[TTI][ARM][MVE] Refine gather/scatter cost model

[anwel]

Summary

This patch refines the gather/scatter cost model, but also changes the TTI function getIntrinsicInstrCost to as an additional parameter accept the instruction that is evaluated, which is needed for the gather/scatter cost evaluation.
The latter did require trivial changes in some non-ARM backends to adopt the new parameter.

Gathers can be extending, so if their result is sext or zext extended, the extend can be integrated. This ensures that in these cases the cost model looks at the type after extension and calculates the cost accordingly.
Similarly, a trunc of a value can be integrated into the scatter that stores it.

[RKSimon]

llvm/include/llvm/Analysis/TargetTransformInfo.h

968–970

Please can you add the doxygen param entry?

1011–1013

Explain what I is

[dmgreen]

The fact that the cost of zext/sext's change so much with context is pretty annoying. This "per single instruction" costmodel is having trouble keeping up. We might need to eventually do something about that, come up with something better at a higher level. I'm not sure what that would look like though.

This looks like a nice improvement over what was already there.

llvm/include/llvm/Analysis/TargetTransformInfo.h

1016–1017

Is it possible to keep I last and give it a default argument of nullptr? Or does that cause other problems?

llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp

897

check -> Check

907

only -> Only

921

I don't think this I->getNumUses() == 0 is needed?

llvm/test/Analysis/CostModel/ARM/mve-gather-scatter-cost.ll

177–187

Can you move these into gep_v4i16. And add a trunc to the scatters. Maybe something like:

%res6 = call <4 x i16> @llvm.masked.gather.v4i16.v4p0i16(<4 x i16*> %gep5, i32 2, <4 x i1> %mask, <4 x i16> undef)
%res6sext = sext <4 x i16> %res6 to <4 x i32>
%res6trunc = trunc <4 x i32> %res6sext to <4 x i16>
call void @llvm.masked.scatter.v4i32.v4p0i32(<4 x i16> %res6trunc, <4 x i32*> %gep5.2, i32 4, <4 x i1> %mask)

If I haven't got that wrong, those scatters should be cheap(ish). It's probably worth adding similar v4i8 and v8i8 tests in too.

[samparker]

llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp

910

Could we just return the ScalarCost if this isn't true?

921

Do we really need to check the uses of a store?

[anwel]

llvm/include/llvm/Analysis/TargetTransformInfo.h

968–970

Yes, thanks for the reminder.

1016–1017

I have a slight aversion against multiple default arguments in one function, as they might lead to confusion.
I've checked though that C++ warns in this case if no VF but an instruction is passed, so as it reduces the amount of necessary changes I'll take your advice and add a default value.

llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp

921

No we don't :)

[anwel]

Implemented some of the changes suggested in comments.

Added the new instruction parameter to the getIntrinsicInstrCost version that only gets types, too, so the functions stay roughly in-sync and templates putting T instead of Value/Type are unaffected.

[anwel]

Moved around some tests and add two more.
Also made sure that extends or truncs to 'incomplete' vector types are not considered valid candidates to be merged into the gather.

[anwel]

llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp

910

Not always. A v8i16 gather that for some reason is extended to v8i32 would be checked here, too, and fail - although it's perfectly fine to build that gather

llvm/test/Analysis/CostModel/ARM/mve-gather-scatter-cost.ll

177–187

Have been moved. I moved those for v8i8, too, and created a set for v4i8.

[dmgreen]

LGTM

llvm/include/llvm/Analysis/TargetTransformInfo.h

1025

Add an I doxygen comment again?

[anwel]

llvm/include/llvm/Analysis/TargetTransformInfo.h

1025

Yes of course ;)

llvm/include/llvm/Analysis/TargetTransformInfo.h

int getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
                             unsigned AddressSpace) const;
 
   /// \return The cost of Gather or Scatter operation
   /// \p Opcode - is a type of memory access Load or Store
   /// \p DataTy - a vector type of the data to be loaded or stored
   /// \p Ptr - pointer [or vector of pointers] - address[es] in memory
   /// \p VariableMask - true when the memory access is predicated with a mask
   ///                   that is not a compile-time constant
   /// \p Alignment - alignment of single element
+  /// \p I - the optional original context instruction, if one exists, e.g. the
+  ///        load/store to transform or the call to the gather/scatter intrinsic
   int getGatherScatterOpCost(unsigned Opcode, Type *DataTy, Value *Ptr,
-                             bool VariableMask, unsigned Alignment) const;
+                             bool VariableMask, unsigned Alignment,
+                             const Instruction *I = nullptr) const;
 
   /// \return The cost of the interleaved memory operation.
   /// \p Opcode is the memory operation code
   /// \p VecTy is the vector type of the interleaved access.
   /// \p Factor is the interleave factor
   /// \p Indices is the indices for interleaved load members (as interleaved
   ///    load allows gaps)
   /// \p Alignment is the alignment of the memory operation
public:
   ///  ((v0+v2), (v1+v3), undef, undef)
   int getArithmeticReductionCost(unsigned Opcode, Type *Ty,
                                  bool IsPairwiseForm) const;
   int getMinMaxReductionCost(Type *Ty, Type *CondTy, bool IsPairwiseForm,
                              bool IsUnsigned) const;
 
   /// \returns The cost of Intrinsic instructions. Analyses the real arguments.
   /// Three cases are handled: 1. scalar instruction 2. vector instruction
   /// 3. scalar instruction which is to be vectorized with VF.
+  /// I is the optional original context instruction holding the call to the
+  /// intrinsic
   int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
                             ArrayRef<Value *> Args, FastMathFlags FMF,
-                            unsigned VF = 1) const;
+                            unsigned VF = 1,
+                            const Instruction *I = nullptr) const;
 
   /// \returns The cost of Intrinsic instructions. Types analysis only.
   /// If ScalarizationCostPassed is UINT_MAX, the cost of scalarizing the
   /// arguments and the return value will be computed based on types.
-  int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
-                            ArrayRef<Type *> Tys, FastMathFlags FMF,
-                            unsigned ScalarizationCostPassed = UINT_MAX) const;
+  /// I is the optional original context instruction holding the call to the
+  /// intrinsic
+  int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, ArrayRef<Type *> Tys,
+                            FastMathFlags FMF,
+                            unsigned ScalarizationCostPassed = UINT_MAX,
+                            const Instruction *I = nullptr) const;
 
   /// \returns The cost of Call instructions.
   int getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) const;
 
   /// \returns The number of pieces into which the provided type must be
   /// split during legalization. Zero is returned when the answer is unknown.
   unsigned getNumberOfParts(Type *Tp) const;
 
virtual int getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
                                 Type *CondTy, const Instruction *I) = 0;
   virtual int getVectorInstrCost(unsigned Opcode, Type *Val,
                                  unsigned Index) = 0;
   virtual int getMemoryOpCost(unsigned Opcode, Type *Src, MaybeAlign Alignment,
                               unsigned AddressSpace, const Instruction *I) = 0;
   virtual int getMaskedMemoryOpCost(unsigned Opcode, Type *Src,
                                     unsigned Alignment,
                                     unsigned AddressSpace) = 0;
-  virtual int getGatherScatterOpCost(unsigned Opcode, Type *DataTy,
-                                     Value *Ptr, bool VariableMask,
-                                     unsigned Alignment) = 0;
+  virtual int getGatherScatterOpCost(unsigned Opcode, Type *DataTy, Value *Ptr,
+                                     bool VariableMask, unsigned Alignment,
+                                     const Instruction *I = nullptr) = 0;
   virtual int getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
                                          unsigned Factor,
                                          ArrayRef<unsigned> Indices,
                                          unsigned Alignment,
                                          unsigned AddressSpace,
                                          bool UseMaskForCond = false,
                                          bool UseMaskForGaps = false) = 0;
   virtual int getArithmeticReductionCost(unsigned Opcode, Type *Ty,
                                          bool IsPairwiseForm) = 0;
   virtual int getMinMaxReductionCost(Type *Ty, Type *CondTy,
                                      bool IsPairwiseForm, bool IsUnsigned) = 0;
   virtual int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
                                     ArrayRef<Type *> Tys, FastMathFlags FMF,
-                      unsigned ScalarizationCostPassed) = 0;
+                                    unsigned ScalarizationCostPassed,
+                                    const Instruction *I) = 0;
   virtual int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
-         ArrayRef<Value *> Args, FastMathFlags FMF, unsigned VF) = 0;
+                                    ArrayRef<Value *> Args, FastMathFlags FMF,
+                                    unsigned VF, const Instruction *I) = 0;
   virtual int getCallInstrCost(Function *F, Type *RetTy,
                                ArrayRef<Type *> Tys) = 0;
   virtual unsigned getNumberOfParts(Type *Tp) = 0;
   virtual int getAddressComputationCost(Type *Ty, ScalarEvolution *SE,
                                         const SCEV *Ptr) = 0;
   virtual unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) = 0;
   virtual bool getTgtMemIntrinsic(IntrinsicInst *Inst,
                                   MemIntrinsicInfo &Info) = 0;
public:
   int getMemoryOpCost(unsigned Opcode, Type *Src, MaybeAlign Alignment,
                       unsigned AddressSpace, const Instruction *I) override {
     return Impl.getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, I);
   }
   int getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
                             unsigned AddressSpace) override {
     return Impl.getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
   }
-  int getGatherScatterOpCost(unsigned Opcode, Type *DataTy,
-                             Value *Ptr, bool VariableMask,
-                             unsigned Alignment) override {
+  int getGatherScatterOpCost(unsigned Opcode, Type *DataTy, Value *Ptr,
+                             bool VariableMask, unsigned Alignment,
+                             const Instruction *I = nullptr) override {
     return Impl.getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask,
-                                       Alignment);
+                                       Alignment, I);
   }
   int getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy, unsigned Factor,
                                  ArrayRef<unsigned> Indices, unsigned Alignment,
                                  unsigned AddressSpace, bool UseMaskForCond,
                                  bool UseMaskForGaps) override {
     return Impl.getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
                                            Alignment, AddressSpace,
                                            UseMaskForCond, UseMaskForGaps);
   }
   int getArithmeticReductionCost(unsigned Opcode, Type *Ty,
                                  bool IsPairwiseForm) override {
     return Impl.getArithmeticReductionCost(Opcode, Ty, IsPairwiseForm);
   }
   int getMinMaxReductionCost(Type *Ty, Type *CondTy,
                              bool IsPairwiseForm, bool IsUnsigned) override {
     return Impl.getMinMaxReductionCost(Ty, CondTy, IsPairwiseForm, IsUnsigned);
    }
-  int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, ArrayRef<Type *> Tys,
-               FastMathFlags FMF, unsigned ScalarizationCostPassed) override {
+   int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
+                             ArrayRef<Type *> Tys, FastMathFlags FMF,
+                             unsigned ScalarizationCostPassed,
+                             const Instruction *I) override {
      return Impl.getIntrinsicInstrCost(ID, RetTy, Tys, FMF,
-                                      ScalarizationCostPassed);
+                                       ScalarizationCostPassed, I);
    }
    int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
-       ArrayRef<Value *> Args, FastMathFlags FMF, unsigned VF) override {
-    return Impl.getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF);
+                             ArrayRef<Value *> Args, FastMathFlags FMF,
+                             unsigned VF, const Instruction *I) override {
+     return Impl.getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF, I);
    }
   int getCallInstrCost(Function *F, Type *RetTy,
                        ArrayRef<Type *> Tys) override {
     return Impl.getCallInstrCost(F, RetTy, Tys);
   }
   unsigned getNumberOfParts(Type *Tp) override {
     return Impl.getNumberOfParts(Tp);
   }
   int getAddressComputationCost(Type *Ty, ScalarEvolution *SE,

llvm/include/llvm/Analysis/TargetTransformInfoImpl.h

public:
   }
 
   unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
                                  unsigned AddressSpace) {
     return 1;
   }
 
   unsigned getGatherScatterOpCost(unsigned Opcode, Type *DataTy, Value *Ptr,
-                                  bool VariableMask,
-                                  unsigned Alignment) {
+                                  bool VariableMask, unsigned Alignment,
+                                  const Instruction *I = nullptr) {
     return 1;
   }
 
   unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
                                       unsigned Factor,
                                       ArrayRef<unsigned> Indices,
                                       unsigned Alignment, unsigned AddressSpace,
                                       bool UseMaskForCond = false,
                                       bool UseMaskForGaps = false) {
     return 1;
   }
 
   unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
                                  ArrayRef<Type *> Tys, FastMathFlags FMF,
-                                 unsigned ScalarizationCostPassed) {
+                                 unsigned ScalarizationCostPassed,
+                                 const Instruction *I) {
     return 1;
   }
   unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
-            ArrayRef<Value *> Args, FastMathFlags FMF, unsigned VF) {
+                                 ArrayRef<Value *> Args, FastMathFlags FMF,
+                                 unsigned VF, const Instruction *I) {
     return 1;
   }
 
   unsigned getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) {
     return 1;
   }
 
   unsigned getNumberOfParts(Type *Tp) { return 0; }

llvm/include/llvm/CodeGen/BasicTTIImpl.h

if (UseMaskForGaps)
           BinaryOperator::And, MaskVT);
 
     return Cost;
   }
 
   /// Get intrinsic cost based on arguments.
   unsigned getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
                                  ArrayRef<Value *> Args, FastMathFlags FMF,
-                                 unsigned VF = 1) {
+                                 unsigned VF = 1,
+                                 const Instruction *I = nullptr) {
     unsigned RetVF = (RetTy->isVectorTy() ? RetTy->getVectorNumElements() : 1);
     assert((RetVF == 1 || VF == 1) && "VF > 1 and RetVF is a vector type");
     auto *ConcreteTTI = static_cast<T *>(this);
 
     switch (IID) {
     default: {
       // Assume that we need to scalarize this intrinsic.
       SmallVector<Type *, 4> Types;
default: {
       return ConcreteTTI->getIntrinsicInstrCost(IID, RetTy, Types, FMF,
                                                 ScalarizationCost);
     }
     case Intrinsic::masked_scatter: {
       assert(VF == 1 && "Can't vectorize types here.");
       Value *Mask = Args[3];
       bool VarMask = !isa<Constant>(Mask);
       unsigned Alignment = cast<ConstantInt>(Args[2])->getZExtValue();
-      return ConcreteTTI->getGatherScatterOpCost(
-          Instruction::Store, Args[0]->getType(), Args[1], VarMask, Alignment);
+      return ConcreteTTI->getGatherScatterOpCost(Instruction::Store,
+                                                 Args[0]->getType(), Args[1],
+                                                 VarMask, Alignment, I);
     }
     case Intrinsic::masked_gather: {
       assert(VF == 1 && "Can't vectorize types here.");
       Value *Mask = Args[2];
       bool VarMask = !isa<Constant>(Mask);
       unsigned Alignment = cast<ConstantInt>(Args[1])->getZExtValue();
-      return ConcreteTTI->getGatherScatterOpCost(Instruction::Load, RetTy,
-                                                 Args[0], VarMask, Alignment);
+      return ConcreteTTI->getGatherScatterOpCost(
+          Instruction::Load, RetTy, Args[0], VarMask, Alignment, I);
     }
     case Intrinsic::experimental_vector_reduce_add:
     case Intrinsic::experimental_vector_reduce_mul:
     case Intrinsic::experimental_vector_reduce_and:
     case Intrinsic::experimental_vector_reduce_or:
     case Intrinsic::experimental_vector_reduce_xor:
     case Intrinsic::experimental_vector_reduce_v2_fadd:
     case Intrinsic::experimental_vector_reduce_v2_fmul:
public:
   }
 
   /// Get intrinsic cost based on argument types.
   /// If ScalarizationCostPassed is std::numeric_limits<unsigned>::max(), the
   /// cost of scalarizing the arguments and the return value will be computed
   /// based on types.
   unsigned getIntrinsicInstrCost(
       Intrinsic::ID IID, Type *RetTy, ArrayRef<Type *> Tys, FastMathFlags FMF,
-      unsigned ScalarizationCostPassed = std::numeric_limits<unsigned>::max()) {
+      unsigned ScalarizationCostPassed = std::numeric_limits<unsigned>::max(),
+      const Instruction *I = nullptr) {
     auto *ConcreteTTI = static_cast<T *>(this);
 
     SmallVector<unsigned, 2> ISDs;
     unsigned SingleCallCost = 10; // Library call cost. Make it expensive.
     switch (IID) {
     default: {
       // Assume that we need to scalarize this intrinsic.
       unsigned ScalarizationCost = ScalarizationCostPassed;

llvm/lib/Analysis/TargetTransformInfo.cpp

int TargetTransformInfo::getMaskedMemoryOpCost(unsigned Opcode, Type *Src,
   int Cost =
       TTIImpl->getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
   assert(Cost >= 0 && "TTI should not produce negative costs!");
   return Cost;
 }
 
 int TargetTransformInfo::getGatherScatterOpCost(unsigned Opcode, Type *DataTy,
                                                 Value *Ptr, bool VariableMask,
-                                                unsigned Alignment) const {
+                                                unsigned Alignment,
+                                                const Instruction *I) const {
   int Cost = TTIImpl->getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask,
-                                             Alignment);
+                                             Alignment, I);
   assert(Cost >= 0 && "TTI should not produce negative costs!");
   return Cost;
 }
 
 int TargetTransformInfo::getInterleavedMemoryOpCost(
     unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices,
     unsigned Alignment, unsigned AddressSpace, bool UseMaskForCond,
     bool UseMaskForGaps) const {
   int Cost = TTIImpl->getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
                                                  Alignment, AddressSpace,
                                                  UseMaskForCond,
                                                  UseMaskForGaps);
   assert(Cost >= 0 && "TTI should not produce negative costs!");
   return Cost;
 }
 
 int TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
-                                    ArrayRef<Type *> Tys, FastMathFlags FMF,
-                                    unsigned ScalarizationCostPassed) const {
+                                               ArrayRef<Type *> Tys,
+                                               FastMathFlags FMF,
+                                               unsigned ScalarizationCostPassed,
+                                               const Instruction *I) const {
   int Cost = TTIImpl->getIntrinsicInstrCost(ID, RetTy, Tys, FMF,
-                                            ScalarizationCostPassed);
+                                            ScalarizationCostPassed, I);
   assert(Cost >= 0 && "TTI should not produce negative costs!");
   return Cost;
 }
 
 int TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
-           ArrayRef<Value *> Args, FastMathFlags FMF, unsigned VF) const {
-  int Cost = TTIImpl->getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF);
+                                               ArrayRef<Value *> Args,
+                                               FastMathFlags FMF, unsigned VF,
+                                               const Instruction *I) const {
+  int Cost = TTIImpl->getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF, I);
   assert(Cost >= 0 && "TTI should not produce negative costs!");
   return Cost;
 }
 
 int TargetTransformInfo::getCallInstrCost(Function *F, Type *RetTy,
                                           ArrayRef<Type *> Tys) const {
   int Cost = TTIImpl->getCallInstrCost(F, RetTy, Tys);
   assert(Cost >= 0 && "TTI should not produce negative costs!");
int TargetTransformInfo::getInstructionThroughput(const Instruction *I) const {
   case Instruction::Call:
     if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
       SmallVector<Value *, 4> Args(II->arg_operands());
 
       FastMathFlags FMF;
       if (auto *FPMO = dyn_cast<FPMathOperator>(II))
         FMF = FPMO->getFastMathFlags();
 
-      return getIntrinsicInstrCost(II->getIntrinsicID(), II->getType(),
-                                        Args, FMF);
+      return getIntrinsicInstrCost(II->getIntrinsicID(), II->getType(), Args,
+                                   FMF, 1, II);
     }
     return -1;
   default:
     // We don't have any information on this instruction.
     return -1;
   }
 }
 

llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.h

public:
   unsigned getInliningThresholdMultiplier() { return 11; }
 
   int getInlinerVectorBonusPercent() { return 0; }
 
   int getArithmeticReductionCost(unsigned Opcode,
                                  Type *Ty,
                                  bool IsPairwise);
   template <typename T>
-  int getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
-                            ArrayRef<T *> Args, FastMathFlags FMF,
-                            unsigned VF);
+  int getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy, ArrayRef<T *> Args,
+                            FastMathFlags FMF, unsigned VF,
+                            const Instruction *I = nullptr);
   int getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
                             ArrayRef<Type *> Tys, FastMathFlags FMF,
-                            unsigned ScalarizationCostPassed = UINT_MAX);
+                            unsigned ScalarizationCostPassed = UINT_MAX,
+                            const Instruction *I = nullptr);
   int getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
                             ArrayRef<Value *> Args, FastMathFlags FMF,
-                            unsigned VF = 1);
+                            unsigned VF = 1, const Instruction *I = nullptr);
   int getMinMaxReductionCost(Type *Ty, Type *CondTy,
                              bool IsPairwiseForm,
                              bool IsUnsigned);
   unsigned getUserCost(const User *U, ArrayRef<const Value *> Operands);
 };
 
 class R600TTIImpl final : public BasicTTIImplBase<R600TTIImpl> {
   using BaseT = BasicTTIImplBase<R600TTIImpl>;

llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp

int GCNTTIImpl::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
   }
 
   return BaseT::getArithmeticInstrCost(Opcode, Ty, Opd1Info, Opd2Info,
                                        Opd1PropInfo, Opd2PropInfo);
 }
 
 template <typename T>
 int GCNTTIImpl::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
-                                      ArrayRef<T *> Args,
-                                      FastMathFlags FMF, unsigned VF) {
+                                      ArrayRef<T *> Args, FastMathFlags FMF,
+                                      unsigned VF, const Instruction *I) {
   if (ID != Intrinsic::fma)
-    return BaseT::getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF);
+    return BaseT::getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF, I);
 
   EVT OrigTy = TLI->getValueType(DL, RetTy);
   if (!OrigTy.isSimple()) {
-    return BaseT::getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF);
+    return BaseT::getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF, I);
   }
 
   // Legalize the type.
   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, RetTy);
 
   unsigned NElts = LT.second.isVector() ?
     LT.second.getVectorNumElements() : 1;
 
   MVT::SimpleValueType SLT = LT.second.getScalarType().SimpleTy;
 
   if (SLT == MVT::f64)
     return LT.first * NElts * get64BitInstrCost();
 
   if (ST->has16BitInsts() && SLT == MVT::f16)
     NElts = (NElts + 1) / 2;
 
   return LT.first * NElts * (ST->hasFastFMAF32() ? getHalfRateInstrCost()
                                                  : getQuarterRateInstrCost());
 }
 
 int GCNTTIImpl::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
-                                      ArrayRef<Value*> Args, FastMathFlags FMF,
-                                      unsigned VF) {
-  return getIntrinsicInstrCost<Value>(ID, RetTy, Args, FMF, VF);
+                                      ArrayRef<Value *> Args, FastMathFlags FMF,
+                                      unsigned VF, const Instruction *I) {
+  return getIntrinsicInstrCost<Value>(ID, RetTy, Args, FMF, VF, I);
 }
 
 int GCNTTIImpl::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
                                       ArrayRef<Type *> Tys, FastMathFlags FMF,
-                                      unsigned ScalarizationCostPassed) {
+                                      unsigned ScalarizationCostPassed,
+                                      const Instruction *I) {
   return getIntrinsicInstrCost<Type>(ID, RetTy, Tys, FMF,
-                                     ScalarizationCostPassed);
+                                     ScalarizationCostPassed, I);
 }
 
 unsigned GCNTTIImpl::getCFInstrCost(unsigned Opcode) {
   // XXX - For some reason this isn't called for switch.
   switch (Opcode) {
   case Instruction::Br:
   case Instruction::Ret:
     return 10;
unsigned GCNTTIImpl::getUserCost(const User *U,
   }
   case Instruction::Call: {
     if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
       SmallVector<Value *, 4> Args(II->arg_operands());
       FastMathFlags FMF;
       if (auto *FPMO = dyn_cast<FPMathOperator>(II))
         FMF = FPMO->getFastMathFlags();
       return getIntrinsicInstrCost(II->getIntrinsicID(), II->getType(), Args,
-                                   FMF);
+                                   FMF, 1, II);
     } else {
       return BaseT::getUserCost(U, Operands);
     }
   }
   case Instruction::ShuffleVector: {
     const ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
     Type *Ty = Shuffle->getType();
     Type *SrcTy = Shuffle->getOperand(0)->getType();

llvm/lib/Target/ARM/ARMTargetTransformInfo.h

public:
 
   int getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy, unsigned Factor,
                                  ArrayRef<unsigned> Indices, unsigned Alignment,
                                  unsigned AddressSpace,
                                  bool UseMaskForCond = false,
                                  bool UseMaskForGaps = false);
 
   unsigned getGatherScatterOpCost(unsigned Opcode, Type *DataTy, Value *Ptr,
-                                  bool VariableMask, unsigned Alignment);
+                                  bool VariableMask, unsigned Alignment,
+                                  const Instruction *I = nullptr);
 
   bool isLoweredToCall(const Function *F);
   bool isHardwareLoopProfitable(Loop *L, ScalarEvolution &SE,
                                 AssumptionCache &AC,
                                 TargetLibraryInfo *LibInfo,
                                 HardwareLoopInfo &HWLoopInfo);
   bool preferPredicateOverEpilogue(Loop *L, LoopInfo *LI,
                                    ScalarEvolution &SE,

llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp

int ARMTTIImpl::getInterleavedMemoryOpCost(
 
   return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
                                            Alignment, AddressSpace,
                                            UseMaskForCond, UseMaskForGaps);
 }
 
 unsigned ARMTTIImpl::getGatherScatterOpCost(unsigned Opcode, Type *DataTy,
                                             Value *Ptr, bool VariableMask,
-                                            unsigned Alignment) {
+                                            unsigned Alignment,
+                                            const Instruction *I) {
+  using namespace PatternMatch;
   if (!ST->hasMVEIntegerOps() || !EnableMaskedGatherScatters)
     return BaseT::getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask,
-                                         Alignment);
+                                         Alignment, I);
 
   assert(DataTy->isVectorTy() && "Can't do gather/scatters on scalar!");
   VectorType *VTy = cast<VectorType>(DataTy);
 
   // TODO: Splitting, once we do that.
-  // TODO: trunc/sext/zext the result/input
 
   unsigned NumElems = VTy->getNumElements();
   unsigned EltSize = VTy->getScalarSizeInBits();
   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, DataTy);
 
   // For now, it is assumed that for the MVE gather instructions the loads are
   // all effectively serialised. This means the cost is the scalar cost
   // multiplied by the number of elements being loaded. This is possibly very
   // conservative, but even so we still end up vectorising loops because the
   // cost per iteration for many loops is lower than for scalar loops.
   unsigned VectorCost = NumElems * LT.first;
   // The scalarization cost should be a lot higher. We use the number of vector
   // elements plus the scalarization overhead.
   unsigned ScalarCost =
       NumElems * LT.first + BaseT::getScalarizationOverhead(DataTy, {});
 
-  // TODO: Cost extended gathers or trunc stores correctly.
-  if (EltSize * NumElems != 128 || NumElems < 4)
-    return ScalarCost;
   if (Alignment < EltSize / 8)
     return ScalarCost;
 
+  unsigned ExtSize = EltSize;
+  // Check whether there's a single user that asks for an extended type
+  if (I != nullptr) {
+    // Dependent of the caller of this function, a gather instruction will
+    // either have opcode Instruction::Load or be a call to the masked_gather
+    // intrinsic
+    if ((I->getOpcode() == Instruction::Load ||
+         match(I, m_Intrinsic<Intrinsic::masked_gather>())) &&
+        I->hasOneUse()) {
+      const User *Us = *I->users().begin();
+      if (isa<ZExtInst>(Us) || isa<SExtInst>(Us)) {
+        // only allow valid type combinations
+        unsigned TypeSize =
+            cast<Instruction>(Us)->getType()->getScalarSizeInBits();
+        if (((TypeSize == 32 && (EltSize == 8 || EltSize == 16)) ||
+             (TypeSize == 16 && EltSize == 8)) &&
+            TypeSize * NumElems == 128) {
+          ExtSize = TypeSize;
+        }
+      }
+    }
+    // Check whether the input data needs to be truncated
+    TruncInst *T;
+    if ((I->getOpcode() == Instruction::Store ||
+         match(I, m_Intrinsic<Intrinsic::masked_scatter>())) &&
+        (T = dyn_cast<TruncInst>(I->getOperand(0)))) {
+      // Only allow valid type combinations
+      unsigned TypeSize = T->getOperand(0)->getType()->getScalarSizeInBits();
+      if (((EltSize == 16 && TypeSize == 32) ||
+           (EltSize == 8 && (TypeSize == 32 || TypeSize == 16))) &&
+          TypeSize * NumElems == 128)
+        ExtSize = TypeSize;
+    }
+  }
+
+  if (ExtSize * NumElems != 128 || NumElems < 4)
+    return ScalarCost;
+
   // Any (aligned) i32 gather will not need to be scalarised.
-  if (EltSize == 32)
+  if (ExtSize == 32)
     return VectorCost;
   // For smaller types, we need to ensure that the gep's inputs are correctly
-  // extended from a small enough value. Other size (including i64) are
+  // extended from a small enough value. Other sizes (including i64) are
   // scalarized for now.
-  if (EltSize != 8 && EltSize != 16)
+  if (ExtSize != 8 && ExtSize != 16)
     return ScalarCost;
 
   if (auto BC = dyn_cast<BitCastInst>(Ptr))
     Ptr = BC->getOperand(0);
   if (auto *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
     if (GEP->getNumOperands() != 2)
       return ScalarCost;
     unsigned Scale = DL.getTypeAllocSize(GEP->getResultElementType());
     // Scale needs to be correct (which is only relevant for i16s).
-    if (Scale != 1 && Scale * 8 != EltSize)
+    if (Scale != 1 && Scale * 8 != ExtSize)
       return ScalarCost;
     // And we need to zext (not sext) the indexes from a small enough type.
-    if (auto ZExt = dyn_cast<ZExtInst>(GEP->getOperand(1)))
-      if (ZExt->getOperand(0)->getType()->getScalarSizeInBits() <= EltSize)
+    if (auto ZExt = dyn_cast<ZExtInst>(GEP->getOperand(1))) {
+      if (ZExt->getOperand(0)->getType()->getScalarSizeInBits() <= ExtSize)
         return VectorCost;
+    }
     return ScalarCost;
   }
   return ScalarCost;
 }
 
 bool ARMTTIImpl::isLoweredToCall(const Function *F) {
   if (!F->isIntrinsic())
     BaseT::isLoweredToCall(F);

llvm/lib/Target/Hexagon/HexagonTargetTransformInfo.h

bool enableInterleavedAccessVectorization() {
     return true;
   }
 
   unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract);
   unsigned getOperandsScalarizationOverhead(ArrayRef<const Value*> Args,
             unsigned VF);
   unsigned getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type*> Tys);
   unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
-            ArrayRef<Value*> Args, FastMathFlags FMF, unsigned VF);
+                                 ArrayRef<Value *> Args, FastMathFlags FMF,
+                                 unsigned VF, const Instruction *I);
   unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
-            ArrayRef<Type*> Tys, FastMathFlags FMF,
-            unsigned ScalarizationCostPassed = UINT_MAX);
+                                 ArrayRef<Type *> Tys, FastMathFlags FMF,
+                                 unsigned ScalarizationCostPassed = UINT_MAX,
+                                 const Instruction *I = nullptr);
   unsigned getAddressComputationCost(Type *Tp, ScalarEvolution *SE,
             const SCEV *S);
   unsigned getMemoryOpCost(unsigned Opcode, Type *Src, MaybeAlign Alignment,
                            unsigned AddressSpace,
                            const Instruction *I = nullptr);
   unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
             unsigned AddressSpace);
   unsigned getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index,
             Type *SubTp);
   unsigned getGatherScatterOpCost(unsigned Opcode, Type *DataTy, Value *Ptr,
-            bool VariableMask, unsigned Alignment);
+                                  bool VariableMask, unsigned Alignment,
+                                  const Instruction *I);
   unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
             unsigned Factor, ArrayRef<unsigned> Indices, unsigned Alignment,
             unsigned AddressSpace, bool UseMaskForCond = false,
             bool UseMaskForGaps = false);
   unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
             const Instruction *I);
   unsigned getArithmeticInstrCost(
       unsigned Opcode, Type *Ty,

llvm/lib/Target/Hexagon/HexagonTargetTransformInfo.cpp

 }
 
 unsigned HexagonTTIImpl::getCallInstrCost(Function *F, Type *RetTy,
       ArrayRef<Type*> Tys) {
   return BaseT::getCallInstrCost(F, RetTy, Tys);
 }
 
 unsigned HexagonTTIImpl::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
-      ArrayRef<Value*> Args, FastMathFlags FMF, unsigned VF) {
-  return BaseT::getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF);
+                                               ArrayRef<Value *> Args,
+                                               FastMathFlags FMF, unsigned VF,
+                                               const Instruction *I) {
+  return BaseT::getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF, I);
 }
 
 unsigned HexagonTTIImpl::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
-      ArrayRef<Type*> Tys, FastMathFlags FMF,
-      unsigned ScalarizationCostPassed) {
+                                               ArrayRef<Type *> Tys,
+                                               FastMathFlags FMF,
+                                               unsigned ScalarizationCostPassed,
+                                               const Instruction *I) {
   if (ID == Intrinsic::bswap) {
     std::pair<int, MVT> LT = TLI.getTypeLegalizationCost(DL, RetTy);
     return LT.first + 2;
   }
   return BaseT::getIntrinsicInstrCost(ID, RetTy, Tys, FMF,
-                                      ScalarizationCostPassed);
+                                      ScalarizationCostPassed, I);
 }
 
 unsigned HexagonTTIImpl::getAddressComputationCost(Type *Tp,
       ScalarEvolution *SE, const SCEV *S) {
   return 0;
 }
 
 unsigned HexagonTTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src,

 }
 
 unsigned HexagonTTIImpl::getShuffleCost(TTI::ShuffleKind Kind, Type *Tp,
       int Index, Type *SubTp) {
   return 1;
 }
 
 unsigned HexagonTTIImpl::getGatherScatterOpCost(unsigned Opcode, Type *DataTy,
-      Value *Ptr, bool VariableMask, unsigned Alignment) {
+                                                Value *Ptr, bool VariableMask,
+                                                unsigned Alignment,
+                                                const Instruction *I) {
   return BaseT::getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask,
-                                       Alignment);
+                                       Alignment, I);
 }
 
 unsigned HexagonTTIImpl::getInterleavedMemoryOpCost(unsigned Opcode,
       Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices,
       unsigned Alignment, unsigned AddressSpace, bool UseMaskForCond,
       bool UseMaskForGaps) {
   if (Indices.size() != Factor || UseMaskForCond || UseMaskForGaps)
     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,

llvm/lib/Target/PowerPC/PPCTargetTransformInfo.h

public:
   int getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
                                  unsigned Factor,
                                  ArrayRef<unsigned> Indices,
                                  unsigned Alignment,
                                  unsigned AddressSpace,
                                  bool UseMaskForCond = false,
                                  bool UseMaskForGaps = false);
   unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
-            ArrayRef<Value*> Args, FastMathFlags FMF, unsigned VF);
+                                 ArrayRef<Value *> Args, FastMathFlags FMF,
+                                 unsigned VF, const Instruction *I = nullptr);
   unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
-            ArrayRef<Type*> Tys, FastMathFlags FMF,
-            unsigned ScalarizationCostPassed = UINT_MAX);
+                                 ArrayRef<Type *> Tys, FastMathFlags FMF,
+                                 unsigned ScalarizationCostPassed = UINT_MAX,
+                                 const Instruction *I = nullptr);
 
   /// @}
 };
 
 } // end namespace llvm
 
 #endif

llvm/lib/Target/PowerPC/PPCTargetTransformInfo.cpp

int PPCTTIImpl::getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
   // vector (except that the first shuffle can take two incoming vectors
   // because it does not need to take itself).
   Cost += Factor*(LT.first-1);
 
   return Cost;
 }
 
 unsigned PPCTTIImpl::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
-      ArrayRef<Value*> Args, FastMathFlags FMF, unsigned VF) {
-  return BaseT::getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF);
+                                           ArrayRef<Value *> Args,
+                                           FastMathFlags FMF, unsigned VF,
+                                           const Instruction *I) {
+  return BaseT::getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF, I);
 }
 
 unsigned PPCTTIImpl::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
-      ArrayRef<Type*> Tys, FastMathFlags FMF,
-      unsigned ScalarizationCostPassed) {
+                                           ArrayRef<Type *> Tys,
+                                           FastMathFlags FMF,
+                                           unsigned ScalarizationCostPassed,
+                                           const Instruction *I) {
   if (ID == Intrinsic::bswap && ST->hasP9Vector())
     return TLI->getTypeLegalizationCost(DL, RetTy).first;
   return BaseT::getIntrinsicInstrCost(ID, RetTy, Tys, FMF,
-                                      ScalarizationCostPassed);
+                                      ScalarizationCostPassed, I);
 }
 
 bool PPCTTIImpl::canSaveCmp(Loop *L, BranchInst **BI, ScalarEvolution *SE,
                             LoopInfo *LI, DominatorTree *DT,
                             AssumptionCache *AC, TargetLibraryInfo *LibInfo) {
   // Process nested loops first.
   for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
     if (canSaveCmp(*I, BI, SE, LI, DT, AC, LibInfo))

llvm/lib/Target/SystemZ/SystemZTargetTransformInfo.h

int getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
                                  ArrayRef<unsigned> Indices,
                                  unsigned Alignment,
                                  unsigned AddressSpace,
                                  bool UseMaskForCond = false,
                                  bool UseMaskForGaps = false);
 
   int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
                             ArrayRef<Value *> Args, FastMathFlags FMF,
-                            unsigned VF = 1);
-  int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
-                            ArrayRef<Type *> Tys, FastMathFlags FMF,
-                            unsigned ScalarizationCostPassed = UINT_MAX);
+                            unsigned VF = 1, const Instruction *I = nullptr);
+  int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, ArrayRef<Type *> Tys,
+                            FastMathFlags FMF,
+                            unsigned ScalarizationCostPassed = UINT_MAX,
+                            const Instruction *I = nullptr);
   /// @}
 };
 
 } // end namespace llvm
 
 #endif

llvm/lib/Target/SystemZ/SystemZTargetTransformInfo.cpp

 static int getVectorIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy) {
   if (RetTy->isVectorTy() && ID == Intrinsic::bswap)
     return getNumVectorRegs(RetTy); // VPERM
   return -1;
 }
 
 int SystemZTTIImpl::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
                                           ArrayRef<Value *> Args,
-                                          FastMathFlags FMF, unsigned VF) {
+                                          FastMathFlags FMF, unsigned VF,
+                                          const Instruction *I) {
   int Cost = getVectorIntrinsicInstrCost(ID, RetTy);
   if (Cost != -1)
     return Cost;
-  return BaseT::getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF);
+  return BaseT::getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF, I);
 }
 
 int SystemZTTIImpl::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
                                           ArrayRef<Type *> Tys,
                                           FastMathFlags FMF,
-                                          unsigned ScalarizationCostPassed) {
+                                          unsigned ScalarizationCostPassed,
+                                          const Instruction *I) {
   int Cost = getVectorIntrinsicInstrCost(ID, RetTy);
   if (Cost != -1)
     return Cost;
-  return BaseT::getIntrinsicInstrCost(ID, RetTy, Tys,
-                                      FMF, ScalarizationCostPassed);
+  return BaseT::getIntrinsicInstrCost(ID, RetTy, Tys, FMF,
+                                      ScalarizationCostPassed, I);
 }

llvm/lib/Target/X86/X86TargetTransformInfo.h

public:
   int getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
                          const Instruction *I = nullptr);
   int getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index);
   int getMemoryOpCost(unsigned Opcode, Type *Src, MaybeAlign Alignment,
                       unsigned AddressSpace, const Instruction *I = nullptr);
   int getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
                             unsigned AddressSpace);
   int getGatherScatterOpCost(unsigned Opcode, Type *DataTy, Value *Ptr,
-                             bool VariableMask, unsigned Alignment);
+                             bool VariableMask, unsigned Alignment,
+                             const Instruction *I);
   int getAddressComputationCost(Type *PtrTy, ScalarEvolution *SE,
                                 const SCEV *Ptr);
 
   unsigned getAtomicMemIntrinsicMaxElementSize() const;
 
   int getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
                             ArrayRef<Type *> Tys, FastMathFlags FMF,
-                            unsigned ScalarizationCostPassed = UINT_MAX);
+                            unsigned ScalarizationCostPassed = UINT_MAX,
+                            const Instruction *I = nullptr);
   int getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
                             ArrayRef<Value *> Args, FastMathFlags FMF,
-                            unsigned VF = 1);
+                            unsigned VF = 1, const Instruction *I = nullptr);
 
   int getArithmeticReductionCost(unsigned Opcode, Type *Ty,
                                  bool IsPairwiseForm);
 
   int getMinMaxReductionCost(Type *Ty, Type *CondTy, bool IsPairwiseForm,
                              bool IsUnsigned);
 
   int getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,

llvm/lib/Target/X86/X86TargetTransformInfo.cpp

int X86TTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
 
   return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, I);
 }
 
 unsigned X86TTIImpl::getAtomicMemIntrinsicMaxElementSize() const { return 16; }
 
 int X86TTIImpl::getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
                                       ArrayRef<Type *> Tys, FastMathFlags FMF,
-                                      unsigned ScalarizationCostPassed) {
+                                      unsigned ScalarizationCostPassed,
+                                      const Instruction *I) {
   // Costs should match the codegen from:
   // BITREVERSE: llvm\test\CodeGen\X86\vector-bitreverse.ll
   // BSWAP: llvm\test\CodeGen\X86\bswap-vector.ll
   // CTLZ: llvm\test\CodeGen\X86\vector-lzcnt-*.ll
   // CTPOP: llvm\test\CodeGen\X86\vector-popcnt-*.ll
   // CTTZ: llvm\test\CodeGen\X86\vector-tzcnt-*.ll
   static const CostTblEntry AVX512CDCostTbl[] = {
     { ISD::CTLZ,       MVT::v8i64,   1 },
if (ISD != ISD::DELETED_NODE) {
     if (ST->is64Bit())
       if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy))
         return LT.first * Entry->Cost;
 
     if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy))
       return LT.first * Entry->Cost;
   }
 
-  return BaseT::getIntrinsicInstrCost(IID, RetTy, Tys, FMF, ScalarizationCostPassed);
+  return BaseT::getIntrinsicInstrCost(IID, RetTy, Tys, FMF,
+                                      ScalarizationCostPassed, I);
 }
 
 int X86TTIImpl::getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
                                       ArrayRef<Value *> Args, FastMathFlags FMF,
-                                      unsigned VF) {
+                                      unsigned VF, const Instruction *I) {
   static const CostTblEntry AVX512CostTbl[] = {
     { ISD::ROTL,       MVT::v8i64,   1 },
     { ISD::ROTL,       MVT::v4i64,   1 },
     { ISD::ROTL,       MVT::v2i64,   1 },
     { ISD::ROTL,       MVT::v16i32,  1 },
     { ISD::ROTL,       MVT::v8i32,   1 },
     { ISD::ROTL,       MVT::v4i32,   1 },
     { ISD::ROTR,       MVT::v8i64,   1 },
if (ISD != ISD::DELETED_NODE) {
     if (ST->is64Bit())
       if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy))
         return LT.first * Entry->Cost;
 
     if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy))
       return LT.first * Entry->Cost;
   }
 
-  return BaseT::getIntrinsicInstrCost(IID, RetTy, Args, FMF, VF);
+  return BaseT::getIntrinsicInstrCost(IID, RetTy, Args, FMF, VF, I);
 }
 
 int X86TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
   static const CostTblEntry SLMCostTbl[] = {
      { ISD::EXTRACT_VECTOR_ELT,       MVT::i8,      4 },
      { ISD::EXTRACT_VECTOR_ELT,       MVT::i16,     4 },
      { ISD::EXTRACT_VECTOR_ELT,       MVT::i32,     4 },
      { ISD::EXTRACT_VECTOR_ELT,       MVT::i64,     7 }
for (unsigned i = 0; i < VF; ++i)
         getVectorInstrCost(Instruction::ExtractElement, SrcVTy, i);
 
   return MemoryOpCost + MaskUnpackCost + InsertExtractCost;
 }
 
 /// Calculate the cost of Gather / Scatter operation
 int X86TTIImpl::getGatherScatterOpCost(unsigned Opcode, Type *SrcVTy,
                                        Value *Ptr, bool VariableMask,
-                                       unsigned Alignment) {
+                                       unsigned Alignment,
+                                       const Instruction *I = nullptr) {
   assert(SrcVTy->isVectorTy() && "Unexpected data type for Gather/Scatter");
   unsigned VF = SrcVTy->getVectorNumElements();
   PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
   if (!PtrTy && Ptr->getType()->isVectorTy())
     PtrTy = dyn_cast<PointerType>(Ptr->getType()->getVectorElementType());
   assert(PtrTy && "Unexpected type for Ptr argument");
   unsigned AddressSpace = PtrTy->getAddressSpace();
 

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

unsigned LoopVectorizationCostModel::getVectorIntrinsicCost(CallInst *CI,
   Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
   assert(ID && "Expected intrinsic call!");
 
   FastMathFlags FMF;
   if (auto *FPMO = dyn_cast<FPMathOperator>(CI))
     FMF = FPMO->getFastMathFlags();
 
   SmallVector<Value *, 4> Operands(CI->arg_operands());
-  return TTI.getIntrinsicInstrCost(ID, CI->getType(), Operands, FMF, VF);
+  return TTI.getIntrinsicInstrCost(ID, CI->getType(), Operands, FMF, VF, CI);
 }
 
 static Type *smallestIntegerVectorType(Type *T1, Type *T2) {
   auto *I1 = cast<IntegerType>(T1->getVectorElementType());
   auto *I2 = cast<IntegerType>(T2->getVectorElementType());
   return I1->getBitWidth() < I2->getBitWidth() ? T1 : T2;
 }
 static Type *largestIntegerVectorType(Type *T1, Type *T2) {
unsigned LoopVectorizationCostModel::getGatherScatterCost(Instruction *I,
   Type *ValTy = getMemInstValueType(I);
   Type *VectorTy = ToVectorTy(ValTy, VF);
   const MaybeAlign Alignment = getLoadStoreAlignment(I);
   Value *Ptr = getLoadStorePointerOperand(I);
 
   return TTI.getAddressComputationCost(VectorTy) +
          TTI.getGatherScatterOpCost(I->getOpcode(), VectorTy, Ptr,
                                     Legal->isMaskRequired(I),
-                                    Alignment ? Alignment->value() : 0);
+                                    Alignment ? Alignment->value() : 0, I);
 }
 
 unsigned LoopVectorizationCostModel::getInterleaveGroupCost(Instruction *I,
                                                             unsigned VF) {
   Type *ValTy = getMemInstValueType(I);
   Type *VectorTy = ToVectorTy(ValTy, VF);
   unsigned AS = getLoadStoreAddressSpace(I);
 

llvm/test/Analysis/CostModel/ARM/mve-gather-scatter-cost.ll

;
   ret i32 0
 }
 
 define void @gep_v4i32(i32* %base, i16* %base16, i8* %base8, <4 x i32> %ind32, <4 x i16> %ind16, <4 x i1> %mask)  {
 ; CHECK-LABEL: 'gep_v4i32'
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gep1 = getelementptr i32, i32* %base, <4 x i32> %ind32
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %res1 = call <4 x i32> @llvm.masked.gather.v4i32.v4p0i32(<4 x i32*> %gep1, i32 4, <4 x i1> %mask, <4 x i32> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: call void @llvm.masked.scatter.v4i32.v4p0i32(<4 x i32> %res1, <4 x i32*> %gep1, i32 4, <4 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %indzext = zext <4 x i16> %ind16 to <4 x i32>
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gep2 = getelementptr i32, i32* %base, <4 x i32> %indzext
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %res2 = call <4 x i32> @llvm.masked.gather.v4i32.v4p0i32(<4 x i32*> %gep2, i32 4, <4 x i1> %mask, <4 x i32> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: call void @llvm.masked.scatter.v4i32.v4p0i32(<4 x i32> %res2, <4 x i32*> %gep2, i32 4, <4 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %indsext = sext <4 x i16> %ind16 to <4 x i32>
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gep3 = getelementptr i32, i32* %base, <4 x i32> %indsext
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %res3 = call <4 x i32> @llvm.masked.gather.v4i32.v4p0i32(<4 x i32*> %gep3, i32 4, <4 x i1> %mask, <4 x i32> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: call void @llvm.masked.scatter.v4i32.v4p0i32(<4 x i32> %res3, <4 x i32*> %gep3, i32 4, <4 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gepu = getelementptr i32, i32* %base, <4 x i32> %ind32
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 36 for instruction: %resu = call <4 x i32> @llvm.masked.gather.v4i32.v4p0i32(<4 x i32*> %gepu, i32 1, <4 x i1> %mask, <4 x i32> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 36 for instruction: call void @llvm.masked.scatter.v4i32.v4p0i32(<4 x i32> %resu, <4 x i32*> %gepu, i32 1, <4 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gepos = getelementptr i8, i8* %base8, <4 x i32> %indzext
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %geposb = bitcast <4 x i8*> %gepos to <4 x i32*>
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %resos = call <4 x i32> @llvm.masked.gather.v4i32.v4p0i32(<4 x i32*> %geposb, i32 4, <4 x i1> %mask, <4 x i32> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: call void @llvm.masked.scatter.v4i32.v4p0i32(<4 x i32> %resos, <4 x i32*> %geposb, i32 4, <4 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gepbs = getelementptr i16, i16* %base16, <4 x i32> %indzext
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gepbsb = bitcast <4 x i16*> %gepbs to <4 x i32*>
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %resbs = call <4 x i32> @llvm.masked.gather.v4i32.v4p0i32(<4 x i32*> %gepbsb, i32 4, <4 x i1> %mask, <4 x i32> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: call void @llvm.masked.scatter.v4i32.v4p0i32(<4 x i32> %resbs, <4 x i32*> %gepbsb, i32 4, <4 x i1> %mask)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret void
 ;
   %gep1 = getelementptr i32, i32* %base, <4 x i32> %ind32
   %res1 = call <4 x i32> @llvm.masked.gather.v4i32.v4p0i32(<4 x i32*> %gep1, i32 4, <4 x i1> %mask, <4 x i32> undef)
   call void @llvm.masked.scatter.v4i32.v4p0i32(<4 x i32> %res1, <4 x i32*> %gep1, i32 4, <4 x i1> %mask)
 
   %indzext = zext <4 x i16> %ind16 to <4 x i32>
   %gep2 = getelementptr i32, i32* %base, <4 x i32> %indzext
   %res2 = call <4 x i32> @llvm.masked.gather.v4i32.v4p0i32(<4 x i32*> %gep2, i32 4, <4 x i1> %mask, <4 x i32> undef)
   call void @llvm.masked.scatter.v4i32.v4p0i32(<4 x i32> %res2, <4 x i32*> %gep2, i32 4, <4 x i1> %mask)
 
   %indsext = sext <4 x i16> %ind16 to <4 x i32>
   %gep3 = getelementptr i32, i32* %base, <4 x i32> %indsext
   %res3 = call <4 x i32> @llvm.masked.gather.v4i32.v4p0i32(<4 x i32*> %gep3, i32 4, <4 x i1> %mask, <4 x i32> undef)
   call void @llvm.masked.scatter.v4i32.v4p0i32(<4 x i32> %res3, <4 x i32*> %gep3, i32 4, <4 x i1> %mask)
 
   ; unaligned
   %gepu = getelementptr i32, i32* %base, <4 x i32> %ind32
   %resu = call <4 x i32> @llvm.masked.gather.v4i32.v4p0i32(<4 x i32*> %gepu, i32 1, <4 x i1> %mask, <4 x i32> undef)
   call void @llvm.masked.scatter.v4i32.v4p0i32(<4 x i32> %resu, <4 x i32*> %gepu, i32 1, <4 x i1> %mask)
 
   ; 1 scale
   %gepos = getelementptr i8, i8* %base8, <4 x i32> %indzext
   %geposb = bitcast <4 x i8*> %gepos to <4 x i32*>
   %resos = call <4 x i32> @llvm.masked.gather.v4i32.v4p0i32(<4 x i32*> %geposb, i32 4, <4 x i1> %mask, <4 x i32> undef)
   call void @llvm.masked.scatter.v4i32.v4p0i32(<4 x i32> %resos, <4 x i32*> %geposb, i32 4, <4 x i1> %mask)
 
   ; bad scale (but doesn't really matter because i32)
   %gepbs = getelementptr i16, i16* %base16, <4 x i32> %indzext
   %gepbsb = bitcast <4 x i16*> %gepbs to <4 x i32*>
   %resbs = call <4 x i32> @llvm.masked.gather.v4i32.v4p0i32(<4 x i32*> %gepbsb, i32 4, <4 x i1> %mask, <4 x i32> undef)
   call void @llvm.masked.scatter.v4i32.v4p0i32(<4 x i32> %resbs, <4 x i32*> %gepbsb, i32 4, <4 x i1> %mask)
   ret void
 }
 
 define void @gep_v4f32(float* %base, i16* %base16, i8* %base8, <4 x i32> %ind32, <4 x i16> %ind16, <4 x i1> %mask)  {
 ; CHECK-LABEL: 'gep_v4f32'
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %gep1 = getelementptr float, float* %base, <4 x i32> %ind32
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %res1 = call <4 x float> @llvm.masked.gather.v4f32.v4p0f32(<4 x float*> %gep1, i32 4, <4 x i1> %mask, <4 x float> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: call void @llvm.masked.scatter.v4f32.v4p0f32(<4 x float> %res1, <4 x float*> %gep1, i32 4, <4 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %indzext = zext <4 x i16> %ind16 to <4 x i32>
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %gep2 = getelementptr float, float* %base, <4 x i32> %indzext
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %res2 = call <4 x float> @llvm.masked.gather.v4f32.v4p0f32(<4 x float*> %gep2, i32 4, <4 x i1> %mask, <4 x float> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: call void @llvm.masked.scatter.v4f32.v4p0f32(<4 x float> %res2, <4 x float*> %gep2, i32 4, <4 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %indsext = sext <4 x i16> %ind16 to <4 x i32>
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %gep3 = getelementptr float, float* %base, <4 x i32> %indsext
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %res3 = call <4 x float> @llvm.masked.gather.v4f32.v4p0f32(<4 x float*> %gep3, i32 4, <4 x i1> %mask, <4 x float> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: call void @llvm.masked.scatter.v4f32.v4p0f32(<4 x float> %res3, <4 x float*> %gep3, i32 4, <4 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %gepu = getelementptr float, float* %base, <4 x i32> %ind32
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 36 for instruction: %resu = call <4 x float> @llvm.masked.gather.v4f32.v4p0f32(<4 x float*> %gepu, i32 1, <4 x i1> %mask, <4 x float> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 36 for instruction: call void @llvm.masked.scatter.v4f32.v4p0f32(<4 x float> %resu, <4 x float*> %gepu, i32 1, <4 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gepos = getelementptr i8, i8* %base8, <4 x i32> %indzext
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %geposb = bitcast <4 x i8*> %gepos to <4 x float*>
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %resos = call <4 x float> @llvm.masked.gather.v4f32.v4p0f32(<4 x float*> %geposb, i32 4, <4 x i1> %mask, <4 x float> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: call void @llvm.masked.scatter.v4f32.v4p0f32(<4 x float> %resos, <4 x float*> %geposb, i32 4, <4 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gepbs = getelementptr i16, i16* %base16, <4 x i32> %indzext
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gepbsb = bitcast <4 x i16*> %gepbs to <4 x float*>
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %resbs = call <4 x float> @llvm.masked.gather.v4f32.v4p0f32(<4 x float*> %gepbsb, i32 4, <4 x i1> %mask, <4 x float> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: call void @llvm.masked.scatter.v4f32.v4p0f32(<4 x float> %resbs, <4 x float*> %gepbsb, i32 4, <4 x i1> %mask)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret void
 ;
   %gep1 = getelementptr float, float* %base, <4 x i32> %ind32
   %res1 = call <4 x float> @llvm.masked.gather.v4f32.v4p0f32(<4 x float*> %gep1, i32 4, <4 x i1> %mask, <4 x float> undef)
;
   ret void
 }
 
 define void @gep_v4i16(i16* %base, <4 x i32> %ind32, <4 x i16> %ind16, <4 x i1> %mask)  {
 ; CHECK-LABEL: 'gep_v4i16'
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gep1 = getelementptr i16, i16* %base, <4 x i32> %ind32
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 36 for instruction: %res1 = call <4 x i16> @llvm.masked.gather.v4i16.v4p0i16(<4 x i16*> %gep1, i32 2, <4 x i1> %mask, <4 x i16> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 36 for instruction: call void @llvm.masked.scatter.v4i16.v4p0i16(<4 x i16> %res1, <4 x i16*> %gep1, i32 2, <4 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %indzext = zext <4 x i16> %ind16 to <4 x i32>
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gep2 = getelementptr i16, i16* %base, <4 x i32> %indzext
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 36 for instruction: %res2 = call <4 x i16> @llvm.masked.gather.v4i16.v4p0i16(<4 x i16*> %gep2, i32 2, <4 x i1> %mask, <4 x i16> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 36 for instruction: call void @llvm.masked.scatter.v4i16.v4p0i16(<4 x i16> %res2, <4 x i16*> %gep2, i32 2, <4 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %indsext = sext <4 x i16> %ind16 to <4 x i32>
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gep3 = getelementptr i16, i16* %base, <4 x i32> %indsext
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 36 for instruction: %res3 = call <4 x i16> @llvm.masked.gather.v4i16.v4p0i16(<4 x i16*> %gep3, i32 2, <4 x i1> %mask, <4 x i16> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 36 for instruction: call void @llvm.masked.scatter.v4i16.v4p0i16(<4 x i16> %res3, <4 x i16*> %gep3, i32 2, <4 x i1> %mask)
+
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction:   %gep5 = getelementptr i16, i16* %base, <4 x i16> %ind16
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction:   %res5 = call <4 x i16> @llvm.masked.gather.v4i16.v4p0i16(<4 x i16*> %gep5, i32 2, <4 x i1> %mask, <4 x i16> undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 2 for instruction:   %res5zext = zext <4 x i16> %res5 to <4 x i32>
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction:   %res5trunc = trunc <4 x i32> %res5zext to <4 x i16>
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction:   call void @llvm.masked.scatter.v4i16.v4p0i16(<4 x i16> %res5trunc, <4 x i16*> %gep5, i32 4, <4 x i1> %mask)
+
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction:   %res6 = call <4 x i16> @llvm.masked.gather.v4i16.v4p0i16(<4 x i16*> %gep5, i32 2, <4 x i1> %mask, <4 x i16> undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 2 for instruction:   %res6sext = sext <4 x i16> %res6 to <4 x i32>
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction:   %res6trunc = trunc <4 x i32> %res6sext to <4 x i16>
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction:   call void @llvm.masked.scatter.v4i16.v4p0i16(<4 x i16> %res6trunc, <4 x i16*> %gep5, i32 4, <4 x i1> %mask)
+;
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret void
 ;
   %gep1 = getelementptr i16, i16* %base, <4 x i32> %ind32
   %res1 = call <4 x i16> @llvm.masked.gather.v4i16.v4p0i16(<4 x i16*> %gep1, i32 2, <4 x i1> %mask, <4 x i16> undef)
   call void @llvm.masked.scatter.v4i16.v4p0i16(<4 x i16> %res1, <4 x i16*> %gep1, i32 2, <4 x i1> %mask)
 
   %indzext = zext <4 x i16> %ind16 to <4 x i32>
   %gep2 = getelementptr i16, i16* %base, <4 x i32> %indzext
   %res2 = call <4 x i16> @llvm.masked.gather.v4i16.v4p0i16(<4 x i16*> %gep2, i32 2, <4 x i1> %mask, <4 x i16> undef)
   call void @llvm.masked.scatter.v4i16.v4p0i16(<4 x i16> %res2, <4 x i16*> %gep2, i32 2, <4 x i1> %mask)
 
   %indsext = sext <4 x i16> %ind16 to <4 x i32>
   %gep3 = getelementptr i16, i16* %base, <4 x i32> %indsext
   %res3 = call <4 x i16> @llvm.masked.gather.v4i16.v4p0i16(<4 x i16*> %gep3, i32 2, <4 x i1> %mask, <4 x i16> undef)
   call void @llvm.masked.scatter.v4i16.v4p0i16(<4 x i16> %res3, <4 x i16*> %gep3, i32 2, <4 x i1> %mask)
+  
+  ; result zext
+  %gep5 = getelementptr i16, i16* %base, <4 x i16> %ind16
+  %res5 = call <4 x i16> @llvm.masked.gather.v4i16.v4p0i16(<4 x i16*> %gep5, i32 2, <4 x i1> %mask, <4 x i16> undef)
+  %res5zext = zext <4 x i16> %res5 to <4 x i32>
+  %res5trunc = trunc <4 x i32> %res5zext to <4 x i16>
+  call void @llvm.masked.scatter.v4i16.v4p0i16(<4 x i16> %res5trunc, <4 x i16*> %gep5, i32 4, <4 x i1> %mask)
+  
+  ; result sext
+  %res6 = call <4 x i16> @llvm.masked.gather.v4i16.v4p0i16(<4 x i16*> %gep5, i32 2, <4 x i1> %mask, <4 x i16> undef)
+  %res6sext = sext <4 x i16> %res6 to <4 x i32>
+  %res6trunc = trunc <4 x i32> %res6sext to <4 x i16>
+  call void @llvm.masked.scatter.v4i16.v4p0i16(<4 x i16> %res6trunc, <4 x i16*> %gep5, i32 4, <4 x i1> %mask)
+  
+  ret void
+}
+
+define void @gep_v4i8(i8* %base, <4 x i8> %ind8, <4 x i1> %mask)  {
+; CHECK-LABEL: 'gep_v4i8'
+
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction:   %gep5 = getelementptr i8, i8* %base, <4 x i8> %ind8
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction:   %res5 = call <4 x i8> @llvm.masked.gather.v4i8.v4p0i8(<4 x i8*> %gep5, i32 2, <4 x i1> %mask, <4 x i8> undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction:   %res5zext = zext <4 x i8> %res5 to <4 x i32>
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction:   %res5trunc = trunc <4 x i32> %res5zext to <4 x i8>
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction:   call void @llvm.masked.scatter.v4i8.v4p0i8(<4 x i8> %res5trunc, <4 x i8*> %gep5, i32 4, <4 x i1> %mask)
+
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction:   %res6 = call <4 x i8> @llvm.masked.gather.v4i8.v4p0i8(<4 x i8*> %gep5, i32 2, <4 x i1> %mask, <4 x i8> undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction:   %res6sext = sext <4 x i8> %res6 to <4 x i32>
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction:   %res6trunc = trunc <4 x i32> %res6sext to <4 x i8>
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction:   call void @llvm.masked.scatter.v4i8.v4p0i8(<4 x i8> %res6trunc, <4 x i8*> %gep5, i32 4, <4 x i1> %mask)
+;
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret void
+;
+  
+  ; result zext
+  %gep5 = getelementptr i8, i8* %base, <4 x i8> %ind8
+  %res5 = call <4 x i8> @llvm.masked.gather.v4i8.v4p0i8(<4 x i8*> %gep5, i32 2, <4 x i1> %mask, <4 x i8> undef)
+  %res5zext = zext <4 x i8> %res5 to <4 x i32>
+  %res5trunc = trunc <4 x i32> %res5zext to <4 x i8>
+  call void @llvm.masked.scatter.v4i8.v4p0i8(<4 x i8> %res5trunc, <4 x i8*> %gep5, i32 4, <4 x i1> %mask)
+  
+  ; result sext
+  %res6 = call <4 x i8> @llvm.masked.gather.v4i8.v4p0i8(<4 x i8*> %gep5, i32 2, <4 x i1> %mask, <4 x i8> undef)
+  %res6sext = sext <4 x i8> %res6 to <4 x i32>
+  %res6trunc = trunc <4 x i32> %res6sext to <4 x i8>
+  call void @llvm.masked.scatter.v4i8.v4p0i8(<4 x i8> %res6trunc, <4 x i8*> %gep5, i32 4, <4 x i1> %mask)
+  
   ret void
 }
 
-define void @gep_v8i16(i16* %base, i8* %base8, i32* %base32, <8 x i32> %ind32, <8 x i16> %ind16, <8 x i1> %mask)  {
+define void @gep_v8i16(i16* %base, i8* %base8, i32* %base32, <8 x i32> %ind32, <8 x i16> %ind16, <8 x i8> %ind8, <8 x i1> %mask)  {
 ; CHECK-LABEL: 'gep_v8i16'
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gep1 = getelementptr i16, i16* %base, <8 x i32> %ind32
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 136 for instruction: %res1 = call <8 x i16> @llvm.masked.gather.v8i16.v8p0i16(<8 x i16*> %gep1, i32 2, <8 x i1> %mask, <8 x i16> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 136 for instruction: call void @llvm.masked.scatter.v8i16.v8p0i16(<8 x i16> %res1, <8 x i16*> %gep1, i32 2, <8 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %indzext = zext <8 x i16> %ind16 to <8 x i32>
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gep2 = getelementptr i16, i16* %base, <8 x i32> %indzext
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %res2 = call <8 x i16> @llvm.masked.gather.v8i16.v8p0i16(<8 x i16*> %gep2, i32 2, <8 x i1> %mask, <8 x i16> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: call void @llvm.masked.scatter.v8i16.v8p0i16(<8 x i16> %res2, <8 x i16*> %gep2, i32 2, <8 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %indsext = sext <8 x i16> %ind16 to <8 x i32>
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gep3 = getelementptr i16, i16* %base, <8 x i32> %indsext
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 136 for instruction: %res3 = call <8 x i16> @llvm.masked.gather.v8i16.v8p0i16(<8 x i16*> %gep3, i32 2, <8 x i1> %mask, <8 x i16> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 136 for instruction: call void @llvm.masked.scatter.v8i16.v8p0i16(<8 x i16> %res3, <8 x i16*> %gep3, i32 2, <8 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 136 for instruction: %resu = call <8 x i16> @llvm.masked.gather.v8i16.v8p0i16(<8 x i16*> %gep2, i32 1, <8 x i1> %mask, <8 x i16> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 136 for instruction: call void @llvm.masked.scatter.v8i16.v8p0i16(<8 x i16> %resu, <8 x i16*> %gep2, i32 1, <8 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gepos = getelementptr i8, i8* %base8, <8 x i32> %indzext
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %geposb = bitcast <8 x i8*> %gepos to <8 x i16*>
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %resos = call <8 x i16> @llvm.masked.gather.v8i16.v8p0i16(<8 x i16*> %geposb, i32 2, <8 x i1> %mask, <8 x i16> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: call void @llvm.masked.scatter.v8i16.v8p0i16(<8 x i16> %resos, <8 x i16*> %geposb, i32 2, <8 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gepbs = getelementptr i32, i32* %base32, <8 x i32> %indzext
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gepbsb = bitcast <8 x i32*> %gepbs to <8 x i16*>
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 136 for instruction: %resbs = call <8 x i16> @llvm.masked.gather.v8i16.v8p0i16(<8 x i16*> %gepbsb, i32 2, <8 x i1> %mask, <8 x i16> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 136 for instruction: call void @llvm.masked.scatter.v8i16.v8p0i16(<8 x i16> %resbs, <8 x i16*> %gepbsb, i32 2, <8 x i1> %mask)
+
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 10 for instruction:   %indzext4 = zext <8 x i16> %ind16 to <8 x i32>
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction:   %gep4 = getelementptr i16, i16* %base, <8 x i32> %indzext4
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 2 for instruction:   %indtrunc = trunc <8 x i32> %ind32 to <8 x i16>
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 8 for instruction:   call void @llvm.masked.scatter.v8i16.v8p0i16(<8 x i16> %indtrunc, <8 x i16*> %gep4, i32 2, <8 x i1> %mask)
+
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 8 for instruction:   %res = call <8 x i16> @llvm.masked.gather.v8i16.v8p0i16(<8 x i16*> %gep4, i32 2, <8 x i1> %mask, <8 x i16> undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 10 for instruction:   %ressext = sext <8 x i16> %res to <8 x i32>
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 2 for instruction:   %restrunc = trunc <8 x i32> %ressext to <8 x i16>
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 8 for instruction:   call void @llvm.masked.scatter.v8i16.v8p0i16(<8 x i16> %restrunc, <8 x i16*> %gep4, i32 4, <8 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret void
 ;
+  ; no offset ext
   %gep1 = getelementptr i16, i16* %base, <8 x i32> %ind32
   %res1 = call <8 x i16> @llvm.masked.gather.v8i16.v8p0i16(<8 x i16*> %gep1, i32 2, <8 x i1> %mask, <8 x i16> undef)
   call void @llvm.masked.scatter.v8i16.v8p0i16(<8 x i16> %res1, <8 x i16*> %gep1, i32 2, <8 x i1> %mask)
 
+  ; offset zext
   %indzext = zext <8 x i16> %ind16 to <8 x i32>
   %gep2 = getelementptr i16, i16* %base, <8 x i32> %indzext
   %res2 = call <8 x i16> @llvm.masked.gather.v8i16.v8p0i16(<8 x i16*> %gep2, i32 2, <8 x i1> %mask, <8 x i16> undef)
   call void @llvm.masked.scatter.v8i16.v8p0i16(<8 x i16> %res2, <8 x i16*> %gep2, i32 2, <8 x i1> %mask)
 
+  ; offset sext
   %indsext = sext <8 x i16> %ind16 to <8 x i32>
   %gep3 = getelementptr i16, i16* %base, <8 x i32> %indsext
   %res3 = call <8 x i16> @llvm.masked.gather.v8i16.v8p0i16(<8 x i16*> %gep3, i32 2, <8 x i1> %mask, <8 x i16> undef)
   call void @llvm.masked.scatter.v8i16.v8p0i16(<8 x i16> %res3, <8 x i16*> %gep3, i32 2, <8 x i1> %mask)
 
   ; unaligned
   %resu = call <8 x i16> @llvm.masked.gather.v8i16.v8p0i16(<8 x i16*> %gep2, i32 1, <8 x i1> %mask, <8 x i16> undef)
   call void @llvm.masked.scatter.v8i16.v8p0i16(<8 x i16> %resu, <8 x i16*> %gep2, i32 1, <8 x i1> %mask)
 
   ; 1 scale
   %gepos = getelementptr i8, i8* %base8, <8 x i32> %indzext
   %geposb = bitcast <8 x i8*> %gepos to <8 x i16*>
   %resos = call <8 x i16> @llvm.masked.gather.v8i16.v8p0i16(<8 x i16*> %geposb, i32 2, <8 x i1> %mask, <8 x i16> undef)
   call void @llvm.masked.scatter.v8i16.v8p0i16(<8 x i16> %resos, <8 x i16*> %geposb, i32 2, <8 x i1> %mask)
 
   ; bad scale
   %gepbs = getelementptr i32, i32* %base32, <8 x i32> %indzext
   %gepbsb = bitcast <8 x i32*> %gepbs to <8 x i16*>
   %resbs = call <8 x i16> @llvm.masked.gather.v8i16.v8p0i16(<8 x i16*> %gepbsb, i32 2, <8 x i1> %mask, <8 x i16> undef)
   call void @llvm.masked.scatter.v8i16.v8p0i16(<8 x i16> %resbs, <8 x i16*> %gepbsb, i32 2, <8 x i1> %mask)
+
+  ; trunc scatter
+  %indzext4 = zext <8 x i16> %ind16 to <8 x i32>
+  %gep4 = getelementptr i16, i16* %base, <8 x i32> %indzext4
+  %indtrunc = trunc <8 x i32> %ind32 to <8 x i16>
+  call void @llvm.masked.scatter.v8i16.v8p0i16(<8 x i16> %indtrunc, <8 x i16*> %gep4, i32 2, <8 x i1> %mask)
+  
+  ; ext result to <8 x i32>
+  %res = call <8 x i16> @llvm.masked.gather.v8i16.v8p0i16(<8 x i16*> %gep4, i32 2, <8 x i1> %mask, <8 x i16> undef)
+  %ressext = sext <8 x i16> %res to <8 x i32>
+  %restrunc = trunc <8 x i32> %ressext to <8 x i16>
+  call void @llvm.masked.scatter.v8i16.v8p0i16(<8 x i16> %restrunc, <8 x i16*> %gep4, i32 4, <8 x i1> %mask)
+  
   ret void
 }
 
 define void @gep_v8f16(half* %base, i8* %base8, i32* %base32, <8 x i32> %ind32, <8 x i16> %ind16, <8 x i1> %mask)  {
 ; CHECK-LABEL: 'gep_v8f16'
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %gep1 = getelementptr half, half* %base, <8 x i32> %ind32
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 136 for instruction: %res1 = call <8 x half> @llvm.masked.gather.v8f16.v8p0f16(<8 x half*> %gep1, i32 2, <8 x i1> %mask, <8 x half> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 136 for instruction: call void @llvm.masked.scatter.v8f16.v8p0f16(<8 x half> %res1, <8 x half*> %gep1, i32 2, <8 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %indzext = zext <8 x i16> %ind16 to <8 x i32>
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %gep2 = getelementptr half, half* %base, <8 x i32> %indzext
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %res2 = call <8 x half> @llvm.masked.gather.v8f16.v8p0f16(<8 x half*> %gep2, i32 2, <8 x i1> %mask, <8 x half> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: call void @llvm.masked.scatter.v8f16.v8p0f16(<8 x half> %res2, <8 x half*> %gep2, i32 2, <8 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %indsext = sext <8 x i16> %ind16 to <8 x i32>
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %gep3 = getelementptr half, half* %base, <8 x i32> %indsext
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 136 for instruction: %res3 = call <8 x half> @llvm.masked.gather.v8f16.v8p0f16(<8 x half*> %gep3, i32 2, <8 x i1> %mask, <8 x half> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 136 for instruction: call void @llvm.masked.scatter.v8f16.v8p0f16(<8 x half> %res3, <8 x half*> %gep3, i32 2, <8 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 136 for instruction: %resu = call <8 x half> @llvm.masked.gather.v8f16.v8p0f16(<8 x half*> %gep2, i32 1, <8 x i1> %mask, <8 x half> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 136 for instruction: call void @llvm.masked.scatter.v8f16.v8p0f16(<8 x half> %resu, <8 x half*> %gep2, i32 1, <8 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gepos = getelementptr i8, i8* %base8, <8 x i32> %indzext
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %geposb = bitcast <8 x i8*> %gepos to <8 x half*>
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %resos = call <8 x half> @llvm.masked.gather.v8f16.v8p0f16(<8 x half*> %geposb, i32 2, <8 x i1> %mask, <8 x half> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: call void @llvm.masked.scatter.v8f16.v8p0f16(<8 x half> %resos, <8 x half*> %geposb, i32 2, <8 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gepbs = getelementptr i32, i32* %base32, <8 x i32> %indzext
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gepbsb = bitcast <8 x i32*> %gepbs to <8 x half*>
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 136 for instruction: %resbs = call <8 x half> @llvm.masked.gather.v8f16.v8p0f16(<8 x half*> %gepbsb, i32 2, <8 x i1> %mask, <8 x half> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 136 for instruction: call void @llvm.masked.scatter.v8f16.v8p0f16(<8 x half> %resbs, <8 x half*> %gepbsb, i32 2, <8 x i1> %mask)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret void
 ;
+
+  ; no offset ext
   %gep1 = getelementptr half, half* %base, <8 x i32> %ind32
   %res1 = call <8 x half> @llvm.masked.gather.v8f16.v8p0f16(<8 x half*> %gep1, i32 2, <8 x i1> %mask, <8 x half> undef)
   call void @llvm.masked.scatter.v8f16.v8p0f16(<8 x half> %res1, <8 x half*> %gep1, i32 2, <8 x i1> %mask)
 
+  ; offset zext
   %indzext = zext <8 x i16> %ind16 to <8 x i32>
   %gep2 = getelementptr half, half* %base, <8 x i32> %indzext
   %res2 = call <8 x half> @llvm.masked.gather.v8f16.v8p0f16(<8 x half*> %gep2, i32 2, <8 x i1> %mask, <8 x half> undef)
   call void @llvm.masked.scatter.v8f16.v8p0f16(<8 x half> %res2, <8 x half*> %gep2, i32 2, <8 x i1> %mask)
 
+  ; offset sext
   %indsext = sext <8 x i16> %ind16 to <8 x i32>
   %gep3 = getelementptr half, half* %base, <8 x i32> %indsext
   %res3 = call <8 x half> @llvm.masked.gather.v8f16.v8p0f16(<8 x half*> %gep3, i32 2, <8 x i1> %mask, <8 x half> undef)
   call void @llvm.masked.scatter.v8f16.v8p0f16(<8 x half> %res3, <8 x half*> %gep3, i32 2, <8 x i1> %mask)
 
   ; unaligned
   %resu = call <8 x half> @llvm.masked.gather.v8f16.v8p0f16(<8 x half*> %gep2, i32 1, <8 x i1> %mask, <8 x half> undef)
   call void @llvm.masked.scatter.v8f16.v8p0f16(<8 x half> %resu, <8 x half*> %gep2, i32 1, <8 x i1> %mask)
 
   ; 1 scale
   %gepos = getelementptr i8, i8* %base8, <8 x i32> %indzext
   %geposb = bitcast <8 x i8*> %gepos to <8 x half*>
   %resos = call <8 x half> @llvm.masked.gather.v8f16.v8p0f16(<8 x half*> %geposb, i32 2, <8 x i1> %mask, <8 x half> undef)
   call void @llvm.masked.scatter.v8f16.v8p0f16(<8 x half> %resos, <8 x half*> %geposb, i32 2, <8 x i1> %mask)
 
   ; bad scale
   %gepbs = getelementptr i32, i32* %base32, <8 x i32> %indzext
   %gepbsb = bitcast <8 x i32*> %gepbs to <8 x half*>
   %resbs = call <8 x half> @llvm.masked.gather.v8f16.v8p0f16(<8 x half*> %gepbsb, i32 2, <8 x i1> %mask, <8 x half> undef)
   call void @llvm.masked.scatter.v8f16.v8p0f16(<8 x half> %resbs, <8 x half*> %gepbsb, i32 2, <8 x i1> %mask)
+
+  ret void
+}
+
+define void @gep_v8i8(i8* %base, <8 x i8> %ind8, <8 x i1> %mask)  {
+; CHECK-LABEL: 'gep_v8i8'
+
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 18 for instruction:   %indzext = zext <8 x i8> %ind8 to <8 x i32>
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction:   %gep5 = getelementptr i8, i8* %base, <8 x i32> %indzext
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 8 for instruction:   %res5 = call <8 x i8> @llvm.masked.gather.v8i8.v8p0i8(<8 x i8*> %gep5, i32 2, <8 x i1> %mask, <8 x i8> undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 2 for instruction:   %res5zext = zext <8 x i8> %res5 to <8 x i16>
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction:   %res5trunc = trunc <8 x i16> %res5zext to <8 x i8>
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 8 for instruction:   call void @llvm.masked.scatter.v8i8.v8p0i8(<8 x i8> %res5trunc, <8 x i8*> %gep5, i32 4, <8 x i1> %mask)
+
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 8 for instruction:   %res6 = call <8 x i8> @llvm.masked.gather.v8i8.v8p0i8(<8 x i8*> %gep5, i32 2, <8 x i1> %mask, <8 x i8> undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 2 for instruction:   %res6sext = sext <8 x i8> %res6 to <8 x i16>
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction:   %res6trunc = trunc <8 x i16> %res6sext to <8 x i8>
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 8 for instruction:   call void @llvm.masked.scatter.v8i8.v8p0i8(<8 x i8> %res6trunc, <8 x i8*> %gep5, i32 4, <8 x i1> %mask)
+
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret void
+;
+
+  ; result zext
+  %indzext = zext <8 x i8> %ind8 to <8 x i32>
+  %gep5 = getelementptr i8, i8* %base, <8 x i32> %indzext
+  %res5 = call <8 x i8> @llvm.masked.gather.v8i8.v8p0i8(<8 x i8*> %gep5, i32 2, <8 x i1> %mask, <8 x i8> undef)
+  %res5zext = zext <8 x i8> %res5 to <8 x i16>
+  %res5trunc = trunc <8 x i16> %res5zext to <8 x i8>
+  call void @llvm.masked.scatter.v8i8.v8p0i8(<8 x i8> %res5trunc, <8 x i8*> %gep5, i32 4, <8 x i1> %mask)
+  
+  ; result sext
+  %res6 = call <8 x i8> @llvm.masked.gather.v8i8.v8p0i8(<8 x i8*> %gep5, i32 2, <8 x i1> %mask, <8 x i8> undef)
+  %res6sext = sext <8 x i8> %res6 to <8 x i16>
+  %res6trunc = trunc <8 x i16> %res6sext to <8 x i8>
+  call void @llvm.masked.scatter.v8i8.v8p0i8(<8 x i8> %res6trunc, <8 x i8*> %gep5, i32 4, <8 x i1> %mask)
+  
   ret void
 }
 
 define void @gep_v16i8(i8* %base, i16* %base16, <16 x i8> %ind8, <16 x i32> %ind32, <16 x i1> %mask)  {
 ; CHECK-LABEL: 'gep_v16i8'
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gep1 = getelementptr i8, i8* %base, <16 x i32> %ind32
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 528 for instruction: %res1 = call <16 x i8> @llvm.masked.gather.v16i8.v16p0i8(<16 x i8*> %gep1, i32 1, <16 x i1> %mask, <16 x i8> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 528 for instruction: call void @llvm.masked.scatter.v16i8.v16p0i8(<16 x i8> %res1, <16 x i8*> %gep1, i32 2, <16 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 74 for instruction: %indzext = zext <16 x i8> %ind8 to <16 x i32>
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gep2 = getelementptr i8, i8* %base, <16 x i32> %indzext
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %res2 = call <16 x i8> @llvm.masked.gather.v16i8.v16p0i8(<16 x i8*> %gep2, i32 2, <16 x i1> %mask, <16 x i8> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: call void @llvm.masked.scatter.v16i8.v16p0i8(<16 x i8> %res2, <16 x i8*> %gep2, i32 2, <16 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 74 for instruction: %indsext = sext <16 x i8> %ind8 to <16 x i32>
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gep3 = getelementptr i8, i8* %base, <16 x i32> %indsext
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 528 for instruction: %res3 = call <16 x i8> @llvm.masked.gather.v16i8.v16p0i8(<16 x i8*> %gep3, i32 2, <16 x i1> %mask, <16 x i8> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 528 for instruction: call void @llvm.masked.scatter.v16i8.v16p0i8(<16 x i8> %res3, <16 x i8*> %gep3, i32 2, <16 x i1> %mask)
+
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gepbs = getelementptr i16, i16* %base16, <16 x i32> %indzext
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gepbsb = bitcast <16 x i16*> %gepbs to <16 x i8*>
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 528 for instruction: %resbs = call <16 x i8> @llvm.masked.gather.v16i8.v16p0i8(<16 x i8*> %gepbsb, i32 2, <16 x i1> %mask, <16 x i8> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 528 for instruction: call void @llvm.masked.scatter.v16i8.v16p0i8(<16 x i8> %resbs, <16 x i8*> %gepbsb, i32 2, <16 x i1> %mask)
+
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 74 for instruction:   %indzext4 = zext <16 x i8> %ind8 to <16 x i32>
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction:   %gep4 = getelementptr i8, i8* %base, <16 x i32> %indzext
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 10 for instruction:   %indtrunc = trunc <16 x i32> %ind32 to <16 x i8>
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 16 for instruction:   call void @llvm.masked.scatter.v16i8.v16p0i8(<16 x i8> %indtrunc, <16 x i8*> %gep4, i32 2, <16 x i1> %mask)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret void
 ;
+  ; no offset ext
   %gep1 = getelementptr i8, i8* %base, <16 x i32> %ind32
   %res1 = call <16 x i8> @llvm.masked.gather.v16i8.v16p0i8(<16 x i8*> %gep1, i32 1, <16 x i1> %mask, <16 x i8> undef)
   call void @llvm.masked.scatter.v16i8.v16p0i8(<16 x i8> %res1, <16 x i8*> %gep1, i32 2, <16 x i1> %mask)
 
+  ; offset zext
   %indzext = zext <16 x i8> %ind8 to <16 x i32>
   %gep2 = getelementptr i8, i8* %base, <16 x i32> %indzext
   %res2 = call <16 x i8> @llvm.masked.gather.v16i8.v16p0i8(<16 x i8*> %gep2, i32 2, <16 x i1> %mask, <16 x i8> undef)
   call void @llvm.masked.scatter.v16i8.v16p0i8(<16 x i8> %res2, <16 x i8*> %gep2, i32 2, <16 x i1> %mask)
 
+  ; offset sext
   %indsext = sext <16 x i8> %ind8 to <16 x i32>
   %gep3 = getelementptr i8, i8* %base, <16 x i32> %indsext
   %res3 = call <16 x i8> @llvm.masked.gather.v16i8.v16p0i8(<16 x i8*> %gep3, i32 2, <16 x i1> %mask, <16 x i8> undef)
   call void @llvm.masked.scatter.v16i8.v16p0i8(<16 x i8> %res3, <16 x i8*> %gep3, i32 2, <16 x i1> %mask)
 
   ; Bad scale
   %gepbs = getelementptr i16, i16* %base16, <16 x i32> %indzext
   %gepbsb = bitcast <16 x i16*> %gepbs to <16 x i8*>
   %resbs = call <16 x i8> @llvm.masked.gather.v16i8.v16p0i8(<16 x i8*> %gepbsb, i32 2, <16 x i1> %mask, <16 x i8> undef)
   call void @llvm.masked.scatter.v16i8.v16p0i8(<16 x i8> %resbs, <16 x i8*> %gepbsb, i32 2, <16 x i1> %mask)
+
+  ; trunc scatter
+  %indzext4 = zext <16 x i8> %ind8 to <16 x i32>
+  %gep4 = getelementptr i8, i8* %base, <16 x i32> %indzext
+  %indtrunc = trunc <16 x i32> %ind32 to <16 x i8>
+  call void @llvm.masked.scatter.v16i8.v16p0i8(<16 x i8> %indtrunc, <16 x i8*> %gep4, i32 2, <16 x i1> %mask)
+  
   ret void
 }
 
 define void @gep_v16i8p(<16 x i8*> %base, i32 %off, <16 x i1> %mask)  {
 ; CHECK-LABEL: 'gep_v16i8p'
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %gepbs = getelementptr i8, <16 x i8*> %base, i32 %off
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 528 for instruction: %resbs = call <16 x i8> @llvm.masked.gather.v16i8.v16p0i8(<16 x i8*> %gepbs, i32 2, <16 x i1> %mask, <16 x i8> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 528 for instruction: call void @llvm.masked.scatter.v16i8.v16p0i8(<16 x i8> %resbs, <16 x i8*> %gepbs, i32 2, <16 x i1> %mask)