnalyzeReturnForVectorFloat(ArgsFlags);
     bool Return = CCState::CheckReturn(ArgsFlags, Fn);
     OriginalArgWasFloat.clear();
     OriginalArgWasF128.clear();
+    OriginalArgWasFloatVector.clear();
     return Return;
   }
 
   bool WasOriginalArgF128(unsigned ValNo) { return OriginalArgWasF128[ValNo]; }
   bool WasOriginalArgFloat(unsigned ValNo) {
       return OriginalArgWasFloat[ValNo];
   }
+  bool WasOriginalArgVectorFloat(unsigned ValNo) const {
+    return OriginalArgWasFloatVector[ValNo];
+  }
+  bool WasOriginalRetVectorFloat(unsigned ValNo) const {
+    return OriginalRetWasFloatVector[ValNo];
+  }
+
   bool IsCallOperandFixed(unsigned ValNo) { return CallOperandIsFixed[ValNo]; }
   SpecialCallingConvType getSpecialCallingConv() { return SpecialCallingConv; }
 };
 }
 
 #endif

llvm/trunk/lib/Target/Mips/MipsCCState.cpp

static bool originalTypeIsF128(Type *Ty, const SDNode *CallNode) {
   const ExternalSymbolSDNode *ES =
       dyn_cast_or_null<const ExternalSymbolSDNode>(CallNode);
 
   // If the Ty is i128 and the function being called is a long double emulation
   // routine, then the original type is f128.
   return (ES && Ty->isIntegerTy(128) && isF128SoftLibCall(ES->getSymbol()));
 }
 
+/// Return true if the original type was vXfXX.
+static bool originalEVTTypeIsVectorFloat(EVT Ty) {
+  if (Ty.isVector() && Ty.getVectorElementType().isFloatingPoint())
+    return true;
+
+  return false;
+}
+
+/// Return true if the original type was vXfXX / vXfXX.
+static bool originalTypeIsVectorFloat(Type * Ty) {
+  if (Ty->isVectorTy() && Ty->isFPOrFPVectorTy())
+    return true;
+
+  return false;
+}
+
 MipsCCState::SpecialCallingConvType
 MipsCCState::getSpecialCallingConvForCallee(const SDNode *Callee,
                                             const MipsSubtarget &Subtarget) {
   MipsCCState::SpecialCallingConvType SpecialCallingConv = NoSpecialCallingConv;
   if (Subtarget.inMips16HardFloat()) {
     if (const GlobalAddressSDNode *G =
             dyn_cast<const GlobalAddressSDNode>(Callee)) {
       llvm::StringRef Sym = G->getGlobal()->getName();
void MipsCCState::PreAnalyzeCallResultForF128(
     const TargetLowering::CallLoweringInfo &CLI) {
   for (unsigned i = 0; i < Ins.size(); ++i) {
     OriginalArgWasF128.push_back(
         originalTypeIsF128(CLI.RetTy, CLI.Callee.getNode()));
     OriginalArgWasFloat.push_back(CLI.RetTy->isFloatingPointTy());
   }
 }
 
-/// Identify lowered values that originated from f128 arguments and record
-/// this for use by RetCC_MipsN.
+/// Identify lowered values that originated from f128 or float arguments and
+/// record this for use by RetCC_MipsN.
 void MipsCCState::PreAnalyzeReturnForF128(
     const SmallVectorImpl<ISD::OutputArg> &Outs) {
   const MachineFunction &MF = getMachineFunction();
   for (unsigned i = 0; i < Outs.size(); ++i) {
     OriginalArgWasF128.push_back(
         originalTypeIsF128(MF.getFunction()->getReturnType(), nullptr));
     OriginalArgWasFloat.push_back(
         MF.getFunction()->getReturnType()->isFloatingPointTy());
   }
 }
 
-/// Identify lowered values that originated from f128 arguments and record
+/// Identify lower values that originated from vXfXX and record
+/// this.
+void MipsCCState::PreAnalyzeCallResultForVectorFloat(
+    const SmallVectorImpl<ISD::InputArg> &Ins,
+    const TargetLowering::CallLoweringInfo &CLI) {
+  for (unsigned i = 0; i < Ins.size(); ++i) {
+    OriginalRetWasFloatVector.push_back(
+        originalTypeIsVectorFloat(CLI.RetTy));
+  }
+}
+
+/// Identify lowered values that originated from vXfXX arguments and record
 /// this.
+void MipsCCState::PreAnalyzeReturnForVectorFloat(
+    const SmallVectorImpl<ISD::OutputArg> &Outs) {
+  for (unsigned i = 0; i < Outs.size(); ++i) {
+    ISD::OutputArg Out = Outs[i];
+    OriginalRetWasFloatVector.push_back(
+        originalEVTTypeIsVectorFloat(Out.ArgVT));
+  }
+}
+/// Identify lowered values that originated from f128, float and sret to vXfXX
+/// arguments and record this.
 void MipsCCState::PreAnalyzeCallOperands(
     const SmallVectorImpl<ISD::OutputArg> &Outs,
     std::vector<TargetLowering::ArgListEntry> &FuncArgs,
     const SDNode *CallNode) {
   for (unsigned i = 0; i < Outs.size(); ++i) {
-    OriginalArgWasF128.push_back(
-        originalTypeIsF128(FuncArgs[Outs[i].OrigArgIndex].Ty, CallNode));
-    OriginalArgWasFloat.push_back(
-        FuncArgs[Outs[i].OrigArgIndex].Ty->isFloatingPointTy());
+    TargetLowering::ArgListEntry FuncArg = FuncArgs[Outs[i].OrigArgIndex];
+
+    OriginalArgWasF128.push_back(originalTypeIsF128(FuncArg.Ty, CallNode));
+    OriginalArgWasFloat.push_back(FuncArg.Ty->isFloatingPointTy());
+
+    OriginalArgWasFloatVector.push_back(FuncArg.Ty->isVectorTy());
     CallOperandIsFixed.push_back(Outs[i].IsFixed);
   }
 }
 
-/// Identify lowered values that originated from f128 arguments and record
-/// this.
+/// Identify lowered values that originated from f128, float and vXfXX arguments
+/// and record this.
 void MipsCCState::PreAnalyzeFormalArgumentsForF128(
     const SmallVectorImpl<ISD::InputArg> &Ins) {
   const MachineFunction &MF = getMachineFunction();
+
   for (unsigned i = 0; i < Ins.size(); ++i) {
     Function::const_arg_iterator FuncArg = MF.getFunction()->arg_begin();
 
     // SRet arguments cannot originate from f128 or {f128} returns so we just
     // push false. We have to handle this specially since SRet arguments
     // aren't mapped to an original argument.
     if (Ins[i].Flags.isSRet()) {
       OriginalArgWasF128.push_back(false);
       OriginalArgWasFloat.push_back(false);
+      OriginalArgWasFloatVector.push_back(false);
       continue;
     }
 
     assert(Ins[i].getOrigArgIndex() < MF.getFunction()->arg_size());
     std::advance(FuncArg, Ins[i].getOrigArgIndex());
 
     OriginalArgWasF128.push_back(
         originalTypeIsF128(FuncArg->getType(), nullptr));
     OriginalArgWasFloat.push_back(FuncArg->getType()->isFloatingPointTy());
+
+    // The MIPS vector ABI exhibits a corner case of sorts or quirk; if the
+    // first argument is actually an SRet pointer to a vector, then the next
+    // argument slot is $a2.
+    OriginalArgWasFloatVector.push_back(FuncArg->getType()->isVectorTy());
   }
 }

llvm/trunk/lib/Target/Mips/MipsCallingConv.td

class CCIfOrigArgWasF128<CCAction A>
     : CCIf<"static_cast<MipsCCState *>(&State)->WasOriginalArgF128(ValNo)", A>;
 
 /// Match if this specific argument is a vararg.
 /// This is slightly different fro CCIfIsVarArg which matches if any argument is
 /// a vararg.
 class CCIfArgIsVarArg<CCAction A>
     : CCIf<"!static_cast<MipsCCState *>(&State)->IsCallOperandFixed(ValNo)", A>;
 
+/// Match if the return was a floating point vector.
+class CCIfOrigArgWasNotVectorFloat<CCAction A>
+    : CCIf<"!static_cast<MipsCCState *>(&State)"
+                "->WasOriginalRetVectorFloat(ValNo)", A>;
 
 /// Match if the special calling conv is the specified value.
 class CCIfSpecialCallingConv<string CC, CCAction A>
     : CCIf<"static_cast<MipsCCState *>(&State)->getSpecialCallingConv() == "
                "MipsCCState::" # CC, A>;
 
 // For soft-float, f128 values are returned in A0_64 rather than V1_64.
 def RetCC_F128SoftFloat : CallingConv<[

 ]>;
 
 // Only the return rules are defined here for O32. The rules for argument
 // passing are defined in MipsISelLowering.cpp.
 def RetCC_MipsO32 : CallingConv<[
   // Promote i1/i8/i16 return values to i32.
   CCIfType<[i1, i8, i16], CCPromoteToType<i32>>,
 
-  // i32 are returned in registers V0, V1, A0, A1
-  CCIfType<[i32], CCAssignToReg<[V0, V1, A0, A1]>>,
+  // i32 are returned in registers V0, V1, A0, A1, unless the original return
+  // type was a vector of floats.
+  CCIfOrigArgWasNotVectorFloat<CCIfType<[i32],
+                                        CCAssignToReg<[V0, V1, A0, A1]>>>,
 
   // f32 are returned in registers F0, F2
   CCIfType<[f32], CCAssignToReg<[F0, F2]>>,
 
   // f64 arguments are returned in D0_64 and D2_64 in FP64bit mode or
   // in D0 and D1 in FP32bit mode.
   CCIfType<[f64], CCIfSubtarget<"isFP64bit()", CCAssignToReg<[D0_64, D2_64]>>>,
   CCIfType<[f64], CCIfSubtargetNot<"isFP64bit()", CCAssignToReg<[D0, D1]>>>

llvm/trunk/lib/Target/Mips/MipsISelLowering.h

public:
 
     MVT getScalarShiftAmountTy(const DataLayout &, EVT) const override {
       return MVT::i32;
     }
 
     bool isCheapToSpeculateCttz() const override;
     bool isCheapToSpeculateCtlz() const override;
 
+    /// Return the register type for a given MVT, ensuring vectors are treated
+    /// as a series of gpr sized integers.
+    virtual MVT getRegisterTypeForCallingConv(MVT VT) const override;
+
+    /// Return the register type for a given MVT, ensuring vectors are treated
+    /// as a series of gpr sized integers.
+    virtual MVT getRegisterTypeForCallingConv(LLVMContext &Context,
+                                              EVT VT) const override;
+
+    /// Return the number of registers for a given MVT, ensuring vectors are
+    /// treated as a series of gpr sized integers.
+    virtual unsigned getNumRegistersForCallingConv(LLVMContext &Context,
+                                                   EVT VT) const override;
+
+    /// Break down vectors to the correct number of gpr sized integers.
+    virtual unsigned getVectorTypeBreakdownForCallingConv(
+        LLVMContext &Context, EVT VT, EVT &IntermediateVT,
+        unsigned &NumIntermediates, MVT &RegisterVT) const override;
+
+    /// Return the correct alignment for the current calling convention.
+    virtual unsigned
+    getABIAlignmentForCallingConv(Type *ArgTy, DataLayout DL) const override {
+      if (ArgTy->isVectorTy())
+        return std::min(DL.getABITypeAlignment(ArgTy), 8U);
+      return DL.getABITypeAlignment(ArgTy);
+    }
+
     ISD::NodeType getExtendForAtomicOps() const override {
       return ISD::SIGN_EXTEND;
     }
 
     void LowerOperationWrapper(SDNode *N,
                                SmallVectorImpl<SDValue> &Results,
                                SelectionDAG &DAG) const override;
 

llvm/trunk/lib/Target/Mips/MipsISelLowering.cpp

static bool isShiftedMask(uint64_t I, uint64_t &Pos, uint64_t &Size) {
   if (!isShiftedMask_64(I))
     return false;
 
   Size = countPopulation(I);
   Pos = countTrailingZeros(I);
   return true;
 }
 
+// The MIPS MSA ABI passes vector arguments in the integer register set.
+// The number of integer registers used is dependant on the ABI used.
+MVT MipsTargetLowering::getRegisterTypeForCallingConv(MVT VT) const {
+  if (VT.isVector() && Subtarget.hasMSA())
+    return Subtarget.isABI_O32() ? MVT::i32 : MVT::i64;
+  return MipsTargetLowering::getRegisterType(VT);
+}
+
+MVT MipsTargetLowering::getRegisterTypeForCallingConv(LLVMContext &Context,
+                                                      EVT VT) const {
+  if (VT.isVector()) {
+      if (Subtarget.isABI_O32()) {
+        return MVT::i32;
+      } else {
+        return (VT.getSizeInBits() == 32) ? MVT::i32 : MVT::i64;
+      }
+  }
+  return MipsTargetLowering::getRegisterType(Context, VT);
+}
+
+unsigned MipsTargetLowering::getNumRegistersForCallingConv(LLVMContext &Context,
+                                                           EVT VT) const {
+  if (VT.isVector())
+    return std::max((VT.getSizeInBits() / (Subtarget.isABI_O32() ? 32 : 64)),
+                    1U);
+  return MipsTargetLowering::getNumRegisters(Context, VT);
+}
+
+unsigned MipsTargetLowering::getVectorTypeBreakdownForCallingConv(
+    LLVMContext &Context, EVT VT, EVT &IntermediateVT,
+    unsigned &NumIntermediates, MVT &RegisterVT) const {
+
+  // Break down vector types to either 2 i64s or 4 i32s.
+  RegisterVT = getRegisterTypeForCallingConv(Context, VT) ;
+  IntermediateVT = RegisterVT;
+  NumIntermediates = VT.getSizeInBits() < RegisterVT.getSizeInBits()
+                         ? VT.getVectorNumElements()
+                         : VT.getSizeInBits() / RegisterVT.getSizeInBits();
+
+  return NumIntermediates;
+}
+
 SDValue MipsTargetLowering::getGlobalReg(SelectionDAG &DAG, EVT Ty) const {
   MipsFunctionInfo *FI = DAG.getMachineFunction().getInfo<MipsFunctionInfo>();
   return DAG.getRegister(FI->getGlobalBaseReg(), Ty);
 }
 
 SDValue MipsTargetLowering::getTargetNode(GlobalAddressSDNode *N, EVT Ty,
                                           SelectionDAG &DAG,
                                           unsigned Flag) const {

 // ---
 // i32 - Passed in A0, A1, A2, A3 and stack
 // f32 - Only passed in f32 registers if no int reg has been used yet to hold
 //       an argument. Otherwise, passed in A1, A2, A3 and stack.
 // f64 - Only passed in two aliased f32 registers if no int reg has been used
 //       yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is
 //       not used, it must be shadowed. If only A3 is available, shadow it and
 //       go to stack.
+// vXiX - Received as scalarized i32s, passed in A0 - A3 and the stack.
+// vXf32 - Passed in either a pair of registers {A0, A1}, {A2, A3} or {A0 - A3}
+//         with the remainder spilled to the stack.
+// vXf64 - Passed in either {A0, A1, A2, A3} or {A2, A3} and in both cases
+//         spilling the remainder to the stack.
 //
 //  For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack.
 //===----------------------------------------------------------------------===//
 
 static bool CC_MipsO32(unsigned ValNo, MVT ValVT, MVT LocVT,
                        CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags,
                        CCState &State, ArrayRef<MCPhysReg> F64Regs) {
   const MipsSubtarget &Subtarget = static_cast<const MipsSubtarget &>(
       State.getMachineFunction().getSubtarget());
 
   static const MCPhysReg IntRegs[] = { Mips::A0, Mips::A1, Mips::A2, Mips::A3 };
+
+  const MipsCCState * MipsState = static_cast<MipsCCState *>(&State);
+
   static const MCPhysReg F32Regs[] = { Mips::F12, Mips::F14 };
 
+  static const MCPhysReg FloatVectorIntRegs[] = { Mips::A0, Mips::A2 };
+
   // Do not process byval args here.
   if (ArgFlags.isByVal())
     return true;
 
   // Promote i8 and i16
   if (ArgFlags.isInReg() && !Subtarget.isLittle()) {
     if (LocVT == MVT::i8 || LocVT == MVT::i16 || LocVT == MVT::i32) {
       LocVT = MVT::i32;
static bool CC_MipsO32(unsigned ValNo, MVT ValVT, MVT LocVT,
 
   // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following
   // is true: function is vararg, argument is 3rd or higher, there is previous
   // argument which is not f32 or f64.
   bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1 ||
                                 State.getFirstUnallocated(F32Regs) != ValNo;
   unsigned OrigAlign = ArgFlags.getOrigAlign();
   bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8);
+  bool isVectorFloat = MipsState->WasOriginalArgVectorFloat(ValNo);
 
-  if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) {
+  // The MIPS vector ABI for floats passes them in a pair of registers
+  if (ValVT == MVT::i32 && isVectorFloat) {
+    // This is the start of an vector that was scalarized into an unknown number
+    // of components. It doesn't matter how many there are. Allocate one of the
+    // notional 8 byte aligned registers which map onto the argument stack, and
+    // shadow the register lost to alignment requirements.
+    if (ArgFlags.isSplit()) {
+      Reg = State.AllocateReg(FloatVectorIntRegs);
+      if (Reg == Mips::A2)
+        State.AllocateReg(Mips::A1);
+      else if (Reg == 0)
+        State.AllocateReg(Mips::A3);
+    } else {
+      // If we're an intermediate component of the split, we can just attempt to
+      // allocate a register directly.
+      Reg = State.AllocateReg(IntRegs);
+    }
+  } else if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) {
     Reg = State.AllocateReg(IntRegs);
     // If this is the first part of an i64 arg,
     // the allocated register must be either A0 or A2.
     if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3))
       Reg = State.AllocateReg(IntRegs);
     LocVT = MVT::i32;
   } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) {
     // Allocate int register and shadow next int register. If first

llvm/trunk/lib/Target/Mips/MipsRegisterInfo.cpp

eliminateFrameIndex(MachineBasicBlock::iterator II, int SPAdj,
   MachineFunction &MF = *MI.getParent()->getParent();
 
   DEBUG(errs() << "\nFunction : " << MF.getName() << "\n";
         errs() << "<--------->\n" << MI);
 
   int FrameIndex = MI.getOperand(FIOperandNum).getIndex();
   uint64_t stackSize = MF.getFrameInfo().getStackSize();
   int64_t spOffset = MF.getFrameInfo().getObjectOffset(FrameIndex);
+  unsigned alignment = MF.getFrameInfo().getObjectAlignment(FrameIndex);
 
   DEBUG(errs() << "FrameIndex : " << FrameIndex << "\n"
                << "spOffset   : " << spOffset << "\n"
-               << "stackSize  : " << stackSize << "\n");
+               << "stackSize  : " << stackSize << "\n"
+               << "alignment  : " << alignment << "\n");
 
   eliminateFI(MI, FIOperandNum, FrameIndex, stackSize, spOffset);
 }
 
 unsigned MipsRegisterInfo::
 getFrameRegister(const MachineFunction &MF) const {
   const MipsSubtarget &Subtarget = MF.getSubtarget<MipsSubtarget>();
   const TargetFrameLowering *TFI = Subtarget.getFrameLowering();

llvm/trunk/test/CodeGen/Mips/cconv/vector.ll

+; RUN: llc < %s -march=mips -mcpu=mips32 -disable-mips-delay-filler | FileCheck %s --check-prefixes=ALL,MIPS32,MIPS32EB
+; RUN: llc < %s -march=mips64 -relocation-model=pic -mcpu=mips64 -disable-mips-delay-filler | FileCheck %s --check-prefixes=ALL,MIPS64,MIPS64EB
+; RUN: llc < %s -march=mips -mcpu=mips32r5 -mattr=+fp64,+msa -disable-mips-delay-filler | FileCheck %s --check-prefixes=ALL,MIPS32R5,MIPS32R5EB
+; RUN: llc < %s -march=mips64 -relocation-model=pic -mcpu=mips64r5 -mattr=+fp64,+msa -disable-mips-delay-filler | FileCheck %s --check-prefixes=ALL,MIPS64R5
+; RUN: llc < %s -march=mipsel -mcpu=mips32 -disable-mips-delay-filler | FileCheck %s --check-prefixes=ALL,MIPS32,MIPS32EL
+; RUN: llc < %s -march=mips64el -relocation-model=pic -mcpu=mips64 -disable-mips-delay-filler | FileCheck %s --check-prefixes=ALL,MIPS64,MIPS64EL
+; RUN: llc < %s -march=mipsel -mcpu=mips32r5 -mattr=+fp64,+msa -disable-mips-delay-filler | FileCheck %s --check-prefixes=ALL,MIPS32R5,MIPS32R5EL
+; RUN: llc < %s -march=mips64el -relocation-model=pic -mcpu=mips64r5 -mattr=+fp64,+msa -disable-mips-delay-filler | FileCheck %s --check-prefixes=ALL,MIPS64R5
+
+
+
+; Test that vector types are passed through the integer register set whether or
+; not MSA is enabled. This is a ABI requirement for MIPS. For GCC compatibility
+; we need to handle any power of 2 number of elements. We will test this
+; exhaustively for combinations up to MSA register (128 bits) size.
+
+; First set of tests are for argument passing.
+
+define <2 x i8> @i8_2(<2 x i8> %a, <2 x i8> %b) {
+; ALL-LABEL: i8_2:
+; MIPS32EB-DAG: srl ${{[0-9]+}}, $5, 24
+; MIPS32EB-DAG: srl ${{[0-9]+}}, $4, 24
+; MIPS32EB-DAG: srl ${{[0-9]+}}, $5, 16
+; MIPS32EB-DAG: srl ${{[0-9]+}}, $4, 16
+
+; MIPS32EL: addu $1, $4, $5
+
+; MIPS32R5-DAG: sw $4
+; MIPS32R5-DAG: sw $5
+
+; MIPS64EB-DAG: dsrl ${{[0-9]+}}, $5, 56
+; MIPS64EB-DAG: dsrl ${{[0-9]+}}, $4, 56
+; MIPS64EB-DAG: dsrl ${{[0-9]+}}, $5, 48
+; MIPS64EB-DAG: dsrl ${{[0-9]+}}, $4, 48
+
+; MIPS64EL-DAG: sll ${{[0-9]+}}, $4, 0
+; MIPS64EL-DAG: sll ${{[0-9]+}}, $5, 0
+
+; MIPS64R5-DAG: sd $4
+; MIPS64R5-DAG: sd $5
+
+  %1 = add <2 x i8> %a, %b
+  ret <2 x i8> %1
+}
+
+; Test that vector spilled to the outgoing argument area have the expected
+; offset from $sp.
+
+define <2 x i8> @i8x2_7(<2 x i8> %a, <2 x i8> %b, <2 x i8> %c, <2 x i8> %d,
+                        <2 x i8> %e, <2 x i8> %f, <2 x i8> %g) {
+entry:
+
+; MIPS32EB-DAG: srl ${{[0-9]+}}, $4, 24
+; MIPS32EB-DAG: srl ${{[0-9]+}}, $5, 24
+; MIPS32EB-DAG: srl ${{[0-9]+}}, $6, 24
+; MIPS32EB-DAG: srl ${{[0-9]+}}, $7, 24
+
+; MIPS32EL-DAG: andi ${{[0-9]+}}, $4, 65280
+; MIPS32EL-DAG: andi ${{[0-9]+}}, $5, 65280
+; MIPS32EL-DAG: andi ${{[0-9]+}}, $6, 65280
+; MIPS32EL-DAG: andi ${{[0-9]+}}, $7, 65280
+
+; MIPS32-DAG: lbu ${{[0-9]+}}, 16($sp)
+; MIPS32-DAG; lbu ${{[0-9]+}}, 17($sp)
+; MIPS32-DAG: lbu ${{[0-9]+}}, 20($sp)
+; MIPS32-DAG: lbu ${{[0-9]+}}, 21($sp)
+; MIPS32-DAG: lbu ${{[0-9]+}}, 24($sp)
+; MIPS32-DAG: lbu ${{[0-9]+}}, 25($sp)
+
+; MIPS32R5-DAG: sw $4, {{[0-9]+}}($sp)
+; MIPS32R5-DAG: sw $5, {{[0-9]+}}($sp)
+; MIPS32R5-DAG: sw $6, {{[0-9]+}}($sp)
+; MIPS32R5-DAG: sw $7, {{[0-9]+}}($sp)
+
+; MIPS32R5-DAG: lbu ${{[0-9]+}}, 40($sp)
+; MIPS32R5-DAG: lbu ${{[0-9]+}}, 41($sp)
+; MIPS32R5-DAG: lbu ${{[0-9]+}}, 42($sp)
+; MIPS32R5-DAG: lbu ${{[0-9]+}}, 43($sp)
+; MIPS32R5-DAG: lbu ${{[0-9]+}}, 44($sp)
+; MIPS32R5-DAG: lbu ${{[0-9]+}}, 45($sp)
+; MIPS32R5-DAG: lbu ${{[0-9]+}}, 46($sp)
+; MIPS32R5-DAG: lbu ${{[0-9]+}}, 47($sp)
+; MIPS32R5-DAG: lbu ${{[0-9]+}}, 48($sp)
+; MIPS32R5-DAG: lbu ${{[0-9]+}}, 49($sp)
+; MIPS32R5-DAG: lbu ${{[0-9]+}}, 50($sp)
+; MIPS32R5-DAG: lbu ${{[0-9]+}}, 51($sp)
+
+; MIPS64EB-DAG: dsrl ${{[0-9]+}}, $4, 48
+; MIPS64EB-DAG: dsrl ${{[0-9]+}}, $5, 48
+; MIPS64EB-DAG: dsrl ${{[0-9]+}}, $6, 48
+; MIPS64EB-DAG: dsrl ${{[0-9]+}}, $7, 48
+; MIPS64EB-DAG: dsrl ${{[0-9]+}}, $8, 48
+; MIPS64EB-DAG: dsrl ${{[0-9]+}}, $9, 48
+; MIPS64EB-DAG: dsrl ${{[0-9]+}}, $10, 48
+
+; MIPS64R5-DAG: sd $4, {{[0-9]+}}($sp)
+; MIPS64R5-DAG: sd $5, {{[0-9]+}}($sp)
+; MIPS64R5-DAG: sd $6, {{[0-9]+}}($sp)
+; MIPS64R5-DAG: sd $7, {{[0-9]+}}($sp)
+; MIPS64R5-DAG: sd $8, {{[0-9]+}}($sp)
+; MIPS64R5-DAG: sd $9, {{[0-9]+}}($sp)
+; MIPS64R5-DAG: sd $10, {{[0-9]+}}($sp)
+
+  %0 = add <2 x i8> %a, %b
+  %1 = add <2 x i8> %0, %c
+  %2 = add <2 x i8> %1, %d
+  %3 = add <2 x i8> %2, %e
+  %4 = add <2 x i8> %3, %f
+  %5 = add <2 x i8> %4, %g
+  ret <2 x i8> %5
+}
+
+define <4 x i8> @i8_4(<4 x i8> %a, <4 x i8> %b) {
+; ALL-LABEL: i8_4:
+; MIPS32-DAG: srl ${{[0-9]+}}, $5, 24
+; MIPS32-DAG: srl ${{[0-9]+}}, $4, 24
+; MIPS32-DAG: srl ${{[0-9]+}}, $5, 16
+; MIPS32-DAG: srl ${{[0-9]+}}, $4, 16
+; MIPS32-DAG: srl ${{[0-9]+}}, $5, 8
+; MIPS32-DAG: srl ${{[0-9]+}}, $4, 8
+
+; MIPS32R5-DAG: sw $4
+; MIPS32R5-DAG: sw $5
+
+; MIPS64-DAG: sll ${{[0-9]+}}, $4, 0
+; MIPS64-DAG: sll ${{[0-9]+}}, $5, 0
+
+; MIPS64R5-DAG: sll ${{[0-9]+}}, $4, 0
+; MIPS64R5-DAG: sll ${{[0-9]+}}, $5, 0
+
+  %1 = add <4 x i8> %a, %b
+  ret <4 x i8> %1
+}
+
+define <8 x i8> @i8_8(<8 x i8> %a, <8 x i8> %b) {
+; ALL-LABEL: i8_8:
+; MIPS32-NOT: lw
+; MIPS32-DAG: srl ${{[0-9]+}}, $7, 24
+; MIPS32-DAG: srl ${{[0-9]+}}, $6, 24
+; MIPS32-DAG: srl ${{[0-9]+}}, $7, 16
+; MIPS32-DAG: srl ${{[0-9]+}}, $6, 16
+; MIPS32-DAG: srl ${{[0-9]+}}, $7, 8
+; MIPS32-DAG: srl ${{[0-9]+}}, $6, 8
+; MIPS32-DAG: srl ${{[0-9]+}}, $5, 24
+; MIPS32-DAG: srl ${{[0-9]+}}, $4, 24
+; MIPS32-DAG: srl ${{[0-9]+}}, $5, 16
+; MIPS32-DAG: srl ${{[0-9]+}}, $4, 16
+; MIPS32-DAG: srl ${{[0-9]+}}, $5, 8
+; MIPS32-DAG: srl ${{[0-9]+}}, $4, 8
+
+; MIPS32R5-DAG: sw $4
+; MIPS32R5-DAG: sw $5
+; MIPS32R5-DAG: sw $6
+; MIPS32R5-DAG: sw $7
+
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $5, 56
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $4, 56
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $5, 48
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $4, 48
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $5, 40
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $4, 40
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $5, 32
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $4, 32
+; MIPS64-DAG: sll $[[R0:[0-9]+]], $4, 0
+; MIPS64-DAG: sll $[[R1:[0-9]+]], $5, 0
+; MIPS64-DAG: srl ${{[0-9]+}}, $[[R1]], 24
+; MIPS64-DAG: srl ${{[0-9]+}}, $[[R0]], 24
+; MIPS64-DAG: srl ${{[0-9]+}}, $[[R1]], 16
+; MIPS64-DAG: srl ${{[0-9]+}}, $[[R0]], 16
+; MIPS64-DAG: srl ${{[0-9]+}}, $[[R1]], 8
+; MIPS64-DAG: srl ${{[0-9]+}}, $[[R0]], 8
+
+; MIPS64R5-DAG: sd $4
+; MIPS64R5-DAG: sd $5
+
+  %1 = add <8 x i8> %a, %b
+  ret <8 x i8> %1
+}
+
+define <16 x i8> @i8_16(<16 x i8> %a, <16 x i8> %b) {
+; ALL-LABEL: i8_16:
+; MIPS32-DAG: lw ${{[0-9]+}}, 16($sp)
+; MIPS32-DAG: lw ${{[0-9]+}}, 20($sp)
+; MIPS32-DAG: lw ${{[0-9]+}}, 24($sp)
+; MIPS32-DAG: lw ${{[0-9]+}}, 28($sp)
+; MIPS32-DAG: srl ${{[0-9]+}}, $7, 24
+; MIPS32-DAG: srl ${{[0-9]+}}, $6, 24
+; MIPS32-DAG: srl ${{[0-9]+}}, $7, 16
+; MIPS32-DAG: srl ${{[0-9]+}}, $6, 16
+; MIPS32-DAG: srl ${{[0-9]+}}, $7, 8
+; MIPS32-DAG: srl ${{[0-9]+}}, $6, 8
+; MIPS32-DAG: srl ${{[0-9]+}}, $5, 24
+; MIPS32-DAG: srl ${{[0-9]+}}, $4, 24
+; MIPS32-DAG: srl ${{[0-9]+}}, $5, 16
+; MIPS32-DAG: srl ${{[0-9]+}}, $4, 16
+; MIPS32-DAG: srl ${{[0-9]+}}, $5, 8
+; MIPS32-DAG: srl ${{[0-9]+}}, $4, 8
+
+; MIPS32R5-DAG: lw ${{[0-9]+}}, 16($sp)
+; MIPS32R5-DAG: lw ${{[0-9]+}}, 20($sp)
+; MIPS32R5-DAG: lw ${{[0-9]+}}, 24($sp)
+; MIPS32R5-DAG: lw ${{[0-9]+}}, 28($sp)
+; MIPS32R5-DAG: insert.w $w[[W0:[0-9]+]][0], $4
+; MIPS32R5-DAG: insert.w $w[[W0]][1], $5
+; MIPS32R5-DAG: insert.w $w[[W0]][2], $6
+; MIPS32R5-DAG: insert.w $w[[W0]][3], $7
+
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $7, 56
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $6, 56
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $7, 48
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $6, 48
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $7, 40
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $6, 40
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $7, 32
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $6, 32
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $5, 56
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $4, 56
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $5, 48
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $4, 48
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $5, 32
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $4, 32
+
+; MIPS64R5-DAG: insert.d $w[[W0:[0-9]+]][{{[0-9]}}], $4
+; MIPS64R5-DAG: insert.d $w[[W0]][{{[0-9]}}], $5
+; MIPS64R5-DAG: insert.d $w[[W1:[0-9]+]][{{[0-9]}}], $6
+; MIPS64R5-DAG: insert.d $w[[W1]][{{[0-9]}}], $7
+
+  %1 = add <16 x i8> %a, %b
+
+  ret <16 x i8> %1
+}
+
+define <2 x i16> @i16_2(<2 x i16> %a, <2 x i16> %b) {
+; ALL-LABEL: i16_2:
+; MIPS32: addu    $[[R0:[0-9]+]], $4, $5
+; MIPS32: andi    $[[R1:[0-9]+]], $[[R0]], 65535
+; MIPS32: srl     $[[R2:[0-9]+]], $5, 16
+; MIPS32: srl     $[[R3:[0-9]+]], $4, 16
+; MIPS32: addu    $[[R4:[0-9]+]], $[[R3]], $[[R2]]
+; MIPS32: sll     $2, $[[R4]], 16
+
+; MIPS32R5-DAG: sw $4
+; MIPS32R5-DAG: sw $5
+
+; MIPS64-DAG: sll ${{[0-9]+}}, $5, 0
+; MIPS64-DAG: sll ${{[0-9]+}}, $4, 0
+
+; MIPS64R5-DAG: sll ${{[0-9]+}}, $4, 0
+; MIPS64R5-DAG: sll ${{[0-9]+}}, $5, 0
+
+  %1 = add <2 x i16> %a, %b
+  ret <2 x i16> %1
+}
+
+define <4 x i16> @i16_4(<4 x i16> %a, <4 x i16> %b) {
+; ALL-LABEL: i16_4:
+; MIPS32-DAG: srl ${{[0-9]+}}, $7, 16
+; MIPS32-DAG: srl ${{[0-9]+}}, $6, 16
+; MIPS32-DAG: srl ${{[0-9]+}}, $5, 16
+; MIPS32-DAG: srl ${{[0-9]+}}, $4, 16
+
+; MIPS32R5-DAG: sw $4
+; MIPS32R5-DAG: sw $5
+; MIPS32R5-DAG: sw $6
+; MIPS32R5-DAG: sw $7
+
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $4, 48
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $5, 48
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $4, 32
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $5, 32
+
+; MIPS64R5-DAG: sd $4
+; MIPS64R5-DAG: sd $5
+
+  %1 = add <4 x i16> %a, %b
+  ret <4 x i16> %1
+}
+
+define <8 x i16> @i16_8(<8 x i16> %a, <8 x i16> %b) {
+; ALL-LABEL: i16_8:
+; MIPS32-DAG: lw ${{[0-9]+}}, 16($sp)
+; MIPS32-DAG: lw ${{[0-9]+}}, 20($sp)
+; MIPS32-DAG: lw ${{[0-9]+}}, 24($sp)
+; MIPS32-DAG: lw ${{[0-9]+}}, 28($sp)
+; MIPS32-DAG: srl ${{[0-9]+}}, $7, 16
+; MIPS32-DAG: srl ${{[0-9]+}}, $6, 16
+; MIPS32-DAG: srl ${{[0-9]+}}, $5, 16
+; MIPS32-DAG: srl ${{[0-9]+}}, $4, 16
+
+; MIPS32R5-DAG: lw ${{[0-9]+}}, 16($sp)
+; MIPS32R5-DAG: lw ${{[0-9]+}}, 20($sp)
+; MIPS32R5-DAG: lw ${{[0-9]+}}, 24($sp)
+; MIPS32R5-DAG: lw ${{[0-9]+}}, 28($sp)
+; MIPS32R5-DAG: insert.w $w[[W0:[0-9]+]][0], $4
+; MIPS32R5-DAG: insert.w $w[[W0]][1], $5
+; MIPS32R5-DAG: insert.w $w[[W0]][2], $6
+; MIPS32R5-DAG: insert.w $w[[W0]][3], $7
+
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $6, 48
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $7, 48
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $6, 32
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $7, 32
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $4, 48
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $5, 48
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $4, 32
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $5, 32
+
+; MIPS64R5-DAG: insert.d $w[[W0:[0-9]+]][{{[0-9]}}], $4
+; MIPS64R5-DAG: insert.d $w[[W0]][{{[0-9]}}], $5
+; MIPS64R5-DAG: insert.d $w[[W1:[0-9]+]][{{[0-9]}}], $6
+; MIPS64R5-DAG: insert.d $w[[W1]][{{[0-9]}}], $7
+
+  %1 = add <8 x i16> %a, %b
+  ret <8 x i16> %1
+}
+
+define <2 x i32> @i32_2(<2 x i32> %a, <2 x i32> %b) {
+; ALL-LABEL: i32_2:
+; MIPS32-DAG: addu    $2, $4, $6
+; MIPS32-DAG: addu    $3, $5, $7
+
+; MIPS32R5-DAG: sw $4
+; MIPS32R5-DAG: sw $5
+; MIPS32R5-DAG: sw $6
+; MIPS32R5-DAG: sw $7
+
+; MIPS64-DAG: sll     ${{[0-9]+}}, $4, 0
+; MIPS64-DAG: sll     ${{[0-9]+}}, $5, 0
+
+; MIPS64R5-DAG: sd $4
+; MIPS64R5-DAG: sd $5
+
+  %1 = add <2 x i32> %a, %b
+
+  ret <2 x i32> %1
+}
+
+define <4 x i32> @i32_4(<4 x i32> %a, <4 x i32> %b) {
+; ALL-LABEL: i32_4:
+; MIPS32-DAG: lw ${{[0-9]+}}, 16($sp)
+; MIPS32-DAG: lw ${{[0-9]+}}, 20($sp)
+; MIPS32-DAG: lw ${{[0-9]+}}, 24($sp)
+; MIPS32-DAG: lw ${{[0-9]+}}, 28($sp)
+; MIPS32-DAG: addu $2
+; MIPS32-DAG: addu $3
+; MIPS32-DAG: addu $4
+; MIPS32-DAG: addu $5
+
+; MIPS32R5-DAG: lw ${{[0-9]+}}, 16($sp)
+; MIPS32R5-DAG: lw ${{[0-9]+}}, 20($sp)
+; MIPS32R5-DAG: lw ${{[0-9]+}}, 24($sp)
+; MIPS32R5-DAG: lw ${{[0-9]+}}, 28($sp)
+; MIPS32R5-DAG: insert.w $w[[W0:[0-9]+]][0], $4
+; MIPS32R5-DAG: insert.w $w[[W0]][1], $5
+; MIPS32R5-DAG: insert.w $w[[W0]][2], $6
+; MIPS32R5-DAG: insert.w $w[[W0]][3], $7
+
+; MIPS64-DAG: sll     ${{[0-9]+}}, $4, 0
+; MIPS64-DAG: sll     ${{[0-9]+}}, $5, 0
+; MIPS64-DAG: sll     ${{[0-9]+}}, $6, 0
+; MIPS64-DAG: sll     ${{[0-9]+}}, $7, 0
+; MIPS64-DAG: dsrl    ${{[0-9]+}}, $4, 32
+; MIPS64-DAG: dsrl    ${{[0-9]+}}, $5, 32
+; MIPS64-DAG: dsrl    ${{[0-9]+}}, $6, 32
+; MIPS64-DAG: dsrl    ${{[0-9]+}}, $7, 32
+  %1 = add <4 x i32> %a, %b
+  ret <4 x i32> %1
+}
+
+define <2 x i64> @i64_2(<2 x i64> %a, <2 x i64> %b) {
+; ALL-LABEL: i64_2:
+; MIPS32-DAG: lw ${{[0-9]+}}, 16($sp)
+; MIPS32-DAG: lw ${{[0-9]+}}, 20($sp)
+; MIPS32-DAG: lw ${{[0-9]+}}, 24($sp)
+; MIPS32-DAG: lw ${{[0-9]+}}, 28($sp)
+; MIPS32-DAG: addu $2
+; MIPS32-DAG: addu $3
+; MIPS32-DAG: addu $4
+; MIPS32-DAG: addu $5
+
+; MIPS32R5-DAG: lw ${{[0-9]+}}, 16($sp)
+; MIPS32R5-DAG: lw ${{[0-9]+}}, 20($sp)
+; MIPS32R5-DAG: lw ${{[0-9]+}}, 24($sp)
+; MIPS32R5-DAG: lw ${{[0-9]+}}, 28($sp)
+; MIPS32R5-DAG: insert.w $w[[W0:[0-9]+]][0], $4
+; MIPS32R5-DAG: insert.w $w[[W0]][1], $5
+; MIPS32R5-DAG: insert.w $w[[W0]][2], $6
+; MIPS32R5-DAG: insert.w $w[[W0]][3], $7
+
+; MIPS64-DAG: daddu $2, $4, $6
+; MIPS64-DAG: daddu $3, $5, $7
+
+; MIPS64R5-DAG: insert.d $w[[W0:[0-9]+]][{{[0-9]}}], $4
+; MIPS64R5-DAG: insert.d $w[[W0]][{{[0-9]}}], $5
+; MIPS64R5-DAG: insert.d $w[[W1:[0-9]+]][{{[0-9]}}], $6
+; MIPS64R5-DAG: insert.d $w[[W1]][{{[0-9]}}], $7
+
+  %1 = add <2 x i64> %a, %b
+  ret <2 x i64> %1
+}
+
+; The MIPS vector ABI treats vectors of floats differently to vectors of
+; integers.
+
+; For arguments floating pointer vectors are bitcasted to integer vectors whose
+; elements are of GPR width and where the element count is deduced from
+; the length of the floating point vector divided by the size of the GPRs.
+
+; For returns, integer vectors are passed via the GPR register set, but
+; floating point vectors are returned via a hidden sret pointer.
+
+; For testing purposes we skip returning values here and test them below
+; instead.
+@float_res_v2f32 = external global <2 x float>
+
+define void @float_2(<2 x float> %a, <2 x float> %b) {
+; ALL-LABEL: float_2:
+; MIPS32: mtc1 $7, $f[[F0:[0-9]+]]
+; MIPS32: mtc1 $5, $f[[F1:[0-9]+]]
+; MIPS32: add.s $f[[F2:[0-9]+]], $f[[F1]], $f[[F0]]
+; MIPS32: swc1 $f[[F2]]
+; MIPS32: mtc1 $6, $f[[F3:[0-9]+]]
+; MIPS32: mtc1 $4, $f[[F4:[0-9]+]]
+; MIPS32: add.s $f[[F5:[0-9]+]], $f[[F4]], $f[[F3]]
+; MIPS32: swc1 $f[[F5]]
+
+; MIPS32R5-DAG: sw $4
+; MIPS32R5-DAG: sw $5
+; MIPS32R5-DAG: sw $6
+; MIPS32R5-DAG: sw $7
+
+; MIPS64-DAG: sll $[[R0:[0-9]+]], $4, 0
+; MIPS64-DAG: sll $[[R1:[0-9]+]], $5, 0
+; MIPS64-DAG: mtc1 $[[R0]], $f{{[0-9]+}}
+; MIPS64-DAG: mtc1 $[[R1]], $f{{[0-9]+}}
+; MIPS64-DAG: dsrl $[[R2:[0-9]+]], $4, 32
+; MIPS64-DAG: dsrl $[[R3:[0-9]+]], $5, 32
+; MIPS64-DAG: sll $[[R4:[0-9]+]], $[[R2]], 0
+; MIPS64-DAG: sll $[[R5:[0-9]+]], $[[R3]], 0
+; MIPS64-DAG: mtc1 $[[R4]], $f{{[0-9]+}}
+; MIPS64-DAG: mtc1 $[[R5]], $f{{[0-9]+}}
+
+; MIPS64R5-DAG: sd $4
+; MIPS64R5-DAG: sd $5
+
+  %1 = fadd <2 x float> %a, %b
+  store <2 x float> %1, <2 x float> * @float_res_v2f32
+  ret void
+}
+
+@float_res_v4f32 = external global <4 x float>
+
+; For MSA this case is suboptimal, the 4 loads can be combined into a single
+; ld.w.
+
+define void @float_4(<4 x float> %a, <4 x float> %b) {
+; ALL-LABEL: float_4:
+; MIPS32-DAG: mtc1 $4
+; MIPS32-DAG: mtc1 $5
+; MIPS32-DAG: mtc1 $6
+; MIPS32-DAG: mtc1 $7
+; MIPS32-DAG: lwc1
+; MIPS32-DAG: lwc1
+; MIPS32-DAG: lwc1
+; MIPS32-DAG: lwc1
+
+; MIPS32R5-DAG: lw $[[R1:[0-9]+]], 16($sp)
+; MIPS32R5-DAG: insert.w $w[[W0:[0-9]+]][0], $[[R1]]
+; MIPS32R5-DAG: lw $[[R2:[0-9]+]], 20($sp)
+; MIPS32R5-DAG: insert.w $w[[W0]][1], $[[R2]]
+; MIPS32R5-DAG: lw $[[R3:[0-9]+]], 24($sp)
+; MIPS32R5-DAG: insert.w $w[[W0]][2], $[[R3]]
+; MIPS32R5-DAG: lw $[[R4:[0-9]+]], 28($sp)
+; MIPS32R5-DAG: insert.w $w[[W0]][3], $[[R4]]
+
+; MIPS32R5-DAG: insert.w $w[[W1:[0-9]+]][0], $4
+; MIPS32R5-DAG: insert.w $w[[W1]][1], $5
+; MIPS32R5-DAG: insert.w $w[[W1]][2], $6
+; MIPS32R5-DAG: insert.w $w[[W1]][3], $7
+
+; MIPS64-DAG: sll $[[R0:[0-9]+]], $4, 0
+; MIPS64-DAG: sll $[[R1:[0-9]+]], $5, 0
+; MIPS64-DAG: mtc1 $[[R0]], $f{{[0-9]+}}
+; MIPS64-DAG: mtc1 $[[R1]], $f{{[0-9]+}}
+; MIPS64-DAG: dsrl $[[R2:[0-9]+]], $4, 32
+; MIPS64-DAG: dsrl $[[R3:[0-9]+]], $5, 32
+; MIPS64-DAG: sll $[[R4:[0-9]+]], $[[R2]], 0
+; MIPS64-DAG: sll $[[R5:[0-9]+]], $[[R3]], 0
+; MIPS64-DAG: mtc1 $[[R4]], $f{{[0-9]+}}
+; MIPS64-DAG: mtc1 $[[R5]], $f{{[0-9]+}}
+; MIPS64-DAG: sll $[[R6:[0-9]+]], $6, 0
+; MIPS64-DAG: sll $[[R7:[0-9]+]], $7, 0
+; MIPS64-DAG: mtc1 $[[R6]], $f{{[0-9]+}}
+; MIPS64-DAG: mtc1 $[[R7]], $f{{[0-9]+}}
+; MIPS64-DAG: dsrl $[[R8:[0-9]+]], $6, 32
+; MIPS64-DAG: dsrl $[[R9:[0-9]+]], $7, 32
+; MIPS64-DAG: sll $[[R10:[0-9]+]], $[[R8]], 0
+; MIPS64-DAG: sll $[[R11:[0-9]+]], $[[R9]], 0
+; MIPS64-DAG: mtc1 $[[R10]], $f{{[0-9]+}}
+; MIPS64-DAG: mtc1 $[[R11]], $f{{[0-9]+}}
+
+; MIPS64R5-DAG: insert.d $w[[W0:[0-9]+]][{{[0-9]}}], $4
+; MIPS64R5-DAG: insert.d $w[[W0]][{{[0-9]}}], $5
+; MIPS64R5-DAG: insert.d $w[[W1:[0-9]+]][{{[0-9]}}], $6
+; MIPS64R5-DAG: insert.d $w[[W1]][{{[0-9]}}], $7
+
+  %1 = fadd <4 x float> %a, %b
+  store <4 x float> %1, <4 x float> * @float_res_v4f32
+  ret void
+}
+
+@double_v2f64 = external global <2 x double>
+
+define void @double_2(<2 x double> %a, <2 x double> %b) {
+; ALL-LABEL: double_2:
+; MIPS32-DAG: sw $7
+; MIPS32-DAG: sw $6
+; MIPS32-DAG: ldc1
+; MIPS32-DAG: ldc1
+; MIPS32:     add.d
+; MIPS32-DAG: sw $5
+; MIPS32-DAG: sw $4
+; MIPS32-DAG: ldc1
+; MIPS32-DAG: ldc1
+; MIPS32:     add.d
+
+; MIPS32R5-DAG: lw $[[R1:[0-9]+]], 16($sp)
+; MIPS32R5-DAG: insert.w $w[[W0:[0-9]+]][0], $[[R1]]
+; MIPS32R5-DAG: lw $[[R2:[0-9]+]], 20($sp)
+; MIPS32R5-DAG: insert.w $w[[W0]][1], $[[R2]]
+; MIPS32R5-DAG: lw $[[R3:[0-9]+]], 24($sp)
+; MIPS32R5-DAG: insert.w $w[[W0]][2], $[[R3]]
+; MIPS32R5-DAG: lw $[[R4:[0-9]+]], 28($sp)
+; MIPS32R5-DAG: insert.w $w[[W0]][3], $[[R4]]
+
+; MIPS32R5-DAG: insert.w $w[[W1:[0-9]+]][0], $4
+; MIPS32R5-DAG: insert.w $w[[W1]][1], $5
+; MIPS32R5-DAG: insert.w $w[[W1]][2], $6
+; MIPS32R5-DAG: insert.w $w[[W1]][3], $7
+
+; MIPS64-DAG: dmtc1 $6, $f[[R0:[0-9]+]]
+; MIPS64-DAG: dmtc1 $4, $f[[R1:[0-9]+]]
+; MIPS64-DAG: add.d $f[[R2:[0-9]+]], $f[[R1]], $f[[R0]]
+; MIPS64-DAG: dmtc1 $7, $f[[R3:[0-9]+]]
+; MIPS64-DAG: dmtc1 $5, $f[[R4:[0-9]+]]
+; MIPS64-DAG: add.d $f[[R5:[0-9]+]], $f[[R4]], $f[[R3]]
+
+; MIPS64R5-DAG: insert.d $w[[W0:[0-9]+]][{{[0-9]}}], $4
+; MIPS64R5-DAG: insert.d $w[[W0]][{{[0-9]}}], $5
+; MIPS64R5-DAG: insert.d $w[[W1:[0-9]+]][{{[0-9]}}], $6
+; MIPS64R5-DAG: insert.d $w[[W1]][{{[0-9]}}], $7
+
+  %1 = fadd <2 x double> %a, %b
+  store <2 x double> %1, <2 x double> * @double_v2f64
+  ret void
+}
+
+; Return value testing.
+; Integer vectors are returned in $2, $3, $4, $5 for O32, $2, $3 for N32/N64
+; Floating point vectors are returned through a hidden sret pointer.
+
+@gv2i8 = global <2 x i8> <i8 1, i8 2>
+@gv4i8 = global <4 x i8> <i8 0, i8 1, i8 2, i8 3>
+@gv8i8 = global <8 x i8> <i8 0, i8 1, i8 2, i8 3, i8 4, i8 5, i8 6, i8 7>
+@gv16i8 = global <16 x i8> <i8 0, i8 1, i8 2, i8 3, i8 4, i8 5, i8 6, i8 7, i8 8, i8 9, i8 10, i8 11, i8 12, i8 13, i8 14, i8 15>
+
+@gv2i16 = global <2 x i16> <i16 1, i16 2>
+@gv4i16 = global <4 x i16> <i16 0, i16 1, i16 2, i16 3>
+@gv8i16 = global <8 x i16> <i16 0, i16 1, i16 2, i16 3, i16 4, i16 5, i16 6, i16 7>
+
+@gv2i32 = global <2 x i32> <i32 0, i32 1>
+@gv4i32 = global <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+
+@gv2i64 = global <2 x i64> <i64 0, i64 1>
+
+define <2 x i8> @ret_2_i8() {
+; ALL-LABEL: ret_2_i8:
+; MIPS32-DAG:   lhu $2
+; MIPS32R5-DAG: lhu $2
+
+; FIXME: why is this lh instead of lhu on mips64?
+
+; MIPS64-DAG:  lh $2
+; MIPS64-DAG:  lh $2
+  %1 = load <2 x i8>, <2 x i8> * @gv2i8
+  ret <2 x i8> %1
+}
+
+define <4 x i8> @ret_4_i8() {
+; ALL-LABEL: ret_4_i8:
+; MIPS32-DAG:   lw $2
+; MIPS32R5-DAG: lw $2
+
+; MIPS64-DAG:   lw $2
+; MIPS64R5-DAG: lw $2
+
+  %1 = load <4 x i8>, <4 x i8> * @gv4i8
+  ret <4 x i8> %1
+}
+
+define <8 x i8> @ret_8_i8() {
+; ALL-LABEL: ret_8_i8:
+; MIPS32-DAG:   lw $2
+; MIPS32-DAG:   lw $3
+
+; MIPS32R5: copy_s.w $2, $w[[W0:[0-9]+]]
+; MIPS32R5: copy_s.w $3, $w[[W0]]
+
+; MIPS64-DAG:   ld $2
+; MIPS64R5-DAG: ld $2
+  %1 = load <8 x i8>, <8 x i8> * @gv8i8
+  ret <8 x i8> %1
+}
+
+define <16 x i8> @ret_16_i8() {
+; ALL-LABEL: ret_16_i8:
+; MIPS32-DAG: lw $2
+; MIPS32-DAG: lw $3
+; MIPS32-DAG: lw $4
+; MIPS32-DAG: lw $5
+
+; MIPS32R5-DAG: copy_s.w $2, $w[[W0:[0-9]+]][0]
+; MIPS32R5-DAG: copy_s.w $3, $w[[W0]][1]
+; MIPS32R5-DAG: copy_s.w $4, $w[[W0]][2]
+; MIPS32R5-DAG: copy_s.w $5, $w[[W0]][3]
+
+; MIPS64-DAG: ld $2
+; MIPS64-DAG: ld $3
+
+; MIPS64R5-DAG: copy_s.d $2
+; MIPS64R5-DAG: copy_s.d $3
+
+  %1 = load <16 x i8>, <16 x i8> * @gv16i8
+  ret <16 x i8> %1
+}
+
+define <2 x i16> @ret_2_i16() {
+; ALL-LABEL: ret_2_i16:
+; MIPS32-DAG:   lw $2
+
+; MIPS32R5-DAG: lw $2
+
+; MIPS64-DAG:   lw $2
+
+; MIPS64R5-DAG: lw $2
+  %1 = load <2 x i16>, <2 x i16> * @gv2i16
+  ret <2 x i16> %1
+}
+
+define <4 x i16> @ret_4_i16() {
+; ALL-LABEL: ret_4_i16:
+; MIPS32-DAG: lw $2
+; MIPS32-DAG: lw $3
+
+; MIPS32R5-DAG: copy_s.w $2, $w[[W0:[0-9]+]]
+; MIPS32R5-DAG: copy_s.w $3, $w[[W0]]
+
+; MIPS64-DAG:   ld $2
+; MIPS64R5-DAG: ld $2
+  %1 = load <4 x i16>, <4 x i16> * @gv4i16
+  ret <4 x i16> %1
+}
+
+define <8 x i16> @ret_8_i16() {
+; ALL-LABEL: ret_8_i16:
+; MIPS32-DAG: lw $2
+; MIPS32-DAG: lw $3
+; MIPS32-DAG: lw $4
+; MIPS32-DAG: lw $5
+
+; MIPS32R5-DAG: copy_s.w $2, $w[[W0:[0-9]+]][0]
+; MIPS32R5-DAG: copy_s.w $3, $w[[W0]][1]
+; MIPS32R5-DAG: copy_s.w $4, $w[[W0]][2]
+; MIPS32R5-DAG: copy_s.w $5, $w[[W0]][3]
+
+; MIPS64-DAG: ld $2
+; MIPS64-DAG: ld $3
+
+; MIPS64R5-DAG: copy_s.d $2
+; MIPS64R5-DAG: copy_s.d $3
+
+  %1 = load <8 x i16>, <8 x i16> * @gv8i16
+  ret <8 x i16> %1
+}
+
+define <2 x i32> @ret_2_i32() {
+; ALL-LABEL: ret_2_i32:
+; MIPS32-DAG: lw $2
+; MIPS32-DAG: lw $3
+
+; MIPS32R5-DAG: copy_s.w $2, $w[[W0:[0-9]+]]
+; MIPS32R5-DAG: copy_s.w $3, $w[[W0]]
+
+; MIPS64-DAG:   ld $2
+; MIPS64R5-DAG: ld $2
+
+  %1 = load <2 x i32>, <2 x i32> * @gv2i32
+  ret <2 x i32> %1
+}
+
+define <4 x i32> @ret_4_i32() {
+; ALL-LABEL: ret_4_i32:
+; MIPS32-DAG: lw $2
+; MIPS32-DAG: lw $3
+; MIPS32-DAG: lw $4
+; MIPS32-DAG: lw $5
+
+; MIPS32R5-DAG: copy_s.w $2, $w[[W0:[0-9]+]][0]
+; MIPS32R5-DAG: copy_s.w $3, $w[[W0]][1]
+; MIPS32R5-DAG: copy_s.w $4, $w[[W0]][2]
+; MIPS32R5-DAG: copy_s.w $5, $w[[W0]][3]
+
+; MIPS64-DAG: ld $2
+; MIPS64-DAG: ld $3
+
+; MIPS64R5-DAG: copy_s.d $2, $w[[W0:[0-9]+]]
+; MIPS64R5-DAG: copy_s.d $3, $w[[W0]]
+
+  %1 = load <4 x i32>, <4 x i32> * @gv4i32
+  ret <4 x i32> %1
+}
+
+define <2 x i64> @ret_2_i64() {
+; ALL-LABEL: ret_2_i64:
+; MIPS32-DAG: lw $2
+; MIPS32-DAG: lw $3
+; MIPS32-DAG: lw $4
+; MIPS32-DAG: lw $5
+
+; MIPS32R5-DAG: copy_s.w $2, $w[[W0:[0-9]+]][0]
+; MIPS32R5-DAG: copy_s.w $3, $w[[W0]][1]
+; MIPS32R5-DAG: copy_s.w $4, $w[[W0]][2]
+; MIPS32R5-DAG: copy_s.w $5, $w[[W0]][3]
+
+; MIPS64-DAG: ld $2
+; MIPS64-DAG: ld $3
+
+; MIPS64R5-DAG: copy_s.d $2, $w[[W0:[0-9]+]]
+; MIPS64R5-DAG: copy_s.d $3, $w[[W0]]
+
+  %1 = load <2 x i64>, <2 x i64> * @gv2i64
+  ret <2 x i64> %1
+}
+
+@gv2f32 = global <2 x float> <float 0.0, float 0.0>
+@gv4f32 = global <4 x float> <float 0.0, float 0.0, float 0.0, float 0.0>
+
+define <2 x float> @ret_float_2() {
+entry:
+; ALL-LABEL: ret_float_2:
+
+; MIPS32-DAG: swc1 $f{{[0-9]+}}, 0($4)
+; MIPS32-DAG: swc1 $f{{[0-9]+}}, 4($4)
+
+; MIPS32R5-DAG: swc1 $f{{[0-9]+}}, 0($4)
+; MIPS32R5-DAG: swc1 $f{{[0-9]+}}, 4($4)
+
+; MIPS64: ld $2
+
+; MIPS64R5: ld $2
+
+  %0 = load <2 x float>, <2 x float> * @gv2f32
+  ret <2 x float> %0
+}
+
+define <4 x float> @ret_float_4() {
+entry:
+; ALL-LABEL: ret_float_4:
+
+; MIPS32-DAG: swc1 $f{{[0-9]+}}, 0($4)
+; MIPS32-DAG: swc1 $f{{[0-9]+}}, 4($4)
+; MIPS32-DAG: swc1 $f{{[0-9]+}}, 8($4)
+; MIPS32-DAG: swc1 $f{{[0-9]+}}, 12($4)
+
+; MIPS32R5: st.w $w{{[0-9]+}}, 0($4)
+
+; MIPS64-DAG: ld $2
+; MIPS64-DAG: ld $3
+
+; MIPS64R5-DAG: copy_s.d $2, $w{{[0-9]+}}[0]
+; MIPS64R5-DAG: copy_s.d $3, $w{{[0-9]+}}[1]
+
+  %0 = load <4 x float>, <4 x float> * @gv4f32
+  ret <4 x float> %0
+}
+
+@gv2f64 = global <2 x double> <double 0.0, double 0.0>
+
+define <2 x double> @ret_double_2() {
+entry:
+; ALL-LABEL: ret_double_2:
+
+; MIPS32-DAG: sdc1 $f{{[0-9]+}}, 8($4)
+; MIPS32-DAG: sdc1 $f{{[0-9]+}}, 0($4)
+
+; MIPS32R5: st.d $w{{[0-9]+}}, 0($4)
+
+; MIPS64-DAG: ld $2
+; MIPS64-DAG: ld $2
+
+; MIPS64R5-DAG: copy_s.d $2, $w{{[0-9]+}}[0]
+; MIPS64R5-DAG: copy_s.d $3, $w{{[0-9]+}}[1]
+
+  %0 = load <2 x double>, <2 x double> * @gv2f64
+  ret <2 x double> %0
+}
+
+; Test argument lowering and call result lowering.
+
+define void @call_i8_2() {
+entry:
+; ALL-LABEL: call_i8_2:
+; MIPS32EB-DAG: addiu $4
+; MIPS32EB-DAG: addiu $5
+; MIPS32-NOT: addiu $6
+; MIPS32-NOT: addiu $7
+
+; MIPS32R5-DAG: lhu $4, {{[0-9]+}}($sp)
+; MIPS32R5-DAG: lhu $5, {{[0-9]+}}($sp)
+
+; MIPS32R5: jal
+; MIPS32R5: sw $2, {{[0-9]+}}($sp)
+
+; MIPS32R5-DAG: sb ${{[0-9]+}}, 1(${{[0-9]+}})
+; MIPS32R5-DAG; sb ${{[0-9]+}}, %lo(gv2i8)(${{[0-9]+}})
+
+; MIPS64EB: daddiu $4, $zero, 1543
+; MIPS64EB: daddiu $5, $zero, 3080
+
+; MIPS64EL: daddiu $4, $zero, 1798
+; MIPS64EL; daddiu $5, $zero, 2060
+
+; MIPS64R5-DAG: lh $4
+; MIPS64R5-DAG: lh $5
+
+; MIPS32: jal i8_2
+; MIPS64: jalr $25
+
+; MIPS32EB-DAG: srl $[[R0:[0-9]+]], $2, 16
+; MIPS32EB-DAG: sb $[[R0]]
+; MIPS32EB-DAG: srl $[[R1:[0-9]+]], $2, 24
+; MIPS32EB-DAG: sb $[[R1]]
+
+; MIPS32EL: sb $2
+; MIPS32EL: srl $[[R0:[0-9]+]], $2, 8
+; MIPS32EL: sb $[[R0]]
+
+; MIPS64EB: dsrl $[[R4:[0-9]+]], $2, 48
+; MIPS64EB: sb $[[R4]]
+; MIPS64EB: dsrl $[[R5:[0-9]+]], $2, 56
+; MIPS64EB: sb $[[R5]]
+
+; MIPS64EL: sll $[[R6:[0-9]+]], $2, 0
+; MIPS64EL: sb $[[R6]]
+; MIPS64EL: srl $[[R7:[0-9]+]], $[[R6]], 8
+; MIPS64EL: sb $[[R7]]
+
+; MIPS64R5: sd $2
+
+  %0 = call <2 x i8> @i8_2(<2 x i8> <i8 6, i8 7>, <2 x i8> <i8 12, i8 8>)
+  store <2 x i8> %0, <2 x i8> * @gv2i8
+  ret void
+}
+
+define void @call_i8_4() {
+entry:
+; ALL-LABEL: call_i8_4:
+; MIPS32: ori $4
+; MIPS32: ori $5
+; MIPS32-NOT: ori $6
+; MIPS32-NOT: ori $7
+
+; MIPS32R5-DAG: lw $4, {{[0-9]+}}($sp)
+; MIPS32R5-DAG: lw $5, {{[0-9]+}}($sp)
+
+; MIPS64: ori $4
+; MIPS64: ori $5
+
+; MIPS64R5: lw $4
+; MIPS64R5: lw $5
+
+; MIPS32: jal i8_4
+; MIPS64: jalr $25
+
+; MIPS32: sw $2
+
+; MIPS32R5-DAG: sw $2
+
+; MIPS64: sw $2
+; MIPS64R5: sw $2
+
+  %0 = call <4 x i8> @i8_4(<4 x i8> <i8 6, i8 7, i8 9, i8 10>, <4 x i8> <i8 12, i8 8, i8 9, i8 10>)
+  store <4 x i8> %0, <4 x i8> * @gv4i8
+  ret void
+}
+
+define void @call_i8_8() {
+entry:
+; ALL-LABEL: call_i8_8:
+
+; MIPS32: ori $6
+; MIPS32: ori $4
+; MIPS32: move  $5
+; MIPS32: move  $7
+
+; MIPS32R5-DAG: ori $6
+; MIPS32R5-DAG: ori $4
+; MIPS32R5-DAG: move  $5
+; MIPS32R5-DAG: move  $7
+
+; MIPS64EB: daddiu $4, ${{[0-9]+}}, 2314
+; MIPS64EB: daddiu $5, ${{[0-9]+}}, 2314
+
+; MIPS64EL: daddiu $4, ${{[0-9]+}}, 1798
+; MIPS64EL: daddiu $5, ${{[0-9]+}}, 2060
+
+; MIPS32: jal i8_8
+; MIPS64: jalr $25
+
+; MIPS32-DAG: sw $2
+; MIPS32-DAG: sw $3
+
+; MIPS32R5-DAG: sw $2
+; MIPS32R5-DAG: sw $3
+
+; MIPS64: sd $2
+; MIPS64R5: sd $2
+
+  %0 = call <8 x i8> @i8_8(<8 x i8> <i8 6, i8 7, i8 9, i8 10, i8 6, i8 7, i8 9, i8 10>, <8 x i8> <i8 12, i8 8, i8 9, i8 10, i8 6, i8 7, i8 9, i8 10>)
+  store <8 x i8> %0, <8 x i8> * @gv8i8
+  ret void
+}
+
+define void @calli8_16() {
+entry:
+; ALL-LABEL: calli8_16:
+; MIPS32-DAG: sw  ${{[0-9]+}}, 28($sp)
+; MIPS32-DAG: sw  ${{[0-9]+}}, 24($sp)
+; MIPS32-DAG: sw  ${{[0-9]+}}, 20($sp)
+; MIPS32-DAG: sw  ${{[0-9]+}}, 16($sp)
+
+; MIPS32: ori $4, ${{[0-9]+}}, {{[0-9]+}}
+; MIPS32: ori $7, ${{[0-9]+}}, {{[0-9]+}}
+; MIPS32: move  $5, ${{[0-9]+}}
+; MIPS32: move  $6, ${{[0-9]+}}
+
+; MIPS32R5-DAG: copy_s.w $4, $w{{[0-9]+}}
+; MIPS32R5-DAG: copy_s.w $5, $w{{[0-9]+}}
+; MIPS32R5-DAG: copy_s.w $6, $w{{[0-9]+}}
+; MIPS32R5-DAG: copy_s.w $7, $w{{[0-9]+}}
+
+; MIPS32R5-DAG: sw  ${{[0-9]+}}, 28($sp)
+; MIPS32R5-DAG: sw  ${{[0-9]+}}, 24($sp)
+; MIPS32R5-DAG: sw  ${{[0-9]+}}, 20($sp)
+; MIPS32R5-DAG: sw  ${{[0-9]+}}, 16($sp)
+
+; MIPS64-DAG: daddiu $4
+; MIPS64-DAG: daddiu $5
+; MIPS64-DAG: daddiu $6
+; MIPS64-DAG: daddiu $7
+
+; MIPS64R5-DAG: copy_s.d $4
+; MIPS64R5-DAG: copy_s.d $5
+; MIPS64R5-DAG: copy_s.d $6
+; MIPS64R5-DAG: copy_s.d $7
+
+; MIPS32: jal i8_16
+; MIPS64: jalr $25
+
+; MIPS32-DAG: sw $5, 12(${{[0-9]+}})
+; MIPS32-DAG: sw $4, 8(${{[0-9]+}})
+; MIPS32-DAG: sw $3, 4(${{[0-9]+}})
+; MIPS32-DAG: sw $2, %lo(gv16i8)(${{[0-9]+}})
+
+; MIPS32R5-DAG: insert.w $w[[W0:[0-9]+]][0], $2
+; MIPS32R5-DAG: insert.w $w[[W0]][1], $3
+; MIPS32R5-DAG: insert.w $w[[W0]][2], $4
+; MIPS32R5-DAG: insert.w $w[[W0]][3], $5
+; MIPS32R5-DAG: st.w $w[[W0]]
+
+; MIPS64-DAG: sd $3
+; MIPS64-DAG: sd $2
+
+; MIPS64R5-DAG: insert.d $w[[W0:[0-9]+]][0], $2
+; MIPS64R5-DAG: insert.d $w[[W0:[0-9]+]][1], $3
+
+  %0 = call <16 x i8> @i8_16(<16 x i8> <i8 6, i8 7,i8 6, i8 7,i8 6, i8 7,i8 6, i8 7,i8 6, i8 7,i8 6, i8 7, i8 6, i8 7, i8 9, i8 10>, <16 x i8> <i8 7, i8 9,i8 7, i8 9,i8 7, i8 9,i8 7, i8 9,i8 7, i8 9,i8 7, i8 9,i8 12, i8 8, i8 9, i8 10>)
+  store <16 x i8> %0, <16 x i8> * @gv16i8
+  ret void
+}
+
+define void @calli16_2() {
+entry:
+; ALL-LABEL: calli16_2:
+
+; MIPS32-DAG: ori $4
+; MIPS32-DAG: ori $5
+
+; MIPS32R5-DAG: lw $4
+; MIPS32R5-DAG: lw $5
+
+; MIPS64: ori $4
+; MIPS64: ori $5
+
+; MIPS64R5-DAG: lw $4
+; MIPS64R5-DAG: lw $5
+
+; MIPS32: jal i16_2
+; MIPS64: jalr $25
+
+; MIPS32: sw $2, %lo(gv2i16)
+
+; MIPS32R5: sw $2, %lo(gv2i16)
+
+; MIPS64: sw $2
+
+; MIPS64R6: sw $2
+
+  %0 = call <2 x i16> @i16_2(<2 x i16> <i16 6, i16 7>, <2 x i16> <i16 12, i16 8>)
+  store <2 x i16> %0, <2 x i16> * @gv2i16
+  ret void
+}
+
+define void @calli16_4() {
+entry:
+; ALL-LABEL: calli16_4:
+; MIPS32-DAG: ori $4
+; MIPS32-DAG: ori $5
+; MIPS32-DAG: ori $6
+; MIPS32-DAG: move $7
+
+; MIPS32R5-DAG: ori $4
+; MIPS32R5-DAG: ori $5
+; MIPS32R5-DAG: ori $6
+; MIPS32R5-DAG: move $7
+
+; MIPS64-DAG: daddiu $4
+; MIPS64-DAG: daddiu $5
+
+; MIPS64R5-DAG: ld $4
+; MIPS64R5-DAG: ld $5
+
+; MIPS32: jal i16_4
+; MIPS64: jalr $25
+
+; MIPS32-DAG: sw $3, 4(${{[0-9]+}})
+; MIPS32-DAG: sw $2, %lo(gv4i16)(${{[0-9]+}})
+
+; MIPS32R5-DAG: sw $3, 4(${{[0-9]+}})
+; MIPS32R5-DAG: sw $2, %lo(gv4i16)(${{[0-9]+}})
+
+; MIPS64: sd $2
+; MIPS64R5: sd $2
+
+  %0 = call <4 x i16> @i16_4(<4 x i16> <i16 6, i16 7, i16 9, i16 10>, <4 x i16> <i16 12, i16 8, i16 9, i16 10>)
+  store <4 x i16> %0, <4 x i16> * @gv4i16
+  ret void
+}
+
+define void @calli16_8() {
+entry:
+; ALL-LABEL: calli16_8:
+
+; MIPS32-DAG: sw  ${{[0-9]+}}, 28($sp)
+; MIPS32-DAG: sw  ${{[0-9]+}}, 24($sp)
+; MIPS32-DAG: sw  ${{[0-9]+}}, 20($sp)
+; MIPS32-DAG: sw  ${{[0-9]+}}, 16($sp)
+
+; MIPS32-DAG: ori $4, ${{[0-9]+}}, {{[0-9]+}}
+; MIPS32-DAG: ori $5, ${{[0-9]+}}, {{[0-9]+}}
+; MIPS32-DAG: move  $6, ${{[0-9]+}}
+; MIPS32-DAG: move  $7, ${{[0-9]+}}
+
+; MIPS32R5-DAG: copy_s.w $4, $w{{[0-9]+}}
+; MIPS32R5-DAG: copy_s.w $5, $w{{[0-9]+}}
+; MIPS32R5-DAG: copy_s.w $6, $w{{[0-9]+}}
+; MIPS32R5-DAG: copy_s.w $7, $w{{[0-9]+}}
+
+; MIPS32R5-DAG: sw  ${{[0-9]+}}, 28($sp)
+; MIPS32R5-DAG: sw  ${{[0-9]+}}, 24($sp)
+; MIPS32R5-DAG: sw  ${{[0-9]+}}, 20($sp)
+; MIPS32R5-DAG: sw  ${{[0-9]+}}, 16($sp)
+
+; MIPS64-DAG: daddiu $4
+; MIPS64-DAG: daddiu $7
+; MIPS64-DAG: move $5
+; MIPS64-DAG: move $6
+
+; MIPS64R5-DAG: copy_s.d $4, $w[[W0:[0-9]+]][0]
+; MIPS64R5-DAG: copy_s.d $5, $w[[W0]][1]
+; MIPS64R5-DAG: copy_s.d $6, $w[[W1:[0-9]+]][0]
+; MIPS64R5-DAG: copy_s.d $7, $w[[W1]][1]
+
+; MIPS32: jal i16_8
+; MIPS64: jalr $25
+
+; MIPS32-DAG: sw $5, 12(${{[0-9]+}})
+; MIPS32-DAG: sw $4, 8(${{[0-9]+}})
+; MIPS32-DAG: sw $3, 4(${{[0-9]+}})
+; MIPS32-DAG: sw $2, %lo(gv8i16)(${{[0-9]+}})
+
+; MIPS32R5-DAG: insert.w $w[[W0:[0-9]+]][0], $2
+; MIPS32R5-DAG: insert.w $w[[W0]][1], $3
+; MIPS32R5-DAG: insert.w $w[[W0]][2], $4
+; MIPS32R5-DAG: insert.w $w[[W0]][3], $5
+; MIPS32R5-DAG: st.w $w[[W0]]
+
+; MIPS64: sd $3
+; MIPS64: sd $2
+
+; MIPS64R5-DAG: insert.d $w[[W2:[0-9]+]][0], $2
+; MIPS64R5-DAG: insert.d $w[[W2]][1], $3
+
+  %0 = call <8 x i16> @i16_8(<8 x i16> <i16 6, i16 7, i16 9, i16 10, i16 6, i16 7, i16 9, i16 10>, <8 x i16> <i16 6, i16 7, i16 9, i16 10, i16 12, i16 8, i16 9, i16 10>)
+  store <8 x i16> %0, <8 x i16> * @gv8i16
+  ret void
+}
+
+define void @calli32_2() {
+entry:
+; ALL-LABEL: calli32_2:
+
+; MIPS32-DAG: addiu $4
+; MIPS32-DAG: addiu $5
+; MIPS32-DAG: addiu $6
+; MIPS32-DAG: addiu $7
+
+; MIPS32R5-DAG: addiu $4
+; MIPS32R5-DAG: addiu $5
+; MIPS32R5-DAG: addiu $6
+; MIPS32R5-DAG: addiu $7
+
+; MIPS64: daddiu $4
+; MIPS64: daddiu $5
+
+; MIPS64R5-DAG: ld $4
+; MIPS64R5-DAG: ld $5
+
+; MIPS32: jal i32_2
+; MIPS64: jalr $25
+
+; MIPS32-DAG: sw $2, %lo(gv2i32)(${{[0-9]+}})
+; MIPS32-DAG: sw $3, 4(${{[0-9]+}})
+
+; MIPS32R5-DAG: sw $2, %lo(gv2i32)(${{[0-9]+}})
+; MIPS32R5-DAG: sw $3, 4(${{[0-9]+}})
+
+; MIPS64: sd $2
+
+; MIPS64R5: sd $2
+
+  %0 = call <2 x i32> @i32_2(<2 x i32> <i32 6, i32 7>, <2 x i32> <i32 12, i32 8>)
+  store <2 x i32> %0, <2 x i32> * @gv2i32
+  ret void
+}
+
+define void @calli32_4() {
+entry:
+; ALL-LABEL: calli32_4:
+
+; MIPS32-DAG: sw  ${{[0-9]+}}, 28($sp)
+; MIPS32-DAG: sw  ${{[0-9]+}}, 24($sp)
+; MIPS32-DAG: sw  ${{[0-9]+}}, 20($sp)
+; MIPS32-DAG: sw  ${{[0-9]+}}, 16($sp)
+
+; MIPS32-DAG: addiu $4
+; MIPS32-DAG: addiu $5
+; MIPS32-DAG: addiu $6
+; MIPS32-DAG: addiu $7
+
+; MIPS32R5-DAG: sw  ${{[0-9]+}}, 28($sp)
+; MIPS32R5-DAG: sw  ${{[0-9]+}}, 24($sp)
+; MIPS32R5-DAG: sw  ${{[0-9]+}}, 20($sp)
+; MIPS32R5-DAG: sw  ${{[0-9]+}}, 16($sp)
+
+; MIPS32R5-DAG: addiu $4
+; MIPS32R5-DAG: addiu $5
+; MIPS32R5-DAG: addiu $6
+; MIPS32R5-DAG: addiu $7
+
+; MIPS64-DAG: daddiu $4
+; MIPS64-DAG: daddiu $6
+; MIPS64-DAG: daddiu $5
+; MIPS64-DAG: move $7
+
+; MIPS64R5-DAG: copy_s.d $4, $w[[W0:[0-9]+]][0]
+; MIPS64R5-DAG: copy_s.d $5, $w[[W0]][1]
+; MIPS64R5-DAG: copy_s.d $6, $w[[W1:[0-9]+]][0]
+; MIPS64R5-DAG: copy_s.d $7, $w[[W1]][1]
+
+; MIPS32: jal i32_4
+; MIPS64: jalr $25
+
+; MIPS32-DAG: sw $5, 12(${{[0-9]+}})
+; MIPS32-DAG: sw $4, 8(${{[0-9]+}})
+; MIPS32-DAG: sw $3, 4(${{[0-9]+}})
+; MIPS32-DAG: sw $2, %lo(gv4i32)(${{[0-9]+}})
+
+; MIPS32R5-DAG: insert.w $w[[W0:[0-9]+]][0], $2
+; MIPS32R5-DAG: insert.w $w[[W0]][1], $3
+; MIPS32R5-DAG: insert.w $w[[W0]][2], $4
+; MIPS32R5-DAG: insert.w $w[[W0]][3], $5
+; MIPS32R5-DAG: st.w $w[[W0]]
+
+; MIPS64-DAG: sd $2
+; MIPS64-DAG: sd $3
+
+; MIPS64R5-DAG: insert.d $w[[W0:[0-9]+]][0], $2
+; MIPS64R6-DAG: insert.d $w[[W0:[0-9]+]][1], $3
+
+  %0 = call <4 x i32> @i32_4(<4 x i32> <i32 6, i32 7, i32 9, i32 10>, <4 x i32> <i32 12, i32 8, i32 9, i32 10>)
+  store <4 x i32> %0, <4 x i32> * @gv4i32
+  ret void
+}
+
+define void @calli64_2() {
+entry:
+; ALL-LABEL: calli64_2:
+
+; MIPS32-DAG: sw  ${{[0-9a-z]+}}, 28($sp)
+; MIPS32-DAG: sw  ${{[0-9a-z]+}}, 24($sp)
+; MIPS32-DAG: sw  ${{[0-9a-z]+}}, 20($sp)
+; MIPS32-DAG: sw  ${{[0-9a-z]+}}, 16($sp)
+
+; MIPS32-DAG: addiu $4
+; MIPS32-DAG: addiu $5
+; MIPS32-DAG: addiu $6
+; MIPS32-DAG: addiu $7
+
+; MIPS32R5-DAG: copy_s.w $4, $w{{[0-9]+}}
+; MIPS32R5-DAG: copy_s.w $5, $w{{[0-9]+}}
+; MIPS32R5-DAG: copy_s.w $6, $w{{[0-9]+}}
+; MIPS32R5-DAG: copy_s.w $7, $w{{[0-9]+}}
+
+; MIPS32R5-DAG: sw  ${{[0-9]+}}, 28($sp)
+; MIPS32R5-DAG: sw  ${{[0-9]+}}, 24($sp)
+; MIPS32R5-DAG: sw  ${{[0-9]+}}, 20($sp)
+; MIPS32R5-DAG: sw  ${{[0-9]+}}, 16($sp)
+
+; MIPS64: daddiu $4
+; MIPS64: daddiu $5
+; MIPS64: daddiu $6
+; MIPS64: daddiu $7
+
+; MIPS64R5: daddiu $4
+; MIPS64R5: daddiu $5
+; MIPS64R5: daddiu $6
+; MIPS64R5: daddiu $7
+
+; MIPS32: jal i64_2
+; MIPS64: jalr $25
+
+; MIPS32-DAG: sw $5, 12(${{[0-9]+}})
+; MIPS32-DAG: sw $4, 8(${{[0-9]+}})
+; MIPS32-DAG: sw $3, 4(${{[0-9]+}})
+; MIPS32-DAG: sw $2, %lo(gv2i64)(${{[0-9]+}})
+
+; MIPS32R5-DAG: insert.w $w[[W0:[0-9]+]][0], $2
+; MIPS32R5-DAG: insert.w $w[[W0]][1], $3
+; MIPS32R5-DAG: insert.w $w[[W0]][2], $4
+; MIPS32R5-DAG: insert.w $w[[W0]][3], $5
+; MIPS32R5-DAG: st.w $w[[W0]]
+
+; MIPS64-DAG: sd $3
+; MIPS64-DAG: sd $2
+
+; MIPS64R5-DAG: insert.d $w[[W0:[0-9]+]][0], $2
+; MIPS64R6-DAG: insert.d $w[[W0:[0-9]+]][1], $3
+
+  %0 = call <2 x i64> @i64_2(<2 x i64> <i64 6, i64 7>, <2 x i64> <i64 12, i64 8>)
+  store <2 x i64> %0, <2 x i64> * @gv2i64
+  ret void
+}
+
+declare <2 x float> @float2_extern(<2 x float>, <2 x float>)
+declare <4 x float> @float4_extern(<4 x float>, <4 x float>)
+declare <2 x double> @double2_extern(<2 x double>, <2 x double>)
+
+define void @callfloat_2() {
+entry:
+; ALL-LABEL: callfloat_2:
+
+; MIPS32-DAG: addiu $4, $sp, 24
+; MIPS32-DAG: addiu $6, $zero, 0
+; MIPS32-DAG: lui $7
+
+; MIPS32R5-DAG: addiu $4, $sp, 24
+; MIPS32R5-DAG: addiu $6, $zero, 0
+; MIPS32R5-DAG: lui $7
+
+; MIPS64: dsll $4
+; MIPS64: dsll $5
+
+; MIPS64R5-DAG: copy_s.d $4, $w{{[0-9]+}}
+; MIPS64R5-DAG: copy_s.d $5, $w{{[0-9]+}}
+
+; MIPS32: jal float2_extern
+; MIPS64: jalr $25
+
+; MIPS32-DAG: lwc1 $f[[F0:[0-9]+]], 24($sp)
+; MIPS32-DAG: lwc1 $f[[F1:[0-9]+]], 28($sp)
+
+; MIPS32-DAG: swc1 $f[[F1]], 4(${{[0-9]+}})
+; MIPS32-DAG: swc1 $f[[F0]], %lo(gv2f32)(${{[0-9]+}})
+
+; MIPS32R5-DAG: lwc1 $f[[F0:[0-9]+]], 24($sp)
+; MIPS32R5-DAG: lwc1 $f[[F1:[0-9]+]], 28($sp)
+
+; MIPS32R5-DAG: swc1 $f[[F1]], 4(${{[0-9]+}})
+; MIPS32R5-DAG: swc1 $f[[F0]], %lo(gv2f32)(${{[0-9]+}})
+
+; MIPS64: sd $2
+
+; MIPS64R5: sd $2
+
+  %0 = call <2 x float> @float2_extern(<2 x float> <float 0.0, float -1.0>, <2 x float> <float 12.0, float 14.0>)
+  store <2 x float> %0, <2 x float> * @gv2f32
+  ret void
+}
+
+define void @callfloat_4() {
+entry:
+; ALL-LABEL: callfloat_4:
+
+; MIPS32: sw ${{[0-9]+}}, 36($sp)
+; MIPS32: sw ${{[0-9]+}}, 32($sp)
+; MIPS32: sw ${{[0-9]+}}, 28($sp)
+; MIPS32: sw ${{[0-9]+}}, 24($sp)
+; MIPS32: sw ${{[0-9]+}}, 20($sp)
+; MIPS32: sw ${{[0-9]+}}, 16($sp)
+; MIPS32: addiu $4, $sp, 48
+; MIPS32: addiu $6, $zero, 0
+; MIPS32: lui $7
+
+; MIPS32R5: copy_s.w $6, $w{{[0-9]+}}
+; MIPS32R5: copy_s.w $7, $w{{[0-9]+}}
+; MIPS32R5: sw ${{[0-9]+}}, 36($sp)
+; MIPS32R5: sw ${{[0-9]+}}, 32($sp)
+; MIPS32R5: sw ${{[0-9]+}}, 28($sp)
+; MIPS32R5: sw ${{[0-9]+}}, 24($sp)
+; MIPS32R5: sw ${{[0-9]+}}, 20($sp)
+; MIPS32R5: sw ${{[0-9]+}}, 16($sp)
+; MIPS32R5: addiu $4, $sp, 48
+
+; MIPS64-DAG: dsll $4
+; MIPS64-DAG: dsll $5
+; MIPS64-DAG: dsll $6
+; MIPS64-DAG: dsll $7
+
+; MIPS64R5-DAG: copy_s.d $4, $w{{[0-9]+}}
+; MIPS64R5-DAG: copy_s.d $5, $w{{[0-9]+}}
+; MIPS64R5-DAG: copy_s.d $6, $w{{[0-9]+}}
+; MIPS64R5-DAG: copy_s.d $7, $w{{[0-9]+}}
+
+; MIPS64: jalr $25
+; MIPS32: jal
+
+; MIPS32-DAG: lwc1 $f{{[0-9]+}}, 48($sp)
+; MIPS32-DAG: lwc1 $f{{[0-9]+}}, 52($sp)
+; MIPS32-DAG: lwc1 $f{{[0-9]+}}, 56($sp)
+; MIPS32-DAG: lwc1 $f{{[0-9]+}}, 60($sp)
+
+; MIPS32R5: ld.w $w{{[0-9]+}}, 48($sp)
+
+; MIPS64-DAG: $2
+; MIPS64-DAG: $3
+
+; MIPS64R5-DAG: insert.d $w[[W0:[0-9]+]][0], $2
+; MIPS64R5-DAG: insert.d $w[[W0:[0-9]+]][1], $3
+
+  %0 = call <4 x float> @float4_extern(<4 x float> <float 0.0, float -1.0, float 2.0, float 4.0>, <4 x float> <float 12.0, float 14.0, float 15.0, float 16.0>)
+  store <4 x float> %0, <4 x float> * @gv4f32
+  ret void
+}
+
+define void @calldouble_2() {
+entry:
+; ALL-LABEL: calldouble_2:
+
+; MIPS32-DAG: sw ${{[0-9a-z]+}}, 36($sp)
+; MIPS32-DAG: sw ${{[0-9a-z]+}}, 32($sp)
+; MIPS32-DAG: sw ${{[0-9a-z]+}}, 28($sp)
+; MIPS32-DAG: sw ${{[0-9a-z]+}}, 24($sp)
+; MIPS32-DAG: sw ${{[0-9a-z]+}}, 20($sp)
+; MIPS32-DAG: sw ${{[0-9a-z]+}}, 16($sp)
+
+; MIPS32-DAG: addiu $4, $sp, [[R0:[0-9]+]]
+; MIPS32-DAG: addiu $6, $zero, 0
+; MIPS32-DAG: addiu $7, $zero, 0
+
+; MIPS32R5-DAG: copy_s.w $4, $w{{[0-9]+}}
+; MIPS32R5-DAG: copy_s.w $5, $w{{[0-9]+}}
+; MIPS32R5-DAG: copy_s.w $6, $w{{[0-9]+}}
+; MIPS32R5-DAG: copy_s.w $7, $w{{[0-9]+}}
+
+; MIPS32R5-DAG: sw  ${{[0-9]+}}, 36($sp)
+; MIPS32R5-DAG: sw  ${{[0-9]+}}, 32($sp)
+; MIPS32R5-DAG: sw  ${{[0-9]+}}, 28($sp)
+; MIPS32R5-DAG: sw  ${{[0-9]+}}, 24($sp)
+; MIPS32R5-DAG: sw  ${{[0-9]+}}, 20($sp)
+; MIPS32R5-DAG: sw  ${{[0-9]+}}, 16($sp)
+
+; MIPS64-DAG: dsll $5
+; MIPS64-DAG: dsll $6
+; MIPS64-DAG: dsll $7
+; MIPS64-DAG: daddiu $4
+
+; MIPS64R5-DAG: copy_s.d $4, $w{{[0-9]+}}
+; MIPS64R5-DAG: copy_s.d $5, $w{{[0-9]+}}
+; MIPS64R5-DAG: copy_s.d $6, $w{{[0-9]+}}
+; MIPS64R5-DAG: copy_s.d $7, $w{{[0-9]+}}
+
+; MIPS32: jal double2_extern
+; MIPS64: jalr $25
+
+; MIPS32-DAG: ldc1 $f[[F0:[0-9]+]], 48($sp)
+; MIPS32-DAG: ldc1 $f[[F1:[0-9]+]], 56($sp)
+
+; MIPS32-DAG: sdc1 $f[[F1]], 8(${{[0-9]+}})
+; MIPS32-DAG: sdc1 $f[[F0]], %lo(gv2f64)(${{[0-9]+}})
+
+; MIPS32R5: ld.d $w[[W0:[0-9]+]], 48($sp)
+; MIPS32R5: st.d $w[[W0]], 0(${{[0-9]+}})
+
+; MIPS64-DAG: sd $2
+; MIPS64-DAG: sd $3
+
+; MIPS64R5-DAG: insert.d $w[[W0:[0-9]+]][0], $2
+; MIPS64R5-DAG: insert.d $w[[W0:[0-9]+]][1], $3
+
+  %0 = call <2 x double> @double2_extern(<2 x double> <double 0.0, double -1.0>, <2 x double> <double 12.0, double 14.0>)
+  store <2 x double> %0, <2 x double> * @gv2f64
+  ret void
+}
+
+; The mixed tests show that due to alignment requirements, $5 is not used
+; in argument passing.
+
+define float @mixed_i8(<2 x float> %a, i8 %b, <2 x float> %c) {
+entry:
+; ALL-LABEL: mixed_i8:
+
+; MIPS32-DAG: mtc1 $5, $f{{[0-9]+}}
+; MIPS32: andi $[[R7:[0-9]+]], $6, 255
+; MIPS32: mtc1 $[[R7]], $f[[F0:[0-9]+]]
+; MIPS32: cvt.s.w $f{{[0-9]+}}, $f[[F0]]
+
+; MIPS32-DAG: mtc1 $4, $f{{[0-9]+}}
+; MIPS32-DAG: lwc1 $f{{[0-9]+}}, 16($sp)
+; MIPS32-DAG: lwc1 $f{{[0-9]+}}, 20($sp)
+; MIPS32-DAG: add.s $f0, $f{{[0-9]+}}, $f{{[0-9]+}}
+
+; MIPS32R5: andi $[[R0:[0-9]+]], $6, 255
+; MIPS32R5: sw $[[R0]], {{[0-9]+}}($sp)
+; MIPS32R5: sw $[[R0]], {{[0-9]+}}($sp)
+; MIPS32R5-DAG: sw $5, {{[0-9]+}}($sp)
+; MIPS32R5-DAG: sw $4, {{[0-9]+}}($sp)
+
+; MIPS64EB-DAG: sll $[[R0:[0-9]+]], $4, 0
+; MIPS64EB-DAG: mtc1 $[[R0]], $f{{[0-9]+}}
+; MIPS64EB: sll $[[R6:[0-9]+]], $5, 0
+; MIPS64EB: andi $[[R7:[0-9]+]], $[[R6]], 255
+; MIPS64EB: mtc1 $[[R7]], $f[[F0:[0-9]+]]
+; MIPS64EB: cvt.s.w $f{{[0-9]+}}, $f[[F0]]
+
+; MIPS64EB-DAG: dsrl $[[R1:[0-9]+]], $4, 32
+; MIPS64EB-DAG: sll $[[R2:[0-9]+]], $[[R1]], 0
+; MIPS64EB-DAG: mtc1 $[[R2:[0-9]+]], $f{{[0-9]+}}
+
+; MIPS64EB-DAG: sll $[[R3:[0-9]+]], $6, 0
+; MIPS64EB-DAG: mtc1 $[[R3]], $f{{[0-9]+}}
+; MIPS64EB-DAG: dsrl $[[R4:[0-9]+]], $6, 32
+; MIPS64EB-DAG: sll $[[R5:[0-9]+]], $[[R4]], 0
+; MIPS64EB-DAG: mtc1 $[[R5:[0-9]+]], $f{{[0-9]+}}
+
+; MIPS64EL-DAG: dsrl $[[R1:[0-9]+]], $4, 32
+; MIPS64EL-DAG: sll $[[R2:[0-9]+]], $[[R1]], 0
+; MIPS64EL-DAG: mtc1 $[[R2:[0-9]+]], $f{{[0-9]+}}
+
+; MIPS64EL: sll $[[R6:[0-9]+]], $5, 0
+; MIPS64EL: andi $[[R7:[0-9]+]], $[[R6]], 255
+; MIPS64EL: mtc1 $[[R7]], $f[[F0:[0-9]+]]
+; MIPS64EL: cvt.s.w $f{{[0-9]+}}, $f[[F0]]
+
+; MIPS64EL-DAG: dsrl $[[R4:[0-9]+]], $6, 32
+; MIPS64EL-DAG: sll $[[R5:[0-9]+]], $[[R4]], 0
+; MIPS64EL-DAG: mtc1 $[[R5:[0-9]+]], $f{{[0-9]+}}
+
+; MIPS64EL-DAG: sll $[[R0:[0-9]+]], $4, 0
+; MIPS64EL-DAG: mtc1 $[[R0]], $f{{[0-9]+}}
+; MIPS64EL-DAG: sll $[[R3:[0-9]+]], $6, 0
+; MIPS64EL-DAG: mtc1 $[[R3]], $f{{[0-9]+}}
+
+; MIPS64R5: sll $[[R0:[0-9]+]], $5, 0
+; MIPS64R5: andi $[[R1:[0-9]+]], $[[R0]], 255
+; MIPS64R5: sd $4, {{[0-9]+}}($sp)
+; MIPS64R5: sd $6, {{[0-9]+}}($sp)
+
+  %0 = zext i8 %b to i32
+  %1 = uitofp i32 %0 to float
+  %2 = insertelement <2 x float> undef, float %1, i32 0
+  %3 = insertelement <2 x float> %2, float %1, i32 1
+  %4 = fadd <2 x float> %3, %a
+  %5 = fadd <2 x float> %4, %c
+  %6 = extractelement <2 x float> %5, i32 0
+  %7 = extractelement <2 x float> %5, i32 1
+  %8 = fadd float %6, %7
+  ret float %8
+}
+
+define <4 x float> @mixed_32(<4 x float> %a, i32 %b) {
+entry:
+; ALL-LABEL: mixed_32:
+
+; MIPS32-DAG: mtc1 $6, $f{{[0-9]+}}
+; MIPS32-DAG: mtc1 $7, $f{{[0-9]+}}
+; MIPS32-DAG: lwc1 $f{{[0-9]+}}, 28($sp)
+; MIPS32-DAG: lwc1 $f{{[0-9]+}}, 24($sp)
+; MIPS32-DAG: swc1 $f{{[0-9]+}}, 0($4)
+; MIPS32-DAG: swc1 $f{{[0-9]+}}, 4($4)
+; MIPS32-DAG: swc1 $f{{[0-9]+}}, 8($4)
+; MIPS32-DAG: swc1 $f{{[0-9]+}}, 12($4)
+
+; MIPS32R5: insert.w $w[[W0:[0-9]+]][0], $6
+; MIPS32R5: insert.w $w[[W0:[0-9]+]][1], $7
+; MIPS32R5: lw $[[R0:[0-9]+]], 16($sp)
+; MIPS32R5: insert.w $w[[W0:[0-9]+]][2], $[[R0]]
+; MIPS32R5: lw $[[R1:[0-9]+]], 20($sp)
+; MIPS32R5: insert.w $w[[W0:[0-9]+]][3], $[[R1]]
+; MIPS32R5: lw $[[R0:[0-9]+]], 24($sp)
+
+; MIPS64-DAG: sll ${{[0-9]+}}, $6, 0
+; MIPS64-DAG: dsrl $[[R0:[0-9]+]], $4, 32
+; MIPS64-DAG: sll $[[R1:[0-9]+]], $[[R0]], 0
+; MIPS64-DAG: mtc1 $[[R1]], $f{{[0-9]+}}
+; MIPS64-DAG: sll $[[R2:[0-9]+]], $4, 0
+; MIPS64-DAG: dsrl $[[R3:[0-9]+]], $5, 32
+; MIPS64-DAG: sll $[[R4:[0-9]+]], $[[R3]], 0
+; MIPS64-DAG: mtc1 $[[R4]], $f{{[0-9]+}}
+; MIPS64-DAG: mtc1 $[[R2]], $f{{[0-9]+}}
+; MIPS64-DAG: sll $[[R6:[0-9]+]], $5, 0
+; MIPS64-DAG: mtc1 $[[R6:[0-9]+]], $f{{[0-9]+}}
+
+; MIPS64R5: insert.d $w[[W0:[0-9]+]][0], $4
+; MIPS64R5: insert.d $w[[W0]][1], $5
+; MIPS64R5: sll $[[R0:[0-9]+]], $6, 0
+; MIPS64R5: fill.w $w{{[0-9]+}}, $[[R0]]
+
+  %0 = uitofp i32 %b to float
+  %1 = insertelement <4 x float> undef, float %0, i32 0
+  %2 = insertelement <4 x float> %1, float %0, i32 1
+  %3 = insertelement <4 x float> %2, float %0, i32 2
+  %4 = insertelement <4 x float> %3, float %0, i32 3
+  %5 = fadd <4 x float> %4, %a
+  ret <4 x float> %5
+}
+
+
+; This test is slightly more fragile than I'd like as the offset into the
+; outgoing arguments area is dependant on the size of the stack frame for
+; this function.
+
+define <4 x float> @cast(<4 x i32> %a) {
+entry:
+; ALL-LABEL: cast:
+
+; MIPS32: addiu $sp, $sp, -32
+; MIPS32-DAG: sw $6, {{[0-9]+}}($sp)
+; MIPS32-DAG: sw $7, {{[0-9]+}}($sp)
+; MIPS32-DAG: lw ${{[0-9]+}}, 48($sp)
+; MIPS32-DAG: lw ${{[0-9]+}}, 52($sp)
+
+; MIPS32R5-DAG: insert.w  $w0[0], $6
+; MIPS32R5-DAG: insert.w  $w0[1], $7
+; MIPS32R5-DAG: lw  $[[R0:[0-9]+]], 16($sp)
+; MIPS32R5-DAG: insert.w  $w0[2], $[[R0]]
+; MIPS32R5-DAG: lw  $[[R1:[0-9]+]], 20($sp)
+; MIPS32R5-DAG: insert.w  $w0[3], $[[R1]]
+
+; MIPS64-DAG: sll ${{[0-9]+}}, $4, 0
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $4, 32
+; MIPS64-DAG: sll ${{[0-9]+}}, $5, 0
+; MIPS64-DAG: dsrl ${{[0-9]+}}, $5, 32
+
+; MIPS64R5-DAG: insert.d  $w0[0], $4
+; MIPS64R5-DAG: insert.d  $w0[1], $5
+
+  %0 = uitofp <4 x i32> %a to <4 x float>
+  ret <4 x float> %0
+}
+
+define <4 x float> @select(<4 x i32> %cond, <4 x float> %arg1, <4 x float> %arg2) {
+entry:
+; ALL-LABEL: select:
+
+; MIPS32-DAG: andi ${{[0-9]+}}, $7, 1
+; MIPS32-DAG: andi ${{[0-9]+}}, $6, 1
+; MIPS32-DAG: lw $[[R0:[0-9]+]], 16($sp)
+; MIPS32-DAG: andi ${{[0-9]+}}, $[[R0]], 1
+; MIPS32-DAG: lw $[[R1:[0-9]+]], 20($sp)
+; MIPS32-DAG: andi ${{[0-9]+}}, $[[R0]], 1
+
+; MIPS32R5-DAG: insert.w $w[[W0:[0-9]+]][0], $6
+; MIPS32R5-DAG: insert.w $w[[W0]][1], $7
+; MIPS32R5-DAG: lw $[[R0:[0-9]+]], 16($sp)
+; MIPS32R5-DAG: lw $[[R1:[0-9]+]], 20($sp)
+; MIPS32R5-DAG: insert.w $w[[W0]][2], $[[R0]]
+; MIPS32R5-DAG: insert.w $w[[W0]][3], $[[R1]]
+; MIPS32R5-DAG: slli.w $w{{[0-9]}}, $w[[W0]]
+
+; MIPS64-DAG: sll $[[R0:[0-9]+]], $6, 0
+; MIPS64-DAG: mtc1 $[[R0]], $f{{[0-9]+}}
+; MIPS64-DAG: dsrl $[[R1:[0-9]+]], $6, 32
+; MIPS64-DAG: sll $[[R2:[0-9]+]], $[[R1]], 0
+; MIPS64-DAG: mtc1 $[[R2]], $f{{[0-9]+}}
+
+; MIPS64-DAG: sll $[[R3:[0-9]+]], $7, 0
+; MIPS64-DAG: mtc1 $[[R3]], $f{{[0-9]+}}
+; MIPS64-DAG: dsrl $[[R4:[0-9]+]], $7, 32
+; MIPS64-DAG: sll $[[R5:[0-9]+]], $[[R4]], 0
+; MIPS64-DAG: mtc1 $[[R5]], $f{{[0-9]+}}
+
+; MIPS64-DAG: sll $[[R6:[0-9]+]], $8, 0
+; MIPS64-DAG: mtc1 $[[R6]], $f{{[0-9]+}}
+; MIPS64-DAG: dsrl $[[R7:[0-9]+]], $8, 32
+; MIPS64-DAG: sll $[[R8:[0-9]+]], $[[R7]], 0
+; MIPS64-DAG: mtc1 $[[R8]], $f{{[0-9]+}}
+
+; MIPS64-DAG: sll $[[R9:[0-9]+]], $9, 0
+; MIPS64-DAG: mtc1 $[[R9]], $f{{[0-9]+}}
+; MIPS64-DAG: dsrl $[[R10:[0-9]+]], $9, 32
+; MIPS64-DAG: sll $[[R11:[0-9]+]], $[[R10]], 0
+; MIPS64-DAG: mtc1 $[[R11]], $f{{[0-9]+}}
+
+; MIPS64-DAG: sll $[[R12:[0-9]+]], $4, 0
+; MIPS64-DAG: andi ${{[0-9]+}}, $[[R12]], 1
+; MIPS64-DAG: dsrl $[[R13:[0-9]+]], $4, 32
+; MIPS64-DAG: sll $[[R14:[0-9]+]], $[[R13]], 0
+; MIPS64-DAG: andi ${{[0-9]+}}, $[[R14]], 1
+
+; MIPS64-DAG: sll $[[R15:[0-9]+]], $5, 0
+; MIPS64-DAG: andi ${{[0-9]+}}, $[[R15]], 1
+; MIPS64-DAG: dsrl $[[R16:[0-9]+]], $5, 32
+; MIPS64-DAG: sll $[[R17:[0-9]+]], $[[R16]], 0
+; MIPS64-DAG: andi ${{[0-9]+}}, $[[R17]], 1
+
+; MIPS64R5-DAG: insert.d $w{{[0-9]+}}[0], $8
+; MIPS64R5-DAG: insert.d $w{{[0-9]+}}[1], $9
+; MIPS64R5-DAG: insert.d $w{{[0-9]+}}[0], $6
+; MIPS64R5-DAG: insert.d $w{{[0-9]+}}[1], $7
+; MIPS64R5-DAG: insert.d $w{{[0-9]+}}[0], $4
+; MIPS64R5-DAG: insert.d $w{{[0-9]+}}[1], $5
+
+  %cond.t = trunc <4 x i32> %cond to <4 x i1>
+  %res = select <4 x i1> %cond.t, <4 x float> %arg1, <4 x float> %arg2
+  ret <4 x float> %res
+}

llvm/trunk/test/CodeGen/Mips/ctlz-v.ll

 ; RUN: llc  < %s -march=mipsel -mcpu=mips32r2 | FileCheck %s -check-prefix=MIPS32
 ; RUN: llc  < %s -march=mips64el -mcpu=mips64r2 | FileCheck %s -check-prefix=MIPS64
 
 declare <2 x i32> @llvm.ctlz.v2i32(<2 x i32>, i1)
 
 define <2 x i32> @ctlzv2i32(<2 x i32> %x) {
 entry:
 ; MIPS32: clz     $2, $4
 ; MIPS32: clz     $3, $5
 
-; MIPS64-DAG: sll $[[A0:[0-9]+]], $4, 0
-; MIPS64-DAG: clz $2, $[[A0]]
-; MIPS64-DAG: sll $[[A1:[0-9]+]], $5, 0
-; MIPS64-DAG: clz $3, $[[A1]]
+; MIPS64-DAG: dsrl $[[A0:[0-9]+]], $4, 32
+; MIPS64-DAG: sll $[[A1:[0-9]+]], $[[A0]], 0
+; MIPS64-DAG: clz $[[R0:[0-9]+]], $[[A1]]
+; MIPS64-DAG: dsll $[[R1:[0-9]+]], $[[R0]], 32
+; MIPS64-DAG: sll $[[A2:[0-9]+]], $4, 0
+; MIPS64-DAG: clz $[[R2:[0-9]+]], $[[A2]]
+; MIPS64-DAG: dext $[[R3:[0-9]+]], $[[R2]], 0, 32
+; MIPS64-DAG: or $2, $[[R3]], $[[R1]]
 
   %ret = call <2 x i32> @llvm.ctlz.v2i32(<2 x i32> %x, i1 true)
   ret <2 x i32> %ret
 }
 

llvm/trunk/test/CodeGen/Mips/cttz-v.ll

 ; MIPS32-DAG: subu    $3, $[[R4]], $[[R8]]
 
 ; MIPS64-DAG: sll     $[[A0:[0-9]+]], $4, 0
 ; MIPS64-DAG: addiu   $[[R0:[0-9]+]], $[[A0]], -1
 ; MIPS64-DAG: not     $[[R1:[0-9]+]], $[[A0]]
 ; MIPS64-DAG: and     $[[R2:[0-9]+]], $[[R1]], $[[R0]]
 ; MIPS64-DAG: clz     $[[R3:[0-9]+]], $[[R2]]
 ; MIPS64-DAG: addiu   $[[R4:[0-9]+]], $zero, 32
-; MIPS64-DAG: subu    $2, $[[R4]], $[[R3]]
-; MIPS64-DAG: sll     $[[A1:[0-9]+]], $5, 0
-; MIPS64-DAG: addiu   $[[R5:[0-9]+]], $[[A1]], -1
-; MIPS64-DAG: not     $[[R6:[0-9]+]], $[[A1]]
-; MIPS64-DAG: and     $[[R7:[0-9]+]], $[[R6]], $[[R5]]
-; MIPS64-DAG: clz     $[[R8:[0-9]+]], $[[R7]]
-; MIPS64-DAG: jr      $ra
-; MIPS64-DAG: subu    $3, $[[R4]], $[[R8]]
+; MIPS64-DAG: subu    $[[R5:[0-9]+]], $[[R4]], $[[R3]]
+; MIPS64-DAG: dsrl    $[[R6:[0-9]+]], $4, 32
+; MIPS64-DAG: sll     $[[R7:[0-9]+]], $[[R6]], 0
+; MIPS64-DAG: dext    $[[R8:[0-9]+]], $[[R5]], 0, 32
+; MIPS64-DAG: addiu   $[[R9:[0-9]+]], $[[R7]], -1
+; MIPS64-DAG: not     $[[R10:[0-9]+]], $[[R7]]
+; MIPS64-DAG: and     $[[R11:[0-9]+]], $[[R10]], $[[R9]]
+; MIPS64-DAG: clz     $[[R12:[0-9]+]], $[[R11]]
+; MIPS64-DAG: subu    $[[R13:[0-9]+]], $[[R4]], $[[R12]]
+; MIPS64-DAG: dsll    $[[R14:[0-9]+]], $[[R13]], 32
+; MIPS64-DAG: or      $2, $[[R8]], $[[R14]]
 
   %ret = call <2 x i32> @llvm.cttz.v2i32(<2 x i32> %x, i1 true)
   ret <2 x i32> %ret
 }
 

llvm/trunk/test/CodeGen/Mips/return-vector.ll

entry:
   %call = call <2 x float> (...) @f2()
   %v0 = extractelement <2 x float> %call, i32 0
   %v1 = extractelement <2 x float> %call, i32 1
   %add1 = fadd float %v0, %v1
   ret float %add1
 
 ; CHECK-LABEL:        call_f2:
 ; CHECK:        call16(f2)
-; CHECK-NOT:    lwc1
-; CHECK:        add.s    $[[R2:[a-z0-9]+]], $[[R0:[a-z0-9]+]], $[[R1:[a-z0-9]+]]
+; CHECK:        addiu $4, $sp, [[O0:[0-9]+]]
+; CHECK-DAG:    lwc1 $f[[F0:[0-9]]], [[O0]]($sp)
+; CHECK-DAG:    lwc1 $f[[F1:[0-9]]], 20($sp)
+; CHECK:        add.s    $f0, $f[[F0]], $f[[F1]]
+
 }
 
 
 define double @call_d2() {
 entry:
   %call = call <2 x double> (...) @d2()
   %v0 = extractelement <2 x double> %call, i32 0
   %v1 = extractelement <2 x double> %call, i32 1
   %add1 = fadd double %v0, %v1
   ret double %add1
 
 ; CHECK-LABEL:        call_d2:
 ; CHECK:        call16(d2)
-; CHECK-NOT:    ldc1
-; CHECK:        add.d    $[[R2:[a-z0-9]+]], $[[R0:[a-z0-9]+]], $[[R1:[a-z0-9]+]]
-}
-
+; CHECK:        addiu $4, $sp, [[O0:[0-9]+]]
+; CHECK-DAG:    ldc1 $f[[F0:[0-9]+]], 24($sp)
+; CHECK-DAG:    ldc1 $f[[F1:[0-9]+]], [[O0]]($sp)
+; CHECK:        add.d    $f0, $f[[F1]], $f[[F0]]
 
+}
 
 ; Check that function returns vector on stack in cases when vector can't be
 ; returned in registers. Also check that vector is placed on stack starting
 ; from the address in register $4.
 
 
 define <8 x i32> @return_i8() {
 entry:

 entry:
   %vecins1 = insertelement <4 x float> undef,    float %a, i32 0
   %vecins2 = insertelement <4 x float> %vecins1, float %b, i32 1
   %vecins3 = insertelement <4 x float> %vecins2, float %c, i32 2
   %vecins4 = insertelement <4 x float> %vecins3, float %d, i32 3
   ret <4 x float> %vecins4
 
 ; CHECK-LABEL:        return_f4:
-; CHECK-DAG:    lwc1    $[[R0:[a-z0-9]+]], 16($sp)
-; CHECK-DAG:    swc1    $[[R0]], 12($4)
+; CHECK-DAG:    lwc1    $f[[R0:[0-9]+]], 16($sp)
+; CHECK-DAG:    swc1    $f[[R0]], 12($4)
 ; CHECK-DAG:    sw      $7, 8($4)
 ; CHECK-DAG:    sw      $6, 4($4)
 ; CHECK-DAG:    sw      $5, 0($4)
+
 }
 
 
 define <4 x double> @return_d4(double %a, double %b, double %c, double %d) {
 entry:
   %vecins1 = insertelement <4 x double> undef,    double %a, i32 0
   %vecins2 = insertelement <4 x double> %vecins1, double %b, i32 1
   %vecins3 = insertelement <4 x double> %vecins2, double %c, i32 2

 
 define <2 x float> @return_f2(float %a, float %b) {
 entry:
   %vecins1 = insertelement <2 x float> undef,    float %a, i32 0
   %vecins2 = insertelement <2 x float> %vecins1, float %b, i32 1
   ret <2 x float> %vecins2
 
 ; CHECK-LABEL:        return_f2:
-; CHECK:        mov.s   $f0, $f12
-; CHECK:        mov.s   $f2, $f14
+; CHECK-DAG:    sw   $5, 0($4)
+; CHECK-DAG:    sw   $6, 4($4)
 }
 
 
 define <2 x double> @return_d2(double %a, double %b) {
 entry:
   %vecins1 = insertelement <2 x double> undef,    double %a, i32 0
   %vecins2 = insertelement <2 x double> %vecins1, double %b, i32 1
   ret <2 x double> %vecins2
 
 ; CHECK-LABEL:        return_d2:
-; CHECK:        mov.d   $f0, $f12
-; CHECK:        mov.d   $f2, $f14
+; CHECK-DAG:    ldc1 $f[[F0:[0-9]]], 16($sp)
+; CHECK-DAG:    sdc1 $f[[F0]], 8($4)
+; CHECK-DAG:    mtc1 $6, $f[[F1:[0-9]+]]
+; CHECK-DAG:    mtc1 $7, $f
+; CHECK-DAG:    sdc1 $f[[F0]], 0($4)
+
 }